The Analysis of Linear Partial Differential Operators

Classics in Mathematics
Lars Hormander The Analysis of Linear Partial

Differential Operators I
Springer
Berlin
Heidelberg
New York
HongKong
London
Milan
Paris
Tokyo
Lars Hormander
The Analysis of Linear Partial

Differential Operators I
Distribution Theory
and Fourier Analysis
Reprint of the 2nd Edition 1990
Springer
Lars Hormander
Department of Mathematics
University of Lund
Box 118
SE-22100 Lund, Sweden
e-mail: lvh@maths.lth.se
Originally published as Vol. 256 in the series:

Grundlehren dermathematischen Wissenschaften in 1983 and 1990,
and thereafter as a Springer Study Edition in 1990
Library of Congress Cataloging.in·Publication Data
HOrmander, Lars.
The analysis of linear partial differential operato", I L. HOrmander.
p. cm. - (Classics in mathematics, ISSN 1431-0821)
Originally published: 2nd ed. Berlin; Now Yark : Springer-Verlag, c1990- .
Includes bibliographical ref"""""", and index...
Contents: 1. Distribution theoIy and Fourier analysis
ISBN-13: 978-3-540-00662-6
1. Differential equations. Partial 2. Partial differential operators. l Title. II. Series.
QA377.H5782003
515',7242-dc21
2003050516
Mathematics Subject Classification (2000): 46F, 46E, 26A, 42A, 35A, 35J, 35L
ISSN 1431-0821
ISBN-13: 978-3-540-00662-6 e-ISBN-13: 978-3-642-61497-2
DOl: 10.1007/978-3-642-61497-2
This work is subject to copyright. All rights are reserved. whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or
parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its
current version, and permission for use must always be obtained from Springer-Verlag. Violations are liable
for prosecution under the German Copyright Law.
Springer-Verlag Berlin Heidelberg New York

a member of Bertels mannSpringer Science+Business Media GmbH
@ Springer-Verlag BerlinHeidelberg 2003
The use of general descriptive names, registered names. trademarks etc. in this publication does not imply, even
in the absence of a specific statement, that such names are exempt from the relevant protective laws and
regulations and therefore free for general use.
Printed on acid-free paper 41/3142ck-543210

Preface to the Second Edition
The main change in this edition is the inclusion of exercises with answers
and hints. This is meant to emphasize that this volume has been written
as a general course in modern analysis on a graduate student level
and not only as the beginning of a specialized course in partial differen-
tial equations. In particular, it could also serve as an introduction to
harmonic analysis. Exercises are given primarily to the sections of gen-
eral interest; there are none to the last two chapters.
Most of the exercises are just routine problems meant to give some
familiarity with standard use of the tools introduced in the text. Others
are extensions of the theory presented there. As a rule rather complete
though brief solutions are then given in the answers and hints.
To a large extent the exercises have been taken over from courses
or examinations given by Anders Melin or myself at the University
of Lund. I am grateful to Anders Melin for letting me use the problems
originating from him and for numerous valuable comments on this
collection.
As in the revised printing of Volume II, a number of minor flaws
have also been corrected in this edition. Many of these have been called
to my attention by the Russian translators of the first edition, and
I wish to thank them for our excellent collaboration.
Lund, October 1989 Lars Hormander

Preface
In 1963 my book entitled "Linear partial differential operators" was

published in the Grundlehren series. Some parts of it have aged well
but others have been made obsolete for quite some time by techniques
using pseudo-differential and Fourier integral operators. The rapid de-
velopment has made it difficult to bring the book up to date. Howev-
er, the new methods seem to have matured enough now to make an
attempt worth while.
The progress in the theory of linear partial differential equations
during the past 30 years owes much to the theory of distributions
created by Laurent Schwartz at the end of the 1940's. It summed up a
great deal of the experience accumulated in the study of partial differ-
ential equations up to that time, and it has provided an ideal frame-
work for later developments. "Linear partial differential operators" be-
gan with a brief summary of distribution theory for this was still un-
familiar to many analysts 20 years ago. The presentation then pro-
ceeded directly to the most general results available on partial differ-
ential operators. Thus the reader was expected to have some prior fa-
miliarity with the classical theory although it was not appealed to ex-
plicitly. Today it may no longer be necessary to include basic distribu-
tion theory but it does not seem reasonable to assume a classical
background in the theory of partial differential equations since mod-
ern treatments are rare. Now the techniques developed in the study
of singularities of solutions of differential equations make it possible
to regard a fair amount of this material as consequences of extensions
of distribution theory. Rather than omitting distribution theory I have
therefore decided to make the first volume of this book a greatly ex-
panded treatment of it. The title has been modified so that it indicates
the general analytical contents of this volume. Special emphasis is put
on Fourier analysis, particularly results related to the stationary phase
method and Fourier analysis of singularities. The theory is illustrated
throughout with examples taken from the theory of partial differential
equations. These scattered examples should give a sufficient know-
ledge of the classical theory to serve as an introduction to the system-
VIII Preface
atic study in the later volumes. Volume I should also be a useful in-
troduction to harmonic analysis. A chapter on hyperfunctions at the
end is intended to give an introduction in the spirit of Schwartz distri-
butions to this subject and to the analytic theory of partial differential
equations. The great progress in this area due primarily to the school
of Sato is beyond the scope of this book, however.
The second and the third volumes will be devoted to the theory of
differential equations with constant and with variable coefficients re-
spectively. Their prefaces will describe their contents in greater detail.
Volume II will appear almost simultaneously with Volume I, and Volume
III will hopefully be published not much more than two years later.
In a work of this kind it is not easy to provide adequate refer-
ences. Many ideas and methods have evolved slowly for centuries, and
it is a task for a historian of mathematics to uncover the development
completely. Also the more recent history provides of course consider-
able difficulties in establishing priorities correctly, and these problems
tend to be emotionally charged. All this makes it tempting to omit ref-
erences altogether. However, rather than doing so I have chosen to
give at the end of each chapter a number of references indicating re-
cent sources for the material presented or closely related topics. Some
references to the earlier literature are also given. I hope this will be
helpful to the reader interested in examining the background of the
results presented, and I also hope to be informed when my references
are found quite inadequate so that they can be improved in a later
edition.
Many colleagues and students have helped to improve this book,
and I should like to thank them all. The discussion of the analytic
wave front sets owes much to remarks by Louis Boutet de Monvel, Pierre
Schapira and Johannes Sjostrand. A large part of the manuscript was
read and commented on by Anders Melin and Ragnar Sigurdsson in
Lund, and Professor Wang Rou-hwai of Jilin University has read a
large part of the proofs. The detailed and constructive criticism given
by the participants in a seminar on the book conducted by Gerd
Grubb at the University of Copenhagen has been a very great help.
Niels JI1lrgen Kokholm took very active part in the seminar and has
also read all the proofs. In doing so he has found a number of mis-
takes and suggested many improvements. His help has been invalu-
able to me.
Finally, I wish to express my gratitude to the Springer Verlag for
encouraging me over a period of years to undertake this project and
for first rate and patient technical help in its execution.
Lund, January 1983 Lars Hormander

Contents
Introduction . . . . . .
Chapter I. Test Functions 5

Summary . . . . . . 5
1.1. A review of Differential Calculus 5
1.2. Existence of Test Functions 14
1.3. Convolution . . . . . . . . . 16
1.4. Cutoff Functions and Partitions of Unity 25
Notes . . . . . . . . . . . . . . . . . 31
Chapter II. Definition and Basic Properties of Distributions 33

Summary . . . . . 33
2.1. Basic Definitions . . . . . . . . . 33
2.2. Localization . . . . . . . . . . . 41
2.3. Distributions with Compact Support 44
Notes . . . . . . . . . . . . . . . . 52
Chapter III. Differentiation and Multiplication by Functions 54

Summary . . . . . . . . . . 54
3.1. Definition and Examples 54
3.2. Homogeneous Distributions 68
3.3. Some Fundamental Solutions 79
3.4. Evaluation of Some Integrals 84
Notes . . . . . . . 86
Chapter IV. Convolution 87

Summary . . . . . 87
4.1. Convolution with a Smooth Function 88
4.2. Convolution of Distributions . . . 100
4.3. The Theorem of Supports 105
4.4. The Role of Fundamental Solutions 109
X Contents
4.5. Basic I! Estimates for Convolutions 116

Notes . . . . . . . . . . . . . . . 124
Chapter V. Distributions in Product Spaces 126

Summary . . . . . . . 126
5.1. Tensor Products 126
5.2. The Kernel Theorem 128
Notes . . . . . . . . . 132
Chapter VI. Composition with Smooth Maps 133

Summary . . . . . . . . . . 133
6.1. Definitions . . . . . . . . 133
6.3. Distributions on a Manifold 142
6.4. The Tangent and Cotangent Bundles 146
Notes . . . . . . . . . . . . . . . 156
Chapter VII. The Fourier Transformation 158

Summary . . . . . . . . . . . . . 158
7.1. The Fourier Transformation in !/' and in !/" 159
7.2. Poisson's Summation Formula and Periodic Distributions 177
7.3. The Fourier-Laplace Transformation in If' 181
7.4. More General Fourier-Laplace Transforms 191
7.5. The Malgrange Preparation Theorem . . 195
7.6. Fourier Transforms of Gaussian Functions 205
7.7. The Method of Stationary Phase 215
7.8. Oscillatory Integrals 236
7.9. H(S)' I! and HOlder Estimates 240
Notes . . . . . . . . . . . . . 248
Chapter VIII. Spectral Analysis of Singularities 251

Summary . . . . . . . . . . . . . . . 251
8.1. The Wave Front Set . . . . . . . . 252
8.2. A Review of Operations with Distributions 261
8.3. The Wave Front Set of Solutions of Partial Differential
Equations. . . . . . . . . . . . . . 271
8.4. The Wave Front Set with Respect to CL . . . . • 280
8.5. Rules of Computation for WFL • . . • . • . • • 296
8.6. WFL for Solutions of Partial Differential Equations 305
8.7. Microhyperbolicity 317
Notes . . . . . . . . . . . . . . . . . . . . . . 322
Contents XI
Chapter IX. Hyperfunctions 325

Summary . . . . . . . 325
9.1. Analytic Functionals 326
9.2. General Hyperfunctions 335
9.3. The Analytic Wave Front Set of a Hyperfunction 338
9.4. The Analytic Cauchy Problem . . . . . . . . 346
9.5. Hyperfunction Solutions of Partial Differential Equations 353
9.6. The Analytic Wave Front Set and the Support 358
Notes . . . . . . . . . . 368
Exercises 371
Answers and Hints to All the Exercises. 394
Bibliography. 419
Index . . . . 437
Index of Notation. 439

Introduction
In differential calculus one encounters immediately the unpleasant fact

that not every function is differentiable. The purpose of distribution
theory is to remedy this flaw; indeed, the space of distributions is essen-
tially the smallest extension of the space of continuous functions where
differentiation is always well defined. Perhaps it is therefore self evident
that it is desirable to make such an extension, but let us anyway discuss
some examples of how awkward it is not to be allowed to differentiate.
Our first example is the Fourier transformation which will be
studied in Chapter VII. If v is an integrable function on the real line
then the Fourier transform Fv is the continuous function defined by
00
(Fv)(~)= S riX~v(x)dx, ~E1R.

-00
It has the important property that
(1) F(Dv)=MFv, F(Mv)= -DFv
whenever both sides are defined; here Dv(x) = -idv/dx and Mv(x)
=xv(x). In the first formula the multiplication operator M is always
well defined so the same ought to be true for D. Incidentally the
second formula (1) then suggests that one should also define F for
functions of polynomial increase.
Next we shall examine some examples from the theory of partial
differential equations which also show the need for a more general
definition of derivatives. Classical solutions of the Laplace equation
(2)
or the wave equation (in two variables)
(3)
are twice continuously differentiable functions satisfying the equations
everywhere. It is easily shown that uniform limits of classical solutions
2 Introduction
of the Laplace equation are classical solutions. On the other hand, the
classical solutions of the wave equation are all functions of the form
(4) v(x, y)= f(x+ y)+g(x- y)
with twice continuously differentiable f and g, and they have as

uniform limits all functions of the form (4) with f and g continuous.
All such functions ought therefore to be recognized as solutions of (3)
so the definition of a classical solution is too restrictive.
Let us now consider the corresponding inhomogeneous equations
(5) 82uj8x 2 + 8 2uj8y2 =F,

(6) 82vj8x 2 -8 2vj8y2 =F
where F is a continuous function vanishing outside a bounded set. If
F is continuously differentiable a solution of (6) is given by
(7) V(x,y)= H -F(~,rJ)d~drJj2.
~-y+lx-~I<O
However, (7) defines a continuously differentiable function v even if F

is just continuous. Clearly we must accept v as a solution of (6) even if
second order derivatives do not exist. Similarly (5) has the classical
solution
(8) u(x, y) =(4n)- 1 HF(~, rJ) log«x - ~)2 +(y- rJ)2) d~ drJ
provided that F is continuously differentiable. Again (8) defines a

continuously differentiable function u even if F is just continuous, and
we should be able to accept u as a solution of (5) also in that case.
The difficulties which are illustrated in their simplest form by the
preceding examples were eliminated already by the concept of weak
solution which preceded distribution theory. The idea is to rewrite the
equation considered in a form where the unknown function u is no
longer differentiated. Consider as an example the equation (6). If u is a
classical solution it follows that
(6)'
for every continuous function cjJ vanishing outside a compact, that is,
closed and bounded, set. Conversely, if (6)' is fulfilled for all such cjJ
which are say twice continuously differentiable then (6) is fulfilled. In
fact, if (6) were not satisfied at a point (xo, Yo) we could take cjJ non-
negative and 0 outside a small neighborhood of (xo, Yo) and conclude
that (6)' is not fulfilled either. For such "test functions" cjJ we can
integrate by parts twice in the left-hand side of (6)' which gives the
equivalent formula
Introduction 3
(6)"
Summing up, if u is twice continuously differentiable then (6) is
equivalent to the validity of (6)" for all test functions </J, that is, twice
continuously differentiable functions </J vanishing outside a compact
set. However, (6)" has a meaning if u is just continuous, and one calls
u a weak solution of (6) when (6)" is valid.! It is easily verified that
the flaws of the classical solutions pointed out above disappear if one
accepts weak continuous solutions.
The function F is uniquely determined by u when (6)" is fulfilled.
However, for an arbitrary continuous function u there may be no
continuous function F such that
(9)
can be written in the form
(10) L(</J) = HF</Jdxdy.
Distribution theory goes beyond the definition of weak solutions by
accepting for study expressions L of the form (9) even when they are
not of the form (10). A distribution is any such expression which
depends linearly on a test function </J (and its derivatives). When it can
be written in the form (10) it is identified with the function F. It turns
out that one can extend the basic operations of calculus to distri-
butions; in particular differentiation is always defined.
Let us also consider some examples of similar expressions occur-
ring in physics. First consider a point mass with weight 1 at a point a
on the real axis. This can be considered as a limiting case of a mass
distribution with uniform density 1/2e in the interval (a-e, a+e) as
e --> O. The corresponding functional is
a+.
L.(</J)= S </J(x)dx/2e.
a-.
When e --> 0 we have L.( </J) --> </J( a), so L( </J) = </J( a) should represent the
unit mass at a. This interpretation is of course standard in measure
theory.
Next we consider a dipole at 0 with moment 1. This may be
defined as the limit of the pointmass 1/f> at f> and -1/f> at 0 as f> --> O.
Thus we must consider the limit of the functional
I Note that differential equations appear naturally in a weak form in the calculus of
variations.
4 Introduction
when !5 ----> 0, which is M (cjJ) = cjJ' (0). This functional is therefore the
appropriate description of the dipole.
It is possible to pursue this development and define distributions
as limits of functionals of the form (10). However, we shall not do so
but rather follow the path suggested by the definition of weak de-
rivatives. This is the original definition of Schwartz and it has the
advantage of avoiding the question which limits define the same
distribution.
The formal definition of distributions is given in Chapter II after
properties of test functions have been discussed at some length in Chap-
ter I. Differentiation of distributions is then studied in Chapter III; it
is shown in Section 4.4 that we have indeed obtained a minimal exten-
sion of the space of continuous functions where differentiation is always
possible. In Chapters IV, V, VI we extend convolution, direct product
and composition from functions to distributions. Chapter VII is devot-
ed to Fourier analysis of functions and distributions. The choice of
material differs a great deal from standard texts since it is dictated
by what is required in the later parts. The method of stationary phase
is given a particularly thorough treatment. In Chapter VIII we discuss
the Fourier analysis of singularities of distributions. This turns out
to be a local problem so it can be discussed also for distributions
on manifolds. The phrase "singularity" above is deliberately vague;
in fact we shall consider singularities both from a C"" and from an
analytic point of view. The results lead to important extensions of the
distribution theory in Chapters III-VI. For instance, one can define
the restriction of a distribution u to a submanifold Y if u has no singular-
ity at a normal to Y. Many applications to regularity and uniqueness
of solutions of differential equations are also given. The analytic theory
is continued in Chapter IX which is devoted to hyperfunctions. These
are defined just as distributions but with real analytic test functions.
The main new difficulty stems from the fact that there are no such
test functions vanishing outside a compact set.
Chapter I. Test Functions
Summary
As indicated in the introduction one must work consistently with
smooth "test functions" in the theory of distributions. In this chapter
we have collected the basic facts that one needs to know about such
functions. As an introduction a brief summary of differential calculus
is given in Section 1.1. It is written with a reader in mind who has
seen the material before but perhaps with different emphasis and
different notation. The reader who finds the presentation hard to
follow is recommended to study first a more thorough modern treat-
ment of differential calculus in several variables, and experienced
readers should proceed directly to Section 1.2. In addition to the basic
indispensible facts we have included in Sections 1.3 and 1.4 some
more refined constructions which will be useful some time in this
book but are not important for the main theme. The reader in a hurry
may therefore wish to omit Section 1.3 from Theorem 1.3.5 on and
also Theorem 1.4.2, Lemma 1.4.3 and the rest of Section 1.4 from
Theorem 1.4.6 on.
1.1. A Review of Differential Calculus

At first we shall consider functions of a single real variable but permit
values in a Banach space. Thus let I be an open interval on the real
line IR and let V be a Banach space with norm denoted I II. A map f:
I --+ V is called differentiable at xEI, with derivative f'(X)E V, if
(1.1.1) 11(f(x+h)- f(x»/h- f'(x)II--+O when h--+O.
We can write (1.1.1) in the equivalent form
(1.1.1)' IIf(x+h)- f(x)- f'(x)hll =o(lhl) when h--+O.
If V=IRn and we write f=(fl, ... ,jn) then differentiability of f is of
course equivalent to differentiability of each component fj. For vector
6 1. Test Functions
valued functions the mean value theorem must be replaced by the

following
Theorem 1.1.1. If f: I --+ V is differentiable at every point in I, then

(1.1.2) Ilf(y) - f(x) II ~Iy-xl sup{II1'(x + t(y -x))II, O~t ~ I}; x,YEI.
Proof Let M>sup{II1'(x+t(y-x))II, O~t~ l} and set
E= {t; O~t~l, Ilf(x+t(y-x))- f(x) II ~Mtlx- yl}.

E is closed since f is continuous, and OEE, so E has a largest element
s. If t > sand t - s is sufficiently small we have
IIf(x+t(y-x))- f(x) II
~ Ilf(x+t(y-x))- f(x+s(y-x))11 + Ilf(x+s(y-x))- f(x) II
~ MI(t-s)(y-x)1 + Msly-xl = Mtly -xl.
Hence s = 1 which proves the theorem.
Remark. If f is just continuous in [x, y] and differentiable in the

interior we obtain (1.1.2) with supremum for 0 < t < 1 as a limit of
(1.1.2) applied to smaller intervals. If VEV an application of (1.1.2) to
x --+ f(x) - xv gives
(1.1.2)' Ilf(y)- f(x)- v(y-x)11 ~ Iy -xl sup 111'(x +t(y-x)) -vII

0<1< 1
which is often more useful, particularly with v = 1'(x).
Corollary 1.1.2. Let f be continuous in I and differentiable outside a

closed subset F where f=O. If XEF and 1'(y)--+O when I . . . . F3y--+x,
then 1'(x) exists and is equal to O.
Proof If YEF then f(y)- f(x) =0. Otherwise let z be the point In
F n [x, y] closest to y. Then (1.1.2)' gives
IIf(y) - f(x) II = Ilf(y) - f(z)11 ~ Iy-zl sup 1I1'(z+ t(y-z))11

0<1< 1
which is o(ly-xl) as y--+x.
Example 1.1.3. If P is a polynomial and f(x)=P(1/x)e-11X, x>O, f(x)

= 0, x ~ 0, then f is continuous. The derivative for x =1= 0 is of the same
form with P(l/x) replaced by (P(l/x) - P'(1/x))/x 2 so 1'(0) exists and is
equal to O.
1.1. A Review of Differential Calculus 7
Let U be another Banach space, X an open subset of U. If f is a

function from X to V we define differentiability by an analogue of
(1.1.1)' :
Definition 1.1.4. f is called differentiable at XEX if there is an element

f'(x)EL(U, V) such that
(1.1.1)" Ilf(x + h) - f(x) - f'(x) hll =o(llhll), h --+ o.
Here L(U, V) is the space of continuous linear transformations from U
to V, which is a Banach space with the norm
IITII= sup IITxll, TEL(U, V).
IIxll < 1
By C 1 (X, V) we denote the set of continuously differentiable functions

from X to V, that is, the set of functions f which are differentiable at
every point and for which X 3X --+ f'(x)EL(U, V) is continuous.
If f is just differentiable at every point on the line segment [x, y]

={x+t(y-x); 0~t~1} then (1.1.2)' gives for every TEL(U, V)
(1.1.2)" IIf(y)- f(x)- T(y-x)11 ~ Ily-xll sup 11f'(x+t(y-x»- TIl
0<1< 1
for f(x+t(y-x»-Tt(y-x) is differentiable in t on [0,1] with de-

rivative
f'(x + t(y - x»(y - x) - T(y - x).
Theorem 1.1.5. If hE
C 1 (X, V) and h --+ f, h' --+ g locally uniformly in X,
then fEC 1 (X, V) and f'=g.
Proof If we apply (1.1.2)" to h with T= f;(x) we obtain when j --+ 00

IIf(y)- f(x)-g(x)(y-x)11 ~ Ily-xll sup IIg(x+ t(y-x»-g(x)11
0<1<1
which proves that f is differentiable at x with f'(x)=g(x). Since g is

continuous this proves the theorem.
Theorem 1.1.6. If f and g are continuous functions in X with values in

V and L(U, V) respectively, and t--+ f(x+ty) is for all x, YEU differen-
tiable with respect to t with derivative g(x+ty)y when X+tYEX, then
fECI and f' = g. It suffices to make the hypothesis for all y in a set Y
c U with closed linear hull equal to U.
Proof (1.1.2)' gives for small Ilyll

I f(x + y) - f(x) - g(x) yll ~ Ilyll sup IIg(x + t y) - g(x) II
0<1< 1
8 I. Test Functions
which proves the first statement. To prove the second one it suffices
to show that the set of all y for which the hypothesis holds is linear
and closed. This follows from (1.1.2)'; the details are left for the
reader.
If f is a linear map V ---+ V then f is of course differentiable and

f'(x) = f for every x. More generally, let VI'"'' Vk be Banach spaces
and L(V1 , ... , Vk; V) the space of multilinear maps
VI x ... X Vk3(X 1 , ... ,Xk)---+ f(x 1 , ... ,Xk)EV
which are continuous, that is,
Ilfll= sup Ilf(x 1 , .. ·,xk)ll<oo.
IIXjl1 <1
With this norm L(V1 , ... , Vk; V) is a Banach space. The map
VI EB .. ·EB Vk3(X I , .. · ,Xk)---+ f(x 1 , ... ,Xk)EV
is differentiable for every (x I ' ... , x k ), and the differential is
VI EB ... EBVk 3(YI' .. ·,h)---+ f(Yl' x 2 , ... ,xk )
+ f(x 1 , Y2'" .,Xk )+ .. · + f(x 1 , X2 ,· .. ,Xk_l' Yk)'
Another important example of a C I map is for two Banach spaces
V and V the map f taking an invertible element TEL(U, V) to its
inverse T-IEL(V, V). That T- I is an inverse means that
TT- 1=id v , T- 1T=id u·
If SEL(V,V) we have (T+S)T-I=idv+ST- I so if IISIIIIT- 1 11<1 a
right inverse of T + S is given by
00
T-1(id v+ST- 1 )-1 = LT- 1( -ST-1t

o
In the same way we see that it is a left inverse. Thus f(T) = T- 1 is
defined in an open set and
IIf(T + S) - f(T) + T- 1ST-111 ~ I T- 11 3 1ISI1 2 /(l-IISIIII T- 1 11)
so f is differentiable at T and 1'(T) S = - T- 1ST- I. Thus f is con-
tinuous so l' is continuous.
Next we shall discuss composite functions. Let X as before be an
open subset of the Banach space V, f a map from X to a Banach
space V with range contained in an open set Y where another map g
to a third Banach space W is defined. If f is differentiable at x and g
is differentiable at Y = f(x) then h = g 0 f is differentiable at x and
(1.1.3) h'(x) = g'(y) 1'(x) (the chain rule).
1.1. A Review of Differential Calculus 9
The proof is obvious. From (1.1.3) it follows that hE C l if gE C l and

fECI.
The differential f' may be viewed as a map
X x U3(X, t ) L (f(x),f'(X)t)EY x V
which is linear in the second component which should be thought of
as a tangent direction. Then the chain rule says that given a com-
mutative diagram
with 1, gEC I we obtain hECI and another commutative diagram
Instead of the notation f' one often writes d1, particularly when f
is real valued. If f is defined in an open set in 1R.n and we write t
= L tjej where ej is the j'h unit vector, then
(df)(t) = (dj)(L tjej)= L tjdf(e j)= L 8f18x j t j .
But tj=(dx)(t) so we can write this equation in the form
df = L8f18xjdx j .
By the chain rule this formula remains valid if Xj are in fact functions
of ZEZ and both f and Xj are regarded as functions on Z so that both
sides are linear functions there. This is called the invariance of the
differential.
Next we prove the inverse function theorem which shows that
differential calculus does accomplish its goal of reducing the study of
fairly general equations to linear ones.
Theorem 1.1.7. Let X be open in U and fECI(X, V), and let XoEX,
f(xo} = Yo' For the existence of g E C I (Y, U), where Y is a neighborhood
of Yo, such that a) fog=identity near Yo or b) gof=identity near Xo or
c) f o g = identity near Yo and go f = identity near x o, it is necessary and
sufficient that there is a linear map AEL(V, U) such that respectively
a)' f'(xo)A=id y b)' Af'(xo)=id u c)' f'(xo)A=id y , Af'(xo)=id u .
The condition c)' is equivalent to bijectivity of f'(x o} and it implies that
g is uniquely determined near Yo' If V(U) is finite dimensional then
a)'{b)') is equivalent to surjectivity (injectivity) of f'(x o).
10 I. Test Functions
Proof The necessity is an immediate consequence of the chain rule

(1.1.3). In the proof of the sufficiency we first observe that if f ° gl
=id near Yo and g2of=id near Xo then gl =g 2ofo g 1 =g2 near Yo
which proves uniqueness in case c) and reduces the proof to existence
in cases a) and b). If we replace f by f 0 A resp. A 0 f we see that it
suffices to study the case where U = V and f'(xo)=id. Choose b>O so
that
11f'(x)-idll <t when Ilx-xoll ~b.
For Ilxj-xoll ~b,j=1,2, we have by (1.1.2)"
(1.1.4) Ilf(x 1) - f(x 2)-(X 1-x 2)11 ~ Ilxl -x211/2.

Hence f is injective in {x; Ilx-xoll ~b}. To solve the equation f(x)
= y when Ily - Yo II < bl2 we set
(1.1.5) Xk=X k _ 1 + y- f(X k_ I ), k=1,2, ...
as long as this leads to points with Ilxk - Xo II < b. We have
Ilxl-xoll = lIy-Yoll <bI2.

If k>l and Ilxj-xoll <b whenj<k then
Ilxk -Xk_lll = Ilx k_ 1 - f(X k- 1 ) -(Xk_ 2 - f(x k_ 2»11

~ Ilx k_ 1 - Xk_ 211/2 < bl2k
by (1.1.4). Hence
k
Ilxk -xoll <b L 2- j <b
1
so x k is defined for all k and is a Cauchy sequence. For the limit x we

have Ilx - Xo II < b, and when k -+ 00 we obtain f(x) = Y from (1.1.5).
To prove that the inverse g(y)=x, which is now defined when
IIY-Yoll <b/2, belongs to C 1 we set
g(y)=x, g(y+k)=x+h.
This means that f(x+h)=y+k and that f(x)=y. Hence

k= f(x +h) - f(x) = f'(x) h +o(llhll).
In view of (1.1.4) we have IIk-hll < Ilh1211, hence IIh11/2< Ilkll <2I1hll.
Since 11f'(x)-lll <2 we obtain
h= f'(X)-l k + o(llkll)
which proves that g is differentiable and that g'(y) = f'(g(y»- \ which
is a continuous function of y.
1.1. A Review of Differential Calculus II
We shall now define differentials of higher order and the space

Ck(X, V) where as before X is an open subset of the Banach space V
but k is now an integer >1. This is done by induction so fECk(X, V)
if fEC 1 (X, V) and f'EC k- 1 (X, L(V, V)). The differential f" of f' is
called the second differential of f, and so on,
f(k)EC(X,L(V,L(U, ... ,L(U, V))).
The vector space L(V, L(U, ... , L(V, V))) is isomorphic as a Banach
space to the space L( V, ... , V; V) = V( V, V) of k-linear maps from V to
V. In fact, we always have
L(U, L(Vl' ... , Vj; V))=L(U, VI' ... , Vj; V)
for if f is an element in the space on the left and the assertion is
known when j is replaced by j -1, then
V x VI X ... X V j 3(X, Xl'"'' Xj) -+ f(x)(x 1 , ... ,Xj)EV
is an element in the space on the right, and all its elements can be so
obtained. The correspondence is obviously linear and norm preserv-
mg.
By I!'.(V, V) we shall denote the space of symmetric k linear forms
from V to V, that is, forms such that the value at (x I ' ... , x k ) is not
changed by a permutation of Xl' ... ,Xk .
Theorem 1.1.S. If fECk then Pk) is a symmetric multilinear form, that

is, the order of differentiation is not essential.
Proof Set LlyF(x)=F(x+y)-F(x). By repeated use of (1.1.2)" we ob-

tain if L is a multilinear form
IILI Yk" .LlyJ(x) - L(Yl'"'' Yk)11
~ sup IILl Lly1f'(x + t Yk) Yk - L(YI"'" Yk)11
yk _ 1 ' ..
O<t< I
If we choose L = f(k)(X) it follows that

(1.1.6)
This determines Pk ) completely of course, and since Llyk ... LlyJ(x) does
not depend on the order of the differences it follows that Pk)(X) is
symmetric.
From (1.1.3) it follows at once by induction that h=gofECk if f,

gEC\ for if this is proved for smaller values of k we have g' 0 fEC k - 1
and j' E Ck - l, and the composition of linear maps

L(V, W) x L(lJ, V) --+ L(U, W)
is continuous, hence infinitely differentiable. The inverse function
theorem (Theorem 1.1.7) also remains valid with C l replaced by C k
throughout. In fact, the map
L(U, V)3T--+ T-lEL(V, U)
is in Ck where it is defined, with kth differential
(Sl' ... , Sk) --+ (-I(L T- l Sil T- l Si2 T- l ... Sik T- l
summed over all permutations of 1, ... , k. In the proof of Theorem
1.1.7 we therefore obtain inductively that gECk if fECk, by using that
g'(y) = j'(g(y))-l.
If f ECk in a neighborhood of the line segment [x, x + y] then we
have Taylor's formula
k-l
(1.1.7) f(x+y)= LfW(x;y, ... ,y)/j!
o
1
+ Jf(k)(X+ ty; y, ... , y)(I- t)k-l dt/(k-I)!.
o
This follows inductively by partial integration since
d
dJ(k-ll(X+t y ; y, ... , y)= Pk)(x+ty; y, ... , y).
In particular we obtain
k
(1.1.8) Ilf(x+y)-LfW(x;y, ... ,y)/j!ll=o(llyllk) as y--+O.
o
Let us now assume that f is defined in an open set X eRn. It
follows from Theorem 1.1.6 that fECk if and only if all the partial
derivatives
o of
OXil ... oX ij
of order j~k are defined and continuous. By Theorem 1.1.8 we know

that the order of differentiation does not matter. With the notation OJ
=%xj we can therefore write these partial derivatives in the form
O~I ... o:nf = oaf
where a=(a l , ... , an) is a multi-index, that is, an n-tuple of non-nega-
tive integers, and oa=o~I ... o:n. We set lal=LlXj , which is the order of
differentiation. Note that
1.1. A Review of Differential calculus 13
j<k)(X; y, ... ,y)= L k! aaf(x)ya/ex.!

lal=k
where ex.! = ex. 1! ... ex."! and ya= y~' ... y~n. Hence Taylor's formula can be
written
(1.1.7)' f(x+ y)= L aaf(x)ya/ex.!
lal <k
1
+kJ (1-tt- 1 L aaf(x Hy)ya/ex.! dt.
o lal=k
The following application is often useful:
Theorem 1.1.9. If fECk(B) where B={XE1R.",lxl<R} and k~l, then

f(x)- f(O) = L xi,h(x) where ,hECk-1(B), aa ,h(O)=aa aJ(O)/(l + Iex.D, and
(1.1.9) suplaa ,h1~suplaaaJI, 1ex.I<k.

B B
Proof Using (1.1.7)' with k = 1 we can take

1
,h(x) = S (aJ)(tx)dt,
o
and jj E C k - 1 since the integrand is in Ck - 1. The estimate (1.1.9) is
obvious.
Any linear differential operator

P= L aa(x)aa
lal~m
can be written in the form P(x, a) where

P(x, e)= L aa(x)ea
lal~m
is a polynomial in eE1R.". By convention coefficients are put to the left

of differentiations. Note that
P(x, e)=e-<x.~) Pe<x.~).
Leibniz' formula for the differentiation of a product takes the form
(1.1.10) P(x, o)(uv)= L (p(a) (x, a) u)aav/ex.!
where p(a)(x, e) = a~p(x, e). Indeed, u -+ P(x, a)(uv) is a differential op-
erator for fixed v, and
e
e- <x.~) P(x, a)(e<x·~)v)=P(x, +a)v= LP(a)(x, e) aav/ex.!
by Taylor's formula. We shall refer to (1.1.10) as Leibniz' formula.
1.2. Existence of Test Functions

For an open set X in JR" we shall denote by Ck(X) the space of k
times continuously differentiable complex valued functions in X if k is
a non-negative integer, and we set
with intersection taken for all finite k. In the introduction we have

seen the need for functions of the following kind:
Definition 1.2.1. By C~(X) we denote the space of all UECk(X) with

compact support. The elements of Cg'(X) are called test functions.
Definition 1.2.2. If UE C(X) then the support of u, written supp u, is the

closure in X of the set {XEX; u(x)=!=O}, that is, suppu is the smallest
closed subset of X such that U ~ 0 in X, supp u.
If the definition of a function UEC~(X) is extended to JR" by

letting u=O in JR", X we obtain of course a function in C~(JRn). Thus
we may regard C~(X) as a subspace of q(JR") which increases with
X. For an arbitrary subset M cJRn we may also define C~(M) as the
set of elements in C~(1R") with support contained in M. When k=O
we sometimes omit k in the notation.
Lemma 1.2.3. There exists a non-negative function ¢ E Cg' (JRn) with

¢(O»O.
Proof Let f(t)=exp( -lit), t>O and f(t) =0, t~O. From example
1.1.3 we know that fEC1(JR) and repeating the argument gives
fECOO(JR). Hence
¢(x)=f(1-lxI 2 ),
"
IxI 2 =LX;,
1
has the required properties.
By translation and change of scales we obtain the non-negative

Cg' function
(1.2.1)
which is positive at Xo and has support in the ball of radius () with
center at xo. We can now prove a fact already alluded to in the
introduction.
1.2. Existence of Test Functions 15
Theorem 1.2.4. If f, gEC(X) and

(1.2.2) Jf¢dx= Jg¢dx, ¢EC't;'(X)
then f =g.
Proof If h= f-g we have

(1.2.3) Jh¢dx=O, ¢EC't;'(X).
Taking real and imaginary parts we find that h may be assumed real
valued provided that ¢ is taken real valued. If h(xoH=O then we can
take ¢EC'O(X) non-negative with ¢(xoHO and support so close to
Xo that ¢h has a constant sign which contradicts (1.2.3). Hence h=O
identically as claimed.
A more general but less elementary result of the same kind is
Theorem 1.2.5. If f, g are locally integrable functions in X and (1.2.2) is

valid, then f = g almost everywhere in X.
Proof Again it suffices to show that if h satisfies (1.2.3) then h = 0

almost everywhere. To do so we use Lebesgue's theorem stating that
lim t- n J Ih(x)-h(y)ldy=O
!~o Ix-YI<!
for almost every x. With ¢EC't;' having support in the unit ball and
J¢dx=l, we can write for XEX and small t
h(x)= Jh(x)¢«x - y)/t) dy/t"
= J(h(x) - h(y» ¢«x - y)/t) dy/t n+ Jh(y) ¢«x - y)/t) dy/tn.
The last integral vanishes by hypothesis and the preceding one tends
to 0 with t for almost all x, which proves that h(x)=O almost every-
where.
Remark. The theorem is also an immediate consequence of Theorem

1.3.2.
Lemma 1.2.3 is all one needs to get distribution theory started.

However, a number of more subtle constructions of test functions will
be needed to develop the theory, and we shall discuss them in the
following two sections. As an indication of how rich the space C't;'(K)
is we prove already now a classical theorem of Borel which will
sometimes be useful.
Theorem 1.2.6. For j=O, 1, ... let fjEC'[;(K) where K is a compact subset
of Rn, and let I be a compact neighborhood of °
in R. Then one can
find fEC'[;(K x I) such that
oif(x,t)/oti=fj(x), t=O, j=0,1, ....
Proof Choose gEC~(l) so that di (g(t)-1)/dti=0 when t=O for j

=0,1, .... For example, if (-e, e)cI we can take ¢ according to
Lemma 1.2.3 with support in (0, e) and J¢(t)dt= 1, and let g be the
solution ofg'(t)=¢(-t)-¢(t) with support in I. Now
gi(x, t)=g(t/ei)tifj(x)/j!
is in C~(K x I), and
(1.2.4) lo~gi(x,t)I~2-i if lod~j-1,
provided that ej is sufficiently small. In fact taking t/ej as new variable
we see that
lo~gj(x, t)1 ~ C~,jej-~t
where ()(t is the order of differentiation with respect to t. For small ej

we obtain (1.2.4) since ()(t <j. Hence the modified Taylor series
co
f(x, t)= L: gj(x, t)
o
is uniformly convergent. So are the series obtained by differentiation.
In view of Theorem 1.1.5 it follows that f E C~ (K x I) and that
iJif(x,t)/otj=L:ojgk(x,t)/otj=fj(x) when t=O.
1.3. Convolution
If u and v are in C(Rn) and either one has compact support, then the
convolution u* v is the continuous function defined by
(1.3.1) (u*v)(x)=Ju(x-y)v(y)dy, xERn.
Thus u*v is a superposition of translates of u taken with the weights
v(y) dy, so we can expect u* v to inherit properties of u such as
differentiability. On the other hand, taking x - y as a new integration
variable in (1.3.1) we obtain
(1.3.2)
1.3. Convolution 17
so the properties of v should also be inherited. The reason for the

commutativity (1.3.2) is perhaps more clear if we note that (1.3.1)
implies
(1.3.1)' J(U* v)cj>dx = JJU(x) v(y) cj>(x + y)dxdy, cj>E cg(1Rn),
whi.ch conversely implies (1.3.1) by Theorem 1.2.4. Now (1.3.1)' shows
that (1.3.2) just expresses the commutativity of addition in ]R.n. Simi-
larly the associativity leads to
(1.3.3)
if all except one of the continuous functions u, v, w have compact
support. The direct verification of (1.3.3) from (1.3.1) is of course an
easy exercise. Taking w = 1 we find that
(1.3.4) J(u*v)dx= Judx Jvdx
when U and v have compact support.
If UEC l and VECo, either one having compact support, we can
differentiate under the integral sign in (1.3.1) and obtain that u* VE C l
and
(1.3.5)
By the commutativity (1.3.2) we could differentiate on the factor v
instead if VEC 1 • If UECj and VECk it follows therefore that U*VECj+k
and that
(1.3.6)
The preceding conclusions can be strengthened in various ways.
For example it is clear that u*v is a continuous function if UEC o and
v is just locally integrable. If UE Cb we conclude for such v that
u* VE cj. Summing up, we have found
Theorem 1.3.1. If uECb then U*VECj if vEL~oc' and U*VECHk if

VECk.
The convolution can be used to approximate functions by more

differentiable ones, a technique known as regularization:
Theorem 1.3.2. Let O~cj>EC'~, Jcj>dx=1. If uECb(]R.n) it follows that

U",=U*cj>EC~(]R.n). When suppcj> ---+ {O} we have
(1.3.7) sup lo"u - o"u",1 ---+ 0 if lod ~j;

if vEll(]R.n) then V",EC""(]R.n) and v", ---+ v in II norm if p< 00.
Proof In view of (1.3.6) and Theorem 1.3.1 we just have to prove

(1.3.7) when IX = O. If Iyl < b in supp ¢ we obtain
lu(x)-u",(x)1 = IS (u(x)-u(x - y))¢(y)dyl ~ sup lu(x) -u(x - y)1
lyl<6
and this tends to 0 with b since u is uniformly continuous. Since

I v",11 LP ~ IlvilLP and cg is dense in IJ' the last statement follows at once.
It is sometimes useful to know more precisely how the regular-
izations converge, and we give a result of that kind:
Theorem 1.3.3. Let VECi(IR") and ¢ECg<'(IR"), S¢dx=1. Then

(1.3.8) u(x, t)= Sv(x-ty)¢(y)dy
is in Coo in {(x, t)EIRn+ 1, t* O} and ti'u(x, t) is in Ck+ i(IR"+ 1) for every

non-negative integer k. When t = 0 we have
(1.3.9) 8;(ti'u(x, t))=O if i<k; 8~(tkU(X, t))=k!v(x).
Proof Since u(x, t) -+ v(x) when t -+ 0, by Theorem 1.3.2, we obtain

(1.3.9) from Taylor's formula as soon as we know that tku(x, t)ECk+i.
This we shall prove for every ¢ECg<'. That UE Ci follows at once if we
differentiate under the integral sign in (1.3.8). When k*O we may
therefore assume the statement proved already for smaller values of k.
For t=l=O we have
ti'u(x, t)= t k Itl-" Sv(y)¢«x - y)/t)dy.
If we differentiate and change variables back again it follows that

Sv(x - ty)8;¢(y)dy,
8;(ti'u(x, t)) = ti'-l
8t (tk U(X, t))=ti'-l Sv(x-ty)«k-n)¢(y)-l>A¢(y))dy.
In view of Corollary 1.1.2 this is also true when t = 0 so the first order
derivatives of ti'u(x, t) are in C k - 1 + i which proves the theorem.
Corollary 1.3.4. Given arbitrary UiE Ck - i (IR"), 0 ~j ~ k, one can find

UE Ck(IRn+ 1) so that
(1.3.10) 8{u(x, t)=u/x) when t=O, j=O, "', k.
Proof For r=O, ... , k we shall prove inductively that one can find u
=ur satisfying (1.3.10) when O~j~r. We can take UO(x, t)=uo(x). If
U,-l has already been chosen we must find ur = U,-l + U so that
8!U=0, j<r, 8~U=v,=u,-8~ur-l when t=O.

1.3. Convolution 19
Since VrECk-r we obtain a function U with these properties from

Theorem 1.3.3.
Regularization of a function does not increase the support very
much, for we have
(1.3.11) supp u*vcsupp u+supp V= {x + Y; XESUPP U, yESUpp v}.
This is an immediate consequence of (1.3.1) or (1.3.1)'.
When applying Theorem 1.3.2 we must of course appeal to the
existence of test functions proved in Lemma 1.2.3. However, test
functions on JR can also be constructed by convolutions starting from
the simplest step functions, and we shall discuss this now since it
leads to important quantitative information.
Set
Ha(x)=a- 1 when O<x<a and Ha(x)=O otherwise.
If u is a continuous function then
x
u*Ha(x)=a- 1 Ju(x-t)dt=a- J u(t)dt
1
o x-a
is in C I and the derivative is (u(x)-u(x-a))ja, so U*HaECk+ 1 if

UEC k.
Theorem 1.3.5. Let a o ~ a l ~ ... be a positive sequence with
and set
uk=Hao*···*Hak·
Then UkE C~ - I (JR) has support in [0, a] and converges as k -+ 00 to a
function UE C~ (JR) with support in [0, a] such that Judx = 1 and
(1.3.12) IU(k)(x)1 ~t Jlu(k+ l)(x)1 dx ~ 2kj(ao ... ak), k =0,1, ....
Proof UI vanishes except in [0, ao + al]' increases with slope 1ja oa l in

[0, al]' is constant in [ai' ao] and decreases linearly to in (a o, ao
+a l ), so u l is continuous. Hence UkEC~-1 by the remarks preceding
°
the proof, and the support is in [O,ao+ ... +ak] by (1.3.11). With the
notation
(rau)(x)=u(x-a)
(which later on will be recognized as a convolution) we have
UkUl_ n
j-I 1
- o -(1-r
ai a )(HaJ*···*Hak ),
I
if j~k-1.
Fig. I. Uk for k;£6 with ao =l, aj = 1.5/(j(j+ 1))
Since SH a dx = 1 it follows from (1.3.4) that

(1.3.12)' Slu~)1 dx ~ 2j j(a o ... aj_ 1)'
lu~>i:~:;2j j(a o ... a),
for a convolution u*v can always be estimated by sup lui S Ivldx. Now
IU k+m-uml = lum*H am + * ... *H am + k -urnl
1
~(am+1 + ... +am+k)suplu~1

~2ao 1aj1(arn +1 + ... +am +k )
by the proof of Theorem 1.3.2, so Um has a uniform limit u. So has the
derivative
and so on. By Theorem 1.1.5 it follows that UE C", and (1.3.12)' gives
(1.3.12). Since Sukdx= 1 for every k by (1.3.4), we have Sudx= 1.
Only in the last step was it important that I aj was assumed

convergent. However, if it diverges then the limit u must be identically
O. This will follow from our next lemma which shows how precise the
construction in Theorem 1.3.5 really is.
Lemma 1.3.6. If UEC m (( - ro, TJ) vanishes on the negative half axis, and
if aj are positive decreasing numbers with T~a1 + ... +am , then
(1.3.13) lu(x)I~I 2 2j supa 1 ... ajluU)(y)1 if x~T.
jEJ Y< x
Here J = {j; 1 ~j~m and aj+1 <a j or j=m}.
Proof The formula

1.3. Convolution 21
20
17
15
Fig. 2. Z and the moves allowed when m=20, J = {I, 3, 7,15,17, 20}
used in the proof of Theorem 1.3.5, gives for j ~ m -1 that

(1.3.14)
Starting from the case j=O this allows us to write u as a sum of terms
of the form
(1.3.15) U>
ta kt ". ta k i Ha 1 * ... *H aJ.*a 1 ". aJ.u; i~m, j~m,
and akv ~av' v ~ i.
(The last condition guarantees that tak ••• t ak . is a translation by at
least a 1 + ... + ai .) We shall say that (1'.3.15) 'is a legitimate term of
type (i,}) and represent it by the point (i,j)ElR 2. In any term with
(i,}}EZ,
Z = {(i,}); O~i<m, O~j <m, aj + 1 ~ai+ tl
we apply (1.3.14) and obtain two new terms. Since aj + 1 ~ ai + 1 the one
which comes from the second term in (1.3.14) is legitimate of type
(i+l,j), and (i+l,j)EZ unless i+l=m, for ai+2~ai+1' The other term
is legitimate of type (i,j + 1), and we have (i,j + I)EZ unless j + 1 = m,
thus j + 1 EJ, or else a j + 2 < ai + 1; since a i + 1 ~ aj + 1 this implies i + 1 <j
+2 and aj + 2 <aj + 1 , hencej+1EJ. After a finite number of steps we
conclude that u is a sum of terms of the form (1.3.15) with either i~m
or jEJ and i + 1 ~j. (In Fig. 2 (i,}) is a free arrowhead.) The number of
terms of type (i,j) is at most 2i + j since each term corresponds to an

increasing path from (0,0) to (i,}) among lattice points, so there are i
+j steps and at most two alternatives for each of them. Since
I2i + j <2 2j
i<j
and the terms with i=m vanish in (0, T), we have proved (1.3.13).
If we drop the assumption T~al + ... +am in Lemma 1.3.6 and set
c=T/(a1+ ... +am ), we can apply Lemma 1.3.6 to u(cx) which is
defined for x ~ a l + ... + am. This gives
(1.3.13)' lu(x)1 ~ I (4c)j sup a 1 ... aj luU)(y)1 if x ~ T.
jEJ Y<X
If
(1.3.16) luU)(x)I~Kj/(al ... a) whenjEJ and x~T
then it follows that

(1.3.17)
In fact,
""
I(4cK)i=4cK/(1-4cK)<8cK if 8cK<1.
1
If we apply (1.3.17) to the integral of the function U in Theorem 1.3.5

taking K = 2, we also conclude that u could not have been given
support in an interval shorter than I a/16, for 8c/(1- 8c) < 1 if
c< 1/16. Apart from a fixed scale factor the construction in Theorem
1.3.5 is therefore quite precise. If I aj = 00 and (1.3.16) is valid for all j
we may also conclude that u == O. In fact, we shall deduce the complete
Denjoy-Carleman theorem.
Let M 0 = 1, M l' M 2' ... be a sequence of positive numbers, and
denote by CM([a, b]) the set of all uEC""([a, b]) such that for some K
(1.3.18)
Definition 1.3.7. eM is called quasi-analytic if uU)(x)=O for every j at

some xE[a,b] implies u=o in [a,b] when uECM([a,b]).
Let
(1.3.19) L.= inf MIlk
J k?;'j k
be the largest increasing minorant of Mf1k and let Mj be the largest

logarithmically convex minorant of the sequence M j ,
1.3. Convolution 23
(1.3.20) Mj* =inf {Mj' M1, - jl /(I-k l Mf-kl/(l-k l when k<j <l}.
Thus Mj2~Mj_IMj+l,j>0, which shows that Mj>O for allj unless
Mj* = 0 for every j > O. Moreover,
(1.3.21)
is a decreasing sequence. (We define aj = + 00 if M; =0.)
Theorem 1.3.8 (Denjoy-Carleman). The following conditions are equiva-
lent:
(i) CM is quasi-analytic.
00
(ii) L I/L j = 00.

o
L (Mj)- Iii = 00.

00
(iii)
o
00
(iv) L a = 00.
j
1
Proof First we consider the rather uninteresting case where

Lj~L< 00 for every j, that is, limMWi~L for some sequence kr-+ 00.
Letting I run through this sequence in (1.3.20) and taking k = 0 there
we obtain
IiMo?;,Mj*=Mo/a l ... aj?;,Mo/(a l ... ai_l)a:- j - l , i~j,
and this implies ai-I ~L. Thus the conditions (ii), (iii), (iv) are fulfilled.
If u satisfies (1.3.18) and uU)(xo)=O for every j, then Taylor's formula
gives for every xE[a, b]
Ju(x)J ~Kki+ 1 JX-xoJkiMk/k j ! -+ 0 when j -+00.
Thus (i) is also valid.

Assume now that L j -+ 00 when j -+ 00, that is, that M:/ k -+ 00 as
k -+ 00. Then the points (k, log M k ) will lie above lines with arbitrarily
high slope so Mj is positive and aj -+ O. Thus J = U; aj+ I < aj} is an
infinite set where the graph of log M: as a function of k has a corner.
This implies that Mj = Mj when jEi; analytically this follows since for
k<j<1
M1,-jl/(I-kl MP-kl/(I-kl?;,M:(l-il/(I-klMTu-kl/(I-kl
> M:"a\'-j)U-kl/(I-kla~(I-jlU-k)/(l-kl > M:" (a.la. )t >M:"
= J J J+ 1 = J JI' J+ 1 J.
Since
M:=M~/a1 ... ak~akk
it follows that M]'j ~ l/a j when jEJ. If k <jEi and k ~ i <j implies ifj:J
then ai=aj for k~i~j so it follows that Li~Lj~l/aj=l/ai' hence
Lj~l/aj for every j. This proves that (iv)=>(ii). Since 10gMj is a

convex function of j and vanishes when j = 0 the slope j- 1 log Mj* of
the chord from (0, 0) is increasing, that is, (Mnl/i is increasing. Hence
Lj~(Mj)l/i which proves that (ii) => (iii). If (iv) is false we know from
Theorem 1.3.5 that there is a function u=l=O in Cg'([a,b]) with
luU)(x)1 ~Kj+ l/a l ... aj=Kj+ 1 Mj* ~Kj+l Mj
so (i) is not valid. Thus (i) => (iv). On the other hand, if (iv) is valid we
can apply (1.3.17) with m equal to an element of J. The set J in
(1.3.16), defined in Lemma 1.3.6, is then the set of elements ~m in J, so
(1.3.16) follows from (1.3.18) since M j =Mj=l/a l ... aj when jEJ.
Hence (iv) => (i). The remaining proof that (iii) => (iv) follows from
Lemma 1.3.9 (Carleman's inequality). If aj>O then

00 00
(1.3.22) ~)ala2'" an)l/n~eLan'

1 1
Proof With cj to be chosen later we have by the inequality between

geometric and arithmetic means
n
(a l .. , an)l/n =(c l ... cn)-l/n(c l a l '" Cnan) l/n ~(Cl ... cn)-l/nn- l L cmam •
1
00 00
L(al .. ·an)l/n~ L cmam/n(n+l)=Lcmam/m~Leam

l~m~n 1 1
which proves (1.3.22).

When Mn=n! then Taylor's formula gives for JECM([a,b]) that
00
J(x) = L JU)(y)(x - y)j/j!

o
if x,YE[a,b] and Ix-yl<r. Thus Theorem 1.3.8 is completely elemen-
tary in this case. Note that the Taylor series is then absolutely
convergent for all XE<C with Ix - yl < r and it gives an analytic con-
tinuation of J to
{ZE<C; Iz- yl <r for some YE[a, b]}.
Conversely, if F is an analytic function in a complex neighborhood of
[a, b] then the restriction of F to [a, b] is in CM([a, b]) by Cauchy's
inequalities. Accordingly this class is called the class of real analytic
Junctions. It is obvious that the preceding remarks can be applied also
to functions of several variables.
1.4. Cutoff Functions and Partitions of Unity

In distribution theory one often has to replace a function by one with
compact support without changing it on a large compact set. This is
done by multiplication with a "cutoff function" as constructed in the
following
Theorem 1.4.1. If X is an open set in JR." and K is a compact subset,

then one can find CPE C'(;'(X) with 0 ~ cP ~ 1 so that cP = 1 in a neigh-
borhood of K.
Proof Choose e > 0 so small that

(1.4.1) Ix-YI~4e when xEK, YECX,
and let v be the characteristic function of
K 2 ,={Y; Ix-YI~2e for some xEK}.
According to Lemma 1.2.3 we can find a non-negative function
XEC'(;'(B) where B is the unit ball, such that JXdx = 1. Then X,(x)
=c"x(x/e) has support in the ball {x; Ixl<e} and JX,dx=l, so
cp=v*X,EC'(;'(K 3 ,)
by Theorem 1.3.1 and (1.3.11), and l-cP=(I-v)*X, vanishes in K, by
(1.3.11). This proves the theorem.
For future reference we also note that

lo'cpl ~ JIc'x,1 dx = e-I,I JIC'xl dx.
Thus
(1.4.2)
where C, only depends on ex, n and the choice of the norm. Using
Theorem 1.3.5 it is possible to give a still more precise result which
we mention for the sake of completeness since it is sometimes useful.
However, the reader can jump to Theorem 1.4.4 without loss of
continuity.
Theorem 1.4.2. Let X be an open set in the n dimensional vector space
V with norm I I and let K be a compact subset. If
d=inf{IIx-yII; XECX,YEK}
L dj < d, then one can find

00
and dj is a positive decreasing sequence with

1
CPEC'(;'(X) with O~cp~l, equal to 1 in a neighborhood of K, so that
(1.4.3) 1</J(k)(X; Y1' ... , Yk)1 ~ Ck IIY111 ... IIYklild1 ... dk ; k= 1,2, ....
Here C depends only on the dimension n.
Proof. Assume first that V =1R.n and that Ilxll = max Ixjl. Let u be the
function in Theorem 1.3.5 with a j =d j+ 1 and set h(t)=u(t+ Ld)2).
Then we have Itl~Lldjl/2 if tEsupph, and for every j
flhUl(t)ldt~2jld1 ... dj' fh(t)dt=1.
We can now apply the proof of Theorem 1.4.1 with e= 1 and
x(x) = h(x 1) ... h(x n),
taking for v the characteristic function of
{y; Ilx-yll ~d12 for some xEK}.
It follows that
laa</JI ~flaaxldx ~2Ialld1 ... dlal
so introducing the differentials instead we have
Wk)(x; Y1'···' Yk)1 ~(2n)k Ily 1 1 .. . IiYk Illd 1· .. dk,

which proves Theorem 1.4.2 in this case. The passage to a Euclidean
norm and from there to an arbitrary norm can be made by the
following lemma which gives (1.4.3) with C = 2n 2 •
Lemma 1.4.3. If K is a convex symmetric body in the n dimensional

vector space V, then one can find an ellipsoid B with center at and °
BcKcByn.
Proof. Let B be an ellipsoid of maximal volume contained in K and

choose coordinates so that B is the unit ball. We have to show that K
cByn. To do so we assume that K contains a point (t,O, ... ,O) with
t > yn and prove that B cannot be maximal then. The tangent cone to
B with vertex at (t, 0, ... ,0) touches B where Xl = lIt, so the part of B
where Ix 11> lIt is in the interior of K because of the convexity and
symmetry with respect to 0. Now consider the ellipsoid
(1-e)n-1xi+y2/(1-e)~I, y=(x~+ ... +x;yt,
which has the same volume as B. The inequality may be written

xi + y2 -((n-l)xi - y2) e+ 0(e 2) < 1,
that is,
(xi + y2 -1)(1 + e) < e(nxi -1) + 0(e 2).
Fig. 3
For small 8>0 the right-hand side is negative when IX11«1/t

+ 1/Vn)/2 so this part of the ellipsoid is in the interior of K. But this
is also true of the remaining part since Ix 1 1;;;(1/t+!/Vn)/2>1/t. (See
Fig. 3.) Hence B is not maximal.
An equivalent way of stating Lemma 1.4.3 is of course that for any
norm I I in V one can find a Euclidean norm I I such that
Ixl~lIxll~lxlvn, XEV.
In order to make conclusions about a distribution from local

hypotheses, it is necessary to write arbitrary test functions as sums of
test functions of small support.
Theorem 1.4.4. ~et X1,.;,;"Xk b~ open sets in JRn and let 4>EC(f(V Xj).
Then one can find 4> j EC O (Xj)' } = 1, ... , k, such that .
(1.4.4)
If 4>;;; 0 one can take all 4> j ;;; o.
Proof We can choose compact sets K I' ... , K k with K j C X j so that
k
supp 4> C UK j• (In fact, every point in supp 4> has a compact neigh-
1
borhood contained in some Xj. By the Borel-Lebesgue lemma a finite
number of such neighborhoods can be chosen which cover all of
supp 4>. The union of those which belong to X j is a compact set K j
cXj") Using Theorem 1.4.1 we now choose I/IjEC(f(X) with O~I/Ij~ 1
and 1/1 j = 1 in K j. Then the functions
4>1 =4>1/1 1 ,4>2=4>1/12(1-1/11)' ···,4>k=4>l/Ik(l-l/Il)···(l-l/Ik-l)
have the required properties since
n(1-1/1)=0
k k
"'i4>j-4>= -4>
1 1
because either 4> or some 1 -1/1j is zero at any point.

28 L Test Functions
Combining Theorems 1.4.4 and 1.4.1 we obtain
Theorem 1.4.5. Let Xl' ... ' X k be open sets in 1Rnand K a compact
k
subset of UXj. Then one can find <PjEC'(;(X) so that <Pj?;'O and
k 1
L <Pj::;; 1 with equality in a neighborhood of K.
1
Remark. In Theorems 1.4.4 and 1.4.5 we could allow an infinite

number of open sets Xj while in the conclusion only finitely many <Pj
are not identically O. In fact, by the Borel-Lebesgue lemma a finite
number of the sets Xj suffice to cover supp <P and K respectively.
The functions <P j in Theorem 1.4.5 are said to be a partition of
unity at K subordinate to the covering of K by the sets Xj. Infinite
partitions of unity are sometimes required, and for the sake of com-
pleteness we shall give one simple and one fairly intricate example.
Theorem 1.4.6. For every neighborhood X of the cube
- XE 1Rn·1
K -{ , Xj 1<1 ·-1 , ... ,n}
=2,j-
one can find a non-negative function <PEC'(;(X) such that

L<P(x-g)=l
where g runs over all lattice points, that is, points with integer coor-
dinates.
Proof By Theorem 1.4.1 we can choose i/lEC'(;(X) so that 0::;;t/t::;;1
and t/t = 1 on K. Hence
L
If'(X) = t/t(x - g)
is a periodic Coo function with If'(x)?;, 1 everywhere, so <p(x)
=t/t(x)/If'(x) has the required properties.
Theorem 1.4.6 gives a partition of unity subordinate to the cover-
ing of 1Rn by the translates X + {g} of X. We shall now construct
partitions of unity corresponding to a covering by convex symmetric
neighborhoods which may vary both in size and in shape. It is
convenient to describe the neighborhoods in terms of the correspond-
ing norms.
Definition 1.4.7. If X is an open set in a finite dimensional vector
space V and I IIx for every XEX is a norm in V, then we shall say that
we have a slowly varying metric in X if
(1.4.5) XEX, Ily-xllx< 1 implies YEX and Ilvlly::;; CIIvll x, VEY,
where C?;, 1 is independent of x, y, v.
Note that if CIIx - yllx < 1 then Ilx - ylly < 1 so

Ilvllx~ CIIvlly, VEv'
Replacing the norm I Ilx by C II Ilx we have therefore

(1.4.5)'
which we shall assume from now on.
Example 1.4.8. Let X be an open set in V and d(x) a Lipschitz con-

tinuous function, positive in X and zero in V'-. X, with
Id(x)-d(y)l~ Ilx- yll if X, YEX,
where II II is a fixed norm in V. Then
IIvll x =2I1vll/d(x)
defines a slowly varying metric in X, for Ilx - yllx < 1 means that
Ilx-YIl <d(x)/2 which implies Id(x)-d(Y)I<d(x)/2 and d(x)/2
< d(y) < 2d(x). For any given open set X =t= V we may always
take for d(x) the distance to F = CX defined by
d(x) = infllx - YII.
YEF
We shall now show how a slowly varying metric gives rise first to
a covering and then to a subordinate partition of unity.
Lemma 1.4.9. Let e<l/C and choose a maximal sequence X l 'X 2 ' ... in
X such that
(1.4.6) Ilxl'-x.k~e when v=t=,u.
Then the balls
B~={x; IIx-x.llxv <R},
where eC<R<l, are a covering of X, and no point belongs to more
than N=(2C 2 /e+l)n different B~.
Proof The existence of a maximal sequence follows from the fact that
a set satisfying (1.4.6) is necessarily discrete. If XEX cannot be added
to the sequence then either Ilx-x.llxv <e or else Ilx-x.llx<e for some
v, and in the latter case it follows that Ilx-x.llxv < Ce so xEB~ in any
case. If xEB~ we obtain since R < 1
Ilx -x.1Ix < C
and if x E B! also then
e~ Ilxl' -x.k~ C Ilxl' -x.lIx·
Hence the balls

{y; Ily-x.llx<e/2C}
are disjoint when xEB~, and they are all contained in the ball
{y; Ilx-Yllx<C+e/2C}
so the number of such v cannot exceed the ratio (2 C 2/e + I)" of the
measures.
Choose e = 1/2 C and for every v a non-negative function
t/I.EC~(B!) which is 1 in B: and for a given decreasing sequence dj
L
with dj = 1 has the estimate
1t/I~k)(X; YI' ... , Yk)1 ~ C~ IIYIllxv·· .llh k/dl·· .dk·
This is possible by Theorem 1.4.2 with C I equal to three times the
constant in (1.4.3). Since IIYllxv~CIIYllx for every xEsuppt/l. we obtain
1t/I~k)(X; YI'···' Yk)1 ~(CC I)k IlyIllx·· .llykllxldl· ..dk·
As in the proof of Theorem 1.4.4 we now introduce
<P.=t/I.(1-t/lI)···(I-t/l.-I)
and obtain L <P. = 1 in X. No point is in the support of more than N
factors t/I". Now note that if for some norm we have at x
1.fj(k)(X, YI' ... ' Yk)1 ~A~ Ilylll ... IIYkll/dl·· .dk,
for k = 0,1, ... and j = 1, 2, then the same estimate is valid for fJ2 with
the constant Al +A2 instead. This follows from the fact that dj is
decreasing and from the rules for differentiating a product. Hence
1<p~k)(X; YI' ... , Yk)1 ~ (N CC I)k IIYIll x·· .llh Ilxldl· . .dk.
We have now proved
Theorem 1.4.10. For any slowly varying metric in the open set X in the
n dimensional vector space V one can choose a sequence x. E X such that
the balls
B.={x; Ilx-x.IIxv <1}
form a covering of X for which the intersection of more than N =(4C 3
+ 1)" balls B. is always empty. Moreover, for any decreasing sequence
dj with Ld j =l one can choose non-negative <p.EC~(B.) with L<P.=l
in X so that for all k
(1.4.7)
where C is the constant in (1.4.5) and C d3 is that in (1.4.3) so it only
depends on n.
Notes 31
Remark. Without changing the proof one may allow dj to be functions

of x provided that they vary slowly with respect to the metric, that is,
(1.4.5)'
The following corollary is sometimes a useful supplement to Theo-

rem 1.4.1.
Corollary 1.4.11. Let Fo and FI be two closed sets in JR.". Then one can
find 4>EC OO (C(FonFI )) such that 4>=0 near Fo,-(FonFI ), 4>=1 near
FI '-(FonFI ) and, with the notation in Theorem 1.4.10,
(1.4.8)
Here
d(x)=max(d(x, Fo), d(x, FI )), d(x, Fj)=min Ilx- yl!.
YEFj
The support of 4> is bounded if FI is compact.
Proof d(x) is Lipschitz continuous with Lipschitz constant 1, so Ilvllx

=21Ivll/d(x) is a slowly varying metric in C(FonFI ) (Example 1.4.8). A
ball
(1.4.9) {y; Ilx-Yllx~l}
cannot meet both Fo and FI , for if d(x) =d(x, Fj) then Ilx - yll
~ d(x, Fj)/2, hence d(y, Fj) ~ d(x, Fj)/2 in the ball. If we form the
partition of unity in Theorem 1.4.10 we can therefore take 4> equal to
the sum of all 4>. with support meeting Fl. If FI is compact then
d(x, Fo) < 2d(x, FI ) when Ixl is large enough. Thus d(x) < 2d(x, F1 ) so
(1.4.9) cannot meet Fl then. Hence supp 4> will be bounded.
Notes
As pointed out in the summary the introductory Section 1.1 contains
only classical material and we shall not discuss its history. The test
functions used in Section 1.2 also have a long tradition in the calculus
of variations. Thus Theorem 1.2.4 is very close to the de Bois Rey-
mond lemma found in all introductory texts on this topic. Theorem
1.2.6 is due to Borel [1]. More refined extension theorems are given
later as Corollary 1.3.4 and Theorem 2.3.6.
The construction of infinitely differentiable functions by means of
repeated convolutions used in the proof of Theorem 1.3.5 has ancient
roots in harmonic analysis. It was used explicitly and systematically

by Mandelbrojt [1,2] who attributes the idea to unpublished work by
H.E. Bray. The method reappeared in the work by Ehrenpreis [3] on
convolution operators and has been used frequently in the theory of
partial differential equations since then (cf. Hormander [21, 27, 30]).
Cohen [2] observed that the proof by Bang [1] of the Denjoy-
Carleman theorem can be phrased in similar terms. We use a variant
of his approach combined with arguments from Mandelbrojt [2] in
our proof. Quasi-analyticity will play a role in Section 8.4. Perhaps it
is appropriate to mention here that the problem of characterizing
quasi-analytic classes originates from the theory of partial differential
equations (see Hadamard [1, p. 37]).
Continuous partitions of unity on compact sets were defined by
Dieudonne [1], and the term Dieudonne decomposition was actually
current for a while. The construction in the proof of Theorem 1.4.4 is
taken from the third edition of the classical monograph on Riemann
surfaces by Weyl [2] where the main change was actually the use of
partitions of unity. However, the covering lemma 1.4.9 and the par-
tition of unity in Theorem 1.4.10 are essentially from Whitney [1]
which shows that very sophisticated partitions of unity were used
several years before Dieudonne's note. Whitney's lemma occurs in the
original form as Lemma 2.3.7 below. A more general version was stated
by Treves [5]. The full generality given in Section 1.4 is needed in
the general theory of pseudo-differential operators (Chapter XVIII).
The combination of these ideas with the repeated convolutions of
Section 1.3 is probably new in its full generality but special cases can
be found for example in Andersson [1].
Chapter II. Definition and Basic Properties
of Distributions
Summary
In the introduction we have seen how various difficulties in the theory
of partial differential equations and in Fourier analysis lead one to
extend the space of continuous functions to the space of distributions.
In Section 2.1 we make the definition explicit and precise, using the
properties of test functions proved in Chapter I. The weak topology in
the space of distributions is also introduced there. The notion of
support is extended to distributions in_ Section 2.2 and it is shown
there that distributions may be defined locally provided that the local
definitions are compatible. In addition it is proved that if u is a
distribution then there is a unique way to define u(cp) for all CPEC oo
with supp un supp cp compact. The problem of estimating u(cp) in
terms of the derivatives of cP on the support of u only is discussed at
some length in Section 2.3. The deepest result is Whitney's extension
theorem (Theorem 2.3.6). We shall rarely need the results which follow
from it so the reader might prefer to skip the section from Theorem
2.3.6 on.
2.1. Basic Definitions

In the introduction we were led to consider expressions such as
(2.1.1)
where faE C(X) and the sum is finite. Here X is an open set in ]R".
However, the same form L may have many different representations
of this kind, so in the following definition we just keep an obvious
property of the expression (2.1.1).
Definition 2.1.1. A distribution u in X is a linear form on Cg'(X) such

that for every compact set K c X there exist constants C and k such
that
34 II. Definition and Basic Properties of Distributions
(2.1.2) lu(</J)I~ C L sup 10'" </JI , </JEC't'(K).

l"'l~k
The set of all distributions in X is denoted by 2}/(X). If the same

integer k can be used in (2.1.2) for every K we say that u is of order
~ k, and we denote the set of such distributions by 2}/k(X). Their
union 2}~(X) = U2}'k(X) is the space of distributions of finite order.
That u is a linear form on C't'(X) means of course that u is a

function from C't'(X) to <C such that
u(a</J + bl/l) = au(</J) + bu(I/I); a,bE<C; </J,I/IEC't'(X).

Remark. The reason for the traditional notation 2}/(X) is that Laurent
Schwartz used the notation 2}(X) instead of C't'(X).
The sum (2.1.1) defines a distribution of order k if fa = 0 when
loci> k, and it defines a distribution whenever the sum is locally finite,
that is, on every compact set there are only a finite number of
functions fa which do not vanish identically. We shall see later on
that all distributions are in fact of the form (2.1.1).
Example 2.1.2. If XoEX then u(</J)=o"'</J(x o) defines a distribution of

order loci. That the order is not smaller follows if we choose I/IEC't'
with 1/1(0)=1 and set </Jo(x)=(x-xo)"'I/I((x-xo)/b), for U(</Jil)=OC! and
sup IiJP </J 01 ~ C bl",I-lfll -+ 0 when b -+ 0 if 1131 < loci·
More generally, if XjEX is a sequence of points with no limit point in

X, and if ocj are multi-indices, then
is a distribution in X because a compact subset can only contain

finitely many Xj. By the first part of the example we have uE2}~(X) if
and only if locjl is bounded; the order is then maxlocJ
The continuity condition in Definition 2.1.1 guarantees that u
behaves well when acting on functions depending on parameters:
Theorem 2.1.3. If </J(x, Y)ECoo(X x Y) where Y is an open set in lRm, and

if there is a compact set K eX such that </J(x, y)=O when x¢K, then
Y -+ u(</J(., y»
is a Coo function of y if UE2}'(X), and
o~u(</J(., y»=u(o~</J(., y».

2.1. Basic Definitions 35
Proof For fixed YEY we have by Taylor's formula

cp(x, y + h) = cp(x, y) + L h/}cp(x, y)/oYj + t/I(x, y, h),
supI0~t/I(x,y,h)I=O(lhI2) as h~O, \::let..
x
Hence
u(cp(., y+h))=u(cp(., y))+ L hju(ocp(., y)/oy} + O(IW)
which shows that y --+ u(cp(., y)) is differentiable and that
o
-0 u(cp(., y)) = u(ocp(., y)/oYj).
Yj
Iteration of this result proves the theorem.
The continuity condition (2.1.2) is often stated as a sequential
continuity:
Theorem 2.1.4. A linear form u on Cg"(X) is a distribution if and only

if u(cpj)~O when j~ CIJ for every sequence CPjECg"(X) converging to 0
in the sense that suploacp)~O for every fixed et. and suppcpjcK for all
j and some fixed compact set K c X.
Proof It is obvious that (2.1.2) implies the condition in the theorem.

On the other hand, suppose that there is a compact set K c X such
that (2.1.2) is not valid for any C and k. Taking C = k = j we conclude
that for some cP jE Cg"(K) we have
lu(cp}l>j L suploacpjl.
lal~j
Since this condition is not changed if CPj is multiplied by a constant

factor, it is no restriction to assume that u(cpj) = 1. Then we have
10acpjl~1/j when j~Iet.1 so cpr.... O although u(cpj) does not converge to
O. Thus the condition in the theorem is not fulfilled either, which
proves the assertion.
A third equivalent form of the continuity condition is the follow-
ing one:
Theorem 2.1.5. A linear form u on Cg"(X) is a distribution if and only

if there exist functions PaE C(X) such that
(2.1.3)
a
and on each compact set in X all but a finite number of the functions
Pa vanish identically. One can take Pa=O when 1et.I>k if and only if u is
of order ~k.
Proof The sufficiency of (2.1.3) is obvious. On the other hand, let UE

~'(X). Take an increasing sequence of compact sets KjcX such that
every compact subset of X belongs to some K j , and choose XjEC~(X)
with Xj=1 in K j (Theorem 1.4.1). Writing !/Ij=Xj-Xj-l ifj>1 and !/II
=Xl we have
00
ef>= L!/I/P if ef>EC~(X);

1
the sum is of course finite. Since the support of !/Ijef> is contained in

that of !/I j' Definition 2.1.1 gives with suitable Cj and increasing k j
00 00
lu(ef»I~Llu(!/Ijef»I~L Cj sup supIO"("'jef»I.

1 1 l"l~kj
If we note that L 2 - j = 1 and differentiate the product, we obtain

lu(ef»I~sup2jCj2kj sup suplofJ!/Ijo"ef>I.
j l<>+fJl~kj
The functions
p,,(x)=sup sup 2j+kjCjloll!/l)
j 1<>+ III ~kj
are continuous in X, for ofJ!/Ij=O on K; when j>i so only a finite

number of j and f3 must be considered. In addition p" = 0 on K; if
IIXI > k;. Since
2j+kjCjloll!/ljO"ef>I~lp"o"ef>1 if IIX+f3I~kj'
we obtain lu(ef»I~supsuplp"o"ef>I, which implies (2.1.3).

"
We shall now derive a dual form of (2.1.3) which is quite close to
(2.1.1). Let C(X) be the space of continuous functions in X tending to
o at the boundary, that is, which are arbitrarily small outside compact
subsets. This is a Banach space with the maximum norm, and the dual
space is the space of bounded measures in X. Let B be the space of all
arrays U={u,,} with U"EC(X) for every IX and the norm
II UII = Lsuplu"l.
If L is a linear form on B with norm ~ 1, then
L(U)= LS u"d/-l"
where dJl" is a measure in X with total mass at most 1. Now (2.1.3)
means that the map
is a linear form with norm at most 1 defined on a linear subspace of

B. By the Hahn-Banach theorem it can be extended to a linear form L
of norm at most 1 on all of B, which means that there are measures
Ila with JIdll,,1 ~ 1 such that
u(¢) = L({p,,8"¢}) = I Jp,,8"¢dll,,·
(2.1.1 )'
where the supports of the measures dv" are locally finite.
As in the introduction we identify the space of continuous func-
tions in X with a subspace of .@/(X) by assigning to each continuous
function f the distribution
(2.1.4) C'O(X)3¢--+ Jf¢dx
which we also denote by f. This is legitimate since Theorem 1.2.4
shows that two functions defining the same distribution are identical.
More generally we can make this identification when fEL~oc(X), the
space of functions which are integrable on compact subsets of X
modulo those which vanish almost everywhere. In fact, Theorem 1.2.5
shows that functions defining the same distribution are in the same
equivalence class. We can also identify arbitrary measures with distri-
butions of order 0, for we have
Theorem 2.1.6. If UE.@lk(X) we can in a unique way extend u to a linear

form on ct(X) such that (2.1.2) remains valid for all ¢ECt(K) and
some constant C.
Proof. It follows from Theorem 1.3.2 that for every ¢ECt(X) we can
find a sequence ¢vEC'O(X) with support in a fixed compact neigh-
borhood K of supp ¢, so that
(2.1.5) I supI8"(¢-¢v)I--+O, v--+ 00.
l"l~k
Thus we must define u(¢) =lim U(¢v). This limit exists, for (2.1.5)
implies in view of (2.1.2) that when v, 11--+ 00 we have
IU(¢v)-u(¢/l)I=lu(¢v-¢/l)I~C I supI8"(¢v-¢)I--+O.
1,,1 ~k
That the limit is independent of the sequence chosen follows at once
by mixing two sequences. If we apply (2.1.2) to ¢v and let v--+ 00 we
conclude that (2.1.2) is valid for all ¢ECt with support in the interior
of K, so the theorem is proved.
Since a measure can be defined to be a linear form on C8(X) with

the continuity property (2.1.2) for k=O, we have now identified ~'O(X)
with the space of measures in X. If an integrable function f is first
identified with the measure f dx, as is customary in integration theory,
and f dx is then identified with a distribution, the result will of course
be the same as if we identify f with a distribution directly.
A positive distribution is always a measure:
Theorem 2.1.7. If u is a distribution in X with u(¢)~O for all non-

negative ¢EC~(X), then u is a positive measure.
Proof We have to show that u is of order O. To do so we note that for

any compact set K eX Theorem 1.4.1 gives a function XEC~(X) with
o~ X~ 1 and X= 1 on K. Then
xsupl¢I±¢~O
if ¢EC'~(K) is real valued. By hypothesis it follows that

u(x)supl¢1 ±u(¢)~O
or equivalently that
(2.1.6) lu(¢)1 ~u(x)supl<pl, ¢EC~(K).
e
If we apply this to Re ei6 ¢ when is real and choose so that ei8 u( ¢) e
is real, we obtain (2.1.6) for complex valued ¢ also, hence UE~'o .
.@'(X) is obviously a vector space with the natural definition of

addition and multiplication by complex numbers,
(a l u l + a 2 u 2 )(¢) = a l u l (¢) + a 2 u 2 (¢);
¢EC~(X), UjE~'(X), ajECC.
We shall always use the weak topology in ~'(X) (also called the weak*
topology), that is, the topology defined by the semi-norms
~'(X)3U -+ lu(¢)I,
where ¢ is any fixed element of C~(X). Thus ui -+ u means that
ui (¢) -+ u(¢)
for every ¢ E C~(X). Occasionally we 'shall need the following com-
pleteness property:
Theorem 2.1.8. If u j is a sequence in ~'(X) and

(2.1.7) u(¢)= limuj(¢)
j--+oo
exists for every cf>EC';(X), then UE.@'(X). Thus Uj--+U in .@'(X) as

j--+ 00. Moreover, (2.1.2) is valid for all Uj with constants C and k
independent of j, and uj(cf>j)--+u(cf» if cf>j--+cf> in C';(X).
Proof When K is a compact subset of X the space C';(K) is a Fn~chet

space with the topology defined by the semi-norms
11cf>lIa=suplaacf>l, cf>EC';(K).
(The completeness is a consequence of Theorem 1.1.5.) (2.1.2) is valid
for u j (with constants C and k which maya priori depend on j), so uj
restricted to C';(K) is a continuous linear form on C';(K). For fixed
cf>EC';(K) it follows from (2.1.7) that the sequence uj(cf» is bounded.
Hence the principle of uniform boundedness (the Banach-Steinhaus
theorem) shows that (2.1.2) is valid for all uj with constants C and k
independent of j. When j --+00 we obtain (2.1.2) for the limit u. If
cf>j --+ cf> in C';(X) we have supp cf>j C K for some compact subset K of
X and all j. Hence uj(cf>j-cf» --+ 0 by the uniformity of (2.1.2), which
proves that uj(cf>j) --+ u(cf».
By Cauchy's convergence principle for <C the existence of the limit

(2.1.7) means precisely that uicf»-uk(cf»--+O whenj, k--+oo. Hence an
equivalent statement of Theorem 2.1.8 is that every sequence uj in
.@'(X) such that uj-uk --+ 0 when j, k --+00 must have a limit U in
.@'(X).
Theorem 2.1.9. If UjE.@'(X), Uj~O and Uj--+u in .@'(X), then u~O and
uj--+u in the weak topology of measures, that is, Uj(cf»--+u(cf» for every
cf>ECg(X).
Proof That u~O is obvious, and Theorem 2.1.7 shows that uj and U
are measures. In addition the proof of Theorem 2.1.7 gives for every
compact set K c X a uniform bound
luicf»I~CKSUplcf>l, cf>ECg(K), Vj.
Since uicf» --+ u(cf» for cf>EC';(K) it follows that this remains true for
all cf> in the closure F of C';(K) in Cg(K). By Theorem 1.3.2 F
contains all continuous functions with support in the interior of K,
which proves the theorem.
We shall now give some simple examples to indicate what conver-

gence of distributions means in concrete cases. They also illustrate the
fact, to be proved in Chapter IV, that every distribution is the limit
of a sequence of functions in CO' .
Example 2.1.10. Let vEC~(1R") and set u.(x)=c"v(x/s). Then

J J
U.(cf» = c"v(x/s)cf>(x)dx= v(x)cf>(sx)dx --> cf>(0) v(x)dx, J
if cf>EC~(1R"). (Compare Theorem 1.3.2.)
Example 2.1.11. Let wEC~(1Rn), assume that Jxaw(x)dx=O when

lod<k and set u.(X)=C"-kW(X/S). Then we have for cf>EC~(1Rn)
U.(cf» = Jcf>(sx)s-kw(x)dx= J( L siai-kxaaacf>(O)/od +O(s))w(x)dx

iai;i;k
--> I aacf>(O)Jxaw(x)dx/od.
iai=k
Example 2.1.12. If ur(x)=tNeirX, xE1R, where N is a positive integer,

then
J J
Ut(cf» = t Neirx cf>(x) dx = i t N- 1 eitx cf>'(x) dx = ...
=iN+lt-1Jeirxcf>(N+l)(x)dx-->O as t-->oo, cf>EC~(1R).
Example 2.1.13. If ur(x)=teitX, x>O and Ur(x) =0, x~O, then
Jteitxcf>(x)dx=icf>(O)+i J eitxcf>'(x)dx
00 00
Ut(cf» =
o 0
00
=icf>(O)-cf>'(O)/t- Jeirxcf>"(x)dx/t --> icf>(O), as t -->00,

o
if cf>EC~(1R).
Example 2.1.14. Let Ut(x)=tl/keitX", xE1R, where k is an integer >1. To
determine the limit we first examine
x
F(x)= Jeiykdy.
o
When x>O we shift the integration to a segment of the line argz=n/2k
and a circular are, on which Imzk~clzlk-l Imz where c>O.1t follows
that
Je-ykdy
00
F(x)-->e"i/2k as x --> 00.
o
Hence
00
F(x)-->_e"i/2k Jrykdy as x-->-oo ifkiseven

o
<Xl
F(x)-->-e-"i/ 2k Je- yk dy as x-->-oo ifkisodd.

o
2.2. Localization 41
Now we have if cPECo(1R.)

J
ut(cP) = t l /kF'(xtl/k)cP(x)dx = - JF(x tl/k)cP'(x)dx
o 00
--+ -F(-oo) J cP'(x)dx-F(oo) J cP'(x)dx=(F(oo)-F(-oo))cP(O).

-00 0
Here
J e-ykdy
00
F( 00)- F( - 00)=2 cos n/2k if k is odd

o
J rykdy
00
=2e"i/2k if k is even,
o
which is not O. (In terms of the r function defined in Section 3.2 the
integral is equal to r«l + k)/k).) Note the contrast with the case k = 1.
2.2. Localization
If Y eX e1R.n and UE.@'(X), we can restrict u to a distribution u y in Y
by setting
Uy(cP) = u(cP), cPECo(Y).
Our next purpose is to prove the less trivial fact that a distribution is
determined by the restrictions to the sets in an open covering:
Theorem 2.2.1. If UE.@'(X) and every point in X has a neighborhood to

which the restriction of U is 0, then U = O.
Proof If cPECo(X) we can for every XESUPPcP find an open neigh-

borhood Y e X such that the restriction of U to Y is O. By the Borel-
Lebesgue lemma we can choose a finite number of such open sets lj
e X which cover supp cP. But according to Theorem 1.4.4 we can then
write cP=I.cPj where cPjECo(lj). Thus u(cPj)=O which proves that u(cP)
= I. u(cPj) =0.
Theorem 2.2.1 makes it natural to extend Definition 1.2.2 as fol-
lows:
Definition 2.2.2. If UE.@'(X) then the support of u, denoted supp u, is

the set of points in X having no open neighborhood to which the
restriction of u is O.
Thus X ....... supp U is the open set of points having a neighborhood

in which U vanishes, so U vanishes in X . . . . supp U by Theorem 2.2.1,
and X ....... supp U contains every open set where U vanishes. Thus we
have
(2.2.1) U(<p) = 0 if UEgc'(X), <pEC't'(X) and suppullsupp¢=0.
Closely related to the notion of support is the notion of singular

support:
Definition 2.2.3. If UEgc'(X), then the singular support of u, denoted

sing supp u, is the set of points in X having no open neighborhood to
which the restriction of U is a Coo function.
Since every point in X ....... sing supp U has a neighborhood where U is
a Coo function, it follows from Theorem 2.2.1 that the restriction of U to
X ....... sing supp U is a Coo function. This is not true for any larger open
set than X ....... sing supp u.
We shall now supplement Theorem 2.2.1 by an existence theorem:
Theorem 2.2.4. Let Xi' iEI, be an arbitrary family of open sets in 1R.n,
and set X=UX i. If UiEgc'(X i) and ui=u j in Xi"X j for all i,jEI, then
there exists one and only one UEgc'(X) such that ui is the restriction of
U to Xi for every i.
Proof The uniqueness is precisely Theorem 2.2.1 and the existence

follows from essentially the same proof. If U is a distribution with the
required properties we must have
(2.2.2) u(¢)=Iui(<Pi) if ¢=I¢i' ¢iEC't'(Xi)
and the sum is finite. By Theorem 1.4.4 every <pEC't'(X) can be
written as such a sum, and we shall prove that I Ui(¢i) is independent
of how it is chosen. This will follow if we show that I ¢i = 0 implies
L Ui(¢i)=O. Set K = Usupp ¢i' which is a compact subset of X, and
use Theorem 1.4.5 to choose functions t/lkEC't'(Xk) such that It/lk=l
in K and the sum is finite. Then we have t/lk<PiECO(Xk"Xi) so
Ui(t/lk<Pi) =Uk(t/lk¢J
Hence
Having proved that (2.2.2) defines uniquely a linear form on

Co(X) we must show that it has the continuity properties required of
2.2. Localization 43
a distribution. Choose any compact set K c X and as before functions

t/lkEC't;'(Xk) with Lt/lk=1 on K and the sum finite. If fjJEC't;'(K) we
have fjJ= LfjJt/lk with fjJt/lkEC't;'(Xk) so (2.2.2) gives
U(fjJ) = L Uk(fjJt/lk)·
Since (2.1.2) is valid for Uk and the maximum of the derivatives of fjJt/lk
can be estimated in terms of those of fjJ, we conclude that
lu(fjJ)I~C L suplaafjJl, fjJEC't;'(K),

lal~m
which completes the proof. Note that it shows that if UiE~'k for every
i then UE~'k.
Remark. If U·E~'(X), v = 1, 2, ... and u· restricted to Xi has a limit UiE

~'(Xi) for every i, then the proof shows that u· has a limit in ~'(X).
In fact, if we apply (2.2.2) to each u· we find that u·(fjJ) has a limit u(fjJ)
satisfying (2.2.2). The continuity of u follows from the second part of
the proof of Theorem 2.2.4 or else from Theorem 2.1.8.
If uEL\oc(X), the form

u(c/» = SufjJdx
is of course meaningful for every fjJE COO(X) such that
(2.2.3) SUppUIlSUpPc/>~X.
(We write A~X when A is compact and contained in X.) We shall

now prove that the domain of an arbitrary distribution can be ex-
tended in this way.
Theorem 2.2.5. Let UE~'(X) and let F be a relatively closed subset of

X containing supp u. Then there is one and only one linear form u on
{fjJ; fjJECOO(X), FllsupPc/>~X} such that
(i) u(c/»=u(c/» if fjJEC't;'(X),
(ii) u(c/» =0 if c/>ECOO(X) and Fllsuppc/>=0.
The domain of it is of course largest when F = supp u, but we need

the uniqueness statement also for other sets F.
Proof a) Uniqueness. Let fjJECOO(X) and let FllsuppfjJ=K be a com-

pact subset of X. By Theorem 1.4.1 we can find t/lEC't;'(X) so that t/I
=1 in a neighborhood of K. Then we have fjJ=fjJO+fjJI where fjJo
=t/I¢EC'~(X) and cPI =(1- t/I)cP, so that F nsupp cPI =0. Using (i) and
(ii) we obtain
u(cP) = u(cPo)+ u(cP I) = u(cPo),
which proves the uniqueness of U.
b) Existence. We have seen in a) that every cP E COO(X) with
F nsuppcP compact can be written cP =cPo +cPI with cPoEC~(X) and
F n supp cP 1 = 0. If cP = ¢~ + cP'l is another such decomposition, then X
=cPo-¢~EC~(X) and FnsuPPx=FnsuPP(cPl-cP'l)=0 so it follows
from (2.2.1) that O=u(X)=u(cPo)-u(cP~). Setting u(cP)=u(cPo) therefore
gives a unique definition of a linear form u which obviously has the
required properties.
From now on we write u(¢) instead of u(cP) and thus consider u(cP)
as defined for all UEg&'(X) and all cPECOO(X) satisfying (2.2.3). In view
of the symmetry of (2.2.3) we shall sometimes write (u, cP) instead of
u(¢).
2.3. Distributions with Compact Support
If UE~'(X) has compact support we have seen that u(cP) can be

defined for all cPECOO(X). When t/lEC~(X) and t/I= 1 in a neigh-
borhood of supp u, we have
u(¢)=u(t/I¢)+u«1-t/l)cP)=u(t/lcP), cPECOO(X).
Hence it follows from (2.1.2) that
(2.3.1) lu(¢)1 ~ C L sup laacPl, cPECOO(X),
lal;iik K
where K is the support of t/I and C, k are constants. Conversely,

suppose that we have a linear form v on COO(X) such that for some
constants C and k and some compact set LeX
(2.3.2) Iv(¢)I~C L sup laa¢l, cPECOO(X).
lal;iik L
Then the restriction of v to C~(X) is a distribution u with support

contained in L. Since it follows from (2.3.2) that v(cP)=O if LnsuppcP
=0, we obtain from Theorem 2.2.5 that v(cP) = u(cP) for every ¢ECOO(X).
Hence we have proved
Theorem 2.3.1. The set of distributions in X with compact support is

identical with the dual space of COO(X) with the topology defined by the
semi-norms
¢ -+ I sup lo"¢I,
l"l~k K
where K ranges over all compact subsets of X and k over all integers
~O.
Schwartz used the notation @,,(X) for the space Coo(X) equipped
with this topology. Accordingly the space of distributions with com-
pact support in X is denoted by @"'(X). From the proof of Theorem
2.3.1 it follows that @"'(X) can be identified with the set of distri-
butions in @"'(1R") with supports contained in X. We may therefore use
the notation @"'(A) also when A is an arbitrary subset of JR" to denote
the set of distributions in @"'(JR") with supports contained in A. We
write @"'k(A)=@"'(A)(l~'k(JR").
The smallest k which can be used in (2.3.1) is of course the order
of the distribution u. For K one can take any neighborhood of supp u
but usually not the support itself.
Example 2.3.2. Let K be a compact set in JR" which is not the union
of finitely many compact connected sets. Then one can find uE@"'(K)
of order 1 so that (2.3.1) is not valid for any C and k. In fact, the
hypothesis means that we can find a sequence of disjoint non-empty
compact subsets K j of K such that K'-.(K1u ... uKj) is compact.
Choose xjEK j , let Xo be a limit point of {xJ and set
u(¢)= I mi¢(x)-¢(xo))
where mj is a positive sequence such that
Imjlxj-xol=1, Imj=oo.
Such a sequence exists since liminflxj-xol=O. Then
lu(¢)1 ~sup WI
so u is a distribution. On the other hand, if (2.3.1) is valid and we
choose ¢ E COO equal to 1 in a neighborhood of K 1 u ... u K j and 0
near K '-. (K 1 u ... u K j ), hence at x o, then we obtain
I
i~j
mi~C
which is a contradiction when j -+ 00.
Although (2.3.1) is not in general valid with K =supp u we can

prove that the left-hand side must vanish then if the right-hand side
does:
Theorem 2.3.3. If UEtf' is of order ~k and if 4>ECk,

(2.3.3) 8a4>(x) =0 for lal~k and XESUPPU,
then it follows that u( 4» = o.
Recall that u(4)) was defined for all 4>EC~ in Theorem 2.1.6, and as
in Theorem 2.2.5 we have a unique extension to all 4>ECk, for which
u(4))=O when suppunsupp4>=0. The estimate (2.3.1) is valid for all
4> E Ck if K is a neighborhood of supp u.
Proof of Theorem 2.3.3. By Theorem 1.4.1 and the remarks after it we

can choose X.ECO' so that X.=1 in a neighborhood of suppu, X.=O
outside
M.={y; Ix-yl~e for some XESUppU}
and
18a.x.I~Ce-la.l, lod~k.
Since suppu and supp(1-X')4> do not intersect, we have

u(4)) = u(4) X.) + u(4)(1- X.» = u(4) X.)
so using (2.3.1) we obtain
lu(4))I~C I supI8a.(4)x.)I~C' I supI8a.4>118 Px.1
lal~k la.l+IPI~k
~ C" I Bla.H sup 1811.4>1.
la.l~k M,
To show that the right-hand side tends to 0 with B we must prove that
(2.3.4) BlaH SUp 1811.4>1 ..... 0 when B ..... 0 if lad ~ k.
M,
By the definition of M. we can for every YEM. choose XESUPP u so

that Ix - yl ~ B. This gives (2.3.4) for IIXI = k since 8a 4> is uniformly
continuous and vanishes on Supp u. If lal < k and YEM. we obtain by
Taylor's formula
1
I811. 4> (y)1 ~ k I I sup l(d/dt)k-la.l (811. 4» (x + t(y - x»1
( - a)! 0<1<1
for the derivatives of 4> of order < k vanish at x. If the differentiation
is carried out we obtain a sum of terms involving a factor (y - x)p,
IPI = k -Ial, and a derivative of 4> of order k, which proves (2.3.4).
An important consequence of Theorem 2.3.3 is the following
Theorem 2.3.4. If u is a distribution of order k with support equal to

{y}, then u has the form
(2.3.5) u(</»= L aaaa</>(y), </>ECk •

lal~k
Proof If we expand </> in a Taylor series

</>(x) = L aa</>(y)(x - ynoc! + t/J(x)
lal~k
we have aat/J(y)=O when locl~k so u(t/J)=O by Theorem 2.3.3. Hence

(2.3.5) follows with aa = u«. - ynoc !). (. denotes the variable.)
Note that Theorem 2.3.4 explains why the limits in Examples

2.1.10,2.1.11,2.1.13,2.1.14 had to be of the form (2.3.5).
There is a result similar to Theorem 2.3.4 when the point is
replaced by a subspace, but the proof is somewhat more complicated:
Theorem 2.3.5. Let x=(x', x") be a splitting of the variables in JR." in

two groups. If u is a distribution in JR." of order k with compact support
contained in the plane x' = 0, then
(2.3.6) u(</»= L ua(</>a)
lal~k
where ua is a distribution of compact support and order k -loci in the x"

variables, oc = (ex', 0) and
Proof. If CPEC oo and we form the Taylor expansion in x',

cp(x) = L aa</>(o, x")x'ajex! + tP(x)
la'i ~k,a" = 0
then a«tP(x)=O when x'=O and locl~k so u(tP)=O. Thus u(cp) has the
form (2.3.6) with
ua(t/J) = u(t/J(x")x'ajex !).
What is not obvious is that ua is of order k -Iexl and not just of order
k. To prove that we observe that ua(t/J)=u(cp) for all </>ECoo such that
(2.3.7) cp(x)=t/J(x")x''''jex!+O(lx'lk+ 1) as x'-+O.
By repeated application of Corollary l.3.4 to one x' variable at a time,
we find that for any t/JECk-I",1 one can find CPECk satisfying (2.3.7).
The proof shows that if t/J E COO one can find cp E COO so that for any
given compact set K c JR."
L suplaYcpl~c L suplapt/JI.
IYI~k K IPI~k-lal
Hence we obtain
IUa(I/I)I~C' L sup I(JPI/I I, I/IEC,(?,
IPI;iik-lal
which completes the proof.
For the sake of completeness we shall finally discuss a more

general method for extending functions than Corollary 1.3.4 and ex-
amine its implications for the structure of distributions with given
support.
If UECk (1R.") we denote its Taylor polynomial of order k at y by
uk(x, y) or u(x, y) for short,
u(x, y) = L (Jau(y)(x - y)ala!.
lal;iik
We denote the remainder term by R(x,y), thus u(x)=u(x,y)+R(x,y).
Note that
(J~R(x,y)=(Jau(x)- L (Ja+Pu(y)(x-y)PIP!
IPI;iik-lal
is determined by the derivatives of u at x and at y only. By Taylor's
formula the quotient 1(J~R(x, y)l/lx - ylk-1a l is continuous in 1R." x 1R." if
it is defined as 0 on the diagonal.
Theorem 2.3.6 (Whitney's Extension Theorem). Let K be a compact set

in 1R." and u"" lal~k, continuous functions on K. Set for lal~k
V",(x,y)=lua(x)- L ua+p(y)(x-y)PIP!lIx-yllal-k
IPI;iik-I",1
when x, YEK and X9=Y, V",(x,x)=O when XEK. If V'" is continuous on
K xK when lal~k, it is possible to find VEC k(1R.") with (Jav(x)=u",(x),
xEK, lal ~k. One can then choose v so that with C depending only on K
(2.3.8) L sup I(J"'vl ~ C( L sup V", + L sup lual).
l"'l;iik lal;iik KxK lal;iik K
We have already observed the necessity of the condition in the

theorem, and the proof of sufficiency will depend on the following
Lemma 2.3.7. There exists a partition of unity 1 = L <Pj in CK such that

no point is in the support of more than N functions <Pj' the diameter of
the support of <P j is at most twice its distance to K, and
(2.3.9) I(Ja <p}x) I ~ C",d(x)-I"'I, xf/;K,
where d(x) is the distance from x to K.
Proof This is a special case of Theorem 1.4.1 0 with the metric chosen
as in Example 1.4.8 where F = K.
Proof of Theorem 2.3.6. Choose for every j a point YjEK with minimal
distance to supp cPj. With
u(x, y) = I u",(y)(x - yntx!
1a:I~k
denoting the expected Taylor expansion of v at YEK, we set

vex) = I cPix)u(x, Y), x¢K; vex) = Uo(X), XEK.
Choose x*EK with Ix-x*l=d(x). If x is in the support of cPj then the
distance from supp cPj to K is at most d(x), the diameter is at most
2d(x), so Ix - Yjl ~ 3d(x) and Ix* - Y) ~ 4d(x). If Y is an arbitrary point
in K then
(2.3.10) Ivex) - u(x, y)1 = o(lx - Yl k)
uniformly in y. If XEK this follows from the continuity of Uo(x, y).
Otherwise we use the fact that
Iv(x)-u(x, y)1 ~I cPix) lu(x, Yj) -u(x, Y)I·
Here IX-Yjl~3d(x)~3Ix-YI when cPj(xHO. When Iyl~k we have
(2.3.11) lo~(u(x, Yl) - u(x'Y2))1
=0((lx-y11+IYl -Y 21)k-lYl) if Yl' Y2 EK,
for o~(u(x, Yl) - u(x, Y2)) is a polynomial in x of degree k -Iyl and
lo~o~(u(x, Yl) -u(x, Y2))x=y,1

= Up+iYl' Y2) IYl - Y2Ik-lfJl-lyl =O(IYl - Y2Ik-lfJ+YI).
(2.3.11) with y=O gives (2.3.10); the general case will be useful later
on.
From (2.3.10) it follows that v is continuous and if k~ 1 also that v
is differentiable at any point YEK, with differential dx u(x, Y)lx=y' When
x¢K and k>O we obtain by differentiation
(2.3.12)
Set v.(x) = I cPj(X)O.u(x, Y), x¢K, and v.(x) = ua:v(x), xEK, where tx.
=(0, ... ,1,0, ... ) with 1 just in the v-th place. If our contentions are
already proved with k replaced by k-l then V.ECk - 1 • If we prove
that the first sum in (2.3.12) and its derivatives of order ~k-l tend
to 0 when x -+ K, it follows in view of Corollary 1.1.2 (in n variables)
that ovv is continuous and in fact in Ck- \ and that o"v = u" on K,
1Q(I~k.
Now we have for x¢K when 1/31+lyl~k and /3*0
L oPCPj(x)o~u(x, y)= L oPcp/x)o~(u(x, Yj)-u(x, x*)),
because LOPcp/x)=O when /3*0. Recalling that Ix-y)~3d(x) we
obtain from (2.3.11) and (2.3.9)
L oPcp/x)o~u(x, Y) = o(d(x)-IPI+k-IYI)
which proves the assertion on the first sum in (2.3.12). The preceding
estimate and a trivial bound for
L cp/x)o~u(x, Y)
give, if we pay attention to constants also, that
L lo"v(x)1 ~ C(1 + d(X))k( L sup V" + L sup IU"I).
l"l~k l"l~k KxK l"l~k K
Now if we drop from the definition of v the terms for which supp CPj
°
has distance > 1 from K, we do not change v where d(x) < 1 but we
make v(x) = when d(x) > 3, so (2.3.8) follows.
Corollary 2.3.8. If U is a distribution of order k and compact support K,

then
(2.3.13) lu(cp)I~c( L sup lo"cp(x)- L o"+Pcp(y)(x-y)PI/3!1
1"I~kX.YEK.x*y If/I~k-I"I
·lx-yll"l-k+ L suplo"cpl), CPEC OCJ •

l"l~k K
Proof. If we apply Theorem 2.3.6 with u" = o"cp IK we obtain a function

v with u(v-cp)=O, by Theorem 2.3.3. Hence
lu(cp)1 = lu(v)1 ~ C L sup lo"vl

l"l~k
and (2.3.13) therefore follows from (2.3.8).
Example 2.3.2 shows that the first sum on the right-hand side
cannot be omitted if K is not a finite union of compact connected
sets. By Banach's theorem we also know that every continuous linear
form with respect to the seminorm in the right-hand side of (2.3.13) is
continuous with respect to the seminorm L
sup lo"cpl if and only if
1"I~k K
the first sum in (2.3.13) can be estimated by the second one. A
necessary condition for this is given in the following
Theorem 2.3.9. Let K be a compact connected set and assume that

(2.3.14) sup Il/I(x)-l/I(y)l/lx-yl~C L suplaal/ll, l/IEC oo •
x,YEK,x*y lal::;;l K
Then there is a constant C' such that any two points x, YEK can be
joined by a rectifiable curve in K with length ~ C'lx - YI.
Proof Fix two points x o, YoEK. If X is a connected open neigh-

borhood of K we denote by dey) the infimum of the length of poly-
gonal arcs from Xo to y contained in X. If u(y) = min (d(y), d(yo)) then
u(xo)=O and u(yo)=d(yo), and we have
(2.3.15) lu(x)-u(y)I~lx-yl if [x,y]c:X.
Define ulj> according to Theorem 1.3.2 with ¢ of so small support that
ulj> is defined in a neighborhood of K. In view of (2.3.15) we have
IUIj>(x) - u",(y)1 ~ Ix - yl
in a neighborhood of K when Ix - yl is small, hence lai ulj>l ~ 1 on K.
With l/I = ulj> we obtain from (2.3.14)
IUIj>(xo)-ulj>(Yo)1 ~ Ixo - Yol C(d(yo) +n).
Letting supp ¢ -+ {OJ we conclude that
d(Yo)~lxo-yol C(d(Yo)+n).
When Ixo - Yol ~ 1/2 C it follows that d(Yo)~2 n Clxo - Yol.

For any &>0 the set
K,={x; Ix-yl<& for some YEK}
is a connected neighborhood of K since K is connected. Hence it
contains an arc oflength <2nClxo-Yol+& from Xo to Yo' Represent-
ing it as a function of the arc length we obtain when &-+ 0 a curve of
length ~2nClxo-Yol from Xo to Yo which is contained in K, provided
that Ixo - Yol ~ 1/2 C. Since K is compact and connected this proves
the theorem.
The conditions given in Example 2.3.2 and Theorem 2.3.9 are also
sufficient:
Theorem 2.3.10. Let K be a compact set in JR.n with finitely many
connected components such that any two points x, y in the same com-
ponent can be jOined by a rectifiable curve in K of length ~ Clx - YI. If
U is a distribution of order k with supp u c: K, it follows then that
(2.3.16) lu(¢)1 ~ C L sup laa¢l, ¢Eck(JR.n).

lal::;;k K
Proof Let s --> x(s) be a curve in K with x(O) = y and arc length s. Then
(2.3.17) 1t:,(s)1 ~ C 1'-1«1 L sup loll ¢I, IIXI ~ k,
11I1=k K
if
t:,(s)=o«¢(x(s))- L o«+II¢(y)(x(s)_y)Ii/f3!.
11I1~k-I«1
This is obvious when IIXI = k. If IIXI < k and (2.3.17) is already proved for
derivatives of higher order, we conclude that
Idt:,(s)/dsl~ Cnsk-I«I-l L suploll¢l.
llil=k K
Since t:,(O) =0 we obtain (2.3.17) with C replaced by Cn. If d(x, y) is

the infimum of the lengths of curves from y to x in K then (2.3.17)
gives
(2.3.17), lo«¢(x)- L o«+II¢(y)(x_y)Ii/f3!1
llil~k-I"I
~ Cd(x, yf-'·' L suploll¢l.

IfJl=k K
When d(x,y)~Clx-yl the estimate (2.3.16) follows from (2.3.13) and

(2.3.17),
Theorem 2.3.11. Let K be a compact set in JRn with finitely many
connected components and assume that any two points in the same
component can be joined by a rectifiable curve in K of length at most
Clx- ylY. If u is a distribution of order k with suppucK and if m is
an integer with m 'Y ~ k, then
(2.3.18) lu(¢)1 ~ C L suplo"¢I, ¢E Coo.
1«I~m K
Proof We use (2.3.13) combined with (2.3.17)" now with k replaced by

m, noting that since 'Y ~ 1
Ix_Yly(m-I«I)~lx_Ylk-I«1 if Ix-YI<1.
Sets satisfying the hypotheses in Theorem 2.3.11 are sometimes

called regular in the sense of Whitney.
Notes
As mentioned in the introduction the definition of distributions used
here is that of Schwartz [1]. One of its advantages is that it suggests
naturally the proof of existence theorems for differential equations by
Notes 53
duality (the Hahn-Banach theorem). Such ideas had already led So-
bolev [1] rather far towards a distribution theory.
Other spaces of test functions give different spaces of distributions.
In Chapter VII we shall define a space !/' with @"'(1R")c:!/'c:~'(1R")
by using as test functions a subspace !/ of Coo(1R") restricted by global
conditions. Any non-quasi-analytic class of functions can be used as
test functions for a distribution space with properties similar to those
of Schwartz distributions. This has been done by Beurling [1] (see
also Bjorck [1]). One can also use quasi-analytic classes of test func-
tions but localization is much harder then. The typical real analytic
class will be studied in Chapter IX. Also spaces of entire analytic
functions can be used as test functions for special purposes (see
Gelfand and Silov [1]), but one is then rather far removed from the
intuitive notion of a generalized function.
The topology in Cg"(X) defined by the semi-norms in the right-
hand side of (2.1.3) is the inductive limit of the topology in Cg"(K)
when the compact set K increases to X, so it is a !l'!F topology. (See
Dieudonne-Schwartz [1].) We have avoided this terminology in order
not to encourage the once common misconception that familiarity with
!l'!F spaces is essential for the understanding of distribution theory.
The convenient explicit form of the semi-norms in (2.1.3) has been
adopted from Garding and Lions [1], and it will occasionally be im-
portant later. (For an example see Section 10.7.)
The problem of estimating u(cP) by means of cP and its derivatives
in supp u only was discussed in Schwartz [1] where Example 2.3.2 and
Theorems 2.3.10, 2.3.11 are given in an only slightly different form.
Corollary 2.3.8 and Theorem 2.3.9 were proved in Glaeser [1] (see
also Hormander [5]). The main point is of course the extension
theorem 2.3.6 of Whitney [1]. His results are actually stronger and
cover also the extension of COO functions. In that case the extended
function v does not depend linearly on the given data ua • A linear
extension of COO functions from a half space to the whole space has
been given by Seeley [2].
Chapter III. Differentiation
and Multiplication by Functions
Summary
Our motivations for distribution theory came largely from the limi-
tations of the classical notion of differentiability. In this chapter we
shall see that differentiation of distributions is indeed always possible.
In addition we shall discuss multiplication. This operation on the other
hand is not always defined unless one factor is smooth.
Differentiation of distributions and multiplication by smooth func-
tions is defined in Section 3.1. As examples we discuss differentiation
of functions with simple discontinuities which leads us to the Gauss-
Green formula, and to Cauchy's integral formula. As an application of
the latter we digress to discuss boundary values in the distribution
sense of analytic functions. As further illustration of multiplication
and differentiation of distributions we discuss homogeneous distri-
butions at some length in Section 3.2. Fundamental solutions of some
classical second order differential operators are constructed in Section
3.3. In Section 3.4 finally we have collected some computations of
integrals, particularly of Gaussian functions, which are needed in
those constructions.
3.1. Definitions and Examples

If u is a continuous function such that 0kU is defined everywhere and
continuous, we obtain by partial integration
f(OkU) ¢dx= - Juok¢dx, ¢EC'O(X).
(Compare with the weak derivatives in the introduction.) If I is a
continuous function then
J(fu)¢dx=Ju(f¢)dx, ¢EC'O(X),
where I ¢ is another test function if IE COO. The following definition is
therefore an extension of the classical one:
3.1. Definitions and Examples 55
Definition 3.1.1. If UE.@'(X) we set

(3.1.1)
and if jECOO(X) we define
(3.1.2) (fu)(c/J) = u(fc/J) , c/JEC~(X).
It is clear that (3.1.1) and (3.1.2) define distributions 0kU and ju

with support contained in supp u, and that the maps u -+ Ok u and
u -+ ju are continuous.
Remarks. 1) From the uniqueness part of Theorem 2.2.5, with F

= supp u, we conclude that (3.1.1) and (3.1.2) are valid for all
c/JECOO(X) with suppunsuppc/J~X.
2) The product ju is defined for all u, jE.@'(X) with
(3.1.3) sing supp u n sing supp j = 0.
In fact, X is covered by open subsets Y such that UECOO(y) or
jECOO(Y), so the product is defined in Y by (3.1.2) if jECOO(y) and by
(3.1.2)' (fu)(c/J) = j(uc/J), c/JEC~(Y),
if UE COO(y). In case j and u are both in COO(y) the two definitions

agree with the pointwise definition of the product uf, so we have a
uniquely defined product (fU)yE.@'(Y). The restriction to Z c Y is of
course (fu>z. By the localization Theorem 2.2.4 it follows that there is
a unique distribution jUE.@'(X) with restriction (fu}y to Y for every Y.
3) If jECOO(X), UE.@'(X) and
(3.1.3)' supp u n supp j ~X
then the definitions at the end of Section 2.2 mean that

<u,f) = u(f) = (fu)(1)
for u(f)=u(c/Jj)=(fu)(c/J)=(fu)(1) if c/JEC~(X) is equal to 1 near
supp U n supp f In view of the preceding remark we can therefore use
this formula to extend the definition of <u,f) to all u, jE.@'(X)
satisfying (3.1.3) and (3.1.3)'.
4) From Theorem 2.1.6 it follows that (3.1.2) also defines a distri-
bution jUE.@'k(X) if jECk(X) and UE.@,k(X). The preceding remarks
are of course applicable also to this situation although more awkward
to state.
As for smooth functions one can interchange the order of differen-

tiation for distributions,
56 III. Differentiation and Multiplication by Functions
8j 8ku=8 k8j u, UEgc'(X).

