Triebel_titelei 29.7.

2008 13:49 Uhr Seite 1

EMS Tracts in Mathematics 7

Triebel_titelei 29.7.2008 13:49 Uhr Seite 2

EMS Tracts in Mathematics

Editorial Board:
Carlos E. Kenig (The University of Chicago, USA)
Andrew Ranicki (The University of Edinburgh, Great Britain)
Michael Röckner (Universität Bielefeld, Germany, and Purdue University, USA)
Vladimir Turaev (Indiana University, Bloomington, USA)
Alexander Varchenko (The University of North Carolina at Chapel Hill, USA)

This series includes advanced texts and monographs covering all fields in pure and applied
mathematics. Tracts will give a reliable introduction and reference to special fields of
current research. The books in the series will in most cases be authored monographs,
although edited volumes may be published if appropriate. They are addressed to graduate
students seeking access to research topics as well as to the experts in the field working
at the frontier of research.

1 Panagiota Daskalopoulos and Carlos E. Kenig, Degenerate Diffusions

2 Karl H. Hofmann and Sidney A. Morris, The Lie Theory of Connected Pro-Lie Groups
3 Ralf Meyer, Local and Analytic Cyclic Homology
4 Gohar Harutyunyan and B.-Wolfgang Schulze, Elliptic Mixed, Transmission and
Singular Crack Problems
5 Gennadiy Feldman, Functional Equations and Characterization Problems on Locally
Compact Abelian Groups
,
6 Erich Novak and Henryk Wozniakowski, Tractability of Multivariate Problems.
Volume I: Linear Information
Triebel_titelei 29.7.2008 13:49 Uhr Seite 3

Hans Triebel

Function Spaces and

Wavelets on Domains
Triebel_titelei 29.7.2008 13:49 Uhr Seite 4

Author:

Hans Triebel
Friedrich-Schiller-Universität Jena
Fakultät für Mathematik und Informatik
Mathematisches Institut
07737 Jena, Germany
E-mail: [email protected]

2000 Mathematics Subject Classification: 46-02; 46E35, 42C40; 42B35, 28A80

Key words: Function spaces, wavelets, bases and frames, extension problem, thick domains,
cellular domains

978-3-03719-019-7

The Swiss National Library lists this publication in The Swiss Book, the Swiss national bibliography,
and the detailed bibliographic data are available on the Internet at http://www.helveticat.ch.

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, re-use of illustrations, recitation, broad–
casting, reproduction on microfilms or in other ways, and storage in data banks. For any kind of use
permission of the copyright owner must be obtained.

© 2008 European Mathematical Society

Contact address:
European Mathematical Society Publishing House
Seminar for Applied Mathematics
ETH-Zentrum FLI C4
CH-8092 Zürich
Switzerland
Phone: +41 (0)44 632 34 36
Email: info @ems-ph.org
Homepage: www.ems-ph.org

Typeset using the authors’ TEX files: I. Zimmermann, Freiburg

Printed in Germany

987654321
Preface

This book may be considered as a continuation of the monographs [T83], [T92], [T06].
Now we are mainly interested in spaces on domains in Rn , related wavelet bases and
wavelet frames, and extension problems. But first we deal in Chapter 1 with the usual
spaces on Rn , periodic spaces on Rn and on the n-torus T n and their wavelet expansions
under natural restrictions for the parameters involved. Spaces on arbitrary domains are
the subject of Chapter 2. The heart of the exposition are the Chapters 3, 4, where
we develop a theory of function spaces on so-called thick domains, including wavelet
expansions and extensions to corresponding spaces on Rn . This will be complemented
in Chapter 5 by spaces on smooth manifolds and smooth domains. Finally we add in
Chapter 6 a discussion about desirable properties of wavelet expansions in function
spaces introducing the notation of Riesz wavelet bases and frames. This chapter deals
also with some related topics, in particular with spaces on cellular domains.
Although we rely mainly on [T06] we repeat basic notation and a few classical
assertions in order to make this text to some extent independently understandable and
usable. More precisely, we have two types of readers in mind:
researchers in the theory of function spaces who are interested in wavelets as
new effective building blocks, and
scientists who wish to use wavelet bases in classical function spaces in diverse
applications.
Here is a guide to where one finds basic definitions and key assertions adapted to the
second type of readers:

1. Classical Sobolev spaces Wpk .Rn /, Sobolev spaces Hps .Rn /, classical Besov
s
spaces Bpq .Rn / and Hölder–Zygmund spaces C s .Rn / on the Euclidean n-spaces
R : Definition 1.1, Remark 1.2, p. 2.
n

2. Wavelets in Rn : Section 1.2.1, p. 13.

3. Wavelet bases for spaces on Rn : Theorem 1.20, p. 15.

4. Spaces on the n-torus T n : Definition 1.27, Remark 1.28, p. 21.

5. Wavelet bases for spaces on T n : Theorem 1.37, p. 26.

6. Spaces on arbitrary domains  in Rn : Definitions 2.1, 5.17, Remark 2.2, pp. 28,
29, 147.

7. u-wavelet systems in domains  in Rn : Definitions 2.4, 6.3, pp. 32, 179.

8. u-Riesz bases and u-Riesz frames: Definition 6.5, Section 6.2.2, pp. 180, 202.
vi Preface

9. Wavelet bases in L2 ./ and Lp ./ in arbitrary domains  in Rn : Theorems 2.33,

2.36, 2.44, pp. 49, 53, 59.
10. Classes of domains  in Rn and their relations: Definitions 3.1, 3.4, 5.40, 6.9,
Proposition 3.8, pp. 70, 72, 75, 168, 182.
11. Wavelet bases in E-thick domains (covering bounded Lipschitz domains) 
in Rn : Theorems 3.13, 3.23, Corollary 3.25, pp. 80, 89, 91.
12. Spaces, frames and bases on manifolds: Definitions 5.1, 5.5, 5.40, Theorems 5.9,
5.37, pp. 133, 135, 136, 164, 168.
13. Frames and bases on domains: Definition 5.25, Theorems 5.27, 5.35, 5.38, 5.51,
6.7, 6.30, 6.32, 6.33, pp. 152, 153, 162, 166, 175, 181, 196, 197, 198.

Formulas are numbered within chapters. Furthermore in each chapter all defini-
tions, theorems, propositions, corollaries and remarks are jointly and consecutively
numbered. Chapter n is divided in sections n:k and subsections n:k:l. But when
quoted we refer simply to Section n:k or Section n:k:l instead of Section n:k or Sub-
section n:k:l. If there is no danger of confusion (which is mostly the case) we write
s s s s s s s s
Apq , Bpq , Fpq , …, apq …(spaces) instead of Ap;q , Bp;q , Fp;q , …, ap;q …. Similarly for
aj m , j m , Qj m (functions, numbers, cubes) instead of aj;m , j;m , Qj;m etc. References
are ordered by names, not by labels, which roughly coincides, but may occasionally
cause minor deviations. The numbers behind I in the Bibliography mark the page(s)
where the corresponding entry is quoted (with the exception of [T78]–[T06]).
It is a pleasure to acknowledge the great help I have received from David Edmunds
(Brighton) who looked through the manuscript and offered many comments.

Jena, Summer 2008 Hans Triebel

Contents

Preface v

1 Spaces on Rn and T n 1
1.1 Definitions, atoms, and local means . . . . . . . . . . . . . . . . . . . 1
1.1.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.3 Local means . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Spaces on Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.1 Wavelets in L2 .Rn / . . . . . . . . . . . . . . . . . . . . . . . . 13
s
1.2.2 Wavelets in Apq .Rn / . . . . . . . . . . . . . . . . . . . . . . . 14
s
1.2.3 Wavelets in Apq .Rn ; w/ . . . . . . . . . . . . . . . . . . . . . 17
1.3 Periodic spaces on R and T n . . . . .
n
. . . . . . . . . . . . . . . . . 19
1.3.1 Definitions and basic properties . . . . . . . . . . . . . . . . . 19
s;per
1.3.2 Wavelets in Apq .Rn / . . . . . . . . . . . . . . . . . . . . . . 23
s
1.3.3 Wavelets in Apq .T n / . . . . . . . . . . . . . . . . . . . . . . . 26

2 Spaces on arbitrary domains 28

2.1 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.1.1 Function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.1.2 Wavelet systems and sequence spaces . . . . . . . . . . . . . . 30
2.2 Homogeneity and refined localisation spaces . . . . . . . . . . . . . . 33
2.2.1 Homogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.2 Pointwise multipliers . . . . . . . . . . . . . . . . . . . . . . . 35
2.2.3 Refined localisation spaces . . . . . . . . . . . . . . . . . . . . 36
2.3 Wavelet para-bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.1 Some preparations . . . . . . . . . . . . . . . . . . . . . . . . 41
s;rloc
2.3.2 Wavelet para-bases in Fpq ./ . . . . . . . . . . . . . . . . . 43
2.3.3 Wavelet para-bases in Lp ./, 1 < p < 1 . . . . . . . . . . . . 46
2.4 Wavelet bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.4.1 Orthonormal wavelet bases in L2 ./ . . . . . . . . . . . . . . 48
s;rloc
2.4.2 Wavelet bases in Lp ./ and Fpq ./ . . . . . . . . . . . . . 53
2.5 Complements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.5.1 Haar bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.5.2 Wavelet bases in Lorentz and Zygmund spaces . . . . . . . . . 60
2.5.3 Constrained wavelet expansions for Sobolev spaces . . . . . . . 65

3 Spaces on thick domains 69

3.1 Thick domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.1.2 Classes of domains . . . . . . . . . . . . . . . . . . . . . . . . 69
viii Contents

3.1.3 Properties and examples . . . . . . . . . . . . . . . . . . . . . 73

3.2 Wavelet bases in Axpq
s
./ . . . . . . . . . . . . . . . . . . . . . . . . 77
z s
3.2.1 The spaces Fpq ./ . . . . . . . . . . . . . . . . . . . . . . . . 77
3.2.2 The spaces Axpqs
./ I . . . . . . . . . . . . . . . . . . . . . . . 79
3.2.3 Complemented subspaces . . . . . . . . . . . . . . . . . . . . 83
3.2.4 Porosity and smoothness zero . . . . . . . . . . . . . . . . . . 85
3.2.5 The spaces Axpqs
./ II . . . . . . . . . . . . . . . . . . . . . . . 89
3.3 Homogeneity and refined localisation, revisited . . . . . . . . . . . . . 91
3.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.3.2 Homogeneity: Proof of Theorem 2.11 . . . . . . . . . . . . . . 92
s;rloc
3.3.3 Wavelet bases in Fpq ./, revisited . . . . . . . . . . . . . . 95
3.3.4 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

4 The extension problem 101

4.1 Introduction and criterion . . . . . . . . . . . . . . . . . . . . . . . . 101
4.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.1.2 A criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.2 Main assertions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.2.1 Positive smoothness . . . . . . . . . . . . . . . . . . . . . . . 103
4.2.2 Negative smoothness . . . . . . . . . . . . . . . . . . . . . . . 105
4.2.3 Combined smoothness . . . . . . . . . . . . . . . . . . . . . . 106
4.3 Complements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.3.1 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.3.2 Constrained wavelet expansions in Lipschitz domains . . . . . . 112
4.3.3 Intrinsic characterisations . . . . . . . . . . . . . . . . . . . . 117
4.3.4 Compact embeddings . . . . . . . . . . . . . . . . . . . . . . . 123

5 Spaces on smooth domains and manifolds 130

5.1 Wavelet frames and wavelet-friendly extensions . . . . . . . . . . . . 130
5.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.1.2 Wavelet frames on manifolds . . . . . . . . . . . . . . . . . . . 132
5.1.3 Wavelet-friendly extensions . . . . . . . . . . . . . . . . . . . 139
5.1.4 Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.1.5 Wavelet frames in domains . . . . . . . . . . . . . . . . . . . . 151
5.2 Wavelet bases: criterion and lower dimensions . . . . . . . . . . . . . 158
5.2.1 Wavelet bases on manifolds . . . . . . . . . . . . . . . . . . . 158
5.2.2 A criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.2.3 Wavelet bases on intervals and planar domains . . . . . . . . . 161
5.3 Wavelet bases: higher dimensions . . . . . . . . . . . . . . . . . . . . 163
5.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.3.2 Wavelet bases on spheres and balls . . . . . . . . . . . . . . . . 164
5.3.3 Wavelet bases in cellular domains and manifolds . . . . . . . . 167
5.3.4 Wavelet bases in C 1 domains and cellular domains . . . . . . . 172
 Contents ix

5.4 Wavelet frames, revisited . . . . . . . . . . . . . . . . . . . . . . . . 174

5.4.1 Wavelet frames in Lipschitz domains . . . . . . . . . . . . . . . 174
5.4.2 Wavelet frames in ."; ı/-domains . . . . . . . . . . . . . . . . . 177

6 Complements 178
6.1 Spaces on cellular domains . . . . . . . . . . . . . . . . . . . . . . . 178
6.1.1 Riesz bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
6.1.2 Basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . 181
6.1.3 A model case: traces and extension . . . . . . . . . . . . . . . 185
6.1.4 A model case: approximation, density, decomposition . . . . . 188
6.1.5 Cubes and polyhedrons: traces and extensions . . . . . . . . . . 192
6.1.6 Cubes and polyhedrons: Riesz bases . . . . . . . . . . . . . . . 196
6.1.7 Cellular domains: Riesz bases . . . . . . . . . . . . . . . . . . 197
6.2 Existence and non-existence of wavelet frames and bases . . . . . . . 199
6.2.1 The role of duality, the spaces Bpq
s
.Rn / . . . . . . . . . . . . . 199
6.2.2 The non-existence of Riesz frames in exceptional spaces . . . . 202
6.2.3 Reinforced spaces . . . . . . . . . . . . . . . . . . . . . . . . 204
6.2.4 A proposal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
6.3 Greedy bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
6.3.1 Definitions and basic assertions . . . . . . . . . . . . . . . . . 210
6.3.2 Greedy Riesz bases . . . . . . . . . . . . . . . . . . . . . . . . 212
6.4 Dichotomy: traces versus density . . . . . . . . . . . . . . . . . . . . 215
6.4.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
6.4.2 Traces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
6.4.3 Dichotomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
6.4.4 Negative smoothness . . . . . . . . . . . . . . . . . . . . . . . 226
6.4.5 Curiosities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
6.4.6 Pointwise evaluation . . . . . . . . . . . . . . . . . . . . . . . 228
6.4.7 A comment on sampling numbers . . . . . . . . . . . . . . . . 232
6.5 Polynomial reproducing formulas . . . . . . . . . . . . . . . . . . . . 237
6.5.1 Global reproducing formulas . . . . . . . . . . . . . . . . . . . 237
6.5.2 Local reproducing formulas . . . . . . . . . . . . . . . . . . . 239
6.5.3 A further comment on sampling numbers . . . . . . . . . . . . 240

Bibliography 243
Symbols 251
Index 255
Chapter 1
Spaces on Rn and T n

1.1 Definitions, atoms, and local means

1.1.1 Definitions

We use standard notation. Let N be the collection of all natural numbers and N0 D
N [ f0g. Let Rn be Euclidean n-space, where n 2 N. Put R D R1 , whereas C is the
complex plane. Let S.Rn / be the usual Schwartz space and S 0 .Rn / be the space of all
tempered distributions on Rn . Furthermore, Lp .Rn / with 0 < p  1, is the standard
quasi-Banach space with respect to the Lebesgue measure in Rn , quasi-normed by
Z 1=p
kf jLp .R /k D
n
jf .x/j dx
p
(1.1)
Rn

with the natural modification if p D 1. As usual, Z is the collection of all integers;

and Zn where n 2 N, denotes the lattice of all points m D .m1 ; : : : ; mn / 2 Rn with
mj 2 Z. Let Nn0 , where n 2 N, be the set of all multi-indices,
P
˛ D .˛1 ; : : : ; ˛n / with ˛j 2 N0 and j˛j D jnD1 ˛j . (1.2)

If x D .x1 ; : : : ; xn / 2 Rn and ˇ D .ˇ1 ; : : : ; ˇn / 2 Nn0 then we put

x ˇ D x1ˇ1    xnˇn (monomials): (1.3)

If ' 2 S.Rn / then

Z
n=2
O
'./ D .F '/./ D .2/ e ix '.x/ dx;  2 Rn ; (1.4)
Rn

denotes the Fourier transform of '. As usual, F 1 ' and ' _ stand for the inverse
Fourier transform, given by the right-hand side of (1.4) with i in place of i . Here
x denotes the scalar product in Rn . Both F and F 1 are extended to S 0 .Rn / in the
standard way. Let '0 2 S.Rn / with

'0 .x/ D 1 if jxj  1 and '0 .y/ D 0 if jyj  3=2; (1.5)

and let
'k .x/ D '0 .2k x/  '0 .2kC1 x/; x 2 Rn ; k 2 N: (1.6)
Since
1
X
'j .x/ D 1 for x 2 Rn ; (1.7)
j D0
2 1 Spaces on Rn and T n

the 'j form a dyadic resolution of unity. The entire analytic functions .'j fO/_ .x/ make
sense pointwise in Rn for any f 2 S 0 .Rn /.

Definition 1.1. Let ' D f'j gj1D0 be the above dyadic resolution of unity.
(i) Let
0 < p  1; 0 < q  1; s 2 R: (1.8)
s
Then Bpq .Rn / is the collection of all f 2 S 0 .Rn / such that

X
1 1=q
kf jBpq
s
.Rn /k' D 2jsq k.'j fO/_ jLp .Rn /kq <1 (1.9)
j D0

.with the usual modification if q D 1/.

(ii) Let
0 < p < 1; 0 < q  1; s 2 R: (1.10)
s
Then Fpq .Rn / is the collection of all f 2 S 0 .Rn / such that

 X
1
ˇ ˇq 1=q 
 
kf jFpq
s
.Rn /k' D  2jsq ˇ.'j fO/_ .  /ˇ jLp .Rn / < 1 (1.11)
j D0

.with the usual modifications if q D 1/.

Remark 1.2. The theory of these spaces may be found in [T83], [T92], [T06]. In
particular these spaces are independent of admitted resolutions of unity ' according to
(1.5)–(1.7) (equivalent quasi-norms). This justifies our omission of the subscript ' in
(1.9), (1.11) in the sequel. We remind the reader of a few special cases and properties
referring for details to the above books, especially to [T06], Section 1.2.
(i) Let 1 < p < 1. Then

Lp .Rn / D Fp;2
0
.Rn / (1.12)

is a well-known Paley–Littlewood theorem.

(ii) Let 1 < p < 1 and k 2 N0 . Then

Wpk .Rn / D Fp;2

k
.Rn / (1.13)

are the classical Sobolev spaces usually equivalently normed by

 X 1=p
kf jWpk .Rn /k D kD ˛ f jLp .Rn /kp : (1.14)
j˛jk

This generalises (1.12).

(iii) Let  2 R. Then  _
I W f 7! hi fO (1.15)
 1.1 Definitions, atoms, and local means 3

with hi D .1 C jj2 /1=2 , is an one-to-one map of S.Rn / onto itself and of S 0 .Rn /
s
onto itself. Then I is a lift for the spaces Apq .Rn / with A D B or A D F and s 2 R,
0 < p  1 (p < 1 for the F -scale), 0 < q  1:
s
I Apq .Rn / D Apq
s
.Rn / (1.16)

(equivalent quasi-norms). With

Hps .Rn / D Is Lp .Rn /; s 2 R; 1 < p < 1; (1.17)

one has
Hps .Rn / D Fp;2
s
.Rn /; s 2 R; 1 < p < 1; (1.18)
and
Hpk .Rn / D Wpk .Rn /; 1 < p < 1; k 2 N0 : (1.19)
Nowadays one calls the Hps .Rn /Sobolev spaces (sometimes fractional Sobolev spaces
or Bessel-potential spaces) with the classical Sobolev spaces Wpk .Rn / as a special case.
(iv) We denote
C s .Rn / D B11
s
.Rn /; s 2 R; (1.20)
as Hölder–Zygmund spaces. Let

.1h f /.x/ D f .x C h/  f .x/; .lC1

h
f /.x/ D 1h .lh f /.x/; (1.21)

where x 2 Rn , h 2 Rn , l 2 N, be the iterated differences in Rn . Let 0 < s < m 2 N.

Then
kf jC s .Rn /km D sup jf .x/j C sup jhjs jm h f .x/j; (1.22)
x2Rn

where the second supremum is taken over all x 2 Rn and h 2 Rn with 0 < jhj  1,
are equivalent norms in C s .Rn /.
(v) This assertion can be generalised as follows. Once more let 0 < s < m 2 N
and 1  p; q  1. Then
Z 1=q
sq n q dh
kf jBpq .R /km D kf jLp .R /k C
s n n
jhj kh f jLp .R /k
m
jhj1 jhjn
(1.23)
and
Z 1 1=q
n q dt
kf jBpqs
.Rn /km D kf jLp .Rn /kC t sq sup km h f jLp .R /k (1.24)
0 jhjt t

(with the usual modification if q D 1) are equivalent norms in Bpq s

.Rn /. These are
the classical Besov spaces. We refer to [T92], Chapter 1, and [T06], Chapter 1, where
one finds the history of these spaces, further special cases and classical assertions. In
addition, (1.23), (1.24) remain to be equivalent quasi-norms in
 
s
Bpq .Rn / with 0 < p; q  1 and s > n p1  1 C : (1.25)
4 1 Spaces on Rn and T n

1.1.2 Atoms

We give a detailed description of atomic representations of the spaces introduced in

Definition 1.1 for two reasons. First we need these assertions later on. Secondly, atoms
and local means, which are the subject of Section 1.1.3, are dual to each other where
the natural smoothness assumptions and the cancellation conditions change their roles.
Let Qj m be cubes in Rn with sides parallel to the axes of coordinates, centred at
2 m with side-length 2j C1 where m 2 Zn and j 2 N0 . If Q is a cube in Rn and
j

r > 0 then rQ is the cube in Rn concentric with Q and with side-length r times of the
side-length of Q. Let j m be the characteristic function of Qj m .
Definition 1.3. Let 0 < p  1, 0 < q  1. Then bpq is the collection of all
sequences ˚
 D j m 2 C W j 2 N0 ; m 2 Zn (1.26)
such that
1  X
X q=p 1=q
k jbpq k D jj m jp < 1; (1.27)
j D0 m2Zn

and fpq is the collection of all sequences  according to (1.26) such that
 X 1=q 
 
k jfpq k D  2j nq=p jj m j m .  /jq jLp .Rn / < 1 (1.28)
j;m

with the usual modifications if p D 1 and/or q D 1.

Remark 1.4. Note that the factor 2j nq=p in (1.28) does not appear if one relies on the
p-normalised characteristic function j.p/
m .x/ D 2
n.j 1/=p
j m .x/. This is the usual
way to say what is meant by fpq . But for what follows the above version seems to be
more appropriate. Of course bpp D fpp .
Definition 1.5. Let s 2 R, 0 < p  1, K 2 N0 , L 2 N0 and d  1. Then the
L1 -functions aj m W Rn 7! C with j 2 N0 , m 2 Zn , are called .s; p/-atoms if

supp aj m  d Qj m ; j 2 N0 ; m 2 Zn I (1.29)

there exist all .classical/ derivatives D ˛ aj m with j˛j  K such that

n
jD ˛ aj m .x/j  2j.s p /Cj j˛j ; j˛j  K; j 2 N0 ; m 2 Zn ; (1.30)

and Z
x ˇ aj m .x/ dx D 0; jˇj < L; j 2 N; m 2 Zn : (1.31)
Rn

Remark 1.6. No cancellation (1.31) for a0;m is required. Furthermore, if L D 0

then (1.31) is empty (no condition). If K D 0 then (1.30) means aj m 2 L1 and
n
jaj m .x/j  2j.s p / . Of course, the conditions for the above atoms depend not only
 1.1 Definitions, atoms, and local means 5

on s and p, but also on the given numbers K; L; d . But this will only be indicated
when extra clarity is required. Otherwise we speak about .s; p/-atoms instead of
.s; p/K;L;d -atoms. Let as usual
   
1 1
p D n 1 and pq D n 1 (1.32)
p C min.p; q/ C

where bC D max.b; 0/ if b 2 R.

Theorem 1.7. (i) Let 0 < p  1, 0 < q  1, s 2 R. Let K 2 N0 , L 2 N0 , d 2 R

with
K > s; L > p  s; (1.33)
and d  1 be fixed. Then f 2 S 0 .Rn / belongs to Bpqs
.Rn / if, and only if, it can be
represented as
X1 X
f D j m aj m (1.34)
j D0 m2Zn

where the aj m are .s; p/-atoms .more precisely .s; p/K;L;d -atoms/ according to Defi-
nition 1.5 and  2 bpq . Furthermore

kf jBpq
s
.Rn /k  inf k jbpq k (1.35)

are equivalent quasi-norms where the infimum is taken over all admissible representa-
tions (1.34) .for fixed K; L; d /.
(ii) Let 0 < p < 1, 0 < q  1, s 2 R. Let K 2 N0 , L 2 N0 , d 2 R with

K > s; L > pq  s; (1.36)

and d  1 be fixed. Then f 2 S 0 .Rn / belongs to Fpqs

.Rn / if, and only if, it can be
represented by (1.34) where aj m are .s; p/-atoms .more precisely .s; p/K;L;d -atoms/
according to Definition 1.5 and  2 fpq . Furthermore,

kf jFpq
s
.Rn /k  inf k jfpq k (1.37)

are equivalent quasi-norms where the infimum is taken over all admissible representa-
tions (1.34) ( for fixed K; L; d ).

Remark 1.8. Recall that dQj m are cubes centred at 2j m with side-length d 2j C1
where j 2 N0 and m 2 Zn . For fixed d with d  1 and j 2 N0 there is some
overlap of the cubes dQj m where m 2 Zn . This makes clear that Theorem 1.7 based
on Definition 1.5 is reasonable. Otherwise the above formulations coincide essentially
with [T06], Section 1.5.1. There one finds also some technical comments how the
convergence in (1.34) must be understood. Atoms of the above type go back essentially
to [FrJ85], [FrJ90]. But more details about the complex history of atoms may be found
in [T92], Section 1.9.
6 1 Spaces on Rn and T n

1.1.3 Local means

We wish to derive estimates for local means which are dual to atomic representations
according to Theorem 1.7 as far as smoothness assumptions and cancellation properties
are concerned. First we collect some definitions. Let Qj m be the same cubes as in the
previous Section 1.1.2.
Definition 1.9. Let A 2 N0 , B 2 N0 and C > 0. Then the L1 -functions kj m W Rn 7!
C with j 2 N0 , m 2 Zn , are called kernels .of local means/ if
supp kj m  CQj m ; j 2 N 0 ; m 2 Zn I (1.38)
there exist all .classical/ derivatives D ˛ kj m with j˛j  A such that

jD ˛ kj m .x/j  2j nCj j˛j ; j˛j  A; j 2 N0 ; m 2 Zn ; (1.39)

and Z
x ˇ kj m .x/ dx D 0; jˇj < B; j 2 N; m 2 Zn : (1.40)
Rn

Remark 1.10. No cancellation (1.40) for k0;m is required. Furthermore if B D 0 then

(1.40) is empty (no condition). If A D 0 then (1.39) means kj m 2 L1 and jkj m .x/j 
2j n . Compared with Definition 1.5 for atoms we have different normalisations in (1.30)
and (1.39) (also due to the history of atoms). We adapt the sequence spaces introduced
in Definition 1.3 in connection with atoms to the above kernels.
Definition 1.11. Let s 2 R, 0 < p  1, 0 < q  1. Then bNpq
s
is the collection of
all sequences  according to (1.26) such that
X
1  X q=p 1=q
k jbNpq
n
s
kD 2j.s p /q jj m jp < 1; (1.41)
j D0 m2Zn

and fNpq
s
is the collection of all sequences  according to (1.26) such that
 X 1=q 
 
k jfNpq
s
kD 2jsq jj m j m .  /jq jLp .Rn / < 1 (1.42)
j;m

with the usual modification if p D 1 and/or q D 1.

Remark 1.12. The notation bpq s s
and fpq (without bar) will be reserved for a slight
modification of the above sequence spaces in connection with wavelet representations.
One has bNpp
s
D fNpp
s
.
Definition 1.13. Let f 2 Bpq s
.Rn / where s 2 R, 0 < p  1, 0 < q  1. Let kj m
be the kernels according to Definition 1.9 with A > p  s where p is given by (1.32)
and B 2 N0 . Then
Z
kj m .f / D .f; kj m / D kj m .y/ f .y/ dy; j 2 N0 ; m 2 Zn ; (1.43)
Rn
 1.1 Definitions, atoms, and local means 7

are local means, considered as a dual pairing within .S.Rn /; S 0 .Rn //. Furthermore,
˚
k.f / D kj m .f / W j 2 N0 ; m 2 Zn : (1.44)

Remark 1.14. We justify the dual pairing (1.43). According to [T83], Theorems 2.11.2,
s
2.11.3, one has for the dual spaces of Bpp .Rn / that
sCp
s
Bpp .Rn /0 D Bp0 p0 .Rn /; x 2 R; 0 < p < 1; (1.45)

where
1 1
C 0 D1 if 1  p  1 and p 0 D 1 if 0 < p < 1: (1.46)
p p
Since kj m 2 C A .Rn / (L1 .Rn / if A D 0) has compact support one obtains that
kj m 2 Buv
A"
.Rn / for any " > 0 and 0 < u; v  1. By
s
Bpq .Rn /  Bpp
s"
.Rn / and s
B1q .Rn /  Bpq
s;loc
.Rn / (1.47)

locally for any s 2 R, " > 0, 0 < p < 1 and 0 < q  1 one has by (1.45) and
A > p  s that (1.43) makes always sense as a dual pairing. This applies also to
f 2 Fpqs
.Rn / since
s
Fpq .Rn /  Bp;max.p;q/
s
.Rn /: (1.48)
But (1.43) can also be justified for f 2 Bpq
s
.Rn / and f 2 Fpq
s
.Rn / as in [T06],
Section 5.1.7, by direct arguments.
s
In Section 1.2 we characterise the spaces Bpq .Rn / and Fpq
s
.Rn / in terms of wavelets.
Since wavelets are special atoms one has by Theorem 1.7 the desired estimates from
above. But wavelets can also be considered as kernels kj m of corresponding local
means. This gives finally the needed estimates from below which will be reduced to
the following theorem which might be considered as the main assertion of Section 1.1.
Theorem 1.15. (i) Let 0 < p  1, 0 < q  1, s 2 R. Let kj m be kernels according
to Definition 1.9 where A 2 N0 , B 2 N0 with

A > p  s; B > s; (1.49)

and C > 0 are fixed. Let k.f / be as in (1.43), (1.44). Then for some c > 0 and all
f 2 Bpq
s
.Rn /,
kk.f / jbNpq
s
k  c kf jBpq
s
.Rn /k: (1.50)
(ii) Let 0 < p < 1, 0 < q  1, s 2 R. Let kj m and k.f / be the above kernels
where A 2 N0 , B 2 N0 with

A > pq  s; B > s; (1.51)

and C > 0 are fixed. Then for some c > 0 and all f 2 Fpq
s
.Rn /,

kk.f / jfNpq
s
k  c kf jFpq
s
.Rn /k: (1.52)
8 1 Spaces on Rn and T n

Proof. Step 1. We prove (i). By Remark 1.14 the local means kj m .f / with f 2
s
Bpq .Rn / make sense. Let
1 X
X
f D rl arl .x/; f 2 Bpq
s
.Rn /; (1.53)
rD0 l2Zn

be an optimal atomic decomposition according to Theorem 1.7 with arl as in (1.29)–

(1.31) where
K D B > s and L D A > p  s: (1.54)
We recall that an atomic decomposition is called optimal if for some suitably chosen
fixed D > 0 and all f 2 Bpqs
.Rn /,
k jbpq k  D kf jBpq
s
.Rn /k: (1.55)
For j 2 N0 we split (1.53) into

X
j 1
X
f D fj C f j D  C  : (1.56)
rD0 rDj C1

Then
Z Z Z
kj m .y/ f .y/ dy D kj m .y/ fj .y/ dy C kj m .y/ f j .y/ dy: (1.57)
Rn Rn Rn

Let r  j and let l 2 lrj .m/ where C and d in

˚
lrj .m/ D l 2 Zn W CQj m \ d Qrl 6D ; (1.58)

have the same meaning as in (1.38), (1.29). Then card lrj .m/  1 and
ˇZ ˇ
n ˇ ˇ
j.s p /ˇ ˇ
2 ˇ n kj m .y/ arl .y/ dy ˇ
R
X Z
n
c2 j.s p /
sup jD arl .x/j

jkj m .y/j  jy  2j mjB dy
x Rn (1.59)
jjDB
n
 c 0 2.j r/.s p / 2rB 2jB 2j n 2j n
n
D c 0 2.j r/.s p B/ :
Hence one obtains for any fixed " > 0 that (modification if p D 1)
ˇZ ˇp
ˇ ˇ X
j X
n
j.s p
2 /p ˇ ˇ
kj m .y/ fj .y/ dy ˇ  c
n
jrl jp 2.j r/.s p BC"/p : (1.60)
ˇ
R n
rD0 l2l j .m/
r

For fixed r; j with r  j and l 2 Z one has n

˚
card m 2 Zn W l 2 lrj .m/  2.j r/n : (1.61)
 1.1 Definitions, atoms, and local means 9

Then it follows from (1.60) that

ˇ ˇp
X ˇZ ˇ
n
j.s p
2 /p ˇ kj m .y/ fj .y/ dy ˇˇ
ˇ
m2Zn R n

X
j
n X
c 2.j r/n 2p.j r/.s p BC"/ jrl jp (1.62)
rD0 l2Zn

X
j X
Dc 2p.j r/.sBC"/ jrl jp :
rD0 l2Zn

Let r > j . Then one obtains by (1.54) for l 2 lrj .m/ that
ˇZ ˇ
n ˇ ˇ
j.s p / ˇ ˇ
2 ˇ n kj m .y/ arl .y/ dy ˇ
R
X Z
n
c2 j.s p /
sup jD kj m .x/j

jarl .y/j  jy  2r ljA dy
x Rn (1.63)
jjDA
n
 c 0 2.j r/.s p / 2j nCAj 2rArn
n
D c 0 2.j r/.s p CnCA/ :

By (1.56), (1.57) it follows that for any fixed " > 0,

ˇZ ˇp  X p
n ˇ ˇ X
j.s p / ˇ ˇ n
.j r/.s p CnCA"/p
2 ˇ kjm .y/ f j
.y/ dy ˇ  c 2 jrl j :
Rn r>j j
l2lr .m/
(1.64)
If p  1 then one obtains
ˇZ ˇp
ˇ ˇ X X
n
/p ˇ ˇ c n
2j.s p
ˇ kj m .y/ f j
.y/ dy ˇ 2.j r/.s p CnCA"/p jrl jp :
Rn r>j j
l2lr .m/
(1.65)
If 1 < p < 1 (modification if p D 1) then it follows from (1.64), Hölder’s inequality
and card lrj .m/  2n.rj / that
ˇZ ˇp
ˇ
/p ˇ
ˇ
2
n
j.s p j ˇ
ˇ n kj m .y/ f .y/ dy ˇ
R
X n X
c 2.j r/.s p CnCA"/p 2.rj /.p1/n jrl jp
r>j j
(1.66)
l2lr .m/
X X
Dc 2.j r/.sCA"/p jrl jp :
r>j j
l2lr .m/

Summation over m 2 Zn results for all p with 0 < p  1 (modification if p D 1)

10 1 Spaces on Rn and T n

in
ˇ ˇp
X ˇZ ˇ X X
2
n
j.s p /p ˇ kj m .y/ f .y/ dy ˇˇ  c
j
2.j r/.sp CA"/p jrl jp :
ˇ
m2Zn R n
r>j l2Zn
(1.67)
Now one obtains by (1.62), (1.67) and (1.49) for some ~ > 0,
ˇ ˇp
X ˇZ ˇ 1
X X
2
n
j.s p /p ˇ ˇ
kj m .y/ f .y/ dy ˇ  c 2.j r/~p jrl jp : (1.68)
ˇ
m2Zn R n
rD0 l2Zn

Using (1.43), (1.41), (1.55) based on (1.27) we obtain by standard arguments that

kk.f / jbNpq
s
k  c k jbpq k  c D kf jBpq
s
.Rn /k: (1.69)

This proves (1.50)

Step 2. The proof of part (ii) of the theorem will be based on the vector-valued
maximal inequality
1 
 X  1=q   X
1 1=q 
 
n   
 M jgk j .  /
w q=w
jLp .R /  c  jgk .  /jq jLp .Rn / (1.70)
kD0 kD0

due to [FeS71] where

0 < p < 1; 0 < q  1; 0 < w < min.p; q/: (1.71)

Here M is the Hardy–Littlewood maximal function,

Z
.M g/.x/ D sup jQj1 jg.y/j dy; x 2 Rn ; (1.72)
Q

where the supremum is taken over all cubes centred at x. A short proof may also
be found in [Tor86], pp. 303–305. By Remark 1.14 the local means kj m .f / with
f 2 Fpq
s
.Rn / make sense. Let
1 X
X
f D rl arl .x/; f 2 Fpq
s
.Rn /; (1.73)
rD0 l2Zn

be an optimal atomic decomposition according to Theorem 1.7 with arl as in (1.29)–

(1.31) where
K D B > s and L D A > pq  s: (1.74)
An atomic decomposition is called optimal if for some suitably chosen fixed D > 0
and all f 2 Fpq
s
.Rn /,
k jfpq k  D kf jFpqs
.Rn /k: (1.75)
 1.1 Definitions, atoms, and local means 11

We rely again on the splitting (1.56), (1.57). Let r  j . We assume q < 1 (if q D 1
one has to modify what follows appropriately). Then the counterpart of (1.59), (1.60)
is now given by
ˇZ ˇq
ˇ ˇ
jsq
2 j m .x/ ˇ ˇ kj m .y/ fj .y/ dy ˇˇ
Rn
X
j X (1.76)
rn
c jrl jq 2 p q j m .x/ 2.j r/.sBC"/q
rD0 l2l j .m/
r

P
where x 2 Rn . For fixed j; r, and l the summation j m .x/ over those m 2 Zn with
l 2 lrj .m/ is comparable with rl .x/ and can be estimated from above by its maximal
function. Hence we obtain for any w > 0,
X ˇZ ˇq
ˇ ˇ
2jsq ˇ
j m .x/ ˇ kj m .y/ fj .y/ dy ˇˇ
m2Zn Rn
(1.77)
X
j X  rn 
.j r/.sBC"/q
c 2 M jrl jw 2 p w rl .  / .x/q=w :
rD0 l2Zn

This is the counterpart of (1.62). Let now r > j . Then it follows as in (1.63), (1.64)
that
ˇZ ˇ
ˇ ˇ X X
js
2 j m .x/ ˇˇ kj m .y/ f .y/ dy ˇˇ  c j m .x/
j
2.j r/.sCACn/
n
2r p jrl j:
Rn r>j j
l2lr .m/
(1.78)
For x 2 Rn with j m .x/ D 1 and 0 < w < 1 the last factor can be estimated by
 X n
w X nw
2r p jrl j  2r p jrl jw
j j
l2lr .m/ l2lr .m/
Z X nw
j n
c2 rn
2 2jn
2r p jrl jw rl .y/ dy (1.79)
Rn j
l2lr .m/
 X nw

 c 0 2.rj /n M 2r p jrl jw rl .  / .x/:
j
l2lr .m/

Assuming again q < 1 one obtains for any fixed " > 0 that
ˇZ ˇq
ˇ ˇ
jsq ˇ
2 j m .x/ ˇ kj m .y/ f .y/ dy ˇˇ
j
Rn
X n
 X nw
 (1.80)
c 2.j r/.sCACn w "/q M 2r p jrl jw rl .  / .x/q=w :
r>j j
l2lr .m/
12 1 Spaces on Rn and T n

Since n
A C s > pq D n (1.81)
min.1; p; q/
one may choose w with w < min.1; p; q/ and " > 0 in (1.80) such that
n
~ DACsCn  " > 0: (1.82)
w
With r D j C t we have
X ˇZ ˇq
ˇ ˇ
2 jsq
j m .x/ ˇˇ kj m .y/ f .y/ dy ˇˇ
j

m2Zn
n R
1
X X  X nw
 (1.83)
c 2t~q M 2.j Ct/ p jj Ct;l jw j Ct;l .  / .x/q=w :
tD1 m2Zn j
l2lj Ct .m/

Summation over j in (1.77) with s  B C " < 0 and in (1.83) gives by (1.42) that
 X
1 X
   
 q=w 1=q 
kk.f / jfNpq
rn
s
kc M jrl jw 2 p w rl .  / jLp .Rn /
rD0 l2Zn
1  1   q=w 1=q 
X  X X X 
t~ 0 
Cc 2  M ::: jLp .R /
n

tD1 j D0 m2Zn j
l2lj Ct .m/
(1.84)
for some 0 < ~ 0 < ~. By construction, the lattice 2j Zn refined by l 2 ljjCt .m/ is
related to 2j t Zn . Since w < min.1; p; q/ one can apply (1.70). Then one obtains
by (1.28) and (1.75) that

kk.f / jfNpq
s
k  c k jfpq k  c D kf jFpq
s
.Rn /k: (1.85)

This proves (1.52).

Remark 1.16. We refer in this context also to [Kyr03]. In particular we took over from
this paper the idea to decompose the lattice 2j t Zn in connection with (1.83), (1.84)
into clusters around 2j Zn . These clusters disappear after the vector-valued maximal
inequality has been applied. According to Definition 1.9 the kernels kj m have compact
supports. This was of some use for us. But there is little doubt that one can replace
the compactness assumption (1.38) by a sufficiently strong decay. We refer again to
[Kyr03].
Remark 1.17. We followed closely [Tri07c] (as always throughout Chapter 1). In
particular, Theorem 1.15 coincides essentially with [Tri07c], Theorem 15, and its proof.
But the first formulation of Theorem 1.15 was given in [Tri07b], Theorem 36 (without
proof), where we used it in connection with para-wavelet bases in domains and sampling
numbers.
 1.2 Spaces on Rn 13

1.2 Spaces on Rn
1.2.1 Wavelets in L2 .Rn /

Recall that this book might be considered as the continuation of [T06]. There we dealt
in Section 3.1, with wavelet bases (isomorphisms) in the spaces
s
Bpq .Rn /; s
Fpq .Rn / where s 2 R; 0 < p; q  1; (1.86)
(with p < 1 for the F -spaces) based on a weaker version of Theorem 1.15. We wish
to improve these assertions using now Theorem 1.15. But first we fix some notation
and recall a few well-known facts.
We suppose that the reader is familiar with wavelets in Rn of Daubechies type and
the related multiresolution analysis. The standard references are [Dau92], [Mal99],
[Mey92], [Woj97]. A short summary of what is needed may also be found in [T06],
Section 1.7. We give a brief description of some basic notation. As usual, C u .R/ with
u 2 N collects all (complex-valued) continuous functions on R having continuous
bounded derivatives up to order u (inclusively). Let

F 2 C u .R/; M 2 C u .R/; u 2 N; (1.87)

be real compactly supported Daubechies wavelets with
Z
M .x/ x dx D 0 for all v 2 N0 with v < u.
v
(1.88)
R

Recall that F is called the scaling function (father wavelet) and M the associated
wavelet (mother wavelet). We extend these wavelets from R to Rn by the usual tensor
procedure. Let
G D .G1 ; : : : ; Gn / 2 G 0 D fF; M gn ; (1.89)
which means that Gr is either F or M . Let
G D .G1 ; : : : ; Gn / 2 G j D fF; M gn ; j 2 N; (1.90)
which means that Gr is either F or M where * indicates that at least one of the
components of G must be an M . Hence G 0 has 2n elements, whereas G j with j 2 N
has 2n  1 elements. Let
j
Y
n
‰G;m .x/ D 2j n=2 j
Gr .2 xr  mr /; G 2 G j ; m 2 Zn ; (1.91)
rD1

where j 2 N0 . We always assume the F and M in (1.87) have L2 -norm 1. Then

˚ j
‰G;m W j 2 N0 ; G 2 G j ; m 2 Zn (1.92)
is an orthonormal basis in L2 .Rn / .for any u 2 N/ and
1 X X
X
j n=2 j
f D j;G
m 2 ‰G;m (1.93)
j D0 G2G j m2Zn
14 1 Spaces on Rn and T n

with Z
j
j;G
m D j;G
m .f /D2 j n=2
f .x/ ‰G;m .x/ dx
Rn (1.94)
 j 
D 2j n=2 f; ‰G;m
j
in the corresponding expansion, adapted to our needs, where 2j n=2 ‰G;m are uniformly
bounded functions (with respect to j and m).

s
1.2.2 Wavelets in Apq .Rn /

One may ask whether the orthonormal basis (1.92) in L2 .Rn / remains an (uncondi-
tional) basis in other spaces on Rn . First candidates are Lp .Rn / with 1 < p < 1 but
also related (fractional) Sobolev spaces and classical Besov spaces. Something may
be found in the above-mentioned books [Dau92], [Mal99], [Mey92], [Woj97]. One
may also consult [T06], Remarks 1.63, 1.65, pp. 32, 34, for more details and further
references. An extension of this theory to all spaces
s
Apq .Rn / where either A D B or A D F and s 2 R, 0 < p; q  1; (1.95)

with p < 1 for the F -spaces may be found in [Kyr03], [Tri04] and [T06], Section 3.1.3,
Theorem 3.5, p. 154. Basically one identifies the wavelets in (1.91) with atoms and
with kernels of local means as described in Theorems 1.7 and 1.15. But the sharp
version of Theorem 1.15 was not available at this time. We had to rely on a weaker
version (from the point of view of applications to wavelets) and obtained wavelet bases
and isomorphisms for all spaces in (1.95) under somewhat unnatural restrictions for
the required smoothness u 2 N in (1.87) of the underlying wavelets.
Now we return to this topic and adapt first the sequence spaces in Definition 1.11
incorporating the additional parameter G 2 G j according to (1.89), (1.90).
Definition 1.18. Let s 2 R, 0 < p  1, 0 < q  1. Then bpq
s
is the collection of
all sequences ˚
m 2 C W j 2 N0 ; G 2 G ; m 2 Z
 D j;G j n
(1.96)
such that
X
1 X  X q=p 1=q
n
j.s p /q
k jbpq
s
k D 2 jj;G
m j
p
<1 (1.97)
j D0 G2G j m2Zn

s
and fpq is the collection of all sequences (1.96) such that
 X 1=q 
 
k jfpq
s
kD 2jsq jj;G
m j m .  /j
q
jLp .Rn / < 1 (1.98)
j;G;m

with the usual modification if p D 1 and/or q D 1.

 1.2 Spaces on Rn 15

Remark 1.19. One may consult Remark 1.12 as far as notation is concerned. Recall
that j m is the characteristic function of the cube Qj m as introduced in Section 1.1.2
and used in Definitions 1.3 and 1.11.
As justified at the beginning of [T06], Section 3.1.3, we may abbreviate the right-
hand side of (1.93) by X
j n=2 j
j;G
m 2 ‰G;m (1.99)
j;G;m

since the conditions for the sequences  always ensure that the corresponding series
converges unconditionally at least in S 0 .Rn / (which means that any rearrangement
converges in S 0 .Rn / and has the same limit). Local convergence in Bpq

.Rn / means
convergence in Bpq .K/ for any ball K in R . Similarly for Fpq .R /. Recall that p
 n  n

and pq are given by (1.32).

Otherwise we use standard notation naturally extended from Banach spaces to
quasi-Banach spaces. In particular fbj gj1D1  B in a separable complex quasi-Banach
B space is called a basis if any b 2 B can be uniquely represented as
1
X
bD j bj ; j 2 C (convergence in B): (1.100)
j D1

A basis fbj gj1D1 is called an unconditional basis if for any rearrangement  of N

(one-to-one map of N onto itself) fb .j / gj1D1 is again a basis and
1
X
bD  .j / b .j / (convergence in B) (1.101)
j D1

for any b 2 B with (1.100). Standard bases of (separable) sequence spaces as consid-
ered in this book are always unconditional. We refer to [AlK06] for details about bases
in Banach (sequence) spaces.
j
Theorem 1.20. (i) Let 0 < p  1, 0 < q  1, s 2 R and let ‰G;m be the wavelets
(1.91) based on (1.87), (1.88) with

u > max.s; p  s/: (1.102)

Let f 2 S 0 .Rn /. Then f 2 Bpq

s
.Rn / if, and only if, it can be represented as
X
j n=2 j
f D j;G
m 2 ‰G;m ;  2 bpq
s
; (1.103)
j;G;m

unconditional convergence being in S 0 .Rn / and locally in any space Bpq


.Rn / with
 < s. The representation (1.103) is unique,
 j 
m D2
j;G j n=2
f; ‰G;m (1.104)
16 1 Spaces on Rn and T n

and
˚  j 
I W f 7! 2j n=2 f; ‰G;m (1.105)
is an isomorphic map of Bpqs
.Rn / onto bpq
s
. If, in addition, p < 1, q < 1, then
˚ j
‰G;m is an unconditional basis in Bpq .R /.
s n

(ii) Let 0 < p < 1, 0 < q  1, s 2 R, and

u > max.s; pq  s/: (1.106)

Let f 2 S 0 .Rn /. Then f 2 Fpq

s
.Rn / if, and only if, it can be represented as
X
j n=2 j
f D j;G
m 2 ‰G;m ;  2 fpq
s
; (1.107)
j;G;m

unconditional convergence being in S 0 .Rn / and locally in any space Fpq 

.Rn / with
 < s. The representation (1.107) is unique with (1.104). Furthermore˚I in (1.105)
j
is an isomorphic map of Fpq
s
.Rn / onto fpq
s
. If, in addition, q < 1, then ‰G;m is an
unconditional basis in Fpq .R /.
s n

Proof. Step 1. This theorem coincides with the corresponding assertions in [Tri04] and
[T06], Section 3.1.3, Theorem 3.5, pp. 153–156, under the more restrictive smoothness
and cancellation assumptions (1.87), (1.88) with
   
2n n 2n n
u > max s; C s and u > max s; C s (1.108)
p 2 min.p; q/ 2
s
for Bpq .Rn / and Fpq
s
.Rn /, respectively. But all technicalities such as the unconditional
convergence of (1.103), (1.107), the uniqueness (1.104), the isomorphic map (1.105),
the use of duality, (1.45), (1.46), the unconditional bases, can be taken over verbatim.
Hence it remains to justify that the improved assumptions (1.102) and (1.106) are
natural and sufficient.
Step 2. Let f be given by (1.103). Then
n
ajGm D 2j.s p / 2j n=2 ‰G;m
j
; G 2 Gj ; (1.109)
s
are atoms in Bpq .Rn / according to Definition 1.5 and Theorem 1.7 (i) with K D L D u
s
(up to unimportant constants). Having the different normalisations for bpq in (1.97)
and for bpq in (1.27) in mind one obtains by Theorem 1.7 (i) that f 2 Bpq .Rn / and
s

kf jBpq
s
.Rn /k  c k jbpq
s
k (1.110)

where the extra summation over G does not influence this argument. Here c in (1.110)
is independent of f . Conversely, if f 2 Bpq
s
.Rn /, then

j
kjGm D 2j n=2 ‰G;m ; j 2 N 0 ; G 2 G j ; m 2 Zn ; (1.111)
 1.2 Spaces on Rn 17

s
are kernels of corresponding local means in Bpq .Rn / according to Definitions 1.9, 1.13
based on (1.91), and Theorem 1.15 (i) with A D B D u (again neglecting unimportant
constants). The modification
˚
k.f / D kjGm .f / W j 2 N0 ; G 2 G j ; m 2 Zn (1.112)

of (1.44) is immaterial for the application of Theorem 1.15 (i). One obtains

kk.f / jbpq
s
k  c kf jBpq
s
.Rn /k (1.113)

for some c > 0 and all f 2 Bpq

s
.Rn /. By (1.104) one has k.f / D  and one obtains
by (1.110) and (1.113) that

k jbpq
s
k  kf jBpq
s
.Rn /k: (1.114)

Now one obtains part (i) of the above theorem in the same way as in [Tri04], [T06].
Step 3. The proof for the F -spaces is the same now relying on Theorems 1.7 (ii)
and 1.15 (ii).

Remark 1.21. Again we followed closely [Tri07c].

s
1.2.3 Wavelets in Apq .Rn ; w/

Based on [HaT05] we extended in [T06], Section 6.2, Theorem 1.20 with u as in

(1.108) to some weighted spaces. There is no difficulty to apply the corresponding
arguments to the above improved version with u according to (1.102), (1.106). The
main reason for giving an explicit formulation comes from the application to construct
wavelet bases for periodic spaces which is the subject of Section 1.3.