In fact, if c{>EC't'(X) the definition means that
(8 j 8ku)(c{» = - (8 ku)(8 j c{» = u(8 k8j c{» = u(8 j 8kc{»
= (8 k8j u)(c{».
We can therefore use the notation 8a u for partial derivatives as we
have done in the case of functions. Iteration of (3.1.1) gives
(3.1.1)'
The usual rule for differentiation of a product remains valid,
(3.1.4) 8k(fu)=(8d)u+ f(8 k u); fECoo(X), UEgc'(X);
for this means that
-u(f8kc{» = u((8d) c{» - u(8 k(f c{»), c{>E C't'(X) ,
which is true since 8k(fc{»=(8kf)C{>+f8kc{>.
Example 3.1.2. The function H(x)=1 for x>O, H(x)=O for x~O, on 1R
is called the Heaviside function. The derivative is by definition
00
H'(c{» = -H(c{>') = - Sc{>'(x)dx=c{>(O).

o
One defines the Dirac measure ba at aE1R" by
that is, as the unit mass at a. With this notation H' = boo The de-
rivatives of the Dirac measure are
(8 a b.)(c{» =( _1)la l 8ac{>(a) , c{>E COO,
so Theorem 2.3.4 means that linear combinations of ba and its de-
rivatives are the only distributions with support at a.
Theorem 3.1.3. If u is a function in the open set Xc 1R which is in

C1(X'- {x o}) for some XoEX, and if the function v which is equal to u'
for X=FXo is integrable in a neighborhood of x o, then the limits
u(xo±O)= lim u(x)
x-xo±o
exist and
u' = v +(u(xo +O)-u(xo -0)) bxo .
Proof If xo<y and the interval [xo,y] belongs to X then

y
u(x)=u(y)- Sv(t)dt, Xo <x<y,
x
which shows that u(xo +0) exists. Similarly u(xo -0) exists. We have
u'(¢) = -u(¢')=lim - S u(x)¢'(x)dx
' .... +0 Ix-xol>'
= lim (u(xo + s) ¢(xo + s) - u(xo - s) ¢(xo - s) + S v(x) ¢(x) dx),
, .... + 0 Ix-xol >,
if ¢EC't'(X). This proves the theorem.
Note that since u'-v=O in X,-{x o} it is clear that u'-v must

be a distribution with support at x o, hence a linear combination of t>xo
and its derivatives, which explains qualitatively the form of the con-
clusion.
Theorem 3.1.3 shows that by pointwise differentiation we may miss
something essential at a discontinuity. However, this is not the case
for the distribution derivative:
Theorem 3.1.4. If UE~'(X) where X is an open interval on 1R and if u'

= 0, then u is a constant.
Proof That u' = 0 means that

-u(¢')=O, ¢EC't'(X).
If t/lE C't'(X) then the equation ¢' = t/I has the unique solution
x
¢(x) = S t/I(t) dt
-00
vanishing to the left of the support of t/I. It is in C't'(X) if and only if

it vanishes to the right of it also, that is, if
00
l(t/I) = S t/I(x) dx
-00
is equal to O. Thu~ u(t/I)=O if l(t/I)=O, so u(t/I)=Cl(t/I) for some

constant C. In fact, if t/loECO'(X) is chosen with 1(t/lo)=1, then the
integral of t/I- l(t/I) t/lo is 0 so
0= u(t/I- l(t/I) t/I 0) = u(t/I) - l(t/I) u(t/lo)·
Hence u = C = u( IjJ 0) as claimed.
Corollary 3.1.5. If UE~'(X), where X c1R, and u' +au= fEC(X) where
aECOO(X), then UEC1(X). Thus u' +au= f in the classical sense.
Proof Assume first that a = O. Since f has a primitive function

VEC1(X) and (u-v),=u'-v'=f-f=O, it follows from Theorem 3.1.4
that u - v = C, so UE Cl. For a general a we let E be a solution =1= 0 of
the equation E' =Ea; we can take E =exp S adx, so EECoo • Now we
have
d
dx (Eu)=Eu' +E'u=E(u' +au)=EfEC,
so EUE C 1, hence UE C 1.
The corollary remains valid if u=(u1, ... ,Uk) and f=(/1' ... '/k)
have k components and a is a k x k matrix of COO functions. In the
preceding proof we just have to let E be an invertible k x k matrix
such that E' =Ea. Since ordinary differential equations of higher order
can be reduced to first order systems, we can now prove
Corollary 3.1.6. If UE~'(X), X dR, and if

U<m)+a m_ 1u(m-1)+ ... +aou= fEC(X),
where the coefficients ajECoo(X), then UEcm(x) so the equation is
fulfilled in the classical sense.
Proof With uj=u U - 1), 1 ~j~m, we have the equations

u~+am_1um+ ... +aOu1 =f, uj-uj+ 1 =0, l~j<m,
so UjEC1(X) by the preceding remark on Corollary 3.1.5. Hence

UEC m.
We shall now pass to several variables, proving first an analogue

of Theorem 3.1.4.
Theorem 3.1.4'. Let UE~'(Y x I) where Y is an open set in lRn - l and I

an open interval on lR. If 8nu=0 then
u(</J) = S uo(</J(·, xn)) dx n, </JE C'O(Y x I),
where UoE~'(Y). Thus u is a distribution Uo in x'=(x 1, ... ,Xn_ 1) inde-
pendent of x n.
Proof Choose t/loEC'O(I) with St/lo(t)dt=l and define

uo(X)=u(Xo); XEC'O(Y), Xo(x)=X(x')t/lo(x n)·
It is obvious that UoE~'(Y). If </JE c;:''O(Y x I) and
(I </J)(x') = S </J(x', xn) dx n,
we have just as in the proof of Theorem 3.1.4

¢(X) - (I ¢ )(X') I{I o(Xn) = on (]I
where (]IECg'(Y x I). Hence
u(¢) = u((I ¢)o)= uo(I ¢)= Ju o(¢(" xn)) dX n
where the last equality follows from Theorem 2.1.3 applied to
-00
We shall now prove a weak but useful analogue of Corollary 3.1.5.
Theorem 3.1.7. If u and f are continuous functions in X eJRn and OjU

= f in the sense of distribution theory, then 0ju(x) exists at every XEX
and is equal to f(x).
Proof. We may assume that j = n and that X = Y x I as in

Theorem 3.1.4', for the assertion is local. Let T be a fixed point in I
and set Xn
v(X) = Jf(x', t)dt.

T
Then on(u-v)= f - f =0 in the sense of distribution theory, so Theo-

rem 3.1.4' and its proof shows that u(x)-v(x)=w(x') where w is a con-
tinuous function. Since v (x) + w(x') is pointwise differentiable with
respect to Xn with derivative J, the theorem is proved.
We shall now extend Example 3.1.2 by differentiating the charac-

teristic function of an open set in JRn with a C l boundary.
Definition 3.1.8. If Y e X are open sets in JRn then Y is said to have

C l boundary in X if for every boundary point XoEX of Y one can
find a C l function p in a neighborhood X 0 of Xo such that
p(xo)=O, dp(xoHO, YnXo={xEX o; p(X) <O}.
Using a partition of unity it is easy to show that there exists a
function pEC1(X) such that p=O, dp=l=O on the boundary OY of Y in
X and YnX={xEX; p(x)<O}. We leave the proof as an exercise for
the reader. Right now it is more important for us to note that if say
op/ox l =1=0 at Xo then the C l map
(Xl"'" Xn) --+(p(X), X2 ,···, xn)
from a neighborhood of Xo has a C l inverse. This implies that p =0 is
equivalent to Xl =1{I(x') where I{IEC l and x'=(x 2 , ... ,xn). If op/oxl,§O
say at Xo then Y is defined by Xl ~ l{I(x') in a neighborhood X~ of xo'
If Xy is the characteristic function of Y, then 0jXy is supported by

oY. To compute 0jXy in X~ we observe that if hEC oo vanishes in
( - 00, 0) and equals 1 in (1, (0), and Y n X~ is defined by Xl> "'(x') then
.-0
Xy=limh((xi -I/I(X'))/B) in X~,
pointwise and in the sense of distributions, so
._0
0jXy = lim vF((x 1 - I/I(X'»/B)/B
in the latter sense if v=(I, -o",/ox'). Thus we have if 4>EC~(X~)
._0
<OjXy, 4» = lim Svjh'((x 1- I/I(X'»/B) C 14>(X) dx
= SVj(x') 4>("'(x'), x') dx'
(cf. Theorem 1.3.2). If we observe that the Euclidean surface element

dS on OY is (1 + W12)t dx' and that v/(1 + W12)t = n is the interior unit
normal of 0Y, this means that
(3.1.5)
in X~ and therefore in all of X.

If f =(fl' ... ,f,,) is a vector field with components in C~(X), then
(3.1.5) means that
Sdiv f dx = <Xy, L 0jl;) = - L <OjXy,f) = - oYS<f. n) dS.
y
The formula
(3.1.6) Sdivfdx= - S <f. n)dS
y oY
is the Gauss-Green formula. By Theorem 1.3.2 it is valid for every
fEq(X). It is in fact easy to verify (3.1.6) for all continuous f with
compact support in X such that div f. defined in the sense of distribu-
tion theory happens to be in V (Y). However, we leave this as an
exercise for the reader. The formulas (3.1.5) and (3.1.6) are completely
equivalent, so if we had assumed (3.1.6) known we would have ob-
tained a slightly shorter proof of (3.1.5).
Theorem 3.1.9. Let YeX be open subsets of lR" such that Y has a C I
boundary oY in X, and let UEC 1 (X). If Xy denotes the characteristic
function of Y, dS the Euclidean surface measure on Y and n the
interior unit normal there, then
°
(3.1.7)
Proof If UECOO(X) we obtain (3.1.7) from (3.1.5) and (3.1.4). By Theo-

rem 1.3.2 any UECbeX) is the limit of a sequence of functions in
Cjf(X) converging in q(X), so (3.1.7) follows for all such u, hence in
general since the formula has a local character.
Our proof of (3.1.5) remains valid if the function", is Lipschitz
continuous; the normal n is just defined almost everywhere with
respect to dS then. In particular this means that we can use (3.1.6) for
cubes, where it is of course perfectly elementary. This is useful in the
proof of the following result.
Theorem3.1.10. Let P=Laj(x)l\+b(x), where ajEC l and bECo, be a

first order linear partial differential operator in X c]R". If U is differenti-
able at every point in X and there is a continuous function f such that
Pu(x) = f(x) for every XEX, then Pu = f in the sense of distribution
theory, which is meaningful since 8jUE~'1 and ajEC l .
Proof Let cf>ECjf(X). We must prove that

Ju(bcf>- L 8/aj cf»)dx= Jfcf>dx.
For a cube I with sides parallel to the coordinate axes we set
F,,(I)=F(I)= J(fcf>+u(L 8j (a j cf»-bcf»)dx+ Ju(a, n)cf>dS
I M
where n is the interior unit normal and (a, n) the Euclidean scalar
product. By (3.1.6) Fu(1) =0 if UEC 1. For general U we first claim only
that
(3.1.8) lim F(1)/m(1) =0
xoEI,m(I)~O
for every fixed XoEX. In the proof we denote by v the first order
Taylor expansion of u at x o, thus v(x)-u(x)=o(lx-xol). Then
Fu (1) = Fu (I) - Fv (1)
= J(f-Pv)cf>dx+ J(u-v)(L 8/a j cf»-bcf»dx+ J (u-v)(a, n) cf> dS
I I OJ
where Fv is defined with f replaced by Pv. Now

f(x)-Pv(x)~O as x~xo, suplu-vl=o(s(1))
I
where s(1) is the length of the edge of I. This proves (3.1.8).

(3.1.8) implies that F(1) = 0 for every I. In fact, starting from any
cube lowe can divide 1o into 2" cubes with sides of half the length,
and for one of these cubes II we must have
!P(Il)1 ~ IF(I o)I/2n

because F(I) is additive, that is, F(Io) is the sum of F(Il) when II
varies over the smaller cubes. Thus we can find a shrinking sequence
of cubes I j with m(Ij+ 1) = m(IN2n, such that
IF (Io)jfm(I 0) ~ IF(I l)jfm(I 1) ~ IF(I 2)I/m(I 2) ~ ....
They all have a point Xo in common so the limit is 0 by (4.1.8). Hence
F(Io)=O. When 10 is taken so large that supp¢cIo, the theorem is
proved.
The proof of the preceding theorem is taken from the classical

proof of the fact that if u is a differentiable function in Xc <C = lR 2
and du is proportional to dz at every point, then u is an analytic
function. If z = x + i y we can for every differentiable function v in <C
write
dv = ov/oxdx + ov/oydy = ov/ozdz + ov/ozdz
where
%z =t(i3/ox - iO/oy), %z =t(i3/ox + ii3/oy).
The condition on. u is therefore that ou/oz=O at every point and we
have shown that this implies that au/a z= 0 in the sense of distribution
theory. In Section 4.4 we shall see that such distributions are analytic
functions in the usual sense, in particular that they are in COO. Howev-
er, already now we shall make some further remarks on analytic
functions and the operator a/oz.
First we note that if Y eX c <C are open sets such that Y has C 1
boundary in X, then
(3.1.9) 2 S o¢/ozdxdy= - i S ¢d(x+iy), ¢EC~(X),
Y oY
where the C 1 curve aY is oriented so that Y is to the left of it. To
prove this we observe that if s is the arc length on aY then
(dx/ds, dy/ds) is the unit tangent and (-dy/ds, dx/ds) is the interior
unit normal. Taking f =(¢, i¢) in (3.1.6) we therefore obtain
2 S o¢/ozdxdy= - S (-¢dy/ds+i¢dx/ds)ds
Y oY
=-i S ¢(dx+idy).
oY
The formula (3.1.9) can also be written
(3.1.10) <OXy/OZ, ¢) =i/2 S ¢(dx+idy), ¢Eq(X).
oY
Let (E Y and apply (3.1.9) to </J{x, y)/{z - () with Y replaced by

Y 'D. where D. is a disc of radius B with center at ( and B is small.
Then
2 J o</J{x,y)/oz{z-()-1dxdy
y,D e
=- i J </J{x, y){z - ()-1 dz + i J </J{x, y){z - ()-1 dz,

ay a~
where oD. is oriented as the boundary of D. and not as that of Y,D •.

Throughout we have written z = x + i y. Since
J </J{x,y){z-()-1dz=</J{O J (z-()-1dz+O{B),
aD, aD,
letting B~ 0 gives Cauchy's integral formula
(3.1.11) </J{O= -n- 1 Jo</J{x, y)/oz{Z-()-1 dxdy
y
+(2ni)-1 J</J{x,y){Z-()-1 dz, </JECb{X), (EY.

ay
{Note that (x, y) ~ (x + i Y - ()- 1 is locally integrable.) In particular,

when Y=X there is no curve integral and (3.1.11) means that for the
function E,{x,y)=n- 1{z-()-1 we have
(3.1.12) oE,/oz = (j,.
The complete formula (3.1.11) follows from (3.1.12) and (3.1.10), for
8{Xy E,)/oz = (j, + E,oXy/oz,
in view of (3.1.4) and the localization principles in Section 2.2 since
one factor is smooth at every point.
We shall now discuss the existence of boundary values of analytic
functions in the sense of distribution theory.
Theorem 3.1.11. Let I be an open interval on 1R and let

Z={ZE<C; RezEI, O<Imz<y}
be a one sided complex neighborhood. If f is an analytic function in Z
such that for a non-negative integer N
If{z)1 ~ C{Im z)- N, ZEZ,
then f{. +iy) has a limit foE~'N+1{I) as y~O, that is,
lim
y~+o
Jf{x+iy)</J{x)dx=<fo, </J), </JEC~+ 1 (I).
Proof If N > 0 we choose some ZoE Z and introduce the integral

z
F(z) = I f(C)d', ZEZ,
zo
along some path in Z. The integral is an analytic function inde-

pendent of the path, F'(Z) = f(z), and if the path from Zo to Z is taken
first horizontal, then vertical, we obtain if 1 is bounded as we may
assume
IF(z)l;£ C 1(lmz)1-N if N>l,
IF(z)l;£ - C log 1m z+ C 1 if N =1.
In case N = 1 it follows that the integral of F is continuous in i since
log t is an integrable function of t near O. In any case we obtain after
N +1 integrations that f(Z)=G(N+1)(Z) where G is continuous in i
and analytic in Z. This proves that
lim f(. + iy)=lim dN+1G(. + iy)/dx N+1 =dN+1G(. )/dxN+1
y-o y-+O
in ~'N+ 1(1) which proves the theorem.
Another proof which gives a useful formula for fo is obtained as

follows. Set for 4>EC~+1(1)
<P(x, y)= L 4>U>(x)(iy)i/j!.
i~N
This is the Nth partial sum of the Taylor expansion in y which an

analytic extension of 4> would have if it did exist. Then <P(x, 0) = 4> (x)
and
2iJ<P/8Z=(iJ/iJx + iiJ/iJy) (N+ 1)(x)(iyt/N!.
Fix Ywith O<Y<y. IfO<y<y-Y we obtain from (3.1.9) applied to
4>(z)f(z+iy) that
I 4>(x, O)f(x +iy)dx- I <P(x, Y)f(x + i Y + iy)dx

=2i I I J(z+iy)iJ4>/iJzd)'(z)
O<lrnz<Y
where d)' is the Lebesgue measure in <C. Writing z = x + it Y, 0 < t < 1

gives
(3.1.13) I 4>(x)f(x+iy)dx= I 4>(x, Y)f(x+iY +iy)dx

1
+ I I f(x+ itY + iy)4>(N+ l)(x)(iYt+ 1 t N/N!dxdt.
o
There is a uniform bound for the integrand in the double integral as
y --+0 so
I cp(x)f(x+iy)dx ~ I iP(x, Y)f(x+iY)dx

1
+ I I f(x+ it Y)cp(N+ l)(x)(iy)N+ 1 t N/N! dxdt.
o
We shall often use the notation f(x + iO) for the distribution limit
just defined. Similarly we shall write f(x - iO) for such a limit from
the lower half plane.
Theorem 3.1.12. Let I be an open interval on IR and set

Z± ={ZE<C; RezEI, 0< ±Imz<y}.
If f is an analytic function in Z=Z+ uZ- such that
If(z)I~Cllmzl-N, ZEZ,
then the repeated integral
(3.1.14) F(cp) = J(I f(x + iy)cp(x, y)dx)dy, CPE C~(Z u I),
exists and defines a distribution in fiJ' N(Z u I) such that
i
(3.1.15) <oF/oz,cp) ='2 <f(. +iO)-f(. -iO), cp(. , 0),
We have F= fin Z.
Proof The first proof of Theorem 3.1.11 shows that F=G(N) where
IG(z)1 ~ C'log (C' /Iz!), hence GEL1. Partial integration gives
F(cp)=( _1)N I I G(x+ iy)oN cp(x, y)/ox Ndxdy, cpEC~(Z u I),
which proves the first statement. If CPEC~+ l(ZuI) then
<oF /oz, cp) = -<F, ocp/az) = -lim JJ f(x + iy)ocp(x, y)/ozdxdy
<_0 iyi><
=lim ~(J f(x + ie)cp(x, e)dx- I f(x - ie) cP(x, -e) dx)
H02
i
~'2<f(. +iO)-f(· -iO), cp(. , 0),
Example 3.1.13. (x+iO)-1_(x-iO)-1=-2nU)0 by (3.1.15) since

o(1/z)/oz=nb o,o by (3.1.12). (Here bo,o is the Dirac measure in <C at
0.)
When f(. +iO)=f(. -iO) we have of/oz=O. As already men-
tioned, we shall show in Section 4.4 that this implies that F is defined
by an analytic function. Thus f is the restriction to Z of an analytic

function in Z u I if and only if the two limits coincide.
The hypothesis in Theorem 3.1.11 cannot be relaxed very much:
Theorem 3.1.14. If f is analytic in the rectangle Z defined in Theorem

3.1.11 and limf(. +iy) exists in fi)'k(I), then
y-o
If(z)1 ~ C(lm Z)-k-l, ZEZ',
if Z' is the product of an interval J ~ I and (0, '1/2), say.
Proof Choose e/>EC~(lR2) equal to 1 near Z' and with support in

{z; RezEI, Ilmzl <'I}. Cauchy's integral formula applied to fe/> in the
set 1m Z > 1m '/2 gives if ,= e + i'1 EZ'
fm= -n- 1 IS f(x+iy)oe/>(x,y)/oz(z-,)-ldxdy
Y>~/2
+ (2ni)-1 Se/>(x, '1/2) (x - e - i'1/2)-1 f(x + i'1/2)dx.

The hypothesis implies a uniform bound for f(. + i y) in fi)'k when
0<y<y/2, so the last integral can be estimated by
C1 I supIO~(e/>(x,'1/2)(x-e-i'1/2)-1)I~C21'1I-k-1.
l"l~k
The double integral is even bounded. This proves the theorem.
Theorem 3.1.11 has an analogue for analytic functions of several

variables:
Theorem 3.1.15. Let X be an open set in lRn , r an open convex cone in

lRn, and set for some '1>0
Z={ZE<c n ; RezEX, ImzEr, Ilmzl<y}.
If f is an analytic function in Z such that
(3.1.16) If(z)I~Cllmzl-N, ZEZ,
then f(. + iy) has a limit foEfi)'N+ l(X) as r3Y -+ O. If fo =0 then f =0.
Proof Choose a fixed YEr with IYI<y. We may assume that Ol/I
since the theorem is trivial otherwise, and then we can choose C so
that
(3.1.17) t~Cly+tYI if YEr and t>O.
It is sufficient to prove this when t = 1 and then we just have to note
that - Y¢:f since 0 would otherwise be in r. Now set for e/>EC~+ l(X)
(3.1.18) cI>(x,y)= L aa¢(x)(iYnrY.!.

lal~N
Then we have if yET and lyl+IYI<y

(3.1.19) J¢(x) f(x + iy)dx = JcI>(x, Y) f(x+ iy+ iY)dx
+(N+l) JJ f(x+iy+itY) L aa¢(x)(iy)a/rY.!tNdxdt.
O<t<1 lal=N+l
In fact, the formula is clearly invariant under linear changes of vari-

ables so we may assume that Y lies along the positive x 1 axis. Then
the formula (3.1.19) is obtained by applying (3.1.13) for fixed x 2 ' ... , xn
and integrating afterwards with respect to these variables. By (3.1.16)
and (3.1.17) we have tNlf(x+iy+itY)I~C' so the integrand in the
double integral has an integrable majorant. Hence f(. + iy) converges
in ~/N+ 1 to fo where
(3.1.20) (fo,¢)=JcI>(x, Y)f(x+iY)dx

+(N+l) JJf(x+itY) L aa¢(x)(iy)a/rY.!tNdxdt.
o<t< 1 lal=N+ 1
Assume now that fo=O. Take YET and ¢ECO'(X) and form
F(w)= J¢(x)f(x+wy)dx= J¢(x-Rewy)f(x+ i 1m wy)dx.
This is an analytic function of w when 0 < 1m wlyl < y and
IRe wllyl <d, the distance from supp ¢ to ax. When 1m w -+ 0 we
know that F and all its derivatives tend to 0 since fo = O. Hence F
remains analytic when IRe wllyl < d if we define F = 0 when 1m w < O.
By the uniqueness of analytic continuation it follows that F=O identi-
cally. Hence f =0 in Z, which completes the proof.
Remark. If fo = 0 in an open non-void subset X 0 of X we still obtain

f=O in Z when RexEX o' Hence f=O in Z and fo=O in X if X is
connected.
Finally we shall prove an analogue of Theorem 3.1.4 for multipli-
cation.
Theorem 3.1.16. If UE~/(X) and xju=O, j=l, ... ,n, then u=c(jo for
some constant c.
Proof The support of u is at O. If ¢ECOO(X) we can by Theorem 1.1.9

write
L
¢(x) = ¢(O) + Xj¢j(x), ¢jE Coo,
and obtain
u(¢) =¢(O) u(l)+ L (Xj u, ¢j) =c ¢(O).
3.2. Homogeneous Distributions

If a is a complex number with Re a > -1 then the function on IR
x~ =xa if x>O, x~ =0 if x~O,
is locally integrable so it defines a distribution. (We define log x to be

real when x> 0 and this defines x a uniquely when x> 0.) It is clear
that
(3.2.1)
and by Theorem 3.1.3 we have
d
(3.2.2) ~xa =axa- 1 if Rea>O.
dx + +
We want to extend the definition of x~ to all aECC, as a distribution,

so that these properties are preserved as far as possible. That some
exception must occur is clear though for if a=O then the left-hand
side of (3.2.2) is H'(x) = c5 0 (Example 3.1.2) and the right-hand side
must be O.
For ¢EC~(IR) the function
00
a-->Ia(</»=<x~, </»= J xa¢(x)dx

o
is analytic when Re a > -1 for the differential is
00
da Jx alog x¢(x)dx.
o
Now (3.2.2) means that
(3.2.2)'
so for Re a > -1 and any integer k > 0 we have
(3.2.3)
The right-hand side is analytic for Re a > - k -1 except for simple
poles at -1, - 2, ... , - k. If a is not a negative integer we can thus
define I a(</» by analytic continuation with respect to a, or equivalently
by (3.2.3) with any k> -1- Re a. By (3.2.3) Ia then defines a distribu-
tion of order ~k. We shall denote it by x~. At a= -k the residue of
the function a-->Ia(¢) is
lim (a + k)I a(¢) =( _l)k I oWk»)/(l- k) ... ( -1) = </>(k- l)(O)/(k -i)!
a-+ ~k
so
(3.2.4)
Subtracting the singular part we obtain as a + k = 8 --+ 0

la( ¢) - ¢(k- 1)(O)/«k -I)! 8)
J
00
=(_l)k (x'-1)¢(k)(x)/«8+1-k) ... 8)dx

o
+ ¢(k- l)(O)(1/«k -1- 8) ... (1- 8)) -l/(k -1) !)/8 --+
-I (logx)¢(k)(x)dx/(k-1)!+¢(k-l)(O) Ct 1
1/j)/(k-1)!.
Thus we define
J(logx)¢(k)(x)dx/(k-1)!
00
(3.2.5) X~k(¢)= -
Ct
o
1
+¢(k-l)(O) l/j )/(k-l)!.
The relation (3.2.1) or equivalently

(3.2.1)' <x:,x¢)=<xa/\¢)
follows for all aE<C since it is valid when Re a > -1, hence by analytic
continuation when a is not a negative integer, and the value of both
sides when a = - k is obtained by letting a --+ - k after subtracting a
term C/(a+k), which must be the same on both sides. Also (3.2.2)
follows if a is not a negative integer or O. If k is a non-negative integer
then (3.2.4) gives
lim (!£x: +kx:- 1 ) = lim (a+k)x:- 1

a~-k dx a~-k
= lim (a+k+ l)x: =( -l)kb~)/k!,
a--+-k-l
so dropping terms of the form C b~) /( a + k) which must cancel, we

obtain
d
(3.2.2)" dx X~k = _kX~k-l +( -ltb~)/k!.
This follows also directly from (3.2.5) by an easy calculation.

The preceding argument is essentially due to Marcel Riesz. There
is an older method for defining the distribution x:'
due to Hadamard,
which consists of omitting first a neighborhood of the singularity at O.
Thus one forms, now for any aE<C
00
H a.,(¢)= Jxal{>(x)dx, ¢ECO'(1R.).
Assume that a is not a negative integer and let k be the smallest

integer ~ 0 such that k + Re a> -1. If we integrate by parts k times
and use Taylor's formula to express ¢Ul(e) in terms of derivatives of ¢

at 0, we obtain an identity of the form
00
(3.2.6) Ha .• (¢) =( _1)k J xa+k¢(kl(x)/«a + 1) ... (a+ k»dx

o
k-1
+ L: A j ¢U)(0)ea+1+j +o(1), e-+O.
o
There can be no other decomposition of the form
Ha,.(¢)=Bo+ L: Bj e- Aj +o(1), e -+ 0,
where the sum is finite and ReAj~O, A/tO. In fact, we have
Lemma 3.2.1. If Co, ... , Ck and A1' ... , Ak are different complex numbers
with Re Aj~O and Aj=l=O, then
k
Co + L: Cje- Aj -+ 0, e -+ 0,
1
implies that C o = ... = Ck=O.
Proof. Assume first that all Aj are purely imaginary. Replace e by erl
°
and let e -+ through a sequence such that e- Aj has a limit Yj for
every j. Then IYjl = 1 and
C o + L: CjYje-Ajl=O
for all real t, hence for all complex t. When t -+ 00 on the imaginary
axis one term dominates so this is not possible unless all Cj = 0. If
°
max Re Aj = (J > in the general case, we have
L: Cje(q-Aj) -+ 0, e -+ 0,
ReAj=q
so all the coefficients here must vanish which completes the proof.
+j ~ °
By Lemma 3.2.1 the terms in the expansion (3.2.6) with Re a + 1
are therefore uniquely determined, so it is legitimate to discard
the singular terms and define the finite part of the integral
00
Jxa¢(x)dx to be
o
JxQ+k¢(k)(x)/«a+1) ... (a+k»dx.
00
(_1)k
o
But this agrees with our previous definition of <x,,+, ¢) by (3.2.3). If a
= - k is a negative integer the procedure is somewhat different,
= _ _1_
(k-1)!
S(10gx)cp(k)(x)dx+cp(k-1)(0) (kf 1/j)l/(k-1)!
0 Y' 1
/~2 AjCPU)(0)ei+1-k_(~~t)! cp(k-1)(0)+o(I).
*
To define the finite part we must discard not only linear combinations
of powers e- A. with Re A~ 0, A 0, but also a multiple of log e. This
can be justified by an analogue of Lemma 3.2.1 and gives (3.2.5) again.
However, the notion of finite part is now more delicate for if we
replace e by 2e say it will change.
The function x~ is homogeneous of degree a for Re a > -1. This
means that for t > 0
co co
<x~,Cp>= Jxacp(x)dx=t" Jxacp(tx)tdx=t"<x~,CPt>
o 0
where CPt (x) = tcp(tx). The analytic continuation of the two sides must
agree, so
(3.2.7)
a not a negative integer. If a= -k we obtain from (3.2.5)
co
t-k<X~k, CPt> = - J10gxcp(k)(tx)d(tx)/(k-1)!
Ct
o
1
+ cp(k- 1)(0) I/j )/(k -I)!
00 dx
=<x~\CP>+logt! cp(k) (x) (k-1)!'
Hence
(3.2.8)
so the homogeneity is partly lost.

In addition to x~ we shall also have to use the function
x~ =0 if x>O, x~ =Ixl a if x<O,
where a> -1. This is the reflection of x~ with respect to the origin,
<x~, cp> = <x~, <$) where cp(x) = cp( - x),
so it is clear that all we have said about the definition of x~ for arbi-
trary complex a remains true for x~.
By Theorem 3.1.11 the function za, defined in ([>,JR _ as e"logz where

logz is real if zE1R+, has boundary values (x±io)a on the real axis
from the upper and lower half planes. When Re a> 0 it is clear that
(3.2.9)
Now «x ± io)a, cjJ> is for every test function cjJ an entire analytic
function of a since it is the limit of entire analytic functions. Hence
(3.2.9) remains valid when a is not a negative integer. When a --+ - k,
where k is a positive integer, we have by (3.2.4) and the derivation of
(3.2.5)
and since
e h i "=(-1)k(1+ni(a+k)+O(a+k)2) as a--+-k,
we obtain when a --+ - k
e h i " x: + (j~- I)I«k -I)! (a + k))+ n j(j~- I)I(k -I)! --+ ( _l)k X~k.
When a --+ - k in (3.2.9) we must have cancellation of the singular
terms and it follows that
(3.2.10) (x ± iOt k= X~k + ( _l)k X:::k ± n i( _l)k (j~-I) I(k -I)!.
In particular
(3.2.11)
which agrees with Example 3.1.13 if k = 1. In fact (3.2.11) follows for
any k from the case k = 1 by differentiation since
d
(3.2.12) dx (x±io)a=a(x±iO)a-l.
The average of the two distributions in (3.2.10) is sometimes denoted

by ,,!-k, thus
(3.2.10)' ,,!-k = «x + iO)-k +(x - iO)- k)/2 =X~k +( _l)k X:::k.
From (3.2.12) and the obvious fact that x(x±iO)"=(x±io)a+1 we
obtain
(3.2.12)'
By (3.2.5) we have
00
,,!-l(cjJ)=X~ l(cjJ-cP)= - f log x(cjJ'(x) + cjJ'( -x))dx

o
00
= - f (log Ixl)cjJ'(x)dx.
-00
Hence
d
(3.2.13) ~- 1 = dx log Ixl.
Since cfJ(x)-cfJ( -x)=O(x) as x -+ 0, an integration by parts also gives

00
~-l(cfJ)= J(cfJ(x)-cfJ( -x))dx/x=lim J cfJ(x)dx/x.

o .~O Ixl>'
The last integral, where a symmetric neighborhood of the singularity
tending to 0 has been removed, is called a principal value. Thus
(3.2.14) <~-I,cfJ>=lim J cfJ(x)dx/x=PVJcfJ(x)dx/x, cfJEq .
•~o Ixl>'
The problems we have encountered in the discussion of x':. when a
is a negative integer were caused by the factor a in (3.2.2). By a
change of normalizations they can be made to disappear. First note
that (3.2.2)' assumes a particularly simple form if cfJ' = - cfJ, that is, cfJ(x)
=e- x • This is not a function of compact support but it decreases so
fast at + 00 that the proof of (3.2.2)' is valid for it. Set
00
(3.2.15) r(a)= Jxa-1e-xdx, Rea>O,

o
which in our old notation is Ia_ 1 (e-"). Then (3.2.2)' means that
(3.2.16) r(a+l)=ar(a) if Rea>O.
Using (3.2.16) we can extend r(a) analytically to a meromorphic
function in <C with simple poles at the integers ~O, and (3.2.16)
remains valid outside the poles. The residue at an integer - k ~ 0 is
lim (a+k)r(a)= lim r(a+k+ 1)/a(a+ 1) ... (a+k-l)
a--k a--k
=r(I)/( -k) ... (-1)=( -Ink!

which is of course just (3.2.4) with k replaced by k + 1, acting on e- x .
In Section 3.4 we shall prove that
r(a)r(1- a) = nisin (na).
This implies that the r-function has no zeros, so the quotient defined
by
(3.2.17) X':. =x':./r(a+l), Rea>-1
is analytic when Re a > -1. Since (3.2.2)' gives, when combined with
(3.2.16),
(3.2.2)'" x':. (cfJ') = -X':.-I(cfJ)
it is now clear that X':. can be continued analytically to all aE<C so

that dx':.ldx = X':.-l. Noting that X~ = H we obtain
(3.2.17)'
We shall now carry some of the preceding results over to IRn. First
note that if uEL~oc(lRn"o) is homogeneous of degree a, that is, u(tx)
=t"u(x) when X=FO and t>O, then
and conversely this implies that u is homogeneous. If Re a > - n then

u is integrable in a neighborhood of 0 because with polar coordinates
x = rw, Iwl = 1, we have
lu(rw)I=,-Realu(w)1 and dX=r"-ldrdw
where dw is the surface measure on the unit sphere. In that case u

defines a distribution in IR nand (3.2.18) is valid when ¢ E CO'(IRn).
Definition 3.2.2. A distribution u in IR n" 0 is called homogeneous of

degree a if (3.2.18) is valid. If u is a distribution in IRn and (3.2.18) is
valid for all ¢ECO'(lRn) then u is said to be homogeneous of degree a
in IRn.
The problem which we shall discuss is the extension of homo-

geneous distributions from IR n" 0 to IR n, which as we know from the
case n = 1 is not always possible. However, we shall first rephrase
(3.2.18). If we differentiate with respect to t using Theorem 2.1.3 and
put t = 1 it follows that
(3.2.19)
where Jl= 'f.xjO j is the radial vector field. We shall prove that (3.2.19)
implies
00
(3.2.20) u(",)=O if ",ECO'(lRn"o) and Jr"+n-l "'(rx)dr=O;

o
since the integral is a homogeneous function of x of degree - a - n we
may take Ixl = 1 here. (3.2.20) follows from (3.2.19) if we show that the
equation
(a+n)¢+Jl¢='"
has a solution ¢E CO'(lRn" 0). With polar coordinates this equation
can be written
a
- (r"+"¢(rw» = I/I(rw) r"+n-1
ar
so the solution which vanishes for small r is also zero for large r. This
proves (3.2.20), and since
00
S r"+n-1(¢(rx)-t"+n¢(rtx»dr=0, ¢ECg"(lRn"O),
o
we see that (3.2.18), (3.2.19) and (3.2.20) are equivalent. If we note that
<u, A¢> = L <xju, aj¢> = - L <ajXjU, ¢> = -<AU, ¢> -n<u, ¢>
we can also write (3.2.19) in the form
(3.2.19)' Au=au
(Euler's identity for homogeneous "functions"). In particular we find
using Corollary 3.1.5 that homogeneous distributions when n = 1 are
just multiples of ixi a on each half axis. It is also clear that if 1/1 is a
°
homogeneous COO function in lR"" of degree b then I/Iu is homo-
geneous of degree a + b. Since
AajU= ajAU-ajU =(a-1) aju
differentiation lowers the degree of homogeneity by one unit.
Theorem 3.2.3. If uEE!&'(lR"" 0) is homogeneous of degree a, and a is

not an integer ~ -n, then U has a unique extension uEE!&'(lR") which is
homogeneous of degree a. If P is a homogeneous polynomial then (Pu)·
=PU, and if a=F1-n then (aju)"=aju. The map
E!&' (lR" " 0) 3 U --+ UEE!&' (lRn)
is continuous.
Proof a) Uniqueness. The difference between two homogeneous exten-

sions is supported at 0, hence a linear combination of derivatives of
0 0 , Now aa 0 0 is homogeneous of degree - n -iai since
<
<aa8 0 , ¢t> = t"+ lal aa8 0 , ¢ >, ¢E Cg"(lRn),
so a homogeneous distribution with support at

geneous of integer degree ~ - n.
°
must be homo-
b) Existence. If u is a function and ¢ECg"(lRn"O) then

co
<u,4»= S Ju(w)ra +n - 14>(rw)drdw.
1"'1 = 1 0
This suggests that for arbitrary ¢ E Cg"(lR") we should introduce

(3.2.21)
It follows from (3.2.7) that RA> is a homogeneous function of degree

-n-a. By Theorem 2.1.3 Ra is a continuous map from C~(K) to
coo(IRn" 0) for every compact set K c]Rn.
Choose a fixed function IjJEC~(IRn"o) such that
00
(3.2.22) JljJ(tx) dt/t = 1, x =1=0.

o
It suffices to take ljJ(x) as a function of Ixl so that (3.2.22) is valid for
one x. Then IjJRacf>EC~(JR.n"o) and
J ta+n-lljJ(tx)(R.cf»(tx) dt
00
Ra(IjJRacf>)(x) =
o
00
J
=(R.cf>)(x) ljJ(tx) dt/t = (Racf»(x).
o
Hence it follows from (3.2.20) that u(IjJRacf» is always independent of
the choice of IjJ and that u(IjJR a4»=u(4» if cf>EC~(JR.n"o). Thus
(3.2.23) <u, cf» = <u, IjJRacf», 4>EC~(1Rn),
defines a distribution u in 1Rn which extends u. The map u ~ u is

obviously continuous. Since
(R a4>t)(x) = <r"/n- 1, tncf>(rtx» = tn- 1- (a+n- 1) Racf>(x) = t-a Racf>(x)
by (3.2.7) again, it follows that <u,4>t)=t- a<u,cf», so u is homo-

geneous.
Finally we note that (Pu)"-Pu and (Oju)"-OjU are homogeneous
distributions of degree a + degree P and a -1 respectively and sup-
ported by 0, so they must be zero by our hypotheses. This completes
the proof of the theorem.
(3.2.23) stil1 defines an extension u of u if a is an integer - n - k :;;:;

- n, but it may depend on the choice of IjJ then. u may also fail to be
homogeneous, for (3.2.8) gives
(R_n_kcf>t)(x) = <r:;:k-l, t"4>(trx»
=t"-I+k+ 1 «r:;:k-l,4>(rX»-IOgt :~ cf>(rX)/k!lr=o)
= rk+n(R_n_kcf>(X) -log t<Pk(x»
where <Pk is the homogeneous part of degree k in the Taylor expan-
sion of cf>. Thus
(3.2.24) <u,4»=t- k - n <u,cf>t)+logt L <U,xIXIjJ)OIX4>(O)/rx!.
IIXI=k
Any other extension of u is of the form U=u+ ~>aaa(jo where the

sum is finite. Replacing u by U does not change the logarithmic term
in (3.2.24) but introduces a new term
I(1-t JaH )aaaa(jo
which can only be 0 if aa = 0 when lod =1= k. Thus the weakened homo-
geneity property (3.2.24) cannot be improved. In particular there exists
a homogeneous extension if and only if
(3.2.25) <u,xaljl) =0 when 11X1=k= -n-a.
That (3.2.25) is independent of the choice of ljI is clear from the
preceding result but can also be seen as follows. If v is homogeneous
of degree - n in JR."" 0 then v(ljI) is independent of the choice of ljI
satisfying (3.2.22), in virtue of (3.2.20), so we may write
S(v)=<v, ljI)
for such ljI. If v is a continuous function then introducing polar
coordinates gives
S(v)= IS v(w)ljI(wr)dwdrjr= J v(w)dw
JwJ=1 JwJ=1
so S(v) is the integral of v over the unit sphere. (This has nothing to
do with the Euclidean metric really, we could equally well integrate
the Kronecker form
I( _l~-IV(X)xjdxIA ... Adxj_1Adxj+1A ... Adxn
over the C 1 boundary of any neighborhood of 0.) In (3.2.25) xau is
homogeneous of degree - n, so (3.2.25) can be written
(3.2.25)' S(Xa u) =0 when 11X1+degreeu= -no
We can also rewrite (3.2.24) as follows

(3.2.24)' <u,¢)=t-k-n<u,¢t)+logt I s(xau)aa¢(O)jlX!.
JaJ=k
Note that if u is homogeneous of degree a and a is not an integer
which is ~ - n then (3.2.23) can be written in the form
(3.2.23)' <u, ¢) =S(uRa ¢)=S(u <t~+n-l, ¢(t.)), ¢ECO'(IRn),
for uRa¢ is then homogeneous of degree a-n-a= -no However,
when a = - n - k is an integer ~ - n then (3.2.23) depends on the
choice of ljI. Our choice of u with a fixed function ljI guarantees that
(3.2.26) (Pu)"=Pu
if P is a homogeneous polynomial. In fact, if the degree is m then

<t~k- 1, P(tx) ¢(tx) = P(x)<r:.- k- 1, ¢(tx)
by (3.2.1)', hence
<PIi, ¢) = <Ii, P¢) = <u, "'(R_n_kP¢) = <u, ",PR- n+ m - k¢)
=<Pu, ",R_ n+ m _ k¢) = «Pu)', ¢).
(3.2.25) is automatically fulfilled if u is homogeneous of degree a=
°
- n - k where k is an integer ~ and u has parity opposite to k. By
this we mean that (3.2.18) is strengthened to
(3.2.18)' <u,¢)=sgnt t"<u, ¢t) if ¢ECg'(lRn"O),
¢r<x)=tn¢(tx),
for every t =!= 0. It is of course sufficient to assume (3.2.18)' for t = -1
in addition to (3.2.18). If u is a function then (3.2.18)' means for t=-1
that
f u(x)¢(x)dx=( _1)1+a+n f u(x)¢( -x)dx=( _1)I+k f u( -x) ¢(x) dx,
that is, u(x)=(_I)k+lU(_X).
(3.2.18)' always implies (3.2.25). In fact, if '" is even and satisfies
(3.2.22), and if ¢(x)=xa",(x), locl=n, then ¢_I=(-I)n+k¢=(-I)Q¢,
hence (3.2.18)' gives <u, ¢) = - <u, ¢), that is, <u, ¢) = 0. Thus u has a
homogeneous extension. We claim that there is a unique extension Ii
satisfying (3.2.18)' for all ¢ECg'(lRn) and that it is given by
(3.2.23)"
Here f'+n-l is defined by (3.2.10)'. The uniqueness is obvious, for if loci

=k then
so the usually undetermined part of the extension has the wrong

parity. The second part of (3.2.1 0)' gives (recall that a + n = - k)
<.t.- k- 1, ¢(t.) =<t~k-I, ¢(t.)+( _1)k-l¢( -t.).
If U is the extension of u defined by (3.2.23) then (3.2.23)" means that

2<1i, ¢) = < u, ¢) +( _1)k+n-l< U, ¢ -I)' ¢ECg'(lRn).
Hence (3.2.23)" does define a distribution. If ¢ECg'(lRn"O) then the

right-hand .side is equal to 2<U,¢) by (3.2.18)' so Ii=U in lRn"O.
Finally we obtain (3.2.18)' with u replaced by Ii for all ¢ECg'(lRn) and
t= -1, since
Summing up, we have now proved

Theorem 3.2.4. If uE!1d'(lRn"o) is homogeneous of integer degree a=
-n-k~ -n, then u has an extension uE!1d'(lRn) satisfying (3.2.24)'. This
determines u apart from a linear combination of derivatives of order k
of <5 0 , A consistent choice of extension can be made so that (3.2.26) is
fulfilled for every homogeneous polynomial P. A homogeneous extension
exists if and only if (3.2.25)' is valid. If u satisfies (3.2.18)' then there is
a unique extension U with the same property for every ¢EC~(lRn). It is
given by (3.2.23)".
Remark. If U is homogeneous of integer order a = - n - k > - nand

satisfies (3.2.18)' then we also have (3.2.23)" for the unique homo-
geneous extension.
We shall refrain from discussing the difference 0jU-(Oju)' in gener-
al because it depends on the choice of t/I. However, one useful case
where t/I does not matter is the following one.
° °
Theorem 3.2.5. Let U l ' ... , Un E!1d' (IR.n" 0) all be homogeneous of degree
1 - n in IR n" and let I OJ U j = there. Then it follows that
IOjUj =c<5 o, c= IS(ujt/lj)
where t/lix)=x/lxI2, Ixl denoting the Euclidean metric.
Proof We know that IOjU j is homogeneous of degree -n and sup-
ported by 0, so IOjU j =c<5 o for some c. If ¢EC~(lRn) and ¢(O) = 1 then
c= I (OjU j, ¢) = - I (u j, 0j¢)'
Choose ¢(x)=x(lxl) where XEC~(IR) and X=1 near 0. Then
-OJ¢ =t/lj(x) t/I(x) , t/I(x) = -x'(lxl)lxl,
and since
00 00
Jt/I(tx) dt/t= - Jx'(tlxl) Ixl dt= 1, x,*O,

o 0
the condition (3.2.22) is fulfilled so -(uj,Oj¢)=S(t/lju) by the de-

finition of S.
3.3. Some Fundamental Solutions

In Section 3.1 we saw that Cauchy's integral formula is closely related
to the fact that oE/oz=<5 o if E=1/nz. Integration of a function on IR
means convolving it with the Heaviside function H, and dH/dx=<5 o.
These are two examples of fundamental solutions:
Definition 3.3.1. A distribution EE.@'(Rn) is called a fundamental

solution of the differential operator P = I aa oa with constant (com-
plex) coefficients if PE=b o.
Fundamental solutions are very important in the study of exis-

tence and regularity of solutions of differential equations. Such appli-
cations will be discussed in Section 4.4 after we have studied con-
volution of distributions. It will be convenient then to have available
the examples of fundamental solutions which we shall give now.
Theorem 3.3.2. Put E(x)=(2n)-1 loglxl if xER 2 "O and for n>2
E(x) = -lxI 2 - n/(n-2)c n, xERn"o,
where Ixl is the Euclidean norm and Cn the area of the unit sphere. Then
ojE is defined by the locally integrable function xjlxl-n/cn and
n
(3.3.1 ) AE=IOJE=b o·
1
Proof If cfJEC(f(Rn) then

<OjE,cp) = -<E,o/P)= -lim S E(x)ojcp(x)dx
.~o Ixl>'
= Scp(x)ojE(x)dx+lim S E(x)cfJ(x)xilxl dS
.~o Ixl~.
by Gauss' formula. The surface integral is O(e) if n > 2 and O(eJog l/e)
if n = 2 so the limit is 0. Thus ajE is defined by the locally integrable
function ojE(x) which for n > 2 also follows from Theorem 3.2.3. For
x~O we have
so Theorem 3.2.5 and the fact that S(I xJjlxl n+ 2 Cn) = 1 gives
AE= IOjojE=b o.
We could also make this conclusion without appealing to Theorem
3.2.5 :
<AE, cp) = <E, A cp) = lim S (EA cfJ - cfJAE) dx

.~o Ixl>.
= lim S div(E grad cfJ - cfJ grad E) dx
.~O Ixl>.
= lim S <cfJ grad E - E grad cfJ, x/lxl) dS = cfJ(O) .

• ~O Ixl~.
Note that when n=2 we have Ll =40 2/ozoz and

40E/oz=z/lzI 2n = 1/nz
so we get back our old result that 0(1/nz)/oz= boo
Next we consider the heat equation in lRn+ 1 :
Theorem 3.3.3. Denote the variables in IRn+ 1 by (x, t)EIRn x IR and set
E(x, t)=(4nt)-n/2 exp( -lxI 2 /4t), t>O, E(x, t)=O, t~O.
Then E is locally integrable in IRn+l, EECoo(IRn+l"O), and
(3.3.2)
Proof That E is COO in IR n+ 1 ,,0 follows from Corollary 1.1.2 as in the

closely related Example 1.1.3. By (3.4.1)" below the integral of E(x, t)
°
with respect to x is equal to 1 when t > 0, so E is locally integrable
and defines a distribution. When t > we have
oE/oxj = -xjE/2t, LlxE= -nE/2t+lxI 2 E/4t 2 =oE/ot
so (a;ot-Llx)E is supported by 0. When </JEC~ we have
«%t-Llx)E, </J) = -<E, o</J/ot+LlA)

=lim S -E(x, t)(o</J/ot+LlA)dxdt=limS E(x, e) </J(x, e)dx
£-0 t>e £-0
.-0
= lim S E(x, 1) </J(V; x, e) dx = ci>(0)
by bounded convergence. The theorem is proved.
We shall now consider the closely related Schrodinger operator

iO /0 t + Ll x or more generally operators of the form
n
(3.3.2)' L=%t- L AjkojOk
j.k~ 1
where the symmetric matrix A = (A jk) is constant and det A 0. In *

analogy to Theorem 3.3.3 we try to find a fundamental solution of the
form
E(x, t)= ct- n/2 exp( - <Bx, x)/t)
for t > °where B is another symmetric matrix. Then
ojE= -2E(Bx))t, 0jokE= -2EBjk/t+4E(Bx)iBx)k/t2,
LE =(2 TrBA -n/2) E/t+«Bx, x) -4<ABx, Bx») E/t 2.
To make this vanish we must take B so that 4BAB=B, that is, B

= A - 1/4. Since E increases exponentially as t ---> + 0 unless Re B ~ 0 we
must make this assumption. Now Re(Bx,x)=Re(Ay,y) if x=2Ay
so this is equivalent to assuming that Re A ~ O. However, even then
we do not necessarily get a locally integrable function if n> 1, so we
need the following
Theorem 3.3.4. If B is a symmetric non-singular matrix with ReB~O

then
(3.3.3) (nt)-n I 2(detB)t Jr<Bx.x>lt¢(x)dx--->¢(O) as t---> +0,
if ¢ECg'(JRn). For any even integer k~n we have
(3.3.4) t- n/2 lJ r<Bx.X>lt¢(x)dxl~Ck I (JIO'''¢ldx+suplO'''¢I).
lal ;;;k
Here (detB)t is defined as explained in Section 3.4.
We postpone the proof a moment and give an immediate con-

sequence:
Theorem 3.3.5. Let A be a symmetric n x n matrix with Re A ~ 0 and

det A =F O. Then
J(4nt)-n /2(det A)-t dt(J r
00
(3.3.5) (E, ¢) = <A - l x. x >/4t ¢(x, t) dx),

o
¢ECg'(JRn+l), defines a distribution in JRn+ 1 of order ~n+1, and
(3.3.6)
if L is defined by (3.3.2)'.
The proof is an obvious modification of that of Theorem 3.3.3
when Theorem 3.3.4 is available. We therefore leave the proof for the
reader and pass to the main step in the proof of Theorem 3.3.4.
Lemma 3.3.6. If B satisfies the hypotheses of Theorem 3.3.4 then

(3.3.7) IJr<Bx.x>lt¢(x)dxl~Cl I N(oa¢), ¢Ec~j(JRn),
lal;;;; 2j
provided that oa¢(o)=o when lal<2j. Herej=0,1, ... ,0<t<1 and
N(t/I)=suplt/ll + Jlt/lldx.
Proof (3.3.7) is obvious when j=O so we assume j>O and that (3.3.7)
has already been proved for smaller values of j. We can write
¢(X) = I Xk¢k(X)
1
where ¢kEC~j-l, OIZ¢k(O) =0 if I(XI + 1 <2j and

I I N(OIZ¢k)~ C I N(OIZ¢).
k IIZI;;; 2j- 1 IIZI;;; 2j
In 'fact, when Ixi >t we could take ¢~ =xk¢~xI2, and when Ixi < 1 we
could use the functions ¢Z given by Theorem 1.1.9. If XEC~ has
support in the unit ball and X(x) = 1 when Ixi < then to
¢k(X) = X(x) ¢Z(x) +(1- X(x)) ¢t(x)
will have the required properties. Writing C =B- 1 12 we have
xk = I CkiO<Bx, x)lox i ,
so an integration by parts gives
Je- (Bx,x)/t Xk¢k(X) dx = t I J
Cki e- (Bx,x)/t o¢kloXidx.
Since o¢kloXi satisfies the hypotheses of the lemma with j replaced by

j -1, the estimate (3.3.7) now follows by the inductive hypothesis.
Proof of Theorem 3.3.4. First note that the proof of (3.3.7) is valid for
all ¢E c 2 j with derivatives tending fast to 0 at 00 but not necessarily
of compact support. In particular it is valid for functions such as
exp( - <Ax, x») where A is a diagonal matrix with positive diagonal
elements Aj • By (3.4.1.)" we have
Je- (Bx,x)/t e- (AX,x) dx =(n tt/2(det(B + At))-t.
If we differentiate with respect to Aj and put Aj = 1 afterwards, it
follows that
Je- (Bx,x)/t x2lZe-lxl2 dx = O(~/2+ IIZI), t ~ O.
(nt)-n/2 Jr(Bx,x)/te-lxI2dx~(detB)-t, t~O.
In addition
Jr (Bx,x)/t x lZ e-l xI2dx =0 if some (Xj is odd.
Now we write for the function ¢ in Theorem 3.3.4
¢(x) = T(x) r !xl' + tjJ(x)
where T is a polynomial of degree k -1 and tjJ vanishes of order k at
O. This means just that T is the Taylor polynomial of ¢(x) elxl2 . Since
I N(OlZtjJ)~ C I N(OIZ¢)
11Z1;;;k 11Z1;;;k
and the coefficients of T can be estimated by the derivatives of <p at 0,

the estimate (3.3.4) follows if we use (3.3.7) with 2j=k and <p replaced
by 1/1. Using Taylor polynomials of order k + 2 instead, we obtain
(3.3.3).
In Chapter VII we shall see that the preceding results are more
easily accessible by means of the Fourier transformation.
3.4. Evaluation of Some Integrals

To avoid interruption of the main argument we have postponed some
elementary computations to this section. The main point is the study
of the integral
where IxI2=xi+ ... +x;. HCn is the area of the unit sphere sn-lclRn
then introduction of polar coordinates gives
When n=2 this is equal to n so
(3.4.1)
- 00
and it follows that in general

(3.4.2)
When n=3 this gives 4n=2nt/r(t) so 2r(i)=nt, or
(3.4.3) r(t) = nt.
Thus
(3.4.4) c 2n =2nn/(n-l)!, C2n + 1 =2 n+ 1 n n/«2n-l) ... 3.1).
The volume Cn of the unit ball in lRn is cn/n, which follows for
example from Gauss' formula applied to the radial vector field x.
Thus
(3.4.5) C 2n =nn In!, C 2n + 1 = 2n+ 1 nn /«2n + 1)(2n -1) ... 3.1).
From (3.4.1) we obtain by a change of variables
00
(3.4.1)' J e- Qt2
dt=(n/a)t, a>O.
-00
3.4. Evaluation of Some Integrals 85
More generally, if A is a symmetric positive definite n x n matrix, then

(3.4.1 )"
IR"
This is an immediate consequence of (3.4.1)' if A has diagonal form,

and we can always give A diagonal form by an orthogonal transfor-
mation. Now the set H of symmetric matrices A with Re A positive
definite is an open convex set in the n(n + 1)/2 dimensional complex
vector space of symmetric n x n matrices. If A E H then det A =1= 0 since
Ax = 0, x E<C" implies
O=Re<Ax,x)=«ReA)x,x), hence x=O.
Since H is convex it follows that there is a unique analytic branch of
H3A-->(detA)-l: such that (detA)-l:>O when A is real. Both sides of
(3.4.1)" are analytic when A E H so it follows that (3.4.1)" is valid for
all AEH.
(det A)-l: is also uniquely defined by continuity when A is in the
closure of H, for if det A =1= 0 we have two analytic branches of the
square root in a neighborhood of A, only one of which agrees with
the definition chosen in H. We shall compute (det A)-l: when A = iB is
purely imaginary and non-singular. To do so we may assume that B
has diagonal form for a real orthogonal transformation does not
change (detA)-l: when AEH since it does not when A is real. 'Thus we
assume that <Bx,x)=IbjxJ with all bj=l=O. Then
det(iB+d)= Il(s+ib)
and the square root which is positive when all b j vanish has the
argument
!-Iarg(e+ib)
where each term lies between - nl2 and n12. When e --> 0 this con-
verges to
!-I~sgnbj'
Now I sgn b j is the signature sgn B of B, so we have proved that
(3.4.6) (det( iB))-l: = Idet BI-l: exp(n i(sgn B)/4).
Occasionally we shall also need some properties of the r function.
First we observe that if a> 0 and b > 0 then
(3.4.7) X':.-l * x:- 1 =B(a, b) X':.+b- 1
where
1
(3.4.8) B(a,b)=Sta - 1 (I-t)b- 1 dt
o
is called the beta function. Taking the scalar product with e- t in both
sides of (3.4.7) we obtain by the definition of the r function
r(a)r(b) =B(a, b) r(a+b),
that is,
(3.4.9) B(a, b)=r(a)r(b)/r(a+b).
Hence (3.4.7) may be written, with the notation (3.2.17),
(3.4.10)
which will follow by analytic continuation for all a, bE~ when convo-
lution has been defined for distributions in Section 4.2.
Taking t = 1/(1 + s) as new variable we obtain when 0< a < 1
1 ()()
B(a, 1- a) = St"- 1(1_ t)-adt= S s-a(1 + s)- 1ds.
o 0
Integrating s-a/(l+s) in ~ slit along 1R.+, first from R-iO to 0, then

to R + iO and finally along the circle lsi = R to R - iO, we obtain when
R --+ 00
B(a,1-a)(1-e- 27t ;a)=2nie- 7t ;a,
which can be written in the form
(3.4.11) r(a)r(1-a)=B(a,l-a)=n/sin(na).
By analytic continuation this remains valid when a is not an integer.
Notes
In Section 3.1 we just rewrote some basic real and complex analysis in
the language of distribution theory. The discussion of boundary values
of analytic functions will be continued in Chapter VIII where it plays
an important role. The discussion of x~ in Section 3.2 goes back to
Hadamard [1] and to Marcel Riesz [1]. Homogeneous distributions
in several variables were also considered by them, and they were
studied at great length by Gelfand and Silov [2]. The reader can find
more information and examples there. Sections 3.3 and 3.4 are also
entirely classical in contents. For a proof that a product with reason-
able algebraic properties cannot always be d.efined we refer to
Schwartz [3].
Chapter IV. Convolution
Summary
In Section 1.3 we defined the convolution U 1 * u 2 of two continuous

functions U 1 and U 2 , one of which has compact support. The de-
finition can be applied without change if U 1 E~' (resp. Iff/) and u 2 E C;;'
(resp. COO); we have u1 * u 2 E COO then. Section 4.1 is devoted to such
convolutions. As in the case of functions this is an efficient method to
approximate distributions by COO functions. It can often be used to
extend statements concerning smooth functions to distributions, par-
ticularly when translation invariant questions are concerned. As ex-
amples of this we give in Section 4.1 a discussion of convex, sub-
harmonic and plurisubharmonic functions.
Convolution of two distributions U 1 , U 2 one of which has compact
support is defined in Section 4.2 so that the associativity
is preserved. It is then elementary to see that

supp(u 1 *u 2 )cSUPpu 1 +suppu 2 ·
Section 4.3 is devoted to the proof of the theorem of supports which
states that when U 1 and U 2 both have compact supports, then there is
equality if one takes convex hulls of the supports. The standard proofs
of this depend on analytic function theory, and we shall return to
them later on. The reader might therefore prefer to wait for Section
7.3 rather than studying the end of the proof of Theorem 4.3.3.
Section 4.4 is intended to present the basic methods used to derive
results on existence and smoothness of solutions of constant coef-
ficient partial differential equations from the properties of a funda-
mental solution. This is an important application of the convolution.
The final Section 4.5 is then devoted to IJ' estimates for convolutions.
In addition to estimates related to HOlder's inequality we prove po-
tential estimates of the Hardy-Littlewood-Sobolev type and derive
88 IV. Convolution
from them relations between the If (or Holder) classes of a function

and its derivatives. These are basic particularly in the study of elliptic
differential equations. They will be supplemented in Section 7.9 when
we have Fourier analysis at our disposal.
4.1. Convolution with a Smooth Function
The convolution u * ¢ of a distribution uE~'(1Rn) and a function

¢ EC;;' (1Rn) is defined by
(u * ¢)(x) = u(¢(x -.))
where the right-hand side denotes u acting on ¢(x - y) as a function of
y. If u is a continuous function this agrees of course with the previous
definition (1.3.1), and the properties of the convolution proved in Sec-
tion 1.3 remain valid for distributions:
(4.1.1) supp(u * ¢) csupp u+supp ¢.

For any multi-index 0: we have
(4.1.2) o"(u * ¢) =(o"u) * ¢ =u * (o"¢).
Proof It follows from Theorem 2.1.3 that u * ¢E COO and that

o"(u * ¢)=u * o"¢.
This proves (4.1.2) for the second equality in (4.1.2) follows at once
from the definition of o"u. To prove (4.1.1) we note that u*¢(x)=O
unless x - yEsupp ¢ for some yEsupp u, which means that XESUPP u
+ supp ¢. This is a closed set since supp ¢ is compact. The theorem is
proved.
Commutativity of the convolution would not make sense in the

present asymmetric setup, but we shall prove the important asso-
ciativity:
Theorem 4.1.2. If uE~'(1Rn) and ¢, l/!EC;;'(1R n), then

(u*¢)*l/!=u*(¢*l/!).
The proof is an easy consequence of the following
Lemma 4.1.3. If cfJECb(1Rn) and I/IECg(R n), then the Riemann sum
(4.1.3) L cfJ(x-kh)hnl/l(kh)
keZ"
converges to cfJ * I/I(x) in Cb when h -> O.

J'.roof The support of the sum (4.1.3) is contained in the compact set
supp cfJ + supp 1/1. Since the function (x, y) -> cfJ(x - y) I/I(y) is uniformly
continuous, the sum (4.1.3) converges uniformly to cfJ * I/I(x) when
h -> O. Differentiating the sum at most j times we obtain the same
conclusion for the derivatives since aa(cfJ * 1/1) = (aacfJ) * 1/1 when jor:j ~j.
This proves the lemma.
Proof of Theorem 4.1.2. With the assumptions in the theorem we

obtain from Lemma 4.1.3
u * (cfJ * 1/1) (x) = lim u(L cfJ(x - • - kh)hnl/l(kh»
h~O
= lim Du * cfJ)(x -kh)hnl/l(kh)= S(u * cfJ)(x - y)l/I(y)dy

h~O
which proves the statement even for any I/IE cg.
We can now prove an analogue of Theorem 1.3.2 for distributions:
Theorem 4.1.4. Let 0 ~ cfJ E C,(;, ScfJ dx = 1. If u E~' (R n) it follows that u'"
=u*cfJEC"(Rn) and that u",->u in ~'(Rn) as suppcfJ-> {O}.
Proof To clarify the computations we note that u(l/I) = u * t/f(O) if

I/IEC,(;(R n) and t/f(x) = 1/1 ( -x). This gives
u",(I/1) = u'" * t/f(O) = u* cfJ * t/f(O) = u( 1/1 in C'(; as supp cfJ -> {O} so
it follows that u",(I/1) -> u(l/I) as claimed.
Theorem 4.1.4 shows that ~'(Rn) could have been defined by

completion of CO(R n) or even of C<Xl(Rn) in the manner suggested by
examples from physics given at the end of the introduction. This is
also true for ~'(X) if X is any open set in R n:
Theorem 4.1.5. If UE~'(X) there is a sequence UjEC'(;(X) such that

uj -> u in ~'(X).
Proof Choose a sequence XjEC'(;(X) such that on any compact subset

of X we have Xj = 1 for all large j. Then choose cfJjE C'(; (Rn) satisfying
90 IV. Convolution
the hypothesis of Theorem 4.1.4 with so small support that

(4.1.4) supp tPj + supp Xj c: X
and Ixi < I/j if XESUPPtPj. Since Xj uEtC'(X)c:tC'(1R") we can form
uj=(Xju) * tPj
and obtain a function in CO'(X) by (4.1.4) and (4.1.1). If I/IE CO' (X) we
have as in the proof of Theorem 4.1.4
u/I/I) = (Xju)(C$j* 1/1) = u(X/ C$j* 1/1)).
Since sUPP C$j* 1/1 belongs to any neighborhood of supp 1/1 for large j,
we have x/C$j*I/I)=C$j*1/I then, and it follows that u/I/I) --+ u(l/I) as
stated.
Remark. That CO'(X) is dense in !!&'(X) follows also from the Hahn-
Banach theorem since the dual space of !!&'(X) (with the weak to-
pology) is CO'(X) by an elementary fact concerning weak topologies.
Also note that formal rules of computation such as (3.1.4) follow for
distributions by means of Theorem 4.1.5 when they are known for Coo
functions.
If UE!!&'(X) and tPECO'(lRn), the convolution u*tP is defined in
(4.1.5) {x; X-YEX if YESUPPtP}
which is close to X when tP has small support. With obvious modifi-
cations all properties proved above when X =lRn remain valid.
Regularization by convolution can often be used to reduce ques-
tions concerning distributions to smooth functions. We shall give
some important examples.
Theorem 4.1.6. If u, VE!!&'(X) where X is an open interval on lR then
u'~O if and only if u is defined by an increasing function, and V"~O
if and only if v is defined by a convex function, that is, a continuous
function with
(4.1.6) v(tx+(I-t)y)~tv(x)+(I-t)v(y); O<t<l; X,YEX.
Proof a) Assume u, VEC OO • Then u'~O implies

1
U(X+ Y)-U(X)= Y Ju'(x+ty)dt~O if y~O,
o
and since v' is therefore increasing we conclude that
1
(V(X+y)-v(x))ly= JV'(x+ty)dt
o
is an increasing function of y if y ~ O. Hence

v(x+ty)-v(x)~t(v(x+y)-v(x» if O~t~1 and X,X+YEX
which is equivalent to convexity. The converse of these conclusions is
obvious. Also note that ifO~l/I+EC;;'(1R.+), jl/l+dx=l, then
u*l/I:(x)= j U(X-ey)l/I+(y)dy
is an increasing function of x and a decreasing function of e (where it
is defined), while if O~l/IeEC;;'(1R.), jl/ledx=1 and l/Ie is even we have
that
v*l/I~(x)= j V(X-ey)l/Ie(y)dy= j (V(X-eY)+V(X+ey))l/Ie(y)dy
y>O
is a convex function of x and an increasing function of e when e > 0

since
d
dy (v(x+ y)+v(x- y»=v'(x+ y)-v'(x- y)~O if y~O.
b) In general we choose ¢ as in Theorem 4.1.4 and form the

regularizations u'" = u * ¢, vtI> = V*¢. Assume u' ~ 0, v" ~ O. Then u¢
= u' * ¢ ~ 0, v~ = v" * ¢ ~ 0 so uti> * l/I: (x) is an increasing function of x
and a decreasing function of 6 while vtl>*l/I=(x) is a convex function of
x which increases with 6. Letting supp </> -+ {O} we conclude that u* l/I:
and v*l/I~ have the same properties, so when 610 we obtain
u*l/I:ruo, v*l/I~lvo
where Uo is increasing, Vo satisfies (4.1.6), and
<u,X)=juoxdx, <v,X)=jvoxdx, O~XEC;;'(X).
It follows that U o is finite everywhere because it would otherwise be
+ 00 in an interval, and Vo is finite since by (4.1.6) it would otherwise
be - 00 in an interval. Now (4.1.6) implies continuity for'
v(x) - v(x- y) ~(v(x +hy) - v(x))jh ~ v(x + y) - v(x), Ihl < 1,
so U o and Vo have the stated properties. Conversely, if u and v are
defined by increasing and convex functions respectively, then so are uti>
and vtl>. Thus u4,~O and v~~O which gives u'~O and v"~O when
supp </> -+ {O}.
Theorem 4.1.7. If VE.@'(X) where X is open in 1R.n and if

(4.1.7) L LYjYkal\v~O for all YE1R. n,
then v is defined by a continuous function satisfying (4.1.6) on every line
segment in X and conversely. One calls v a convex function.
92 IV. Convolution
Proof We may assume that X is convex. If VEC OO then (4.1.7) means

precisely that
d2
dt 2 v(x+ty)~O when X+tYEX,
so the statement follows from Theorem 4.1.6. If O~t/!EC'f/ is an even

function with Jt/! dx = 1, then
J
v*t/!.(x) = v(x-ey)t/!(y)dy
is also a convex function and it increases with e. If v is just known to
be in .@'(X) we can now argue exactly as in the proof of Theorem
4.1.6. v</> satisfies (4.1.7) so v</>*t/!. is a convex function which increases
with e, hence v*t/!. is convex and increases with e. The decreasing
limit Vo as e!O defines v and satisfies (4.1.6) so Vo is finite everywhere
and upper semicontinuous. This implies continuity since for suf-
ficiently small Iyl
v(x+hy)-v(x)~h(v(x)-v(x- y))~ - Ch, O<h< 1.
The converse is obvious.
We shall now prove an analogue of the second part of Theorem

4.1.6 for several variables.
Theorem 4.1.8. Let X be an open set in JR n• If UE.@'(X) is real and

Au ~O it follows that u is defined by a subharmonic function u o , that is,
an upper semi-continuous function with values in [ - 00, (0) such that
M(x, r)= J uo(x+rw)dw/c n , cn = J dw,

Iwl= 1 Iwl= 1
is an increasing function of r for XEX and O~r<d(x, CX), the distance

from x to ax. Conversely, if Uo is an upper semi-continuous function
with values in [ - 00, (fJ) which is not identically - 00 in any component
of X and if uo(x)~M(x,r) when r<d(x,CX), then uoEL~oc(X) and
L1u~O for the distribution u defined by U o' The function Uo is uniquely
determined by u at every point. If K is a compact subset of X then the
maximum principle is valid,
sup U o = sup u o.
aK K
Proof Assume first that UEC OO and that Au~O. Let O<r<R and set
v(x)=O. Ixl>R; v(x)=e(R)-E(x), r<lxl <R;
v(x)=e(R)-e(r), Ixl <r;
where E is defined in Theorem 3.3.2, E (x) = e(lxl). Then v is con-

tinuous so grad v is the function which is
grad v= -gradE, r<lxl<R

and 0 otherwise. Again by Theorem 3.1.9
Llv=div grad v=< -grad E, -x/lxl)dS R +< -grad E, x/lxl)dS.

=dSR/CnRn-l_dS./cnrn-l
where dS. and dS R are the Euclidean surface measures on the spheres
Ixl=r and Ixl=R. When d(x, CX»R we have since Llu~O and v~O
o~(Llu)*v(x)=u*LI v(x) = M(x, R) - M(x, r)
which proves that M (x, r) is increasing for r > 0, hence for r ~ 0 by the
continuity. Note that if 0 ~ I/IE C~, S1/1 dx = 1, and 1/1 is a function of Ix I
then
u*I/I.(x) = S u(x-ey)l/I(y)dy
is an increasing function of e. This follows if we introduce polar

coordinates.
On the other hand, we have if UE COO
S u(x+rw)dw= S (u(x)+r L WAU(X)+r; L WjWAOkU(X) + O(r 3 ))dw
= cn(u(x) + r2 LI u(x)/2n + O(r 3 )).
In fact, S wjwkdw=O ifj=l=k, and

S wldw=n- 1 SLwldw=cn/n.
Hence
LI u(x)=lim 2n(M(x, r) -u(x))/r2 ~O if u(x)~M(x, r) .
• -0
Now let UE £&' (X) and LI u ~ O. If ¢ satisfies the conditions in

Theorem 4.1.4 then Uq,=U*¢ECoo and Lluq,~O where uq, is defined.
Hence
Uq,*I/1.
is an increasing function of e so letting supp ¢ --> {OJ we conclude that
U * 1/1. is an increasing function of e, and
S (u*I/I.)(x+rw)dw
is an increasing function of r in the interval [0, a) where it is defined.
Hence
94 IV. Convolution
where U o is upper semicontinuous, M(x, r) is an increasing function of

r when O~r<d(x, eX), and
(u,X>=Juo(x)X(x)dx if O~XEC~(X).
Hence uoELloc(X).
Assume now that U o is an upper semicontinuous function in a
connected open set X such that uo(x)~M(x, r) for O<r<d(x, e X) and
U o $ - 00. If uo(x) > - 00 it follows that U o is integrable in the ball
with center x and any radius r<d(x, eX). The open set XocX of
points such that U o is integrable in a neighborhood is therefore closed,
for if XEX is a limit point of Xo then we can find YEX o with uo(Y»
- 00 and Ix - yl <d(y, e X). Hence X 0 =X so uoELloc. If t/I is chosen as
above then
uo(x) ~uo*t/I.(x)
by the mean value property, and
._0
lim uo*t/I.(x) ~ uo(x)
since U o is upper semi continuous. Hence

Uo * t/I .(x) --+ uo(x) as e --+ 0
so U o is determined by the corresponding distribution. Since uo*t/I.
inherits the mean value property from u o' we have
O~ .,1 (u o* t/I.) =(.,1 uo)* t/I. --+ .,1 U o as e --+ O.
When proving the maximum principle we may assume that sup U o
K
=0. Then uo(x)=O for some XEK since U o is upper semicontinuous. If
r is the distance from x to oK then
O=uo(x)~ J uo(x+rw)dw/ J dw.
I"'I~ 1 I"'I~ 1
For some Wo we have x + rWoEoK by the definition of r. If uo(x

+rwo)<O then uo<O in a neighborhood of x+rwo, and since u~O in
K it follows that Juo(x + rw) dw < O. This is a contradiction proving
that uo(x+rwo)=O, hence supuo=O. The proof is complete.
oK
For later reference we also give a property of subharmonic func-

tions with a closely related proof:
Theorem 4.1.9. Let Vj be a sequence oj subharmonic junctions in a

connected open set X cIRn, which have a uniform upper bound on any
compact set. Then
a) if Vj does not converge to - 00 uniformly on every compact set in

X then there is a subsequence vjk which is convergent in L~oc(X).
b) if v is a subharmonic function and vj -t V in E0' (X), then vj -t V in
L\oc(X),
(4.1.8) lim Vj(x)~v(x), XEX,
j~ 00
with the two sides equal and finite almost everywhere. More generally,
(4.1.9) lim sup(Vj-f)~sup(v-f)
j~ 00 K K
for every compact set K e X and every continuous function f on K.
Proof a) By hypothesis one can find jk and x k such that all Xk belong
to a compact subset of X and Vjk(X k) is bounded. We may assume that
x k -t XoEX and to simplify notation that jk = k. If B is a closed ball
eX with center Xo it follows that S vkdx is bounded from below. In
B
fact, for large k there is a closed ball Bk with center at x k such that
BeBkeX and m(Bk)-tm(B). Hence
S vkdx= S vkdx- S vkdx~m(Bk)vk(xk)- S vkdx
B Bk Bk" B Bk" B
is bounded from below, which implies that S IVkl dx is bounded since
B
vk has a uniform upper bound on B. If XEB then the mean value of vk
over a ball with center at x and radius r is an increasing function of r
for which we have a bound when r is small. Hence we obtain a bound
for the Ll norm of vk over any such ball contained in X. The
argument used in the proof of Theorem 4.1.8 to show that uoEL~oc(X)
now gives that Vk is bounded in L~oc(X). Hence there is a subsequence
vjk converging weakly as a measure to a limit v with J v = lim J vik ~ 0
(Theorem 2.1.9). By Theorem 4.1.8 v is a subharmonic function, so b)
will prove that vjk -t v in L~oc'
b) Choose!/J. as in the proof of Theorem 4.1.8. Then
Vj(x)~Vj*!/J.(x) -t v*!/J.(x)
uniformly on compact sets in X as j -t 00, if t5 IS small enough. If
O~XEC;;' then
S(V*!/J.(x) + t:-vix))x(x)dx
-t S(v*!/J.(x)+t:-v(x))X(x)dx, j -tOO,
and if e > 0 the integrand is positive for large j. Hence

lim S Iv-vjIXdx~2 S Iv*!/J,\+e-vlx dx .
J~ 00
96 IV. Convolution
Since e and b are arbitrary it follows that v ~ Vj in L~oc' By Dini's

theorem
sup (v j - f) ~sup (vj*t/I;;-f)
K K
~sup(v*t/I;;-f)~sup(v-f)+e, b<b.,
K K
which proves (4.1.8) and (4.1.9). If O~XECO'(X) we obtain by Fatou's

lemma __
J J
lim vjxdx~lim vjX dx = VX dx J
so using (4.1.8) we conclude that lim vix) = v (x) almost everywhere
when X(x) > O. The proof is complete.
Example 4.1.10. If f is an analytic function in X de then log If I IS

subharmonic and in any component where f $ 0 we have
A log If I=2n L mjb zj ·
where Zj are the zeros and mj the multiplicities.
Proof Near a point where f =FO there is an analytic branch g of logf,
hence log If I= Re g is harmonic. In a neighborhood of Zj we can write
f(z)=(z-z)mjg(z), log If(z)1 =mj log Iz-zjl + log Ig(z)1
where g(Zj)=FO. By Theorem 3.3.2 it follows that A log Ifl=2nmj bzj
there.
If N is a positive integer then log IflllN =N- l log If I is of course

also subharmonic with Laplacean 2n L m)N bz . We can approximate
any measure by such measures so it is quite pla~sible that subharmonic
functions can be approximated by functions of the form log IflllN
in X c<c. We shall have more to say about this topic in Section 15.1.
Theorem 4.1.11. Let X be an open set in <c n • Every real UE.@'(X) such
that
(4.1.10)
can be defined by a plurisubharmonic function uo , that is, an upper
semicontinuous function such that <C3 t ~ u(z + tw) is subharmonic where
it is defined, for arbitrary z, WE<c n • Conversely, every such function U o
which is not identically - 00 in a component of X is in L~oc and defines
a distribution satisfying (4.1.10). The function U o is uniquely determined
by u.
Proof The statement is obvious if UE COO. Note that (4.1.10) gives

L1u~O
if we choose w equal to any basis vector and add. The approximation

by regularization used in the proof of Theorem 4.1.8 is therefore
applicable without change to prove the theorem in general. This is left
as an exercise for the reader.
As an example we shall now discuss u = log If I when f is analytic

$0 in an open set X c:«::n. Near a point where f *0 we can choose an
analytic branch g of logf, so log If I= Re g and (4.1.10) is then zero.
Thus the sum in (4.1.10) is supported by the zero set. Let 0 be a zero,
of order k say. In the Taylor expansion
f=fk+fk+1 + ...
the homogeneous part J" of order k is not identically 0 then, and we
shall compute (4.1.10) near 0 when J,,(wH=O. This is sufficient since a
quadratic form (also with distribution coefficients) is determined by a
finite number of its values. Let w be the Zl axis and set z' =(Z2' ... , zn).
For fixed z'
where Cj(z') are the small zeros of f(C,z')=O and mj the multiplicities,
which are ~k since okflo~ *0. The "counting" measure on the right
depends continuously on z' since log If(z l' z')1 does (in the distribution
topology by Theorem 4.1.9), and for ¢ECj{' with support close to we
have
°
(4.1.11)
where dA is the Lebesgue measure in «::n -1.

Now we distinguish two cases,
(i) m 1=k, C1(z') is analytic and f(Z)=(Zl -C(z,»k g(z), g*O.
(ii) m 1 < k except for z' in a null set.
One of these must occur. In fact, a zero of order k is a zero of
ak- 1f(z)/oz~-l
so the implicit function theorem shows that z 1 = C(z') for some ana-
lytic function C. Now either
f = afloz 1 = ... = ak-1f1az~-1 =0 when z 1 = C(z'),
and then we have case (i), or else one of these functions is not
identically zero and then we have case (ii). In case (ii) we omit from
the integration in (4.1.11) the set where m 1 =k. Repeating the argu-
ment a finite number of times, until the order of the zero cannot be
lowered any longer because one comes across case (i), we obtain
98 IV. Convolution
Theorem 4.1.12. Let f be an analytic function $0 in an open set

X c<c n. Let Z be the set of zeros Zo of f such that in a neighborhood
of Zo we can write
f(z)=g(z)h(z)m
where g(zo)=FO, h(zo)=O and ah(zo)/aZ=FO, and let dS be the Euclidean
surface measure on Z. Then dS is locally integrable in X and
n
(4.1.12) L Wj\1\a 2 10g Ifl/azjMk ="2J.l(w)dS
where near Zo
J.l(w)=ml<w, h')IZ/lh'1 2.
In particular A log Ifl=2nmdS.
Proof It only remains to note that (4.1.12) follows from (4.1.11) in case
(i) above since dS=(1 +la'daz'12)dA(Z').
Every plurisubharmonic function in X c<c n is of course subhar-
monic as a function in Xc IR. 2n, but it has additional properties:
Theorem 4.1.13. If u is plurisubharmonic in X c<c n and J.l is the positive

measure ,1 u, then
r2- 2n dJ.l(z) J
Iz-" <r
is an increasing function of r for 0 < r < d (', eX).
Proof We may assume that '=0. a) Assume first that UECO and that
u is a function of Izl only. Write u(z) = F(lzI2). Then FE Coo and the
plurisubharmonicity of u means that
F'(lzI 2)lwI 2+ F"(lzI 2) I<w, z)1 2 ~O,
that is, F'(s)~O and F'(s)+sF"(s)~O. Thus sF'(s) is a positive increas-
ing function of s. (Note that this condition is independent of n.) Now
,1 u=4(nF'(s) + sF" (s)), s=lzI2,
JJ Audxdy=4 J (nF'(s)+sF"(s))d(C 2n sn)

Izl <r 0
where C 2n is the volume of the unit ball in IR. 2n. Hence

r 2- 2n JJ ,1udxdy=4nC2n r2F'(r2)=4nC2n_2r2 F'(r 2)
Izl<r
where the last equality follows from (3.4.5). This is an increasing
function. We shall use later on that
(4.1.13) ( JJ Au/2ndxdy)/(r2n-2 C 2n _ 2)=2r2F'(r 2).
Izl <r
b) Assume still that UEC OO but not that u is a function of Izl only.
If U is a unitary transformation, then u( U z) is also plurisubharmonic
and the Laplacean is (,1 u)(U z). Hence
v(z)= Ju(U z)dU,
where dU is the Haar measure on the unitary group, is a Coo function
of Izl only, and
J ,1vdxdy= J ,1udxdy
Izl<r Izl<r
so the assertion follows from the case a) already studied. If t/JEC~(1R)
is a decreasing function of r2 it follows that
r 2- 2n J,1ut/J(lxl/r)dx= J -dt/J(R)r 2- 2n J ,1udx
R>O Ixl<Rr
is an increasing function of r.
c) For a general u we form the regularizations u'" according to
Theorem 4.1.4. Then u'" -> u in P)' so ,1 u'" -> ,1 u in P)'. Hence
r2- 2n Jt/J(lxl/r)d1L(x)
is an increasing function of r. Letting t/J increase to the characteristic
function of (-1, 1) we have proved the theorem.
It is customary and convenient to use the normalization
(4.1.14) e(u,r,()=(r2n-2C2n_2)-1 J ,1u/2n
Iz-,I <r
which by Theorem 4.1.13 is an increasing function of r.
Proposition 4.1.14. If u is a plurisubharmonic function such that eU is
homogeneous of degree k~O, then e(u,r,O)=k.
Proof By averaging over the unitary group as in part b) of the proof
of Theorem 4.1.13 we reduce the proof to the case where u is a
function of Izl, thus u(z)=kloglzl+C. Then F(s)=(k/2)logs+C so
e(u, r, 0) = k by (4.1.13).
Theorem 4.1.15. If f is an analytic function in X c(Cn, and not identi-

cally 0 in any component, then
(4.1.15) e(log If I, r, () -> k when r -> 0, if (EX,
where k is the order of the zero at (. If f is a polynomial then the limit
as r -> 00 is the degree of the polynomial.
Proof We may assume that (=0. The definition of k means that
f(z)=fk(z)+O(lzlk+l), z->O,
100 IV. Convolution
where I" is a homogeneous polynomial of degree 0 which IS not

identically O. Put
F.(z) = f(rz)/rk.
Then F. converges locally uniformly to I" as r --+ 0, so log IF.I --+ log II"I
in fi)' in view of Theorem 4.1.9. Hence the same is true for the
Laplaceans and we conclude as in part c) of the proof of Theorem
4.1.13 that when r --+ 0
e(log IF.I, 1,0) --+ e(log Ifkl, 1,0) = k
where the last equality is a consequence of Proposition 4.1.14. Now
e(log IF.I, 1,0) = e(log If I, r, 0)
which follows immediately if log If I is replaced by a smooth approxi-
mation. This proves (4.1.15). If f is a polynomial we can let r --+ 00
instead and obtain as limit of F. the homogeneous part of highest
degree, which proves the last statement.
4.2. Convolution of Distributions

To define the convolution of two distributions we shall use the prop-
erties of the convolution
CO'(IRn)3c/> --+ u*cPEcoo(IR n)
defined in Section 4.1 when UEfi)'(IR n). It is obvious in view of (4.1.2)
that u* c/>j --+ 0 in coo(IRn) if cPj --+ 0 in CO'(IR n). (See Theorems 2.3.1 and
2.1.4 for the definition of convergence.) If hEIRn we define the trans-
lation operator 'h by ('hcP)(x)=cP(x-h) (which is convolution by (\)
and obtain
U*('hc/» = 'h(U* cPl·
Thus u* commutes with translations. Conversely, we have
Theorem 4.2.1. If U is a linear map from CO'(IRn) to C"(IRn) which is

continuous in the sense that U cPj --+ 0 in C(IRn) when cPj --+ 0 in CO'(IR n),
and if U commutes with all translations, then there exists a unique
distribution U such that U cP=u*c/>, c/>ECO'(IR n).
Proof If such a distribution exists we must have u(;P) = U cP(O). (We

recall the notation ;P(x) = c/>( -x).) Now the linear form
C0'3cP --+ (U ;P)(O)
4.2. Convolution of Distributions 101
is by hypothesis a distribution u. From the fact that (U 1/»(O)=(u*I/>)(O)

we obtain by replacing I/> by 'hI/> and using the commutativity with 'h
that
(U 1/>)( - h) =('h U 1/»(0) =(U 'hl/>)(O) = (U*'hl/>)(O) = (u* 1/>)( - h)
which proves that UI/>=u*l/>, I/>EC~(lRn). The proof is complete.
If uE~"(lRn) it follows from (4.1.1) that I/> --+ u*1/> is a continuous

map from C~(lRn) to C~(lRn), that is, sequences converging to 0 are
mapped to other such sequences. The convolution u * I/> is also defined
for arbitrary I/>Ec oo (lR n) then and gives a continuous map from
c oo (lRn) to Coo (lRn).
There is a unique way to define the convolution of two distri-
butions U 1 and u 2 , one of which has compact support, so that the
associativity
(u 1 *U 2 )*I/>=U 1 *(u z *1/»
remains valid for I/>EC~(lRn). In fact, the mapping
C~(lRn)31/> --+ U1 *(uz*l/»
is linear, translation invariant and continuous because it is the com-
position of two such mappings. Hence there is a unique uE~'(lRn)
such that
(4.2.1)
Definition 4.2.2. The convolution u 1 *u 2 of two distributions

U 1 and u 2
one of which has compact support is defined to be the unique distri-
bution u such that (4.2.1) is valid.
By Theorem 4.1.2 the definition is consistent with our original one

when u 2 E C~, and a simple modification of Theorem 4.1.2 shows that
it is also consistent with our earlier definition when u 1 EC'(lRn) and
u 2 Ec oo (lRn). Somewhat more generally we have
Theorem 4.2.3. If U t E~'k(lRn), u 2 ECt(lRn) (or U t EC'k(lRn), u 2 EC k (lRn»

then u t *u 2 is the continuous function x --+ u t (u 2 (x- .».
Proof If this function is denoted by u then the proof of Theorem 4.1.2
shows without change that when t/JEC~, U t E~'k(lRn), u 2 ECt(lRn) then
u*t/J=u t *(uz*t/J)·
This proves the first part of the statement and the other follows in the
same way.
102 IV. Convolution
By its definition the convolution is associative, that is,

u 1*(U 2*U 3)=(U 1*U 2)*U 3
if all the distributions uj except at most one have compact support.
Theorem 4.2.4. The convolution is commutative, that is,
if one of the distributions u 1 , u 2 has compact support. We have

(4.2.2) sUPP(U 1 *U 2 )c:suPpu 1 +suppu 2 .
Proof To prove that two distributions V 1 and V 2 are equal it suffices
to show that
v 1*(¢*I/I)=v 2 *(qJ*I/I) when ¢, I/IEC~.
For then we have (Vl *¢)*I/I=(v 2 *¢)*1/1 by Theorem 4.1.2, hence V 1*¢
=v 2 *¢ and so V 1=v 2 • Now we have
(u 1*u 2 )*(¢*I{!)=u 1*(u 2 *(¢*I{!))=u 1*((u 2 *¢)*I{!)
=U 1*(I{!*(u 2 *¢))=(u 1*I/I)*(u z*¢)
where in addition to Theorem 4.1.2 we have used the commutativity
of convolution of functions. In the same way we obtain
(u 2 *u 1)*(¢* 1/1) = (u 2 *u1)*(I{!* ¢) =(u 2 * ¢)*(u 1* I{!)
=(u 1*u 2 )*(¢*I{!)
which proves that U1*u 2 =U Z *U 1. To prove the last statement we
choose ¢ as in Theorem 4.1.4 and note that since (u 1*u z )*¢
=U 1 *(u z *¢) we have
supp ((u 1 *u 2 )*¢)c:supp U 1 +supp U z +supp ¢
by (4.1.1). When supp ¢ -+ {OJ it follows that (4.2.2) is valid. The
theorem is proved.
If U z has compact support it follows from (4.2.2) that U 1 * u2 is

determined in a neighborhood of x if U 1 is known only in a neigh-
borhood of {x} -supp u z . Hence the convolution U1*u z is defined in
{x; x - YEX for all yESUpp u z }
if u 1 E.@'(X).
Theorem 4.2.5. If U1 and Uz are distributions in IR", one of which has

compact support, then
(4.2.3) sing sUPP (u 1*uz)c:sing supp u 1 + sing supp u2.
4.2. Convolution of Distributions 103
Proof Assume u 2 Etf', choose t/lEC,(? equal to 1 near singsuppu 2, and

set U2 =V 2 +W2 where v 2 =t/lu 2 and w2 =(1-t/l)u 2EC,(? Then
u 1*W 2EC oo and U1*v 2 is a Coo function in
{x; {x} -supp v 2 C Csing supp ud·
This means that
sing supp u 1*u 2 = sing supp u 1* v2 c sing supp u 1+ supp V 2
and since supp v2 c supp t/I can be taken as close to sing supp u 2 as we
wish, we obtain (4.2.3).
Differentiation can be interpreted as a convolution, for we have

(4.2.4) aau=(aabo)*u, uE.@'(IRn).
In fact, using (4.1.2) we obtain for ljJEC,(?(IR n)
(oau) * ljJ = u*(aaljJ) = u*(b o * (aaljJ)) = u*(aabo)*ljJ
which proves (4.2.4). Note in particular that convolution with 15 0 is the
identity operator. If u1 and u2 are two distributions, one of which has
compact support, it follows that
(4.2.5)
In fact, if the differentiations are rewritten as convolutions with aab o
this follows from the associativity and commutativity of the con-
volution. More generally, if
P=~L... aa aa ,
where aaE<C and the sum is finite, is a partial differential operator
with constant coefficients, then
(4.2.4)' U E.@' (IR n),
We shall now prove a local form of Theorem 4.2.3 containing

Theorem 4.2.5.
Theorem 4.2.6. Let U1E.@'(IRn) and U2E&'''(IR n). Then U 1 *U 2 is in C k in a

neighborhood of x if for every yESUpp u 2 one can find an integer j"i?, 0
and an open neighborhood ~ of y such that u 1E.@'j({X} -~) and
U2ECk+j(~) or u 1 ECk+j({x} -~) and U2E.@'j(~).
Proof We may assume in the proof that x=O for otherwise we just
have to make a translation of u 1. If ljJ, t/lE C'(? (~) it follows from
Theorem 4.2.3 and (4.2.5) that (c»Ul)*(t/lu2)EC~(IRn). Choose an open
104 IV. Convolution
covering WI' ... , WN of supp U 2 so fine that W.n W" =1=0 implies
w.uW"ct;, for some y. Let cP.EC;;'(WJ and IcP.=1 in a neigh-
borhood of supp u 2. Then u 2= I cP.U2 and
u 1 *u 2 =((1- I cP,,)U 1)*U2+ I (cP"u 1)*(cP. U2)'

'."
The terms in the sum where supp cP. n supp cP" =1= 0 are in Ck since
supp cP. u supp cP" c t;, for some y. The other terms vanish in a neigh-
borhood of 0 by (4.2.2). This proves the theorem.
The convolution u l *u 2 can be defined in many cases where nei-

ther u l nor u 2 has compact support. What one needs in just that the
map
(4.2.6)
is proper, that is, that the inverse image of each compact set is
compact. This condition is very natural if we look at the convolution
of functions defined by (1.3.1)'. Assume now that uj are distributions
and that (4.2.6) is proper. If X is a bounded open set in lR" it follows
that for some constant C
(4.2.7) XESUPP u l , yESUpp u2 , X+ YEX => Ixi ~ C and Iyl ~ C.
Then the restriction of (cPl Ul)*(cP2U2) to X is independent of the
choice of cPjEC;;'(lR") provided that cPj= 1 in a neighborhood of
{x; Ixi ~ q. In fact, suppose that we change cPl by adding a function
l/J 1 vanishing near this ball. By (4.2.2)
supp (l/J 1 U1 )*(cP2 u 2) csupp (l/J I Ul ) +supp (cP2 u 2)
which contains no point in X by (4.2.7), so
((cPl +l/JI)U l )*(cP2 U2)=(cPl Ul)*(cP2 U2) in X.
We take this as definition of U1 *U 2 in X and note that Theorem 2.2.4
shows that these local definitions together define u l *u 2 in !0'(lR").
More generally, u 1 *u z is defined in X if X is the largest open set such
that (4.2.6) is proper if lR" is replaced by X; and (4.2.2) remains valid.
As a~ example we observe that if uj E!0'(lR), j= 1, 2, and
suppujclR+, then Ul *U 2 is defined and suPP(U l *U 2 )clR+. More
generally, let rclRn be a closed convex cone which is proper in the
sense that it does not contain any straight line. Then
{(x, Y)Er x r, Ix + yl ~ q
is bounded, for if (x j , Y)Er x rand Ix j + Yjl ~ c, IXjl-+ 00, we can pass
to a subsequence such that x)lxjl-+ XEr, hence Y)lxjl-+ -XEr. Then
the straight line 1Rx lies in r which is a contradiction. For every such
cone the convolution of distributions in
{uE.@'(1R"), supp ucr}
is therefore always defined and makes this set an algebra.
4.3. The Theorem of Supports
In this section we shall prove an inclusion opposite to (4.2.2) for the

support of a convolution. This is not possible without restrictions, for
if we take u 1 = 1 and u 2 with Ju 2 dx = 0 then u 1 *u 2 = O. We shall
therefore have to assume that both U 1 and U 2 have compact supports.
Even so, if U 1 is the characteristic function of a bounded open set X,
and supp u 2 cB for some ball B, then
supp u 1 *u 2 c(oX)+B,
which does not contain X + supp u2 if B is small enough. This forces
us to take convex hulls of the supports, and we digress to discuss this
concept before returning to the main topic.
Definition 4.3.1. If E is a compact set in 1R" then chE is the closed

convex hull of E, that is, the intersection of all closed convex sets
containing E.
Equivalently, chE is the closure of the convex set of centers of

gravity
By the Hahn-Banach theorem, if y¢:chE then one can separate y from

chE by a hyperplane <x, 0 =c, so that <x, 0 <c if xEchE but
<y, O>c. Set
(4.3.1) HE(~)=SUp<x, 0 = sup <x, 0, ~E1R".
xeE XEchE
We have seen that y¢:chE implies <y,O>Hd~) for some~, so chE is

the set of all xE1R" such that
(4.3.2)
One calls HE the supporting function of E. Since it is the supremum of
linear functions we have
(4.3.3) HE(~+'1) ~HEm + H E('1), HE(t~)= tHE(~);
t~O, ~, '1E1R".
106 IV. Convolution
HEm < 00 for every ~ E 1R" since E is bounded, and it follows from
(4.3.3) that HE is convex.
Theorem 4.3.2. For every convex positively homogeneous function H in

1R" there is precisely one convex compact set K such that H = H K' in
fact,
(4.3.4)
Proof If K is convex and H = H K we have already proved that K

must be given by (4.3.4). All that remains is therefore to prove that
HK=H if K is defined by (4.3.4). By the definition of K we have HK~H.
Set
G is a closed convex set since H is convex and therefore continuous. If

t/E1R" it follows that there is a half space containing G with (t/, H(t/))
on its boundary. Thus we have for some (y, t)E1Rn+ 1" 0 and aE1R
<y,O+t't'~a if (~,'t')EG; <y,t/)+tH(t/)=a.

Here 't' can be replaced by any larger number so t ~ O. If t = 0 we
would obtain <y, 0 ~ a for every ~ E 1R" which is impossible since y 0
then. Hence t>O. Since H is positively homogeneous, the set G is
*
invariant under multiplication by positive scalars which implies that
a~O and that
<ylt,O+H(~)~O, ~E1R".
Hence X= - yitEK and

H K(t/) ~ <x, t/) = H(t/) - alt ~ H(t/).
This completes the proof.
If K 1 and K2 are compact sets in 1R", the supporting function of

the sum
is obviously given by
HKl+K2=HKl +HK2 ·
If K is a compact set and tE1R, we set
tK = {tx; xEK}
and obtain HtK(~)=HK(t~), that is,
HtK(~)=tHKm if t>O, HtK(~)= -tH K( -~) if t<O.
Finally, if K" is any family of compact sets contained in a fIxed

compact set and K is the closure of the union, then
HK=supHK«'
"
Conversely, if the right-hand side is fInite for every ~, it is the
supporting function of a convex compact set K' containing K" for
every IX, and the equality is therefore valid. Thus there is an easy way
to translate operations on sets to operations on supporting functions.
Now we state the main result in this section, the theorem of
supports:
(4.3.5)
Proof. That the left-hand side is contained in the right-hand side

follows from (4.2.2) so it suffIces to show that
(4.3.6)
In doing so we may assume that UjECO'. In fact, if ljJ is chosen as in
Theorem 4.1.4, and (4.3.6) is known for the convolution of smooth
functions, then
supp(u 1 * ljJ) + supp(u 2 * ljJ) c: ch supp(u 1 * ¢ * u 2 * ¢)
c: ch supp(u 1 * u 2) + ch supp ¢ * ¢,
and as supp ¢ - to} we obtain (4.3.6). We shall now prove that (4.3.6)
follows for U jE CO' if we assume the special case
(4.3.7) 2suppuc:chsuppu*u, UECO'(IR n)
which will be proved afterwards.
Set for fixed U 1, u 2 ECO'(IRn) and j = 0,1, ...

Kj=ch U
degp'P2;;ij
SUPP(P1 UI)* (P2 U 2)
where PI' P2 are polynomials. Then Ko=chsuppu l *U 2' and (4.3.6)

will follow if we prove that Kj=K o for every j. In fact, this means
that
SUPP(PI U1)*(P2 u 2)c:K o
for arbitrary polynomials Pj' hence for arbitrary entire analytic func-
tions Pj since they are limits of their Taylor polynomials. Now take
Pj(x)=E(x-x j , t)
108 IV. Convolution
where E is defined as in Theorem 3.3.3 and Xj are points with

uj(xj)=!=O. When t'\,O it follows then that
supp U 1 (x 1) U 2 (X 2 ) (\, * (jX2 c K o,
hence suppu 1 +suppu z cK o.
To prove that Kj=K o for every j it suffices to show that
(4.3.8) 2K j cK j _ 1+Kj+1' j>O.
In fact, if H j is the supporting function of K j , we obtain from (4.3.8)
Hj-Hj_1~Hj+1-Hj' j>O,
hence Hj+k"?;H j _ 1 +(k+1)(H j -H j _ 1 ). Since H j+k is bounded by the
sum of the supporting functions of supp U 1 and of supp u z , it follows
that Hj-Hj_1~O. But Hj_l~Hj since K j _ 1cK j , so H j =Hj_ 1 for
every j > 0 as claimed.
(4.3.8) means that if deg P1 + deg Pz ~j then
2SUPP(P1 U1)*(P2 U2)cK j _ 1+Kj+1
or by (4.3.7) that
(4.3.9) SUPP(P1 u 1) * (P2 u 2) * (P1 u 1) * (pz u 2) c K j-1 + Kj+ l'
When proving (4.3.9) we may assume that P1(X)=x vq1(X), If we write
q 2(X) = xvpz(x) then
Xv((q 1 U1) * (pz U 2 )) =(P1 U 1) * (pz U 2) +(q 1 U 1) * (qz u 2 )
has support in K j _ 1 , so the convolution with (P 1 U 1 )*(P Z u z) has sup-
port in K j _ 1 +K j cK j _ 1+K j +1. Moreover,
(q 1U 1) * (qz U z) * (Pl U 1) * (pz U z) =((q 1U 1) * (pz U z)) * ((P1 U1 ) * (qz u z))
has support in K j _ 1 + Kj+ l ' so this completes the proof of (4.3.9).
The remaining proof of (4.3.7) is based on the following
Lemma 4.3.4. If UE Cg'(1Rn) and u(x) = u( - x) then

(4.3.10)
Proof For any gECg' we have
(4.3.11) Ilgll i2 = g* g(O)
so both sides of (4.3.10) are equal to U * u * U * u(O).
Lemma 4.3.5. If K is a compact set in IRn then

(4.3.12) suplul~Cllai ... a;uIIL2' uECg'(K).
Proof By Taylor's formula applied to each variable we have
u(x)= S (X 1-Y1)···(X n-Yn)oi···o;u(y)dy

Yj<Xj
End of proof of Theorem 4.3.3. Combination of (4.3.10) and (4.3.12)

gives in view of (4.3.11)
Il u lli2 ~suplu * ill ~ C1!(01" .on u) * (0 1" .0n u)llL2
=CI!oi ... o;u*ullL2, uEC~(K).
Now replace u by u~,
u~(x)=e(X'~)u(x)
and note that
u~ * u~(x) = e(X,O u * u(x),
oa(u~ * u~)(x)=e(X,O(~ + onu * u)(x).
Then we obtain
Se2(X,Olu(xWdx~ C' L 1~12n-lal(S e2(x'~)loau * ul 2dx)t
lal;;;2n
~ C"(1 + 1~lfneH(~)
where H is the supporting function of supp u * u. If we replace ~ by t ~
and let t --'> + 00, it follows that
2<x, O~Hm if u(x)=l=O.
Hence 2<x, O~Hm when XESUPPU, which proves (4.3.7) and Theo-
rem 4.3.3.
4.4, The Role of Fundamental Solutions
Let us first recall that a differential operator with constant coefficients

in 1R.n is a finite sum
P=Laa oa
with aaE<C. According to Definition 3.3.1 a distribution EE~'(1R.n) is
called a fundamental solution of P if
(4.4.1 )
Fundamental solutions were constructed for some operators in Sec-
tion 3.3. We shall prove later (Theorem 7.3.10) that every operator has
a fundamental solution.
110 IV. Convolution
The importance of fundamental solutions is due to the following

two consequences of (4.2.5)':
(4.4.2) E*(Pu)=U, UEC'(JRn),
(4.4.3) P(E* f)= f, fEC'(JRn).
Thus convolution with E is both a left and a right inverse of P. From
(4.4.3) it follows that the equation Pu = f has a solution for every
fEC'(IR n), and (4.4.2) makes it possible to obtain information on say
the singularities of U from those of Pu.
Theorem 4.4.1. If P has a fundamental solution with sing supp E = {OJ

and X is any open set in JRn, then
(4.4.4) sing supp U= sing supp Pu, UE~'(X).
Proof It is always true that sing supp Pu c sing supp u. If UEC' we

obtain from (4.4.2) and Theorem 4.2.5 that
sing supp U= sing supp E * (Pu) c sing supp Pu
so the assertion is valid when UEC'. If I{!EC~(X) is equal to 1 10 an
open subset Y, it follows that
Y n sing supp Pu= Y nsingsupp P(I{!u)
= Y nsingsupp I{!u= Y nsing supp U
Examples of operators for which Theorem 4.4.1 is applicable are

the Cauchy-Riemann, Laplace and heat operators for which funda-
mental solutions with singularity only at 0 were given in Section 3.3.
Note that if we take X =JR n and u equal to a fundamental solution E,
we see that (4.4.4) implies that singsuppE={O}. In Section 11.1 we
shall characterize the operators to which Theorem 4.4.1 is applicable;
they are called hypoelliptic. By repeating the proof of Theorem 4.4.1
we shall now show that an analogue of the StieItjes-Vitali theorem is
valid for all of them:
Theorem 4.4.2. Let P satisfy the hypothesis in Theorem 4.4.1. If

UjEC"'(X), Puj=O, and uj"-+u in ~'(X) when j-+ ro, then UEC"'(X) and
uj-+u in C"'(X).
Proof Since Pu=limPuj=O we have UEC"'(X) by Theorem 4.4.1. We

may therefore assume that U= O. If I{! is chosen as in the proof of
Theorem 4.4.1 then !j=P(I{!u)-+O in C' and supp!jcsuppdl{!cX'- Y.
On a compact set KeY we have

o' uj(x) = o' E* Jj(x) = Jj(o· E(x - .» --+0 uniformly as j --+ 00,
For the Cauchy-Riemann and the Laplace operators we have a

fundamental solution E which is real analytic in ]Rn, {O}, that is,
analytic in a neighborhood in (Cn. One can then improve Theorem
4.4.2:
Theorem 4.4.3. Assume that P has a fundamental solution E which is

real analytic in ]Rn, O. For every open X e]Rn one can then find an
open neighborhood Z e (Cn such that every solution in X of the equation
Pu=O can be extended to an analytic function in Z. If Puj=O in X
and uj--+u in .@'(X), then the extension of uj-u to Z converges uni-
formly to 0 on compact sets in Z.
Proof Let Zo be an open set in (Cn such that Zon]Rn=]Rn, {O} and E
is analytic in ZOo Choose Y, t/J as in the proof of Theorem 4.4.1 and
set f = P(t/Ju). Then
u(x)=E*f(x)=SE(x-y)f(y)dy, XEY,
and the right-hand side is defined and analytic in
(4.4.5) {z; RezEY and Z-YEZ o if YEsuppdt/J}
which is a neighborhood of Y in (Cn. We can choose Zo so that
x+iYEZo=>x+itYEZ o, O~t~l,
which implies the same property for the set (4.4.5). An analytic func-
tion in the set (4.4.5) is then uniquely determined by its restriction to
real arguments, so letting t/J and Y vary we obtain an analytic con-
tinuation of u to the union Z of all the open sets (4.4.5). This proves
the first statement and the second one follows in the same way.
In Chapters VII and VIII we shall show that Theorem 4.4.3 is

applicable if and only if P is elliptic, that is, m denoting the order of P
L a.~·=l=O, O=l=~E]Rn.
1·I=m
Corollary 4.4.4. Assume that P satisfies the hypothesis in Theorem 4.4.3

and that X is a connected open set in ]Rn. If UE.@'(X) and Pu=O it
follows then that u=O if u vanishes in an open non-empty subset Yof
X.
112 IV. Convolution
Proof This follows by the uniqueness of analytic continuation.
We can now prove an extension of the classical approximation

theorem of Runge:
Theorem 4.4.5. Assume that P satisfies the hypothesis in Theorem 4.4.3,

and let Y eX eRn be open sets such that X" Y is not a disjoint union
FuK where K is compact and non-empty, and F is closed in X. Every
solution UECOO(y) of the equation Pu=O is then a limit in COO(y) of
restrictions to Y of solutions of the same equation in X.
Proof By the Hahn-Banach theorem we must show that if WEC'(Y) is

orthogonal to all solutions of the equation Pu = 0 in X, then W is
orthogonal to all solutions in Y. Set
<E, ¢>=<E, (p)
where ($(x)=¢( -x). Thus E(x)=E( -x) for X=FO so E is analytic then.
We have t P E= {) where t P = ~:aa( - 0)(1. is the transpose of P, that is,
<tpv, u> = <v, Pu> if VEt3"'(Y), UE COO(y).
Now set
Then tpv=w. If we show that VEt3"'(Y), then

<w, u> = <tpv, u> =<v, Pu> =0
if Pu = 0 in Y so the theorem will be proved.
v is an analytic function outside M = supp w. If xfj: X then
oav(x)=;<o~E(. -x), w>=O
for every IX since
~o~E(y-x)=O if YEX.
Hence v = 0 in every component 0 of CM which contains some point
in CX. If 0 is a component of CM which is contained in X and is
bounded, then K = 0 n (X" Y) eX" Y is bounded and closed, because
OOeOMeYeX. Hence K is compact and F=COn(X" Y) is closed
in X so K = 0, hence 0 e Y. It follows that v = 0 in every component of
CM which is not contained in Y, with the exception of the unbounded
component when it is a subset of X. The theorem is therefore proved
unless X contains a neighborhood of infinity.
By repeated use of the part of the theorem proved now we con-
clude that every solution of Pu = 0 in a ball B is the limit of solutions
u j in Rn. In fact, if B=BoeBl cB 2 c ... is a sequence of concentric
balls with radius ~ 00, then we can find uJ with PuJ =0 in Bl so that
uJ ~ u in B, and successively solutions u~ in Bk such that
\u~+1-u~\<2-j-k in B k_ 1.
It follows then that
exists in JR" and that Puj=O. Moreover,

\u.-u
J
1 \ <2-j
J =
in Bo
which proves the assertion.
Returning to the proof above we find that if \y \< R in supp w then
v(x)=<E(. -x), w)=O, \x\<R,
because y~E(y-x) is the limit in C""({y; \y\ <R}) of solutions in ]R.n,
which are orthogonal to w by assumption. Hence v = 0 in the un-
bounded component also, which completes the proof.
Remark. The approximation theorem is not valid if the hypothesis on

X'-.. Y in Theorem 4.4.5 is not fulfilled, for a sequence of solutions u j
in X which converges in Y must also converge in the open set Y uK
=X'-..F then. The reason is that we can take ¢EC'fj(X'-..F) equal to 1
a
in a neighborhood of K and obtain in a
uj=¢uj=E*}j
where }j=P(¢u) has support in suppd¢cX'-..(FuK)=Y, so uj con-
verges to a solution in XuO. However, if xoEK the solution u(y)=
E(y-x o) of the equation Pu=O in Y cannot have an extension U to
YuK, for w=u-U would then vanish in YuK'-..{xo} which is ab-
surd since Pw=(jxo in YuK. (We assume that P is not a constant.)
Finally we shall prove an existence theorem which exploits both
(4.4.2) and (4.4.3):
Theorem 4.4.6. Let P satisfy the hypothesis in Theorem 4.4.3, let X be

an open set in ]R.n and fE~'(X). Then one can find UE~'(X) so that
Pu=f.
Proof Let Xj be the set of points XEX with \x\ <j and distance > I/j
to CX. Then the hypothesis in Theorem 4.4.5 is fulfilled for Y = X j. In
fact, for a point XEK the distance to a point YEC X must exceed I/j
since the segment between y and x contains points in Xj' and \x\ <j
since x lies on an interval with end points in Xj. Choose ¢jEC'fj(X)
equal to 1 in Xj and set vj=E*(¢J). Then (4.4.3) gives
Pvj=¢J=f in Xj.
114 IV. Convolution
We must now correct Vj to obtain a limit as j -) 00. To do so we

observe that vj +i -v j satisfies the equation P(Vj+i-V)=O in Xj. By
Theorem 4.4.5 we can therefore find WjEC<Xl(X) with Pwj=O so that
IVj+i-Vj-wjl<2-j in X j _ i .
Then
u= lim (Vj- L Wi)
j-oo i<j
exists in ~'(X) and satisfies the equation Pu = f In fact, on X k we

have when j>k+ 1
j-i k
vj - L Wi= L (Vi+i-Vi-Wi)+Vk+i - Lw;.

i<j k+i 1
The sum converges uniformly on X k to a solution v of the equation

Pv=O. This completes the proof, for Pvk +1 = fin X k.
It is now easy to prove that all distributions are of the form (2.1.1):
Theorem 4.4.7. For every fE~'(X), Xc1R n, one can find faEC(X) such
that
f(¢)= LSfaoa¢dx, ¢EC~(X),
and the sets suppfa are locally finite. If fE~; the sum can be taken
finite.
Proof Since x~/m! is a fundamental solution of (d/dx)m+ 1 on 1R, if m

is a non-negative integer, the product
E(x) =x7+ ... x:+/(m!)n
is a fundamental solution of P=(Ol ... 0n)m+i in 1Rn, and EECk ifm>k.
If f Ec9"k(1Rn) it follows that u = E *f E C(1R n) satisfies the equation Pu
=f To prove the theorem we proceed as in the proof of Theorem
2.1.5, choosing first a partition of unity l=L.pj in X. Let XjEC~(X)
be equal to 1 near the support of .pj. If mj is larger than the order of
xjf we have just seen that one can find UjE C(1Rn) such that
(oi ... ontj+ 1 U j = XJ It follows that
(J. ¢)= L (Xj!. .pj¢) = L (U j, ±(Ol ... 0n)m j+i(.pj¢).

If we carry out the differentiations we obtain the desired repre-
sentation.
As an application of Theorem 4.4.7 we shall prove an extension of

Corollary 3.1.6:
Theorem 4.4.8. Let X = Y x I c]R" where Y is an open set in 1Rn - 1 and

I is an open interval on 1R, and assume that UE~'(X) satisfies a
differential equation of the form
0;:'u+a m _ 10;:'-1 u+ ... +aou= f
where a j is a differential operator in x' =(x 1, ... , x n_ 1) with coefficients
in C<Xl(X) and f is a continuous function of xnE1 with values in ~'(Y).
Then it follows that u is a c m function of xnE1 with values in ~'(Y).
Proof If we allow u to be vector valued and the coefficients to be

square matrices the proof reduces to the case m = 1 just as that of
Corollary 3.1.6. We assume this from now on. Shrinking Y and I if
necessary we can by Theorem 4.4.7 write u as a finite sum
(4.4.6) u= L o~u_
I-I ~/1.
where U~EC(X), and the proof of Theorem 4.4.7 also shows that
f = L O~ fa, faE C(X).
1~1~/1.'.~n=O
(We just regard xn as a parameter and apply the proof in the other
variables.) If u~ = 0 when IXn > 0 then u and on u = f - a o u are continuous
with values in ~'(Y), Otherwise let v be the smallest integer such that
u~=O when IlXnl>v. Then we have
0n u = f -aou= Lo~f_- Laoo-u~= L o~v~=on LO~ ~

where v~ and ~ are continuous functions and v. = 0 when IXn> v, ~ = 0
when IXn> V -1. In fact, we obtain the expression involving v~ by
commuting the coefficients of ao through the differentiations suc-
cessively; recall that ao only involves x' derivatives. To pass to the
representation with ~ we just take out a factor an if IXn > 0 and take a
primitive function of v~ with respect to Xn if IXn = 0, If w = u - L o~ ~ we
have 0nw=O. Hence
w= LO~w_(x')
by Theorem 3.1.4' and Theorem 4.4.7. Now we obtain
U= LO~w<x(x')+ L0<X~
where Q(n ~ V -1 in the sum. Iterating the argument v times we obtain
a similar representation where IXn=O in every term, and the theorem is
proved.
If there exists a representation (4.4.6) of u with U<x continuous in

Y x I the proof gives more:
116 IV. Convolution
Theorem 4.4.8'. Assume in addition to the hypotheses in Theorem 4.4.8

that u can be extended to a distribution in Y x J where J is an open
neighborhood of I, and that f is continuous in I with values in £&'(Y).
Then u is in cm in I with values in £&'(Y).
Note that Theorem 3.1.11 is essentially only the special case of

solutions of the Cauchy-Riemann equation. The importance of Theo-
rem 4.4.8' is that it allows us to interpret boundary conditions such as
u=u o when xn=xO
if Xo is an end point of I and UoE£&'(Y). The theorem also gives us a
unique way of extending u to Y x 1R by defining u as a function of xn
to be 0 outside I.
4.5. Basic IJ Estimates for Convolutions

In this section we shall use the notation
Ilullp=(J lu(x)IPdx)l/P
if uEI1'(1Rn ), l;;:;;p < 00. To avoid convergence questions we shall usu-
ally assume UEC o' We write Ilulloo=suplul then. The following IS
essentially Holder's inequality, which is really the special case k = 2.
Theorem 4.5.1. If u 1, ... , UkEC o then

(4.5.1 )
if
(4.5.2)
Proof Let us first consider the exceptional case when Pj = 00 for some
j, say Pl=oo. Then Pj=l whenj=l=l and (4.5.1) is obvious. Writing
l/Pj=tj and vj=lujlPi it is otherwise clear that (4.5.1) is equivalent to
the inequality
(4.5.3) v~'* ... *v~k(O);;:;;l when O;;:;;VjEC O' Jv j dx=l,
where 0;;:;; tj;;:;; 1 and ~)j = k -1. The left-hand side is a convex func-
tion of t since IlVj(xli = exp (~)j log v/x) is convex. We have
(t1' ... , t k)= L (1- t)'ej
4.5. Basic 11 Estimates for Convolutions 117
where ej is the vector with r

coordinate 0 and the others equal to 1.
Since 0 ~ 1- tj ~ 1 and L (1- t) = 1, the inequality follows from the
cases where t = ej which were discussed at the beginning of the proof.
Corollary 4.5.2. If 1 ~ Pj ~ 00, j = 1, ... , k, and

l/Pl + ... +1/Pk=k-1 + l/q, 1 ~q~ 00,
then
(4.5.4)
Proof If u is the convolution on the left and VEC o, then

lu* v(O)1 ~ IIu 1 1lpl ... Iluk lip. Ilv Il q" l/q + l/q' = 1,
by (4.5.1) with k + 1 factors now. The converse of Holder's inequality
gives (4.5.4).
In particular, we have
(4.5.5)
if 1 ~p~q~ 00 and kE~, l/r= 1 + 1/q-1/p. If q= 00 then the converse

of Holder's inequality shows that (4.5.5) cannot be valid unless kE~.
The same conclusion follows for P= 1 if we observe that by Holder's
inequality and its converse (4.5.5) is equivalent to
lu*v*k(O)1 ~ C IlulipIlvllq,
which means that (4.5.5) remains valid with q and P replaced by P'
and q' if l/p + l/p' = 1, l/q + l/q' = 1. This means in particular that we
cannot apply our result (4.5.5) to any homogeneous k when r< 00, for
if say
(4.5.6) k(y) = lyl-n1a, yEIR",
then Jk(y)'dy diverges at 00 ifr/a~l and at 0 ifr/a~1. However, we
shall now prove the Hardy-Littlewood-Sobolev inequality which states
that (4.5.5) remains valid for this k as if k were in La, except for the
extreme cases where we know that it must fail:
Theorem 4.5.3. If l<a<oo and l<p<q<oo,

(4.5.7) l/p+ l/a= 1 + l/q,
then, with ka defined by (4.5.6),
(4.5.5)' Ilka*ullq~ Cp,a Ilullp, UEC o'
The proof will be based on a few lemmas. We write 1/a+1/a'=1.
llS IV. Convolution
Lemma 4.5.4. If 1 ~p<a' then

(4.5.8)
Proof For every R > 0 we have

Ika*u(x)l~ S lyl-nlalu(x-y)ldy+ S lyl-nlalu(x-y)ldy
/y/<R /y/>R
~ qRn-nla IlulL", + Rnlp' -nla Ilull p),
for
S lyl-np'lady= C 1 R n- np'la
/y/>R
for a finite C1 since p' > a. If we balance the terms by choosing R so

that Rnlp= Ilullp/liulloo then W la ' = Ilull:/a' Ilull~pla' and (4.5.8) follows.
Next we need a fundamental covering lemma of Calderon and

Zygmund:
Lemma 4.5.5. Let uELl(]R.n) and let s be a number >0. Then we can
write
(4.5.9)
where all terms are in Ll,
00
(4.5.10) Ilvlll +I Ilwklll ~31Iulll'

1
(4.5.11) Iv(x)I~2ns almost everywhere,

and for certain disjoint cubes I k
(4.5.12)
00
(4.5.13) sI m(lk)~ Ilu11 1 •

1
If u has compact support, the supports of v and all Wk are contained in

a fixed compact set.
Proof Divide the whole space ]R.n into a mesh of cubes of volume
> s- 1 Slui dx. The mean value of lui over every cube is thus < s.
Divide each cube into 2n equal cubes, and let 111' 1 12 , 113 , ... be those
(open) cubes so obtained over which the mean value of lui is ~s. We
have
(4.5.14) sm(llk)~ S luldx<2 nsm(llk)'
II.
4.5. Basic 11 Estimates for Convolutions 119
For if Ilk was obtained by subdivision of the cube I, the construction

gives
sm(Ilk)~ J luldx~J luldx<sm(I)=2n sm(Ilk)'
Ilk I
We set
(4.5.15) v(x) = J udyjm(I lk)'
Ilk
W Ik(X) = U(X) - V(X),

=0,
Next we make a new subdivision of the cubes which are not
among the cubes I I k' select those new cubes 121 , 122 "" over which
the mean value of lui is ~s and extend the definitions (4.5.15) to these
cubes. Continuing in this way we obtain disjoint cubes Ijk and func-
tions wjk ' which we rearrange as a sequence. If the definition of v is
completed by setting v(x)=u(x) when xrt0= U I k, it is clear that
(4.5.9) holds. To prove (4.5.10) we first note that
J(Ivl+lwkl)dx~3 J luldx.
Ik Ik
Since the cubes are disjoint, w k vanishes outside I k and v = u in CO, we

immediately get (4.5.10). Further, (4.5.11) follows from (4.5.14) if XEO.
On the other hand, if xrtO, there are arbitrarily small cubes contain-
ing x over which the mean value of lui is < s. Hence lu(x)1 ~ s at every
Lebesgue point in CO, that is, almost everywhere. (4.5.12) follows
from the construction, and adding the inequalities (4.5.14) we obtain
(4.5.13) since the cubes are disjoint. The proof is complete.
The reason for wanting W k to have integral 0 is the following lemma:
Lemma 4.5.6. If WELl has support in a cube I, Jwdx=O, and if 1* is

the doubled cube, with the same center and twice the side, then
(4.5.16) (J Ika*wladx)l/a~ Ca Ilwll l ·

CI <
Proof We may assume that the center of I is at 0, and we denote the

side by L. By the mean value theorem
Ik a* w(x)1 = lJ(ka(x - y)- ka(x)) w(y)dYI ~ CLlxl- l - n/a Ilwll l , xrtI*,
and this proves (4.5.16) since
(J Ixl-a-ndx)l/a~ C/L.
Ixl>L
120 IV. Convolution
Lemma 4.5.7. The operator k.* is of weak type 1, a in the sense that
(4.5.17) m{x; Ik.*u(x)l>t}t"~C.llull~, UEV.
Proof Assume that I u 111 = 1 and decompose u by means of Lemma

4.5.5. Then we have by (4.5.8), with p = 1 now,
Ik.*vl~CS1/•.
Define s so that Cs 1/·=t/2. Then Ik.*u(x)1 >t implies I Ik.*wkl >t/2.

I:
If 0= U we have by (4.5.13), (4.5.10) and (4.5.16)
m(0)~2n/s, S (I Ik.*Wkl)·dx~ c,
co
which means that
m{x; I Ik.*wkl > t/2} ~2n/s + C(t/2)-· < C't-·
so (4.5.17) is proved.
(4.5.17) is a substitute for (4.5.5)' which remains true when p=1.
Using an argument of Marcinkiewicz we shall now prove that to-
gether with Lemma 4.5.4 it proves Theorem 4.5.3:
Proof of Theorem 4.5.3. Assume that Ilullp= 1. If

m(t)=m{x; Ik.*u(x)l>t}
then 00 00
Ilk.*ull:= - S fldm(t)=q S fl- 1m(t)dt

o 0
so we must estimate met). With a number s to be chosen later we split u,
u=v+W, where V=U when lul~s, W=U when lul>s.

Then we have by (4.5.8) and (4.5.7)
Ilk.*vlloo~ Cs 1-p/.· = Csp/q,
and we choose s so that
Csp/q = t/2.
Then Ika * w(x) I> t/2 if Ik. * u(x) I> t, so (4.5.17) gives
m(t)~m{x; Ik.*w(x)1 >t/2} ~ C't-·llwll~,
and we obtain by Minkowski's inequality
Ilka*ull:~ C" S fl- 1-.( S luldx)"dt
lui >s
~C"(S( S fl- 1- adt)1/al u(x)ldx)".
«Iu(xll
4.5. Basic IJ Estimates for Convolutions 121
The integral with respect to t is proportional to t'I- a since q>a, and

when s = lu(x)1 this is proportional to
lu(x)l(q-a)p/q = lu(x)la p/p'.
Altogether we have therefore
Ilka*ull:~ C 3 (J lu(xW+P/P' dx)a= C (J lu(x)IPdx)a= C

3 3,
which completes the proof of (4.5.5)'.
As an application we shall now prove the Sobolev embedding

theorems, for which Theorem 4.5.3 was in fact originally intended.
First we give a local form.
Theorem 4.5.8. Let UE.@'(X) where X is an open set in IRn, and assume
that 0jUEn;oc(X), j=l, ... ,n, where l<p<n. Then it follows that
UE~oc(X) if
(4.5.18) lip = llq + lin.
Proof Let E be the fundamental solution of the Laplacean given in

Theorem 3.3.2 and set Ej=ojE=xjlxl-nlcn' Then
IEix)1 ~ Ixl- n / a Ic n, lla = I-lin,

so Theorem 4.5.3 gives, Co being dense in IJ',
(4.5.19)
Now choose XECg"(X) equal to 1 in a large open subset Y of X. Then
xu=E*,1 (xu)= I Ej*Oj(XU)
= I Ej*(X(OjU)) + I Ej*(uojx).
Here Ej*(X(oju))ELq by (4.5.19) and Ej*(UOjX)ECOO(Y) by Theorem
4.2.5, so uEL~oc(Y), which proves the theorem.
Next we give a global version of Theorem 4.5.8:
Theorem 4.5.9. Let UE.@'(IRn) and assume that 0juEIJ'(R n), j=l, ... ,n
where l<p<n. Then there is a constant C such that u-CEH(IRn)
where q is determined by (4.5.18).
Proof With the notation used in the proof of Theorem 4.5.8 we set
v=IEj*ojUEH(R n)
122 IV. Convolution
and must prove that 0kV = 0kU, k = 1, ... , n, which implies that v - U is a
constant. Here IJ'3j --+ Ej *j EH denotes the continuous extension of
the convolution map defined on Co. Choose XE Cgo (JR") with 0 ~ X~ 1
and X= 1 in a neighborhood of O. With Ej(x) = x(ex)E/x) we have if
WEIJ'
Ej*w --+ Ej*w in H(JR") when e --+ O.
Since IIEi*wllq~ Cllwllp with C independent of e it suffices to prove
this when WECo. Then we have Ej*w=Ej*w on any compact set
when e is small, and IEj*wl ~ IEjl * IwlELq so the statement follows by
dominated convergence. Hence we have with convergence in ~'
v=lim IEj*ojU' Ok v=lim IEj*Ok 0ju=lim IOjEj*OkU,
l;-O 8-0 £-0
Here
I OjEj= x(ex),1 E +e I Xj(ex)Ej= <>0 +e I Xj(ex)Ej
where Xj=OjX, The H norm of (Xiex)E)*OkU has a uniform bound as
8--+0, so 0kV=OkU as claimed.
When p increases to n then the exponent q increases to 00, and the
preceding result breaks down in the limiting case. However, when
p > n we have a substitute result based on a supplement to Theorem
4.5.3.
Theorem 4.5.10. Assume that kEC 1 (lRn ,,0) is homogeneous oj degree

- nja, let 1 ~ p ~ 00 and assume that
(4.5.20) 0<y=n(1-1ja-1jp)< 1.
Then we have
(4.5.21) sup Ihu(x)-k*u(y)llx- yl-Y ~ C Ilulb,
x*y
U E IJ' (1R n) II tff' (1R n).
Proof The convolution is a continuous function for

(4.5.22) (J Ik(y)IP' dy)l/p' ~ CR(n-np'/a)/p' = CRY.
IYI<R
In proving (4.5.21) we assume y=O and set h=lxl. Then
h u(x) - h u(O) = J(k(x - y) - k( - y» u(y)d y
and we split the integral in the part with Iyl <2h and that with
Iyl > 2h. The first part is < C' hY Ilulip by (4.5.22). In the second part we
use that Ik(x+y)-k(y)I~Chlyl-l-n/a by the mean value theorem (see
also the proof of Lemma 4.5.6) which is in IJ" at 00 since
n- p'(1 + nja)=p'n(1jp' -1ja)-p' =p'(y -1)<0.
4.5. Basic If Estimates for Convolutions 123
Hence
( f Ik(x - y) - k( - y)IP' dy)l/P' :;:;; C hY
Iyl > 2h
The following is an analogue of Theorem 4.5.9.
Theorem 4.5.11. Let UE.@'(1R.n) and assume that OjuEI!'(1R. n), j= 1, ... , n,
where p > n. Then u is a continuous function and with y = 1 - nip we
have
(4.5.23) sup lu(x) -u(y)l/lx- ylY:;:;; C I Ilojuli p '
x*y
Proof. The modified convolution
v(x) = fI (E/x- y)-E/ - Y»Oju(y)dy

defines a continuous function by the proof of Theorem 4.5.10, and the
estimate (4.5.23) is valid with v instead of u in the left hand side. Here
the notations are the same as in the proof of Theorem 4.5.9. We can
repeat the proof there to show that 0kV=OkU for every k if we note
that the L p ' norm of BX/BX)Ej is O(B 1 - y ) by (4.5.22). This proves the
theorem.
The local version of Theorem 4.5.11 which follows is proved just

as Theorem 4.5.8, with reference to Theorem 4.5.10 instead of Theo-
rem 4.5.3.
Theorem 4.5.12. Let UE.@'(X) where X is an open set in 1R.n, and assume
that 0jUE.u;oc(X), j = 1, ... , n, where p > n. Then it follows that U is
H older continuous of order y = 1 - nip, that is,
(4.5.24) sup lu(x)- u(y)l/lx - ylY < 00 if K r& X.
x=t=y;x,yeK
By repeated use of the preceding theorems we obtain the full

Sobolev embedding theorems:
Theorem 4.5.13. Let UE.@'(X) where X is an open set in 1R.n and assume
that aaUEn;oc(X) when lod =m. Here m is a positive integer and
l<p<oo: If lod<m then
(i) OaUEVloC(X) if q<oo and 1/p:;:;;1Iq+(m-itxl)/n
(ii) oa u is Holder continuous of order y if 0 < y < 1 and
lip:;:;; (m -1a:1- y)/n.
124 IV. Convolution
Proof (i) follows from Theorem 4.5.8 by induction for decreasing lad·
To prove (ii) it is by Theorem 4.5.l2 sufficient to show that aja«uEL~oc
whenj= 1, ... , n if nlq= 1-y. But then we have
lip ~ 1/q + (m -lad -l)/n
so this follows from (i) or the hypothesis.
The global result is parallel but with equality in the conditions (i),
(ii), and we omit the statement.
Notes
The applications of regularization to convex, subharmonic and pluri-
subharmonic functions in Section 4.l are mainly intended to illustrate
the use of regularization, but some of the results are required in
Chapters XV and XVI. For a more thorough discussion of convex
functions and sets the reader might consult the classical text by
Bonnesen and Fenchel [1]. We shall continue the study of sub-
harmonic and plurisubharmonic functions in Chapter XVI using po-
tential representation formulas. There is a recent monograph by Hay-
man and Kennedy [1] on subharmonic functions, and the reader
could consult Lelong [1] for the theory of plurisubharmonic func-
tions. Theorem 4.1.12 actually goes back to Poincare [1], and Theo-
rem 4.l.l5 is due to Lelong [2].
In Section 4.2 we followed Gelfand and Shilov [1] in the de-
finition of the convolution. The definition of Schwartz will be given in
Section 5.1. The theorem of supports in Section 4.3 is due to Titch-
marsh [1] in one dimension. The simple extension to n variables was
given by Lions [1]. Most proofs depend more or less on analytic
function theory (see Section 16.3). Mikusinski [1] gave an argument
reproduced here which reduces the proof to the case of a convolution
with equal factors. The theorem is then a consequence of the Paley-
Wiener theorem (Theorem 7.3.l). We give a direct elementary proof
using only convolutions. A similar argument occurs in Mikusinski
[2].
The observations in Section 4.4 on the use of fundamental so-
lutions will be developed very systematically in Chapters X and XI so
we refrain from discussing the results here. The estimates in Section
4.5 are due to Hardy and Littlewood [1] when n = 1. Sobolev [2] gave
a rather difficult reduction of the n dimensional case to the one
dimensional case by means of spherical symmetrization. Later on
Notes 125
DuPlessis [1] has observed that there is a very simple reduction by

means of the inequality between geometric and arithmetic means.
Here we have chosen proofs in the spirit of Zygmund [1] which will
be applicable also in a related context in Section 7.9. The main point
is a covering lemma due to Calderon and Zygmund [1], which is
actually due to F. Riesz [1] in the one dimensional case. The reader
should consult Nirenberg [4] for a thorough discussion of results
related to the Sobolev embedding theorems. A very simple derivation
from the Hardy-Littlewood maximal theorem can be found in Hed-
berg [1].
Chapter V. Distributions in Product Spaces
Summary
We were not able to define the product of arbitrary distributions in
Chapter III. However, as we shall now see this can always be done
when they depend on different sets of variables. Thus to arbitrary
distributions UjE!?)'(X}, Xj open in IRnj U=1,2), we define in Section
5.1 a product U1 Q9U 2 E!?)'(X 1 x X 2) in Xl x X 2 cIRnl+n2. In case U j are
functions this is the function Xl x X 23(X 1, x 2) -+U 1(X 1)U 2(X 2).
On the other hand, a function K E C(X 1 X X 2) can be viewed as the
kernel of an integral operator :%,
(:% U)(X 1) = JK(Xl' X2)U(x 2)dx 2,
mapping Co(X 2) to C(X 1) say. It is not easy to characterize the
operators having such a kernel. However, the analogue in the theory
of distributions is very satisfactory. It is called the Schwartz kernel
theorem and states that the distributions K E!?)' (X 1 xX 2) can be iden-
tified with the continuous linear maps :% from C~(X2) to !?)'(X 1 )
which they define. This will be proved in Section 5.2. We shall return to
this topic in Section 8.2. A rather precise classification of singularities
will then allow us to discuss the regularity of :% U and its definition
when U is not smooth.
5.1. Tensor Products

If Xj is an open set in IRnj,j=1,2, and ifujEC(Xj), then the function
u 1 Q9U 2 in Xl xX 2 cIRnl+n2 defined by
(u 1® U2)(X 1, x 2) =U 1(x 1)U 2(X 2), XjEXj'
is called the direct (or tensor) product of U 1 and U 2 • To extend the
definition to distributions we observe that u1 ® u 2 E C(X 1 X X 2) and
5.1. Tensor Products 127
that
J
H(U1 ® U2)(cf>1 ® cf>2) dX1 dX2 = U1 cf>1 dxd U2 cf>2 dX2, cf>jEC'O(XJ
Theorem 5.1.1. If UjE~'(X j), j = 1, 2, then there is a unique distribution

UE&&'(X 1 xX 2) such that
(5.1.1) U(cf>l ®cf>2)=U 1(cf>1)U 2(cf>2)' cf>jEC'O(X}.
We have
(5.1.2) u(¢) = U 1[u 2(cf>(x 1, x 2))] =u 2[u 1(¢(Xl' x 2))],
¢EC'O(X1 xX 2 ),
where uj acts on the following function as a function of Xj only. If
UjEC', j = 1, 2, the same formula is valid for cf> E Coo. The distribution U is
called the tensor product and one writes u = u 1® u2.
Proof. a) Uniqueness. We must show that ifuE2)'(X 1 xX 2) and if

U(¢1 ®¢2)=0 for ¢jEC'O(X},
then u=O. To do so we take I/IjEC'O(1Rnj) with I/Ij'?;O, Jl/ljdx j =l, and
Ixjl;;:?; 1 if XjESUpp I/Ij. With
'P,;(x 1 , x 2) =e- n1 - n2 1/11 (x t!e)1/12 (x2/e)
we know that u*IJ'.-+u in &&'(Y) if Y~XlxX2 (Theorem 4.1.4).
However, u*IJ'.=O in Y for small I', since 1J'.(X 1 -Yl' X 2 -Y2) is the
product of a function of Y1 and one of Y2' Hence u=o in Y and
therefore in X.
b) Existence of u and (5.1.2). Let K j be a compact subset of Xj.
Then
IUj(cf>}I;;:?; C jLsupla~¢),
1~I~kj
Iq,(x 1) = U 2(cf>(X 1 , .))

is in C'O(K 1) by Theorem 2.1.3, and
a~1 Iq,(x 1 ) =u2(a~1 ¢(xl' .)).
Hence
so U1(I q,) is defined and
lu1(Iq,)I;;:?;C 1 C 2 L supla~~a~~cf>(Xl,X2)1.
I~JI ~kj
128 V. Distributions in Product Spaces
Writing u(cf» = u 1(I,p) we obtain a distribution satisfying (5.1.1) and the

first part of (5.1.2). In the same way we obtain a distribution satisfying
(5.1.1) and with the second property in (5.1.2). By the uniqueness both
conditions (5.1.2) must therefore be valid. The remaining statement
follows in the same way; note that
(5.1.3)
Example 5.1.2. If baj is the Dirac measure at ajEXj' then ba ,®ba2 is

the Dirac measure ba at a=(a 1, a 2)EX 1 xX 2' Theorem 2.3.5 can now
be stated as follows: If UE~I(X1 xX 2) is of order k and if suppuc
Xl x {a 2}, a2EX 2, then
u= L
ua ®OXb a2
lal~k
where UaE~lk-lal(x 1) and (X is a multi-index corresponding to the X 2

variables.
The direct product allows us to justify the definition (1.3.1)' of the
convolution in general. In fact, if u 1 , u2E~I(IR.n) and either one has
compact support, then
(5.1.4) (U 1®U 2)(cf>(X l +x 2))=(U l * u 2)(cf» , cf>EC~(IR.n),
for if cP(x) = cf>( - x) we have
uicf>(x 1+x 2))=(U 2 * cP)( -Xl)
so the left-hand side is u 1* (u 2 * cP)(O) =(u 1 * u 2) * cP(O). (5.1.4) could also
have been taken as definition of convolution. However, it was con-
venient to have convolution available in the proof of Theorem 5.1.1.
5.2. The Kernel Theorem
Every function K E C(X 1 X X 2) defines an integral operator f from

Co(X 2) to C(X 1) by the formula
(fcf>)(X1)=JK(X1,X2)cf>(X2)dx2' cf>EC o(X 2), X1EX 1·
We shall now show that the definition can be extended to arbitrary
distributions K if cf> is restricted to C~ and f cf> is allowed to be a
distribution. To do so we start from the observation that when
KEC(X1 xX 2 )
(5.2.1)
Theorem 5.2.1 (The Schwartz kernel theorem). Every KE~'(X 1 X X 2)
defines according to (5.2.1) a linear map f from C~(X 2) to ~'(X 1)
which is continuous in the sense that $'cf>j~O in ~'(Xl) if cf>j~O in

Cg'(X 2)' Conversely, to every such linear map $' there is one and only
one distribution K such that (5.2.1) is valid. One calls K the kernel of
$'.
Proof. If KE~'(Xl xX 2) then (5.2.1) defines a distribution $'cf> since
ljJ~K(ljJ®cf» is continuous; $' is continuous since cf>~K(ljJ®cf» is
continuous. To prove the converse we first note that the uniqueness is
identical to the uniqueness in Theorem 5.1.1. To prove the existence
we observe that for any compact sets K j c Xj there are constants C, ~
such that
(5.2.2) 1($'cf>,ljJ)I~C L suplaaljJl L suplaPcf>I;
lal ~N, IPI ~N2
ljJEC'~(Kl)' cf>ECg'(K 2).
In fact, by hypothesis the bilinear form
Cg'(K 1) x Cg'(K2)3(ljJ, cf» ~ ($' cf>, ljJ)
is continuous with respect to cf> (resp.ljJ) for fixed ljJ (resp. cf», and
every separately continuous bilinear form in a product of Frechet
spaces is continuous.
Let lj~Xj' choose the compact sets K j as neighborhoods of ~
and set for (X1,X2)EYl x Y2 and small B>O
(5.2.3) K,(x 1 , x 2) = c·,-n 2 ($' ljJ 2 ((X 2 - .)/B), ljJ I ((Xl - .)/B)
where ljJj are chosen as in the proof of Theorem 5.1.1. Note that if we
already knew that there is a distribution K satisfying (5.2.1) then K,
would be K * 'I', and therefore converge to K as B ~ O. Our program is
now to show that K, does have a limit in ~'(Yl x Y2) when B ~ 0 and
then to show that (5.2.1) is fulfilled for the limit.
(5.2.3) is well defined when B is smaller than the distance from lj
to CKj, and by (5.2.2) we have with P,=Nl +N2 +n 1 +n 2
(5.2.4) IK,(Xl,X2)I~CB-" if XjElj, j=I,2.
We shall prove that K, has a limit in ~'''+I(Yl x Y2 ) as B~O by using
an argument which is very close to the proof of Theorem 3.1.11. Note
that if ljJEcoo(lRn ) then
(5.2.5)
a a
'" (B-nljJ(X/B)) = L ~(B-nljJ/X/B)), ljJ/X) = -xjljJ(x).
vB VXj
In fact, by the homogeneity
a
B", (B-nljJ(X/B)) + L Xj ~a (B-nljJ(X/B)) = - nB-nljJ(X/B)
vB VXj
which implies (5.2.5). Now it follows from the continuity (5.2.2) that
we may differentiate with respect to e or Xj in (5.2.3), and by (5.2.5)
this gives
where Xv runs over all coordinates of (X 1 ,X 2 ). Here V. is defined by

replacing 1/11 or 1/12 by the product with -xv, so (5.2.4) is valid for V •.
Repeating this process we conclude that
K~)(X1,X2)=aj K.(X 1,X 2)/aei
is a sum of derivatives of order j of functions having a bound of the
form (5.2.4), so ff K~) is bounded in £&,j(Y1 x Y2) for every j. With a
fixed small band e -+ 0 we now use Taylor's formula
~ 1
K.= L (e- bY K~ljj! +(e- b)~+ 1 S Krr+~(~)_d)(1-t)~/J1! dt.
Since ° °
(1- tt /(b + t(e - b))p ~ b-~
it follows for <PE C~+ 1 (Y1 x Y2 ) that when e -+ 0
<K., <P) -+ <Ko, <P)

~ 1
=L(-b)j<K~ljj!,<P)+(_W+1 S <K~t!:),(1-tnJ1!dt,
° °
where KoE£&'~+ 1 (Y1 x Y2).
Let cPjEC'~(Yj) and form
<K., cP1 ® cP2) = SSK.(X1' X2)cP1 (x 1)cP2(x 2)dx 1dx 2·
With the notation Vlj,.(xj)=e-njl/li -x/e) we have
SSK.(X1' X2)cP1 (X 1)cP2(x 2)dx 1dX 2

=ss <fIP2,.(· -X 2)cP2(X 2),1P1,.(. -X 1)cP1(x 1»dx 1dx 2·
Replacing the integral by a Riemann sum first we conclude as in the
proof of Lemma 4.1.3 that the integration can performed "under the
sign", hence
<K., cP1 Q9cP2)=<f(cP2*~2")' cP1 *~1")'
Since cPj* ~j" -+ cPj in Cg'(Yj) when e -+ 0, it follows from (5.2.2) that
the right-hand side converges to <f
cP2' cP1) when e -+ O. Thus
<Ko, cP1 Q9cP2)=<fcP2' cP1) if cPjECg'(Yj),

and since Yj are arbitrary relatively compact subsets of Xj' this com-
pletes the proof.
Example 5.2.2. The kernel of the identity map %: C~(X) ~ C~(X),

where X is an open set in JR", is the distribution
<K, cP) = ScP(x, x)dx, cPEC~(X x X),
with support in the diagonal {(x, x), XEX}.
Theorem 5.2.3. The kernel of a continuous map %: C~(X) ~ .@'(X) is
supported by the diagonal if and only if
(5.2.6)
where aaE.@'(X) and the sum is locally finite.
Proof For the operator (5.2.6) we have
<% <p, tfJ) = L <aa' (fyaqJ)tfJ)
so the kernel is given by
<K, cP) = L <aa' ~cP(x, Y)lx~y)
which is obviously supported by the diagonal. Conversely, if the
kernel K of % is supported by the diagonal, it follows from Theorem
2.3.5 that K has the preceding form, which proves the theorem.
The preceding operators preserve supports; more generally, we
have
Theorem 5.2.4. If K E.@' (X 1 X X 2) and % is the corresponding operator,
then
(5.2.7) supp %ucsupp K osuppu, UEC~(X2)'
Here supp K eX 1 X X 2 is considered as a relation acting on

supp ucX 2 . Thus
suppK 0 M ={x l EX l ; 3X 2 EM, (Xl' X2 )ESUppK}.
This is a closed set when M is compact, for supp K is closed.
Proof Assume that X dsupp K supp u. Then there is a neighborhood
0
V of Xl such that Vn(suppK o supp u)=0. If VEC~(V) then

(supp v ® u) n supp K = 0
which proves that <%u, v) =0, hence %u=O in V.
Example 5.2.5. If f: X 1 ~ X 2 is a continuous map and % <p = <p 0 f,
<pEC~(X 2)' then the kernel is given by
<K, cP) = ScP(x, f(x))dx, cPE C~(X 1 xX 2)'

so the support is in the graph of f
The operator (5.2.6) has a natural extension to all <PEe!' if the
coefficients aa E Coo. General sufficient smoothness conditions for the
existence of such extensions will be given in Chapter VIII, but we give

an elementary example now:
Theorem 5.2.6. If K EC oo (Xl xX 2) then the map :% defined by (S.2.1)
has a continuous extension from @"'(X 2 ) to C oo (X1 ),
(S.2.8) :%u(x 1 )=u(K(x 1 , .)), UE@"'(X 2 ),X 1 EX1 •
Conversely, every continuous linear map :% from @"'(X 2 ) to C oo (X1 ) is
defined in this way by a kernel KEC oo (Xl x X 2 ).
Proof. If KECOO it follows from Theorem 2.1.3 that (S.2.8) defines a
map @"'(X 2 )--+ COO (X 1)' and the continuity is a consequence of Theo-
rem 2.1.8. Conversely, if we are given a continuous map :%:
@"'(X 2 )--+Coo(X1 ) then
K(.,x 2 )=:%C\2' X2 EX 2 ,
is a continuous function of X 2 with values in COO (Xl)' Taking differ-
ence quotients we find that K is continuously differentiable in x 2 ,
<y, ax,) K(., x 2 ) = -:% <y, a) b X2 '
Repeating the argument gives KEC oo (Xl xX 2 ). We obtain (S.2.8) since

this is true for finite linear combinations of Dirac measures and they
are dense in @"'.
Notes
The tensor product was defined in Schwartz [1], and the kernel
theorem was announced shortly afterwards in Schwartz [2]. In both
cases the main point is the decomposition of test functions in Xl X X 2
into sums of tensor products of test functions in Xl and in X 2' Thus
the topological tensor product of CO'(X1) and CO'(X 2) is involved,
and Schwartz [4] gave a proof emphasizing this aspect. Ehrenpreis
[4] published a more elementary proof where the decomposition was
made by Fourier series expansion (see also Gask [1]). Here we have
used instead the fact that a regularization of any test function in Xl
X X 2 by a product of two test functions in the corresponding spaces
JR"' and JRn 2 can be considered as a superposition of tensor products
of test functions in Xl and in X 2' (To be able to use this argument we
had to define convolution before the tensor product.)
There is an interesting addendum to Theorem S.2.3 due to Peetre
[2]: If :% is any linear map CO'(X) --+ COO(X) with supp:%ucsuppu,
UECO'(X), then :% is a differential operator with COO coefficients, that
is, (S.2.6) is valid with aaECoo. Note that no continuity is assumed; it
follows from the restriction on the supports.
Chapter VI. Composition with
Smooth Maps
Summary
If f is a map 1Rn ~ 1Rm then a function u in 1Rm can be pulled back to
a function u 0 f in 1Rn, the composition. In Section 6.1 we show that
this operation can be defined for all distributions u if f E COO and the
differential is surjective. (In Section 8.2 we shall find that the com-
position can be defined for more general maps f when the location of
the singularities of u is known in a rather precise sense.) As an
example we discuss in Section 6.2 how powers of real quadratic forms
can be used to construct fundamental solutions for homogeneous
second order differential operators with real coefficients. In Section 6.3
we use the fact that distributions can be composed with diffeomor-
ph isms to define distributions on Coo manifolds simply as distri-
butions in the local coordinates which behave right when the coor-
dinates are changed. In Section 6.4 we continue the discussion of
manifolds by giving a short review of the calculus of differential forms
on a manifold, ending up with the Hamilton-Jacobi integration theory
for first order differential equations. These results will not be used
until Chapter VIII, and the geometrical notions related to the Hamil-
ton-Jacobi theory will be discussed in much greater depth in Chapter
XXI.
6.1. Definitions
Let Xj be an open set in 1Rnj , j=1,2, and let f: Xl ~X2 be a Coo
map. We wish to extend the definition of the composition
CO(X 2)3U~uofECO(Xd
to distributions u so that the map
~'(X 2)3U ~ uofE~I(Xl)
134 VI. Composition with Smooth Maps
is continuous. If this is possible it follows from Theorem 4.1.5 that it

can only be done in one way. However, we must put conditions on f
Theorem 6.1.1. If ujof-+O in §'(X I ) for every sequence Uj ECg'(X 2 )

such that uj -+ 0 in £0'(X 2), then f is open, that is, f(V) is open in X 2 if
VeX I is open. If UECg'(X 2 ) implies uofECg'(XI), then f is proper.
Proof The second statement is obvious, for if K is a compact subset

of X 2 we can choose u equal to 1 on K and obtain U 0 f = 1 on
f I (K) which must therefore be compact if U 0 f E Cg' (X I)' To prove
the first statement we assume that f is not open. Assume for example
that OEX j , f(O) =0, and that V is a compact neighborhood of 0 in XI
such that f(V) is not a neighborhood of 0 in X 2' Choose a sequence
y/H(V) so that Yj -+ 0, and set Iyjl=ej . With 0~uECg'(lRn2) equal to 1
in the unit ball, we set u,(y)=e-n'u(y/e) and obtain u,of(x)"?;c n,
when C Ix 1< e. Since u, 0 f "?; 0 it follows in view of Theorem 2.1.9 that
u,of does not converge to 0 in §'(XI ) as e-+O. Now
Vj=U'j - I aj,aoaoYj
lal~1l
tends to 0 in £0' (X 2) when j -+ 00 if n2 - n l + .u"?; 0 and aj,a are suitably

chosen. In fact,
<vj,¢)=ej'-n, J¢(ejy)u(y)dy- I aj,a(-o)a¢(Yj)
lal ~Il
=e]>-n, J I oa¢(Yj)(ejy - yl/a! u(y)dy + 0(e;2-n,+I'+ I)
lal~1l
- I aj,a( _a)IX¢(y)
IIXI~1l
so we just have to take

aj,IX = ( _1)la l ej'2- n, J(ejy - y)"u(y)dy/a!.
If we replace aIXO yj by a smooth approximation with support in Cf(V)
we obtain a sequence v;-+O in ,~'(X2) such that vjof=u,jOf in V.
This does not converge to 0 in §' (XI) so the theorem is proved.
If f is open then f'(x) is surjective for all x in an open dense

subset of XI' On the other hand, by the inverse function theorem f is
open if f'(x) is surjective for every x. We shall now prove that one
can then define the composition with f
Theorem 6.1.2. Let Xj e lR n j, j = 1, 2, be open sets, and f: X I -+ X 2 a Coo

map such that f'(x) is surjective for every XEX I . Then there is a unique
continuous linear map f*:§'(X 2 )-+§'(X 1 ) such that f*u=uof when
6.1. Definitions 135
UE CO(X 2). It maps !'.d,k(X 2) into .@'k(X 1 ) for every k. One calls f* u the
pullback of u by f
Proof As already observed, the uniqueness follows from Theorem

4.1.5. To prove the existence we choose for any fixed XoEX 1 a Coo map
g: XI --+ lR n, -n" for example a linear map, such that the direct sum
ffBg,
has a bijective differential at x o. By the inverse function theorem there

is an open neighborhood Y1 eX 1 of X o such that the restriction of
fEB g to Y1 is a diffeomorphism on an open neighborhood Y2 of
(f(xo),g(x o)). We denote the inverse by h. If UECO(X 2 ) and CPECli'(Y1)
then a change of variables gives
J(f* u)cpdx = Ju(f(x))cp(x)dx = Ju(y')cp(h(y)) Idet h'(y)1 dy
where we have written y=(y',y")ElRn2EBlRnl-n2. Hence
(6.1.1) (f* u)(cp) = (u ® 1)(<1», <1>(y) = cp(h(y)) Idet h' (y)l.
Here 1 is the function 1 in lR n ,-n2. If UE.@'(X 2 ) and we choose
UjE Cli'(X 2) so that uj --+ U in .@'(X 2)' it follows in view of the remark
after Theorem 2.2.4 that f*u j converges in .@'(X) to a distribution
f* u defined by (6.1.1) in Y1 • Thus (6.1.1) gives a local definition of
f* u, and the continuity of the map U --+ f* u follows at once from
(6.1.1). The theorem is proved.
Remark. The proof shows that if f E ck+ 1 only, then f* is well defined
and continuous in .@,k. In fact, cp --+ <1> is continuous from C~ to C~.
(We need an extra derivative for f since det h' involves one derivative
of f)
Since we have defined f*u by continuous extension from the case
of functions u, it is clear that the usual rules of computation remain
valid:
(6.1.2) oJ*u= L 0jJ,J*Ok U, UE.@'(X 2 ) (the chain rule),
(6.1.3) f*(au) =(f* a)(f*u); aECoo(X 2 ), UE.@'(X 2 )·
Here f is assumed to satisfy the hypotheses of Theorem 6.1.2. If in
addition we have a COO map g: X 2 --+ X 3 with surjective differential,
then
(6.1.4) (go f)* U= f* g* u, UE.@'(X 3 ),
In practice it is often convenient to use the notation u(f), u f or 0
even u(f(x)) instead of f*u since (6.1.2)-(6.1.4) look more familiar

then. However, one must always keep in mind then that one IS
referring to an extension of the pointwise definition given by (6.1.1).
Example 6.1.3. If f is a diffeomorphism Xl - X 2 between open sets in

1Rn then f*(\=ldetf'(x)l-lbx where f(x)=y. This follows from (6.1.1)
with h=f-l.
Example 6.1.4. If Mtx=tx, xE1R n, t>O, then

(Mt* u)(cjJ) = u( cjJ(. /t)/tn).
Thus (3.2.18) with t replaced by l/t means that
Mtu= f!1u in 1Rn " 0,
which is just the usual definition of homogeneity of degree a.
Theorem 6.1.5. If p is a real valued function in COO(X), X c1Rn , and if

Ip'I=(Ilop/oxi)t=l=O when p=O, then p*bo=dS/lp'l where dS is the
Euclidean surface measure on the surfact: {x; p(x)=O}.
Proof Let p(xo)=O and assume for example that op(XO)/ox 1 =1=0. Then
we can apply (6.1.1) in a neighborhood with
h- 1 (x) =(p(x), x 2, ... , x n).
Then h(O, Y2' ... , yn)=(I/I(Y2' .. " Yn), Y2' .. " Yn) lies on the surface p =0,
and we have for cjJECo(Y) if Y is a small neighborhood of Xo
(p* bo , cjJ> = S(cjJ!iOl pi) 0 h(O, Y2' ... , yn)dY2 ... dYn'
Since p(I/I, Y2' ... , Yn) =0 we have for j =2, ... , n
0lPOl/l/OYj + OjP =0.
Hence
Ip'I=loIPIM, M=(l+t(Ol/l/Oyyt
Since dS=MdY2'" dYn with the parameters Y2'" Yn, this proves the
theorem.
From (6.1.2) it follows that if H is the Heaviside function then

OjP* H =(ojP)p*b o =(ojP)/lp'ldS,
which means that we have given another proof of the Gauss-Green
formula (3.1.5). One calls p* bo a simple layer on the surface p = 0, and
its derivatives are called multiple layers. They are essentially the pull-
backs by p of the derivatives of boo In fact, let
L=La",(x)o'"
be a differential operator with Coo coefficients and define V so that LO
=L and V+ 1 = [V, p], that is,
V+l u=V pu-pVu.
It is clear that V are differential operators of decreasing order. Then
we have
(6.1.5) L(up*t5o)=L(Vu)P*t5~)/k!, UEC oo ,
k
where the sum is locally finite. By a change of variables we reduce the

proof to the case where p(x)=x 1 , and we may assume that L=o~
then. Since V=o~-km!/(m-k)! the formula (6.1.5) becomes Leibniz'
formula for the differentiation of a product then.
6.2. Some Fundamental Solutions

Let A be a real non-singular quadratic form in 1Rn , thus oA/ox=I=O
when x =1= O. Then A *f is defined in 1Rn" 0 if f E~' (1R). When f is
homogeneous of degree a then f(A)=A*f is homogeneous of degree
2a, for if Mtx=tx for xE1Rn , t>O, we have
Mt* A*f = (A Mt)*f =(t 2 A)*f =A*mi>f = t 2a A*f
where ms denotes multiplication by s on 1R. Unless 2a is an integer
~ - n it follows from Theorem 3.2.3 that f(A) has a unique extension to
1Rn which is homogeneous of degree 2a, and we shall use the notation
f(A) for the extension also, when no ambiguity can arise.
We can write
A(x)= L ajkxjxk
with ajk = akj , and introduce the differential operator
B(0) = L bjkojOk
where (b jk ) is the inverse of (a jk ). We shall now compute B(o)A*f in
1Rn" 0 when f is homogeneous of degree a on 1R. Since
oJ (A) = oA/oxJ'(A),
ojoJ(A) = 2 ajkf' (A) + oA/oxjoA/oxJ" (A)
and L bjkoA/oxj = 2x k, we have
B(o)f(A) =2nf'(A) + 4Af"(A) = g(A)
where by (3.2.19)'
g=2nf' + 4tf" =(2n+4(a -1))1'.
This is 0 if a=(2-n)/2, so B(8)f(A) vanishes in 1Rn"O then. It follows
that B(8)f(A) is homogeneous of degree -n and supported by {O},
hence it is a multiple of bo which will now be determined.
Theorem6.2.1. Let the signature of A be (n+, n_), that is, n+ +n_ =n

and A is positive (negative) definite in some n+ (n_) dimensional plane.
If cn is the area of the unit sphere in 1Rnand n > 2, then
(6.2.1) B(8)(A ± iO)(2-n)/2 =(2 - n)cnldet AI- t e hin -/ 2 bo,
(6.2.1)' B(8)A* X<;'-n)/2 = ±4n(n-2)/2 sin (nn±/2)ldet AI-t bo.
Proof It is sufficient to verify that

(6.2.1)+ B(8)(A + iO)(2-n)/2 = (2 - n)c nIdet AI- t r"in-/2 bo .
In fact, complex conjugation gives the other case of (6.2.1). The two
cases of (6.2.1), are interchanged if A is replaced by -A. By (3.2.9) we
have .. . .
x~(e,,,a-e-"'")=(x-iO)ae"'"-(x+iO)ae-,,,a, Re a>O.
Since r(a+ l)r( -a)= -n/sin(na) and X~ =x~/r(a+ 1), it follows that
X~ =ir( -a)/2n((x-iO)a e"ia_(x + io)ar"ia), Re a>O, a¢'lL+,
and by analytic continuation this is extended to all a¢'lL +. If we note
that
(2-n)cJ(n/2-1)/n= _4n(n-2)/2
and
nin_/2+ni(l-n/2)=ni(l-n+/2),
then (6.2.1)' is a consequence of (6.2.1).
First we shall prove that
(6.2.1)" B(8)A(2-n)/2 =(2 - n)cn(det A)-t bo
when Re A is positive definite and (det A)- t is defined as in Section
3.4. Here A(2-n)/2 is of course the homogeneous extension of
A(x)(2-n)/2 where -n/2 <argA(x) <n/2. By the uniqueness of analytic
continuation it suffices to prove (6.2.1)" when A is positive definite.
We can then choose a linear bijection T in 1Rn such that the pullback
T* A of A by T is the Euclidean metric form. (6.2.1)" is then a
consequence of Theorem 3.3.2, for
IdetAI-tT*b o = IdetAI-t Idet TI- 1 bo =b o
so the two sides have the same pullback under T by Theorem 3.3.2.
Assume now that A is real and non-degenerate and apply (6.2.1)"

to A.(x) = -iA(x)+elxI2, e>O. Then
(det A.)-t -+ Idet AI- t e"i(sgnA)/4
(cf. (3.4.6)), B. -+ iB, and
A~2-n)/2 =(rlti/2(A + ei Ixj2))(2-n)/2 -+ i- 1eltin /4 (A + iO)(2-n)/2
in ~'(lRn"o) by Lemma 6.2.2 below. From Theorem 3.2.3 it follows
that we have convergence in ~'(lRn) for the homogeneous extensions,
so (6.2.1)+ follows from (6.2.1)".
Lemma 6.2.2. Let F be a COO function in X x J where Xc lRn is an

open set and J is a neighborhood of 0 in lR. Let f, I and Z be as in
Theorem 3.1.11. If F(x,O)EI and 13F(x,O)/13x=l=O when XEX, and
F(x, e)EZ when XEX, O<eEJ, it follows that f(F(., e))-+F(., 0)* f(. + iO)
as e-+ +0.
Proof. The statement is local and invariant under coordinate changes

so we may assume that F(x, 0) = Xl. From the proof of Theorem 3.1.11
we know that f = G(N+ 1) where G is analytic in Z and continuous in
Z. If CPECg"(X) it follows by partial integrations with respect to Xl
that as e-+O
Jcp(x) f(F(x, e)) dx = Jcp(x) G(N + (F(x, e)) dx
1)
= JG(F(x, e))(13 1( - 13F(x, e)/13x 1)-lt+ 1 cp(x) dx

-+ JG(F(x, 0))( -Ol)N+ 1 cp(x) dx
=JG(X 1)(-Ol)N+1 cp(x)dx

= J<f(· + iO), cp(., X') dx'.
This completes the proof.
If we divide (6.2.1) by the constant in the right-hand side of (6.2.1)

we obtain a fundamental solution E of B(o) which is homogeneous of
degree 2 - n. (When n + = n _ = 1 it is easy to see that (6.2.1)' remains
valid, so it gives a fundamental solution also in this case.) Let us now
consider the special case of the wave operator in lRn + 1
D=c- 202/ot 2 -A
where tElR, c is the speed of light, and A is the Laplacean in xElRn. In
our earlier notation this is the operator B(o) if
By (6.2.1)' a fundamental solution is given by (n~l)
E =11;<1 ~n)/2 e/4A * X~ ~n)/2.

It has support in the double cone where A ~ O. Let
E+ =2E when t>O, E+ =0 when et<lxl.
The two definitions agree in the overlap of the two regions and define
a distribution which is homogeneous of degree I-n in JRn +1.,O. We
use the same notation for the homogeneous extension to the whole
space. Then
where E ~ = s* E +' S denoting reflection with respect to the origin.

Now DE+ = DE~ =(jo, so both are fundamental solutions supported
*
respectively by the forward cone where t ~ 0 and the backward cone
where t ~ O. When n is odd and 1 the support is in the boundary of
the cone by (3.2.17)'.
We can compute A * X~ - n)/2 when t > 0 by using (6.1.1) with h
equal to the inverse of
(t, x) ~ (A, x).
Thus
h(s, x) =((s+ Ixl2)t Ie, x), Idet h'l =(2e)~ 1 (s+ IxI2)~t.
It follows that
(6.2.4) (E +, ¢) = 11;<1 ~n)/24 ~ 1(x~ ~n)/2, 4J),
(6.2.5)
if t >0 in supp ¢. Set
(6.2.6) ci>(t, r) = rn~ 2 S ¢(t, rw)dw.
Iwl ~ 1
Then introduction of polar coordinates in (6.2.5) gives
4J(s) = Sci>((s + r2)! Ie, r) r dr(s + r2)~t

=e S ci>(t,(C 2 t 2 _S)!)dt.
s <C 2 t 2
As stated by Theorem 4.4.8 E + is thus a continuous function of t > 0

with values in 19"(JRn),
(6.2.4)' (E + (t), tjJ) = nP -")/2 c/4 (x~ ~n)/2, 1[J((c 2 t 2 - .)t»,
tjJECOO(JR"), t>O,
(6.2.6)' ti/(r)=,n~2 S tjJ(rw)dw; ti/(ir) =0, rEJR.
Iwl ~ 1
(E+(t),I/J>=c/2 J I/J(x)dx, n=1,

Ixl <et
(6.2.7) (E+(t), I/J> =c/2n J I/J(X)(C 2 t 2 -lxI 2)-tdx, n=2,
Ixl<et
(E+ (t), I/J> =n- V c/4(d/ds)"-1 ",(stLe2t2, n=2 v+ 1.
By Theorem 4.4.8' we know that E+(t) and E~(t)=atE+(t) have limits
in ~' when t --> + 0, and
> J (E+ (t), ¢(t,.» dt,

00
(E+, ¢ = ¢ECg'(1Rn+ 1),

o
for the right-hand side is homogeneous of degree I-n and the
equality is valid when ¢ECg'(1Rn+ L ,O). Since E+ is a fundamental
solution it follows that E+(+O)=O, E~(+0)=C2bo.
Theorem 6.2.3. The wave operator has a unique fundamental solution

E+(resp.E_) with support in the forward cone where ct~lxl (resp. the
backward cone where c t ~ -Ix!). When n is odd the support is in the
boundary of the cone if n =!= 1.
Proof Only the uniqueness remains to be proved. We shall prove

more, that there is no fundamental solution other than E + with
support in the half space t~O. Any other fundamental solution must
be of the form E + + u where 0 u = 0 and t ~ 0 in supp u. But then
u=b* u=(DE+)* u=E+ * Du=O
where the computations are legitimate since
supp E + x supp U3«X, t), (y, s» --> (x + y, t + s)
is proper. (See the end of Section 4.2.) In fact, a bound for t + S implies
a bound for t and s since t ~ 0 and s ~ 0, and since Ixl ~ c t a bound for
x + y gives a bound for x and y then. The proof is complete.
The fundamental solutions in Theorem 6.2.3 are called the ad-

vanced and retarded fundamental solutions respectively, while those given
by normalization of (6.2.1) are called the Feynman fundamental so-
lutions. The fundamental solutions lead quickly to the solution of the
Cauchy problem for the wave equation:
Theorem 6.2.4. For arbitrary ¢O'¢lEcoo(1Rn) and f ECoo (1R:+ 1), 1R:+1
={(t,x); t~O, xE1Rn}, the Cauchy problem
(6.2.8)
and
u=cPo, oU/ot=cP1 when t=O,
has a unique solution uEC OO (1R:+ 1), and it is given by
t
(6.2.9) u(t,.) =c- 2 E+ (t) * cP1 +C- 2 E~ (t) * cPo + SE+ (t-s) * I(s, .)ds.
o
Proof If 1=0, cPo =cP1 =0 then (6.2.8) implies Duo =0 in 1Rn+ 1 if U o
=U in 1R:+1 and uo=O in 1R"+1 '-..1R:+ 1. The proof of Theorem 6.2.3
shows that U o must then be equal to 0, so the uniqueness is proved.
Now (6.2.9) defines a solution in C OO (1R:+ 1) of (6.2.8). Indeed, that
UE COO follows from the fact that E(t) and all its t derivatives are
continuous with values in C' when t~O. Since E+( +0)=0, E~( +0)
= c2 00 we have
t t
%t SE + (t -s) * I(s,.) ds = SE'+ (t - s) * I(s,.) ds,
o 0
t t
0 2 / ot 2 SE + (t-s) * I(s,.) ds = SE'~ (t - s) * I(s,.) ds+ c2 00 * I(t, .).
o 0
The equation Du=1 follows now since DE+ =0, t>O. The boundary
conditions in (6.2.8) are obtained if we also note that E'~(O)=
c 2 L1 E + ( + 0) = o. The proof is complete.
We shall return to the Cauchy problem in Chapter XII.
6.3. Distributions on a Manifold
The definition of composition of distributions with diffeomorphisms

allows us to define distributions on arbitrary COO manifolds. First we
recall the definition of manifolds.
Definition 6.3.1. An n-dimensional manifold is a Hausdorff space with

countable basis in which each point has a neighborhood homeomor-
phic to some open set in 1R". A COO structure on a manifold X is a
family ff of homeomorphisms 1(, called local coordinate systems, of
open set X K C X on open sets g K C 1R" such that
i) If 1(, 1(' EfIF, then the map
(6.3.1)
(between open sets in 1R.n) is infinitely differentiable. (This is then true

of the inverse map also.)
ii)
iii) If Ko is a homeomorphism of an open set X 0 c X on an open
set in 1R.n and the map
KKo 1: Ko(X 0 nX K ) -+ K(X 0 nX K)
as well as its inverse is infinitely differentiable for every KE§, it

follows that KoE.9F.
A manifold with a COO structure is called a Coo manifold. The sets

X are called coordinate patches and the cartesian coordinates of
K
K(X), XEX K , are called local coordinates in X K •

The condition iii) in Definition 6.3.1 is in a way superfluous. For if
ff satisfies i) and ii) we can extend ff in one and only one way to a
family ff' satisfying i), ii) and iii). In fact, the only such family ff' is
the set of all homeomorphisms K' of open subsets X of X on open K'
subsets of lR.n such that (6.3.1) and its inverse are infinitely differen-
tiable for every KE.9F. The simple verification is left for the reader.
(Clearly every extension of ff satisfying i) is contained in this family
ff'. That ff' satisfies i) and ii) and contains ff follows from the fact
that ff satisfies i) and ii).) A Coo structure can thus be defined by an
arbitrary family ff satisfying i) and ii) only, but if the condition iii) is
dropped there are many families defining the same structure. Such a
family is called a Coo atlas, and two atlases are called equivalent if
they define the same Coo structure.
Definition 6.3.2. Let X be a Coo manifold. A function U defined in X

will be said to be in Ck(X) or in ~oc(X) if for every coordinate system
the composite function (K-1)*U defined by
(K- 1)* u(X)=U(K-1(X)), XEX K ,
We leave as an exercise for the reader to verify that it is sufficient

to require that U 0 K- 1 be in CkC~ K) or in ~oc(X K) for every K in an
atlas. Also note that if v is a function with compact support in XK and
we set
U=v0 K in X"' U = 0 elsewhere,
it follows that UECk(X) if and only if VECk(X,,); the analogous state-

ment is valid for ~oc. We shall somewhat incorrectly denote U by v 0 K.
To motivate our definition of a distribution in X we shall now

give a slightly different description of Ck(X). Thus let UECk(X) and set
UK=UOK- 1
where K is an arbitrary coordinate system. UK is then in Ck(X K)' and
since for any other coordinate system we have U= UK 0 K = UK' 0 K' in
X" n X"" it follows that
(6.3.2)
Conversely, if to every coordinate system K we are given a function u"
in X" in such a way that (6.3.2) is valid for any two coordinate
systems K and K', then there exists one and only one function U in X
such that U,,=UOK- 1 for every K, and UECk(X) if and only if
UK E Ck(X J for every K. In analogy with this description of a function
in Ck(X) as a system of functions U"ECk(X,J satisfying (6.3.2), we can
introduce distributions on a manifold:
Definition 6.3.3. Let X be a ex: manifold. If to every coordinate

system K in X we are given a distribution U"Eg()'(X J such that
(6.3.3)
we call the system UK a distribution u in X. The set of all distributions
in X is denoted by g()'(X). Similarly we define g()'k(X).
It is convenient to use the notation u" = U 0 K- 1 as in the case

discussed above where U was a continuous function on X. Thus !£'(X)
now appears as an extension of Co(X) if we identify a function
UECO(X) with the system U,,=UOK- 1 . The following theorem shows in
particular that Definition 6.3.3 coincides with our previous one if X is
an open subset of 1R n.
Theorem 6.3.4. Let ,?, be an atlas for X. If for every KE!#' we have a
distribution U"Eg()'(X,,), and (6.3.3) is valid when K and K' belong to .'iT, it
follows that there exists one and only one distribution UE!0'(X) such
that U 0 K- 1 = u" for every KE?
Proof Let tjJ be an arbitrary COO coordinate system in X. We first note

that there exists one and only one distribution UIj/ in ~'(X Ij/) such that
for every K
UIj/=(KotjJ-l)*U" in tjJ(Xlj/nX,JcXIj/'
In fact, this follows from Theorem 2.2.4 in view of the hypotheses in
Theorem 6.3.4. In particular, if tjJE!#' we have UIj/=ulj/ since we may
choose K = 1/1. Furthermore, it is immediately verified that the distri-

butions U", satisfy (6.3.3) for any two coordinate systems K and K'.
Hence they define a distribution U with the required properties. This
proves the theorem, for the uniqueness of U is a trivial consequence of
the proof.
The reader may have asked himself why we did not define !'.&'(X)
as the space of continuous linear forms on CO'(X). The reason for this
is that if fEC(X) and ¢ECO'(X) we have no invariant way of integrat-
ing f ¢ in order to identify f with such a linear form. However, we
would have obtained something rather close to !'.&'(X). In fact, let U be
a continuous linear form on CO'(X). Then u defines a distribution
U"E!'.&'(X,J by
(We define ¢oK as 0 outside X".) If ¢ECO'(K'(X" n X"'» then

u",(¢) =U(¢oK')=U(¢ 01/1 10K) =U,,(¢o 1/1 I)
where 1/11 =K'oK- I. In view of (6.1.1) this means that
(6.3.4)
where 1/1 = K K,-I. Conversely, assume given distributions u" in X"
0
satisfying (6.3.4) for all K in an atlas ~ Choose a partition of unity 1

2:
= Xj with XjE CO'(X,,) for some KjE~ Then
U(¢)=2:<u"j,(Xj¢) OKj l), ¢ECO'(X),
is a continuous linear form on CO'(X). We leave as an exercise for the
reader to verify by means of Theorem 6.1.2 that U" = u" for every
KE~ The continuous linear forms U on CO'(X) can thus be identified
with the systems U"E!'.&'(X,,) satisfying (6.3.4). They are called distribu-
tion densities. If all u" are continuous (COO) we say that U is continuous
(COO). When U is a distribution and ¢ is a COO density it follows from
(6.3.3) and (6.3.4) that ¢u is a distribution density, so
<u, ¢) =<¢u, I)
is defined if ¢ in addition has compact support. Again we leave as an
exercise for the reader to verify that this identifies !'.&'(X) with the
space of continuous linear forms on CO' densities.
As soon as a strictly positive COO density a is chosen in X, the map
u---+au
identifies distributions with distribution densities and functions with
densities. This is the case if X clRn where the Lebesgue measure is
used to give such an identification. Another important case is any
Riemannian manifold.
6.4. The Tangent and Cotangent Bundles
Having introduced the notion of COO manifold we shall now discuss

some basic facts on differential calculus on manifolds. These will not
be used until Chapter VIII. Throughout we denote by X a COO
manifold.
Definition 6.4.1. The vector space T,,(X) of tangent vectors to X at x is

the space of real distribution densities t in X of order one with
support at x and t(1)=0.
To justify the definition we observe that if X c1Rn then Theorem

2.3.4 gives for some t1, ... ,tnE1R
n
t(¢) = I tjo¢(x)jOXj
1
so t(cjJ) is the derivative of cjJ at x in the direction t, which it is natural

to consider as a tangent vector at x. If X is a general manifold and x
is in the coordinate patch X" then ~(¢) = t( ¢ K) has the form
0
t"(cjJ)=ItjOjcjJ(KX), cjJEC~(X,J.
Thus Tx(X) is always a vector space of dimension n = dim X, and we
have identified U T,,(X) with X" x1Rn. If xEX"nX", then
xeXpe
where f =K' 0 K- 1: K(X" nX",)-+ K'(X" nX",). Thus

tj' = I odjt~, that is, tIC' = !'(KX) tIC.
This is a COO map so the atlas consisting of the maps

U T"(X)3t-+(K(X),t")EX,,x1Rn
xeXK
makes T(X) a Coo manifold. These observations are summed up by

saying that T(X) is a vector bundle over X with fiber dimension n:
Definition 6.4.2. A Coo real vector bundle over X with fiber dimension
N is a Coo manifold V with
(i) a Coo map n: V -+ X called the projection,
(ii) a vector space structure in each fiber v" = n-1(x),
(iii) local isomorphisms between V and the product of open sub-
sets of X and 1RN.
Explicitly condition (iii) means that for each XEX there is an open
neighborhood Y and a COO map 1/1 of Vy=n-l(y) onto YxIR N such
that 1/1-1 is also In Coo, I/I(~)={x} xIRN for every XEY, and the
composed map
~->{X} xIRN ->IRN
is .a linear isomorphism. In the case of T(X) we have defined isomor-

phisms
n- 1 XIC->XIC xIRn->x lC xIRn
with these properties. One calls T(X) the tangent bundle of X.
Let V be any vector bundle over X and choose an open covering
{XJiEI of X such that for each i there is a Coo map I/Ii of n- 1 (X i ) onto
Xi x IR N with the properties listed above. Then
gij = I/Ii I/Ij 1
0
can be regarded as a Coo map from X;rlXj to the group GL(N,IR) of

invertible N x N matrices with real entries, and we have
(6.4.1) gijgji=identity in XinX j;
gijgjkgki=identity in XinXjnX k.
A system of such N x N matrices gij with Coo coefficients is called a
system of transition matrices. One can recover the bundle V from
them by forming the set V' of all (i, x, t)EI x X x IR N such that XEX i,
and defining (i, x, t) to be equivalent to (i', x', t') if X=X' and t' = gi,;t. It
follows from (6.4.1) that this is an equivalence relation, and it is easily
proved that the quotient of V' by this equivalence relation is a vector
bundle. It is isomorphic to V if gij were obtained from local triv-
ializations of the vector bundle V as explained above. We shall
sometimes find it convenient to look at a vector bundle in this way
which is directly suitable for calculations.
A vector bundle is thus a family of vector spaces Vx , XEX, varying
smoothly with x. If Y c X then a section u of V over Y is a map
Y3Y-> U(Y)E v" that is, U is a map from Y to V with nou=identity.
Since V is a Coo manifold the set Ck(y, V) of C k sections is well de-
fined for k=O,l, ... or k=oo. If we have a covering X=UX i as above
with local trivializations I/Ii of l'Ix, then Ui=l/IiouECk(YnXi,IRN)
and
(6.4.2) ui = gijuj in Y n Xi n Xj'
Conversely, any system UiE Ck(y n Xi' IRN) satisfying (6.4.2) defines a
section of the vector bundle. We can therefore also define say the
space of distribution sections !i&'(Y, V) (if Y is open) as the space of all
systems UiE.@'(Y n Xi' JR N) satisfying (6.4.2). We could also allow the

distributions u i to be complex valued which strictly speaking means
that we complexify the bundle V. (The definition of complex vector
bundles is obtained by substituting <C for JR in Definition 6.4.2.)
Examples. 1) A line bundle Q over X is defined by taking
gK'K=ldet(KoK'-IYIOK' in XKnX K,
for arbitrary coordinate patches X K and X K , in X. The sections of Q
are then the densities introduced in Section 6.3.
2) A Coo section U of T(X) over Y is called a Coo vector field in Y.
If ¢EC 1 (y), then ¢~u(¢) is a first order differential operator in Y
with no constant term, that is, U annihilates constants.
The map from T(X)lx to X K X JRn above is a special case of a very

K
general construction: If XI' X 2 are Coo manifolds and f: X 1 ~ X 2 is a

C'X) map, then we have a map f*: T(X I) ~ T(X 2) with
f* Tx(XI)~Tf(x)(X2) defined by
U*t)(¢)=t(¢of), ¢EC~(X2)' tET(XI)'
If local coordinates are introduced in Xl and in X 2' then the matrix

of f* is just the Jacobian matrix with respect to these coordinates, so
we shall sometimes use the notation f' instead of f*. In particular, if
f: JR~X is a curve this defines the tangent vector f'E1f(X) as the
image of d/dt on JR.
The dual V' = U V; of a vector bundle V is of course also a vector
x
bundle since any trivialization of V defines a trivialization of V'. The
transition matrices of V' are t gij 1. In particular, the dual T*(X) of
T(X) is a vector bundle called the cotangent bundle. For any
¢ E Ck(X), k > 0, the differential
T,,(X)3t~ t(¢)
at x is an element d¢(x)ETx*(X), so ¢ defines a C k - 1 section d¢ (or

¢') of T*(X). An arbitrary section of T*(X) is called a one form.
We shall now look at T*(X) in local coordinates. If
X K3X ~ K(X)=(Xl' ... ,Xn)EXKcJR n
is a local coordinate system, we have identified T(X)lx K with X x JR n

K
so that (x, t) is the tangent vector

Now <t, 0 = L: t/Jjoxj (L: ~jXj)' so in the corresponding trivialization

of T*(X) the form L: ~jdxj at x corresponds to (x, ~), ~E1R.n. If Xj and
~ j are regarded as functions on X X 1R. n then L: ~ jdx j can also be
K
considered as a differential form there. Its value on the tangent vector

t to XK X 1R.n is (11:* t, 0 if n is the projection X Ie X 1R.n --+ X Ie' Thus a
one form w on T*(X) is invariantly defined by letting its value on the
tangent vector t at YET*(X) be
<t, w) = <n* t, y).
In the standard local coordinates in T*(X) we have w = L: ~jdxj' One

calls w the canonical one form on T*(X).
If f: Xl --+ X 2 is a Ck map, we obtain a map
f*: C k- l (T*(X 2 ))--+ Ck-l(T*(X l )),
called the pullback of the one form by j, if we define for a one form u
in X 2 and a tangent vector tE T(X)
<tJ*u) =<f*t, u).
For example, if ¢EC1(X) then ¢' is a map X --+ T*(X) which we can
use to pull back the canonical one form w from T*(X) to X. As is
obvious in the local coordinate representation of w we obtain
(6.4.3) (¢')*w=d¢.
This is in fact an alternative characterization of the one form w.

lf Y is a COO submanifold of X then the conormal bundle N(Y) of
Y is defined as
{yE T*(X), n* y = yE Y and y vanishes on 1'y (Y)}.
If we introduce local coordinates x = (x l ' ... , xn) in X such that Y is

defined by Xl = ... =xk=O then N(Y) is defined by Xl = ... =Xk= ~k+l = ...
=~n=O in the corresponding coordinates in T*(X). Thus N(Y) is a
vector bundle of total dimension n. The definition of w or the coor-
dinate representation of w both show at once that the restriction of w
to N(Y) is equal to O. N(Y) is the dual of the bundle on Y with fiber
~(X)/~(Y) which is called the normal bundle. We shall hardly ever
use it and will therefore often drop the prefix in conormal when no
confusion seems likely.
Another important vector bundle over X is the bundle Ak T*(X)
for which the fiber at x consists of the alternating multilinear forms
on Tx(X). If ~l' ... , ~kET*(X)x then
T,,(X)k3(tl' ... , t k) --+ det <t;, ~j)~.j=1
is an element in Ak T*(X)x denoted by ~11\ •.• 1\ ~k' and the whole fiber
is spanned by such elements. A bilinear product
AkT*(X)x x Al T*(X)x --+ A k + 1 T*(X)x,
also denoted by 1\, can be uniquely defined so that the product of
~ 1 1\ ... 1\ ~k and 11t 1\ ... 1\ I'll is ~ 1 1\ ... 1\ ~k I\1h 1\ ... 1\ 1'/1' A Coo section
of Ak T*(X) is called a k form. It can always be written as a linear
combination of forms of the special type
(6.4.4)
even with fl' ... ,fk chosen among local coordinates. There is a unique
first order differential operator d from k forms to k + 1 forms such that
dUodfl 1\ ... 1\ dJ;J = dfo 1\ dfl 1\ ... 1\ dfk;
we have = 0, and for any k form f with df = 0, k > 0, one can locally
d2
find a k -1 form u such that f = duo (COO functions are considered as 0
forms.) If t/I: Y --+ X is a Coo map and f is a form on X, then t/I* f can
be defined just as in the case of 1 forms, and we have
dt/l* f = t/I* df
since this is true for 0 forms (the invariance of the differential) and
consequently also for forms of the type (6.4.4).
One calls X oriented if an atlas is given such that for any 1(, 1(' in
the atlas the Jacobian of I( 01(' - 1 is positive where it is defined. If X is
of dimension nand f is an n form in X, we have
(l(-l)*f=fKdxl 1\ ... I\dxn in XK
where I(X)=(Xl' ... ,xn ). Since
h,=(dett/l')t/I*h in 1('(XKnXK,), t/I=l(ol('-1
for arbitrary local coordinates we obtain the transformation law
(6.3.4) when I( and 1(' are in the atlas defining the orientation, for the
Jacobian is then positive. (Note that on the other hand every n form f
which is different from 0 at every point defines an orientation where
the atlas consists of all I( such that fK > O. One says that X is oriented
by f > 0.) Hence f defines a distribution density, which as linear form
on C~ is denoted by
fiJ --+ S fiJf, fiJEC~(X).
X
If we have instead a k formf and an oriented k dimensional sub-

manifold Y of X then a distribution density with support in Y is
defined by
fiJ--+JfiJf, fiJEC~(X).
Y
The right-hand side is interpreted by pulling 4J and / back to Y by

means of the inclusion map Y ~ X. If / is a k -1 form then
0= f d(4Jf) = f 4Jd/+ f d4J 1\/
Y Y Y
which is clear if 4J has support in a coordinate patch. Suppose now

that Me Y is an open subset with smooth boundary aM and that M
is compact. Letting 4J approach the characteristic function of M we
then obtain as in Section 3.1
(6.4.5) f d/ = f f (Stokes' formula.)
M ilM
Here aM is oriented by taking a local coordinate system Y1' ... , Yk for

Y with Y 1 = 0 on oM and Y 1 < 0 in M; if this is in the positive atlas for
M then Yz, ... 'Yk is in the positive atlas for oM. This ends our brief
review of the differential calculus on manifolds; the reader not pre-
viously familiar with this topic should consult a detailed exposition or
supply the missing details. However, we shall make some additional
comments on how the cotangent bundle occurs in the study of differ-
ential operators.
If X is a manifold and E, F two (complex) vector bundles on X,
then a linear differential operator from sections of E to sections of F
is a linear map COO (X, E) --+ COO (X, F) which is a differential operator
in the usual sense in terms of local coordinates in X and local frames
for E and F. Thus let (Xl' ... , xn) be local coordinates in XI( c X and
let e1, ... , eN (resp. /1' ... JM) be sections of E and F forming bases at
every point in XI(. Then the condition is that in XI(
(Pu)j= L ~kUk
ifu=Luke k is a section of E and PU=L(Pu)j~; here
~k=L~ka(x)oa
~
is a differential operator in the local coordinates. This condition is

obviously independent of the choice of local coordinates and of local
frames for E and F. We say that P is of order ~m if lod~m in the
sum. Then the principal symbol p(y) of P at YET*(X) with local
coordinates x, ~ is defined as the linear map from Ex to Fx , x=ny,
which in terms of the bases ek , ~ is given by
(6.4.6) (p(Y)U)j = L ~k~(X)~~Uk·
1~I=m
The definition is in fact independent of the choice of local coordinates

and bases in the bundle, for it means that if 4JECOO(X) and (x,4J'(x»
= y, then for any section u of E we have at x

(6.4.6), p(y)u= lim erne-t'" P(et'" u),
t~oo
and the right hand side is invariantly defined. (Later on we shall

modify p slightly by a factor i'" but this is not important here.) To
prove the equivalence of (6.4.6) and the coordinate free definition
(6.4.6)' we just note that when (6.4.6)' is evaluated in terms of local
coordinates then we must let m derivatives fall on the exponential to
get a non-zero contribution.
If P is a scalar differential operator, that is, a differential operator
from COO(X) to Coo (X), then p is a function defined on T*(X). The
zeros of p (outside the zero section 0 of the cotangent bundle) are
called characteristics of P. A surface in X defined by </J = c is called
characteristic (at x) if p(x, </J'(x)) =0 when ¢(x)=c (at x), that is, the
conormal bundle (at x) is in the characteristic set. We shall now discuss
the problem of integrating the characteristic equation p(x, ¢'(x)) =0,
that is, the construction of functions with all level surfaces characteris-
tic. In this discussion we allow p to be any real valued Coo function
on T*(X), thus we drop the property of the principal symbol that it is
a homogeneous polynomial in each fiber.
Denote by ~ the section ¢' of T*(X). We must satisfy two con-
ditions:
(i) the section must lie in the zero set of p,
(ii) ~ must be of the form ¢'.
Now we know from (6.4.3) that (ii) implies
d¢=~*w.
Hence
(6.4.7)
Conversely, if (6.4.7) is fulfilled then ~*w is locally of the form d¢ and
we obtain locally a solution of the characteristic equation p(x, ¢'(x))
= O. The differential form
(6.4.8) a=dw
is called the symplectic form of T*(X) and plays a fundamental role in
what follows. (Later on we shall devote the entire Chapter XXI to a
study of the geometry to which it leads.) In the standard local coor-
dinates x, ~ in T*(X) we have
(6.4.8)' a=Ld~jl\dxj
which means that for two tangent vectors to T*(X) with the coor-
dinates (t', -r') and (til, -r") the symplectic form is
(6.4.8)" tj
t'.'
}
1='" (r~t~' - t'.r~') =(t" r') - (t' r").
1..-}}}} , ,
This is a non-degenerate bilinear form, that is, it vanishes for all

(til, r") if and only if (t', r') = O. Since a = dw we have of course da = O.
We can now reinterpret (6.4.7). Let S= {(x, e(x))} c T*(X) be the graph
of the section. It is an n dimensional manifold parametrized by x, and
(6.4.7) means precisely that the restriction of a to S is equal to 0, that
is, the pullback of a to S by the inclusion S<--+ T*(X) is equal to O. In
other words, the tangent plane of S is at every point y orthogonal to
itself with respect to the bilinear form a. Since a is non-degenerate the
annihilator of a k-dimensional subspace of TiT* (X)) with respect to
a is necessarily of dimension 2n - k, so it follows that the tangent
plane of S must be its own annihilator.
Summing up, it follows from (i) and (ii) that the n dimensional
manifold S = {(x, e(x))} c T*(X) satisfies the conditions
(i)' p=OonS
(ii)' the restriction of a to S is equal to 0, that is, the tangent plane
of S is its own annihilator with respect to a at every point in S.
Conversely, if S satisfies (i)', (ii)' and the restriction of the pro-
jection n to S is a diffeomorphism on X, then S is a section satisfying
(i) and (ii). Thus (i)', (ii)' represent a slightly generalized, geometrical
form of our problem.
That p=O on S implies that dp=O on the tangent planes of S. To
exploit this we define a vector field Hp on T*(X) by
(6.4.9) (t,dp) =a(t, Hp), tET(T*(X)),
which is possible since a is non-degenerate. In terms of local coor-
dinates (6.4.9) means if t=(tx, t~) and Hp=(hx' h~) that
(tx' op/ox) +(t~, op/oO =(t~, hx ) - (tx' h~),
that is, hx = op/oe, h~ = - op/ox. Thus
(6.4.10)
in terms of local coordinates. One calls H p the Hamilton vector field
of p. Now we obtain from (i)'
aCt, Hp)=O if YES and tE~(S),
so H p must be in the tangent plane of S in view of (ii)'. Hence S is

generated by integral curves of H p' that is, solutions of the Hamilton
equations
(6.4.11)
in local coordinates. Note that (6.4.11) implies
so p is constant on such a curve. If p vanishes at one point of the

curve it follows that p vanishes identically on it; the curve is then
called a bicharacteristic of p.
Theorem 6.4.3. Assume that So is an n -1 dimensional manifold

c T*(X) such that
(a) p=O on So,
(b) the restriction of a to So is equal to 0, that is, the tangent space
of So is self orthogonal with respect to a,
(c) Hp is not a tangent to So at any point of So.
Then the union of the bicharacteristics of p starting on So defines an
n dimensional manifold satisfying (i)' and (ii)' in some neighborhood of So.
Proof Let 4>(t)y be the solution of the Hamilton equations (6.4.11) at

time t when 4>(0)')'=1'. This map is defined for small t. Then
So xIR3(Y, t) --+ 4>(t)yET*(X)
is well defined in a neighborhood of So x {O}, and the differential is

injective there by hypothesis for its image contains both H p and T(So)
so the dimension is n by condition (c). The image of a suitably small
neighborhood U is therefore an n dimensional manifold S. Since
a(t,Hp)=<t,dp)=O for tET(So), and since a(Hp, Hp)=O, it follows
from (b) that (ii)' is satisfied at So, and (i)' is true since S is a union of
bicharacteristics. It remains to prove that (ii)' is valid in S. To do so it
suffices to show that
(6.4.12) 4> (t)* a = a
for the fact that a restricted to S vanishes at So then shows that this is
also true at 4>(t)So, hence in S.
To verify (6.4.12) we first observe that

4>(s + t) = 4>(s) 4> (t).
°
If we verify that the derivative of 4> (t)* a with respect to t is 0 when t
= it will therefore follow that this is true for any t, hence that
(6.4.12) is valid. In evaluating the derivative when t=O we use local
coordinates. Since
4>(t)(x, ~)=(x+top(x, Wo~, ~ -top(x, Wox)+O(t 2 ),
we have
<P(t)* a = 'I (d~j- td op/oxj) /\ (dx j + tdop/a~) + 0(t 2 )
=a+ t('Id~j /\ dOP/o~j-dop/axj /\ dxj)+ 0(t 2 )
=a- td 2 p+0(t 2 )=a+0(t 2 ).
This completes the proof of (6.4.12) and of Theorem 6.4.3. At the same
time we have proved that the solution operators <P(t) of the Hamil-
tonian equations are canonical transformations:
Definition 6.4.4. A Coo map <P from 0 1 c:: T*(X) to O 2 c:: T*(X) is
called canonical (or sympletic) if <P* a = a.
°
Note that <P* w n= wn and that wn is a 2 n form on T* (X) which is
nowhere since wn=n! d~1 /\dXl /\ ... /\d~n/\dxn in local coordinates.
Thus wn can be used to orient T*(X), and we conclude that the differ-
ential of a canonical transformation is always bijective.
Theorem 6.4.3 solves a geometrical form of the initial value prob-
lem for the equation p(x, ¢'(X» =0. The following is a more con-
ventional analytical version of the result:
Theorem 6.4.5. Let p be a real valued Coo function in a neighborhood of

(0, I'/)ET*(lR n) such that
(6.4.13) p(O,I'/)=o, op(O,I'/)/Ol'/n"4=O
and let tjJ be a real valued Coo function in IRn-l such that
otjJ(O)/oxj=I'/j, j<n.
Then there is in a neighborhood of OElR a unique real valued solution
n
¢ of the equation
(6.4.14) p(X, ¢'(X» =0
satisfying the boundary condition
(6.4.15) ¢(X',O)=tjJ(X'), o¢(O)/ox=I'/.
Here x'=(x 1 , ... ,xn_ 1 ).
Proof By the implicit function theorem the equation

p(x', O, otjJ(x')/ox', ~n)=O
has a unique solution ~n=~n(x') with ~n(O)=I'/n' provided that Ix'i is
small. Then
is an n -1 dimensional manifold satisfying (a) and (b) in Theorem

6.4.3. Condition (c) is fulfilled since the xn component oploe n of Hp is
not O. For the surface S given by Theorem 6.4.3 the projection n
restricted to S has therefore a surjective differential at (0,1]), so S is a
section of T*(]Rn) near O. Hence we obtain a solution </> of (6.4.14)
near 0 such that (x, </>'(X))ES, in particular
o</>(x', O)lox' = oljJ lox', o</>(O)lox = 1].
Thus cjJ(x', 0) = ljJ(x') + C so </> - C has the required properties. The
uniqueness follows from the discussion preceding Theorem 6.4.3 of the
various ways of stating (6.4.14), so this completes the proof.
Sometimes one wants to solve two equations

p(x, cjJ' (x)) = 0, q(x, cjJ' (x)) = 0
at the same time. Since the integral curves of the Hamilton field H p
must lie in the surface S defined by </>' and since along these curves
(6.4.16)
it follows that the Poisson bracket
(6.4.17)
must also vanish in S. The argument can be repeated to show that the
repeated Poisson brackets {p, {p, q}}, ... must vanish in S. When {p, q}
vanishes identically on {(x, e); p(x, e)=q(x, e)=O} and dp, dq are
linearly independent it is not hard to modify the proof of Theorem
6.4.3 to show that one can find S satisfying (i)' for p and q as well as
condition (ii)', so that S passes through a given manifold So of dimen-
sion n - 2 satisfying (a), (b) and a suitable form of (c). We shall not
pursue this further since a thorough discussion will be given in Chap-
ter XXI. However, the notion of Poisson bracket will be important in
Chapter VIII, and it should be kept in mind that by (6.4.16) the
Poisson bracket of two functions on T*(X) is invariantly defined.
Notes
As indicated at the end of Section 6.3 one can define E&'(X) when X is
a manifold as the dual of the space of COO densities of compact
support. This is a special case of the theory of currents of de Rham
[lJ, which also contains a study of distribution valued differential
forms of arbitrary degree. The results in Section 6.1 are thus essen-
Notes 157
tially contained in de Rham's theory, for composition with a map f

having surjective differential can locally be split into a tensor product
with the function 1 in some new variables and a change of variables
in the domain of f. The formula (6.1.5) is from John [5] though.
Just as we recalled differential calculus in Section 1.1 we have
summed up in Sections 6.3 and 6.4 the notions of manifold, tangent
bundle, cotangent bundle as well as general vector bundle, and the
calculus of differential forms. This should suffice to define our no-
tation but the reader previously unfamiliar with these topics will at
least have to look elsewhere for a detailed presentation of differential
forms. Besides the book of de Rham [1] he might consult for example
Warner [1] or Sternberg [1].
In Section 6.4 we also discussed the Hamilton-Jacobi integration
theory for first order non-linear differential equations. We chose a
geometrical presentation which leads to the notions of Hamilton
vector field, bicharacteristic and Poisson bracket which will all be
required for some results in Section 8.5. They form the basis for the
symplectic geometry which pervades a large part of the modern
theory of linear partial differential operators with variable coefficients.
We shall have to devote the entire Chapter XXI to developing all the
symplectic geometry required in our program.
The reader can very well postpone the study of Sections 6.3 and
6.4 until he reaches part III of this book. It is largely to make this
possible that we chose to define composition of distributions with
maps without reference to distributions on manifolds although the
reversed order would have been preferable conceptually.
Chapter VII. The Fourier Transformation
Summary
The Fourier transformation of a function UEV is defined by
u(~)= Jri(x·~>u(x)dx.
In Section 7.1 we extend the definition to all uE[I", the space of
temperate distributions, which is the smallest subspace of !!}' contain-
ing L1 which is invariant under differentiation and multiplication by
polynomials. That this is possible is not surprising since the Fourier
transformation exchanges differentiation and multiplication by coor-
dinates. (See also the introduction.) It is technically preferable though
to define [1" as the dual of the space [I' of rapidly decreasing test
functions. After proving the Fourier inversion formula and basic rules
of computation, we study in Section 7.1 the Fourier transforms of L2
functions, distributions of compact support, homogeneous distri-
butions and densities on submanifolds. As an application fundamental
solutions of elliptic equations are discussed. Section 7.2 is devoted to
Poisson's summation formula and Fourier series expansions. We re-
turn to the Fourier-Laplace transform of distributions with compact
support in Section 7.3. After proving the Paley-Wiener-Schwartz theo-
rem we give applications such as the existence of fundamental so-
lutions for arbitrary differential operators with constant coefficients,
Asgeirsson's mean value theorem and Kirchoffs formulas for so-
lutions of the wave equation. The Fourier-Laplace transform of distri-
butions which do not necessarily have compact support is studied in
Section 7.4. In particular we compute the Fourier-Laplace transform
of the advanced fundamental solution of the wave equation. The
Fourier transformation gives a convenient method for approximating
COO functions by analytic functions. This is used in Section 7.5 to
prove the Malgrange preparation theorem after we have recalled the
classical analytical counterpart of Weierstrass.
Section 7.6 is devoted to the Fourier transform of Gaussian func-
tions and the convolution operators which they define. This prepares
7.1. The Fourier Transformation in .'l' and in .'l" 159
for a rather detailed discussion in Section 7.7 of the method of sta-

tionary phase, which is a fundamental tool in the study of pseudo-
differential and Fourier integral operators in Chapters XVIII and
XXV. The Malgrange preparation theorem plays an essential role in
many of the proofs. As an application of the simplest form of the
method of stationary phase we introduce in Section 7.8 the notion of
oscillatory integral. This gives a precise meaning to equations such as
b(~)=(21t)-n f ei(x.odx
and will simplify notation later on. In Section 7.9 finally we continue
the proof of IJ' estimates for convolution operators started in Section
4.5. Applications are given concerning the regularity of solutions of
elliptic differential equations with constant coefficients. Although the
results are very important in the study of non-linear elliptic differen-
tial equations they will not be essential in this book so the reader can
skip Section 7.9 without any loss of continuity.
7.1. The Fourier Transformation in [/ and in [/'

The purpose of Fourier analysis in lR" is to decompose arbitrary
functions into usually continuous sums of characters. By a character
one means an eigenfunction for the translations, that is, a function f
such that for every YElRn
f(x + y)= f(x)c(y), x ElRn ,
for some c(y). If f(O)=O we conclude that f vanishes identically.
Excluding this uninteresting case we can normalize f so that f(O) = 1.
Then x=O gives f(y)=c(y), hence
(7.1.1) f(x+y)=f(x)f(y), f(O) = 1.
Assuming that f is continuous we obtain if gEC~ and ffgdy=l
f(x) = f f(x+ y)g(y)dYECoo
(Theorem 1.3.1). Differentiation of (7.1.1) with respect to y gives when
y=O
oJ=aJ,
and since f(0) = 1 it follows that
(7.1.2) f(x)=exp <x, a), <x, a) = L" xjaj.
1
160 VII. The Fourier Transformation
Conversely, the exponential (7.1.2) satisfies (7.1.1) so we have de-

termined all continuous characters.
Which characters are needed for the expansion of a given function
u depends on the properties of u. We shall mainly consider functions
and distributions which are fairly well behaved at 00 and shall only
use bounded characters then, that is, take a purely imaginary in
(7.1.2).
Definition 7.1.1. If fELl(JRn) then the Fourier transform f is the

bounded continuous function in IRn defined by
(7.1.3)
If f also happens to be integrable one can express f in terms of f

by Fourier's inversion formula
(7.1.4)
so f(e) is the density of the character ei(x.;) in the harmonic decom-
position of f To study the Fourier transform and in particular to
prove (7.1.4) we shall first consider functions in a subset of COO
containing CO".
Definition 7.1.2. By Y or Y(IRn) we denote the set of all ljJEcoo(IRn)

such that
(7.1.5) sup IxfJ oaljJ(x)1 < 00
x
for all multi-indices rx. and p. The topology in Y defined by the semi-
norms in the left-hand side of (7.1.5) makes Y a Frechet space.
The importance of the class Y is due to the following result, where
we use the notation Dj = -ioj which is much more convenient as soon
as the Fourier transformation is involved. Note that DjYe Y, XjY
eY, and that YeV.
Lemma 7.1.3. The Fourier transformation ljJ -+ ¢ maps Y continuously

into Y. The Fourier tr~nsform of DjljJ is ej¢(e), and the Fourier
transform of xjljJ is - DjljJ.
Proof Differentiation of (7.1.3) gives

Da¢(e)= Sri(X,O( -xYljJ(x)dx
and is legitimate since the integral obtained is uniformly convergent.
Hence ¢ECoo and Da¢ is the Fourier transform of (_x)aljJ. Integrating
by parts we also obtain
7.1. The Fourier Transformation in [I' and in [1" 161
(7.1.6)
These operations are legitimate since 41EY. Hence
sup I~.B Da$(~)1 ~ C sup (1 + Ixlt+ IID.B« -x)a41(x»1
x
where C=J(1+lxl)-n- dx, so the Fourier transformation maps Y

1
continuously into Y. When 0(=0 we obtain from (7.1.6) that ~fl$ is

the Fourier transform of Dfl41, which completes the proof.
Lemma 7.1.4. If T: Y --+ Y is a linear map such that

TDj41=DjT41, TXj41=xjT41, j=1, ... ,n, 41EY,
then T 41 = c 41 for some constant c.
Proof If 41EY and 41(y) = 0 then we can write
41 (x) = L (Xj - Yj)41j(x)
where 41jEY. In fact, Theorem 1.1.9 shows that we can do so with
41jECOO, and for x=I=y we could take 41ix)=41(x)(xj-Yj)lx-YI-2 which
behaves at OCJ as a function in Y. Combining these two choices by a
partition of unity we obtain 41j with the desired properties. Hence
T41(x)=I.(xj-y)T41ix)=0 if x=y.
It follows that for all 41EY
T 41 (x) = c(x) 41 (x)
where c is independent of 41. Taking 41 =1= 0 everywhere we obtain
CECoo. Now
0=DjT41-TDj41=(Djc)41, 41EY,
so c must be a constant.
Theorem 7.1.5. The Fourier transformation F: 41 --+ $ is an isomorphism

of Y with inverse given by Fourier's inversion formula (7.1.4).
Proof By Lemma 7.1.3 F2 maps Y into Y and anticommutes with Dj

and Xj' With the notation R41(x)=41( -x) we conclude from Lemma
7.1.4 applied to T=RF2 that RF 2=c. To determine c we can take
41(x)=exp( -lxI2j2), which is a function in Y. Then (xj+iD j) 41=0 so
(-Dj+i~)$(~)=0,j=1, ... ,n. Hence $=c 1 41 where c 1 =$(0)=(2nt/ 2
by (3.4.1),. Thus F2 41 =ci 41, so c=ci =(2 nt which completes the proof.
Another interesting determination of the constant c is as follows.

Assume n = 1, which is sufficient since we can otherwise take for f a
product fl (x 1) ... f,,(xn)' Splitting the integral (7.1.3) at 0 and integrat-

ing by parts we obtain f(~) = F+ (~) + F_ where m
00 00
F+ (~) = Jf(x)e-iX~dx = f(O)/i~ +f'(0)/(i~)2 + Jf"(x)e-ix~dx/(i~)2

o 0
=f(0)/i~+O(I~I-2) as ~--+oo, Im~~O,
o
F_(~)= J f(x)e-ix~dx= - f(O)/i~ +O(l~1-2) as ~ --+ 00, 1m ~;?; O.
-00
Assume fEC,{; which makes F_ and F+ entire analytic functions. If")lR

is the circle ICI =R oriented counterclockwise and )Ii are the half
circles in the upper and lower half planes respectively, then Cauchy's
integral formula gives
Jf(~)d~= lim J F+WdC- J F_WdC
'Vi
R-+oo 1;'
= lim J f(O)/iC dC =2nf(0).

R-oo YR
The constants 2n in Cauchy's integral formula and in the inversion

formula are therefore "the same", and one is often free to choose
between using the Fourier inversion formula or Cauchy's integral
formula.
Instead of relying on Lemma 7.1.4 we could also have verified
directly that Fourier's inversion formula must be valid for some
constant c in place of (2nt. What is involved is computing the double
integral
Since the double integral does not converge absolutely, the order of
integration cannot be inverted so we must introduce a factor which is
a function of ~ to produce convergence. Thus choose I/IEY' with 1/1(0)
= 1 and note that by dominated convergence
J$(~)ei<x.~) d~ =lim
__ 0 JI/I(e~)$(~)ei<X.~) d~
=lim HI/I(e~)¢(y)ei<X-Y.~) d~dy .

• _0
In this absolutely convergent double integral we integrate first with

respect to ~ and obtain
J$(~)ei<x.~) d~ =lim
._0 J¢(y)t/I((y - x)/e)dy/gn
=lim J¢(x +ez)t/I(z)dz = ¢(x) Jt/I(z)dz.
• -0
We shall now prove some fundamental properties of the Fourier

transformation on !7.
7.1. The Fourier Transformation in [/ and in [/' 163
Theorem 7.1.6. If 4> and t/J are in [/', then

(7.1.7)
(7.1.8)
(7.1.9)
(7.1.10)
--
J$t/Jdx= J4>tP dx,
J4>l/idx=(2n)-" J$$dx
~~
4>*t/J =4>t/J,
q;Ifi =(2n)-"$*tP.
(Parseval's formula),
Proof Both sides of (7.1.7) are equal to the double integral

JJ4>(x)t/J(~)e-i<x·~>dxd~.
To prove (7.1.8) we set x=(2n)-ntP and obtain using the Fourier
inversion formula
xm =(2n)-" JtP(x)ei<x.Odx=t/J(~).
Hence (7.1.8) follows if we apply (7.1.7) with t/J replaced by X. The
proof of (7.1.9) is as elementary as that of (7.1.7) and is left for the
reader. To obtain (7.1.10) finally we note that the Fourier transform of
;fifo is (2n)"4>(-x)t/J(-x) and that the Fourier transform of $*tP is
(2n)" 4>( -x)(2n)"t/J( -x) in view of (7.1.9) and the Fourier inversion
formula. The proof is complete.
Definition 7.1.7. A continuous linear form u on [/' is called a tem-

perate distribution. The set of all temperate distributions is denoted
by [/".
The restriction of a temperate distribution to C~(lR") is obviously

a distribution in ~'(IR"). We can in fact identify [/" with a subspace of
~'(lRn) since the following lemma shows that a distribution UE[/"
which vanishes on C~(lR") must also vanish on !I'.
Lemma 7.1.8. C~ is dense in !I'.
Proof Let 4>E[/' and take t/JEC~ such that t/J(x)=l when Ixl~1. Put
4>.(x) = 4>(x) t/J(ex). Then it is clear that 4>.E C~, and since
4>.(x)-4>(x)=4>(x)(t/J(ex)-l)=O if Ixl<l/e,
we conclude that 4>.-+4> in [/' when e-+O.
It is obvious that Iff' c [/". Other examples of elements in [/" are

measures dll such that for some m
W+ Ixl)-mldll(X)1 < 00.
In particular, this implies that lJ'(1R n) c [/' for every p. It is also clear
that [/' is closed under differentiation and under multiplication by
polynomials or functions in Yo
Definition 7.1.9. If UE[/', the Fourier transform U is defined by

(7.1.11)
It follows from Lemma 7.1.3 that UE[/', and since the proof of
(7.1.7) is valid for all cP, t/JELl the preceding definition agrees with
(7.1.3) if fELl.
Fourier's inversion formula as proved in Theorem 7.1.5 states that

¢=(2ntcp if cPE[/, cp(x) = cP( -x).
If U is in [/' we obtain
Here it is of course the composition of u and x ~ - x. Thus. we have
Theorem 7.1.10. The Fourier transformation is an isomorphism of [/'

(with the weak topology), and Fourier's inversion formula li=(2ntu is
valid for every UE[/'.
In particular, fELl and JEV then the inversion formula (7.1.4) is

valid for almost every x.
Theorem 7.1.11. If uEL2(1Rn) then the Fourier transform U is also in

L2(1Rn) and Parsevars formula (7.1.8) is valid for all cP, t/JEL2.
Proof Choose a sequence UjEC~ such that Uj~U in L2 norm. Then

Iluj - uk lli2 = (2n)n IIUj -Uk IIi2 ~ 0
by (7.1.8) which is already proved in Yo In view of the Riesz-Fischer
theorem it follows that there is a function UEL2 with Uj~ U in L2,
and U =u by the continuity of the Fourier transformation in [/'. Now
both sides of (7.1.8) are continuous functions of cP and t/J in the L2
norms, so (7.1.8) is valid for arbitrary L2 functions.
If UElJ' and 1 ;£p;£2, we can write U as the sum of a function in L2

and one in Ll, so the Fourier transform is in L7oc' A better result
follows from the Riesz-Thorin convexity theorem:
Theorem 7.1.12. If T is a linear map from lJ" n lJ'2 to H' n H2 such

that
7.1. The Fourier Transformation in ;? and in ;?' 165
(7.1.12)
and if l/P=t/Pi +(1-t)/P2' l/q=t/qi +(1-t)/q2' for some tE(O, 1), then
(7.1.12)' IITfllq~MtiMi-t Ilfll p, fEIfI(1If2.
Proof We may assume P < 00 for otherwise Pi = P2 = 00 and (7.1.12)'
follows then from Holder's inequality. The method of proof is similar
to that of Theorem 4.5.1. First we write (7.1.12) in the form
I<Tf, g)1 ~ Mjllfllp)gllqj' l/qj+ l/qj= l.
If O~F, GEV and IIFIIL' = IIGIIL' = 1 then the absolute value of
is ~1 when Rez=O or Rez=l provided that Ifol~l, Igol~1. If we

take for F and G functions which take only finitely many values, it is
clear that cP is analytic and bounded when 0 ~ Re z ~ 1. By the
Phragmen-LindelOf theorem the bound 1 is also valid in the interior,
thus
I<T(foF i /P), goG i /q')l ~Mti Mi -t, l/q + l/q' = l.
This proves (7.1.12)' for a dense subset of IfI(1 If2. By hypothesis both
sides are continuous in Ifl (1 If2 which completes the proof.
Theorem 7.1.13 (Hausdorff-Young). If fEIf and 1 ~p~2 then JEIf',

l/p + l/p' = 1, and
(7.1.13) IIJIILP' ~(2n)n/p' IlfIILP'
Proof (7.1.13) follows from Parseval's formula when p=2 and is

obvious when p = 1, so it follows in general from Theorem 7.1.12.
In Section 7.6 we shall see that the Fourier transform of a function

in If may have positive order if p > 2.
Already Lemma 7.1.3 indicates that the Fourier transformation
exchanges local smoothness properties and growth properties at 00.
Another case of this is the following
Theorem 7.1.14. The Fourier transform of a distribution UEtS"(1R.n) is the

function
(7.1.14)
The right-hand side is also defined for every complex vector ~Eccn and
is an entire analytic function of ~, called the Fourier-Laplace transform
ofu.
Proof. If <pEC~(1Rn) we have

u(<p) = u(¢) =(ux®<p~)(e-i<X.~»= S<p@u(e- i<.. 0) d~
by Theorem 5.1.1 or rather its analogue for distributions of compact
support and test functions in Coo. It follows that is equal to the u
function (7.1.14) which is a Coo function in {:n by Theorem 2.1.3. It
satisfies the Cauchy-Riemann equations since we may differentiate
with respect to ~ directly on the exponential, and this proves the
theorem.
Note that u*~<··O=u(Oei( .. O. The properties of the entire ana-

lytic function u will be discussed further in Section 7.3.
Theorem 7.1.15. If U l Eg' and UzE~', it follows that u l * u2Eg' and

that the Fourier transform of U l * u2 is ul u2.
The product is defined since Theorem 7.1.14 shows that U2 EC oo .
Proof. If <PEC~ we have by definition of the convolution

(u l *u 2)(<P)=U 1 *u z *cP(O)=U l (U 2*<P)·
The right-hand side is a continuous linear form on Y, for if <PEg and
u 2 is of order k then u2 * <PEg and
L supIX~afJU2*<pI~Cj L suplx~afJ<pl.
IHfJl~j IHfJl~j+k
To compute the Fourier transform of u l * U2 we note that

(u 1 * u 2 )(¢) = u l (u 2 * ¢), <PEg.
If <PE C~ then u * ¢ is the Fourier transform of u <p, for this is
2 2
Se-i<X'~>u2(e-i( .. ~» <p(~) d~ =uz(¢(x + .»

by Theorem 5.1.1. Hence
(u l * U 2 )(¢)=U l (U Z <P) =(u l U2 )(<p) , <pEC~,
which proves the statement.
Theorem 7.1.15 contains the extension to g' of the basic proper-

ties of the Fourier transformation observed in Lemma 7.1.3,
(7.1.15) UEg',
for differentiation can be written as convolution with a derivative of
the Dirac measure at O. We can also obtain (7.1.15) by differentiation
7.1. The Fourier Transformation in 9' and in 9" 167
of
(7.1.16)
Alternatively we could also argue that (7.1.15) must be valid in fIJ'
since it is valid in the dense subset fIJ and both sides are continuous
in fIJ'. This argument shows also that if T is a linear bijection
lR" --+ lR" then
(7.1.17)
Theorem 7.1.16. If uEfIJ'(lR") is a homogeneous distribution of degree a,

u
then is homogeneous of degree - a - n.
Proof If M t denotes multiplication by then M~u=t"u so (7.1.17)

gives for t > 0
Remark. We shall see below (Theorem 7.1.18) that every homogeneous

UE.@'(lR n ) is automatically in fIJ'(lR").
Example 7.1.17. With the notation in Section 3.2 the Fourier trans-
form of X~ is e+ i1t (a+ 1)/2(~ +io)-a-l and that of (x ± iO)a is
2ne±i1taI2X±a-l(~) for every aECC. If k is an integer then the Fourier
transform of ~-l-k is ni-l-k(sgn~)~k/k! if k~O and 2ni- 1- kc5(-k-l)
if k<O.
The second statement follows from the first by the Fourier in-
version formula, and the third follows from the second and (3.2.10)',
(3.2.17)'. To prove the first statement we observe that when B > 0 and
Rea> -1 the Fourier transform of e-exx:.(x) is
f xae-x(Hi~)dx/r(a+ 1)=(B+ i~)-a-l f zae-zdz/r(a+ 1)

00 00
~ --+
o 0
where the last integral is taken on the ray with direction B+i~ and ~
is defined in CC slit along lR_ so that 1a = 1. In view of the Cauchy
integral formula the integral can be taken along lR+ so it is equal to
r(a+ 1). Hence the Fourier transform is
~ --+(B+i~)-a-l =e- i1t (a+ 1)/2(~ _ie)-a-l.
(Note that this explains (3.4.10).) When e --+ 0 it follows that the Fou-
rier transform of ~+ has the stated form, that is,
<X:', cl» =e- i1t (a+ 1)/2«~ - io)-a-l, ¢), ¢EfIJ,
if Re a > - 1. Both sides are entire analytic functions of a so the

equality must hold for all aE<C. The remaining part of the first
statement follows when x and ~ are replaced by - x and -~.
We shall now consider homogeneous distributions in JR n or in

JR n " o.
Theorem 7.1.1S. If UE~'(JRn) and the restriction to JRn"o is homo-

geneous of degree a, then UE[/'. If in addition UEcoo(JRn"o) then
UE COO(JR n" 0).
Proof Choose'" E Cg'(JRn " 0) satisfying (3.2.22). Since

00
'" o(x) = 1 - S"'(x/t) dt/t

1
is in Cg' we can write
00
u(</J) = U("'o</J) + Su(",(./t)</J)dt/t

1
00
=U("'o</J)+ S u(",</J(t.)) t"+n-1dt, </JEC~.
If Ko=supp"'o, K=supp"', and k is the order of U in a neigh-

borhood, then
L supID-cf>I+ S L tlal+Rea+n-lsupIDa</Jldt)
00
lu(cf»I~c(
lal::;;k Ko 1 I-I::;;k tK
~ C' L L suplxfJ D-cf>I, </JE Cg'(JR n),

I-I::;;k IfJl::;;lal+M
if M is an integer >Rea+n. This proves that UE[/'. Assume now that
UEC OO in JRn"O. If Rea< -n then u
is continuous, for u="'ou+
(1 - '" 0) u where the first term has compact support and the second
is integrable. In general we conclude that DfJ ~a which is the Fourier u,
transform of ( - x)fJ D"u, is a continuous function if Re a -Ial + IPI < - n.
Since this is true for any IPI if lal > Re a + IPI + n, we obtain
UE coo(JRn" 0).
Theorems 7.1.16 and 7.1.18 show that the Fourier transformation

is a bijection of homogeneous distributions of degree a in JR n which
are COO in JR n" 0 on such distributions of degree - n - a. Let us now
consider the Fourier transform U of a distribution u satisfying
(3.2.24)', that is,
u=t-k-nM!/tu+logt L S(x"u)(-1)kiNio/aL
lal~k
7.1. The Fourier Transformation in g and in g' 169
By (7.1.17) we have in 1R"

V =t-kM~ V +Q logt
where Q is the homogeneous polynomial of degree k
(7.1.18)
(Recall that Sx denotes integration over the unit sphere.) Thus
Vo= V +Qlogl.1
is homogeneous of degree k, for
t-kM~Vo=t-kM~V +Q(logt+logl.l)= V +Qlogl.l= V o.
lt follows that Vo is a bounded function in a neighborhood of 0, and
homogeneous of degree k and COO in 1R",,0; we have
(7.1.19)
We shall apply this to construct fundamental solutions, but first we
must introduce some terminology.
Definition 7.1.19. A homogeneous polynomial P(~) in n variables, with

complex coefficients, is called elliptic if P(~) =1= 0, 0=1= ~ E1R". An in-
homogeneous polynomial is called elliptic if the homogeneous part of
highest degree is elliptic.
Theorem 7.1.20. If P is a homogeneous elliptic polynomial of degree m

in 1R", then P(D) has a fundamental solution E such that
E=Eo-Q(x) log Ixl.
Here Eo is homogeneous of degree m-n and COO in 1R",,0, and Q is a
polynomial which is identically 0 when n > m and is defined when n;?; m
by
(7.1.20) Q(x)=S~«ix, Om-"/P(~))(2n)-"/(m-n)!.
Proof We define E so that £=(l/PY with the notation in Theo.Iems

3.2.3 and 3.2.4. Then PE=l"=b, hence P(D)E=(j, and (2n)"£=£ has
the properties just described.
Remark. Using the analyticity of P it is easy to show that Eo is

analytic in 1R"" O. We shall prove this in a general context in Section
8.6. (See also a remark after Theorem 7.1.22.)
Theorem 7.1.20 shows that every homogeneous elliptic operator

with constant coefficients satisfies the hypothesis in Theorem 4.4.1.
However, the main difficulty in the proof of Theorem 7.1.20, the

definition of (liP)"' can be avoided in this kind of application. Indeed,
the proofs of Theorems 4.4.1 and 4.4.2 work equally well if one just
has a parametrix:
Definition 7.1.21. If P(D) is a differential operator with constant coef-

ficients, then E E~' (1R. ft) is called a parametrix of P(D) if
P(D) E = 15+ OJ, OJECoo(IRft).
When constructing a parametrix one does not have to pay atten-

tion to the definition of liP at 0 since the Fourier transform of a
distribution of compact support is always in COO. More generally, we
have
Theorem 7.1.22. Every elliptic operator P(D) with constant coefficients

has a parametrix E which is a Coo function in IRft" O.
Proof We can write

P(~)=Pm(~)+Pm-l(~)+ ... +Po
where 1j is homogeneous of degree ~ and Pm(~) oF 0 when ~ oF O. Then
IPm(~)I~c>O when 1~1=1, so the homogeneity gives
It follows that for some constants C and R

IP(~)I~ IPm(~)I-IPm- 1(~)1- ... ~cl~lm - c(1~lm-l + ... + 1)~cl~lm/2
when ~ EIR ft and I~I ~ R. Since a derivative of order k of 1IP(~) is of the
form Q(~)IP(~)k+l with Q of degree ~(m-1)k, as is immediately
verified by induction, we conclude that when I~I > R
(7.1.21)
Choose XEC~(IRft) equal to 1 in {~; lei <R}. Then (l-X@)/P@
is a bounded COO function, hence the Fourier transform of a distribu-
tion EEg'. We have P(D)E=b+OJ where w= -x. Hence OJEg, and as
in the proof of Theorem 7.1.18 it follows from (7.1.21) that DP x a E is
continuous when 1131-10(1- m < - n. The proof is complete.
Remark. The error term OJ obtained in the proof has an analytic

extension to<cn by Theorem 7.1.14. It is also easy to show by deform-
ing the integration contour into the complex domain that E has an
analytic extension to a conic neighborhood of IRft"O. We leave the
7.1. The Fourier Transformation in Y and in Y' 171
proof as an exercise for the reader since more general results will be
given in Section 8.6. - The proof of Theorem 7.1.22 remains valid for
all P such that p(a)( ~)IP(~) --+ 0 for all o('~ 0 when ~ --+ 00 in 1Rn. Here
p(a)(~) = oa P(~). One just has to show that p(a)(~)IP(~) = O(I~I-Ialc) for
some C > 0 as ~ --+ 00 and then verify that Em = IIP(~) inherits the
same property for all 0(. (Thus c is independent of 0(.) The converse is
also true. We refer to Chapter XI for these and many related results.
We shall now determine the Fourier transforms of the homo-
geneous distributions studied in Section 6.2.
Theorem 7.1.23. Under the assumptions in Theorems 6.2.1 the Fourier

transform of (A±iO)(2-n)/2 is (n-2)cnldetAI-te+7tin-/2(B=FiO)-1, and
that of A*x~-n)/2 is
2n(n-2)j2Idet AI-t(ie±7tin±/2(B _ iO)-l - ie+ 1tin ±/2(B + iO)-l).
Proof If A had positive definite real part instead, then (6.2.1)" would
be valid. Taking Fourier transforms we would then find that the
Fourier transform of A(2-n)/2 is equal to
(n - 2) cn(det A)-tIBm
outside O. By Theorem 3.2.3 and Theorem 7.1.16 this must then be

true in the whole space. Now assume that A is just real and non-
degenerate, and apply this result to A.= -iA+elxI 2, e>O. As in the
proof of Theorem 6.2.1 we obtain when e --+ 0 that the Fourier trans-
form of
is equal to IdetAI-t e7ti (SgnA)/4(n-2)cn times the limit of l/iF.(~) where

F.(~) is the quadratic form with matrix
(a jk + eib jk )- 1 = (b jk ) -ei(b jk )2 + 0(e 2).

The limit is i- l(B(~) - iO)-l. In fact, since we have functions homo-
geneous of degree - 2 > - n it is sufficient to verify this in 1Rn'- 0 (by
Theorem 3.2.3), and then it follows from Lemma 6.2.2. We have
therefore proved the first statement in Theorem 7.1.23. The others are
then derived as at the beginning of the proof of Theorem 6.2.1.
We shall now compute the Fourier transform of a distribution UE

~'(1R n)
such that for some integer k
(7.1.22) (U, ¢) =sgn tt-k(u, ¢(t.), ¢ECg'(1Rn), t::j::O.
This means that u is homogeneous of degree - n - k and has parity

opposite to k (cf. (3.2.18)'). By Theorem 3.2.4 restricting to 1Rn ' - 0 gives
a bijection of such distributions in IR n on the distributions in IR n '-. °

having the same property when 4>EC~(IRn'-.o). Moreover, we have by
(3.2.23)"
(u, 4» =S(u(rk-l, 4>(t.» )/2, 4>E C~(IRn).
The formula extends by continuity to all 4>EY'. To compute (11,4»
=(u, ¢) we first calculate (Ck-l, ¢(tx» in terms of 4> when
x = (1, 0, ... , 0). Then ¢(t, 0, ... , 0) is the Fourier transform of
S4>(~l,~')d~' where ~'=(~2""'~n)' The Fourier transform of r k- 1 IS
2ni- 1 - k(Jk where
(7.1.23) (Jir)=2-1(sgn7:)7:k/k!, k=0,1, ... ;
(Jk(7:)=O(-k-l), k=-1,-2, ...
by example 7.1.17. Hence
(rk-l, ¢(tx» = 2n i- 1 - k J
«(Jk(7:), 4>(7:, ~'» d~'
=2ni- 1 - k«(Jk(x, 0), 4>(~».
Note that (J k is homogeneous of degree k so the last expression is

homogeneous in x of degree k. The orthogonal invariance shows that
it is equal to (l-k-l, ¢(tx» for every x=l=O, so we obtain for all 4>EY
(7.1.24)
or formally
(7.1.24)'
In Sections 8.2 and 12.6 we shall give a precise meaning to such

formulas. Note that when k<O they show that 11 is determined at ~ by
the restriction of u to a neighborhood of the orthogonal plane; for
k~O this is true for the singularities only. In Chapter XII we shall
return to (7.1.24) which is the essence of the Herglotz-Petrovsky
formula for the fundamental solution of hyperbolic equations. To
have a reference then we sum up the preceding results:
Theorem 7.1.24. If uEY'(IRn) satisfies (7.1.22), that is, u is homogeneous

of degree - n - k and of parity opposed to the integer k, then the
Fourier transform is given by (7.1.24) with (Jk defined in (7.1.23).
We shall close this section by studying Fourier transforms of

densities (simple layers) on submanifolds. First recall that the Fourier
transform of 0 0 is the Lebesgue measure dx, so by Fourier's inversion
formula the Fourier transform of dx is (2n)n oo . More generally:
7.1. The Fourier Transformation in ~ and in ~' 173
Theorem 7.1.25. If VeRn is a linear subspace and V.i is the orthogonal

space, then the Fourier transform of the Euclidean surface measure in V
is (2n)dimY times the Euclidean surface measure in V.i.
Proof By an orthogonal transformation we can make V defined by x"

=(xk+ l' ... , Xn)=O. Set x'=(x 1 , ••• , Xk)' Then
(7.1.25) s¢(x', O)dx' = (2n)k S cjJ(O, x")dx", cjJE!I',
for if <P(x') = S cjJ(x', x")dx", then <PE!I'(Rk), &(~')=¢(~', 0), so (7.1.25) is
just Fourier's inversion formula for <P. The theorem is proved.
Let dS y and dS Y1 be the surface measures in V and in V.i, and let

uoEL2(dS y). Then u=UodSyE!I"(Rn) and the Fourier transform is
uo(O with the coordinates just used. Hence Parseval's formula gives if
cjJECo(Rn )
R k - nS lu(~W ¢(~/R) d~ = S luo(~'W ¢(~'/R, ~") d~
~ S luo(~'W d~' S cjJ(O, ~") d~" = (2 n)k S IU ol2 dS y S ¢ dS y1, R ~ 00.
yl
We shall now extend these facts to general surfaces. As indicated by
the preceding formula, it is convenient to change notation and let k
be the codimension of V instead.
Theorem 7.1.26. Let K be a compact subset of a C 1 manifold M eRn of
codimension k, with Euclidean surface area denoted by dS. If u=uodS
where Uo is supported by K and square integrable with respect to dS,
then
(7.1.26) S lu(~Wd~~CRkSluoI2dS, R>O,
I~I<R
where C is independent of u and R.
Proof After using a partition of unity we may assume that in a

neighborhood of K the manifold M is of the form
x" =h(x'); x' =(x l , ... , Xn_k)EIR n - \ x" =(xn _ k+ 1"'" xn)ERk.
Here hECl. Then dS=a(x')dx' where a is a positive continuous func-
tion; U o is a function of x' and
u(~)= S e-i«x'. n+(h(x'J. nJ uo(x')a(x') dx'.
For a fixed ~" Parseval's formula gives
S lu(~Wd~' =(2n)n-k S IU ol2a 2dx' =(2n)"-k S IUol2 adS.
Integration with respect to ~" for I~"I <R leads to (7.1.26).
174 VIr. The Fourier Transformation
In spite of the simplicity of the proof, the estimate (7.1.26) is

optimal:
Theorem 7.1.27. Let uE!I", uEL~oc and assume that

(7.1.27) lim sup I lu(~)12 d~/Rk < 00.
R-oo I~I<R
If the restriction of u to an open subse/;"- X of 1R" is supported by a C l

submanifold M of codimension k, then it is an L2 density uodS on M
and
(7.1.28) I IUol2dS~Climsup I lu(~Wd~/Rk
M R-oo I~I<R
where C only depends on n.
Proof Choose an even function XECO'(1R") with support in the unit

ball and I Xdx = 1. The Fourier transform of u. = u* X.' where X.(x)
=C"X(X/B), is U(~)i(B~). Hence, with I II denoting L2 norm,
Ilu.1I 2=(2n)-" I lu(~W li(B~W d~ ~ C' B- kK(e)

where
K(e)= sup (2n)-" I lu(~Wd~/R\
.R>l I~I<R
00
C' = sup li(~W +

I~I < 1
L1 I~I
sup
> 2j - 1
li(~W2kj.
The intersection of supp u. with a compact subset of X belongs to the

set M. of points at distance < e from supp u eM when e is small. If
l/IEC'(j(X) it is geometrically evident and easily proved by a change of
variables that
e- k I Il/I(xWdx-+C k I Il/I1 2 dS, e-+O,
ME suppu
where Ck is the volume of the unit ball in 1R k. Hence

l<u,l/I)12=liml<u.,l/I)12~C"
_0
I
wwu_o
Il/II 2 dSlimK(e).
This proves that there is an L2 density U o dS on M such that for

l/IEC'(j(X)
<u, l/I)= I uol/ldS, IluoI2dS~C"limsup I lu(~Wd~/Rk.
R-oo I~I<R
The proof is complete.
The following is a substitute for Parseval's formula:

7.1. The Fourier Transformation in [/' and in [/" 175
Theorem 7.1.28. Let 4>EC o(JR n). If u=uodS is an L2 density with

compact support on a C 1 manifold M e1R. n of codimension k, then
(7.1.29) lim
R~oo
Jlu(~W4>(~/R)d~/Rk
=(2n)"-k J luo(xW( J 4>(~)d(J(~))dS(x)
M Nx
where dS is the Euclidean surface element on M and d(J is the Euclidean

integration element in the normal plane Nx of M at x, passing through
O.
Proof In view of Theorem 7.1.26 it suffices to prove (7.1.29) when U o
is continuous and 4>EC~(1R.n). Let 4>(~)=cP(~), thus <PEg;, and set
<PR(x)=Rn-k<p(Rx).
Then the Fourier transform of <PR is R-k4>(~/R) so the Fourier
transform of u*(~/R)R-k. Hence by Parseval's formula,
Jlu(~W 4>(~/R)d~/Rk =(2n)"(u* <PR, u).
It remains to compute the limit of u*<P R on M. In the formula
u* <PR(x) = Juo(y)<P(R(x- y))R"-kdS(y), xEM,
we may assume without restriction that M is given by x" = h(x') as in
the proof of Theorem 7.1.26, for the integral tends rapidly to zero
outside the support of u o' With the notation used there we can write
u* <PR(x', h(x')) = Juo(y') <P(R(x' - y'), R(h(x') - h(y')))Rn-k a(y') dy'
= Juo(x' - y'/R) <P(y', R(h(x')-h(x' - y'/R)))a(x' - y'/R)dy'
--+ uo(x') J<P(y', h'(x')y')a(x')dy'
by dominated convergence. Clearly the convergence is locally uniform.