Definition 1.22. Let w be a real C 1 function in Rn such that for all 2 Nn0 and
suitable positive numbers c ,

jD  w.x/j  c w.x/ for all x 2 Rn ; (1.115)

and
 ˛=2
0 < w.x/  c w.y/ 1 C jx  yj2 for all x 2 Rn ; y 2 Rn ; (1.116)

and some constants c > 0 and ˛  0. Let ' D f'j gj1D0 be a resolution of unity
according to (1.5)–(1.7).
s
(i) Let p, q, s as in (1.8). Then Bpq .Rn ; w/ is the collection of all f 2 S 0 .Rn /
such that
X
1
 q 1=q
kf jBpq
s
.Rn ; w/k D 2jsq w.'j fO/_ jLp .Rn / <1 (1.117)
j D0
18 1 Spaces on Rn and T n

.with the usual modification if q D 1/.

s
(ii) Let p, q, s as in (1.10). Then Fpq .Rn ; w/ is the collection of all f 2 S 0 .Rn /
such that
  1 1=q 
 X 
kf jFpq
s
.Rn ; w/k D w 2jsq j.'j fO/_ .  /jq jLp .Rn / < 1 (1.118)
j D0

.with the usual modification if q D 1/.

Remark 1.23. These weighted spaces (avoiding local singularities) play a role in the
theory of function spaces and had been studied in detail in [ET96], Chapter 4 (mainly
based on [HaT94a], [HaT94b]) and [T06], Chapter 6, where one finds also references
to the literature. Here we restrict our attention to an extension of Theorem 1.20 to these
spaces. For this purpose one has first to modify the sequence spaces in Definition 1.18
using the same notation as there.

Definition 1.24. Let w be the same weight as in Definition 1.22 and let s 2 R, 0 <
p  1, 0 < q  1. Then bpq s
.w/ is the collection of all sequences (1.96) such that
X
1 X  X q=p 1=q
n
k jbpq
s
.w/k D 2j.s p /q w.2j m/p jj;G
m j
p
<1
j D0 G2G j m2Zn
(1.119)
s
and fpq .w/ is the collection of all sequences (1.96) such that
 1=q 
 X jsq 
k jfpq
s
.w/k 
D j
2 w.2 m/ jm j m .  /j
q j;G q n 
jLp .R / < 1 (1.120)
j;G;m

with the usual modifications if p D 1 and/or q D 1.

Remark 1.25. This extends Definition 1.18 to the weighted case and coincides with
[T06], Definition 6.11. We extend now Theorem 1.20 to the weighted case in the same
s
understanding as described there now with respect to the spaces Apq .Rn ; w/ where
either A D B or A D F .

Theorem 1.26. Let Apq s

.Rn ; w/ with A D B or A D F be the spaces as introduced
in Definition 1.22 and let apq
s
.w/ be the related sequence spaces according to Defini-
tion 1.24.
j
(i) Let 0 < p  1, 0 < q  1, s 2 R and let ‰G;m be the wavelets (1.91) based
on (1.87), (1.88) with
u > max.s; p  s/: (1.121)
Let f 2 S 0 .Rn /. Then f 2 Bpq
s
.Rn ; w/ if, and only if, it can be represented as
X
j n=2 j
f D j;G
m 2 ‰G;m ;  2 bpq
s
.w/; (1.122)
j;G;m
 1.3 Periodic spaces on Rn and T n 19

unconditional convergence being in S 0 .Rn / and in any space Bpq


z with  < s
.Rn ; w/
1
z
and w.x/w .x/ ! 0 if jxj ! 1. Furthermore, the representation (1.122) is unique,
 j 
m D2
j;G j n=2
f; ‰G;m (1.123)

and ˚  
j
I W f 7! 2j n=2 f; ‰G;m (1.124)
is an isomorphic map of s
Bpq .Rn ; w/ onto bpqs
.w/. If, in addition, p < 1, q < 1,
˚ j
then ‰G;m is an unconditional basis in Bpq .Rn ; w/.
s

(ii) Let 0 < p < 1, 0 < q  1, s 2 R and

u > max.s; pq  s/: (1.125)

Let f 2 S 0 .Rn /. Then f 2 Fpq

s
.Rn ; w/ if, and only if, it can be represented as
X
j n=2 j
f D j;G
m 2 ‰G;m ;  2 fpq s
.w/; (1.126)
j;G;m

unconditional convergence being in S 0 .Rn / and in any space Fpq

z with  < s
.Rn ; w/
1
z
and w.x/w .x/ ! 0 if jxj ! 1. Furthermore, the representation (1.126) is unique
with (1.123) and I in (1.124) is an isomorphic map of Fpqs
.Rn ; w/ onto fpqs
.w/. If, in
˚ j
addition, q < 1, then ‰G;m is an unconditional basis in Fpq .R ; w/.
s n

Proof. This theorem with u as in (1.108) coincides with [T06], Theorem 6.15,
pp. 270–71. The corresponding proof reduces the weighted case to the unweighted
one using localisation and interpolation. But this works also for u as in (1.121) and
(1.125) relying now on Theorem 1.20.

1.3 Periodic spaces on Rn and T n

1.3.1 Definitions and basic properties

Let ˚
T n D x D .x1 ; : : : ; xn / 2 Rn W 0  xj  1 (1.127)
be the n-torus (adapted to wavelets and Z ) where opposite points are identified in the
n

usual way. Then periodic distributions f 2 D 0 .T n / can be represented as

X
f D am e i2 mx ; x 2 T n ; (1.128)
m2Zn

where the Fourier coefficients fam g  C are of at most polynomial growth,

jam j  c .1 C jmj/~ for some c > 0, ~ > 0 and all m 2 Zn : (1.129)

20 1 Spaces on Rn and T n

s
The theory of periodic distributions and related periodic spaces Bpq .T n / and Fpqs
.T n /
has some history which is not the subject of what follows. We rely on [ST87], Chapter 3,
and [T83], Chapter 9, with a reference to [Tri83]. We wish to complement these
considerations by the periodic counterpart of the wavelet isomorphism according to
Theorem 1.20. First we collect what we need in the sequel.
We extend f 2 D 0 .T n / given by (1.128) with (1.129) from D 0 .T n / to S 0 .Rn /,
X
extper f D am e i2 mx ; x 2 Rn : (1.130)
m2Zn

Then extper f 2 S 0 per .Rn /, where

˚
S 0 .Rn / D h 2 S 0 .Rn / W h.   k/ D h for all k 2 Zn
per
(1.131)

are the periodic distributions on Rn . Recall that h 2 S 0 per .Rn / if, and only if, h can be
represented as
X
hD am e i2 mx ; x 2 Rn ; (1.132)
m2Zn

with (1.129) for some c > 0 and ~ > 0. We have

X  
.F h/./ D am F e i2xm ./
m2Zn
X
D am .F 1/.  2 m/ (1.133)
m2Zn
X
D .2/n=2 am ı2 m ;
m2Zn

where ı2 m is the ı-distribution with respect to the off-point 2 m. Let f'j g be the
same dyadic resolution of unity as in Definition 1.1. Then
X
'j hO D am .2/n=2 'j .2 m/ ı2 m (1.134)
m2Zn

and X
 _
'j hO .x/ D am 'j .2 m/ e i2 mx : (1.135)
m2Zn

Let
w .x/ D .1 C jxj2 /=2 ; 2 R; (1.136)
be an admitted weight according to Definition 1.22 (with ˛ D j j) and let Apq
s
.Rn ; w /
be the corresponding spaces. Let

.Rn ; w / \ S 0
per
s;per
Apq .Rn ; w / D Apq
s
.Rn /: (1.137)
 1.3 Periodic spaces on Rn and T n 21

If h 2 S 0 per .Rn / is given by (1.132) then one has by (1.135) and (1.117), (1.118) that

kh jBpq
s;per
.Rn ; w /k
X 1  X q 1=q
jsq  n  (1.138)
D 2 w .x/ am 'j .2 m/ e i2 mx
jLp .R /
j D0 m2Zn

and
kh jFpq
s;per
.Rn ; w /k
 X
1 ˇ X ˇq 1=q 
 jsq ˇ i2 mx ˇ
 (1.139)

D w .x/ 2 ˇ am 'j .2 m/ e ˇ n 
jLp .R /:
j D0 m2Zn

These spaces are non-trivial if, and only if, w 2 Lp .Rn /.

s
The spaces Apq .T n / will be introduced in the standard way where we refer for
further details to the above literature, especially to [ST87], Chapter 3.
Definition 1.27. Let ' D f'j g be again the above dyadic resolution of unity. Let s 2 R
and 0 < q  1.
(i) Let 0 < p  1. Then
˚
s
Bpq .T n / D f 2 D 0 .T n / W kf jBpq
s
.T n /k' < 1 ; (1.140)
X
1  X q 1=q
 
kf jBpq
s
.T n /k' D 2jsq  am 'j .2 m/ e i2 mx jLp .T n / (1.141)
j D0 m2Zn

where f is given by (1.128).

(ii) Let 0 < p < 1. Then
˚
s
Fpq .T n / D f 2 D 0 .T n / W kf jFpq
s
.T n /k' < 1 ; (1.142)
 1 ˇq 1=q 
 X jsq ˇˇ X i2 mx ˇ

kf jFpq
s
.T n /k' 
D 2 ˇ am 'j .2 m/ e ˇ n 
jLp .T /
j D0 m2Zn
(1.143)
where f is given by (1.128).
Remark 1.28. This coincides with [ST87], Definition, pp. 162–63. Basic assertions
parallel to the corresponding theory on Rn may be found in [ST87], Section 3.5. In
particular, the above spaces are independent of admitted '. Furthermore,

Wpk .T n / D Fp;2
k
.T n /; k 2 N0 ; 1 < p < 1; (1.144)

are the usual periodic Sobolev spaces and

C s .T n / D B11
s
.T n /; s > 0; (1.145)
22 1 Spaces on Rn and T n

are the usual periodic Hölder–Zygmund spaces. All classical spaces on Rn of Sobolev–
Besov type as discussed in Remark 1.2 have periodic counterparts and related intrinsic
norms. A corresponding list may be found in [ST87], Section 3.5.4, pp. 167–69. By the
support properties of 'j the sums over Zn in (1.141), (1.143) and in (1.138), (1.139)
are trigonometrical polynomials and the corresponding quasi-norms make sense. They
are connected by the extension operator extper from D 0 .T n / into S 0 per .Rn / according
to (1.130).
Theorem 1.29. Let s 2 R, 0 < p  1 .p < 1 for the F -spaces/, 0 < q  1. Let
s;per
w be as in (1.136) with <  pn . Let Apq .Rn ; w / be as in (1.137). Then ext per is
an isomorphic map from
s
Bpq .T n / onto Bpq
s;per
.Rn ; w / (1.146)
and from
s
Fpq .T n / onto Fpq
s;per
.Rn ; w /: (1.147)

The spaces on the right-hand sides of (1.146), (1.147) are independent of .

Proof. Obviously,
[ ˚
Rn D Ql ; Ql D x 2 Rn W lj  xj  lj C1 : (1.148)
l2Zn

Then one obtains for p < 1,

 X p
 
w .x/ am 'j .2 m/ e i2 mx jLp .Rn /
m2Zn
 X p Z
 n 
D am 'j .2 m/ e i2 mx
jLp .T / .1 C jxj2 /p=2 dx (1.149)
m2Zn Rn
 X p
 
 am 'j .2 m/ e i2 mx jLp .T n / :
m2Zn

This proves (1.146) .modification if p D 1/. Similarly one obtains (1.147).

Remark 1.30. The above isomorphism had been used in [T83], Section 9, and [Tri83]
to transfer properties of function spaces on Rn to the corresponding spaces on T n . The
spaces on the right-hand sides of (1.146), (1.147) are independent of . This suggests
the following definition.
Definition 1.31. Let S 0per .Rn / be as in (1.131). Let s 2 R, 0 < p  1 .p < 1 for
the F -spaces/, 0 < q  1 and <  pn . Then
s;per
Bpq .Rn / D Bpq
s;per
.Rn ; w / D Bpq
s
.Rn ; w / \ S 0per .Rn / (1.150)

and
s;per
Fpq .Rn / D Fpq
s;per
.Rn ; w / D Fpq
s
.Rn ; w / \ S 0per .Rn /: (1.151)
 1.3 Periodic spaces on Rn and T n 23

s;per
1.3.2 Wavelets in Apq .Rn /

Let L 2 N. As a consequence of the multi-resolution analysis one can replace F and

M in (1.87), (1.88) by

F./ D F .2  / 2 C u .R/; M./ D M .2  / 2 C u .R/; u 2 N; (1.152)

L L L L

j
and ‰G;m in (1.91) by

Y
n
 
.j CL/n=2
j;L
‰G;m .x/ D2 Gr 2j CL xr  mr ; G 2 G j ; m 2 Zn : (1.153)
rD1

We choose (and fix) L such that

0;L
supp ‰G;0  fx 2 Rn W jxj < 1=2g; G 2 G 0 D fF; M gn : (1.154)

Since L 2 N is fixed once and for all in Section 1.3 we simplify our notation and write
j j;L
‰G;m instead of ‰G;m given by (1.152), (1.153). Then one has
˚
j
supp ‰G;0 D supp ‰G;0
0
.2j  /  x 2 Rn W jxj < 2j 1 (1.155)

for j 2 N0 and G 2 G j . The adapted wavelet expansions (1.93), (1.94) read now as
follows,
1 X X
X
.j CL/n=2 j
f D j;G
m 2 ‰G;m (1.156)
j D0 G2G j m2Zn

with  
.j CL/n=2 j
m D m .f / D 2
j;G j;G
f; ‰G;m ; (1.157)
based on the newly interpreted orthonormal basis in (1.92). Let
˚
Pjn D m 2 Zn W 0  mr < 2j CL ; j 2 N0 ; (1.158)

be the 2.j CL/n lattice points in 2.j CL/ T n . Then

j
X j X j
‰G;m;per .x/ D ‰G;m .x  l/ D ‰G;mC2j CL l .x/ (1.159)
l2Zn l2Zn

with j 2 N0 and m 2 Pjn are the periodic extensions of the distinguished wavelets
with off-points 2j L m 2 T n . With u 2 N as in (1.87), (1.88) one has
j
‰G;m;per 2 C u .Rn / \ Apq
t;per
.Rn /; t < u; (1.160)

with A D B or A D F according to Definition 1.31. In particular it makes sense to

t;per
ask whether these periodic wavelets may serve as a basis in the spaces Apq .Rn / in
analogy to Theorem 1.26. First we need the corresponding sequence spaces.
24 1 Spaces on Rn and T n

s;per
Definition 1.32. Let s 2 R, 0 < p  1, 0 < q  1. Then bpq is the collection of
all sequences ˚
D j;Gm 2 C W j 2 N0 ; G 2 G ; m 2 Pj
j n
(1.161)
such that
X
1 X  X q=p 1=q
n
k jbpq
s;per
kD 2j.s p /q j m j
j;G p
<1 (1.162)
j D0 G2G j m2Pjn

s
and fpq is the collection of all sequences (1.161) such that
 
 X jsq ˇ ˇq 1=q 
k jfpq
s;per
k D  2 ˇ j;G
m j m .  /ˇ jLp .T n / < 1 (1.163)
j;G;m

with the usual modification if p D 1 and/or q D 1, where j m is the characteristic

function of a cube with the left corner 2j L m and of side-length 2j L (a subcube
of T n ).
Remark 1.33. This is the periodic counterpart of Definitions 1.18, 1.24. With a D b
s;per
or a D f let apq
s
.w/ and apq be the corresponding sequence spaces. Let
j;G
j;G
m D j;G
m D mC2j CL l for j 2 N0 ; m 2 Pjn ; l 2 Zn : (1.164)

Then one obtains

k japq
s;per
k  k japq
s
.w /k; <  pn ; (1.165)
j;per
in the same way as in (1.149). For the corresponding restriction ‰G;m ,
j;per j
‰G;m .x/ D ‰G;m;per .x/; x 2 T n; (1.166)

of the periodic wavelets in (1.159), (1.160), one has the following basic assertion.
Proposition 1.34. Let u 2 N. Then
˚ j;per
‰G;m W j 2 N0 ; G 2 G j ; m 2 Pjn (1.167)

is an orthonormal basis in L2 .T n /.
j
Proof. From the corresponding assertion for ‰G;m in (1.92), now modified by (1.153),
and (1.155) it follows that the functions in (1.167) are orthonormal. Let f 2 L2 .T n /.
Then the restriction to T n of the expansion of

g D extper f 2 L2 .Rn ; w /; <  n2 ; (1.168)

according to Theorem 1.26 is an expansion of f by the system (1.167). (One may also
consult (1.176) below.) Hence the system (1.167) is complete.
 1.3 Periodic spaces on Rn and T n 25

Remark 1.35. This is the periodic counterpart of (1.92)–(1.94). The periodic coun-
terpart of Theorem 1.20 is the subject of Section 1.3.3. First we deal with the periodic
spaces in Rn according to Definition 1.31. Let
Z
 j;per  j;per
f; ‰G;m  D f .x/ ‰G;m .x/ dx (1.169)
Tn

be the usual dual pairing in .D.T n /; D 0 .T n //, appropriately interpreted. (Recall that
j;per
the functions ‰G;m are real.)
j
Theorem 1.36. Let ‰G;m;per with j 2 N0 , G 2 G j and m 2 Pjn be the periodic
wavelets (1.159) in R with u 2 N in (1.160).
n

(i) Let 0 < p  1, 0 < q  1, s 2 R and

u > max.s; p  s/: (1.170)

s;per
Let f 2 S 0 .Rn /. Then f 2 Bpq .Rn / if, and only if, it can be represented as
X
j;G .j CL/n=2 j
f D m 2 ‰G;m;per ; 2 bpq
s;per
; (1.171)
j;G;m

unconditional convergence being in S 0 .Rn / and in any space Bpq


.Rn ; w / with  < s
and <  p . Furthermore, this representation is unique,
n

 j;per 
j;G
m D 2.j CL/n=2 f; ‰G;m  (1.172)

according to (1.166), (1.169), and

˚  j;per 
I W f 7! 2.j CL/n=2 f; ‰G;m  (1.173)
s;per s;per
is an isomorphic map of Bpq .Rn / onto bpq . If, in addition, p < 1, q < 1, then
˚ j s;per
‰G;m;per is an unconditional basis in Bpq .Rn /.
(ii) Let 0 < p < 1, 0 < q  1, s 2 R and

u > max.s; pq  s/: (1.174)

s;per
Let f 2 S 0 .Rn /. Then f 2 Fpq .Rn / if, and only if, it can be represented as
X
j;G .j CL/n=2 j
f D m 2 ‰G;m;per ; 2 fpqs;per
; (1.175)
j;G;m

unconditional convergence being in S 0 .Rn / and in any space Fpq


.Rn ; w / with  < s
and <  p . Furthermore, this representation is unique with (1.172) and I in (1.173)
n

s;per s;per ˚ j
is an isomorphic map of Fpq .Rn / onto fpq . If, in addition, q < 1, then ‰G;m;per
s;per
is an unconditional basis in Fpq .Rn /.
26 1 Spaces on Rn and T n

Proof. Step 1. If f 2 S 0 .Rn / is represented by (1.171) or (1.175) then it follows from

(1.164), (1.165) and Theorem 1.26 that f 2 Apq s
.Rn ; w / with <  pn . Furthermore,
s;per
f 2 S 0per .Rn / and, hence, f 2 Apq .Rn / according to Definition 1.31.
s;per
Step 2. If f 2 Apq .Rn / then one has for j 2 N0 , G 2 G j , m 2 Pjn and l 2 Zn
that .in the usual interpretation/
Z
 j  j
f; ‰G;mC2 j CL l D f .x/ ‰G;m .x  l/ dx
R n
Z
j
D f .x  l/ ‰G;m .x  l/ dx
Rn
Z
D j
f .x/ ‰G;m .x/ dx (1.176)
Rn
Z
j;per
D f .x/ ‰G;m .x/ dx
Tn
 j;per 
D f; ‰G;m 
where we used that f is periodic in Rn and (1.159), (1.166), (1.169). One can justify the
formal calculation in (1.176) by duality as in Rn and a limiting argument approximating
f by trigonometrical polynomials. Expanding now f according to Theorem 1.26 than
one obtains by (1.176), (1.165) the expansions (1.171), (1.175) with (1.172).

s
1.3.3 Wavelets in Apq .T n /
˚ j;per
By Proposition 1.34 the system ‰G;m in (1.167) is an orthonormal basis in L2 .T n /.
By Theorem 1.29,
s
extper Apq .T n / D Apq
s;per
.Rn / (1.177)
is an isomorphic map for all admitted parameters. This gives the possibility to transfer
Theorem 1.36 from the periodic spaces on Rn to the corresponding spaces on T n .
˚ j;per
Theorem 1.37. Let ‰G;m be the orthonormal basis in L2 .T n / according to Propo-
sition 1.34 with u 2 N as in (1.160).
(i) Let 0 < p  1, 0 < q  1, s 2 R, and
u > max.s; p  s/: (1.178)
Let f 2 D 0 .T n /. Then f 2 Bpq
s
.T n / if, and only if, it can be represented as
X
j;G .j CL/n=2 j;per
f D m 2 ‰G;m ; 2 bpq
s;per
; (1.179)
j;G;m

unconditional convergence being in D 0 .T n / and in any space Bpq


.T n / with  < s.
Furthermore, this representation is unique,
.j CL/n=2
 j;per 
m D2
j;G
f; ‰G;m  ; (1.180)
 1.3 Periodic spaces on Rn and T n 27

and ˚  j;per 
I W f 7! 2.j CL/n=2 f; ‰G;m  (1.181)
s;per
is an isomorphic map of Bpqs
.T n / onto bpq . If, in addition, p < 1, q < 1, then
˚ j;per
‰G;m is an unconditional basis in Bpq s
.T n /.
(ii) Let 0 < p < 1, 0 < q  1, s 2 R, and

u > max.s; pq  s/: (1.182)

Let f 2 D 0 .T n /. Then f 2 Fpq

s
.T n / if, and only if, it can be represented as
X
2.j CL/n=2 ‰G;m ;
j;per
f D j;G
m 2 fpq
s;per
; (1.183)
j;G;m

unconditional convergence being in D 0 .T n / and in any space Fpq


.T n / with  < s.
Furthermore, this representation is unique with (1.180) and I in (1.181) is an iso-
s;per ˚ j;per
morphic map of Fpq s
.T n / onto fpq . If, in addition, q < 1, then ‰G;m is an
unconditional basis in Fpq
s
.T n /.
Remark 1.38. As mentioned this assertion is an immediate consequence of Theo-
rem 1.36. We used the same notation as there. The idea to periodise wavelet bases on
R to construct (periodic) wavelet bases on the 1-torus T , preferably for L2 .T / appears
in the literature several times. We refer to [Dau92], pp. 304–05, [Mal99], Section 7.5.1,
pp. 282–83, and [Woj97], Section 2.5, pp. 38–42. Although there is a huge literature
about periodic wavelets we could not find papers dealing with wavelet bases in the
above setting. We followed here essentially [Tri07c].
Chapter 2
Spaces on arbitrary domains

2.1 Basic definitions

2.1.1 Function spaces

We fix some notation. Let  be an arbitrary domain in Rn . Domain means open set.
Then Lp ./ with 0 < p  1 is the standard quasi-Banach space of all complex-
valued Lebesgue measurable functions f in  such that
Z 1=p
kf jLp ./k D jf .x/j dx
p
(2.1)


(with the natural modification if p D 1) is finite. As usual, D./ D C01 ./ stands
for the collection of all complex-valued infinitely differentiable functions in Rn with
compact support in . Let D 0 ./ be the dual space of all distributions in . Let
g 2 S 0 .Rn /. Then we denote by gj its restriction to ,

gj 2 D 0 ./ W .gj/.'/ D g.'/ for ' 2 D./: (2.2)

With A D B or A D F the spaces Apq

s
.Rn / have the same meaning as in Definition 1.1
Definition 2.1. Let  be an arbitrary domain in Rn with  6D Rn and let

0 < p  1; 0 < q  1; s 2 R; (2.3)

with p < 1 for the F -spaces.

(i) Then
˚
s
Apq ./ D f 2 D 0 ./ W f D gj for some g 2 Apq
s
.Rn / ; (2.4)

kf jApq
s
./k D inf kg jApq
s
.Rn /k; (2.5)
where the infimum is taken over all g 2 Apq
s
.Rn / with gj D f .
(ii) Let ˚
AQpq
s x D f 2 Apq
./ s
.Rn / W supp f  x : (2.6)
Then ˚
AQpq
s
./ D f 2 D 0 ./ W f D gj for some g 2 AQpq
s x ;
./ (2.7)

kf jAQpq
s
./k D inf kg jApq
s
.Rn /k (2.8)

where the infimum is taken over all g 2 AQpq

s x with gj D f .
./
 2.1 Basic definitions 29

Remark 2.2. Part (i) is the usual definition of Apq

s
./ by restriction. The spaces
Qs x s n
Apq ./ are closed subspaces of Apq .R /. One has an one-to-one correspondence
between AQpq
s x and AQpq
./ s
./, written in a somewhat sloppy way as

AQpq
s x D AQpq
./ s
./; (2.9)
if, and only if,
fh 2 Apq
s
.Rn / W supp h  @g D f0g: (2.10)
This is the case if  is a bounded Lipschitz domain, 0 < p; q  1 and s > p , (1.32),
since j@j D 0 and Apq .Rn /  Lloc Qs
1 R /. The spaces Apq ./ with
s n
s < 0 will be not
1 
of interest for us. We remark that (2.10) does not hold if s < n p  1 because ı
s
and its translates are elements of Apq .Rn /. But for general domains  it may happen
that (2.10) is not valid even for spaces Apq s
.Rn / with 1  p; q  1 and s > 0. We
give an example looking for spaces on domains  with j j > 0, where D @. Let
frl W l 2 Ng be the set of all rational numbers with 0 < rl < 1 and let Il be open
intervals centred at rl such that Il  .0; 1/. Let
1
[ 1
X
D Il with jIl j < 1: (2.11)
lD1 lD1

Then
1
[
D Œ0; 1 n Il with j j > 0: (2.12)
lD1
Let
1 < p < 1; 1  q  1; 0 < s < 1=p: (2.13)
Then l 2 s
Fpq .R/where l is the characteristic function of Il and one obtains by
s
the homogeneity properties of Fpq .R/ according to [T01], Corollary 5.16, p. 66 (also
discussed in Section 2.2 below, especially (2.46)) that
1
kl jFpq
s
.R/k  jIl j p s ; l 2 N: (2.14)
Let  be the characteristic function of . Then one has by (2.14) that
1
X 1
k jFpq
s
.R/k  c jIl j p s < 1 (2.15)
lD1

if the jIl j are chosen sufficiently small. Hence  2 Fpqs

.R/ and also  2 Fpqs
.R/.
This disproves (2.10). We refer in this context also to [T06], Section 1.11.6, where we
discussed assertions of type (2.9) in the case of bounded Lipschitz domains.
Remark 2.3. Definition 2.1 covers as special cases of Bpq s s
./ and Fpq ./ the corre-
sponding restrictions of the classical spaces discussed in Remark 1.2. This is obvious
for the Lebesgue spaces
Lp ./ D Fp;2
0
./; 1 < p < 1: (2.16)
30 2 Spaces on arbitrary domains

But for the other classical Sobolev–Besov spaces some care is necessary, in particular
in connection with intrinsic norms. This is closely related to the extension problem
which is the subject of Chapter 4. In case of the classical Sobolev spaces Wpk ./ we
refer to Definition 2.53 and Remark 2.54 below where we discuss these questions in
some detail. Otherwise we remind the reader of some intrinsic norms for the classical
Besov spaces
s
Bpq ./; 1  p  1; 1  q  1; s > 0; (2.17)
in bounded Lipschitz domains  (formally introduced in Definition 3.4 below). Let
lh f be the differences in Rn according to (1.21) where l 2 N and h 2 Rn . Let for
x 2 ,
(
.lh f /.x/ if x C kh 2  for k D 0; : : : ; l,
.lh; f /.x/ D (2.18)
0 otherwise;
s
be the differences adapted to . Then Bpq ./ with p, q, s as in (2.17) can be equiva-
lently normed by
Z 1 1=q
 sq q dt
kf jBpq ./km D kf jLp ./kC
s
t sup kh; f jLp ./k
m
(2.19)
0 jhjt t

where s < m 2 N. This is the counterpart of (1.24) for bounded Lipschitz domains
. There is also an appropriate extension of this assertion to
 
s
Bpq ./; 0 < p; q  1; s > n p1  1 C : (2.20)

Details and references may be found in [T06], Theorem 1.118, Remark 1.119, pp. 74–76.

2.1.2 Wavelet systems and sequence spaces

Recall that (arbitrary) open sets  in Rn are called domains. We always assume that
 6D Rn (hence D @ 6D ;). If 1 and 2 are two sets in Rn then
˚
dist. 1 ; 2 / D inf jx 1  x 2 j W x 1 2 1 ; x 2 2 2 : (2.21)

We rely on the well-known Whitney decomposition as it may be found in [Ste70],

Theorem 3, p. 16; Theorem 1, p. 167, adapted to our needs. Let
0
Qlr  Qlr
1
; l 2 N0 I r D 1; 2; : : : ; (2.22)

be concentric (open) cubes in Rn with sides parallel to the axes of coordinates, centred
at 2l mr for some mr 2 Zn with the respective side-lengths 2l and 2lC1 such that
0
for suitable pairwise disjoint cubes Qlr ,
[
D 0
Qlr and dist.Qlr1
; /  2l if l 2 N; (2.23)
l;r
 2.1 Basic definitions 31

complemented by dist.Q0r 1
; /  c for some c > 0. For adjacent cubes Qlr 0
, Ql00 r 0
0
one may assume that jl  l j  1. We used Whitney decompositions of this type in
[T06], Section 4.2.3 (based on [Tri07a]) and in [Tri07b] to construct so-called wavelet
para-bases in some B-spaces and F -spaces in . The elements of these para-bases
preserve the structure of the Rn -wavelets according to (1.91). But now we wish to
convert these para-bases into bases. But this requires some additional multi-resolution
arguments and local orthonormalisation. As a consequence one is losing the one-to-
one relation between Whitney cubes and the index sets for the related bases. This
suggests to give priority to the bases and to describe their properties and relations to
wavelets (localisation, smoothness, cancellation) in qualitative terms. For this purpose
we introduce the counterparts of the wavelet systems (1.91) and the sequence spaces
in Definition 1.18 in Rn now based on the following approximate lattices Z in .
Let  be an arbitrary domain in Rn with  6D Rn (hence D @ 6D ;) and let for
some positive numbers c1 ; c2 ; c3 ,
˚
Z D xrj 2  W j 2 N0 I r D 1; : : : ; Nj (2.24)

x D N [ f1g such that

where Nj 2 N

jxrj  xrj0 j  c1 2j ; j 2 N0 ; r 6D r 0 ; (2.25)

and
[
Nj
  
dist B xrj ; c2 2j ;  c3 2j ; j 2 N0 : (2.26)
rD1

Here B.x; %/ is a ball centred at x 2 Rn and of radius % > 0. It is always assumed that
the positive numbers c1 ; c2 ; c3 are sufficiently small in dependence on  such that for
any j 2 N0 there are points xrj with (2.24)–(2.26) (as suggested by Nj 2 N). x Let for
j 2 N0 and G 2 G 0 D fF; M gn ,

Y
n
 
.j CL/n=2
j;L
‰G;m .x/ D2 Gr 2j CL xr  mr ; m 2 Zn ; (2.27)
rD1

be as in (1.153), based on the dilated real compactly supported Daubechies wavelets

(1.152),
L
F D F .2  /
L
2 C u .R/; L
M D M .2  /
L
2 C u .R/; u 2 N; (2.28)

and (1.88), where we choose L 2 N such that " > 0 in

supp L
F  ."; "/; supp L
M  ."; "/ (2.29)

is small (as specified later on). Let Fx D fF; : : : ; F g 2 fF; M gn .

32 2 Spaces on arbitrary domains

Definition 2.4. Let  be an arbitrary domain in Rn with  6D Rn and let Z be as

in (2.24)–(2.26). Let L 2 N and u 2 N as in (2.27), (2.28), and (1.88). Let K 2 N,
D > 0 and c4 > 0. Then
˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj where Nj 2 N x (2.30)

with  
supp ˆjr  B xrj ; c2 2j ; j 2 N0 ; (2.31)

B.xrj ; c2 2j / as in (2.26), is called a u-wavelet system .with respect to / if it consists

of the following three types of functions.
(i) Basic wavelets
0;L
ˆ0r D ‰G;m for some G 2 fF; M gn ; m 2 Zn : (2.32)

(ii) Interior wavelets

j;L
ˆjr D ‰G;m ; j 2 N; dist.xrj ; /  c4 2j ; (2.33)

for some G 2 fF; M gn , m 2 Zn .

(iii) Boundary wavelets
X
ˆjr D j
dm;m j;L
0 ‰x
F ;m0
; j 2 N; dist.xrj ; / < c4 2j ; (2.34)
jmm0 jK

j
for some m D m.j; r/ 2 Zn and dm;m0 2 R with

X j j
jdm;m0j  D and supp ‰Fj;L
x ;m0  B.xr ; c2 2
j
/: (2.35)
jmm0 jK

Remark 2.5. By construction all wavelets ˆjr are real. Just as for wavelet bases in
Rn and T n only the number u 2 N is of interest. All other numbers L; K; D and the
constants c1 ; : : : ; c4 are technical ingredients. It comes out that they depend on u, but
not on . This may justify the above notation. It is our first aim to show that for any
u 2 N there are u-wavelet systems which are also orthonormal bases in L2 ./. As in
case of Rn and T n one can prove afterwards that the complete orthonormal u-wavelet
systems are also unconditional bases in some function spaces of B-type and F -type.
For this purpose one needs the counterparts of the sequence spaces in Definitions 1.18
and 1.32. Let jr be the characteristic function of the ball B.xrj ; c2 2j / in (2.31).

Definition 2.6. Let  be an arbitrary domain in Rn with  6D Rn and let Z be as in

(2.24)–(2.26). Let s 2 R, 0 < p  1, 0 < q  1. Then bpq s
.Z / is the collection
of all sequences
˚
x
 D jr 2 C W j 2 N0 I r D 1; : : : ; Nj ; Nj 2 N; (2.36)
 2.2 Homogeneity and refined localisation spaces 33

such that
X
1 X
Nj q=p 1=q
n
j.s p /q
k jbpq
s
.Z /k D 2 jjr jp < 1; (2.37)
j D0 rD1

s
and fpq .Z / is the collection of all sequences (2.36) such that

 1 Nj 1=q 
 X X jsq j 
k jfpq
s
.Z /k D  2 jr jr .  /jq jLp ./ < 1 (2.38)
j D0 rD1

.obviously modified if p D 1 and/or q D 1/.

Remark 2.7. There is no need to bother about the seeming ambiguity about the admitted
distributions of the points xrj 2 Z , the radius c2 2j of the balls in (2.31), or the
replacement of the balls by suitable cubes in connection with the characteristic functions
in (2.38). All this can be justified by the related observations in [T06], Section 1.5.3
(at the expense of equivalent quasi-norms, p < 1). Furthermore,
s
fpp .Z / D bpp
s
.Z /; s 2 R; 0 < p  1:

Remark 2.8. The wavelet expansions according to Theorem 1.20 for spaces on Rn and
their periodic counterparts in Theorem 1.36 rely on the lattices 2j Zn . We extended
this theory in [T06], Section 4.2 (based on [Tri07a]) and [Tri07b] to domains in Rn
preserving essentially this lattice structure. We obtained distinguished frames, called
para-bases, where the main part, being a basis in its closed hull, comes from the
above basic wavelets (2.32) and the above interior wavelets (2.33). The remainder
part preserves the lattice structure but spoils the property of being a basis. Now we
wish to convert this remainder part into a basis in its closed hull as indicated at the
beginning of this Section 2.1.2. This is reflected by the boundary wavelets in (2.34).
However it seems to be reasonable to abandon the lattice structure. This may justify
the introduction of u-wavelet systems according to Definition 2.4 now based on the
approximate lattices in Z .

2.2 Homogeneity and refined localisation spaces

2.2.1 Homogeneity

Recall that this book may be considered as the continuation of [T06] (and [T01]). We
rely on the results obtained there. We proved in [T01], Theorem 5.14, pp. 60–61,
for bounded C 1 domains and in [T06], Proposition 4.20, pp. 208–09, for bounded
Lipschitz domains that some spaces Fzpq s
./ according to Definition 2.1 (ii) have the
refined localisation property which will be the subject of Section 2.2.3. This is closely
34 2 Spaces on arbitrary domains

s
connected with local homogeneity assertions for some spaces Fpq as proved first in
s
[T01], Corollary 5.16, p. 66. This had been extended in [CLT07] to some spaces Bpq .
Now we return to this subject in greater generality both for its own sake but also as a
crucial ingredient for the study of refined localisation spaces below.
Let p and pq be the same numbers as in (1.32) and let either

U D fx 2 Rn W jxj < g or U D fx 2 Rn W jxr j < g (2.39)

(balls or cubes) where 0 <  < 1. Let Apq

s
./ and AQpq
s
./ be the spaces as
introduced in Definition 2.1.
Definition 2.9. Let U be either the balls or the cubes according to (2.39). Then
 Fz
s
if 0 < p < 1, 0 < q  1, s > pq ;
pq .U /
Fxpq
s
.U / D 0
Fpq .U / if 1 < p < 1, 1  q < 1, s D 0; (2.40)
s
Fpq .U / if 0 < p < 1, 0 < q  1, s < 0;

and
 Bz
s
if 0 < p  1, 0 < q  1, s > p ;
pq .U /
Bxpq
s
.U / D 0
Bpq .U / if 1 < p < 1, 0 < q  1, s D 0; (2.41)
s
Bpq .U / if 0 < p  1, 0 < q  1, s < 0:

Remark 2.10. One observation about the above restrictions for the domains and the
parameters seems to be in order. In Section 2.2.3 we introduce the refined localisation
s;rloc
spaces Fpq ./ for arbitrary domains  in Rn for the same parameters as in (2.40).
It turns out that
s;rloc
Fpq .U / D Fxpq
s
.U /: (2.42)
This equality with  in place of U is not true in general for arbitrary domains, but for
so-called E-thick domains which include balls and cubes. We refer to Theorem 3.28.
This may justify the above definition including the restrictions of the parameters.
Theorem 2.11. (Homogeneity property) Let U be either the balls or the cubes ac-
cording to (2.39) and let Axpq
s
.U / with A D B or A D F be one of the spaces from
Definition 2.9. Then

kf .  / jAxpq .U1 /k  s p kf jAxpq

n
s s
.U /k (2.43)

where the equivalence constants are independent of  with 0 <   1 and of f 2

Axpq
s
.U /.
Remark 2.12. The case

kf .  / jFxpq .U1 /k  s p kf jFxpq

n
s s
.U /k; 0 <   1; (2.44)

with
0 < p  1; 0 < q  1; s > pq ; (2.45)
 2.2 Homogeneity and refined localisation spaces 35

.where q D 1 if p D 1 and Fx11 s

D Bx11
s
/ is covered by [T01], Corollary 5.16,
p. 66. By Definition 2.1 one can reformulate (2.44) as

supp f  Ux ;
n
kf .  / jFpq
s
.Rn /k  s p kf jFpq
s
.Rn /k; (2.46)

where again 0 <   1 and s; p; q are as in (2.45). This local homogeneity is a

cornerstone of what follows. It will be used to prove the existence of wavelet bases
s;rloc
in refined localisation spaces Fpq ./ with (2.45) for arbitrary domains  in Rn . In
Section 3.2 we consider wavelet bases in the other spaces covered by Definition 2.9.
Afterwards we prove in Section 3.3.2 the homogeneity assertion (2.43) for these spaces.
A direct proof of (2.43) (without using wavelet expansions) for the spaces

Bxpq
s
.U / with 0 < p; q  1; s > p ; (2.47)

has been given recently in [CLT07].

2.2.2 Pointwise multipliers

In [T01], Section 5.17, p. 67, we used the homogeneity property (2.46) with (2.45) for
some adapted multiplier assertions. This can now be extended to all spaces covered by
Theorem 2.11. Let C ~ .Rn / with ~  0 be the usual spaces, normed by
X
kg jC ~ .Rn /k D sup jD ˛ g.x/j if ~ 2 N0 (2.48)
x2Rn
j˛j~

and
X jD ˛ g.x/  D ˛ g.y/j
kg jC ~ .Rn /k D kg jC Œ~ .Rn /k C sup (2.49)
jx  yjf~g
j˛jDŒ~

if 0 < ~ D Œ~ C f~g with Œ~ 2 N0 and 0 < f~g < 1, where the latter supremum
is taken over all x 2 Rn , y 2 Rn with 0 < jx  yj  1. Let s 2 R, 0 < p  1
(p < 1 for the F -spaces), 0 < q  1, and ~ > max.s; p  s/. According to [T92],
Corollary, p. 205, there is a positive number c such that

kgf jApq
s
.Rn /k  c kg jC ~ .Rn /k  kf jApq
s
.Rn /k (2.50)

for all g 2 C ~ .Rn / and all f 2 Apqs

.Rn / where again A D B or A D F (pointwise
multiplier property). This assertion is sharp in the following sense: For any 0 < ~ <
max.s; p  s/ there are functions g 2 C ~ .Rn / which are not pointwise multipliers
s
for Apq .Rn /, [T92], Remark 3, p. 206, with a reference to [T83], Corollary, p. 143.
s
Otherwise pointwise multipliers in the spaces Apq .Rn / are not the subject of this book.
In addition to [T83], [T92] one finds more recent results and related references in [T06],
Section 2.3. We are only interested in a local assertion combining Theorem 2.11 with
(2.50).
36 2 Spaces on arbitrary domains

Theorem 2.13. Let Axpq

s
.U / be one of the spaces according to Definition 2.9 with
0 <   1. Let ~ > max.s; p  s/. Then

kgf jAxpq
s
.U /k  c kg.  / jC ~ .Rn /k  kf jAxpq
s
.U /k (2.51)

where c is a positive constant which is independent of f 2 Axpq

s
.U /,  with 0 <   1,
and
g 2 C ~ .Rn / with supp g  U2 : (2.52)

Proof. By (2.43) and (2.50) we have

kgf jAxpq .U /k   p s kg.  /f .  / jAxpq

n
s s
.U1 /k
(2.53)
 c  p s kg.  / jC ~ .Rn /k  kf .  / jAxpq
n
s
.U1 /k;

where we used some standard properties of function spaces in unit balls or unit cubes
such as extension and multiplication with suitable cut-off functions. Re-transformation
of (2.53) gives (2.51).

2.2.3 Refined localisation spaces

Let  be an arbitrary domain in Rn and let Qlr 0 1

, Qlr be the Whitney cubes according
to (2.22), (2.23) and the other properties described there. Let % D f%lr g be a related
resolution of unity with

supp %lr  Qlr

1
; jD  %lr .x/j  c 2ljj ; x 2 ; 2 Nn0 ; (2.54)

for some c > 0 and

1 X
X
%lr .x/ D 1 if x 2 : (2.55)
lD0 r
s
Let F11 D B11
s
. Let pq be the same number as in (1.32).

Definition 2.14. Let  be an arbitrary domain in Rn with  6D Rn . Let % D f%lr g be

the above resolution of unity. Then
˚
s;rloc
Fpq ./ D f 2 D 0 ./ W kf jFpq
s;rloc
./k% < 1 (2.56)

with
X
1 X 1=p
kf jFpq
s;rloc
./k% D k%lr f jFpq
s
.Rn /kp (2.57)
lD0 r

when
0 < p  1; 0 < q  1; s > pq ; (2.58)
 2.2 Homogeneity and refined localisation spaces 37

.q D 1 if p D 1/ and
X
1 X 1=p
kf jFpq
s;rloc
./k% D k%lr f jFpq
s 1
.Qlr /kp (2.59)
lD0 r

when )
either 1 < p < 1; 1  q < 1; s D 0;
(2.60)
or 0 < p  1; 0 < q  1; s < 0
.again q D 1 if p D 1/.
Remark 2.15. The parameters p, q, s in (2.58), (2.60) are the same as in (2.40)
complemented by p D q D 1, covered by (2.41). Obviously, %lr f with f 2 D 0 ./
1
is extended outside of Qlr by zero. This is justified by (2.54). Then both (2.57) and
(2.59) make sense. In case of (2.57), (2.58) one can replace Fpq s
.Rn / by Fzpq
s
.Qlr1
/,
hence
X1 X 1=p
kf jFpq ./k% D
s;rloc
k%lr f jFzpq
s 1
.Qlr /kp : (2.61)
lD0 r

This follows from Definition 2.1 and Remark 2.2. Since supp %lr f  Qlr 1
one may
s 1 s n
ask whether one can replace Fpq .Qlr / in (2.59), (2.60) by Fpq .R /. But this is not the
case. We return to this point in Remark 2.22 below.
Theorem 2.16. Let  be an arbitrary domain in Rn with  6D Rn . Then the spaces
s;rloc
Fpq ./ according to Definition 2.14 are quasi-Banach spaces. They are independent
of % D f%lr g .equivalent quasi-norms/.
Proof. Recall that (2.57) can be replaced by (2.61). Then the independence of the
s;rloc
spaces Fpq ./ of % D f%lr g follows from the pointwise multiplier assertion in
Theorem 2.13 applied to g D %lr with (2.54) and a possible second resolution of unity.
s;rloc
Afterwards one obtains by standard arguments that Fpq ./ is a quasi-Banach space.

Remark 2.17. We used only the pointwise multiplier property according to Theo-
rem 2.13. In other words, if Axpq s
is one of the spaces covered by Theorem 2.13,
Definition 2.9, then the above arguments can also be applied to spaces quasi-normed
by
X 1 X 1=v
k%lr f j Axpq
s 1
.Qlr /kv ; 0 < v  1: (2.62)
lD0 r

But it is not clear whether these more general spaces are of any interest. In case of the
s;rloc
spaces Fpq ./ as introduced in Definition 2.14 the situation is somewhat different.
The first step was taken in [T01], Theorem 5.14, where we proved for bounded C 1
domains  in Rn that
s;rloc
Fpq ./ D Fzpq
s
./; 0 < p; q  1; s > pq ; (2.63)
38 2 Spaces on arbitrary domains

(q D 1 if p D 1). Here Fzpq s

./ are the same spaces as introduced in Definition 2.1
and commented in Remark 2.2. We called this the refined localisation property of the
spaces Fzpq
s
./. We extended this assertion in [T06], Proposition 4.20, Remark 4.21,
to bounded Lipschitz domains. Now we took the refined localisation property in Defi-
s;rloc
nition 2.14 as the motivation to introduce the spaces Fpq ./ with (2.57), (2.58) for
arbitrary domains. On the one hand we prove (2.63) later on in Proposition 3.10 for
so-called E-thick domains. On the other hand, (2.63) cannot be expected for arbitrary
domains, even if one assumes that j j D 0 with D @ such that one has at least
(2.9), (2.10). For example, if

 6D Rn ; x D Rn
 and j j D 0; (2.64)

then Fzpq s
./ D Fpqs
.Rn /. However if p < 1, q < 1, then D./ is dense in
s;rloc s
Fpq ./. This contradicts (2.63) if s > n=p (then Fpq .Rn / ,! C.Rn /). Later on
we prove in Theorem 3.28 and Remark 3.29 for E-thick domains  that
s;rloc
Fpq ./ D Fpq
s
./ if 0 < p; q  1; s < 0; (2.65)
s;rloc
(q D 1 if p D 1). This justifies the definition of the spaces Fpq ./ with (2.59),
(2.60).
s;rloc
First we characterise Fpq ./ with s > pq in terms of the ball means of differ-
ences in R ,
n

 Z 1=u
n
M
d t;u f .x/ D t jM
h f .x/j
u
dh ; x 2 Rn ; t > 0; (2.66)
jhjt

where 0 < u  1 (usual modification if u D 1) and where M h

are the usual
differences according to (1.21). By [T92], Theorem 3.5.3, p. 194, one can characterise
s
the spaces Fpq .Rn /, s > pq , in terms of these ball means. We refer also to [T06],
Section 1.11.9. Let  be again an arbitrary domain in Rn with  6D Rn , hence
D @ 6D ;, and let
 
ı.x/ D min 1; dist.x; / ; x 2 : (2.67)

For M 2 N let ~ with 0 < ~ < 1 and c > 0 be numbers such that

B.x; M t /  ; dist.B.x; M t /; /  c ı.x/; (2.68)

for all x 2  and all 0 < t  ~ ı.x/. Recall that B.x; r/ stands for a ball centred at
x 2 Rn and of radius r > 0.
s;rloc
Theorem 2.18. Let  be an arbitrary domain in Rn with  6D Rn and let Fpq ./
be the spaces according to Definition 2.14 with

0 < p  1; 0 < q  1 .with q D 1 if p D 1/; s > pq : (2.69)

 2.2 Homogeneity and refined localisation spaces 39

(i) Let
n n
max.1; p/ < v  1; s > (2.70)
p v
.which means v D 1 if p D 1/. Then
s;rloc
Fpq ./ ,! Lv ./ (2.71)

.continuous embedding/.
(ii) Let 0 < u < min.1; p; q/ and s < M 2 N in (2.66). Let ~ be as above. Then
s;rloc
f 2 Lv ./ with v as in (2.70) belongs to Fpq ./ if, and only if,
 Z 1=q 
 ~ı.  /
dt 
 sq
f .  /q jLp ./ s
 C kı .  /f jLp ./k < 1
M
 t d t;u (2.72)
0 t

.equivalent quasi-norms/.

Proof. Recall the well-known continuous embedding

s
Fpq .Rn / ,! Lv .Rn /; s n
p
>  nv : (2.73)

Then (2.71) follows from (2.57), p  v, and an obvious counterpart of (2.57) for
Lv ./. Part (ii) is essentially covered by [T01], Corollary 5.15, p. 66, and the under-
lying proof.

Remark 2.19. The rather long and complicated proof of (2.63) for bounded C 1
domains in Rn was just based on characterisations of Fpq s
.Rn / in terms of ball means
and one obtained equivalent quasi-norms of type (2.72) as a by-product. Now we
s;rloc
defined Fpq ./ via refined localisation. Then the corresponding arguments in [T01]
can be used without changes. As indicated above we prove later on (2.63) for the larger
class of E-thick domains by means of wavelet bases.

Corollary 2.20. Let  be an arbitrary domain in Rn with  6D Rn and let

k;rloc
Wpk;rloc ./ D Fp;2 ./; k 2 N; 1 < p < 1: (2.74)

Then Wpk;rloc ./ is the collection of all f 2 Lp ./ such that

X
kf jWpk;rloc ./k D kı kCj˛j D ˛ f jLp ./k
j˛jk
X (2.75)
 kD ˛ f jLp ./k C kı k f jLp ./k
j˛jDk

is finite .equivalent norms/.