But the integral is the integral of <P over the tangent plane T of M at
(x', h(x')), taken with respect to the Euclidean area, so by Theorem
7.1.25 it is equal to
(2n)-k J ¢(~)d(JR)·
T"
The formula (7.1.29) follows at once.
Theorem 7.1.28 also occurs in a different guise which will be useful

in Chapter XIV.
Theorem 7.1.29. Let FEC 1 (1R.n) be real valued, let vEL2 and denote by
~ the Fourier transform of eitFv, tE1R.. If 4>EC°(1R.n)nLOO(1R.n) then
J
(7.1.30) (2n)-n 1~(~W4>(~/t)d~ --+ JIv(xW 4> (F'(x)) dx, t --+ 00.
Proof First we assume that VECg' and that </>EY'. Then </>(~/t)J.-;(~) is
the Fourier transform of the convolution of tnl/J(t.) and ve itF, if rj/=</>.
The product by the complex conjugate of ve itF is
vex) e-itF(x) Svex - y)eitF(X-Y)tnl/J(t y)dy
= vex) Svex - y/t)eit(F(X-Y/t)- F(X))I/J(y)dy
---+lv(xW Sr i (y,F'(x)I/J(y)dy=lv(x)1 2 </>(F'(x)), t---+oo.
We also have a majorant CJv(x)1 so (7.1.30) follows. If </> = 1 we obtain

(7.1.30) from Parseval's formula which also extends (7.1.30) to all </> in
the closure of Y' in the maximum norm, that is, all </>ECO(JR) converg-
ing to 0 at 00. If 0~</>~1, </>ECg' and </>=1 in {F'(x); XESUppV}, it
follows that
SIJ.-;(~W(1- </>(~/t))dx ---+ 0, t ---+ 00.
Hence (7.1.30) is valid for every bounded </> vanishing in a sufficiently

large compact set. Thus we have proved (7.1.30) for all
</>ECO(JRn)nLOO(JRn) if VECg'. Since Cg' is dense in L2 it follows that
(7.1.30) is valid for all vEL2.
The following corollary is essentially identical to Theorem 7.1.28

with k = 1. The case of a higher codimension k can be discussed in the
same way.
Corollary 7.1.30. Let J.-; be defined as in Theorem 7.1.29 and let

4>EC o(IRn+l), Then
(7.1.31) (2n)-n IS 1J.-;(~W4>(~/R, t/R)d~ dt/R

t~O
---+ SS Iv(xW4>(tF'(x), t)dtdx.

t~O
Proof With s= t/R as new variable the left-hand side can be written
If we replace t by R, F by sF and </>(~) by 4>(~, s) in Theorem 7.1.29 it

follows that the inner integral converges boundedly to
(2n)" SIv(xW4>(sF'(x), s) dx.
This implies (7.1.31) since the integration with respect to s is taken

over a finite interval.
7.2. Poisson's Summation Formula

and Periodic Distributions
In Section 7.1 we determined the Fourier transform of the Lebesgue

measure on any linear subspace of ]R.n. Our first purpose here is to
determine the Fourier transform of the sum of the Dirac measures at
the points in a discrete subgroup of ]R.n.
Theorem 7.2.1. If U a is the sum of Dirac measures
ua= I bag, OoFaE]R.,

gEZ»
Proof. Since bag * U a= U a if g E ?In, we have

(ei(ag")-l)ua=O, gE?l".
All the factors do not vanish except at points in (2 n/a)?l". At the
u
origin for example a must just be a multiple of the Dirac measure,
for (sinaxp)ua=O,j=l, ... ,n, so this follows from Theorem 3.1.16 if
u
we take sinax/2 as new variables. Now a is invariant under trans-
lations in (2n/a)?l" because e21ti(.,g)/aua=ua' This implies that ua is a
measure with the same mass at every point in (2 n/a) ?In, thus
Explicitly this means that

(7.2.1)
or if we replace 1> by a translation of 1>
I ¢(ag) ei(ag,x) = CaI 1>(2ng/a + x), 1> E.9'.
Now integrate both sides for O<xj<2n/a, j= 1, ... , n. All terms with
g oF 0 drop out on the left-hand side then and we obtain
¢(O)(2n/a)"=caS1>(x)dx=c a¢(O),
Note that the preceding argument did not use Fourier's inversion
formula. Since Theorem 7.2.1 implies that
aa =(2 n/a)" a" ua= (2 n)" ua
we obtain another determination of the constant in Fourier's in-
version formula which has the advantage that the constant is directly
related to the volume of the period of the exponentials ei(x,g), gE71n .

(7.2.1) can now be written
(7.2.1)' I.c1i(ag) = (2n/at L 4> (2n g/a), 4> E /7,
and is called Poisson's summation formula. Note that as a ..... O or 00 we
obtain as special cases the Fourier transform of 1 and c5 o.
Let us now consider a distribution u which is periodic with period
1 in each variable, that is,
u(x-g)=u(x), gE71n •
Such a distribution is automatically temperate. For let 4> be a
function in CO' (1R.n) with
(7.2.2) L4>(x-g)=l
(see Theorem 1.4.6). If 1/1 E CO' then
<u, 1/1) = L <u, 1/14>(. -g) = L <u, 1/1(. +g)4»
by the periodicity. Since
/731/1 ..... 4> L 1/1(. +g)E CO'
is continuous, this proves that u is temperate. To determine <u,I/I)
=<u,.{i) we apply Poisson's summation formula to I/I(.)e-i(x,.) which
gives
4> (x) L .(i(x + g) = (2nt L I/I(2n g) e- h i (X,g) 4>(x).
Hence
where
cg=<u,4>e- hi (.,g».
Note that if u is a continuous function then
Ju 4> e- 21ti (.,g) dx = Jue- 21ti(.,g) dx
I
where l={x; O~xj<l} is the unit cube, for we can just integrate over
the integer translations of 1 and sum using the periodicity of u and
(7.2.2). A general periodic u can be regarded as a distribution on the
torus T"=1R."/71n and as a limiting case we just have the integral over
the torus then. Thus
(7.2.3) u=(2n)" L cgc5 21tg ,
(7.2.4) C =)(2ng)=O(lgl k) if uEfi)'k.
By Fourier's inversion formula

(7.2.5)
with convergence in Y'. Thus we have recovered the basic facts on
Fourier series. If UE Ck then cg=;:;¢(2ng)=O(lgl- k) so (7.2.5) is uni-
formly convergent if k > n.
Theorem 7.2.2. If UECk(]Rn), k>n, and U is periodic with period 1 in

each variable, then (7.2.5) is valid with uniform convergence and with the
coefficients given by (7.2.4).
That cg must be given by (7.2.4) follows at once by integration of

(7.2.5) over the torus, so the contents of the theorem are just the
existence of such a series expansion. This obvious determination of
the Fourier coefficients is of course the reason behind the determi-
nation of the constant in Fourier's inversion formula after Theorem
7.2.1.
Theorem 7.2.3. If U E L2(rn) then (7.2.5) is valid with the coefficients

(7.2.4) and convergence in L2(Tn). Parseval's formula is valid,
(7.2.6) JluI 2dx = D CgI2.
Tn
Conversely, if I Icgl2 < <Xl then (7.2.5) converges in L2(Tn), and the
Fourier coefficients of the sum u are equal to Cg.
Proof If uEcn+1(Tn) we have (7.2.5) with absolute convergence.

Hence
Jlul 2 dx= J ICgCg,e21ti(x,g-g'>dx=IICgI2.
Tn Tn
Hence the map L2(rn) 3U-+ {C g } E [2(zn) is isometric. The range con-
tains the dense subset [1 (zn) so the map is unitary.
Closely related to Poisson's summation formula is the formula

(7.2.7) I f(ag/k)=(k/atI J f(x)e- 21ti (g,x>k/adx
which is valid if f E cn+ 1 (]Rn) is periodic with period a in each vari-

able and k is a positive integer. It follows if we observe that
F(x)= I f(x+ag/k)
is periodic with period a/k and has the Fourier coefficients in the right-
hand side of (7.2.7). When a = k-+ <Xl we obtain formally Poisson's
formula again, and one can justify this limiting procedure.
We shall now indicate a slightly different path to Poisson's sum-

mation formula when n = 1, which is parallel to our proof of Fourier's
inversion formula by Cauchy's integral formula. Let t/J E [/ and form
<iff, u) = <t/J, u)
00
if/(0)/2 + L if/(2 n k) =
1
where u is the measure with mass 1/2 at 0 and 1 at 2nk when k is a

positive integer. Since u is the limit of u e-£X as e--+ + 0, we have with
[/' convergence
u(~)=lim (1/2+ Ie-2"ik~-2"£k)

£--+0 1
= lim (1/2 + e- 2"i~-2"£/(I_e-2"i~-2"£))

£~O
= lim 21. cot n(~ -ei) =~ cot n(~ - iO).

£~O Z 2z
Hence
<
00
(7.2.8) if/(0)/2 + L if/(2 n k) = cot n(. - iO), t/J/2i)

1
and replacing t/J by I/f we obtain

-1
(7.2.8)' if/(O)l2+ L if/(2nk) = -(cotn(. +iO), t/J/2i).
-00
Now we have 00
cotn(~-iO)-cotn(~+iO)=2i L (\<e)
-00
by Example 3.1.13 since n cot n z has a simple pole with residue 1 at

every integer. Adding (7.2.8) and (7.2.8)' we obtain Poisson's sum-
mation formula.
The fact that n cot n z has a pole with residue 1 at every integer
can often be used to compute sums of the form L t/J(k) directly as the
sum of residues of t/J(z) n cot nz. Take for example t/J(z)=I/(z-w)
where w is not an integer, and integrate (2ni)-1 t/J(z)ncotnz around a
rectangle with sides given by IRe zl =N + 1/2 or 11m zl =N where N is
an integer. By Cauchy's integral formula the integral is equal to
N
- L (w-k)-l+ ncotnw
-N
when N is large. The integral of z -1 cot n z around the rectangle is 0

since the integrand is an even function so that contributions from
opposite points cancel. Now cot n z is uniformly bounded and we have
Z-l_(Z-W)-l =O(N- 2 ) on the rectangle so we conclude that the
7.3. The Fourier-Laplace Transformation in 8' 181
integral--+O as N --+ 00. Hence we obtain the familiar formula

N
1tcot1tW= lim I (w-k)-l,
N_oo -N
which is closely related to (3.4.11).
7.3. The Fourier-Laplace Transformation in C'

We have seen in Theorem 7.1.14 that the Fourier transform 11 of any
distribution u E tff' (IR n) can be extended to an entire analytic function
in (Cn called the Fourier-Laplace transform of u. We shall now discuss
its properties in greater detail. If u E V and H is the supporting
function of supp u, defined in (4.3.1), then it is clear that
(7.3.1) lu(()1 ~ Iluliv exp H(Im O.
Theorem 7.3.1 (Paley-Wiener-Schwartz). Let K be a convex compact

subset of IRn with supporting function H. If u is a distribution of order
N with support contained in K, then
(7.3.2)
Conversely, every entire analytic function in (Cn satisfying an estimate
of the form (7.3.2) is the Fourier-Laplace transform of a distribution
with support contained in K. If u E C~ (K) there is for every N a
constant CN such that
(7.3.3)
Conversely, every entire analytic function in (Cn satisfying (7.3.3) for
every N is the Fourier-Laplace transform of a function in C~(K).
Proof. To prove the necessity of (7.3.2) we choose x~ E C~(K~) where
K~={x+y; xEK, lyl~c5}, so that Xa=1 in K0j2 and ID<XX~I~C<xc5-I<x1
(see (1.4.2)). Then we have if u E tff,N (K)
I sup ID<X(x~e-i<. -'»1
lu(Ol = lu(x~e-i<-'O)I ~ C
1<xI;&iN
~ C' exp(H(ImO+c5IIm (I) I c5- I<xI(1 +IWN-1<x 1.
1<xI;&iN
The estimate (7.3.2) follows when we take c5 = 1/(1 + IW. To prove
(7.3.3) when u E C~(K) it suffices to note that by (7.3.1)
I(<X u(Ol ~ IID<Xullv expH(Im O.
Next we prove the sufficiency of (7.3.3). Thus let U be an entire

analytic function satisfying (7.3.3) for every N. Then the restriction of
U to JRn is the Fourier transform of the COO function
u(X)=(2n)-n Jei<x.~) U(~)d~
so all we have to prove is that supp u c K. To do so we note that for

any choice of 11 E JR n
(7.3.4)
for the rapid decrease of U at infinity allows us to shift integration in
the direction 11 in JRn to a parallel in the complex plane. (Alter-
natively, differentiation with respect to I1j under the integral sign is
equivalent to i times differentiation with respect to ~ j' and the integral
of a derivative is 0.) Estimation of (7.3.4) by means of (7.3.3) with
N =n+ 1 gives
lu(x)l;;; e-<x.~)+H(~) C n + 1 J(1 + IW- n - 1 d~.
If we replace 11 by tl1 and let t~ + CXl it follows that u(x) =0 unless

H(I1)~<X, 11). But if this is true for every 11 we have XEK by Theorem
4.3.2, which proves that suppucK.
To prove the sufficiency of (7.3.2) we first note that the restriction
to JR" of a function U satisfying (7.3.2) is the Fourier transform of a
distribution u E !I". Choose <p E CO' with support in the unit ball and
J<pdx=l, and set <p~(x)=(j-n<p(xj(j). Then the Fourier transform of
u * <p~ is U$~ which has an entire analytic extension U $~ such that
N=1,2, ....
In fact, (7.3.3) is valid for $~ with H replaced by the supporting

function of the ball with radius b and center at O. From the results
already proved it follows therefore that suppU*<P~cK6' When b~O
we have U*<P6~U which proves that suppucK. The proof is com-
plete.
Note that the special case of Theorem 4.3.3 where u 1 =U 2 in (4.3.5)

is an immediate consequence of Theorem 7.3.1. We shall give some
further applications to differential operators with constant coefficients.
P(D) be the differential operator obtained when

-iOjox j •
'j
Let P be a polynomial with complex coefficients in ('1' ... , 'n) and let
is replaced by
Theorem 7.3.2. If fEI9"(JR") then the equation
(7.3.5) P(D)u= f
7.3. The Fourier-Laplace Transformation in C' 183
has a solution UE$'(1R.n) if and only if J(Wpm is an entire function.

The solution is then uniquely determined and
(7.3.6) ch suppu=ch suppf.
Proof. If (7.3.5) has a solution UE $' we obtain
P«()u(o=Jm
by taking the Fourier-Laplace transform on both sides, so J(WP«() is
the entire function u«(). The other half of the proof requires a lemma.
Lemma 7.3.3. If h(z) is an analytic function of z E <C and p(z) is a

polynomial with leading coefficient a, then
lah(O)1 ~max Ih(z) p(z)l.
Izl= 1
Proof. Set q(z) =z'" p(l/z) where m is the degree of p and p is obtained
from p by conjugation of the coefficients. Then q(O)=ii and by the
maximum principle
lah(O)1 = Iq(O) h(O)1 ~max Iq(z) h(z)1 = max Ip(z) h(z)1

Izl=1 Izl=1
which proves the lemma.
End of proof of Theorem 7.3.2. We can choose the coordinates so that

the coefficient a of (~ in pm is not 0, when m is the degree of P.
Then P«(1 +z, (2' ... ) has leading coefficient a so the lemma gives if g
= JIP is entire
lallg«()I~ sup IP«(1 +z, (2' .. ·)g«(1 +z, (2' .. ·)1
Izl= 1
= sup IJ«(1 +z, (2' .. ·)1·
Izl= 1
When J satisfies (7.3.2) we obtain the same estimate for g but with
u
another constant, so g = where U E $' and c h supp u c c h supp f
=ch suppPu. The opposite inclusion is trivial so we obtain (7.3.6),
which is also a consequence of Theorem 4.3.3.
As an application we shall now prove a refinement of the Asgeirs-

son mean value theorem.
Theorem 7.3.4. Let u(x, y) be a continuous solution of the equation

(L1 x -L1)u=O
in a neighborhood of K = {(x, y); x, y E1R.n; Ixl + Iyl ~R}. Then we have
f u(x,O)dS(x)= f u(O, y)dS(y).

Ixl=R IYI=R
If n is odd and =l= 1 it suffices to assume that u is a solution near oK.
Proof. By regularization the proof is reduced to the case where u E Coo.
Let f be the measure
f(¢)= f ¢(x,O)dS(x)- J ¢(O, y)dS(y).
Ixl=R IYI=R
Then
J(~, q)=G«~, O)-G«q, q»)
where G is an entire function of one complex variable, for an even
entire function of one complex variable z is an entire function of Z2.
Now 1
G(z)-G(w)=(z-w)f G'(w+t(z-w))dt
o
so (G(z)-G(w))/(z-w) is an entire function in (C2. Hence
J(~, q)/«~, O-(q, q»)
is an entire function and therefore the Fourier transform of a distribu-
tion Jl with support in K such that (Ax-Ay)Jl= -f. Hence
-<I, u) =(Jl, (Ax-Ay)u) =0
if u is a solution in a neighborhood of K. The stronger statement for
odd n requires a more detailed study of Jl. Since Jl =E *f by (4.4.2)
where E is one of the fundamental solutions in Theorem 6.2.1 (with
n+ =n_ =n) we have by Theorem 4.2.5
sing supp JlC {(x+x', y+ y'); Ixl =Iyl, (x', y')ESUppf}.
Thus Ix'I+Iy'I=R and Ix'I=O or 1y'1=0. By the triangle inequality,
this implies Ix+x'I+ly+y'I~R and since suppJ.LcK it follows that
singsupp J.LC {(x, y); Ixl +IYI =R} =oK.
If n is odd, n =l= 1, then (6.2.1)' gives a fundamental solution E such that
Ixl = Iyl if (x, y) Esupp E, so the same inclusion is valid for the support
then. This proves the theorem. Note that for even n we have
E(x,y)(_1)1+n/2~0 when Ixl~lyl
for the fundamental solution given by (6.2.1). Since Jl(x,y)=E*f(x,y)

has the sign (_1)1+n/2 when Ixl+IYI<R, the two mean values
having opposite signs the Asgeirsson theorem fails for a regularization
of E.
The special case of Theorem 7.3.4 when u is independent of y is
just the mean value property of harmonic functions (cf. Theorem
4.1.8). Let us next consider a solution of the wave equation

(Ll x -a 2 /at 2 )u=0
in lR 3+ \ assuming that UE c 1 near the double cone defined by Ix I
+ltl~R. We can apply Asgeirsson's theorem to
and obtain
R
(7.3.7) 2n J u(O, t)dt=R J u(Rw,O)dS(w).
-R 1<01= 1
Differentiation with respect to R gives
2n(u(0, R)+u(O, -R»= J (u(Rw, O)+R<u~(Rw, 0), w» dS(w)
1<01= 1
and if we apply (7.3.7) to au/at instead we have
2n(u(0, R)-u(O, -R»=R J u;(Rw,O)dS(w).
1<01= 1
Elimination of u(O, - R) gives after change of notation Kirchoffs
formula
1
(7.3.8) u(O, t) = - J (u(tw, 0) + t( <u~(tw, 0), w) + u;(tw, 0))) dS(w).
4n 1<01= 1
This is of course a special case of (6.2.9).
After this digression to show that a general result like Theorem

7.3.2 may contain some quite specific information, we prove an ap-
proximation theorem related to Theorem 4.4.5.
Definition 7.3.5. A solution u of the differential equation P(D)u=O in

lR" is called an exponential solution if it can be written in the form
u(x)= f(x)ei<x.o
where ( E (Cn and f is a polynomial.

Theorem 7.3.6. If X clRn is convex, then the closed linear hull in
cOO(X) of the exponential solutions of the equation P(D)u=O consists
of all its solutions in COO (X).
Proof. By the Hahn-Banach theorem we must show that if v Eth"'(X) is

orthogonal to the exponential solutions then v is orthogonal to all
solutions in cOO(X). To do so it is sufficient to show that v(WP( - () is
an entire function, for then Theorem 7.3.2 shows that v = P( - D) j1 for
some j1 E th'" (X), hence
<v, u) =<P( -D)j1, u) =<j1, P(D)u) =0
if u E Coo (X) and P(D) u = 0. The proof is therefore completed by the

following
Lemma 7.3.7. If VEt9"(JR n) is orthogonal to all exponential solution of

P(D)u=O, then v(WP( -0 is an entire analytic function.
Proof. Choose a fixed vector 0=1=0 such that P( -to-O is not inde-
pendent of t for any'. This is true in particular if Pm (0) =1=0 where Pm is
the principal part of P, that is, the homogeneous part of highest
degree. From the hypothesis it follows then that v(tO+WP( -to-O is
an analytic function of t for fixed ,. In fact, if P( -to-,) considered
as a polynomial in t has a zero of order k for t = to, we obtain by
differentiating the identity
P(D) e- i (x,t6+0 =P( -to - 0 e- i (x,t6+0
with respect to t that
P(D)«x,0>i e -i(x,t o 8+ 0 )=0, j<k.
Hence v«x, 0>i e -i(x,t o6+ 0 ) =0, j<k, which means that v(tO+O has a
zero of order k at least at to. For every' we can now define
F(O=lim v(tO+WP( -to-,).
t~O
That F is entire follows from Weierstrass' preparation theorem (Theo-

rem 7.5.1) but we can also give a direct proof. For a fixed choose '0
for all , in a neighborhood of Hence'0'
r so that P(-tO-'oHO when Itl=r. Then P(-tO-'HO when Itl=r
F(O=(2ni)-1 J v(tO+WP(-tO-,)dt!t
'0'
111=r
is an analytic function of , in a neighborhood of The proof is
complete.
There is a result parallel to Theorem 7.3.1 which describes the

convex hull of the singular support.
Theorem 7.3.8. Let uEt9"(JRn) and let K be a convex non-empty compact

subset of JRn with supporting function H. In order that sing supp u c K
it is necessary and sufficient that there exists a constant N and a
sequence of constants Cm such that for m = 1,2, ...
(7.3.9)
Proof. To show that (7.3.9) is necessary we split u into a sum U=U 1

+u 2 where suppu 1 cK 1 /m and U 2 ECg'(JRn). (Here K~ is defined as in
7.3. The Fourier-Laplace Transformation in ~' 187
the proof of Theorem 7.3.1.) If U is of order N -1 it follows from

Theorem 7.3.1 that
1121 (e)l;:;:; Cm (1 + IW N -1 eH(Im~)+IImWm.
Hence
lu 1 ml;:;:;C m (1+IClte H (Im,) if IIm(l;:;:;mlog(l(l+1),
m
and by (7.3.3) we have 12 2 e-H(Im')-.O when e-.
00 in this set. The
estimate (7.3.9) is therefore valid with some larger Cm'
To prove the sufficiency of (7.3.9) we shall make a change of

integration contour as in the proof of Theorem 7.3.1 but it has to be
chosen differently in order to fit the set where (7.3.9) is applicable.
Thus define r" to be the cycle
lR.n3~-.e(~) = ~ + i'1log(1 + I~ 12).
An immediate computation gives that del" ... "den=F(~)d~l" ...
"d~n where F-.1 as ~-.oo. We have IImCl<21'1llog(1+ICI) on r".
Choose <P~ as in the proof of Theorem 7.3.1. Since 12m $lJm is rapidly
decreasing we obtain by application of Stokes' formula to the homo-
topic cycles Fo and r"
(7.3.10) U * 4>1J(x) =(2n)-n Sei(x,O 12m $~(e) del" ... "den'
r"
(Actually we are only changing the integration contour in one direc-
tion so this could be done by means of Cauchy's integral formula.) On
r" we have
(7.3.11) lei(x.o uml;:;:; Cqexp(H(ImO-<x, 1m 0)(1 +IClt
= C q (1 +1~12)H(q)-(x,q)(1 +IClt.
If xortK we can choose '1 so that
H('1) -<x, '1) < -1
for all x in a neighborhood X ° of xo' If we replace '1 by t'1 and 2t > n
+N it follows from (7.3.11) that the integral in (7.3.10) is absolutely
convergent for XEX 0 even if the decreasing factor $IJ is omitted. Since
$ie) = $(b 0-. 1 boundedly as b-'O we conclude that the restriction of
U to X 0 is the function
(7.3.12) u(x)=(2n)-n S ei (x,o u md e 1 " .. · "den, XEX o,

r,"
if 2t > n + N. If 2t > n + N + j the integral (7.3.12) remains absolutely
and uniformly convergent when XEX 0 after at most j differentiations
under the integral sign. Hence UECi(Xo) for every j, so uEC""(CK) as
was to be proved.
The following application to differential operators IS parallel to

(7.3.6).
Theorem 7.3.9. If UE«C'(JR") then

(7.3.13) ch sing suppu = ch sing suppP(D)u.
Proof Set f =P(D)u and let H be the supporting function of

singsuppf Then (7.3.9) is valid for J and follows for by means of u
Lemma 7.3.3, just as in the proof of Theorem 7.3.2. Hence the left-
hand side of (7.3.13) is contained in the right-hand side, and the
opposite inclusion is obvious.
In particular Theorem 7.3.9 states that UE COO if UE«C' and

P(D)UECClO. This is not true for arbitrary differential operators with
variable Coo coefficients. For example, xdu/dxEC ClO if U is a function
which is 0 for x<O, equal to 1 for O<x< 1 and goes smoothly to 0 for
1 < x < 2. There is no analogue of the theorem of supports (Theo-
rem 4.3.3) with supports replaced by singular supports which is valid
for arbitrary distributions. In fact, one can find U 1, U2 E«C'(JR") so that
U 1*U 2 ECoo but neither factor is in Coo. An example is u1(x)=X(x)/(x
+iO), u2 (x)=X(x)/(x-iO) if XEC~(JR). However, (7.3.13) gives such a
result when the support of one factor is a point; it is also valid when
the support of one factor is finite. These matters will be studied
further in Section 16.3.
It follows from Theorem 7.3.1 that fairly general Laplace trans-

forms define distributions in JRn. To motivate the definition we recall
that
<U, <p) =(2n)-n(u, ($( - .), if UEf/' and <PEfI'.
Given a measure d/1 in (Cn we now try to define UE~'(JRn) by the
similar formula
(7.3.14) u(<P) =(2n)-" J($( - () d/1m, <PE Co (JRn).
This is a valid definition if for some m ~ 0 and C, N
(7.3.15) 11m" ~mlog(l +1(1)+ C when (Esuppd/1,
(7.3.16) J(1 + I(I)-N Id/1ml < 00.
In fact, if <PECo({x; Ixl ~R}) we have by (7.3.3) or rather by the
derivation of (7.3.3) from (7.3.1)
(1+I(lt+kl($ml~Ck,ReRllm'l L supID a 4>l.

lal~N+k
If k?;Rm the exponential can be estimated by (1 +l(l)k in suppdJl.

Hence the integral (7.3.14) converges and
Iu(4)) I ~ C R L sup IDa¢l, 4>ECO'({X; Ixl ~R}).
lal~N+k
Thus (7.3.14) defines a distribution which is of order ~N +Rm+ 1

when Ixl <R. It is of order N if 11m" is bounded in suppdJl.
If P is a polynomial, then
<P(D)u, ¢) =<u,P( -D)¢)
and the Fourier-Laplace transform of P( -D)¢ is P( -()cPm. Hence
(7.3.17) <P(D)u, ¢) =(2n)-" f cP( - OP(OdJlm·
As an application we shall now show that every differential operator
with constant coefficients has a fundamental solution. We must then
choose the measure d Jl so that the integral on the right-hand side of
(7.3.17) is equal to the integral of cP over JR".
Theorem 7.3.10. For every polynomial P =1= 0 in n variables one can find
a distribution EE~~(JR") such that P(D)E=<>.
We prepare the proof with two lemmas used to construct the

measure d Jl.
Lemma 7.3.11. If 4>ECO'(<C") and

(7.3.18) 4>(e i9 () = 4>m, eEIR; f 4>«()dA«() = 1,
where dA is the Lebesgue measure in <C", then
(7.3.19)
for any entire analytic function F.
Proof By Cauchy's integral formula

fF«( ei9) d e=2nF(0).
If we multiply by 4>«() and integrate, (7.3.19) follows from (7.3.18).
With a fixed positive integer m we denote by Pol(m) the complex

vector space of polynomials of degree ~ m in n variables and by
PoIO(m) the vector space with the origin removed. A norm in Pol(m) is
given by Q~Q(O) where
(7.3.20)
Note that this function of ¢ is bounded from below if Q =F 0, for some

derivative of Q is a constant =F O.
Lemma 7.3.12. For every ball Z c <en with center at 0 one can find a
non-negative function eliE COO (PolO(m) x <en) such that
(i) eli(Q, () is absolutely homogeneous of degree 0 with respect to Q.
(ii) eli(Q, () satisfies (7.3.18) for fixed Q and vanishes when' rf;Z.
(iii) there is a constant C such that
(7.3.21) Q(O)~ C IQ(oI if eli(Q, ()=FO.
If F is analytic in Z and Q is a polynomial of degree ~m it follows

that
Q(O) IF(O)I ~ C z S!P(O Q(')I dA(O.
z
Proof For a fixed Qo the existence of such a eli is quite obvious, for we
can find WE<e n such that Qo(w)=FO. We can then choose r>O so that
rWEZ and Qo(zw)=FO when Izl =r, which excludes at most m values of
r. If 'I' ~ 0 and 'I' has support near rw and integral 1, then
eli(() = S'I' (e iO () d 8/2 n

has the required properties. The same eli can be used for all
QEPoIO(m) such that aQ is close to Qo for some aE<e. Piecing such
local constructions together by a partition of unity in the projective
space of PolO(m), that is, using a partition of unity where the terms are
absolutely homogeneous functions of degree 0, we obtain eli with the
stated properties. The last statement follows from (7.3.19) and (7.3.21)
for
Q(O) IF(O)I = II F(O Q(O) eli(Q, () dA(01 ~ C z I IF(O Q(oI dA(O.
z
Proof of Theorem 7.3.10. Let ~ be the polynomial ,->P(¢+() ob-
tained by translation of P and set
(7.3.22) E(</J) =(2n)-n Id ¢S $( -¢ - O/P(¢ + () eli(~, ()d ,1.(0.
This is of the form (7.3.14) and
SO + I(I)-N Idfl(Ol ~Sd¢SO +I¢ + (I)-N leli(~, ()1/1~(o1 dA(O
~ CJd¢S(1 + I¢ + (I)-N P(¢)-l dA(O< 00
z
if N>n. (Recall that P(¢) is bounded from below.) Since 1m' is
bounded in suppdfl the distribution E is of order at most n+ 1. From
(7.3.17) and (7.3.19) we obtain
<P(D)E, ep) =(2n)-n Sd~S $( -~ -OcP(~, odAm

=(2n)-n S$( -~) d ~ = ep(O),
From the preceding construction one can extract quite precise regu-
larity properties of E both for fixed P and as a function of P. This
will be done in Section 10.2.
7.4. More General Fourier-Laplace Transforms
For distributions u with compact support we have defined the

Fourier-Laplace transform by
(7.4.1)
Now it may be possible to define am at least in some subsets of (Cn
for more general distributions u. For fixed 11 = 1m ( we can at least
define (7.4.1) as a distribution in ~ = Re( if e<"~> uE[I". We shall
therefore start by studying for given UE~'(1R.n) the set
(7.4.2)
Lemma 7.4.1. If VE[I" and epEC oo has bounded derivatives of all orders,
then epVE[I". If l/IE[I' then the Fourier transform of l/Iv is the Coo
function
(7.4.3)
Proof The first statement follows since multiplication by 4> is a

continuous map in !/. The second follows from Theorem 7.1.14 if
l/IEC'O. In general we just choose using Lemma 7.1.8 a sequence
l/IkEC'O with l/Ik-+l/I in [I' and obtain l/IkV-+l/Iv in [1", hence f;v-+Ifi;
in [1", and e-i< .. ~> l/Ik -+e-i< .. o l/I in Y, locally uniformly in ~. Thus
(7.4.3) is the distribution limit of the corresponding function with l/I
replaced by l/Ik' which proves the statement since (7.4.3) defines a Coo
function by the first part of the proof.
It follows at once from the lemma that the set r" defined by (7.4.2)
is convex. In fact, if 111' 112Er" and we set u~=e<"~>u, then we have for
11 = tl11 +(1-t)rJ2' 0 < t < 1,
where
¢(x) = e<x.~> /(e<x.~I> + e<X'~2»
is bounded by 1 and has bounded derivatives of all orders. Now

assume that r., has a non-empty interior r.,0. Choose '10"'" '1nEr.,°
affinely independent. Then
where
t/I~(x) = e<x,~> l'i. e<x,~j>
is in Y' if '1 is in the interior of the simplex spanned by '10"'" '1n' In
fact,
t/I~(x) =('i.e<x,~j-~»-l
and if 0 is in the interior of the convex hull of '1 j - '1, then

Ixl ~ C max <x, '1j-'1),
so t/I~ and its derivatives are exponentially decreasing. This is true
uniformly when '1 is in a compact subset of the interior of the simplex.
It follows that u~ is then a COO function
u~(~) =L <u~j' e-i<.,~+i~> /Le<·,~j»
j
of (~, '1). It is of course analytic since the Cauchy-Riemann equations

(%~v + i 0/0 '1 v) u~(~) =0 follow immediately by differentiation.
Theorem 7.4.2. If UE.@' (1R.n) then (7.4.2) defines a convex set r.,. If the
interior r.,0 is not empty then there is an analytic function in 1R.n+ ir.,° u
such that the Fourier transform of e<"~>u is u(. +i'1) for all '1 Er.,°. For
every compact set Me r.,0 there is an estimate
(7.4.4)
Conversely, if r is an open convex set in 1R.n and U is an analytic
function in 1R.n + ir with bounds of the form (7.4.4) for every M ~ r,
then there is a distribution u such that e<"~> UEY" and has Fourier
transform U(. + i'1) for every 1]EI:
Proof. Only the last statement remains to be proved. Let u~ be the

inverse Fourier transform of U(. + i1]). Then ou~/oYJv is the inverse
Fourier transform of
o U(~ + i1])/o1]v = io U(~ + i1])/o~V'
so oU~/01]v=xvu~ and e-<x,~>u~=u is independent of 1]. The proof is
complete.
Now assume that u has support in a convex closed set K which is

no longer assumed to be compact. We can still define the supporting
function by
HK(~)=SUp<X, 0,
xeK
and it is a convex, positively homogeneous function with values in

( - 00, 00]. However, H K may not be continuous, but as the supremum
of a family of continuous functions it is lower semi-continuous. Con-
versely, if H is a function with these properties, then there is precisely
one convex closed set K such that H = H K' and
K = {x; <x,~) ~H(~), ~E1Rn}.
The proof of Theorem 4.3.2 gives this with no essential change, for the
lower semi-continuity of H means precisely that {(r, ~)E1Rn+ 1 ;
r~H(~)} is closed.
Now it is clear that when suppucK we have
'7 E I;. =;> '7 + 8EI;. if H K(8) < 00,
and {8;HK(8)<00} is a convex cone. In fact, if XEC OO (1R) is 1 on
(- 00,1/2) and 0 on (1,00), then
(7.4.5) e(x,9)-HK(fI)=e(x,fI)-HK(fI) X«x, 8) -HK(8)) near supp u
and on the right-hand side we have a function of x with bounded
derivatives.
Theorem 7.4.3. If u satisfies the hypotheses in Theorem 7.4.2 and in
addition supp u c K, then
(7.4.6)
where M is a compact subset of I;.0. Conversely, if there is some '7 for
which (7.4.6) is valid, then supp u c K if K is closed and convex.
Proof To prove (7.4.6) we set, = ~ + i(8 + '7). Then we have H K(8) < 00
and
by (7.4.5). For the function on which u~ acts the derivatives of order k

are bounded by Ck(l + 1Wk eHK(fI). Hence (7.4.6) follows from the
estimates for u~ in the proof of Theorem 7.4.2. Assume now that (7.4.6)
is valid for some '7. If <pECD' and h is the supporting function of
supp <p, then
(2n)n I<u~, <p)1 =IJu(~ +i(8+'7))t$( -~ -i8)d~1

~ C J(1 + 1~I + 1
81)N ~K(fI)+h(-fl)(1 + I~ 1+ 181)-N-n-l d~
~ C' eHK(9)+h(-9).
Replacing 0 by to we obtain <u~,¢)=O if HK(O)+h(-O)<O. Hence

u=O in a neighborhood of x if HK(O)-<x,O)<O for some 0. If
XESUPPU it follows that <X,O)~HK(O) for every 0, that is, suppucK.
This completes the proof which is of course just a slight variation of
the second part of that of Theorem 7.3.1.
Remark. In Theorem 7.4.2 we discussed 11 in the interior of 1Rn + if,;

only. However, the continuity of the Fourier transformation in [/"
shows that if '1Er,. . . . . r,.0 then the Fourier transform of e<"~) U is the limit
in [/' of ~-+u(~+i(l-t)'1+itO) as t-+O, if 0 belongs to a compact
subset M of r,.0.
As an application we shall now compute the Fourier transform of

the advanced fundamental solution E of the wave operator
o =c- 2 fPjBt 2-A
in 1Rn + 1, constructed in Section 6.2. The support is in
K = {(t, x); ct~ Ixl},
H K(r, ~)=sup(tr + <x, O)=sup(tr +ct I~ I),
K t>O
so HK(r,~)=O ifr+cl~I~O and HK(r,~)=oo ifr+cl~I>O. It follows

that the Fourier-Laplace transform E(r,0 must be analytic when
1m r < - c 11m". Since
it follows that
('2-r2jc 2)E=1, if ,2=<',0,
(7.4.7)
there. The Fourier transform of E is therefore the limit (Lemma 6.2.2)
lim (~2_(r_ie)2jc2)-1 =(~2_r2jc2+iO)-1 if r>O
£_+0
=(~2_r2jc2_iO)-1 if r<O;
at the origin it is determined by the homogeneity.
It is also easy to construct E starting from (7.4.7). We must then
show that <',0 -r 2jc 2*0 when Imr< -c 11m". To do so we ob-
serve that the equation
q(s)=(Re' +s Im,)2 -(Rer +s Im't')2jc 2 =0
has real roots then for the quadratic form ~2_'t'2jc2 would otherwise
be negative in a real two dimensional plane. Hence Iq(i)1 is at least as
large as the absolute value of the leading coefficient in q, that is,
(7.4.8)
7.5. The Malgrange Preparation Theorem 195
From (7.4.8) and Theorem 7.4.3 it follows at once that the right-hand
side of (7.4.7) is the Fourier-Laplace transform of a distribution E+
with support in the forward cone {(t,x);ct~lxl} so we obtain another
construction of the forward fundamental solution. It has a much
wider scope than the earlier one. (See Section 12.5.)
7.5. The Malgrange Preparation Theorem
Decomposition of the Fourier transform of a function <PEg by a

partition of unity gives a representation of <p as a sum of entire
analytic functions. This simple observation will be used in this section
to derive the Malgrange preparation theorem from the classical
analytical analogue which we first recall.
Theorem 7.5.1 (The Weierstrass preparation theorem). Let f(t, z) be an

analytic function of (t, Z)E<C1+n in a neighborhood of (0,0) such that
(7.5.1) f=of/ot= ... =Ok-lf/Ot"-l =0, okf/otk,*O at (0,0).
Then there is a unique factorization
(7.5.2)
where aj and c are analytic in a neighborhood of 0 and (0,0) re-
°
spectively, c(O, 0)"4= and aiO)=O.
Proof. Choose r>O so that f is analytic at (t, 0) when It I<2 rand f(t, 0)
,*0 when 0<ltl<2r. Then choose 15>0 so that f(t,z) is analytic when
Itl<3r/2, Izl<b and f(t,z)"4=O when Itl=r, Izl<b. For every z with
IZI<b the equation f(t,z)=O has then precisely k roots tj with Itjl<r
(counted with multiplicity). If there is a factorization (7.5.2) we must
therefore have
k
tk+ak_l(z)tk-l + ... +ao(z)= TI(t-t)
1
= exp«2n i)-l S «0 f(s, z)/os)/ f(s, z)) log(t - s) ds).

Isl=r
Here the last equality assumes that Itl>r so that the logarithm has an
analytic branch when Is I< r. The exponential is an analytic function of
t and z and a polynomial in t so using for example the Lagrange
interpolation formula we conclude that aiz) are analytic functions of
z. If
p(t, z) = tk+ ak_ 1 (z) tk- 1 + ... + ao(z)
the quotient f(t,z)/p(t,z) is analytic when Itl~r for fixed z, Izl<b.

Hence
f(t,z)/p(t,z)=(2ni)-1 J f(s,z)p(S,Z)-l(S-t)-l ds
101=,
if Itl<r and Izl<b. The right-hand side is then an analytic function of
(t, z), which completes the proof.
By the Weierstrass preparation theorem one often means the fol-

lowing more general result, also known as the division theorem or the
Weierstrass formula.
Theorem 7.5.2. If f satisfies (7.5.1) and g is analytic in a neighborhood

of (0,0) in (C1+n then
k-l
(7.5.3) g(t,z)=q(t,z)f(t,z)+ L tjrj(z)
o
where q and rj are analytic near (0,0) and o. Both the quotient q and the
remainder are uniquely determined.
Proof By Theorem 7.5.1 we may assume that f is a polynomial in t of

degree k. If g is an analytic function when It I< 2 rand Izi < 15, and if
f(t,z)=FO when Itl~r, Izl<b, then (7.5.3) implies
(7.5.4) q(t,z)=(2ni)-1 J g(s,z)f(S,Z)-l(S-t)-lds,
101='
It I<r, Izl<b,
for the integral
J (ki l si r/ z))f(s,Z)-l(S-t)-l ds
101=' 0
is equal to 0 since the integration may be made over an arbitrarily

large circle and the integrand is O(lsl-2) as s-+oo. Hence (7.5.3)
implies
k-l
(7.5.5) L ti r/z)=(2ni)-1 J g(s,z)(f(s, z)-f(t,z))/(s-t)ds/f(s, z).
o 101=,
The quotient (f(s, z) - f(t, z))/(s - t) is a polynomial of degree k -1 in s
and t, so the right hand side does define a polynomial in t for all z.
The coefficients are analytic functions of z for Iz 1< b. Since (7.5.4) also
defines an analytic function when Izl<b and Itl<r, the existence and
uniqueness of the decomposition (7.5.3) is proved.
We shall now discuss the corresponding results for COO functions.
The difficulty is then that zeros may be lost. For example, t2+X has
two real zeros when x<O but none when x>O. To be able to keep
track of the zeros we shall therefore start by examining the division
theorem for functions analytic in thin strips around the real axis when
f is a polynomial. We omit the parameters x but insist on uniform
estimates instead. Thus let
(7.5.6) p(t)=tk +ak_ 1 tk- 1+ ... +ao
be a polynomial in tE<C of fixed degree k. Assume that L lajl < 1, and
let g be a bounded analytic function in the strip 11m t I< e. We want to
make a division
k-1
(7.5.7) g(t)=q(t)p(t)+ L tjrj
o
so that q and rj have bounds in terms of M=sup{lg(t)I;llmtl<e} in a
smaller strip Ilmtl<ce. In addition the decomposition must depend
analytically on p if we do not change p very much. To achieve this we
first choose j with 1 ~j ~ k + 1 so that
(7.5.8) p(tHO when Illmtl-ej/(k+2)1 <e/(2(k+2)).
This is possible since p has only k zeros. If W is a bounded open set

and g is analytic in W, p is a polynomial with no zero on OW, then the
proof of Theorem 7.5.2 gives if OWEC 1
g(t) =q(t) p(t) + r(t), tEW,
where
q(t)=(2ni)-1 Jg(s)p(S)-l(S-t)-lds, tEW,
oro
k-1
r(t)= L tj rj =(2ni)-1 J g(s)«p(s)-p(t))/(s-t))p(S)-l ds.
o oro
Note that r is determined by the additional fact that r(t)=O if p(t)=O

and tfj:w. When g is analytic for 11m tl <e and (7.5.8) is valid, we
obtain
(7.5.9) q(t)=(2ni)-1 Jg(s)p(S)-l(S-t)-l ds,
Yt
k-l
(7.5.10) L t' r. = (2 n i) - 1 Jg(s)«p(s) - p(t))/(s - t)) p(s)- 1 ds.
o
Here Ilmtl<ej/(k+2) in (7.5.9), Y is the boundary of the rectangle

{s;llmsl<ej/(k+2),IResl<2} and Yt is the boundary of the union of
this rectangle and a congruent one with center at Re t. The bounds
(7.5.11) Iq(t)I~CMe-k-l when Ilmtl<e/(2(k+2)); Ir.I~CMe-k;
are obvious since Isl<1 when p(s)=O, hence Ip(s)l~cek on oY and oYt.
We shall now examine what happens when the coefficients of pare

changed. Let
(7.5.12)
Since (7.5.8) implies
Ip(t,a)l>ce" if IImtl=ej/(k+2)
we still have
Ip(t, b)1 > ce"j2 if IImtl = ej/(k + 2), la -bl <c 1 e".
We can therefore use the formulas (7.5.9), (7.5.10) with the same j to
divide g(t) by p(t,b) for all b with la-bl<c 1 e". Hence
k-l
(7.5.7)' g(t)=q(t,b)p(t,b)+ L tjrj(b), Ib-al<c 1 e"
o
where q(t,b) and rj(b) are analytic in b also. When Ib-al <c 2 e",
c 2 <c 1 , it follows from Cauchy's inequalities that
(7.5.11)' Io~ o~ q(t, b)1 ~ CapM e- a- 1 - k(IPI+ 1), IImtl <ej(2(k + 2));
Iof rj(b) I~ CpM Ck(IPI+ 1).
To piece the preceding local solutions of (7.5.7)' together we shall
use the partition of unity in Theorem 1.4.6 for CCk = IR. 2k. If K is the
diameter of the support of the function <P in that lemma, then
1= L <PG(a), <PG(a) = <p(K ajc 2 ek - G), aECCk •

The diameter of sUPP<PG is at most C 2 ek • For any lattice point G in
IR. 2k with
(7.5.13)
we choose a point aG in the intersection and apply the preceding
construction with a=aG. This gives qG(t, b) and rj,G(b) satisfying (7.5.7),
and (7.5.11)' when bESUPP<PG' It follows that
q£(t, b, g) = L <PG(b) qG(t, b), rj(b, g) = L <PG(b) rj,G(b)
with summation over all G satisfying (7.5.13) has the properties in the
following lemma where b is regarded as a variable in IR. 2k.
Lemma 7.5.3. For every bounded analytic function g in {tECC;IImtl<e}

one can find q£(t, b, g)ECOO(IR. x IR. 2k) and rj(b, g)ECoo (IR. 2k) depending
linearly on g such that
k-l
(7.5.7)" g(t)=q£(t,b,g)p(t,b)+ L tjrj(b,g), if Ll bj l<1,
o
7.S. The Malgrange Preparation Theorem 199
(7.5.11)" l ~o-q8(t
I b "
b g)I~C
-,._
Me-«-1-le(I_I+1)'
lofrj(b,g)1 ~ C_M B-"0_1+1).
Here M = sup {I g(s) I; IIms 1< B} and the constants are independent of B.
We shall now eliminate the hypothesis that g is analytic by using
the decomposition of COO functions into sums of analytic functions
mentioned at the beginning of this section.
Theorem 7.5.4. For every ge9'(R.) one can find q(t,b,g)eCoo(1RxlR. 2lc)
and r}(b,g)eCoo(R2k) depending linearly on g such that (7.5.7)" is valid
with ql replaced by q and rj replaced by r}, and
(7.5.14) lo~Ofq(t, b,g)1 ~ C,._Hlgl +lg(V)l)dt, v=3 +«+k(lPI + 1),
lof rj(b, g)1 ~ C_ J (lgi +lgMI) dt, v=2+kOPI + 1).
Proof. Choose a function I/IeCO'(lR.) such that 0~I/I~1 and I/I(t)=1
for Itl<1, I/I(t) =0 for Itl>2. Set
go(t) = (21t)-foJ g(t) I/I(t) eil ' dt,
git)=(21t)-1 J g(t)(1/1(2-}t) - I/I(21-}t»eil • dt, j= 1, 2, ....
It is then clear that g} is analytic and that g = L g} in 9'. Since
we have ItVg(t)I~Jlg(V)(t)ldt
Igi t)I~ C2}(1-V)JOgl +lg(V)l)dt if IImtl <Bj =2- j •

Hence an application of Lemma 7.5.3 gives
Ie-I
git)=q8J(t,b,gj)p(t,b)+ L tirtJ(b,gj)' Llbj l<1,
o
lo~Of q8J(t, b, g})1 ~ C,._2-}(V-l-,.-1-lc(I_I+ 1» Higi + Ig(V)l)dt,
IOf rtJ(b, g)1 ~ C_2-}(V-l- le O_I+ 1» J(lgi + Ig(V)I)dt.
With v chosen as in the theorem we obtain a convergent geometric
series on the right-hand side and conclude that
00 00
q(t, b, g) = L q8J(t, b, gj),

1j(b, g) = L rtJ(b, g})
}.o }.o
are Coo functions with the stated properties.
Remark. If g(t, x)e9'(1R. 1+n) then

q(t, b, g(., x»e COO(1R 1 +n+2lc)
and
In fact, differentiation with respect to x may be performed directly on

g by the continuity and linearity with respect to g.
We are now ready to prove the part of the Malgrange preparation
theorem which corresponds to Theorem 7.5.1.
Theorem 7.5.5. Let f(t, x) be a COO function of (t, x)e1R 1+11 near (0,0)
which satisfies (7.5.1). Then there exists a factorization
(7.5.2), f(t, x) = c(t, x)(tk + ak _ 1(x) t"-1 + ... +ao(x))
°
where a; and c are COO functions near and (0,0) respectively, c(O, 0) 9=
and aiO) =0. When f is real the factorization can be chosen real.
°
Proof. We may assume that feCr: since the statement is local. By
Theorem 7.5.4 and the remark after its proof we have in a neigh-
borhood of (0,0,0)
k-l
f(t,x)=Q(t,x, b)p(t, b)+ L ti Rix, b)
o
where Q and R j are in Coo. Taking x=b=O we obtain
k-l
f(t,O)=Q(t,O,O)t"+ L tiRio, 0).
o
Hence Ri(O,O)=O and Q(O,O,O)9=O by (7.5.1). If we differentiate with
°
respect to b and put x = 0, b = afterwards, we obtain when b = °
k-l k-l
O=dbQ(t,O,b)tk+Q(t,O,O) L tidbi + L tidbRi(O,b).
o 0
k-l
Since Q(O, 0, 0)9=0 we have ao= ... =ak _ 1 =0 if Q(t, 0,0) L tjai=O(t~.
o
Hence the differential of the map (b o, ... , bk_1)-+(R o, ... , R k - 1) is bijec-
tive at (0,0). By the implicit function theorem it follows that the
equations
Rix,b)=O, j=O, ... ,k-l,
define Coo functions bix),j=O, ... ,k-l, in a neighborhood of 0. Since
f(t, x) = Q(t, x, b(x))p(t, b(x))
°
and b(O) = we have obtained (7.5.2)'. If f is real we can take Q and R
real when be1Rk and apply the implicit function theorem with be1Rk
instead of be<r k = 1R 2k, which completes the proof.
The following division theorem is analogous to Theorem 7.5.2.

Note that no uniqueness is valid in Theorems 7.5.5 and 7.5.6.
Theorem 7.5.6 (The Malgrange preparation theorem). If f satisfies the

hypothesis in Theorem 7.5.5 and g(t, x) is a Coo function in a neigh-
borhood of (0,0) then
k-l
(7.5.3)' g(t,x)=q(t,x)f(t,x)+ L tir/x)
o
where q and rj are Coo functions in a neighborholUL-of-(O, 0) and O.
Proof. By Theorem 7.5.5 we may assume that

k-l
f(t,x)=tk+ L tia/x)=p(t, a (x».
o
Assuming as we may that ge CO' we can take
q(t, x)= q(t, a(x), g(., x», rj(x) = r/a(x), g(., ~»
with the notation in Theorem 7.5.4.

Note that Theorem 7.5.5 is actually the special case of Theo-
rem 7.5.6 with g(t, x) = tk. In the proof of Theorem 7.5.6 it was only the
application of Theorem 7.5.5 which required shrinking the neigh-
borhood of (0,0). More precisely, if geC OO (1R. 1 +n) then (7.5.3)' follows
in a neighborhood of (0,0) independent of g.
The Malgrange preparation theorem is highly non-trivial even
when k = 1. This is in fact the case which we shall use most frequently.
In Section 7.7 we shall need the extension of it to several t variables.
Theorem 7.5.7. Let Jj(t, x), j = 1, ... , m, be complex valued Coo functions
in a neighborhood of 0 in 1R.m + n with Jj(O, 0) = 0, j = 1, ... , m, and
detaJj(O,O)/atk=FO. If geC OO in a neighborhood of (0,0) we can then
find qj(t,x)eC OO at (0,0) and r(x)eC OO at 0 so that
m
g(t, x) = Lqj(t, x) Jj(t, x) + r(x).
1
Proof. When m = 1 this is a special case of Theorem 7.5.6. We can

always label the functions Jj so that afl(O,O)!at l =FO. Writing t'
=(t2' ... , t n) we then obtain from Theorem 7.5.6 with k= 1
g(t, x) =ql (t, X)fl (t, x) + h(t', x),
Jj(t, x) =q/t, x)fl(t, x)+F/t', x), j=2, ... ,m.
Since dJj=qjdfl +dFj at 0, the differentials of F2, ... ,Fm with respect
to t' are linearly independent at O. If the lemma is already proved for
m -1 variables we obtain
m
h(t', x) = 2:Pj(t', x) lj(t', x)+ r(x),
2
g(t,x)= (ql(t,X)- ~Pj(t"x)qit,X»)fl(t,x) +~Pj(t"X)./j(t,x)+r(x)

To state an analogue of Theorem 7.5.5 in this situation we in-
troduce the ideal I = I (f1> ... '/m) generated by fl"" '/m. This is the set
of all COO functions g in a neighborhood of 0 such that
m
g(t, x) =2:qP, x)./j(t, x)
1
in a neighborhood of 0 for some qjeC oo •

Lemma 7.5.S. If F1 , ... ,FmeI(fl' ... '/m) and dF1 , ... ,dFm are linearly in-
dependent at 0, fl = ... = fm = 0 at 0, then
Proof. By hypothesis
m
F; =I,glj./j, i = 1, ... , m.
1
Since ./j(0,0)=0 we have dF;= 2:gljd./j at O. Hence detg1j=l=O so
./j= 2:(g-lJjl F;eI(F1, ... ,Fm)·

Here is now an extension of Theorem 7.5.5 with k = 1.
Theorem 7.5.9. If fl' "',/m satisfy the hypotheses in Theorem 7.5.7 then
for some 1je coo vanishing at O.

Proof. Theorem 7.5.7 applied to the coordinates tj gives
tj = 2:qlj(t, x)./j(t, x) + T,(x),
and it follows from Lemma 7.5.8 that tl - T,(x) can be used as gener-
ators.
We shall now study the indetermination of the functions T,(x). It is

clear that one can add to them any function of x which is in I, so we
need to characterize such functions.
Lemma 7.S.10. If R(x)eI(tl - Tl (x), ... , tm - Tm(x» then

(7.5.15) IR(x)I~CNIImT(x)IN, N=I,2, ...
in a neighborhood of O.
Proof. There exist Coo functions RIZ(x) and QIZ(t, x) such that
(7.5.16) R(x)= L: (t-T(x)tRIZ(x)/od +'L: (t-T(x)tQIZ(t,x)/oc!
O<IIZI<N IIZI-N
in a fIxed neighborhood of (0,0) for N = 1, 2, .... For N = 1 this follows

from the hypothesis that ReI. If (7.5.16) is known for one value of N
it follows with N replaced by N + 1 if we divide QIZ by t 1 - Tl (x), ... ,
tm - Tm(x) using Theorem 7.5.7. (As observed after the proof of
Theorem 7.5.6 the neighborhood can be chosen independent of N.)
With Ixl small we set
t=ReT(x)+sImT(x), O~s~l.
Then it follows from (7.5.16) that for O~s~ 1

IR(x) - L: (s - i)IIZI(Im T(X»IZ RIZ(x)/oc! I~ CN 11m T(x)I N.
O<IIZI<N
Any interpolation formula shows now that the coeffIcients of the
polynomial in s - i on the left-hand side have similar bounds, which
gives (7.5.15) with another constant.
Lemma 7.S.11. Let F~O be Lipschitz continuous in a neighborhood of 0

in JRn. If geC"" and Igl ~ CNF Nfor every N, it follows that
IDlZgl~CN'IZFN, for all rJ.,N.
Proof. Since F(X+X')~CF(x) iflx'I~F(x), we obtain
I L: g(IZ)(X)X'IZF(x)IIZI/rJ.!I~C~F(xt, Ix'I~I,
IIZI<N
if g(x + F(x) x') is expanded by means of Taylor's formula. Hence we

have an upper bound
Ig(lZ) (x) F(x)11Z1 I~ CN'IZF(x)N, 1rJ.1 <N,
for the coeffIcients. Since N is arbitrary the lemma is proved.
Theorem 7.S.12. If I=I(t1-T1(x), ... ,tm-Tm(x» where T1(0)= ...

= Tm(O) = 0 and ifR(x)e Coo then the following conditions are equiva-
lent:
(i) REI at (0,0).
n
(ii) (7.5.1S) is valid in a neighborhood of O.
(iii) R E [00 = IN at (0,0). More precisely, there is a neighborhood V
of (0,0) and functions q.. Ecoo(JR.,"+n) such that for every N
R(x)= L q.. (t,x)(t-T(x))"/od, (t,X)EV.
I"I=N
Proof (i) ~ (ii) by Lemma 7.5.10. From (ii) it follows by Lemma 7.5.11
that for all {3 and v we have in a neighborhood of 0
IR<")(x)I;:;:;CN,pP, F=Ll.fjI2, Jj=t j -1j(x).
Hence
R=(LJj~/F)N R= L j'"q ..
I.. I=N
where q.. = c.. J" F-I .. I R when F=l=O. Now qa. and all of its derivatives
converge to 0 when F -+0, so repeated application of Corollary 1.1.2
shows that q.. ECoo if q.. is defined as 0 when F=O. Since (iii) obviously
implies (i) the proof of the theorem is complete.
In Section 7.7 we shall also need the following application of Theo-
rem 7.S.6 which shows that if f is real valued, then the normal form
in Theorem 7.S.5 can also be reached by a change of variables without
multiplying by a function.
Theorem 7.5.13. Let f(t, x) be a real valued COO function near (0,0) in
JR.l +n such that
(7.5.17) f=af/at= ... =ak-1f/atk- 1=0,
akflat!' > 0 at (0,0).
Then one can find a real valued COO function T(t, x) with
T=O, aT/at>O at (0,0)
and COO functions aj(x) vanishing at 0 such that
k-2
(7.S.18) f(t, x)=Tk/k+ L aix) Tj
o
in a neighborhood of (0, 0).
Proof. Since f(t,O)=t!'c(t)/k where CECoo and c(O) >0, we can in-
troduce tC(t)l/k as a new variable. Thus we may assume that f(t, 0)
=tk/k. If n> 1 and the theorem is already proved for smaller values of
n, we may if x' =(x 1 , ... , xn _ 1) assume that
k-2
f(t, x', 0) =tk/k+ L bj(x')ti.
o
Set
k-2
F(t, x, a) = f(t, x) + I ajt j .
o
Then F(t, x, 0) = f(t, x) and we want to let t and a vary with Xn so that
F(t, x, a) remains constant. This means that we want to have
(7.5.19) (8f/8t+ kfjajtj-l) dt/dx n+ 8f/8x n+ kf da/dxntj=O.
Now Theorem 7.5.6 gives the decomposition
8f/8x n=q(t, x, a) (8f/8t+ kf jajt j - 1) + kf r/x, a) t j

in a neighborhood of 0, and (7.5.19) is fulfilled if
(7.5.20) dt/dx n= -q(t, x, a), da/dxn =
- r/x, a).
We integrate these equations with initial data t = T and a = A when xn
=0 and obtain Coo functions t(T, x, A) and a(x, A) defined near 0 such
that t(T,x,A)=T and a(x,A)=A when xn=O. In a neighborhood of 0
we have
k-2
f(t(T, x, A), x) + I a/x, A) t(T, x, A)j
o
k-2 k-2
=f(T, x', 0)+ I AjTj=Tk/k+ I (Aj+bj(x'))Tj.
o 0
Since the Jacobian matrices 8a/8A and 8(t, x, a)/8(T, x, A) are equal to
the identity when Xn = 0 we can in a neighborhood of 0 express A as a
function of x and a(x, A), and T as a function of t(T, x, A), x and
a(x, A). Hence
k-2 k-2
f(t,x)+ I ajtj=T(t,x,a)k/k+ I (A/x, a)+bix'))T(t, x, ay.
o 0
When we put a = 0 the theorem is proved.
7.6. Fourier Transforms of Gaussian Functions

If A is a symmetric n x n matrix with complex elements, it is clear that
the function
(7.6.1) lR.n3x~exp( -<Ax,x)/2)
is in 5/" if and only if <Re Ax, x) ~O, xEIRn.

Theorem 7.6.1. If A is a non-singular symmetric matrix and ReA~O

then the Fourier transform of (7.6.1) is another Gaussian function
(7.6.2)
where B =A -1 and the square root is defined as explained in Section
3.4. If A = - iAo where Ao is real, symmetric and non-singular then
(7.6.2) becomes
(7.6.2)' 1R."3~ --+ (2n)"/2Idet Aol-t exp (n i(sgn A 0)/4 - i(Aij 1~, 0/2).
Proof. Ifu(x)=exp(-(Ax,x)/2), then

(7.6.3)
Conversely, if u is a distribution satisfying these equations then the
derivatives of uexp(Ax,x)/2) are 0, so u is a constant multiple of
(7.6.1). If we pass to Fourier transforms in (7.6.3) we obtain
~ju=-i(AD)A j=l,oo.,n,
or if we multiply by B
i(B~)ju=DA j=l,oo.,n.
It follows that
U= C exp( -(B~, ~)/2).
When ReA is positive definite we have C=(2nt/2(detA)-±
= (2nt/2(det B)t by (3.4.1)", and when Re A is just non-negative the
same result follows if we replace A by A + d and let e --+ O. The special
case (7.6.2)' is obtained from (3.4.6).
Remark. The first part of the proof could of course be replaced by a

shift of the integration contour. - The quadratic forms (Ax, x) and
(B~,O are said to be dual. Since (B~,0=(X,02/(Ax,x) when
ox(x,02/(Ax,x)=0, this is an invariant correspondence between
quadratic forms in dual spaces.
Theorem 3.3.5 is now perfectly obvious, for E(x, t) is for t>O just
the inverse Fourier transform of exp- t(A~, 0 so it tends to <5 in !!"
as t --+ O. The solution of the Cauchy problem
(7.6.4) ou/ot+(AD,D)u=O, t>O; u(.,O)=VE!!'(IR.")
which is given by
u(x, t) = JE(x - y, t)v(y)dy
has the Fourier transform
U(~, t)=v(~) exp( - t (A~, 0)
with respect to x. This solution is also immediately obtained by

taking Fourier transforms with respect to x in (7.6.4).
We shall now study operators of this form in some detail to
prepare for some applications in Chapter XVIII. If A is any quadratic
form in 1R.n with Re A ~ 0 we write
e-A(D)v=(2n)~n Jei<x.o~A(~)v@d~, vEff,
for the inverse Fourier transform of e~A(~)v(~). It is clear that e~A(D)

maps ff into ff, ff' into ff' and L2 into L2 continuously. Note that
If we make a linear change of coordinates it is therefore clear that

e~A(D) is transformed to e~A'(D) if A'(D) is the differential operator
A(D) expressed in the new coordinates.
Theorem 7.6.2. If Re A ~ 0 we have for every integer k ~ 0

(7.6.5) Ile~A(D)u_ L (-A(D))ju/j!IIL2~ IIA(D)kUlk!IIL2' uEff.
j<k
Proof If we take Fourier transforms on both sides, the inequality

reduces by Parseval's formula to the inequality
lew - L ~/j!I~lwlklk!, Rew~O
j<k
which follows from Taylor's formula.
To pass to maximum norms in (7.6.5) we need an elementary case

of Sobolev's inequalities (Section 4.5). We shall give a direct proof and
a statement which will be useful also for later reference.
Lemma 7.6.3. If s is an integer > nl2 then

(7.6.6) lu(xW ~ C J (L IDau(yW + lu(yW)dy, uEff.
Ix~YI< 1 lal~s
Proof By Theorem 7.1.22 the operator ( - LI)S has a parametrix E with

support in the open unit ball. Thus (- LI)S E = t5 + w, WE CO', so
1~12s Em is a bounded function. Hence I~a E(~)I ~ C(l + IW~sEL2 if lod
=s. Since
the estimate (7.6.6) follows by Cauchy-Schwarz' inequality.

If we apply (7.6.5) to Dau when IIXI = s also, we obtain in view of

(7.6.6) since e-A(D) commutes with Da
(7.6.7) sup le-A(D)u(x)- L (-A(D))iu(x)/jW
j<k
~ C/k! L IIA(D)kDa u lli2, UEY.

lal~'
Hence e-A(D)u(x)-u(x) ~ 0 uniformly when IIAII ~O, and the differ-

ence is O(IIA Ilk) for every k outside supp u. We shall now prove a
more precise estimate there.
Lemma 7.6.4. If Re A ~ 0, IIA II ~ 1, and the Euclidean distance d(x) from
x to supp u is at least 1, then we have for every integer k ~ 0
(7.6.8) le-A(D)u(x)1 ~ Ck IIA Ilk+Sd(x)-k L sup IDaul, uEY(1Rn).
lal ~k+ 2.
Proof We may assume that x=O, thus u(y)=O when IYI<d=d(O).
Since Iy)llyl ~ 1/01 for some j if y =1= 0, we can choose a partition of
unity 1 = L Xj(y) in 1Rn" 0 so that XjE Coo is homogeneous of degree 0
and <tj , y) ~ IYI/2 01 when yESUpp Xj with ± tj equal to one of the
basis vectors. Since
L L IDaXjul ~ Ck L IDaul,
j lal ~k+ 2. lal ~k+ 2.
the estimate (7.6.8) will follow for u if it is known for each Xju. In
proving (7.6.8) we may therefore assume that
(7.6.9) <t,Y)~IYI/201 if YESUPPU,
where t is a unit vector. Now we have if B is the dual form of A
rA(D)u(O)=c JrB(Y)/4u(y)dy,
<y, t) e- B(Y)/4=2A(t, -iJ/iJy) rB(y)/4.
Hence repeated integrations by parts give for je=O, 1, ...
e-A(D)u(O) = 2j c JrB(y)/4(A(t, iJ/iJy)<y, t) - lyU(Y) dy
=2 j e- A(D)(A(t, iJ/iJy)<y, t)-lyu(y)ly=o'
If we now apply (7.6.7) with k = 0 it follows that
Ir A(D)u(O)j2 ~4.i C L IIDa(A(t, iJ/iJy)<y, t) -1)ju(y)lli2'
lal~'
If we recall (7.6.9) and that Iyl ~ d ~ 1 in supp u, we obtain if 2j> n

le-A(D)u(O)j2~ C j dn - 2j IlAII2,i( I suplDa u l)2.
lal~s+j
If we take j = k + s, the estimate (7.6.8) is proved.

Theorem 7.6.5. Let ReA~O, IIAII ~ 1, and let s be an integer >n/2.

Then we have for every integer k ~ s that e-A(D) u is continuous and
(7.6.10) IrA(D)u(x)- I (-A(D»ju(x)/j!1 ~ CkllAll k I suplDaul
j<k lal<;s+2k
if UE CS+ 2k and the right-hand side is finite. At Euclidean distance d(x)

from supp u greater than 1, the stronger bound (7.6.8) is valid.
Proof Assume first that UEC~. Let x=O and choose XEC~ equal to 1
in the unit ball. If we apply (7.6.7) to XU and (7.6.8) to (1- X) u it
follows that (7.6.10) is valid for x =0, hence for every x. When
UE C~+ 2k we can apply this result to the regularizations of u or rather
their differences to conclude that e-A(D)u is continuous and that
(7.6.8), (7.6.10) remain valid for u. If only UE CS+ 2k and the right hand
side of (7.6.10) is finite we apply this result to uR=X(./R)u and
conclude when R -> 00 that rA(D)uELco and that (7.6.8), (7.6.10) hold
almost everywhere. (Recall that rA(D) is continuous in !fl, hence
e-A(D)uR . . . . e-A(DJ u in ,'I".) But e-A(D)uREC and e-A(D)(u-uR) ........ O in LCO
on every compact set by (7.6.8) so it follows that rA(D)u is in fact
continuous.
Note that when ReA is not posItlve definite it is in no way

obvious a priori that e-A(D)u must be a continuous function when the
derivatives of U are bounded but do not tend to 0 at 00. Theorem
7.6.5 can be regarded as a statement on the Cauchy problem (7.6.4).
As already mentioned the preceding results will be important in
Chapter XVIII. The local properties of r A(DJ will be developed fur-
ther in Section 18.4. Another aspect of the preceding results is the
method of stationary phase which will be studied in Section 7.7.
The following application of Theorem 7.6.1 was mentioned after
Theorem 7.1.13.
Theorem 7.6.6. If (j is a distribution of order ~k for every uEll(JR")

then
(7.6.11) k ~ n(1/2 -l/p).
Proof The assumption implies if K is a compact set in JR" that

(7.6.12) l<u,cp>I~ClluIILP I supID"cpl;
1"1 <;k
UEll, cpEC~(K).
In fact, <u, cp) is continuous with respect to UEll for fixed cpEC~(K)
since <u,cp)=<u,¢) and ¢EY'. By hypothesis <u,cp) is also con-
tinuous with respect to ljJ in the Ck norm for fixed uEIJ'. A separately
continuous bilinear form in the product of a Frechet space and a
metrizable space is always continuous, so (7.6.12) follows. Choose now
U * 0 in g' so that U has compact support K, and apply (7.6.12) to ut
and CPt where
Ut(e) = u(e) eitl~12, cptm = ut(e)
and t is a large positive number. Then ut is the convolution of U and
ccn/2e-ilxI2/4t, with a constant c which can be obtained from (7.6.2),
so we have if p>2
Ilutliv = lIullv, Ilutll Loo ~ct-n/21Iullu·
Hence
Ilutllip = Jlutl P - 21utl2d X ~ Cu t-n(p- 2)/2.
The left-hand side of (7.6.12) is a constant when u = up cP = CPt' so we
have
1 ~ C't"(l/p-t)t!'
for large t, which proves (7.6.11). (Note that (7.6.11) is void if p~2.)
In Section 7.9 we shall prove that UE~'k for every UEIJ' if there is
strict inequality in (7.6.11).
We shall now discuss two examples which show that Theorem
7.6.1 can sometimes be used to solve differential equations with vari-
able coefficients also. The first example is the construction of a fun-
damental solution for the Kolmogorov equation
(7.6.13) iJ2ujoX2+xoujoy-oujot=0, t>O;
u(t,.)-+b(xo.Jlo) as t-+O.
Assume that a Fourier transform U with respect to (x, y) exists and
behaves well as t-+O. Then we obtain by taking Fourier transforms in
(7.6.13)
-e 2 U -y/oUjoe -oU/ot=O, t>O.
Thus dU = - e U dt if dy/=O and de =y/dt, so
2
t
U(t, e+ Y/t, Y/) = U(O, e, Y/) exp Jo -(e + y/s)2ds
= U(O, e, y/)exp( -e 2t-ey/t -y/ 2t j3) .
2 3
. h B=
WIt (2t
t2
so the exponential is the Fourier transform of

(x, y) -+ J!. /(2nt 2) exp( - X2 /t + 3xy/t2 - 3 y2/t3).
By inverting the Fourier transformation we obtain when U(O, e, 11)
=e-ixO~-iYOlJ
u(t, x, y-tx)=3 t /(2nt 2) exp( -(x-xof/t

+ 3(x - xo)(Y - yo)/t 2 - 3(y - Yof /t 3).
Writing u(t,x,y)=E(t,x,y;xo,yo) we have
E(t, x, y; x o, Yo) = 3t /(2nt 2) exp( - (x _XO)2 /t

+ 3(x -xo)(y + tx - yo)/t 2 - 3(y+tx- YO)2/t 3).
It is now easy to reverse the argument to show that E is a COO

function when t>O and satisfies (7.6.13). The map
Cg'(lR 3)3/ -+ E/ECOO(1R 3),

E/(t,x,y)= I E(t-s,x,y;xo,Yo)/(S,xo,yo)dxodYods
s<t
is a twosided fundamental solution of the Kolmogorov operator in

(7.6.13), that is, a left and right inverse of the operator.
Our second example is the construction of a fundamental solution
for the operator -D~+2DoDl +xiD~. Because of translation in-
variance in Xo and X2 it suffices to find a solution u of the equation
(7.6.14)
We want u to have support where xo>O which will lead to solution of

the Cauchy problem. Let U be the Fourier transform with respect to
x 2 , and the Fourier-Laplace transform with respect to Xo which we
expect to exist when 1m eo <0. Then we obtain the equation
(7.6.15)
For the Eq. (7.6.14) without the xi D~ term the fundamental solution
has support where Xl <a, 2XO+Xl -a>O, so it is natural to expect that
u shall vanish when Xl >a. Integration of (7.6.15) gives when Xl <a
U = C(e o, e 2) exp(i(e~xl - xi e~/3)/2eo)

and if this together with U =0 for Xl >a shall satisfy (7.6.15) we must
have 2e o Ci exp(i(e~a-a3 e~/3)/2eo)= 1, hence
U(eO'x l , e 2)=( -i/2eo)exp(i(e~(Xl -a)-e~(xi -a 3)/3)/2e o), Xl <a.
212 VIr. The Fourier Transformation
Note that the real part of the exponent is

Im';o«a-xJ/2+,;~(a3-xi)/61';012);:;;0 when Im';o<O.
If we now use Theorem 7.6.1 to invert the Fourier transform with
respect to x 2 we obtain the function
V(';O' Xl' xJ =( - i/2 ';0)(3 i';o/(a 3 -xi))t (2 n)-t exp iE,

E = ';0«x 1- a) + 3x~/(xi - a 3))/2.
Here (i';o)t is considered as lying in the right half plane, so (i';o)-t

=r1ti/4';ot when -n<arg';o<O. By Example 7.1.17 this is the Fourier-
Laplace transform with respect to Xo of xo}/r(1/2)=xo}/nt. The
factor eiE gives a tran"Jation so we obtain
u(x) =V3(2n)- 1((2xO +x1-a)(a 3 -xf) - 3x~)- t

when xo>O, Xl <a, and the expression under the square root is
positive. Otherwise u=O. The integral with respect to x 2 is inde-
pendent of Xo and Xl so this is a locally integrable function. It is an
easy exercise to go back through the calculations now to show that u
does satisfy (7.6.14).
In a discussion of Gaussian functions it would not be natural to

omit the central limit theorem although it is outside our main topic:
Theorem 7.6.7. Let Ji be a positive measure in JR" such that
SdJi=l, SlxI 2 dJi<oo, SxdJi=O.

Suppose that no hyperplane through 0 contains the support of Ji and set
Ajk= S xjxkdJi(x).
A is then a positive definite symmetric matrix. If Jiv is V"/2 times the
composition of the v-fold convolution Ji* ... * Ji with the map X --+ xvi, it
follows that in the weak topology of measures
Jiv --+ Idet 2nAI- t exp( - <A -1 X, X) /2).
Proof Since
L.AjkYjYk= S<x, y)2dJi(x)
is positive unless <x, y) =0 when XESUPP Ji, it is clear that A is
positive definite. The convolution of two positive measures 11' and Ji"
with total mass 1 is defined by
(11' * Ji")(¢) = IS ¢(x + y) dJi'(x) dJi"(y)
if 4> is a bounded continuous function. As observed in Section 5.1 this

agrees with our earlier definition when J1." and 4> have compact sup-
port. It is clear that J1.' *J1." has total mass 1 and that the Fourier
transform is P' p". Hence the Fourier transform of J1.. is (see (7.1.17))
p.(e) = p{eNvt.
Since SdJ1..(x) = 1 for every v it follows that IP.WI ~ 1. We may differ-
entiate twice under the integral sign in
p(e) = Se- i (x. o dJ1.(x)
and conclude that PEC 2 , DjDkP(O)=Ajk' DjP(O) =0. Hence
p(e) = 1-(Ae, 0/2 +o(leI 2 )
so
p(eNv) = 1-(Ae, 0/2v+o(1/v)=exp( -(Ae, 0/2v+o(1/v))
uniformly on any compact set. Hence it follows by dominated con-
vergence that
P. --+exp( -(Ae, 0/2) in g".
Now the continuity in g" of the inverse Fourier transformation gives
J1.. --+ Idet2nAI-t exp( -(A -1 x, x)/2)
in g". But all J1.. are positive measures so we have convergence in the
weak topology of measures by Theorem 2.1.9.
So far we have only discussed the Fourier transform of exp Q when

Q is a quadratic polynomial. We shall end this section by a brief
study of the simplest cubic polynomial Q in one variable. To conform
with standard notation we consider an inverse Fourier transform:
Definition 7.6.8. The Airy function Ai(x) on 1R is the inverse Fourier

transform of e --+ exp(ie 3 /3).
A priori A i is just in g", but we shall prove that A i is a Coo

function defined by
(7.6.16) Ai(x)=(2n)-1 S exp(i,3/3+i'x)d,.
Im{~~>O
To do so we observe that for real x we have if' = ~ + il'/

Re(i,3/3+i,x)= _~21'/+1'/3/3_l'/x.
This proves that the integral converges and is independent of 1'/.
Moreover,
exp(i(¢+i1])3j3)--+exp(i¢3j3) in [I" as 1]--+ +0,

so (7.6.16) follows when '1--+ +0. Since ¢2'1 grows faster than any
linear function of ¢ when ¢ --+ 00 the integral in (7.6.16) converges for
all XE<C. Hence A i(x) is an entire analytic function of XE<C. It satisfies
the Airy differential equation
(7.6.17) Ai"(x)-xAi(x)=O
because
From the equation it follows that Ai"(O)=O. A change of integration

contour gives
J exp(i(3j3)d(=Ren- 1Je"i/6 r
()()
Ai(0)=(2n)-1 t 3/3dt
R+i 0
= 3-* r(1j3)j2n,
Ai'(0)=(2n)-1 Ji(ei~3/3d(= -3 t r(2j3)j2n.
If w is a cubic root of unity, that is, w 3 = 1, then it is clear that
x --+ Ai(wx) is another solution of (7.6.17) with the same value at 0 but
with the derivative wAi'(O). Hence any two of these solutions form a
basis for the solutions of Airy's differential equation. The linear re-
lation between the three of them is
(7.6.18) LwAi(wx)=O
which follows since the value of this solution at 0 is A i(O) L w = 0 and
the derivative is Ai' (0) w 2 = o. L
When x --+ + 00 we obtain an asymptotic expansion for the Airy
function by choosing '1 in (7.6.16) so that the derivative of (3 j3 + (x
vanishes at i'1, that is, '12 =x. By Taylor's formula we have then
so we obtain
J e-~2v'X+i~3/3d¢.
()()
(7.6.19) Ai(x)=exp(-2xtj3)(2n)-1
-()()
As in the proof of Theorem 7.6.2 we expand ei~3/3 in a finite Taylor

series and obtain an asymptotic series with the terms
J e-~2v'X( _1)k¢ 6k 3- 2k j(2k)! d¢

()()
- 00
=( _1)k 3- 2k r(3 k+t)f(2k)! X-(6k+ 1)/4.

Fig. 4. Ai(x) and Ai'(x)j2
Hence
L( -9)-k r(3k+t)/(2k)!x- 3kl2
00
(7.6.20) A i(x)~(2n)-1 exp( - 2xt/3) x- t

o
which means that the difference between A i(x) and a partial sum is
smaller in absolute value than the first term left out. The preceding
result is not only valid for positive x. If e > 0 we still have (7.6.19)
when larg xl < n - e, and (7.6.20) remains valid although the error must
be estimated by the next term with Vxreplaced by Re Vx.
Note that
Ai(x) is exponentially decreasing when largxl «n-e)/3, oscillatory
when argx= ±n/3, and exponentially increasing when (n+e)/3
<Iargxl<n-e. To see what happens when argx is close to -n we
can use (7.6.18) which gives in particular when r>O
(7.6.21) Ai( -r)=n--!r*(sin(trt + n/4) + O(r--f)).
This proves that there is a zero of A i( - r) with 2rt /3 + n/4 close to nn
when n is a large positive integer. In fact there is one for n = 1, 2, ...
and these are all the zeros of the Airy function. (See Fig. 4 where the
dotted curves give the leading term in the asymptotic expansion.)
7.7. The Method of Stationary Phase
In this section we shall make a systematic study of the asymptotic

behavior of integrals of the form
Sueirof dx
where f and u are smooth, Imf ~ 0 and w ~ + 00. If UE Cg', f E COO is

real valued and f' =1= 0 in supp u, then
Jueiwf dx=O(w- k), w~ 00,
for every k. This follows from Lemma 7.1.3 if u has so small support
that f can be taken as a coordinate in a new local coordinate system
in a neighborhood of supp u. We can reduce to this situation by
decomposing u with a partition of unity. In the following theorem we
elaborate the integration by parts in Lemma 7.1.3 to obtain uniform
bounds and to cover the case of complex valued f also.
Theorem 7.7.1. Let K e1R." be a compact set, X an open neighborhood

of K and j, k non-negative integers. If uEq(K), fEC k +1 (X) and
Imf~O in X, then
(7.7.1) Wi+klJ u(x)(lmf(x))ieiwf(X)dxl
~C L sup ID a ul(lf'1 2 +Imf)l aI!2-k, w>O.

lal~k
Here C is bounded when f stays in a bounded set in Ck +l(X). Jilihen f

is real valued, the estimate (7.7.1) reduces to
(7.7.1), wkIJu(x)eiwf(X)dxl~C L supIDa ullf'l'a l-2k, w>O.

lal~k
Proof. When k=O the assertion is obvious since tir! is bounded for
t>O. In the proof we may therefore assume that k>O and that (7.7.1)
is already proved for smaller values of k as well as for smaller values
of j and the same k if j > O. It may also be assumed that
N=1f'12+Imf
is positive in K, for if (7.7.1) is proved then we can always replace f
by f+ei in (7.7.1) and let e~ +0.
To be able to raise j or integrate by parts we observe that

n
Nu=u L loflox vI2+ulmf,
1
hence
n
(7.7.2) U= Luvofloxv+u o Imf
1
if NUv=uollox v when v =1=0 and Nuo=u. Since

iwofloxveiwf =ov eiwf
we obtain after an integration by parts

j u(Imf)ieiwf dx = j uo(Imf)i+ 1 ffwf dx
+i/w L(j(o.uv)(Imf)ieiWf dx+ jju.(ovlmf)(ImfY-leiwf dx) .
•
With the notation
lul",= L 1000ui
1"1=,,,
we have by the inductive hypothesis
(7.7.3) Wi+klj u(Imf)iffWf dxl
~ C sup (~: (Iuol", + .tl1uvl",+ 1) N"'/2-k+ 1

+ ",ta ~ lu.o Im fl",N",/2-k).
v
To estimate the right-hand side we need a lemma.
Lemma 7.7.2. If gEC 2 ( -(j)j) is non-negative, then

(j2Ig'(OW ~g(O)(g(O) + 2 sup (j2Ig"(X)I).
Ixld
Proof The proof is reduced to the case (j = 1 if x/(j is introduced as a

new variable. We may also assume g(O) = 1. By Taylor's formula
1+g'(0)x+Mx2/2~g(x)~0, Ixl<l; M=suplg"l.
If M~2 we take x 2 =1 and obtain Ig'(0)1~1+M/2~(1+2M)t since
M2/4~M, and if M>2 we take x 2 =2/M and obtain Ig'(0)1~2/lxl
=(2M)t. This proves the lemma.
End of proof of Theorem 7.7.1. Lemma 7.7.2 applied to Imf gives

(7.7.4) lo.lmfI2~Clmf, lovNl2~CN in K.
We shall now prove that when JJ~k we have
(7.7.5) Nt L IUvl" + N IUol" ~ C L lulrN(r-

n '"
",)/2 •
v= 1 r= 0
This follows from the definition of Nand u. if JJ =0 so we may

assume that 0 < JJ ~ k and that the estimate is proved for smaller
values of JJ. Let letl = JJ and apply 0" to the equations
Nu.=uoJjox v when V=FO, Nuo=u.
Estimating all terms except N 8a uv in the left-hand side by means of

(7.7.5) we obtain when v=l=O in view of (7.7.4)
N IUvl1' ~ c(INI1Iuvll'_ 1+ IUvll'_ 2 + ... + IUvi o + lull' If I1+ lull'_ 1+ ... + lul o)
~ C(Ntluvll'_l + IUvll'_ 2 + ... + lull' Nt + lull'_l + ... )
L lulrN(r-l'+ 1)/2
I'
~ C"
o
which gives (7.7.5) then. Similarly
I'
~ CL lul r N(r-I')/2
o
which completes the proof of (7.7.5). From (7.7.5) and (7.7.4) we obtain
I'
(7.7.6) luv8vImfll' ~ CL lul r N(r-I')/2, v =1=0, J1~ k.
o
If the estimates (7.7.5) and (7.7.6) are used in the right-hand side of
(7.7.3) we obtain (7.7.1).
Theorem 7.7.1 shows that the integral

Sueirof dx, W > 0,
is a rapidly decreasing function of w if UECg', fECoo, Imf~O, and
there is no point in supp u with Imf = 0 and f' = o. Thus the essential
contributions must always come from points where the phase f is real
and stationary. This is the basis for the stationary phase method which
in addition describes the contributions from the simplest types of such
critical points. The special case where f is a quadratic form was
discussed in Section 7.6 but we restate the result in a form which is
more convenient here.
Lemma 7.7.3. Let A be a symmetric non-degenerate matrix with

1m A ~ O. Then we have for every integer k > 0 and integer s> n/2
(7.7.7) IS u(x) eiro (Ax,x)/2dx -(det(wA/2ni))-t 7;.(w)1
~Ck(IIA-111/w)n/2+k L IIDaullu, UE[/,
lal~2k+.
k-1
(7.7.8) 7;.(w)= L (2iw)-i<A- 1 D,D)i u(0)/j!.
o
Proof By Theorem 7.6.1 the Fourier transform of x -exp iw<Ax, x)/2

is
~ -exp( - i<A -1~, O/2w)(2n)"/2(det(iw-1 A -1»t
=exp( - i<A -1~, O/2w)(det(wA/2ni))-t.

Hence
Seiro (Ax.X)/2 u(x) dx =(det(wA/2n i))-t e-i(A -lD.D)/2ro U(0)
so (7.7.7) follows from (7.6.7).
Remark. If B is another symmetric matrix with ImB~O we obtain by

replacing A with A + tB in (7.7.7) and expanding the t derivative of the
integral that
(7.7.9) (2iw)-i [:t «A+tB)-1 D,D)iu(O)/j!-tTr B(A+tB)-1]
=(2iw)-i-1«A + tB)-1 D, D)i+ 1(iw<Bx, x) u(x)/2)x~ o/U + 1)!

when t is small. This is of course an identity which remains valid for
all symmetric matrices with A + tB non-singular.
The calculation of sums like (7.7.8) is occasionally simplified by
rewriting them as if part of the exponential in (7.7.7) were a factor of
u. The algebraic contents of this are described in the following lemma.
Lemma 7.7.4. Let A be a symmetric non-singular matrix, and let B be a

symmetric matrix such that det{A + tB) is independent of t. This means
precisely that A - 1 B is nilpotent, that is, (A - 1 B)k = 0 for some k. With
the same k we have when w - 00
(7.7.10) L (2iw)-i«A+B)-1D,D)i u(0)/j!
i<N
- L (2iw)-i<A -1 D, D)i(eiro (BX,x)/2 u)(0)/j! =O(W- N / k).
i<N
Proof If Ai are the eigenvalues of A -1 B then
det(A+tB}=detA det(I +tA -1 B)=detA f1 (1 +tAJ

This is independent of t if and only if all Ai are 0, that is, A - 1 B is
nilpotent. If (A - 1 B}k = 0 we have
k-1
(A+tB}-1=(/+tA- 1B}-1A- 1= L (-tA- 1B)"A- 1.
• ~o
If we expand the polynomial «A + tB}-1 D, D)iu(O) in t by Taylor's
formula and use (7.7.9) we can write the first sum in (7.7.10) in the
form
N-l j(k-l)
(7.7.11) L L (2iw)-j-v<A- 1 D,D)j+v«iw<Bx,x)/2)V u)(0)/U+v)!v!.
j= 0 v= 0
By Taylor's formula the second sum in (7.7.10) is equal to
L (2iw)-j <A- 1 D,D)j«iw<Bx,x)/2)'u)(0)/j! v!.

j<N
The terms with v> j vanish so we can replace j by j + v where j ~ 0,

v~O andj+v<N. This means that we have the terms in (7.7.11) withj
+v<N. For the missing terms j+v~N and v~j(k-l), hence N~jk.
This means that they are O(W- N1k ) which completes the proof.
Using Taylor's formula it is easy to extend (7.7.7) to more general

phase functions:
Theorem 7.7.5. Let K elRn be a compact set, X an open neighborhood

of K and k a positive integer. If UEC~k(K), fEC 3k + 1 (X) and Imf~O
in X, Imf(x o) =0, f'(xo)=O, detf"(xoHO, 1'=1=0 in K-.....{x o} then
(7.7.12) II u(x) eicof(X)dx - eiCO /(XO)(det(wf"(x o)/2n i))-t L w- j Ljul

j<k
~ Cw- k L suplDaul, W>O.
lal;;02k
Here C is bounded when f stays in a bounded set in C 3k + l(X) and

Ix-xol/lf'(x)1 has a uniform bound. With
gxo(x) = f(x)- f(xo)-<f"(xo)(x-x o), x-x o)/2
which vanishes of third order at Xo we have
Lju= L L i-j2-v<f"(XO)-lD,D)V(g~ou)(xo)jp,!v!.
V-/J=j 2v~3/J
This is a differential operator of order 2j acting on u at xo. The

coefficients are rational homogeneous functions of degree - j in
f"(x o), ... , p 2i+ 2)(X O) with denominator (detf"(x o))3 j . In every term the
total number of derivatives of u and of f" is at most 2j.
Proof. First note that since

f'(x) = f'(x)- f'(x o) = f"(xo)(x -xo)+O(lx -x OI2 )
we have
Ix-xol ~ 1If"(x o)-lll(lf'(x)1 + Clx -x O I2 ).
Hence Ix-xol~211f"(xO)-llllf'(x)1 if Ix-xol is small enough, so the

hypothesis made on f in (7.7.12) implies that Ix-xol/lf'(x)1 is bound-
ed near xo' To have C bounded we have just added a condition
which excludes that f' is close to 0 at some other point in K.
If D"'u(xo) =0, lexl <2k, then Taylor's formula gives
ID"'u(x)I~Clx-xoI2k-laIM, M= I suplDllul.
IIlI ~2k
Since Ix-xol/lf'(x)1 is bounded it follows from Theorem 7.7.1 that
II ueirof dxl~ CMw- k.

If PEC,(;, is equal to 1 in a neighborhood of Xo and u 1 is the product
of p and the Taylor expansion of order 2k at Xo of any UEC~k(K) we
can apply this result to U o = u - ul . Hence it remains only to estimate
I Ul eirof dx. All derivatives of u l can be estimated by M, and the
support is close to xO'
Assuming as we may that f(xo)=O we set
f.(x) = (f"(xo)(x -x o), x -x o)/2 +sgxo(x).

Thus fl(X)=f(x) and fo(x) is a quadratic form in x-x o with Imfo~O.
If supp p is sufficiently close to Xo we have a uniform bound for
Ix-xolllf;(x)1 when O~s~l and XESUppp. Differentiating
l(s) = I U 1(x) eirofs(X)dx
2k times we obtain
1(2k)(s) = (iW)2k I u 1 (x) gxo(X)2k eirofs(x) dx.
When lexl ~ 3 k we have

IDa(U1g;:)(x)llx-xollal-6k~ CM
so it follows from Theorem 7.7.1 with k replaced by 3k that
11(2k)(s)1 ~ CMw-k, O~s~l.
By Taylor's formula
11(1)- I 1(/l)(O)I.u!I~ sup 11(2k)(s)II(2k)!
/l<2k 0<s<1
which has the desired bound. If Gxo is the Taylor expans~on of gxo of
order 3 k then
g~o - G~o = O(lx - XOI3k + 3/l- 2) = O(lx - XOI2k+ 2/l)
and this is a function in C 3k c Ck + I' so another application of Theo-

rem 7.7.1 gives
11(1')(0)- S(iwGxJl'u l eiwfodxl~ CMw- k •
If we apply Lemma 7.7.3 to the integral it follows that
IS uei wf dx-(det(wf"(x o)/2ni))-t L L (2iw)-V

1'<2k v~l'+k
x <f"(XO)-l D, D)V(iwGxo)l'u(xo)/v!/-L!I ~ CMw- k •
To get a non-zero term in the sum we must have 2 v ~ 3/-L. When V-/-L
=j this implies 3j=3v-3/-L~v, 2j=2v-2/-L~/-L so there is just a finite
number of terms for each j. Altogether 3/-L derivatives are required to
remove the zero of Gxo at x o, and 2v-3/-L=2j-/-L derivatives then
remain. If /-L=O we obtain 2j derivatives acting on u. If /-L>O we have
altogether 2j -/-L derivatives at most acting on u and on (G~~)Il, that is, at
most 2j derivatives acting on u and on f", which completes the proof.
Remark. When f is real one can also prove Theorem 7.7.5 by using
the Morse lemma to change variables so that f(x) - f(x o) becomes a
quadratic form. This gives a result of the same form as Lemma 7.7.3
when feC 2k + 2 +s, and a similar improvement can be made when f is
complex. However, we are primarily interested in the COO case and
have chosen a proof which assumes higher regularity of f than neces-
sary but gives an effective method for computing the quite com-
plicated expressions L j •
In practice the functions u and f usually depend on parameters.

As we shall now see this causes no problem when f is real valued.
For the sake of simplicity the result is stated in the Coo case only.
Theorem 7.7.6. Let f(x, y) be a real valued Coo function in a neigh-

borhood of (0,0) in 1RHm. Assume that f;(O, 0)=0 and that f;~(O,O) is
° °
non-singular, with signature u. Denote by x(y) the solution of the
equation f;(x, y) = with x(O) = given by the implicit function theorem.
Then there exist differential operators Lf,j,y in x of order 2j with the
properties in Theorem 7.7.5 such that when uECO'(K), K close to (0,0),
(7.7.13) IS u(x, y) eiwf(x'Y)dx - eiwf(x(y)'Y)ldet(wf;~(x(y), y)/2n)l-t
k-l
x e"ia/4 L Lf,i,Yu(x(y), y)w-jl ~ Cw- L k supID~u(x, Y)I.
o lal~2k x
The theorem is an immediate consequence of Theorem 7.7.5. It is

of course sufficient to sum for j + nl2 < k. Sometimes it is advan-
tageous to sum the terms in a different order as in Lemma 7.7.4. We

give an example which will be useful in Section 18.1.
Theorem 7.7.7. Let f be a real valued function in c oo (1R."+m). If K is a
compact subset of 1R 2n + m and uECO'(K), k~n, then
(7.7.14) IS u(x, ~,y) eiw(f(x.y)- <x.O)dxd~
k-n
- eiwf (O'Y)(2n/w)n L <iDx!w, D~)v(eiwry(X)u(x, ~,y))x= o.~=f;,(o.y)1
o
~ Cw-(k+n)/2 L supID~.~u(x,~, Y)I.
l"I~2k x.~
Here ry(x)=f(x,y)-f(O,y)-<fJO,y), x) vanishes of second order when

x=O.
In the sum in (7.7.14) the ~ derivatives must act on u, and x de-
rivatives acting on eiwry(x) bring out with w a derivative of ry vanishing
at x = 0. Another x derivative must act on it to give a non-zero
contribution. This shows that the terms in the sum are 0(w- V / 2 ), for at
°
most v/2 derivations bring out a factor w.
Proof The differential of the phase f(x, y) - <x,
° °
with respect to x
and ~ is equal to precisely when x = and ~ = f;(O, y). The Hessian
1,11
there is the sum of the Hessian B = ( ~x
0)° of f(x, y) and the
Hessian A= (~I ~I) of -<x,O. Note that IdetAI=1, sgnA=O, A
= A-I and (A - 1 B)2 = 0. The asymptotic expansion of the integral in
(7.7.14) is according to Theorem 7.7.6 given by
eiwf (O.Yl(2n/wt L (2iw)- v «A + B)- 1 D, D) v(eiWRy(X)u)/v!x= 0
v<k-n
where Ry(x)=rv(x)-<f;~(0,y)x,x)/2 and the form «A+B)-1 D,D)=

-2<Dx,D~)-<f;~(0,y)D~,D~) is of order 1 in x. As observed above
the terms left out in the sum are 0(w-(k-nl/2). By Lemma 7.7.4 and the
same argument the difference between the sum and
L (2iw)-V<A -1 D, D)V(eiwry(X)U)/V!
v<k-n
can be estimated by Cw-(k-n)/2 L ID~.~ul which proves (7.7.14).

l"I~2k
Remark. It is easy to give a direct proof similar to that of Theo-

rem 7.7.5.
When the phase f is allowed to be complex valued there is a
difficulty already in stating an analogue of Theorem 7.7.6, for the
equation f;(x,y)=O no longer defines a critical point xEIRn when
y=l=O. To cope with this problem we can use the Malgrange prepara-
tion theorem proved in Section 7.5. Assume that f(x, Y)E Coo in a
neighborhood of (0, 0) in 1R.n + m and that
(7.7.15) Imf~O; Imf(O,O)=O, f~(O,O)=O, detf~~(O,O)=l=O.
Using Theorem 7.5.9, where the variables (t, x) correspond to (x, y)

now, we can find XiY)EC OO at 0 in 1R.m so that
1 = 1(0 flox 1, ... , ofloxn) = l(x1 - X 1 (y), ... , xn - Xn(y)).
If X(y) is real then x -> f(x, y) has a critical point at X(y). Otherwise
there is no critical point near 0, but it is suggestive to think of X(y) as
a critical point which has become complex although this is only a
figure of speech. Repeated use of Theorem 7.5.7 gives (see the proof
of Lemma 7.5.10)
(7.7.16) f(x,y)= L f"(y)(x-X(yW/a!modl N
lal<N
for some f"E COO. Since ~f(x, y) - f"(Y)El we can think of f"(y) as the
value of o:;f(x, y) at the critical point. When lal = 1 we have in particu-
lar fa(Y)El, hence fa(Y)El oo by Theorem 7.5.12. We can therefore
choose f"=0 in (7.7.16) when lal = 1, and f - fOEI2.
Lemma 7.7.8. The hypothesis (7.7.15) implies that near 0, for some
C>O,
(7.7.17)
Neither fO nor X are uniquely determined. However, Theorem

7.5.12 shows that another choice would only change the two sides by
terms which are OOlm XIN) for every N, and this does not affect the
validity of (7.7.17) near 0 since X (0) =0.
Proof of Lemma 7.7.8. By hypothesis and (7.7.16) we have

O~lmf(x,y) =Im(f°(y)+ L f"(y)(x-X(y)na!)+O(lx-X(yW).
lal=2
Choose x=ReX(y)-tllmX(y)1 with tE1R. n, Itl<1. If X(y) is real we

obtain (7.7.17) at once. Otherwise we have with '1=lmX(y)/llmX(y)1
O~lm(jD(y) + IlmX(yW L f"(y)(t+ i'1)a/a!)+ O(llm X(yW)·
lal=2
Hence
1m fO(y) ~ 11m X(yW(sup - 1m L f"(y)(t+ i'1)a/a! - C IlmX(y)I).
111<1 lal=2
Now r(y) -+ r(0)=8~j(0, 0) when y -+ O. If Ajk=8 2j(0)/8x j 8xk then

det A =1= 0 and 1m A ~ 0 by (7.7.15). The proof is therefore completed by
the following
Lemma 7.7.9. If A is an invertible symmetric n x n matrix with ImA ~O

then there is a positive constant c such that, with t, '7E1R",
(7.7.18) sup - 1m <A(t+ i'7), t+ i'7) ~c, 1'71 = 1.
1/1< 1
Proof For compactness reasons it suffices to prove this for a fixed '7.
Write A=Ai +iA 2. Then
-lm<A(t+i'7), t+i'7)=<A 2'7, '7) -<A 2t, t) -2<Ai t, '7).
If <A z '7, rJ) =1=0 we just take t=O. If <A 2 rJ, rJ) =0 then A 2rJ =0 since A2
is semi-definite, so Ai rJ =1= 0 since A'7 =1= O. Then we can take t = - EA 1 rJ
with a small positive E.
By analytic continuation from the real case we shall now extend

the stationary phase formula in Theorem 7.7.6 to the complex case.
Set as above
j:~(O, O)=A =Ai + iA 2.
Then detA=I=O and we have det(A i +zA 2)=I=O when Imz=I=O. In fact, if
XE<c n and (Ai +zA 2)x=0 then
O=lm«Ai +zAz)x, x) =lmz<A 2 x, x).

This implies A 2 x=0 if 1m Z =1= 0, for A2 is semi-definite. Hence A i x=O
so Ax=O which implies x=O. The equation det(A i +zA 2)=0 can
have at most n real roots z. Thus we can choose AE[O,I] so that
det(A i +AA 2)=1=O. Then
det(A i +zA 2) =1=0, ZEZ,
if Z is a sufficiently small neighborhood of the line segment [A,IJ
between A and i. Now introduce
F(x,y, z)=Rej(x, y)+zlmj(x, y), ZEZ.
If x and yare sufficiently small then detF~~(x,y,z)=I=O, ZEZ, and we
have
ImF(x,y,z)~O if ZEZ+ ={ZEZ; Imz~O}.
Let K be a small compact neighborhood of (0,0) in 1Rn+ m and let
uEC~(K). To obtain an asymptotic expansion of
s(ro, y)= Ju(x, y)eirof(X'Y)dx
when W -> + 00 we shall more generally consider

S(w, y, z)= Su(x, y)eiWF(X,Y,Z)dx, ZEZ+.
An expansion of S is given by Tht:orem 7.7.6 if ZEZ o =Z n1R. Since

ImF(x,y,z)~O we also have a uniform bound for S when ZEZ+. This
will lead to an asymptotic expansion of s(w,y)=S(w,y,i) after we
have examined the analyticity of the expansion (7,7.13). When ZEZ o
the terms in the asymptotic expansion are of the form
(7.7.19)
evaluated at the critical point defined by F~=O. Here LF,j is a differ-
ential operator in x of order 2j such that the coefficients multiplied by
(detF~~)3j are polynomials of degree 3nj-j in D~F, 2;;;;lor:!;;;;2j+2. The
square root of the determinant is of course defined as explained in
Section 3.4.
To extend the definition of (7,7.19) to Z + we shall consider the
ideal I generated by of(x, y, z)/oxj,j = 1, ... , n, in a neighborhood of
M = {(a, 0, tA +(1- t)i); 0;;;; t;;;; 1}
in 1Rn+m x <C. Since F~=O on M, it follows from Theorem 7.5.9 by a
partition of unity in Z that there exist generators of the form
xj-X/y, z). Here Xj satisfy the Cauchy-Riemann equation approxi-
mately; more generally we have
Lemma 7.7.10. If tp(x, y, z) is a Coo function in a neighborhood of M

and 8'1' /8z = 0, then
tp(x, y, Z)= I I/I/x, y, z)8F(x, y, z)/8x j + tp0(y, z)
where 0 '1'0 /ozEIoo, hence
l0tp°/8zl;;;; CNIImX(y, z)I N, N = 1, 2, ...
for small YE1R n and Z in a neighborhood of [A, i].
Proof Since 8'1' /8z = 0 F /oz = ° we obtain by differentiation

8'1'° /oz = - I 81/1)8z0F /8x j EI.
Hence the lemma follows from Theorem 7.5.12.
We can take Z so small that the lemma is always applicable when

ZEZ. Now denote by S/w, y, z) the function obtained from (7.7.19)
when every derivative o~F and o~u is replaced by a function FOC(y, z) or
uP(y, z) in the same residue class mod I. When ZEZ o these are pre-
cisely the terms in the asymptotic expansion given by Theorem 7.7.6,
hence
(7.7.20) IS(w, y, Z)- L siw, y, z)1 ~ CN W- N - n/ 2 , ZEZ O '
j<N
It is clear that S(w, y, z) is an analytic function of z, and
(7.7.21)
by Lemma 7.7.8 and Lemma 7.7.10, for
IlmXl2ve-crollrnxl2 ~ Cvw- v.
Choose Z so that the boundary of Z+ is piecewise in C 1 and set
Sj(w,y,z)=(2ni)-1 S Siw,y,()d(/((-z), ZEZ+

iJZ+
Sj is then analytic in Z+, and Cauchy's integral formula (3.1.11) shows

that
ISj-Sjl~Cj,vw-V, v=1,2, ....
Thus (7.7.20) remains valid with Sj replaced by Sj. It is obvious that
we have a bound
IS(w,y,Z)- L Sj(w,y,z)I~C,
j<N
ZEZ+,
for ImFO(y, z)~O, ZEZ+. It follows that there is a constant y>O such
that
(7.7.22) Is(w,y)- L Sj(w,y,i)I~Cw-Y(N+n/2),
j<N
for we have the following elementary lemma:
Lemma 7.7.11. Let Z + be simply connected. Then there exists a positive
harmonic function y(z) in Z + such that if g is analytic and Igl ~ 1 in Z +
while Igl~e<l on Zo it follows that Ig(z)l~eY(Z), ZEZ+.
Proof We can choose a conformal mapping Z -+ w(z) of Z+ on the
upper half plane in <C mapping Zo to (-1,1). Set g(z)=G(w(z)). Then
it follows from the maximum principle that
log IG(w)I-(log e)(arg(w-a) -arg(w+a))/n ~O, 1m w>O,
if 0 < a < 1, for this is true at the boundary. Letting a -+ 1 we obtain
Ig(z)1 ~eY(z) where y(z) = (arg (w(z) -1) - arg (w(z) + l))/n.
It follows from (7.7.22) that

(7.7.22)' Is(w,y)- L Siw,y,i)I~Cw-(N+nj2).
j<N
In fact, this follows from (7.7.22) if we sum for all j < v with v so large
that y( v + n12) > N + n12, for it is clear that the terms with j ~ N can be
estimated by the right hand side of(7.7.22)'. Hence we have proved
Theorem 7.7.12. Let f(x, y) be a complex valued Coo function in a

neighborhood of (0,0) in lRHm, satisfying (7.7.15), and let uECg'(K)
where K is a small neighborhood of (0,0) in 1R.n+m. Then
where for functions G(x, y) the notation GO(y) stands for a function of
y only which is in the same residue class modulo the ideal generated by
ofloxj,j=l, ... ,n. When y=O the definition of the square root is as in
Section 3.4 and for small y =l= 0 it is determined by continuity.
So far we have not really proved that (7.7.23) is valid for every
choice of the representatives for the residue classes but only those
which were obtained in the proof. However, this is a consequence of
the following result.
Proposition 7.7.13. Let the hypotheses of Theorem 7.7.12 be fulfilled and

let f O, GO be in the same residue classes as f O, GO. Then we have
(7.7.24)
Proof Integration of eZ from z to w shows that

I~ - eWI ~ Iz - wi exp max (Re z, Re w).
Hence it follows from Lemmas 7.5.10 and 7.7.8 that for every N
IGOei())fo _Goei())lol~ CNw IImXIN e-c"'IImXI2~ C~Wl-NI2,
which proves the proposition.
The importance of the preceding results will be manifest in Chap-

ter XVIII and particularly in Chapter XXV which rests entirely on
them. (Theorem 7.7.12 will only be required in Section 25.4.) However,
we shall interrupt the flow of technical results to give a simple
application already here.
Theorem 7.7.14. Let u=adS be a Cg' density on a Coo hypersurface ~

in 1R.n of total curvature K =l= 0, dS denoting Euclidean surface measure.
Then
(7.7.25) It(n-l)/2u(t~)_ L a(x) IK(x)l-t(2n)(n-l)/2 e-i(x,t O -ltia/41 ~ Cit;

1~1=1, t>1.
Here the sum is taken over all XESUPPU where ~ is normal to E, and (J
denotes the number of centers of curvature at x in the direction ~ minus
the number of centers of curvature in the direction -~.
Proof. If n(x) is normal to E at x, then the Gauss map

E,3X ~ n(x)ESn- 1 is a local diffeomorphism since the total curvature
is not O. We can assume that u has so small support that E can be
parametrized in a neighborhood by parameters tE1R.n - 1 and the
Gauss map is a diffeomorphism. Writing a dS = b(t)dt we have
u(t~) = Se- it(x(t),O b(t)dt, I~I = 1.
The phase function f(t,~) = - <x(t), 0 has a critical point as a func-
tion of t if ~ is a normal to E at x(t), and then the Hessian 1;;' =
- <x" (t), 0 is the second fundamental form with respect to the normal
direction -~. Thus the eigenvalues with respect to the first fundamen-
tal form
L gjk dtjdtk = Idx(t)jZ
are the principal curvatures K l' ... , Kn _ l' so the signature is the
number of centers of curvature in the direction - ~ minus the number
in the direction ~, and
det(1;;')/det (gjk) = TI Ki=K.
Hence
Idet (tf" /2n)l- t =(2n/t)(n-l)/2IKI-t Idet (gjk)l- t.
Since b=aldet(gjk)l t the estimate (7.7.25) follows from (7.7.13) when
k>1 +n/2.
Theorem 7.7.14 is much more precise than Theorems 7.1.28 and
7.1.29 when it is applicable, but those results make no hypothesis on
the curvature.
Corollary 7.7.15. Let X e1R.n be a bounded open convex set with Coo
boundary of strictly positive curvature. If X is the characteristic func-
tion of X then
(7.7.26) 1)/2 Xm_(2n)(n-l)/2(K(x+)-tr i (x+,O+Xi(n+ 1)/4
11~I(n+
+ K(x_)-t e-i(x-,O-lti(n+ 1)/4)1 < c/I~I, I~I > 1,

where x+' x_ are the points on oX where the exterior normal is~, -~.
Proof. By (3.1.5) we have OjX = njdS where n is the interior normal on

X. Thus i~jXm is the Fourier transform of the density njdS for j = 1,
... , n. If we write down (7.7.25) for these densities, the estimate (7.7.26)
follows.
The following estimate for the number of lattice points in convex

sets in JRn is a classical application of Corollary 7.7.15.
Theorem 7.7.16. Let X have the properties in Corollary 7.7.15 and

assume that OEX. If N(t) is the number of points in lLnntX, t>O, then
(7.7.27) IN(t)-tnm(X)1 ~ crn- 2n/(n+ 1), t> 1.
Proof. We may assume n> 1. The characteristic function of t X is

X(x/t), so
N(t)= L X(g/t).
gEZ"
It is tempting to apply Poisson's summation formula here, but X is not

smooth enough. Let us therefore form a regularization
x.(x) = SX(x - ey)¢(y)dy
with O~ ¢EC~, S¢ dy = 1. If supp ¢ c (X n ( - X»/2 then
X(x(l + e» ~ X.(x) ~ X(x(l- e», 0 < e < 1.
This implies that
(7.7.28) N(t/(l +e» ~L X.(g/t) ~ N(t/(l- e».
Poisson's summation formula can now be applied and it gives
L x.(g/t) = rn L X.(2ngt) = rn L x(2ngt)$(2ng at).
The term with g=O is tnm(X), so it is the main term in (7.7.27). By
(7.7.26) we can estimate the sum for g=l=O by
crn L Igtl-(n+l)/2(1+lgetl)-N
g*O
where N may be taken arbitrary. We want e to be so small that et is
small too. Then the second factor plays no role when we sum for
Igl < l/et, and when Igl> l/et we replace it by Igetl- N• Comparing the
two sums obtained with the corresponding integrals we can estimate the
sum by
This allows us to rewrite (7.7.28) in the form

rn(l-e)nm(X)- Ce-(n-l)/2 <N(t) <rn(l +e)nm(X) + Ce-(n-l)/2.
To minimize the difference between the estimates we choose e now so
that rne=c(n-l)/2, that is, e=t- 2n/(n+l), which gives (7.7.27).
So far we have only studied the asymptotic behavior of the in-

tegral
JueiOJJ dx
extended over the whole space. We shall now discuss what happens if
we only integrate over an open set X with COO boundary oX. The
results will not be used in this book so we shall be quite brief. Since
they are only local we can make a local change of variables such that
X becomes the half space in JRn defined by xn<O. We write x'
=(x 1, ... , x n_ 1)·
Theorem 7.7.17. Let fEcoo(JRn) and Imf~O when xn<O, and let
UECg'(1Rn) have support close to 0. Then
(i) if of(O)/ox' ,*0 we have
J ue iwJ dx = O(w- N ), w --+ + ro, N = 1, 2, ....
Xn<O
(ii) if of(O)/ox'=O, det 02f(0)/ox'ox' ,*0 and of(O)/ox,*O, we have

J L ajw- j
00
u(x)tfOJJ(X)dx ~tfwJ(O)w-(n+ 1)/2
Xn<O °
where ao = (det (f::x' (0)/2n z))- t(iof(0)/OX n)-1 u(O).
(iii) if of(O)/ox=O but det0 2f(0)/ox' ox',*O and det0 2f(0)/oxox,*0
we have
J u(x)eiwJ(X)dx ~eiwJ(0)w-n/2 L bi w- il2
00
Xn<O 0
where bo = (det (f:~(0)/2ni))- t u(0)/2.
Proof. (i) follows at once from Theorem 7.7.1 for the integral with
respect to x' is rapidly decreasing uniformly with respect to x n • In
cases (ii) and (iii) we have by Theorem 7.7.12 an asymptotic expansion
I(xn' w) = Ju(x)eiOJJ(X)dx' ~ L w- (n-1)/2-iuixJeiOJJO(Xn).
°
Here ImfO(xn)~O for small xn~O, and UiECOO({X n; xn~O}) has sup-
port close to and uo(O)=u(O) (det (gx' (0)/2 n i))-t. Since
f(x)-r(x n)EI(oflox 1, ... , oflOXn_ 1 )2,
we have ofloxn-dfo/dxnEI, so fO'(O) = of(O)/ox n,*0 in case (ii). Hence
J(w)= J° UO(xn)eiOJJO(Xn) dXn

-00
=uo(O)eiOJ[O(O)/iwfO,(O)+ J° i/w(uo/fo,)' eiOJJOdxn·

-00
The partial integration can be continued and the other terms in the
expansion of l(xn, w) can be discussed in the same way which proves
the statement (ii). In case (iii) we note that if Lj=dol/oxj at 0 then
n-l
<f"(0) x, x) = jO" (0) x; + L ajkLix)Lk(X),
j,k= 1
Hence jO" (0) =1= 0, Lj(x',O) are linearly independent and

detl:~(O)=(det(oLj/oxJj,k~ 1)2 jO"(O) det ajk
= jO"(O) detl:: x ' (0).
It is clear that if c= jO"(O) then 1m c~O.
If veCg'(JR) an asymptotic expansion of
o
S eiOW2j2v(t)dt
-00
is obtained by writing V=VO+V1 where vo(t) = (v(t)+ v( -t))/2 is even

and v 1(t)=(v(t)-v(-t»/2 is odd. From the Taylor expansion it is easy
to see that v 1(t)=tW(t 2) where we Coo, w(V)(0)/V!=V(2v+1)(0)/(2v+1)!.
Hence
° 00 00
S ei())ct 2j2 v(t)dt = S ei())ct2j2vo(t)dt/2 - S ei ())CS w(2s)ds
-00 -00 0
00
-!(cw/27tI}-t L (2iw c)-v D2v v(O)/v!