40 2 Spaces on arbitrary domains

Proof. This follows by elementary reasoning from (2.57) and the well-known Paley–
Littlewood assertion
k
Fp;2 .Rn / D Wpk .Rn /; k 2 N; 1 < p < 1: (2.76)

Remark 2.21. Some thirty years ago we dealt in [T78], Chapter 3, with problems of
this type for (fractional) Sobolev spaces and Besov spaces.
Remark 2.22. Finally we return to the question why one cannot replace Fpqs 1
.Qlr / in
s n 1
(2.59) by Fpq .R / although %rl f has compact support in Qlr (as we did erroneously
in [T01], Section 5.20, p. 68). This can be reduced to a corresponding homogeneity
s
problem (2.43) for spaces Apq with s < 0. Let

.x/ > 0 if x 2 U1 D fy W jyj < 1g and supp D US1 ; (2.77)

be a C 1 bump function in Rn . Let

n
s;p
j .x/ D 2j.s p / .2j x/; j 2 N0 : (2.78)

With U j D U2j one obtains by (2.43) for 0 < p; q  1,

s;p
k j jBpq
s
.Rn /k  k jBpq
s
.Rn /k  1 if s > p (2.79)

and
s;p
k j jBpq
s
.U j /k  k jBpq
s
.Rn /k  1 if s < 0: (2.80)
We used the B-counterpart of (2.46). If s < 0 then one can rely on equivalent quasi-
s
norms for the spaces Bpq .Rn / in terms of local means with non-negative kernels as
may be found in [T06], Corollary 1.12, p. 11. One finds (after some calculations) that

k s;p
j jBpq
s
.Rn /k  2j jsj ; j 2 N0 ; s < 0: (2.81)

To provide a better understanding of the behaviour of js;p in Bpq

s
.Rn / and in Bpq
s
.U j /
s n
we first remark that (2.78) are normalised atoms in Bpq .R / if s > p . This is in good
agreement with (2.79). If s < 0 then one needs moment conditions for the atoms. To
obtain (2.80) one can complement js;p in U 2j nU j by suitable functions generating
the needed moment conditions. This explains (2.80). However in case of (2.81) there
is no such possibility and the outcome shows that js;p are no longer normalised atoms
s
in Bpq .Rn / with s < 0.
 2.3 Wavelet para-bases 41

2.3 Wavelet para-bases

2.3.1 Some preparations

Recall that this book should be considered as the continuation of [T06] and that we rely
on the assertions obtained there. In [T06], Section 4.2, we dealt with wavelet para-bases
in some spaces Apq s
./ and AQpq
s
./ according to Definition 2.1 in bounded Lipschitz
domains in R (based on [Tri07a]). The starting point was the refined localisation
n

property
Fzpq
s
./ D Fpq
s;rloc
./; 1 < p; q  1; s > 0; (2.82)
s;rloc
(with q D 1 if p D 1), [T06], Proposition 4.20, p. 208, where Fpq ./ has
the same meaning as in Definition 2.14. But afterwards we needed only this refined
localisation property combined with the homogeneity (2.43) for the spaces in the first
line of (2.40), hence (2.44), (2.45). This makes clear that there is no problem to extend
s;rloc
the considerations in [T06], Section 4.2, to the refined localisation spaces Fpq ./
according to (2.56)–(2.58). Nevertheless we give a careful description which will be
the starting point for the construction of wavelet bases according to Definition 2.4 in
spaces of this type.
Again let  be an arbitrary domain in Rn with  6D Rn (hence D @ 6D ;). We
modify the Whitney decomposition (2.22), (2.23) as follows. Let
0
Qlr  Qlr
1
 Qlr
2
 Qlr ; l 2 N0 I r D 1; 2; : : : ; (2.83)
be concentric open cubes in Rn with sides parallel to the axes of coordinates centred
at 2l mr for some mr 2 Zn with the respective side-lengths 2l , 5  2l2 , 6  2l2 ,
2lC1 . According to the Whitney decomposition there are disjoint cubes Qlr 0
of this
type such that [
D 0
Qlr and dist .Qlr ; /  2l (2.84)
l;r

if l 2 N and r D 1; 2; : : : , complemented by dist.Q0r ; /  c for some c > 0. We

may assume that jl  l 0 j  1 for any two admitted cubes Qlr1
, Ql10 r 0 , having a non-
empty intersection. We follow the construction of wavelet para-bases given in [T06],
Section 4.2.4, now for arbitrary domains. We adapt the real wavelets in (2.27), (2.28),
based on (1.87), (1.88), by
Y
n
 
.j CL/n=2
j
‰G;m .x/ D2 Ga 2j CL xa  ma ; G 2 fF; M gn ; m 2 Zn ; (2.85)
aD1

where L 2 N0 is fixed such that

j
supp ‰G;m  Qlr if 2j L m 2 Qlr
2
for l 2 N0 and j  l; (2.86)
and
2Lj m 2 Qlr
2 1
if Qlr j
\ supp ‰G;m 6D ; for l 2 N0 and j  l; (2.87)
42 2 Spaces on arbitrary domains

for admitted cubes according to (2.83), (2.84). The number L 2 N0 depends on F ,

M and in particular on u in (1.87), (1.88), but it has nothing to do with  and is
assumed to be fixed. It specifies " in (2.29). Let fF; M gn and fF; M gn be as in
(1.89), (1.90). As said we follow [T06], Section 4.2.4, where one finds further details
and explanations. We will be brief. Let for j 2 N0 ,
Sj;1 D fF; M gn  fm 2 Zn W 2j L m 2 Qlr
2
for some l < j , some rg (2.88)

be the main index set (with S0;1 D ;) and

Sj;2 D fF; M gn  fm 2 Zn W 2j L m 2 Qjr
2
for some rgnSj;1 (2.89)
be the residual index set. Let
1
[ 1
[
S 
DS ;1
[S ;2
; S ;1
D Sj;1 ; S ;2
D Sj;2 ; (2.90)
j D1 j D0

and let ˚ j
‰ 1; D ‰G;m W .j; G; m/ 2 S ;1 (2.91)
be the main wavelet system and
˚ j
‰ 2; D ‰G;m W .j; G; m/ 2 S ;2 (2.92)
j
be the residual wavelet system where ‰G;m are given by (2.85)–(2.87). We discuss
briefly the outcome. The arguments in [T06], Section 4.2.4, are based on the refined
localisation property (2.82), the homogeneity according to Theorem 2.11 and the reduc-
tion to wavelet expansions in Rn as described in Theorem 1.20 for the spaces Apq s
.Rn /
n
and in Section 1.2.1 for L2 .R /. Having these ingredients the quality of the domain 
s;rloc
is irrelevant. This will be specified in Section 2.3.2 for the spaces Fpq ./. But this
applies also to Lp ./ with 1 < p < 1. One has in particular
L2 ./ D L.1/ .2/
2 ./ ˚ L2 ./; (2.93)
with
L.1/
2 ./ D span‰
1;
and L.2/
2 ./ D span‰
2;
: (2.94)
By construction, ‰ 1;
is an orthonormal basis on L.1/
2 ./.
Any element of the main
wavelet system ‰ 1; is orthogonal to any element of the residual wavelet system. The
pairwise scalar products of the elements of ‰ 2; generate a band-limited matrix caused
by the mild overlapping of the supports of elements belonging to adjacent cubes. In
any case, ‰ 2; is locally finite and the cardinal number of elements of ‰ 2; with
supports intersecting a given compact subset of  is finite. It is one of our main goals
to find u-wavelet systems according to Definition 2.4 which are orthonormal bases in
L2 ./. Then we identify the interior wavelets in (2.33) with ‰ 1; and convert ‰ 2; by
a specific orthonormalisation procedure into the basic and boundary wavelets (2.32),
(2.35). But this will be shifted to Section 2.4. First we extend the theory of wavelet
para-bases for the spaces Fzpq
s
./ in bounded Lipschitz domains  in Rn according to
(2.82) to more general spaces. We follow at least partly [Tri07b].
 2.3 Wavelet para-bases 43

s;rloc
2.3.2 Wavelet para-bases in Fpq ./

Let  be an arbitrary domain in Rn and let Qlr 0 1

, Qlr be the Whitney cubes according
to (2.83) (in immaterial modification of (2.22)). Let % D f%lr g be a related resolution
of unity with (2.54), (2.55). We are interested in wavelet expansions for the spaces
s;rloc
Fpq ./ according to Definition 2.14 with (2.56)–(2.58). We rely on the L-version
s
of the wavelet expansion for Fpq .Rn / as described at the beginning of Section 1.3.2
applied to
X1 X
f D %lr f; %lr f 2 Fpq s
.Rn /: (2.95)
lD0 r

We fix L 2 N so that we have (2.86), (2.87) for the wavelets in (2.85). The crucial
point is the wavelet expansion of %lr f according to [T06], Proposition 4.17, p. 206,
based on the homogeneity property (2.43) for the spaces Fxpqs
.U / in the first line in
(2.40). But afterwards one can follow the arguments in [T06] verbatim. One obtains
intrinsic wavelet expansions
X
j n=2 j
f D j;G
m .f / 2 ‰G;m (2.96)
.j;G;m/2S 

s;rloc
for f 2 Fpq ./. We do not repeat the details. But we fix the outcome. Recall that
j
the functions ‰G;m are real. One has
Z
j
m .f / D 2
j;G j n=2
f .x/ ‰G;m .x/ dx; .j; G; m/ 2 S ;1 ; (2.97)
Rn

and
Z
j
j;G
m .f /D2 j n=2
%jm .x/ f .x/ ‰G;m .x/ dx; .j; G; m/ 2 S ;2 ; (2.98)
Rn

where %jm are some real C 1 functions with

supp %jm  Qjr ; jD  %jm .x/j  c 2j jj ; 2 Nn0 ; (2.99)
and f is extended outside of  by zero. Here Qjr has the same meaning as in (2.83)
with r D r.m/ as in (2.89). Let j m be the characteristic function of Qjr with
s;
r D r.m/ in (2.88), (2.89). We need the -version of fpq
s
in Definition 1.18. Let fpq
be the collection of all sequences
˚
m 2 C W .j; G; m/ 2 S
 D j;G 
(2.100)
such that
 X ˇ ˇq 1=q 
 
k jfpq
s;
kD 2jsq ˇj;G
m jm .  / ˇ jLp ./  (2.101)
.j;G;m/2S 

s
is finite. Recall that F11 D B11
s
.
44 2 Spaces on arbitrary domains

Theorem 2.23. Let  be an arbitrary domain in Rn with  6D Rn . Let 0 < p  1,

0 < q  1 .with q D 1 if p D 1/ and pq < s < u 2 N. Let
˚ j
‰G;m W .j; G; m/ 2 S  D ‰ 1; [ ‰ 2; (2.102)

be the above real intrinsic wavelet system (2.91), (2.92) based on (1.87), (1.88) and
s;rloc
(2.85)–(2.87). Let v be as in (2.70), (2.71). Then f 2 Lv ./ is an element of Fpq ./
if,and only if, it can be represented by
X
j n=2 j
f D j;G
m 2 ‰G;m ;  2 fpq s;
: (2.103)
.j;G;m/2S 

Furthermore,
kf jFpq
s;rloc
./k  inf k jfpq
s;
k (2.104)
s;rloc
where the infimum is taken over all representations (2.103). Any f 2 Fpq ./ can
be represented by (2.96) with (2.97), (2.98) and

k.f / jfpq
s;
k  kf jFpq
s;rloc
./k (2.105)

.equivalent quasi-norms/.
Proof. This is an extension of [T06], Theorem 4.22 (i), p. 212, from Fzpq s
./ D
s;rloc
Fpq ./, where  is a bounded Lipschitz domain and where p; q; u D k are re-
stricted by [T06], (4.112), p. 212, to the above situation. Now we rely on the improved
wavelet expansion Theorem 1.20 in Rn , which results in s < u and its L-version men-
tioned above. Otherwise one can follow the arguments in [T06]. In particular, (2.103)
is an atomic representation for f in Rn (no moment conditions are required) and after
multiplication with %lr also for %lr f . One obtains
X
kf jFpq
s;rloc
./kp  k%lr f jFpq
s
.Rn /kp  c k jfpq k :
s; p
(2.106)
l;r

s;rloc
The crucial point is the representability of f 2 Fpq ./ by (2.96), (2.105) with
(2.104) as a consequence. But this is the same as in [T06] (reduction to Rn , and
clipped together via refined localisation).
Remark 2.24. We add a comment about the convergence of (2.103). First we remark
that (2.103) converges unconditionally in S 0 .Rn /. If p < 1 and v < 1 in (2.70),
(2.71) then (2.103) converges absolutely (and hence unconditionally) in Lv ./. If
s;rloc
p < 1, q < 1, then (2.103) converges unconditionally in Fpq ./. If p < 1, q D
;rloc
1, then one has at least unconditional convergence in Fpp ./ with pq <  < s.
Recall that F11s
D C s . If  is bounded and p D q D 1 then (2.103) converges
unconditionally in C ;rloc ./ with 0 <  < s. If  is unbounded and p D q D 1
then one has unconditional convergence in C ;rloc . \ KR / where KR is a ball centred
at the origin and of radius R .! 1/.
 2.3 Wavelet para-bases 45

Remark 2.25. By (2.88)–(2.90) and (2.86) we have for any .j; G; m/ 2 S  that
j
supp ‰G;m  Qlr   (2.107)

for some l  j and some r. Then (2.96)–(2.98) is an intrinsic stable frame repre-
sentation, where stable refers to (2.103), (2.104). What is meant by a frame will also
be discussed later on in Remark 5.11 in greater detail. As mentioned in connection
with (2.93), (2.94) the main part ‰ 1; with .j; G; m/ 2 S ;1 consists of pairwise
orthogonal Daubechies wavelets. It is an unconditional basis (p < 1, q < 1) on the
s;rloc;1
complemented subspace Fpq ./ spanned by ‰ 1; ,
s;rloc
Fpq ./ D Fpq
s;rloc;1
./ ˚ Fpq
s;rloc;2
./ (2.108)

where the corresponding projection operators can be constructed by (2.96) (with S ;1
in place of S  ) and (2.105). As for the residual part we have the orthogonality (2.93),
(2.94) and
s;rloc;2
Fpq ./  C u ./: (2.109)
s;rloc;2
Furthermore, Fpq ./ is locally finite-dimensional (which means that for any
j
bounded domain ! with ! x  , there are only finitely many functions ‰G;m with
j
.j; G; m/ 2 S ;2 such that supp ‰G;m \! x 6D ;). This may justify calling (2.102) a
para-basis (with the usual notational caution if p < 1, q D 1 or even p D q D 1).

Remark 2.26. The proof of Theorem 1.20 is based on the observation that the Dau-
bechies wavelets may serve simultaneously as atoms and as kernels of local means
satisfying the needed moment conditions. This is also the case for the above wavelets
j
‰G;m if .j; G; m/ 2 S ;1 , hence G 2 fF; M gn (main wavelet system (2.91)) because
we have always (1.88). But this does not apply to the scaled down scaling functions
‰Fjx ;m with G D Fx D .F; : : : ; F /. As a consequence one has now only a para-basis.
However it will be one of the main aims of what follows to convert (2.102) into an
orthonormal basis which is a u-wavelet system according to Definition 2.4. Then the
basic wavelets (2.32) and the interior wavelets (2.33) coincide with the above main
0
wavelet system ‰ 1; in (2.91), complemented by some ‰G;m , whereas the boundary
wavelets (2.34), (2.35) are obtained by local orthonormalisation from the remaining
elements of the residual wavelet system ‰ 2; in (2.92).

Remark 2.27. One has always the continuous embedding

s;rloc
Fpq ./ ,! Fzpq
s
./ (2.110)

for p, q, s as in Theorem 2.23, where Fzpqs

./ are the spaces introduced in Defini-
tion 2.1. This follows from (2.103), (2.104) and the atomic arguments in the proof of
Theorem 2.23. It will be one of the main points of what follows to show that
s;rloc
Fpq ./ D Fzpq
s
./; 0 < p; q  1; s > pq ; (2.111)
46 2 Spaces on arbitrary domains

(q D 1 if p D 1) for huge classes of domains, the so-called E-thick domains,

including bounded C 1 domains and bounded Lipschitz domains. We mentioned a
special case of (2.111) in (2.82). This observation will pave the way to develop a
theory of wavelet bases for large classes of the spaces AQpq
s s
./ and Apq ./ according
to Definition 2.1 in E-thick domains. This is the subject of Chapter 3.

2.3.3 Wavelet para-bases in Lp ./, 1 < p < 1

One may ask whether Theorem 2.23 can be extended to other refined localisation spaces
according to Definition 2.14 having the homogeneity property (2.43) (which are two
of the main ingredients of the above proof with a reference to [T06]). But this causes
some trouble. To obtain (2.106) we used that (2.103) is an atomic decomposition in Rn
(after appropriate normalisation). Since s > pq no moment conditions for the atoms
are needed. But if s  0 then Theorem 1.7 requires some moment conditions (1.31).
j
The elements ‰G;m 2 ‰ 1; of the main wavelet system (2.91) fit in this scheme, but
not the elements ‰Fjx ;m 2 ‰ 2; with j 2 N of the residual wavelet system (2.92)
briefly mentioned in Remark 2.26. But there is a remarkable exception.
Based on the well-known Paley–Littlewood property
0
Fp;2 .Rn / D Lp .Rn /; 1 < p < 1; (2.112)
one has by Definitions 2.1, 2.14 that
0
Fp;2 ./ D Fzp;2
0 0;rloc
./ D Fp;2 ./ D Lp ./; 1 < p < 1; (2.113)
0;
for arbitrary domains  with  6D Rn . We need fp;2 according to (2.100), (2.101)
with s D 0, q D 2.
Theorem 2.28. Let  be an arbitrary domain in Rn with  6D Rn . Let 1 < p < 1
and u 2 N. Let ˚ j
‰G;m W .j; G; m/ 2 S  D ‰ 1; [ ‰ 2; (2.114)
be the same intrinsic wavelet system as in (2.102) based on (1.87), (1.88) and
(2.85)–(2.87). Then Lp ./ is the collection of all locally integrable functions f
.either with respect to Rn or to / which can be represented by
X
j n=2 j 0;
f D j;G
m 2 ‰G;m ;  2 fp;2 : (2.115)
.j;G;m/2S 

Furthermore,
0;
kf jLp ./k  inf k jfp;2 k; (2.116)
where the infimum is taken over all representations (2.115). Any f 2 Lp ./ can be
represented by (2.96) with (2.97), (2.98) and
0;
k.f / jfp;2 k  kf jLp ./k (2.117)
.equivalent norms/.
 2.3 Wavelet para-bases 47

Proof. Obviously, Lp ./ has the refined localisation property, hence (2.106) with Lp
s
in place of Fpq on the left-hand side, and the homogeneity (2.43) with Lp instead
x0
of Apq . Then one obtains (2.96) with (2.97), (2.98) and (2.117). Let f be given by
s
(2.115). In contrast to the atomic representation (2.103) in Fpq .Rn / with s > pq
where no moment conditions for the corresponding atoms are required the situation is
now different. According to Theorem 1.7 one needs for atomic representations
1 X
X
f D j m aj m (2.118)
j D0 m2Zn

in Lp .Rn / D Fp;2
0
.Rn / with 1 < p < 1 first moment conditions
Z
aj m .x/ dx D 0; j 2 N; m 2 Zn : (2.119)
Rn

We split (2.115) according to (2.88)–(2.92) into the main part and the residual part
X X
f D 2j n=p j;G
m aj m C
G
2j n=p j;G G
m aj m
.j;G;m/2S ;1 .j;G;m/2S ;2 (2.120)
D f1 C f2

with
1 1
ajGm D 2j n. p  2 / ‰G;m
j
: (2.121)
j
If ‰G;m 2 ‰ 1; then ajGm are normalised atoms in Lp .Rn / D Fp;2 0
.Rn / satisfying
(2.119) with ajGm in place of aj m . We split (2.100) naturally in 1 and 2 ,
˚
l D j;G
m 2 C W .j; G; m/ 2 S
;l
; l D 1; 2: (2.122)

Then it follows by Theorem 1.7 that

 
kf1 jLp ./k D kf1 jLp .Rn /k  c 1 jfp;2
0; 
: (2.123)

Furthermore, for the residual part f2 one has

X  2 0; p  2 0; p
kf2 jLp ./kp  c 2j n jj;G     jf  ;
m j   jfpp
p
p;2
.j;G;m/2S ;2
(2.124)
where we used (2.101) and the structure of S ;2 discussed in Remark 2.25. This proves
 
kf jLp ./k  c  jfp;2 0; 
: (2.125)

Together with (2.117) one has (2.116).

48 2 Spaces on arbitrary domains

Remark 2.29. In other words, one obtains for Lp ./, 1 < p < 1, in arbitrary
domains  in Rn with  6D Rn intrinsic wavelet para-bases of any smoothness u 2 N
in terms of the Littlewood–Paley assertion (2.117). In particular one has common
wavelet para-bases for all spaces covered by Theorems 2.23 and 2.28 as long as s < u.
Later one we convert these wavelet para-bases into orthonormal bases which are u-
wavelet systems according to Definition 2.4. We refer to Section 2.4.2.
Remark 2.30. One may ask what happens if (2.114) is not based on the smooth
functions F , M according to (1.87), (1.88), but on related Haar functions as described
in [T06], Proposition 1.54, p. 28. Then we obtain suitably adapted systems of type Hn
as in [T06], Theorem 1.58, p. 29. By this theorem and the above considerations one has
intrinsic Haar bases (not only para-bases) for Lp ./. This observation can be extended
s;rloc s;rloc
to those spaces Fpp ./ D Bpp ./ which are covered both by Theorem 2.23 and
[T06], Theorem 1.58, p. 29 , in particular
 
n 1
< p < 1; p < s < min 1; : (2.126)
nC1 p
We return to this point later on and refer to Section 2.5.1.

2.4 Wavelet bases

2.4.1 Orthonormal wavelet bases in L2 ./

Again let  be an arbitrary domain in Rn with  6D Rn . Recall that domain means

open set. So far we have by Theorems 2.23 and 2.28 wavelet para-bases
˚ j
‰G;m W .j; G; m/ 2 S  D ‰ 1; [ ‰ 2; (2.127)
s;rloc
in some refined localisation spaces Fpq ./ and in Lp ./ with 1 < p < 1. Here
1; 2;
‰ and ‰ are the main wavelet system and the residual wavelet system according
to (2.88)–(2.92), based on (2.85)–(2.87). These two systems are orthogonal to each
other and they span L2 ./,
L2 ./ D L.1/ .2/
2 ./ ˚ L2 ./; L.l/
2 ./ D span ‰
l;
; (2.128)
(2.93), (2.94), where l D 1; 2. Furthermore, ‰ 1; is an orthonormal basis of L.1/2 ./.
A few words about the elements of the residual set ‰ 2; have been said after (2.94) and
in Remark 2.25. There is the temptation to apply Schmidt’s very classical orthonor-
malisation as it may be found in textbooks about Hilbert spaces, for example [Yos80],
Chapter III, Section 5, p. 88, or [Tri92], pp. 86–87, to ‰ 2; . This will be of some
help for us locally. But globally it is of little use beyond L2 ./ since it destroys the
localisation (support property) of its elements
Y
n
 
j
‰G;m .x/ D 2.j CL/n=2 Ga 2j CL xa  ma ; .j; G; m/ 2 S ;2 : (2.129)
aD1
 2.4 Wavelet bases 49

To circumvent these difficulties we use some (one-dimensional) multiresolution ar-

guments, the product structure of (2.129) and a further refinement of the Whitney
decomposition (2.83). Let now
0
Qlr  Qlr
1
 Qlr
2
 Qlr
3
 Qlr ; l 2 N0 ; r D 1; 2; : : : ; (2.130)

be concentric open cubes in Rn with sides parallel to the axes of coordinates centred at
2l mr for some mr 2 Zn with the respective side-lengths 2l ; 5  2l2 ; 6  2l2 ; 7 
2l2 ; 2lC1 . In other words, we insert now the cubes Qlr 3
in the construction at the
0
beginning of Section 2.3.1. Otherwise, Qlr are the same pairwise disjoint cubes as
there with (2.84). Recall that jl  l 0 j  1 for two cubes Qlr
1
, Ql10 r 0 having a non-empty
intersection. We fix now L 2 N0 in (2.129) such that we have (2.86) with Qlr 3
in place
of Qlr , hence
j
supp ‰G;m  Qlr
3
if 2j L m 2 Qlr
2
for l 2 N0 and j  l; (2.131)

whereas (2.87) remains unchanged. This refinement ensures that the supports of the
0 0
wavelets related to cubes Qlr and QlC2;r 0 are disjoint. Instead of this explicit con-
struction one could simply say that L is chosen sufficiently large (but fixed once and
for all).
Definition 2.31. Let  be an arbitrary domain in Rn with  6D Rn and let u 2 N.
Then ˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj x
with Nj 2 N (2.132)
is called an orthonormal u-wavelet basis in L2 ./ if it is both an u-wavelet system
according to Definition 2.4 and an orthonormal basis in L2 ./.
Remark 2.32. In Definition 2.4 we marked not only the required smoothness u 2 N in
the wavelets (2.27) based on (2.28), (2.29) but also L. Now we adopt the position that
L is chosen sufficiently large as described and fixed. This may justify our omission of
L in what follows. Similarly as in (2.91), (2.92) we put
˚
ˆ1; D interior wavelets ˆjr (2.133)

and ˚
ˆ2; D basic and boundary wavelets ˆjr : (2.134)
Theorem 2.33. Let  be an arbitrary domain in Rn with  6D Rn . For any u 2 N
there are orthonormal u-wavelet bases in L2 ./ according to Definition 2.31.
Proof. Step 1. Let
˚ j
‰G;m W .j; G; m/ 2 S  D ‰ 1; [ ‰ 2; (2.135)

be the same wavelet system as in Theorem 2.28 based on (1.87), (1.88) and (2.85)–
(2.87) now modified by (2.131). We have (2.128). Then

ˆ1; D ‰ 1; is an orthonormal basis in L.1/

2 ./: (2.136)
50 2 Spaces on arbitrary domains

Furthermore, L.2/ .1/

2 ./ is orthogonal to L2 ./. Hence we must find an orthonormal
system ˆ2; according to (2.134) such that it spans L.2/ 0
2 ./. The basic wavelets ˆr in
;1
(2.32) are the elements ‰G;m in (2.92), based on (2.89) (recall that S0 D ;). They
0

fit in the above scheme and need not to be considered. Hence it is sufficient to care
for the elements with j 2 N in (2.92). These are boundary elements. In other words
one must find an orthonormalisation of the elements of ‰ 2; with j 2 N resulting in
boundary wavelets according to Definition 2.4 (iii).
Step 2. First we deal with the one-dimensional model case  D .1; 0/. Let Qlr 0
l
in (2.130) be an interval centred at 2 r for some suitable negative integer r and of
side-length 2l . The related residual wavelets according to (2.129) are given by
l
m .x/ D 2.lCL/=2 G .2
lCL
x  m/; m 2 Z; (2.137)
with G D F (scaling function) or G D M (wavelet). Of interest is the case
0
where the right endpoint of Qlr , hence xl;r D 2l r C 2l1 , is the left endpoint of an
0
admitted Whitney interval QlC1;r 0 . Then some functions
 lCL   lC1CL 
G 2 x  m and F 2 x  m0 (2.138)
centred near xl;r may not be orthogonal. (Recall that we always assume that L is large
ensuring that everything is local near xl;r ). The decisive argument comes from the
multiresolution property
 lCL  X G  
G 2 xm D cl;m;t F 2lC1CL x  t ; (2.139)
t2Z

where only terms with

   
supp F 2lC1CL  t  supp G 2lCL  m (2.140)
G
are of interest. Otherwise one has cl;m;t D 0. Hence (2.139) is a finite sum where only
terms with
ˇ l1L ˇ
ˇ2 t  2lL mˇ  c 2Ll ; hence jt  2mj  2c; (2.141)
for some c > 0 are needed. Furthermore, replacing x by 2x in (2.139) it follows that
G
cl;m;t D cm;t
G G
is independent of l. The coefficients cl;m are also translation invariant,
cm;t D cmCm0 ;tC2m0 . Hence all coefficients in (2.139) coincide with finitely many
G G

real numbers. Let Zm be the collection of all t 2 Z with (2.141). Let G D F in

(2.137), (2.139) and let temporarily D F . Then one obtains by (2.137), (2.139)
that
X X 0 00
m D dm;t tlC1 C dm;t tlC1 D mlC1 C mlC1 ;
l
(2.142)
t2Z0m t2Z00
m

0
where Zm D Z0m [ Z00m is decomposed such that mlC1 collects all terms on the right-
00
hand side of (2.139) with (2.140) and tlC1 2 ‰ 2; . The remaining terms mlC1 are
 2.4 Wavelet bases 51

0
not only orthogonal to mlC1 by the orthogonality of the translated scaling functions,
but also to all other terms of ‰ 1; and ‰ 2; not involved in (2.139). Of course, one can
replace G D F in (2.137), (2.139) in the above arguments by G D M . The resulting
00
functions mlC1 are orthogonal to all other functions in ‰  D ‰ 1; [ ‰ 2; , but they
may be not orthogonal among themselves. But now one can apply the orthogonalisation
procedure mentioned above between the formulas (2.128) and (2.129). In this way one
eliminates possible linear dependences (stepping from frames to bases). Then one
obtains functions which are orthogonal among themselves and to all other functions
from ‰  not involved in this local procedure. This can be done at any point 2l r C
2l1 . Together with the unchanged elements of ‰ 2; one obtains an orthonormal
system in
L.2/
2 ./ with  D .1; 0/ (2.143)
which can be identified with (2.134) collecting basic and boundary wavelets according
G
to Definition 2.4 (i), (iii). The above discussion about the coefficients cl;m;t in (2.139)
justifies (2.35). Together with (2.136) one obtains an orthonormal basis in L2 ./
with  D .1; 0/ which is an u-wavelet basis according to Definition 2.31. Since
everything is local this proves also the theorem for arbitrary domains  in R.
Step 3. The corresponding assertion for arbitrary domains  in Rn with n  2
j
can be reduced to the 1-dimensional case. Let ‰G;m with j 2 N be an element of
the residual system ‰ 2; according to (2.129) based on the Whitney cubes in (2.130)
(excluding again the basic wavelets). We assume that the right face with respect to
the x1 -direction, say, x1 D 2l r C 2l1 as in the above 1-dimensional model case,
0
is part of the left face of an admitted Whitney cube QlC1;r 0 . Then one can apply the
1-dimensional orthogonalisation procedure with respect to the x1 -factors according to
j
Step 2. By the product structure of ‰G;m in (2.129) it follows from Fubini’s theorem that
this is also an n-dimensional orthogonalisation. Applying (2.142) to the x1 -direction
one obtains after multiplication with the remaining n  1 directions terms of type
0
lC1
t
l
.x1 / ‰G;m 0 .x /; x 0 D .x2 ; : : : ; xn /: (2.144)
Using again the multiresolution property one obtains the .n  1/-dimensional counter-
part of (2.139), X
0 G;n lC1 0
0 .x / D
l
‰G;m dl;m 0 ;t 0 ‰ x 0 .x /;
F ;t
(2.145)
t 0 2Zn1

with a counterpart of (2.140). Here ‰FlC1 0 x

x ;t .x / with F D .F; : : : ; F / are the same
dilated scaling functions as in Definition 2.4. Inserted in (2.144) one obtains a linear
combination of terms
0
lC1
t .x1 / ‰FlC1
x ;t 0 .x /: (2.146)
However it is not guaranteed that each element in (2.146) is an element in the wavelet
0
expansion with respect to QlC1;r 0 . Decomposing (2.145) as in (2.142) there may
remain linear combinations of terms
0
lC1
t .x1 / ‰FlC1
x ;t 0 .x / (2.147)
52 2 Spaces on arbitrary domains

which are orthogonal to all other functions of ‰ 2; , involved in the above procedure or
not, and also to ‰ 1; . However this undesirable effect can only happen near xv -faces
0 with v D 2; : : : ; n, hence near edges (with respect to x1 and xv ). In order
0
of QlC1;r
to clarify what is going on we assume first n D 2. Of course in the above arguments x1
and x2 can change their roles making clear that the indicated effect can happen only in
corner points. Now one can proceed with the orthogonalisation as described above and
orthogonalise the remaining functions locally around isolated corners. Then they are
orthogonal among themselves and to all other functions in ‰ 2; and ‰ 1; . They have
the desired structure (2.34), (2.35) as boundary wavelets. For n  3 one can apply this
argument iteratively as follows. If the wavelets considered for orthogonalisation are
not near a corner-point of the related cube then there is at least one harmless direction
of coordinates (playing the role of x2 at the beginning of this step). Orthogonalisation
for the remaining .n  1/ directions may be assumed by induction. Finally one can
orthogonalise the remaining functions locally around isolated corners as described
above in case of n D 2. The discussion about the coefficients in (2.139) has an n-
dimensional counterpart being independent of dilation and translations. This ensures
the existence of the numbers K and D in (2.34), (2.35).
Remark 2.34. Both in case of the para-bases (2.102) based on (2.88)–(2.92) on the
one hand and in case of u-wavelet bases according to Definitions 2.4, 2.31 on the other
hand one tries to preserve as much as possible from the Rn -wavelets as described in
Section 1.2.1 when switching from Rn to domains  in Rn . It is quite clear that one has
to pay a price. The Rn -wavelets depend heavily on translations and dyadic dilatations.
There are no counterparts for domains (with exception of wavelets on the n-torus T n
as considered in Section 1.3). There is a huge literature about wavelets on domains.
We refer to [Mal99], [CDV00], [CDV04], [Dah97], [Dah01], [Coh03], [HoL05]. But
as far as we can see the constructions suggested there have little in common with
our approach. In case of para-bases one has now the residual wavelet system ‰ 2; in
(2.92) and in particular the scaled down scaling functions ‰Fjx ;m (in the notation used in
Definition 2.4) which spoil the orthogonality. First steps in this direction had been done
in [T06], Section 4.2 (based on [Tri07a]). This has been extended in [Tri07b]. Using
the rather particular properties of wavelets (multiresolution analysis, product structure)
one can restore the lost orthogonality at the expense of the lattice structure. These are
the boundary wavelets in Definition 2.4 (iii). Then one obtains bases which preserve
at least in a qualitative way typical ingredients of wavelets such as localisation, dyadic
refinements and also cancellations as far as basic and interior wavelets are concerned.
This is presented here for the first time. It is one of the main aims of this exposition to
develop this theory and its far-reaching consequences.
Remark 2.35. We fix the outcome such that it can be used as a starting point of what
follows beyond L2 . Let  be an arbitrary domain in Rn with  6D Rn and let u 2 N.
For suitably chosen numbers L 2 N, K 2 N, D > 0, positive constants c1 ; : : : ; c4
and approximate lattices Z there are orthonormal real u-wavelet bases in L2 ./,
˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj where Nj 2 N;x (2.148)
 2.4 Wavelet bases 53

according to Definitions 2.4, 2.31 and Theorem 2.33. Then f 2 L2 ./ can be ex-
panded as
X1 XNj 1 X
X Nj
  j
f D f; ˆr ˆr D
j
jr 2j n=2 ˆjr (2.149)
j D0 rD1 j D0 rD1
with Z
jr D jr .f /D2 j n=2
f .x/ ˆjr .x/ dx; (2.150)
Rn

adapted notationally to our later needs with the real L1 -normalised functions 2j n=2 ˆjr
(where L can be neglected). To justify the integration over Rn in (2.150) one may as-
sume that f is extended outside of  by zero. Of course,
1 X
X Nj
2j n jjr j2 D kf jL2 ./k2 : (2.151)
j D0 rD1

Recall that both the basic wavelets ˆ0r in (2.32) with G 2 fF; M gn and the interior
wavelets ˆjr in (2.33) have the cancellation property
Z
x ˇ ˆjr .x/ dx D 0 if jˇj < u: (2.152)
Rn

This follows from (1.88) and (2.27).

s;rloc
2.4.2 Wavelet bases in Lp ./ and Fpq ./

We obtained the orthonormal u-wavelet basis in Theorem 2.33 converting the resid-
ual wavelet system ‰ 2; in (2.92) into the basic and the boundary wavelets (2.32),
(2.34), (2.35) by strictly local arguments. This makes clear that one can now transfer
s
Theorems 2.23 and 2.28 from wavelet para-bases to u-wavelet bases. Let fpq .Z /
loc
be the sequence spaces as introduced in Definition 2.6. As usual L1 ./ collects all
complex-valued Lebesgue-measurable functions in  such that
Z
jf .x/j dx < 1 for any bounded domain ! with ! x  : (2.153)
!

Theorem 2.36. Let  be an arbitrary domain in Rn with  6D Rn . Let 1 < p < 1

and u 2 N. Let
˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj x
with Nj 2 N; (2.154)
be an orthonormal u-wavelet basis in L2 ./ according to Theorem 2.33 and Defini-
tion 2.31. Then Lp ./ is the collection of all f 2 Lloc
1 ./ which can be represented
as
X1 X Nj
f D jr 2j n=2 ˆjr ;  2 fp;2
0
.Z /: (2.155)
j D0 rD1
54 2 Spaces on arbitrary domains

Furthermore, fˆjr g is an unconditional basis in Lp ./. If f 2 Lp ./ then the

representation (2.155) is unique with  D .f /,
Z
 
jr .f / D 2j n=2 f; ˆjr D 2j n=2 f .x/ ˆjr .x/ dx (2.156)


and ˚  
I W f 7! .f / D 2j n=2 f; ˆjr (2.157)
is an isomorphic map of Lp ./ onto fp;2
0
.Z /,

kf jLp ./k  k.f / jfp;2

0
.Z /k (2.158)
.equivalent norms/.
Proof. We split f given by (2.155) as in (2.120) into f D f1 Cf2 , where f1 collects the
terms with the interior wavelets (2.33), and f2 the terms with the basic and the boundary
wavelets (2.32), (2.34), (2.35). We have (2.152) with ˇ D 0 for the terms in f1 . Then
we are in the same position as in the proof of Theorem 2.28. Hence, (2.155) converges
unconditionally in Lp .Rn / and hence in Lp ./. Furthermore by construction of the
orthonormal u-wavelet basis in L2 ./, an element f can be represented by (2.115)
if, and only if, it can be represented by (2.155). If f 2 D./ then this representation
is unique with (2.150), and hence (2.156) (recall that the ˆjr ’s are real and compactly
supported in ). Since D./ is dense in Lp ./ one obtains by completion the unique
representation (2.155), (2.156), and also (2.158). By the above comments it follows
that I in (2.157) is the indicated isomorphic map.
Remark 2.37. One may consider (2.158) as a Paley–Littlewood characterisation of
Lp ./ in terms of wavelet bases. As mentioned in Remark 2.30 there is a counterpart
of the above theorem in terms of -adapted Haar bases. We refer to Section 2.5.1
s
Recall that fpq .Z / are the sequence spaces according to Definition 2.6.
s;rloc
Theorem 2.38. Let  be an arbitrary domain in Rn with  6D Rn . Let Fpq ./ be
the spaces as introduced in Definition 2.14 with (2.56)–(2.58),
0 < p  1; 0 < q  1; s > pq ; (2.159)

.q D 1 if p D 1/. Let fˆjr g in (2.154) be an orthonormal u-wavelet basis in L2 ./

according to Theorem 2.33 and Definition 2.31 with s < u 2 N. Let v be as in (2.70),
s;rloc
(2.71). Then f 2 Lv ./ is an element of Fpq ./ if, and only if, it can be represented
as
X1 X Nj
f D jr 2j n=2 ˆjr ;  2 fpq s
.Z /: (2.160)
j D0 rD1
s;rloc
Furthermore if f 2 Fpq ./ then the representation (2.160) is unique with  D .f /
s;rloc
as in (2.156) and I in (2.157) is an isomorphic map of Fpq ./ onto fpq
s
.Z /,

kf jFpq
s;rloc
./k  k.f / jfpq
s
.Z /k (2.161)
 2.5 Complements 55

.equivalent quasi-norms/. If p < 1, q < 1 then fˆjr g is an unconditional basis in

s;rloc
Fpq ./.
Proof. By the same arguments as in the proof of Theorem 2.23 it follows that (2.160)
is an atomic decomposition in Rn (no moment conditions are needed). One has as a
s;rloc s;rloc
consequence f 2 Fpq ./. Furthermore since f 2 Fpq ./ can be represented by
(2.103) it can also be represented by (2.160). By (2.70), (2.71) it is also a representation
in Lv ./. Then it follows from Theorem 2.36 that this representation is unique with
(2.156). Finally one obtains the remaining assertions from Theorem 2.23 and the above
comments.
s;rloc
Remark 2.39. To what extent the spaces Fpq ./ with (2.159) in arbitrary domains 
in R are of self-contained interest is not so clear. On the one hand we have satisfactory
n

characterisations of these spaces in Theorem 2.18 and Corollary 2.20. On the other
hand we described in Remark 2.17 the little history of these spaces with (2.63) as the
s;rloc
main assertion. The motivation to deal with the spaces Fpq ./ comes from the
z s
question under which circumstances the more natural spaces Fpq ./ have the refined
localisation property, expressed by (2.63). This cannot be true for general domains,
but it will be valid in so-called E-thick domains. This is one of the main points of the
following Chapter 3. We refer in particular to Proposition 3.10.

2.5 Complements
2.5.1 Haar bases

We complement the above considerations by a few topics which are a little bit outside
the main stream of this exposition. We will be brief. First we add some comments
about Haar bases. The classical forerunners of the Daubechies wavelets (1.87), (1.88)
on R are the Haar wavelet hM ,
1 if 0  x  1=2,
hM .x/ D 1 if 1=2 < x  1, (2.162)
0 if x 62 Œ0; 1 ,
and the scaling function hF , hF .x/ D jhM .x/j (the characteristic function of the
interval Œ0; 1 ). This very first example of a multiresolution analysis resulting in an
orthonormal basis in L2 .R/ goes back to Haar, [Haar10]. Otherwise the n-dimensional
machinery described in Section 1.2.1 can be applied. In particular,
˚ j
HG;m W j 2 N0 ; G 2 G j ; m 2 Zn (2.163)
with
j
Y
n
HG;m .x/ D 2j n=2 hGr .2j xr  mr /; G 2 G j ; m 2 Zn ; (2.164)
rD1
56 2 Spaces on arbitrary domains

and j 2 N0 , is an orthonormal basis in L2 .Rn /. The corresponding expansion is given

by
1 X X
X
j n=2 j
f D j;G
m 2 HG;m (2.165)
j D0 G2G j m2Zn

with
 j 
m D m .f / D 2
j;G j;G j n=2
f; HG;m : (2.166)
s s
This is the counterpart of (1.93), (1.94). Let bpq and fpq be the same sequence spaces
as in Definition
 1.18.
 Then Theorem 1.20 can be complemented as follows. Recall
that p D n p1  1 C .
j
Theorem 2.40. Let HG;m be the Haar wavelets (2.164).
(i) Let 1 < p < 1. Then f 2 S 0 .Rn / is an element of Lp .Rn / if, and only if, it
can be represented as
X
j n=2 j
f D j;G
m 2 HG;m ;  2 fp;2
0
; (2.167)
j;G;m

unconditional convergence being in Lp .Rn /. The representation (2.167) is unique with

j;G j;G
m D m .f / as in (2.166). Furthermore,
˚  j 
I W f 7! 2j n=2 f; HG;m (2.168)
˚ j
is an isomorphic map of Lp .Rn / onto fp;2
0
and HG;m is an unconditional basis in
Lp .Rn /.
(ii) Let 0 < p < 1 and
 
1
p < s < min 1; : (2.169)
p

Then f 2 S 0 .Rn / is an element of Bpp

s
.Rn / if, and only if, it can be represented as
X
j n=2 j
f D j;G
m 2 HG;m ;  2 bpp
s
; (2.170)
j;G;m

unconditional convergence being in Bpp

s
.Rn /. The representation (2.170) is unique
j;G j;G
with m D m .f / as in ˚ (2.166). Furthermore I in (2.168) is an isomorphic map
j
of Bpp .R / onto bpp and HG;m is an unconditional basis in Bpp
s n s s
.Rn /.

Proof. Step 1. Part (i) is a famous result of real analysis. A proof may be found in
[Woj97], Section 8.3. It goes back to J. Marcinkiewicz, [Mar37] (one-dimensional
case), who in turn relied on [Pal32]. The above formulation is a little bit pedantic. But
we wanted to keep the assertion in the context of Theorem 1.20.
 2.5 Complements 57

Step 2. The proof of Theorem 1.20 relies on atomic representations as described in

Theorem 1.7 and the identification of j;G
m .f / in (2.166) with local means according
to Theorem 1.15. We modify (1.111) by
j
kjGm D 2j n=2 HG;m ; j 2 N0 ; G 2 G j ; m 2 Zn : (2.171)

Then we can apply Theorem 1.15 with A D 0 and B D 1, what requires p < s < 1.
The atomic Theorem 1.7 cannot be applied immediately. If s < 1=p then characteristic
s
functions of cubes are elements of Bpp .Rn / D Fpp
s
.Rn /. If

1
0 < p < 1; p < s <  < ; (2.172)
p
then one can apply the homogeneity assertion (2.43) based on the first lines in (2.40),
(2.41) or (2.47), covered by [CLT07]. One obtains that
j j
kHG;m jBpp

.Rn /k  2j.s/ kHG;m jBpp
s
.Rn /k: (2.173)

Now it follows from [T06], Theorem 2.13, that (2.170) is an expansion by non-smooth
atoms (after correct normalisation). Hence one can also apply Step 2 of the proof
of Theorem 1.20 as long as p < s < 1=p. Condition (2.169) covers the needed
restrictions for s. As in Step 1 of the proof of Theorem 1.20 we refer for technicalities
to [T06], Section 3.1.3, Theorem 3.5.

Remark 2.41. We refer to [T06], Theorem 1.58, ˚ jRemark 1.59, pp. s29–30, where we
discussed in detail under which circumstances HG;m is a basis in Bpq .Rn /. In partic-
ular the restriction (2.169) is natural. We gave here a new independent proof of part (ii)
of the above theorem for ˚ two reasons. First the assertions in [T06], Theorem 1.58, do
j
not cover that the basis HG;m is unconditional and that I in (2.168) is an isomorphic
s
map onto bpp . Secondly the above arguments are of the same type as in the proof of
Theorem 1.20. One may ask whether part (ii) of the above theorem can be extended to
s
other spaces Bpq .Rn / and Fpq
s
.Rn /. This is surely possible, but not done so far. The
sharp assertions about local means according to Theorem 1.15 fit in such a scheme.
Also the needed homogeneity (2.43) is available. But one has to extend the theory of
s
non-smooth atoms as treated in [T06], Theorem 2.13, from Bpp .Rn / to Bpq
s
.Rn / and
s
Fpq .Rn /. The generalisation of the above theorem from Bpp s
.Rn / to Bpq
s
.Rn / can also
be done by real interpolation.

Our main concern in Chapter 2 is not with spaces on Rn , but with spaces on arbitrary
domains  in Rn . Taking Theorem 2.40 as the Haar version of Theorem 1.20 we are
now in the same position as in case of the Daubechies wavelets in . Some assertions
0
are even simpler now. Let Qlr be the same pairwise disjoint open Whitney cubes as
in (2.22), (2.23) and let lr be their characteristic functions. First we complement
s
Theorem 2.16. To be consistent with our previous notation we prefer now Fpp instead
of Bpp . Recall that Fpp D Bpp .
s s s
58 2 Spaces on arbitrary domains

Proposition 2.42. Let  be an arbitrary domain in Rn with  6D Rn . Let 0 < p < 1,

 
1
p < s < min 1; ; (2.174)
p

and
n n
max.1; p/ < v < 1; s > : (2.175)
p v
s;rloc
Let Fpp ./ be the same space as in Definition 2.14. Then
˚
Fpps;rloc
./ D f 2 Lv ./ W kf jFpps;rloc
./k < 1 (2.176)

with
X
1 X 1=p
kf jFpp
s;rloc
./k D klr f jFpp
s
.Rn /kp ; (2.177)
lD0 r

.equivalent quasi-norms/.

Proof. We have (2.70), (2.71) which justifies to replace D 0 ./ in (2.56) by Lv ./
in order to avoid technical problems caused by the multiplication with characteristic
functions. Otherwise we are in the same position as in the proof of Theorem 2.16. We
need the pointwise multiplier assertion according to Theorem 2.13 for the characteristic
functions lr . As there one can reduce this question by homogeneity to the problem
of whether the characteristic function of the unit cube is a pointwise multiplier in
s
Bpp .Rn / D Fpp s
.Rn / with p, s as above. But this is a well-known assertion and may
be found in [T83], Theorem 2.8.7, p. 158.

Now we can complement the wavelet para-bases and wavelet bases originating from
Daubechies wavelets by Haar bases adapted to  (being again an arbitrary domain in
Rn with  6D Rn ). Let Qlr 0
be the open Whitney cubes according to (2.22), (2.23)
of side-length 2 now with the left corner 2l mr for some mr 2 Zn (in immaterial
l

modification of the previous setting). Let hM as in (2.162) and hF .x/ D jhM .x/j as
before. We extend (2.164) to G 2 fF; M gn , hence

j
Y
n
HG;m .x/ D 2j n=2 hGa .2j xa  ma /; G 2 fF; M gn ; m 2 Zn ; (2.178)
aD1

and j 2 N0 , which is the counterpart of (2.85). We extend Qlr

0
with (2.22), (2.23) by
its left and lower faces,
1
Qlr D 2l mr C 2l fx 2 Rn W 0  xa < 1g ; 1
Qlr D Qlr
0
: (2.179)

Then the adapted modifications of (2.88)–(2.92) are given by

Sj;1 D fF; M gn  fm 2 Zn W 2j m 2 Qlr

1
for some l < j , some rg; (2.180)
 2.5 Complements 59

Sj;2 D fF; M gn  fm 2 Zn W 2j m 2 Qjr

1
for some rg; (2.181)
1
[ 1
[
S  D S ;1 [ S ;2 ; S ;1 D Sj;1 ; S ;2 D Sj;2 ; (2.182)
j D1 j D0

and
˚ j
H  D HG;m W .j; G; m/ 2 S  : (2.183)

The Haar system inside an admitted cube Qlr 0

is the scaled down by the factor 2l
and translated Haar system for the unit cube. By construction H  is the orthonormal
0
union of these Haar systems in all admitted cubes Qjr . One obtains in particular the
following complement of Theorem 2.33.

Proposition 2.43. Let  be an arbitrary domain in Rn with  6D Rn . Then H 

according to (2.183) is an orthonormal basis in L2 ./.

Proof. This follows from the corresponding assertion for the unit cube in Rn and the
above constructions and comments.

˚ jNow we are in the same position as in Section 2.3. But instead of the para-basis
‰G;m in (2.102) we have now the orthonormal basis H in (2.183). Additional local
orthogonalisations as in Section 2.4 are not necessary now. As in (2.100), (2.101) let
˚
fpqs;
D  W k jfpq s;
k<1 (2.184)

with
   X 1=q 
 jf s;  D  2jsq jj;G

m j m .  /j jLp ./;
q
pq  (2.185)
.j;G;m/2S 

where
˚
 D j;G
m 2 C W .j; G; m/ 2 S

: (2.186)
If 0 < p D q < 1 then
   X 1=p
 jf s;   n
2j.s p /p jj;G
m j
p
pp : (2.187)
.j;G;m/2S 

Theorem 2.44. Let  be an arbitrary domain in Rn with  6D Rn . Let H  be the

orthonormal Haar basis in L2 ./ according to (2.183).
(i) Let 1 < p < 1. Then Lp ./ is the collection of all f 2 Lloc
1 ./ which can be
represented as
X
j n=2 j 0;
f D j;G
m 2 HG;m ;  2 fp;2 ; (2.188)
.j;G;m/2S 
60 2 Spaces on arbitrary domains

unconditional convergence being in Lp ./. Furthermore, H  is an unconditional

basis in Lp ./. If f 2 Lp ./ then the representation (2.188) is unique with  D
.f /,
Z
 j  j
j;G
m .f / D 2j n=2
f; HG;m D 2j n=2
f .x/ HG;m .x/ dx (2.189)


and ˚  j 
I W f 7! .f / D 2j n=2 f; HG;m (2.190)
0;
is an isomorphic map of Lp ./ onto fp;2 ,
 
f jLp ./k  k.f / jf 0;  (2.191)
p;2

.equivalent norms/.
(ii) Let 0 < p < 1,  
1
p < s < min 1; (2.192)
p
s;rloc
and v as in (2.175). Let Fpp ./ be the same space as in Definition 2.14. Then
s;rloc
Fpp ./ is the collection of all f 2 Lv ./ which can be represented by
X
j n=2 j
f D j;G
m 2 HG;m ;  2 fpp
s;
; (2.193)
.j;G;m/2S 

s;rloc
according to (2.184)–(2.187), unconditional convergence being in Fpp ./. Fur-
s;rloc s;rloc
thermore, H is an unconditional basis in Fpp ./. If f 2 Fpp ./ then the


representation (2.193) is unique with  D .f / as in (2.189) and I in (2.190) is an

s;rloc s;
isomorphic map of Fpp ./ onto fpp ,
   
f jF s;rloc ./  .f / jf s;  (2.194)
pp pp

.equivalent quasi-norms/.
Proof. Based on Theorem 2.40 and Propositions 2.42, 2.43 one can follow the argu-
ments from Sections 2.3, 2.4.

2.5.2 Wavelet bases in Lorentz and Zygmund spaces

s
We mainly deal in this book with the spaces Apq ./ on domains, their wavelet rep-
resentations and related extension problems. But we outline a few related topics. So
far we have wavelet bases for the Lebesgue spaces Lp with 1 < p < 1 in Rn , cov-
ered by Theorems 1.20, 2.40 (i) with Lp .Rn / D Fp;2
0
.Rn /, and in arbitrary domains 
with  6D R according to Theorems 2.36, 2.44 (i). We wish to extend these wavelet
n

representations to the
Lorentz spaces Lp;q ./; 1 < p < 1; 1  q < 1; (2.195)
 2.5 Complements 61

and the

Zygmund spaces Lp .log L/a ./; 1 < p < 1; a 2 R: (2.196)

We restrict ourselves to a description and refer for details, proofs, and further references
to [Tri08].
Let  be either Rn or an arbitrary domain in Rn with  6D Rn , let j j be the
Lebesgue measure of a (Lebesgue measurable) set in Rn and let f be a complex-
valued Lebesgue measurable, a.e. finite function in . Then

f .%/ D jfx 2  W jf .x/j > %gj; %  0; (2.197)

is the distribution function of f and f  defined on Œ0; 1/ by

f  .t / D inff% W f .%/  t g; t  0; (2.198)

its (decreasing) rearrangement. Details and properties may be found in [BeS88],

[EdE04] and [Har07]. The Lorentz spaces Lp;q ./ and the Zygmund spaces
Lp .log L/a ./ can be defined for all 0 < p; q  1, a 2 R. But we restrict our-
selves to those spaces which are of interest for us.
Definition 2.45. Let 1 < p < 1.
(i) Let  be either Rn or an arbitrary domain in Rn with  6D Rn . Let 1  q < 1.
Then ˚
Lp;q ./ D f 2 Lloc 1 ./ W kf jLp;q ./k < 1 (2.199)
with Z 1=q
jj  q dt

kf jLp;q ./k D t 1=p
f .t / : (2.200)
0 t
(ii) Let  be an arbitrary domain in Rn with jj < 1. Let a 2 R. Then
˚
Lp .log L/a ./ D f 2 L1 ./ W kf jLp .log L/a ./k < 1 (2.201)

with
Z jj 1=p
ap 
kf jLp .log L/a ./k D .1 C j log t j/ p
f .t / dt : (2.202)
0

Remark 2.46. The above Lorentz spaces Lp;q ./ and Zygmund spaces Lp .log L/a ./
are Banach spaces (although (2.200) and (2.202) are only respective equivalent quasi-
norms). Recall that Lp;q ./ can be obtained by real interpolation
 
Lp;q ./ D Lp0 ./; Lp1 ./ ;q ; p1 D 1 p0
C p1 ; (2.203)

where 1 < p0 < p < p1 < 1, [T78], Section 1.18.6. The above Zygmund spaces
Lp .log L/a ./ can also be reduced by extrapolation to Lebesgue spaces LpQ ./ with
1 < pQ < 1. Here one needs jj < 1. This may be found in [ET96], Section 2.6.2,
62 2 Spaces on arbitrary domains

with a reference to [EdT95]. It is just this reduction of the above Lorentz and Zygmund
spaces to Lebesgue spaces Lp ./ with 1 < p < 1 which paves the way to transfer
wavelet representations for the Lebesgue spaces to Lorentz and Zygmund spaces. As
said we restrict ourselves here to a description referring for details to [Tri08]. We
concentrate ourselves first to u-wavelet bases in domains  6D Rn and incorporate
afterwards  D Rn and Haar bases. We need the straightforward extension of Defini-
tion 2.6 using the same notation as there. In particular, jr is the characteristic function
of the ball B.xrj ; c2 2j / in (2.31).
Definition 2.47. Let  be an arbitrary domain in Rn with  6D Rn .jj < 1 for the
Zygmund spaces/ and let Z be as in (2.24)–(2.26). Let 1 < p < 1, 1  q < 1,
a 2 R. Then ƒpq .Z / is the collection of all sequences
˚
x
 D jr 2 C W j 2 N0 I r D 1; : : : ; Nj ; Nj 2 N; (2.204)

such that
 X Nj
1 X 1=2 
 
k jƒp;q .Z /k D  jjr jr .  /j2 jLp;q ./ < 1 (2.205)
j D0 rD1

and ƒp .log ƒ/a .Z / is the collection of all sequences (2.204) such that

 X Nj
1 X 1=2 
 
k jƒp .log ƒ/a .Z /k D  jjr jr .  /j2 jLp .log L/a ./ < 1:
j D0 rD1
(2.206)
Remark 2.48. Recall that

Lp;p ./ D Lp .log L/0 ./ D Lp ./; 1 < p < 1: (2.207)

Similarly one has

ƒp;p .Z / D ƒp .log ƒ/0 .Z / D fp;2

0
.Z / (2.208)

according to the above definition and (2.38).

Theorem 2.49. Let 1 < p < 1 and u 2 N.
(i) Let 1  q < 1 and let  be an arbitrary domain in Rn with  6D Rn . Let
˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj x
with Nj 2 N; (2.209)

be an orthonormal u-wavelet basis in L2 ./ according to Theorem 2.33 and Defini-

tion 2.31. Then Lp;q ./ is the collection of all f 2 Lloc
1 ./ which can be represented
as
X1 X Nj
f D jr 2j n=2 ˆjr ;  2 ƒp;q .Z /: (2.210)
j D0 rD1
 2.5 Complements 63
˚
Furthermore, ˆjr is an unconditional basis in Lp;q ./. If f 2 Lp;q ./ then the
representation (2.210) is unique with  D .f /,
Z
 
r .f / D 2
j j n=2
f; ˆr D 2
j j n=2
f .x/ ˆjr .x/ dx (2.211)


and ˚  
I W f 7! .f / D 2j n=2 f; ˆjr (2.212)
is an isomorphic map of Lp;q ./ onto ƒpq .Z /,

kf jLp;q ./k  k.f / jƒp;q .Z /k (2.213)

.equivalent quasi-norms/.
(ii) Let a 2 R and let  be an arbitrary domain in Rn with jj < 1. Then
Lp .log L/a ./ is the collection of all f 2 L1 ./ which can be represented as

1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 ƒp .log ƒ/a .Z /: (2.214)
j D0 rD1

˚
Furthermore, ˆjr is an unconditional basis in Lp .log L/a ./. If f 2
Lp .log L/a ./ then the representation (2.214) is unique with (2.211) and I in (2.212)
is an isomorphic map of Lp .log L/a ./ onto ƒp .log ƒ/a .Z / with a counterpart of
(2.213).
Remark 2.50. This is the extension of Theorem 2.36 from Lebesgue spaces to Lorentz
and Zygmund spaces. Basically one proves the above theorem for the Lorentz spaces
by the interpolation (2.203) of corresponding Lebesgue spaces (and its sequence coun-
terpart). Similarly for the Zygmund spaces using extrapolation. But this requires some
care and is shifted to [Tri08].
Although it is quite clear that there is a counterpart of the above theorem in terms
of Haar bases, generalising Theorems 2.40 (i), 2.44 (i) it seems to be reasonable to
give an explicit formulation. Let H  be the orthonormal Haar basis according to
Proposition
˚ 2.43 and (2.183). We incorporate now Rn identifying H  in this case with
Rn j
H D HG;m according to (2.163). We need the Lorentz and Zygmund counterparts
s;
of the sequence spaces fpq in (2.184), (2.186) originating from (2.101), (2.100) with
s D 0 and q D 2. Let  be an arbitrary domain in Rn (where  D Rn is now admitted)
and let 1 < p < 1, 1  q < 1. Then ƒpq 
is the collection of all sequences
˚
 D j;G
m 2 C W .j; G; m/ 2 S

(2.215)

such that
 X 1=2 
 
k jƒpq

kD jj;G
m j m .  /j
2
jLp;q ./ < 1: (2.216)
.j;G;m/2S 
64 2 Spaces on arbitrary domains

a with 1 < p < 1, a 2 R, jj < 1, is the collection of all

Similarly ƒp .log ƒ/
sequences (2.215) such that
 X 1=2 
 
k jƒp .log ƒ/
a k D  jj;G
m j m .  /j
2
jLp .log L/a ./ < 1:
.j;G;m/2S 
(2.217)
This is the adapted counterpart of Definition 2.47. Now Theorems 2.44 (i) and 2.49
can be complemented as follows.
Corollary 2.51. Let 1 < p < 1.
(i) Let 1  q < 1 and let  be an arbitrary domain Rn .where  D Rn is
admitted/. Let
˚ j
H  D HG;m W .j; G; m/ 2 S  (2.218)
be the orthonormal Haar basis in L2 ./ according to Proposition 2.43 and (2.183)
interpreted in the indicated way if  D Rn . Then Lp;q ./ is the collection of all
f 2 Lloc
1 ./ which can be represented as
X
j n=2 j
f D j;G
m 2 HG;m ;  2 ƒpq

: (2.219)
.j;G;m/2S 

Furthermore, H  is an unconditional basis in Lp;q ./. If f 2 Lp;q ./ then the

representation (2.219) is unique with  D .f / as in (2.189) and I in (2.190) is an
isomorphic map of Lp;q ./ onto ƒpq 
.
(ii) Let a 2 R and let  be an arbitrary domain in Rn with jj < 1. Then
Lp .log L/a ./ is the collection of all f 2 L1 ./ which can be represented by
X
j n=2 j
f D j;G
m 2 HG;m ;  2 ƒp .log ƒ/
a: (2.220)
.j;G;m/2S 

Furthermore, H  is an unconditional basis in Lp .log L/a ./. If f is contained in

Lp .log L/a ./ then the representation (2.220) is unique with  D .f / as in (2.189)
and I in (2.190) is an isomorphic map of Lp .log L/a ./ onto ƒp .log ƒ/ a.

Remark 2.52. The proof of Theorem 2.49 in [Tri08] extends Theorem 2.36 by linear
interpolation, some extrapolation and duality from Lebesgue spaces to Lorentz and
Zygmund spaces. Similarly one obtains Corollary 2.51 from Theorem 2.44 (i). Or-
thogonal wavelet bases in more general rearrangement invariant spaces on Rn have
been considered in [Soa97] using sub-linear real interpolation. One may ask whether
our arguments can be based on more general interpolation assertions in order to extend
Theorem 2.49 and Corollary 2.51 to other Orlicz spaces and rearrangement invariant
spaces. Zygmund spaces may be called logarithmic Lp -spaces. There are correspond-
ing logarithmic Sobolev spaces and several extensions based on logarithmic interpo-
lation spaces, closely connected with extrapolation arguments. This theory started in
[EdT95], [ET96], Section 6.2, and has been developed afterwards in [EdT98], [CFT04],
 2.5 Complements 65

[CFMM07]. It remains to be seen whether one can use these techniques also in con-
nection with what follows here to construct wavelet bases in logarithmic versions of
s
function spaces of type Apq .

2.5.3 Constrained wavelet expansions for Sobolev spaces

s
Later on we deal in Section 4.3.2 with constrained wavelet expansions in spaces Apq ./
with 0 < p; q  1 (p < 1 for F -spaces) and s > 0 in Lipschitz domains  in Rn .
There one finds also further discussions about this topic. We describe here a somewhat
curious counterpart for Sobolev spaces in arbitrary domains.
Definition 2.53. Let  be an arbitrary domain in Rn with  6D Rn . Let 1 < p < 1
and k 2 N. Then
˚
Wpk ./ D f 2 D 0 ./ W kf jWpk ./k < 1 ; (2.221)
X
kf jWpk ./k D kD ˛ f jLp ./k: (2.222)
j˛jk

Remark 2.54. Recall that

k
Fp;2 .Rn / D Wpk .Rn /; 1 < p < 1; k 2 N: (2.223)
k
Since Fp;2 ./, denoted temporarily as Wpk ./, is defined according to (2.4), (2.5) by
restriction one has the continuous embedding

Wpk ./ ,! Wpk ./; 1 < p < 1; k 2 N: (2.224)

The question whether these two spaces coincide might be called the non-linear exten-
sion problem. But more interesting is the (linear) extension problem, which means the
search for linear bounded operators

ext W Wpk ./ ,! Wpk .Rn / with ext f j D f (2.225)

or
Ext W Wpk ./ ,! Wpk .Rn / with Extf j D f; (2.226)

where gj has the same meaning as in (2.2). Later on we deal in detail with (linear)
s
extensions of type (2.225) for all spaces Apq ./. Some of the references given there
apply also to (2.226). For the spaces Wp ./ one has the Sobolev embedding
k

Wpk ./ ,! Lq ./; k n

p
  nq ; 1 < p  q < 1; (2.227)

obtained by restriction of a corresponding assertion on Rn . It is well known that the

two spaces in (2.224) coincide for large classes of domains. This applies in particular
to bounded Lipschitz domains, where one has also (2.225), (2.226). For arbitrary
66 2 Spaces on arbitrary domains

domains  in Rn one cannot expect that the two spaces in (2.224) coincide and also
the counterpart

Wpk ./ ,! Lq ./; k n

p
  nq ; 1 < p  q < 1; (2.228)

of (2.227) need not hold for all q, but maybe for some q (near p). Curiously enough
one can address questions of type (2.228) in terms of wavelet expansions.

So far we have for the Sobolev spaces Wpk;rloc ./ according to (2.74) both the
equivalent norms in Corollary 2.20 and the orthonormal u-wavelet bases with u > k
in Theorem 2.38. Nothing like this can be expected for Wpk ./ in arbitrary domains.
However there is a curious substitute.
The elements ˆjr of any orthonormal u-wavelet basis in L2 ./ according to Theo-
rem 2.33 and Definitions 2.31, 2.4 have compact supports in . This applies also to the
j
elements HG;m of the Haar basis H  in Proposition 2.43 and (2.183). In particular,

j
D ˛ ˆjr and D ˛ HG;m ; ˛ 2 Nn0 ; (2.229)

have also compact supports in . (It does not matter whether one looks at (2.229) as
elements of S 0 .Rn / or D 0 ./.) If f 2 Wpk ./ according to Definition 2.53 then one
can apply Theorem 2.36 (and similarly Theorem 2.44 (i)) to each D ˛ f 2 Lp ./ with
j˛j  k,
1 X
X Nj
D f D
˛
jr .D ˛ f / 2j n=2 ˆjr ; (2.230)
j D0 rD1

where
   
jr .D ˛ f / D 2j n=2 D ˛ f; ˆjr D .1/j˛j 2j n=2 f; D ˛ ˆjr : (2.231)

It may happen that D ˛ ˆjr are no longer regular distributions, but this does not matter.
Let for k 2 N,
˚
.f /k D jr .f /k 2 RC W j 2 N0 I r D 1; : : : ; Nj (2.232)

with X ˇ ˇ
jr .f /k D 2j n=2 ˇ f; D ˛ ˆj ˇ: (2.233)
r
j˛jk

We use the same notation as in Theorem 2.36.

Theorem 2.55. Let  be an arbitrary domain in Rn with  6D Rn . Let 1 < p < 1,

k 2 N and u 2 N. Let
˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj x
with Nj 2 N; (2.234)
 2.5 Complements 67

be an orthonormal u-wavelet basis in L2 ./ as in Theorem 2.33 and Definition 2.31.

Then Wpk ./ according to Definition 2.53 is the collection of all f 2 Lp ./ which
can be represented as

1 X
X Nj
f D jr .f / 2j n=2 ˆjr ; .f /k 2 fp;2
0
.Z /: (2.235)
j D0 rD1

Furthermore,
kf jWpk ./k  k.f /k jfp;2
0
.Z /k: (2.236)

Proof. This is an immediate consequence of Theorem 2.36 combined with (2.230)–

(2.233).

Remark 2.56. The wavelet expansion (2.155), (2.158) with .f / 2 fp;2 0
.Z / is now
the subject of the additional constraints .f / 2 fp;2 .Z /. This may justify to call
k 0

(2.235) a constrained wavelet expansion of f 2 Wpk ./. It gives the possibility to de-
cide whether a function f 2 Lp ./ or f 2 Lloc k
1 ./ belongs to Wp ./ independently
of the quality of .

Remark
˚ j 2.57. As mentioned above one can replace the orthonormal u-wavelet basis
ˆr in L2 ./ by the Haar basis H  . In other words, f 2 Lp ./ belongs to Wpk ./
if, and only if, it can be represented as
X
j n=2 j 0;
f D j;G
m .f / 2 HG;m ; .f /k 2 fp;2 ; (2.237)
.j;G;m/2S 

and  
kf jWpk ./k  .f /k jfp;2
0; 
(2.238)
in generalisation of Theorem 2.44. Here
˚
.f /k D j;G
m .f /k 2 RC W .j; G; m/ 2 S

(2.239)

with X ˇ ˇ
ˇ f; D ˛ H j ˇ:
m .f /k D 2
j;G j n=2
G;m (2.240)
j˛jk

j
If j˛j  1 then D ˛ HG;m are ı-distributions and their derivatives on faces and edges of
cubes.

Remark 2.58. If  is a bounded Lipschitz domain then there is a linear extension

operator according to (2.226), Wpk ./ D Wpk ./, and (2.227) with Wpk ./ in place
of Wpk ./. One obtains by Theorem 2.55 and (2.227) that

k.f / jfq;2
0
.Z /k  c k.f /k jfp;2
0
.Z /k (2.241)
68 2 Spaces on arbitrary domains

for all p; q in (2.227). Let  be an arbitrary domain in Rn with  6D Rn . Then one

has
Wpk ./ ,! Lq ./; k  pn   nq ; 1 < p  q < 1; (2.242)
if, and only if, there is a number c > 0 with (2.241) for all f 2 Lp ./. (Recall that
f 2 Lp ./ is an element of Wpk ./ if the right-hand side of (2.241) is finite.) But it
is not clear (to the author) how (2.241) is related to the geometry of the, say, connected
domain  (beyond bounded Lipschitz domains).
Remark 2.59. According to Definition 2.1 we deal here with spaces Apq s
./ defined
by restriction, including the special case Wp ./ in the above context. But at least for
k

Sobolev spaces the intrinsic version according to Definition 2.53 is natural and makes
sense. It attracted a lot of attention since the days of Sobolev, [Sob50], including
questions of type (2.226) and (2.242). The related standard reference nowadays is
Maz’ya’s book [Maz85].
Chapter 3
Spaces on thick domains

3.1 Thick domains

3.1.1 Introduction

Recall that open sets in Rn are denoted as domains. In Definition 2.1 we introduced
in arbitrary domains  with  6D Rn the spaces Apq
s
./ and AQpq
s
./ as subspaces of
0
D ./. So far we obtained in Theorem 2.36 common orthonormal u-wavelet bases for
all spaces
Lp ./ D Fp;2
0
./ D Fzp;2
0 0;rloc
./ D Fp;2 ./; 1 < p < 1; (3.1)
in arbitrary domains based on (2.113). On the one hand there is little hope to extend
s
this assertion to other spaces Apq ./ and AQpq
s
./. But on the other hand we obtained
in Theorem 2.38 for the refined localisation spaces
s;rloc
Fpq ./; 0 < p  1; 0 < q  1; s > pq ; (3.2)
(q D 1 if p D 1) in arbitrary domains  satisfactory orthonormal u-wavelet bases.
As described in Remark 2.17 we have so far
s;rloc
Fpq ./ D Fzpq
s
./; 0 < p; q  1; s > pq ; (3.3)
(q D 1 if p D 1) if  is a bounded Lipschitz domain in Rn . It is our first aim to
extend this assertion to E-thick domains which are more natural for questions of this
type. For this purpose we introduce first several classes of domains which will play a
crucial role in what follows.

3.1.2 Classes of domains

Tacitly we always assume that a domain  in Rn is not empty,  6D ;. Let l.Q/ be the
side-length of a (finite) cube Q in Rn with sides parallel to the axes of coordinates and
let jLj be the length of a rectifiable curve (closed path) L in Rn . Let I be an (arbitrary)
index set. Then
ai  bi for i 2 I (equivalence) (3.4)
for two sets of positive numbers fai W i 2 I g and fbi W i 2 I g means that there are two
positive numbers c1 and c2 such that
c1 ai  bi  c2 ai for all i 2 I: (3.5)
1 2
Let dist. ; / be as in (2.21).
70 3 Spaces on thick domains

Definition 3.1. Let  be a domain (D open set) in Rn with  6D Rn and D @.

(i) Then  is said to be an ."; ı/-domain, 0 < " < 1, 0 < ı < 1, if it is connected
and if for any x 2 , y 2  with jx  yj < ı there is a curve L  , connecting x
and y such that jLj  "1 jx  yj and

dist.z; /  " min.jx  zj; jy  zj/; z 2 L: (3.6)

(ii) Then  is said to be E-thick .exterior thick/ if one finds for any interior cube
Qi   with

l.Qi /  2j ; dist.Qi ; /  2j ; j  j0 2 N; (3.7)

a complementing exterior cube Qe  c D Rn n with

l.Qe /  2j ; dist.Qe ; /  dist.Qi ; Qe /  2j ; j  j0 2 N: (3.8)

(iii) Then  is said to be I -thick .interior thick/ if one finds for any exterior cube
Qe  c with

l.Qe /  2j ; dist.Qe ; /  2j ; j  j0 2 N; (3.9)

a reflected cube Qi   with

l.Qi /  2j ; dist.Qi ; /  dist.Qe ; Qi /  2j ; j  j0 2 N: (3.10)

(iv) Then  is said to be thick if it is both E-thick and I -thick.

Remark 3.2. The equivalence constants in (3.7)–(3.10) are independent of j . In other
words, a domain  is called E-thick if for any choice of positive numbers c1 , c2 , c3 , c4
and j0 2 N there are positive numbers c5 , c6 , c7 , c8 such that one finds for each interior
cube Qi   with

c1 2j  l.Qi /  c2 2j ; c3 2j  dist.Qi ; /  c4 2j ; (3.11)

j  j0 , an exterior cube Qe  c with

c5 2j  l.Qe /  c6 2j ; c7 2j  dist.Qe ; /  dist.Qi ; Qe /  c8 2j ;

(3.12)
j  j0 . One may think about the Whitney cubes according to (2.22), (2.23) identifying
0
Qlr or Qlr1
with Qi in (3.7) or (3.11) (where now l D j ). One checks quite easily
that  is E-thick if, and only if, one has (3.12) for some positive c5 , c6 , c7 , c8 for the
indicated Whitney cubes .with related positive c2 , c2 , c3 , c4 /. Similarly one can detail
(3.9), (3.10) in case of I -thick domains.
Remark 3.3. In this Chapter 3 we are mainly interested in E-thick domains which
prove to be the natural class of domains in connection with wavelet bases in function
spaces. In the next Chapter 4 we use these results to study the so-called extension
 3.1 Thick domains 71

problem. Then I -thick domains are coming in naturally. On the one hand by Proposi-
tion 3.6 below ."; ı/-domains are special I -thick domains. On the other hand it is well
known that ."; ı/-domains in the plane R2 play not only a crucial role in the theory
of quasi-conformal mappings in R2 but also in connection with the question of the
extendability of the Sobolev spaces Wpk ./ according to Definition 2.53 to R2 , hence
the problem of whether there exists a linear and bounded extension operator Ext as
indicated in (2.226). What is meant by ."; ı/-domains  in Rn can be seen in Fig-
ure 3.1. In particular (3.6) ensures that with L there is also a surrounding croissant-like

L
y
L
x


Figure 3.1

L
y
Qi
x

e
Q

Figure 3.2

subdomain L  . According to [Jon81] for any ."; ı/-domain  in Rn there exists

a linear and bounded extension operator

Extpk W Wpk ./ ,! Wpk .Rn /; 1  p  1; k 2 N; (3.13)

with Extpk f j D f , where Definition 2.53 is naturally extended to p D 1 and p D 1.

But it is not a common extension operator as needed in many applications. However
the method in [Jon81] has been modified recently in [Rog06] such that one obtains
common extension operators of type (3.13). The considerations in [Jon81] have been
72 3 Spaces on thick domains

extended in [Chr84], [Miy93] to spaces of fractional order and in [See89], [Miy98] to

the above spaces
s
Fpq ./ with 0 < p < 1, 0 < q  1, s > pq ; (3.14)

according to Definition 2.1. Some assertions beyond ."; ı/-domains may also be found
in [Ry00]. In the plane R2 the property to be an ."; ı/-domain is not only sufficient for
the extendability of intrinsically defined Sobolev spaces but (with some mild additional
assumptions) also necessary. The related literature may be found in [Jon81]. However
the situation is not so simple and a detailed discussion may be found in [Maz85],
Section 1.1.16, §1.5. Our intentions are different. According to Definition 2.1 all spaces
considered in this book are introduced by restriction. In case of irregular domains this
may change the situation substantially. We refer in this context to Remark 2.54. We
return in Chapter 4 to the extension problem but in contrast to the above-mentioned
literature in the framework of wavelet bases. First we complement the above notation
about sets and classes of domains as follows.

Recall that balls in Rn centred at x 2 Rn and of radius r > 0 are denoted as B.x; r/.
Furthermore for 2  n 2 N,

Rn1 3 x 0 7! h.x 0 / 2 R (3.15)

is called a Lipschitz function (on Rn1 ) if there is a number c > 0 such that

jh.x 0 /  h.y 0 /j  c jx 0  y 0 j for all x 0 2 Rn1 , y 0 2 Rn1 : (3.16)

Definition 3.4. (i) Let n 2 N. A closed set in Rn is said to be porous if there is a

number with 0 < < 1 such that one finds for any ball B.x; r/ centred at x 2 Rn
and of radius r with 0 < r < 1, a ball B.y; r/ with

B.y; r/  B.x; r/ and B.y; r/ \ D ;: (3.17)

(ii) Let 2  n 2 N. A special Lipschitz domain .C 1 domain/ in Rn is the collection

of all points x D .x 0 ; xn / with x 0 2 Rn1 such that

h.x 0 / < xn < 1; (3.18)

where h.x 0 / is a Lipschitz function according to (3.15), (3.16) .a bounded C 1 function/.

(iii) Let 2  n 2 N. A bounded Lipschitz domain .C 1 domain/ in Rn is a bounded
domain  in Rn where the boundary D @ can be covered by finitely many open
balls Bj in Rn with j D 1; : : : ; J , centred at such that

Bj \  D Bj \ j for j D 1; : : : ; J ; (3.19)

where j are rotations of suitable special Lipschitz domains .C 1 domains/ in Rn .

 3.1 Thick domains 73

Remark 3.5. Of course we always assume that bounded (Lipschitz or C 1 ) domains

are not empty. Later one we need the porosity of the boundary D @ of some
0
E-thick domains in connection with spaces Apq ./ of smoothness zero. A detailed
discussion about porosity (also called ball condition) especially in connection with
fractal measures may be found in [T01], Section 9.16–9.19, pp. 138–141. If one
knows (3.17) only for balls B.x; r/ centred at x 2 , then one has (3.17) for all balls
with =2 in place of , hence is porous. In Definition 5.40 we introduce so-called
cellular domains, a class of domains between Lipschitz and C 1 .

3.1.3 Properties and examples

We wish to compare the diverse types of domains introduced in the preceding Sec-
tion 3.1.2 and to illustrate the situation by a few examples. Let  be a domain (D open
x ı is the largest domain ! with !  .
set) in Rn . Then ./ x Of course,   ./x ı . We
use the notation as introduced in Definitions 3.1, 3.4.
Proposition 3.6. (i) Let  be an ."; ı/-domain in Rn . Then  is I -thick. Furthermore,
@ is porous and j@j D 0.
(ii) Let  6D Rn be an arbitrary domain. Then one has the decomposition of Rn
into three disjoint sets
x and @.Rn n/
Rn D  [ @ [ .Rn n/ x  @: (3.20)

Furthermore,
x if, and only if, ./
@ D @.Rn n/ x ı D : (3.21)
x D  and R n
(iii) If  is an E-thick domain in R then ./
n ı n x is I -thick.
x 6D Rn , then Rn n
(iv) If  is an I -thick domain in Rn and  x is E-thick.

Proof. Step 1. We prove (i). Let  be an ."; ı/-domain and let Qe be as in Figure 3.2,
with (3.9) where j0 2 N is assumed to be sufficiently large. Recall that  is connected.
Then there is a point x 2  with

dist.x; @/  dist.x; Qe /  2j (3.22)

and a second point y 2  with jx  yj  2j . As indicated in Figure 3.2, one finds in
the croissant-domain L a cube Qi with the desired properties. This argument proves
also the porosity of @. Then one obtains j@j D 0 from [T01], Sections 9.16, 9.17,
pp. 138–39, where porosity was called ball condition.
Step 2. Sine x D  [ @ the decomposition in (3.20) is obvious. The second
assertion in (3.20) follows from
x 
@.Rn n/ x and x \  D ;:
@.Rn n/
x ı D  then x 2 @.Rn n/.
If x 2 @ for a domain  with ./ x Together with (3.20)
n x n x x ı . Then
one obtains @ D @.R n/. Conversely let @ D @.R n/ and x 2 ./
74 3 Spaces on thick domains

either x 2  or x 2 @ D @.Rn n/. x But the latter is not possible. Hence x 2  and
x D . This proves (ii).
./ ı

Step 3. If  is E-thick and x 2 @ then it follows x 2 @.Rn n/ x and by (3.20)

that @ D @.Rn n/.x Now one obtains ./ x ı D  by (3.21). Furthermore, Rn nx is
I -thick. This proves (iii). Part (iv) follows from Definition 3.1.
Remark 3.7. The first assertion in part (i) of the above proposition is a cornerstone
of the considerations in [Jon81], [See89] in connection with the extension problems
(3.13), (3.14). In good agreement with Remark 3.2 the arguments in these papers rely
on Whitney cubes with universal equivalence constants. But there exist connected I -
thick domains which are not ."; ı/-domains. Let 0 < ˛ < 1 and let locally   R2
below the curve x2 D jx1 j˛ , Figure 3.3. Then  is I -thick, but not an ."; ı/-domain

Qe


Qi

Figure 3.3


Qi

Qe

Figure 3.4

(the indicated connecting curves are too long). Furthermore if  is locally above the
curve x2 D jx1 j˛ as indicated in Figure 3.4, then  is E-thick but not an ."; ı/-domain
(possible croissants are not fat enough). But there is an other difference between E-
thick, I -thick domains on the one hand and C 1 , Lipschitz, ."; ı/-domains on the other
 3.1 Thick domains 75

hand. There are E-thick domains  and I -thick domains  such that j@j > 0 with
the consequences discussed so far in Remark 2.2. We give an example and return for
this purpose to the construction (2.11), (2.12). Let Il be the same intervals as there. Let
Jl0 with l 2 N be intervals in .0; 1/ of length "  2l and dist.Jl0 ; 0/  2l . Similarly
Jl1 are corresponding intervals approaching 1. Let " > 0 be small and
1
[  
D Il [ Jl0 [ Jl1 : (3.23)
lD1

Then  is I -thick with

@ D Œ0; 1 n and j@j > 0: (3.24)

To construct a corresponding E-thick domain we use the domain  in (2.11) now

written as
1
[ 1
[
D Il D Il0 ; (3.25)
lD1 lD1

where Il0 are disjoint intervals. We decompose each Il0 into

Il0 D Il1 [ fxlk g1

kD1 [ Il
2
(3.26)

where Il1 is the union of disjoint intervals Il;k

1
of length  2k jIl0 j, k 2 N. Similarly
for Il2 . This can be done in such a way that Il1 is E-thick with the related exterior
intervals in Il2 and vice versa. Then
1
[ 1
[
1 D Il1 and 2 D Il2 (3.27)
lD1 lD1

are E-thick and

0 < j@j  j@1 j C j@2 j: (3.28)
Hence there are E-thick domains ! with j@!j > 0.
Thick domains, ."; ı/-domains and bounded Lipschitz domains have the same
meaning as in Definitions 3.1 and 3.4.
Proposition 3.8. (i) Any bounded Lipschitz domain is thick.
(ii) Any connected bounded Lipschitz domain is an ."; ı/-domain.
(iii) The classical snowflake domain  in the plane R2 according to Figure 3.5, is
a thick ."; ı/-domain.

Proof. Parts (i) and (ii) are more or less obvious. If  is the indicated snowflake
domain then it is also and ."; ı/-domain as can be seen elementary. Then it follows by
Proposition 3.6 (i) that  is I -thick. It remains to prove that  is also E-thick. This
76 3 Spaces on thick domains

Qe
B Qi

C 
j
23 3  3j

Figure 3.5.

follows from a careful examination of the indicated construction in Figure 3.5. Starting
at level 3j it follows by geometrical reasoning that the following construction remains
inside of the triangle ABC , whereas the indicated strip S , vertically translated by 3j ,
remains free of points of @. Hence for Qi with l.Qi /  3j as shown one finds in
this strip a complementing cube Qe with l.Qe /  3j .

Remark 3.9. The examples in Remark 3.7 make clear that E-thick and I -thick domains
may be rather bizarre. Even the more regular ."; ı/-domains cover some domains with
fractal boundary. In connection with the above snowflake domain we refer to [Maz85],
Section 1.5.1, pp. 70–71, where one finds an alternative construction of a snowflake-
like ."; ı/-domain in the plane R2 based on squares instead of equilateral triangles.
Otherwise we believe that (exterior and interior) thick domains are at the heart of
the matter for problems of the theory of function spaces on domains as considered in
this book: wavelet characterisations, linear extensions to Rn , and intrinsic descriptions.
However there is a large variety of similar notation adapted to (other) specific questions.
 3.2 Wavelet bases in Axpq
s
./ 77

Based on [TrW96] and [ET96], Section 2.5, we describe in Remark 4.29 below what
is meant by (exterior and interior) regular domains and what can be said about atomic
representations for related function spaces in these domains. Furthermore we refer to
the corkscrew property for non-tangentially accessible domains introduced in [JeK82]
in order to study boundary value problems for second order elliptic equations. Details
and generalisations (domains of class S) may be found in [Ken94], pp. 4, 8. In
connection with Sobolev and Poincaré inequalities for function spaces (preferably
Sobolev spaces) conditions of the above type play a role resulting in John domains
and plump domains. Details, references, examples and discussions may be found in
[EdE04], Section 4.3.

3.2 Wavelet bases in Axpq

s
./

3.2.1 The spaces Fzpq

s
./

In arbitrary domains  in Rn with  6D Rn we have the orthonormal u-wavelet bases

˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj x
with Nj 2 N (3.29)

in L2 ./ according to Theorem 2.33 based on Definitions 2.4, 2.31. We extended this
s;rloc
assertion in Theorem 2.38 to the refined localisation spaces Fpq ./ with

0 < p  1; 0 < q  1; s > pq ; (3.30)

(q D 1 if p D 1) as introduced in Definition 2.14 with (2.56)–(2.58). In particular,

s;rloc
any f 2 Fpq ./ can be uniquely represented as
1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 fpq
s
.Z /; (3.31)
j D0 rD1

s
where fpq .Z / are the sequence spaces according to Definition 2.6 and
Z
 
jr D jr .f / D 2j n=2 f; ˆjr D 2j n=2 f .x/ ˆjr .x/ dx: (3.32)


It is the main aim of Section 3.2 to extend these wavelet representations (bases, isomor-
phisms onto related sequence spaces) to some spaces AQpq s
./ and Apqs
./ according
to Definition 2.1 where  is now assumed to be E-thick. First we prove by direct argu-
ments that the spaces Fzpq
s
./ with (3.30) have the same wavelet representations (3.31),
s;rloc
(3.32) as the spaces Fpq ./ with the consequence that these two spaces coincide.
Proposition 3.10. Let  be an E-thick domain in Rn according to Definition 3.1 (ii).
Let
0 < p  1; 0 < q  1; s > pq ; (3.33)
78 3 Spaces on thick domains

.q D 1 if p D 1/. Let Fpq s;rloc

./ be as in Definition 2.14, (2.56)–(2.58), and Fzpq
s
./
z z
be as in Definition 2.1 (ii) .with F11 ./ D B11 .//. Then
s s

s;rloc
Fpq ./ D Fzpq
s
./: (3.34)

Proof. As indicated above we prove that Fzpqs

./ has the same wavelet representation
˚
as Fpq ./ with (3.34) as a consequence. Let f be given by (3.31) where ˆjr is an
s;rloc

u-wavelet system according to Definition 2.4 with u > s. After correct normalisations
s
one has an atomic representation in Fpq .Rn / (no moment conditions are needed). Then
it follows from Definition 2.6 and Theorem 1.7 that

kf jFzpq
s
./k  kf jFzpq
s x
./k D kf jFpq
s
.Rn /k  c k jfpq
s
.Z /k: (3.35)

Conversely let f 2 Fzpq

s
./. If p < 1 then it follows from (2.71) with 1 < v < 1
and Theorem 2.36 that f can be represented as
1 X
X Nj
f D jr .f / 2j n=2 ˆjr (3.36)
j D0 rD1

at least in Lv ./. If p D 1 and, hence, also v D 1, then one has (3.36) at least
locally. But this is sufficient also in case of p D q D 1 (recall that all spaces are
considered as subspaces of D 0 ./). We wish to apply Theorem 1.15 (ii) with A D 0
and u D B > s to the kernels

krj D 2j n=2 ˆjr ; j 2 N0 I r D 1; : : : ; Nj : (3.37)

As for the correct normalisation we refer to Definition 2.4, based on (2.27) and (1.39).
In case of the basic wavelets (2.32) and the interior wavelets (2.33) these kernels have
the required moment conditions as a consequence of u > s and (1.88). This need not
to be the case for the boundary wavelets in Definition 2.4 (iii). Let krj be given by
(3.37) where ˆjr is a boundary wavelet according to (2.34), (2.35). We may assume
that j  j0 where j0 is sufficiently large. Then

supp krj  Qi (3.38)

for a suitable interior cube with (3.7). Let Qe be a related complementing cube with
(3.8). Then there is a function kVrj 2 C u .Rn / with supp kVrj  Qe such that kQrj ,

kQrj .x/ D krj .x/ C kVrj .x/; x 2 Rn ; (3.39)

is an admitted kernel having in particular the required moment conditions. The exis-
tence of such complementing functions kVrj is quite plausible but not obvious. A detailed
construction may be found in [TrW96], pp. 665–66. Let g 2 Fzpq s x
./ with gj D f
and
kg jFpq
s
.Rn /k D kg jFzpq
s x
./k  2 kf jFzpq s
./k: (3.40)
 3.2 Wavelet bases in Axpq
s
./ 79

x one has
Since supp g  
Z Z
kQrj .x/ g.x/ dx D krj .x/ f .x/ dx: (3.41)
Rn 

Now we can apply Theorem 1.15. One obtains that

k.f / jfpq
s
.Z /k  c kg jFpq
s
.Rn /k  c 0 kf jFzpq
s
./k: (3.42)

Together with (3.35) it follows that Fzpq

s
./ is the collection of all f which can be
uniquely represented by (3.36) with

kf jFzpq
s
./k  k.f / jfpq
s
.Z /k: (3.43)

But according to Theorem 2.38 this is the same as for the corresponding representation
s;rloc
in Fpq ./. This proves (3.34).

3.2.2 The spaces Axpq

s
./ I
s
We are now ready for the main result of Chapter 3. Let Apq ./ and AQpq
s
./ be the
spaces introduced in Definition 2.1 for arbitrary domains  in R with  6D Rn ,
n

considered as subspaces of D 0 ./. Let

   
1 1
p D n 1 and pq D n 1 ; 0 < p; q  1; (3.44)
p C min.p; q/ C

with bC D max.b; 0/ if b 2 R, be the same numbers as in (1.32).

Definition 3.11. Let  be an E-thick domain in Rn according to Definition 3.1 (ii).

 Fz
Then
s
pq ./ if 0 < p < 1, 0 < q  1, s > pq ,
Fxpq
s
./ D 0
Fpq ./ if 1 < p < 1, 1  q < 1, s D 0, (3.45)
s
Fpq ./ if 0 < p < 1, 0 < q  1, s < 0,
and  Bz s
if 0 < p  1, 0 < q  1, s > p ,
pq ./
Bxpq
s
./ D 0
Bpq ./ if 1 < p < 1, 0 < q  1, s D 0, (3.46)
s
Bpq ./ if 0 < p  1, 0 < q  1, s < 0:

Remark 3.12. Recall that all spaces are considered as subspaces of D 0 ./. As always
s;rloc
 6D Rn . By Proposition 3.10 the refined localisation spaces Fpq ./, introduced in
Definition 2.14 in arbitrary domains  in R with  6D R , coincide with the spaces
n n

Fxpq
s
./ D Fzpq
s
./ if  is E-thick and s > pq . This assertion has a little history
mentioned briefly in Remark 2.17. In Definition 3.4 we recalled what is meant by
80 3 Spaces on thick domains

bounded Lipschitz domains and bounded C 1 domains in Rn . In [T01], Theorem 5.14,

p. 60–61, we proved (3.34) for bounded C 1 domains in Rn , called the refined locali-
sation property for Fzpq
s
./. We obtained in [T06], Proposition 4.20, p. 208, the same
z s
assertion for Fpq ./ in bounded Lipschitz domains under the additional restriction

1 < p; q  1 (q D 1 if p D 1); s > 0:

This is now valid for all p, q, s in (3.33) since bounded Lipschitz domains are E-thick
according to Proposition 3.8 (i). However in [T06], p. 209, Remark 4.21, footnote, we
already indicated that such an extension might be possible. Now we have this refined
localisation property of Fzpq
s
./ in all E-thick domains. This class of domains seems
to be natural for questions of this type. In other words, beyond E-thick domains the
s;rloc
refined localisation spaces Fpq ./ according to (2.56)–(2.58) seem to be the right
way to look at spaces of this type.
s s
Let bpq .Z / and fpq .Z / be the same sequence spaces as in Definition 2.6. One
word about the convergence of the series in the following theorem seems to be in order.
The u-wavelet system fˆjr g according to Definition 2.4 in  can also be considered as
an u-wavelet system in Rn . Hence instead of saying that a related series converges in
D 0 ./ one could equally say that this series converges in S 0 .Rn /. Similarly one can
likewise say that such a series converges locally in Axpq
 
./ or locally in Apq .Rn /. In
particular, if  is bounded then convergence locally in Axpq

./ means convergence in
Axpq ./. Hence as far as convergences and other technicalities are concerned (duality


etc.) we are in the same position as in Theorem 1.20 where one finds corresponding
comments in front of this theorem and in the first step of its proof.
Theorem 3.13. Let  be an E-thick domain in Rn according to Definition 3.1 (ii). Let
for u 2 N, ˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj x
with Nj 2 N; (3.47)
be an orthonormal u-wavelet basis in L2 ./ according to Theorem 2.33 and Defini-
tions 2.31, 2.4.
(i) Let Fxpq
s
./ be the spaces in (3.45) and let

u > max.s; pq  s/; s 6D 0: (3.48)

Then f 2 D 0 ./ is an element of Fxpq

s
./ if, and only if, it can be represented as

1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 fpq
s
.Z /; (3.49)
j D0 rD1

unconditional convergence being in D 0 ./ and locally in any space Fxpq


./ with
 < s. Furthermore, if f 2 Fxpq ./ then the representation (3.49) is unique with
s

 D .f /,  
jr D jr .f / D 2j n=2 f; ˆjr ; (3.50)
 3.2 Wavelet bases in Axpq
s
./ 81

and ˚  
I W f 7! .f / D 2j n=2 f; ˆjr (3.51)
˚
is an isomorphic map of Fxpq s
./ onto fpq
s
.Z /. If, in addition, q < 1 then ˆjr is
an unconditional basis in Fxpq s
./.
x
(ii) Let Bpq ./ be the spaces in (3.46) and let
s

u > max.s; p  s/; s 6D 0: (3.52)

Then f 2 D 0 ./ is an element of Bxpq
s
./ if, and only if, it can be represented as
1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 bpq
s
.Z /; (3.53)
j D0 rD1

unconditional convergence being in D 0 ./ and locally in any spaces Bxpq 

./ with
 < s. Furthermore, if f 2 Bxpq ./ then the representation (3.53) is unique with
s

 D .f / as in (3.50) and I in (3.51) is an isomorphic map of Bxpq

s
./ onto bpqs
.Z /.
˚ j
x
If, in addition, p < 1, q < 1, then ˆr is an unconditional basis in Bpq ./.
s

Proof. Step 1. Part (i) with s > pq is covered by Proposition 3.10. The proof given
there applies also to the spaces Bxpq
s
./ with s > p . Hence it remains to prove the
theorem for the spaces with s < 0.
Step 2. We prove (i) for the third line in (3.45) in three steps. Let in particular
s < 0 and u > pq  s: (3.54)
Recall that the interior wavelets in (2.33) have the cancellation property (2.152). Then
it follows by Theorem 1.7 (ii) that
n
arj D 2j.s p / 2j n=2 ˆjr (3.55)
s
are atoms in Fpq .Rn / if ˆjr are basic wavelets or interior wavelets according to Defini-
tion 2.4 (neglecting immaterial constants). As for the boundary wavelets ˆjr in (2.34),
(2.35) we are in a similar situation as in (3.38), (3.39). By the same arguments one can
complement

ˆjr with supp ˆjr  Qi by V̂ jr 2 C u .Rn / with supp V̂ jr  Qe

such that arj ,

 
arj .x/ D 2j.s p / 2j n=2 ˆjr .x/ C V̂ jr .x/ ;
n
x 2 Rn ; (3.56)
s
is an atom in Fpq .Rn / having the needed moment conditions. Let now f be given by
(3.49). Then
X1 XNj
n
gD 2j.s p / jr arj ;  2 fpq
s
.Z /; (3.57)
j D0 rD1
82 3 Spaces on thick domains

s
is an atomic decomposition of g in Fpq .Rn / with gj D f . One has by Theorem 1.7
that
kf jFpq
s
./k  kg jFpq s
.Rn /k  c k jfpq
s
.Z /k: (3.58)
Step 3. Next we assume that f 2 Fpq
s
./ can be represented by
1 X
X Nj
f D jr .f / 2j n=2 ˆjr : (3.59)
j D0 rD1

As in (3.36), (3.37) we interpret

   
jr .f / D 2j n=2 f; ˆjr D f; krj ; krj D 2j n=2 ˆjr ; (3.60)

as local means with respect to the kernels krj . Again the kernels krj have the desired
normalisation (1.39). One can apply Theorem 1.15 with A D u > pq  s and B D 0.
In particular, no moment conditions are required. Then it follows that
k.f / jfpq
s
.Z /k  c kf jFpq
s
.Rn /k: (3.61)
Let g 2 Fpqs
.Rn / with gj D f . Then .f / D .g/ and one has (3.61) with g in
place of f on the right-hand side. Taking the infimum over all such g one arrives at
(3.61) with  on the right-hand side instead of Rn . Together with (3.58) one obtains
k.f / jfpq
s
.Z /k  kf jFpq
s
./k (3.62)
for all f 2 Fpqs
./ which can be represented by (3.59).
Step 4. In other words, it remains to prove that any f 2 Fpq
s
./ can be represented
by (3.59). Let 1 < p < 1 and 0 < q < 1. Then S./ D S.Rn /j is dense in
s
Fpq ./. Since
S./  Lp ./ ,! Fpq s
./ (3.63)
s
it follows that Lp ./ is also dense in Fpq ./. According to Theorem 2.36 any f 2
Lp ./ can be represented by (3.59). Let f 2 Fpq s
./ and let ffl g  Lp ./ be an
approximating sequence. By (3.62) with fl in place of f and Fatou’s lemma one has
.f / 2 fpqs
.Z /. Then it follows from Step 2 that
1 X
X Nj
gD jr .f / 2j n=2 ˆjr 2 Fpq
s
./: (3.64)
j D0 rD1

By the same arguments as in Rn one obtains f D g, where details may be found in

[T06], p. 155. This proves that any f 2 Fpq
s
./ can be represented by (3.59). By
s s"
Fp1 ./ ,! Fpp ./; " > 0; (3.65)
one has a corresponding assertion for Fp1s
./ with 1 < p < 1, s < 0. If p  1 then
one obtains the desired representability from the continuous embedding
s
Fpq 
./ ,! Frr ./; s n
p
D   nr ; 1 < r < 1: (3.66)
 3.2 Wavelet bases in Axpq
s
./ 83

This covers all cases and proves part (i) of the above theorem.
Step 5. The proof of part (ii) for the B-spaces is the same at least as long as p < 1.
Let p D 1. If  is bounded then the representability (3.59) follows from
s
B1q s
./ ,! Bpq ./; 0 < p < 1: (3.67)
The question of the representability (3.59) in D 0 ./ is a local matter (it must be tested
against ' 2 D./). Then it follows from the local nature of ˆjr and the above
considerations that f 2 B1q
s
./ can also be represented by (3.59) if  is unbounded.
Step 6. By the above arguments it follows that I is the indicated isomorphism. In
particular if p < 1, q < 1 then fˆjr g is an unconditional basis.
Remark 3.14. So far we have a satisfactory assertion for the spaces Apq
0
./ of smooth-
ness zero only in case of Lp ./ D Fp;2 ./, 1 < p < 1. We refer to Theorem 2.28.
0

We return to spaces of smoothness zero in Section 3.2.4. Then it will be clear that one
needs some further properties of the underlying domain  to extend the above theorem
to spaces of smoothness zero.

3.2.3 Complemented subspaces

Theorem 3.13 gives the possibility to interpret AQpq

s
./ with A D B or A D F as a
s n
complemented subspace of Apq .R /. This will be of some use later on in Chapter 4 in
connection with the extension problem.
As usual T W A ,! B means that T is a linear and bounded (continuous) operator
(map) from the quasi-Banach space A into the quasi-Banach space B. If T D id is a
linear and bounded embedding then we often shorten id W A ,! B by A ,! B.
A subspace B of a quasi-Banach space A is called complemented if there is a linear
and bounded operator P , called a projection, with
P W A ,! A; PA D B; P 2 D P: (3.68)
If b 2 B and b D P a for some a 2 A, then
P b D P 2 a D P a D b; b 2 B: (3.69)
Hence P , restricted to B, is the identity and B D fa 2 A W P a D ag: In particular B
is a closed subspace of A. This extends naturally well-known properties of projections
in Banach spaces to quasi-Banach spaces.
The spaces Fzpq
s
./ with s > pq and Bzpq s
./ with s > p are subspaces of
0
D ./. If  is E-thick then one has the unique representations (3.49), (3.53) and
the isomorphic map I in (3.51) onto corresponding sequence spaces. In the proof of
Proposition 3.10 and in Step 1 of the proof of Theorem 3.13 we considered (3.36) as
an atomic representation in Rn (after correct normalisation of the Rn -atoms ˆjr , where
no moment conditions are needed). In particular,
kf jFzpq
s
./k  kf jFpq
s
.Rn /k; f 2 Fzpq
s
./; (3.70)
84 3 Spaces on thick domains

represented by (3.36), with p, q, s as in (3.33). Similarly for Bzpq

s
./. Let temporarily
Qs Qs s
idApq ./ be the spaces Apq ./ interpreted in this way as closed subspaces of Apq .Rn /.
Recall that AQpq ./
s x has been introduced in Definition 2.1 as a closed subspace of
s
Apq .Rn /.
Proposition 3.15. Let  be an E-thick domain in Rn according to Definition 3.1 (ii).
Let AQpq
s x and AQpq
./ s
./ be the spaces introduced in Definition 2.1 (ii). Let idAQpq
s
./
be as above.
(i) Let 0 < p < 1, 0 < q  1, s > pq . Then idFzpq s
./ is a complemented
z s x
subspace both of Fpq .R / and of Fpq ./. Furthermore,
s n

Fzpq
s x
./ D idFzpq
s
./ ˚ fh 2 Fpq
s
.Rn / W supp h  @g: (3.71)

(ii) Let 0 < p  1, 0 < q  1, s > p . Then idBzpq s

./ is a complemented
z s x
subspace both of Bpq .R / and of Bpq ./. Furthermore,
s n

Bzpq
s x
./ D idBzpq
s
./ ˚ fh 2 Bpq
s
.Rn / W supp h  @g: (3.72)

Proof. Let f 2 Fzpq s

./ with p, q, s as in part (i) be given by (3.36) now considered
s
as an atomic representation in Fpq .Rn /. If ˆjr are boundary wavelets then the related
kernels kQrj have the same meaning as in (3.39), (3.38), complemented by

kQrj D 2j n=2 ˆjr ; j 2 N0 ; (3.73)

for basic and interior wavelets ˆjr . In analogy to (3.36) with (3.50) we put

1 X
X Nj
Pf D ƒjr .f / 2j n=2 ˆjr ; f 2 Fpq
s
.Rn /; (3.74)
j D0 rD1

with Z
ƒjr .f / D 2j n=2 kQrj .x/ f .x/ dx: (3.75)
Rn
s
On the one hand, (3.74) is an atomic decomposition in Fpq .Rn / to which Theorem 1.7
s
can be applied. On the other hand, (3.75) are local means in Fpq .Rn / having the
necessary moment conditions according to Theorem 1.15. Both together result in

kPf jFpq
s
.Rn /k  c kƒ.f / jfpq
s
.Z /k  c 0 kf jFpq
s
.Rn /k: (3.76)

s
Hence, P is a linear and bounded operator in Fpq .Rn /. Furthermore, Pf 2 idFzpq
s
./.
z
If f 2 idFpq ./ then (3.74) coincides with (3.59), (3.60). Hence Pf D f . In
s

particular one has

P 2 D P in Fpq
s
.Rn /; s
PFpq .Rn / D idFzpq
s
./; (3.77)
 3.2 Wavelet bases in Axpq
s
./ 85

s
and (3.70). Hence P is a projection of Fpq .Rn / onto idFzpq
s
./. Furthermore, one has

P fh 2 Fpq
s
.Rn / W supp h  @g D 0 (3.78)

and the decomposition (3.71). This proves part (i). The proof of part (ii) is the same.