°
L (iwc/2)-V D2v+
00
+(WC)-1 1 v(0)v!/(2v+ 1)!

°
The proof of Theorem 7.7.5 shows that we have a similar expansion if
the phase function is replaced by 1°, and this completes the proof in
case (iii).
When I is real one can study the effect of parameter dependence

in the preceding situation by means of Theorem 7.5.13 with k=2. This
leads to Fresnel integrals. However, we leave this for the reader since
we shall discuss a similar but more complicated problem below.
Finally we shall discuss briefly the asymptotic behavior of
Su(x) ei())j(X) dx
when I is real valued but has a degenerate critical point Xo' From
now on integration will always be taken over the whole space so there
will be no boundary problems to investigate. The simplest case is
when the rank of I" is n -1 at x o ' We can choose coordinates so that
f:: x'
is non-degenerate, if x' = (x l ' ... , Xn _ 1)' An asymptotic expansion
of the integral with respect to x' is obtained from Theorem 7.7.6. We
are then left with an integral with respect to x n • After changing
notation our problem is therefore to study the asymptotic behavior of
J u(t)eirof(t) dt
00
-00
when UE Cg' has support close to 0 and l' (0) = 1" (0) = O.
If l' has a zero of finite order k - 1 at 0 we can write
f(t) = f(O) + ta(t), a(O) =1=0.
Introducing tla(t)11/k as a new variable and changing notation again,
we are led to studying the integral
Ju(t)eirotk dt
or its complex conjugate. First assume that k = 3. Since eirotl is the
Fourier transform of Ai(-r(3w)-t)(3w)-t (see Section 7.6) we have
Ju(t)eirotl dt= Ju(r)Ai(r(3w)-t)(3w)-tdr.
If we introduce a finite part of the Taylor series expansion of A i, it
follows when UE Cg' that
(7.7.29) Ju(t)eicotl dt
= L AiUl (0)/j!(3w)-U+ 1)/32nDi u(0)+O(w-(N+1)/3).
i<N
The derivatives Ai(j)(O) can be expressed in terms of the r function.
Also for arbitrary k the Fourier transform of eitk is easily seen to be
an entire function if we write it as an integral over a contour where
it is negative. With UECg' we obtain for odd k
(7.7.30) Ju(t)eicotk dt = L 2k- 1r(U + 1)/k) (Sin (k -1)U + 1)n) w- U+ l)/k

i<N 2k
x (- D)iu(O)/j! + O(W-(N+ l)/k), W -+ 00,
and for even k
(7.7.31) Ju(t)eirotk dt = L 2k- 1r«2j + 1)/k)e"i(2i+ 1)/2k W-(2i+ l)/k
i<N
X (iD) 2i U(0)/(2j)! + O(w- (2N+ l)/k), w -+ 00.
The simple details of the proof are left as an exercise for the reader.
Note that in (7.7.30) the terms for which k divides j + 1 are missing.
All this is quite elementary, but the situation becomes rather
complicated if there are parameters y present, for the multiplicity of
the zeros of !,'(t, y) as a function of t may vary with y. We shall only
discuss the case k = 3 where the Airy function turns up and leave the
extension to general k for the reader.
Theorem 7.7.1S. Let f be a real valued COO function near 0 in 1R 1+n

such that af/at=a 2 f/at 2 =0 but a 3 flat 3 *0 at O. Then there exist Coo
real valued functions a(y), b(y) near 0 such that a (0) =0, b(O) = f(O) and
Ju(t, y)eirof(t,Y)dt,...., eirob(Y)(Ai(a(y)wt)w-t L uov(Y)w-

00
(7.7.32) V
o
+ Ai'(a(y)wf)w-t L u lv (y)w- V ),
00
o
provided that UECg' and suppu is sufficiently close to O. Here UjvECg'.
Proof Replacing t by - t if necessary we may assume that

a3f(0)lat 3>0. Application of Theorem 7.5.13 with k=3 gives a func-
tion T(t, y) with T(O, 0)=0, aT(t, O)lat>o such that
f(t, y) = T3/3 + a(y) T + b(y),
where a, bECoo. Introducing T as a new integration variable instead of
t we obtain when supp u is sufficiently close to 0
J(w, y) = Ju(t, y)eirof(t'Y)dt = Jv(T, y)e iro (T3/ 3+a(Y) T +b(y)) dT

where VECg' and suppv is close to O. Choose XECg'(1R) equal to 1 in
a neighborhood of 0 and with small support. Then we have
J(w, y) = JX(T)v(T, y) eiro (T3/ 3+a (y)T +b(Y))dT.
Here we divide v by T2 + a(y) using Theorem 7.5.6,
v(T, y)=(T 2 +a(y))q(T, y)+ r1(y)T +ro(Y).
After an integration by parts we then obtain
J(w, y) = ilw Jd~ (X(T)q(T, y))eiw (T3/ 3+a(y)T+b(Y))dT
+ Jx(T)(r1(y) T+ ro(y))e iro (T3/3+ a (Y)T+b(Y))dT.

The first integral on the right hand side is of the same form as the one
we started with apart from the factor 1/w in front. We have
J J
eiW (T3/ 3+a (y)T)X(T)dT= ei(T3/3+a(Y)wtT)x(Tw-t)dTlwt
=2nA i(a(y)wt)w- t - Jei(T3/ 3 +a(y)rot T)(1_ X(Tw-t))dTlw t .
In the last integral the integration outside the support of X(Tw- t )
should be moved into the upper half plane, and it is easily seen by
means of Theorem 7.7.1 that it is rapidly decreasing as w -. + 00.
Similarly
Jeiw (T3/3+a(y)T) X(T) TdT -(2:n:/z) A i'(a(y) wi") w- i
is rapidly decreasing. Iteration of this argument proves the theorem.
It is easy to extend Theorem 7.7.18 to several integration vari-

ables:
Theorem 7.7.19. Let f(x, y) be a real valued Coo function near 0 in

1R.n + m such that
(7.7.33) f: (0,0) = 0, rankf:~(O, 0) = n -1,
<X, %x)3f(0, 0)=1=0 if O=l=XEKerf:~(O,O).
Then there exist real valued COO functions a (y), b(y) near 0 such that
a (0) = 0, b(O) = /(0) and
(7.7.32)' Ju(x, y)eirof(X'Y)dx '" eiWb (Y)w-(n-l)/2(A i(a(y)w1-)w-t
x f uo.(Y)w-· + Ai'(a(y)w1-)w- f U1V(y)W-
o
t
0
v),
provided that UECO' and suppu is sufficiently close to O. Here UjvECO'.
Proof. Let us first verify the invariance of the last condition in (7.7.33).
To do so we let x ~ x(s) be any smooth curve with x (0) = 0 and x' (0)
=X. When s=O we have
d 3f(x(s), 0)/ds 3 = <X, O/OX)3 f(O, 0)+ 3<f:~(0, 0) X, xl/(O).
The second term vanishes since X E Ker f:~(O, 0). Thus the last part of
(7.7.33) means that
(7.7.33),
and this condition is independent of the choice of local coordinates.
Now we may assume the coordinates labelled so that
det(02f10XiOX)~,j~ 2 =1=0 at (0,0).
Then the equations of (x, y)/ox j =0,j=2, ... , n, determine Xj as Coo
functions X j(X l ' y), j = 2, ... , n. If we introduce x j - X j as new variables
instead of Xj' j = 2, ... , n, we have reduced the proof to the case where
of(x,y)/ox j =0,j=2, ... ,n, when x 2= ... =xn=0. By Theorem 7.7.6 the
integral (7.7.32)' with respect to the variables x 2, ... , xn only is of the
form
where U has an asymptotic expansion in powers of l/w as w --+ 00.

lf we apply Theorem 7.7.18 to each term in the expansion, we obtain
(7.7.32)'. (Note that we can take X=(I,O, ... ,O) because
iPf(O, O)/aXl OXj=O for allj.)
In view of (7.6.20) the right-hand side of (7.7.32)' is rapidly decreas-

ing when a(y) > O. On the other hand, it follows from (7.6.21) that
(7.7.32)' agrees with the expansion given by Theorem 7.7.6 (with two
critical points) when a(y) <0. When a(y)=O we have an asymptotic
expansion in powers of w- t , and this remains true in any zone where
a(y)w~ is bounded. However, the amplitude quickly becomes very
small when this quantity grows positive. In Section 12.2 we shall see
that these expansions represent in optics the behavior of a wave in
respectively the shadow region, the illuminated region and the pen-
umbra at a caustic. The role of the Airy function is to describe the
transition between the different types of asymptotic expansions.
7.8. Oscillatory Integrals
Theorem 7.7.1 allows one to define the useful notion of oscillatory

integrals. Let X cJRft be open and let r be an open cone in
X x (JRN,- {On for some N. This means that r is invariant under
multiplication by positive scalars of the component in JRN. We shall say
that a function fjJECOO(r) is a phase function in r if
(i) fjJ(x, te)=tfjJ(x, e) if (x, e)Er, t>O.
(ii) ImfjJ~O in r,
(iii) dfjJ*O in r.
We wish to show that an integral of the form
(7.8.1) f d4>(X, 8) a(x, e) de
defines a distribution in X even if a is large, provided that a oscillates

more slowly than the factor ei 4>.
Definition 7.S.1. Let m, p, (j be real numbers with O<p ~ 1 and

0~(j<1. Then we denote by S;,iX xJR N) the set of all aECOO(X xJR N)
such that for every compact set K c X and all a:, {3 the estimate
(7.8.2) ID~D~a(x, e)1 ~ Ca,fl, K(1 + lel)m- pial + dllll,
xEK, eEJR N,
is valid for some constant Ca,fl,K'
The elements of S;, ~ are called symbols of order m and type p, fJ.
The best possible constants in (7.8.2) are semi-norms in S;,~ which
make it a Frechet space.
Theorem 7.S.2. Let ¢ be a phase function in the open cone reX x 1R. N
and let F be a closed cone cru(X x {O}). Then the functional I",
defined by
(7.8.3) I",(au) = Jei",(X,8)a(x, 8) u(x)dxd8
when the integral is absolutely convergent can be extended in a unique
way to all aE U S;jX x 1R. N) with support in F and all u ECg' (X), so
m,p,~
that I",(au) is a continuous linear function of aES;,~ for every fixed

UECg'(X), mE1R., pE(O, 1] and fJE[O, 1). The linear form
I""a: u --+ I",(au)
is a distribution of order ~ k if aE S;. ~ and
m-kp < -N, m-k(1-fJ)<-N
Proof Choose XECg'(1R.N) so that X(8)=1 when 181<1 and X(8)=0
when 181> 2, and set (see the proof of Theorem 7.5.4)
Xv(8)=X(2- 8)-X(2 1 - 8),
V V v>O; Xo(8)=X(8).
Then we have
00
IXv=1 and 2v-l~181~2Hl when 8ESUPPX v' v =1=0.

o
If aES;,d and XEK~X we obtain
ID~D8Xv(8) a(x, 8)1 ~ Ca,p,K(1 + 181t- plal+dlPI
since IDaXv(8)1~ Ca(1 +1(1)-la l with a constant independent of v. Hence
the series L Xv a converges to a in S;:
d if m' > m, for at most two terms
in the series have overlapping supports. If there is an extension of I",
with the required properties, it follows that it must be given by
(7.8.4)
The theorem will therefore be proved if we show that the series on the
right-hand side converges and that the sum has the properties listed in
the theorem.
To do so we write for v~O
I",(Xv+ 1 au) = J~"'(X,8)Xl (2- v 8) a(x, 8) u(x) dxd8

=2Nv JeiW </>(X,8)Xl(8)a(x,w8) u(x) dxd8
where w=2v. By hypothesis y=max(D, I-p) < 1, and if

ID~D6a(x, 8)1 ~M(1 + 18I)m-plal+bIPI, 11l+ PI ~k, xEK,
we obtain
ID~8a(x,2V8)I~CM2v(m+Ylan if xEK, 1/2<181<2, 11l1~k.
Hence Theorem 7.7.1 gives the estimate
II</>(Xv+ 1 au)1 ~ CM2 v(N+m+ yk-k) I suplDaul, UEC~(K).
lal ~k
When (1- y) k> N + m we conclude that (7.8.4) converges and that
u~I</>(au) is a distribution of order ~k as stated.
It is convenient to use the notation (7.8.3) for I</>(au) even when the
integral is not convergent. The extended definition of (7.8.3) will be
called an oscillatory integral. For the distribution u~I</>(au) the no-
tation Sei</>(x. 8) a(x, 8) d8
will often be used.
An important example of an oscillatory integral is the formula
(7.8.5) S ei(x.8)d8=D o(x)(2nt·
JR."
By definition we must for the proof consider the oscillatory integral

Hei(x.8)u(x)dxd8, uEC~(lR").
If XEC~ is equal to 1 in a neighborhood of 0 we have X(./t)~1 in

S7.0 for any m>O as t~ 00. Thus the oscillatory integral is the limit
of the convergent double integral
Hei(x.8) X(8/t) u(x) dxd8= St" x( - tx) u(x) dx
x(
= S -x) u(x/t)dx ~u(O) S -x)dx x(
= u(O) (2 n)" X(O) = u(O)(2n)".
This proves (7.8.5) which is thus another way of expressing Fourier's
inversion formula. That the left-hand side of (7.8.5) is a distribution
which is singular only at 0 is also a consequence of the following
theorem.
Theorem 7.8.3. For the distribution I</>.a defined by (7.8.3) we have
singsupp I</>.ac {XEX; cp~(x,8)=O for some (x,8)EF}=S.
The restriction of I</>. a to X . . . . S is the COO function
(7.8.6) x~ Sei </> (x. 8) a(x, 8)d8
which is defined for fixed x as an oscillatory integral.
Proof The definition of S means that </J(x,8) is a phase function of 8

when x is fixed in X . . . . S, so the oscillatory integral in (7.8.6) is defined
then. It is a continuous function of x, for the proof of the existence of
the oscillatory integral shows that it is the limit, locally uniformly
with respect to x, of the COO function
Seit/>(X,6)a(x, 8) X(8/t) de.
The derivative with respect to x is
Seit/>(x, 6)(i</J~(x, e) a(x, e) + a~(x, 8)) x(8/t) de
which converges to the oscillatory integral obtained by differentiating
(7.8.6) under the integral sign. Thus the function (7.8.6) is in C 1 (X . . . . S)
and the derivative may be computed by formal differentiation under
the integral sign. Since </J~(x, e) a(x, e)(1-X(e))ES;,~ 1 we may repeat
the procedure and conclude that (7.8.6) defines a function in
COO(X . . . . S). This function is equal to the distribution u-+It/>(au) there,
for ifuEC;;'(X . . . . S) we have
It/>(au) = limS u(x)dx(S ei t/>(X,6)a(x, 8) x(e/t) de)
t~oo
= Su(x)dx(S ei t/>(X,6)a(x, e)de).

The arguments used in the preceding proof show quite generally
that one may operate on oscillatory integrals as with standard in-
tegrals; differentiation can be performed under the integral sign, or-
ders of integration can be interchanged and so on. We leave for the
reader to contemplate this extension of integral calculus and give
instead an example.
Example 7.8.4. The Cauchy problem

(7.8.7) c- 2 iJ 2 E/iJt 2 -AE=0 in 1R 1 + n,
E=O, iJE/iJt=()o when t=O,
has the solution
E(t, x) =(2n)-n S(ei(ctl~1 + (x, 0) - ei(-ctl~l+ (x, ~») d~/2il~lc.
The integral is absolutely convergent when I~I < 1, and when I~I > 1 it
is defined as the difference between two oscillatory integrals with
phase functions <x,
0 ±ctl~l. The differential with respect to x is
never 0, so E(t, x) is a COO function of t with values in ~'(1Rn). We
obtain (7.8.7) by differentiating under the integral sign if we recall
(7.8.5). By Theorem 7.8.3 we have
sing supp E c {(t, x); x ± ct~/I~I = 0 for some ~ =l= O}
={(t,x); Ixl=cltl}.
This is the double light cone. We leave as an exercise for the reader to
verify that 2c 2 E=E+ -E_ where E+ and E_ are the advanced (re-
tarded) fundamental solutions in Theorem 6.2.3.
7.9. H(s)' IJ' and Holder Spaces
In Theorem 7.1.11 we proved that the Fourier transformation maps L2

onto itself. By Theorem 7.1.13 we also know that If is mapped into
If' if 1 ~ p < 2. However, the range is much smaller than If' then for
Theorem 7.6.6 shows that the Fourier transform of If' contains distri-
butions of positive order when p' > 2. We shall now round off these
results by studying some spaces of distributions which are closely
related to L2 and therefore possible to keep track of when using the
Fourier transformation.
Definition 7.9.1. If s is a real number, then H(S)(lRn) denotes the space

of all UE[J>' such that uEL~, the L2 space with respect to the measure
(1 + leI 2 )'de/(2nt. In H(s) we define the norm
(7.9.1) Ilull(s) =«2n)-n S lu(eW(1 + leI 2)Sde)t.
The Fourier transformation is an isomorphism H(S) --+ L~ since
L~c[J>'.We shall now give a description of H(s) which does not refer
to U. By Theorem 7.1.11 we have H(o)=L2 , and it is obvious that H(s)
decreases when s increases. Let s be a positive integer for a moment.
Expanding (1+leI 2 )s=(1+ei+ ... +e;Y we then obtain
(7.9.2) Ilullt.)= L ca iiDa ulli2
lal ~s
where ca are polynomial coefficients. Thus H(S) is the space of all uEI!
such that D a UEL2 when lad ~s. (If L2 is replaced by If here we obtain
the general Sobolev spaces usually denoted by UI,/.) Next let O<s< 1.
Then
(7.9.3) Ilullt.) ~(2n)-n Slu(e)1 2 (1 + lel 2S) de ~ 2I1 u llt.),
and we shall also prove that
(7.9.4) (2n)-n S lu(e)1 2 (1 + lel 2s) de
= S lul 2 dx+AsSS lu(x)-u(YWlx- yl-n- 2s dxdy.
This implies that H(s) consists of all L2 functions for which the right-
hand side of (7.9.4) is finite. In view of (7.9.1) the estimates (7.9.3) are
7.9. H(,), IJ and Holder Spaces 241
equivalent to the elementary inequalities

(1 +1~12)S~1 +1~12S~2(1 +1"2),, O<s<1.
(The second one is trivial and the first follows if we divide by (1 + 1~12)s
and note that aS~a if O~a~l and O<s<1.) To prove (7.9.4) we use
Parseval's formula and obtain
Sf lu(x)-u(YWlx- yl-n- 2S dxdy= Hiu(x+ y) "':'u(yWlxl- n- 2s dxdy
=(2n)-n Hiei(X'~> - Wlxl-n-2sla(~Wdxd~
for the Fourier transform of y-->u(x+y)-u(y) is a(~)(ei(x,O_l). Now

A; 1 = 1~1-2s Siei(X,O -11 2 Ixl- n- 2sdx
is independent of ~, for it is obviously a function of I~I only and if we

replace x by tx we find that the value at ~ is equal to the value at t~.
This proves (7.9.3) and (7.9.4). (It is easy to see that AsIs(1- s) has
finite limits as s --> 0 or s --> 1.) If 0 < s < 1 and k is a positive integer it
follows now as in our discussion of H(k) above that H(Hk) consists of
all uEL2 such that Da UE L2 when loci ~k and
(7.9.5)
The norm Ilull(Hk) in H(Hk) is equivalent to
L IIDaull p + L (HiDa u(x)-Da u(y)llx-yl-n- 2s dxdy)t.
lal;:>k lal=k
Having described the H(s) spaces with s~O we observe that !/ is

dense in L;, hence in H(s)' for every s. If UE!/' it follows that
is finite if and only if uEH(s) and then it is equal to lIull(s)' Thus H(s) is
the dual of H(_s) which gives a description also when s<O.
The space CY of Holder continuous functions in 1Rn of order
YE(O, 1) was defined in Theorem 4.5.12 to be the space of continuous
functions such that for every compact set K
sup lu(x) - u(y)l/lx - ylY < 00.
x,yeK
We shall say that UEC k + Y, where k is a positive integer and 0<1'< 1, if

UEC k and for every compact set K
sup IDIXu(x) - DIXu(y)l/lx - ylY < 00, loci = k.
x,yeK
It is clear that the subspace c~+ Y of elements in Ck+ Y with compact

support satisfies (7.9.5) if s < y. Thus
(7.9.6) Ck+YcH
o (s)
if O<s<k+y.
If uEH(_s) and O<s<k+y we obtain
I(u, 4»1 ~ Ilull(_s)II4>II(s)~ c L supID"4>(x) -D"4>(y)I/lx - ylY,

l"l=k
when 4>EC'~(K) where K is a compact set in ]Rn. From the obvious
inclusion C~ c H (k) we also conclude that
(7.9.7) H(_S)C~/k if s~k and k is an integer.
Next we shall relate the Fourier transform I? of H(S) to IJ'spaces.
Lemma 7.9.2. L~cH if and only if q=2 and s~O or 1~q<2 and
s> n(l/q -1/2).
Proof The case q = 2 is obvious. It is also clear that L~ c H implies

q ~ 2, for all functions in L~ are not in L10c when q > 2. When 1 ~ q < 2
we obtain by Holder's inequality
SIvlqd~ = S(lvI2)q/2 d~ ~(S Iv1 2 (1 + 1~12)Sd~)q/2(J (1 + 1~12)-qs/(2-q)d~)1-q/2
~ C(SlvI2(1 + 1~12)Sd~~/2
if 2qs/(2-q»n, that is, s>n(1/q-1/2). If s=n(1/q-1/2) we can take
v=(1+1~12)-n/2q(log(2+1~I))-a and obtain VEH if and only if qa>l,
and vEL~ if and only if 2a > 1. When we take aE(1/2, l/q) we find that
L~ is not contained in H.
Theorem 7.9.3. The Fourier transform of H(s) is contained in H if

1~q<2 and s>n(1/q-l/2). The Fourier transform of IJ' is contained in
H(_S) if 2<p~ 00 and s>n(1/2-1/p).
Proof The first statement follows immediately from Lemma 7.9.2. To

prove the second one let UEIJ' and note that when 4>Eff'
I(u, 4»1 =I(u, 4»1 ~ Ilulbll4>ll v ' ~ C114>II(s)
by the first part of the proof, so uEH(_s)'
If we combine the second part of Theorem 7.9.3 with (7.9.7)

we conclude that for p>2 the Fourier transform of 11' is in ~/j if
j> n(1/2 -l/p). (Recall that by Theorem 7.6.6 this would be false if
j < n(1/2-1/p).) We have actually proved a great deal more for there
is a considerable margin in the inclusion (7.9.7).
7.9. H(,), IJ and HOlder Spaces 243
A particularly important special case of Theorem 7.9.3 is the

following Bernstein theorem:
Corollary 7.9.4. The Fourier transform of H(S) is contained in L1 if

s>nI2, and H(S) is then contained in the space of continuous functions
on 1R." tending to 0 at 00.
Corollary 7.9.4 is of course a slightly stronger version of Lemma

7.6.3. In estimates such as (7.6.10) we could therefore have used H(S)
norms for any s > nl2 instead of the smallest integer s > nl2 as we
actually did.
Our discussion so far shows that one cannot express the IJ norm
of u very well in terms of the Fourier transform u. To prove con-
tinuity of maps in IJ spaces one can therefore seldom use Fourier
transforms except in L2. However, we shall now prove some rather
precise estimates supplementing those in Section 4.5 by combining the
methods used there with the Fourier transformation in L2.
Theorem 7.9.5. Let kE9"(1R.") and assume that kELloe'

(7.9.8) L J IRlaIDak(~Wd~/R"~C<oo, R>O,
lal ~s RI2 < I~I < 2R
where s is an integer > n12. Then it follows that for 1 < p < 00
(7.9.9)
In addition
(7.9.10)
Proof Choose a function ljJEC~({~; 1~1~2}) which is equal to 1 when

I~I ~ 1. Then we have for ~ =1= 0
L (ljJ(2-j~)-ljJ(21-j~)),
00
(7.9.11) 1=
-00
which we shall use to decompose k. If we set

kR(~) =(ljJ(~) -ljJ(2~)) k(R~)
it follows from (7.9.8) that (with another C)
(7.9.12) L JIDakR(~Wd~~C.
lal ~s
Hence sup IkRI ~ C' by Lemma 7.6.3, so Ik(R ~)I ~ C' when I~I = 1, which
means that
(7.9.13)
when ~ =FO. Since we have assumed that kEL~oc it follows that kEL'XJ.
Parseval's formula now gives (7.9.9) when p = 2 with C 2 = C'. More-
over, Corollary 7.9.4 shows that kR is the Fourier transform of a
function kREV with IlkRIILl ~ C" and kRECoo. More precisely we have
Slk R(xW(1 + IxI2)Sdx~ C3
so Cauchy-Schwarz' inequality gives
(7.9.14) S IkR(X)ldx~ C 4 t"/2-s.
Ixl>t
Bounds of the same form are valid for ~ikR' hence for DiR' so we
have
(7.9.15)
which implies
(7.9.16)
We are now ready to prove the analogue of (4.5.16),
(7.9.17) Slhwldx~CSlwldx ifwECg'(I)andSwdx=O.
P*
Here I is a cube and 1* the "doubled cube" as in Lemma 4.5.6. We
may assume in the proof that the center is at 0 and that the norm in
lRn is the maximum norm so that I is defined by Ixl <t and 1* by
Ixl<2t. Since the Fourier transform of kR(Rx)R n is kR(~/R) it follows
from (7.9.11) that
00
k= L k2i(2i.)2ni
-00
with g' convergence. Now (7.9.14) and (7.9.16) give since supp weI
S RnlkR(·R)*wldx~Slwldx S IkR(Rx)1 d(Rx) ~ CSlwl dx(tRt/ 2- s,

xO* IxJ>t
J RnlkR(·R)* wi dx ~HI(kR«X - y)R)-kR(xR») w(y)IRndxdy
x¢I*
~ C SlwldxtR.
Hence the triangle inequality gives

S Ihwldx~CSlwldx( L (2i t)n/2-s+ L 2it)~C"Slwldx
x¢I* 2it?,1 . 2it<1
We shall now prove (7.9.10). In doing so we decompose u accord-
ing to Lemma 4.5.5 with s replaced by r (since s has a different
meaning now). All terms are in L2. Since (7.9.9) is already proved
7.9. H(S)' IJ and HOlder Spaces 245
when p = 2 we have
-r 2 m{x; Ihv(x)I>-r/2} ~4I1hvlli2~ CIIvlli2~ C'-rllvIl LI.
If 0 = UIt then -rm(O)~2"lIullL' by (4.5.13), and (7.9.17) gives
m{x; x¢O'L Ihwix)1 >-r/2}rr/2 ~ S Llhw) dx ~ C JLlwjl dx ~ 3 quilL"
CO
Since Ik*u(x)I~-r unless Ik*v(x)I>-r/2 or XEO or x¢O and Llhwix)1

>-r/2, we have proved the weak type estimate (7.9.10).
It suffices to prove (7.9.9) when UEC,/:,. If (7.9.9) is known for some
p then it follows for the conjugate exponent p', l/p + l/p' = 1. In fact
Ihu*v(O)1 =lhv*u(O)1 ~ IIhvllLPliull v ' ~ CIIvllvllullv'
when u, VEC,/:" This implies k*UElf' and that (7.9.9) is valid with p
replaced by p'. Thus we may assume 1 <p<2. The proof is then a case
of the Marcinkiewicz interpolation theorem like that of Theorem
4.5.3. For -r>0 we write u=u.+U, where u,=u when lul<-r and l!..=u
when lui ~ -r. Both terms are then in L2 and
m{x; Ihu(x)1 >-r} ~m{x; Ihu,(x)I>-r/2} +m{x; Ih l!..(x)I>-r/2}
~ C(-r- 2 I1u,lIi2 +c 111 U,IIL')'
by (7.9.9) with p=2 and by (7.9.10). Hence
00
IIhullf.,=p J -rP-Im{x; Ihu(x)1 >-r} d-r

o
~C( Sf lu(xW-rP - 3 dxd-r+ Sf lu(x)I-rP - 2 dxd-r)
lu(x)1 <, lu(x)1 ~.
= C«2-p)-I +(p_1)-I) Jlu(x)IPdx
A partly parallel but much more elementary argument gives es-

timates in Holder spaces also. Set for 0 < y < 1
IUly = sup lu(x) - u(Y)I/lx - ylY.
x*y
Theorem 7.9.6. If k satisfies the hypotheses of Theorem 7.9.5 and

O<y<l, then
(7.9.18)
Proof. By using the partition of unity (7.9.11) we can decompose u,

U= LU j , Uj(~)=~(~/2j)u(e).
Here we have written $(~)=t/I(~)-t/l(2~), so cjJEY. Explicitly
uix) = Ju(x - 2- j y) cjJ(y)dy= J(u(x - 2- j y) -u(x» cjJ(y)dy

since
JcjJ(y)dy=$(O)=O.
Hence
(7.9.19)
and since
we also have
(7.9.20)
Choose XECO'(Rn,o) equal to 1 in supp$. Then
k(~) ui~) = k(~) X(~j2j) ui~)·

The estimate (7.9.12) is valid for k(R~)x(~), uniformly in R, so this is
the Fourier transform of a function kR with uniformly bounded L1
norm. With R=2j we have
huj=RnkR(R.)*u j .
Hence it follows from (7.9.19), (7.9.20) that
(7.9.19)' Ik*u) ~ C'2- yj lul y,
(7.9.20)' Ihujl ~ C'2(1- Yli luly'
Combining (7.9.20)' with the mean value theorem we obtain
Ihu(x) -hu(y)1 ~I Ihuix) -huiy)1

~ C'l u ly(2 I 2- yj + Ix - yl I 2(1- y)j)
2 -i< Ix-yl 2 -i~ Ix-YI
~ C"IUlylx - ylY j(y(1- y».

The proof is complete.
The following is a typical application of Theorems 7.9.5 and 7.9.6

in the theory of partial differential equations.
Theorem 7.9.7. Let P(D) be an elliptic differential operator in Rn of

order m. If X eRn is an open set and UE~'(X), then P(D)uED;oc(X)
implies D"uED;oc(X) when 11X1=m, if 1<p<oo, and P(D)UEcY(X) im-
plies D"UEcY(X) when 11X1=m if O<y<l.
7.9. H(s» 11' and HOlder Spaces 247
Proof Let E be the parametrix constructed in Theorem 7.1.22. Since

BEe''''' and (7.1.21) gives
1~1l D"B(~)I ~ C"IlI~IIIlI-m-I"1
the hypotheses of Theorems 7.9.5 and 7.9.6 are fulfilled by Dil E when
IPI = m. Let Y be a relatively compact open subset of X and choose
XECg'(X) equal to 1 in Y. Then
P(D)(xu) = f1 + f2
where f1 = XP(D) UEIY (resp. CY) and f2 = 0 in Y. Thus
xu+w*(xu)=E*f1+E*f2
so D"u-(D"E)*f1 is a Coo function in Y. Since D"E*f1EIY (resp. 0)
when locl=m by Theorems 7.9.5 and 7.9.6, the proof is complete.
The conclusion in Theorem 7.9.7 is of course also valid when

locl<m but combination of Theorem 7.9.7 and Theorem 4.5.13 gives a
better result then. The following theorem shows that it is not possible
to extend Theorem 7.9.7 to the excluded limiting cases. The result
justifies a remark made in the introduction concerning the flaws in
the classical notion of solution of a differential equation.
Theorem 7.9.8. Let X be an open set in JR. n, n> 1, and let P(D) be a
differential operator with constant coefficients of order m > O. Then one
can find UEc:;-l(X) . . . . C~(X) so that P(D) UE cg(X).
Proof We may assume that OEX. Let K={x; Ixl~R}cX be a com-

pact ball with center at O. The set of all UEC~-l(K) with
P(D)UEC8(K) is a Banach space with the norm
supIP(D)ul+ L supID"ul.
l"l<m
If all such functions are in C~ it follows from Banach's theorem that
L suplD"ul ~ C(sup IP(D) ul+ L supID"ul), uEC~(K).
l"l~m l"l<m
If Pm is the principal symbol of P we conclude that
(7.9.21) L supID"ul~C'(suplPm(D)ul+ L supID"ul), uEC~(K).

lal~m l"l<m
Let U E COO(JR. n) be a solution of the equation Pm(D) U = 0, choose

XECg'(K) so that X(x) = 1 when Ixl<R/2 and set
uix) =2- mi (XU)(2 i x).
We have
supID"'u) ~ C2(!",!-m)j, loci ~m,
and R/2~2j Ixl ~R in suppPm(D) uj so these supports are disjoint. If we
N
apply (7.9.21) to u = L uj it follows that
o
N L ID"'U(O)I~ C.
!"'!~m
When N --+ 00 we conclude that D'" U (0) = 0, loci = m, for every U E COO
satisfying the equation Pm(D) U = O. Taking U(x) = ei(x. 0 we find that
Pm(O = 0 implies (= O. This is not possible when n> 1, which proves
the theorem.
Remark. If Lemma 7.3.7 is used at the end of the proof one can
conclude that for a given oc with loci =m there is a function UEC,;;-l
with P(D) UE cg and D"'u¢; cg unless Pm(D) is a multiple of D"'. By a
simple category argument it follows that u can be chosen so that
D"'u¢; cg for every oc with loci = m such that Pm(D) is not a multiple of
D"'.
Notes
The basic facts on the Fourier transformation in Section 7.1 go back

to Fourier in a more or less precise form. However, the idea of
Schwartz to start from the dense function space [/' meant a great
simplification also of the classical foundations of the subject. The dual
definition of the Fourier transformation in [/" absorbed a number of
earlier generalizations, in particular that of Bochner [1], while pre-
serving the classical ease of calculation. Apart from this and the
precision derived from the Lebesgue integral the first result in Sec-
tion 7.1 dating from this century is Theorem 7.1.13 for which the
Fourier series analogue is due to Young for even integer pi and to
Hausdorff [lJ in general. Theorem 7.1.12 and its application to the
Hausdorff-Young theorem for Fourier series as well as Fourier in-
tegrals is due to M. Riesz [2], but our proof of Theorem 7.1.12 is due
to Thorin [1]. The best possible constant in Theorem 7.1.13 has been
found by Beckner [1]. Fourier transforms of homogeneous distri-
butions were discussed in much greater detail by Gelfand and Shilov
[2J and by Garding [4]. Theorem 7.1.24 is very close to the Herglotz-
Petrovsky formula as given in Atiyah, Bott and Garding [lJ (see
Section 12.6). Fourier transforms of densities on manifolds in JR" occur
Notes 249
frequently in scattering theory and will be useful in such contexts in

Chapter XIV. Theorem 7.1.26 to Corollary 7.1.30 can essentially be
found in Hormander [32] and Agmon and Hormander [1] with
references to earlier literature.
The term Paley-Wiener-Schwartz theorem for Theorem 7.3.1 is well
established although perhaps not quite appropriate; Paley and Wiener
[1] actually proved a case of Theorem 7.4.2 dealing with Fourier-
Laplace transforms of L2 functions on a half line and Schwartz
[5] proved Theorems 7.4.2 and 7.4.3 in full. Theorems 7.3.2 and 7.3.6
are from Malgrange [1]. Asgeirsson [1] proved Theorem 7.3.4 in a
quite different way, the refinement is due to Lewy [2]. Theorem
7.3.8 was first stated and proved in the predecessor of this book, but
the main idea comes from Ehrenpreis [2] (see also Malgrange [1] and
Hormander [14]). Theorem 7.3.10 is due to Ehrenpreis [1] and Mal-
grange [1] who proved it by duality. A constructive proof was already
attempted by Cauchy [1]; unfortunately it involves Fourier trans-
forms of exponentially increasing functions but this flaw is not hard to
correct. For large classes of differential operators fundamental so-
lutions were constructed long ago (see e.g. Fredholm [1], Herglotz
[1], Zeilon [1]). A general construction yielding good regularity proper-
ties was given in the predecessor of this book. (See Agranovich [1]
for an earlier one.) The improved version here was published by
Hormander [29].
For a history of the Weierstrass preparation theorem and criticism
of current terminology and general trends in mathematics we refer to
Siegel [1]. For earlier proofs of the Malgrange preparation theorem
the reader should consult Malgrange [6], Mather [1] and Nirenberg
[3]. The proof given here is close to that of Mather [1]. Theorem
7.5.13 is due to Chester, Friedmann and Ursell [1] and Levinson [1]
in the analytic case. The theory of unfolding of singularities has given
a great simplification and an extension to COO by means of the
Malgrange preparation theorem. (See Guillemin and Schaefer [1],
Duistermaat [1] and the references there.) We have proved the earlier
results here in this spirit.
The Fourier transform of Gaussians is of course classical. As
examples we have given the fundamental solution of the Kolmogorov
equation (see Kolmogorov [1]) which is important in the theory of
Brownian motion. The operator (7.6.14) is one of the simplest normal
forms of hyperbolic operators with double characteristics (cf.
Hormander [36]). The Airy function was introduced by Airy [1] to
study light near a caustic (see Section 12.2). The proof by Stokes [1]
of the asymptotic expansion of A i might be considered as a fore-
runner of the method of stationary phase in Section 7.7. This has been
very popular with the physicists in this century under the name of the
(J)WKB method for (Jeffreys), Wentzel, Kramers, Brillouin, or some-
times, from a somewhat different analytical point of view, the name
of the saddle point method. We refer to Froman [1] for a systematic
discussion. For several variables the method seems to have appeared
first in Hlawka [1] who proved Theorems 7.7.14, 7.7.16 and Corollary
7.7.15. The presentation here is close to Hormander [26, 34] up to
and including Theorem 7.7.6 although we have avoided using the
Morse lemma here. It can be found in a suitable form in Hormander
[26]. The extension to complex valued phase functions is due to
Melin and Sjostrand [1]. They used the notion of almost analytic
continuation, which is a systematic development of the arguments
used here in the proof of Theorem 3.1.15. The Malgrange preparation
theorem was proved with such techniques by Nirenberg [3]. It is
therefore not surprising that we have been able to replace the almost
analytic machinery by the Malgrange preparation theorem. However,
the reader should be aware that the methods of Melin and Sjostrand
are more precise from the point of view of the number of derivatives
required. Theorem 7.7.18 goes back to Airy [1]. A complete proof in
the analytic case was given by Chester, Friedmann and Ursell [1]; see
also Ludwig [2]. Here we have followed the simpler and more general
modern approach of Guillemin and Schaeffer [1], Duistermaat [1]
who built on the progress in singularity theory by Thorn, Arnold and
others.
The notion of oscillatory integral in Section 7.8 was introduced as
here in Hormander [26]. It will be convenient particularly in Chapter
XXV to be able to use this suggestive notation which is of course
common in applied mathematics without precise definitions.
As already mentioned in the text the spaces H(s) are special cases
of the spaces of Sobolev [2] when s is a non-negative integer. For
negative integers s they arose in the theory of partial differential
equations in connection with duality methods (see e.g. Lax [2]) and
for half integers in connection with boundary problems (see e.g. Aron-
szajn [1]). They have been very much generalized and studied during
the last decades. Examples of these more general Besov spaces will be
encountered in Chapters XIV and XXX but we refer to Peetre [4] for
a general discussion. Theorem 7.9.5 is essentially due to Mihlin [1]. In
the form given here it was proved in Hormander [13]. However, the
prototype is the theorem of M. Riesz [3] on conjugate functions and
its n dimensional generalization by Calderon and Zygmund [1] whose
proof is followed here. We refer the reader who wants to study these
matters further to Stein [1]. In the theory of linear partial differential
equations to which this book is devoted we shall have very few
occasions to refer to the IJ or Holder theory at all.
Chapter VIII. Spectral Analysis
of Singularities
Summary
In Chapter VII we have seen that a distribution u of compact support
is smooth if and only if the Fourier transform Ii is rapidly decreasing.
If u is not smooth we can use the set of directions where Ii is not
rapidly decreasing to describe which are the high frequency com-
ponents of u causing the singularities. This analysis turns out to have
an invariant and local character. For a distribution UE~'(X) on a Coo
manifold X we are therefore led to define a set
WF(u) c T*(X) '- 0
with projection in X equal to sing supp u, which is conic with respect
to multiplication by positive scalars in the fibers of T*(X). We call it
the wave front set of u by analogy with the classical Huyghens
construction of a propagating wave. In this construction one assumes
that the location and oriented tangent plane of a wave is known at
one instant of time and concludes that at a later time it has been
translated in the normal direction with the speed of light. The data
are thus precisely rays in the cotangent bundle.
The advantages of the notion of wave front set are manifold. First
of all it allows one to extend a number of operations on distributions.
For example, the restriction of UE~'(X) to a submanifold Y of X can
always be defined when the normal bundle of Y does not meet WF(u),
that is, high frequency components of u remain of high frequency after
restriction to Y. Secondly, differential operators and to some extent
their fundamental solutions are local even with respect to the wave
front set. This leads to important simplifications in their study known
as microlocal analysis.
Section 8.1 gives the basic definitions of the wave front set and
some important examples. In Section 8.2 we then reconsider the
operations defined in Chapters III-VI from our new point of view.
Thus we obtain extended definitions of composition and multipli-
cation as well as more precise information on the singularities of the
252 VIII. Spectral Analysis of Singularities
results of these operations. In Section 8.3 we prove the simplest facts

on the wave front set of solutions of linear partial differential equa-
tions, in particular that the wave front set is included in the union of
the characteristic set and the wave front set of the right-hand side.
Note that since the characteristic set c T*(X)'-.. 0 usually projects
onto X it is not possible to give a satisfactory statement of this result
without the notion of wave front set. When the principal part is real
and has constant coefficients we also show that the wave front set is
invariant under the bicharacteristic flow, which in the case of the
wave equation reduces to the Huyghens construction above and so
justifies our terminology.
One can also consider a stricter classification of singularities, such
as the set sing SUPPA U of points where U is not a real analytic function.
This set too admits a spectral decomposition to a set WFA(u)c
T*(X)'-.. 0, which is defined in Section 8.4 and studied in Sections
8.5 and 8.6. In particular this notion allows one to state a more
precise form of the uniqueness of analytic continuation: If a distri-
bution vanishes on one side of a C 1 hypersurface Y and th~ nor-
mal of Y at y is not in WFA(u), then u vanishes in a neighborhood
of y. In other words, the normals of the boundary of supp u must be
in WFA(u) where the boundary is in C 1 • In Section 8.5 we also give a
notion of normal to the boundary of a general closed set making this
statement valid in general. This concept is discussed geometrically at
some length in preparation for some later applications. The first
comes already in Section 8.6 where various generalizations of the
theorem of Holmgren on unique continuation of solutions to partial
differential equations with analytic coefficients are given.
In Section 8.7 finally we discuss the analytic wave front set for
distributions obtained as limits of F(x + i y)-l where F is analytic and
y -+ 0 in a cone such that the zeros of F are only encountered in the
limit. The results are useful in the study of the Cauchy problem in
Chapter XII.
8.1. The Wave Front Set
If VEC'(JR") we can decide whether v is in C~ by exammmg the

behavior of the Fourier transform (j at 00. In fact, if VEC~(JR") then
(8.1.1)
by Lemma 7.1.3. Conversely, if (8.1.1) is fulfilled then VEC~ by
Fourier's inversion formula (7.1.4). (See also Theorem 7.3.1.) For a
8.1. The Wave Front Set 253
general vEg' we have defined sing supp v as the set of points having
no neighborhood where v is in Coo (Definition 2.2.3). Similarly we can
introduce the cone l'(v) of all '1 EIRn " 0 having no conic neigh-
borhood V such that (8.1.1) is valid when eEv. It is clear that l'(v) is
then a closed cone in lRn " 0, and we have l'(v) =0 if and only if
VECO".
While sing supp v only describes the location of the singularities,
the cone l'(v) describes only the directions of the high frequencies
causing them. We can combine the two types of information by using
the following lemma.
Lemma 8.1.1. If ¢ECO"(lR n) and vEg'(lRn) then

(8.1.2) l'(¢v)cl'(v).
Proof The Fourier transform of u=¢v is the convolution
u(~)=(2n)-n S<p('1)v(e-'1)d'1
where <pE[!>. For some M~O we have
Iv(~)1 ~ C(1 + I~DM.
Let 0< c< 1 and split the integral into the parts where 1111 < clei and
I'll ~ cI~I. In the second case I~ -111 ~ (1 + c- 1) 1'11. Hence
(8.1.3) (2n)n lu( ~)I ~ sup IV('1)111 <p11 L'
I~- ~I < cl~1
+C S 1$('1)1(1 +C- 1)M(1 +1'1I)Md'1.

1~I>cl~1
If r is an open cone where (8.1.1) is valid and r 1 is a closed cone

cru {O} we can choose c so that 11Er if eEr1 and Ie -111 <clel, for
this is obviously possible when lei = 1. Since 1111 ~ (1- c) lei it follows
from (8.1.3) and (8.1.1) that u is rapidly decreasing in r 1• In fact, we
have for N~O
(8.1.3)' sup(l + leltlu(e)1 ~(l-c)-N sup IV(11) I(1 + 1'1lt I <PilL'
~ r
+ C(1 + C-1)N+M S1<P(11)1(1 + 111I)N+M d11·
The lemma is proved.
If X is an open set in lRn and UE~/(X), we set for XEX

(8.1.4) l'x(u) = n l'(¢u); ¢ECO"(X), ¢(x)=l=O.
'"
From Lemma 8.1.1 it follows that
(8.1.5) l'(¢u)---+l'X<u) if ¢ECO"(X), ¢(x) ,*,0 and supp¢---+{x}.
In fact, if V is an open cone =>Ex(u), the compactness of the unit

sphere shows that we can find ¢ l' ... , ¢ jE C'(/(X) with
When ¢EC'(/(X) and supp¢ is so close to x that ¢1 ... ¢j'*0 there, we

can write ¢=l/I¢l ... ¢j with l/IEC'(/(X) and obtain from (8.1.2)
This proves (8.1.5) since by definition E(¢u)=>Ex(u) when ¢(x)=FO.

In particular (8.1.5) implies that EAu)=0 if and only if ¢UECoo for
some ¢ EC'(/(X) with ¢(x) =l= 0, that is, x¢sing supp u.
Definition 8.1.2. If u E.@'(X), then the closed subset of X x (1R n" 0)

defined by
WF(u) = {(x, e)EX x (1R n " 0); eEEAu)}
is called the wave front set of u. The projection in X is sing supp u.
The set WF(u) is conic in the sense that it is invariant under

multiplication of the second variable by positive scalars. It could
therefore be considered as a subset of X X sn-1 where sn-1 is the unit
sphere.
Proposition 8.1.3. If uE/iS"(1R n) then the projection of WF(u) on the

second variable is E(u).
Proof. The projection W is contained in E(u) by the definition of

W F (u). It is closed since the intersection with the unit sphere is the
projection of a compact set in 1Rn x sn- 1. If V is a conic neighborhood
of W then every xE1R n has a neighborhood Ux such that
E(¢u)c V if ¢EC'(/(Ux).
We can cover suppu by a finite number of such neighborhoods UXj
and choose ¢jEC'(/(Ux) with E ¢j = 1 near supp u. But then it follows
that
E(u)=E(I¢ju)C UE(¢ju)c V
which proves the proposition.
Proposition 8.1.3 shows that WF(u) contains all information in

sing supp u and in E(u). However, the projection in Proposition 8.1.3
is of limited interest since it is not invariant under a change of
variables.
Theorem 8.1.4. If X is an open set in 1R" and S a closed conic subset of

X x (1R"" 0) then one can find UE.@'(X) with WF(u)=S.
Proof It is sufficient to prove the statement when X =1R" for other-

wise we can apply this case to the closure of S in 1R" x (1R"" 0).
Choose a sequence (Xk' (Jk)ES with l(Jkl = 1 so that every (x, (J)ES with
I(JI = 1 is the limit of a subsequence. Let CPEC~ and <p(0) = 1. Then
u(x) = L k- 2 cp(k(x - x k»eik3 (X, Ok)
00
(8.1.6)
I
is a continuous function in 1R", and we shall prove that WF(u)=S.

First we prove that WF(u)cS. If (xo, ~o)¢S we can choose an open
neighborhood U of Xo and an open conic neighborhood V of ~o such
that
(8.1.7) (U x V)nS=0.
Write u=u I +u 2 where u l is the sum of the terms in (8.1.6) with xdU
and U 2 the sum of terms with XkE U. Then u l E Coo in a neighborhood
U I of Xo because all but a finite number of terms vanish in UI if
0 1 cU. Now
(8.1.8)
Here (Jk¢V because of (8.1.7). If VI is another conic neighborhood of

~o and VI cVu{O} then 1~-'1I~c(l~I+II'/D when ~EVI and I'/¢v, for
some c> 0, since this is true when I~ I+ I'll = 1. Thus
I~ -k 3(Jkl ~ c(I~1 + k3)~cl~lt k, ~EVI'
and since <pEY it follows that 122 is rapidly decreasing in VI' Thus
(xo, ~o) is not in WF(u).
Now let (xo, ~o)ES. Choose XEC~ equal to 1 near xo' To prove
that (xo, ~o)EWF(u) we must show that fit cannot decrease rapidly in
a conic neighborhood of ~o' To do so we first observe that
X(x)cp(k(x - x k» = CPk(k(x - x k»
where CPk(X) = X(x/k + xk)cp(x) belongs to a bounded set in Y. The
Fourier transform of Xu is a sum of the form (8.1.8) with cP replaced
by CPk' If Xk is close to Xo and k is large then CPk=CP, and we obtain for
any N
IXU(P(Jk)1 ~k-"-2 - CN L j-"-2(1k 3(Jk _j3(J)jj)-N.
j*k
Here
so the sum is O(k-N).lfwe choose N>n+2 we obtain for large k that

Ixu(P Ok)1 ~ k- n - 2/2
if x k is close to xo' Since (xo, ¢o/I¢ol) is a limit point of the sequence
(Xk' Ok) it follows that XU cannot decrease rapidly in a conic neigh-
borhood of ¢° and the theorem is proved.
We shall now determine the wave front set for some classes of
distributions which occur very frequently.
Theorem S.1.S. Let V be a linear subspace of JR" and u=uodS, where

UoECoo(V) and dS is the Euclidean surface measure. Then
WF(u)=suppu x (V.L" 0).
Proof If XE Cg' then

(XU)(¢) = S e-i<x·~>X(x)uo(x)dS(x).
v
If we write ¢ = ¢' + f' where ¢' E V and ¢" E V.L, then this is a rapidly
decreasing function of ¢' which does not vanish on any open set
unless Xu = O. Hence XU does not decrease rapidly in any open cone
meeting V.L unless xu=O, but there is rapid decrease in every cone
where I¢ I~ CI n
This proves the assertion.
It would have been sufficient to prove Theorem 8.1.5 for dS itself,

for we have always
(8.1.9) WF(au)cWF(u) if aEC OO •
This follows at once from the definition. Another important general
fact is that for all ()(
(8.1.10) WF(D"u)cWF(u).
To prove this we take XECg' equal to 1 near x and Xl ECg' equal to 1
in a neighborhood of supp X. Then we have
1:AD"u) c1:(XD"u) =1:(XD"Xl u)c1:(D"Xl U)c1:(Xl u).
When supp Xl ..... {x} it follows that (8.1.10) is valid. Summing up, we
have
(8.1.11) WF(Pu)c WF(a)
if P is any linear differential operator with COO coefficients.
Next we shall examine the boundary values of analytic functions
as defined in Theorem 3.1.15. There r is an open convex cone. Let
S.l. The Wave Front Set 257
(8.1.12)
be the dual cone. It is closed, convex and proper, that is, it contains no
straight line, for r would otherwise be contained in a hyperplane and
lack interior points. Conversely, every closed convex proper cone ro
is the dual cone of precisely one open convex cone r. It is defined by
(8.1.13)
The proof by the Hahn-Banach theorem is very close to that of

Theorem 4.3.2 and is left for the reader. Instead we shall prove
Theorem 8.1.6. If the hypotheses of Theorem 3.1.15 are fulfilled, then

(8.1.14)
where ro is the dual cone of r.
Proof If ¢EC~(X) the representation (3.1.20) of <fo, ¢) is valid with

N replaced by any integer v ~ N provided that N is also replaced by v
in the definition (3.1.18) of QJ. Hence
(8.1.15) (#o)(e)=<foe-i(.,o, ¢) = SQJ(x, Y)f(x+ iY)e-i<X+iY'~)dx

+(v+l) SS f(x+itY)e-i<X+itY,~) I aa¢(x)(iynrx!tVdxdt.
O<t<l lal~v+1
When <Y, 0 < 0 it follows that
(8.1.16) l#o(e)I~C"',v(e<Y'o+ I et(Y,OtV-Ndt)
= C""v(e(Y'O +(v-N)! <- Y, ON-v-1).

The right-hand side in OOeI N - v - 1) in a conic neighborhood of any
point in the half space <Y, 0 <0, Hence };(¢fo)c{e; <Y, 0 ~O} for
every YEr with IYI<y, so };(¢fo)crO which proves the theorem.
The hypotheses in the theorem can be weakened in various ways.

In particular it is sufficient to assume f analytic for ZEX 1 +ir1 and
IImzl small when X1~X and t1 cru{O}. We could also have added
to fo l:J. COO term since this does not affect WF(fo)' A converse result
is then valid, and it will be proved in Section 8.4. We shall also prove
then that Theorem 8.1.6 remains valid when singularities are defined
as points of non-analyticity.
To prepare for a discussion of the wave front set for homogeneous
distributions we shall now prove a modification of Lemma 8.1.1.
Lemma 8.1.7. If VE!/' then WF(v)c1R.n x F where F is the limit cone

ofsuppvat 00, consisting of all limits of sequences tjXj with XjESUPPV
and 0< tj -+0.
r r n F = {o}
°
Proof F is obviously closed. For every closed cone
we can choose I: > and C so that
with
I~ -111 ~I:I~I if ~Er, 11 ESUPP V and I~I > C.

In fact, we could otherwise choose ~jEr and 11 j ESUPP V so that
l~j-11jl < I~jl/j and I~jl > j. The sequence '1il~jl will then have a limit point
()ErnF with 101=1 which is a contradiction. If cPECg'(JR.n) then the
Fourier transform of u=cPv is (2n)-nQj*v. Choose I/tEcoo(JR.n) so that
1/t(~)=1 when 1~1>1 and I/tW=O when 1~1<1/2. Then <PR(~)
=Qj(~)I/t(~/R) is equal to QjW when I~I~R, hence
(2n)nu(~)=t\(<PR(~-'1)) if ~Er and R~I:I~I, I~I>C.
Since VE!/' it follows that for some N, C', C" we have when ~Er,
I~I> C,R~el~1
lu(~)I~C' I supl'1<%D:<PR(~-'1)1
IHPI;;;N
~ C"(1 +1~lt I sup 1'1<%DPQj('1)I.
IHPI~N Iql>R/2
If we choose R=I:I~I the right-hand side is rapidly decreasing since

QjE!/. This proves the lemma.
Theorem 8.1.8. If UE~'(JR.n) is homogeneous in JR.n,o then
(8.1.17) (x,~)EWF(u)¢>(~, -x)EWF(u) if ~ =1=0 and x =1=0,
(8.1.18) XESUPPU ¢> (0, -x)EWF(u) if x =1=0,
(8.1.19) ~ESUppU ¢> (0, ~)EWF(u) if ~ =1=0.
°
Proof Assume first that u is homogeneous in JR.n. To prove (8.1.17) it
is sufficient to show that if Xo =1= 0, ~ 0 =1= then
(8.1.17)'
u
for is also homogeneous and (8.1.17)' applied to gives the reversed u
implication since ~=(2n)nu. Choose XEC~(JR.n) equal to 1 in a
neighborhood of ~o and I/tEC~(JR.n) equal to 1 in a neighborhood of
Xo so small that
(8.1.20) (supp I/t x supp x)n WF(u) =0.
We have to estimate the Fourier transform of v = XU in a conic
neighborhood of -xo' Let l/t(x)=l when Ix-xol<2r and consider
v( - tx) when Ix - xol < rand t is large. If u is homogeneous of degree

a in JRn then
v( -tx)=X*u( -tx)=(u, X( -tx+ .)=ta+n(u, X(t(. -x))).

Set I/Iu = U o and (l-I/I)u = u 1 • Then I" (u o) n supp X =0 by Proposition
8.1.3 and (8.1.20). Hence
J
(u o, X(t(. - x))) = uo<e)X(~/t)ei<x.~> d~/t"
is rapidly decreasing as t -+ 00, for t N 120 (t~) X(~) is bounded for every
N. Moreover,
(u 1 , X(t(. -x))) =(u, (l-I/I)X(t(. -x)))
is also rapidly decreasing, for

y -+ tN(l-l/I(y))X(t(y-x))
is bounded in Y for any N. In fact, Ix - xol < r by hypothesis, and
ly-xol>2r in supp(l-I/I(y)), hence t~tly-xl/r and lyl~ly-xl+
IXol + r. Since XEY this completes the proof of (8.1.17)'.
In general it follows from (7.1.19) and (7.1.18) that we can write u
=w+wo+Q(D)Wl' u=w+wo+Q(~)Wl where w is homogeneous,
suppwoC:{O}, w1 (x)=lxl-"/c" when x*o, Wl(~)= -logl~l, and Q is a
polynomial. Since u - wand 12 - ware in Coo except at the origin, we
obtain (8.1.17) in general.
To prove (8.1.18) we first observe that since a=(2n)"u it follows from
°
Lemma 8.1.7 with v=u that x¢suppu => (0, -x)¢WF(u). Assume now
that (0, -xo)¢WF(u). Choose XECg' equal to 1 at so that the Fourier
transform of Xu is rapidly decreasing in a conic neighborhood r of
°
-x o' Adding to u a term with support at does not affect (8.1.18) so
we may assume that u is homogeneous of degree a in JR" unless a =
-n-k and (3.2.24) is valid for an integer k~O. Hence the Fourier
transform of Xu at tx is
x*u(tx)=(u, X(. +tx) =t"(u, ¢t(' +x) +logt L ca(aaX)(tx)/a!
lal~k
where ¢t(x)=t"X(tx) and the sum should be omitted unless k= -n-a

is an integer ~O. When XEr the left-hand side tends rapidly to as
t -+ 00, and so does the sum. Thus
°
(u,¢t(' +x)=u*¢t(x)-+O in r as t-+oo.
The convolution converges to (2n)"u in Y'(lR"). Hence U=O in r so
Xo ¢supp u and (8.1.18) is proved.
If u and therefore 12 is homogeneous in JR" then (8.1.19) follows if
(8.1.18) is applied to U. If u is not homogeneous then u(t.)_ta u is a
distribution =1= 0 supported by 0 for some t > O. Hence (O,~) is in

WF(u) for every ~ =1= 0, and ~ ESUPP a since a= U + V where U is of the
form (7.1.19) with Uo and Q homogeneous, Q$O and V is a poly-
nomial. The proof is complete.
Our final example concerns the distributions defined by oscillatory
integrals in Section 7.8.
Theorem 8.1.9. For the distribution
A= Jei c/>(.,8)a(., e)de
defined in Theorem 7.8.2 we have
(8.1.21) W F(A) c {(x, ¢~(x, e)); (x, e)EF and ¢~(x, e) =O}.
Before the proof we observe that ¢~(x, e) = 0 implies ¢(x, e) = 0

since ¢ is homogeneous of degree 1 with respect to e. By hypothesis
Im¢~O so it follows that Im¢~(x, e)=o. Thus ¢~(x, e) is real in
(8.1.21).
Proof Let t/lE C'~(X). Then the definition of A means that

f,4.(~)= JJei (c/>(x,8)-(X,O)t/I(x)a(x, e)dxde
as an oscillatory integral. We want to show that this is rapidly
decreasing in any closed cone VelR n which does not intersect
{¢~(x, e); (x, e)EF, xESUppt/l, ¢~(x, e)=O}.
Then we have for some c > 0
(8.1.22) I~ -¢~(x, e)1 + lell¢~(x, e)1 ~ c(I~1 + leI)
if (x, e)EF, XESUPP t/I, ~EV.
To prove (8.1.22) we first observe that ¢~(x, e) and lei ¢~(x, e) are
continuous in F with the value 0 when e=o. By the homogeneity it
suffices to prove (8.1.22) when I~ I+ lei = 1. By the compactness we only
have to show then that the left-hand side is never 0 when (x, e)EF,
xESUppt/l, ~EV. If e=o we have 1~-¢~(x,e)l=l, and when e=l=O,
¢~(x, e) = 0 we have ~ =1= ¢~(x, e) since ~ EV, which proves (8.1.22).
Expressing the oscillatory integral by means of the partition of
unity in e used in the proof of Theorem 7.8.2 we have
f
f,4(~)= JJei (c/>(x,8)- (x, mt/l(x)Xv(e)a(x, e)dxde.
o
Each term is in fI'. With R=2 v - l the terms with v=l=O can be written
(8.1.23) RN JJe (Rc/>(x,8)- (x, ~»t/I(x) Xl (e)a(x, R e)dxde.
i
If cJ>(X, O)=(ReP(x, O)-<x, ¢»/(R+I¢I) and ¢EV we have by (8.1.22)

1cJ>~1 + 1cJ>~1 ~ c(R 101 + I¢I)/(R + I¢I) ~ c
in the support of l/I (x) XI (O)a(x, RO). With y=max(1-p,b)<1 we have
ID~Del/l(X)Xl (O)a(x, RO)I ~ C"pRm+Y(IIXI+IPil.
By Theorem 7.7.1 it follows that (8.1.23) is estimated for large k by

CkRm+N+kY(R + I¢I)-k ~ CkR- 1 1¢lm+N+ 1+(y-l)k if ¢EV.
L2 XA (¢)
00
Since 1- v = 2 we conclude that is rapidly decreasing in V,

1
which completes the proof of the theorem.
Theorem 8.1.5 is a very special case of Theorem 8.1.9. In fact, let

M be a COO manifold in Rn defined near a point xoEM by
ePl(X)= ... = ePk(X) =0
where dePl' ... ,dePk are linearly independent at xo' If aEC~(Rn) has
support near Xo then
k
S
A = a(.) ei <P(.,8)dO, eP(x, 0) = L eP/X)Oj
1
is by (7.8.5) equal to
(2nta(x)b(ePl' ... , ePk)
°
where b is the b "function" at in Rk. This is an arbitrary smooth
density on M with support near xo' Theorem 8.1.9 gives now
WF(A)c {(x, eP~(x, 0»; eP/x)=O,j = 1, ... , k, xEsuppa}.
In the right-hand side we have the conormal bundle of M at supp a.
(We shall see in Section 8.2 that equality is valid.) We leave as an
exercise for the reader to apply Theorem 8.1.9 to the distributions in
Example 7.8.4. Theorem 8.1.9 is in fact one of the basic results leading
to Lagrangian distributions (see Chap. XXV).
8.2. A Review of Operations with Distributions
Singularities made it impossible to give a general definition of multi-

plication of distributions and composition with maps. We shall now
show that the definition of both operations can be extended when one
takes into account the more refined description of the singularities
given by the wave front set. As in Section 6.1 we shall always define
such operations by continuous extension from the smooth case, so the

first point to discuss is the topology in the space of distributions with
a given bound for the wave front set.
Let X be an open set in IR n, let r be a closed cone in X x (IR n '-... 0)
and set
~~(X)= {UE~I(X), WF(u) cr}.
Lemma 8.2.1. A distribution UE~'(X) is in ~~(X) if and only if for

every cPEC'g'(X) and every closed cone V cIRn with
(8.2.1 ) r n(supp ¢ x V) =0
we have
(8.2.2)
Proof. (8.2.2) implies that (x, ~)¢WF(u) if ¢(x)=!=O and ~ is in the

interior of V, so the condition is sufficient. On the other hand, if
UE~~(X) and ~EL(cPU) then (x, ~)Er for some XESUPP ¢, by Proposi-
tion 8.1.3 and (8.1.9), so ~¢V by (8.2.1). This proves (8.2.2).
Definition 8.2.2. For a sequence UjE~~(X) and UE~~(X) we shall say

that Uj-+U in ~~(X) if
(i) Uj -+ U in ~I(X) (weakly)

(ii) supl~INlfzJ(~)-¢Ul~)I-+O, j -+ 00,
v
for N = 1, 2, ... if cP EC'g'(X) and V is a closed cone in IR n such that
(8.2.1) is valid.
Since (i) implies that ¢Uj -+ ¢U uniformly on every compact set and
N is arbitrary in (ii), we can replace (ii) by
(ii)' sup supl~INI¢Ul~)1 < 00, N = 1,2, ....

j ~EV
The following is an extension of Theorem 4.1.5.
Theorem 8.2.3. For every UE~~(X) there is a sequence UjE C'g'(X) such
that U j -+ U in ~~(X).
Proof As in the proof of Theorem 4.1.5 we take uj=(Xju)*¢j where

a) XjE C'g'(X) and Xj = 1 on any compact set in X for large j,
b) O~cPjEC'g'(IRn), J¢jdx=l, and sUPPcPj is so small that (4.1.4)
holds.
Then we know already that UjEC~(X) and that Uj--+U in .@'(X). If

¢ and V satisfy (8.2.1) we can find I/IEC~(X) equal to 1 in a neigh-
borhood of supp ¢ and a closed cone W with interior containing
V . . . . 0 so that
r n(supp 1/1 x W)=0.
For largej we have ¢uj=¢Wj where wj=¢j*(l/Iu), hence
Iw) =1$) Ifttl ~Ifttl.
Since I~I is rapidly decreasing in W, the proof of Lemma 8.1.1 gives
(ii)' (see (8.13)'), and the theorem is proved.
Theorem 8.2.4. Let X and Y be open subsets of 1Rm and 1Rn respectively
and let f: X --+ Y be a COO map. Denote the set of normals of the map
by
N f = {(f(x), 1])E Y X 1Rn; 1'(x) 1] =O}.
Then the pullback f*u can be defined in one and only one way for all
UE.@'{Y) with
(8.2.3)
so that f* U=Uo f when UECOO and for any closed conic subset r of
Y x (1Rn ....... 0) with r n Nf = 0 we have a continuous map f*: .@~{ Y) --+
.@f*T(X),
(8.2.4) f* r = {(x, 1'(x) 1]); (f{X),1])Er}.
In particular we have for every UE.@'{Y) satisfying (8.2.3)
WF(f* u) c f* WF(u).
Proof Define f*u=uof when UECOO(y). By Theorem 8.2.3 the theo-

rem will be proved if we show that f* maps sequences UjEC OO con-
verging in '@~(Y) to sequences converging in .@j*r{X), First we shall
just prove convergence in .@'(X). If UEC~{Y) and XEC~(X) we have
by Fourier's inversion formula applied to U
(8.2.5) J
<f*u, X> =(2n)-n 11{1])/ X<1]) d1],
IX<1]) = Jx{x)ei(f(X),'I)dx.
Let XoEX, set yo=f{x o), I;o={1]; (Yo,1])Er} and choose

a) a closed conic neighborhood V of I;o in 1R n. . . . 0 such that
l' (x o) '1 =1= 0, 1] E V,
b) a compact neighborhood Yo of Yo such that V is a neigh-
borhood of I; for every yE Yo; such a neighborhood exists since r is
closed,
c} a compact neighborhood X 0 of Xo with f(X o} in the interior of

Yo and 1'(x} 11=1=0 ifxEXo and l1EV.
Choose cPEC~(Yo} equal to 1 on f(X o}. Then (8.2.5) is valid when
XE C~(X 0) for every UE COO(y) if U is replaced by cPu in the right-hand
side. Since d<f(x),l1>=<dx, 1'(x) 11> and
1111 ~ CI1'(x) 111 if XESUPP X and l1E V,
it follows from Theorem 7.7.1 that
(8.2.6)
If UjECOO(y) and uj-tu in .@~(Y) we have
l<Puj(l1)1 ~ C~(1 + 1111)-N, 11 If V, N = 1, 2, ...
and for some M (cf. the proof of Theorem 2.1.8)
l¢Uj(l1)1 ~ C(1 + 111I)M, l1E1R·.
Hence dominated convergence in V and in CV gives
<f*uj, X> - t (2n)-· S¢U(11) I il1) dl1,
so f*u j converges in .@' to a limit independent of the sequence
chosen. (See the remark after Theorem 2.2.4.) We denote the limit by
f*u.
In the proof of the continuity of f*: .@~(Y)-t.@f'r(X) we set xf*u j
=v j . Then (8.2.5) with X replaced by xe-i< .. ~> gives
Vj(¢)=(2n)-· S¢U j(l1)Iil1, ¢)dl1,
I x(l1, ¢)= Sx(x)ei«f(x).q>-<x·~»dx.
Let W be an open conic neighborhood of 1'(x o) I;.o =(f*r)xo' We may
assume the neighborhoods V and X 0 above chosen so that
1'(X)l1EW if XEX o and l1EV.
Then we have for some e>O
11'(X}11-¢I~e(l¢I+1111) if XEX o, l1EVand ¢lfW
for the left hand side cannot vanish when I¢I + 1111 = 1. Hence Theorem
7.7.1 gives for any N
(8.2.6)' IIx(11,~)I~CN(1+1~1+1111)-N if~lfWandl1EV.
If 11 If V we use another obvious consequence of Theorem 7.7.1,

(8.2.6)"
Summing up the preceding estimates we obtain when ~ If W
IVj(~)1 ~ C~(J (1 + I~I + 1111)M -N dl1 +(1 + IW- N J l¢Uil1)I(1 + Il1lt dl1).
v tV
Since ljJu j satisfies condition (ii)' (after Definition 8.2.2) in CV it
follows that
......-::-:--..
sup supl~IN Ixf* U 8)1 < r:IJ, N = 1,2, ...
j ~~w
if W is a conic neighborhood of (f* r)xo and supp X is sufficiently

close to xo. By a partition of unity it follows now that f*u j -+ f*u in
fl)'rr, which completes the proof.
If X is a COO manifold and uEfl)'(X) we can now define

WF(u)cT*(X)-.....O so that the restriction to a coordinate patch XI(
is equal to K* WF(uo K- 1 ). In fact, when f is a diffeomorphism be-
tween open sets in lR.n it follows from Theorem 8.2.4 that WF(f*v)
is the pullback of WF(v) considered as a subset of the cotangent
bundle. Hence the preceding definition is independent of the choice
of local coordinates. It is clear that WF(u) is a closed subset of
T*(X)-..... 0 which is conic in the sense that the intersection with the
vector space Tx*(X) is a cone for every XEX. Indeed, these are local
properties inherited from the local coordinate patches.
If E is a COO vector bundle over X and uEfl)'(X, E), we define
WF(u) locally as UWF(u i ) where (u 1 , •.• , UN) are the components of u
with respect to a local trivialization of E. Passage to another local
trivialization only means that (u 1 , ••. , UN) is multiplied by an invertible
COO matrix so the definition is independent of the choice of local
trivialization.
Example 8.2.5. If u is a COO density on a COO submanifold Y of the

manifold X, then
WF(u) = {(x, ~)ET*(X); XESUPPU, ~ =1=0 and Tx(Y), < 0 =O}.
In fact, with suitable local coordinates this is just Theorem 8.1.5. Thus
the wave front set is the restriction to supp u of the normal bundle
N(Y)={(y, ~), YEY, (l;'(Y), 0 =O}
with the zero section removed.
Example 8.2.6. For the distributions (A ±iO)(2-n)/2 in (6.2.1) we have
(8.2.7) WF((A ±iO)(2-n)/2) = {(x, tdA(x)), X =1= 0, A(x) =0, (~O} u To* -.. . O.
In fact, since WF((t ± iO)a) c {(O, .), • ~ O} by Theorem 8.1.6, it follows
from Theorem 8.2.4 that the left-hand side is contained in the right-
hand side when x =t= 0, and this is trivial when x = o. Since

B(a) (A ± i 0)(2 -n)/2 = Cb o,
it follows from (8.1.11) and the fact that c=t=O, WF(bo)=To* ...... O (by
Theorem 8.1.5) that the left-hand side of (8.2.7) contains To* ...... O. Now
(A±iO)(2-n)/2 is not in Coo at any x=t=O where A(x)=O which proves
that there is equality in (8.2.7). If we recall from the proof of Theo-
rem 6.2.1 how x"± is written as a linear combination of (t ± i 0)" when
af/;:7L+, it follows that in lRn ...... O
(8.2.8) WF(A* x<;-n)/2) = {(x, tdA(x)); A(x)=O, x =t= 0, t=t=O}.
Since the wave front set is closed it must at the origin contain
(8.2.9) ((O,dA(x)); A(x)=O,x=t=O}={(O,~); ~=t=O,B(~)=O},
and when n+ (n_) is odd the argument above shows that it contains
To* ...... o. It will follow from Theorem 8.3.1 that the wave front set at 0
is given by (8.2.9) when n+ (n_) is even. For the advanced fundamen-
tal solution of the wave operator the preceding argument gives at
once that the wave front set is the normal bundle of the forward light
cone with the origin removed, together with To* ...... O.
In Chapter VI we could never pull a distribution back to a
manifold of lower dimension. However, this is sometimes allowed by
Theorem 8.2.4 and we list an important special case:
Corollary 8.2.7. Let X be a manifold and Y a submanifold with normal
bundle denoted by N(Y). For every distribution u in X with WF(u)
disjoint with N(Y) the restriction ul Y to Y is a well defined distribution
on Y, the pullback by the inclusion Y t:....+ X.
Example 8.2.8. Let Z be another submanifold of X and u a COO
density on Z. Then WF(u)cN(Z), and N(Z)nN(Y) is contained in
the zero section if and only if XEZ n Y and ~ETx* orthogonal to Tx(Z)
and ~(Y) implies ~ =0. This means that ~(Z)+ ~(Y)= TJX), that is,
that Z and Y have a transversal intersection. The restriction of u to Y
is then defined. It is a density in Y nZ. In fact, we may choose the
coordinates locally so that X =lR n, Y is defined by x' = 0 and Z by x"
=0 where x' denotes the first n' coordinates and x" the next n"
coordinates. Write x = (x', x", x"') where x'" are thus coordinates in
Z n Y. Then u = a(x', x"') b(x") which is the limit in .@~(Z) (X) when
e-+O of a(x',x"')cfJ(X"/e)e- n", if cfJEC~(lRn'') and JcfJ(x")dx"=1. The
restriction to Y is the function a(O, x"') cfJ(X"/e) en" which converges in
.@'(Y) to the density a(O, x"') b(x") on Z n Y considered as a sub-
manifold of Y.
We are now able to define the product of some pairs of distri-

butions which have singularities at the same point. To do so we first
observe that if u and v are functions in X then the product u(x) v(x) is
the restriction to the diagonal of the tensor product u(x) v(y) defined
for (x, Y)EX x X. Thus we shall first examine the tensor product.
Theorem 8.2.9. If uE2)'(X), vE2)'(Y) then

(8.2.10) WF (u (8) v) c: (WF(u) x WF(v))u«supp u x {O}) x WF(v))
u (WF(u) x (supp v x {O})).
Proof If uEtt'(1Rm) and vEtt'(]R.n) then the Fourier transform of UQ9V is

Ifu=l=O and v =1=0 it is clear that
U(~)V(I1).
E (uQ9v) c: (E(u) x E(v)) u ({O} x E(v)) u(E(u) x {O}).

To prove (8.2.10) we just have to apply this to ¢u and t/lv where ¢
and t/I are in COO with supports close to x and y respectively.
Theorem 8.2.10. If u, VE2)'(X) then the product uv can be defined as

the pullback of the tensor product u(8)v by the diagonal map b: X -+
X x X unless (x, ~)EWF(u) and (x, -~)EWF(v) for some (x, ~). When the
product is defined we have
(8.2.11 ) WF(uv)c: {(x, ~ +11); (x, ~)EWF(u)
or ~=O,(X,I1)EWF(v) or 11=0}.
Proof For the diagonal map b(x) = (x, x) from X to X x X we have

b'(x) t=(t, t) if tET,,(X),
tb'(x)(~, 11)=~+11 if~, I1ET,,*(X).
The theorem IS therefore an immediate consequence of Theorems
8.2.4, 8.2.9.
Example 8.2.11. If u and v are COO densities on submanifolds Y and Z

of X intersecting transversally, then uv is a COO density on YnZ. In
fact, uQ9v is a Coo density on Y x Z which intersects the diagonal in
X x X transversally so the statement follows from Example 8.2.8.
Since pullbacks have been defined here by continuous extension of

composition of functions, it is clear that the preceding definition of
multiplication extends the multiplication of distributions and smooth
functions defined in Chapter II. Similarly all standard rules of calculus
remain valid for the extended operations defined in this section; we
leave for the reader to fill in these obvious details.
Finally we shall discuss WF(ffu) when ff is a linear transfor-

mation from C~(Y) to ~'(X). For the sake of simplicity in statements
we restrict ourselves to open sets X and Y in Euclidean spaces.
Theorem 8.2.12. Let Xc 1Rnand Yc 1Rm be open sets and let K E

~'(X X Y). Denote the corresponding linear transformation from c~(Y)
to ~'(X) by ff. Then we have
WF(ffu)c {(x, ¢); (x, y, ¢, O)EWF(K) for some YESUppU}, UEc~(y).
Proof Let XoEX, choose XEc~(X) with X(xo)=l and set

Kl =(X®U)KE@"'(XX Y).
The Fourier transform of Xff u is K1 (¢, 0). Now Proposition 8.1.3 gives
l"(K 1 ) c {(¢, 11); (x, y, ¢, I1)EWF(K) for some XESUPP X, YESUppU}.
Hence it follows that
I'(Xffu)c{¢; (x,y,¢,O)EWF(K) for some XESUPPX and YESUppU}.
When SUPP X -+ {xo} the theorem follows.

The proof shows that ff maps c~(M) continuously into ~~(X) if
M is a compact subset of Y and
r={(x, ¢); (x, y, ¢, O)EWF(K) for some YEM}.
For the union of all such sets we shall use the notation
WF(Kh={(X,¢); (x,y,¢,O)EWF(K) for some YEY}.
It is of course not necessarily a closed set. If it is empty then ff is a
continuous map from c~(Y) to COO (X).
The first part of the following theorem is essentially dual to

Theorem 8.2.12.
Theorem 8.2.13. There is a unique way of defining ffUE~'(X) for every

UE@"'(Y) with WF(u)n WF'(K)y=0, where
WF'(K)y = {(y, 11); (x, y, 0, -11)EWF(K) for some XEX}
so that the map @"'(M)n~~=lU-+ffuE~'(X) is continuous for all com-
pact sets Me Y and all closed conic sets r disjoint with WF'(K)y. We
have
(8.2.12) WF(ff u) c WF(K)x u WF'(K) 0 WF(u)
where
WF'(K) = {(x, y, ¢, 11); (x, y, ¢, -11)E WF(K)}
is considered as a relation mapping sets in T*(Y),

T*(X), 0.
° to sets in
Proof Let I{IECO'(Y) be equal to 1 in a neighborhood of M. When

UE CO' and the support is contained in this neighborhood then
%u=%(ul{l)=.Yt;.I{I
where .Yt;. has the kernel
Ku=K(1@u).
If UE~'(Y) we have by Theorem 8.2.9
WF(1@u)={(x,y,0, 1'/); (y, I'/)EWF(u)}.
The product Ku is therefore defined for every UE~~(Y) if r is disjoint
with WF'(K}y, and Theorem 8.2.10 also gives
WF(KJ c {(x, y, ~, 1'/ + 1'/'); (y, I'/)E WF(u) and (x, y, ~, I'/')E WF(K)}
u WF(K)u WF(1@u).
It is clear that the map ~~3U -+ KuE~'(X X Y) is continuous. Setting
%u=.Yt;.I{I therefore gives a continuous extension of % to all UE~~
with support close to M. The uniqueness of such an extension follows
from Theorem 8.2.3. Using Theorem 8.2.12 we obtain
WF(%u)c{(x,~); (x,y,~, -I'/)EWF(K)
for some (y, I'/)EWF(u)} u WF(Kh,
Let now Xc]R.n, Yc]R.m and Zc]R.P be open sets and let KIE
~'(X x Y), K2E~'(Y x Z). Assume that the projection
(8.2.13) supp K 2 3(y, z) -+ z
is proper, that is, the inverse image of any compact set is compact. If
UECO'(Z) it follows then that Jt;UEt&"(Y), and by Theorem 8.2.12 we
have
If we assume that
(8.2.14)
then the composition ~ Jt; is defined as a continuous map
0
CO'(Z)-+~'(X). Thus it has a Schwartz kernel KE~'(XXZ). When Kl

and K 2 are smooth then
K(x, z)= SK1(x, y)K 2(y, z)dy.
In general the kernel K is also obtained by pulling the tensor product

K)®KzEf»'(XX Yx YxZ) back to Xx YxZ with the map
(x, y, z) -> (x, y, y, z). The normal set of this map is {(x, y, y, z; 0,1'/,
-I'/,O)} so it follows from (8.2.14) that the pullback is defined. Finally
the pullback is integrated with respect to y over Y, which gives K and
an estimate for WF(K) if one also recalls Theorem 8.2.12. This leads
to the following theorem for which the reader should have no difficul-
ty at all in supplying the missing and rather repetitive details of proof.
Theorem 8.2.14. When (8.2.14) is valid and the projection (8.2.13) is

proper then the composition ~ 0 ~ is defined and for the kernel K we
have
WF'(K)c WF'(K)) 0 WF'(Kz)u(WF(K))x x Z x {O})
u(X x {O} X WF'(Kz}z).
To illustrate the preceding results we shall consider convolution by

a distribution kEf»'(Rn). This has as kernel the distribution K ob-
tained by pulling back k with the map
R n x R n 3(X, y)->x- YERn.
Theorem 8.2.4 gives
WF(K)c{(x,y, (, -(); (x-y, ()EWF(k)).
For any constant c we have k= fc* K where fc(x)=(x+c, c), thus
1.,'(x)«(, 1'/)=(. Hence Theorem 8.2.4 also gives
WF(k)c {(x, 0; (x+c, c, (, -()EWF(K))
so there is in fact equality,
(8.2.15) WF(K) = {(x, y, (, - e); (x - y, ()E WF(k)}.
Since the two frequency components vanish simultaneously it follows
that convolution with k maps C~ into Coo and has a continuous
extension to a map tff' -> !!i/. Furthermore, we have
(8.2.16) WF(hu)c{(x+y,();(X,()EWF(k) and (y,e)EWF(u)}, uEtff'.
This improves Theorem 4.2.5 a great deal. (A direct proof of (8.2.16) is
easily obtained from Theorem 4.2.5 and the obvious fact that E(hu)
is contained in E(k)rlE(u) when k, uEtff'.)
8.3. The Wave Front Set of Solutions of Partial Differential Equations 271
8.3. The Wave Front Set of Solutions

of Partial Differential Equations
A differential operator with COO coefficients of order m in an open set

X e lRn is of the form
(8.3.1) P=P(x, D)= I a~(x)D~.
1~I~m
The principal part (or symbol) Pm is defined by

(8.3.2) Pm(x, ~)= I a~(x) ~~.
1~I=m
Note that the definition differs from that in Section 6.4 by a factor i"'.
Corresponding to (6.4.6)' we have
(8.3.2)'
If X is a Coo manifold then a differential operator of order m on X is

by definition an operator which has the form (8.3.1) in local coor-
dinate systems. From (8.3.2)' it follows that the principal symbol is
invariantly defined in the cotangent bundle. We shall now prove a
weak converse of (8.1.11).
Tbeorem 8.3.1. If P is a differential operator of order m with Coo

coefficients on a manifold X, then
(8.3.3) WF(u) e Char Pu WF(Pu), UE~/(X),
where the characteristic set Char P is defined by

(8.3.4) Char P= {(x, ~)ET*(X)'- 0, Pm(x,~) =o}.
Corollary 8.3.2. If P is elliptic, that is, Pm(x,~)=1= 0 in T*(X) '- 0, then

WF(u) = WF(Pu), UE~/(X).
Hence
singsuppu=singsuppPu, UE~/(X).
Proof of Theorem 8.3.1. We have stated the result for a manifold but
it is purely local so we may assume that X elRn in the proof. If
Pm(X O ' ~ 0) =1= 0 we can choose a neighborhood U e X of Xo and an open
cone V3~0 such that
(8.3.5)
for some C. Later on another condition will be imposed on U and V.
Choose a fixed <pEC't'(U) with <p(x o) = 1. To estimate q;;;(~) when ~EV
we first note that if

tpv= I( -D)"(a"v),
that is, tP is the formal adjoint of P, then Pu = f means that
<u, tpv) =<f. v), VEC~(X).
----..
We would like to find v so that the left-hand side is ¢u(~), that is,
tpv(x)=¢(x) e-i(x.o.
For large ~ an approximate solution is e-i(x.~> ¢(x)/P"'(x, ~). To improve
it we set
which gives the equation for v the form

w-Rw=¢.
Here R =R 1 + ... +Rm and R)~lj is a differential operator of order ~j
which is a homogeneous function of ~ of degree O. In fact, to obtain a
term in R which is homogeneous of degree - j we must let m - j
derivatives fall on the exponential e-i(x.~> and have no more than j
left which can act on w. By (8.3.5) all x derivatives of the coefficients
of R)~P are bounded in U x V. Formally the equation w-Rw=¢ is
satisfied by W= IRk¢. However, the sum is unlikely to converge, so
we take instead a large partial sum
wN= I Rk¢.
Then we have
wN-RwN=¢-R N¢
and RN is a sum of terms each containing a factor I~I-k for some
k ~ N. The preceding equation means that
tp(x, D)(e-i<x'~>WN(X)/Pm(x, ~)) =e-i<X.~>(¢ _RN ¢).
Hence
(8.3.6) q;u(O=u(e-i( .. ~> RN ¢) +f(e-i<··~>wN/Pm(.' m, ~E V.
If the distribution u is of order p. in a neighborhood of supp ¢ then

the first term on the right-hand side of (8.3.6) can be estimated by
C I supID"(e-i("~>RN¢)I~CNI~II'-N, 1~1~1.
I"I~I'
Here N - p. is as large as we please. If (xo, ~o)¢ WF(f), it follows from

(8.1.3)' that we can choose the neighborhood U of Xo and the conic
neighborhood V of ~ ° such that for some integer M and k = 1,2, ...
supl~lkliffml~Ck I sup ID"t/t1 , t/tEC~(U).
V 1"I~k+M
Taking ljJ =wN/Pm("~) we conclude that the second term on the right
hand side of (8.3.6) is O@ -k) as ~ -+ 00 in V. Hence
¢U@=O(I~I-k), ~EV, k= 1,2, ...

which proves Theorem 8.3.1.
Theorem 8.3.1 allows us to complete the proof that the wave front
set of A* X<;-n)/2 (defined in Theorem 6.2.1) is given by (8.2.9) at 0
when n+ (n-.:) is even. In fact, we know already that (8.2.9) is a lower
bound, and since B(D) A * X<;-n)/2 =0 when n+ (nJ is even it follows
that (8.2.9) is also an upper bound.
With the notation in Theorem 6.2.1 we note that
for a suitable choice of c+, c_. If we write ~ =2tAx, that is, x

=(2t)-1 B~ in (8.2.7) we have
(8.3.7)
The difference E+ -E_ satisfies the equation B(D)(E+ -E_)=O and
In fact, this follows from (8.3.7) when x =1= 0, for the two terms have
disjoint wave front sets then, and at 0 we can use the argument just
given for X<;-n)/2. By a translation we obtain solutions of the equation
B(D) u = 0 with WF(u) containing any desired point in Char B. Howev-
er, not every subset of Char B can be the wave front set of a solution.
To prove this we first take vEtC'(1Rn ) and set B(D)v=g. Then
so (8.2.16) and (8.3.7) give

WF(v)c WF(g)u {(x + tB'(~), ~);
(x, ~)EWF(g), t >0, ~ =1=0, B(~) =O}.

Using E _ instead of E + gives the same inclusion with t < 0 instead. If
(x, ~)EWF(v)' WF(g)
it follows from Theorem 8.3.1 that B(~) = 0, and the preceding in-
clusions show that one can find t _ and t + so that t _ < 0 < t + and
(x-t± B'(~), ~)EWF(g). This leads to
Theorem 8.3.3. Let B be a real non-singular quadratic form in 1R.n, let X

be an open set in 1R.n and UE.@'(X) a solution of the equation B(D)u=f
If (x, ~)EWF(u), WF(f) then B(~)=O and

Ix {~} c WF(u)
if I c X is a line segment containing x with direction B' (~) such that
I x {~} does not meet WF(f).
Thus singularities of u with frequency ~ propagate with fixed

frequency in the direction B'(~) in X until they meet the singularities
of f
Proof That B@=O follows from Theorem 8.3.1. Choose ¢EC'g'(X) so

that ¢(x) = 1 and Ln supp ¢ cI if L is the line through I. Then v
=¢uEtf' and
B(D) v =¢B(D)u+ w=¢f +w
where suppwcsuppd¢. Since
(L x {~})n WF(B(D)v)=(L x {~})n WF(w)
it follows from the discussion preceding the statement of the theorem
that there are points z ± EL on either side of x such that
(z±, ~)EWF(w), hence
z±ELnsuppd¢ and (z±, ~)EWF(u).

If y + and y _ are arbitrary points in the interior of I on different sides
of x we can choose ¢ so that Ln supp d¢ is as close to {y +, Y _} as we
wish. Hence (y±, ~)EWF(u) which proves the theorem.
In a moment we shall prove that Theorem 8.3.3 is valid for much

more general differential operators with constant coefficients although
we do not have quite so explicit fundamental solutions to work with
then. However, we give first an example of a solution of the wave
equation in IR. 4 which indicates that Theorem 8.3.3 gives all con-
ditions which WF(u) must satisfy when B(D)u=O.
Example 8.3.4. There exists a solution UE.@'(lR.4) of the wave equation

o u=(c- 2 8 2 /8t 2 -Ax) u =0
such that for a given y with Iyl = 1
WF(u) = {(t, cty; sc, -sY), tEIR., s*O}.
To construct u we change notation and let E +, E _ be the advanced
and retarded fundamental solutions (see Section 6.2). These are pro-
portional to (j(C 2 t 2 _lxI 2) when t~O, so for the solution Eo=E+ -E_
of OEo=O we have by Example 8.2.5, Theorem 8.2.4 and Theorem

8.3.1
WF(Eo)c {(t, ctx, sc, -sx); tE1R, SE1R . . . . 0, xE1R 3, Ixl = 1}.
Let f be a positive C~ density on the line L through 0 with direction
(l,cy), and set u=Eo*f Example 8.2.5 gives
WF(f)c {(t, cty;" ~); tE1R, ,+c<y, 0 =O}.
Now the tangent plane ,+ c<y, 0 = 0 of the characteristic cone ,2
_c21~12=0 meets the cone only when ("~) is proportional to (c, -y)
so (8.2.16) gives
(8.3.8) WF(u)c {(t, cty; sc, -sy); tE1R, S9=O}.
By (4.2.2) we have {(t,x); ct=<x,y)}nsuppucL. If t is so large that
u = f *E + in a neighborhood of (t, ct Y)EL it follows from (6.2.7) that
the total mass of the measure u at distance ;;;;b from (t,cty) is at least
Cb 2 for some C>O. Hence (t,ctY)Esingsuppu. Since u is real valued
the wave front set is symmetric with respect to the origin in the
frequency variable. Hence there is equality in (8.3.8) for a missing
point would make the left-hand side empty by Theorem 8.3.3.
By another convolution it would be easy to construct a solution
with "one half' of the wave front set above. We leave this for the
reader since a general result of this kind will be proved below (Theo-
rem 8.3.8).
We shall now extend Theorem 8.3.3 to general differential oper-

ators with constant real coefficients and non-singular characteristic
set:
Definition 8.3.5. A differential operator P(D) with constant coefficients

in 1Rn is said to be of real principal type if the principal symbol Pm is
real and
(8.3.9)
Since P~(~)=O implies mPm(~)=<P~(~),O=O it would be sufficient

to assume (8.3.9) when Pm(~)=O.
By Theorem 7.1.23 the Fourier transform of the fundamental so-
lution E± of B(D) used in the proof of Theorem 8.3.3 is (B(~HiO)-l.
According to Lemma 6.2.2 this is the limit as B ---+ =+= 0 of B( ~
+iBV(~))-l if v is a vector field with <B'(~), v(~)>O. We shall take
this as a guide for the construction of fundamental solutions in the
general case, but the presence of lower order terms will force us to go
a finite distance into the complex domain. With P of real principal
type we set
v(~) =P~(~)I~II -m.
This vector field is homogeneous of degree 0 with respect to ~. In the

following lemma we give a lower bound for P in the direction iv(~)
from ~.
Lemma 8.3.6. There exist positive constants t, C I , C 2 , C 3 such that

(8.3.10) ImP(~+itv(~)+iV)
~ C I (1 +1~lr-1 +<P~(~), V) - C 2 (1V1 + l)IVI(I~1 + IVlr- 2
if ~ERn,I~I~C3,vE]Rn.
Proof Taylor's formula gives
ImP(~ +itv(~)+ iV)~ <P~(~), tv@+ V) - C(t+ !v1)2(1~1 + t+!Vlr- 2

- C(I~I +t+ !VIr-I.
Here we have by (8.3.9) for some c>O
<P~@, v(~) ~cl~lm-l.
If t is fixed so that tc>C, we obtain (8.3.10) when V=O; an obvious

estimate of the terms involving V gives (8.3.10) in general.
Theorem 8.3.7. If P(D) is of real principal type then one can find
E±E~I(Rn) and W± Ecoo(Rn) such that P(D)E± =<>+W± and
(8.3.11) WF(E±)c{(tP~(~), ~); t~O,Pm(~)=O, ~*O}uTo*""'O.
Proof It is no restriction to assume that m > n + 1, for if E ± has the

properties stated in the theorem for ,1kP(D), then ,1kE± has these
properties for P(D). (,1 is the Laplace operator.) Replacing P by - P
interchanges E + and E _ so it suffices to construct E _. Guided by the
second order case as indicated above we let r be the chain
(8.3.12)
where C 3 and t are given by Lemma 8.3.6. Noting that by (8.3.10)
ImPm~ C 1(1 +IRe (It+ 1 on r we set
(8.3.13) E_(x)=(2n)-n Jei (x,OjP(()d(1 A ... Ad(n' xERn.

r
In terms of the parameters ~ l' ... , ~ n on r we have explicitly
d(l A ... Ad(n=Jd~l A ... Ad~n'
J =D(~ 1 + itV 1(~), ... , ~n + itvn(~))fD(~ l' ... , ~n) --+ 1 at 00,
so the integral (8.3.13) is locally absolutely and uniformly convergent.

When ¢EC't' we have with l/t=P( -D)¢
(P(D)E _, ¢) = (E_, l/t) =(2n)-n JJ l/t(x)ei(x. O /P(()dxd(l A ... A d(n.
T
Here we integrate first with respect to x and use that t/I(-()=P(()$(-Q.

Now F(() d( 1 A .•. A d(n is a closed differential form for every
analytic function F, since dF is a linear combination of d( l ' ... , d(n.
Thus $( - ()d( 1 A ... A d(n is a closed differential form which decreases
rapidly at infinity, so Stokes' formula gives
(P(D)E _, ¢) =(2n)-n S $( - ()d(l A ..• A d(n
T
=(2n)-n S $( - ~)d~ -(2n)-n S $( - ()d( 1 A .•. A d(".

To
Here ro is the chain
1R.n3~~~+itvo(~), 1~I~C3'
where Vo is a smooth extension of v from I~I = C3 to I~I ~ C 3, which

makes ruro homotopic to 1R.". Thus P(D)E_ =b+w_ where
w_(x)= -(2n)-n S ei(x' O d(l A ..• Ad(n
To
is an entire analytic function since ro is compact. Hence Theorem

8.3.1 shows that Pm@=O if (x, ~)EWF(E_) and x*O.
To complete the proof we must show that (xo, ~o)¢WF(E_) if
xo¢1R._P~(~o). This condition means precisely that we can find VE1R."
with IVI = 1 and
(8.3.14)
Choosing a conic neighborhood W of ~ 0 such that for some c > 0
(P~(~), V»cl~lm-l, ~EW,
we obtain from (8.3.10) when ~EW
1m P(~ + itv(~) + is V)
~ C1 (1 +1~lt-l +cl~lm-ls_ Cz(s+ l)s(l~1 +s)m-Z
~Cl(I+I~l)m-l
if 0 < s< eI~ I and I~ I is large enough. Replacing V by e V we have

(8.3.15) 1m P(~ + itv(~) + is V)~ C1 (1 + 1~lt-l,
~EW, O~s~I~I, I~I ~ C~.
Choose XEC (1R." . . . . 0) homogeneous of degree 0 with support in W

OO
so that 0 ~ X ~ 1, and X= 1 in a conic neighborhood Wo of ~ o. If

(x, V) >0, which is true in a neighborhood of x o, we obtain using

Stokes' formula
(8.3.13)' E_(x)=(2n)-n J ei(x,P/P«()d(lA ... Ad(n
r'urQ
where r' is the chain

JRn3~ -> ~ + itv(~) + iI~1 X(~) V, I~I ~ C'3'
and r~ is the union of the part of r where C 3 < I~I < C'3 with the chain
{(~, s); I~I = C'3' O<s< C~} -> ~ + itv(~) + isX(~)V
with suitable orientations. The contribution to (8.3.13)' when (Er~ or
Re (EWo is an analytic function of x when (x, V) >0. If M is a
measurable conic set contained in a closed proper convex cone G,
then the wave front set of the function
x-> J ei(x,o/P«()d(lA ... Ad(n' (x,V»O,
{Er', Re{EM
is contained in {(x,~); (x,V»O, ~EG}. This follows from Theorem

8.1.6. In fact, replacing x by z = x + i Y we obtain a bounded analytic
function when Ixl is bounded, (x, V) > 0, and y is in the interior of the
dual cone of G, for
Re i(z, 0= -(x, im O-(y, ReO~ -t(x, v(~»< Clxl.
We can cover CWo with a finite number of such cones G which do not
contain ~o, so it follows that (xo, ~o)¢ WF(E _). The proof is complete.
Repetition of the proof of Theorem 8.3.3 gives now
Theorem 8.3.3/. Let P(D) be of real principal type. If uEEd/(X), P(D)u

= f and (x, ~)EWF(u),- WF(f), then Pm(~)=O and
I x {O eWF(u)
if I eX is a line segment containing x with direction l!:m such that
I x {~} does not meet WF(f).
Finally we shall give a general version of Example 8.3.4.
Theorem 8.3.8. Let P(D) be of real principal type, O=!=~EJRn and Pm(~)
=0. Then one can find uEcm (1R.n) such that P(D)UEcoo(JRn) and
(8.3.16) WF(u)={(tP~(~),s~); tEJR, s>O}.
Proof Set L=JRP~(~) and let fF be the set of all UEcm(JRn) with
pUEc oo (1R.n), uECOO(CL) and WF(u)eJRn x (1R.+ ~). The theorem states
that there is an element uEff which is not in Coo, for uEff implies
WF(u)cIRP~ x IR+ ~
and by Theorem 8.3.3' UEC OO if the inclusion is strict. Now ff is a
Frechet space with the seminorms
(i) supID"ul, 1Q(I~m, K a compact subset ofIR",
K
(ii) sup ID"ul, Q( arbitrary, K a compact subset of CL,
K
(iii) sup ID" P(D)ul, Q( arbitrary, K a compact subset of IR",
K ...-....
(iv) supl'1I N I¢u('1)I, N=l, 2, ... , ¢EC(f(IR").
erN
Here rN is a sequence of conic neighborhoods of ~ in IR" shrinking to
IR+~. We need only use a countable number of compact sets K and
functions ¢ since the semi-norms (iv) can be estimated by the corre-
sponding ones with ¢ replaced by a function t/I which is 1 in supp ¢.
(See the proof of Lemma 8.1.1.) The proof of completeness is an
exercise for the reader. If ff c cm+ 1 then the closed graph theorem
shows that the inclusion ff "-+ cm + 1 is continuous. Thus one can find
N, ¢EC(f(IR"), Kl ~IR" and K2~C L so that
(8.3.17) I ID"u(O)1 ~ C { I sup ID"ul + I sup ID"ul

l"l=m+ 1 1,,1 ~m K, 1,,1 ~N K2
+ I sup ID" P(D)ul + sup (1 + 1'1lt Iqm('1)I, uEff.

I"I~N K, erN
To show that (8.3.17) is not valid we need to construct approxi-
mate solutions of the equation Pu =0 concentrated close to L, thus
away from K 2 • To make the last term small the Fourier transform of
u should be concentrated close to the direction ~. It is therefore
natural to set for t > 0
Then
P(D)ut(x) = eit(x,~) P(D + t~)Vt(X)
=('"-1 eit(x,O (t p~j)(~)DjVt + Pm - 1(~)Vt + ... )
where terms indicated by dots contain a negative power of t, and p;!)

=8j Pm • A formal solution
Vt =V O +t- 1 V1+ ...
may be found by solving the first order equation
(8.3.18) "
Lvo = I P;!)(~)Djvo + Pm - 1(~)Vo =0
1
and then successively equations

(8.3.19)
where Jj is determined by Vo, •.. , Vj-I. The support of Vo is a cylinder
with the axis in the direction ~(e); we can choose Vo with vo(O) = 1 and
support close to L by prescribing such values on a plane 1: orthogonal
to p~(e). If the other functions Vj are determined by the boundary
condition vj=O on 1:, it is clear that suppvjcsuPPVo for j'*'O. For
v/ = "L. v.t-
J
j
j<M
the third sum on the right-hand side of (8.3.17) is O(tm- I - M + N ). The

last term is rapidly decreasing when t -+ 00 since
¢U,(11)= I t- j (¢V)(11- t e)
j<M
°
and t+ 1111 ~ CI11-tel when 11 FfrN • The first sum on the right-hand side
of (8.3.17) is O(tm), the second sum is for an appropriate choice of
vo, but the left-hand side grows as tm + I since e '*' O. If we take M = N
this is a contradiction which completes the proof.
The preceding argument can of course be given a more con-

structive look by summing a very lacunary sequence of the functions
u/ as in the proof of Theorem 8.1.4. However, making the proof of the
closed graph theorem explicit in this way tends to hide the idea of the
proof. - In Chapter X we shall prove that the equation P(D) v = f has
a solution VEC""(JRn) for every fEC""(JRn). If we take f =P(D)u and
replace u by u-v we obtain a solution of the homogeneous equation
P(D)u=O satisfying (8.3.16).
The fact that we have restricted ourselves to discussing operators
of real principal type with constant coefficients does not mean that
results such as Theorem 8.3.3' and Theorem 8.3.8 are not available
when the coefficients are variable. We shall return to these matters
in Chapter XXVI after introducing the appropriate tools.
8.4. The Wave Front Set with Respect to CL

Let Lk be an increasing sequence of positive numbers such that La = 1
and
(8.4.1)
8.4. The Wave Front Set with Respect to CL 281
for some constant C. If X e 1Rn is an open set we shall denote by

CL(X) the set of all UECOO(X) such that for every compact set K eX
there is a constant C K with
(8.4.2) ID"u(x)1 ~ CdCKLI"I)I"I, xEK,
for all multi-indices tx. (This notation, used from now on, differs
slightly from the standard one used in Section 1.3.) When Lk=k+ 1
this means that CL(X) is the set of real analytic functions in X. The
class C L with L k =(k+1)a, a>l, is called the Gevrey class of order a.
It occurs quite frequently in the theory of partial differential equa-
tions.
Proposition 8.4.1. CL(X) is a ring which is closed under differentiation. If

f: y -+ X is an analytic map from the open set Y e1Rm to the open set
X e1R n, then composition with f defines a map f*: CL(X) -+ CL(y).
Proof Since Lk is increasing we obtain by Leibniz' rule

sup ID"(uv)1 ~ Ci(2C K L I"I)I"1
K
if u and v satisfy (8.4.2). Thus C L is a ring. That CL is closed under

differentiation follows from the inequality
(8.4.3) (Lj+l)j+l~(CLy+l~C2j+1JJj
which is a consequence of the second part of (8.4.1). To prove the last

statement we note that the derivatives of f*u at y of order k are the
same as the derivatives of
z -+ L (D"u) (f(y)) (if(z) - if(y))"/tx!
1,,1 ;:ik
when z = y. The right-hand side is an analytic function of z when YEK
and z is complex with Iy-zl <r sufficiently small. It can then be
estimated by
k
C L (CLk)I"I(C'r)I"I/tx!=CL(nCC'rLkY/j!.
l"l;:ik 0
Now JJk/j! ~ Vk/k! by the first part of (8.4.1), so this sum can be
estimated by
k
CL\jk! L (nCC'r)j <2CVk/k!
o
if r ~ 1/(2 n CC'). By Cauchy's inequalities we conclude that
ID"f*u(Y)I~2Cr-kVk if Itxl=k, YEK,
Proposition 8.4.1 shows that we can define CL(X) by means of

local coordinate systems when X is a real analytic manifold. (This
means that an atlas for X is given such that the maps (6.3.1) are all
real analytic.)
For any distribution UE.@'(X) we define sing SUPPL U to be the
smallest closed subset of X such that U is in CL in the complement.
(When CL is the real analytic class we shall use the notation
sing SUPPA u.) The purpose of this section is to show how one can
make a spectral analysis of this set parallel to Sections 8.1 and 8.2. A
new difficulty occurs when
(8.4.4)
for then the class CL is quasi-analytic by the Denjoy-Carleman theo-

rem (Theorem 1.3.8) so one cannot choose cutoff functions in C L •
(Multiplication by Coo functions not in CL may of course increase
sing SUPPL u.) However, this difficulty can be circumvented by using
Theorem 1.4.2 to choose test functions with adequate bounds for
derivatives up to a certain order only. This leads to a description of
sing SUPPL in terms of Fourier transforms:
Proposition 8.4.2. Let XoEX dR" and UE.@'(X). Then UEC L in a neigh-
borhood of Xo if and only if for some neighborhood U of Xo there is a
bounded sequence UNEC'(X) which is equal to U in U and satisfies
(8.4.5)
for some constant C.
Proof a) Necessity. Let UEC L when Ix-xol<3r and choose XNEC'O so

that XN=l when Ix-xol<r, XN=O when Ix-xol>2r and
(8.4.6)
where C1 does not depend on N. This is possible by Theorem 1.4.2
with dk =r/2N for k~N. With UN=XNU we obtain from (8.4.2) and
(8.4.6) since N~LN and LIIXI~LN when lexl~N
IDlXuNI ~ CK(C K+ C1)N LNN, lexl =N.
Hence
lelXuN(e)1 ~ C~+ 1 J!iN, lexl =N,
and since lei ~nt max lejl this proves (8.4.5).
b) Sufficiency. In U we have
(8.4.7)
when N = I0(1 + n + 1, for ~~UN(~) is then integrable by (8.4.5). That UN is

bounded in C' implies by the Banach-Steinhaus theorem that
IUN(~)I~C(l+I~I)M, N=1,2, ...
where C and M are independent of N. (See the proof of Theorem
2.1.8.) We use this to estimate the integrand in (8.4.7) when I~I <LN
but use the estimate (8.4.5) when I~I > LN. This gives that in U
ID~u(x)1 ~ C3(C3LI~I+'+ l)I~I+'+M
and repeated use of (8.4.3) shows now that UECL(U).
Proposition 8.4.2 suggests the following definition:
Definition 8.4.3. If X dR" and UE~'(X) we denote by WFdu) the
complement in X x (lR"' 0) of the set of (x o, ~o) such that there is a
neighborhood U cX of x o, a conic neighborhood r of ~o and a
bounded sequence UNEC'(X) which is equal to U in U and satisfies
(8.4.5) when ~Er. When CL is the analytic class we use the notation
WFA(U).
By definition WFdu) is a closed subset of X x (1R', 0). The follow-
ing lemma shows that UN can always be chosen as products of u and
suitable cutoff functions, obtained by regularizing those in the proof
of Proposition 8.4.2.
Lemma 8.4.4. Let UE~'(X) and let K be a compact subset of X, F a

closed cone c1R' such that WFL(u)n(KxF)=0. If XNEC~(K) and for
all 0(
(8.4.6)'
it follows that XNU is bounded in C,M if U is of order M in a neigh-
borhood of K, and we have
(8.4.5)'
Proof The bounded ness of XNU is obvious since XN is bounded in C~.

Let xoEK, ~oEF' {a} and choose U, r, UN according to Definition
8.4.3. IfsuPPXNcU then XNU=XNUN. By hypothesis UN satisfies (8.4.5)
in r, and luN ( ~)I ~ C(1 + I~ I)M for fixed C, M ~ 0. From (8.4.6)' it
follows that
IXN(l1)1 ~ CN+ 1 (LN/(I 11 I + LN)t (1 + 1111)-·-l-M.
If we apply (8.1.3)' with v=u N and l/>=XN' the estimate (8.4.5)' is
proved if F is replaced by a closed conic neighborhood of ~o con-
tained in the interior of r apart from the origin. Since F can be
covered by a finite number of such neighborhoods it follows that

(8.4.5)' is valid if supp XN C V for a sufficiently small neighborhood V
of Xo' We can cover K by such neighborhoods Vj, j=l, ... ,J, and
choose XN,jEC'~(V) so that I XN,j= 1 in K and XN,j satisfies (8.4.6)' for
j = 1, ... , J. To do so we just have to regularize any partition of unity
by a function t/I N chosen according to Theorem 1.4.2 so that it satis-
fies (8.4.6). If XNEC~(K) satisfies (8.4.6)' it is clear that XN,jXN also
satisfies (8.4.6)' with some other constants. Hence (8.4.5)' is valid with
XN replaced by XN,jXN' Since I XN,jXN = XN this completes the proof.
Lemma 8.4.4 and Proposition 8.4.2 give easily
Theorem 8.4.5. The projection of WFL(u) in X is equal to singsuPPLu

if uE!0'(X).
Proof a) If UEC L in a neighborhood of it follows from Proposition
Xo
8.4.2 that (xo,~o)¢WFdu), b) Assume that (xo,~o)
~oElR.n,,{o}.
¢WFL(U) for all ~oElR.n,,{o}. Then we can choose a compact
neighborhood K of Xo so that WFL(u)n(K xlR.n)=0. By Lemma 8.4.4
there is a sequence XNEC~(K) which is equal to 1 in a neighborhood
V of Xo such that XNU is bounded in Iff' and satisfies (8.4.5). Hence
Xo ¢sing SUPPL u by Proposition 8.4.2.
The condition (8.4.6)' is satisfied by any fixed function in CL with
support in K. If C L is non-quasianalytic we can therefore simplify
Definition 8.4.3 to the existence of a fixed distribution v which is
equal to u in a neighborhood of Xo and has a Fourier transform
satisfying (8.4.5) in a conic neighborhood of ~ o' This is parallel to
Definition 8.1.2, so we obtain
Theorem 8.4.6. For all U and L we have WF(u) C WFL(u) C WFA (u);
moreover, if £j-:i:.£j then WFL"(u)c WFdu).
The conditions (8.4.6)' remain valid if we multiply all XN by the

same function in CL(X). This gives
Theorem 8.4.7. WFdau)c WFL(u) if aECL(X) and uE!0'(X).

It is obvious that
WFL(au/axj)c WFL(u)
for (8.4.5) implies in view of (8.4.3)
l~jUN+ 1 (~)I-:i:. CI~I(CLN+ dl~lt+ 1 -:i:. C'(C' LN/lelt.
If we combine this with Theorem 8.4.7 we obtain
if UE~'(X) and
P(x, D)u = L aa(x)Da
is a differential operator with coefficients in CL(X).

We could now proceed to study WFL(u) using arguments com-
pletely analogous to those in Sections 8.1 to 8.3. However, to avoid
such boring repetitions we shall use an alternative approach which
has the advantage that it is also applicable in the study of hyper-
functions in Chapter IX. The first step is just an improvement of
Theorem 8.1.6.
Theorem 8.4.8. If the hypotheses of Theorem 3.1.15 are fulfilled, then

(8.4.8)
where rO is the dual cone of r.
Proof Let X 1 ~ X 0 ~ X be open sets and choose using Theorem 1.4.2 a
sequence ¢vEC'O(Xo) with ¢v= 1 in Xl such that
IDa¢vl~(Cl(V+1))lal, lal~v+1.
As in (3.1.18) we set
qJv(x, y)= L aa¢v(x)(iyna!
lal ~v
and have by (8.1.15) for a fixed YEr with IYI<y
q;);;(~)= JqJv(x, Y)f(x+iY)e-i<x+iY,Odx
+(v+ 1) H f(x + it Y)e-i<X+itY,~) L aa¢.(x)(i yna! tVdxdt.
0<,<1 1/XI~v+l
For .u~v+1 we have with IYl 1 = L Ilil
I L a/X¢v(x)(iY)/X/a!I~C\'(v+1)I'IYli/.u!
I/XI~I'
so it follows with C 2 = 2 ectlYIt that
IqJv(x, Y)I~C~+1, l(v+1) L a/X¢v(x)(iY)/X/a!l~C~+l.

lal~v+ 1
The estimate (8.1.16) is therefore replaced by

-----
I¢Jo(~)1 ~ C;+l(e<Y,O +(v-N)! (- y, ON-v-l), (Y, 0 <0.
Set fv=<PN+v-JO' When <Y,O< -cl~1 for a fixed c we obtain for

some C 4
(8.4.9)
fore-cl~1 ~ v! (cl~l)-v.
Now f oE2lJIN+1 by Theorem 3.1.15. If we choose the sequence <Pv
bounded in C~+1 it follows that 1. is bounded in C'N+1. Thus we
have proved that
WFA(fo)eX x {~; <Y, O~O},
and since Y has an arbitrary direction in r, this proves the theorem.
We shall now show that it is possible to associate with any
distribution uEY"(1Rn) an analytic function U in the convex tube
Q={ZEa::: n ; IImzl< I}
(Euclidean norm) so that
(8.4.10) U= S U(. +iw)dw.
1"'1= 1
The interest of such a decomposition is clear if we note that for the
boundary value U(. +iw)=lim U(. +irw) we must have
r/' 1
WFA(U(. + iw))c1Rn x 1R_ w,

by Theorem 8.4.8, so one should be able to determine if (x, -W)E
WFA(u) by just looking at the behavior of U near x+iw.
Assuming that U is well behaved at infinity we denote by O(~, y)
the Fourier transform of U(x+iy) with respect to x and observe that
the Cauchy-Riemann equations 8Uj8x j +i8Uj8Yj=O give
~jO + 8 Oj8Yj =0, j=l, ... , n.
Thus O(~, y)= Uo(~)e-<Y'~>, and (8.4.10) leads to the condition u(~)
=I(~)Uo(~),
(8.4.11) I(~)= S e-<""~>dw.

1"'1= 1
When n=l we have I(~)=2cosh~, and when n>l we have I(~)
=Io«~, Of)
1
(8.4.11 )' I O(p)=C n _ 1 S (1_t 2 )(n-3)/2 etP dt
-1
where cn - 1 is the area of sn-2. This is a multiple of the Bessel

function J(n_2)/2(ip) divided by p(n-2)/2 so the asymptotic behavior is
well known. For the convenience of the reader we give a direct proof
of what we need.
Lemma 8.4.9. 10 is an even analytic function in CC such that for every

e>O
(S.4.12) I o(P) = (2n)(n-1)/2 e" P- (n -1)/ 2(1 + 0 (1/ p))
if P -+00, largpi <n/2 -e.
There is a constant C such that for all PECC
(S.4.12)' II o(p)1 ~ C(1 + Ipl)-(n-1)/2 e IRe pl .
Proof. Since Io('p)=Io(p) we may assume that 0~argp<n/2 when

proving (S.4.12) and (S.4.12)'. To prove (S.4.12) we take the integration
from -1 to + 1 in (S.4.11)' along the short sides of a right triangle
with hypothenuse (-1,1) and (t-1)p <0 on the side through 1. Writ-
ing 1-t 2 =(1-t)(1+t) and taking s=(l-t)p as integration variable
on this side we obtain since 1 + t = 2 - s/P
J(2S)(n-3)/2 e -s ds + 0(1/ p).

00
e- P I o(p)p(n-1)/2 = cn _ 1
o
(S.4.12) follows in view of (3.4.2). If we integrate along the half lines
(± 1 - t) p < 0 instead, the estimate (S.4.12)' is obtained.
Now we introduce the function K correspondending to decom-
position of u = 8,
(S.4.13)
It follows at once from (S.4.12) that the integral converges and defines
an analytic function in Q.
Lemma 8.4.10. K(z) is an analytic function in the connected open set

Q={ZECC n ; (z,z)¢(-oo, -l]}::::>Q.
For any closed cone r cO such that (z, z) is never ~O when ZEr-..... {OJ
there is some c>O such that K(z)=0(e- c1zl ) when z-+oo in We have r.
for real x and y
(S.4.14) IK(x + iy)1 ~K(iy)
=(n -1)! (2n)-n(1-lyl)-n(l + O(l-lyl)), Iyl)" 1.
Proof Introducing polar coordinates we obtain K(z)=Ko«z,z)t)

where Ko is the even analytic function
Ko(w) =(2n)-nII exp (ipW1 w)/I o(p) pn-1 dwdp
00
=(2n)-n JIo(ipw)/Io(p)pn- 1dp, IImwl<l, WECC.

o
When W= -it, 1/2<t<1 then

()()
Ko( -it)=(2n)-n S Io(pt)/Io(p)pn-1dp

o
()()
=(2n)-n S e- p(I-t)(1 +O(l/(p + 1»)t-(n-l)/2 pn-1dp

o
=(n-1)! (2n)-n(1-t)-n(1 +O(l-t»
which proves (8.4.14). To study the analyticity of K we note that if w
= it, -1 < t < 1, then Cauchy's integral formula gives
()()
Ko(w) = (2n)-n S Io(ipw(l + is})/I o(p(1 + is})(l + ist pn-l dp

o
for any real s. In fact, 10 +0 in the right half plane since the Bessel
function J. has only real zeros when v> -1 (see Hurwitz [1]). The
right-hand side is an analytic function K.(w) in the set
Zs={w; IRe(iw(l+is»I<l}={w; IsRew+Imwl<l}.
Z s is a strip with boundary lines passing through ± i and slope - s.
We have Ks(w}=Ks'(w} in Zs"Zs' since this is a convex set contain-
ing an interval on the imaginary axis where we know that the equality
is valid. Hence the functions Ks define together an analytic extension
of Ko to <C,-{it;tE1R., Itl~l}, so K is analytic in Q. The set Q is
connected for the component of1R.n contains all x+iy with (x,y)+O
since X+itYEQ when O~t~l, and Q is contained in the closure of
this set.
It remains to prove that Ko(w) is exponentially decreasing when
Iwl<C Rew. When Rew>l we have
IKo(w)1 = i(2n)-n
()()
I I o(iplwI2}/I o(pw}wnpn-l dPI
~ C1 S e- pReW «l + plwl)/(l + plwI 2 »(n-l)/2Iwl npn- 1dp ~ C 2 lwl n •

o
Furthermore, Ko is exponentially decreasing on 1R. since
K(x)e<x'~)=(2n)-n Sei<x.O/I(~+irf)d~
is bounded for xE1R.n and small 1111. Hence the Phragmen-Lindelof
theorem gives IKo(w}1 ~ C3e-CRewlwln if Rew> 1. The proof is com-
plete.
Theorem 8.4.11. If uE9'I/(1R.n} and U =K*u, where K is defined by
(8.4.13), then U is analytic in Q= {z; 11m zl < 1} and for some C, a, b
(8.4.15)
The boundary values U (. + iw) are continuous functions of WES' - I

with values in 51" (JR"), and
(8.4.16) <u,cP)=S <U(. +iw),cP)dw, cPE5I'.
Conversely, if U is given satisfying (8.4.15) then (8.4.16) defines a
distribution UE5I" with U=K*u. We have for any L
(8.4.17) (JR" x S·-I) n WFL(u)
={(x,w); Iwl=l, U is not in C L at x-iw},
and an analogous description is valid for WF(u).
That U is in C L at x - iw means of course that for some neigh-

borhood V of x-iw and some constant C we have
Iaaz U(z)1 <C
=
I + lal Dill
lal if ZEV and IImzl < 1.
For the real analytic class this means that U can be continued
analytically to a full neighborhood of x - iw.
Proof of Theorem 8.4.11. That UE!I" means that for some a, b

IU(cP)1 ~ C I sup IxaDPcPl, cP EY.
lal ;;ia, 11I1;;ib
By Cauchy's inequalities and (8.4.14) we have for every f3
IDP K(x + iy)1 ~ C p (1-lyl)-·-IPl e -c lx l, Iyl < 1,
which gives (8.4.15) with b replaced by b + n. The Fourier transform of
U(x + iy) is e-<Y,012(~)/I(~), hence continuous when Iyl ~ 1 with values
in Y', and we obtain (8.4.16) by the definition of I.
Conversely, assume given an analytic function U satisfying (8.4.15).
Set Uy=U(. +iy) when IYI<1. By the proof of Theorem 3.1.15 the
limit Uw of Uy when y -4 w, Iwl = 1, exists in Y'. If b is an integer ;;;; 0
then
Uw(cP) = S U(x) I allcP(X)( -iw)P/f3!dx
IPI ;;ib
+(b+1) IS U(x+i(l-t)w) I afl cP(x)(-iw)P/f3!tb dxdt,
0<1<1 IPI=b+1
Thus Uw is a continuous function of w with values in Y', and

IUw(cP)I~C I S(1+lxl)al aP cP(x)ldx.
IPI ;;ib+ I
It follows that (8.4.16) defines a distribution uE5I" with a similar
estimate. We have Ow=e-<"w> 00' hence 12=100, so U =u*K as
claimed.
To prove (8.4.17) we first assume that (x o, Wohf W Fdu) and that

IW ol=1. We must show that U=K*UEC L at xo-iwo. By hypothesis
we can find r>O, a bounded sequence uNEtC'(JR") equal to U when
Ix-xol<r and a conic neighborhood r of Wo such that
(8.4.18)
Recall that the boundedness of UN implies that for fixed C and M~O
(8.4.19)
Set U=UN+V N. Then U=K*uN+K*v N and

K*vN(z)=(K(z- .), VN)·
Now K(x+iy-t) is well defined when lyI2<1+lx-tI 2, so it is well

defined and rapidly decreasing with all derivatives when It-xol~r if
(8.4.20)
It follows that K *VN is analytic and uniformly bounded in compact

subsets of the set defined by (8.4.20), which is a neighborhood of Xo
-iwo·
The Fourier transform of K*u N(. +iy) is e-<y·ouN(~)/I(~), so
(8.4.12) gives
ID~ K *uN(x + iy)1 ~ C Se-<Y.~>-I~I(l + IW("-1)/21~~llaN(~)ld~.
Using (8.4.18) we can estimate the integral when ~Er and I~I>LN by
CN+IL~ S 1~11"1+("-1)/2-Nd~~C~+1I!j. if locl~N-2n,lyl<1.
I~I>LN
The integral when ~Er and I~I~LN has the bound

cIJ~I+("-1)/2+M+"< C N+ 1LN loci ::::;;_N -M -2n,
N = 1 N'
in view of (8.4.19). It remains to examine

C S e-<Y.~>-I~I(l +1~\)M+(n-l)/21~11"ld~.
~¢T
Choose e>O so that (w o, 0 «1-2e)I~1 when ~rtr. Then

(8.4.21) -(y, 0 -I~I < -el~1 if ~rtr and Iy+wol <e.
Hence we obtain if N = loci + 2n + M

C S e-<Y.~>-I~I(l +1~I)M+(n-l)/21~P"ld~ ~ C~N! ~ C~L~,
~¢T
8.4. The Wave Front Set with Respect to C L 291
Summing up, it follows by repeated use of (8.4.3) that

IDa U(z)1 < clal+2n+M+ 1Dal+ 2n + M < Clal+1 Dal
= 1 lal+2n+M= 2 lal
in the intersection of Q and a neighborhood of xo-iwo.
The other inclusion in (8.4.17) follows from the next lemma.
Lemma 8.4.12. Let dJ.L be a measure on sn-1 and r an open convex

cone such that
(y,W)<O when O*YEf, wEsuppdJ.L.
If U is analytic in Q and satisfies (8.4.15) then
F(z) = f U(z + iw) dJ.L(w)
is analytic and IF(z)I~C'(l+IRezl)allmzl-b when ImZEr and Ilmzl is
small enough. For every measure dJ.L on sn-1 we have
(8.4.22) WFL(UI')c:{(x,e); -e/leIESUppdJ.L and U¢C L at x-ie/leD,
(8.4.22)' WF(Up)c:{(x,e); -e/leIESUppdJ.L and U¢C oo at x-ie/leD.
Here Up= SU(. +iw)dJl(w).
Proof If wEsuppdJl and ImzEr, we have for some c>O

Ilm(z+iwW = 1 +2(w, 1m z) + 11m Zl2
< 1-2cllmzl + IImzl2 < l-cllmzl
if IImzl<c. Hence l-IIm(z+iw)l>cllmzl/2 when Ilmzl<c, which
proves the first statement. By Theorem 8.4.8 it follows that
WFA (UI')c:1R" x r e
where r e is the dual cone of r, and it is obvious that
singsuPPL UI'C:{x; U is not in CL at x+iw for some wEsuppdJl}.
To prove (8.4.22) we write dJl=IdJlj where suppdJ.Lj is contained in
the intersection of suppdJ.L and a narrow open convex cone Vj. Apply-
ing the result just proved with dJ.L replaced by dJ.Lj and r replaced by
the interior of the dual cone - Vje we obtain
WFL(Up)C:U {(x, e); -e/leIEij, U¢C L at x+iw for some WEV;l.

j
If -e/lel¢suppdJl or UEC L at x-ie/lei we can choose the covering

*
so that - e/lel ¢ ij for every j or for all j 1 while U E CL at x + iw for
every WEV1 . In both cases it follows that (x, ~)¢WFL(U,J which proves
(8.4.22). The proof of (8.4.22)' is of course identical. This completes the
proof of Lemma 8.4.12 and of Theorem 8.4.11.
Corollary 8.4.13. Let r 1 , •.. , Ij be closed cones in 1Rn ....... 0 with union
1Rn. . . . O. Any uE9'''(1Rn) can then be written U= 2>j where ujEY' and
WFL(uj)e W~(u)n(1Rn x Ij).
If U= L uj is another such decomposition then uj = uj + I Ujk where
ujkEY', ujk = -u kj and k
Proof. If <P j is the characteristic function of lj . . . . (lj n(~ u ... u lj_ 1))
we have I <Pj= 1. With U =K*u and Uj=K*(uj-u) we have I 0;=0,
and it follows from Theorem 8.4.11 and Lemma 8.4.12 that we can
take
Uj = JU(. -iw) <p/w)dw,
ujk = JUj(. - iw) <Pk(w)dw - JUk(· - iw) <pj(w)dw.
Next we improve Theorem 8.1.4.
Theorem 8.4.14. If X e1Rn is open and S is a closed conic set in

X x (1R n ....... O) then one can find UE~'(X) with WF(u) = WFL(u)=S for
every L.
Proof It is sufficient to prove the statement when X =1R n, and we

only have to verify for the chosen U that
WF(u)= W~(u)=S.
Let (x k ,8k ) be a sequence without repetitions which IS dense in

{(x, 8)ES; 181 = I}. With K defined by (8.4.13) we set
00
U(z) = I 3 -k K«z - x k - i8 k)/2), 11m zl < 1.

I
Since
IK«z-x k -i8k)/2)1 ~K(i(lmz-8k)/2)~ C(l-IImzl)-n
it is clear that U is an analytic function satisfying (8.4.15). Noting that
00
I 3- k=3- N /2 and that It8k+81~t+1 ifI81=1, t>O, we obtain

N+l
IU(Xk- it8k)1 > 3 -kIK( - i(t+ 1) 8J2)1/2

-I Lr j K«Xk-Xj- i(t8k + 8j ))/2)I-HXl, t /' 1.
j<k
Hence U is not even bounded in Q near any point in

S'={(X, -0); (X,O)ES and 101=1}.
On the other hand, it is clear that U is analytic near any point in
(lRnX sn-1),-s'. By Theorem 8.4.11 this completes the proof of the
theorem.
Next we prove a converse of Theorems 8.1.6 and 8.4.8.
Theorem 8.4.15. Let UE~'(X), X dR.n, and assume that WF;:.(u)cX x r a

(resp. WF(u) cX x Fa) where r a is the dual of an open convex cone r.
If X1~X and r 1 is an open convex cone with closure cru{O}, then
one can find a function F analytic in {x+iy; XEX 1 , YEr1, Iyl<y}, such
that
and the limit of F(. +iy) when y-+O in r 1 differs from u by an element
in CL(X 1) (resp. COO (X 1)).
Proof Set v=xu where XEC~(X) is equal to 1 in Xl. If V=K*v is

defined as in Theorem 8.4.11 we have V E CL at every point in Xl
+i(sn-1 n C(_r a )). Choose McSn - 1 open with r a nsn - 1 cM and M
in the interior of rt. Then v=v 1+V2 where
v 1= S V(.+im)dm
-w¢M
belongs to CL in X 1 and v2 is the boundary value of the analytic

function
F(z)= S V(z+im)dm, ImzEr1 , IImzl<y·
-wEM
Lemma 8.4.12 completes the proof.
Remark. It follows from Corollary 8.4.13, Theorem 8.4.15 and Theo-

rem 8.4.8 that WFL(u) (resp. WF(u)) is the intersection of WFA(u -u 1)
for al.l u 1ECL (resp. COO). Thus the notions WF;:.(u) and WF(u) can be
derived from WFA(U). - For the analytic case the statement of Theo-
rem 8.4.15 can be simplified: the restriction of u to Xl is the bound-
ary value of an analytic function F with the stated properties.
Corollary 8.4.16. If UE~'(X) where X is an interval on IR. and if XoEX

is a boundary point ofsuppu, then (xo, ±l)EWFA(u).
Proof Assume for example that (x o, -1)¢ WFA(U). Then we can find F
analytic in Q={z; Imz>O, Iz-xol<r} with boundary value u. There
is an interval Ic(xo-r, xo+r) where u=O. By Theorem 3.1.12 (and

Theorem 4.4.1) F can be extended analytically across I so that F = 0
below I. Thus the uniqueness of analytic continuation gives F =0,
hence u = 0 in (xo - r, Xo + r). This contradicts that Xo is a boundary
point of suppu and proves the corollary. (See also Theorem 3.1.15 for
a more general form of the proof.)
Note that the corollary can be phrased as a uniqueness theorem:

If we know that WFA(u) does not contain Tx*(lR),O for any XEX then
u must vanish identically if u vanishes in an open set. We shall extend
this statement to several variables in the next section.
Finally we prove an analogue of Lemma 8.1.7 and of Theorem
8.1.8.
Lemma 8.4.17. If UE//' then WFA(u)clR n x F where F is the limit cone

u
of supp defined in Lemma 8.1.7.
Proof. The Fourier transform of u*K is z2(e)/I(e). If r is an open cone

with tnF={O} we can choose a closed cone F' with F,{O} in its
interior and r nF' = {O}. Then we have for some c< 1
<y, 0 ~clyllel if YEr, eEF'.
Hence Lemma 8.4.9 shows that e-<Y'~>/I@ and all its derivatives are e
e
bounded if yE - r, Iyl < 2/(1 + c), and EF'. Since supp u is contained in
the union of F' and a compact set, it follows that u(e)e-<Y'~>jI(e) is in
//' when yE - r, lyl < 2/(1 + c). By Theorem 7.4.2 it follows that u*K
has an analytic continuation to {z; 1m ZE - r, 11m zi < 2/(1 + Hence cn.
WFA(u) c Cr by Theorem 8.4.11 which proves the lemma.
Theorem 8.4.18. If UE//'(lRn) is homogeneous in lRn,O then

(8.4.23) (x, ~)EW1Hu)-=(~, -x)EWli(u) if ~ =1=0, X =1=0,
(8.4.24) XESUPPU-=(O, -x)EWFL(u), X=1=0,
(8.4.25) ~ESUppU-=(O, ~)EWli(u), ~ =1=0.
Proof. Since ~=(27t)nu it follows from Lemma 8.4.17 that

X¢suppu => (0, -x)¢Wli(u), if x=I=O.
On the other hand, Theorem 8.1.8 gives
(0, -X)¢Wli(u) => (0, -x)¢WF(u)=>x¢suppu
which proves (8.4.24). If u is homogeneous then (8.4.25) is (8.4.24)
applied to U. Otherwise suppu=lR n and (0, ~)EWF(u)c Wli(u) for
8.4. The Wave Front Set with Respect to C L 295
every eas in the proof of (8.1.19). Hence (8.4.25) is true. Arguing

exactly as in the proof of (8.1.17) we also conclude that (8.4.23) follows
if we prove that
(8.4.23)'
if xo=l=O, eo=l=O and U is homogeneous in ]R.". The proof will essentially
be a repetition of that of (8.1.17)', with cutoff functions chosen more
carefully.
Choose compact neighborhoods K and K in ]R.· ....... O of Xo and eo
such that
(8.4.26) (K x K)n W&.(U) =0.
By Theorem 1.4.2 we can find a sequence XNECO'(K) equal to 1 in a

fixed neighborhood of eo such that (8.4.6)' is valid for every 0(. We
shall estimate the Fourier transform of VN=XNU in a conic neigh-
borhood of -x o. The homogeneity of U gives as in the proof of
(8.1.17)'
VN( -tx)=t"+·<u, XN(t(. -x))).
Choose r>O and t/lNECO'(K) so that t/lN(X) = 1 when Ix-xol <2r and
t/lN satisfies (8.4.6)', and set UON=t/lNu, u1N=(1-t/lN)U, Then
10 = de/t",

and by Lemma 8.4.4 and (8.4.26)
luON(te)1 ~ C( CLN/t)N, eESUPP XN'

Hence
Moreover,
11 = /l
~ C' L sup t1lli-l"ljy" D/I XN(y)l.
1"+/11 ;>/llyl >tr
Now (8.4.6)' implies that for some constant C
lylN ID/I XN(y)1 ~ C( CLN)N, IPI ~ J.L.
Hence we obtain the estimate
If J.l' is an integer ~O and ~J.l+a+n it follows in view of(8.4.3) that

IvN(x)1 ~ C 2 (C 2 L N_I',/l x lt-I"
if x is in the cone generated by {x; Ix+xol <r}. This means that
(~o, -xO)¢WFL(u), and (8.4.23)' is proved.
8.5. Rules of Computation for W FL
We shall now prove analogues for WFL of the results on WF In

Section 8.2, starting with an analogue of the basic Theorem 8.2.4.
Theorem 8.5.1. Let X and Y be open subsets of lR" and lR m respectively

and let f: X -+ Y be a real analytic map with normal set Nf . Then we
have
Proof First assume that there is an analytic function (J> in

Q={y' +iy"; y'EY, Y"Er, ly"l <Y},
where r is an open convex cone, such that

14i(y'+iy")I~Cly"I-N in Q,
U= lim (J>(. +iy).

r3y~O
This implies that WFA(u)cYxro. Let XoEX and assume that

1'(xo) 11 =1=0, l1Ero, {O}. Then 1'(x o)rO is a closed, convex, proper
cone. We claim that
To give another expression for the right-hand side we let r 1 be an

open convex cone with closure contained in r u {O} and 1'(X o)11 =1=0,
l1Er~, {O}. Then 1'(xo)r~ is a closed convex cone with dual cone
{hElR"; <h, 1'(X o)11) ~ 0, l1EI;.°} = {h; f'(Xo)hEi;J,

which implies that
1'(xo)I;.°={~;<h,O~0 if f'(Xo)hEi;.}.
Since ~ cr u {O} we obtain when ~)" r that
T(xo)ro={~; <h, O~O if f'(Xo)hEr}.
8.5. Rules of Computation for WFL 297
Thus let hEIR" and j'(Xo)hEF. Then Imf(x+ieh)EF for small e>O
°
if x is in a sufficiently small neighborhood X of xo. In X we have °
f*u= lim 4J(f(. +i8h)).
£_+0
In fact, the proof of Theorem 3.1.15 shows that 4J(f(. +i8h)+iy) is a

continuous function of (8, Y)EIR.+ X t, near 0, with values in E0'. Let-
ting 8 --> 0 first we obtain the left-hand side of the equation and letting
Y --> 0 first we obtain the right-hand side. By Theorem 8.4.8 we now
obtain
WFA(f*u)l xo C {(xo, ~); (h, 0 ~O}

Using Corollary 8.4.13 and Theorem 8.4.15 we can write a general
u as a finite sum I U j where each term is either a C L function in a
neighborhood of f(x o) or else satisfies the hypotheses above with
some cone Ij such that Ijo is small and intersects WFL(u)lf(xo)' By
hypothesis T(x o)'1=1=O when (f(x o), '1)EWFL(u) so we conclude that
and this implies (8.5.1).
Remark. Theorem 8.5.1 shows in particular that WFL(u) can be de-

fined as a subset of T* (X)" 0 if X a real analytic manifold.
Theorem 8.5.2. Theorem 8.2.9 remains valid with WF replaced by WFv
The proof is an obvious modification of that of Theorem 8.2.9. In

particular, Theorems 8.5.1 and 8.5.2 show that for a non-zero analytic
density u on a real analytic submanifold WFA(u) is the normal bundle.
Example 8.2.6 is also immediately obtained with W FA in the left hand
side of (8.2.7). Combining Theorems 8.5.1 and 8.5.2 we conclude:
Theorem 8.5.3. If u, VE@'(X) and (x, ~)EWFL(U) implies (x, -~)¢

WFL(v) then the product uv is defined and
WFL(UV)C{(X, ~ +'1); (x, ~)EWFL(U)

or ~=O,(x,'1)EWFL(V) or'1=O}.
When proving an analogue of Theorem 8.2.12 we begin with a

special case which fits the notations in Theorem 8.4.11 well.
Theorem 8.5.4. Let uElg"(lR n), split the coordinates in lRn into two
groups x' = (x l, ... , x n') and x" = (xn' + l' ... , x n), and set
Ul (x') = SU(X', x")dx"
in the sense defined in Section 5.2. Then
WFL(Ul)C {(x', ~'); (x', x", ~', O)EWFdu) for some x"}.
Proof With the notation in Theorem 8.4.11 we have
(U,c/J®t/J)= S (V(.+iw),c/J®t/J)dw,
lwl~ 1
c/JEC~(lRn'), t/JEC~(lRn-n').
Take t/J(x") = X(h") where X = 1 in the unit ball, and let b ~ O. Since V
is exponentially decreasing at infinity it follows then that
(U l ,¢)= S (V(.+iw),c/J®l)dw= S (Vl(·+iw'),c/J)dw

lrol~l lrol~l
where
V l (z') = S V(Z', x")dx"= S V (Z', x" +iy")dx",
is an analytic function when 11m zll < 1 which is bounded by
C(l-IIm zll)-N, If Iw~1 = 1 and (x', x", w~)¢ W FL(U) for every
x"ElR n- n' then Vl EC L at x'-iw~. Hence Lemma 8.4.12 implies
that (x', W~)¢WFL(Ul)'
The following statement is parallel to Theorem 8.2.12 but essen-
tially equivalent to Theorem 8.5.4.
Theorem 8.5.4'. Let Xc lR n, Y c lR m be open sets and K Ef0' (X X Y) be

a distribution such that the projection supp K ~ X is proper. If
UECL(y) then
WFL(%u)c {(x, ~); (x, y,~, O)EWFL(K) for some YESUppU}.
Here % is the linear operator with kernel K.
Proof Replacing K by K (1 ® u) we may assume that U= 1. Without

changing K over a given compact subset of X we may replace K by a
distribution of compact support, and then the statement is identical to
Theorem 8,5.4.
The following is an analogue of Theorem 8.2.13 and the notations

employed are obvious modifications of those used in Theorem 8.2.13.
8.5. Rules of Computation for W FL 299
Theorem 8.5.5. If UEC'(Y) and WFL(u)n WF{(K)y=0 then

WFL(%u)c WFL(K)xu(WF{(K)o WFdu)).
Proof By Theorem 8.5.2 we have

WFL(l (8) u)c {(x, y, 0,1/); (y,1/)EWFL(u)}.
If Ku=K(l<8)u) it follows Theorem 8.5.3 that
WFL(K u ) c{(x, y, (, 1/+1/'); (y, 1/)EWFdu), (x,y, (, 1/')EWFL(K)}
u WFL(K)u WFd1(8) u).
Since % u is the integral of Ku over Y, an application of Theorem 8.5.4
completes the proof.
The applications to convolutions indicated at the end of Section

8.2 obviously carryover to WFL with no change. We shall not dwell
on this any longer but shall instead prove an n dimensional analogue
of Corollary 8.4.16 concerning the uniqueness of analytic conti-
nuation.
Theorem 8.5.6. Let UE.@'(X), X c]R", and assume that f is a real valued
real analytic function in X and XO a point in supp u such that
(8.5.3) df(x°)=!=O, f(x)~f(xO) if XESUppU.
Then it follows that

(8.5.4)
Proof Replacing f by f(x)-lx-xOI 2 we may assume that

f(x) < f(xO) if x°=l=xESUppU.
Since df(x°)=j=O we may take f as a coordinate locally, so we may
assume that f(x)=x n and that xO=O. Choose a neighborhood Y of
OE1R.n - 1 so that Yx{O}~X. Since suppun(Yx{O})={O} we can
choose an open interval I c1R. with DEI so that
YxI~X and (oYxI)nsuppu=0.
If a(x') is an entire analytic function of x'=(x1, ... ,Xn _ 1) then Theo-
rem 8.5.4' (with X x Y, x and y replaced by I x Y, xn and x') gives that
Ua(x n) = Ju(x)a(x')dx'
y
is well defined as a distribution in I and that

WFA(Ua)c{(xn'(n); (x',xn,O,(n)EWFA(u) for some X'E¥}.
Here (x', xn) must be close to 0 if xn is small. If, say (0, en)¢WFA(U), en
=(0, ... ,0,1), then we can choose I so that (x,en)¢WFA(U) if xEYxI.
Hence (xn, l)¢WFA(Ua) if xnEI, so Corollary 8.4.16 gives that Ua=O in
I, because Ua=O when xn>O. Thus, ifu 1 is U restricted to YxI,
for all real analytic a and all cP E C't;' (1). Since a is free to vary in a
dense subset of coo(lR.n-l) it follows from Theorem 5.1.1 that u=o in
Y x I. This contradiction proves (8.5.4).
Theorem 8.5.6 obtains a more suggestive form if one introduces

the normal set of any closed set Fe X:
Definition 8.5.7. If F is a closed subset of a C 2 manifold X then the

exterior normal set Ne(F)c T*(X)'-O is defined as the set of all (X O, eO)
such that XO E F and there is a real valued function f E C 2 (X) with
df(xO) = eo =t= 0 and
(8.5.3)' f(x)~f(xO) when XEF.
It would be sufficient to require f to be defined in a neighborhood

U of xc, for if O~CPEC~(U) is equal to 1 near XO we can replace f by
cp(x)f(x) + (1- cp(x» f(xO). Thus the definition of Ne(F) is entirely local.
In local coordinates we can always replace f by f2(X)-lx-xOI 2 where
f2 is the second order Taylor expansion at xc, so f can always be
taken analytic and strictly smaller than f(xO) when XO =t= XEF. The
following proposition shows that there is always a large normal set if
8F=t=0.
Proposition 8.5.8. For every closed subset F of the C 2 manifold X the

projection of Ne(F) in X is dense in 8F. If xOEF, fECI(X), df(xO)
=e°=t=O and f(x)~f(xO) when xEF, then (xO,eO)ENe(F). If XcIR" and
Y is a convex open set eX . . . . F, XO EF n 8 Y, we have (XO, eO)ENe(F) for
some eo with <x -
xc, eO) > 0, XE Y.
Proof Since the assertions are local we may assume that Xc IR". If
yEX . . . . F and ZEF has minimal Euclidean distance to y, thus
Iz-yI2_lx-yI2~0, xEF,
then (z, y-z)ENe(F). If xOE8F we can choose a sequence y.EX . . . . F

converging to XO and obtain a sequence (z., (.)ENe(F) with z. -+ xc,
which proves the first statement. If fECI and f(x)~f(xO) when xEF,
8.5. Rules of Computation for W FL 301
we choose yv=xo+f'(XO)/v on the outer normal and obtain if zv=xo
1f'(x°)jvI 2 ~ Iw v-f'(xO)/vI 2, f(xO +wv)~ f(xO).

Thus vlwvI2/2~<wv' f'(xO»~o(lwvl), so vWv ~ 0 as v ~oo and
v(Yv -z.)= f'(XO)-vw v ~ f'(XO)
which proves that (XO, f'(xO))ENe(F).

Next assume that Y is convex and that Y~X, FnoY={xO}, Fn Y
=0. Let OE Y. Then the homogeneous function f of degree 1 such that
Y = {x;f(x) < 1} is convex, for if f(x) + f(y)=M then f(x+ y)~M since
(x+ y)jM = (x/f(x))f(x)jM +(y/f(y))f(y)jMEY.
Let O~XECg', JXdx= 1, and set 1.= f*x, where x.(x) = en x (x/e).
Then we have If-I.I < C e since f is Lipschitz continuous. If we note
that 1-f ~ 0 in F with equality only at XO we conclude that the
maximum of 1 -I. in F is ~ C e and is attained at a point x, such that
x, ~ XO when e ~ O. Since I. is convex we have if ~.= -I.' (xJ/1 f.' (x.)1
(x., ~,)ENe(F); <x-x., ~,> ~O when I. (x) < f,(x.).
If ~o is a limit point of ~. when e ~ 0 it follows that (XO, ~O)ENe(F)
and that <x-xo, ~o> >0 if XEY.
If Y is any convex open set eX ....... F with XO E F n 0 Y we can apply
the preceding result to the interior of the convex hull of XO and any
compact subset K of Y with interior points. Then the last statement in
the proposition follows when K)' Y.
In what follows we also use the notation
for the interior normal set of F and N(F)=Ne(F)uN;(F) for the whole
normal set. The closure in T*(X) . . . . O will be denoted by N(F). Theo-
rem 8.5.6 can now be restated as follows:
Theorem 8.5.6'. For every UE.@'(X) we have
(8.5.5) N(suppu)c WFA(U).
The importance of Theorem 8.5.6' will be enhanced in Section 8.6
where we prove that if U satisfies a differential equation P(x, D)u=O
with analytic coefficients, then W FA (u) is contained in the characteris-
tic set of P. Thus the principal symbol p(x,~) vanishes on WFA(u), so
it must vanish on N(suppu) by (8.5.5). We shall now examine the
purely geometrical consequences of having such a function. Recall (see
Section 6.4) that if p is a real valued function in COO(T*(X)"O) and

p(XO, ~O)=O, then the Hamilton equations
dx/dt = op(x, Wo~, de/dt = - op(x, Wox
°
with initial data (x, e) = (XO, ~O) when t = define a curve through
(XO, eO) on which p remains equal to 0. It is called a bicharacteristic
(strip); the projection in X is called a bicharacteristic curve. It is non-
singular if 0p (x, woe
=1= 0.
Theorem 8.5.9. Let F be a closed subset of X and assume that

pECOO(T*(X)"O) is real valued and vanishes on Ne(F). If (XO, ~O)ENe(F)
it follows that a neighborhood of (XO, eO) on the bicharacteristic strip
t ---> (x(t), ~ (t» for p through (XO, ~O) remains in Ne(F). If p~(XO, eO) =1=
there is a function tPECOO(X) such that for some e>O
°
tP (x (t» =0, dtP(x(t» = (t)e if It I< e,
and tP(x) <
x¢=r.
°when x is in a neighborhood of r = {x(t), Itl < e} in F but
Proof We may assume that X e]Rn since the result is local. Choose
fECoo with f(xO)=O, df(xO)=eo, and f(x)<O if x°=l=xEF. Using
Theorem 6.4.5 we can find a solution cjJ of the Cauchy problem
(8.5.6) ocjJ/ot + p(x, cjJ~) =0, cjJ(O, x) = f(x),
for It I< (j and x in a compact convex neighborhood W of XO where

df(x)=l=O. We have cjJ~=I=O in (-(j,(j)xW, and cjJ~(t,x)=e, cjJ;(t,x)=
- p(x,~) on the curves
(8.5.7) dx/dt=op(x, Wo~, de/dt= -op(x, Wox;
(x,e)=(y,!,(y», t=O.
When y =xo we have in particular cjJ~(t, x(t» = e(t), and cjJ(t, x(t» is
independent of the choice of f
If (j is small enough then
M(t)=maxcjJ(t,x» max cjJ(t, x), Itl<(j,
FnW Fnl1W
for this is true when t=O. Thus the maximum M(t) is attained at a
point xtEF in the interior of W. We claim that M(t) =0, It I< (j, which
is true by hypothesis when t=O. To prove this we observe that since
Isupu -sup vi ~sup lu -vi for all bounded functions u, v, we have
IM(t) - M(s)1 ~max IcjJ(t, x) - cjJ(s, x)l.
FnW
8.5. Rules of Computation for WFL 303
We may replace F n W by {xl' x s } here for this does not change the
maximum values M(t) and M(s). Now (XI' ¢~(t, X,))ENe(F) and
(x s' ¢~(s, xs))ENe(F), hence P(XI' ¢~(t, XI)) = p(xs' ¢~(s, x s)) =0. Thus it
follows from (8.5.6) that ¢;(t, XI) = ¢~(s, x s) =0, so Taylor's formula
gives
¢(t, XI) - ¢(s, XI) = O«t _S)2), ¢(t, x s) - ¢(s, x s) = O«t _S)2).
This means that M(t)-M(s)=0«t-S)2) so M'(t)=O and M(t)=M(O)
=0.
Now replace f(x) by g(x) = f(x)-lx-xOI 2 in the preceding argu-
ment. Thus we let t/J be the solution of the equation
iJt/J/iJt+p(x,t/J~)=O, t/J(O,x)=g(x).
For sufficiently small band W the results proved for ¢ are also valid
for t/J, and ¢(t,x)-t/J(t,x) is a strictly convex function of XEW. When
X = x(t) we know that ¢(t, x(t)) = t/J(t, x(t)) and that ¢~(t, x(t))
=t/J~(t, x(t))=W), so the convexity gives ¢(t, x»t/J(t, x) when x*x(t).
Since ¢(t, x) and t/J(t, x) both have the maximum value zero in W n F,
it must be attained at x(t) and only there in the case of t/J (t, x). Hence
(x(t), ~(t))ENe(F) and ¢(t, x(t))=O.
Assume now that p~(xo, ~O)*O. The equation ¢;(T,x)=O defines a
COO function T(x) in a neighborhood of XO with T(xO)=O unless
0= ¢;;(O, XO) = - <P~(XO, ~O), ¢;~(O, XO»
= <p~, p~> + <f;~p~, P~>.
In that case we just replace f by g. Note that ¢;(t, x (t)) = - p(x(tH (t))
=0 because (x(t), ~(t))EN.(F), so T(x(t))=t. We may assume that
¢(t,x)<O if xEWnF and x*x(t). Then cP(x)=¢(T(x),x) has the
required properties.
Corollary 8.5.10. Let F be a closed subset of X and set

.h;-={pECoo(T*(X)-.....O); p=O on Ne(F)}.
Then .h;- is an ideal in Coo (T* (X) -.. . 0) which is closed under Poisson
brackets, that is, if P and q are in .h;- then .h;- contains the Poisson
bracket
Proof Only the last assertion needs verification. We may assume that
p is real valued. If (XO, ~O)ENe(F) then the bicharacteristic strip
t -+ (x(t), ~ (t)) for p through (XO, ~O) is in Ne(F) for small t by Theorem
8.5.9, hence
o=dq(x(t), ~(t))/dt = {p, q}(x(t), ~ (t))
which proves the corollary.
If p 1> ••• , Pk E.AlF and d PI' ... , d Pk are linearly independent at

(XO, eO)ENe(F), then repeated use of Theorem 8.5.9 gives a k dimensional
manifold through (XO, eO) contained in Ne(F). The restriction of the
one form (~, dx) and therefore of the symplectic form to any manifold
XcNe(F) is equal to O. In fact, if (XO, ~O)ENe(F) we can choose f with
f-:i:,f(xO) in F so that (eO,dx)=df(x) in the tangent space of E at
(XO, eO). This is zero since f has a maximum in E at (XO, eO). Hence
k-:i:,n, and since Ne is conic it follows that k<n if apda~, ... , apJa~
are linearly independent.
We shall now discuss the dual objects of Ne(F) briefly. This is not
really related to the main theme of the chapter, but the proofs contain
arguments similar to the proof of Proposition 8.5.8 and throw more
light on the set Ne(F). Moreover, the result will be important in
Chapter XXVI. It confirms the geometrically plausible fact that an
integral curve of a vector field cannot leave a closed set without
pointing out of it at some boundary point.
Theorem 8.5.11. Let v be a C 1 (or just Lipschitz continuous) vector

field in X, and let F be a closed subset of X. Then the following
conditions are equivalent:
(i) Every integral curve of dxjdt=v(x(t)), O-:i:,t-:i:,T, with x(O)EF is
contained in F.
(ii) (v(x), 0 -:i:, 0 for all (x, e)E Ne(F).
Proof (i) => (ii). Let fEC 2 , df(x)=e and assume that f restricted to F
has a local maximum at x. Locally we can solve the equation dyjdt
=v(y(t)) with y(O)=x. Since (i) gives that f(x)~f(y(t)) for small t>O,
the derivative at t=O of f(y(t)) must be -:i:,0, that is, (v(x), 0-:i:,0. To
prove that (ii) => (i) we may assume that X =1R.n and begin by proving
an elementary lemma:
Lemma 8.5.12. Let F be a closed set in 1R.n and set
f(x)=min Ix-zl 2
ZEF
where I I is the Euclidean norm. Then we have

f(x+ y)= f(x)+f' (x, y)+o(iyi),
f'(x, y)=min {(2y, x-z); zEF, Ix-zI 2 = f(x)}.
Proof We may assume in the proof that x = O. Set
q.(y)=inf { - 2(y, z); zEF, Izl-:i:,(f(O))t + B}.

q. is homogeneous of degree 1, and q./, qo as dO. The limit is

therefore uniform on the unit sphere, so
qo(y)~q.(y)~qo(y)-c.lyl, c.--+O as 8--+0.
Now
ly-zl2 =lzl2 -2(y, z) +lyl2
which gives immediately
f(y)~f(O)+qo(Y) + lyl2.
On the other hand, when Iyl ~8j2 the minimum in the definition of
fey) is assumed for some z with Izl~f(0)t+8j2, so
f(y)~f(0)+q.(Y)+lyI2, Iyl ~8j2.
The lemma is proved.
Proof of Theorem 8.5.11. With the notation in (i) and Lemma 8.5.12
we have if t< T
lim (f(x(s» - f(x(t»)j(s - t) = f'(x(t), v(x(t»).
s-t+O
Since the result to be proved is local we may assume that for all x
and y
Iv(x)-v(y)1 ~ Clx- YI.
When ZEF and Ix(t)-zI2 = f(x(t» we have
2(v(x(t», x(t) -z) = 2(v(z), x(t) - z) - 2( v(z) - v (x (t», x(t) - z).
The last term is ~2Cf(x(t». From the proof of Proposition 8.5.8 we
recall that if f(x(t» > 0 then (z, x(t) - Z)ENe(F) for every z such that
Ix(t)-zI2=f(x(t». Hence the first term on the right is ~O by con-
dition (ii) so the right-hand derivative of f(x(t» is ~ 2 C f(x(t». Thus
the right-hand derivative of f(x(t» e- 2et is ~ 0 so this is a decreasing
function by a simple modification of the proof of Theorem 1.1.1. (Note
that f is continuous.) If f(x(O»=O we obtain f(x(t»=O, O<t< T, as
claimed.
8.6. WFL for Solutions of Partial Differential Equations
If
P(x, D) = L a,,(x)D"
l"l~m
is a differential operator with coefficients in CL(X), we have proved

that
UE~'(X).
When a" are real analytic this is also an easy consequence of Theorem
8.4.8, Lemma 8.4.2 and Theorem 8.4.5. Making that assumption we
shall now prove a converse similar to Theorem 8.3.1.
Theorem 8.6.1. If P(x, D) is a differential operator of order m with real

analytic coefficients in X, then
(8.6.1) WFL(u) c Char Pu WFL(Pu), UE~'(X),
where the characteristic set Char P is defined by (8.3.4).
Proof We shall repeat the proof of Theorem 8.3.1 but make a more
careful choice of cutoff functions. We must prove that if (xo, ~o) does
not belong to the right-hand side of (8.6.1) and ~o =1= 0, then
(xo, ~o)¢ W FL(U), The hypothesis means that we can choose a compact
neighborhood K of Xo and a closed conic neighborhood V of ~o in
JR n '- 0 such that
(8.6.2) Pm(x, ~)=I=O in K x V
(8.6.3) (K x V)n WFdPu) =0.
Using Theorem 1.4.2 we now choose a sequence XNECg'(K) equal to 1

in a fixed neighborhood U of Xo such that for every Q(
(8.6.4)
Then the sequence UN = X2NU is bounded in 8' and equal to U in U.

The theorem will be proved if we show that (8.4.5) is valid in V when
I~ I> N, for (8.4.5) follows from the boundedness of UN when
I~I~N~LN'
To estimate u @ in V we must solve the equation
N
(8.6.5)
approximately. Writting v=e-i<x'~>w/Pm(x,~) and noting that the

principal symbol of Ip is Pm(x, -~), we obtain instead of (8.6.5) an
equation of the form
(8.6.6)
where Rjl~lj is a differential operator of order less than or equal to j
with analytic coefficients which are homogeneous of degree 0 with
respect to ~ when ~EV and XEK. Formally a solution would be given
by 00
L.o
R kX2N'
However, we must not introduce derivatives of very high order so we

set
WN= I Rj, ... Rjk X2N'
j, + ... + jk;g;N-m
A simple calculation gives

wN-Rw N=X2N- I Rj, ... R jk X2N=X2N- eN'
i1 + ... + jk>N -m;?;i2+ ... + ik
This means that

rp(X, D)(e-i<x.~> WN(X, WPm(X, ~)) = e- i(X,O(X2N(X) - eN(x, ~)).
With integrals denoting action of distributions we obtain

(8.6.7) Ju(x)X2N(x)e-i(X,~> dx = JU(X) eN (X, ~)e-i(x,O dx
+ Jf(x)e-i(x,~> wN(x, WPm(x,~) dx.
Here f =P(x, D)u. To estimate the right-hand side of (8.6.7) we first
prove a simple lemma.
Lemma 8.6.2. There is a constant C' such that, if j = j 1 + ... +jk and j
+!PI~2N,
(8.6.8) !DJlRj, ... RjkX2N!~C'N+INj+IJlI!~!-j, ~EV.
Proof By the homogeneity it suffices to prove the lemma when !~! = 1.

All coefficients occurring in some R j when !~! = 1, ~E V, have a fixed
bound in a fixed complex neighborhood of K, so the lemma is a
consequence of the following one.
Lemma 8.6.3. Let K be a compact set in 1R.n and K' a neighborhood of

K in <en. If al, ... ,a j _ l are analytic and !a l !<1, ... ,!a j _ 1 !<1 in K',
j~N, we have
(8.6.9)
Proof By Cauchy's inequalities we have for some r>O
IDlXa11 ~ lal! r- 11X1 in K,

and (8.6.4) gives
!DIXXNI~Co(CoN)IIXI~CoeNC~lla!! if IIX!~N,
for NllXl/lal! ~ eN. Now it is clear that Dil a l ... a j _ l DijXN is a sum of
terms of the form (DlXlal) ... (DlXj-laj._l)DlXjXN where
If there are Ck1 .. kj terms with IIXII=kl, ... ,llXjl=kj' the left-hand side
of (8.6.9) can be estimated by
CoeN(max(C o, l/r))NI Ckl ... kjkl!'" k j !.
Since the derivative Dik in (8.6.9) operates on all the following factors,
it is easy to see that
I Ckl ... kjX~l ... xyj =(X 1 + ... +X)(X2 + ... +x) ... Xj'
It follows that
I Ckl ... kjkl!"· k j ! =s"·s (Xl + ... +X)(X2 + ... +x) ... xje-(Xl+ .. +Xj)dx
° °
=(2j -I)!! ~(2N)j.
(The integral is computed by taking Xl + ... +x j , x 2 + ... +x j , ... as

new variables.) This completes the proof of the lemma.
End of the proof of Theorem 8.6.1. If M is the order of u in a neigh-

borhood of K, we can estimate the first term on the right-hand side of
(8.6.7) for large Nand I,;I>N, ';EV, by
C I (l+I,;I)M-lalsupIDaeN(x,m
lal;;:;M x
The number of terms in eN cannot exceed 2N, and each term can be
estimated by means of (8.6.8), which gives the bound
C11,;IM+m-N cN+ 1 NN+M2N.
If N is replaced by N + m + M this is an estimate of the desired form

(8.4.5) even for the analytic class. To estimate the last term in (8.6.7)
we observe that (8.6.8) gives
(8.6.1 0)
We have a similar bound for W N 1,;lm/Pm(x, ';). The proof is therefore

completed by the following lemma.
Lemma 8.6.4. Let fE fifi' (X), let K be a compact subset of X and V a

closed COlJe c lR.n" 0 such that
WFdf)n(K x V)=0.
If WNEC~)(K) and (8.6.10) is fulfilled, then
...............
(8.6.11) IWNf(,;)I~C~+l(LN_M_n/I,;I)N-M-n if ';EV, I,;I>N, N>M +n.
Here M is the order of f in a neighborhood of K.
8.6. W FL for Solutions of Partial Differential Equations 309
Proof By Lemma 8.4.4 we can find a sequence fN which is bounded in

C'M and equal to f in a neighborhood of K so that
Ih('1)1 ~ C(CLN/I '1 It, '1 EW,

where W is a conic neighborhood of V. Then wNf=wNfN" N'=N-M
-no Since
by (8.6.10), it follows from (8.1.3) that
1WJ'(~)I~C~+l((LN';I~I)N' +NNI~ln+M-N), ~EV, I~I>N.
Since N'~LN' this proves (8.6.11).
Combination of Theorems 8.6.1 and 8.5.6' gives
Theorem 8.6.5 (Holmgren's uniqueness theorem). If UE~'(X) is a

solution of a differential equation P(x, D)u =0 with analytic coefficients,
then the principal symbol Pm(x, ~) must vanish on N (supp u). Thus u = 0
in a neighborhood of a non-characteristic c 1 surface if this is true on
one side.
The last statement follows in view of Proposition 8,5.8. (Recall

from Section 6.4 that a C 1 surface with normal ~ at x is non-
characteristic at x if Pm(x, ~)=t=0.) If P is elliptic (cf. Corollary 8.3.2)
then the theorem states that N(suppu) is empty, so suppu has no
boundary point in X. If X is connected and u =0 near a point in X, it
follows that u=o in X. A stronger unique continuation theorem is
obtained if we use Corollary 8.5.10 also:
Theorem 8.6.6. Let P(x, D) be a differential operator with analytic

coefficients and let ~ be the smallest subset of COO(T* X-.....O) which
contains all Coo functions vanishing on Char P and is closed under
Poisson brackets. If UE~' (X) and P(x, D)u = 0 it follows then that all
functions in ~ must vanish on N (supp u).
In particular, if the functions in ~ have no common zeros then we

conclude that u vanishes identically if X is connected and u vanishes
in an open set. If u vanishes on one side of a C 1 surface with normal
~ at x, then u vanishes in a neighborhood of x unless all functions in
~ vanish at (x, ~). This is an improvement of the classical uniqueness
theorem of Holmgren as the following example shows:
Example 1. If P(x,~)=~i+xg~+",+X;_l~; then ~l' Xl~2"",Xn_l~n

vanish on Char P. Taking Poisson brackets we obtain
{~l' Xl ~2} =~2' {~2' X2~3} =~3' ... , {~n-l' Xn_ l ~n} =~n
so the functions in <'C have no common zeros.
Example 2. If P(x,~)=xgi+~~+~~ then <'C contains ~2'~3,X2' and

since {~2' x 2 } = 1E<'C there are no common zeros. However, the so-
lutions of P(x, D)u = 0 need not be analytic. In fact,
ut(x)=exp (rx3 +iXl r2 -x~r2/2)
is a solution for every r. Hence
I ut(x)e-tdr
00
u(x)=
o
is a Coo solution when IX31 < 1, but u is not real analytic since
00
D~u(O)= I r2k e - t dr=(2k)!.

o
For differential operators with constant coefficients forming Pois-
son brackets is of no use, for the Poisson bracket of any two
functions of ~ is O. The following is then a partial converse of Theo-
rem 8.6.5.
Theorem 8.6.7. Let the plane <x, N) =0, NEJR n, be characteristic with
respect to the differential operator P(D), that is, Pm(N) =0. Then there
exists a solution u of the equation P(D)u=O such that UEcoo(JRn) and
suppu={x; <x,N)~O}.
Proof Let P = Pm + Pm _ 1 + ... + Po where 1j is homogeneous of degree j

and Pm =1=0. With a fixed vector ~ such that Pm(~) =1=0 we shall study the
solutions of the equation
(8.6.12) P(sN +t~)=O
for large s. To do so we set t = ws which reduces (8.6.12) to an
algebraic equation in wand l/s,
Pm(N + w~) + ... + (l/st- k Pk(N + w~) + ... = O.
When l/s=O this algebraic equation in w is not identically satisfied
since Pm(~)=I=O but it is true for w=O since Pm(N) =0. Hence it follows
from Lemma A1.3 in the appendix to Volume II that for some
integer p the equation (8.6.12) has a solution which is an analytic
function of (l/s)l/p in a neighborhood of the origin and vanishes at the
origin. This means that (8.6.12) has a solution
L CiS-I/P)i
00
(8.6.13) t(S)=S
I
analytic for IsI/PI > M where M is a constant. Thus we have with a

constant C
(8.6.14) It(s)1 ~ C Isll -lip, lsi> (2M)P.
Now choose a number p such that 1-1/p (2M)P
J
i1"+oo
(8.6.15) u(x)= ei(x,sN+t(s)Oe-(sli)Pds.
i't-oo
Here we define (s/i)P so that it is real and positive when s is on the

positive imaginary axis, and we choose a fixed branch of slip in the
upper half plane. The integral is convergent and independent of 't, for
when x is in a fixed bounded set we have in view of (8.6.14)
(8.6.16) Re(i(x, sN + t(s) 0 -(s/i)P)
~ -'t(x, N) + Clxll~llsll-I/P -lslP cos (np/2)
~ -'t(x, N) -clsl P
if 0 < c < cos (n p/2) and lsi is large. This estimate also shows that, when
x is in a compact set, the integral·,(8.6.15) is uniformly convergent
even after an arbitrary number of differentiations with respect to x.
Hence UECoo and using (8.6.12) we conclude that P(D)u=O. From
(8.6.16) we also obtain
J e-clulP du.
00
lu(x)1 ~e-t(x,N>
-00
Hence it follows when 't -+ + 00 that u(x) =0 if (x, N) >0. (Compare

the proof of Theorem 7.3.1.)
When (x, N) < 0 we can replace the integration contour in (8.6.15)
by the negatively oriented boundary of the set
{s; lsi < (2M)P or Ims<O, IResl < (2M)P}.
The integrand is then exponentially decreasing and remains so for all

complex x with (Rex, N) <0. Hence u is analytic in the half space
{XElR. n ; (x, N) <O} and does not vanish identically by Fourier's in-
version formula. (Note that u(x) is the Fourier transform of
exp( - (s/i)P) when (x, N) = t and (x, 0 = 0.) This proves that supp u
is the closure of the half space. The theorem is proved.
Remark. Separating the integration in (8.6.15) for Res>O and for

Res<O we can write u(x)=u+(x)+u_(x) where u+(u_) is the boundary
value of a function analytic when <1m Z, N) > 0 (resp. <1m Z, N) < 0).
This proves that WFA(u) = {(x, tN); <x, N) =O} is the normal bundle
of the boundary of the support.
The following theorem gives a useful summary of the results in the

constant coefficient case.
Theorem 8.6.8. Let Xl and X 2 be open convex sets in 1R." such that
Xl eX 2, and let P(D) be a differential operator with constant
coefficients. Then the following conditions are equivalent:
(i) Every UE.@'(X 2) satisfying the equation P(D)u=O in X 2 and
vanishing in Xl must also vanish in X 2'
(ii) Every hyperplane which is characteristic with respect to P and
intersects X 2 also intersects Xl'
Proof (i) => (ii) Assume that n is a characteristic hyperplane which

does not intersect Xl' Let H be the half space bounded by n which
does not intersect Xl' By Theorem 8.6.7 we can find a solution u of
the equation P(D)u=O with suppu=H, so (i) shows that HnX 2 =0,
hence nnX2=0.
(ii) => (i). Let Y2 be a point in X 2. Choose a point Y1EXl and

denote by I the line segment between Y1 and Y2' We can find an open
convex set X ~ Xl such that every characteristic plane intersecting I
also meets X. In fact, if xoEI and ~oE1R.", Pm(~O) =0, I~ol = 1, we can
choose an open ball 17~Xl which meets the plane <x-xo, ~o>=O
and consequently meets every characteristic plane with normal close
to ~o passing through a point near xo' By the Borel-Lebesgue lemma
a set X with the required properties can therefore be constructed by
taking the convex hull of a finite number of open balls 17 ~ Xl'
Let 1"; be the interior of the convex hull of X and Yt = Yl + t(Y2
-Yl)' O~t~1. For small t we have Yt EX 1, hence 1";cX1 and u=O in
1";. Let T be the supremum of all tE[O, 1] such that u=O in 1";. Then u
=0 in YT and h¢X1 • If n is a supporting plane of YT then n is non-
characteristic if hEn since n intersects I but not X, and if h¢n then
n n ¥reX ~ Xl' Hence it follows from Proposition 8.5.8 and Theorem
8.6.5 that aYT n supp u = 0. Hence T = 1, and u = 0 in a neighborhood
of the arbitrarily chosen point Y2 EX 2' This proves condition (i).
Corollary 8.6.9. If the support of a solution UE.@'(1R.") of the equation
P(D)u = 0 is contained in a half space with non-characteristic boundary,
then u=O.
Proof Every characteristic plane intersects the half space.
Corollary 8.6.10. Let Nl and N2 be real vectors such that

(8.6.17) Pm(rlNl+r2N2)=l=0 when rl>O and r2~0.
Set Xa),az = {x; <x, N j ) <aj,j= 1, 2}, where a l and a 2 are real numbers
or + 00. If UE!!&'(Xa),a,) satisfies the equation P(D)u=O and if u=O in
Xc,az for some c<a l then u=O in Xa1,az'
Proof The normal of a plane which does not intersect Xc,az must be a
linear combination with non-negative coefficients of Nl and N 2 • If the
plane is characteristic, the normal must therefore be proportional to
N2 by (8.6.17). Hence (ii) in Theorem 8.6.8 is fulfilled.
Corollary 8.6.11. Let X be an open proper convex cone with vertex y

such that no hyperplane through y which is characteristic with respect to
P(D) intersects X only at y. Every UE!!&'(X) satisfying the
equation P(D)u=O and vanishing outside a bounded subset of X must
then vanish in all of x.
Proof Since X is proper we can find a plane n through y which meets

X only at y. We can apply Theorem 8.6.8 with X 2 =X and Xl equal
to the intersection of X and a suitable half space with boundary
parallel to n. The hypothesis means that in every characteristic plane
containing a point in X 2 there is a half ray which lies entirely in X 2'
Hence it meets Xl so the corollary follows from Theorem 8.6.8.
In the proof of Theorem 8.3.7 we actually saw that w_ was an

entire analytic function, and if the reference to Theorem 8.1.6 in the
proof is replaced by a reference to Theorem 8.4.8, we obtain
Theorem 8.6.12. If P(D) is of real principal type, then one can find
E±E!!&'(JRn) and analytic w± such that P(D)E± =t>+w± and
(8.6.18) WFA(E±) c {(tP~W, ~); t~O, Pm(~) =0, ~ =l=0} u To*'- {OJ.
We can now prove an analogue of Theorem 8.3.3'.
Theorem 8.6.13. Let P(D) be of real principal type. If UE!!&'(X), P(D)u

= f and (x, ~)EWFdu),- WFL(f), then Pm(~)=O and
Ix {O c WFL(u)
if I c X is a line segment with direction P~W containing x such that
(I x {~})n WFL(f) =0.
Proof Without restricting the generality we may of course assume

that uErff'. Let U and F be the analytic functions in {z; IImzl<l}
corresponding to U and f as in Theorem 8.4.11. Since U = K * u and F
= K * f we have P(D) U = F. In particular, if Iwl = 1 then
(8.6.19) P(D) U(. -iw)=F(. -iw).
If there is some yEl such that (y, e)rt WFL(u) then there is a neigh-
borhood Wof - e/lel in sn-1 such that
u 1= S U(. +iw)dwECL(V), f1 = S F(. +iw)dwECL(V1)

W W
where V is a neighborhood of y and V1 is a neighborhood of 1. We

have (x, e)E WFL(U 1) since (x, e)rt WFL(u - u 1), and P(D) u 1= k
We may assume that x-y=tp':(e) for some t>O. Set
Now choose a cutoff function XEC~ which is 1 near the segment

[y,x] so that l_rlsuppdxcV. We can choose X so that X(x)
=",«x,'1»), XEV, for some "'ECce and some '1 with <p':(e),'1) *0, thus
'1 is linearly independent of e. By Theorem 8.5.1 this implies that
WFL(v)lv C Vx 1R.'1,
WFL(P(D)v)II_ cVx1R.'1.
It follows from Theorem 8.5.5 that the analogue for WFL of (8.2.16) is
valid, so for v=E+*P(D)v-w+*v we have
Since v=u 1 in a neighborhood of x this is a contradiction proving

that (y, e)E WFL(U). The proof is complete.
Corollary 8.6.14. Let P(D) be of real principal type, UE~'(X) and

P(D)u=O. If (x,e)EN(suppu) and lcX is an interval containing x on
the line through x with direction p':(e), then 1 C supp u.
Proof By Theorem 8.5.6' we have (x, ¢)E WFA(u) so I x {¢} c WFA(u) by

Theorem 8.6.13. This proves the statement.
Corollary 8.6.14 should be compared with Theorem 8.5.9 which is far

more general but only local. Theorem 8.6.13 and Corollary 8.6.14 can
be extended to operators with real analytic coefficients, but the proofs
require additional technical tools then. Instead we shall prove an
extension of Theorem 8.6.1 to convolution equations which will be
useful in Section 12.9.
If JlEY"(JR.n) we can define a characteristic set as follows. First we
let r be the set of all ¢oEJR.n...... {OJ such that there is a complex conic
neighborhood V of ¢o and an analytic function iP in Vc
= {C EV, "I > C} for some C such that iP (i. = 1 in Vc "JR.n and
(8.6.20)
for some Nand C 1 • We shall denote by Char Jl the complement of r

in JR.n ...... {OJ.
Theorem 8.6.15. If JlEY"(1R.n) and uEt9"(1R.n), then
(8.6.21)
Proof We shall use the interpretation of W FA in Theorem 8.4.11. With

the notation in that theorem we must show that u*K(z) is analytic at
xo-i¢o if ¢o¢;CharJl, l¢ol=1 and (xo,¢o)¢;WFA(f), f=Jl*u. Choose
V and iP as above so that (8.6.20) is valid and iP (i. = 1 in Vc "JR.n. Let
W' and W" be closed conic neighborhoods of ¢o in JR.n ...... {OJ such that
W" is contained in the interior of W' and W' c V. Choose XE COO with
O~X~ 1 equal to 1 in a neighborhood of W~'c and supp XC W;c so that
X is homogeneous of degree 0 when I¢I > 3 C. Then the Fourier trans-
form of u * K(. + iy), Iyl < 1, can be decomposed as follows
If we introduce the inverse Fourier transforms
K 1 (z) =(2n)-n J(1- x(¢))ei<z. ~> /I@d¢,
K 2 (z)=(2n)-n JX(¢)iP(¢)ei<z'~>/I@d¢
which are rapidly decreasing when Re z --+ 00, 11m zl < 1, it follows that
(8.6.22)
It is clear that K 1 remains analytic when 11m z + ~ 01 is sufficiently

small, so K1 *u(z) is analytic at xo-i~o. To study the properties of K2
we shall follow the proof of Lemma 8.4.10 although we must now
work in all variables and apply Stokes' formula. Let Xl (~) be a COO
function with support in W~~ which is 1 in W;~ for another conic
neighborhood W'II of ~ 0 and is homogeneous of degree 0 for I~ I> 4 C.
We want to move the integration to the cycle (x = Re z)
where 0 < b ~ 1 is chosen so small that we do not leave Vc when

~ ESUPP Xl. To estimate the integrand we shall use Lemma 8.4.9 and
the inequality
Re(i(x+ iy, ~ + i'1) -(~ + i'1, ~ + i'1)t)
~ - (x, '1) - (y, 0 _(1~12 -1'112)t
valid when 1'11<1~1. (This follows from the fact that Rew2~(Rew)2.)
When '1=pl~lx(1+lxI2)-t, 0~p~1, we obtain the estimate
The parenthesis is a convex function of p which is -1 for p = 0 and

-(1+lxI2)t when p=1. Hence
Re(i(x+iy, ~ +i'1)-(~ +i'1, ~+i'1)t)

~-I~I(l-p+ p(l +lxI2)t)_(y, o.
Using Stokes' formula as just indicated it follows that for some b > 0
there is an analytic continuation of K 2 to
{z; 11m zl < I-b +b(l + IRe ZI2)t, 11m z + ~ol <b}
where the second restriction as in the discussion of K 1 comes from
the set where the integration contour has not been deformed. An
integration by parts shows that K 2 is rapidly decreasing at infinity in
this set.
The properties of K2 show that the boundary value K 2*/(. -i~o)
is equal to the convolution of I and the boundary values K 2 (. -i~o)
which are analytic except at O. Write 1=11 +12 where 11 EtC' and 12
vanishes when Ix-xol<r, say. Then
. K2*/2(Z)=f2(K2('-Z))
is analytic when z is so close to xo-i~o that K(. -z) is uniformly in

g when Ix-xol~r. By Theorem 8.4.8 we have WFA (K 2 (. -i~o))
e{(O,t~o),t>O}. (Cf. Lemma 8.4.12.) If we recall that (xo'~O)¢WFA(fl)
it follows by the analogue of (8.2.16) for WFA that
xo¢singsuPPA 11 * Kz<. - i~o)'
Hence K * u is analytic at Xo - i ~ 0 which completes the proof.
Remark. The theorem remains valid if uE~'(1Rn) and

supp Jl x supp U3(X, y)-+x + y
is proper. In fact, Jl * U is then defined (see Section 4.2) and for any x
we have Jl*u=Jl*(</Ju), u=</Ju in a neighborhood of x if <pEer} is
equal to 1 in a sufficiently large set. If we apply (8.6.21) to </Ju we
conclude that (8.6.21) is valid for the fiber at x.
8.7. Microhyperbolicity
If F is a real valued real analytic function in the open set X e1R n and
8 is a real vector such that <8,F'(x)>O when F(x)=O, then
F8- 1 = lim l/F(. +ie8)
l:_+O
is a well defined distribution with

WFA(F8- 1) = {(x, tF'(x)); F(x)=O, t>O}.
In fact, if r is an open convex cone such that
<y,F'(x)>O, O~YEt, XEXo~X,
then it follows from Taylor's formula that
(8.7.1) Iyl ~ ClF(x + iy)1 if XEX 0' YEr and lyl is small.
The statement is therefore a consequence of Theorem 8.4.8. Now a
weaker form of (8.7.1) may be valid also when F has critical points. A
typical example is the Lorentz form F(x)=xi -x~ - ... -x;. By (7.4.8)
F(Y)~IF(x+iY)I; x,YE1Rn ;
so (8.7.1) is valid with lyl replaced by lyl2 if t . . . {OJ is contained in the
open light cone. The following terminology is motivated by this exam-
ple.
Definition 8.7.1. A real analytic function F in the open set X cIRn is

called microhyperbolic with respect to 8EIRn if there is a positive
continuous function t(x) in X such that
(8.7.2) F(x+it8HO, if O<t<t(x), XEX.
In the following discussion of the local properties of F we may

shrink X so that t is bounded from below in X by a positive constant
and then replace e by a multiple to achieve that
(8.7.2)' F(x+it8HO if 0<t~1, XEX.
To simplify notation we also assume that OEX and study F near 0.
Lemma 8.7.2. If F satisfies (8.7.2)' and F(t8) has a zero of order m

exactly when t = 0, then
F(x) =Fo(x)+O(lxl m + 1)
where Fo is a homogeneous polynomial of degree m and
(8.7.3)
Proof Let y be a fixed real vector and set

g(t, s)=F(t8+sy).
Then g(t,0)=crm+0(tm + 1 ), c=l=O, and we claim that
g(t, s)=O(ltl +Isl)m at (0,0).
If this is not true then the largest 2 such that
g(t, s) = O(ltl + Isll)m at (0,0)
is a rational number with 1/m~2<1. Write 2=p/q where p and q are
positive relatively prime integers, and consider the limits
gt (w) = lim g(wlsll, s)!Jslml.
s_±O
If atisk is a term in g(t,s) withj+k/2=m then q divides m-j. Hence
gt (w)=cw m +( ± 1)P c 1 wm - q +( ± 1)2 P C2 wm - 2q + ...
where all ci are not equal to 0. From (8.7.2)' it follows that 1m w ~ °

for the zeros of g~(w), for if g(wlsll, s)=F(Re wlsl l 8+sy+ i 1m wlsll8)=0
° °
and s is small it follows that 1m w~O. Now we can find a number
z =1= such that g~ (w) = if wq = ( ± 1)P z. All such w cannot lie in a half
plane unless q=2 and p is even, which contradicts that 1 ~p<q.
Hence 2 = 1, and since y is arbitrary we conclude that F(x)
Fo(x) = lim F(BX)/B m

.~o
is a homogeneous polynomial of degree m. By the argument above it

follows from (8.7.2)' that
Fo(x+w8):j=0 if xE1R n and 1m w>O.
Hence Fo(x+w8)=( -It Fo( -x-w8):j=0 if XE1R n and Imw<O,
Remark. In the proof that g(t, s) = O(ltl + Islt it would have been
sufficient to assume that for small real sand small It I
g(t,s)=O = Imt<Clsl.
In fact, for the zeros of g(wlsl". s)!lslmJ. we have 1m wlslJ. ~ Clsi then.
Since A< 1 it follows that 1m w ~ 0 if w is a zero of gt. This obser-
vation will be useful in Chapter XII.
Lemma 8.7.3. Let Fo be a homogeneous polynomial satisfying (8.7.3).

Then the component T of 8 in {xE1R"; Fo(x)=FO} is a convex cone. The
zeros t of Fo(x+ty) are real if XE1R" and YET; they are then negative
if and only if x is also in T. The coefficients of Fo(x)/Fo(8) are real.
Proof a) The zeros t of Fo(x+t8) are all real for every real x, for if
Fo(x+t8)=0 then Fo(x+Ret8+iImt8)=0, hence Imt=O by (8.7.3).
This implies that the quotient Fo(x)/Fo(8) of the lowest and the high-
est coefficient in this polynomial in t is real.
b) Set
Then T6 is open, and 8ET6 since the zeros are -1 when x = 8. If Xo is

in the closure of T6 then Fo(xo+t8)=0 =
t~O, so XOET6 if Fo(xo):j=O.
Thus T6 is open and closed in {xE1R n, Fo(x)=FO}, so T6 contains the
component T of 8 there.
c) If xET8 then
Fo(BX+(1-B)8)=~Fo(x+(I-B)B-18)=F0 if O<B~l.
Hence Fo =F 0 on the line segment between x and 8. In particular

T6 c T so these cones are identical.
d) If YET and B>O is fixed, then
Ey={XE1Rn; Fo(x+iB8+isy)=0 = Res<O}
is open, and OEEy since Fo(iB8+ isy) = (is)m Fo(B8/s+ y)=O implies
s<O. If x is in the closure of Ey then Fo(x+iB8+isy)=0 implies
Res::;::;O, and Res=O would contradict (8.7.3). Hence XEEy so Ey=Rn

and
if xER n, YEr and e >0.
Since r is open it follows that
Fo(x+iYHO if xERn, YEr.
Thus Fo(x+ty) has only real zeros then (see a)). Since the component
of yin {xERn; Fo(x)=I=O} is equal to r, it follows from b), c) that the
zeros are negative if and only if XEr, and then the line segment
between x and y also lies in r. The proof is complete.
Homogeneous polynomials satisfying (8.7.3) are called hyperbolic

with respect to 8. We shall resume their study in Section 12.4. Howev-
er, what we need to prove now is that F has essentially the properties
proved for Fo. The proof will be based on the idea in part d) of the
proof of Lemma 8.7.3.
Lemma 8.7.4. If r l is a closed cone contained in r u {O} then one can

find b > 0 such that
(8.7.4) blylm::;::;IF(x+iy)1 if YErI' XER n, Iyl<b, Ixl<b.
Proof Let K be a convex compact subset of r containing 8, which
generates a cone containing r l and 8. Since
F(rz)/r m ~ Fo(z) if r ~ 0
and Fo =1= 0 in K, we can choose r > 0 so that
(8.7.5) F(tyHO if tE<C, O<ltl::;::;r, YEK.
Note that t=O is a zero of order m. With YEK consider the equation
F(x+ie8+isy)=0.
When (x,e)ERn + I and Ixl+lel is small enough there are exactly m
roots s with Isl<r, for there is none with Isl=r and when X=8=0
there are m roots s = O. If x = 0 and y = 8 the roots are s = - 8 so they
are negative if 8> 0, as we now assume. If s is a root with Re s = 0
then F(x-yIms+i88)=0 which contradicts our hypothesis (8.7.2)'. It
follows that for small Ixl+8, 8>0, and YEK, we have m roots with
Res<O, Isl<r. Letting 8--+0 we conclude that F(x+isy) has m zeros
with Res::;::;O and Isl<r.
If f(z) is an analytic function in {ZE<C; Izi ::;::;r} and has m zeros
Zl ' ... , Zm in the disc, then
IIi (Z-Zj)I/'f(Z)' ::;::;(2rr Iwl=r

1
sup 1/lf(w)1
by the maximum principle. When Re Zj ~ 0, j = 1, ... , m, it follows that

1/lf(z)I~(2r/Rez)m sup 11/f(w)l, if Izl<r, Rez>O.
Iwl=r
If we apply this to F(x+izy) it follows for small XElR n and YEK that
1/IF(x+isy)1 ~(2r/s)m sup 1/IF(x+iwy)l, O<s<r,

Iwl=r
and this proves (8.7.4).
Weare now ready to prove the main result of the section:
Theorem 8.7.5. Let F be real analytic in the open set X clR n and
microhyperbolic with respect to 8ElRn. If XEX we denote by F,,(y) the
lowest homogeneous part in the Taylor expansion of y -+ F(x + y). Then
F,,(8HO and the component rx of 8 in {YElRn; Fx(y),*O} is an open
convex cone. If r; is the dual cone then
P={(x, 0; XEX, ~ErxO} c T*(X)
is closed. The limit
Fe- 1 = lim F(. + ie8)-1
£-+0
exists in !?&'(X), and
(8.7.6)
The canonical one form OJ = <~, d x) vanishes in r o in the sense that
if t -+ (x (t), ~(t))E ro is a Coo curve then <~(t), x' (t) = 0.
Proof The existence of the limit and the inclusion (8.7.6) are con-
sequences of Theorem 8.4.8 and Lemma 8.7.4. If (x o, ~o)¢rO then we
can find YoErxo so that <Yo, ~o) <0. By Lemma 8.7.4 F is also micro-
hyperbolic with respect to Yo in a neighborhood U of x o, hence
F,,(Yo)'*O when XEU in view of Lemma 8.7.2. But this implies that
(x, ~)¢rO if XEU and <Yo, 0<0, so r o is closed.
If I is an open interval on JR and I3t~(x(t), ~(t)) is a Coo curve
contained in r o then the degree of Fx(t)(Y) is locally constant in an
open everywhere dense set. In fact, if J is an open interval c I and
m is the minimum of the degree of Fx(t) when tEJ, then the degree
of Fx(t) is equal to m for all t in an open interval J' c J because it
is upper semi-continuous. By Taylor's formula we have for small y
F(x(t)+y)= L RAt,y)y"', tEJ',
l"'l=m
where R",EC oo . Differentiation with respect to t gives
<F' (x (t)+ y), x'(t) = L aR",(t, y)/aty"'.
l"'l=m
Hence <0 Fx(t)(y)/o y, x'(t» =0, for the Taylor expansion of the left-hand
side starts with this polynomial of order m -1. This means that Fx(t)
+IRx'(t)=Fx(t), so <e,x'(t»=O if eEr'~:(t), which proves that
<e(t), x'(t» =0, tEJ'. Hence <e(t),x'(t»=O, tEl, for this is proved in
a dense subset.
Remark. It is not possible to have an inclusion (8.7.6) where ro has

smaller convex fibers. In fact, assume that ro is a closed convex proper
conic neighborhood of the fiber of W FA (F9- I) at Xo. Then it follows
from Theorem 8.4.15 and the remark after its proof that there is a
function G with boundary value F9- 1 which is analytic in
(8.7.7) {x+iy; Ix-xol<c5, IYI<c5,YEr~}
for some c5>0. Now continuous analytic functions are uniquely de-
termined by their boundary values, and Theorem 3.1.15 shows that
this is still true when the boundary values are assumed in the distribu-
tion sense. Thus G is an analytic continuation of l/F, so F =1= 0 in the
set (8.7.7). Hence
F:o(Z) =lim F(xo + ez)/em
.~o
has no zero with 1m z in the interior of 1:.0°, so J:oo c f Xo and r.o:::::J FO

xo·
We shall now discuss the example
F(x)=xi-x~- ... -x;, 8=(1,0, ... ,0),
mentioned at the beginning of the section. Then Theorem 8.7.5 gives
WFA (F9- I )c {(x, tF'(x», F(x)=O, tX I >O} u {(O, y); Yl ~O, F(y)~O}.
On the other hand, WFA (F9- 1) must contain the first set on the right-
hand side since sing supp Fo- I is the set of zeros of F. Hence it also
contains the closure which is the boundary of the second set apart
from O. However, when n=4 there is nothing else in WFA (F9- 1 ). To
prove this we observe that by (7.4.7) the Fourier transform of FO- 1 is a
multiple of c5(ei - e~ - ... - e!) restricted to e1> o. Hence it follows
from Theorem 8.4.18 that (0,Y)EWFA (F9- 1 ), y=l=O, is equivalent to
YI >0 and F(y)=O which proves the assertion.
Notes
That singularities should be classified according to their spectrum was
recognized independently and from different points of view by several
Notes 323
mathematicians around 1970. The first was perhaps Sato [3,4J (see
also Sato-Kawai-Kashiwara [1]) who introduced and studied for hy-
perfunctions u a set SS(u) (called the singular support) which is our
WFA(u) in the case of distributions. As proved by Bony [3J it is also
equal to the essential support of Bros and Iagolnitzer (see Section 9.6
and Iagolnitzer [1]). WF(u) was first defined by Hormander [25J by
means of pseudo-differential operators. This definition, which will be
given in Section 18.1 below, was in fact more or less implicit in
standard methods for localization by means of such operators. The
equivalent definition of WF(u) used here comes from Hormander
[26] where the results of Section 8.2 were also proved. In Section 8.4
we start with the definition of W Fdu) given in Hormander [27] but
shift to equivalent definitions closely related to those of Sato by
means of an analytic decomposition of the b function. This is quite
similar to the decomposition of b used in Sato-Kawai-Kashiwara [1,
p.473] and Bony [3], but the analyticity of the decomposition is an
essential advantage. This was pointed out to us by Louis Boutet de
Monvel; see also the related exposition by his student Lebeau [1]
and the survey by Schapira [2].
The wave front set was introduced by Hormander [25] to
simplify the study of the propagation of singularities. Note that results
like Theorem 8.3.3' on the wave front set are entirely local and
therefore easier to prove than the corresponding weaker results on
singsuppu. Indeed, these state in the simplest form that if P(D)UECOO
and OEsingsuppu then JRP~(e)csingsuppu for some e with Pmm=O.
This was first proved by Grusin [lJ who constructed a fundamental
solution with singular support contained in any "half' of the bichar-
acteristic cone obtained by projecting Char P in JRn. (The method was
extended to the analytic case by Andersson [1].) The fundamental
solution must be adapted to the distribution u being studied. Here on
the other hand we have just needed two natural fundamental solutions
E± (with properties more or less classical in quantum electrodynamics
in the case of the Klein-Gordon equation). Particularly in the analytic
case and for variable coefficients this eliminates considerable technical
difficulties. Conceptually it is of course a great advantage that one
knows unambiguously in which direction a singularity described by a
point in WF(u) is going to travel. For the sources of Example 8.3.4
and Theorem 8.3.8 see Zerner [1,2] and Hormander [24].
The results on differential operators in Sections 8.3 and 8.6 are
merely intended as examples. The third part of this book will mainly
be devoted to the study of WF(u) for solutions of (pseudo-)differential
equations. In the analytic case there is also a vast theory of W FA (u),
usually even for hyperfunction solutions. We refer the reader to Sato-
Kawai-Kashiwara [1], Kashiwara [1], Sjostrand [1, 2] and the ref-

erences given there.
The Holmgren uniqueness theorem (Theorem 8.6.5) was proved by
Holmgren [1] in a special case and by John [1] in full generality for
classical solutions. The key to the proof is a result on analyticity of
integrals over non-characteristic surfaces depending on a parameter
for solutions of differential equations. This was used by John [1] to
prove analyticity of solutions of elliptic equations and related reg-
ularity results. As observed in Hormander [27] and independently by
Kawai (see Sato-Kawai-Kashiwara [1, 470-473]) one can now reverse
the order and deduce uniqueness theorems from microlocal regularity
theorems. The purpose of this was to prove uniqueness theorems
related to Theorem 8.6.13 in the case of characteristic boundaries.
Unique continuation across a surface E at a characteristic point where
E is strictly convex with respect to the corresponding tangential
bicharacteristic was proved in the predecessor of this book by
geometrical arguments combined with the Holmgren theorem. Suc-
cessively refined geometrical arguments were then given by Bony
[1,2] and Hormander [28]. They are now superseded by Theorem
8.5.9 which is due to Sjostrand [1]. One of the original results of
Bony is presented as Theorem 8.6.6. The construction in Example 2
following it is due to Baouendi and Goulaouic [1]. (Hypoellipticity of
such operators will be proved in Chapter XXII where further ref-
erences are given.) Theorem 8.6.7 is from Hormander [1]; it was
proved in Hormander [9] that the null solutions are dense in all
solutions in COO({x; <x,
N) >O}) if P(D) has no non-characteristic
factor. Theorem 8.6.8 - Corollary 8.6.11 are close to results of John
[1] and were proved as stated here in the predecessor of this book.
Further relations between suppu and WFA(u) will be discussed in
Section 9.6.
K.G. Andersson [1] introduced the notion of local hyperbolicity
with respect to e which is the conjunction of microhyperbolicity with
respect to e and - e, and Ggrding [5] continued his study. Microhy-
perbolicity was defined as here by Kashiwara and Kawai [1] who used
the local Bochner tube theorem (see Komatsu [1]) to prove the
crucial Lemma 8.7.4. The reader is referred to Chapter XII for further
information in this context.
Chapter IX. Hyperfunctions
Summary
We defined .@'(X) as the space of continuous linear forms on C;;'(X).
This is by no means the most general concept of its kind, for a larger
space of distributions is obtained if C;;'(X) is replaced by a dense
subspace with a stronger topology. An example is the space of ele-
ments of compact support in CL (defined in Section 8.4) provided that
it does not contain just the 0 function, that is,
~)/Lk<oo.
The study of the dual space of distributions is then fairly similar to
that of .@'(X).
The situation is rather different in the quasi-analytic case where
L l/Lk= 00.
No analogue of C;;' (X) is then available but we may regard CL(X) as
a substitute for C<Xl(X). The dual space of CL(X) can be taken as the
elements of compact support in a distribution theory preserving many
of the features of .@'(X) but differing in some respects. The largest
space of distributions is obtained when C L is the real analytic class. It
was introduced in a different way by Sato who coined the term
hyper function for its elements. In this chapter we shall give an in-
troduction to the theory of hyperfunctions in a manner which follows
Schwartz distribution theory as closely as possible.
Section 9.1 is devoted to the study of hyperfunctions of compact
support. In particular we give an elementary proof of the crucial and
non-trivial fact that there is a good notion of support. The general
definition of hyperfunctions can then be given in Section 9.2 along
lines first proposed by Martineau. Section 9.3 is devoted to the wave
front set with respect to analytic functions of a hyperfunction and the
definition of operations such as multiplication. This is done rather
quickly for most proofs in Sections 8.4 and 8.5 were chosen so that
326 IX. Hyperfunctions
they are applicable to hyperfunctions after a few basic facts have been
established.
Section 9.4 is devoted to the existence of analytic solutions of
analytic differential equations. In addition to the classical Cauchy-
Kovalevsky theorem precise information on bounds and existence
domains is given. These are applied in Section 9.5 to prove some basic
facts on hyperfunction solutions of analytic differential equations. In
Section 9.6 finally we present the Bros-Iagolnitzer definition of WFA(u)
and prove as an application a theorem of Kashiwara on the relation
between suppu and WFA(u) similar to Holmgren's uniqueness theo-
rem.
9.1. Analytic Functionals
If K is a compact set in JR n then a distribution u E@"'(K) is a linear

form on coo(JRn) such that, if OJ is a neighborhood of K,
lu(4))I;;:;; c'" L sup ID"4>I, 4> E COO (JRn).
l"l~k '"
One can extend u(4)) by continuity to all cfJ E COO (OJ). (See Theorem
2.3.1 and the remarks after its statement.) Since the derivatives of an
analytic function can be estimated in a compact set by the maximum
of the modulus in a neighborhood, the following definition is quite
analogous:
Definition 9.1.1. If K c(Cn is a compact set, then A'(K), the space of

analytic functionals carried by K, is the space of linear forms u on the
space A of entire analytic functions in (Cn such that for every neigh-
borhood OJ of K
(9.1.1) lu(4))I;;:;;C,,,suplcfJl, cfJEA.
'"
Example. u(cfJ)=La"D"4>(O)/cd is an analytic functional carried by 0 if
and only if la,,1 ~ C.el"l for every e > O. (The sufficiency follows from
Cauchy's inequalities and the necessity by taking 4>(z) = z".) u is not a
distribution unless the sum is finite.
It would suffice to consider only polynomials cfJ in the definition,

for every entire analytic function is locally uniformly the sum of its
Taylor series. Note that A'(K) is a Frechet space with the best con-
stants C",(u) as semi-norms.
In contrast to what one might expect from the analogy with iC'(K)
it is not always true that uEA'(K 1 )nA'(K 2) implies uEA'(K I nK2).
For example,
1
u(¢)= J¢(z)dz, ¢EA(<C 1),
o
has any C 1 curve from 0 to 1 as a minimal carrier. However, we
shall prove that this is true if K l' K 2 e 1R.n, which is the case we are
mainly interested in. As a first step in this direction we shall prove
that if uEA'(K) and Ke1R.n, then u(¢) can be defined for every ¢
which is analytic just in a neighborhood of K.
Proposition 9.1.2. Let K eR.n be a compact set, and set for e >0
K.={ZE<c n; IRez-xl+2IImzl~e for some xEK}.
For every ¢ which is analytic in a neighborhood V of K. one can then

find a sequence ¢ j E A such that
sup I¢ j - ¢I--+O, j --+ 00.

K£
Proof. Choose XEC~(V nR.n) equal to 1 near K.n1R. n, and set
where E j is the normalized Gaussian function
E j(x)=U!n)n/2 exp( - j<x, x»).
Since E j E A it is clear that ¢ j E A. The set K. is defined by
K.={z; 2IImzl~F(Rez)}, F(y)=e-minlx-yl.

XEK
By the triangle inequality IF(y)-F(y')I~ly-y'I. If zo=xo+iYoEK. it

follows that K. contains the chains
when
E/z -
°~t~ 1, and they have the same boundary. Since the form
0 ¢W d'l /\ ... /\ d'n is closed in K. we obtain by Stokes' formula
¢j(z) = J Ej(z-x) X(x) ¢(x)dx
Ix-xol > 21yol
+ J E/z-')¢Wd'I/\···/\d(n·
T(z o.l)
We take z=zo and observe that

Re- <zo-x, zo-x) = -Ixo _x1 2 + IY ol2 ~ - 3Ixo-xI2/4
if Ix-xol>2IYol,
Re-<zo-(, zo-O = -lxo-xI2+lx-xoI2/4= -3Ix- x oI2/4
if (EF(zo, 1), Re(=x.
Since 1<p(o-cf>(zo)I~Clx-xol in the integrals we conclude that
Icf> izo) - cf>(zo) JE/zo - x) X(x) dxl
~ Co JU/nt/2 e-3jlx-xoI2/4Ix_xol dx ~ C 1 j-t.
Now
1- JEj(zo - x)x(x)dx= J(1- X(x»E/zo - x)dx
is exponentially decreasing as j-HXJ since Re<zo-x, zo-x) has a
positive lower bound when Zo E K, and x E supp (1- X). This completes
the proof.
With the notation in the proposition we have by (9.1.1) if U E A'(K)

lu(cf»-U(cf>k)l~ C sup Icf>j-cf>kl--O as j, k--oo
KE
so we can define
U(cf» = lim u(cf>).
j-+ co
Since the sets K. form a fundamental system of neighborhoods of K,

it is clear that we have now obtained a unique definition of u(cf» for
every cf> which is analytic in a neighborhood OJ of K, and (9.1.1) remains
valid for all cf> analytic in OJ.
We shall now associate with every uEA'(K) a regularization from

which U can be conveniently reconstructed. Let E be the fundamental
solution of the Laplace operator in JR n+ 1 given in Theorem 3.3.2 and
set
P=i3E/i3Xn+1 =x n +1 IX I-n-1/cn+1'
Here X=(x,X n+1)EJRn+1. If UE Cg(K), K compact in JRn, then
J
U(X)=P * (u® (j)(X)=xn+ 1 u(Y)(lx - yI 2 +X;+ l)-(n+ 1)/2 dy/c n+1
is odd as a function of Xn+ l' harmonic outside K x {O} and converges
to ±u/2 as x n+1--±0. Thus u is the jump across the plane xn+1 =0 of
the harmonic function U, which is also clear from the fact that
,1 U =,1 P * (u ® (j) =U ® i3n+ 1 (j.
Very little has to be changed in this discussion if u E A' (K):
Proposition 9.1.3. If K is a compact set in JRn and uEA'(K), then

(9.1.2)
is a harmonic function in JRn+ 1'-.. (K x {O}) which is odd as a function
of Xn+ l' If cJ> is a harmonic function in JRn+ 1 and X E C~(JRn+ 1) is
equal to 1 in a neighborhood of K x {OJ, then
(9.1.3)
Proof. U(X) is defined when XrtKx{O} for y-+P(X-(y, 0)) is then

analytic in a neighborhood of K. The continuity of u implies that U is
continuous and that we may compute the derivatives of U by differen-
tiating on P in (9.1.2). Since P is harmonic outside 0 it follows that U
is harmonic. To prove (9.1.3) we note that if YEJRn+l and X=1 near
Y then
SP(X - Y) LI (XcJ>)(X)dX = <LlP(X - Y), XcJ>(X) = <on+ 1 by, XcJ»
= - On+ 1 cJ>(Y).
By the uniqueness of analytic continuation this remains true for all Y
in a complex neighborhood of K x {OJ in <cn + 1, and the left-hand side
is then the uniform limit of the corresponding Riemann sums. Setting
Y =(y, 0) and letting u operate on each term in the Riemann sum, we
obtain (9.1.3).
(9.1.3) determines u completely, for we have
Lemma 9.1.4. For every entire analytic function ¢ in <c n there is a

unique entire analytic function cJ> in <c such that
n+ 1
n+l
(9.1.4) L 02cJ>/OZ;=0, and cJ>=O,On+lcJ>=¢ when zn+l=O.
1
For every R> 1 there is a constant CR such that

(9.1.5) 1cJ>(Z,Zn+l)I~CRlzn+ll sup I¢(QI.
I_-zi <Rlzn+d
Proof. If cJ> satisfies (9.1.4) then

u(t, x)=cJ>(z+iZn+ lX, Zn+ It); tEJR, xEJRn;
satisfies the wave equation 02 ujot 2 = LI xu, and
u=O, oujot=Zn+ 1 ¢(z+izn+ lX) when t=O.
Hence it follows from Theorem 6.2.4 that
cJ>(z, Zn+ 1) =Zn+ 1<E + (1), ¢(z + iZn+ 1.)
where E+(1) is a distribution with support in {XE1R.n; Ixl~1}, given

explicitly by (6.2.4)'. This proves the uniqueness and (9.1.5). On the
other hand, if we define cI> by this formula we obtain an entire
analytic function which satisfies (9.1.4) when zn+ 1 >0 and zji is real.
But the entire function I,02cI>joz; must vanish identically if it van-
ishes in this set, so the lemma is proved.
Remark. A direct proof can also be made by estimating the terms in

the power series expansion
00
cI>(z, zn+l)= I,z;!;/( _A)k¢(Z)j(2k+ 1)!.

o
As already pointed out Lemma 9.1.4 implies that the map from
analytic functionals to harmonic functions defined by (9.1.2) is in-
jective. Using Lemma 9.1.4 we shall now prove that it is also sur-
jective.
Proposition 9.1.5. If K is a compact set in 1R.n and U is a harmonic

function in 1R.n+ 1 '- (K x {O}) which is odd as a function of Xn-t l' then
there is a unique uEA'(K) such that (9.1.3) is valid when XEC~(1R.n+l),
X= 1 near K x {O} and cI> is any harmonic function in 1R.n + 1. We have
where H is harmonic in 1R.n + 1, and H vanishes identically if and only if

U~O at 00.
Proof. The right-hand side of (9.1.3) is independent of the choice of X,

for it is equal to 0 if XEC~(1R.n+l'-(Kx{O})). For any (bO we may
therefore choose X so that every point in supp X has distance < (j to
K x {O}, and we can always take X even as a function of xn+ l'
Then (9.1.3) is automatically true if cI> is even as a function of xn+ l'
When ¢ is a polynomial in (Cn we now define
(9.1.6) u(¢)= - JU A(XcI»dX
where cI> is given by Lemma 9.1.4. Taking R=4j3 in the lemma we
obtain
IA (xcI»1 ~ C~ sup I¢I,
K76
for if Ix-YI2+X;+1 <(j2 for some YEK then Iz-xl~4Ixn+tlj3 implies

Iz-yl<7(jj3, hence zEK7~' This proves that uEA'(K). For reasons of
continuity it follows now that (9.1.3) is valid for every entire harmonic
cI> which is odd with respect to Xn+ l ' hence for all entire harmonic cI>.
9.1. Analytic FunctionaIs 331
For the harmonic function in JR"+ 1 ....... (K x {O}) defined by

U1 (X) = Uy(P(X - (y, 0»)
we know from Proposition 9.1.3 that (9.1.3) is valid with U replaced
by U 1 • Writing H = U1 - U we obtain
JH A (X 4» dX =0
for every entire harmonic 4>. Now choose Xl E C~(JRn+ 1) so that Xl = 1
in a neighborhood of Kx{O} and X=l in SUPPX1' Then (l-Xl)H
=H1 ECOC> and for all exponential solutions 4> of A in JR"+1 we have
0= JH 1 A (X 4» dX = J(A H l)4>dX.
Hence it follows from Theorem 7.3.2 and Lemma 7.3.7 that AH 1 =Af
for some fE C~. This means that H 1 - f is a harmonic function which
is equal to H outside a compact set, and therefore outside K x {O}.
Thus H has been extended to a function which is harmonic in JR"+1.
Since H = 0 is equivalent to H -+0 at 00 by the maximum principle,
and since U1-+0 at 00, the last statement in the proposition follows.
We obtain (9.1.3) for every harmonic function in JR"+ 1 since this is
true with U replaced by U1 and since
JH A(X4»dX =0.
We are now ready to prove some important facts on the elements
of
A'(JR")= U A'(K).
KelRn
Theorem 9.1.6. IfuEA'(JRn) then there is a smallest compact set KcJR"

such that uEA'(K); it is called the support of u.
Proof. Let K be the intersection of all compact sets K' cJR" such that
uEA'(K'). By Proposition 9.1.3 a harmonic function in JRn+l . . . . (K'
X {O}) is defined by (9.1.2). It is uniquely determined by its restriction
to the complement of the plane x"+ 1 =0. The functions obtained for
different choices of K' must therefore agree in their common domain
of definition and give together a harmonic function in JRn+ 1 ....... (K
x {O}). Hence uEA'(K) by Proposition 9.1.5.
Next we shall prove a completeness theorem for analytic function-

also In order to prepare for the construction of boundary values of
analytic functions in Section 9.3 we shall then consider some analytic
functionals which are carried by compact sets close to JR" but not
contained in JRn. This requires another look at Propositions 9.1.3 and
9.1.5. If uEA'(KJ where K, is defined in Proposition 9.1.2, then (9.1.2)
defines a harmonic function U in the complement of

K.={XEIRn+l; Ix-YI2+X;+1~e2 for some YEK}.
This will follow if we just show that
(9.1.7) Re(x-z, x-Z)+X;+l>O if X¢K. and zEK•.
The left-hand side is equal to
Ix-RezI 2 +x;+1-IImzI 2
and for some Y E K we have
IRez- yl +2lImzl ~e.
Since X ¢ K. it follows that
e2 <lx-YI2+x;+1'
so the triangle inequality gives
(Ix- Rezl 2+x;+ l)t>e-IRez- yl ~ IImzl
which proves (9.1.7). On the other hand, we actually saw in the proof
of Proposition 9.1.5 that U harmonic outside K. implies uEA'(K 7 .).
We are now ready to prove the crucial completeness result:
Theorem9.1.7. Let Ko and K be compact sets with KocKclR n , let

ujEA,(<c n ) and assume that
(i) For any compact neighborhood V of K in (Cn we have ujEA'(V)
for large j.
(ii) For any compact neighborhood Vo of Ko in (Cn we have
Uj-UkEA'(Vo) for large j, k.
Then one can find uEA'(K) so that for any compact neighborhood Vo of Ko
(iii)
Condition (iii) determines U uniquely modA'(K o).
Proof. If (iii) is fulfilled with U replaced by v also, then U- v E A' (Vo)

for every Vo so u-vEA'(K o) as claimed. Thus we only have to prove
the existence of u. Choose a sequence er--+O so that with the notation
in Proposition 9.1.2
(i)'
(ii)' Uj-UkEA'(K o,.) if k~j.
As we have just seen it follows from (i)' that

UiX) = U jy P(X -(y, 0))
is a harmonic function outside K oj , and fro~ (ii)' that Uj - Uk has a

harmonic extension to the complement of KO,oj if k > j. By Runge's
approximation theorem (Theorem 4.4.5) we can approximate Uj+ 1
- Uj in CKo,oj l?y functions harmonic in the complement Q of K o
x {OJ. In fact, KO,Oj is not the union of two disjoint non-empty
compact sets one of which is disjoint with K o x {OJ, for it is a union
of balls with center in K o x {OJ. Let
Mj={XE1R.n+1; IXI~j, Ix-YI2+x;+1~2ef for all YEK o}'
This is a compact subset of the complement of Ko,oj and it increases
to Q when j-+ 00. We can therefore choose lj harmonic in Q so that
(9.1.8) IUj+ 1- Uj-ljl ~2-j in M j .
(Strictly speaking Uj + 1 - Uj should be replaced by the harmonic ex-
tension to the complement of Ko,oj') Since lj can be replaced by
(lj(x, Xn+ 1) -lj(X, - Xn+ 1))/2 in (9.1.8), we can take lj odd as a func-
tion of xn+ l '
It follows from (9.1.8) that the limit
U =lim(Uj - V1- ... -lj-1)

00
= Uj - V1-· .. -lj-1 + I(Uk+1 - Uk - v,,)

j
exists and is harmonic outside K x {OJ, for the sum is harmonic

outside K o •. by (9.1.8), and the other terms are harmonic outside Ko J..
'~J
Let U be the corresponding element in A'(K). Since U - Uj is har-

monic outside Ko,oj we have u-uj EA'(K o,70) which proves (iii).
The following theorem is a substitute for the existence of partitions

of unity:
Theorem 9.1.8. If K l ' ... , Kr are compact subsets of 1R.nand

UE A'(K 1 U ... U K r ), then one can find UjE A'(K) so that
U=U 1+ ... +ur·
Proof. It is sufficient to prove the statement when r = 2. The function

U defined by (9.1.2) is harmonic outside K1 uK 2 where Kj=K j x {OJ.
The theorem will be proved if we can split U into a sum
U=U 1+U2
where Uj is harmonic outside Kj • To do so we choose using Corollary
1.4.11 a function = 1 near K1 '-.(K1 nK 2). Then
U1 =4>U-v, U2 =(1-4»U+v
have the required properties ifvEc oo (IRn+1'-.(K 1nK 2 )) and
Llv=Ll(4)U).
Here we define 4>U=O near K2'-.(K1nK2) and (l-4»U=O near K1
'-.(K1nK2)' Since Ll(4)U) vanishes near (K1 UK2)'-.(K1nK2) it is a
COO function outside K1 nK 2 • The existence of v is therefore a con-
sequence of Theorem 4.4.6. (Note that this is based on another appli-
cation of Runge's approximation theorem.)
Finally we note that if U EIf' (JR n) then U defines an element in
A'(JR n) with the same support. In fact, the harmonic function
U(x,x n + 1) defined by (9.1.2) has the~' limit ±u/2 as Xn+ 1--+ ±O, so
continuation of U as a harmonic function is only possible outside
supp U x to}. Thus we have an injection preserving supports
If' (JR n) "-+ A' (JR n).
The operations defined for distributions in Chapters III to VII
carryover easily to A'(JRn). We shall just recall them briefly and leave
all details for the reader.
a) If uEA'(JR n) then ojuEA'(JRn) can be defined by
(OjU)(4» = -u(Oj4»
when 4> is an entire function, for sup IOj4>1 can be estimated by the
(fJ
supremum of 14>1 over a neighborhood of w.

b) If uEA'(K), KeJRn, and f is analytic in a neighborhood w of
K, then we define the product fu by
(ju)(4» = u(j 4»
when 4> is analytic in w. Here it is of course important that Proposi-
tion 9.1.2 allowed us to extend U to all functions analytic in w.
c) If U EA'(JRn) and v E A'(JRm) then U Q9 v E A'(JRn+m) is defined by
(UQ9v)(4» =ux (v y (4>(x, y))) = Vy (u x (4) (x, y)))
when 4> is a polynomial in (Cn+m. The second equality is obvious then
and the first is a definition; it is clear that we obtain an analytic
functional supported by supp U x supp v.
d) If KEA'(JRnxJRm) and v is analytic in a neighborhood of the
projection of supp K in JRm, then % v E A' (JR n) is defined by
(% v)(4)) =K(4) Q9 v)
for every entire function 4> in (Cn.
e) If U E A' (K), where K is a compact set in 1Rn, and if f is a real

analytic bijection of an open set wc1Rn on a neighborhood of K with
inverse h, then the pullback f* U E A' (f* K) is defined by
(f* u)(4)) = u«4> h) Idet h'!)

0
when 4> is analytic in w.

f) If u, V E A' (1R n) then c), d), e) allow us to define U * v by letting
the pullback of U ® v by the map (x, y)--+(y, x - y) act on the function
1.
g) If uEA'(K), K compact in 1Rn, then the Fourier-Laplace trans-
form
am=U(exp -i(., 0)
is an entire analytic function such that for every e > 0
laml ~ C. exp(HK(Im ')+eIW, (E(Cn.
This follows at once from the definition of A'(K). Conversely, every

entire function satisfying these estimates is the Fourier-Laplace trans-
form of a unique element in A'(K). The uniqueness follows from the
fact that
U(P) = P( - D) 12(0)
for every polynomial P. The existence proof will be given as Theorem
15.1.5, and the result will not be used in the meantime.
9.2. General Hyperfunctions

We want to define hyperfunctions in 1Rn in such a way that they are
locally equivalent to analytic functionals with compact support in 1Rn.
This will be done in two steps.
Definition 9.2.1. If Xc 1Rn is open and bounded we define the space of

hyperfunctions B(X) in X by
(9.2.1) B(X) =A'(X)jA'(oX).
The reader might object here that this does not give the desired
result in the case of distributions, for
tC'(X)/tC'(OX)L-+!0'F(X).
However, the definition will be justified in a moment when we prove
that the analogue of the Localization Theorem 2.2.4 is valid.
If u, vEA'(X) and u-vEA'(aX), then X nsuppu=X nsuppv since

supp u e supp V u ax and supp v e supp U u ax. Thus it is legitimate to
define the support of the class u· of u in B(X) by
suppu'=X nsuppu.
If YeX and X, Yare open and bounded, we can for every
uEA'(X) find vEA'(Y) so that Ynsupp(u-v)=0. This follows from
Theorem 9.1.8 applied to u and the compact sets Y and (X'- Y). The
class v· of v in B(Y) is uniquely determined by the class u· of u in
B(X) and is called the restriction of u· to Y. Note that the restriction
of u· to Y is 0 if and only if Ynsuppu=0. As in the case of
distributions the support of a hyperfunction is therefore the smallest
closed set such that the restriction of the hyperfunction to the comple-
ment is equal to O. The definition of sing supp u is also extended to
hyperfunctions with no change.
We can now state and prove the localization theorem.
Theorem 9.2.2. Let Xj be open sets in JR." with bounded union X. If

ujEB(X) and for all i,j we have ui=u j in XinX j (that is, the re-
strictions are equal) then there is a unique u E B(X) such that the
restriction of u to Xj is equal to U j for every j.
Proof. The uniqueness is clear for if v has the same property as u then
the support of u - v is empty so u - v = O. To prove the existence we
begin with the case of just two open sets X 1 and X 2' Choose
~j E A'(X j) defining u j for j = 1, 2. The support of Vi - V 2 is contained
CX 2)'
III
(Xl uX 2)'-(X 1 nX 2)e(CX 1 nX 2)U(X 1 n
so Theorem 9.1.8 gives a decomposition
Vi - V 2 = Vi - V2 , Vi EA'(CX i nX 2)' V2 EA'(X 1 n CX 2)'
Now
V= Vi - Vi = V 2- V2 EA'(X 1 uX 2 )
defines an element in B(X 1 U X 2) which restricts to u j in Xj for j

=1,2.
Next we assume that we have countably many sets Xj' j = 1,2, ....
Repeated use of the special case just proved gives a sequence Vj in
B(X 1 u ... u X j) with restriction ui to Xi for i ~j. Let
l-jEA'(X 1 u ... uX j ) be in the class of Vj' Since supp(l-j- Y,;)eX
'- (X 1 U ... U X) whe.? k > j, it follows from Theo~m 9.1.7 that there is
an element VEA'(X) such that supp(V-l-j)eX,-(Xiu ... uX j) for
every j. Hence the class u of V has the desired restrictions.
If we have more than countably many Xj we just choose count-

ably many of them with the same union and· then a corresponding u.
The uniqueness established at the beginning of the proof shows that
the restriction of U to Xj is then uj for every j.
It follows from Theorem 2.2.4 and the remarks at the end of
Section 9.1 that we have an injection ~'(X)--+B(X). Let us also note
here that the elements with compact support in B(X) can be identified
with the elements in A' (lR") having support in X. In fact, let U E A' (X)
and assume that the class u· has compact support K eX. Then
supp U e K U 0 X so Theorem 9.1.8 gives a decomposition
u=u 1 +u 2 , u 1 EA'(K), u2 EA'(oX)
which is unique since K and ax are disjoint. This means that u· =u~
for a unique u 1 EA'(K).
It is easy to extend the operations on A'(lR") discussed at the end
of Section 9.1 to operations on B(X). First it is clear that if X and Y
are bounded open sets in lRn and f is a real analytic diffeomorphism
of a neighborhood of Y on a neighborhood of X, then we obtain a
bijection
1*: B(X)--+B(Y)
from the bijections
f*: A'(X)--+A'(y) and 1*: A'(aX)--+A'(ay).
The easy proof that
(fg)* =g*f*
is left for the reader.
We can now define B(X) for any real analytic manifold X. First
we choose an atlas !#" of analytic diffeomorphisms of coordinate
patches X ,/t:::::, X on open sets XK ~ lR" such that K has an analytic
extension to a neighborhood of the closures. Then
KK'-l: K'(XKnXK,)--+K(XKnX K ,), K, K'E!#",
has an analytic extension to a neighborhood of the closures, so
(K K'-l)*: B(K(X K n X K')) --+ B(K'(X K n X K'))
is defined. We can therefore define a hyperfunction U E B(X) as a
collection of hyperfunctions UK E B(X K) for every K E!#" such that (6.3.3)
is valid. The easy but tedious proof that B(X) is independent of the
choice of atlas and that it agrees with our previous definition when
X ~ lR" is left for the conscientious reader.
The notion of support and restriction carryover immediately to

the general case. A final justification of Definition 9.2.1 is given by
Theorem 9.2.3. If X is a real analytic manifold and Y an open subset

then every uEB(Y) is the restriction to Y of a hyperfunction vEB(X)
with support in Y.
Proof Let Ie X" -> X" be a coordinate system E ff on X. Then

u"EB(K(YnX.,)) is the class of an element UEA'(K(YnX,J) which
also defines a hyperfunction VEB(X,,) since K(YnX")cX,,. The re-
striction of V to K(YnX,,) is equal to u". The desired extension of u to
YuX" is now obtained immediately if to an atlas for Y with coor-
dinate patches ~ Y we add the coordinate system K with the hyper-
function V. Continuing in this way we can successively extend u to all
of X. (If X is not countable one should use Zorn's lemma but we are
not interested in such generality.)
The extension of Theorem 9.2.2 to a real analytic manifold X with

open subsets Xj is immediate. So is the definition of the product fu of
a hyperfunction u EB(X) by a function f which is real analytic in a
neighborhood of supp u, as well as the definition of the tensor pro-
duct.
9.3. The Analytic Wave Front Set of a Hyperfunction
Definitions 8.1.2 and 8.4.3 of WF and WFL make no sense for hyper-
functions but it is possible to use the equivalent characterization in
Theorem 8.4.11. For the sake of brevity we shall only discuss WFA .
With K still denoting the analytic function in {z; 11m Zl2 < 1 + IRe Z12}
constructed in Lemma 8.4.10 we first prove an analogue of a part of
Theorem 8.4.11.
Proposition 9.3.1. If uEA'(JR") then

U(z)=K * u(z)=utK(z -t)
is an analytic function in
Z={z; IImzI 2 <1+IRez-tI 2, tESUppU}.
If X is a bounded open neighborhood of supp u in JR" then
(9.3.1) u(4))=lim J JU(x+irw)4>(x)dxdw, 4>EA.
r_1 1"'1= 1 X
For any function U which is analytic when 11m zl < 1 and any bounded
open set X let 1"(U, X) be the set of all y E JR" with Iyl = 1 such that U
is analytic at x+iy for every XEOX. Then
J
U;(¢) = U(x+iy)¢(x)dx, ¢EA,
x
is in A'(X) if YE1R", IYI<l, and U; can be defined for all YEL(U,~)
so that U; remains a continuous function of y with values in A'(X).
Thus
U;(¢) = JUc!(¢)djl(w)=limHX U(x+irw)¢(x)dxdjl(w),
r~l
¢ EA,
defines an element of A'(X) for every measure djl with support in

L(U,X).
Proof. K (z - t) is an analytic function of t in a complex neighborhood

of supp u if z E Z. Hence U is defined in Z, and U is analytic since
.derivatives of U can be put on K. If ¢ E A and r < 1 then
J JU(x+irw)¢(x)dxdw=u(f/J r)
1"'1= 1 x
f/Jr(t)= J JK(x+irw-t)¢(x)dxdw
1"'1= 1 x
= J JK(t+irw-x)¢(x)dxdw.
1"'1= 1 x
(Recall that K is even.) By Theorem 8.4.11 f/J r-+¢ in .@'(X) as r-+1.
Choose X E C;;' (X) with 0 ~ X ~ 1 so that X = 1 in a neighborhood of
suppu. Then
JK(t+irw-x)¢(x)dx
x
= J K(t+irw-z)¢(z)dZ 1 A .•• Adz., O<e<l,
Y("" £)
by Stokes' formula, ')I(w, e) denoting the chain

X3x-+x+iex(x)w.
Letting r-+ 1 we conclude that f/Jr(t) has an analytic limit in a complex
neighborhood of supp u. The limit is necessarily equal to ¢(t), which
proves (9.3.1).
With the same notation we obtain in the second part of the
proposition
U;(¢)= J U(z+iy)¢(z)dZ 1 A ..• Adz.
y(w,£)
if Iyl < 1 and Iy+ewl < 1. This proves the asserted continuity in
{y; IYI= 1, <y, w) < -1/2, U is analytic at x+iy if XEX, X(x)=O}.
Since w ES·- 1 is arbitrary and X can be chosen equal to 1 on any

compact subset of X, the proof is complete.
Suppose with the notation in the first part of the proposition that
there is a point XoESUpp u such that U is analytic at Xo + iw when
Iwl = 1. Then there is a compact neighborhood M eX of Xo such
that U is analytic at M +is·- 1 • Hence
u(4J)=Ulc",M(4J)+ J JU(x+iw)4J(x)dxdw.
Iwl= 1 M
The restriction of u to the interior of M is therefore the analytic

function
J U(x+iw)dw, XEM.
Iwl= 1
Definition 9.3.2. If uEA'(1R·) then WFA(u) is the set of all (x, ~)E1R·
x (1R........ 0) such that U = K * u is not analytic at x - i ~/I~I.
We have just proved that the projection of WFA(u) in 1R. is

sing SUPPA u. Since WFA(u) is determined at x by the local properties
of u, the definition is immediately extended to general hyperfunctions
in an open subset of 1R•. To prove that the results of Sections 8.4 to
8.6 can be extended to hyperfunctions we shall work consistently with
boundary values of functions analytic in tube domains. The following
is an analogue of Theorem 3.1.15 and Theorem 8.4.8.
Theorem 9.3.3. Let X be an open set in 1R., r a connected open cone in

1R........ to} and f an analytic function in an open set Z e<C· such that for
every open set Xl ~X and closed convex cone r;, eru to} we have for
some y>O
Then
(i) there is an element fx I E A' (X 1) independent of r 1 and uniquely
determined modA'(oX 1) such that the analytic functional
A34J~fx,(4J)- Jf(x+iy)4J(x+iy)dx
XI
is carried by any given neighborhood of oX 1 in <C. when y E r1 and Iyl is
small enough. Thus fx, defines uniquely a hyperfunction in B~ 1)'
(ii) If X 2 eX 1 is another open set then fx, - fX2 E A'(X 1 ....... X 2) so
there is a unique fxEB(X) such that fx and fx, have the same restric-
tion to X 1 for every Xl'
(iii) If If(x+iy)I~Clyl-N, XEX 1 , YEll, Iyl<y, as in Theorem
3.1.15 then fx restricted to X 1 is equal to the distribution limit which
exists by Theorem 3.1.15.
(iv) If f can be continued analytically to a neighborhood ofax 1

then
¢---+S f(x+iy)cp(x)dx,
converges in A'(X 1)' when T1 3 y---+O, to an element satisfying the con-
°
dition (i) above.
(v) If fx = in some non-empty Xl e X and Z is connected, then f
=fx=O.
(vi) WFAUx) eX x(P'-.. {O}).
Proof. (i) The analytic functional
cp---+ S f(x+iY1)CP(x+iY1)dx- S f(x+iyz)cp(x+iYz)dx
Xl Xl
is carried by aX 1+i[Y1'YZ]. In fact, ifaX t is smooth it follows from

Stokes' formula that the difference is equal to the integral of the
closed form f(z) ¢(z)dz 1 1\ ... 1\ dz" over the chain ax 1 + i [Yz, Y1]' and
we can approximate X 1 arbitrarily closely by an open set with Coo
boundary. (i) is thus a consequence of Theorem 9.1.7. To prove (ii) we
just have to observe that the analytic functional
S f(x+iy)¢(x+iy)dx
is carried by Xl '-.. X Z + i y, for this implies that fXl - fX2 is carried by

any neighborhood of Xl '-.. X z:::J a X 1 U a X z. (iii) Let fo be the distribu-
tion limit in .@'(X 1). If I/IEA then I/Ifo is the distribution limit of 1/11,
so (3.1.20) gives
<cpj~, 1/1)= S <P(x, y)(l/If)(x+iy)dx
+(N+l) IS (I/If)(x+ity) I a"cp(x)(iY)"/rx!tNdxdt.
0<1<1 ~ l"I~N+1
Here cpEC~(X1)' <P is defined by (3.1.18) and YET1, Iyl<y. The

formula (3.1.20) extends fo to a distribution in C'(X 1) if we integrate
only over X 1 for arbitrary ¢ E C~ (IR n). With ¢ E C~ (X 1) equal to 1 in
a large compact subset of X 1 it follows then from the preceding
formula that
1/1---+ <fo, 1/1)- S f(x+iy)l/I(x+iy)dx
is carried by a Xl + i [0, y]. Hence the uniqueness in (i) gives fo

-fx,EA'(aX1) so fx=fo in Xl' (iv) The existence of the limit fo in
A' (X 1) follows exactly as in the proof of Proposition 9.3.1 so we leave
this for the reader. Since
S f(x+iy)¢(x+iy)dx---+fo(cp), ¢EA,
it follows from (i) that fx 1 - fo is carried by a X l'

To prove (v) and (vi) we denote by Ry the difference in (i) which is

carried by any neighborhood of ax 1 when Iyl is small. Then
K *fx,(z)- S K(z-t-iy)f(t+iy)dt=RyK(z-.)
x,
when IImzl+IYI<l, YEr1 , Iyl<'l'. Here RyK(z-.) is analytic in any
compact subset of {z;IImzI2<1+IRez-xI2, XEaX 1 } when Iyl is
small. If fx ,EA' (a x 1) and X 2 ~ X 1 it follows that
F,(z) = S K(z-t-iy)f(t+iy)dt
x,
is analytic in a fixed complex neighborhood of X 2+ is· - 1 when yE r1
and Iyl is small. Thus
f(x+iy)= S F,(x+iy+iw)dw
Iwl~ 1
is analytic for x in a fixed complex neighborhood of X 2 when Iyl is

small. Hence f can be analytically extended to a neighborhood of X 2'
This extension is identically 0 in X 2 so f = 0 in Z if Z is connected,
which proves (v). To prove (vi) we observe that
K * fXl(Z)= S K(z-t-iy)f(t+iy)dt+RyK(z-.)
x,
is analytic at X 2 + iw if Iw - yl < 1. Since
Iw- y12= 1-2<w, y)+lyI2
<
this is true if y is replaced by e y for some small e> 0 and w, y) > O.
Hence K*fx,(z) is analytic at X 2 +iw unless <w,y)~O for every
y E r, that is, - w E roo This completes the proof of (vi) and of the
theorem.
The hyperfunction fx E B(X) will be called the boundary value of f

from r. Occasionally the notation fo of Theorem 3.1.15 will be used,
but we shall use the notation br f when we want to emphasize that
the limit is taken from r. This notation assumes tacitly that f is
analytic in a set Z with the properties listed in Theorem 9.3.3. Then f
is called r analytic at X.
There is a converse of (vi) in Theorem 9.3.3 (cf. Theorem 8.4.15):
Theorem 9.3.4. Let X be an open set in JR" and let u E B(X). If

WFA(u)cX x (r°,-O)
where r is an open convex cone in JR" '- 0, then there is a r analytic
function f such that u = br f
Proof. If Xl «::::, X we can choose v e A' (X 1) defining u in Xl' Set V

=K*v, let r 1 be a closed cone eru{O} and choose MeSn - 1 open
with ro nsn-1 eM so that 1\1 is in the interior of r1 o. If Xl «::::, X 0«::::' X it
follows from Proposition 9.3.1 that we can write v = V 1 +V2 where
v 1 (cp)= J V';O(cp)dro, cpeA,
w,-M
is analytic in X 1 and v2 is the boundary value of the analytic function

V2 (z)= J V(z+iro)dro, IImzl</', Imzel;., RezeX o·
-M
If X 2 «::::, X 1 it follows that the restriction of u to X 2 is the boundary

value of a function f analytic in {z; RezeX 2 , Imzer1 , IImzl</'tl
for some /'1 >0. The uniqueness of f implied by Theorem 9.3.3(v)
shows that the functions f obtained for different r 1 and X 2 together
define a r analytic function at X.
An immediate consequence of Theorem 9.3.3 is the classical "edge

of the wedge" theorem:
Theorem 9.3.5. Let f± be analytic in

Z±={x+iy; xeX, ±yer, IYI</'}
where /'>0 and ris an open convex cone. If fo=brf+=b-rf- then
fo is an analytic function which extends both f + and f - .
Proof. If ee ro n ( - r)0 then <y, e> = 0 when y e r so e= 0. Hence it

follows from Theorem 9.3.3(vi) that WFA(fo) is empty, which means
that fo is real analytic. If we subtract the analytic continuation of fo
from f± it follows from Theorem 9.3.3(v) that the difference vanishes
in Z ± when 11m zl is small, and this proves the theorem.
We can supplement Proposition 9.3.1 with an analogue of Lem-
ma 8.4.12:
Lemma 9.3.6. With the notation in Proposition 9.3.1 we have

WFA(U;)e{(x, e); xeX, -e/lelesuppdjl and U
is not analytic at x-ie/IWuaXx(1R n "O).
Proof. Replace the reference to Theorem 8.4.8 in the proof of Lemma

8.4.12 by a reference to Theorem 9.3.3(vi).
We come now to the important decomposition theorem corre-

sponding to Corollary 8.4.13:
Theorem 9.3.7. Let Fl , ... , f'; be closed cones in lRn " {O} with union lR"
,,{O}. If X is a bounded open set in lR" and uEB(X) then U= LU j
where ujEB(X) and WFA(U)c WFA(u)n(X x lj), j= 1, ... , J. If U= Luj
is another such decomposition then uj=u j + LU jk with UjkEB(X), ujk =
-ukj and WFA(ujk)cX x (ljnT,')' k
Proof. Extend U to vEA'(X), set V=K*v and define vjEA'(Y) by

J
Vit/t) = V,;(t/t) ct>iw) dw, t/tEA.
Here Y is a bounded open neighborhood of X, V'; is defined in
Proposition 9.3.1 and ct>. is the partition of unity in the proof of
Corollary 8.4.13. Then IVj=v, and if uj is the restriction of Vj to X
we obtain
WF(u)c WFA(u)n(X x lj)
in view of Lemma 9.3.6. To prove the second part we let wjEA'(X) be
an extension of uj-u j , thus W= LWjEA'(oX). Replacing wl by wl-w
we may assume that LWj=O. Set Jtj=K * Wj and
wjk(t/t) = JJtj~(t/t) ct>k(W) dw- Jw,,~(t/t) ct>iw)dw.
Since
L wjk(t/t)-wj(t/t)=O
k
we obtain for the restrictions ujk of wjk to X
The last part of the theorem is called Martineau's edge of the

wedge theorem. Its significance is perhaps more clear if we take lj
convex and proper and denote by Gj the interior of the dual cone.
Then it follows from Theorem 9.3.4 that uj=bGjfj and uj=bGJ; for
some Gj analytic fj,fj'. Moreover, ujk is the limit from Gj+G k of a
function fjk,fjk= - h.j, thus
bGjfjk=bGj+Gkfjk=bGkfjk= -bGkh. j
which confirms that
It is therefore possible to define B(X) as the set of all (fl' ... ,fJ) such
that fj is Gj analytic at X, identifying with 0 the set of all
(Lfl k' ... , LfJk)

k k
where fjk=-/"j is Gj+G k analytic. (Two Gj analytic functions are

called equal if they have identical Gj analytical restrictions.) We can
also use this observation to extend operations on hyperfunctions as
we did in Section 8.5. Assume for example that X e JR n, Ye JRmare
open sets and h a real analytic map X -+ Y. Set
Nh={(h(x), f/)E Y x JRm; th'(x)f/=O}.