3.2.4 Porosity and smoothness zero

0
On the one hand Definition 3.11 applies also to some spaces Apq ./ of smoothness zero
in E-thick domains. On the other hand in Theorem 3.13 we excluded just these spaces.
With exception of the spaces Lp ./, 1 < p < 1, where we have the satisfactory
Theorem 2.36, spaces of smoothness zero require always some extra considerations.
We need some fractal arguments, characteristic functions as pointwise multipliers and
some more or less sophisticated interpolation. But nothing of this will be treated in
this book in detail and for its own sake. So we will be brief, restricting ourselves to
formulations, outline of proofs and references. First we complement Definition 3.1 (ii)
where we introduced E-thick domains and Definition 3.4 (i) where we said what is
meant by porosity. For this purpose we recall that a locally finite positive Radon
measure in Rn is called isotropic if there is a continuous strictly increasing function h
on the interval Œ0; 1 with h.0/ D 0, h.1/ D 1, and

.B. ; r//  h.r/ with 2 D supp ; 0 < r < 1; (3.79)

(where the equivalence constants according to (3.5) are independent of and r). For
details, discussions, and conditions ensuring that Radon measures with these properties
exist we refer to [T06], Section 1.15.1, pp. 95–97 (where we assumed that D supp
is compact, but this is immaterial). There one finds also references and examples with

h.r/ D r d ; 0  r  1; 0 < d < n (3.80)

as the most distinguished case (d -sets). We call a closed set in Rn uniformly porous if
it is porous according to Definition 3.4 (i) and if there is an isotropic Radon measure
with (3.79).

Definition 3.16. Let  be a domain .= open set/ in Rn with  6D Rn . Let D @.

(i) Then  is called E-porous if there is a number with 0 < < 1 such that one
finds for any ball B. ; r/, centred at 2 and of radius r with 0 < r < 1, a ball
B.y; r/ with

x D ;:
B.y; r/  B. ; r/ and B.y; r/ \  (3.81)

(ii) Then  is called uniformly E-porous if it is E-porous and if is uniformly

porous.
86 3 Spaces on thick domains

Remark 3.17. Porous sets (sometimes also called sets satisfying the ball condition) as
introduced in Definition 3.4 (i) play a role in fractal geometry and also in the theory of
function spaces. We refer to [T06], Definition 9.20, Remark 9.21, p. 393, Remark 2.32,
p. 146, and [T01], Sections 9.16–9.18, pp. 138–141, where one finds also some prop-
erties. In particular as also mentioned in Remark 3.5, from (3.81) one has (3.17) with,
say, =2 in place of . Hence the boundary of an E-porous domain is porous. Al-
though porosity and fractal measures are not the subject of this book we collect a few
properties which are related to our intentions. In particular we wish to say a word how
E-porous and E-thick domains introduced in Definition 3.1 are related to each other.
Proposition 3.18. Let  be a domain .D open set/ in Rn with  6D Rn and D @.
(i) If  is E-porous then  is E-thick and j j D 0.
(ii) If  is E-thick and if there is an isotropic Radon measure with (3.79) and

h.2j /  c 2.n/l h.2j l / for all j 2 N0 ; l 2 N0 ; (3.82)

and some positive numbers c and , then  is uniformly E-porous.

(iii) If  is E-thick and if there is an isotropic Radon measure with (3.79), (3.80)
.which means that is an d -set/, then  is uniformly E-porous and n  1  d < n.
Proof. Step 1. Part (i) follows easily from Definition 3.1 (ii) and [T01], Section 9.17,
pp. 138–39 (as far as j j D 0 is concerned).
Step 2. We prove (ii). According to [T01], Proposition 9.18, pp. 139–40 , (3.82)
is a (necessary and sufficient) criterion for the porosity of with (3.79). Since we
assumed that  is E-thick it follows that  is uniformly E-porous.
Step 3. We prove (iii). If is an d -set with d < n then we have (3.82) with
 D n  d . By (ii) it follows that  is uniformly E-porous. It remains to prove
0 < n  1  d < n. Since  is E-thick we may assume without restriction of
generality that

Qn1 D fx D .x 0 ; 0/ W 0  xj  1I j D 1; : : : ; n  1g   (3.83)

and that on each line fx 0 ; xn g with x 0 2 Qn1 and 0 < xn < 1 there is at least
one point 2 . To avoid a contradiction it follows by elementary reasoning that
d  n  1.
Let Apqs
.Rn / be one of the spaces according to Definition 1.1. Then m 2 L1 .Rn /
is said to be a pointwise multiplier for Apq
s
.Rn / if

f 7! mf s
generates a bounded map in Apq .Rn /: (3.84)

Of course one has to say what this multiplication means. But this is not our topic.
A careful discussion may be found in [RuS96], Section 4.2. We are interested only
in characteristic functions  of domains  as pointwise multipliers. This is a dis-
tinguished case which attracted a lot of attention. One may consult [RuS96], [T06],
Section 2.3, and the literature mentioned there. We rely here especially on [Tri03]. Let
s
Apq ./ and AQpqs
./ be the same spaces as in Definition 2.1.
 3.2 Wavelet bases in Axpq
s
./ 87

Proposition 3.19. Let  be a domain in Rn with  6D Rn . Let D @ be uniformly

porous .which means that is porous according to Definition 3.4 (i) and that there is a
Radon measure with (3.79)/. Let  be the characteristic function of . Then there
is a number ı with 0 < ı < 1 such that  is a pointwise multiplier in
 
s
Fpq .Rn /; 1 < p < 1; 1  q  1; ı p1  1 < s < pı (3.85)

and in 1 
s
Bpq .Rn /; 1 < p < 1; 0 < q  1; ı p
1 <s < ı
p
: (3.86)
Furthermore,
s
Fpq ./ D Fzpq
s
./ (3.87)
with p, q, s as in (3.85), and
s
Bpq ./ D Bzpq
s
./ (3.88)

with p, q, s as in (3.86).
Proof. The multiplier assertions are essentially covered by [Tri03], Corollaries 3, 4,
where one may choose ı D  with  as in (3.82). (There we assumed that  is
bounded, but this is immaterial.) In particular, ext,
(
f .x/ if x 2 ,
.ext f /.x/ D (3.89)
0 if x 2 Rn n;
s
is a linear and bounded extension operator from Fpq ./ with p, q, s as in (3.85) into
s n s s
Fpq .R / and from Bpq ./ with p, q, s as in (3.86) into Bpq .Rn /. Then one obtains
(3.87) and (3.88).
Remark 3.20. Let Apq s
.Rn / be either Fpq
s
.Rn / in (3.85) or Bpq
s
.Rn / in (3.86). Then
one has as a by-product

fh 2 Apq
s
.Rn / W supp h  @g D f0g: (3.90)

Otherwise we refer for technicalities, more general assertions and the related literature
to [RuS96], [T06] and in particular to [Tri03].
Now one can complement Theorem 3.13 by spaces of smoothness zero. All notation
have the same meaning as there. This applies also to diverse technicalities.
Proposition 3.21. Let  be an uniformly E-porous domain in Rn according to Defi-
nition 3.16 (ii) with  6D Rn . Let for u 2 N,
˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj x
with Nj 2 N; (3.91)

be an orthonormal u-wavelet basis in L2 ./ according to Theorem 2.33 and Defini-

tions 2.31, 2.4.
88 3 Spaces on thick domains

(i) Then f 2 D 0 ./ is an element of

0
Fpq ./; 1 < p < 1; 1  q < 1; (3.92)

if, and only if, it can be represented as

1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 fpq
0
.Z /; (3.93)
j D0 rD1

unconditional convergence being in D 0 ./. Furthermore, if f 2 Fpq

0
./ then the
representation (3.93) is unique with  D .f /,
 
jr D jr .f / D 2j n=2 f; ˆjr (3.94)

and ˚  
I W f 7! .f / D 2j n=2 f; ˆjr (3.95)
˚
is an isomorphic map of Fpq
0
./ onto fpq
0
.Z /. Furthermore, ˆjr is an unconditional
basis in Fpq
0
./.
(ii) Then f 2 D 0 ./ is an element of
0
Bpq ./; 1 < p < 1; 0 < q  1; (3.96)

if, and only if, it can be represented as

1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 bpq
0
.Z /; (3.97)
j D0 rD1

unconditional convergence being in D 0 ./. Furthermore, if f 2 Bpq 0

./ then the
representation (3.97) is unique with  D .f / as in (3.94) and I in (3.95)
˚ is an
isomorphic map of Bpq 0
./ onto bpq
0
.Z /. If, in addition, q < 1 then ˆjr is an
unconditional basis in Bpq
0
./.
Proof. By Proposition 3.18 (i) the domain  is E-thick. Then one can apply Theo-
s
rem 3.13 to the spaces Fpq ./ covered by (3.85), (3.87) with s 6D 0 (recall that ı < 1).
If s > 0 is chosen sufficiently small such that ext in (3.89) is an extension operator
s s
both for Fpq ./ and Fpq ./ then the complex interpolation
s
s
Fpq ./; Fpq ./ 1 D Fpq
0
./ (3.98)
2

follows from the well-known Rn -counterpart

s
s
Fpq .Rn /; Fpq .Rn / 1 D Fpq
0
.Rn /; (3.99)
2

where 1 < p < 1, 1  q < 1. By real interpolation one obtains

 s s

Bpq ./; Bpq ./ 1 ;q D Bpq
0
./ (3.100)
2
 3.2 Wavelet bases in Axpq
s
./ 89

where 1 < p < 1, 0 < q  1. There are counterparts for the corresponding se-
s s
quence spaces fpq .Z / and bpq .Z /. As far as complex and real interpolation and
applications of common extension operators are concerned we refer to [T06], Sec-
tion 1.11.8, pp. 69–72, where we dealt with problems of this type in case of bounded
Lipschitz domains. A few comments about interpolation may also be found in Sec-
tion 4.3.1 below. The arguments in the Step 2 of the proof of [T06], Theorem 4.22,
p. 213, show how one can incorporate the desired interpolation formulas for the related
sequence spaces. We used this type of argument also in connection with the proof
of [T06], Corollary 4.28, pp. 216–17. Now the isomorphic mapping properties of I
extend from 0 < jsj small to s D 0. This completes the proof.

Remark 3.22. The proof of Theorem 3.13 relies on the observation that, according to
Theorem 1.7, one does not need moment conditions for atoms if s > p or s > pq ,
and that, according to Theorem 1.15, one does not need moment conditions for kernels
of local means if s < 0. This excludes in particular spaces with s D 0. The above
considerations might be considered as an attempt to circumvent this handicap. One has
to pay a price: We strengthened the natural class of E-thick domains in Theorem 3.13
by the smaller class of uniformly E-porous domains in Proposition 3.21. On the one
hand the notation of porosity (also called ball condition) is closely related to some
problems in the theory of function spaces, especially to the question under which
conditions the characteristic function  of a domain  is a pointwise multiplier in
spaces of smoothness zero. We refer to [T06], Remark 2.32, p. 146, and in particular to
[Tri02], [Tri03]. On the other hand one may ask whether (first) moment conditions for
atoms or kernels of local means for spaces of smoothness zero are really necessary. But
this is the case. It came out quite recently that (first) moment conditions for atoms in
spaces of smoothness zero are indispensable, even for L2 .Rn /. We refer to [Vyb08a].
Corresponding assertions for spaces with 0 < p  1 and s D n p1  1 may be found
in [Sch08].

3.2.5 The spaces Axpq

s
./ II

It seems to be reasonable to summarise the above considerations and to clip together

Theorem 3.13 and Proposition 3.21. We do not repeat the technicalities preceding
s s
Theorem 3.13. Let bpq .Z / and fpq .Z / be the same sequence spaces as in Defini-
tion 2.6.

Theorem 3.23. Let  be an uniformly E-porous domain in Rn according to Defini-

tion 3.16 (ii) with  6D Rn . Let for u 2 N,
˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj x
with Nj 2 N; (3.101)

be an orthonormal u-wavelet basis in L2 ./ according to Theorem 2.33 and Defini-

tions 2.31, 2.4.
90 3 Spaces on thick domains

(i) Let Fxpq

s
./ be the spaces in (3.45) and let

u > max.s; pq  s/: (3.102)

Then f 2 D 0 ./ is an element of Fxpq

s
./ if, and only if, it can be represented as

1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 fpq
s
.Z /; (3.103)
j D0 rD1

unconditional convergence being in D 0 ./ and locally in any space Fxpq


./ with
x
 < s. Furthermore, if f 2 Fpq ./ then the representation (3.103) is unique with
s

 D .f /,  
jr D jr .f / D 2j n=2 f; ˆjr ; (3.104)
and ˚  
I W f 7! .f / D 2j n=2 f; ˆjr (3.105)
˚
is an isomorphic map of Fxpq s
./ onto fpq
s
.Z /. If, in addition, q < 1 then ˆjr is
an unconditional basis in Fxpq s
./.
x
(ii) Let Bpq ./ be the spaces in (3.46) and let
s

u > max.s; p  s/: (3.106)

Then f 2 D 0 ./ is an element of Bxpq

s
./ if, and only if, it can be represented as

1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 bpq
s
.Z /; (3.107)
j D0 rD1

unconditional convergence being in D 0 ./ and locally in any space Bxpq


./ with
x
 < s. Furthermore, if f 2 Bpq ./ then the representation (3.107) is unique with
s

 D .f / as in (3.104), and I in (3.105) is an isomorphic map of Bxpqs

./ onto
˚ j
bpq .Z /. If, in addition, p < 1, q < 1, then ˆr is an unconditional basis in
s

Bxpq
s
./.
Proof. By Proposition 3.18 the domain  is E-thick. Then the above assertion follows
from Theorem 3.13 and Proposition 3.21.
Remark 3.24. If u 2 N is given then I in (3.105) is a common isomorphic map
for all Axpq
s
./ with (3.102), (3.106) onto respective sequence spaces. This gives the
possibility to shift mapping problems between these function spaces to corresponding
assertions between sequence spaces where they often can be treated more easily. We
refer to [DNS06a], [DNS06b], [DNS07] and the literature listed there. In numerical
analysis preference is given (so far) to bounded Lipschitz domains which are covered
by the above considerations.
 3.3 Homogeneity and refined localisation, revisited 91

Corollary 3.25. Let  be either a bounded Lipschitz domain in Rn with n  2 accord-

ing to Definition 3.4 (iii) or the classical snowflake domain from Proposition 3.8 (iii) and
Figure 3.5, p. 76, in R2 . Then  is uniformly E-porous according to Definition 3.16 (ii)
and Theorem 3.23 can be applied.

Proof. By Proposition 3.8 the domain  is E-thick in both cases. If  is a bounded

Lipschitz domain in Rn with n  2 then D @ is an .n  1/-set. The boundary of
the above snowflake domain is the so-called snowflake curve (Koch curve). It is an d -set
4
with d D log
log 3
(Hausdorff dimension). We refer, for example, to [Mat95], pp. 66–68.
Some relevant calculations may also be found in [T06], pp. 359–60. Then it follows
from Proposition 3.18 (iii) that  is uniformly E-porous. In particular, Theorem 3.23
can be applied.

Remark 3.26. There is little doubt that one can find for any d with n  1  d < n
a bounded E-thick domain  in Rn , n  2, such that D @ is an d -set. Then it
follows from Proposition 3.18 (iii) that  is uniformly E-porous and one can apply
Theorem 3.23. We refer to [T97], Section 16.4, p. 123, based on [T97], Theorem 16.2,
Remark 16.3, pp. 120–22, where we constructed bounded star-like domains  such
that D @ is an d -set with n  1 < d < n. However it was noticed by A. Carvalho
that some arguments given there are not correct. He gave a modified proof in [Car06].
The construction produces (presumably) also an E-thick domain. In Theorem 5.43
we complement the above assertions in case of so-called bounded cellular domains,
covering some spaces so far excluded in Theorem 3.23.

3.3 Homogeneity and refined localisation, revisited

3.3.1 Introduction

We return to Section 2.2. For the spaces Axpq s

.U / from Definition 2.9 we claimed the
homogeneity (2.43). According to Remark 2.12 we know this property so far for the
spaces Fxpq
s
.U / with (2.45), hence the first line in (2.40), but not for the other cases. It
will be our first aim to seal this gap now. This will be done in Section 3.3.2. It justifies
afterwards the pointwise multiplier Theorem 2.13 for all cases and the assertions about
the refined localisation spaces from Section 2.2.3. This will not be repeated. In
Proposition 3.10 we proved that
s;rloc
Fpq ./ D Fzpq
s
./; 0 < p; q  1; s > pq ; (3.108)

(q D 1 if p D 1) extending (2.63) from bounded C 1 domains, [T01], and bounded

Lipschitz domains, [T06], to E-thick domains. It is the second aim of the present
Section 3.3 to prove
s;rloc
Fpq ./ D Fpq
s
./; 0 < p; q  1; s < 0; (3.109)
92 3 Spaces on thick domains

(q D 1 if p D 1) for the other spaces in Definition 2.14 if  is E-thick, comple-

mented by a corresponding assertion if s D 0. This will be done in Section 3.3.3. Both
(3.108) and (3.109) (complemented by a corresponding assertion for s D 0) cover also
(2.42) what justifies finally the notation used in (2.40).

3.3.2 Homogeneity: Proof of Theorem 2.11

It does not matter whether U in (2.39) are balls or cubes. To fix the imagination let
us assume that
U D fx 2 Rn W jxj < g; 0 <   1; (3.110)
are balls. Let U D U1 . Let Axpq
s
.U / with A D B or A D F be one of the spaces
according to Definition 2.9. Then we have to prove that

kf .  / jAxpq .U /k  s p kf jAxpq

n
s s
.U /k; 0 <   1; (3.111)

where the equivalence constants are independent of  and of f 2 Axpq

s
.U /. As
mentioned in Remark 2.12 this assertion is known for
s;rloc
Fpq .U / D Fxpq
s
.U / D Fzpq
s
.U / with s > pq ;

hence the first line in (2.40). This crucial observation has been used several times also
after Section 2.2. It is one of the cornerstones of the theory developed in this book. Now
we prove the remaining assertions which we did not use after Section 2.2, especially
not to obtain the wavelet expansions in Theorems 3.13, 3.23 (Corollary 3.25) on which
we now rely.
Step 1. We begin with a preparation. Let Apq s
./ be one of the spaces according
to Definition 2.1 (i) in a domain  in R . Let 0 < c1 < c2 < 1, typically c1 D 1=2
n

and c2 D 2, and let c1 < c < c2 . Let

c  D fy 2 Rn W y D cx; x 2 g : (3.112)

Then
kf .c 1  / jApq
s
.c/k  kf jApq
s
./k; f 2 Apq
s
./; (3.113)
where the equivalence constants depend on c1 ; c2 , but not on c and . This is known
(and can be checked easily) if  D Rn . One obtains this assertion for arbitrary
domains  by restriction according to Definition 2.1 (i). Similarly for (3.113) with
AQpq
s s
./ according to Definition 2.1 (ii) in place of Apq ./.
Step 2. By Step 1 it is sufficient to deal with  D 2k where k 2 N0 . Let
for brevity U2k D U k in (3.111) with U D U 0 (the unit ball). With  D U
in Definition 3.11 one obtains the spaces in Definition 2.9. Hence we have for all
these spaces wavelet representations which we constructed explicitly in the proofs of
Proposition 3.10 (s > pq ), Theorem 3.13 (s 6D 0) and Proposition 3.21 (s D 0). In all
cases we obtained the same type of representation, say, for example (3.49) with (3.50),
 3.3 Homogeneity and refined localisation, revisited 93

based on the orthonormal u-wavelet basis fˆjr g in (3.47). This gives the possibility to
compare such expansions in two domains which are related to each other by a dilation
of type (3.112). This applies in particular to the unit ball U as a reference domain and
the dilated balls U k . In this dyadic situation the functions fˆjr g in U are transformed
into the corresponding functions in U k . But this applies not only to the functions fˆjr g
but also to the complements of the kernels krj in (3.37) by (3.39) and of the atoms
arj in (3.55) by (3.56) if moment conditions are needed. Afterwards one has only
Rn -estimates, both for local means and for atoms based exclusively on Theorems 1.7,
1.15. In other words the equivalence constants for the isomorphic map I in (3.51) can
be chosen in all cases covered by Theorem 3.13, hence s 6D 0, independently of k 2 N.
This applies afterwards also to s D 0 by the arguments in the proof of Proposition 3.21.
In other words it remains to control what happens with the coefficients in the related
expansions if one applies the dilation

Dk W x 7! 2k x; U k D Dk U; k 2 N: (3.114)

We deal with the case s 6D 0 covered by Theorem 3.13. One can incorporate s D 0
afterwards by the arguments in the proof of Proposition 3.21 what will not be done in
detail. Let f 2 Axpqs
.U k / covered by Theorem 3.13 where again A D B or A D F .
Similarly apq .Zk / with a D b or a D f and Zk D ZU k , where Z0 D ZU refers to
s

the unit ball U . Then

kf jAxpq
s
.U k /k  k.f /k japq
s
.Zk /k (3.115)

with .f /k D fjr .f /k g, where

Z
jr .f /k D 2j n=2 f .x/ ˆjr .x/k dx; .j; r/ 2 Zk ; (3.116)
Uk

(appropriately interpreted if s < 0) where ˆjr .  /k refers to the system (3.47) with
 D U k . Let ˆjr .  / D ˆjr .  /0 and jr .f / D jr .f /0 . On the other hand we can
expand f .2k  / 2 Axpq
s
.U / according to, say, (3.49) by
  X  
f 2k x D jr f .2k  / 2j n=2 ˆjr .x/ (3.117)
.j;r/2Z0

with (3.50), hence

Z
 k

jr f .2 / D 2 j n=2
f .2k x/ ˆjr .x/ dx; .j; r/ 2 Z0 : (3.118)
U

By the constructions resulting in Theorem 2.33, based on Definitions 2.4, 2.31, the
wavelet bases for L2 .U / and L2 .U k / are related by

ˆjr .2k x/ D 2k n=2 ˆjr Ck .x/k ; x 2 Uk ; .j; r/ 2 Z0 ; (3.119)

94 3 Spaces on thick domains

with the left-hand side considered as an orthonormal basis in L2 .U / and the right-hand
side considered as an orthonormal basis in L2 .U k /. To stick at the same r it is suitable
to accept an index-shifting in Zk for j from N0 to k C N0 . Then one has by (3.118)
that
Z
 
jr f .2k  / D 2.j Ck/n=2 f .x/ ˆjr Ck .x/k dx D jr Ck .f /k : (3.120)
Uk

Using (3.119), (3.120) one obtains by (3.117) that

X
f .x/ D jr Ck .f /k 2j n=2 ˆjr .2k x/
.j;r/2Z0
X (3.121)
D jr .f /k 2j n=2 ˆjr .x/k ; x 2 U k;
.j;r/2Zk

where we replaced j C k by j in the last line and used the above index-shifting in Zk .
This is the desired representation both for the B-spaces and the F -spaces.
Step 3. After these preparations one can now prove (3.111) with  D 2k and
U D U k as follows. By Theorem 3.13 (ii) and Definition 2.6 we have
   s 
kf .2k  / jBxpq
s
.U /k   f .2k  / jbpq .Z0 /
X Xˇ  
ˇ q=p 1=q (3.122)
D
n
2j.s p /q ˇj f .2k  / ˇp :
r
j r

Inserting (3.120) one obtains by the above-mentioned index-shifting

kf .2k  / jBxpq
n
s
.U /k  2k.s p / k.f /k jbpq
s
.Zk /k
(3.123)
 2k.s p / kf jBxpq
n
s
.U k /k:

This proves (3.111) for the B-spaces. Similarly it follows for the F -spaces from
Theorem 3.13 (i) and Definition 2.6 that
   s 
kf .2k  / jFxpq
s
.U /k   f .2k  / jfpq .Z0 /
 X ˇ   ˇq 1=q 
 
D 2jsq ˇjr f .2k  / jr .  /ˇ jLp .U /
.j;r/2Z0
 X ˇ ˇq 1=q 
 
 2sk  2.j Ck/sq ˇjr Ck .f /k j Ck;r .2k  /ˇ jLp .U / (3.124)
.j;r/2Z0
n 
 X ˇ ˇq 1=q 

 2k.s p /  2jsq ˇjr .f /k j;r .  /ˇ jLp .U k /
.j;r/2Zk
 
2
n
k.s p / f jFxs .U k /:
pq

This proves (3.111) for the F -spaces.

 3.3 Homogeneity and refined localisation, revisited 95

s;rloc
3.3.3 Wavelet bases in Fpq ./, revisited

After the complete proof of Theorem 2.11 we know now that also all the other asser-
tions in Section 2.2 are valid. This applies in particular to the pointwise multiplier
Theorem 2.13 and the properties of the refined localisation spaces according to Defi-
s;rloc
nition 2.14 and Theorem 2.16. For the spaces Fpq ./ with s > pq we have the
wavelet representation in Theorem 2.38. It is the first aim of this Section 3.3.3 to ex-
s;rloc
tend this assertion to other refined localisation spaces Fpq ./ with s  0. Secondly
we obtain as a corollary (3.109) under the assumption that  is E-thick. This can be
rephrased as the refined localisation property of the related spaces Fpq s
./.
We use the same notation and references as in Theorem 2.38.
s;rloc
Theorem 3.27. Let  be an arbitrary domain in Rn with  6D Rn . Let Fpq ./ be
˚ j
the spaces as introduced in Definition 2.14. Let ˆr be an orthonormal u-wavelet
basis in L2 ./ according to Theorem 2.33 and Definitions 2.31, 2.4 with
u > max.s; pq  s/: (3.125)
Then f 2 D 0 ./ is an element of s;rloc
Fpq ./ if, and only if, it can be represented as
1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 fpq
s
.Z /: (3.126)
j D0 rD1
s;rloc
Furthermore, if f 2 Fpq ./ then the representation (3.126) is unique with  D
.f /, Z
 
r .f / D 2
j j n=2
f; ˆr D 2
j j n=2
f .x/ ˆjr .x/ dx (3.127)

.appropriately interpreted/ and
˚  
I W f 7! .f / D 2j n=2 f; ˆjr (3.128)
s;rloc
is an isomorphic map of Fpq ./ onto fpq
s
.Z /,
kf jFpqs;rloc
./k  k.f / jfpq
s
.Z /k (3.129)
˚ j
.equivalent quasi-norms/. If p < 1, q < 1 then ˆr is an unconditional basis in
s;rloc
Fpq ./.
Proof. Step 1. If s > pq (and hence u > s) then the above assertion coincides
essentially with Theorem 2.38. Hence one has to prove the theorem for the remain-
ing spaces with s  0. As far as formulations are concerned one may also consult
Theorems 3.13 (i) and 3.23 (i).
Step 2. Let s  0 and let f be given by (3.126). We decompose f with respect to
1
the Whitney cubes Qlt in (2.22),
1 X
X X
f D flt with flt D jr .l; t / 2j n=2 ˆjr ; (3.130)
lD0 t j;r
96 3 Spaces on thick domains

1
where each flt is a partial sum of (3.126) collecting terms subordinated to Qlt by the
j j
construction in (2.22), (2.23) indicated by r .l; t / which are either r or zero such
that there is an one-to-one relation between
[
fjm g and fjr .l; t /g: (3.131)
l;t

By the homogeneity (2.43) with Axpq

s
D Fpq
s
(now justified) and Theorem 3.23 (i) with
Fxpq D Fpq it follows that
s s

kflt jFpq
s 1
.Qlt k  k.l; t / jfpq
s
.Z /k; (3.132)

where .l; t/ collects the related terms jr .l; t / with equivalence constants which are
independent of l and t . One has
X
kfl;t jFpq
s 1 p
.Qlt k  k jfpq s
.Z /kp (3.133)
l;t

(modification if p D q D 1). Let f%lt g be a resolution of unity as used in Defini-

tion 2.14. Then
%lt f D %lt .flt C C/ (3.134)
where CC indicates some neighbouring terms. By the pointwise multiplier Theo-
rem 2.13 applied to g D %lt one obtains by (2.59) and (3.133), (3.134) that f 2
s;rloc
Fpq ./ and
kf jFpq
s;rloc
./k  c k jfpq
s
.Z /k: (3.135)

By the orthogonality of fˆjr g one has in the same way as before jr D jr .f / with
(3.127).
s;rloc
Step 3. Conversely let f 2 Fpq ./. Then we have %lt f 2 Fpq
s 1
.Qlt / and one
obtains by Theorem 3.23 (i) with Fxpq D Fpq ,
s s

k.%lt f / jfpq
s
.Z /k  k%lt f jFpq
s 1
.Qlt /k: (3.136)

Again it follows from the homogeneity (2.43) that the equivalence constants in (3.136)
are independent of l and t (explicit calculations may be found in Step 3 of the proof in
Section 3.3.2). But then it follows from (2.55), (2.59) that .f / 2 fpq
s
.Z / and

k.f / jfpq
s
.Z /k  c kf jFpq
s;rloc
./k: (3.137)

s;rloc
The representability of f 2 Fpq ./ by (3.126) (with  D .f /) follows by the
above arguments from the representability of %lt f 2 Fpq
s 1
.Qlt /.
s
As before we put F11 D B11
s
.
 3.3 Homogeneity and refined localisation, revisited 97

Theorem 3.28. (i) Let  be an E-thick domain in Rn according to Definition 3.1 (ii).
Let )
either 0 < p  1, 0 < q  1, s > pq ,
(3.138)
or 0 < p  1, 0 < q  1, s < 0

.q D 1 if p D 1/. Let Fpq s;rloc

./ and Fxpq s
./ be the spaces according to Defini-
tions 2.14 and 3.11. Then
Fpqs;rloc
./ D Fxpq
s
./: (3.139)
(ii) Let  be an uniformly porous domain in Rn according to Definition 3.16 (ii).
Then
0;rloc
Fpq ./ D Fpq
0
./; 1 < p < 1; 1  q < 1; (3.140)
0;rloc
where Fpq ./ has the same meaning as in Definition 2.14.

Proof. If s > pq then (3.139) is covered by Proposition 3.10 as a consequence of the
fact that the two spaces have the same wavelet representations. But this applies also to
all other cases, comparing Theorem 3.13, Proposition 3.21 with Theorem 3.27.

Remark 3.29. In particular one obtains (3.109) as a special case, complementing

(3.108).

3.3.4 Duality

Recall the well-known duality assertion

s
Bpq .Rn /0 D Bps n
0 q 0 .R /; s 2 R; (3.141)

with
1 1 1 1
1  p; q < 1; C 0 D C 0 D 1; (3.142)
p p q q
s
where Bpq .Rn / are the spaces according to Definition 1.1. Here (3.141) must be inter-
preted within the dual pairing .S.Rn /; S 0 .Rn // or the dual pairing .D.Rn /; D 0 .Rn //.
This requires that S.Rn / and D.Rn / are dense in Bpq s
.Rn /, what is the case (since we
excluded p D 1 and/or q D 1 in (3.142)). This is a classical assertion which may
be found in [T78], Section 2.6.1, pp. 198–99, and [T83], Theorem 2.11.2, p. 178. It
s
can be extended to spaces with p < 1 and also to the spaces Fpq .Rn /. We refer again
to [T83], Section 2.11. But this is not of interest in the present context. We ask for
counterparts of (3.141) with domains  in place of Rn . There are assertions of this type
for the spaces Bzpqs x
./ according to Definition 2.1 with respect to arbitrary domains 
in [T78], Section 4.8.1, p. 332, with the expected outcome. But this was a little bit too
s x
bold at least as far as the interpretation of Bzpq ./ as a subspace of D 0 ./ is concerned.
In some sense we seize the opportunity and return to this point (after more than thirty
years) using now the above wavelet characterisations.
98 3 Spaces on thick domains

Theorem 3.30. Let Bxpq

s
./ be the same spaces as in (3.46) where  is an E-thick
domain in R according to Definition 3.1 (ii) with
n

1 1 1 1
0 6D s 2 R; 1  p; q < 1; C 0 D C 0 D 1: (3.143)
p p q q

Then D./ is dense in Bxpq

s
./ and

Bxpq
s
./0 D Bxps
0 q 0 ./ (3.144)

are the dual spaces interpreted within the dual pairing .D./; D 0 .//.
Proof. Step 1. Let s; p; q be as in (3.143). It follows from Theorem 3.13 (ii) that I
s
in (3.51) is an isomorphic map onto bpq .Z / normed by (2.37). Then one obtains
by (3.53) that finite linear combinations of ˆjr are dense in Bxpq
s
./. Since ˆjr 2
C u .Rn / has compact support in  it can be approximated within  by D./-functions
(Sobolev’s mollification). Hence D./ is dense in Bxpqs
./.
Step 2. Since D./ is dense in Bxpq ./ with (3.143) it makes sense to interpret
s

its dual within the dual pairing .D./; D 0 .//. We complement bpq s
.Z / in Defini-
N
tion 2.6 by `q .`p j / consisting of all sequences
˚
D jr 2 C W j 2 N0 I r D 1; : : : ; Nj ; x
Nj 2 N; (3.145)

with
X
1 X
Nj q=p 1=q
N
k j`q .`p j /k D j rj
j p
< 1: (3.146)
j D0 rD1

If 1  p; q < 1 then one has

N N
`q .`p j /0 D `q 0 .`p0j / (3.147)

for the dual spaces interpreted as usual as the dual pairing

1 X
X Nj
0 0j N 0 N
h ; iD j
r r ; 2 `q .`p j /; 2 `q 0 .`p0j /: (3.148)
j D0 rD1

N
To adapt I in (3.51) to `q .`p j / we put

J W f 7! .f / D f j
r .f /g; x
Nj 2 N (3.149)

with Z
n n
j
r .f / D 2j.s p C 2 / f .x/ ˆjr .x/ dx (3.150)

(usually interpreted if s < 0). Then

J Bxpq
N
s
./ D `q .`p j /: (3.151)
 3.3 Homogeneity and refined localisation, revisited 99

For the dual operator J 0 one has the isomorphic map

Bxpq
N N
J 0 W `q 0 .`p0j / D `q .`p j /0 onto s
./0 ; (3.152)

interpreting the dual of Bxpqs

./ as a subspace of D 0 ./. In particular one obtains with
f 2 Bxpq
N
s
./ and 0 2 `q 0 .`p0j / that
X Z
n
0 0j j.s p Cn /
hJf; i D r 2
2 f .x/ ˆjr .x/ dx
j;r 
Z X n n (3.153)
D f .x/ 0j
r 2j.s p C 2 / ˆjr .x/ dx
 j;r
0 0
D .f; J /:
Hence X
J0 0
D 0j
r 2
j n=2 j
ˆr (3.154)
j;r

with
j.s p Cn/ n j.s pn0 /
0j
r D2
0j
r D2 0j
r : (3.155)
One obtains
N
k0 jbps0 q 0 .Z /k D k 0
j`q 0 .`p0j /k: (3.156)
Then (3.144) follows from (3.154), (3.156) and Theorem 3.13.
Remark 3.31. We excluded s D 0 in Theorem 3.30 because one needs in this case
that the underlying domain  is not only E-thick but even uniformly E-porous what
is a stronger assumption according to Definition 3.16 and Proposition 3.18 (i).
Corollary 3.32. Let Bxpq
0
./ D Bpq
0
./ be the same spaces as in (3.46) where  is
an uniformly E-porous domain in Rn according to Definition 3.16 (ii). If
1 1 1 1
1 < p < 1; 1  q < 1; C 0 D C 0 D 1; (3.157)
p p q q
then D./ is dense in Bpq
0
./ and
0
Bpq ./0 D Bp00 q 0 ./ (3.158)
are the dual spaces interpreted within the dual pairing .D./; D 0 .//.
Proof. One can argue in the same way as in the proof of Theorem 3.30 using now
Proposition 3.21 (ii) instead of Theorem 3.13 (ii).
Corollary 3.33. The assertions of Theorem 3.30 and Corollary 3.32 apply in particular
to the spaces Bxpq
s
./ with (3.143) and Bxpq0
./ with (3.157) if  is either an uniformly
E-porous domain according to Definition 3.16 or a bounded Lipschitz domain accord-
ing to Definition 3.4 (iii) or the classical snowflake domain from Proposition 3.8 (iii)
and Figure 3.5, p. 76, in R2 .
100 3 Spaces on thick domains

Proof. By Proposition 3.18 E-porous domains are in particular E-thick. Hence one can
apply both Theorem 3.30 and Corollary 3.32. The corresponding assertions for bounded
Lipschitz domains and the classical snowflake domain follow now from Corollary 3.25.

Remark 3.34. The proofs of Theorem 3.30 and Corollary 3.32 rely on the existence
of a common wavelet basis in Bxpqs
./ and its anticipated dual. One can try to apply
this argument to other cases. A distinguished example is the special case
s;rloc
Bpp ./ D Fpp
s;rloc
./; s 2 R; 1 < p < 1; (3.159)

of the refined localisation spaces according to Definition 2.14. We formulate the out-
come.
Corollary 3.35. Let  be an arbitrary domain in Rn with  6D Rn . Then D./ is
s;rloc
dense in the refined localisation spaces Bpp ./ according to Definition 2.14 and
(3.159). Furthermore,
s;rloc
Bpp ./0 D Bps;rloc
0 p0 ./; 1
p
C 1
p0
D 1; (3.160)

are the dual spaces interpreted within the dual pairing .D./; D 0 .//.
Proof. Using Theorem 3.27 one can apply the arguments from the proof of Theo-
rem 3.30 with
s
fpp .Z / D bpp
s
.Z / (3.161)
according to Remark 2.7.
Chapter 4
The extension problem

4.1 Introduction and criterion

4.1.1 Introduction
s
We introduced in Definition 2.1 the spaces Apq ./ on arbitrary domains (D open sets)
in R with  6D R by restriction of Apq .R / to . In particular, Apq
n n s n s
./ is considered
0
as a subset of D ./ and re,

re g D gj W Apq
s
.Rn / ,! Apq
s
./; (4.1)

is the linear and bounded restriction operator according to (2.2). One of the most
fundamental problem in the theory of function spaces is the question of whether there
is a linear and bounded extension operator ext,

ext W Apq
s s
./ ,! Apq .Rn /; ext f j D f: (4.2)

Using (4.1) this can also be written as

re B ext D id s
(identity in Apq ./): (4.3)

A first discussion including relevant references may be found in Remark 3.3. It is mainly
a question about the quality of the underlying domain . The nowadays classical
assertions about extensions may be found in [T92], Theorem 5.1.3, p. 239, where 
is a bounded C 1 domain. The final solution of the extension problem for bounded
Lipschitz domains  in Rn goes back to [Ry99] where V. S. Rychkov constructed an
s
universal extension operator which applies to all spaces Apq ./. For a corresponding
formulation one may also consult [T06], Section 1.11.5. As we indicated in Remark 3.3
beyond Lipschitz domains the extension problem is rather tricky. It may happen that
the constructed operator, for example Extpk in (3.13) is not an extension operator for
other spaces. To ensure that there are extension operators is of interest for its own sake,
but there are applications where one needs common extension operators at least in some
.s; p; q/-regions. In contrast to the literature mentioned above and in Remark 3.3 we
deal with the extension problem in the framework of wavelet bases.

4.1.2 A criterion

Let  be a (non-empty) arbitrary domain (D open set) in Rn with  6D Rn and let

c D Rn n  be its (non-empty) closed complement in Rn . Let s; p; q be as in (2.3)
102 4 The extension problem

with p < 1 for the F -spaces. Then

˚
AQpq
s
.c / D f 2 Apq
s
.Rn / W supp f  c (4.4)
s
is a closed subspace of Apq .Rn /. This extends (2.6) to arbitrary closed (non-empty)
sets D  in R with 6D Rn . At the beginning of Section 3.2.3 we recalled that a
c n

closed subspace of a quasi-Banach space is called complemented if there is a projection

P with (3.68), (3.69). There we explained also that ,! means linear and continuous
embedding.
Theorem 4.1. Let  be an arbitrary .non-empty/ domain in Rn with  6D Rn and let
s
Apq ./ be a space according to Definition 2.1. Then there exists an extension operator
ext with (4.2) if, and only if, AQpq
s
.c / is a complemented subspace of Apq
s
.Rn /.
Proof. Step 1. Let ext be an extension operator for Apq
s
./ according to (4.2), (4.3).
Then
P D ext B re W Apq
s
.Rn / ,! Apq
s
.Rn / (4.5)
is a linear and bounded map. By (4.3) one has

P 2 D ext B re B ext B re D P: (4.6)

s
Hence P is a projection of Apq .Rn / onto PApq
s
.Rn /. By P B ext D ext it follows that
ext maps Apq ./ into PApq .R /. If f D PApq .Rn / then f D extŒre f . Hence ext
s s n s
s s
maps Apq ./ onto PApq .Rn /. From

kf jApq
s
./k  k ext f jApq
s
.Rn /k (4.7)

if follows that
ext W Apq
s
./ () PApq
s
.Rn / (4.8)
is an isomorphic map. Since P is a projection one obtains that Q D id  P is also a
projection. Furthermore, since

Qf j D f j  Pf j D 0; f 2 Apq
s
.Rn /; (4.9)

one obtains that supp Qf  c . Conversely if f 2 Apq s

.Rn / with supp f  c
then Pf D 0 and hence f D Qf . In other words, QApq .Rn / D AQpq
s s
.c /. Hence
AQpq
s
.c / is a complemented subspace of Apq
s
.Rn /.
Step 2. Let AQpq
s
.c / be a complemented subspace of Apq
s
.Rn / and let Q be the
corresponding projection. Let P D id  Q and

ext f D P g; f 2 Apq
s
./; g 2 Apq
s
.Rn /; gj D f: (4.10)

If h 2 AQpq
s
.c / then P h D h  Qh D 0. Hence, ext f is independent of the chosen
g with the indicated properties. In particular, ext is a linear operator, and

k ext f jApq
s
.Rn /k  c inf kg jApq
s
.Rn /k  kf jApq
s
./k: (4.11)
gjDf
 4.2 Main assertions 103

Furthermore,
ext f j D f j  Qf j D f j: (4.12)
Hence ext is a (linear and bounded) extension operator.
Remark 4.2. If Q is a common projection in, say, Bpq
s
.Rn / with
.s; p; q/ 2 RB  R  .0; 1  .0; 1 ; (4.13)
s
then ext is a common extension operator for the corresponding spaces Bpq ./ in the
same region RB . Similarly for the F -spaces, or both B-spaces and F -spaces. This
follows from the construction. Here common means that, say, ext is defined on the
union of the spaces in questions,
[
dom.ext/ D s
Bpq ./ (4.14)
.s;p;q/2RB
s
such that its restriction to each admitted space Bpq ./ has the indicated property.
Remark 4.3. In particular one has always extension operators for the Hilbert spaces
H s ./ D B2;2s
./ with s 2 R in arbitrary domains . (Recall that an infinite-
dimensional Banach space is isomorphic to a Hilbert space if, and only if, every closed
subspace is complemented, [AlK06], Theorem 12.4.4, p. 305). But this does not say
very much. We are looking for common extension operators in some .s; p; q/-regions
as explained in the previous remark. We discussed these questions in Section 4.1.1 and
in Remark 3.3 where one finds also some references.

4.2 Main assertions

4.2.1 Positive smoothness

In (4.2) and Remark 4.2 we explained what is meant by an extension operator and a
common extension operator. As in (3.44) we put
   
1 1
p D n 1 and pq Dn 1 ; 0 < p; q  1: (4.15)
p C min.p; q/ C
s s
The spaces Bpq ./ and Fpq ./ have the same meaning as in Definition 2.1.
Theorem 4.4. Let  be a .non-empty/ I -thick domain in Rn according to Defini-
x 6D Rn and j@j D 0. Then for any u > 0 there is a common
tion 3.1 (iii) with 
extension operator ext u ,
extu W Bpq
s s
./ ,! Bpq .Rn /; 0 < p  1; 0 < q  1; p < s < u; (4.16)
and
extu W Fpq
s s
./ ,! Fpq .Rn /; 0 < p < 1; 0 < q  1; pq < s < u: (4.17)
104 4 The extension problem

x is E-thick. Furthermore since

Proof. One has by Proposition 3.6 (iv) that ! D Rn n 
c D Rn n  D @ [ ! and j@j D 0 (4.18)
s
one obtains by Proposition 3.15 for the above spaces, abbreviated by Apq ,

AQpq
s
.c / D AQpq
s
x D idAQpq
.!/ s
.!/: (4.19)
Here we used (3.20), hence @!  @, j@!j D 0, and s > p , hence Apq
s
.Rn / 
L1 .R /. We deal with the F -spaces. Let for u 2 N,
loc n

˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj (4.20)
be an orthonormal u-wavelet bases now in L2 .!/ as in Theorem 3.13 and let
1 X
X Nj
Qf D ƒjr .f / 2j n=2 ˆjr ; f 2 Fpq
s
.Rn /; (4.21)
j D0 rD1

with Z
ƒjr .f / D 2j n=2 kQrj .x/ f .x/ dx (4.22)
Rn
be the adapted counterpart of (3.74), (3.75). By (4.19) and the same arguments as in
s
the proof of Proposition 3.15 it follows that Q is a projection in Fpq .Rn / and
s
QFpq .Rn / D Fzpq
s
.c /: (4.23)
Now one obtains by Theorem 4.1 and Remark 4.2 that there is a common extension
operator extu for all spaces in (4.17) which can be constructed explicitly by (4.10).
These arguments apply also to the spaces in (4.16).
Corollary 4.5. Theorem 4.4 applies in particular to
(i) bounded Lipschitz domains  in Rn , n  2, according to Definition 3.4 (iii), or
to intervals on R,
x 6D Rn ,
(ii) ."; ı/-domains  in Rn according to Definition 3.1 (i) with 
(iii) snowflake domains  in R , Figure 3.5, p. 76.
2

x D
Proof. By Propositions 3.6, 3.8 the above domains are I -thick with  6 Rn and
j@j D 0. This shows that one can apply Theorem 4.4.
Remark 4.6. The discussions in Section 3.1.3 show that there exist rather bizarre
(disconnected) I -thick domains  even if one assumes that j@j D 0. On the other
hand, domains with inwards cusps as in Figure 3.3, p. 74, are I -thick and one can apply
Theorem 4.4, whereas domains with outwards cusps as in Figure 3.4, p. 74, are not
s
I -thick. Recall that the spaces Apq ./ are defined by restriction. In rough domains
(beyond Lipschitz domains) the question of extendability is different if one deal with
spaces introduced intrinsically. This applies in particular to the Sobolev spaces Wpk ./
according to Definition 2.53. We discussed this question in Remark 3.3 where one finds
also some references.
 4.2 Main assertions 105

4.2.2 Negative smoothness

s
The arguments in the proof of Theorem 4.4 cannot be applied to spaces Apq ./ with
s < 0. On the other hand we obtained in Theorem 3.13 wavelet bases in spaces
s
Apq ./ with s < 0 in a constructive way. This can be interpreted in terms of extension
operators. As before (4.2) and Remark 4.2 say what is meant by common extension
s s
operators. The spaces Bpq ./ and Fpq ./ have the same meaning as in Definition 2.1.
Theorem 4.7. Let  be a .non-empty/ E-thick domain in Rn according to Defini-
tion 3.1 (ii) with  6D Rn . Then for any " with 0 < " < 1 there is a common extension
operator ext " ,

ext" W Bpq
s s
./ ,! Bpq .Rn /; "1 < s < 0; " < p  1; 0 < q  1; (4.24)

and

ext" W Fpq
s s
./ ,! Fpq .Rn /; "1 < s < 0; " < p < 1; " < q  1: (4.25)

Proof. We deal with the F -spaces and rely on the arguments in Step 2 of the proof of
Theorem 3.13. Let
˚ j
ˆr W j 2 N0 W r D 1; : : : ; Nj ; pq  s < u 2 N; (4.26)

be the same orthonormal u-wavelet basis in L2 ./ as in (3.47), (3.54). Let arj be as in
(3.55) (basic and interior wavelets ˆjr ) and as in (3.56) (boundary wavelets ˆjr ). For
" with 0 < " < 1 and given p; q we choose pq C "1 < u 2 N. Then (4.26) applies
to all s with "1 < s < 0. By Theorem 3.13 any f 2 Fpq s
./ can be represented by
(3.59), (3.60), hence
1 X
X Nj
 
f D jr .f / 2j n=2 ˆjr ; jr .f / D 2j n=2 f; ˆjr : (4.27)
j D0 rD1

Then ext" given by

1 X
X Nj
n
ext f D
"
2j.s p / jr .f / arj ; (4.28)
j D0 rD1

is a linear and bounded extension operator. This follows from ext" f j D f and
(3.57), (3.62). Choosing u > "1 C "" D .n C 1/"1  n one has a common
extension operator for all spaces in (4.25). Similarly for the B-spaces where the same
extension operator ext" can also be applied to the spaces in (4.24).
Corollary 4.8. Theorem 4.7 applies in particular to
(i) bounded Lipschitz domains in Rn , n  2, according to Definition 3.4 (iii), and
intervals on R,
(ii) snowflake domains  in R2 , Figure 3.5, p. 76.
106 4 The extension problem

Proof. This follows from Proposition 3.8.

Remark 4.9. By Remark 3.7 E-thick domains  in Rn may be rather bizarre. It
may even happen that j@j > 0. Nevertheless by Theorem 4.7 we have always linear
and bounded extension operators. On the other hand the domains with inwards cusps
as in Figure 3.3, p. 74, are not E-thick, but domains with outwards cusps as in Fig-
ure 3.4, p. 74, are E-thick. Otherwise we refer to our discussion in Remark 4.6 about
intrinsically defined Sobolev spaces Wpk ./.