If U EB(Y) and Nh (\ WFA (u) = 0 we shall define h* U so that as before
(9.3.2)
To do so we take an arbitrary Xo EX and write U in a neighborhood

of h(xo) in the form
where fo is real analytic and 'h'(xo)f/*O if O*I1EGj. Then the proof

of Theorem 8.5.1 shows that fj 0 h is Gj analytic near Xo for any Gj
with closure contained in h'(X O)-l Gj , which is the interior of the dual
cone of 'h'(xo}G'J. Thus we can define
h*u= fooh+ I, bGj(fjo h)
in a neighborhood of xO' That this is independent of the decom-

position of u follows from the remarks above. (Note that if one has
two different coverings JRm,-{o}=ur; then one can pass to the
covering consisting of all Ij occurring in one of them.) The proof of
Theorem 8.5.1 gives (9.3.2) without any change. In particular
WFA(u)eT*(X),-O is now well defined if u is a hyperfunction on a
real analytic manifold X. (We leave as an exercise to verify that h*u
as defined above agrees with h* u as defined in Sections 9.1 and 9.2
when h is an analytic diffeomorphism.)
If u E B(X) and v EB(Y) then u ® v EB(X x Y) is defined and
(9.3.3) WFA(u®v)e(WFA(U) x WFA(v))

U(WFA(U) x (supp V x {O}»u«supp u x {O}) X WFA(V)).
It is sufficient to prove this when XeJRn and YeRm. We can then

use Theorem 9.3.7 to decompose u and v. The statement follows if one
notes that when f and g are Gf and Gg analytic respectively, then f
® g is Gf x Gg analytic. The dual cone of Gf x Gg is G'J x G~, which
reduces to G'J x {O} if Gg=JRm and to {O} x G~ if Gf=JRn.
The rest of Section 8.5 can now be extended without any change
of the proofs to the case of hyperfunctions. We leave this repetition to
the reader.
9.4. The Analytic Cauchy Problem
Already in Lemma 9.1.4 we solved an analytic Cauchy problem for

the Laplacean. The extension of Theorem 8.6.1 to hyperfunctions in
Section 9.5 will require a general existence theorem for the analytic
Cauchy problem with precise information on the existence domain.
Such results will be proved in this section.
Points in (Cn will be denoted by Z = (z l' .•. , zn), and
QR={Z; IZjl <R, j= 1, ... , n}
is the polydisc of radius R with center at O. In this section only we
shall use the notation D'" for the operator (o/Ozl)"" ... (%zn)"'n acting
on analytic functions. We start by solving the simple differential
equation
(9.4.1)
with the boundary conditions
(9.4.2) D~u=O when Zj=O and O~k<fJj; j=l, ... ,n.
Lemma 9.4.1. For any f which is analytic in Q R there is a unique

solution of (9.4.1) which is analytic in Q R and satisfies (9.4.2). We have
(9.4.3) sup lui ~RIPI sup Ifl/fJ!.
aR aR
Proof When n = 1 we obtain from Taylor's formula the unique so-
lution z 1
u(z) = U(t)(z - t)P-1 dt/(fJ -1)! = zP U(tz)(1- t)P-1 /(fJ -1)! dt.
o 0
Taking care of one variable at a time we find for any n that if

fJ1 >0, ... , fJk >0, fJk+ 1 = ... = fJn =0, then
1 1 k
o 0
TI
U(Z) =ZP J'" U(t1 Zl' ... , tkZk, Zk+ 1"'" Zn) (1- t/i-1/(fJj-1)! dt
1
is the unique solution of the boundary problem. It is obvious that

(9.4.3) follows.
A slightly weaker existence theorem is valid for small pertur-

bations of (9.4.1). We take R = 1 for the sake of simplicity.
Theorem 9.4.2. Let fJ be a fixed multi-index, IfJ I= m, and let a"', I I~ m,

Q(
and f be bounded analytic functions in Q 1 with

A =(2n etsup Lla"'l < 1.
a,
9.4. The Analytic Cauchy Problem 347
Then the equation

(9.4.1)' DJl u = L aaDau+f
lal~m
has a unique solution satisfying (9.4.2) in Qt' and

(9.4.3)' sup lui ~(1- A)-l sup Ifl/P!.
The proof requires two elementary lemmas.
Lemma 9.4.3. If v is an analytic function of one complex variable ,

when 1'1 < 1, such that
Iv'ml~C(1-1'1)-a, 1'1<1, and v (0) =0,
where a ~ 1, then it follows that with the same C
Ivml ~ Ca- 1 (1-I'I)-a, 1'1 < 1.
Proof If a> 1 the statement follows from the fact that
r
(1-r)a J(1-t)-adt=((1-r)-(1-r)a)/(a-1)
o
takes its maximum when a(1- r)a-l = 1, for the maximum value
is (1- r)/a < l/a. Letting a-+ 1 we obtain the statement when a = 1.
Lemma 9.4.4. If v is analytic when 1'1 < 1 and

Ivml~C(l-I'I)-a, 1'1<1,
where a ~ 0, it follows with the same C that
Iv'ml~ Ce(l +a)(1-1'1)-a-1, 1'1 < 1.
Proof If O<e<p= 1-1'1 we have IV('l)1 ~ C(p-e)-a when 1'1 -'I <e.
m
Hence Cauchy's inequality gives IVi I ~ C e- 1 (p - e) -a. Choosing e
= plea + 1) we minimize the right hand side and obtain
Iv'(')1 ~ C(a + 1)(1 + l/a)a p-a-l < C e(a+ l)p-a-l.
Proof of Theorem 9.4.2. We shall pass from (9.4.1) to (9.4.1)' by using

Lemma 9.4.1 to solve recursively the equations
(9.4.1)" DJl uv + 1 = L aaDau v+ f
lal~m
starting with uo=O. By Lemma 9.4.1 we have

M =sup lUll ~SUp IfIlP!.
P, D,
Subtracting two successive equations (9.4.1)" and writing Vv = Uv+ 1 - U v

we obtain vo=u 1 and
(9.4.4) DP Vv + 1 = L a~D~vv' v=0,1,2, ....
1~I~m
We claim that with C=A/2 nm

n
(9.4.5) Ivv(z)I~CMTI(1-lzjD-mv, zED 1, v=0,1, ....
1
This is just the definition of M when v=O. If (9.4.5) holds for one
value of v, it follows from (9.4.4) and Lemma 9.4.4 that
n
IDPvv+ 1 (z)1 ~A(2ne)-m C M(em(v+ 1)t TI (l-IZjl)-m(v+ 1), ZED 1 .
1
Since vv+ 1 satisfies (9.4.2) we may apply Lemma 9.4.3 m times which
gIVes
n
Ivv+ 1 (z)1 ~A(2ne)-m eM emTI (l-Izjl)-m(v+ 1), ZED 1 .
1
Thus we have verified (9.4.5) with v replaced by v + 1. When zED~ it

follows from (9.4.5) that 2
IVv(z)1 ~Av M.
Hence u=limuv=Lvv exists and is analytic in Dt , and lul~M/(1-A)
there which proves (9.4.3)'. Since lim D~ U v= D~ u for every 0(, letting
v -> 00 in (9.4.1)" gives (9.4.1)'.
If u is a bounded solution in D1 of (9.4.1)' with j=O and if (9.4.2)
is fulfilled, then Vv = u satisfies (9.4.4). By the preceding proof Vv ->0 in
a neighborhood of 0, hence u=o. If u is just a solution in a neigh-
borhood of 0 we can apply this conclusion to u(Rz) for some R > 0,
replacing a~ by a~(Rz)Rm-I~I. This completes the proof.
Theorem 9.4.2 can be used to solve some mixed problems, but we

specialize now to the Cauchy problem which is all we shall need here.
Since the variable zn will play a distinguished role we introduce the
polydiscs with unequal radii
Theorem 9.4.5 (Cauchy-Kovalevsky). Assume that the coefficients in the

differential equation
(9.4.6) L a~D~u=j
1~I;;;m
are analytic in QR.~ and that the coefficient afJ, /J=(O, ... ,O,m), of Dr;:u
is equal to 1. If
(9.4.7)
and f is bounded and analytic in QR,~, then (9.4.6) has a unique analytic
solution in QR/2,b satisfying the Cauchy boundary conditions
(9.4.8)
For u we have the estimate
(9.4.9) sup lui ~2(R(j)msuplfl.
{JR/2,. {JR,.
Proof When (j = R = 1 the statement follows from Theorem 9.4.2 where

A~1/2. With z'=(zl"",zn_l) the equation (9.4.6) can be written
Iaa(Rz', R (j zn)R -Ial (j-a n Da u(Rz', Rt5 zn) = f(Rz', R (j z.).
After multiplication by R m (jm we have reduced the statement to (j = R

= 1, which completes the proof.
Since 11X1~m and IXn<m in (9.4.7), the condition is automatically

fulfilled for small Rand (j if aa are analytic at 0. Hence we have in
particular a local existence theorem for the homogeneous Cauchy
problem. Instead of the homogeneous Cauchy conditions (9.4.8) we
can pose inhomogeneous ones
(9.4.8)' D~(u-¢)=O when zn=O, j<m,
where ¢ is a given analytic function in QR,b' Writing u=¢+v we get
homogeneous Cauchy conditions for v and the equation
Iaa Dav= f- Iaa Da¢.
Thus we are back in the situation discussed in Theorem 9.4.5. Only

the analyticity of ~ ¢ Izn = 0' j < m, is important for we can replace ¢ by
I D~¢(z',O)z~/j!.
j<m
We shall now give global versions of Theorem 9.4.5. Set
P(z,D)= I if(z)Da
lal;:im
and let
P",(z,O= I aa(z)(a
lal=m
be the principal symbol. If ¢ is an analytic function with ¢ = 0 and

o¢/oz=(o¢/OZt, ... ,0¢/oznHO at zo,
then the equation ¢(z)=O defines an analytic hypersurface at ZOo It is
called non-characteristic if
P",(Z,o¢/ozHO.
'*
If t/J is a real valued C t function with t/J = 0 and d t/J 0 at zo, then the
equation t/J =0 defines a real C t hypersurface 1: at ZOo The tangent
plane
dt/J=0t/J+8t/J=0;
L
o t/J = 0 t/J /0 Zj d Zj' 0 t/J /0 Zj = (0/0 x j - i % y)/2
contains a unique analytic hyperplane defined by Ot/J=O. We call 1:
non-characteristic if this hyperplane is non-characteristic, that is,
Pm(Z, ot/J/OZ) ,*0.
Formally this condition looks just as in the analytic case.
Theorem 9.4.6. Let P(z, D) be an analytic differential operator of order

m in the open set Z C (Cn. If S c Z is an analytic non-characteristic
hypersurface then the Cauchy problem
P(z,D)u=f, Da(u-¢)=O onS when locl<m,
has a unique analytic solution u in a neighborhood of S for arbitrary f
and ¢ which are analytic in Z.
Proof At any point ZoES we can make an analytic change of coor-

dinates so that Zo becomes the origin and S is defined by zn=O. Then
'*
we have the situation in Theorem 9.4.5 after dividing by all. (Note that
all 0 at 0, because S is non-characteristic.) Thus we obtain local
solutions, and since they are unique they fit together to a solution
with the properties stated in the theorem.
The set in which the solution exists can often be enlarged by a

continuity method:
Theorem 9.4.7. Let f be analytic in the open subset Z of (Cn and let u
be an analytic solution of the equation P(z, D) u = f in the open set
ZocZ. If zoEZnoZo and Zo has a C 1 non-characteristic boundary at
zo, then u can be continued analytically as a solution of the equation
P(z, D) u = f in a neighborhood of ZOo
Proof Changing coordinates near Zo we may assume that zo=O and

that Zo is defined near 0 by Rezn<t/J(zt, ... ,zn_t,Imzn) where t/JECl,
1/1(0)=0 and dl/l(O)=O. The Cauchy problem

P(z,D)v=f, D~v=D~u when zn=-e
satisfies the hypothesis of Theorem 9.4.5 in QR.~+(O, ... ,0, -e) if band
e are small and
(9.4.10) -e<l/I(z',O), Iz'I<R.
Hence v gives an analytic continuation of u to a neighborhood of 0 if
Rb/2>e. With b fixed and R=3e/b the condition (9.4.10) is fulfilled
for small e since l/I(z',O)=o(lz'l), and this prove.~ the theorem.
We shall now prove a general global existence theorem similar to
Lemma 9.1.4 which will be important in Section 9.5. As before we
write z' = (z 1'"'' zn_l) and we use the Euclidean norm,
Iz'I2 = Iz 112 + ... + Izn _ 112.
Theorem 9.4.8. Let Z be an open convex set in (Cn such that
(9.4.11) {(z',O); Iz'l <R} cZ c {z; Iznl/e+ Iz'l <R},
and let P(z, D) be an analytical differential operator of order m in Z
such that
(9.4.12)
Then the Cauchy problem
(9.4.13) P(z,D)u=f in Z, D~u=O when zn=O, j<m,
has a unique analytic solution in Z for every f which is analytic in Z.
Proof We can find r>O so that (O,Zn)EZ if IZnl <r. Then the convexity
of Z gives that ZEZ if Iznl/r+ Iz'I/R < 1. If we apply Theorem 9.4.5 to
smaller polydiscs with center in the plane zn = 0, it follows that there
exists a solution of the Cauchy problem in a neighborhood t of
{(z',O);lz'I<R} in Z. We have to show that it can be continued
analytically to an arbitrary w=(w', WJEZ along the straight line from
(w',O). In doing so we may assume that wn>O. Replacing Z by the
intersection with a smaller set of the form {z; IZnl/e+lz'-w'l<a}
where a+lw'I<R, we may even assume that w'=O and that (0, Wn)EZ
when O~wn<eR.
Thus we assume now that (O,Zn)EZ when O~zn<eR. Fix a small
c>O, then a large M and set for O~t~e
Zt= {z; 0< Rez n< t(R _(lz'12 + IMImzn l2 + c)t)}.
In Zt we have IImznl<R/M, 0<Rezn<e(R-(lz'12+c)t), and since Z
contains all z with O<Rezn<e(R-lz'l), Imzn=O, it follows that itcz
if M is large enough. Then we also have Ztci for small t, and
°
{ZEaZ t ; Rezn=O} ci, O<t~e. On the other part of aZ t the analytic
tangent plane is defined by <dz, 0 = where
(9.4.14) "= tZ'(lz'1 2 + IMImznl2 +c)-t, Re'n = L
Hence l"l<t"nl~e"nl, so ~(z,')=FO by (9.4.12). In view of Theo-
rem 9.4.7 it follows that the set of tE[O,e] such that u can be con-
tinued analytically to Zt is open. However, it is obviously closed and
non-void so it must be equal to [0, eJ Continuation is therefore
°
possible to Z, for every c>O, and this completes the proof that
analytic continuation is possible from to (O,zn) ifO<zn<eR.
Corollary 9.4.9. Let r be an open convex cone in <en such that

(9.4.15)
H ere a> 0. Let W be an open neighborhood of °and P(z, D) an analytic
differential operator of order m in W such that
(9.4.16)
For every analytic function f in W nr it is then possible to find an
°
analytic solution of the equation P(z, D) u = f in W' n r for some other
°
neighborhood W' of independent of f. If u is any such solution and f
is analytic in a full neighborhood of then u has an analytic conti-
nuation to a neighborhood of 0.
Proof We can choose e>a and a neighborhood Wo of °such that

(9.4.16)'
Let T=(O, ... ,0,aR) and
Z ={ZE<e n; Iznl/e+ Iz'l <R} n(r-iT).

Z satisfies (9.4.11) since (z', iaR)Er if Iz'l <R, by (9.4.15). If R is
sufficiently small then P(z+iT,D) is analytic in Z and satisfies (9.4.12)
there by (9.4.16)'. Also f(z+iT) is analytic in Z. Hence it follows from
Theorem 9.4.8 that the Cauchy problem with homogeneous Cauchy
data for the equation P(z+iT,D)u(z)=f(z+iT) has an analytic so-
lution in Z. Then P(z,D)u(z-iT)= f(z) in Z +iT, which contains the
intersection of r and a neighborhood of
statement is now proved.
°
since e> a. The first
To prove the second statement we take

and solve the equation

P(z+iT,D)v= f(z+iT), ZEZ,
with Cauchy data
D~v(z',O)=Df.u(z',iaR); j<m,lz'I<R.
If R is sufficiently small this is a Cauchy problem for

m-l
w(z)=v(z)- I z~Df.u(z',iaR)/j!
o
of the form studied in Theorem 9.4.8. Hence a solution exists in Z,
and v(z-iT) is then an analytic continuation of u to a neighborhood
0[0.
9.5. Hyperfunction Solutions

of Partial Differential Equations
If P(x, D) is a differential operator with real analytic coefficients in the

open set X c: 1R" then
(9.5.1) W~(P(x, D) u) c: WFA(u), uEB(X).
This is an immediate consequence of Theorems 9.3.7 and 9.3.4, for if u

=brf then P(x,D)u=brP(z,D)f. We shall now prove a converse (cf.
Theorem 8.6.1):
Theorem 9.5.1. If P(x, D) is a differential operator of order m with real

analytic coefficients in X c:1R.n then
(9.5.2) WFA(u) c: Char Pu WFA(P(x, D) u), uEB(X).
Proof. We must show that if (xo,~o)iWFA(P(x,D)u) and Pm(xO,~o)=l=O

then (xo'~o)iWFA(U). In doing so we may assume that xo=O and that
~o=(O, ... ,O,I). Choose a>O so that
(9.5.3) (0, ~)i WFA(P(x, D) u) if ~n = 1,
(9.5.4) Pm(O, 0 =1= 0 when 1('1 ~ a,
Let 1R.n" 0 = Vlj where lj are proper, closed, convex cones, I

J
~'I < a ~n
in ro and ~oilj for j=l=O. By the first part of Theorem 9.3.7 and (9.5.3)
we can then write in a neighborhood Y of 0

J
P(x,D)u=f=I,fj, WFA(fj)cYxlj;
1
J
U= I, Uj, WFA(uj)c Yx lj.
o
Hence
J
P(x,D)u o + I,(P(x,D)u j - fj)=O.
1
By the second part of Theorem 9.3.7 it follows that

J
P(x, D) Uo = I,foj , WFA(fo) c Y x (ronlj).
1
The closure of Go = {YEJR."; y">aly'l} in JR."" 0 is in the interior of the

dual cone of roo We can choose open convex cones Gj ,j-4=O, with
closure in the interior of the dual cone of lj and ~o¢G'J, for ~o¢lj
when j -4= O. By Theorem 9.3.4
J·o·=b
(J G0+ Gj F.J
where Fj is analytic in th~ intersection of JR." + i( G0 + Gj) = ~ and a
neighborhood of O. Now rj satisfies (9.4.15) since JR."+iG o does, and
(9.4.16) is valid by (9.5.4). Hence it follows from the first part of
Corollary 9.4.9 that we can find ~ analytic in the intersection of ~
and a neighborhood of 0 so that P(z, D) Uj=Fj. We also have Uo
=bGo Uo where Uo is analytic in the intersection of JR."+iG o and a
neighborhood ofO. Thus Theorem 9.3.3(v) gives in JR."+iG o near 0
P(z,D) (Uo-t U=0. j)

J
But then the second part of Corollary 9.4.9 proves that Uo - I, Uj can
be continued analytically to a neighborhood of O. Hence 1
is analytic at O. Since (0, ~o)¢ WFA(b Gj Uj) by Theorem 9.3.3(vi), and

(0, ~O)¢WFA(Uj)' when j-4=O, we obtain (0, ~O)¢WFA(U), The proof is
complete.
The proof gives also a microlocal existence theorem:
Theorem 9.5.2. Let (xo, ~o)E T*(X) " 0 be non-characteristic with respect
to the differential operator P(x, D) with real analytic coefficients. For
every fEB(X) one can then find uEB(X) with

(xo, ~O)¢WFA(P(x,D)u- f).
Proof With Ij defined as in the proof of Theorem 9.5.1 we write

f = IJj, WFA (fj) c X x Ij.
Then fo=bGoFo where Fo is analytic in the intersection of JRn+iG o
and a neighborhood of xo. By Corollary 9.4.9 we can find a solution
U of the equation P(z,D)U=Fo in the intersection of JRn+iG o and
another neighborhood of xo. Taking u = bGo U in a neighborhood of
Xo we have proved the theorem.
In view of Theorem 9.5.1 the Holmgren uniqueness theorem and

its refinements proved in Section 8.6 for distribution solutions of the
equation P(x,D)u=O remain valid for hyperfunction solutions. We
shall prove next that Cauchy data can also be defined on an arbitrary
non-characteristic analytic surface. For the sake of simplicity we shall
assume that it is a plane.
Thus let X be an open subset of JR n and set
X ± ={XEX,Xn~O}, Xo={XEX; xn=O}.
If uEB(X +) satisfies the analytic differential equation P(x,D)u=O in
X + and X 0 is non-characteristic with respect to P, we shall prove that
D! ul xn = oEB(X~) can be defined in a natural way. (Here X~ is X 0
considered as a subset of JRn-l, that is, X 0 =X~ x {O}.) First we let U o
be any hyperfunction in X which is equal to u in X + and 0 in X _.
Then f=P(x,D)u o has support in Xo. We can replace U o by uo-v for
any vEB(X) with supp v c X 0' and this changes f to f - P(x, D) v. There
is natural best choice of v:
Theorem 9.5.3. Assume that P(x, D) has real analytic coefficients and
that X 0 is non-characteristic with respect to P(x, D). Then there is for
every fEB(X) with suppfcXo a unique decomposition
(9.5.5) f =P(x,D)v+ I V/~9D!Jo(xn)
j<m
where vEB(X), suppvcX o, and vjEB(X~)for O~j<m.
Proof Assume first that we have some 1, v, Vj satisfying (9.5.5) with

supp f, supp v and supp Vj x {O} contained in a compact subset K of
X o. Then (9.5.5) is equivalent to
(9.5.6) f(¢) = v('P(x, D) ¢)+ I vi( -DnY ¢(., 0)),
j<m
if ¢ is analytic in a neighborhood of K. By Theorem 9.4.6 we can

choose ¢ so that in a neighborhood of K
(9.5.7) tP(z,D)¢=1/I; (-DnY ¢(., 0) = 1/Ij' j=O, ... ,m-I,
if 1/1 and 1/Ij are arbitrary entire functions of nand n -1 variables
respectively. Hence f=O implies v=O and vj=O for every j. To prove
existence we assume given fEB(X) with suppf=KcXo compact and
define
(9.5.8) V(1/I)+ I vi1/lj)=f(¢).
j<m
If Q is a neighborhood of K in (Cn and Q' c (Cn- 1 is a neighborhood of

K o ={X'E1Rn - 1 ,(x',0)EK} it follows from Theorem 9.4.5 that there is a
neighborhood OJ of K in (Cn where (9.5.7) has a solution ¢ such that
sup I¢I ~ C(sup 11/11 + I sup 11/1).
co Q Q'
Hence
which proves that (9.5.8) defines vEA'(K) and vjEA'(K o) satisfying

(9.5.5). In view of the uniqueness proved first we conclude that
(9.5.9) suppf=suppvuU SUpPVj x to}
)
whenever (9.5.5) holds with f, v, vj of compact support, supp vusupp f

cX o·
For an arbitrary fEB(X) with suppfcXo and any open Y~X~ we
can choosefyEB(X) with suppfycYx{O} andfy=fin XIl(Yx1R). In
fact, we just have to extend the hyperfunction which is equal to f in
X Il(Yx1R) and 0 in X + uX _ u(X Il(C Yx1R». Writing
fy = P(x, D) uy + I ur/i9 D~ (jo(x n),

j<m
by the first part of the proof, we conclude from (9.5.9) that

supp(uy-uz)uU supp(uyj-u z ) x to} csupp(fy- fz)
j
if Z ~ X~. Hence u y and uyj define when Y increases a global solution

of (9.5.5). The uniqueness follows similarly from (9.5.9) if we start from
a solution of (9.5.5) with f = 0 and "cut v, vj off' outside a neigh-
borhood of some point in X o' The proof is complete.
Corollary 9.5.4. If P(x, D) has real analytic coefficients in X and X 0 is

non-characteristic with respect to P(x,D) then any uEB(X +) satisfying
the equation P(x, D) u = 0 has a unique extension Uo to X vanishing in

X _ such that for some vjEB(X~)
(9.5.10) P(x,D)u o = I vj®D~(jo(xJ

j<m
If all Vj vanish then u=O in a neighborhood of Xo'
Proof The extension satisfying (9.5.10) is obtained if we first take an

arbitrary extension U o vanishing in X _, apply Theorem 9.5.3 to f
=P(x, D) Uo and subtract v from Uo' The last statement follows from
Holmgren's uniqueness theorem.
To give another interpretation of (9.5.10) we assume for a moment

that UECm(X+) and set uo=u in X+' uo=O in X_. Then we obtain
D~ Uo =(D~ u)o - i I D~-j-l u(., O)®D~ (jo(x n)

j<k
for k~m by induction. With Ij(x, () denoting the polynomial part of

P(x,()/(~+l it follows if P(x,D)u=O that
P(x,D)u o = -i I Pix,D)ulxn~o®D~ (jo(x n )·

j<m
Thus the extension U o is the same as the one in the Corollary, and
(9.5.11)
These equations can be considered as a system of differential equa-

tions for uj=D~ulxn~o,j<m. If a(x) is the coefficient of D; in P(x,D),
which is =1= 0 in X 0 by assumption, then the r
equation contains a
term -iaum _ j _ 1 and otherwise only Uk with k<m-1-j. Starting
with the Equation (9.5.11) with j = m -1 we can therefore determine
uO,u1, ... ,Um_ 1 successively in terms of vm_l'''''vO' This works for
hyperfunctions as well, so under the hypotheses in Corollary 9.5.4 we
obtain well defined "normal derivatives" ujEB(X~). The definition
agrees with that in Theorem 4.4.8' when both are applicable, for the
discussion above in the Cm case remains valid then by the remarks
following Theorem 4.4.8'.
Thus we have now what is needed to state boundary conditions
for differential equations involving hyperfunctions. The correspon-
dence between analytic functionals and harmonic functions in Sec-
tion 9.1 is just the special case of the Dirichlet problem in a half
space. However, we must stop at this point in our brief introduction
to hyperfunction theory.
9.6. The Analytic Wave Front Set and the Support
We shall begin by giving another equivalent definition of WFA(u) due

to Bros and Iagolnitzer. If ueA'(1R") we define an entire function TA u
depending on a positive parameter A by the Gaussian convolution
(9.6.1) TA u(z)=uy exp(-A(Z-y)2/2), zeCC".
Here Z2 = z~ + ... + z;. Since
(9.6.2) Re-(z- y)2 =(Im(z- y))2 -(Re(z- y))2 ~(Im(z- y))2;
z, yeCC";
we have for every 8> 0
(9.6.3) ITA u(z)1 ~ C. exp(A(IIm zl + 8)2/2).
TA is closely related to the Fourier transformation for
exp(A(i <x, 0 - ~2 /2)) TA u(x + i~)
= u,(exp( - A(X - y)2/2 + i A<y, 0))
is the Fourier transform at -A.~ of u multiplied by e-Alx-yI2/2. This
factor localizes at x in very much the same way as the cutoff
functions used in Lemma 8.4.4. It is therefore natural to expect that
(xo'~O)¢WFA(U) if and only if the right-hand side is O(e- CA ) for some
c>O in a neighborhood of (xo, -~o) as A.--+C(). To prove this we shall
begin by estimating TA u when u is the boundary value of an analytic
function. We shall use the notation d(x, A)= inf Ix - yl for the Eu-
clidean distance from xe1R." to A e1R.". YEA
Proposition 9.6.1. Let Qe1R." be an open convex set with OeD, let r
=1R.+ Q be the convex cone generated by Q and let X e1R." be an open
set. Iff is an analytic function in X +iQ and ueA'(1R.") is equal to bTf
in X, then
(9.6.4) ITAU(z)I~CexpAcP(z), zeK, A>O,
for every compact set K e CC" and every continuous function cP on K
such that
(9.6.5) cP(z) > (11m Zl2 -d(Re z, CX)2)/2, zeK,
(9.6.6) cP(z»d(Imz,Q)2/2 if RezeX, zeK.
If u = 0 in X then (9.6.6) can be omitted.
Proof We may assume that X is bounded since (9.6.5) remains valid if

X is replaced by the intersection with a sufficiently large ball. The
estimate (9.6.4) follows from (9.6.2) and (9.6.5) if Re Z is in a neigh-

borhood of Cx, so we may assume in the proof that ReK cx. Write
u=u l +u o where suppu l cX and suppuocCX. By (9.6.5) and (9.6.2)
Re-(z-y)2/2<<P(z), zEK,
for all y in a complex neighborhood of supp u O' so (9.6.4) is valid for
T).u o. Let Xort::.X. For every yEO the analytic functional
A3</>-+U l (</»- Jf(x+iy) </>(x+ iy)dx

Xo
is carried by (X'Xo)u(oXo+[O,iy]). Hence u l is carried by

M =(X, Xo)u(oXo +[0, iy])u(Xo + {iy}).
For every zoEK we can choose YEO so that
11m Zo - Y12/2 < <P(zo)·
When YEX, X 0 we have by (9.6.5) if X 0 is large enough
(IImzI2-IRez- YI2)/2 < <P(z), zEK.
Hence we obtain in view of the convexity of IIm(z- yW
Re( -(z- y)2/2) < <P(z)
for z in a neighborhood of Zo in K and y in a neighborhood of M.
The estimate (9.6.4) follows in a neighborhood of Zo and by the Borel-
Lebesgue lemma in all of K.
Proposition 9.6.1 combined with the decomposition provided by

Theorem 9.3.7 will easily give that T). u(x + i e) grows more slowly than
exp A. e2 /2 if (x, - e)~ WFA(U), To prove a converse result we must study
how U can be reconstructed from T;. u. We start from Fourier's in-
version formula
</>(0) = lim (2n)-n II ei<Y.~>-EI~1 </>(y)dedy, </>ECO'(1Rn),
E-+O
where the convergence factor rEI~1 guarantees absolute convergence.

e i"
If C= + then
has an analytic square root with positive real part when 1,,1 < lei. When
a>O is so small that alyl < 1 when yESUpp </>, we can shift the in-
e
tegration with respect to to the cycle
e-+c=e+iaylel.
Note that i(y,O=i(y,O-ay 2Iel has real part ~O and that
dC l " .. · "dCn=(l +ia (y, e/lel» del" ... "den
since dl~I=L~/I~ld~j and dl~I/\dl~I=O. Hence
<fJ(O)=1im (2n)-n Jd~ Jexp(i(y, 0 -a y21~1-8V(2)

e~O
·(1 + i a(y, ~/I~I») <fJ(y) dy.
Here y~«y,O+iay21~I+iBvr)!l~1 has imaginary part ;:;:0, is bound-

ed in Coo and has a lower bound independent of 8 and ~ for the
norm of its differential with respect to y. By Theorem 7.7.1 the inner
integral is therefore a rapidly decreasing function of ~, uniformly with
respect to 8, thus
when a >0 is small enough. However, the right-hand side is an

analytic function of a when Re a > 0 so the formula is valid for all
a>O.
With a fixed r>O we set ~= -ArW where WES n - 1 and 1[>0.
Replacing <fJ by <fJ(x +.) we obtain
J
DO
(9.6.7)' <fJ(x)=(2n)-n Jr"l[n- 1 dl[ dwSexpl[(i(x-y,rw)

o Iwl~ 1
-ar(x-y)2) (l+ia(x-y,w»)<fJ(y)dy.
If we choose a= 1/2r then the exponent becomes I[ times
E(x- y, rw)=i (x- y, rw) -(x- y)2/2= -(x - y- irw)2/2-r 2/2,
and it follows that
J
e- ).r2/2 T). <fJ(x - i r w) = e)'E(x - y, rw) <fJ(y) d y,
e- ).r2/2 (w, D) T). <fJ(x - i r w) = I[ JeAE(X-y,rw)(r + i (x - y, w») <fJ(y) dy.
Hence we have for every <fJ E C~
J r" I[n-1 dl[

()()
(9.6.8) <fJ(x)=2- 1(2n)-n

o
. J e- Ar2 /2(1 +(w,D/r I[») T).<fJ(x-irw) dw

Iwl~ 1
where the integrand is 0(1[- N) for every N as I[ ~ 00 and is bounded as

I[~O. The extension to analytic functionals is straightforward:
Proposition 9.6.2. If uEA'(lRn) then
Jr,1./2 An- 1 T;. U(Z)dA,

00
F(z)=2- 1 (2n)-n
o
Je-,1./2 An- 2 O/OZj T,1. U(Z) dA

00
Fj(z) =2- 1 (2n)-n

o
are analytic functions in {ZECCn ; Ilmzl < 1} which remain analytic at Xo
+iwo,lwol=1, iffor some C, c>O
(9.6.9) IT,1.u(x+i~)I~Ce,1./2-c,\ Ix-xol+I~-wol<c, A>O.
By Proposition 9.6.1 this is true for every Wo if xo¢supp u. If X is a
bounded open neighborhood of supp u in 1R.n we have with the notation of
Proposition 9.3.1
(9.6.8)' u(cP)= J F';(cP)dw+iI, J Flro(cP) Wj dw, cP EA.
Iwl~l Iwl~l
Proof. The stated analyticity of F and Fj follows immediately from

(9.6.3) and (9.6.9). To prove (9.6.8)' we let cPoEC~(X) be equal to cP in
a neighborhood of supp u. Let R(<po) be the right-hand side of (9.6.8)'
with cP replaced by cPo. Then R(cPo) is the limit when r-+ 1 of
J J e-,1./2 ;,n- J
00
2- 1 (2n)-n cPo (x) dx 1 dA (1- <w, D/A») T;, u(x+ irw) dw.
o Iwl~ 1
Now the definition of T,1. gives

J<Po(x) T;, u(x+irw)dx=u(T,1.cPo(. -irw)),
JcPo(x) <-W,D/A) T,1.u(x+irw)dx=u«w,D/A) T;.cPo(. -irw)).
Thus R(<po) = lim u(cPo) where
00
cPO,r(x)=2- 1 (2n)-n Je-A/2 An-l dA J (1+<w,D/A»)T,1.cPo(x-irw)dw.

o Iwl~ 1
It follows from (9.6.8) that cPO,r -+cPo uniformly when r-+ 1, for the
integrand is uniformly rapidly decreasing as A-+OO when r~ 1. Since
cPo is analytic in a neighborhood of supp u we have by Prop. 9.6.1
IT). cPo(z- irw)1 ~ Ce,1./2-c,1., O<r~ 1
for some c > 0 and all Z in a complex neighborhood of supp u. Hence
cPo,.(z) is uniformly convergent in such a neighborhood of suppu, and
the limit must be equal to cP by the uniqueness of analytic con-
tinuation. This implies that u(cPo,.)-+u(cP), that is, R(cPo)=u(<p). Thus R
= 0 outside supp u and (9.6.8)' is valid.
Combination of Propositions 9.6.1 and 9.6.2 gives the Bros-Iagol-

nitzer definition of WFA(u) (the essential support in their termi-
nology):
Theorem 9.6.3. Let uEA'(1R.n) and (x o, 'O)E T*(1R.n), O. Then (x o, '0)

~WFA(U) if and only if there is a neighborhood V of xo-i,o and
positive constants C, c such that
(9.6.10) IT;.u(z)I~Ce;'(I~oI2/2-C), ZEV, A>O.
Proof Assume first that (9.6.10) is valid. If O=t=tE1R. and Mt(x)=tx,
xE1R.n, then
t" T;.AMi u)(Z) = T;. u(t z)
for
Taking t =" =
01 we reduce the proof to the case "01 1. Then we have
the representation (9.6.8)' of u where F and Fj are analytic at X o - i '0'
so Lemma 9.3.6 gives that (x o, 'o)~ WFA(U).
Now assume that (x o, 'o)~ WFA(U). Choose closed convex proper
conex ~, ... ,Ij which cover 1Rn ,{O} so that 'o~Ij, j=t=l, and
r 1 n WFA(u)xo =0. By Theorem~ 9.3.7 and 9.3.4 we can choose a neigh-
L
borhood X of X o and ujEA'(X) so that U= uj in X, u 1 is analytic in
X and u j = bG)j, j =t= 1, where Gj is the interior of the dual cone of Ij
and Jj is analytic in X + Wj where Qj is a convex set generating Gj •
We shall apply Proposition 9.6.1 to each uj and to u- LUj ' If ,E1R. n
we have d(" Q)2 < "12
unless 0 is the point in OJ closest to ,. This implies that for every

YEG j "_EyI2~"12
if E>O is small. Hence <"y>~O which means that -,EIj. Since
'o~fj when j =t= 1 it follows that there is a neighborhood K of X o- i '0
and some c>O such that for zEK
"°1 2/2-c>(IIm zI2 -d(Re z, CX)2)/2,
"012/2-c>d(Imz,Q)2/2, j=t= 1.
(9.6.10) is now a consequence of Proposition 9.6.1.
If we make a change of scales the distinction between Qj and Gj in

the preceding proof is suppressed and we obtain a very precise supple-
ment to Proposition 9.6.1.
Proposition 9.6.4. Let uEA'(1R.n) and Wo ={'; (0, ')EWFA(u)}. Set u~
=M: u=u(c5.). If K cccn is a compact set and (Jj a continuous function
on K such that
(9.6.11)
it follows for small 1> that
(9.6.12) ITA. u6(z)1 ~ C 6e A.lb(z), zEK, A.>O.
Proof Take any covering of JRn" {OJ with small convex proper cones
°
as in the proof of Theorem 9.6.3. Then u=u o + ~>j where uo=O in a
°
neighborhood X of and uj=bGjij in X with ij analytic in X +iQ j •
Here Qj is a neighborhood of if IjnWo=0, and Qj generates the
open convex cone Gj with dual lj otherwise. We have U6 = L U j6 where
U 06 vanishes in XI1> and uj6 restricted to Xlf> is the boundary value of
ij(1)z),j=t=O, which is analytic in Xjf>+iQ/f>. For small 1> the estimate
(9.6.12) follows from Proposition 9.6.1 if
4>(z) >d(Im z,Q/1»2j2, zEK, j=t=O, Ijn Wo=t=0.
Since d(Imz,Q/1»'\.d(Imz, G) when 1>--+0, this is true for small f> if
(9.6.13)
Now we have a kind of Pythagorean theorem
d(~,Gj)2+d(~, _lj)2=1~12.
Since Gj n ( -lj) = {OJ and the relation b~tween these cones is sy~
metric it suffices to prove this when ~¢Gj' If ~* is the point in Gj
closest to ~ then <~*-~,l1-~*>~O, l1EGj' hence ~*-~Elj since Gj is
a cone, and <~*-~,~*>=O~<O,~*>, OE-lj. Thus ~-~* is the point
in -lj closest to ~ and (see Fig. 5).
d(~, Gy +d(~, _lj)2=1~ _~*12+1~*12 = 1~12.
Fig. 5
Hence (9.6.13) means that

(9.6.13)' <P(z»(llmzI2-d(lmz, -0 2)/2
where r is the union of the cones lj with lj n Wo =1= 0. Thus r is close
to Wo if all lj are small, so (9.6.13)' follows from (9.6.11). The proof is
complete.
An important feature of (9.6.11) is that the right-hand side is linear

along any outer normal of - WOo To exploit this we need an elemen-
tary lemma.
Lemma 9.6.5. Let a and b be positive numbers and u a subharmonic

function in R={ZE<C; O<Rez<a, Ilmzl<b} such thatfor some e>O
u(z) «max (0, -lmz))2 when zER,
u(z) < -b 2 /3 when Ilmzl<b and O<Rez<e.
With A(x)=(b/3) sinh (n(a-x)/b)/sinh (na/b) itfollows that
(9.6.14) u(x+iy)< -A(x)(2y+A(x)) if Iyl < A (x)/2, O<x<a.
Proof. With 0 < b < b we apply the maximum principle in the rectangle
R~={z;O<Rez<a, -b<lmz<b-b} to
v(x+ iy)=u(x+ iy)- b 2 +bA(x) sin(n(y+b)/b).

v is subharmonic since we have added a harmonic function, and v < 0
near the boundary. Hence v<O in R b • Now sin8>28/n for 0<8<n/2
so ifO<y+b<b/2 we obtain
u(x + iy)< b 2 - 2A(x) (y+ b).
We mInImIZe by taking b=A(x) which is legitimate since O<y

+A(x)<3A(x)/2~b/2 when lyl<A(x)/2. This proves (9.6.14).
The important point in (9.6.14) is that the upper bound is negative
when y = 0; the good estimate at one end of the rectangle propagates
with some decay along the line 1m z = o.
We are now ready to state and prove the main result of this
section relating suppu at a point Xo to WFA(u) at xO. It is due to
Kashiwara but sometimes called the co-Holmgren theorem to under-
line the analogy with Theorem 8.5.6'. In the statement we need the
notion of tangent cone of a set. If xoEM e1Rn then Txo(M) is defined
as the set of limits of sequences tj(Xj-x o) when t j -+ + 00 and xjEM.
(See also the analogous limiting cone at infinity in Lemma 8.1.7.) It is
clear that 'T"o is a closed cone. If ",EC 1 (1R n,1Rm) then we have
tit/l(X)-t/l(xo))-+t/I'(xo) TiftiXj-xo)-+ T. Hence

t/I'(x o) 'T"o(M)c T",(xo)(t/I(M))
which shows in particular that 'T"o(M) c Txo(X) is invariantly defined if
M is a subset of a C 1 manifold X instead oClRn.
Theorem 9.6.6. If uEB(X) and XoEX then
(9.6.15) N(Wo)c awo x Txo(supp u)
ifWo={eETx~(X);(Xo,e)EWFA(U)} considered as a subset of the vector
space 'T,,~(X) with the origin removed.
Proof. The statement is local so we may assume that X =1R n, xo = 0

and uEA'(1Rn). Let (eo, to)ENe(Wo). According to Definition 8.5.7 this
means that eoEWo and that there is a COO function f(e) such that 0
=t=f'(e o) = to and f(e)~f(eo)=O when eEWo' If we prove that
to¢Txo(suppu) or -to¢Txo(suppu) implies eo¢Wo we shall have a
contradiction proving (9.6.15).
Choose a compact neighborhood K of the interval 1= [ - to, toJ
-ie o in <cn. For every fixed e>O we have by Proposition 9.6.4 for ud
=u(b.) if b is small enough
(9.6.16) vb(z)=2A.-llogIT;,ub(z)/Cbl<e+llmzI2-d(Imz, - WO)2,
zEK.
Choose a compact neighborhood Ko of to - ie o or - to - ie o such that
ReKon Txo(suppu)=0. Then it follows from Proposition 9.6.1 that
there is a positive constant c such that for small b
(9.6.17)
if Cb is chosen large enough.
Take any real x, e with Ixi + Ie - e 01 small, f( e) = 0, and consider the
subharmonic function
V(w)=vb(x-ie-wf'(e)), WE<C.
If R = {w; IRe wi < 1, 11m wi < b} and b + Ixi + Ie - e 01 is sufficiently small

we have x-ie-wf'@EK when WER. The distance from Imwf'(e)+e
to the hypersurface f = 0 is ± 1m w If' (e)1 if b is sufficiently small. Hence
the distance to Wo is at least 1m w 1f'(e)l, and we obtain from (9.6.16)
when WER
V(w) <e +(Im wf'(e) + e)2 -(1m wf'(e)f
=e+e 2 +2 1m w<f'(e), 0, 1m w>O,
V(w)<e+ e 2 +2 1m w<f'(e), 0 + 11m wI 2 1f'(eW, 1m w~O.
When WER and Re w is near 1 or -1 we obtain from (9.6.17)

V(w) < -C/2+~2+2 1m w<f'(~), 0
again if Ixl + I~ - ~ 01 + b is sufficiently small. If we apply Lemma 9.6.5

to (V(w)-e-~2-2Imw<f'(~),O)/If'mI2 as a function of 1±w it
follows that for some co> 0 and r > 0 independent of e we have
V(w)~ -Co+e+~2 if Iwl<r and Ixl+I~-~ol<r, f(~)=O.
Taking e = co/2 we obtain

IT,t u,J(x - i~ -wf'(~»1 ~ C. exp(2( -c o/2 + ~2)/2)
if Iwl<r and Ixl+I~-~ol<r, f(~)=0.
This means that IT,t u.(z)1 ~ C' exp(2(~~ - co/3)/2) if Iz + i ~ol is sufficiently
small, and by Theorem 9.6.3 it follows that (0, ~o)¢ WFA(u,J)' The proof
is complete.
All the arguments in Section 8.6 based on Holmgren's theorem
have obvious analogues with the characteristic set replaced by the
tangent cone of the support. In particular, the proof of Theorem 8.6.8
gives without change.
Corollary 9.6.7. Let Wi C W 2 be open convex sets in Tx:(X),- {O} such
that
(i) ~n WFA(u)xo=0
(ii) every hyperplane with normal in Txo(supp u) n (- T"o(supp u»
which intersects W2 also intersects Wi'
Then W2 n WFA(U)xo=0.
Corollary 9.6.8. If (xo, ~O)ENe(supp u) then (xo, ~)EWFA(U) implies that
(xo,~+t~o)EWFA(U) for every tE1R..
Proof If ~ is proportional to ~o the statement follows from

Theorem 8.5.6' (which conversely follows from Corollary 9.6.8 and the
fact that WFA(u)xo cannot be empty). Otherwise, assuming that ~
+t~O¢WFA(U)xo' we take a convex open neighborhood Wi of ~+t~o
such that W i nWFA(u)xo=0 and W2=Wi+1R.~0 does not contain O.
Since Txo(supp u)n( - Txo(supp u» is in the orthogonal plane of ~o and
every hyperplane with normal orthogonal to ~ 0 intersecting W2 must
also meet Wi> the statement follows from Corollary 9.6.7.
Theorem 9.6.6 and Corollary 9.6.8 may be considered as the 0 and
n -1 dimensional cases respectively of the following
Theorem 9.6.6'. Let uEB(X), XoEX, Va linear subspace of T"o(X), and

denote by Txo. v (supp u) the closure of the image of T"o(supp u) in
Txo(X)/V. Then
(9.6.15)' N(Woll(V' + {e}))c a(Woll(V' + {e})) x 'T"o. y(suppu)
if Wo = WFA(u).xo and V' + {e} is any affine subspace of 'T,,!(X) parallel

to the orthogonal space V' of V. Here Wo Il (V' + {e}) is regarded as a
subset of v'+{e}, with the origin removed if eEV', so the normals
belong to the dual space Txo (X)/V of v'.
Proof We may assume that X =JRn, Xo =0, and that V' is defined by
C =(e k + l' ... , en)=O. Assume that eo*O and that eo has a compact
neighborhood K such that e Ol = f(e~) and
(9.6.18)
where e'=(e 2, ... ,eJ and fEC OO (JRk-l). We must show that if either
(1, -af/ae') or (-l,af/an is not in 'T"o.y(suppu) at e~ then eo¢Wo.
Assume for example that
(9.6.19)
Choose M so large that
(9.6.20)
This is possible since we have a compact subset of aK shrinking to
{eEoK; e ~f(n, e" =e~} c CWo
1
by (9.6.18). Let t be the smallest number ~O such that

eEK, e 1 > f(e')+t+Mle" -e~12=e¢Wo.
If eEK and e l =f(e')+t+Mlf'-e~12 then e¢Wo by (9.6.20) if eEaK,
and if eEK-......aK this follows from Theorem 9.6.6 since
(-1, af/af, 2M(e" - e~»¢'T"o(suppu)
by (9.6.19). Thus t=O and eo¢Wo which completes the proof.
We shall end with an application to the regularity of solutions of

boundary problems. As in Section 9.5 we consider a plane boundary.
Thus let X be an open neighborhood of OEJRn and
X ± ={XEX; xn~O}, Xo={XEX; xn=O}.
Let P(x, D) be a differential operator of order m in X with analytic
coefficients such that X 0 is non-characteristic. Set x' = (x l' ..• , X n _ 1)'
Theorem 9.6.9. If uEB(X +) satisfies the equation P(x, D) u = and °

(x~,~~)¢WFA(D~ulx"=o), j=O, ... ,m-1,
then there is an e > ° such that
(X,~)¢WFA(U) ijO<xn<e, Ix'-x~I+I~'-~~I<e.
Note that there is no condition on ~n'
Proof According to Corollary 9.5.4 we can extend u to uoEB(X) so

that U o vanishes in X_and
P(x,D)u o = L v/g)D~bo(xn)=f
j<m
Vj is a linear combination of derivatives of the boundary values of D~ u

so (x~, ~~)¢WFA(Vj)' Thus there is a positive eo such that
(X,~)¢WFA(f) if Ix'-x~I+I~'-~~I<eo and XEX.
If e 1 < eo is sufficiently small and M is sufficiently large we also have
Pm(x,~H=O when Ix'-x~I+I~'-~~I<el' Ixnl<e l , l~nl>M,
for Xo is non-characteristic. Hence (x, ~)¢WFA(UO) for such x, ~ by
Theorem 9.5.1. When xn=O this remains true for all ~n by Corol-
lary 9.6.8. Since WFA(u O) is closed we can find e < el so small that
(X,~)¢WFA(UO) if Ix'-x~I+I~'-~~I<e, IXnl<e,l~nl~M.
This completes the proof since U o = u in X + .
Remark. The analogue of Theorem 9.6.9 with WFA replaced by WF is

false even when P has constant coefficients. In fact, if P(D)=D I D2
+D~+D! we can by Example 8.3.4 choose U oEC 2 so that P(D)uo=O
()()
and singsuppuo is the XI axis. Then u=LajUO(xl,X2,X3,X4-1/j) is

I
in C 2(1R4),U and D4 u are in Coo(1R 3 ) when X4=O, and singsuppu
={(x I ,O,O,X4); X4=O or 1/X4 EZ"+} if aj converges to °
sufficiently
rapidly. Since P(D) u=o we have a counter-example to the COO anal-
ogue of Theorem 9.6.9. (A very general version of this example can be
obtained from Theorem 11.3.1.)
Notes
The Fourier transform of a function f on 1R is the sum of boundary

values of the analytic functions F... and F_ in the upper and lower half
Notes 369
planes defined by the Fourier-Laplace transforms off on lR_ and lR+

respectively. (See also a discussion after Theorem 7.1.5.) These are well
defined if say If(x)I=O(eelxl) for every 8>0. Carleman [2J used this
observation to define generalized Fourier transforms. However, taking
boundary values in a classical sense requires additional limitations on
the growth of f (see Beurling [2,3J). One is therefore led to define
boundary values in an abstract sense by introducing the space of pairs
(F+,F_) of functions F± analytic when ImzzO modulo the pairs (F,
- F) where F is entire. This is how Sato [lJ first defined hyper-
functions on lR, and Sato [2J later extended the idea to lRn. This is
technically cumbersome since an invariant setup involves a relative
cohomology group of degree n. It was pointed out by Martineau [lJ
that a fairly elementary presentation is obtained if one starts instead
from the notion of analytic functional. The crucial existence of a
unique support for a functional carried by a compact set in lR n was
proved by Martineau using quite elementary facts on the cohomology
of the sheaf of germs of holomorphic functions. While we follow
Martineau on the whole in Sections 9.1 and 9.2 we have eliminated
even these prerequisites by using the simple observation that solving a
Dirichlet problem gives an isomorphism between analytic functionals
and certain harmonic functions. The latter can be studied with the
methods developed in the course of the preceding chapters. The ad-
vantage is that we just have to study one equation rather than the
overdetermined Cauchy-Riemann system in several complex variables.
After the basic definitions in Sections 9.1 and 9.2 we study WFA(u)
for hyperfunctions in Section 9.3. As already mentioned in the notes
to Chapter VIII this notion was first introduced by Sato [3]. The
proofs in Sections 8.4 and 8.5 were chosen so that they can be used
with small modifications in the case of hyperfunctions. A crucial
technical tool is Theorem 9.3.7, the second part of which is the edge
of the wedge theorem of Martineau [2].
In Section 9.4 we leave hyperfunctions momentarily to study the
analytic Cauchy problem. After a proof of a local existence theorem
along the same lines as in the predecessor of this book, we give
refinements concerning the existence domain taken from Leray [2],
Zerner [3J and particularly Bony and Schapira [1]. This prepares for
the proof by Bony and Schapira [lJ given in Section 9.5 of the non-
characteristic regularity theorem due to Sato [3]. We also define
Cauchy data of hyperfunction solutions of differential equations on
non-characteristic surfaces following Komatsu [2J and Schapira [1].
Section 9.6 presents the Bros-Iagolnitzer definition of It'FA (u) in the
spirit of Sjostrand [1]. The proof of the Kashiwara theorem (Theo-
rem 9.6.6) is also essentially taken from Sjostrand [1]. For the appli-
cation in Theorem 9.6.9 see also Schapira [3], Sjostrand [2]. A very
broad survey of analytic regularity theory can be found in Sjostrand
[1,2].
The aim of this chapter has just been to give an introduction to
hyperfunction theory which follows Schwartz distribution theory as
closely as possible. The reader who wants to study the subject in
depth should of course turn to the basic paper by Sato-Kawai-Ka-
shiwara [1]. It may then be useful to consult also the introductions
given by Kashiwara [1] and Cerezo-Chazarain-Piriou [1].
Exercises
In the following exercises some standard notation is used without ex-

planation. In particular, H denotes the Heaviside function, the char-
acteristic function of the positive real axis, and oa denotes the Dirac
measure at a. The Fourier transform of u normalized as in Section 7.1
is denoted by u.
Most of the exercises are intended to train the student in the routine
use of the tools developed in the text. A few are extensions of the theory
presented there. As a rule a rather complete though brief discussion is
then given in the answers and hints following the exercises.
Chapter I
Exercise 1.1. Let f E COO(R) be an even function. Prove that there is a
function g E COO(R) such that f(x) = g(X2).
Exercise 1.2. Show that every f E COO(R) can be written in the form
with go and g, in COO(R).

Exercise 1.3. Show that when f E coo(Rn) one can find go, gl in coo(Rn)
such that
Exercise 1.4. Show that when f E coo(Rn) one can find a decomposition
f(x) =
with gi, ...ik E coo(Rn) and k even (odd) if f is even (odd).

372 Exercises
Exercise 1.5. Show that when f E c co (R2) and f (Xl, X2) == f (X2, xt), then
one can find g E C CO (R2) such that
Exercise 1.6. Show that there exist numbers ak and bk, k = 0, 1, ... such
that
co
(ii) L lakbkl < 00, n = 0, 1, ...
k=O
co
(iii) Lakbk = 1, n = 0, 1, ... (iv) bk --+ -00 as k --+ 00.
k=O
Show that if f E CCO(R+) and g E CCO(R) is chosen equal to 1 in (-00,1),

equal to 0 in (2, (0), then
co
f (t) = L akg(bkt)f (bk t), t < 0,
o
gives an extension of f which is in CCO(R). Show that the extension is
in CV(R) if f E CV(R+).
Exercise 1.7. f is a locally bounded real valued function on R with
f(x) = 0 when X < 0 and f(x) = O(XN) for every N > 0 as X --+ O.
Show that there is a function F E CCO(R) such that F(x) = 0, X < 0,
and f(x) :s; F(x), x E R.
Chapter II
Exercise 2.1. For which bEe does there exist a distribution U E ~'k(R)
with restriction X ~ x b to R+ ?
Exercise 2.2. Does there exist a distribution u on R with the restriction
x ~ e l / x to R+ ?
Exercise 2.3. For which a > 0 and bEe does there exist a distribution
u E ~'k(R) with restriction x ~ eix-·x b to R+?
Exercise 2.4. Show that for every f E CI(R+) one can find a real valued
function g E CI(R+) and a distribution u E ~'I(R) with restriction fe ig
to R+!
Chapter II 373
Exercise 2.5. Determine limt-+too ft in f!)' (R) when
Exercise 2.6. Determine the following limits in f!)' (R) :
a) limt 2 xcostx b) limt2 1xlcostx c) lim x-lsintx.

t -+00 t-+oo t--++oo
Exercise 2.7. Find the limit of ft(x) = te itx log Ixl in f!)'(R) as t --+ +00.
Exercise 2.S. For which values of a E R is it true that the functions
ft(x) = ta sin (tx) converge to 0 as t --+ +00 a) as C k functions; b) as
distributions?
Exercise 2.9. Let Uj E Cl(X) where X is an open set in Rn , and assume
that for every compact set K c X there is a constant CK such that
Iuil :$; CK on K for every j. Show that if Uj --+ U in f!)'(X) then U E C(X)
and Uj --+ U uniformly on every compact set in X.
Exercise 2.10. Determine a number a and a distribution U =/= 0 such that
U",(t) = (1 - cos t)"', t E R,
is locally integrable when IX > a and (IX - a)u", --+ U in f!)'(R) when
IX --+ a + O. Show that the limit
exists and determine it.

Exercise 2.11. Prove that if f (x) = (Xl" . xn + ie)-l + (Xl' .. xn - ie)-l,
X ERn, then lime--.ofe exists in f!)'(R n).
Exercise 2.12. Put fw(x) = 1/(x4 + w), X E R, where w belongs to the

open right half plane H in C. Determine homogeneous functions aj(w)
in H, 0 :$; j :$; 2, such that
2
fw - ~>j(W)DU)
o
has a limit in f!)' (R) as H :3 W --+ O.
Exercise 2.13. Put Xe(x) = x(x/e)/e where X E CO'(R) and JXdx = 1.
Determine a constant C and a distribution U such that the distribution
Ke = Xe(xy) + ClogeDo(x,y) --+ U when e --+ +0.
374 Exercises
Exercise 2.14. Define ua,e(x) = 8/(82 + (x - aJe)2), x E R, where a E R

and 8 > O. Determine lime->o Ua,e and lim.-.o Ua,eUb,e when they exist in
.@'(R).
Exercise 2.15. Determine the limit in !?&'(R2 \ 0) as t -+ +00 of
Does the limit exist in !?'&'(R2 )?

Exercise 2.16. Set ftAx) = e-itX(x + i8)-1, x E R, and determine the
following limits in .@'(R):
a) b)
Section 3.1
Exercise 3.1.1. Let f E !?&' (1) where I is an open interval on R. Show
that there is a solution u E !?&'(I) of the differential equation u' = f,
and that the difference between two such primitive distributions is a
constant.
Exercise 3.1.2. Let u E !?&' (I), where I is an open interval c R. Show
that if U has order k > 0, then u' has order k + 1.
Exercise 3.1.3. Let u E !?&' (I) where I is a finite open interval c R.
Show that if u is the restriction to I of a distribution of order k in a
neighborhood of Y, then
lu(cp) I :s; C I sup IcpUlj, cp E C(f(I),

j5.k
and that conversely this estimate implies that there is a measure dJ.l on R
with support in Y such that u is the restriction to I of its kth derivative.
Exercise 3.1.4. Show that if f is a measurable function in (-1,1) and
where m is a positive integer, then there is a distribution FE tB',m([-I, 1])

with restriction f to (-1,1).
Section 3.1 375
Exercise 3.1.5. Does there exist a distribution u E ~'(R) which restricts

to the function x ..... el / x exp (ie l / x ), x > O?
Exercise 3.1.6. For which a E R does there exist a distribution U E ~'k(R)
with the restriction x ..... e l / x exp (itt/X) to R+ ?
Exercise 3.1.7. Prove that the limit
(vp(l/x), q>} = lim

.-+0
1
Ixl>.
q>(X) dx/x, q> E CO'(R),
exists. What is the order of the distribution vp(l/x)?

Exercise 3.1.8. J is an odd locally integrable function on R such that
xJ(x) ~ 1 as x ~ 00. Prove that gt(x) = t 2 f'(tx) has a limit in ~'(R)
and determine it.
Exercise 3.1.9. Determine real numbers at. a2 such that the integral
exists when q> E CO'(R) and u = d2(vp(1/x»/dx2.

Exercise 3.1.10. Define log(x + iO) = log Ixl + niH(-x), x E R, calculate
the derivative, and compare it with vp(l/x).
Exercise 3.1.H. Show that the function
has a limit in ~'(R) when Il ~ 0 and give it in a simple form.

Exercise 3.1.12. Compute the nth derivative of x ..... J (Ixl) when J E
C"(R+).
Exercise 3.1.13. Compute the nth derivative of x ..... IxlJ (x) when
JE C"(R).
Exercise 3.1.14. Let 1 be an open interval c: R, and let a E I. a) Show
that for every J E ~'(/) there is a solution u E ~'(I) to the equation
(x - a)u = J, and that two solutions of this division problem differ by a
multiple of ~a. b) Give a solution when J = ~Jfl.
Exercise 3.1.15. Show that if 1 is an open interval on Rand F E CaV)
has no zero of infinite order, then the equation Fu = g has a solution
u E ~'(/) for every g E ~/(I). Describe the solutions when g = O.
376 Exercises
Exercise 3.1.16. Show that if F E C<Xl(R) and the equation Fu = 1 has

a solution u E ~'(R), then F cannot have a zero of infinite order.
Exercise 3.1.17. Compute xjDJ k) for all integers j ~ 0, k ~ O.
Exercise 3.1.18. Compute / bJk) when / E Ck(R).
Exercise 3.1.19. Determine all primitive distributions of a) H(x);
b) xH(x); c) eXH(x); d) Do; e) vp(l/x).
Exercise 3.1.20. Determine all u E ~'(R) satisfying the equations
a) xu' = bo, b) xu' +u=O, c) x 2u' +u = 0,

d) x 2u' + xu = Do, e) (x-l)u = D~, f) (x 2 - l)u = Do,
g) (exp (2nix) - l)u = 0, h) u" = b~ -2b[, i) xu' = DI - D_I,
j) (x + 1)3 u' + u = b, k) x 4u' + u = 0, I) 2

cos xu' = O.
Exercise 3.1.21. Determine all distributions in R n with x~ u = 0 where

N is a fixed positive integer.
Exercise 3.1.22. Determine all distributions in R2 with (xi - x~)u = 0
and XIX2U = O.
Exercise 3.1.23. Let cp, cp E C<Xl(Rn) be real valued, iJCP/iJxI -=F 0 when
cp = O. When does u E ~'(Rn), cpu = 0, CPiJU/iJxI = 0 imply u = O?
Exercise 3.1.24. Show that UN = lim.-->+o(xi + e + iX2)-N exists in ~'(R2)
for every integer N > O. Calculate / N = iJUN / iJXI + 2ixI iJUN / iJx2,
determine the order of UN and sing supp UN.
Exercise 3.1.25. Assume that 0 :-:;; g E C<Xl(R2) and set u.(x) = e(g(x) -
ie)-2, x E R2, e > O. Prove that lim.-->o u. exists in ~'(R2) if (and only if)
What is the limit?

Exercise 3.1.26. Let / and g be real valued functions in C<Xl(Rn),
n ~ 2, and assume that f'(x) -=F 0, g'(x) -=F 0, x ERn. Show that
lime-->+o H(f(x))/(g(x) + ie) exists in ~'(Rn) if log Ig(x)1 is locally inte-
grable on the surface /-1(0).
Exercise 3.1.27. Set sn(z) = 1 + Z + ... + zn, z E C = R2. Show that
u = limn--><Xl n-Ilog ISn(z) I exists in ~'(R2), and calculate u and i\u.
Exercise 3.1.28. Let u be the characteristic function of the unit disc in
R2. Calculate xiJu/iJx + yiJu/Oy.
Section 3.2 377
Exercise 3.1.29. u is the characteristic function of the unit disc in R2.

Determine the order of the distributions
Exercise 3.1.30. For which positive numbers a and b are the first
derivatives of the characteristic function u of
of order o?
Exercise 3.1.31. Put F(x,y) = x 2 + f(Y), where 0:::;; f E CCO(R). What is
the condition for the existence of a distribution u E ,@'(R2) with Fu = 1.
Exercise 3.1.32. Determine all real valued functions q> E C 1(R+) such
that
x/t, when x :::;; q>(t)
u(t, x) = {
-1, when x > q>(t)
satisfies the equation ou/ot + o(u2 /2)/ox = 0 in the distribution sense
when t > O.
Exercise 3.1.33. Set f(z) = (z -1)-3 10gz, where -n :::;; Imlogz < n.
Determine the limit of x 14 f(x + ie) - f(x - ie) in .@'(R) as e -+ +0.
Section 3.2
Exercise 3.2.1. For a E C define Z(a) = {u E .@'(R);xu' = au} and prove
(i) If u E Z (a) then lxi-au is a constant in R+ \ 0 and in R_ \ O.
(ii) If u E Z(a) and suppu = {O}, then a = -j - 1 and u = Cb~) for
some integer j ~ O.
(iii) Ifu E Z(a) then xu E Z(a+ 1) and u' E Z(a-l).
(iv) The maps fx : Z(a) -+ Z(a - 1) and ~ : Z(a - 1) -+ Z(a) are
bijective and each other's inverses if a =F 0, while fx
maps Z (0)
to the multiples of bo and x maps Z(-I) to the constants.
(v) The dimension of Z (a) is equal to 2.
(vi) If 0 =F u E Z (a) then the order of u is the smallest integer k ~ 0
such that k + Re a + 1 > 0, unless supp u = {O} and the order is
-a-I.
Exercise 3.2.2. For a E C and a positive integer k define
Z(a, k) = {u E .@'(R); (xd/dx - a)ku = O}.

378 Exercises
Prove that
(i) If u E Z (a, k) then lxi-au is a polynomial in log Ixl of degree < k
on R+ and on R_.
(ii) If u E Z (a, k) and supp u = {O}, then a = -j - 1 and u = C~~)
where j is a non-negative integer.
(iii) Ifu E Z(a,k) then xu E Z(a+ l,k) and u' E Z(a-l,k).
(iv) The composition of the maps fx : Z (a, k) --+ Z (a - 1, k) and
x : Z (a - 1, k) --+ Z (a, k) is the sum of a nilpotent map and a
times the identity, so the maps are bijective if a i= o. xfx maps
Z(O,k) onto Z(O,k-l), fxZ(O,k) = {v E E0'(R);xv E Z(O,k-l)};
fxx maps Z(-I,k) onto Z(-I,k - 1), and xZ(-I,k) = {u E
E0'(R);u' E Z(-I,k-l)).
(v) dimZ(a,k) = 2k, and ilJ(Z(O,k) c Z(-I,k) c fxZ(O,k+ 1) with
codimension 1 in each inclusion.
(vi) If 0 i= u E Z (a, k) then the order is the smallest integer m ;;::: °
°
such that m + Rea + 1 > unless suppu = {O} and the order is
-a-I.
Exercise 3.2.3. Determine the dimension of the space of solutions

uE E0' (R) of the differential equation
I
m
aj(xd/dx)ju = 0,
o
where aj are constants and am i= 0.
Exercise 3.2.4. Show that if u E E0,k(Rn) and the restriction of u to Rn \
is homogeneous of degree a and not == 0, then k + Re a + n > 0.
°
Exercise 3.2.5. Determine the order and the degree of homogeneity of
the distributions
u(cp) = fooo fooo (cp(x, y) - cp(-x, y) - cp(x, -y)
+cp(-x,-y))dxdy/xy, cp E COO(R 2 ),
v(qJ) = fooo(qJ~(X'O) -qJ~(-x,O))dx/x, cp E COO (R2).
Section 3.3
Exercise 3.3.1. Calculate Jl. = L\log If I when f is a merom orphic function
in a connected open set Z c C = R2 and f =1= 0.
Section 3.3 379
Exercise 3.3.2. Assuming that A(x) = allxt + 2a12XtX2 + a22x~ =1= 0 and
that Re A (x) ~ 0 when 0 =1= x E R2, determine a constant C such that
Clog A(x) is a fundamental solution of a220r - 2a120t 02 + all 01.
Exercise 3.3.3. Determine a fundamental solution E of A2 in R n , n > 2.
Exercise 3.3.4. Compute Au where u(x) = ealxl/lxl, x E R3; a E C.
Exercise 3.3.5. Compute Au when u(x) = (sin Ixl)/Ixl, x E R3.
Exercise 3.3.6. Determine a fundamental solution of A + a2 in R3.
Exercise 3.3.7. Let A E Ct(R) and A(x) =1= 0 if x =1= O. Prove that
E(qJ) = lim
8-->0
ff qJ(x,y)/(A(x) + iy) dxdy
Ixl>B
= j dx(j qJ(x,y)/(A(x) + iy) dy)
exists when qJ E CJ(R 2) and that E E ~'t(R2). Calculate f = oE/ox +

iA'(x)oE/oy.
Exercise 3.3.8. Prove for every positive integer N the existence of the
limit
UN(qJ) = ~~ j j (x + iy)-N qJ(x,y) dxdy, qJ E Cij(R2),

Ix+iyl>B
and that UN E ~'(R2) is homogeneous of degree -N. Calculate f =

oUN/ox+ioUN/OY·
Exercise 3.3.9. Show that if f is an analytic function in an open
connected set Z c C = R2 and f =1= 0, then the limit
U(qJ) = lim
8-->0
ff f7(X'~»
x + zy
dxdy, qJ E CO'(Z),
If (x+iy) I>B
exists and defines a distribution with f u = 1.

Exercise 3.3.10. Find a fundamental solution E of 0" in R n with support
in the first quadrant, when alllXj are positive.
Exercise 3.3.11. Determine a constant C such that
E (x, t) = 0 otherwise;
380 Exercises
is a fundamental solution of the wave operator c 2 82 /8t 2 - 82 /8x 2 •

Exercise 3.3.12. Determine an entire analytic function F such that
E(x,y) = F(cxy) when x;;::: O,y;;::: 0; E(x,y) = 0 otherwise;
is a fundamental solution of the differential operator 8 2 jax8y -c where

c E C.
Exercise 3.3.13. In R4 with coordinates denoted (t, x), t E R and x E R3,
let u be the characteristic function of the light cone {(t, x); t > lxI}, and
calculate v = 0 U, W = 0 v, where 0 = 8t - A is the wave operator.
Section 4.1
Exercise 4.1.1. Show that if Ue = sgn t X~ (t2 - <:2), 0 =F <: E R, then
lime->o U e exists in §' (R) for any a E C. (The distributions X~ were
defined in Section 3.2. The definition of the composition is obvious
when Re a > -1 and is given by analytic continuation otherwise.)
Calculate the limit when a is a negative integer.
Exercise 4.1.2. Let U be subharmonic in {z E C; Izi < R} and set fl, = Au.
Prove that
1 o
211:
u(re ill ) dB - 2nu(0) =
1Izl<r
r
log -II dfl,(z),
Z
0< r < R.
Exercise 4.1.3. Calculate AI 1m J f (z) I where f (z) = z2 + a, a E R.

Exercise 4.1.4. Let fn(z) = max Re w taken over all w E C with w n = z,
where n is a positive integer. Calculate Afn(z).
Section 4.2
Exercise 4.2.1. Calculate f * f * ... * f (with n factors) if a) f(t) = H(t)
b) f(t) = e-tH(t).
Exercise 4.2.2. Calculate (jJkl * H where k is a positive integer.
Exercise 4.2.3. Let fa be the characteristic function of (0, a) c R, where
a > O. Determine a distribution U a with support on R+ such that
fa * U a = (jo.
Section 4.3 381
Exercise 4.2.4. Calculate (1 * (jo) * Hand 1 * ((jo * H).

Exercise 4.2.5. Prove that ((jh * U - u)/h ~ -u' in ~'(R) as h ~ 0, if
u E ~'(R).
Exercise 4.2.6. Recall that the distributions xi E ~' (R) depend analyti-
cally on A E C and are defined by xi(x) = x" /r(A+ 1), x > 0, xi(x) = 0,
x $; 0, when ReA > -1. Determine xi * x~ for arbitrary A, Jl E C.
Exercise 4.2.7. Solve Abel's integral equation xi * u = f where f is a
given distribution with support on R+ and the solution u is also required
to have its support there.
Exercise 4.2.8. u and v are the surface measures on the spheres {x; Ixl
=
a} and {x; Ixl = b} in R3. Compute the convolution u * v and determine
its singular support.
Section 4.3
Exercise 4.3.1. Show that supp (u * v) = supp u + supp v if u and v are
positive measures in Rn , one of which has compact support.
tt'(R n ) and u * v = 0, it follows from the theorem
Exercise 4.3.2. If u, v E
°
of supports that u = or v = 0. Is this true if only one of the factors u
and v has compact support?
Exercise 4.3.3. Show that if u, v E tt' (Rn) and supp u * v is contained
in an affine subspace V of Rn, then the supports of u and of v are
contained in affine subspaces parallel to V.
Exercise 4.3.4. Let u be the characteristic function of the square in
R2 defined by Ixd < 1, IX21 < 1, and let f = P(iJ)u where P(iJ) is a
differential operator with constant coefficients. Describe the possible
sets supp f which can occur and the corresponding polynomials.
Section 4.4
Exercise 4.4.1. Calculate Jl = L\ log If (z) I where f is analytic outside
[-1,1] and f(z)2 - 2zf(z) + 1 = 0, f(2) > 1. Calculate Jl * E where
E(z) = (2n)~110g Izl.
f (x) = Ixl~5 I,J,k=l ajk XjXk, x E R3 \ 0, where (ajk) is

Exercise 4.4.2. Set
a constant symmetric matrix. a) What is the condition for the existence
382 Exercises
1
of the limit
F(cp) = lim f(x)cp(x) dx
£->0 Ixl>£
for arbitrary cp E CO(R3)? b) Calculate AF when this condition is

fulfilled and prove that F = E * AF where E(x) = -1/(4nlxl).
Exercise 4.4.3. V is a function on R3 such that V (x) ~ 0 as x ~ 00, and
V is harmonic outside a compact set. Show that -4nlxlV(x) ~ (AV, 1),
x ~ 00. Is the hypothesis Vex) ~ 0 essential?
Exercise 4.4.4. Let u E ~' (R) and assume that for some integer k 2 0
we have u * f * f E COO(R) for every f E C§(R). Show that u E COO(R).
Exercise 4.4.5. Let u be a distribution in Rn with compact support, and
assume that fk = (a l ... an)k u is a continuous function for k = 1,2,3, ....
Show that u E C;)(Rn).
Exercise 4.4.6. Show that a differential equation P(d/dx)u = f, where
f E 6"'(R) and P is a polynomial, has a solution u E 6"'(R) if and only
if (f, cp) = 0 for every solution cp of the adjoint differential equation
P(-d/dx)cp = O.
Exercise 4.4.7. Let X be the characteristic function of (-1, 1). Determine
a number a E (0,1) and a function u E CJ(R) such that
X- ba - b_ a = d4 u/dx 4 .
Show that u 2 0, compute I = S u dx and show that
1il I
f dx - f(a) - f (-a) 1:::; I max If(4)1, f E C 4 ([-1, 1]).
Exercise 4.4.8. Let f E 6'" (Rn) and let (1. = ((1.1, ... , (1.n) be a multi-index.
Show that there exists some u E 6"'(Rn ) with aau = f if and only if
(f, x P) = 0 for all multi-indices not satisfying the condition f3 2 (1..
Exercise 4.4.9. Show that if f E 6'" (R2) then the equation Au = f has
a solution u E 6'" (R2) if and only if (f, cp) = 0 when cp(x, y) = (x ± iy)n,
n = 0, 1,2, ...
Section 5.1
Exercise 5.1.1. Show that if u E ~'k, V E ~'l, then u ® v E ~'k+l, and
that u ® v E ~'N implies u E ~'N, v E ~'N unless u or v equals O.
Section 5.2 383
Exercise 5.1.2. Construct for given positive integers k and 1 two dis-
tributions U E .@lk(R) and v E .@II(R) such that U ® v is not of order
k+l-l.
Exercise 5.1.3. Construct for a given positive integer N two distributions
Uo,UI on R which are not of order N - 1 such that UO ® UI is of order
N.
Section 5.2
Exercise 5.2.1. Let f be a continuous function from R to R. Which
operator has the distribution kernel oH(y - f(x»/oy?
Exercise 5.2.2. What is the kernel of the operator
fcp(x) = cp(x) + 1: a(x,y)cp'(y)dy, cp E CO'(R),
where a E C(RZ)?
Exercise 5.2.3. K is a measurable function in X I X X Z where Xj is an
open subset of Rnj, such that
j Xl
IK(X,Y)1 dX::5: A, for almost all y E Xz;
j X2
IK(X,Y)1 dy ::5: B, for almost all x E XI.
Prove that for the corresponding operator f
Section 6.1
Exercise 6.1.1. Calculate ba(cosx) when -1 < a < 1.
Exercise 6.1.2. Calculate U = b~(f) in RZ \ 0 when f(x) = XIXZ.
Exercise 6.1.3. Determine the limit of CPe(x z - yZ)CPe(Y - 1) in .@/(Rz) as
e ---+ +0, where cp E CO', f cp(x) dx = 1, and CPe(t) = qJ(t/e)/e.
384 Exercises
Exercise 6.1.4. Let f,g be real valued functions in COO(X), X open in

Rn , such that df and dg are linearly independent when f = g = O.
Determine U = (j(f,g).
Exercise 6.1.5. Let f,g E coo(Rn) be real valued, df =1= 0 when f = 0
and dg =1= 0 when g = o. Show that if q> E CO'(Rn), q> E CO'(Rn) and
U = (q>(j(f» * (q>(j(g», then sing suppu is contained in
{x + y;x E rl(O) n suppq>,y E g-I(O) n suppq> and

df(x),dg(y) are linearly dependent}.
Give an integral formula for u valid in the complement of this set.

Exercise 6.1.6. Set ut(x) = (f(x) + ie)-l where f E coo(R) is real valued.
Determine the condition on f required for the existence of the limits
u± = limt-+±o U t , and calculate u+ - u_ then.
Exercise 6.1.7. Show that if s > 0 and k is a positive integer, then the
function x ~ (.x2 k - s2k + ie)-l has a limit fs E ~'(R) as e -. +0. Show
that one can find UQ, .•• , Uk E ~' (R) such that
I
k-l
fs - i HI - 2k Uj -. Uk, S -. 0,
o
and determine support and order for these distributions.
Section 6.2
Exercise 6.2.1. In RxRn, with variables denoted (t, x), let 0 = iJ2 /at 2 -l1x
be the wave operator. Calculate the fundamental solution Ek of Ok+l
with support in the forward light cone {(t,x);t ~ Ixl} for every integer
k~O.
Exercise 6.2.2. Find the forward fundamental solution Fa of 0 - a for

every a E C, with notation as in the preceding exercise.
Exercise 6.2.3. Find the forward fundamental solution F of the operator
o + 2boa t + 2 'L7 bjaj + c for arbitrary complex bo, . .. , bn, c, with notation
as in the preceding exercises.
Section 7.1
Exercise 7.1.1. For which even positive integers m and n is f (x) =
exp(xn + i exp(xm» in 9" (R)?
Section 7.1 385
Exercise 7.1.2. Let M be an unbounded subset of Rn. Show that for

every integer m there is a distribution u E 9'" (Rn) with supp u c: M such
that the order of uin the unit ball is > m.
Exercise 7.1.3. Show that if u is a measurable function on Rn and m is a
positive integer, then u E 9'" and uE !!)'m if Jlu(x)1 2(1 + IxI2)-m dx < 00.
Exercise 7.1.4. Prove that if K is a compact subset of Rn and ~j ERn,
I~j - ~kl ~ 1, j 1= k, then
I
00
1
1«p(~j)12 ~ CK J IqJ(x)1 2 dx, qJ E CO'(K),
with CK independent of the sequence ~j.
Exercise 7.1.5. Show that if ~j E R I~j - ~kl ~ 1, j 1= k, and m is a

n,
non-negative integer, then Lajei(X,~j) converges in 9'" to a sum of order

~ m if aj E C and L lajl2(1 + l~jI2)-m < 00.
Exercise 7.1.6. Let u E 9'" (Rn). When does there exist a function f E 9"
such that u = u * f?
Exercise 7.1.7. Show that if u E LP(Rn) and I~I ~ A when ~ E supp U,
then IIu'liv ~ CAliuliv where C only depends on n.
Exercise 7.1.8. Show that if u E Loo(Rn) and ~ E supp U, then one can
find a sequence qJj E 9" such that lu * qJjl ~ 1 and u * qJj(x) ~ ei(x,l;)
uniformly on every compact set.
Exercise 7.1.9. When does a differential equation P(D)u = 0 with
constant coefficients have a solution 1= 0 in a)!!)' b) 9'" c) Iff' d) COO
e) 9".
Exercise 7.1.10. Let f E Ll(Rn) and f * f = f. Find f·
Exercise 7.1.ll. Show that the equation u - u * f = f for a given
f E 9"(Rn) has a solution u E 9" if and only if! 1= 1.
Exercise 7.1.12. Show that if u, v E 9"'(R) have supports on the positive
half axis, then u * v E 9"'(R).
Exercise 7.1.13. Let ua(x) = 1/llogxla when 0 < x < ~, u(x) = 0 when
x < 0 or x > 1, and u E COO when x > 0; here a > O. Determine the
limit of u(~)~(log I~ I)a as ~ ~ 00.
Exercise 7.1.14. What is the Fourier transform of the space Z (a, k) in
Exercise 3.2.2?
Exercise 7.US. Set!Fu = (2n)-n/2u when u E 9"(Rn). Prove that
a) !F4 = I, the identity, and that every u E 9"(Rn) has a unique
386 Exercises
decomposition
b) Show that the differential operators Lvu = XvU + ovu, V = 1, ... , n,

are surjective on 9'(Rn), determine the kernels and show that § Lv Uk =
jk+l Lv Uk for the terms in the decomposition.
Exercise 7.1.16. Show that if K is a continuous function in Rn then the

following conditions are equivalent:
(i) The convolution operator <p f-+ K * <p is positive on CO', that is,
(K * <p,<p) ;;::: 0, <p E CO'(Rn), where L·) is the L2 scalar product.
(ii) L~k=l K(Xj-Xk)tik;;::: 0 for all Xl, ... ,XN ERn and t" ... ,tN E e,
N = 1,2, ....
(iii) K = fi where {l is a positive measure with finite total mass,
(Il, 1) = K(O).
Exercise 7.1.17. Show that if K E 0J'(R n) then the following conditions

are equivalent:
(i) (K * <p, <p) ;;::: 0 for all <p E CO'(Rn).
(ii) K = fi where Il is a positive measure and J(1 + Ixl}-N dll < 00 for
some N.
Exercise 7.1.1S. Let f be a bounded continuous function on R with

J = 0 in a neighborhood of o. Show that the primitive functions u of f
are bounded. What is the support of u?
Exercise 7.1.19. What is the Fourier transform of Rn :3 x f-+ ei (x,8) ?
Exercise 7.1.20. Find the Fourier transform of the following functions
onR: a)xf-+xsinx b)xf-+xsin2 x c)xf-+x-1sinx d)xf-+
(sin x)k (k positive integer) e) x f-+ xH (x) f) x f-+ H (1 + x) + H (1 - x)
g) X f-+ sgnx = H(x) - H(-x) h) x f-+ sin Ixl i) x f-+ 1/(1 + x 2) j)
Xf-+ x/(I+x 2) k) Xf-+ x 3/(I+x 2) 1) Xf-+ arctan x m) Xf-+ x 3/(I+x 4).
Exercise 7.1.21. Use Parseval's formula to calculate the integrals
Exercise 7.1.22. Find the Fourier transform of the function x f-+ Ix2 -11
on R.
Section 7.1 387
Exercise 7.1.23. Find the Fourier transform of the distribution x ~

(X 2 - S2 + iO)-1 on R, where s > O.
Exercise 7.1.24. Calculate the Fourier transform of the function ft(x) =
(cos X - e-itX)/x, x E R, and then e'oo 1ft (x) 12 dx.
Exercise 7.1.25. Determine the limit as & --+ +0 of the fundamental
solution Ee E !/' of the differential operator i&(d/dx)4 + (d/dx)2 + 1 on
R.
Exercise 7.1.26. Find all solutions U E !/(Rn) of the differential equation
n
Au + LXjou/oXj + nu = O.
1
Describe the solutions in !/' (Rn) also.

Exercise 7.1.27. Calculate the Fourier transforms of the following
functions in R2: a) x ~ H(xt}H(X2) b) x ~ x2/((1 + xi)(1 + x~»
c) x ~ xle-"x~ d) <5;(XI) ® e-xV2.
Exercise 7.1.28. Let f be a continuous function on R with f(x) =
i
1/x + O(lxl-2), as x --+ 00. Show that is a function which is continuous
except at the origin where left and right limits exist. Determine the
jump j(+0) - j(-O).
Exercise 7.1.29. Extend the preceding exercise to functions f with
k
f(x) = L ajx-j + O(x- k- I ), x --+ 00,
o
where k is a positive integer.

Exercise 7.1.30. Let f be a continuous function on R with f(x)
a+/x + O(lxl-2) as x --+ ±oo. Prove that g(~) = im
+ (a+ - a_)log I~I
has a limit as ~ --+ ±O, and determine g(+O) - g(-O).
Exercise 7.1.31. For which a E C does the differential equation
xu" + 2u' + (a - x)u = 0
have a solution =1= 0 in !/'(R) such that the limits u(±O) exist?
Exercise 7.1.32. Determine the Fourier transform of R :3 x ~ f (e iX )
where f is an analytic function in a neighborhood of the unit circle.
Work out the special case f(z) = z/((2z - 1)(z - 2» explicitly.
388 Exercises
Exercise 7.1.33. What is the Fourier transform of R 3 x ~ Ixl-a, where

a E C and 0 < Rea < 1.
Exercise 7.1.34. What is the Fourier transform of R 3 x ~ Ix + 11-h
Exercise 7.1.35. Calculate the Fourier transform of fa (x) = lxi-a, x ERn,
where 0 < Rea < n.
Exercise 7.1.36. Show that if u,,(x) = x"lxl-n, x ERn, where lX is a
+
multi-index 0, then u"R) = cn(ia~)" log I~I, and calculate the constant
Cn·
Exercise 7.1.37. What is the Fourier transform of the function R2 3 x ~

A(x)-! where A is a positive definite quadratic form.
Exercise 7.1.38. Calculate the Fourier transform of a function u E
C (R2 \ 0) which is even and homogeneous of degree -1.
Exercise 7.1.39. Find the Fourier transform of
Exercise 7.1.40. Find the Fourier transform of the distribution (x,y) ~

1/(x + iy) in R2.
Exercise 7.1.41. Find the Fourier transform of R3 3 x ~ e~lxl.
Exercise 7.1.42. Let U E tS"(R 3) be the surface measure on the unit
sphere. Find the Fourier transform.
Exercise 7.1.43. Find the Fourier transform of the distribution U =
15b(lxI2 -1) in R3.
Exercise 7.1.44. Find the Fourier transform of the distribution UN III
Exercise 3.1.24.
Exercise 7.1.45. Find the Fourier transform of the distribution UN III
Exercise 3.3.8.
Exercise 7.1.46. Find the Fourier transform of the distribution F in
Exercise 4.4.2.
Section 7.2
Exercise 7.2.1. Develop in Fourier series the function BI on R with
period 1 and BI(x) x -!,0 < x < 1, and deduce the Poisson
summation formula.
Section 7.2 389
Exercise 7.2.2. f is periodic on R with period 2n and f (x) = cos ax,

Ixl :::;; n. Calculate!" + a2f and develop f in Fourier series.
Exercise 7.2.3. Evaluate the sum S(x) = Lf(cosnx)/(1 + n2 ).
Exercise 7.2.4. Find the Fourier transform of the distributions on R
defined by
a) lim tan (x + ie) b) lim (tan (x + ie))2.
e-+I-O e->+O
Exercise 7.2.5. Show that if fELl (R) and <p E 9'(R), then
If * <p(2nn) = If(n)cP(n),
00 00
2n
-00 -00
with absolute convergence on both sides. Deduce that if f,!',!" E Ll(R)

then this remains true with 0, and calculate L~oofz(n).
Exercise 7.2.7. Find the Fourier transform of R 3 x ~ 1/(1 + x 2) and
determine the limit
00
li~ e21l /e(V e/(1 + e2n2) - n).

e....+O 6
-00
Exercise 7.2.8. Find the Fourier transform of <p(x) = (sin x)2 /x2, x E R,
and calculate L':, <p(x + nn).
Exercise 7.2.9. Show that if U E LOO(R) and <p is the function in the
preceding exercise, then the series
I
00
ue(x) = u(x + nn/e)<p(ex + nn)

-00
converges, the range of Ue is contained in the closed convex hull of that

of U, Ue is periodic with period n/e, and
sUppUe C g +'l;~ E supPu,I'lI:::;; 2e},

Iu(x) - Ue(x) I :::;; 2 sup lul(1 - <p(ex)).
Exercise 7.2.10. Show that if U E LOO(R) is real valued and sup lui < 1,
supp U C [-A, A], then u(x) - cos (Ax) has precisely one zero in each
390 Exercises
interval (nnIA,(n+ l)nIA), n E Z; prove that it is simple and that there

are no other zeros in C.
Exercise 7.2.11. Show that if u is a real valued function with u' E L oo ,
sup lu'l < 1, and supp un (-A, A) = (/), and if h(x) = minnEz Ix - 2nnl AI-
n/2A, then h-u has the same sign as h at every maximum or minimum
point of h. Deduce that sup lui < sup Ihl = n12A. (Bohr's inequality.)
Section 7.3
Exercise 7.3.1. Determine all u E $'(R) \ 0 such that every factorization
u = v * w with v, W E $'(R) is trivial in the sense that one factor is a
Dirac measure.
Exercise 7.3.2. Prove that if u E £p(Rn) and uhas compact support, then
u can be extended to an entire analytic function in en such that
Ilu(· + iy)lIv :0:; eH(-y) Ilullv, y ERn,
where H is the supporting function of supp u.

Exercise 7.3.3. Let u E Loo(R), suppu c (-1,1). Prove that u' = K * u
e e
where K(e) = ie, -1 :0:; :0:; 1, K(e) = i(2 - e), 1 :0:; :0:; 3, and K has
period 4. Calculate K and deduce an estimate for sup lu'l in terms of
sup lui.
Exercise 7.3.4. Prove that if u E £p(R) for some p E [1,00] and supp uc
[-A, A] then
II sin at:u' I A + cos at:ullv :0:; lIullv, at: E R.
Exercise 7.3.5. Prove that if P (D), D = -ia I ax is a homogeneous

polynomial then the polynomials h(x) in R n satisfying the differential
equation P(D)h(x) = 0 are dense in the set of all solutions of the
equation in coo(Rn). Prove that the equation P(D)u = f, f E $', has a
solution u E $' if and only if (f, h) = 0 for all such polynomials h.
Section 7.4
Exercise 7.4.1. Let f E L2«0, (0)). Show that the Fourier-Laplace
transform
j (0 = 10 00
e- ix ( f (x) dx
Section 7.4 391
is an analytic function in the half plane where 1m' < 0 and that
1 o
00 If(x)12 dx = sup (2n)-l JIj(~ + i1])1 2 d~
J Ij(~ +
~<o
= lim(2n)-l i1])1 2 d~.

1/-+0
Characterize the functions 1 obtained in this way.

Exercise 7.4.2. Show that if -00 < al < a2 < 00 and f(x)ffljX E L2(R),
j = 1,2, then the Fourier-Laplace transform 1 is analytic in the strip
g; al < 1m, < a2} and
Show that every analytic function 1 in the strip, such that the right-
hand side is finite, is the Fourier-Laplace transform of a function f with
f(x)e ajx E L2, j = 1,2.
Exercise 7.4.3. Find an analytic function F in the strip {z E C; I 1m z I <
1} such that (1 + z2)F (z) is bounded in the strip and
F(x + i - iO) + F(x - i + iO) = b(x).

Determine the Fourier transform G of x 1-+ F(x) and prove that F is
unique.
Exercise 7.4.4. Let A be a positive definite real n x n matrix. Show that
if f(x)e(Ax,x)/2 is in L2(Rn), then the Fourier-Laplace transform 1(0 is
an entire analytic function with
Show also that every entire analytic function such that the left-hand
side is finite is such a Laplace transform.
Exercise 7.4.5. Let cp(z,y) = !(Az,z}+(Bz,y}+!(Cy,y), z E en, y ERn,
where A and C are symmetric matrices, 1m C positive definite and B
non-singular. Show that if u E L2 then
392 Exercises
is an entire analytic function and that
where $(z, z) = maxyeRn - 1m cp(z, y) and dA. is the Lebesgue measure in

en. Show that all entire functions for which the right-hand side is finite
can be obtained in this way. Characterize the functions U obtained if
u E !/' or u E !/" instead.
Exercise 7.4.6. Let u be an entire analytic function in en such that
where a> 0 and b > O. Prove that the Fourier-Laplace transform
U(O = J e-i(X+iy,Ou(x+iy)dx
is independent of y and that
Show that u = 0 if b > a.
Section 7.6
Exercise 7.6.1. Find the Fourier transform of R3 :3 x 1--+ exp i(xi+x~-x~)
Exercise 7.6.2. Let f E CO'(R'). Prov~ that for every t > 0 there is
a function ft E !/' such that ft(~) = f(~)exp(itl~12), and prove that
1ft (x) I :::;; Ct- n/ 2 for x E Rn and t > 1. Use this to decide for which
p E [1,00] that the Fourier transform of LP consists of measures.
Exercise 7.6.3. Find the Fourier transform of the distribution u =
t5(X2 - xi} in R2.
Exercise 7.6.4. Find a fundamental solution E E !/"(R3) of the differen-
tial operator
Exercise 7.6.5. Find a real number a and u E 2&'(R) such that the
sequence fn(x) = n a sin (nx 2 ), n = 1,2, ... has the limit u =1= 0 in 2&'(R) as
n --+ 00.
Section 7.6 393
Exercise 7.6.6. Find a real number a and u E f!fi' (R2) such that the
sequence un{x) = na sin (nxlX2), n = 1,2, ... , has the limit u =F 0 in
f!fi' (R2) as n ~ 00.
Exercise 7.6.7. For which positive real numbers a and which p E [1,00]
is the Fourier transform of fa (X) = (I + x2)-a/2e ix2 in LP?
Exercise 7.6.8. Let p be a polynomial in x E R of degree m > 1 and
real coefficients. Prove that the Fourier transform F of eip(x) is an entire
analytic function, and determine a homogeneous differential equation
of order m - 1 with linear coefficients which it satisfies.
Exercise 7.6.9. Let A be a symmetric n x n matrix with Re A positive
semi-definite and II 1m A II :s:: 1. Prove that if n < Jl < n + 1 and u is
Holder continuous of order Jl in Rn , then e-(AD,D)u is continuous and
with C independent of A and u
sup le-(AD,D)ul :s:: C(Jl- n)-lluI Jl ;

lul Jl = I sup loO:ul + I sup loO:u{x) - oO:u(y)llx - yln- Jl .
io:i,,;n io:i=n xioy
Exercise 7.6.10. Prove that if in addition to the assumptions in the

preceding lemma we know that A{D)ju is Holder continuous of order Jl
for 0 :s:: j :s:: N, then
le-(AD,D)U(X) - I (-(AD,D))ju(x)/j!1 :s:: C(Jl- n)-ll(AD,D)N uIJl/N !.

j<N
Answers and Hints to all the Exercises
Chapter I
1.1. Since f(k)(x) is odd when k is odd we have f(k)(O) = 0 then. By
Theorem 1.2.6 we can choose go E COO(R) with the Taylor expansion
If(2k)(0)f</(2k)!. All derivatives of ft(x) = f(x) - go(x 2) vanish at 0
then. Show that
go(x) + ft (v'x), if x > 0,
g(x) = { .
go (x), If x ::; 0,
has the required properties. Alternatively, prove that if g(x) = f(v'x),
x ~O, then
,
g(n)(x) = 2'-2n 10 (1 - t2)n- f(2n) (tv'x)dt/(n -I)!,
I x> O.
Conclude that g E COO when x ~ 0 and extend g to R.

1.2. Use the preceding exercise.
1.3. Review the solution of the preceding exercises.
1.4. Iterate the result in the preceding exercise.
1.5. Introduce Xl ± X2 as new coordinates and use Exercise 1.3.
1.6. Choose for example bk = _2k and solve the equations (iii) with
n ::; N, k ::; N first, which gives
(N)
ak
= II bj - 1
b·-b
II 1 - bj
b -b·'
0<k<N
- - .
k<j:5.N ] k O:5.j<k k J
Show that laiN) I ::; C2- k2 / 2 and that ak = limN--.oo aiN) exists. (The
procedure is called the Seeley extension; see Seeley [2].)
r:
1.7. Replacing f(x) by sUPt<xf(t) we may assume that f is increasing.
Then take F(x) = f(tx)~(t) dt, x > 0, with 0 ::; qJ E C ((1,2)), o
r qJ(t) dt = 1.
Chapter II 395
Chapter II
2.1. The condition is Re b > -1 - k. Necessity: We must have
I J qJ(x)xb dxl :$; I

k
o
sup IqJUll, qJ E Co((O, 1)).
o
Testing with qJ(x) = qJ(x/<.), qJ E C ((1,2)) shows when <. -+ 0 that
we must have Re b + 1 + k ~ O. In case of equality testing with
qJ(x) ~ IfqJ(2 V x)2 v(ilmb-k l shows that
NI J qJ(x)x bdxl :$; c,
although the left-hand side -+ 00 as N -+ 00. - The sufficiency follows

by taking
U(qJ) = (_I)k .£00 qJ(kl(x)xb+k dx/(b + k) ... (b + 1), qJ E Co(R),
if b =1= -k, ... ,-1; replace xb+k / (b + k) by log x if b = -k < O.

2.2. No. Test as in the preceding exercise!
2.3. The condition is now k(a + 1) + Re b + 1 > o. Modify the solution
of the preceding exercises! Note that the oscillating factor may keep the
order down.
2.4. Choose g with g'(x) increasing so rapidly when x -+ 0 that
f(x)/g'(x) -+ 0 as x -+ 0 and f~ If'(x)/g'(x) Idx < 00, and define
U(qJ) = i'£oo eig(jqJ/g')' dx, qJ E Co (R).
(Verify that f~ Ifld(l/g') < 00.) Note that the amplitude may be very
large if the oscillation is fast!
2.5. The limits are a) nbo, b) 2y'nbo.
2.6. a) 0 b) -2150 c) nbo. (Use that residue calculus gives
f: sin x dx/x = n.)
2.7. -nbo. (Use the preceding exercise.)
2.8. a) For a + k < O. b) For all a.
396 Answers and Hints
2.9. Hint: Prove first that there is a uniform bound for Uj on every
closed ball c K; use Arzela-Ascoli's theorem to extract a convergent
subsequence and identify the limit with u. Conclude that it was not
necessary to take a subsequence.
2.10. a = -!
and u = J2~>52k7t, V = 2v'n~.>\2k+I)7t. (Hint: By the
periodicity it suffices to study u,. when It I < 2n. Taylor expansion of
°
1 - cos t at reduces the first question to a study of (IX - a)t2,. at 0. The
dominating contributions to v occur when cos t = -1, so look at the
Taylor expansion there, which leads to Exercise 2.5 b).)
2.11. Show that every qJ E CO'(R n) can be written in the form
qJ(X) = qJI (x) + ... + qJn(x) + XI ... xnqJ(x)
with CPj E CO'(R n) even in Xj and cP E CO'(Rn).
2.12. ao(w) = nw-3/4/J2, al(w) = 0, a2(w) = nw- I / 4 /v'8, where

Iarg wi < n/2. Hint: It suffices to study Jf w(x)cp(x) dx when cP is
an even test function. Write cp(x) = cp(o) + X2qJ"(0)/2 + qJ(x) and note
that Icp(x)/(x4 +w)1 : : ; Icp (x)/x 4 I which is integrable. Use residue calculus
to handle the other two terms.
2.13. For qJ E CO'(R2) choose a rectangle Q = [-a, a] x [-b, b) containing
the support, and write
j j Xe(xy)cp(x,y)dxdy = II +h +h +h where
h = j hXe (xy)(cp (x, y) - cp(x, 0) - cp(O, y) + cp(O, 0)) dx dy 0, -+
h J' r Xe (xy)(cp (x, 0) - qJ(O, 0)) dx dy 1 (cp(x, O) - qJ(O, 0)) dx/lxl,

= -+
J ~a
h = J' r Xe (xy)(cp (0, y) - qJ(O, 0)) dx dy 1 (qJ(O, y) - qJ(O, 0)) dy /Iyl,
Q
-+
JQ ~d
14 = qJ(O, 0) J1 Q
X(xY/E)dxdY/E = 2cp(0, 0) 1Itl<ab/e
X(t) log (ab/Eltl) dt.
(Verify these claims!) Hence C = 2 and
u=2 j(-log ItI)X(t)dtc5o(x,y)+v,

Chapter II 397
i:
where
V(IP) = (IP(X,O) -IP(O,O» dx/lxl
+ ib b
(IP(O,y) -IP(O,O» dy/lyl + 2log (ab)IP(O,O)
=- fooo (1P~(t, 0) -IP~(-t, 0) + IP~(O, t) -IP~(O, -t» log t dt.

2.14. Ua,e --+ nbo and Ua,eUb,t --+ 2n(b - a)-2bo if a =F b but the limit does
not exist when a = b. (Hint: Look separately at the contributions to
(Ua,eUb,e, IP) when Ix - ayl8l ~ /8, when Ix - b/81 ~ /8, and from the
rest of R.)
2.15. If IP E CO'(R2 \ 0) and lP(x, y) = (r, 0) with polar coordinates then
12111 00
(ft> IP) = t o o sin(tlr2 - 11)(r, O)r drdO
= 12111
00
t sin(tlsl)(v's+1, O)dsdO /2.
o -1
Integration by parts for s :§O gives the limit J<1>(1, O)dO, that is, ft --+ ds,
the arc length measure on the unit circle. If we take instead IP =
lP(x2 + i) where IP E CO'«-l, 1», then
(ft> IP) = 2nt 10 1

sin(t(l - r2»IP(r2)r dr = nt fo
1
sin(ts)lP(l - s)ds
= n[- cos(ts)lP(l - s)]A - n 10 1

cos(ts)IP' (1 - s)ds = -nIP (0) cos t + 0(1).
If IP(O) =F
~'(R2).
° the oscillation as t --+ 00 shows that there is no limit in
2.16. Already the inner limit in a) is equal to 0. If IP E COO and IP(O) = 0,

we can write lP(x) = XIP(x), IP E COO, and see that the second limit of
(ft,t, IP) is also 0. If IP = 1 in (-r, r) then Cauchy's integral formula gives
lim (fr,&> IP) = -2ni+

e-++O
(J + J )e-itxlP(x) dx/x --+ -2ni, t --+ +00.
Ixl>r Ixl=r,Irnx<O
The limit b) is therefore -2nib o; the order of the limits is essential!

Section 3.1
3.1.1. u' = f means that u(cp') = -f(cp), cp E CO'(J). If X E CO'(I) is
fixed with f X dx = 1 then every cp E CO'(I) has a unique decomposition
cp = ax - cp' with a E C and cp E CO'(I); we have a = f cpdx, and
u(cp) = f(cp) + Ca where C = u(X). Now defining u in this way we see at
once for every C that u E g&'(I) and that u' = f.
3.1.2. It is clear that the order is at most k + 1; the solution to the
preceding exercise proves the opposite inequality: if u' is of order k,
then u is of order k - 1.
3.1.3. Use the answer to Exercise 3.1.1 if k > o.
3.1.4. Immediate consequence of the preceding exercise.
3.1.5. Yes; if f(x) = 0, x ~ 0, and f(x) = exp(ie l / x ), x > 0, then we can
take u = ix2 f'.
3.1.6. Elaboration of the preceding answers gives the condition ak ;:::: 1.
3.1.7. The limit is f~(cp(x) - cp(-x» dx/x; the order is 1.
3.1.S. Show first that ft(x) = tf (tx) converges to vp(l/x); then it follows
that gt ~ d(vp(l/x»/dx.
3.1.9. Convergence requires at + 2a2 = 0, which allows two partial
integrations giving the desired result when at +4a2 = 2, that is, al =-2
and a2 = 1. - Generalize this example to higher derivatives of vp(l/x)!
3.1.10. On one hand, log (x + iO) is the limit in Ltoc of log (x + ie), so the
derivative is the limit 1/(x + iO) of l/(x + ie), as e ~ +0. On the other
hand, differentiation of the two terms in the given definition shows that
the derivative is also equal to vp(l/x) - ni~o.
3.1.11. The limit is 2(coshx - 1)/x2 - 2ni~o. Hint: Note that the
expression can be simplified to (eX - l)(x + ie)-2 + (e- X-1)(x - ie)-2.
3.1.12. By induction: f(n) (Ixl)(sgn x)n + 2 L2j+2,;;n f(2j+l)(0)~Jn-2j-2).
3.1.13. By induction:
Ixlf(n) (x) + nsgn xf(n-l) (x) + 2 ~-2(k + l)f(k)(0)~Jn-k-2).
3.1.14. a) Modify the answer to Exercise 3.1.1. b) u = _~y+l) /U + 1).
3.1.15. Use the preceding exercise. When g = 0 we get u = LCaj~y)
where Caj are constants and a is a zero of F of order> j.
3.1.16. Hint: Use test functions cp«x - a)/e) if a is a zero of infinite
order.
Section 3.1 399
3.1.17. (_I)i h~k-i) k !/(k - j)! if j ::;; k and 0 otherwise.
3.1.1S. L~=o(-I)i (;)fU)(O)h~k-j).
3.1.19. a) x+ + C; b) !x! + C; c) (eX -1)H(x) + C;

d) H(x) + C; e) log Ixl + C.
3.1.20. a) U = -ho + C, + C2 H
b) U = c, vp(l/x) + C 2 h
c) u = Cf(x) where f(x) = exp{l/x) when x < 0 and f(x) = 0 when
x~O.
d) u = h~ + C,x::;:' + C2x=' + C3hO
e) u = ho - h~ + Ch,
f) u = -ho + C,h, + C2h_,
g) u=LcA
h) u = H(x) -2(x-l)H(x-l) + C'x+ C2
i) u = H(x -1) + H(x + 1) + C,H(x) + C2
j) u = H(x) exp«(x+ 1)-2-1)/2). (As in c) we have u = 0 in (-00, -1)
and (-1,0), and Exercise 3.1.17 shows that there is no contribution with
support at -1.)
k) u = Cf(x) where f(x) = exp (1/3x 3 ), x < 0 and f(x) = 0 when
x ~ 0 (compare with c».
1) u = LcA,/2+j1t + '[,dj H(n/2 -Ix - U+ l)nl}.
3.1.21. u(rp) = '[,j<N Cj(ohrp(', 0», rp E Co(Rn), where Cj E ~'(Rn-').
3.1.22. u = (Co + C'o, + C202 + C3(oi + O~))ho. Hint: Prove first that
xfu = 0, x~u = 0, deduce that u = L Cl1.ol1.ho with (x, ::;; 2, (X2 ::;; 2, and
use Exercise 3.1.17.
3.1.23. The condition is that rp = rp = 0 implies orp/ox, =1= O. Hint: Note
that 0 = rpo(rpu)/ox, = rporp/ox,u = 0, hence rpu = 0 and orp/ox,u = 0
to prove sufficiency. Choose a Dirac measure to prove necessity.
3.1.24. fN = 0, the order of UN is N -1, and sing SUPPUN = {O}. Hint:
Use that (t + i.x2)-N is a continuous function of t ~ 0 with values in ~'.
Since u, is locally integrable and OUN /OX2 = -iNuN+!. the order of UN
is at most N - 1. Use suitable test functions to show that it cannot be
N-2.
3.1.25. The limit is Lg(x)=o2nih x /Jdetg"(x). Hint: Only zeros of g

are important. If g(O) = 0 then g(x) = Q(x) + O(lxI 3) where Q is a
positive semidefinite quadratic form. If it is positive definite we have
g(x) ~ Q(x)/2 for small Ixi. and taking x/.JB as new variable in the
integral (Us, rp) then gives the answer. To prove necessity consider
Im(u.,cp) with rp ~ 0, conclude that m{x E K;!e::;; g(x)::;; e}::;; Ce for
every compact set K and that if K is a ball with g i= 0 in the interior

then det g"(x) i= 0 at zeros of g on oK. Hence the necessity unless g == 0,
which is obviously excluded.
3.1.26. Hint: This is obvious near any point where f' and g' are
linearly independent or f i= 0 or g i= o. Changing coordinates shows
that it suffices to prove that the limit exists in a neighborhood of 0
when f(x) = Xl, g(O) = 0, and dg is proportional to dx, at 0, hence
olg(O) i= O. Integrate by parts in SXI>O qJ(x)/(g(x) + ill)-' dx and note
that Ilog(g + ill)1 ::5: C + Ilog Igil which is locally integrable both in Rn
and when x, = 0, which allows one to use the dominated convergence
theorem.
3.1.27. u(z) = 0 if Izl ::5: 1, u(z) = log Izl if Izl > 1. (Motivate the
distribution convergence near the unit circle carefully!) Au is the arc
length measure on the unit circle.
3.1.28. This is minus the arc length measure on the unit circle.
3.1.29. The order is 0 in cases a) and c); it is 1 in case b). (Use polar
coordinates in case c) to show that the distribution is -1 times the arc
length measure on the unit circle.)
3.1.30. We must have a > b. If f(x) = x a sin (x-b) then (ou/oy, qJ) =
- S;' qJ(X,j(x» dx, qJ E CO', so ou/oy is always a measure. When
x > 0 then ou/ox = -f'(x)ou/oy has infinite measure near the origin
unless a > b. When this condition is fulfilled verify that ou/ ox =
-f'(x)ou/oy + b(x)H(-y).
3.1.31. Either f vanishes identically or else all zeros of f have finite
order. The sufficiency is close to the one dimensional case (Exercise
3.1.15). To prove necessity test with functions of the form qJ(x)qJ«y-a)/Il)
where a is an endpoint of an interval where f > 0, and estimate the
order of a as a zero of f by means of the order of u as a distribution!
3.1.32. qJ(t) = C.jt - t where C is an arbitrary constant. - Direct
calculation shows that we have a solution outside the curve x = qJ(t).
Taking the jumps into account we obtain qJ(t) = -t or the differential
equation qJ'(t) + (1 - qJ(t)/t)/2 = o.
3.1.33. The limit is 2ni«x _1)-3 H(-x) + b~ (x) + !b, (x». Hint: Take the
Taylor expansion at z = 1.
Section 3.2 401
Section 3.2
3.2.1. Verify (i),(ii),(iii) by direct computation. Then (v),(vi) follow if
Re a > -1. The first part of (iv) is clear. Z (0) consists of functions
constant on each half axis. Now u E Z(-l) means that xu = Co and
u = Covp(l/x)+ ClbO which proves that dimZ(-l) = 2. The other
statements follow now from the first part of (iv). (See also Exercises 2.1
and 3.1.2.)
3.2.2. Argue as in the preceding exercise. To prove (iv) note that since
xtZ(O,k) c Z(O,k - 1) and xtZ(O,l) = {O}, the dimensions of the
spaces show that the inclusion is an equality, which implies the statement
on tZ(O,k). Since txZ(-l,k) c Z(-1,k-1) we conclude from (v) for
lower k that the dimension of Z(-l,k) is at most 2k. We have
xZ(-l, k) ::J xtZ(O, k) = Z(O, k - 1).
The inclusion is strict, for if w,,(x) = (log Ix!)" E Z (0, k + 1) then xw~ =
kWk-l is in Z(0,k)\Z(0,k-1) although w~EZ(-l,k) since we have
(xt + l)ktwk = t(xt)kWk = O. Hence dimZ(-l,k) = 2k and the
other statements follow.
3.2.3. The dimension is 2m. Hint: Write Lajr j = amII(r -Av)k" and use
the preceding exercise.
3.2.4. Hint: Elaborate the solution of Exercise 2.1.
3.2.5. The order of u is 1, the order of v is 2, and the degree of
homogeneity is -2 for both u and v. That the order is at most 2 is
obvious. That the order of v is not 1 follows using test functions of
the form <Pl(x/b)<P2(y/e) where e --.. 0 first and then b --.. O. To prove
that the order of u is 1, note that the integrand can be estimated by
sup l<p'l min (x, y)/xy.
Section 3.3
3.3.1. p. = 2n L mjb zj where Zj are the zeros and poles, with multiplicity
mj and -mj respectively. Hint: log III is harmonic except at the zeros
and poles, and in a neighborhood of such a point Zj we can write
J(z) = (z -Zj)m1g(z) where g is analytic and g(Zj) =1= O.
3.3.2. l/C = 4n(a11a22 -ar2)! with the square root analytic outside R_.
First change the coordinates to make the coefficient a12 vanish. If a11
1
and a22 are then positive, taking Xja1 as new coordinates reduces to the
Euclidean case. The general formula follows by analytic continuation.