4.2.3 Combined smoothness

If  is a thick domain in Rn according to Definition 3.1 (iv) (hence both E-thick

and I -thick) with x 6D Rn and j@j D 0 then we can apply both Theorem 4.4 and
Theorem 4.7 covering all spaces Axpq
s
./ in Definition 3.11 with s 6D 0. However the
extension operators from Theorem 4.4 for spaces with s > 0 and from Theorem 4.7
for spaces with s < 0 are different. There is little hope to find in this way common
extension operators which apply simultaneously to spaces with positive and negative
smoothness s and which include also spaces Apq 0
./ of smoothness s D 0 covered by
Definition 3.11. But there is a way to extend Theorem 4.7 to some spaces with s  0
and to some domains with rough boundary. But one has to strengthen the assumptions
about the underlying domain .
First we recall some notation and assertions from Section 3.2.4. According to
(3.79), (3.80) a closed set in Rn is called an d -set, 0 < d < n, if there is a Radon
measure with
.B. ; r//  r d where 2 D supp ; 0 < r < 1: (4.29)
The equivalence constants in (4.29) are independent of and r (in agreement with
(3.4)). If  is an E-thick domain in Rn with  6D Rn such that its boundary D @
is an d -set then one has by Proposition 3.18 (iii) that n  1  d < n. In case of n D 1
one may incorporate now d D 0. Then (4.29) with d D 0 and the assumption that 
is an E-thick domain on the real line R show that one may assume (without restriction
of generality) that  is an interval.
Theorem 4.10. Let  be a .non-empty/ E-thick domain in Rn according to Defini-
tion 3.1 (ii) with  6D Rn such that its boundary D @ is an d -set satisfying (4.29)
with n  1  d < n .in case of n D 1 and d D 0 this means that  is an interval/.
Then for any " with 0 < " < 1 there is a common extension operator ext " ,
(
" < p  1; 0 < q  1; "1 < s < 0;
ext" W Bpq
s
./ ,! Bpq s
.Rn /; (4.30)
1 < p < 1; 0 < q  1; 0  s < nd p
;
and
(
" < p < 1; " < q  1; "1 < s < 0;
ext W" s
Fpq ./ ,! s
Fpq .Rn /; (4.31)
1 < p < 1; 1  q < 1; 0  s < nd
p
:
 4.2 Main assertions 107

Proof. Let ext" be the same extension operator as in Theorem 4.7. Then (4.24), (4.25)
coincides with the upper lines in (4.30), (4.31). Hence it remains to prove that ext " is
also an extension operator for the corresponding lower lines. For h.r/ D r d one has
(3.82) with  D n  d . By Proposition 3.18 (iii) and Step 2 of its proof it follows that
 is uniformly E-porous and that is porous. Then we can apply Proposition 3.19
specified to d -sets. As remarked in the proof of Proposition 3.19 one may choose
ı D  D n  d in case of d -sets. We deal with the F -spaces. Since  is uniformly
E-porous one has by (3.87) and Theorem 3.23 the representation (4.27) = (3.103),
(3.104), hence
k.f / jfpq
s
.Z /k  kf jFpq s
./k: (4.32)
On the other hand, ext" in (4.28) is an atomic decomposition in Fpq
s
.Rn / and one
obtains as before with a reference to Theorem 1.7 that
k ext" f jFpq
s
.Rn /k  c k.f / jfpq
s
.Z /k: (4.33)
By (4.32) and ext" f j D f it follows that ext" is also an extension operator for the
lower line in (4.31). Similarly for the B-spaces.
Remark 4.11. If  is a bounded Lipschitz domain according to Definition 3.4 (iii)
(or an interval in case of n D 1) then  is E-thick, D @ is an .n  1/-set and
one has 0  s < 1=p in the lower lines of (4.30),  (4.31). In case of the snowflake
4
domain in Figure 3.5, p. 76, one has 0  s < p1 2  log log 3
in the lower lines in (4.30),
(4.31). Otherwise we restricted ourselves in the above theorem to boundaries which
are d -sets according to (4.29). But it is quite obvious that one can generalise (4.29) by
(3.79) with (3.82) for some  > 0. Then one can extend Theorem 4.7 by a counterpart
of Theorem 4.10 using Proposition 3.19 and Theorem 3.23.
Although quite obvious it seems to be desirable to give an explicit formulation
covering both Theorem 4.4 and Theorem 4.10. In Remark 4.2 we said what is meant
by a common extension operator. An extension operator is called universal if it applies
s
to all spaces Apq ./ with s 2 R, 0 < p; q  1 (p < 1 for the F -spaces) in the
understanding of Remark 4.2. The numbers p and pq have the same meaning as in
(4.15).
Corollary 4.12. (i) Let  be a bounded Lipschitz domain in Rn with n  2 according
to Definition 3.4 (iii) or an interval on R. Then there is a universal extension operator.
(ii) Let  be a .non-empty/ thick domain in Rn according to Definition 3.1 (iv)
with  x 6D Rn such that its boundary D @ is an d -set satisfying (4.29) with
n  1  d < n. Then there is for any u > 0 a common extension operator ext u with
extu W Bpq
s s
./ ,! Bpq .Rn /; 0 < p  1; 0 < q  1; p < s < u; (4.34)
extu W s
Fpq ./ ,! s
Fpq .Rn /; 0 < p < 1; 0 < q  1; pq < s < u: (4.35)
Then there is for any " with 0 < " < 1 a common extension operator ext" with
(
" < p  1; 0 < q  1; "1 < s < 0;
ext W Bpq ./ ,! Bpq .R /;
" s s n
(4.36)
1 < p < 1; 0 < q  1; 0  s < nd p
;
108 4 The extension problem
(
" < p < 1; " < q  1; "1 < s < 0;
ext W" s
Fpq ./ ,! s
Fpq .Rn /; (4.37)
1 < p < 1; 1  q < 1; 0  s < nd
p
:

Proof. As mentioned in Section 4.1.1, part (i) goes back to [Ry99]. Theorem 4.10
covers the assertions about ext" . By Proposition 3.18 with  D n  d > 0 one has
j j D 0. Then one can apply Theorem 4.4. This covers the assertions about extu .

Corollary 4.13. Corollary 4.12 (ii) applies to the snowflake domain in Figure 3.5,
4
p. 76, in R2 with d D log
log 3
.
Proof. This follows from Proposition 3.8.

Remark 4.14. A first discussion of the extension problem for function spaces with pos-
itive smoothness on rough domains (beyond Lipschitz) has been given in Remark 3.3.
There one finds also some references which will not be repeated here. But we are
not aware of related assertions in the literature for spaces with negative smoothness as
considered in Theorems 4.7, 4.10 and Corollary 4.12. However we wish to mention
[Miy90] dealing in detail with the extension problem for Hardy spaces Hp  Fp;2 0
with
p < 1 in arbitrary domains. The question of extendability is closely connected with the
problem of intrinsic (quasi-)norms and related characterisations. In case of the classical
Sobolev spaces one may consult Remark 2.54. As far as spaces on Lipschitz domains
are concerned we refer to [T06], Sections 1.11, 1.11.10. But also some of the papers in
Remark 3.3 deal with descriptions of function spaces of positive smoothness in more
general domains. We mention in this context [See89] and the most recent substantial
s
paper [Shv06]. In both papers the intrinsic characterisations of the spaces Apq ./ are
s n
used to prove the existence of extension operators in Apq .R /. The admitted types of
domains are similar to the I -thick domains according to Definition 3.1 (iii). Our way
is different and resulted in Theorem 4.4 in a short, but less constructive proof of the
existence of extension operators. In this book we do not deal with equivalent intrinsic
quasi-norms in general. But we return to this point briefly in Section 4.3.3 and, as far
as classical Sobolev spaces are concerned, in Theorem 4.30.

4.3 Complements
4.3.1 Interpolation

Interpolation is only a marginal topic in this book. Mainly we wish to demonstrate

how overlapping regions of common extension operators, for example ext u and ext "
in Corollary 4.12, can be used to prove far-reaching interpolation results.
We assume that the reader is familiar with the basic assertions of interpolation
theory. Let fA0 ; A1 g be an interpolation couple of complex quasi-Banach spaces.
Then
.A0 ; A1 / ;q ; 0 <  < 1; 0 < q  1; (4.38)
 4.3 Complements 109

denotes, as usual, the real interpolation method based on Peetre’s K-functional. Let

ŒA0 ; A1 ; 0 <  < 1; (4.39)

be the classical complex interpolation method for complex Banach spaces. Basic
assertions about these interpolation methods may be found in [T78] and [BeL76]. One
may also consult [T92], Section 1.6.

Proposition 4.15. Let I be either the real method (4.38) or the complex method (4.39).
Let A0 .Rn / and A1 .Rn / be two admitted spaces according to Definition 1.1 such that

I.A0 .Rn /; A1 .Rn // D A.Rn / (4.40)

results in a space A.Rn / again covered by Definition 1.1. Let  be a bounded Lipschitz
domain according to Definition 3.4 (iii) in Rn , n  2, or an interval in R. Let A0 ./,
A1 ./, A./ be the corresponding spaces on  as introduced in Definition 2.1 (i).
Then
I.A0 ./; A1 .// D A./: (4.41)

Remark 4.16. This coincides essentially with [T06], Section 1.11.8. There one finds
also a proof, further explanations and examples. In particular the classical complex
interpolation method in (4.39) which is naturally restricted to Banach spaces can be
s
extended to a class of quasi-Banach spaces which covers the spaces Apq ./. Then
(4.41) as a consequence of (4.40) remains valid also for this extended complex inter-
polation method which goes back to [MeM00]. Further information may also be found
in [Tri02]. The proof that one can transfer (4.40) from Rn to (4.41) in bounded Lip-
schitz domains  relies on the existence of a universal extension operator as indicated
in Corollary 4.12 (i). This is not available in more general domains. But the method
works for those domains  and those spaces, say, Apq s
./ or AQpq
s
./, for which one
has common extension operators from  to R . Even more, if two .s; p; q/-regions in
n

which one has common extension operators for some spaces Apq s
./ or AQpq
s
./ have a
non-empty overlap then one can combine related interpolation formulas. We describe
a typical example which makes clear what is meant and how to proceed in other cases.
We rely on the common extension operators in Corollary 4.12 (ii).

Theorem 4.17. Let  be a .non-empty/ thick domain in Rn according to Defini-

tion 3.1 (iv) with  x 6D Rn such that its boundary D @ is an d -set satisfying (4.29)
with n  1  d < n .in case of n D 1 and d D 0 this means that  is an interval/.
Let 1 < p < 1 and 0 <  < 1.
(i) Let q0 ; q1 ; q 2 .0; 1 . Let 1 < s0 < s1 < 1 and

s D .1  /s0 C s1 : (4.42)

Then
 s   s 
s1
Bpq0 0 ./; Bpq 1
./ ;q
D Fpq
0
0
s1
./; Fpq 1
./ ;q
D Bpq
s
./: (4.43)
110 4 The extension problem

(ii) Let q0 ; q1 2 .1; 1/, s0 2 R, s1 2 R, and

1 1 
D C : (4.44)
q q0 q1
Let s be given by (4.42). Then
s0
Bpq 0
s1
./; Bpq 1
./ D Bpq
s
./ (4.45)

and
s0
Fpq0
s1
./; Fpq1
./ D Fpq
s
./: (4.46)
Proof. Step 1. Since
s s s
Bp;min.p;q/ ./ ,! Fpq ./ ,! Bp;max.p;q/ ./ (4.47)

it is sufficient to prove (4.43) for the B-spaces. By [T83], Theorem 2.4.2, one has
 s 
Bpq0 0 .Rn /; Bpq
s1
1
.Rn / ;q D Bpq
s
.Rn /: (4.48)

We put temporarily  s 
s1
Bpq0 0 ./; Bpq 1
./ ;q
D B ./: (4.49)
Let 0 < s0 < s1 < u. Then we can apply extu in (4.34) to all three spaces in (4.49)
where we used the interpolation property as far as B ./ is concerned, hence

extu W B ./ ,! Bpq

s
.Rn /: (4.50)

Let re be again the restriction operator, hence re B extu D id. Then one obtains by
(4.50) that
id W B ./ ,! Bpq s
./: (4.51)
Conversely, using (4.48) and the interpolation property applied to re one obtains

id W Bpq
s s
./ ,! Bpq .Rn / ,! B ./: (4.52)

Then (4.51), (4.52) prove (4.43) under the assumption 0 < s0 < s1 .
Step 2. Based on (4.36) one has by the same arguments
 s 
s1
Bpq0 0 ./; Bpq 1
./ ;q D Bpq
s
./ (4.53)

if 1 < s0 < s1 < nd

p
. Note that
˚ 1 
p
; s W 1 < p < 1; 0 < s < nd
p
(4.54)

is a non-empty overlap of the admitted .p; s/-region with the corresponding one from
Step 1. Then one can apply Wolff’s interpolation theorem in [Wol82]. To extend (4.43)
to all admitted p, q, s in part (i) of the theorem one may assume that q0 ; q1 2 .1; 1/
or even q0 D q1 D p. Otherwise one could use afterwards the reiteration theorem
 4.3 Complements 111

of interpolation theory and what has been said so far to extend (4.43) from q0 ; q1 2
.1; 1/ or q0 D q1 D p to all q0 ; q1 2 .0; 1 . In particular we may assume that
S./ D S.Rn /j, the restriction of S.Rn / to , is dense in the spaces considered.
Let for fixed p with 1 < p < 1,
s
Aj D Bpp
j
./; 1 < s1 < 0 < s2 < s3 < nd
p
< s4 < 1; (4.55)

such that
A2 D .A1 ; A3 / 1 ;p
; A3 D .A2 ; A4 / 2 ;p
; (4.56)
covered by the two interpolation regions above. Then one obtains by [Wol82] that

A2 D .A1 ; A4 / 3 ;p
; A3 D .A1 ; A4 / 4 ;p
(4.57)

for naturally calculated 0 < 3 < 4 < 1. This proves (4.43) for the full range of the
admitted parameters.
Step 3. Similarly one can argue for the complex interpolation formulas using a
corresponding assertion for the classical complex interpolation method in [Wol82].
For a more general version of Wolff’s interpolation assertion one may also consult
[JNP83].
Remark 4.18. To fix p in the above theorem is convenient but not necessary. Any
interpolation formula in Rn , real, classical complex, or generalised complex as in
[MeM00], can be transferred from Rn to  as long as one has a common extension
operator. This applies to the admitted .p; s/-regions in Step 1 and Step 2 based on
Corollary 4.12. Similarly one can rely on Theorems 4.4, 4.7 and their Corollaries 4.5,
4.8. However if one wishes to clip together two such admitted .p; s/-regions for some
s
spaces Apq ./ then one needs an overlap as in (4.54) and that the desired interpolation
fits in the scheme of (4.56), (4.57). We describe a further scale of spaces to which the
above method can be applied.
There are several good reasons to ask for a counterpart of Theorem 4.17 with Bxpq
s
./
x s s s
and Fpq ./ as introduced in Definition 3.11 in place of Bpq ./ and Fpq ./. On the
one hand we have the common wavelet characterisations in Theorems 3.13, 3.23. On
the other hand for bounded C 1 domains  and fixed 1 < p < 1, 1  q  1,
˚ s ˚ s
Bxpq ./ W s 2 R and H xp ./ D Fxp;2
s
./ W s 2 R (4.58)

are scales with lifting operators imitating I in (1.15) for spaces on Rn . This can be
found in [T78], Section 4.9.2. There is also a related natural duality theory. An ex-
tension of these assertions, including duality and interpolation, to bounded Lipschitz
domains has been given in [Tri02], Section 3.3 (there are also a few relevant com-
ments in [T06], Section 1.11.6). Now we apply the above interpolation method to the
corresponding spaces. As in Theorem 4.17 we restrict ourselves to typical examples.
Theorem 4.19. Let  be a .non-empty/ E-thick domain in Rn according to Defini-
tion 3.1 (ii) with  6D Rn such that its boundary D @ is an d -set satisfying (4.29)
112 4 The extension problem

with n  1  d < n .in case of n D 1 and d D 0 this means that  is an inter-

val on R). Let 1 < p < 1 and 0 <  < 1. Let Bxpq s
./ and Fxpq
s
./ .now with
x
Fpq ./ D Fpq ./ for all 0 < q  1/ be as in Definition 3.11.
0 0

(i) Let q0 ; q1 ; q 2 .0; 1 . Let 1 < s0 < s1 < 1 and

s D .1  /s0 C s1 : (4.59)

Then
   s 
Bxpq
s0
0
./; Bxpq
s1
1
./ ;q
D Fxpq
0
0
./; Fxpq
s1
1
./ ;q
D Bxpq
s
./: (4.60)

(ii) Let q0 ; q1 2 .1; 1/, s0 2 R, s1 2 R and

1 1 
D C : (4.61)
q q0 q1
Let s be as in (4.59). Then

Bxpq
s0
0
./; Bxpq
s1
1
./ D Bxpq
s
./ (4.62)

and
Fxpq
s0
0
./; Fxpq
s1
1
./ D Fxpq
s
./: (4.63)
Proof. By Proposition 3.18 with  D n  d the domain  is uniformly E-porous with
j j D 0. We follow the arguments from the proof of Theorem 4.17. Let 0 < s0 < s1 :
x By Proposition 3.15
We have (4.48) and replace (as a definition) B in (4.49) by B.
one can identify Bxpqs
./ D Bzpq
s
./ with Bzpqs x
./. Then the extension ext by zero
as in (3.89) is a linear and bounded extension operator from Bxpqs
./ into Bpqs
.Rn /
and we have (4.50) with ext in place of ext u and Bx ./. With the projection P as in
Proposition 3.15 we have re B P B ext D id. Now one can argue as in (4.51), (4.52).
This proves (4.59), (4.60) under the assumption 0 < s0 < s1 . By Theorem 4.10 we
have (4.53) for 1 < s0 < s1 < nd p
. We can apply Proposition 3.19. This gives as
before the desired overlap (4.54). The rest is now the same as in Steps 2 and 3 in the
proof of Theorem 4.17.
Corollary 4.20. Theorems 4.17, 4.19 apply to bounded Lipschitz domains in Rn with
n  2 according to Definition 3.4 (iii), to an interval on R, and to the snowflake domain
in Figure 3.5, p. 76.
log 4
Proof. This follows from Proposition 3.8 with d D log 3
in case of the snowflake curve.

4.3.2 Constrained wavelet expansions in Lipschitz domains

All wavelet expansions for˚function spaces in domains  considered so far are based
on the u-wavelet systems ˆjr according to Definition 2.4. The building blocks ˆjr
 4.3 Complements 113

˚ supports in . sThis suggests that one cannot expect wavelet expansions

have compact
based on ˆjr for spaces Apq ./ having boundary values on D @. Then one
needs wavelets at the boundary in addition. We return to this problem in Chapter 5
below. On the other hand we described in Section 2.5.3 for Sobolev spaces Wpk ./
in arbitrary domains so-called constrained wavelet expansions. These considerations
relied on the specific structure of the norm in (2.222) reducing the question of whether
f 2 Wpk ./ to D ˛ f 2 Lp ./ with j˛j  k. One may ask for counterparts in case
s
of the spaces Apq ./ which can be used to argue as in Section 2.5.3. In arbitrary
domains there is little hope for such a reduction. But the situation is much better if 
is a Lipschitz domain.
Proposition 4.21. Let  be a bounded Lipschitz domain in Rn with n  2 according
to Definition 3.4 (iii) or an interval on R. Let Apq
s
./,
0 < p  1; 0 < q  1; s D  C k with  2 R and k 2 N (4.64)
.p < 1 for the F -spaces/; be the spaces as introduced in Definition 2.1. Then
˚
s
Apq ./ D f 2 Apq 
./ W D ˛ f 2 Apq

./; j˛j  k (4.65)
and X
kf jApq
s
./k  kD ˛ f jApq

./k (4.66)
j˛jk

.equivalent quasi-norms/.
Proof. Recall that for g 2 Apq
s
.Rn /,
X
kg jApq
s
.Rn /k  kD ˛ g jApq

.Rn /k (4.67)
j˛jk

(equivalent quasi-norms). We refer to [T83], Theorem 2.3.8, pp. 58–59. Then one
obtains by Definition 2.1 that
X
kD ˛ f jApq

./k  c kf jApq
s
./k (4.68)
j˛jk

for some c > 0 and all f 2 Apq s

./. As for the converse we rely on Rychkov’s
universal extension operator ext  according to [Ry99], Theorem 4.2, p. 253. By the
convolution structure of this operator it follows from
f 2 Apq
s
./ () g D ext f 2 Apq
s
.Rn / (4.69)
that
D ˛ f 2 Apq

./ () D ˛ g D ext D ˛ f 2 Apq

.Rn /; (4.70)
where 0  j˛j  k. In particular one obtains by
X X
kf jApq
s
./k  kg jApqs
.Rn /k  kD ˛ g jApq

.Rn /k  kD ˛ f jApq

./k
j˛jk j˛jk
(4.71)
the converse of (4.68) and the characterisation (4.65), (4.66).
114 4 The extension problem

Remark 4.22. By (4.67) the counterpart of (4.66) with Rn in place of  is known.

Furthermore characterisations of type (4.65), (4.66) for the half-space
˚
RnC D x 2 Rn W x D .x 0 ; xn /; x 0 D .x1 ; : : : ; xn1 /; xn > 0 (4.72)

in place of  and for bounded C 1 domains  according to Definition 3.4 (iii) may be
found in [T83], Section 3.3.5, pp. 202–03. The arguments used there are similar as in
(4.70). The above proposition remains valid for special Lipschitz domains according
to Definition 3.4 (ii).
After these preparations we are now in a similar position as in Section 2.5.3 where
we dealt with constrained wavelet expansions for Sobolev spaces in arbitrary domains.
Instead of Lp ./ as there we use now

Apq ./; 0 < p  1; 0 < q  1;  < 0; (4.73)

(p < 1 for the F -spaces) in bounded Lipschitz domains  as basic spaces. By

Proposition 3.8 bounded Lipschitz domains are thick and hence E-thick. Then we can
apply Theorem 3.13 to the spaces in (4.73). Let
˚
ˆ D ˆjr W j 2 N0 I r D 1; : : : ; Nj with Nj 2 N; (4.74)

be the same orthonormal u-wavelet basis in L2 ./ according to Theorem 2.33 and
Definitions 2.31, 2.4 as in (3.47) with

p C jj < u 2 N and pq C j j < u 2 N; (4.75)

s s
for the B-spaces and F -spaces, respectively. Let bpq .Z / and fpq .Z / be the same
sequence spaces as in Definition 2.6 and as used in Theorem 3.13 now abbreviated by
s
apq .Z / with a 2 fb; f g. Combining Proposition 4.21 with Theorem 3.13 one can
argue as in Section 2.5.3. In particular if

f 2 Apq
s
./; s D  C k;  < 0; k 2 N; (4.76)

then D ˛ f 2 Apq

./, j˛j  k, can be expanded by

1 X
X Nj
D˛ f D jr .D ˛ f / 2j n=2 ˆjr (4.77)
j D0 rD1

where
jr .D ˛ f / D 2j n=2 .D ˛ f; ˆjr / D .1/j˛j 2j n=2 .f; D ˛ ˆjr / (4.78)
with Z
jr .f /D2j n=2
.f; ˆjr / D2 j n=2
f .x/ ˆjr .x/ dx (4.79)

as before (appropriately interpreted). Similarly as in (2.232), (2.233), we put
˚
.f /k D jr .f /k 2 RC W j 2 N0 I r D 1; : : : ; Nj (4.80)
 4.3 Complements 115

with X ˇ ˇ
jr .f /k D 2j n=2 ˇ f; D ˛ ˆj ˇ: (4.81)
r
j˛jk

Theorem 4.23. Let  be a bounded Lipschitz domain in Rn with n  2 according to

Definition 3.4 (iii) or a bounded interval on R. Let Apq
s
./,
0 < p  1; 0 < q  1; s D  C k with  < 0 and k 2 N; (4.82)
.p < 1 for the F -spaces/ be the spaces as introduced in Definition 2.1. Let ˆ be
the orthonormal u-wavelet basis in L2 ./ with (4.74), (4.75). Then f 2 D 0 ./ is an
element of Apq
s
./ if, and only if, it can be represented as
1 X
X Nj
f D jr .f / 2j n=2 ˆjr ; .f /k 2 apq

.Z /: (4.83)
j D0 rD1

Furthermore,
kf jApq
s
./k  k.f /k japq

.Z /k (4.84)
.equivalent quasi-norms/.
Proof. If f 2 D 0 ./ can be represented by (4.83) then it follows from Theorem 3.13
that D ˛ f 2 Apq
./ for j˛j  k. One has by Proposition 4.21 that f 2 Apq s
./.
Conversely if f 2 Apq s
./ then D ˛ f with j˛j  k can be expanded in Apq 
./ by
(4.77), (4.78). Then (4.83), (4.84) follows from Proposition 4.21.
Remark 4.24. As in Theorem 3.13 the representation (4.83) converges unconditionally
ı
in Apq ./ with ı <  and if, in addition, p < 1, q < 1 in Apq 
./. Since the
j
coefficients r .f / are subject to the constraints .f / 2 apq .Z / it follows that
k 
s
the outcome belongs even to Apq ./ with (4.84). There is the somewhat curious
possibility to reduce the coefficients in (4.78) to linear combinations of the original
coefficients in (4.79) at least if u in (4.75) is sufficiently large. Let again p, q, s be as
in (4.82). Let u  k in addition to (4.75). Then D ˛ ˆjr with j˛j  k are continuous
functions with compact supports in . They can be expanded, say, in L2 ./ by
1 X
X Nl
 
D ˛ ˆjr D D ˛ ˆjr ; ˆlt ˆlt : (4.85)
lD0 tD1
Inserted in (4.78) one obtains
1 X
X Nl
 
jr .D ˛ f / D .1/j˛j 2j n=2 D ˛ ˆjr ; ˆlt .f; ˆlt /
lD0 tD1
1 X
X Nl
j˛j .j l/n=2
 
D .1/ 2 D ˛ ˆjr ; ˆlt lt .f / (4.86)
lD0 tD1
1 X
X Nl
D m˛.j;rIl;t/ lt .f /
lD0 tD1
116 4 The extension problem

where
˚  
Mˆ˛ D m˛.j;rIl;t/ ; m˛.j;rIl;t/ D .1/j˛j 2.j l/n=2 D ˛ ˆjr ; ˆlt (4.87)
are the transition matrices from ˆ in (4.74) to
˚
ˆ˛ D D ˛ ˆjr W j 2 N0 I r D 1; : : : ; Nj ; j˛j  k: (4.88)
Then (4.86) can be abbreviated in the usual way as
.D ˛ f / D Mˆ˛  .f /; j˛j  k: (4.89)
Corollary 4.25. Let  be a bounded Lipschitz domain in Rn with n  2 according
to Definition 3.4 (iii) or a bounded interval on R. Let Apq
s
./ be the same spaces as
in Theorem 4.23 with (4.82). Let ˆ be as above the orthonormal u-wavelet basis in
L2 ./ with (4.74), (4.75) and u  k. Then f 2 D 0 ./ is an element of Apq s
./ if,
and only if, it can be represented by
1 X
X Nj
f D jr .f / 2j n=2 ˆjr (4.90)
j D0 rD1

with X
kMˆ˛  .f / japq

.Z /k < 1 (4.91)
j˛jk

.equivalent quasi-norms/.
Proof. This follows from Theorem 4.23 and the above considerations.
Remark 4.26. At first glance the outcome is somewhat surprising. The entries of the
transition matrices Mˆ˛ in (4.87) depend only on the (sufficiently smooth) u-wavelet
basis ˆ. By the support properties of ˆjr the matrices Mˆ˛ in (4.87) are band-limited.
One needs only a knowledge of the coefficients jr .f / in (4.79) to use the criterion
(4.91) and to decide whether f belongs to Apq s
./ or not. If, say, 1  p < 1 and
s > 1=p then f 2 Apq ./ has boundary values at D @. This seems to contradict
s

the representation (4.90) with (4.91) since all building blocks ˆjr have compact supports
in  and, hence, they vanish at . But one must have in mind that (4.90) converges

only in Apq ./ where  < 0 (and, say, p < 1, q < 1). The constraint (4.91) does
not improve this convergence. Furthermore the entries of the matrices Mˆ˛ in (4.87)
are real and (apparently) violently oscillating. Since  is bounded, (4.90) makes sense
for any polynomial f D P . By (4.86) it follows that
Mˆ˛  .P / D 0 if degree.P / < j˛j  k. (4.92)
One has the impression that the terms with j˛j > 0 in (4.91) are the counterparts
of the terms with mh;
f in (2.19) or corresponding terms involving oscillations as,
for example, in [T92], Section 3.5. But so far it is not so clear what is the use of
Theorem 4.23, Corollary 4.25 and also for the corresponding assertions for the classical
Sobolev spaces in arbitrary domains in Section 2.5.3 (if there is any).
 4.3 Complements 117

4.3.3 Intrinsic characterisations

According to The New Shorter Oxford English Dictionary on historical principles,

Vol. 1 (Clarendon Press, Oxford, 1993), p. 1405, the word intrinsic has been in use in
mathematics since the mid 19th century as not involving reference to external coordi-
nates. In our context one may replace coordinates by quantities to make it sufficiently
nebulous.
The classical Sobolev spaces Wpk ./ with 1 < p < 1 and k 2 N in arbitrary do-
mains  are undoubtedly defined intrinsically by (2.221), (2.222). It is also acceptable
to call the wavelet characterisations of these spaces in Theorem 2.55 intrinsic. On the
s
other hand all spaces of type Apq ./ we are dealing with in this exposition originate
s n
from Apq .R / by restriction or decomposition. This cannot be called intrinsic with
respect to a domain  6D Rn . Then the question arises to find intrinsic descriptions
where, according to the above quotation, the a somewhat vague notation intrinsic refers
to, say, characterisations of these spaces on domains  where all ingredients (building
blocks, quasi-norms etc.) are restricted to . The orthonormal u-wavelet bases
˚
ˆ D ˆjr W j 2 N0 I r D 1; : : : ; Nj (4.93)

in L2 ./ according to Theorem 2.33 and Definitions 2.31, 2.4 fit in this scheme. This
s
applies also to diverse sequence spaces of type apq .Z / with a 2 fb; f g. This justifies
to call the following wavelet representations intrinsic.
s;rloc
(i) The characterisation of the refined localisation spaces Fpq ./ in arbitrary
domains  in R with  6D R according to Definition 2.14 in Theorem 3.27.
n n

(ii) The characterisation of the spaces Axpq

s
./ in E-thick domains  with  6D Rn
according to Definition 3.11 with s 6D 0 in Theorem 3.13.
(iii) The characterisation of the spaces Axpq
s
./ in uniformly E-porous (and hence by
Proposition 3.18 E-thick) domains  with  6D Rn according to Definitions 3.11,
3.16 in Theorem 3.23.
s
(iv) The characterisation of the spaces Apq ./ in bounded Lipschitz domains ac-
cording to Definition 2.1 (i) in Theorem 4.23.

Remark 4.27. These are far-reaching characterisations in terms of common intrinsic

s
wavelet bases, wavelet isomorphisms (onto sequence spaces of type apq .Z /), and
constrained wavelet expansions. Parts (i)–(iii) apply not only to bounded Lipschitz
domains but also to some domains with fractal boundaries covered by Propositions 3.8,
3.18, for example the snowflake domain in Figure 3.5, p. 76. The search for intrinsic
characterisations of function spaces has a long history. But despite the above quotation
from the Oxford English Dictionary it remains a matter of taste what one accepts to be
called intrinsic. We return to this point in Remark 4.29 below. To some extent intrinsic
characterisations are related to the extension problem. But this does not mean that
the existence of a linear and bounded extension operator results automatically in an
118 4 The extension problem

intrinsic characterisation and vice versa. For example, ext u in Theorem 4.4 originating
from (4.10) is hardly more than a non-constructive existence proof. It cannot be used to
s
find an intrinsic characterisation of corresponding spaces Apq ./ in I -thick domains.
Or? On the other hand intrinsic characterisations of function spaces have been used
in literature to construct extension operators. Although this is not our subject here we
add a few further comments and give some references.
Remark 4.28. As far as the classical Sobolev spaces Wpk ./ are concerned we com-
mented on the extension problem in Section 2.5.3 and in Remark 3.3. There one finds
also the relevant literature. Both extension problems and intrinsic characterisations for
the Sobolev spaces and the classical Besov spaces
s
Bpq ./; 1  p; q  1; s > 0; (4.94)

including the Hölder–Zygmund spaces, and some forerunners such as the Slobodeckij
s
spaces Bpp ./, have been studied since the 1950s and early 1960s. In [HaT08], Note
4.6.1, p. 112, we tried to clarify the historical roots, including relevant references. This
will not be repeated here. We only mention that detailed discussions and corresponding
results may be found in the two outstanding Russian books [Nik77], [BIN75]. One may
also consult [T78] as far as the situation in the middle of the 1970s is concerned. For
bounded C 1 domains  the extension problem had been solved for all spaces Apq s
./
around 1990. This applies also to satisfactory intrinsic descriptions of the correspond-
s s
ing spaces Bpq ./ with s > p and Fpq ./ with s > pq in terms of differences,
means of differences and oscillations. It may be found in [T92]. The solution of the
extension problem and intrinsic characterisations in terms of convolutions and local
s
means for all spaces Apq ./ in bounded Lipschitz domains  goes back to [Ry98],
s s
[Ry99]. Intrinsic characterisations both of Bpq ./ and Fpq ./ in bounded Lipschitz
domains for s > p or s > pq in terms of differences and means of differences
may be found in [T06], Theorem 1.118, p. 74, with references to [Dis03], [DeS93]
for the B-spaces. A special case of these characterisations has been mentioned above
s
in Remark 2.3. Intrinsic descriptions and extensions of (anisotropic) spaces Fpq ./
with s > pq in (anisotropic) ."; ı/-domains as in Definition 3.1 (i) are the subject of
[See89]. The most recent and also most advanced paper in this connection is [Shv06]
where the author studies intrinsic characterisations and extensions for spaces

Wpk ; s
Fpq ; s
Bpq with 1  p  1; 1  q  1; s > 0; k 2 N;

on n-sets in Rn as in (3.79) with h.r/ D r n .

Remark 4.29. In order to prepare what follows we first recall of what we said so far
about intrinsic characterisations for spaces on domains. As mentioned above with a
s
reference to [Dis03], [DeS93] one has for the spaces Bpq ./ with (2.17) or, more
generally, (2.20), in bounded Lipschitz domains the equivalent quasi-norms (2.19).
s;rloc
According to Theorem 2.18 the refined localisation spaces Fpq ./ with (2.69), in
particular s > pq , can be described in arbitrary domains by ball means of differences.
As a special case we obtained in Corollary 2.20 a corresponding characterisation for the
 4.3 Complements 119

related Sobolev spaces Wpk;rloc ./. We return to this point below. Some discussions
about the extendability of Sobolev spaces including relevant references may be found in
Remark 3.3, complemented by Remark 3.9. A discussion about the extension problem
in thick domains inclusively related references has been given in Remark 4.14. Finally
s
we recall the rather general intrinsic characterisations of the spaces Apq ./ in terms
of atoms as it has been developed in [TrW96] and described in [ET96], Section 2.5.
Let  be a bounded domain in Rn with  D ./ x ı (which means that  coincides with
the interior of its closure as considered in connection with Proposition 3.6). Then  is
called interior regular if there is a positive number c such that
j \ Qj  c jQj (4.95)
for any cube Q centred at @ with side-length less than 1. It is called exterior regular
if there is a positive number c such that for any cube Q centred at @ with side-length l
less than 1 there exists a subcube Qe with side-length cl and
x
Qe  Q \ .Rn n /: (4.96)
The above domain  is called regular if it is both interior and exterior regular. These
notation are similar but not identical with I -thick, E-thick and thick domains as in-
troduced in Definition 3.1. Somewhat roughly the multiplication of the .s; p/-atoms
in Rn according to Definition 1.5 with the characteristic function of  are called -
.s; p/-atoms. Similarly one adapts the sequence spaces apq in Definition 1.3 to  in the
same way as in Definition 2.6, now denoted as apq .Z /. For details and more precise
formulations we refer to [TrW96] and [ET96], Section 2.5.2. Then it makes sense to
s
ask for intrinsic characterisations of the spaces Apq ./ as introduced in Definition 2.1,
1 X
X Nj
f D jr ajr ;  D fjr g 2 apq .Z / (4.97)
j D0 rD1

with -.s; p/-atoms ajr and

kf jApq
s
./k  inf k japq .Z /k (4.98)
where the infimum is taken over all representations (4.97). This is the question for an
intrinsic -counterpart of Theorem 1.7. The outcome is the following. Let p , pq be
the same numbers as in (1.32). There is a characterisation (4.97), (4.98),
(a) for the spaces
s
Bpq ./; 0 < p  1; 0 < q  1; s > p ; (4.99)
x ı,
in all bounded domains  with  D ./
(b) for the spaces
s
Bpq ./; 0 < p  1; 0 < q  1; s 2 R; (4.100)
in exterior regular domains,
120 4 The extension problem

(c) for the spaces

s
Fpq ./; 0 < p < 1; 0 < q  1; s > pq ; (4.101)

in interior regular domains, and

(d) for the spaces

s
Fpq ./; 0 < p < 1; 0 < q  1; s 2 R; (4.102)

in regular domains.

Greater details and further explanations may be found in [ET96], Section 2.5. As for
proofs we refer to [TrW96]. Nowadays the arguments given there may be simplified
a little bit (but presumably not very much) by using wavelet isomorphisms for the
s
spaces Apq .Rn / according to Theorem 1.20 as the starting point instead of atomic
representations as described in Theorem 1.7. But there is a difference between the
intrinsic wavelet characterisations as recalled in (i)–(iv) at the beginning of this section
and the above descriptions (4.97), (4.98) for the spaces in (a)–(d). On the one hand the
wavelet coefficients jr .f / according to (3.127) are known (more or less) explicitly and
one has to check whether fjr .f /g belongs to some sequence spaces of type apq s
.Z /
or not. But there are no constructive counterparts for optimal coefficients and atoms in
(4.97), (4.98) and one may even question whether these representations are entitled to
be called intrinsic. But this is a little bit like the unsolvable riddle, who was first, the
hen (H) or the egg (E)?

(H) Given f 2 D 0 ./ or f 2 S 0 ./ D S.Rn /j. Decide intrinsically to which

spaces Apq
s
./ it belongs.

(E) Given Apq

s
./. Find an intrinsic description of all of its elements.

Functions have been for ages, are, and will be for ever the favourite subject of mathe-
maticians independently of whether a few of them are recruited to serve as members
of function spaces. The above wavelet characterisations (i)–(iv) fit both in (H) and (E)
whereas (4.97), (4.98) in (a)–(d) are restricted to (E).

Finally we wish to demonstrate how wavelet bases can be used to obtain intrinsic
equivalent norms for Sobolev spaces in E-thick domains which in the case of bounded
smooth domains are more or less known. We introduce the notation

zpk ./ D Fzp;2

W k
./; k 2 N; 1 < p < 1; (4.103)

where Fzp;2
k
./ are the same spaces as in Definition 2.1 (ii). Let as in (2.67),

ı.x/ D min.1; dist.x; //; D @; x 2 : (4.104)

 4.3 Complements 121

Theorem 4.30. Let  be an E-thick domain in Rn according to Definition 3.1 (ii) such
that j@j D 0 and
ˇ˚ ˇ
ˇ x 2  W ı.x/  2j ˇ < 1; j 2 N: (4.105)

Let W zpk ./ with k 2 N and 1 < p < 1 be as in (4.103). Then D./ is dense in
Wzpk ./, the embedding
zpk ./ ,! Lp ./
id W W (4.106)
is compact, and
X
zpk ./k1 D
kf jW kı kCj˛j D ˛ f jLp ./k; (4.107)
j˛jk
X
zpk ./k2 D
kf jW kD ˛ f jLp ./k C kı k f jLp ./k; (4.108)
j˛jDk
X
zpk ./k3 D
kf jW kD ˛ f jLp ./k; (4.109)
j˛jk
X
zpk ./k4 D
kf jW kD ˛ f jLp ./k; (4.110)
j˛jDk

zpk ./.
are equivalent norms in W

Proof. Step 1. By Proposition 3.10 and Corollary 2.20 it follows that (4.107) and
zpk ./.
(4.108) are equivalent norms on W
Step 2. By Theorems 3.13 and 2.36 one has for W zpk ./ and Lp ./ common
˚ j
wavelet bases ˆr and related expansions

1 X
X Nj
f D jr .f / 2j n=2 ˆjr (4.111)
j D0 rD1

with
zpk ./k  k jfp;2
kf jW k
.Z /k; kf jLp ./k  k jfp;2
0
.Z /k: (4.112)

The assumption (4.105) ensures NJ < 1 for any J 2 N. In particular, idJ ,

Nj
X
J X
idJ f D jr .f / 2j n=2 ˆjr ; (4.113)
j D0 rD1

is a linear operator of finite rank. By (2.38) one obtains

zpk ./k;
kf  idJ f jLp ./k  c 2kJ kf jW zpk ./:
f 2W (4.114)
122 4 The extension problem

Now it follows by standard arguments that id in (4.106) is compact. Since each function
ˆjr can be approximated, say, in the norm in (4.107) by functions belonging to D./
(Sobolev’s mollification method) one obtains as a by-product that D./ is dense in
Wzpk ./.
Step 3. Since j@j D 0 one sees by Proposition 3.15 that one can identify W zpk ./
with W x a subspace of Wpk .Rn /. This proves that both (4.109) and
zpk ./,
X
kf jWzpk ./k5 D kf jLp ./k C kD ˛ f jLp ./k (4.115)
j˛jDk

are equivalent norms in W zpk ./. To show that even (4.110) is an equivalent norm it
remains to prove that there is a positive constant c such that
X
kf jLp ./k  c kD ˛ f jLp ./k for all f 2 W zpk ./: (4.116)
j˛jDk

This will be done by contradiction assuming that there is no positive constant c with
zpk ./ with
(4.116). Then one finds a sequence ffj gj1D1  W
X
1 D kfj jLp ./k > j kD ˛ fj jLp ./k: (4.117)
j˛jDk

zpk ./. Since id in (4.106) is compact we may

In particular, ffj g is bounded in W
assume that fj converges in Lp ./,
fj ! f 2 Lp ./; kf jLp ./k D 1: (4.118)
zpk ./ to f . By (4.117) one
Then it follows by (4.117) that ffj g converges even in W
has D f D 0 for j˛j D k. We may assume that  is connected. As a consequence,
˛

f is a polynomial of degree less than k in ,

X
f D aˇ x ˇ 2 W zpk ./; kf jLp ./k D 1: (4.119)
jˇ jk1

Applying again Proposition 3.15 one obtains

zp1 ./;
aˇ 2 W jˇj D k  1: (4.120)
For fixed l 2 N the number of Whitney cubes Qlr 0
needed in (2.83), (2.84) is larger
than c 2l.n1/
for some c > 0 which is independent of l 2 N. This follows also from
the arguments in Step 3 of the proof of Proposition 3.18. Then one has
Z X1
p 0
ı .x/ dx  c 2lp 2l.n1/ 2ln D 1; (4.121)

lD1

where we used p > 1. We apply (4.108) with k D 1 to (4.120) and obtain aˇ D 0

for jˇj D k  1. Iteration gives aˇ D 0 for all jˇj  k  1. Hence f D 0. But this
contradicts (4.119).
 4.3 Complements 123

4.3.4 Compact embeddings

If  is an E-thick domain in Rn with (4.105) then the embedding (4.106) is com-

pact. This observation is based on the wavelet representations (4.111), (4.112) and
the approximation of the identity id by the finite rank operators idJ in (4.114). If one
knows the dimension of the range of idJ then one obtains by (4.114) an estimate of the
so-called approximation numbers of id in (4.106). It is the main aim of what follows
to have a closer look at this type of argument and to make clear what is new compared
with already existing results. But we restrict ourselves to a few comments. First we
recall some definitions and quote related assertions. Let
UA D fa 2 A W ka jAk  1g (4.122)
be the unit ball in the (complex) quasi-Banach space A. If A, B are two (complex)
quasi-Banach spaces then L.A; B/ is the collection of all linear and bounded operators
T W A ,! B quasi-normed by
kT k D supfkT a jBk W a 2 UA g: (4.123)
Furthermore, rank T denotes the dimension of the range of T . Recall that T 2 L.A; B/
is called compact if T UA is precompact in B.
Definition 4.31. Let A; B be quasi-Banach spaces and let T 2 L.A; B/.
(i) Then for all k 2 N the k-th entropy number ek .T / of T is defined as the infimum
of all " > 0 such that
k1
2[
T .UA /  .bj C "UB / for some b1 ; : : : ; b2k1 2 B: (4.124)
j D1

(ii) Then for all k 2 N the k-th approximation number ak .T / of T is defined by

ak .T / D inffkT  Lk W L 2 L.A; B/; rank L < kg: (4.125)
Remark 4.32. Neither entropy numbers nor approximation numbers and their relations
to the spectral theory of compact operators play any role in this book. This may justify
that we refer to [T06], Section 1.10, for further comments, properties, and the relevant
literature. Recall that we explained in (3.4), (3.5) what is meant by the equivalence
ak  bk . Let Bpq s
./ be the spaces as introduced in Definition 2.1 by restriction of
Bpq .R / to . Let p1 C p10 D 1 if 1  p  1 and p 0 D 1 if 0 < p < 1. Recall that
s n

bC D max.b; 0/ for b 2 R.
Theorem 4.33. (i) Let  be an arbitrary bounded domain in Rn . Let p0 ; p1 ; q0 ; q1 2
.0; 1 and
n n
 1 < s1 < s0 < 1; s0  > s1  : (4.126)
p0 p1
Then the embedding
id W Bps00 q0 ./ ,! Bps11 q1 ./ (4.127)
124 4 The extension problem

is compact and
s0 s1
ek .id/  k  n ; k 2 N: (4.128)
(ii) Let  be a bounded Lipschitz domain in R , n  2, according to Defini-
n

tion 3.4 (iii) or a bounded interval in R if n D 1. Let p0 ; p1 ; q0 ; q1 ; s0 ; s1 be as in

part (i) and let
 
1 1
ıC D s0  s1  n  ; (4.129)
p0 p1 C
 
s0  s1 1 1 1 1
D  max  ;  : (4.130)
n 2 p1 p0 2
Let id be as in (4.127). Then for k 2 N,
 either 0 < p0  p1  2;
ıC =n
ak .id/  k if or 2  p0  p1  1; (4.131)
or 0 < p1  p0  1;

ak .id/  k  if 0 < p0 < 2 < p1 < 1;  > 1=2; (4.132)

and
0 ;p /
ıC min.p0 1
ak .id/  k  n  2 if 0 < p0 < 2 < p1 < 1;  < 1=2: (4.133)

Remark 4.34. All equivalences are independent of q0 and q1 . Then it follows from
s s s
Bp;min.p;q/ ./ ,! Fpq ./ ,! Bp;max.p;q/ ./ (4.134)

that one can replace B in (4.127) by F or A 2 fB; F g. Both (4.128) and the corre-
sponding assertions for ak .id/ have a long history. We only mention that for bounded
C 1 domains (4.128) goes back to [EdT89], [EdT92] and [ET96], Section 3.3, and
(4.131), (4.132) to [ET96], Section 3.3, complemented in [Cae98] by (4.133). There
one finds also the history and further references. As for more recent comments one
may also consult [T06], Sections 1.11.2, 1.11.7. The main reason for incorporating the
above theorem is the following. Both for entropy numbers and approximation numbers
it is assumed that  is bounded; in case of the approximation numbers one relies in
addition on Corollary 4.12 (i) ensuring that there is a common extension operator for
the spaces on  into corresponding spaces on Rn . Then multiplication with a suitable
cut-off function shows that it is sufficient to deal with embeddings
˚
id0 W f 2 Bps00 q0 .Rn / W supp f  U ,! Bps11 q1 .Rn / (4.135)

where U D fy 2 Rn W jyj  1g is the unit ball in Rn . Then one can apply the wavelet
isomorphisms in Rn according to Theorem 1.20. This reduces the study of entropy
numbers and approximation numbers for id in (4.127) or id0 in (4.135) to equivalent
problems for sequence spaces. This reduction is quite standard nowadays. It had been
 4.3 Complements 125

used for the first time in [T97] and afterwards in [T01], [T06] and the related literature
mentioned there. We describe the outcome of this reduction and adapt the sequence
s
spaces bpq according to Definition 1.18 and Theorem 1.20 to the notation as used in
[T97], [T01] and more recent in [T06], Section 6.3. Let
˚
 D jr 2 C W j 2 N0 I r D 1; : : : ; Mj ; Mj  2jd ; (4.136)
 M 
for d > 0. With ı  0, 0 < p  1, 0 < q  1 let `q 2j ı `p j be the space of all
sequences  in (4.136) such that
X
1 X
Mj q=p 1=q
  
 j`q 2j ı `pMj  D 2j ıq jjr jp < 1; (4.137)
j D0 rD1

(usual modification if p D 1 and/or q D 1). By Theorem 1.20 it is now quite

clear that the embedding (4.135) and the study of related entropy and approximation
numbers can be reduced to corresponding equivalent questions for related sequence
spaces with d D n. All this has been done in detail in the above-mentioned literature.
But this observation paves the way to apply the above common wavelet bases to study
embeddings of type (4.127) now for more general situations and modified spaces. We
restrict ourselves to a few examples and comments.

Theorem 4.35. Let  be an E-thick domain in Rn according to Definition 3.1 (ii) with
jj < 1. Let p0 ; p1 ; q0 ; q1 2 .0; 1 and
n n
 1 < s1 < s0 < 0; s0  > s1  : (4.138)
p0 p1
Then the embedding
id W Bps00 q0 ./ ,! Bps11 q1 ./ (4.139)
is compact. Furthermore one has (4.128) for the related entropy numbers and (4.131)–
(4.133) with ıC and  as in (4.129), (4.130) for the related approximation numbers.

Proof. Since s1 < s0 < 0 one can apply Theorem 3.13 (ii) to the two spaces in (4.139).
By construction one has Nj  2j n . This reduces the spaces bpq
s
.Z / in (3.53) to the
 j ı Mj 
above sequence spaces `q 2 `p with Mj D Nj in the same way as outlined in
Remark 4.34 in connection with id0 in (4.135). Then one obtains the above theorem
as a corollary of Theorem 4.33.

Remark 4.36. The main point is the direct transfer of (4.139) to corresponding em-
beddings between sequence spaces by using the common wavelet basis according to
Theorem 3.13. One needs that jj < 1 but not that  is bounded. Under these
circumstances a reduction of id in (4.139) to id0 in (4.135) is not always possible. But
this was essential for the proof of Theorem 4.33. By Proposition 3.8 the above theorem
can be applied to the snowflake domain in R2 according to Figure 3.5, p. 76.
126 4 The extension problem

Remark 4.37. One has now two possibilities to extend the assertions about approx-
imation numbers according to Theorem 4.33 (ii) from bounded Lipschitz domains to
more general domains. Let id0 be the same reference embedding as in (4.135). If  is
bounded and if there exists a linear and continuous extension operator

ext W Bps00 q0 ./ ,! Bps00 q0 .Rn / (4.140)

for the source space then one obtains by the arguments indicated in Remark 4.34 that

ak .id/  ak .id0 /; k 2 N; (4.141)

where id is the same embedding as in (4.126), (4.127). Secondly one obtains the same
result if jj < 1 (but  not necessarily bounded) and if both the source space and
the target space have a common wavelet basis. It seems to be reasonable to fix some
assertions which can be obtained in this way as a corollary both to Theorem 4.33,
Remark 4.34 on the one hand and Theorem 4.35 and its proof on the other hand. Let
p and pq be as in (4.15). As before ak are the approximation numbers according to
(4.125).

Corollary 4.38. (i) Let  be a bounded I -thick domain in Rn according to Defini-

tion 3.1 (iii) with j@j D 0. Let p0 ; q0 ; p1 ; q1 ; s0 ; s1 be as in Theorem 4.33 and let in
addition s0 > p0 . Then id in (4.127) is compact with

ak .id/  ak .id0 /; k 2 N; (4.142)

where id0 is the reference embedding (4.135).

(ii) Let  be an E-thick domain in Rn according to Definition 3.1 (ii) with jj < 1.
Let p0 ; q0 ; p1 ; q1 ; s0 ; s1 be as in Theorem 4.33 and let in addition

s0 62 Œ0; p0 and s1 62 Œ0; p1 : (4.143)

Let Bxpq
s
./ be the spaces according to (3.46). Then

id W Bxps00 q0 ./ ,! Bxps11 q1 ./ (4.144)

is compact with (4.142) where id0 has the same meaning as there. One can replace

Bxps00 q0 ./ by Lp0 ./ with 1 < p0 < 1;  pn0 > s1  n

p1
; (4.145)

or
Bxps11 q1 ./ by Lp1 ./ with 1 < p1 < 1; s0  n
p0
>  pn1 : (4.146)

(iii) Let  be an arbitrary domain in Rn with jj < 1. Let p0 , q0 , p1 , q1 , s0 , s1

be as in Theorem 4.33 and let in addition

s0 62 Œ0; p0 q0 ; s1 62 Œ0; p1 q1 ; (4.147)

 4.3 Complements 127

s;rloc
q0 D 1 if p0 D 1 and q1 D 1 if p1 D 1. Let Fpq ./ be the spaces according
to Definition 2.14. Then

id W Fps00q;rloc
0
./ ,! Fps11q;rloc
1
./ (4.148)

is compact with (4.142) in the cases covered by Theorem 4.33 (ii) and id0 as in (4.135).
One can replace

Fps00q;rloc
0
./ by Lp0 ./ with 1 < p0 < 1;  pn0 > s1  n
p1
; (4.149)

or
Fps11q;rloc
1
./ by Lp1 ./ with 1 < p1 < 1; s0  n
p0
>  pn1 : (4.150)

Proof. It is sufficient to complement the arguments in Remark 4.37 by the necessary

references. Since  is bounded one obtains part (i) from Theorem 4.4 and (4.140),
(4.141). Theorem 3.13 complemented by Theorem 2.36 ensure the existence of com-
mon wavelet bases in E-thick domains for the spaces considered. Since jj < 1 one
obtains part (ii) by the same arguments as in the proof of Theorem 4.35. Similarly for
part (iii) where one can rely on Theorem 3.27. Then one needs the F -counterpart of
(4.142). But by Remark 4.34 this is the same as for the B-spaces excluding limiting
cases.

Remark 4.39. By (4.134) one can incorporate the F -spaces in parts (i) and (ii). As
far as spaces of smoothness zero are concerned we restricted the above formulation to
Lp ./ with 1 < p < 1. This can be extended to some other spaces Apq 0
./ under
some restrictions for the parameters and domains. One may consult the theorems
mentioned in the above proof.