Write down the statement explicitly and do the argument in detail!
3.3.3. E{x) = Ixl 4 - n /«4 - n)(4 - 2n)c n ) if n =1= 4; E(x) = -(log r)/{4c4) if
n = 4; here Cn is the area of sn-t. Note that the Laplacian in RH acts
on functions of r = Ixl as iJ; + (n - 1)r- 1iJr • Consider singularities at 0
carefully!
3.3.4. du = a2 u-4nbo. Use the expression for d in the preceding answer
when x =1= 0, and use the Taylor expansion and the known fundamental
solution of d to examine the singularity at O.
3.3.5. du = -u. (Special case of preceding exercise!)
3.3.6. (C+eialxl + C_e-ialxl)/Ixl where C+ + C_ = -1/{4n). When a is real
we can take - cos(alxl) / (4nlxl); when 1m a > 0 the fundamental solution
_eialxl / (4nlxl) is often preferred because of its decrease at infinity.
3.3.7. J = n(sgn{A(+O))-sgn(A(-O))bo. Hint: Note that Jt(y) = 1/{t+iy)
is bounded in ~'1 (R) when t =1= 0, depends continuously on t and has
limits as t --+ ±O with J + - J - = 2nbo. The inner integral is (fA(x), <p{x, .).
- Note that if A is smooth and does not change sign, we get a solution
of a homogeneous differential equation with singular support reduced to
a point.
3.3.8. J = 2n(-iJ/iJz)N- 1bo/(N - 1)! where iJ/iJz = !(OjiJx - iiJ/iJy).
Hint: By Taylor's formula we can write
<p(z) = I ajk zjzk + <p(z)

j+k<N
where ajk = (iJ/iJz)j{iJ/iJz)k<p(O)/j!k! and 1<p{z)1 S;; Clzl N Llal=N sup liJa<pl.
Show that <p may be replaced by <p when integrating over an annulus
{Z;e < Izl < R}, and that (iJUN/iJX + iiJUN/iJy,<P) = 2naN-l,O.
3.3.9. Near a zero a of J of order m we can write J{z) = g(z)m where
g(a) = 0, g'(a) =1= O. With the new variables g(z) we can use the preceding
exercise.
3.3.10. E(x) = il7 x7i - 1H{xj)/(rxj - 1)!'
3.3.11. C = c/2; a change of variables reduces to the preceding exercise.
3.3.12. F(z) = L~ zj /U!)2, a Bessel function. We have a fundamental
solution if F(O) = 1 and zF"(z) + F'{z) = F(z), and this determines the
coefficients of the power series.
3.3.13. v is the measure v(<p) = 2 J<p(lxl, x) dx/lxl supported by the light
cone boundary, and w = 8nbo, which means that v /8n is a fundamental
Section 4.1 403
solution. To prove this first note that if e = t 2 - Ix1 2, then 0 U (e) =

(4eU'(e))' +4U'(e). Letting Ue ...... H we can calculate v(cp) as the limit of
(0 Ue, cp) after taking e as a new variable instead of t in a neighborhood
of a point with t = Ixl containing supp cp. This gives the assertion on v
in R4 \ 0, which is enough since the degree of homogeneity is -2 > -4.
In the same way it follows that w = 0 in R4 \ 0, so w = Cb by the
homogeneity. One obtains C = 8n by testing with a function of t only.
- The tools introduced in Section 6.1 simplify such arguments since the
change of variables is built into the theory.
Section 4.1
4.1.1. If a = -N then the limit is (-I)N((N - 1)!/(2N _ 1)!)b~2N-l).
Hint: Note that U e is orthogonal to even test functions. For an odd
test function cp(t) = tcp(t2), cp E Co'
(cf. Exercise 1.2) we have ue(cp) =
(cp, X~(t - b)), b = 62 . (Motivate by analytic continuation from the easy
case where Re a > 0.) The right-hand side is a convolution, hence a
continuous function of b. Use that X~ = bt-
1) if a = -N.
4.1.2. Hint: Prove this first when U E Coo by using Green's formula and
the fact that ,1 log Izl = 2nbo. Note the special case where U = log If I, f
analytic.
4.1.3. The measure is equal to
2-/x /(x2 + a)bo(y) + 2-/y2/(a - y2)bo(x)

2
where x 2 + a > 0 in the first term and y2 < a in the second one. Hint:
Since the preceding exercise shows that no point carries a positive mass
it suffices to prove this at points where f (z) =1= 0, f' (z) =1= 0, taking J f (z)
as new coordinates. (This will become easier to do in Section 6.1.)
4.1.4. The Laplacian is 2n- 1 sin (n/n)x~-n)/nbo(Y). Hint: fn is continuous
and ofn/oz = z(1-n)/n/2n in C\ iL; use Theorem 3.1.12.
Section 4.2
4.2.1. The convolutions are a) tn- 1H(t)/(n -I)! b) e-ttn- 1H(t)/(n -I)!
4.2.2. The convolution is H(k) = b~k-l).
4.2.3. U a= ~ bka .
4.2.4. The convolutions are 0 and 1. Note that the convolution need not
be associative unless all factors except one have compact support.
4.2.5. (c5 h * u - u)/h, qJ) = (u, (c5-h * qJ - qJ)/h) -+ (u, qJ'), if qJ E Co(R).
4.2.6. The convolution is X~+J1+1, which follows by analytic continuation
from the special case in Section 3.4.
4.2.7. u = X:;:2-.l. * 1 by the preceding exercise since X:;:l = c5o.
4.2.8. The convolution is the function x ~ 2nab/lxl when la-bl :::;; Ixl :::;;
a + band 0 otherwise. The singular support consists of the spheres with
radius a + band la - bl and center at O. To verify this consider first two
continuous functions u(lxl) and v(lxl). The convolution is a continuous
function of Ixl which can be determined by using a radial test function
qJ(x) = qJ(lxI2). Note that Irco + r' w'I 2 = r2 + r'2 + 2rr' (w, w') if wand
w' are unit vectors, and use the fact, known to Archimedes, that the
surface measure of {w; Iwl = 1, (w,w') < t} is 2n(1 + t) if -1 < t < 1.
Section 4.3
4.3.1. If 0 :::;; qJ E Co, then (u * v, qJ) = (u,15 * qJ) where 15 * qJ(x) > 0 if
x = y - z for some y with qJ(Y) > 0 and some z E suppv; if x E suppu,
that is, y E supp u + supp v, it follows that (u * v, qJ) > O.
4.3.2. No, we have for example c50* 1 = O.
4.3.3. Hint: A convex set K is contained in an affine hyperplane with
normal ~ if and only if R 3 t ~ H (t~) is linear, where H is the
supporting function of K. Now apply the theorem of supports.
4.3.4. supp 1 is the square when P (0) =1= 0; when P (0) = 0 it is the
boundary with the interior of the sides parallel to the Xj axis removed
if P(iJ) is divisible by OJ, and it is empty when P = O. Generalize to an
arbitrary polygon!
Section 4.4
4.4.1. I(z) = z + v'z2 -1 =1= 0 outside [-1,1] so Jl. = 0 there; log I/(z)1 is
continuous and
(iJ/ox-,iiJ/oy) log III =1'/1 = 1/Jz 2_1

Section 4.4 405
also in the sense of distribution theory. The distribution boundary values

at x±iO are =Fi/Vl-x2 , hence Jl = iu(x)t5(y) where u(x) = -2i/Vl-x2
if Ixl < 1, u(x) = 0 if Ixl > 1, that is, Jl(<p) = 2J~1 <p(x,0)dx/Vl-x2.
Since v = log If I - Jl ... E is harmonic,
f(z) = 2z(1 + O(z-2», Jl'" E(z) = log Izl J

dJl/2n + O(l/lzl), z --+ 00,
where JdJl = 2n, the maximum principle gives v = log 2 identically, so

Jl'" E(z) = log If(z)/21·
4.4.2. Taking <p = 1 near 0 shows that we must have Li ajj = 0, and
this implies that the limit exists and that
F(<p) = 1
Ixl<R
f(x)(<p(x) - <p(0» dx
if Ixl < R when x E supp<p. Direct computation gives ~f(x) = 0; hence

~F has support at 0 and homogeneity -5, so there is a symmetric
matrix (bjk ) such that
L
~F = bj k O/ht5 o.
We have F = E ... ~F since the difference is harmonic and --+ 0 at 00. If
<p is a function of Ixl equal to Ixl2 near 0 then 0 = (F,~<p) = (~F,<p) =
2 L bjj and we obtain -4nF = 3 L bjkxjxklxl-5, so bjk = -4najk/3 and
~F = -4n L ajk Oj Okt5 0/ 3.
4.4.3. h = V - E ... ~V is harmonic, if E(x) = -1/(4nlxl), and if V --+ 0

at 00 then h --+ 0 there so h = O. This gives the claim, and V == 1 is a
counterexample showing that the hypothesis is essential.
4.4.4. Set F(x) = <p(x)xk+1 H(x)/(k + I)! where <p E Coo(R) is equal to 1
near O. Then F E ctand F(k+2) = 150 + g with g E COO. Hence
d2k+4 (u'" F ... F)/dx 2k+4 = u'" (150 + g) ... (150 + g) = u + 2u'" g + u'" g'" g
where all terms except u are already known to be in Coo.
4.4.5. Express u in terms of fk by convolution with a fundamental
solution.
4.4.6. Use a fundamental solution of P(d/dx) to prove the sufficiency.
4.4.7. a = l/yJ, I = 1/135. Hint: By the preceding exercise a is
determined by (x 2, X - t5 a - t5-a ) = O. Looking successively at sign
changes of u"', u", u' gives u ;;::: 0; we have I = (x 4 /4!, X- t5a - t5-a ). -
The result is the lowest order case of Gauss integration.
4.4.8. Elaborate the similar exercise above on an ordinary differential

equation. - Theorem 7.3.2 gives a general result.
4.4.9. To prove sufficiency examine u = E * f where E(z) = (2n)-llog Izl
is the usual fundamental solution of~. Expand E(z - 0 in a power
series in , when Izl > Rand "I < R where R > sUPcesuppf "I and
conclude that u(z) = 0 when Izl > R.
Section 5.1
5.1.1. Immediate consequence of the definitions.

5.1.2. Try U = I.f'v-22kvei2'x and a similar definition of v; test with
cp(x, y)e- i(x+ y )2'.
5.1.3. Take Uj = If' A~v~}ei}.2HjX where Av = 2v!. The idea is that

I
A~_l ~ AJ for large v which makes one amplitude factor dominate in
any term in the product. (To prove that Uj is not of order N - 1 note
I I
2 .eiJ. 2,+jx were a measure dll then J2n e-iJ. 2Hj x dll = 2nA 2 .
that if "L A~~ r 0 r ~~
which is absurd.)
Section 5.2
5.2.1. The composition cp 1-+ cp 0 f.

5.2.2. o(H(y - x) - a(x,y))/oy.
5.2.3. This is obvious by Fubini's theorem if p = 1 or p = 00. For
1 < p < 00 write IK(x,y)cp(Y)1 = IK(x,y)11-1/p(IK(x,yW/Plcp(y)1) and use
HOlder's inequality.
Section 6.1
6.1.1. cos x = a when x = 2kn ± arccos a with integer k, and then we

have sin x = ±,J1 - a2 • The answer is therefore
00
(1 - a2 )-! Z)c5 2kn+arccosa + c52kn-arccosa).

-00
Section 6.1 407
6.1.2. U = b'(xt} ® (1/x2Ix2i) + (l/xIlxIl) ® b'(X2). Hint: Calculate

v = bO(XlX2) first and note that OIV = X2U, 02V = XlU.
6.1.3. The limit is !(b(1,I) + b(-l,l). (Take x 2 - y2 and y - 1 as local
coordinates.)
6.1.4. U = du / vlf'1 21g'1 2- (f', g')2 where du is the Euclidean surface
measure on L = {x;f(x) = g(x) = O}. Hint: Assume coordinates
labelled so that Yl = f (x), Y2 = g(x), Yj = xb j > 2, is a local coordinate
system at a chosen point xO E L. Then
(U,t:p) = f t:p(x(O,O,y'))IDx/Dyldy', y' = (Y3,···,Yn); t:p E Cg";
ifsupp t:p is close to xo. Here IDy /Dxl = lof /OXlOg/OX2-of/ox2og/0Xll.

At a point where the tangent plane is dXl = dX2 = 0, this is equal to
vlf'1 2 1g'1 2 - (f',g,)2 since of /OXj = og/OXj = 0, j > 2, so we have the
asserted density at such a point. Orthogonal invariance proves that it is
true everywhere.
6.1.5. If Lx = {y; f (y) = g(x - y) = O}, then
u(x) = 1Ex
t:p(y)t:p(x - y) du/\/lf'(Y)1 2Ig'(X - Y)12 - (f'(y),g'(x - y»2,
where du is the surface measure of Lx. (Use the preceding example and
compare with Exercise 4.2.8.)
6.1.6. The condition is that f(x) = 0 implies f'(x) =f. O. (Consider
1m (u B , t:p) with t:p ~ 0.) The limits are then rv± where v± = (t ± iO)-I,
so U+ - u_ = -2niLf(x)=O bx/If'(x)l·
6.1.7. The first assertion follows since t ~ (t + iB)-1 --+ (t + iO)-1 in fi}'
as B --+ +0 and since x ~ x 2k - s2k does not have 0 as a critical value
when s > O. Examination of (fs, t:p) for even test functions t:p shows
after Taylor expansion that the stated formula holds with Uj equal to a
constant times ba2j ), of order 2j and support {O}, when j < k, whereas
Uk is a constant times (d/dx)2k-1 vp (l/x), thus of order 2k and support
equal to R.
Section 6.2
6.2.1. Ek = !n(l-n)/24-k X:+(I-n)/2(A)/kl for t > 0, where A = t 2 _lxI 2 ; Ek
is extended as a homogeneous distribution of degree 2k + 1 - n. - Note
that Ek is a constant times the characteristic function of the forward

light cone if n is odd and k = (n - 1)/2. (Cf. Exercise 3.3.13.)
6.2.2. Fa = ~ akEk with the notation in the preceding exercise. The
sum converges, for the terms with k + (1 - n)/2 > 0 are continuous and
I
00
(A) = 4- kAk /(k!r(k + 1 + (1 - n)/2))

o
converges to an entire analytic function. We have Fa = 0 when t < Ixi
and Fa = 4n(1-n)/2<1>(aA)A(1-n)/2 when t > lxi, and the singularities at
the light cone are described by Lk«n-l)/2 akEk. (Cf. Exercise 3.3.12.)
6.2.3. F = e-huHL; bjxj F-a, a = L~ bJ - b6 + c, with the notation in the

preceding exercise.
Section 7.1
7.1.1. The condition is m ;;::: n. Hint: Compare with exercise 3.1.5. Note
that a function may be in g' although the absolute value is very large.
7.1.2. Hint: Choose a sequence Xj E M with IXjl > j and set u =
L IXjlm+lbxj" Prove that u E g' and derive a contradiction if u is of
order m in the unit ball by looking at 11P12 * u where IP E g and I~I < 4
if ~ E supp QJ.
7.1.3. Hint: Write u(x) = LIIXI:sm XIXUIX(X) with Urx E L2.
7.1.4. Hint: Choose X E CO'(R n) equal to 1 in K and use that IP = XIP

implies (2n)nQJ = X* QJ,
7.1.5. Hint: Reduce to m = 0 and apply the preceding exercise.

7.1.6. supp u must be compact, for the equation is equivalent to (1-j)u =
i
0, and 1 - i= 0 outside a compact set. Conversely, we can always take
for f the inverse Fourier transform of a function in CO' equal to 1 on
supp uif this is a compact set.
7.1.7. Use the answer to the preceding exercise.
7.1.8. Hint: Reduce to the case ~ = 0 by passing to ue-i(x,~). Choose
X E g so that X has support in the unit ball and X(O) i= O. Then
Section 7.1 409
U,s = X,s '" U =1= 0, if X,s (x) = X(!5x)· Now take CPj(x) = CjXl/j(X + Xj) where
ICjl = l/sup IU1/jl and Xj is chosen so that IU"'cpj(O)-11 < 1/j. Conclude
using the preceding exercise that Iu'" CPj(x) - 11 < (1 + Clxl}/j. (The
result is due to Beuding.)
7.1.9. a) and d) P not a constant =1= 0 b) When P has a real zero c)
and e) Only when P = O.
7.1.10. f = O. Since / is continuous and ~ 0 at 00, the equation
/(1- /) = 0 implies / == O.
7.1.11. U = //(1 - j) is in [/ precisely when the denominator never
vanishes. - The statement is also valid with [/ replaced by L 1, but the
sufficiency of the condition is a much harder theorem of Wiener then.
7.1.12. Hint: Use that (u '" v, cp) = (u ® v, ) where (e, '1) = cp(e + '1),
cP E CO'. Replace by X for a suitable X E CCX:> which is 1 in the first
quadrant and vanishes when lei + I'll > 2(1 + Ie + '11}. Generalize to
higher dimensions!
7.1.13. The limit is -i. Hint: We may assume that u is real valued; then
u(-e) = u(e) so we may take e > O. Using (3.1.13) we obtain
use dominated convergence to get the result. Note that u is continuous

but u ~ L 1 if a :::;; 1.
7.1.14. Z(-I-a,k). Use this to simplify the answer to Exercise 3.2.2!
7.1.15. That fF4 = I is a consequence of the Fourier inversion formula.
Let Pk, k = 0, 1,2,3, be the interpolation polynomials defined by Pk(-.) =
(-.4 _ 1)/(4i3k (r - ik» and show that the decomposition holds precisely
when Uk = Pk(fF)u. Show that fF Lv = iLvfF and solve the differential
equation Lvu = f E [/ explicitly. The kernel of Lv is the set of functions
u E [/ such that u(x)~~/2 is independent of xv.
7.1.16. Hint: Approximate L tjb xj by functions cP E CO' to prove that
(i) ==> (ii), and approximate integrals by Riemann sums to prove the
converse. Note that (ii) with N = 2 implies that K(x) = K(-x) and that
IK(x)1 :::;; K(O), so K E [/'. (iii) ==> (i) since
(K '" cP, cp) = (K, cP '" fp) = (fl,I«$1 2), fp(x) = cp(-x).
(i),(ii) ==> (iii), for (K, Icp12) 20, cP E [I' implies (K,X) 20 for all X E Co
with X 2 0 (approximate X by the square of (X + ee- 1xI2 )! E [1'). Hence
K is a positive measure. If CPb(X) = cp(xN)fJ-n , cP E Co' f cpdx = 1, then
so K has finite mass and equality holds. - K is called positive definite

and the result is Bochner's theorem.
7.1.17. Hint: Choose CPb as in the preceding answer. Show that (i)
implies that Kb = K * CPb * (Pb = fib where J1-b is a positive measure with
total mass Kb(O) = O(fJ-N ) if K is of order N in a neighborhood of
O. Show that J1-bl$(e')1 2 = J1-el$(fJW and conclude that J1-b = 1$(fJ')1 2 J1-
where J1- is a measure with fl~l<i/b dJ1-(~) = O(fJ-N ), which gives (ii) with
N replaced by N + 1. The converse is straightforward. - The result is
due to L. Schwartz.
7.1.18. The primitive functions are of course O(x), hence in [1", and
i~u = j, which means that u = (i~)-lj outside the origin, where there
u
may be a pultiple ot:.. the Dirac measu!.e. Thus supp = supp j or
{O} U suppf. Choose K E COO so that i~K(~) = 1 + $(~) where $1 = 0
and $ E Co. Since KUl is integrable for j > 0, it follows that x j K(x)
is continuous and bounded for every j > 0, and since K' - fJ = cP E [I'
it follows that K is bounded at the origin, hence that K EL I . Since
v = K * I is bounded and v' = I + cp * I = I this proves the assertion.
7.1.19. (2n)nfJ(I (by Fourier's inversion formula).
7.1.20. a) n(fJ; - fJ~i) b) ni(2fJ o- fJ 2- fJ~2)/2 c) n times the char-
acteristic function of (-1,1) (use Fourier's inversion formula!) d)
2n(2i)-k~ (J)fJk-2j(-I)j. e) -(~ - iO)-2 f) 2nfJ o + 2sin~/~ g)
-2ivp 1/~ h) vp 1/(~ + 1) - vp 1/(~ -1) i) ne-1el (residue calculus)
j) n sgn ~e-Iel Ii k) 2nifJ o+ni sgn ~e-Iel 1) vp ne-1el li~ (differentiate the
function!) m) -ni sgn ~e-IWv'2 cos(UV2) (residue calculus).
7.1.21. Both integrals are n12. (Use i) and j) in the preceding exercise.)
7.1.22. -2n(fJ~ + fJ o) - 8(~ cos ~ - sin ~)/~3. (Hint: Consider x f-7 x 2 -1
first and then the difference.)
7.1.23. _nis-ie-islel. (Hint: Determine the Fouriertransform of x f-7
(x 2 - S2 + ie)-1 by residue calculus and let e ~ +0.)

7.1.24. jt(~) = -ni(H(~ - 1) + H(~ + 1) - 2H(~ + t)), fR I/t(x)1 2 dx =
n max (1, 21tl - 1).
Section 7.1 411
7.1.25. The limit is -iei1xl /2. Hint: Prove that
where r is the real axis with a neighborhood of +1 (of -1) replaced by

a half circle in the lower (upper) half plane.
7.1.26. The solutions are u(x) = Ce-lxI2/2 with a constant C. The
solutions in !/' are characterized by the fact that el~12/2u is a homoge-
neous distribution of degree 0, so u is the convolution of e-lxI2/2 and an
arbitrary distribution homogeneous of degree -no
7.1.27. a) -1/(~I-iO) ® 1/(~2-iO) b) -in2sgn~2e-led-le21
c) 2nit5o(~d ® exp (-~V4n) d) iffn~le-iedU2.
7.1.28. The jump is -2ni. Note that f(x) = x/(x 2 + 1) + g(x) where g is
integrable, so J(~) - n sgn ~e-I~I /i = g(~) is continuous.
7.1.29. J
= 2nao(jo + g where g E C k- 1 on each closed half axis and g(j)
has the jump 2ni-j- 1aj+l at 0 when 0 ::;; j ::;; k - 1.
7.1.30. g(+O) - g(-O) = -ni(a+ +a_). Hint: By Exercise 7.1.28 it suffices
to study the Fourier transform of f(x) = H(x - 1)/x.
7.1.31. a must be an even integer. Hint: Solve the differential equation
obtained by Fourier transformation, examine the solution at infinity,
and use the preceding exercise.
7.1.32. If f(z) = L:anz n then the Fourier transform is 2nL:a n(jn. In the
special case we have an = -2- lnl /3.
7.1.33. ~ f--+ 1~la-l y'n2 1- a r((1 - a)/2)/r(a/2). Note that the Fourier
transform has to be even and homogeneous of degree a-I, hence
~ f--+ c/~Ia-l, and C can be determined using a Gaussian as test function.
!
Note the special case a = where the value of the constant follows
from Fourier's inversion formula.
7.1.34. el'e~
V 2n/I~I·
7.1.35. Ja = cafn-a where Ca = nn/22n- ar((n - a)/2)/r(a/2). (Argue as

in the case n = 1 in Exercise 7.1.33.)
7.1.36. Cn = -2n n/ 2 /r(n/2). Hint: Use the preceding exercise with a -+ n
or the known fundamental solution of the Laplacian.
7.1.37. ~ f--+ 2nB(~)-! where B(~l, ~2) = A(~2, -~d. Hint: Reduce to the
Euclidean form by a linear transformation.
7.1.38. u(x, y) = 2nu(y, -x), which generalizes the preceding exercise.

Hint: Express u(qJ) = u(c$) in terms of polar coordinates and do the
radial integration using Fourier's inversion formula.
7.1.39. he) = 2n2(el + e~ - 2iele2 + 2e~)-!. Hint: Discuss the Fourier
transform of x t-+ II A(x) first when A is a positive definite quadratic
form and use analytic continuation to pass to forms with positive definite
real part.
7.1.40. 2nl(ie -,,). (Use that we have essentially a fundamental solution
of the Cauchy-Riemann operator.)
7.1.41. e t-+ 8n(1 + leI 2)-2. Hint: If f E Co(R) then the Fourier transform
ofR3 '3 Xt-+ f(lxl) is a function of lei given by
1 00
o
r2f(r) dr 11001=1
e-ir(oo,e) dw = 1
0
00
r2f(r) dr 11
-1
e-irlels2n ds
= ~~ fooo rf(r) sin (riel) dr.

(See the answer to Exercise 4.2.8. There is an analogue of this formula
in every dimension. It involves a Bessel function which is less elementary
for even dimensions.)
7.1.42. u(e) = 4nlel- 1 sin lei. (Use the answer to the preceding exercise.)
7.1.43. u(e) = -ncos lei. Hint: By the second exercise above the Fourier
transform is
(rc5'(r 2 - 1),4n sin (riel) Ilel = 2n(c5' (t - 1), sin (Vile I) lin
7.1.44. UN(e) = 2d((N -1)!)-lle2IN-~ exp(ef/4e2) when e2 < 0 and 0

when e2 ~ O. (Take the Fourier transform in X2 first.)
7.1.45. UN(e,,,) = ni-N22- N(e - i,,)N-l(e + i,,)-I/(N -1)1. (Use the
differential equation established.)
7.1.46. -4nLajkejekl(3IeI 2). (Use the calculation of !iF in the earlier
exercise.)
Section 7.2
7.2.1. Bl (x) = - Ln#l e2ninx 12nin, hence by differentiation 1 - L c5n =
- Ln#l e2ninx , that is, L c5n = L e2ninx , which means that the Fourier
transform of L c5 2nn is L c5n.
Section 7.2 413
7.2.2. f" + a2f = 2asin(an)I~oob(2k+I)7t. Calculation of the Fourier

coefficients gives
00
f(x) = Lasin(an)/n(-lteinx /(a 2 - n2);

-00
a term with a = n should be read as einx /2. Thus f" + a2f =

a sin (an)/nI(-I)n einx, so I(-I)nb n has 2nIb(2k+I)7t as Fourier trans-
form, which again proves Poisson's summation formula. Show on the
other hand how Poisson's summation formula allows one to obtain the
Fourier coefficients without any calculation!
7.2.3. S(x) = 4(ncosh(x -n)/sinhn -1) when 0 ::;; x ::;; 2n. Hint:
S(x) - S" (x) = I~ cos nx = n ~ b2kn - 4 by Poisson's summation
formula. Determine S from this differential equation and the fact that
S is even and periodic.
7.2.4. a) 2ni(bo + 2I~(-I)nb2n); b) -8nI~(-I)nnb2n - 2nbo. Hint:
First expand tan (x + ie) in a power series in e2ix-2e; for b) use that
(tanz)2 = d(tanz)/dz-1.
7.2.5. Note that g(x) = I(f*tp)(x+2nn) converges absolutely and locally
uniformly, and is in Coo with period 2n. The Fourier coefficients are
j(k)fp(k)/2n which gives the first statement. If X E Cij(-n, n) is equal
to 1 near 0, then (H X)' = bo + H X', hence f = f' * (H X) - f * (H X') which
proves the absolute and locally uniform ,fonvergence of If(2nn - x)
when f,f' ELI. When also f" ELI then f(~)(1 + ~2) is bounded, so the
absolute convergence of Ij(n) is trivial. Letting tp - t bo in the usual
way we obtain the second statement.
J
7.2.6. z(e) = niz-Ieizl~l; the sum is -nz- I cot (nz) by Poisson's summa-
tion formula.
7.2.7. The Fourier transform is ~ ~ ne-I~I (cf. Exercise 7.1.20 i)). By
Poisson's summation formula it follows that
L e/(1 + e2n2) = n L e-27t1n/el = n(1 + 2e-27t / e + O(e-47t/e)).
The limit is therefore 2n. - Note the extremely good approximation to
the integral given by the Riemann sums.
7.2.8. fp(~) = max (0, n(1 - 1~1/2)); by Poisson's summation formula
Itp(x + nn) = fp(O)/n = 1.
7.2.9. Note that the partial sums of the series are bounded by sup lui
and that they converge uniformly on compact sets, hence in Y'. The
approximation procedure will be used in the following two exercises.
7.2.10. Hint: Use the preceding exercise to reduce the proof to the case
where u is a trigonometrical polynomial. (A trigonometrical polynomial
L~v ajei).jx of degree v has at most 2v zeros ei).x, and the values can
be prescribed at 2v such points.) Note that the result implies that
-u'(x) + Asin (Ax) has the same sign as sin (Ax) if u(x) = cos (Ax) and
deduce that lu'f 1.12 + lul 2 :s:; sup lul 2 for every u E LOO(R) with suppu c
[-A, A]. (Bernstein's inequality.)
7.2.11. u is bounded by Example 7.1.18. Consider first the case where
u is periodic with period 2nN 1.1, and extend to the non-periodic case
using Exercise 7.2.9.
Section 7.3
7.3.1. u must be a linear combination of (ja and (j~ for some a E R.
Hint: The sufficiency of this condition follows from Exercise 4.3.3, for
example. To prove the necessity note that if u(c) = 0 for some c E C
then we can take v(O = , - c and w= u(O/(' - c), so w must be a
u
Dirac measure and u has the stated form. If has no zero at all then
log u(O is an entire function with real part :s:; C(1 + 11m" + log (1 + I"))
which implies that u(O = Ce- iaC for some a E R, so u = C(ja.
7.3.2. Hint: Assume first that P = 00. By the Paley-Wiener-Schwartz
theorem we have an analytic extension u(z) with
The Phragmen-Lindelof theorem applied to w 1-+ u(x + wy)e-H(-y) 1m w

when 1m w > 0 proves that the absolute value is :s:; Ilullv", proving the
claim when p = 00. If p < 00, put q = pI(p - 1) and apply the result
already proved to u * v where v E U has compact support, IlvlIL" :s:; 1.
7.3.3. K = 4n- 2 L~oo(-I)k(2k - 1)-2(j(2k-l)1t/2. (Compute the Fourier
coefficients of K noting thatK" consists of Dirac measures.) The
u
formula u' = K * u is clear if E CO'«-I, 1)) and follows in general by
regularization since K has finite mass. (It is true when suppu c [-1,1]
since we can apply it to u(tx) and let t --+ 1 - 0.) The total mass of K is
K(I)li = 1, so sup lu'l :s:; sup lui, which is Bernstein's inequality. Equality
occurs when u(x) = ae ix + be-h.
7.3.4. Reduce to A = 1 by a dilation. Then choose K(~) = i~ sina+cosa,
-1 :s:; ~ < 1, so that (~) = K(~)e-icxe is periodic with period 2.
Develop in Fourier series and show that K is a measure with support
Section 7.4 415
{kn - rx;k E Z} with total mass e-iIX K(l) = 1, and argue as in the
preceding exercise.
7.3.5. Hint: Apply Theorems 7.3.6, 7.3.2 and Lemma 7.3.7. If an entire
function h satisfies the equation P(D)h = 0 then P(D)hk = 0 for every
k if hk is the sum of the terms of order k in the Taylor expansion of h,
for all the polynomials P(D)hk have different degrees of homogeneity. -
The result is true if (and only if) every irreducible factor of P vanishes
at the origin, but the proof is harder then (see Malgrange [1]).
Section 7.4
7.4.1. Parseval's formula gives f Ij(~ + il1)1 2 d~ = 2n f;' If(x)fe2x~ dx

which proves the claim. Every analytic function F in the lower half
plane such that IIF(· + il1)11L2 is bounded when 11 < 0 is the Laplace
transform of a function in L2((0, (0)). Use Theorem 7.4.2 or prove
directly that if f~ is the inverse Fourier transform of F(· + il1) then
e-X~f~(x) is independent of 11.
7.4.2. This exercise and the preceding one together are the Paley-Wiener
theorem in the strict sense.
7.4.3. 2F(z) = 1/(eltz / 2 + e- Itz / 2) and G(~) = (ee + e-e)-I. The Fourier
transform of x I--)- F(x + iy) is ~ I--)- e-eYG(~) is, so G(~)(ee + e-e) = 1.
Use residue calculus to get F.
7.4.4. Hint: Express f Ij(~ + il1)f d~ by Parseval's formula.
7.4.5. The first part is a reformulation of the preceding exercise. When
U E [f' one obtains the functions with IU(z)l(l + IzI)Ne-<l>(z,z) bounded
for some N < 0, and when u E [f one obtains those for which this is
true for all N; then we have
which implies U E [f. In the case of fI" we obtain u by duality.

7.4.6. Make an optimal choice of y when estimating U (0. The last
statement follows by using Fourier's inversion formula or Liouville's
theorem. - A distribution theory with the usual properties apart from
the notion of support can be based on the test functions in this exercise.
(See Gelfand and Silov [1,2].)
Section 7.6
7.6.1. ~ ~ n3/2e1Ci/4 exp (-i(~l + ~i - ~~)/4).
7.6.2. The condition is 1 :::;; p :::;; 2. (Cf. Exercise 7.1.13 for a dual
special case.) - The estimate of It follows from the fact that It is the
convolution of I and ce-il~12/4tt-n/2 where c is a constant. By Parseval's
formula we cpnclude that JIftiP dx :::;; Ct n(l-p/2) if p > 2 whereas the
L 1 norm of It is independent of t over any compact set. If p > 2 it
follows by the closed graph theorem and Baire's theorem that one can
find g E LP such that g is not a measure on any open subset of Rn.
On the other hand, if I :::;; p :::;; 2 then the Fourier transform of LP is
contained in 0, where l/p + l/q = 1.
7.6.3. u(~) = ei~U4~2 vn/i~2 with the square root in the right half plane.
Hint: In the integral
we can change the order of integration if we just integrate for - T :::;; t :::;;
T. The integral of the exponential with respect to t can be estimated
by CI~21-! independently of T. When ~2 =1= 0 it converges to the result
given, and we can use the dominated convergence theorem to justify the
result.
7.6.4. E = (4nlx21)-! eix~/4x2-1Ci(sgnx2)/4 /(2n(xl + iX2)) is one solution.
Hint: Take the Fourier transform with respect to X3 and choose a
fundamental solution for the resulting operator, which is essentially the
Cauchy-Riemann operator, so that it is Gaussian in 6.
!
7.6.5. a = and u = ooVn/2. Hint: Express (e inx2 ,qJ) in terms of tp
when qJ E [II is real valued.
7.6.6. a = 2 and u = 2ni';2oo/iJx1oX2. Hint: This time we obtain for real
valued qJ E [II
JJ einxlx2qJ(x) dx = (2nn)-1 JJe-in-I~I~2tp(~) d~

= (2n/n)qJ(O) + (2ni/n 2)olo2qJ(O) + O(n-3).
The first term drops out when we take the imaginary part.
7.6.7. ap> 1. Hint: Prove first that
Section 7.6 417
where r consists of the lines 1m Z = ±!, ± Re z > 1 and the interval

between (1,~) and (-1,-~). Conclude that fa is continuous and that
7.6.S. p'(-D)F(O = ~F(~), which implies that F is an entire analytic

function. Alternatively, prove first that
F (~) = Ir ei(P(x)-xe) dx,
where r consists of an interval from c+i to ci on the imaginary axis and

the half axes {x + c+i; x §O} where a m(±l)m-lc ± > 0 if am is the leading
coefficient in p. (Compare with the discussion of the Airy function.)
7.6.9. Hint: Reduce first to the case where A = iC with C real and of
diagonal form with diagonal elements ±1. Decompose u by the partition
of unity in Theorem 1.4.6 and integrate by parts n times in each term
as in the proof of Lemma 7.6.4. Regularize the terms obtained where u
is differentiated n times and integrate by parts once more in the smooth
part.
7.6.10. Hint: Apply Taylor's formula to e-t(AD,D)u as a function of t.
Bibliography
Agmon, S.: [1] The coerciveness problem for integro-differential forms. 1. Analyse
Math. 6, 183-223 (1958).
[2] Spectral properties of Schrodinger operators. Actes Congr. Int. Math. Nice 2,
679-683 (1970).
[3] Spectral properties of Schrodinger operators and scattering theory. Ann. Scuola
Norm. Sup. Pisa (4) 2, 151-218 (1975).
[4] Unicite et convexite dans les problemes differentiels. Sem. Math. Sup~ No 13,
Les Presses de I'Univ. de Montreal, 1966.
t5] Lectures on elliptic boundary value problems. van Nostrand Mathematical
Studies 2, Princeton, N.J. 1965.
[6] Problemes mixtes pour les equations hyperboliques d'ordre superieur. Coli. Int.
CNRS 117, 13-18, Paris 1962.
[7] Some new results in spectral and scattering theory of differential operators on
JR". Sem. Goulaouic-Schwartz 1978-1979, Exp. II, 1-11.
Agmon, S., A. Douglis and 1. Nirenberg: [1] Estimates near the boundary for solutions
of elliptic partial differential equations satisfying general boundary conditions. I.
Comm. Pure Appl. Math. 12,623-727 (1959); II. Comm. Pure Appl. Math. 17, 35-
92 (1964).
Agmon, S. and 1. Hormander: [1] Asymptotic properties of solutions of differential
equations with simple characteristics. 1. Analyse Math. 30, 1-38 (1976).
Agranovich, M.S.: [1] Partial differential equations with constant coefficients. Uspehi
Mat. Nauk 16:2, 27-94 (1961). (Russian; English translation in Russian Math.
Surveys 16:2,23-90 (1961).)
Ahlfors, 1. and M. Heins: [1] Questions of regularity connected with the Phragmen-
LindelOf principle. Ann. of Math. 50, 341-346 (1949).
Airy, G.B.: [1] On the intensity of light in a neighborhood of a caustic. Trans. Cambr.
Phil. Soc. 6, 379-402 (1838).
Alinhac, S.: [1] Non-unicite du probleme de Cauchy. Ann. of Math. 117,77-108 (1983).
[2] Non-unicite pour des operateurs differentiels a caracteristiques complexes sim-
ples. Ann. Sci. Ecole Norm. Sup. 13, 385-393 (1980).
- [3) Uniqueness and non-uniqueness in the Cauchy problem. Contemporary Math. 27,
1-22 (1984).
Alinhac, S. and M.S. Baouendi: [1] Uniqueness for the characteristic Cauchy problem
and strong unique continuation for higher order partial differential inequalities.
Amer. 1. Math. 102, 179-217 (1980).
Alinhac, S. and C. Zuily: [1] Unicite et non-unicite du probleme de Cauchy pour des
operateurs hyperboliques a caracteristiques doubles. Comm. Partial Differential
Equations 6, 799-828 (1981).
Alsholm, P.K.: [1] Wave operators for long range scattering. Mimeographed report,
Danmarks Tekniske H0jskole 1975.
Alsholm, P.K. and T. Kato: [1] Scattering with long range potentials. In Partial Diff.
Eq., Proc. of Symp. in Pure Math. 23, 393-399. Amer. Math. Soc. Providence, R.1.
1973.
420 Bibliography
Amrein, W.O., Ph.A. Martin and P. Misra: [1] On the asymptotic condition of
scattering theory. Helv. Phys. Acta 43, 313-344 (1970).
Andersson, K.G.: [1] Propagation of analyticity of solutions of partial differential
equations with constant coefficients. Ark. Mat. 8, 277-302 (1971).
Andersson, K.G. and R.B. Melrose: [1] The propagation of singularities along glidings
rays. Invent. Math. 41, 197-232 (1977).
Arnold, V.l: [1] On a characteristic class entering into conditions of quantization.
Funkcional. Anal. i Prilozen. 1, 1-14 (1967) (Russian); also in Functional Anal.
Appl. 1, 1-13 (1967).
Aronszajn, N.: [1] Boundary values of functions with a finite Dirichlet integral. Con-
ference on Partial Differential Equations 1954, University of Kansas, 77-94.
- [2] A unique continuation theorem for solutions of elliptic partial differential
equations or inequalities of second order. J. Math. Pures Appl. 36, 235-249 (1957).
Aronszajn, N., A. Krzywcki and J. Szarski: [1] A unique continuation theorem for
exterior differential forms on Riemannian manifolds. Ark. Mat. 4, 417-453 (1962).
Asgeirsson, L.: [1] Uber eine Mittelwerteigenschaft von Liisungen homogener linearer
partieller Differentialgleichungen 2. Ordnung mit konstanten Koeffizienten. Math.
Ann. 113, 321-346 (1937).
Atiyah, M.F.: [1] Resolution of singularities and division of distributions. Comm. Pure
Appl. Math. 23, 145-150 (1970).
Atiyah, M.F. and R. Bott: [1] The index theorem for manifolds with boundary. Proc.
Symp. on Differential Analysis, 175-186. Oxford 1964.
- [2] A Lefschetz fIxed point formula for elliptic complexes. 1 Ann. of Math. 86, 374-
407 (1967).
Atiyah, M.F., R. Bott and L. Giirding: [1] Lacunas for hyperbolic differential operators
with constant coefficients. 1 Acta Math. 124, 109-189 (1970).
- [2] Lacunas for hyperbolic differential operators with constant coefficients. II. Acta
Math. 131, 145-206 (1973).
Atiyah, M.F., R. Bott and V.K. Patodi: [1] On the heat equation and the index
theorem. Invent. Math. 19,279-330 (1973).
Atiyah, M.F. and I.M. Singer: [1] The index of elliptic operators on compact manifolds
Bull. Amer. Math. Soc. 69, 422-433 (1963).
- [2] The index of elliptic operators. I, III Ann. of Math. 87, 484-530 and 546-604
(1968).
Atkinson, F.V.: [1] The normal solubility of linear equations in normed spaces. Mat.
Sb. 28 (70), 3-14 (1951) (Russian).
Avakumovic, V.G.: [1] Uber die Eigenfunktionen auf geschlossenen Riemannschen
Mannigfaltigkeiten. Math. Z. 65, 327-344 (1956).
Bang, T.: [1] Om quasi-analytiske funktioner. Thesis, Copenhagen 1946, 101 pp.
Baouendi, M.S. and Ch. Goulaouic: [1] Nonanalytic-hypoellipticity for some degen-
erate elliptic operators. Bull. Amer. Math. Soc. 78, 483-486 (1972).
Beals, R.: [1] A general calculus of pseudo-differential operators. Duke Math. 1. 42, 1-
42 (1975).
Beals, R. and C. Fefferman: [1] On local solvability of linear partial differential
equations. Ann. of Math. 97, 482-498 (1973).
- [2] Spatially inhomogeneous pseudo-differential operators 1 Comm. Pure Appl.
Math. 27, 1-24 (1974).
Beckner, W.: [1] Inequalities in Fourier analysis. Ann. of Math. 102, 159-182 (1975).
Berenstein, C.A. and M.A. Dostal: [1] On convolution equations 1 In L'anal. harm.
dans Ie domaine complexe. Springer Lecture Notes in Math. 336,79-94 (1973).
Bernstein, IN.: [1] Modules over a ring of differential operators. An investigation of
the fundamental solutions of equations with constant coefficients. Funkcional. Anal.
i Prilozen. 5:2, 1-16 (1971) (Russian); also in Functional Anal. Appl. 5, 89-101
(1971).
Bibliography 421
Bernstein, LN. and S.I. Gelfand: [1] Meromorphy of the function pl. Funkcional. Anal.
i Priloren. 3: 1, 84-85 (1969) (Russian); also in Functional Anal. Appl. 3, 68-69
(1969).
Bernstein, S.: [1] Sur la nature analytique des solutions des equations aux derivees
partielles du second ordre. Math. Ann. 59, 20-76 (1904).
Beurling, A.: [1] Quasi-analyticity and general distributions. Lectures 4 and 5, Amer.
Math. Soc. Summer Inst. Stanford 1961 (Mimeographed).
- [2] Sur les spectres des fonctions. Anal. Harm. Nancy 1947, Coli. Int. XV, 9-29.
- [3] Analytic continuation across a linear boundary. Acta Math. 128, 153-182 (1972).
Bjorck, G.: [1] Linear partial differential operators and generalized distributions. Ark.
Mat. 6, 351-407 (1966).
Bjork, J.E.: [1] Rings of differential operators. North-Holland Publ. Co. Math. Library
series 21 (1979).
Bochner, S.: [1] Vorlesungen tiber Fouriersche Integrale. Leipzig 1932.
Boman, J.: [1] On the intersection of classes of infinitely differentiable functions. Ark.
Mat. 5, 301-309 (1963).
Bonnesen, T. and W. Fenchel: [1] Theorie der konvexen Korper. Erg. d. Math. u. ihrer
Grenzgeb. 3, Springer Verlag 1934.
Bony, I.M.: [1] Une extension du theoreme de Holmgren sur l'unicite du probleme de
Cauchy. c.R. Acad. Sci. Paris 268,1103-1106 (1969).
[2] Extensions du theoreme de Holmgren. Sem. Goulaouic-Schwartz 1975-1976,
Expose no. XVII.
[3] Equivalence des diverses notions de spectre singulier analytique. Sem. Gou-
laouic-Schwartz 1976-1977, Expose no. III.
Bony, J.M. and P. Schapira: [1] Existence et prolongement des solutions holomorphes
des equations aux derivees partielles. Invent. Math. 17,95-105 (1972).
Borel, E.: [1] Sur quelques points de la theorie des fonctions. Ann. Sci. Ecole Norm.
Sup. 12 (3), 9-55 (1895).
Boutet de Monvel, L.: [1] Comportement d'un operateur pseudo-differentiel sur une
variete it bordo J. Analyse Math. 17,241-304 (1966).
- [2] Boundary problems for pseudo-differential operators. Acta Math. 126, 11-51
(1971).
[3] On the index of Toeplitz operators of several complex variables. Invent. Math.
50, 249-272 (1979).
[4] Hypoelliptic operators with double characteristics and related pseudo-differential
operators. Comm. Pure Appl. Math. 27, 585-639 (1974).
Boutet de Monvel, L., A. Grigis and B. Helffer: [1] Parametrixes d'operateurs pseudo-
differentiels it caracteristiques multiples. Asterisque 34-35,93-121 (1976).
Boutet de Monvel, L. and V. Guillemin: [1] The spectral theory of Toeplitz operators.
Ann. of Math. Studies 99 (1981).
Brezis, H.: [1] On a characterization of flow-invariant sets. Comm. Pure Appl. Math.
23, 261-263 (1970).
Brodda, B.: [1] On uniqueness theorems for differential equations with constant coef-
ficients. Math. Scand. 9, 55-68 (1961).
Browder, F.: [1] Estimates and existence theorems for elliptic boundary value prob-
lems. Proc. Nat. Acad. Sci. 45, 365-372 (1959).
Buslaev, V.S. and V.B. Matveev: [1] Wave operators for the Schrodinger equation with
a slowly decreasing potential. Theor. and Math. Phys. 2, 266-274 (1970). (English
translation.)
Calderon, A.P.: [1] Uniqueness in the Cauchy problem for partial differential equa-
tions. Amer. 1. Math. 80, 16-36 (1958).
[2] Existence and uniqueness theorems for systems of partial differential equations.
Fluid Dynamics and Applied Mathematics (Proc. Symp. Univ. of Maryland 1961),
147-195. New York 1962.
422 Bibliography
- [3] Boundary value problems for elliptic equations. Outlines of the joint Soviet-
American symposium on partial differential equations, 303-304, Novosibirsk 1963.
Calderon, A.P. and R. Vaillancourt: [1] On the boundedness of pseudo-differential
operators. J. Math. Soc. Japan 23, 374-378 (1972).
- [2] A class of bounded pseudo-differential operators. Proc. Nat. Acad. Sci. U.S.A.
69,1185-1187 (1972).
Calderon, A.P. and A. Zygmund: [1] On the existence of certain singular integrals.
Acta Math. 88, 85-139 (1952).
Caratheodory, c.: [1] Variationsrechnung und partielle Differentialgleichungen Erster
Ordnung. Teubner, Berlin, 1935.
Carleman, T.: [1] Sur un probh:me d'unicite pour les systemes d'equations aux derivees
partielles a deux variables independentes. Ark. Mat. Astr. Fys. 26B No 17,1-9 (1939).
[2] L'integrale de Fourier et les questions qui s'y rattachent. Publ. Sci. Inst. Mittag-
Lerner, Uppsala 1944.
[3] Proprietes asymptotiques des fonctions fondamentales des membranes vibrantes.
c.R. Congr. des Math. Scand. Stockholm 1934, 34-44 (Lund 1935).
Catlin, D.: [1] Necessary conditions for subellipticity and hypoellipticity for the
aNeumann problem on pseudoconvex domains. In Recent developments in several
complex variables. Ann. of Math. Studies 100,93-100 (1981).
Cauchy, A.: [1] Memoire sur l'integration des equations lineaires. C.R. Acad. Sci. Paris
8 (1839). In <Euvres IV, 369-426, Gauthier-Villars, Paris 1884.
Cerezo, A., J. Chazarain and A. Piriou: [1] Introduction aux hyperfonctions. Springer
Lecture Notes in Math. 449, 1-53 (1975).
Chaillou, J.: [1] Hyperbolic differential polynomials and their singular perturbations.
D. Reidel Pub!. Co. Dordrecht, Boston, London 1979:
Charazain, J.: [1] Construction de la parametrix du probleme mixte hyperboJique pour
l'equation des ondes. c.R. Acad. Sci. Paris 276, 1213-1215 (1973).
- [2] Formules de Poisson pour les varietes riemanniennes. Invent. Math. 24, 65-82
(1974).
Chazarain, J. and A. Piriou: [1] Introduction a la theorie des equations aux derivees
partielles lineaires. Gauthier-Villars 1981.
Chester, c., B. Friedman and F. Ursell: [1] An extension of the method of steepest
descent. Proc. Cambr. Phil. Soc. 53, 599-611 (1957).
Cohen, P.: [1] The non-uniqueness of the Cauchy problem. O.N.R. Techn. Report 93,
Stanford 1960.
[2] A simple proof of the Denjoy-Carleman theorem. Amer. Math. Monthly 75, 26-
31 (1968).
[3] A simple proof of Tarski's theorem on elementary algebra. Mimeographed
manuscript, Stanford University 1967, 6 pp.
Colin de Verdiere, Y.: [1] Sur Ie spectra des operateurs elliptiques a bicharacteristiques
toutes periodiques. Comment. Math. Helv. 54, 508-522 (1979).
Cook, J.: [1] Convergence to the MlIIller wave matrix. J. Mathematical Physics 36, 82-
87 (1957).
Cordes, H.O.: [1] Ober die eindeutige Bestimmtheit der Losungen elliptischer Differen-
tialgleichungen durch Anfangsvorgaben. Nachr. Akad. Wiss. Gottingen Math.-Phys.
K!. IIa, No. 11,239-258 (1956).
Cotlar, M.: [1] A combinatorial inequality and its application to J} spaces. Rev. Math.
Cuyana 1,41-55 (1955).
Courant, R. and D. Hilbert: [1] Methoden der Mathematischen Physik II. Berlin 1937.
Courant, R. and P.D. Lax: [1] The propagation of discontinuities in wave motion.
Proc. Nat. Acad. Sci. 42, 872-876 (1956).
De Giorgi, E.: [1] Un esempio di non-unicita della soluzione del problema di Cauchy
relativo ad una equazione differenziale lineare a derivate parziali ti tipo parabolico.
Rend. Mat. 14,382-387 (1955).
Bibliography 423
- [2J Solutions analytiques des equations aux derivees partielles a coefficients con-
stants. Sem. Goulaouic-Schwartz 1971-1972, Expose 29.
Deic, V.G., E.L. Korotjaev and D.R. Jafaev: [IJ The theory of potential scattering with
account taken of spatial anisotropy. Zap. Nauen. Sem. Leningrad Otdel. Mat. Inst.
Steklov 73, 35-51 (1977).
Dencker, N.: [lJ On the propagation of singularities for pseudo-differential operators of
principal type. Ark. Mat. 20, 23-60 (1982).
- [2] The Weyl calculus with locally temperate metrics and weights. Ark. Mat. 24, 59-79
(1986).
Dieudonne, J.: [lJ Sur les fonctions continus numeriques definies dans un produit de
deux espaces compacts. c.R. Acad. Sci. Paris 205, 593-595 (1937).
Dieudonne, J. and L. Schwartz: [IJ La dualite dans les espaces (F) et (.!l' F). Ann. Inst.
Fourier (Grenoble) 1, 61-101 (1949).
Dollard, J.D.: [lJ Asymptotic convergence and the Coulomb interaction. J. Math. Phys.
5, 729-738 (1964).
- [2J Quantum mechanical scattering theory for short-range and Coulomb interac-
tions. Rocky Mountain J. Math. 1, 5-88 (1971).
Douglis, A. and L. Nirenberg: [IJ Interior estimates for elliptic systems of partial
differential equations. Comm. Pure Appl. Math. 8,503-538 (1955).
Duistermaat, J.1.: [IJ Oscillatory integrals, Lagrange immersions and unfolding of
singularities. Comm. Pure Appl. Math. 27, 207-281 (1974).
Duistermaat, lJ. and V.W. Guillemin: [IJ The spectrum of positive elliptic operators
and periodic bicharacteristics. Invent. Math. 29, 39-79 (1975).
Duistermaat, J.J. and L. Hormander: [IJ Fourier integral operators II. Acta Math. 128,
183-269 (1972).
Duistermaat, J.1. and J. Sjostrand: [IJ A global construction for pseudo-differential
operators with non-involutive characteristics. Invent. Math. 20, 209-225 (1973).
DuPlessis, N.: [IJ Some theorems about the Riesz fractional integral. Trans. Amer.
Math. Soc. 80, 124-134 (1955).
Egorov, Ju.Y.: [IJ The canonical transformations of pseudo-differential operators. Us-
pehi Mat. Nauk 24:5, 235-236 (1969).
[2J Subelliptic pseudo-differential operators. Dokl. Akad. Nauk SSSR 188, 20-22
(1969); also in Soviet Math. Doklady 10, 1056-1059 (1969).
[3J Subelliptic operators. Uspehi Mat. Nauk 30:2, 57-114 and 30:3, 57-104 (1975);
also in Russian Math. Surveys 30:2, 59-118 and 30:3, 55-105 (1975).
Ehrenpreis, L.: [IJ Solutions of some problems of division I. Amer. J. Math. 76, 883-
903 (1954).
[2J Solutions of some problems of division III. Amer. J. Math. 78, 685-715 (1956).
[3J Solutions of some problems of division IV. Amer. J. Math. 82, 522-588 (1960).
[4J On the theory of kernels of Schwartz. Proc. Amer. Math. Soc. 7, 713-718 (1956).
[5J A fundamental principle for systems of linear differential equations with con-
stant coefficients, and some of its applications. Proc. Intern. Symp. on Linear
Spaces, Jerusalem 1961, 161-174.
[6J Fourier analysis in several complex variables. Wiley-Interscience Publ., New
York, London, Sydney, Toronto 1970.
[7J Analytically uniform spaces and some applications. Trans. Amer. Math. Soc.
101, 52-74 (1961).
[8J Solutions of some problems of division V. Hyperbolic operators. Amer. J. Math.
84, 324-348 (1962).
Enqvist, A: [IJ On fundamental solutions supported by a convex cone. Ark. Mat. 12,
1-40 (1974).
Enss, V.: [lJ Asymptotic completeness for quantum-mechanical potential scattering. I.
Short range potentials. Comm. Math. Phys. 61, 285-291 (1978).
- [2J Geometric methods in spectral and scattering theory of Schrodinger operators.
424 Bibliography
In Rigorous Atomic and Molecular Physics, G. Velo and A. Wightman ed., Plenum,
New York, 198()"'1981 (Proc. Erice School of Mathematical Physics 1980).
Eskin, G.I.: [1] Boundary value problems for elliptic pseudo-differential equations.
Moscow 1973; Amer. Math. Soc. Transl. of Math. Monographs 52, Providence, R.I.
1981.
[2] Parametrix and propagation of singularities for the interior mixed hyperbolic
problem. J. Analyse Math. 32, 17-62 (1977).
[3] General initial-boundary problems for second order hyperbolic equations. In
Sing. in Boundary Value Problems. D. Reidel Publ. Co., Dordrecht, Boston, London
1981, 19-54.
[4] Initial boundary value problem for second order hyperbolic equations with
general boundary conditions I. J. Analyse Math. 40, 43-89 (1981).
Fedosov, B.V.: [1] A direct proof of the formula for the index of an elliptic system in
Euclidean space. Funkcional. Anal. i Prilozen. 4:4, 83-84 (1970) (Russian); also in
Functional Anal. Appl. 4, 339-341 (1970).
Fefferman, C.L.: [1] The uncertainty principle. Bull. Amer. Math. Soc. 9, 129-206
(1983).
Fefferman, C. and D.H. Phong: [1] On positivity of pseudo-differential operators. Proc.
Nat. Acad. Sci. 75, 4673-4674 (1978).
- [2] The uncertainty principle and sharp Garding inequalities. Comm. Pure Appl.
Math. 34, 285-331 (1981).
Fredholm, I.: [1] Sur l'integrale fondamentale d'une equation differentielle elliptique it
coefficients constants. Rend. Circ. Mat. Palermo 25, 346-351 (1908).
Friedlander, F.G.: [1] The wave front set of the solution of a simple initial-boundary
value problem with glancing rays. Math. Proc. Cambridge Phil os. Soc. 79, 145-159 (1976).
Friedlander, F.G. and R.B. Melrose: [1] The wave front set of the solution of a simple
initial-boundary value problem with glancing rays. II. Math. Proc. Cambridge Phi-
los. Soc. 81, 97-120 (1977).
Friedrichs, K.: [1] On differential operators in Hilbert spaces. Amer. J. Math. 61, 523-
544 (1939).
[2] The identity of weak and strong extensions of differential operators. Trans.
Amer. Math. Soc. 55, 132-151 (1944).
[3] On the differentiability of the solutions of linear elliptic differential equations.
Comm. Pure Appl. Math. 6, 299-326 (1953).
[4] On the perturbation of continuous spectra. Comm. Pure Appl. Math. 1, 361-406
(1948).
Friedrichs, K. and H. Lewy: [1] Uber die Eindeutigkeit und das Abhangigkeitsgebiet
der Losungen beim Anfangswertproblem linearer hyperbolischer Differentialglei-
chungen. Math. Ann. 98, 192-204 (1928).
Froman, N. and P.O. Froman: [1] JWKB approximation. Contributions to the theory.
North-Holland Publ. Co. Amsterdam 1965.
Fuglede, B.: [1] A priori inequalities connected with systems of partial differential
equations. Acta Math. 105, 177-195 (1961).
Gabrielov, A.M.: [1] A certain theorem of Hormander. Funkcional. Anal. i Prilozen.
4:2,18-22 (1970) (Russian); also in Functional Anal. Appl. 4,106-109 (1970).
Garding, L.: [1] Linear hyperbolic partial differential equations with constant coef-
ficients. Acta Math. 85, 1-62 (1951).
[2] Dirichlet's problem for linear elliptic partial differential equations. Math. Scand.
1, 55-72 (1953).
[3] Solution directe du probleme de Cauchy pour les equations hyperboliques. Coli.
Int. CNRS, Nancy 1956,71-90.
[4] Transformation de Fourier des distributions homogenes. Bull. Soc. Math. France
89,381-428 (1961).
[5] Local hyperbolicity. Israel 1. Math. 13, 65-81 (1972).
Bibliography 425
- [6] Le probleme de la derivee oblique pour l'equation des ondes. C.R. Acad. Sci.
Paris 285, 773-775 (1977). Rectification c.R. Acad. Sci. Paris 285, 1199 (1978).
- [7] On the asymptotic distribution of the eigenvalues and eigenfunctions of elliptic
differential operators. Math. Scand. 1,237-255 (1953).
Garding, L. and J.L. Lions: [1] Functional analysis. Nuovo Cimento N. 1 del Suppl. al
Vol. (10)14, 9-66 (1959).
Glirding, L. and B. Malgrange: [1] Operateurs differentiels partiellement hypoellip-
tiques et partiellement elliptiques. Math. Scand. 9, 5-21 (1961).
Gask, H.: [1] A proof of Schwartz' kernel theorem. Math. Scand. 8, 327-332 (1960).
Gelfand, I.M. and G.E. Silov: [1] Fourier transforms of rapidly increasing functions
and questions of uniqueness of the solution of Cauchy's problem. Uspehi Mat. Nauk
8:6,3-54 (1953) (Russian); also in Amer. Math. Soc. Transl. (2) 5, 221-274 (1957).
[2] Generalized functions. Volume 1: Properties and operations. Volume 2: Spaces
of fundamental and generalized functions. Academic Press, New York and London
1964, 1968.
Gevrey, M.: [1] Demonstration du theoreme de Picard-Bernstein par la methode des
contours successifs; prolongement analytique. Bull. Sci. Math. 50, 113-128 (1926).
Glaeser, G.: [1] Etude de quelques algebres Tayloriennes. J. Analyse Math. 6, 1-124
(1958).
Godin, P.: [1] Propagation des singularites pour les operateurs pseudo-differentiels de
type principal a partie principal analytique verifiant la condition (P), en dimension
2. C.R. Acad. Sci. Paris 284,1137-1138 (1977).
Gorin, E.A.: [1] Asymptotic properties of polynomials and algebraic functions of
several variables. Uspehi Mat. Nauk 16:1,91-118 (1961) (Russian); also in Russian
Math. Surveys 16:1,93-119 (1961).
Grubb, G.: [1] Boundary problems for systems of partial differential operators of mixed
order. J. Functional Analysis 26,131-165 (1977).
- [2] Problemes aux limites pseudo-differentiels dependant d'un parametre. c.R.
Acad. Sci. Paris 292, 581-583 (1981).
GruSin, V.V.: [1] The extension of smoothness of solutions of differential equations of
principal type. Dokl. Akad. Nauk SSSR 148, 1241-1244 (1963) (Russian); also in
Soviet Math. Doklady 4, 248-252 (1963).
[2] A certain class of hypoelliptic operators. Mat. Sb. 83, 456-473 (1970) (Russian);
also in Math. USSR Sb. 12,458-476 (1970).
Gudmundsdottir, G.: [1] Global properties of differential operators of constant
strength. Ark. Mat. 15, 169-198 (1977).
Guillemin, V.: [1] The Radon transform on Zoll surfaces. Advances in Math. 22, 85-
119 (1976).
[2] Some classical theorems in spectral theory revisited. Seminar on sing. of sol. of
diff. eq.,Princeton University Press, Princeton, N.J., 219-259 (1979).
[3] Some spectral results for the Laplace operator with potential on the n-sphere.
Advances in Math. 27, 273-286 (1978).
Guillemin, V. and O. Schaeffer: [1] Remarks on a paper of O. Ludwig. Bull. Amer.
Math. Soc. 79, 382-385 (1973).
Guillemin, V. and S. Sternberg: [1] Geometrical asymptotics. Amer. Math. Soc. Surveys
14, Providence, R.I. 1977.
Gurevic, 0.1.: [1] Counterexamples to a problem of L. Schwartz. Funkcional. Anal. i
Prilozen. 9:2, 29-35 (1975)(Russian); also in Functional Anal. Appl. 9,116-120 (1975).
Hack, M.N.: [1] On convergence to the Meller wave operators. Nuovo Cimento (10) 13,
231-236 (1959).
Hadamard, J.: [1] Le probleme de Cauchy et les equations aux derivees partielles
lineaires hyperboliques. Paris 1932.
Haefliger, A.: [1] Varietes feuilletees. Ann. Scuola Norm. Sup. Pisa 16, 367-397 (1962).
426 Bibliography
Hanges, N.: [1] Propagation of singularities for a class of operators with double
characteristics. Seminar on singularities of sol. of linear partial diff. eq., Princeton
University Press, Princeton, N.J. 1979, 113-126.
Hardy, G.H. and I.E. Littlewood: [1] Some properties of fractional integrals. (I) Math.
Z. 27, 565-606 (1928); (II) Math. Z. 34,403-439 (1931-32).
Hausdorff, F.: [1] Eine Ausdehnung des Parsevalschen Satzes liber Fourierreihen.
Math. Z. 16, 163-169 (1923).
Hayman, W.K. and P.B. Kennedy: [1] Subharmonic functions I. Academic Press,
London, New York, San Francisco 19"76.
Hedberg, L.I.: [1] On certain convolution inequalities. Proc. Amer. Math. Soc. 36,505-
510 (1972).
Heinz, E.: [1] Uber die Eindeutigkeit beim Cauchyschen Anfangswertproblem einer
elliptischen Differentialgleichung zweiter Ordnung. Nachr. Akad. Wiss. Gottingen
Math.-Phys. Kl. IIa No.1, 1-12 (1955).
Helffer, B.: [1] Addition de variables et applications it la regularite. Ann. Inst. Fourier
(Grenoble) 28:2, 221-231 (1978).
Helffer, B. and J. Nourrigat: [1] Caracterisation des operateurs hypoelliptiques ho-
mogenes invariants it gauche sur un groupe de Lie nilpotent gradue. Comm. Partial
Differential Equations 4:8, 899-958 (1979).
Herglotz, G.: [1] Uber die Integration Iinearer partieller Differentialgleichungen mit
konstanten Koeffizienten I-III. Berichte Sachs. Akad. d. Wiss. 78, 93-126, 287-318
(1926); 80, 69-114 (1928).
Hersh, R.: [1] Boundary conditions for equations of evolution. Arch. Rational Mech.
Anal. 16,243-264 (1964).
- [2] On surface waves with finite and infinite speed of propagation. Arch. Rational
Mech. Anal. 19,308-316 (1965).
Hirzebruch, F.: [1] Neue topologische Methoden in der algebraischen Geometrie.
Springer Verlag, Berlin-Gottingen-Heidelberg 1956.
Hlawka, E.: [1] Uber Integrale auf konvexen Korpern. I. Monatsh. Math. 54, 1-36
(1950).
Holmgren, E.: [1] Uber Systeme von Iinearen partiellen Differentialgleichungen. Ofver-
sigt af Kongl. Vetenskaps-Akad. Forh. 58,91-103 (1901).
- [2] Sur l'extension de la methode d'integration de Riemann. Ark. Mat. Astr. Fys. 1,
No 22, 317-326 (1904).
Hormander, L.: [1] On the theory of general partial differential operators. Acta Math.
94, 161-248 (1955).
[2] Local and global properties of fundamental solutions. Math. Scand. 5, 27-39
(1957).
[3] On the regularity of the solutions of boundary problems. Acta Math. 99, 225-
264 (1958).
[4] On interior regularity of the solutions of partial differential equations. Comm.
Pure Appl. Math. 11, 197-218 (1958).
[5] On the division of distributions by polynomials. Ark. Mat. 3, 555-568 (1958).
[6] Differentiability properties of solutions of systems of differential equations. Ark.
Mat. 3, 527-535 (1958).
[7] Definitions of maximal differential operators. Ark. Mat. 3, 501-504 (1958).
[8] On the uniqueness of the Cauchy problem I, II. Math. Scand. 6,213-225 (1958);
7, 177-190 (1959).
[9] Null solutions of partial differential equations. Arch. Rational Mech. Anal. 4,
255-261 (1960).
[10] Differential operators of principal type. Math. Ann. 140, 124-146 (1960).
[11] Differential equations without solutions. Math. Ann. 140, 169-173 (1960).
[12] Hypoelliptic differential operators. Ann. Inst. Fourier (Grenoble) 11, 477-492
(1961).
Bibliography 427
[13] Estimates for translation invariant operators in 11 spaces. Acta Math. 104, 93-
140 (1960).
[14] On the range of convolution operators. Ann. of Math. 76, 148-170 (1962).
[15] Supports and singular supports of convolutions. Acta Math. 110, 279-302
(1963).
[16] Pseudo-differential operators. Comm. Pure Appl. Math. 18,501-517 (1965).
[17] Pseudo-differential operators and non-elliptic boundary problems. Ann. of
Math. 83, 129-209 (1966).
[18] Pseudo-differential operators and hypoelliptic equations. Amer. Math. Soc.
Symp. on Singular Integrals, 138-183 (1966).
[19] An introduction to complex analysis in several variables. D. van Nostrand
Publ. Co., Princeton, N.J. 1966.
[20] Hypoelliptic second order differential equations. Acta Math. 119, 147-171 (1967).
[21] On the characteristic Cauchy problem. Ann. of Math. 88, 341-370 (1968).
[22] The spectral function of an elliptic operator. Acta Math. 121, 193-218 (1968).
[23] Convolution equations in convex domains. Invent. Math. 4, 306-317 (1968).
[24] On the singularities of solutions of partial differential equations. Comm. Pure
Appl. Math. 23, 329-358 (1970).
[25] Linear differential operators. Actes Congr. Int. Math. Nice 1970, 1,121-133.
[26a] The calculus of Fourier integral operators. Prospects in math. Ann. of Math.
Studies 70, 33-57 (1971).
[26] Fourier integral operators I. Acta Math. 127, 79-183 (1971).
[27] Uniqueness theorems and wave front sets for solutions of linear differential
equations with analytic coefficients. Comm. Pure Appl. Math. 24, 671-704 (1971).
[28] A remark on Holmgren's uniqueness theorem. J. Diff. Geom. 6, 129-134 (1971).
[29] On the existence and the regularity of solutions of linear pseudo-differential
equations. Ens. Math. 17, 99-163 (1971).
[30] On the singularities of solutions of partial differential equations with constant
coefficients. Israel J. Math. 13, 82-105 (1972).
[31] On the existence of real analytic solutions of partial differential equations with
constant coefficients. Invent. Math. 21, 151-182 (1973).
[32] Lower bounds at infinity for solutions of differential equations with constant
coefficients. Israel J. Math. 16,103-116 (1973).
[33] Non-uniqueness for the Cauchy problem. Springer Lecture Notes in Math. 459,
36-72 (1975).
[34] The existence of wave operators in scattering theory. Math. Z. 146, 69-91 (1976).
[35] A class of hypoelliptic pseudo-differential operators with double characteristics.
Math. Ann. 217,165-188 (1975).
[36] The Cauchy problem for differential equations with double characteristics. J.
Analyse Math. 32, 118-196 (1977).
[37] Propagation of singularities and semiglobal existence theorems for (pseudo-)
differential operators of principal type. Ann. of Math. 108,569-609 (1978).
[38] Subelliptic operators. Seminar on sing. of sol. of diff. eq. Princeton University
Press, Princeton, N.J., 127-208 (1979).
[39] The Weyl calculus of pseudo-differential operators. Comm. Pure Appl. Math.
32, 359-443 (1979).
[40] Pseudo-differential operators of principal type. Nato Adv. Study Inst. on Sing.
in Bound. Value Problems. Reidel Publ. Co., Dordrecht, 69-96 (1981).
[41] Uniqueness theorems for second order elliptic differential equations. Comm.
Partial Differential Equations 8, 21-64 (1983).
[42] On the index of pseudo-differential operators. In Elliptische Differentialglei-
chungen Band II, Akademie-Verlag, Berlin 1971, 127-146.
[43] L2 estimates for Fourier integral operators with complex phase. Ark. Mat. 21,
297-313 (1983).
[44] On the subelliptic test estimates. Comm. Pure Appl. Math. 33, 339-363 (1980).
428 Bibliography
Hurwitz, A.: [1] Uber die Nullstellen der Bessel'schen Funktion. Math. Ann. 33, 246-
266 (1889).
lagolnitzer, D.: [1] Microlocal essential support of a distribution and decomposition
theorems - an introduction. In Hyperfunctions and theoretical physics. Springer
Lecture Notes in Math. 449, 121-132 (1975).
Ikebe, T.: [1] Eigenfunction expansions associated with the Schrodinger operator and
their applications to scattering theory. Arch. Rational Mech. Anal. 5, 1-34 (1960).
Ikebe, T. and Y. Saito: [1] Limiting absorption method and absolute continuity for the
Schrodinger operator. 1. Math. Kyoto Univ. 12, 513-542 (1972).
Ivrii, V.Ja: [1] Sufficient conditions for regular and completely regular hyperbolicity.
Trudy Moskov. Mat. Obsc. 33, 3-65 (1975) (Russian); also in Trans. Moscow Math.
Soc. 33, 1-65 (1978).
[2] Wave fronts for solutions of boundary value problems for a class of symmetric
hyperbolic systems. Sibirsk. Mat. Z. 21:4,62-71 (1980) (Russian); also in Sibirian
Math. 1. 21, 527-534 (1980).
[3] On the second term in the spectral asymptotics for the Laplace-Beltrami opera-
tor on a manifold with boundary. Funkcional. Anal. i PriloZen. 14:2, 25-34 (1980)
(Russian); also in Functional Anal. Appl. 14,98-106 (1980).
Ivrii, V.la and V.M. Petkov: [1] Necessary conditions for the correctness of the Cauchy
problem for non-strictly hyperbolic equations. Uspehi Mat. Nauk 29:5, 3-70 (1974)
(Russian); also in Russian Math. Surveys 29:5,1-70 (1974).
Iwasaki, N.: [1] The Cauchy problems for effectively hyperbolic equations (general case).
1. Math. Kyoto Univ. 25, 727-743 (1985).
Jauch, J.M. and 1.1. Zinnes: [1] The asymptotic condition for simple scattering systems.
Nuovo Cimento (10) 11, 553-567 (1959).
Jerison D. and C.E. Kenig: [1] Unique continuation and absence of positive eigenvalues for
Schrodinger operators. Ann. of Math. 121, 463-488 (1985).
John, F.: [1] On linear differential equations with analytic coefficients. Unique con-
tinuation of data. Comm. Pure Appl. Math. 2,209-253 (1949).
[2] Plane waves and spherical means applied to partial differential equations. New
York 1955.
[3] Non-admissible data for differential equations with constant coefficients. Comm.
Pure Appl. Math. 10, 391-398 (1957).
[4] Continuous dependence on data for solutions of partial differential equations
with a prescribed bound. Comm. Pure Appl. Math. 13,551-585 (1960).
[5] Linear partial differential equations with analytic coefficients. Proc. Nat. Acad.
Sci. 29, 98-104 (1943).
Jorgens, K. and J. Weidmann: [1] Zur Existenz der Wellenoperatoren. Math. Z. 131,
141-151 (1973).
Kashiwara, M.: [1] Introduction to the theory of hyperfunctions. In Sem. on micro-
local analysis, Princeton Univ. Press, Princeton, N.J., 1979, 3-38.
Kashiwara, M. and T. Kawai: [1] Microhyperbolic pseudo-differential operators. I. J.
Math. Soc. Japan 27, 359-404 (1975).
Kato, T.: [1] Growth properties of solutions of the reduced wave equation with a
variable coefficient. Comm. Pure Appl. Math. 12, 403-425 (1959).
Keller, J.B.: [1] Corrected Bohr-Sommerfeld quantum conditions for nonseparable
systems. Ann. Physics 4,180-188 (1958).
Kitada, H.: [1] Scattering theory for Schrodinger operators with long-range potentials.
I: Abstract theory. J. Math. Soc. Japan 29, 665-691 (1977). II: Spectral and scatter-
ing theory. J. Math. Soc. Japan 30, 603-632 (1978).
Knapp, A.W. and E.M. Stein: [1] Singular integrals and the principal series. Proc. Nat.
Acad. Sci. U.S.A. 63, 281-284 (1969).
Kohn, J.l.: [1] Harmonic integrals on strongly pseudo-convex manifolds I, II. Ann. of
Math. 78, 112-148 (1963); 79, 450-472 (1964).
Bibliography 429
- [2] Pseudo-differential operators and non-elliptic problems. In Pseudo-differential

operators, CIME conference, Stresa 1968, 157-165. Edizione Cremonese, Roma 1969.
Kohn, J.1. and L. Nirenberg: [1] On the algebra of pseudo-differential operators.
Comm. Pure Appl. Math. 18,269-305 (1965).
- [2] Non-coercive boundary value problems. Comm. Pure Appl. Math. 18, 443-492
(1965).
Kolmogorov, A.N.: [1] Zuflillige Bewegungen. Ann. of Math. 35, 116-117 (1934).
Komatsu, H.: [1] A local version of Bochner's tube theorem. J. Fac. Sci. Tokyo Sect. 1-
A Math. 19,201-214 (1972).
- [2] Boundary values for solutions of elliptic equations. Proc. Int. Conf. Funct. Anal.
ReI. Topics, 107-121. University of Tokyo Press, Tokyo 1970.
Kreiss, H.O.: [1] Initial boundary value problems for hyperbolic systems. Comm. Pure
Appl. Math. 23, 277-298 (1970).
Krzyzailski, M. and 1. Schauder: [1] Quasilineare Differentialgleichungen zweiter Ord-
nung vom hyperbolischen Typus. Gemischte Randwertaufgaben. Studia Math. 6,
162-189 (1936).
Kumano-go, H.: [1] Factorizations and fundamental solutions for differential operators
of elliptic-hyperbolic type. Proe. Japan Acad. 52,480-483 (1976).
Kuroda, S.T.: [1] On the existence and the unitary property of the scattering operator.
Nuovo Cimento (10) 12, 431-454 (1959).
Lascar, B. and R. Lascar: [1] Propagation des singularites pour des equations hyper-
boliques a caracteristiques de multiplicite au plus double et singularites Mas-
loviennes II. J. Analyse Math. 41,1-38 (1982).
Lax, A.: [1] On Cauchy's problem for partial differential equations with multiple
characteristics. Comm. Pure Appl. Math. 9, 135-169 (1956).
Lax, P.D.: [2] On Cauchy'S problem for hyperbolic equations and the differentiability of
solutions of elliptic equations. Comm. Pure Appl. Math. 8, 615-633 (1955).
- [3] Asymptotic solutions of oscillatory initial value problems. Duke Math. 1. 24,
627-646 (1957).
Lax, P.D. and L. Nirenberg: [1] On stability for difference schemes: a sharp form of
Garding's inequality. Comm. Pure Appl. Math. 19,473-492 (1966).
Lebeau, G.: [1] Fonctions harmoniques et spectre singulier. Ann. Sci. Ecole Norm. Sup.
(4) 13,269-291 (1980).
Lelong, P.: [1] Plurisubharmonic functions and positive differential forms. Gordon and
Breach, New York, London, Paris 1969.
- [2] Proprietes metriques des varietes definies par une equation. Ann. Sci. Ecole
Norm. Sup. 67, 22-40 (1950).
Leray, J.: [1] Hyperbolic differential equations. The Institute for Advanced Study,
Princeton, N.J. 1953.
- [2] Uniformisation de la solution du probleme lineaire analytique de Cauchy pres de
la variete qui porte les donnees de Cauchy. Bull. Soe. Math. France 85, 389-429 (1957).
Lerner, N.: [1] Unicite de Cauchy pour des operateurs differentiels faiblement principale-
ment normaux. 1. Math. Pures Appl. 64,1-11 (1985).
Lerner, N. and L. Robbiano: [1] Unicite de Cauchy pour des operateurs de type principal.
1. Analyse Math. 44, 32-66 (1984/85).
Levi, E.E.: [1] Caratterische multiple e problema di Cauchy. Ann. Mat. Pura Appl. (3)
16, 161-201 (1909).
Levinson, N.: [1] Transformation of an analytic function of several variables to a
canonical form. Duke Math. J. 28, 345-353 (1961).
Levitan, B.M.: [1] On the asymptotic behavior of the spectral function of a self-adjoint
differential equation of the second order. Izv. Akad. Nauk SSSR Ser. Mat. 16, 325-
352 (1952).
[2] On the asymptotic behavior of the spectral function and on expansion in
eigenfunctions of a self-adjoint differential equation of second order II. Izv. Akad.
Nauk SSSR Ser. Mat. 19, 33-58 (1955).
430 Bibliography
Lewy, H.: [1] An example of a smooth linear partial differential equation without
solution. Ann. of Math. 66, 155-158 (1957).
- [2] Extension of Huyghen's principle to the ultrahyperbolic equation. Ann. Mat.
Pura App!. (4) 39, 63-64 (1955).
Lions, J.L.: [1] Supports dans la transformation de Laplace. J. Analyse Math. 2,369-380
(1952-53).
Lions, J.L. and E. Magenes: [1] Problemes aux limites non homogenes et applications
I-III. Dunod, Paris, 1968-1970.
Lojasiewicz, S.: [1] Sur Ie probleme de division. Studia Math. 18, 87-136 (1959).
Lopatinski, Ya.B.: [1] On a method of reducing boundary problems for a system of
differential equations of elliptic type to regular integral equations. Ukrain. Mat. Z. 5,
123-151 (1953). Amer. Math. Soc. Transl. (2) 89,149-183 (1970).
Ludwig, D.: [1] Exact and asymptotic solutions of the Cauchy problem. Comm. Pure
Appl. Math. 13,473-508 (1960).
- [2] Uniform asymptotic expansions at a caustic. Comm. Pure Appl. Math. 19, 215-
250 (1966).
Luke, G.: [1] Pseudodifferential operators on Hilbert bundles. J. Differential Equations
12, 566-589 (1972).
Malgrange, B.: [1] Existence et approximation des solutions des equations aux derivees
partieJles et des equations de convolution. Ann. Inst. Fourier (Grenoble) 6, 271-355
(1955-56).
[2] Sur une classe d'operateurs differentiels hypoelliptiques. Bull. Soc. Math. France
85, 283-306 (1957).
[3] Sur la propagation de la regularite des solutions des equations a coefficients
constants. Bull. Math. Soc. Sci. Math. Phys. R.P. Roumanie 3 (53), 433-440 (1959).
[4] Sur les ouverts convexes par rapport a un operateur differentiel. c. R. Acad. Sci.
Paris 254, 614-615 (1962).
[5] Sur les systemes differentiels a coefficients constants. ColI. CNRS 113-122, Paris
1963.
[6] Ideals of differentiable functions. Tata Institute, Bombay, and Oxford University
Press 1966.
Mandelbrojt, S.: [1] Analytic functions and classes of infinitely differentiable functions.
Rice Inst. Pamphlet 29, 1-142 (1942).
- [2] Series adMrentes, regularisations des suites, applications. Coli. Borel, Gauthier-
Villars, Paris 1952.
Martineau, A.: [1] Les hyperfonctions de M. Sato. Sem. Bourbaki 1960-1961, Expose
No 214.
- [2] Le "edge of the wedge theorem" en tMorie des hyperfonctions de Sato. Proc.
Int. Conf. Funct. Anal. ReI. Topics, 95-106. University of Tokyo Press, Tokyo 1970.
Maslov, V.P.: [1] Theory of perturbations and asymptotic methods. Moskov. Gos.
Univ., Moscow 1965 (Russian).
Mather, J.: [1] Stability of COO mappings: I. The division theorem. Ann. of Math. 87,
89-104 (1968).
Melin, A.: [1] Lower bounds for pseudo-differential operators. Ark. Mat. 9, 117-140
(1971).
- [2] Parametrix constructions for right invariant differential operators on nilpotent
groups. Ann. Global Analysis and Geometry 1, 79-130 (1983).
Melin, A. and J. Sjostrand: [1] Fourier integral operators with complex-valued phase
functions. Springer Lecture Notes in Math. 459, 12()-'223 (1974).
[2] Fourier integral operators with complex phase functions and parametrix for an
interior boundary value problem. Comm. Partial Differential Equations 1 :4, 313-
400 (1976).
Melrose, R.B.: [1] Transformation of boundary problems. Acta Math. 147, 149-236
(1981).
Bibliography 431
[2] Equivalence of glancing hypersurfaces. Invent. Math. 37, 165-191 (1976).

[3] Microlocal parametrices for diffractive boundary value problems. Duke Math. 1.
42, 605-635 (1975).
[4] Local Fourier-Airy integral operators. Duke Math. J. 42, 583-604 (1975).
[5] Airy operators. Comm. Partial Differential Equations 3:1,1-76 (1978).
[6] The Cauchy problem for effectively hyperbolic operators. Hokkaido Math. J. To
appear.
- [7] The trace of the wave group. Contemporary Math. 27, 127-167 (1984).
Melrose, R.B. and J. Sjostrand: [1] Singularities of boundary value problems I, II.
Comm. Pure Appl. Math. 31, 593-617 (1978); 35,129-168 (1982).
Mihlin, S.G.: [1] On the multipliers of Fourier integrals. Dokl. Akad. Nauk SSSR 109,
701-703 (1956) (Russian).
Mikusiilski, J.: [1] Une simple demonstration du theoreme de Titchmarsh sur la
convolution. Bull. Acad. Pol. Sci. 7,715-717 (1959).
- [2] The Bochner integral. Birkhauser Verlag, Basel and Stuttgart 1978.
Minakshisundaram, S. and A. Pleijel: [1] Some properties of the eigenfunctions of the
Laplace operator on Riemannian manifolds. Cauad. J. Math. 1,242-256 (1949).
Mizohata, S.: [1] Unicite du prolongement des solutions des equations elliptiques du
quatrieme ordre. Proc. Jap. Acad. 34, 687-692 (1958).
[2] Systemes hyperboliques. J. Math. Soc. Japan 11,205-233 (1959).
[3] Note sur Ie traitement par les operateurs d'integrale singuliere du probleme de
Cauchy. J. Math. Soc. Japan 11, 234-240 (1959).
[4] Solutions nulles et solutions non analytiques. J. Math. Kyoto Univ. 1, 271-302
(1962).
[5] Some remarks on the Cauchy problem. 1. Math. Kyoto Univ. 1, 109-127 (1961).
Meller, C.: [1] General properties of the characteristic matrix in the theory of elemen-
tary particles. I. Kongl. Dansk. Vidensk. Selsk. Mat.-Fys. Medd. 23,2-48 (1945).
Morrey, CB.: [1] The analytic embedding of abstract real-analytic manifolds. Ann. of
Math. 68, 159-201 (1958).
Morrey, CB. and L. Nirenberg: [1] On the analyticity of the solutions of linear elliptic
systems of partial differential equations. Comm. Pure Appl. Math. 10, 271-290
(1957).
Moyer, R.D.: [1] Local solvability in two dimensions: Necessary conditions for the
principle-type case. Mimeographed manuscript, University of Kansas 1978.
Miiller, C: [1] On the behaviour of the solutions of the differential equation LI U
=F(x, U) in the neighborhood of a point. Comm. Pure Appl. Math. 7, 505-515
(1954).
Miinster, M.: [1] On A. Lax's condition of hyperbolicity. Rocky Mountain J. Math. 8,
443-446 (1978).
- [2] On hyperbolic polynomials with constant coefficients. Rocky Mountain J. Math.
8, 653-673 (1978).
von Neumann, J. and E. Wigner: [1] Uber merkwiirdige diskrete Eigenwerte. Phys. Z.
30, 465-467 (1929).
Nirenberg, L.: [1] Remarks on strongly elliptic partial differential equations. Comm.
Pure Appl. Math. 8, 648-675 (1955).
[2] Uniqueness in Cauchy problems for differential equations with constant leading
<;oefficients. Comm. Pure Appl. Math. 10,89-105 (1957).
[3] A proof of the Malgrange preparation theorem. Liverpool singularities I. Sprin-
ger Lecture Notes in Math. 192, 97-105 (1971).
[4] On elliptic partial differential equations. Ann. Scuola Norm. Sup. Pisa (3) 13,
115-162 (1959).
[5] Lectures on linear partial differential equations. Amer. Math. Soc. Regional
Conf. in Math., 17, 1-58 (1972).
432 Bibliography
Nirenberg, L. and F. Treves: [1J Solvability of a first order linear partial differential
equation. Comm. Pure Appl. Math. 16,331-351 (1963).
[2J On local solvability of linear partial differential equations. I. Necessary con-
ditions. II. Sufficient conditions. Correction. Comm. Pure Appl. Math. 23, 1-38 and
459-509 (1970); 24, 279-288 (1971).
Nishitani, T.: [1] Local energy integrals for effectively hyperbolic operators I, II. J. Math.
Kyoto Univ. 24, 623-658 and 659-666 (1984).
Noether, F.: [1J Ober eine Klasse singuliirer Integralgleichungen. Math. Ann. 82, 42-63
(1921).
Olejnik, O.A.: [1J On the Cauchy problem for weakly hyperbolic equations. Comm.
Pure Appl. Math. 23, 569-586 (1970).
Olejnik, O.A. and E.V. Radkevic: [IJ Second order equations with non-negative char-
acteristic form. In Matern. Anal. 1969, ed. R.V. Gamkrelidze, Moscow 1971 (Rus-
sian). English translation Plenum Press, New York-London 1973.
Oshima, T.: [1] On analytic equivalence of glancing hypersurfaces. Sci. Papers College
Gen. Ed. Univ. Tokyo 28, 51-57 (1978).
Palamodov, V.P.: [IJ Linear differential operators with constant coefficients. Moscow
1967 (Russian). English transl. Grundl. d. Math. Wiss. 168, Springer Verlag, New
York, Heidelberg, Berlin 1970.
Paley, R.E.A.C. and N. Wiener: [IJ Fourier transforms in the complex domain. Amer.
Math. Soc. ColI. Publ. XIX, New York 1934.
Pederson, R.: [1] On the unique continuation theorem for certain second and fourth
order elliptic equations. Comm. Pure Appl. Math. 11,67-80 (1958).
- [2J Uniqueness in the Cauchy problem for elliptic equations with double charac-
teristics. Ark. Mat. 6, 535-549 (1966).
Peetre, J.: [1] TMoremes de regularite pour quelques classes d'operateurs differentiels.
Thesis, Lund 1959.
[2] Rectification a l'article "Une caracterisation abstraite des operateurs differentie!s".
Math. Scand. 8,116-120 (1960).
[3] Another approach to elliptic boundary problems. Comm. Pure Appl. Math. 14,
711-731 (1961).
[4J New thoughts on Besov spaces. Duke Univ. Math. Series I. Durham, N.C.
1976.
Persson, J.: [1] The wave operator and P-convexity. Boll. Un. Mat. Ital. (5) 18-B, 591-
604 (1981).
Petrowsky, I.G.: [1J Ober das Cauchysche Problem flir Systeme von partiellen Diffe-
rentialgleichungen. Mat. Sb. 2 (44), 815-870 (1937).
[2J Ober das Cauchysche Problem flir ein System Ii nearer partieller Differentialglei-
chungen im Gebiete der nichtanalytischen Funktionen. Bull. Univ. Moscow Ser. Int.
1, No.7, 1-74 (1938).
[3] Sur l'analyticite des solutions des systemes d'equations differentielles. Mat. Sb. 5
(47), 3-70 (1939).
[4] On the diffusion of waves and the lacunas for hyperbolic equations. Mat. Sb. 17
(59), 289-370 (1945).
[5J Some remarks on my papers on the problem of Cauchy. Mat. Sb. 39 (81), 267-
272 (1956). (Russian.)
Pham The Lai: [1] Meilleures estimations asymptotiques des restes de la fonction
spectrale et des valeurs propres relatifs au laplacien. Math. Scand. 48, 5-31 (1981).
Piccinini, L.C.: [1] Non surjectivity of the Cauchy-Riemann operator on the space of
the analytic functions on Rn. Generalization to the parabolic operators. Bull. Un.
Mat. Ital. (4) 7, 12-28 (1973).
Plis, A.: [1] A smooth linear elliptic differential equation without any solution in a
sphere. Comm. Pure Appl. Math. 14,599-617 (1961).
[2] The problem of uniqueness for the solution of a system of partial differential
equations. Bull. Acad. Pol. Sci. 2, 55-57 (1954).
Bibliography 433
[3] On non-uniqueness in Cauchy problem for an elliptic second order differential

equation. Bull. Acad. Pol. Sci. 11,95-100 (1963).
Poincare, H.: [1] Sur les proprietes du potentiel et les fonctions abeliennes. Acta Math.
22, 89-178 (1899).
Povzner, A.Ya.: [1] On the expansion of arbitrary functions in characteristic functions
of the operator -Au+cu. Mat. Sb. 32 (74), 109-156 (1953).
Radkevic, E.: [1] A priori estimates and hypoelliptic operators with multiple character-
istics. DokI. Akad. Nauk SSSR 187, 274-277 (1969) (Russian); also in Soviet Math.
Doklady 10, 849-853 (1969).
Ralston, J.: [1] Solutions of the wave equation with localized energy. Comm. Pure
Appl. Math. 22, 807-823 (1969).
- [2J Gaussian beams and the propagation of singularities. MAA Studies in Mathe-
, matics 23, 206-248 (1983).
Reed, M. and B. Simon: [1] Methods of modern mathematical physics. III. Scattering
theory. Academic Press 1979.
Rempel, S. and B.-W. Schulze: [1J Index theory of elliptic boundary problems. Aka-
demie-Verlag, Berlin (1982).
de Rham, G.: [1] Varietes differentiables. Hermann, Paris, 1955.
Riesz, F.: [1] Sur I'existence de la derivee des fonctions d'une variable reelle et des
fonctions d'intervalle. Verh. Int. Math. Kongr. ZUrich 1932, I, 258-269.
Riesz, M.: [1J Vintegrale de Riemann-Liouville et Ie probleme de Cauchy. Acta Math.
81, 1-223 (1949).
[2] Sur les maxima des formes bilineaires et sur les fonctionnelles Iineaires. Acta
Math. 49, 465-497 (1926).
[3] Sur les fonctions conjuguees. Math. Z. 27, 218-244 (1928).
[4] Problems related to characteristic surfaces. Proc. Conf. Diff. Eq. Univ. Maryland
1955, 57-71.
Rothschild, L.P.: [1] A criterion for hypoellipticity of operators constructed from
vector fields. Comm. Partial Differential Equations 4:6, 645-699 (1979).
Saito, Y.: [IJ On the asymptotic behavior of the solutions of the Schrodinger equation
(-A +Q(y)_k2) V=F. Osaka J. Math. 14, 11-35 (1977).
- [2] Eigenfunction expansions for the Schrodinger operators with long-range poten-
tials Q(y)=O(lyJ-'), 6>0. Osaka J. Math. 14, 37-53 (1977).
Sakamoto, R.: [1] E-well posedness for hyperbolic mixed problems with constant
coefficients. J. Math. Kyoto Univ. 14,93-118 (1974).
- [2] Mixed problems for hyperbolic equations I. J. Math. Kyoto Univ. 10, 375-401
(1970), II. J. Math. Kyoto Univ. 10,403-417 (1970).
Sato, M.: [1] Theory of hyper functions I. J. Fac. Sci. Univ. Tokyo I, 8,139-193 (1959).
[2] Theory of hyperfunctions II. J. Fac. Sci. Univ. Tokyo I, 8, 387-437 (1960).
[3] Hyperfunctions and partial differential equations. Proc. Int. Conf. on Funct.
Anal. and ReI. Topics, 91-94, Tokyo University Press, Tokyo 1969.
[4] Regularity of hyperfunction solutions of partial differential equations. Actes
Congr. Int. Math. Nice 1970,2,785-794.
Sato, M., T. Kawai and M. Kashiwara: [1] Hyperfunctions and pseudodifferential
equations. Springer Lecture Notes in Math. 287, 265-529 (1973).
Schaefer, H.H.: [1] Topological vector spaces. Springer Verlag, New York, Heidelberg,
Berlin 1970.
Schapira, P.: [1] Hyperfonctions et problemes aux limites elliptiques. Bull. Soc. Math.
France 99,113-141 (1971).
[2] Propagation at the boundary of analytic singularities. Nato Adv. Study Inst. on
Sing. in Bound. Value Problems. Reidel Publ. Co., Dordrecht, 185-212 (1981).
[3] Propagation at the boundary and reflection of analytic singularities of solutions
of linear partial differential equations. Publ. RIMS, Kyoto Univ., 12 Suppl., 441-453
(1977).
434 Bibliography
Schechter, M.: [1] Various types of boundary conditions for elliptic equations. Comm.
Pure App!. Math. 13,407-425 (1960).
- [2] A generalization of the problem of transmission. Ann. Scuola Norm. Sup. Pisa
14, 207-236 (1960).
Schwartz, L.: [1] Theorie des distributions I, II. Hermann, Paris, 1950-51.
[2] Theorie des noyaux. Proc. Int. Congr. Math. Cambridge 1950, I, 220-230.
[3] Sur l'impossibilite de la multiplication des distributions. C. R. Acad. Sci. Paris
239, 847-848 (1954).
[4] Theorie des distributions it valeurs vectorielles I. Ann. Inst. Fourier (Grenoble)
7,1-141 (1957).
[5] Transformation de Laplace des distributions. Comm. Sem. Math. Univ. Lund,
Tome supp!. dedie it Marcel Riesz, 196-206 (1952).
[6] Theorie generale des fonctions moyenne-periodiques. Ann. of Math. 48, 857-929
(1947).
Seeley, R.T.: [1] Singular integrals and boundary problems. Amer. J. Math. 88,781-809
(1966).
[2] Extensions of COO functions defined in a half space. Proc. Amer. Math. Soc. 15,
625-626 (1964).
[3] A sharp asymptotic remainder estimate for the eigenvalues of the Laplacian in a
domain of JR3. Advances in Math. 29, 244-269 (1978).
[4] An estimate near the boundary for the spectral function of the Laplace operator.
Amer. J. Math. 102,869-902 (1980).
[5] Elliptic singular integral equations. Amer Math. Soc. Symp. on Singular Inte-
grals, 308-315 (1966).
Seidenberg, A.: [1] A new decision method for elementary algebra. Ann. of Math. 60,
365-374 (1954).
Shibata, Y.: [1] E-well posedness of mixed initial-boundary value problems with con-
stant coefficients in a quarter space. J. Analyse Math. 37, 32-45 (1980).
Siegel, c.L.: [1] Zu den Beweisen des Vorbereitungssatzes von Weierstrass. In Abhand!.
aus Zahlenth. u. Ana!., 299-306. Plenum Press, New York 1968.
Sjostrand, J.: [1] Singularites analytiques microlocales. Prepublications Universite de
Paris-Sud 82-03.
[2] Analytic singularities of solutions of boundary value problems. Nato Adv. Study
Inst. on Sing. in Bound. Value Prob., Reidel Pub!. Co., Dordrecht, 235-269 (1981).
[3] Parametrices for pseudodifferential operators with multiple characteristics. Ark.
Mat. 12,85-130 (1974).
[4] Propagation of analytic singularities for second order Dirichlet problems I, II,
III. Comm. Partial Differential Equations 5:1, 41-94 (1980),5:2, 187-207 (1980), and
6:5,499-567 (1981).
[5] Operators of principal type with interior boundary conditions. Acta Math. 130,
1-51 (1973).
Sobolev, S.L.: [1] Methode nouvelle it resoudre Ie probleme de Cauchy pour les
equations lineaires hyperboliques normales. Mat. Sb. 1 (43),39-72 (1936).
- [2] Sur un theoreme d'analyse fonctionnelle. Mat. Sb. 4 (46), 471-497 (1938).
(Russian; French summary.) Amer. Math. Soc. Trans!. (2) 34, 39-68 (1963).
Sommerfeld, A.: [1] Optics. Lectures on theoretical physics IV. Academic Press, New
York, 1969.
Stein, E.M.: [1] Singular integrals and differentiability properties of functions. Prince-
ton Univ. Press 1970.
Sternberg, S.: [1] Lectures on differential geometry. Prentice-Hall Inc. Englewood
Cliffs, N .J., 1964.
Stokes, G.B.: [1] On the numerical calculation of a class of definite integrals and
infinite series. Trans. Cambridge Philos. Soc. 9,166-187 (1850).
Bibliography 435
Svensson, L.: [1] Necessary and sufficient conditions for the hyperbolicity of poly-
nomials with hyperbolic principal part. Ark. Mat. 8, 145-162 (1968).
Sweeney, W.J.: [1] The D-Neumann problem. Acta Math. 120,223-277 (1968).
Szego, G.: [1] Beitrage zur Theorie der Toeplitzschen Formen. Math. Z. 6, 167-202
(1920).
Tacklind, S.: [1] Sur les classes quasianalytiques des solutions des equations aux deri-
vees partielles du type parabolique. Nova Acta Soc. Sci. Upsaliensis (4) 10, 1-57 (1936).
Tarski, A.: [1] A decision method for elementary algebra and geometry. Manuscript,
Berkeley, 63 pp. (1951).
Taylor, M.:. [1] Gelfand theory of pseudodifferential operators and hypoelliptic oper-
ators. Trans. Amer. Math. Soc. 153,495-510 (1971).
[2] Grazing rays and reflection of singularities of solutions to wave equations.
Comm. Pure Appl. Math. 29, 1-38 (1976).
[3] Diffraction effects in the scattering of waves. In Sing. in Bound. Value Problems,
271-316. Reidel Publ. Co., Dordrecht 1981.
[4] Pseudodifferentialoperators. Princeton Univ. Press, Princeton, N.J., 1981.
Thorin, 0.: [1] An extension of a convexity theorem due to M. Riesz. Kungl Fys.
Sallsk. Lund. Forh. 8, No 14 (1939).
Titchmarsh, E.C.: [I] The zeros of certain integral functions. Proc. London Math. Soc.
25, 283-302 (1926).
Treves, F.: [1] Solution elementaire d'equations aux derivees partielles dependant d'un
parametre. C. R. Acad Sci. Paris 242, 1250-1252 (1956).
[2] These d'Hormander II. Sem. Bourbaki 135, 2" ed. (Mai 1956).
[3] Relations de domination entre operateurs differentiels. Acta Math. 101, 1-139
(1959).
[4] Operateurs differentiels hypoelliptiques. Ann. Inst. Fourier (Grenoble) 9, 1-73
(1959).
[5] Local solvability in L2 of first order linear PDEs. Amer. J. Math. 92, 369-380
(1970).
[6] Fundamental solutions of linear partial differential equations with constant
coefficients depending on parameters. Amer. J. Math. 84, 561-577 (1962).
[7] Un theoreme sur les equations aux derivees partielles a coefficients constants
dependant de parametres. Bull. Soc. Math. France 90, 473-486 (1962).
[8] A new method of proof of the subelliptic estimates. Comm. Pure Appl. Math.
24,71-115 (1971).
[9] Introduction to pseudodifferential and Fourier integral operators. Volume 1:
Pseudodifferential operators. Volume 2: Fourier integral operators. Plenum Press,
New York and London 1980.
Vauthier, J.: [1] Comportement asymptotique des fonctions entieres de type exponen-
tiel dans (Cn et bornees dans Ie domaine reel. J. Functional Analysis 12, 290-306
(1973).
Vekua, LN.: [1] Systeme von Differentialgleichungen erster Ordnung vom elliptischen
Typus und Randwertaufgaben. Berlin 1956.
Veselic, K. and J. Weidmann: [1] Existenz der Wellenoperatoren flir eine allgemeine
Klasse von Operatoren. Math. Z. 134, 255-274 (1973).
- [2] Asymptotic estimates of wave functions and the existence of wave operators. J.
Functional Analysis. 17,61-77 (1974).
Visik, M.I.: [1] On general boundary problems for elliptic differential equations. Trudy
Moskov. Mat. Obsc. 1, 187-246 (1952) (Russian). Also in Amer. Math. Soc. Trans!.
(2) 24, 107-172 (1963).
Visik, M.1. and G.I. Eskin: [1] Convolution equations in a bounded region. Uspehi
Mat. Nauk 20:3 (123), 89-152 (1965). (Russian.) Also in Russian Math. Surveys
20:3,86-151 (1965).
436 Bibliography
[2] Convolution equations in a bounded region in spaces with weighted norms.

Mat. Sb. 69 (111), 65-110 (1966) (Russian). Also in Amer. Math. Soc. Transl. (2) 67,
33-82 (1968).
[3] Elliptic convolution equations in a bounded region and their applications.
Uspehi Mat. Nauk 22: 1 (133), 15-76 (1967). (Russian.) Also in Russian Math.
Surveys 22: 1, 13-75 (1967).
[4] Convolution equations of variable order. Trudy Moskov. Mat. Obsc. 16, 25-50
(1967). (Russian.) Also in Trans. Moskov. Mat. Soc. 16,27-52 (1967).
[5] Normally solvable problems for elliptic systems of convolution equations. Mat.
Sb. 74 (116),326-356 (1967). (Russian.) Also in Math. USSR-Sb. 3, 303-332 (1967).
van der Waerden, B.L.: [1] Einflihrung in die algebraische Geometrie. Berlin 1939.
- [2] Algebra I-II. 4. Aufl. Springer Verlag, Berlin-Gottingen-Heidelberg 1959.
Wang Rou-hwai and Tsui Chih-yung: [1] Generalized Leray formula on positive complex
Lagrange-Grassmann manifolds. Res. Report, Inst. of Math., Jilin Univ. 8209, 1982.
Warner, F.W.: [1] Foundations of differentiable manifolds and Lie Groups. Scott,
Foresman and Co., Glenview, Ill., London, 1971.
Weinstein, A.: [1] The order and symbol of a distribution. Trans. Amer. Math. Soc.
241, 1-54 (1978).
[2] Asymptotics of eigenvalue clusters for the Laplacian plus a potential. Duke
Math. J. 44, 883-892 (1977).
[3] On Maslov's quantization condition. In Fourier integral operators and partial
differential equations. Springer Lecture Notes in Math. 459, 341-372 (1974).
Weyl, H.: [1] The method of orthogonal projection in potential theory. Duke Math. J.
7,411-444 (1940).
[2] Die Idee der Riemannschen Fliiche. 3. Aufl., Teubner, Stuttgart, 1955.
[3] Uber gewohnliche Differentialgleichungen mit Singularitaten und die zugehori-
gen Entwicklungen willkiirlicher Funktionen. Math. Ann. 68, 220-269 (1910).
[4] Das asymptotische Verteilungsgesetz der Eigenwerte linearer partieIIer Differen-
tialgleichungen (mit einer Anwendung auf die Theorie der Hohlraumstrahlung).
Math. Ann. 71,441-479 (1912).
Whitney, H.: [1] Analytic extensions of differentiable functions defined in closed sets.
Trans. Amer. Math. Soc. 36,63-89 (1934).
Widom, H.: [1] Eigenvalue distribution in certain homogeneous spaces. J. Functional
Analysis 32, 139-147 (1979).
Yamamoto, K.: [1] On the reduction of certain pseudo-differential operators with non-
involution characteristics. J. Differential Equations 26, 435-442 (1977).
Zeilon, N.: [1] Das Fundamentalintegral der aIIgemeinen partiellen linearen Differen-
tialgleichung mit konstanten Koeffizienten. Ark. Mat. Astr. Fys. 6, No 38, 1-32 (1911).
Zemer, M.: [1] Solutions de l'equation des ondes presentant des singularites sur une
droite. C. R. Acad. Sci. Paris 250,2980-2982 (1960).
[2] Solutions singulieres d'equations aux derivees partieIIes. Bull. Soc. Math. France
91,203-226 (1963).
[3] Domaine d'holomorphie des fonctions vi:rifiant une equation aux derivees par-
tieIles. C. R. Acad. Sci. Paris 272, 1646-1648 (1971).
Zuily, C.: [1] Uniqueness and non-uniqueness in the Cauchy problem. Progress in
Math. 33, Birkhauser, Boston, Basel, Stuttgart 1983.
Zygmund, A.: [1] On a theorem of Marcinkiewicz concerning interpolation of oper-
ators.1. Math. Pures Appl. 35, 223-248 (1956).
Index
Adjoint operator 272 Dirac measure 56

Advanced fundamental solution 141 Direct product 126; 127
Airy differential equation 214 Distribution 33
- function 213 - on manifold 144
Analytic functional 326 - density 145
Asgeirsson mean value theorem 183 Dual cone 257
Atlas 143 - - quadratic forms 206
Beta function 86 Edge of the wedge theorem 343; 344

Bicharacteristic (strip) 154; 302 Elliptic operator, polynomial 111; 169;
- curve 302 271
Borel theorem 16 Essential support 322
Exponential solution 185
c1 boundary 59 Feynman fundamental solution 141

Canonical one form 149
Finite part 70
- transformation 155
Fourier transform 160; 164
Carrier of analytic functional 326
- -Laplace transform 165
Cauchy integral formula 63
- problem 141; 349 - inversion formula 161; 164
Cauchy-Kovalevsky theorem 348 Fundamental solution 80
Chain rule 8
Characteristics 152; 271; 350 Gamma function 73; 86
Conormal bundle 149 Gauss-Green formula 60
Convex function 90; 91 Gevrey class 281
- hull 105
Convolution 16; 88; 101 Hamilton vector field 153
Coordinate system, patch 142 Hamilton-Jacobi theory 157
Cotangent bundle 148 Hardy-Littlewood-Sobolev inequality
Critical point 218 117
Cutoff function 25 Heaviside function 56
HOlder continuity 123
Density 145; 148 Holmgren uniqueness theorem 309
Denjoy-Carleman theorem 23 Homogeneous distribution 74
Diagonal 131 Hyperbolic polynomial 320
- map 267 Hyperfunction 335; 337
Differentiable 7 Hypoelliptic operator 11 0
Differential form 148; 150
- operator 13 Inverse function theorem 9
438 Index
Kashiwara-co-Holmgren theorem 364 Regular set 52

Kernel theorem 128 Regularization 17
Kolmogorov equation 210 Retarded fundamental solution 141
Runge theorem 112
Lattice point 28
Leibniz' formula 13 Schwartz kernel theorem 129
Section of vector bundle 147
Sequential continuity 35
Manifold 142; 143
Signature 85
- real analytic 282
Simple layer 136
Measure 38
Singular support 42
Microhyperbolic 318
Slowly varying metric 28
Microlocal analysis 251
Sobolev spaces 240
Multi-index 12
- embedding theorem 123
Multiple layer q6
Stationary phase 215; 218
- points 218
Normal bundle 149 Stieltjes-Vitali theorem 110
of map 263 Subharmonic function 92
- set 300; 301 Support of analytic functional 331
distribution 41
- function 14
Order of distribution 34
- hyperfunction 336
Orientation 150
Supporting function 105
Oscillatory integral 238
Supports, theorem of 107
Symbol 237
Paley-Wiener-Schwartz theorem 181 Symplectic form 152
Parametrix 170 - map 155
Parseval's formula 163
Partition of unity 28
Phase function 236 Tangent bundle 147
Plurisubharmonic function 96 - cone of set 364
Poisson bracket 156 - vector 146
Taylor's formula 12
- summation formula 178
Polydisc 346 Temperate distribution 163
Tensor product 126; 127; 267
Positive distribution 38
Principal part (symbol) 151; 271 Test function 14
Transition matrices 147
- type 275
- value 73 Transpose 112
Transversal intersection 266
Product 55; 267
Proper cone 104; 257 Trivialization of vector bundle 147
- map 104
Pullback of distribution 135; 263 Vector bundle 146
form 149 - field 148
- hyperfunction 345
Wave front set 254; 265; 283; 340

Quasi-analytic class 22
Weak solution 2
- topology of distributions 38
Real analytic function 24; 281 Whitney extension theorem 48
Index of Notation
Spaces of Functions and Distributions
cl 7; 143 tff' 45
Ck 11; 143 tff'k 45
Cy 123; 241 g' 160
Ck0 14 g" 163
Coo
0 14 H(S) 240
CM 22 I I (s) 240
CL 281 L2s 240
Llloc 37; 143 A 326
£2' 34; 144 A'(K) 326
£2' k 34 A (lR")
1 331
£2'F 34 B(X) 335; 337
£2'r 262
Special Symbols
a (multi-index) 12 {j 160; 164

lal (its length) 12 Q 189
a! 13 <u, 44
8j 12 f* (pullback) 134; 149; 263
Dj 160 345
8" 12 f* 148
y" 13 * (convolution) 16; 88; 101
P(x, 8) 13 o (composition) 131; 133
P(D) 182 Char 271; 315
p(")
13 supp 14; 41
8j8z, 8j8z 62 smgsupp 42
f'(x) 7 smgsupPL 282
J<k)(X; YI' ... , Yk) 11 sing SUPPA 282
/::,. 80 WF 254
0 139 WFx 268
11 71; 164 WF; 268
it 108 WE' 268
440 Index of Notation
WFL 283 NJ 263

WFA 283; 340 brU) 342
sgn (signature) 85 w (canonical one form) 149
ch (convex hull) 105 (j (symplectic form) 152
H K (supporting function) 105 ~ 43
TO (dual cone) 257 C (complement)
Ne(F) 300 o (empty set)
N;(F) 301 A '- B (= A nCB)
N(F) 301
Special Distributions
H (Heaviside) 56 (x±iO)" 72
ba 56 X~ 73
PV (principal value) 73 2!;-k 72
x a+ 68; 69 Ai (Airy) 213
x~ 71

The Analysis of Linear Partial Differential Operators

Uploaded by

Copyright:

Available Formats

You might also like

The Analysis of Linear Partial Differential Operators

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Analysis of Linear Partial Differential Operators

Uploaded by

Copyright:

Available Formats

Classics in Mathematics

Lars Hormander The Analysis of Linear Partial

The Analysis of Linear Partial

Reprint of the 2nd Edition 1990

Originally published as Vol. 256 in the series:

Library of Congress Cataloging.in·Publication Data

Springer-Verlag Berlin Heidelberg New York

@ Springer-Verlag BerlinHeidelberg 2003

Printed on acid-free paper 41/3142ck-543210

Lund, October 1989 Lars Hormander

In 1963 my book entitled "Linear partial differential operators" was

Lund, January 1983 Lars Hormander

Chapter I. Test Functions 5

Chapter II. Definition and Basic Properties of Distributions 33

Chapter III. Differentiation and Multiplication by Functions 54

Chapter IV. Convolution 87

4.5. Basic I! Estimates for Convolutions 116

Chapter V. Distributions in Product Spaces 126

Chapter VI. Composition with Smooth Maps 133

Chapter VII. The Fourier Transformation 158

Chapter VIII. Spectral Analysis of Singularities 251

Chapter IX. Hyperfunctions 325

Answers and Hints to All the Exercises. 394

Index of Notation. 439

In differential calculus one encounters immediately the unpleasant fact

(Fv)(~)= S riX~v(x)dx, ~E1R.

It has the important property that

(1) F(Dv)=MFv, F(Mv)= -DFv

with twice continuously differentiable f and g, and they have as

(5) 82uj8x 2 + 8 2uj8y2 =F,

However, (7) defines a continuously differentiable function v even if F

provided that F is continuously differentiable. Again (8) defines a

1.1. A Review of Differential Calculus

valued functions the mean value theorem must be replaced by the

Theorem 1.1.1. If f: I --+ V is differentiable at every point in I, then

Proof Let M>sup{II1'(x+t(y-x))II, O~t~ l} and set

E= {t; O~t~l, Ilf(x+t(y-x))- f(x) II ~Mtlx- yl}.

Hence s = 1 which proves the theorem.

Remark. If f is just continuous in [x, y] and differentiable in the

(1.1.2)' Ilf(y)- f(x)- v(y-x)11 ~ Iy -xl sup 111'(x +t(y-x)) -vII

which is often more useful, particularly with v = 1'(x).

Corollary 1.1.2. Let f be continuous in I and differentiable outside a

IIf(y) - f(x) II = Ilf(y) - f(z)11 ~ Iy-zl sup 1I1'(z+ t(y-z))11

which is o(ly-xl) as y--+x.

Example 1.1.3. If P is a polynomial and f(x)=P(1/x)e-11X, x>O, f(x)

Let U be another Banach space, X an open subset of U. If f is a

Definition 1.1.4. f is called differentiable at XEX if there is an element

By C 1 (X, V) we denote the set of continuously differentiable functions

If f is just differentiable at every point on the line segment [x, y]

for f(x+t(y-x»-Tt(y-x) is differentiable in t on [0,1] with de-

Proof If we apply (1.1.2)" to h with T= f;(x) we obtain when j --+ 00

which proves that f is differentiable at x with f'(x)=g(x). Since g is

Theorem 1.1.6. If f and g are continuous functions in X with values in

Proof (1.1.2)' gives for small Ilyll

If f is a linear map V ---+ V then f is of course differentiable and

T-1(id v+ST- 1 )-1 = LT- 1( -ST-1t

The proof is obvious. From (1.1.3) it follows that hE C l if gE C l and

with 1, gEC I we obtain hECI and another commutative diagram

Proof The necessity is an immediate consequence of the chain rule

Theorem 1.3.1. If uECb then UVECj if vEL~oc' and UVECHk if