Remark 4.40. As in Theorem 4.35 one can replace the approximation numbers ak in
the above corollary by the corresponding entropy numbers ek with (4.128). This is of
interest for parts (ii) and (iii) but not for part (i) where we have by Theorem 4.33 (i) a
better assertion. Furthermore one can ask to which extent one can replace approxima-
tion numbers ak and entropy numbers ek by other s-numbers or diverse types of widths
measuring compactness. In the first line one may think about Kolmogorov numbers and
Gelfand numbers. But there is a plethora of other numbers and widths. The abstract
theory of s-numbers and widths in Banach spaces may be found in [Pie80], Section 11,
[Pie87], Chapter 2, [Pie07], Section 6.2, and [Kon86]. There are some specifications
of the abstract theory to sequence spaces and function spaces. We refer in this context
also to the recent report [Vyb07b], filling also some gaps as far as approximation num-
bers for compact embeddings between function spaces in bounded Lipschitz domains
according to Theorem 4.33 (ii) are concerned. If one wishes to use what is known so far
for some widths, denoted by fhk W k 2 Ng, for the above purposes then the following
ingredients are desirable.
128 4 The extension problem

• The numbers fhk .T /g are defined for compact maps T W A ,! B between

(complex) quasi-Banach spaces and

kT k  h1 .T /      hk .T /     (4.151)

(monotonically decreasing).

• They have the multiplication property

hk .T2 B T B T1 /  kT2 k  hk .T /  kT1 k; k 2 N; (4.152)

where

T1 W A ,! A1 ; T W A1 ,! B1 compact; T2 W B1 ,! B: (4.153)

• One knows hk .id0 / for reference mappings of type (4.135),

˚
id0 W f 2 Aps00 q0 .Rn / W supp f  U ,! Aps11 q1 .Rn /: (4.154)

Afterwards one can switch from a counterpart of Theorem 4.33 (ii) with hk in place
ak (if exists) to the corresponding counterparts of Theorem 4.35 and Corollary 4.38 by
the same arguments as above.

Remark 4.41. So far we replaced the assumption that  is bounded in Theorem 4.33
by jj < 1 in Theorem 4.35 and in the parts (ii) and (iii) of Corollary 4.38. Then
one has the sequence spaces in (4.136), (4.137) with d D n. But there is an elaborated
theory for compact embeddings between sequence spaces of this type for all d > 0
covering at least equivalence assertions for the corresponding entropy numbers. We
refer to [T06], Section 6.3, and the literature given there. If, for example,  is E-thick
with jj D 1 and
ˇ˚ ˇ
ˇ x 2  W ı.x/  2j ˇ  2"j ; j 2 N; (4.155)

for some " > 0, where ı.x/ has the same meaning as in (4.104), then it seems to
be possible to obtain equivalence assertions at least for the entropy numbers of id in
(4.139). As far as sequence spaces are concerned there are far-reaching generalisations
of the above sequence spaces of type (4.136), (4.137) for which one has equivalence
assertions for entropy numbers of related compact embeddings. We refer to [T06],
Remark 6.18, pp. 278–79, where we collected corresponding papers. It remains to be
seen to which extent these results can be used in the above context. One may even ask
for inverse assertions. For this purpose it might be useful to convert the decreasing
sequence fek g of entropy numbers of a compact embedding between two function
spaces into a decreasing (which means non-increasing) function on Œ0; 1/,

e.t / D ek for k  1  t < k; k 2 N: (4.156)

 4.3 Complements 129

Of course, e.t/ ! 0 if t ! 1. Let g.t / be a, say, continuous, positive decreasing

function on Œ0; 1/ with g.t / ! 0 if t ! 1. Then the inverse entropy problem is the
question whether there is a compact embedding, say, of type (4.139) such that

g.t /  e.t /; 0  t < 1: (4.157)

Having (4.128) in mind one may suppose that g.t /  ct K for some c > 0, K > 0.
Similarly one can formulate an inverse approximation problem replacing the entropy
numbers fek g in the above comments by the approximation numbers fak g.
Chapter 5
Spaces on smooth domains and manifolds

5.1 Wavelet frames and wavelet-friendly extensions

5.1.1 Introduction

All wavelet expansions in function spaces on domains  in Rn considered so far

originate from the u-wavelet basis
˚ j
ˆr W j 2 N0 I r D 1; : : : ; Nj x
with Nj 2 N (5.1)

according to Theorem 2.33 and Definitions 2.31, 2.4. The wavelets ˆjr have compact
supports in ,
supp ˆjr  : (5.2)
This makes clear that wavelet expansions as considered, for example, in Theorem 3.13
cannot be expected for spaces Apq s
./ having boundary values at D @. We
discussed this point at the beginning of Section 4.3.2. Nevertheless we obtained in
Theorem 2.55 for the Sobolev spaces Wpk ./ in arbitrary domains and in Theorem 4.23
s
for all spaces Apq ./ in bounded Lipschitz domains so-called constrained wavelet
expansions characterising the corresponding spaces in terms of sequence spaces. But
these are not wavelet expansions converging in the spaces themselves. We have now a
closer look at problems of this type assuming preferably that  is a bounded C 1 domain
in Rn with n  2 according to Definition 3.4 (iii) or a bounded interval  D .a; b/
with 1 < a < b < 1 if n D 1. But at the end of this Chapter 5 we deal briefly with
spaces on so-called cellular domains and comment on spaces in Lipschitz domains and
."; ı/-domains. In Section 6.1 below we return to cellular domains in greater detail.
To avoid any complications in connection with traces we restrict ourselves mostly
to the spaces
s
Apq ./ with s > 0; 1  p < 1; 1  q < 1; (5.3)
according to Definition 2.1. If p < 1 then the trace problem is rather delicate. We
return to this point in Section 6.4 below. Otherwise we assume that the reader is familiar
with basic facts about traces of, say, Sobolev spaces and classical Besov spaces on the
boundary D @ of a smooth domain . We fix some notation. As usual, is
furnished with the .n  1/-dimensional Hausdorff measure (the standard surface
measure). Then Lp . / with 1  p < 1 is the usual Banach space of all complex-
valued -measurable functions f on such that
Z 1=p
kf jLp . /k D jf . /jp .d / (5.4)

 5.1 Wavelet frames and wavelet-friendly extensions 131

is finite. Since we excluded p D 1 and q D 1 in (5.3) the restriction S./ D

S.Rn /j of S.Rn / on  is dense in the space Apq
s
./. If ' 2 S./ then the pointwise
trace '. / D .tr  '/. / with 2 makes sense. Let Apq s
./ be as in (5.3). One
asks whether there is a positive number c such that
k tr  ' jLp . /k  c k' jApq
s
./k for all ' 2 S./. (5.5)
If this is the case then one defines the trace operator tr  ,
tr  W Apq
s
./ ,! Lp . / (5.6)
in the standard way via Cauchy sequences. This is the usual point of view to say what
is meant by traces which is sufficient for our (and almost all) purposes. But one can
give a more direct (albeit quite often less effective) pointwise definition. It is based on
the theory of Lebesgue points asking for which x 2 Rn one has
Z
g.x/ D lim jB.x; r/j1 g.y/ dy; (5.7)
r!0 B.x;r/

where B.x; r/ stands again for a ball centred at x 2 Rn and of radius r and g is
locally integrable in Rn . Let g be the distinguished representative of the corresponding
equivalence class with (5.7) in all points x 2 Rn for which the right-hand side of (5.7)
converges. If one has (5.5) (where one can replace  on the right-hand side by Rn ) then
one obtains for f 2 Apq s
./ and the above distinguished representative g 2 Apq s
.Rn /
with f D gj, that
Z
1
.tr  f /. / D f . / D g. / D lim jB. ; r/j g.y/ dy (5.8)
r!0 B.;r/

-a.e. (up to a set of -measure zero on ). But this is a rather sophisticated observation
based on the theory of capacity which will not be needed here. Some details can be
found in [T01], pp. 260–61, with a reference to [AdH96]. One may also consult in this
context [HeN07].
Let  D . / with 2 D @ be the outer normal at the boundary of the
above bounded C 1 domain  in Rn (bounded interval if n D 1). Let Apq s
./ with

A 2 fB; F g be as in (5.3). If a 2 R then Œa is the largest integer strictly less than a.
Let ˚ @j f 1 
tr s;p
 W f 7! tr  @ j W 0  j  s  p (5.9)
(empty if s  1=p). By the references given later on tr s;p
 ,

Y
Œs1=p
s 1 j
tr s;p

s
Bpq ./ D Bpq p . / (5.10)
j D0

and 
Y
Œs1=p
s 1 j
tr s;p

s
Fpq ./ D Bpp p . / (5.11)
j D0
132 5 Spaces on smooth domains and manifolds

are linear and bounded maps onto the indicated (vector) spaces. Let AQpq
s
./ be the
spaces introduced in Definition 2.1 (ii) with s; p; q and  as above. Again by the
references and the precise interpretations given in Theorem 5.21 and Remark 5.22
below one has 
Y
Œs1=p
s 1 j
z
Bpq ./ D Bpq ./ 
s s
Bpq p . / (5.12)
j D0

and 
Y
Œs1=p
s 1 j
s
Fpq ./ D Fzpq
s
./  Bpp p . / (5.13)
j D0

if
1
1<s 62 N0 ; 1  p; q < 1: (5.14)
p
If 1 < s  1
p
< 0 then (5.13), (5.14) means
s
Apq ./ D AQpq
s
./ (5.15)

where again a reference will be given later on. But it is essentially also covered by
Proposition 3.19 where one may incorporate p D 1 and choose any ı with ı < 1.
Although (5.12), (5.13) require a precise interpretation one can imagine that these
decompositions pave the way to construct wavelet bases and wavelet frames for the
s
spaces Apq ./ under consideration. Wavelet bases for the spaces AQpq
s
./ are covered
by Theorems 3.13, 3.23 and Corollary 3.25. Next one needs wavelet expansions for the

spaces Apq . / on compact C 1 manifolds . This will be done in Section 5.1.2. Then
one has to transfer the wavelet decomposition for spaces on to  which is the subject
of Section 5.1.3 where we construct wavelet-friendly extension operators which might
be of interest for its own sake. The rest of Section 5.1 deals with the combination of
these wavelet expansions resulting in wavelet frames for the spaces in (5.3). In some
cases these frames are bases. This is the subject of Sections 5.2 and 5.3. Finally we
describe in Section 5.4 some alternative constructions.

5.1.2 Wavelet frames on manifolds

Let be a compact metric space furnished naturally with a topology. Let n 2 N.

Then is called a compact n-dimensional C 1 manifold if for some
SMM 2 N there is
an atlas fVm ; m gmD1 consisting of open sets Vm in such that mD1 Vm D and
M

homeomorphic maps m ,

mW Vm () Um D m .Vm /  Rn ; m D 1; : : : ; M; (5.16)
1
of Vm onto connected bounded open sets Um in Rn such that k B m ,
1
k B m W m .Vm \ Vk / () k .Vm \ Vk / if Vm \ Vk 6D ;; (5.17)
 5.1 Wavelet frames and wavelet-friendly extensions 133

are diffeomorphic C 1 maps with positive Jacobians. Here (5.17) are the usual com-
patibility conditions for overlapping open sets Vm . Essentially we are only interested
in compact .n  1/-dimensional C 1 manifolds which are boundaries of n-dimensional
bounded C 1 domains according to Definition 3.4 (iii). Then one has natural atlases.
This may justify that we do not discuss the above definition on an abstract level. The in-
terested reader may consult [BrL75], Section 1.1, [Hel78], §1, or [Tri86], Section 29.1.
We furnish the above n-dimensional compact C 1 manifold with the n-dimensional
Hausdorff measure . Details may be found in [Mat95], Section 4. Of course the im-
age measure m Πrestricted to Vm is equivalent to the Lebesgue measure in Um . As
before, Lp . / with 0 < p  1 is the collection of all complex-valued -measurable
functions on such that
Z 1=p
kf jLp . /k D jf . /j .d /
p
(5.18)


is finite (usual modification if p D 1). Otherwise one can lift via the above mappings
1
m functions, distributions and function spaces from R to . This applies in partic-
n

ular to the space of test functions D. / D C . / and its dual D 0 . /, the space of
1

distributions on . Let fm gM mD1  D. / be a resolution of unity such that

X
M
supp m  Vm and m . / D 1 if 2 : (5.19)
mD1

If f 2 D 0 . / then
1
.m f / B m 2 D 0 .Um /  S 0 .Rn /; m D 1; : : : ; M: (5.20)

Definition 5.1. Let be a compact n-dimensional C 1 manifold. Let A 2 fB; F g and

let s 2 R, 0 < p  1 .p < 1 for the F -spaces/ and 0 < q  1. Then
˚
s
Apq . / D f 2 D 0 . / W .m f / B m
1
2 Apq
s
.Rn /; m D 1; : : : ; M (5.21)

and
X
M
1
kf jApq
s
. /k D k.m f / B m jApq
s
.Rn /k: (5.22)
mD1

Remark 5.2. It follows by standard arguments (diffeomorphic maps, pointwise multi-

pliers) that these spaces are independent (equivalent quasi-norms) of compatible atlases
(local charts) and related resolutions of unity.
We wish to transfer the wavelet expansions for spaces on Rn according Theo-
rem 1.20 to spaces on . For this purpose one needs counterparts of the wavelets in
Theorem 1.20 and of some sequence spaces. It is reasonable to ask for -counterparts
of Definitions 2.4, 2.6. The sequence spaces fit in this scheme. But the cancellations
of the interior wavelets in (2.33), going back to (1.87)–(1.91), are not preserved if mul-
tiplied with C 1 cut-off functions and the subject of diffeomorphic distortions. The
134 5 Spaces on smooth domains and manifolds

cancellations for the wavelets originate from the corresponding cancellations of the
.s; p/-atoms in Rn according to (1.31). But there is a suitable modification. For this
purpose one replaces (1.31), now for all j 2 N0 , by
ˇZ ˇ X
ˇ ˇ
ˇ .x/ a .x/ dx ˇ  D 2j.s pn /.LCn/j sup jD ˛ .y/j; (5.23)
ˇ jm ˇ
dQj m
j˛jL

where the supremum is taken over all y 2 dQj m . Here D is a positive constant. In
other words, we modify Definition 1.5 and the explanations given in Remark 1.6 as
follows.
Let s, p, K, L, d be as in Definition 1.5 and let D > 0. Then the L1 -functions
aj m W Rn 7! C with j 2 N0 , m 2 Zn , are called .s; p/ -atoms .more precisely
.s; p/K;L;d;D -atoms/ if one has (1.29), (1.30), and (5.23) for all functions
having classical derivatives up to order L in dQj m .
In other words, one replaces (1.31) by (5.23) for given D > 0 and all . Expanding
in its Taylor polynomial with remainder term of order L and off-point 2j m then
(5.23) follows from (1.31) for all
D  D.n; L/; n 2 N; L 2 N0 ; (5.24)
where D.n; L/ is some positive constant depending only on n and L (which can be
calculated explicitly). Then any .s; p/K;L;d -atom according to Definition 1.5 and
Remark 1.6 is an .s; p/K;L;d;D -atom if D satisfies (5.24).
Proposition 5.3. Theorem 1.7 remains valid if one replaces .s; p/-atoms .more pre-
cisely .s; p/K;L;d -atoms/ by .s; p/ -atoms .more precisely .s; p/K;L;d;D -atoms with
D  D.n; L/ according to (5.24)/.
Remark 5.4. One has to show that (1.34), (1.35) and (1.34), (1.37) remain valid for
the larger class of .s; p/ -atoms according to the above definitions. This has been done
in detail in [Skr98b]. Multiplications with smooth functions and (local) diffeomor-
phic maps of .s; p/ -atoms preserve essentially the crucial assumptions (1.29), (1.30),
(5.23).
After these preparations one can now introduce the -counterpart of Definitions 2.4
and 2.6. Let be again a compact n-dimensional C 1 manifold. Let B. ; %/ be a ball
centred at 2 and of radius % > 0 (with respect to the metric on ). For some c > 0
let ˚
Z D rj 2 W j 2 N0 I r D 1; : : : ; Nj ; Nj 2 N; (5.25)
with (typically) Nj  2j n such that
 j j

j
r 0 62 B r ; c2 ; j 2 N0 ; r 6D r 0 : (5.26)
This is the counterpart of (2.24), (2.25). Of interest later on are sets Z with
Nj
[  
0
D B j
r ;c 2j ; j 2 N0 ; (5.27)
rD1
 5.1 Wavelet frames and wavelet-friendly extensions 135

for some c 0 > 0. Let C u . / be the usual space of complex-valued functions having
classical derivatives up to order u 2 N inclusively. Since is compact there is no need
to bother about invariantly defined derivatives. We simply write D ˛ f for f 2 C u . /
and X
kf jC u . /k  sup jD ˛ f . /j (5.28)
2
j˛ju

in the interpretation via local charts. Recall that in (5.18) is the n-dimensional
Hausdorff measure on .

Definition 5.5. Let be a compact n-dimensional C 1 manifold and let Z be as in

(5.25), (5.26). Let n 2 N.
(i) Then the collection
˚
ˆ D ˆjr W j 2 N0 I r D 1; : : : ; Nj  C u . / (5.29)

of real functions is called a u-wavelet system if for some positive numbers c1 , c2 ,

 
supp ˆjr  B rj ; c1 2j ; j 2 N0 I r D 1; : : : ; Nj ; (5.30)

and for all ˛ 2 N0 with j˛j  u,

n
jD ˛ ˆjr . /j  c2 2j 2 Cj j˛j ; j 2 N0 I r D 1; : : : ; Nj : (5.31)

(ii) The above u-wavelet system is called oscillating if, in addition, for some c3 > 0,
ˇZ ˇ
ˇ ˇ
ˇ . / ˆr . / .d /ˇˇ  c3 2j 2 j u k jC u . /k;
j n
(5.32)
ˇ


j 2 N0 ; r D 1; : : : ; Nj .

Remark 5.6. This is the counterpart of Definition 2.4 based on (2.32), (2.33), (2.27). As
there we prefer for wavelets (in contrast to atoms) an L2 -normalisation. This explains
the factors 2j n=2 and 2j n=2 D 2j n=2 2j n in (5.31), (5.32). Furthermore we rely on the
same coupling of smoothness and cancellation as in (1.87)–(1.91) expressed by u 2 N.
The above wavelets are more qualitative than there counterparts in Definition 2.4 and
there is no need to distinguish any longer between basic wavelets as in (2.32) and
interior wavelets as in (2.33). It is obviously immaterial whether one requires (5.32)
for j 2 N or j 2 N0 (including the starting wavelets). In terms of local charts, (5.32)
reduces to (5.23) with L D u (having in mind the different normalisations).

Definition 5.7. Let be a compact n-dimensional C 1 manifold and let Z be as in

(5.25), (5.26). Let jr be the characteristic functions for the balls in (5.26). Let s 2 R,
0 < p  1, 0 < q  1. Then bpq s
.Z / is the collection of all sequences
˚
 D jr 2 C W j 2 N0 I r D 1; : : : ; Nj ; Nj 2 N; (5.33)
136 5 Spaces on smooth domains and manifolds

such that
X
1 X
Nj q=p 1=q
n
j.s p /q
k jbpq
s
.Z /k D 2 jjr jp < 1; (5.34)
j D0 rD1

s
and fpq .Z / is the collection of all sequences (5.33) such that

 X Nj
1 X 1=q 
 
k jfpq
s
.Z /k D 2jsq jjr jr .  /jq jLp . / < 1 (5.35)
j D0 rD1

.obviously modified if p D 1 and/or q D 1/.

Remark 5.8. This is the direct counterpart of Definition 2.6 where the quasi-norm
in Lp . / is given by (5.18) independently of local charts. As mentioned there if m
in (5.16) is applied to (image measure) then m Πis equivalent to the Lebesgue
measure in Um . Then one has the same situation as in Remark 2.7 making clear
s
that fpq .Z / is independent of the admitted characteristic functions jr and admitted
atlases. The Hilbert space L2 . / is furnished with the scalar product
Z
.f; g/ D f . / g. / .d /: (5.36)


The corresponding scalar product with Vm in place of can be transferred by (5.16)

to Um , Z
 1
   d
.f; g/Vm D f B m 1 .x/ 
.x/  g B m .x/ dx: (5.37)
Um dx
One may assume that the Radon–Nikodym derivative ddx is a C 1 function in Um
with ddx .x/  c > 0 if x 2 Um . This can be justified by transferring the Lebesgue
measure in Um , multiplied with a smooth resolution of unity, to . Then one obtains a
measure with the desired property which is at least equivalent to the above Hausdorff
measure . We identify with this measure. (Recall that in what follows we are only
interested in C 1 manifolds which are boundaries of bounded C 1 domains according
to Definition 3.4 (iii) where the above assertion can be justified directly).
After these preparations one obtains now the following counterparts of Theo-
rem 1.20 for spaces on Rn and of Theorem 1.37 for spaces on T n . The numbers
p and pq have the same meaning as in (4.15). We use  as in (3.4), (3.5).
Theorem 5.9. Let be a compact n-dimensional C 1 manifold. Let Apq s
. / be the
spaces as introduced
˚ in Definition
˚ 5.1. For any u 2 N there are two oscillating
u-wavelet systems ˆjr and ‰rj according to Definition 5.5 (ii) with the following
properties.
(i) Let s 2 R, 0 < p < 1, 0 < q  1, and

u > max.s; pq  s/: (5.38)

 5.1 Wavelet frames and wavelet-friendly extensions 137

Then f 2 D 0 . / is an element of Fpq

s
. / if, and only if, it can be represented as

1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 fpq
s
.Z /; (5.39)
j D0 rD1

unconditional convergence being in Fpq

s
. / if q < 1 and in Fpq
s"
. / for any " > 0
if q D 1. Furthermore,

kf jFpq
s
. /k  inf k jfpq
s
.Z /k (5.40)

where the infimum is taken over all admissible representations (5.39). Any f 2 Fpq
s
. /
admits the distinguished representation (5.39),
1 X
X Nj
 
f D jr .f / 2j n=2 ˆjr ; jr .f / D 2j n=2 f; ‰rj  ; (5.41)
j D0 rD1

with
kf jFpq
s
. /k  k.f / jfpq
s
.Z /k (5.42)
.u-wavelet frame/.
(ii) Let s 2 R, 0 < p  1, 0 < q  1 and

u > max.s; p  s/: (5.43)

Then f 2 D 0 . / is an element of Bpq

s
. / if, and only if, it can be represented as

1 X
X Nj
f D jr 2j n=2 ˆjr ;  2 bpq
s
.Z /; (5.44)
j D0 rD1

unconditional convergence being in Bpq

s
. / if p < 1, q < 1 and in Bpq
s"
. / for
any " > 0 if p D 1 and/or q D 1. Furthermore,

kf jBpq
s
. /k  inf k jbpq
s
.Z /k (5.45)

where the infimum is taken over all admissible representations (5.44). Any f 2 Bpq
s
. /
admits the distinguished representation (5.41), (5.44), with

kf jBpq
s
. /k  k.f / jbpq
s
.Z /k (5.46)

.u-wavelet frame/.
0
Proof. Step 1. Let A 2 fB; F g and correspondingly ˚ aj 2 fb; f g. Let f 2 D . / be
given by (5.39), (5.44) for some u-wavelet system ˆr with (5.38), (5.43). We rely
on (5.21). One has
X  
1
.m f / B m D jr 2j n=2 m ˆjr B m 1
: (5.47)
j;r
138 5 Spaces on smooth domains and manifolds

Adapting the normalising factors one has by Definition 5.5 that (5.47) is an expansion
in terms of .s; p/ -atoms to which Proposition 5.3 can be applied. Then one obtains
1
that .m f / B m 2 Apq
s
.Rn / and
kf jApq
s
. /k  c k japq
s
.Z /k: (5.48)
Step 2. As for the converse one expands
X j;G  
1 1
.m f / B m D l m f B m 2j n=2 ‰G;l
j
(5.49)
j;G;l
˚ j
according to Theorem 1.20 with respect to the wavelet basis ‰G;l , where
Z
 1
  1

j;G
l
m f B m D 2j n=2 m f B m j
.x/ ‰G;l .x/ dx: (5.50)
Rn
j j
One may assume that supp ‰G;l  Um according to (5.16) for all ‰G;l with
j  1

supp ‰G;l \ supp m B m 6D ;: (5.51)
We transfer (5.49) from Um to . Then (5.37) and summation over m give (5.41).
Using (5.22) one obtains (5.42), (5.45).
Remark 5.10. We proved a little bit more than stated. One has (5.48) for any expan-
sion by appropriately re-normalised .s; p/ -atoms according to Proposition 5.3 and
transferred from Rn to . In other words there is a full counterpart of Theorem 1.7
s
in the version of Proposition 5.3 for spaces Apq . / on compact n-dimensional C 1
manifolds .
Remark 5.11. In (1.100), (1.101) we recalled what is meant by a (unconditional) basis
in quasi-Banach spaces. The notation of a frame in quasi-Banach spaces is not so fixed
but in common use nowadays. We explain our ˚ understanding of this notation taking
Bpqs
. / as a proto-type. A countable system ˆjr is called a (stable) frame in Bpq
s
. /
if it has the following two properties.
s
1. (Stability) There is a natural sequence space, bpq .Z /, such that any f 2 Bpq
s
. /
can be represented by (5.44) with (5.45).
2. (Optimality) There are linear and continuous functionals
jr 2 Bpq
s
. /0 W f 2 Bpq
s
. / 7! jr .f / 2 C; (5.52)
such that any f 2 Bpq
s
. / can be represented by

1 X
X Nj
f D jr .f / 2j n=2 ˆjr ; .f / 2 bpq
s
.Z /; (5.53)
j D0 rD1

with (5.46).
 5.1 Wavelet frames and wavelet-friendly extensions 139

To avoid any misunderstanding we recall that (5.46) means that there are two numbers
0 < C1  C2 < 1 such that

C1 kf jBpq
s
. /k  k.f / jbpq
s
.Z /k  C2 kf jBpq
s
. /k (5.54)

for all f 2 Bpq

s
. /. If the underlying space is a separable Hilbert space, then `2 is the
natural related sequence space. If one starts with a Banach space or even quasi-Banach
space then the search for a suitable sequence space is part of the task. Unconditional
convergence of (5.44) or (5.53) can only be expected if the underlying space is separa-
ble, this means p < 1, q < 1 in case of Bpq s
. /. If the given space is not separable,
for example Bpq . / with max.p; q/ D 1, then convergence takes place in a larger
s
s"
space, for example Bpq . / with " > 0, but the outcome belongs to the given space.
Furthermore (5.52), (5.53) and suitable counterparts for other quasi-Banach spaces
require that the space considered has a sufficiently rich dual space. But this is not
always the case. Since, for example, Lp .Rn /0 D f0g if 0 < p < 1, nothing like (5.53),
(5.52) can be expected. But this unpleasant effect happens also for some other more
interesting spaces. We return to this point in Section 6.2.
Remark 5.12. One may ask whether there are u-wavelet systems according to Defi-
nition 5.5 which are not only frames but even (unconditional) bases. In general this is
not so clear. But it is possible in some cases and for some modified u-wavelet systems.
This will be the subject of Sections 5.2, 5.3.
Remark 5.13. As described in Introduction 5.1.1 it is the main topic of Chapter 5 to find
s
wavelet frames and wavelet bases for spaces Apq ./ using the decompositions (5.12),
(5.13). For this purpose we need the above considerations applied to the compact
.n  1/-dimensional C 1 manifold D @. But it would be desirable to study
wavelets (frames and bases) on more general C 1 manifolds . In [T92], Chapter 7,
s s
and the underlying papers we introduced and studied spaces Bpq . / and Fpq . / on
(non-compact) n-dimensional Riemannian manifolds with positive injectivity radius
and bounded geometry and on Lie groups. The further development of this theory is
mostly due to L. Skrzypczak, [Skr97], [Skr98a], [Skr98b], [Skr03]. In these papers
one finds in particular an elaborated theory of atomic decompositions based on atoms
of the same type as used in Proposition 5.3, where conditions of type (5.23) play a
decisive role. Wavelet frames in these spaces have been studied in the recent paper
[Skr08].

5.1.3 Wavelet-friendly extensions

In Introduction 5.1.1 we outlined our method in this chapter. We reduce the question
s
of wavelet frames and wavelet bases for the spaces Apq ./ in (5.3) to corresponding
assertions for the spaces in the decompositions (5.12), (5.13). By Theorem 3.13 we
have interior wavelet bases for all spaces AQpq
s 
./ of interest. As for the spaces Bpq . /
on the boundary D @ we wish to use the wavelet frames (or related wavelet
140 5 Spaces on smooth domains and manifolds

bases discussed later on) according to Theorem 5.9. One has to shift the wavelets on
the compact .n  1/-dimensional C 1 manifold to the n-dimensional bounded C 1
domain  creating wavelets in  x having all the expected properties of wavelets in
x but no compact supports in . This will be done by constructing wavelet-friendly
,
extension operators which might be of interest for its own sake. For this reasons we
prove the corresponding assertion in full generality. First we adapt some formulations
from Introduction 5.1.1.
If p  1 then one can replace Lp . / in (5.5), (5.6) by L1 . /. Hence we ask for a
positive constant c such that

k tr  ' jL1 . /k  c k' jApq

s
./k for all ' 2 S./: (5.55)

This does not change to say what is meant by traces for the spaces in (5.3) and also
not the discussion in (5.7), (5.8) which is essentially an L1 -matter. But we extend now
s
(5.55) to spaces Apq ./ with p D 1 and/or q D 1 and to spaces with p < 1. For
given r 2 N0 and u 2 N with r < u, we will ask for common trace operators (from 
to ), now denoted by tr r , and for common extension operators (from to ), now
denoted by extr;u
 , in natural .s; p; q/-regions

1
0 < p  1; 0 < q  1; rC C pn1 < s < u; (5.56)
p
where  
1
pn1 D .n  1/ 1 ; 0 < p  1; 2  n 2 N; (5.57)
p C

is the .n  1/-dimensional version of p in (4.15) indicating the dimension n  1

to avoid any misunderstanding. Furthermore, if fBl W l D 1; : : : ; Lg are (complex)
quasi-Banach spaces then the collection of all b D fb1 ; : : : ; bL g with bl 2 Bl and
quasi-normed by
XL Y
L
kbl jBl k is denoted as Bl : (5.58)
lD1 lD1

Instead of tr s;p
 in (5.9) we write now
˚ @j f
tr r W f 7! tr  @ j
W 0j r : (5.59)

Then we obtain by [T83], Theorem 3.3.3, p. 200, that

Y
r
s 1 j
tr r Bpq
s
./ D Bpq p . / (5.60)
j D0

if
1
0 < p  1; 0 < q  1; rC C pn1 < s; (5.61)
p
 5.1 Wavelet frames and wavelet-friendly extensions 141

and
Y
r
s 1 j
tr r s
Fpq ./ D Bpp p . / (5.62)
j D0

if
1
0 < p < 1; 0 < q  1; rC C pn1 < s: (5.63)
p
We add a comment how (5.55) is related to these trace assertions. First we remark that

Lp . / ,! L1 . / if 1  p  1: (5.64)

By (5.60)–(5.63) it follows that it is immaterial to replace (5.5) by (5.55) if p  1. If

0 < p < 1 then one has for ı > 0,
 
1 1 n
sDıC C .n  1/  1 D  n C 1 C ı: (5.65)
p p p
Then it follows from well-known embedding theorems as one may found in [T83],
Theorem 3.3.1, p. 196, that
s 1Cı
Bpq ./ ,! B1;q ./ ,! L1 . /: (5.66)

Again (5.55) makes sense. If p < 1, q < 1, then S./ is dense in Apq s
./ and, as
said, tr  , and then also tr  , are defined by completion. If p D 1 then B1;q
r s
./ with
s > 0 is embedded in the space of continuous functions and tr  makes sense pointwise.
If q D 1 then one has
s s"
Bp;1 ./ ,! Bp;1 ./ for any " > 0: (5.67)

Let " > 0 be small such that one has (5.61) with s  " in place of s. Now one can define
s s" s
tr  for Bp;1 ./ by restriction of tr  for Bp;1 ./ to Bp;1 ./. Hence (5.55) is always
meaningful. As far as (5.60)–(5.63) is concerned we have nothing new to add. The
assertions in [T83], Theorem 3.3.3, p. 200, cover also the existence of corresponding
extension operators. But this is based on Fourier-analytical arguments and of little
use for our purpose. In [T92], Section 4.4, we returned to the trace problem, based
on atoms. Although nearer to our recent intentions it does not cover our needs. For
this reason we construct explicitly a wavelet-friendly extension operator extr;u
 , where
r 2 N0 and u 2 N have the same meaning as in (5.56), (5.61).
Let  be a bounded C 1 domain in Rn according to Definition 3.4 (iii) with
2  n 2 N and let D @ be its .n  1/-dimensional boundary considered as
an .n  1/-dimensional compact C 1 manifold. We recalled at the beginning of Sec-
tion 5.1.2 what is meant by compact C 1 manifolds (without boundaries). We need
some preparations. Let  be the outer normal at and let
(
dist.x; / if x 2 ;
d.x/  (5.68)
 dist.x; / if x 2 Rn n ;
142 5 Spaces on smooth domains and manifolds

be a C 1 function in a tubular neighbourhood

" D fx 2 Rn W dist.x; / < "g (5.69)
of for some (small) " > 0. We may assume that
@k l @k k
d .x/ D 0 if l D
6 k and d .x/ D kŠ: (5.70)
@ k @ k
Then " can be furnished for small " > 0 with curvilinear coordinates
" 3xD D . 0; n/ with 0
2 and n D d. / (5.71)
pointing in the -direction in agreement with (5.70). Let .t / be a real function on R
with
 2 D.R/; supp   ."; "/; .t / D 1 if jt j  "=2: (5.72)
For given L 2 N0 one may assume in addition that
Z
.t / t l dt D 0; l D 0; : : : ; L  1; (5.73)
R
(no condition if L D 0). This is of some service in the following
˚ theorem
˚ for spaces
s
Fpq ./ with q > 0 small. It will not be indicated. Let ˆjl and ‰lj be two real
u-wavelet systems according to Definition 5.5 (i) on the compact .n  1/-dimensional
manifold . Let
Z
j n1
 j j n1
j
l .h/ D 2 2 h; ‰l  D 2 2 h. 0 / ‰lj . 0 / .d 0 / (5.74)


where .; / is the scalar product (5.36) on (recall that ‰lj is real). Then we put for
some r 2 N0 and fgk grkD0  L1 . /,
g D Extr;u
 fg0 ; : : : ; gr g

X
r X Nj
1 X
1 j n1
D l .gk / 2j 2 k
n .2j j 0
n / ˆl . /
kŠ (5.75)
kD0 j D0 lD1

X
r X Nj
1 X
1 j
D l .gk / 2j k 2j n=2 ˆj;k
l
. /
kŠ
kD0 j D0 lD1

with jl .gk / as in (5.74) and

ˆj;k
l
. / D 2j k k
n .2j n/ 2
j=2
ˆjl . 0 /: (5.76)
Each term in (5.75) makes sense. Under the restrictions for gk in the theorem below
s
it follows that (5.75) is an atomic decomposition in the related spaces Apq .Rn /. This
ensures the convergence of (5.75) in the same way as in Theorem 1.7, Remark 1.8 and
the references given there. We will not stress this point in the sequel. In Remark 4.2
we said what is meant by a common extension operator. This will be used, obviously
adapted. Let re D re be the restriction operator according to (4.1).
 5.1 Wavelet frames and wavelet-friendly extensions 143

Theorem 5.14. Let  be a bounded C 1 domain in Rn with n  2 according to

Definition 3.4 (iii) or a bounded interval in R if n D 1. Let Apq
s
./ and Apq
s
. / be the
spaces as introduced in Definition 2.1 (i) on  and in Definition 5.1 on the compact
.n  1/-dimensional manifold D @ .the two endpoints of the interval if n D 1/.
Let r 2 N0 , u 2 N with r < u and " 2 R with 0 < "  1. Let tr r be the trace operator
(5.59) with (5.60), (5.61) and (5.62), (5.63). Let Extr;u
 be given by (5.75) with (5.74),
˚ ˚
(5.76) where ˆjl and ‰lj are the two u-wavelet systems on from Theorem 5.9
 
and L  n 1"  1 in (5.73). Then
ext r;u r;u
 D re B Ext 

with
Ext r;u
 W fg0 ; : : : ; gr g 7! g; (5.77)
according to (5.75) is a common extension operator for
Y
r
s 1 k
extr;u
 W Bpq p s
. / ,! Bpq ./;
kD0 (5.78)
0 < p  1; 0 < q  1; rC 1
p
C pn1 < s < u;
and
Y
r
s 1 k
extr;u
 W Bpp p s
. / ,! Fpq ./;
kD0 (5.79)
0 < p < 1; "  q  1; rC 1
p
C pn1 < s < u;
with
Y
r
s 1 k
tr r B extr;u
 D id; identity in Bpq p . /: (5.80)
kD0

Proof. Step 1. We prove the theorem for the B-spaces. For p, q, s as in (5.78) we need
now only the weaker version
Y
r
s 1 k
tr r W Bpq
s
./ ,! Bpq p . / (5.81)
kD0

of (5.60) based on [T83], Theorem 3.3.3, p. 200. Let for the same p, q, s,
s 1 k
gk 2 Bpq p . /; k D 0; : : : ; r: (5.82)
Since u > s  p1  k > pn1 for k D 0; : : : ; r one can apply Theorem 5.9 to all spaces
˚ ˚
on the right-hand side of (5.82) based on common u-wavelet systems ˆjl and ‰lj .
One has by this assertion and (5.34) that
1 X
X Nj
n1
0
gk . / D jl .gk / 2j 2 ˆjl . 0 /; 0
2 ; (5.83)
j D0 lD1
144 5 Spaces on smooth domains and manifolds

with
X
1 X
Nj q=p 1=q
s 1 k n
k/q
kgk jBpq p . /k  2j.s p
jjl .gk /jp (5.84)
j D0 lD1

(usual modification if p D 1 and/or q D 1),

Z
n1
jl .gk / D 2j 2 gk . 0 / ‰lj . 0 / .d 0 /; (5.85)

0
having in mind that is .n  1/-dimensional. By (5.30), (5.31) (with in place of )
one obtains for ˆj;k
l
in (5.76) that
 0j j

supp ˆj;k
l
 Q j l D B l
; c1 2   "2j ; "2j (5.86)

centred at j
l
D . l0j ; 0/ with diameter  2j and
ˇ ˛ j;k ˇ
ˇD ˆ . /ˇ  c 2j n2 Cj j˛j ; j 2 N0 I l D 1; : : : ; Nj ; (5.87)
l

for all ˛ 2 Nn0 with j˛j  u (in Rn ). With

1 j
j
akl D 2j n=2 ˆj;k
l
and j
kl
D  .gk / 2j k (5.88)
kŠ l
one can re-write (5.75) as X j j
gD kl
akl (5.89)
k;j;l
j s
where akl are L1 -normalised atoms in Bpq .Rn / (no moment conditions are required).
Here one needs that  
1
n 1 D pn D p < s < u; (5.90)
p C
covered by (5.78) and (5.65). Then it follows from Theorem 1.7 and (5.84) that
r X
X 1 X
Nj q=p 1=q
n j p
j.s p /q
kg jBpq
s
.Rn /k c 2 j kl
j
kD0 j D0 lD1
r X
X 1 X
Nj q=p 1=q
n
0
c 2 j.sk p /q
jjl .gk /jp (5.91)
kD0 j D0 lD1
X
r
s 1 k
 c 00 kgk jBpq p . /k:
kD0

In particular, g 2 s
Bpq .Rn /. The coefficients in (5.89), (5.88) depend linearly on gk .
Hence Extr;u
 is (5.77), (5.75) is a linear and bounded map,
Y
r
s 1 k
Ext r;u
 W Bpq p s
. / ,! Bpq .Rn /: (5.92)
kD0
 5.1 Wavelet frames and wavelet-friendly extensions 145

We check that Extr;u

 is an extension operator. Taking the derivative with respect to the
normal  then one obtains by (5.70)–(5.72), (5.75) and (5.83) that
  X1 X Nj
@k g 0 n1

k
. ; 0/ D jl .gk / 2j 2 ˆjl . 0 /
@ j D0
(5.93)
lD1
0
D gk . /; k D 0; : : : ; r:

Hence,
Y
r
s 1 k
tr r B Extr;u
 D id; identity in Bpq p . /: (5.94)
kD0

Since the traces for spaces on Rn and on  are the same one obtains now (5.78), (5.80)
for ext r;u r;u
 D re B Ext  .
Step 2. It is well known that traces for Fpqs
.Rn / and Fpqs
./ on , if exist, are
independent of q, hence they coincide with the traces of Bpp .Rn / and Bpp
s s
./. This
applies in particular to (5.62), (5.63). It is also covered by [T83], Theorem 3.3.3,
p. 200, based on Fourier-analytical arguments. But the main point is not so much the
smoothness of but its porosity according to Definition 3.4 (i). With g as in (5.89) we
re-write the first estimate in (5.91) as

X
r
kg jBpq
s
.Rn /k  c k k jbpq
s
.Q/k (5.95)
kD0

with k D f jkl g and the collection of cubes Q D fQj l g as in (5.86). Since q > " in
 
(5.79) and L  n 1"  1 in (5.73) one obtains by Theorem 1.7 and Proposition 5.3 that
j s
akl in (5.88) are also L1 -normalised atoms in Fpq .Rn /. The counterpart of (5.95) for
s n
the spaces Fpq .R / according to Theorem 1.7 and the L1 -normalised atoms in (5.89)
is given by
X r
kg jFpq .R /k  c
s n
k k jfpqs
.Q/k (5.96)
kD0

with
 X 1=q 
 j 
k jf
k pq
s
.Q/k D  2jsq j kl
j l .  /jq jLp .Rn /; (5.97)
j;l

where j l is the characteristic function of Qj l in (5.86). But this is the point where the
specific nature of Qj l comes in. It follows by [T06], Proposition 1.33, Example 1.35,
p. 19, that
k k jfpq
s
.Q/k  k k jfpp s
.Q/k  k k jbpp s
.Q/k: (5.98)
One may also consult [T06], Proposition 9.22, pp. 393–94, for this type of argument.
This proves (5.79) by reduction to (5.78) with q D p.
146 5 Spaces on smooth domains and manifolds

Remark 5.15. By [T83], Theorem 3.3.3, p. 200, and the references in [T83], Sec-
tions 2.7.2, 3.3.3, it is known since some 30 years that tr r with (5.59)–(5.63) is a
retraction. This means that there is a linear and bounded operator, called co-retraction,
with (5.80) (in the above notation). In particular [T83], Theorem 3.3.3, coincides with
s
[Tri78], Theorem 2.4.2, p. 105. Trace theorems for the classical Besov spaces Bpq ./
with p > 1 and Sobolev spaces Hp ./ D Fp;2 ./, 1 < p < 1, are known since the
s s

late 1950s and 1960s. This may be found in [BIN75], [Nik77], [T78]. The main aim
of the above theorem is not so much to give a new proof of this well-known assertion
but to construct explicitly extension operators in terms of wavelets. This will be used
in what follows.

We recalled at the beginning of Section 3.2.3 what is meant by a complemented

subspace of a given quasi-Banach space and a related projection. Let tr r and extr;u
 be
r;u
as in the above theorem. Then P ,

P r;u D extr;u
 B tr  W Apq ./ ,! Apq ./;
r s s
(5.99)
s
with Apq ./ D Bpq
s s
./ as in (5.78) or Apq ./ D Fpq s
./ as in (5.79). By (5.80) one
has
.P r;u /2 D extr;u r;u
 B tr  B ext  B tr  D P
r r r;u
: (5.100)
Hence P r;u is a projection of Apq
s
./ onto its range, denoted by P r;u Apq
s
./.

Corollary 5.16. Let r 2 N0 and u 2 N with r < u. Let  be a bounded C 1 domain

in Rn with n  2 according to Definition 3.4 (iii) or a bounded interval in R if n D 1.
Let Apq
s
./ be either Bpqs
./ with (5.78) or Fpq s
./ with (5.79). Then P r;u according
r;u
to (5.99) is a projection in Apqs
./ and ext is an isomorphic map onto the range of
P ,
r;u

Yr
‡
s 1 k
extr;u
 Bpq p . / D P r;u Bpq s
./;
kD0 (5.101)
r;u
Yr 1
s p k
ext Bpp . / D P Fpq ./
r;u s

kD0

.equivalent quasi-norms/.

Proof. By (5.100) it remains to prove that extr;u

 generates the indicated isomorphic
s
map. Let Apq ./ D Bpq s
./. We use temporarily the abbreviation

s 1 Y
r
s 1 k
Bpq p . /r D Bpq p . /; (5.102)
kD0

naturally quasi-normed as in (5.58). Then one has by (5.78) and (5.80), (5.99),

s 1
k extr;u
 g jBpq ./k  c kg jBpq . / k
s p r
(5.103)
 5.1 Wavelet frames and wavelet-friendly extensions 147

and
extr;u
 g DP
r;u
extr;u
 g 2P
r;u s
Bpq ./: (5.104)
If f 2 P r;u Bpq
s
./, hence f D P r;u f , then

s 1
f D extr;u
 g with g D tr r f 2 Bpq p . /r (5.105)

and
s 1
kg jBpq p . /r k  c kf jBpq
s
./k D c k extr;u
 g jBpq ./k:
s
(5.106)
Then one obtains the isomorphic map (5.101) for the B-spaces. Similarly for the
F -spaces.

5.1.4 Decompositions

As indicated in Introduction 5.1.1 we wish to clip together the (orthonormal) u-wavelet

bases for the spaces AQpq
s
./ according to Theorem 3.13 with the u-wavelet frames for
related spaces on D @ as described in Theorem 5.9 with Theorem 5.14 as the
decisive vehicle. For this purpose we need the decompositions (5.12)–(5.15) which
we are going to discuss now in greater detail. In the next Section 5.1.5 we combine all
these ingredients to construct wavelet frames (and later on also wavelet bases) in the
s
corresponding spaces Apq ./. First we complement Definition 2.1.

Definition 5.17. Let  be an arbitrary domain in Rn .

ı
(i) Let p, q, s be as in (2.3) with p < 1 for the F -spaces. Then Apq s
./ is the
s
completion of D./ in Apq ./.
(ii) Let u 2 N0 . Then C u ./ is the collection of all complex-valued functions f
having classical derivatives up to order u inclusively in  such that any function D ˛ f
with j˛j  u can be extended continuously to  x and
X
kf jC u ./k D sup jD ˛ f .x/j < 1: (5.107)
x2
j˛ju

Furthermore C./ D C 0 ./ and

1
\
C 1 ./ D C u ./: (5.108)
uD0

ı
Remark 5.18. Whereas the notation Apq s
./ is in common use, the situation for spaces
u
of type C ./ is different. The above version coincides with [HaT08], Definition A.1.
Of course, C u ./ is a Banach space. If  is a bounded C 1 domain in Rn then C u ./
can be recovered from C u .Rn / by restriction in the same way as in Definition 2.1 for
s
the spaces Apq ./. We refer to [HaT08], Theorem 4.1, p. 89.
148 5 Spaces on smooth domains and manifolds

ı
We are mainly interested in the relations between AQpq
s s
./, Apq ./, and subspaces
s
of Apq ./ with zero traces. All that we need is available in the literature.

Proposition 5.19. Let  be a bounded Lipschitz domain in Rn , n  2, according to

Definition 3.4 (iii) or a bounded interval in R if n D 1. Let
  
1 1 1
0 < p < 1; 0 < q < 1; max  1; n 1 <s< (5.109)
p p p

with q  min.p; 1/ for the F -spaces. Then

ı
s
Apq ./ D Apq
s
./ D AQpq
s
./: (5.110)

Remark 5.20. This coincides with a corresponding assertion in [T06], Section 1.11.6,
p. 66, where we referred in turn to [Tri02], Proposition 3.1, p. 494. In [T06], p. 67, we
quoted
ı
s
Apq ./ D AQpqs
./; 0 < p < 1; 0 < q < 1; (5.111)
for bounded C 1 domains  and
 
1 1
s > p D n 1 ; s 62 N0 ; (5.112)
p C p
with a reference to [T01], p. 69–70, and the literature mentioned there. But this is a
somewhat tricky assertion. We need only (5.110) with 1  p < 1.

Let  be a bounded C 1 domain in Rn (a bounded interval in R if n D 1) and let

P be the same projection as in Corollary 5.16 and (5.99). Then Qr;u D id  P r;u
r;u
s
is the projection of Apq ./ onto the kernel of P r;u ,
˚
Qr;u Apq
s
./ D f 2 Apq
s
./ W tr r f D 0 : (5.113)

If b 2 R then Œb  2 Z denotes the largest integer strictly less than b. If p  1


in (5.78), (5.79) then r D s  p1 is the largest admitted r 2 N0 . As before, the
1
Œs p 
corresponding trace operator tr  is given by (5.59),
1
Œs p  ˚ @j f 1 
tr  W f 7! tr  @ j
W 0j  s p
: (5.114)
1
Œs p 1 ;u
Let ext be the same as in (5.78), (5.79) with (5.101).

Theorem 5.21. Let  be a bounded C 1 domain in Rn with n  2 according to

Definition 3.4 (iii) or a bounded interval in R if n D 1. Let D @. Let
(
1 0 < q < 1; B-spaces;
1  p < 1; 1 < s  62 N0 ; (5.115)
p 1  q < 1; F -spaces;
 5.1 Wavelet frames and wavelet-friendly extensions 149

and u 2 N with s < u. Then

ı ˚ Œs 1 
ƒ
Bzpq
s
./ D Bpqs
./ D f 2 Bpq
s
./ W tr  p f D0 ;
(5.116)
ı ˚ Œs 1 
Fzpq
s
./ D Fpq
s
./ D f 2 Fpq
s
./ W tr  p f D0

interpreted as (5.110) if 1 C 1
p
<s< 1
p

Œs 1
. Furthermore,

 ;u Y
1
Œs p 
s 1 k
«
s
Bpq ./ D Bzpq
s
./  ext  p Bpq p . /;
kD0
(5.117)
1
Œs p 
Œs 1 
;u Y s 1 k
s
Fpq ./ D Fzpq
s
./  ext  p Bpp p . /:
kD0


Proof. By Proposition 5.19 we may assume s > p1 . Let r D s  p1 D s 1
p
.
s
Then it follows from (5.101) and (5.113) that Bpq ./ can be decomposed as

˚ Y
r
s 1 k
s
Bpq ./ D f 2 Bpq
s
./ W tr r f D 0  extr;u
 Bpq p . / (5.118)
kD0

with
kf jBpq
s
./k  kP r;u f jBpq
s
./k C kQr;u f jBpq
s
./k: (5.119)
By [T83], Theorem and Corollary 3.4.3, p. 210, and the related proofs one has
˚ ı
f 2 Bpq
s
./ W tr r f D 0 D Bpq s
./ D Bzpq
s
./ (5.120)

and an F -counterpart. This proves (5.116) and also the B-part of (5.117). In case of
the F -spaces one has to use the F -part of (5.101).
Remark 5.22. Quite obviously, (5.117) is the precise version of (5.12)–(5.15). Some-
what in contrast to Theorem 5.14 and Corollary 5.16 we restricted p in Theorem 5.21
to p  1. By the above arguments and [T83], Theorem and Corollary 3.4.3, p. 210,
s
one can extend Theorem 5.21 to Apq ./ with
  (
1 1 0 < q < 1; B-spaces;
0 < p < 1; n 1 <sr < ; (5.121)
p p p  q < 1; F -spaces;

where r 2 N0 . In contrast to the above theorem where for p  1 only the values
s  p1 2 N0 are not covered, one must exclude for p < 1 large s-intervals. This is
not very satisfactory, but unavoidable. In Section 6.4 we discuss which curious effects
s
can happen for traces of spaces Apq .Rn / if p < 1. On the other hand, one can replace
1  q < 1 in (5.115) for the F -spaces by 0 < q < 1. We refer to Corollary 6.24
150 5 Spaces on smooth domains and manifolds

and Remark 6.25. We add some comments about the spaces Apq s
./ with s  p1 2 N0
making clear that assertions of type (5.110), (5.111) and (5.116), (5.117) cannot be
expected. Let again  be a bounded C 1 in Rn and let 0 < p < 1, 0 < q < 1. Then
it follows from [Tri07d], also subject of Section 6.4 below, that
ı

1=p
Bpq ./ D 1=p
Bpq ./ if, and only if, min.1; p/ < q < 1;
ı (5.122)
1=p
Fpq ./ D 1=p
Fpq ./ if, and only if, p > 1, 0 < q < 1:

On the other hand, according to [T01], Sections 5.7, 5.9, pp. 51, 53, and the references
given there one has for
 
1
0 < p < 1; 0 < q < 1; s > p D n 1 (5.123)
p C

the Hardy inequalities

Z
jf .x/jp
sp
dx  c kf jFpq
s
.Rn /kp ; f 2 Fzpq
s
./; (5.124)
 d .x/

where again d.x/ D dist.x; / is the distance of x 2  to , and

Z 1 Z q=p
dt
t sq jf .x/jp dx  c kf jBpq
s
.Rn /kq ; f 2 Bzpq
s
./; (5.125)
0 t t

with t D fx 2  W d.x/ < tg. It follows that the characteristic function of  does
not belong to any of the spaces AQpq
1=p
./ with (5.123), hence

AQpq
1=p
./ 6D Apq
1=p
./; 0 < p < 1; 0 < q < 1; 1
p
> p : (5.126)

This shows that for s D 1=p nothing like (5.110), (5.116) and (5.117) with empty
second factors on the right-hand side, can be expected in cases covered by (5.122),
(5.126). Let
1
0 < p < 1; 0 < q < 1; sDrC > p and r 2 N: (5.127)
p
Then one has
ı

s
Bpq ./ 6D Bzpq
s
./ if min.1; p/ < q < 1;
ı (5.128)
s
Fpq ./ 6D Fzpq
s
./ if p > 1, 0 < q < 1:
This follows from
ırC 1 rC 1
g 2 Apq p ./ and g 62 AQpq p ./ (5.129)
1
with g 2 C ./ according to (5.108) and g.x/  d .x/ near . One obtains the
r

second assertion in (5.129) from (5.124), (5.125) with s D r C p1 . The first assertion
is covered by the arguments in [Tri07d]. This makes clear that assertions of type
 5.1 Wavelet frames and wavelet-friendly extensions 151

(5.116) cannot be expected for the exceptional cases s D r C 1

p
with r 2 N0 . As a
ı
s
consequence we prove in Section 6.2.2 that the spaces Apq ./
in (5.127), (5.128) do
not have an (interior) wavelet frame or basis in contrast to the corresponding spaces
AQpq
s
./ according to Theorem 3.13. Assertions of type (5.126), (5.122) and (5.128) for
s
related classical Besov spaces Bpq with 1 < p; q < 1 and Sobolev spaces Hps D Fp;2 s

with 1 < p < 1 are known since the 1960s and 1970s. This may be found in [T78],
Section 4.3.2, pp. 317–20, and the references given there.

5.1.5 Wavelet frames in domains

As far as wavelet frames are concerned we complete now the programme outlined
in Introduction 5.1.1 and recalled at˚ the beginning of Section 5.1.4. We need the
counterparts of the u-wavelet system ˆjr in Definition 2.4 and of the sequence spaces
in Definition 2.6.
Definition 5.23. Let  be a bounded domain in Rn and let
˚
Z D xlj 2  W j 2 N0 I l D 1; : : : ; Nj ; (5.130)

typically with Nj  2j n , such that for some c1 > 0,

jxlj  xlj0 j  c1 2j ; j 2 N0 ; l 6D l 0 : (5.131)

Let j l be the characteristic function of the ball B.xlj ; c2 2j /  Rn .centred at xlj
and of radius c2 2j / for some c2 > 0. Let s 2 R, 0 < p  1, 0 < q  1. Then
s
bpq .Z / is the collection of all sequences
˚
 D jl 2 C W j 2 N0 I l D 1; : : : ; Nj (5.132)

such that
X
1 X
Nj q=p 1=q
n
k jbpq
s
.Z /k D j.s p
2 /q
jjl jp < 1; (5.133)
j D0 lD1

s
and fpq .Z / is the collection of all sequences (5.132) such that

 X Nj
1 X 1=q 
 
k jfpq
s
.Z /k D  2jsq jjl j l .  /jq jLp ./ < 1 (5.134)
j D0 lD1

.obviously modified if p D 1 and/or q D 1/.

Remark 5.24. Roughly speaking, Z complements Z according to (2.24)–(2.26)
by points near the boundary. This is the almost obvious modification of Definition 2.6.
152 5 Spaces on smooth domains and manifolds

In what follows we are interested in bounded C 1 domains and bounded Lipschitz

domains. If p < 1 then one can replace j l in (5.134) by the characteristic function
of B.xlj ; c20 2j / \  with c20 > 0 (equivalent quasi-norms). This is well known and
covered by [T06], Section 1.5.3, pp. 18–19, and the references given there.
s
Recall that in this Chapter 5 we are mainly interested in spaces Apq ./ accord-
1
ing to (5.3) where  is a bounded C domain. Then related atoms as described in
Definition 1.5, Theorem 1.7 need not to have moment conditions. But it remains to be
a desirable property, also in connection with numerical applications in terms of poly-
nomial approximations. The interior wavelets in (2.33) have this property, for basic
wavelets (2.32) it is not needed, and for the boundary wavelets (2.34) not necessarily
valid. This can be preserved in what follows. But this has to be complemented now by
wavelets of type (5.76). The product structure of these building blocks combined with
a mild modification results in wavelets with some qualitative cancellations of the same
type as in (5.32). Then we speak about oscillation instead of cancellation. This may
x in modification of Definitions 2.4, 5.5. The spaces
justify to introduce wavelets in 
C u ./ have the same meaning as in Definition 5.17.

Definition 5.25. Let  be a bounded domain in Rn with D @ and let Z be as in

(5.130), (5.131). Let u 2 N.
(i) Then
˚
ˆ D ˆjl W j 2 N0 I l D 1; : : : ; Nj  C u ./ (5.135)
x if for some c3 > 0 and c4 > 0,
is called a u-wavelet system in 

x
supp ˆjl  B.xlj ; c3 2j / \ ; j 2 N0 I l D 1; : : : ; Nj ; (5.136)

and
ˇ ˛ j ˇ
ˇD ˆ .x/ˇ  c4 2j n2 Cj j˛j ; j 2 N0 I l D 1; : : : ; Nj ; x 2 ; (5.137)
l

for ˛ 2 Nn0 with 0  j˛j  u.

(ii) The above u-wavelet system is called oscillating if there are positive numbers
c5 , c6 , c7 with c6 < c7 such that

dist.B.xl0 ; c3 /; /  c6 ; l D 1; : : : ; N0 ; (5.138)

and
ˇZ ˇ
ˇ ˇ
ˇ .x/ ˆjl .x/ dx ˇˇ
n
 c5 2j 2 j u k jC u ./k; 2 C u ./; (5.139)
ˇ


for all ˆjl with j 2 N and

dist.B.xlj ; c3 2j /; / 62 .c6 2j ; c7 2j /: (5.140)

 5.1 Wavelet frames and wavelet-friendly extensions 153

Remark 5.26. In other words, the oscillation condition (5.139) is not required for the
terms with j D 0, for which we assume (5.138), and for wavelets ˆjl with j 2 N and,
roughly speaking,
dist.xlj ; /  dist.supp ˆjl ; /  2j : (5.141)

These are the basic wavelets ˆ0l in (2.32) and the boundary wavelets ˆjl in (2.34) with
(2.26), (2.35). On the other hand, one requires (5.139) in particular for all wavelets ˆjl
with supports having non-empty intersection with .

As before we use Apqs

./ with A 2 fB; F g and correspondingly apq s
.Z / with
a 2 fb; f g according to Definitions 2.1 and 5.23. Equivalences  must be understood
as in (3.4), (3.5).

Theorem 5.27. Let  be a bounded C 1 domain in Rn with n  2 according to

Definition 3.4 (iii) or a bounded interval in R if n D 1. Let r 2 N0 and u 2 N
with r < u. Then there is a common oscillating u-wavelet system ˆr according to
Definition 5.25 (ii) for all p, q, s with
(
1 1 0 < q < 1; B-spaces;
1  p < 1; 1<sr < ; (5.142)
p p 1  q < 1; F -spaces;

and A 2 fB; F g such that an element f 2 D 0 ./ belongs to Apq

s
./ if, and only if, it
can be represented as

1 X
X Nj
f D jl 2j n=2 ˆjl ;  2 apq
s
.Z /; (5.143)
j D0 lD1

unconditional convergence being in Apq

s
./. Furthermore,

kf jApq
s
./k  inf k japq
s
.Z /k; (5.144)

where the infimum is taken over all admissible representations (5.143) .equivalent
quasi-norms/. Any f 2 Apq
s
./ can be represented as

1 X
X Nj
f D jl .f / 2j n=2 ˆjl (5.145)
j D0 lD1

where jl .  / 2 Apq

s
./0 are linear and continuous functionals on Apq
s
./ and

kf jApq
s
./k  k.f / japq
s
.Z /k (5.146)

.u-wavelet frame/.
154 5 Spaces on smooth domains and manifolds

Proof. Step 1. Let r D 0 and u 2 N. Then p > 1 if s  0. The orthonormal u-wavelet

basis as used in Theorem 3.23 is oscillating. Then one obtains the desired assertion
from this theorem and Proposition 5.19.
Step 2. Let r 2 N and u 2 N with u > r. First we construct a suitable oscillating
u-wavelet system in x consisting of the interior orthonormal u-wavelet basis in L2 ./
as used in Theorem 3.13, now denoted by
˚
ˆ D ˆj;l
W j 2 N0 I l D 1; : : : ; Nj ; Nj 2 N; (5.147)

(where Nj  2j n ) and a boundary u-wavelet system

˚
ˆ;r D ˆj;k
l
W j 2 NI k D 0; : : : ; r  1I l D 1; : : : ; Nj (5.148)

with D @ (where Nj  2j.n1/ ) specified below. By Definition 2.4 and Re-
mark 5.26 the system ˆ is oscillating and fits in our scheme. The construction of
ˆ;r relies on the decomposition (5.117) with

1
s D r  1; r 2 N; (5.149)
p
and (5.77), (5.75) now with r  1 in place of r. Furthermore we adapt (5.74)–(5.76)
notationally to Definition 5.25 (ii) by shifting j to j C 1, hence j 2 N in place of
j 2 N0 . We give an explicit formulation. Indicating now D @ we let again
˚ ˚
ˆ D ˆj; l
and ‰  D ‰lj; (5.150)

be two real u-wavelet systems according to Theorem 5.9 on the compact .n1/-dimen-
sional manifold , where now j 2 N. Then
Z
 j 1; 
j
l; .h/ D 2j n1
2 h; ‰l 
D2j n1
2 h. 0 / ‰lj 1; . 0 / .d 0 / (5.151)


is the adapted counterpart of (5.74), j 2 N. Similarly we replace (5.75), (5.76) by

g D Extr1;u
 fg0 ; : : : ; gr1 g
1 X N
X
r1 X j
1 j (5.152)
D  .gk / 2j k 2j n=2 ˆj;k . /
kŠ l; l
kD0 j D1 lD1

with
ˆj;k
l
. / D 2j k k
n .2j n/ 2
j=2
ˆjl 1; . 0 /: (5.153)
We choose now  in (5.72) such that
Z 1
.t / t l dt D 0; l D 0; : : : ; u  1: (5.154)
0
 5.1 Wavelet frames and wavelet-friendly extensions 155

Recall that D . 0 ; n / are C 1 curvilinear coordinates in a tubular neighbourhood

of . Then it follows that one has (5.136), (5.137) and (5.139) with ˆj;k
l
in place of
j ;r
ˆl . With ˆ as in (5.148) based on (5.153) one obtains that
˚
ˆr D ˆjl W j 2 N0 I l D 1; : : : ; Nj D ˆ [ ˆ;r ; Nj  2j n ; (5.155)

is an oscillating u-wavelet system according to Definition 5.25 (ii). According to Theo-

rem 5.14 and its proof, (5.152) is an extension operator

Y
r1
s 1 k
Extr1;u
 W Bpq p s
. / ,! Bpq .Rn / (5.156)
kD0

and
X
r1
 1
s p k 
kg jBpq
s
.Rn /k  c gk jBpq . /
kD0
 (5.157)
r1  X
X 1 Nj
X q=p 1=q
n
k/q
 c0 2j.s p
jjl; .gk /jp :
kD0 j D0 lD1

Step 3. We prove that the oscillating u-wavelet system ˆr in (5.155) has the desired
properties. We deal with the B-spaces. The necessary modifications for the F -spaces
are the same as in Step 2 of the proof of Theorem 5.14 where now the additional
assumption (5.73) with " D 1 and L D 0 is empty. Let f 2 D 0 ./ be given by
(5.143) with a D b. Then 2j n=2 ˆjl with either ˆjl D ˆj; l
or ˆjl D ˆj;k
l
are L1 -
normalised atoms in Rn (no moment conditions are needed). We have by Theorem 1.7
that f 2 Bpqs
.Rn / and

k re f jBpq
s
./k  kf jBpq
s
.Rn /k  c k jbpq
s
.Z /k: (5.158)

It remains to prove that any f 2 Bpq s

./ can be represented by (5.145) with ˆjl
according to (5.155) with (5.147) and (5.148), (5.153). We decompose f 2 Bpq
s
./
by
f D P r;u f C Qr;u f (5.159)
with (5.119). By (5.113) and (5.120) one can expand Qr;u f according to Theorem 3.13
by the system ˆ in (5.147),
1 X
X Nj
Q r;u
f D j;
l
.Qr;u f / 2j n=2 ˆj;
l
(5.160)
j D0 lD1

with Z
j;
l
.Qr;u f / D 2j n=2 .Qr;u f /.x/ ˆj;
l
.x/ dx (5.161)

156 5 Spaces on smooth domains and manifolds

and

k .Qr;u f / jbpq

s
.Z /k  kQr;u f jBpq
s
./k  c kf jBpq
s
./k: (5.162)

In modification of (5.99) one obtains by (5.152) and (5.151) that

1 X N  
X
r1 X j
@k f
P r;u
f D j;
l
tr  2j n=2 ˆj;k
l
. / (5.163)
@ k
kD0 j D1 lD1

with    
@k f 1 j @k f
j;
tr 
l
D  tr  2j k : (5.164)
@ k kŠ l; @ k
Recall that is considered as an .n  1/-dimensional compact C 1 manifold. We may
identify the boundary part of Z with Z in (5.34) with n  1 in place of n. Then one
obtains by (5.164),

1 Nj ˇ
X  ˇ 
X ˇ j; @k f ˇˇp q=p
2
n
j.s p /q ˇ tr
ˇ l  k ˇ
j D0
@
lD1

1 Nj ˇ
X  ˇ 
X ˇ j @k f ˇˇp q=p
 2
1
j.sk p  n1
p /q ˇ tr
ˇ l;  k ˇ
(5.165)
j D0
@
lD1
 q
 @k f sk p1 

  tr  jBpq . /
k @
 c kf jBpq
s
./kq ;

where we used Theorem 5.9 and (5.59), (5.60). Now one obtains by (5.159), (5.160),
(5.163) and (5.162), (5.165) the representation (5.145) with

k.f / jbpq
s
.Z /k  c kf jBpq
s
./k: (5.166)

This proves also (5.146). By construction the coefficients jl .f / are linear continuous
s
functionals on Bpq ./.

Remark 5.28. We explained in Remark˚ 5.11 what is meant by a frame. This justifies
to call also the u-wavelet system ˆjl in the above theorem a frame characterised by
the stability (5.144) and the optimality (5.146). The question arises whether this frame
is even a basis. This can be reduced to a corresponding question for the frames on
according to Theorem 5.9. As mentioned ˚ in Remark 5.12 we deal with
˚ these problems
in Sections 5.2, 5.3. The dual system ‰lj of the wavelet frame ˆjl for compact
C 1 manifolds in Theorem 5.9 is again a˚wavelet system of the same type. Nothing
like this can be expected for the system ˆjl in the above theorem. In case of the
 5.1 Wavelet frames and wavelet-friendly extensions 157

boundary functionals jl .f / originating from (5.164) it follows from (5.165) with Rn
in place of  and (5.151) that

jl .f / D .f; ‰lj / with ‰lj 2 Bps n j

0 q 0 .R /; supp ‰ 
l
; (5.167)

in the framework of the dual pairing .S.Rn /; S 0 .Rn //. In particular, ‰lj is a singular
distribution.
Remark 5.29. In Theorems 3.13 and 3.23 we described common wavelet bases for
large regions fs; p; qg and A 2 fB; F g. By Theorem 5.9 we have a similar assertion
for spaces Apqs
. / on compact C 1 manifolds now in terms of wavelet frames. The
wavelet frames in the above Theorem 5.27 depend on r 2 N0 and they are restricted
s
to the .s; p/-strips in (5.142). The problems for spaces Apq ./ with s  p1 2 N0 or
p < 1 are the same as in Remark 5.22.
The above theorem can be extended to some subspaces
˚ @j f
s
Apq ./rN D f 2 Apq
s
./ W tr  @ j
D 0I j D r1 ; : : : ; rk (5.168)
s
of Apq ./ with

rN D fr1 ; : : : ; rk g  f0; : : : ; r  1g; r 2 N;

being integers between 0 and r  1, where p, q, s and A 2 fB; F g as in Theorem 5.27.

Then Apqs
./rN is a complemented subspace of Apq s
./. Let tN D ft1 ; : : : ; tm g be the
complementary integers between 0 and r  1, hence

f0; : : : ; r  1g D fr1 ; : : : ; rk g [ ft1 ; : : : ; tm g; k 2 N; m 2 N; k C m D r; (5.169)

excluding AQpq
s s
./ and Apq ./ in

AQpq
s
./  Apq
s
./rN  Apq
s
./: (5.170)

Corollary 5.30. Let  be a bounded C 1 domain in Rn with n  2 according to

Definition 3.4 (iii) or a bounded interval in R if n D 1. Let r 2 N and u 2 N
with r < u. Let p, q, s be as in (5.142) and A 2 fB; F g. Let rN D fr1 ; : : : ; rk g
be as in (5.169). Then there is a common oscillating u-wavelet system according to
Definition 5.25 (ii) depending only on rN and u such that an element f 2 Apqs
./
rN
belongs to Apq ./ if, and only if, it can be represented as
s

1 X
X Nj
f D jl 2j n=2 ˆjl ;  2 apq
s
.Z /; (5.171)
j D0 lD1

unconditional convergence being in Apq

s
./. Furthermore,

kf jApq
s
./k  inf k japq
s
.Z /k (5.172)
158 5 Spaces on smooth domains and manifolds

where the infimum is taken over all admissible representations (5.171) .equivalent
quasi-norms/. Any f 2 Apq
s
./rN can be represented as

1 X
X Nj
f D jl .f / 2j n=2 ˆjl ; (5.173)
j D0 lD1

where jl .  / 2 Apq

s
./0 are linear and continuous functionals on Apq
s
./ and

kf jApq
s
./k  k.f / japq
s
.Z /k (5.174)

.u-wavelet frame/.
Proof. We indicate the necessary modifications of the proof of Theorem 5.27 restricting
us to the B-spaces. Let ft1 ; : : : ; tm g be as in (5.169). Then the extension operator in
(5.152) must be replaced now by

1 X N
N
Xm X j
1 j
Ext r;u
 fg t1 ; : : : ; g tm g D l; .g tv / 2jtv 2j n=2 ˆj;t
l
v
. / (5.175)
t v Š
vD1 j D1 lD1

with (5.153). With an appropriate modification of ˆ;r in (5.148) one can now follow
the arguments from the proof of Theorem 5.27.
Remark 5.31. Spaces of type (5.168) play a role in boundary value problems for elliptic
operators. It remains to be seen whether one can take advantage of the existence of
oscillating u-wavelet frames in this context.

5.2 Wavelet bases: criterion and lower dimensions

5.2.1 Wavelet bases on manifolds

One may ask under which circumstances the u-wavelet frame ˆr in Theorem 5.27 is not
only a frame but a unconditional basis as describe in (1.100), (1.101) in the respective
spaces. Recall that ˆr is given by (5.155) consisting of ˆ in (5.147) and ˆ;r in
(5.148) (depending on r and u). Whereas ˆ is the interior orthonormal u-wavelet
basis in L2 ./ as used in Theorem 3.13, the system ˆ;r is given by (5.153) based on
the u-wavelet system ˆ in (5.150) on D @. It turns out that the u-wavelet frame
in Theorem 5.27, given by (5.155), is a basis if, and only if, the underlying u-wavelet
frame ˆ according to (5.150) with a reference to Theorem 5.9 is a basis. This will
be justified in Sections 5.2.1, 5.2.2 and applied to concrete situations afterwards. But
first we return to abstract compact n-dimensional C 1 manifolds and to the u-wavelet
frame ˚
ˆ D ˆj; l
W j 2 N0 W l D 1; : : : ; Nj  C u . / (5.176)
 5.2 Wavelet bases: criterion and lower dimensions 159

according to Theorem 5.9 (notationally adapted to our later needs). If, in addition, ˆ
is a basis in L2 . /, then (5.41), (5.42) can be written as

1 X N
X j
 
f D f; ‰lj; 
ˆj;
l
; f 2 L2 . /; (5.177)
j D0 lD1

with  X ˇ  ˇ 1=2
kf jL2 . /k  ˇ f; ‰ j; ˇ2 : (5.178)
l 
j;l

Hence ˆ is a Riesz basis in L2 . /. Recall that we may assume that ˆj;

l
and ‰lj; are
L2 -normalised. A discussion about Riesz bases in (separable complex) Hilbert spaces
may be found in [Gro01], p. 90. In particular, ˆ is the isomorphic image of the
standard basis in `2 . Inserting f D ˆj;
l
in (5.177) one obtains from the uniqueness
the well-known property
(
 j; j 0 ;  1 if j D j 0 , l D l 0 ,
ˆl ; ‰l 0  D (5.179)
0 otherwise:

If ˆ in (5.176) is both a u-wavelet frame according to Theorem 5.9 and a basis in

s
L2 . / then one has the same situation as in Theorem 1.20. Let again Apq . / be either
s s s s s
Bpq . / or Fpq . / and correspondingly apq .Z / be either bpq .Z / or fpq .Z /.
Proposition 5.32. Let be a compact n-dimensional C 1 manifold and let Apq s
. /
˚ u 2 N be both
be the spaces as introduced in Definition 5.1. Let ˆ in (5.176) for


a u-wavelet frame according to Theorem 5.9 with the dual system ‰lj; as above,
and a basis in L2 . /. Let 0 < p  1 .with p < 1 for the F -spaces/, 0 < q  1,
s 2 R, and
u > max.s; p  s/ B-spacesI u > max.s; pq  s/ F -spaces: (5.180)
Then f 2 D 0 . / is an element of Apq
s
. / if, and only if, it can be represented as

1 X N
X j

f D jl 2j n=2 ˆj;

l
;  2 apq
s
.Z /; (5.181)
j D0 lD1

unconditional convergence being in Apqs

. / if p < 1, q < 1, and in Apqs"
. / for
any " > 0 if p D 1 and/or q D 1. The representation (5.181) is unique,
 
jl D jl .f / D 2j n=2 f; ‰lj;  (5.182)
and n   o
I W f 7! 2j n=2 f; ‰lj;  (5.183)
is an isomorphic map of s
Apq / onto apq
. s
.Z /. If, in addition, p < 1, q < 1, then
ˆ in (5.176) is an unconditional basis in Apq
 s
. /.
160 5 Spaces on smooth domains and manifolds

˚
Proof. The smoothness properties both for ˆ and its dual system ‰lj; , and also
the bi-orthogonality (5.179) are essentially the same as in Rn . Then one obtains the
proposition by the same arguments as in the proof of Theorem 1.20 and the references
to [T06] given there.
Remark 5.33. If ˆ in (5.176) is a basis in L2 . / then it is a Riesz basis with (5.177),
(5.178). But it need not to be an orthogonal basis in contrast to all corresponding
˚
wavelet bases in the preceding Chapters 1–3. But the dual system ‰lj; has the
same smoothness and localisation properties as ˆ . This ensures that ˆ is a basis
(isomorphic map) in the same spaces where it is a u-wavelet frame. Examples will be
discussed later on.

5.2.2 A criterion

For bounded C 1 domains in Rn we obtained in Theorem 5.27 common oscillating

u-wavelet frames ˆr for spaces Apq s
./ with (5.142). Here r 2 N (excluding now
r D 0) and u 2 N with r < u are given, where ˆr comes from (5.155) with ˆ as in
(5.147) and ˆ;r as in (5.148), (5.153). This shifts the question of whether ˆr is not
only a frame but even a basis from  to its boundary D @. For all admitted k with
k D 0; : : : ; r  1 we used in (5.148), (5.153) the same boundary system ˆ in (5.150).
But this is not necessary. One may choose for each k an individual admitted boundary
system ˆ , say ˆk . Then one has to replace ˆjl 1; . 0 / in (5.153) by ˆjl;k1; . 0 /
without any changes in the arguments based on (5.148), (5.155). Furthermore one does
not need that the system ˆ or, now, the systems ˆk , are oscillating on according
to Definition 5.5 (ii). These modifications are immaterial for the above considerations,
but they will be of some use later on when we apply the following criterion.
Proposition 5.34. Let  be a bounded C 1 domain in Rn with n  2 as in Defini-
tion 3.4 (iii). Let r 2 N and u 2 N with r < u. Then the oscillating u-wavelet frame
according to Theorem 5.27 and (5.155) is a basis in Apq s
./ with (5.142) if, and only
if, the underlying systems
˚
ˆk D ˆj;
l;k
W j 2 N0 I l D 1; : : : ; Nj  C u . / (5.184)

on D @ with k D 0; : : : ; r  1 according to Definition 5.5 (i) and the above

explanations are bases in L2 . /.
Proof. Step 1. Let ˆk in (5.184) be bases in L2 . /. By Proposition 5.32 they are also
s 1 k
bases in all spaces Bpq p . / of interest, k D 0; : : : ; r  1. In order to prove that ˆr
is also a basis it suffices to show that the representation (5.143) is unique. Let
1 X
X Nj
0D jl 2j n=2 ˆjl ;  2 apq
s
.Z /: (5.185)
j D0 lD1
 5.2 Wavelet bases: criterion and lower dimensions 161

We use the projections Qr;u and P r;u D id  Qr;u as in Step 3 of the proof of
Theorem 5.27. We apply Qr;u to (5.185). By (5.159) and the equality Qr;u P r;u D
Qr;u .id  Qr;u / D 0 one obtains as in (5.160)

1 X N
X j

0D j;
l
2j n=2 ˆj;
l
: (5.186)
j D0 lD1

k
But ˆ is a basis. Hence j;l
D 0. If one applies tr  @@ k with k D 0; : : : ; r  1
to (5.185) then one has the k-term in (5.163) and by (5.153) with n D 0 (in obvious
notation) that
1 X N
X j

ˆjl;k1; . 0 /:
n1
0D j;
l;k
2j 2 (5.187)
j D1 lD1

By assumption, ˆk is a basis in L2 . /. Hence j;

l;k
D 0, and finally jl D 0 in (5.185).
Consequently ˆr is a basis.
Step 2. Let ˆr be a basis. We wish to prove that ˆk in (5.184) are bases in L2 . /.
First we prove that ˆk with k D 0 is a basis in, say, the trace space

"C 1
H " . / D B2;2
"
. / D tr  B2;2 2 ./; 0 < " < 1: (5.188)

Again it is a matter of uniqueness. Let

1 X N
X j

ˆjl;01; . 0 /;
n1
0D j;
l
2j 2  2 b2;2
"
.Z /: (5.189)
j D1 lD1

"C 1
Extending expansions in H " . / by (5.152), (5.153) with k D 0 to B2;2 2 ./ one
"C 1
obtains an admitted expansion in B2;2 2 ./. Since we assume that ˆr is a basis one
has j;
l
D 0. Hence ˆ0 is a basis in H " . /. As a consequence one has the bi-
orthogonality (5.179). But now one is in a similar position as in Proposition 5.32 with
H " . / in place of L2 . /. By the same arguments and references as there it follows
that ˆ0 is a basis in all admitted spaces and, in particular, in L2 . /. Similarly for
k D 1; : : : ; r  1.

5.2.3 Wavelet bases on intervals and planar domains

Domains in the plane R2 are called planar domains. We deal with wavelet bases on
bounded intervals on R D R1 and on bounded planar C 1 domains in R2 according to
s
Definition 3.4 (iii). Let Apq ./ with A 2 fB; F g and correspondingly apq
s
.Z / with
a 2 fb; f g be as introduced in Definitions 2.1 and 5.23.
162 5 Spaces on smooth domains and manifolds

Theorem 5.35. Let  be either a bounded interval on R .n D 1/ or a bounded planar

C 1 domain in R2 .n D 2/. Let r 2 N0 and u 2 N with r < u. Then there is a
common oscillating u-wavelet system ˆr according to Definition 5.25 (ii) for all p, q,
s with
(
1 1 0 < q < 1; B spaces;
1  p < 1; 1<sr < ; (5.190)
p p 1  q < 1; F spaces;

and A 2 fB; F g such that an element f 2 D 0 ./ belongs to Apq

s
./ if, and only if, it
can be represented as

1 X
X Nj
f D jl 2j n=2 ˆjl ;  2 apq
s
.Z /; (5.191)
j D0 lD1

unconditional convergence being in Apq

s
./. The representation (5.191) is unique,

1 X
X Nj
f D jl .f / 2j n=2 ˆjl ; (5.192)
j D0 lD1

where jl .  / 2 Apq

s
./0 are linear and continuous functionals on Apq
s
./ and

f 7! fjl .f /g is an isomorphic map of Apq

s s
./ onto apq .Z / (5.193)

.u-wavelet basis/.

Proof. Step 1. As mentioned in Step 1 of the proof of Theorem 5.27 there is nothing
to prove if r D 0 and u 2 N.
Step 2. Let  be a bounded interval, say the unit interval  D .0; 1/. Let r 2 N,
hence s > 1=p. In this case, (5.152), (5.153) reduces to

X
r1 k
x
gD Extr1;u
 fg0 ; : : : ; gr1 g D .x/ gk .0/ C .1/; (5.194)
kŠ
kD0

where .1/ indicates a corresponding term at x D 1. Now one can argue as in the proof
of Proposition 5.34.
Step 3. The boundary D @ of a bounded C 1 domain in the plane consists of
finitely many closed C 1 curves j . For each such curve there is a tubular neighbour-
hood (strip) ˚
j D x 2 R W dist.x; j / < "
" 2
for some " > 0 (5.195)
and a diffeomorphic map j onto a tubular neighbourhood of
˚
j . j / D y 2 R W jyj D 1 D T ;
2 1
(5.196)
 5.3 Wavelet bases: higher dimensions 163

interpreting the unit circle as the 1-torus T 1 . By Theorem 1.37 there are orthonormal
s
wavelet bases with the desired properties in all spaces Apq .T 1 /. Both these spaces
and also the bases can be transferred by the above diffeomorphic maps j to j . Then
one has a u-wavelet basis on . Application of Proposition 5.34 proves the theorem in
case of planar domains.
Remark 5.36. The above method works in higher dimensions only in exceptional
cases. For example, let  be a tube (torus of revolution) in R3 such that its boundary
D @ can be identified with the 2-torus T 2 . Then one has by Theorem 1.37 a wavelet
s
basis on . Application of Proposition 5.34 gives a wavelet basis for the spaces Apq ./
with (5.142).

5.3 Wavelet bases: higher dimensions

5.3.1 Introduction

In dimensions one and two we have the satisfactory Theorem 5.35. In higher dimen-
sions, n  3, the situation is less favourable. According to Remark 5.36 one has a full
counterpart of Theorem 5.35 for the torus of revolution M3 in R3 with @M3 D T 2
(the 2-torus) with the same exceptional values s  p1 2 N0 . In particular it remains
open whether the distinguished Sobolev spaces
H s .M3 / D B2;2
s
.M3 / with s  1
2
2 N0 (5.197)
have a u-wavelet basis. However it comes out that this is not an artefact produced by
the method but it lies in the nature of the problem. For example, since D.M3 / is dense
in H 1=2 .M3 / it is natural to ask whether there are u-wavelet bases in H 1=2 .M3 / with
compact supports in M3 . The answer is negative. We return to the problem of non-
existence of wavelet bases (and wavelet frames) in Section 6.2.2. Asking for wavelet
bases in balls
Bn D fx 2 Rn W jxj < 1g (5.198)
or on spheres
Sn D fx 2 RnC1 W jxj D 1g (5.199)
the situation is even worse. It comes out that we do not obtain u-wavelet bases by our
method in some exceptional spaces including
H s .Sn / and H s .Bn /; n  3; s 2 N and s  1
2
2 N0 : (5.200)
In particular it remains open whether the classical Sobolev spaces
H k .Sn / D W2k .Sn / and H k .Bn / D W2k .Bn /; n  3; k 2 N; (5.201)
have u-wavelet bases. There is a remarkable difference between the exceptional values
in (5.197) and in (5.200) with n D 3. But it is at least questionable whether this has
164 5 Spaces on smooth domains and manifolds

something to do with the different topology of M3 and S3 or B3 . It depends on our

method which is based on induction by dimension. As for other methods one may
consult the references given at the beginning of Section 5.3.3. We deal in Section 5.3.2
first with spheres and balls. But the method itself works apparently on a much larger
scale. It is based on some surgery cutting manifolds and domains in more handsome
pieces. The resulting wavelet bases are glued together afterwards. One obtains in this
way rather final results for the some classes of domains of the same type as for spheres
and balls. But this requires not only some topological considerations (known as the
domain problem) but also the study of traces and related extensions for some types of
non-smooth domains, the finite union of diffeomorphic distortions of cubes. This will
be done in Sections 5.3.3, 5.3.4 and in greater detail in Section 6.1.

5.3.2 Wavelet bases on spheres and balls

The sphere D Sn in (5.199) is a compact n-dimensional C 1 manifold as introduced

at the beginning of Section 5.1.2. Let Apq s
. / D Apq s
.Sn / be the spaces according
to Definition 5.1. With Z as in (5.25)–(5.27) the sequence spaces apq s
.Z / with
a 2 fb; f g have the same meaning as in Definition 5.7. As for the u-wavelet system
˚
ˆ D ˆj; l
we refer to Definition 5.5 (i) indicating now . In contrast to Theo-
rem 5.9 we do not care any longer whether the corresponding u-wavelet frames are
oscillating. The procedure on which we rely now is characterised by some surgery or
domain decomposition. In tubular neighbourhoods of the cutting surfaces one cannot
guarantee a qualitative cancellation of type (5.32), called oscillation. Although this
property remains valid outside these cutting surfaces this may justify to abandon this
property. Furthermore, the induction by dimension produces additional exceptional
.s; p/ values and it is no longer reasonable to ask for common u-wavelet bases in
s
.s; p/-regions. In other words, we ask for u-wavelet bases in a given space Apq . /
s
and, later on, in Apq ./.
Theorem 5.37. Let D Sn be the n-sphere (5.199) with n  2. Let
1 m
1  p < 1; s>  1; s 62 N0 for m D 1; : : : ; n  1; (5.202)
p p
and 0 < q < 1 for the B-spaces, 1  q < 1 for the F -spaces. Let u 2 N with
u > s. Then there is a u-wavelet system
˚
ˆ D ˆj; l
W j 2 N0 I l D 1; : : : ; Nj  C u . / (5.203)

according to Definition 5.5 (i) such that f 2 D 0 . / belongs to Apq

s
. / if, and only if,
it can be represented as

1 X N
X j

f D jl 2j n=2 ˆj;

l
;  2 apq
s
.Z /; (5.204)
j D0 lD1
 5.3 Wavelet bases: higher dimensions 165

unconditional convergence being in Apq

s
. /. The representation (5.204) is unique,

1 X N
X j

f D jl .f / 2j n=2 ˆj;

l
; (5.205)
j D0 lD1

where jl .  / 2 Apq

s
. /0 are linear and continuous functionals on Apq
s
. / and

f 7! fjl .f /g is an isomorphic map of Apq

s s
. / onto apq .Z / (5.206)

.u-wavelet basis/.
Proof. Step 1. We decompose S2 D @B3 as
˚
S2 D S2C [ S2 [ x 2 R3 W x12 C x22 D 1; x3 D 0 (5.207)

where ˚
S2C D x 2 R3 W jxj D 1; x3 > 0 (5.208)
and similarly S2 . We contract the northern hemisphere S2C
with the North Pole .0; 0; 1/
1
along the meridians such that the C dilation is 1 near the North Pole (no contraction)
and, say, 1=2 at the equator. Then one obtains a diffeomorphic map
˚
W S2C 7! x 2 R3 W jxj D 1; 0 < c < x3  1 D S2c (5.209)

which can be extended to a diffeomorphic map of a tubular neighbourhood of S2C onto

a tubular neighbourhood of S2c . But S2c can be represented as
˚
S2c D x 2 R3 W .x1 ; x2 / 2 Bd ; x3 D h.x1 ; x2 / (5.210)

with
˚
Bd D y 2 R2 W jyj < d for some d > 0 and h 2 C 1 .Bd /: (5.211)

Any wavelet basis for Axpqs

.Bd / according to Theorem 3.23 can be transferred to the
x
corresponding spaces Apq .S2c / and afterwards by the inverse 1 of in (5.209) to
s

Axpq
s
.S2C /. Similarly for Axpq
s
.S2 /. The third set on the right-hand side of (5.207)
is a circle, the equator. But now we are in the same situation as in Theorem 5.35
for planar domains with a reference to Proposition 5.34 which in turn relies on the
wavelet-friendly extension (5.152), (5.153). But this extension works on both sides
of the equator. All other technicalities are covered by Theorem 5.27. This proves the
above assertion for n D 2.
Step 2. Let n D 3. Then
˚
S3 D S3C [ S3 [ x 2 R4 W x12 C x22 C x32 D 1; x4 D 0 (5.212)

is the counterpart of (5.207), (5.208). As in Step 1 one has wavelet bases in Axpq
s
.S3C /
and Axpq
s
.S3 /. The last set in (5.212) can be identified with S2 . We may assume
166 5 Spaces on smooth domains and manifolds

s > 1=p (otherwise there is nothing to prove as mentioned in Step 1 of the proof of
Theorem 5.35). Then one obtains by Step 1 a wavelet basis in
s 1
Bpq p .S2 / if s  2
p
Ds 1
p
 1
p
62 N0 : (5.213)

s 1 k
In other words for the boundary spaces Bpq p .S2 / underlying Proposition 5.34 and
its proof one has the additional restriction (5.213). Hence one needs now not only
1 2
s 62 N0 ; but also s 62 N0 : (5.214)
p p
But otherwise one can argue as in Step 1 where we now benefit from the possibility
that the boundary systems ˆk may be different for different values of k. This proves
the theorem in case of n D 3. The rest is now a matter of induction.

However the arguments on which the extension from S2 to S3 in (5.212), (5.213)

relies can also be used for the step from S2 D @B3 to B3 . By iteration one has the
counterpart of Theorem 5.37 for Bn with n  3 in place of Sn . (In case of n D 2
one has Theorem 5.35.) Although everything is largely parallel to the above theorem
it seems to be reasonable to give an explicit formulation switching to a slightly more
general situation. The boundary D @ of a bounded C 1 domain in Rn consists of
finitely many connected components. We say that such a component is diffeomorphic
to Sn1 D @Bn if there is a diffeomorphic map of a tubular neighbourhood of this
component onto a corresponding tubular neighbourhood of Sn1 D @Bn . Let Apq s
./
with A 2 fB; F g and correspondingly apq .Z / with a 2 fb; f g be as in Definitions 2.1
s 

and 5.23. Equivalences  must be understood as in (3.4), (3.5).

Theorem 5.38. Let  be a bounded C 1 domain in Rn with n  3 according to Defi-

nition 3.4 (iii) such that each of its connected boundary components is diffeomorphic
to Sn1 as explained above. Let
1 m
1  p < 1; s>  1; s 62 N0 for m D 1; : : : ; n  1; (5.215)
p p
and 0 < q < 1 for the B-spaces, 1  q < 1 for the F -spaces. Let u 2 N with
u > s. Then there is an oscillating u-wavelet system
˚
ˆ D ˆj; l
W j 2 N0 W l D 1; : : : ; Nj  C u ./; Nj 2 N; (5.216)

x according to Definition 5.25 (ii) such that f 2 D 0 ./ belongs

.where Nj  2j n / in 
to Apq ./ if, and only if, it can be represented as
s

1 X N
X j

f D jl 2j n=2 ˆj;

l
;  2 apq
s
.Z /; (5.217)
j D0 lD1
 5.3 Wavelet bases: higher dimensions 167

unconditional convergence being in Apq

s
./. The representation (5.217) is unique,

1 X N
X j

f D jl .f / 2j n=2 ˆj;

l
; (5.218)
j D0 lD1

where jl .  / 2 Apq

s
./0 are linear and continuous functionals on Apq
s
./ and
˚
f 7! jl .f / is an isomorphic map of Apq s s
./ onto apq .Z / (5.219)
.u-wavelet basis/.
Proof. Let  D Bn with n  3. As mentioned in Step 1 of the proof of Theorem 5.27
there is nothing to prove if s < 1=p. Let s > 1=p. Then it follows from (5.202)
s 1 k
that one can apply Theorem 5.37 to the needed boundary spaces Bpq p .Sn1 / with
Sn1 D @Bn . Now one can use Proposition 5.34 (in its full generality) as in Step 2
of the proof of Theorem 5.37. Let @ be diffeomorphic to Sn1 as explaind above.
But diffeomorphic distortions of the boundary wavelets in Definition 5.25 (i) result in
wavelets of the same type. Then one obtains again the desired assertion. This can
be extended to boundaries consisting of several components which are diffeomorphic
to Sn1 .
Remark 5.39. We complained in Introduction 5.3.1 that some of the most interesting
function spaces on spheres and balls are not covered by the above Theorems 5.37, 5.38.
The situation is better for the distinguished wavelet frames according to Theorem 5.27.
But even there remain the exceptional cases with s  p1 2 N0 . One can avoid any
exceptional cases if one restricts the wavelet bases on Rn according to Theorem 1.20
to . We return later on in Section 5.4 to this point. But the resulting frames cannot
be converted into u-wavelet bases. There is no counterpart of Proposition 5.34.

5.3.3 Wavelet bases in cellular domains and manifolds

For bounded C 1 domains in Rn we have satisfactory frame representations for the

s
spaces Apq ./ covered by Theorem 5.27. Related assertions for bases are more diffi-
cult, especially in higher dimensions, which means n  3. In contrast to n D 2 where
one has Theorem 5.35, the related higher dimensional counterpart in Theorem 5.38
applies only to very special domains. What about bounded domains in Rn with n  3
where the connected boundary components are not diffeomorphic to Sn1 or T n1 ?
Basically one can try to decompose a domain into simpler standard domains, to deal
with bases in these standard domains, and to glue together the outcome. This is the
so-called domain problem which attracted a lot of attention. But especially the ques-
tion how to glue together bases in standard domains is a rather tricky business. The
most impressive contributions to these problems are still the papers by Z. Ciesielski
and T. Figiel, [CiF83a], [CiF83b], [Cie84] dealing with common spline bases in
168 5 Spaces on smooth domains and manifolds

C k .M /, where k 2 N,
the classical Sobolev spaces Wpk .M /, where k 2 N, 1  p < 1, and
s
the classical Besov spaces Bpq .M /, where 1  p < 1, 1  q < 1, s 2 R,
on compact C 1 manifolds M (with and without boundary). The final goal is reached
after more than 100 pages in [Cie84], Theorem C. There are several proposals in
literature to employ and to modify this approach in the context of wavelet bases. We
refer to [Dah97], Section 10, [Dah01], Section 9, [Coh03], p. 130, with a reference to
the original paper [DaS99], and [HaS04], [JoK07].
We have no final answers. But we wish to look at problems of this type in the
context of our approach. Recall what is meant by a polyhedron (polyhedral domain)
in Rn . In R D R1 open bounded intervals are called polyhedrons. In Rn with n  2 a
set ! is said to be a polyhedron if it is a bounded simply connected Lipschitz domain
such that its boundary @! consists of finitely many faces which are .n1/-dimensional
polyhedrons in .n1/-dimensional hyper-planes in Rn . We do not discuss geometrical
and topological aspects of this definition. Angles resulting from the intersection of
.n  1/-dimensional faces are usually assumed to be non-zero. But for our purpose
this is neither necessary nor suitable (one may think about domain decompositions as
for S3 in (5.212)). However we are more interested in the analytic aspect with the unit
cube
Q D fx 2 Rn W 0 < xr < 1I r D 1; : : : ; ng
as the proto-type of a polyhedron. In Definition 5.17 we introduced the spaces C 1 ./
in domains  in Rn . Recall that a one-to-one map from a bounded domain  in Rn
onto a bounded domain ! in Rn ,

W  3 x 7! y D .x/ 2 !;

is called a diffeomorphic map if

l 2 C 1 ./ and . 1
/l 2 C 1 .!/I l D 1; : : : ; n;
1 1
for the components l of and . /l of its inverse ,
1 1
B D id in  and B D id in !:

A bounded domain  in Rn is said to be diffeomorphic to a bounded domain ! in Rn

if there is a diffeomorphic map of a neighbourhood of  x onto a neighbourhood of !
x
ı
with ! D ./. If is a set in R then
n
is the largest open set in Rn with ı 
(the interior of ).
Definition 5.40. (i) A domain  in Rn , n  2, is said to be cellular if it is a bounded
Lipschitz domain according to Definition 3.4 (iii) which can be represented as
S
L ı
D l with l \ l 0 D ; if l 6D l 0 ; (5.220)
lD1
 5.3 Wavelet bases: higher dimensions 169

such that each l is diffeomorphic to a polyhedron.

(ii) A compact n-dimensional C 1 manifold as introduced at the beginning of
Section 5.1.2, furnished with an atlas fVm ; m gM
mD1 , is said to be cellular if there are
pairwise disjoint open subsets j of with

[
M
D m; and m  Vm ; (5.221)
mD1

such that each m . m / is a cellular domain, m. m/  Um  Rn if n  2 or a

bounded interval if n D 1.
Remark 5.41. Any bounded C 1 domain in Rn according to Definition 3.4 (iii) is
cellular. Based on the above references one can presumably show that

any compact C 1 manifold is cellular.

But we did not check this assertion. We are interested here only on D @, where 
is a bounded C 1 domain.
Proposition 5.42. The boundary D @ of a bounded C 1 domain in Rn with n  2
according to Definition 3.4 (iii) is an .n  1/-dimensional cellular compact manifold.
Proof. Let fVm ; m gM mD1 be an atlas of originating from the special maps according
to Definition 3.4 (ii,iii) with Um D m .Vm /  Rn1 . We choose a suitable cellular
S
domain W1 in U1 with W1  U1 and 1 D 11 .W1 / such that @ 1  M mD2 Vm .
Then application of 2 to V2 n 1 produces in U2 a boundary which fits in the scheme
(assuming that V2 and 1 have a non-empty intersection). Now one can construct a
cellular domain W2 in 2 .V2 n 1 / with a boundary @ 2 of 2 D 21 .W2 / covering
parts of @.V2 n 1 / such that the remaining parts of @.V2 n 1 / and of @ 2 are contained
S
in M mD3 Vm . Iteration gives the desired assertion.

Recall that the characteristic function C of RnC is a pointwise multiplier for

s
Apq .Rn /; s 2 R; 0 < p  1; 0 < q  1; (5.222)

(p < 1 for the F -spaces) if, and only if,

   
1 1 1
max n 1 ; 1 <s < : (5.223)
p p p
A proof may be found in [RuS96], Section 4.6.3, p. 208. We refer for the history of this
substantial assertion to [RuS96], p. 258. This assertion extends to cubes, polyhedrons,
their diffeomorphic images and hence also to cellular domains according to Defini-
tion 5.40 (i). Since cellular domains are bounded Lipschitz domains one obtains by
Corollary 3.25 that one can apply Theorem 3.23. We complement now this assertion.
s
Let Apq ./ with A 2 fB; F g be as before and apq s
.Z / with a 2 fb; f g as used in
170 5 Spaces on smooth domains and manifolds

Theorem 3.23 with a reference to Definition 2.6. We stick at orthonormal u-wavelet

bases in L2 ./ as the starting point, but we do not care any longer about optimal values
of u and, as a consequence, common bases in some .s; p/-regions. This can be done
following the proofs but the outcome is no longer very convincing.
Theorem 5.43. Let  be a cellular domain in Rn according to Definition 5.40 (i). Let
for u 2 N, ˚ j
ˆl W j 2 N0 I l D 1; : : : ; Nj (5.224)
be an orthonormal u-wavelet basis in L2 ./ according to Theorem 2.33 and Defi-
nitions 2.31, 2.4. Let Apq
s
./ be the spaces as introduced in Definition 2.1. Let
0 < p  1 .with p < 1 for the F -spaces/, 0 < q  1, and
 
1 n
 1 < s < min ; (5.225)
p n1
.which means s < 1=p if n D 1/. Let u > u.Apq s
/ where u.Bpq
s
/ D u.s; p/ and
u.Fpq / D u.s; p; q/ are sufficiently large natural numbers. Then f 2 D 0 ./ is an
s

element of Apq
s
./ if, and only if, it can be represented as

1 X
X Nj
f D jl 2j n=2 ˆjl ;  2 apq
s
.Z /; (5.226)
j D0 lD1

unconditional convergence being in D 0 ./ and in any space Apq 

./ with  < s.
Furthermore, if f 2 Apq ./ then the representation (5.226) is unique where jl .  / 2
s
s
Apq ./0 are linear and continuous functionals on Apq
s
./ and
˚
f 7! jl .f / is an isomorphic map of Apq
s s
./ onto apq .Z / (5.227)

.u-wavelet basis/.
Proof. We apply Theorem 3.23 to the upper lines in (3.45), (3.46) with p < 1 and
q  min.1; p/ in case of F -spaces, hence
 
AQpq
s
./; 0 < p < 1; 0 < q  1; s > p D n p1  1 C : (5.228)

Furthermore the characteristic function  of  is a pointwise multiplier in the corre-

sponding spaces Apqs
.Rn / if, in addition, s < 1=p. This requires p1 < n1
n
. One has
by Definition 2.1 that

AQpq
s
./ D Apq
s
./ if 0 < 1
p
< n
n1
; p < s < 1
p
: (5.229)

A cellular domain is in particular a bounded Lipschitz domain. Hence one can ap-
ply the extended complex interpolation method according to Proposition 4.15 and
Remark 4.16. One obtains
s
Fpq ./ D Fps00q0 ./; Fps11q1 ./ ; 0 <  < 1; (5.230)
 5.3 Wavelet bases: higher dimensions 171

with
1 1  1 1 
s D .1  /s0 C s1 ; D C ; D C : (5.231)
p p0 p1 q q0 q1
If 0 < p < 1, 0 < q  1 and s as in (5.225) then one finds spaces Fps00q0 ./ which
fit in (5.229), q0  min.1; p0 / and

Fps11q1 ./; 0 < p1 < 1; 0 < q1  1; s1 < 0; (5.232)

covered by the third line in (3.45) with (5.230), (5.231). We apply Theorem
˚ j 3.23 to
s0 s1
Fp0 q0 ./ and Fp1 q1 ./ with the same underlying u-wavelet basis ˆl in L2 ./.
On the sequence side one has
s
fpq .Z / D fps00q0 .Z /; fps11q1 .Z / : (5.233)

This follows from the comments and references in the proof of Proposition 3.21. Then
the interpolation property for the extended complex method ensures that the isomorphic
map (3.104), (3.105) between the corner spaces in (5.230), (5.233) generates also an
s s
isomorphic map between Fpq ./ and fpq .Z /. Similarly one can argue for the B-
spaces.
s
By Theorem 5.9 we have wavelet frames for all spaces Apq . / on compact n-
1
dimensional C manifolds. It seems to be a tricky business to convert these wavelet
frames into wavelet bases. For n-spheres Sn one has Theorem 5.37 which in turn
resulted in Theorem 5.38. Now we complement these assertions as follows. Let
s
apq .Z / be the sequence spaces according to Definition 5.7.
Corollary 5.44. Let be a compact n-dimensional cellular C 1 manifold according
to Definition 5.40 (ii) and let Apq
s
. / be the spaces as introduced in Definition 5.1. Let
   
1 1 1
0 < p < 1; 0 < q < 1; max n 1 ; 1 <s < : (5.234)
p p p
˚
Then there is a u-wavelet system ˆjl according to Definition 5.5 (i) such that f 2
D 0 . / is an element of Apq
s
. / if, and only if, it can be represented as

1 X
X Nj
f D jl 2j n=2 ˆjl ;  2 apq
s
.Z /; (5.235)
j D0 lD1

unconditional convergence being in Apq

s
. /. Furthermore, if f 2 Apq s
. / then the
j 0
representation (5.235) is unique where l .  / 2 Apq . / are linear and continuous
s

functionals on Apq
s
. / and
˚
f 7! jl .f / is an isomorphic map of Apq s s
. / onto apq .Z / (5.236)

.u-wavelet basis/.
172 5 Spaces on smooth domains and manifolds

Proof. Let be decomposed according to Definition 5.40 (ii) with (5.221). Then
m . m / is cellular in R . With p, q, s as in (5.234) the characteristic function 
n m
s n
of m . m / is a pointwise multiplier in Apq .R / with (5.222), (5.223). Now one can
apply Theorem 5.43 to
1
m .  / .m f / B m ./ in m . m /;

in the notation of Definition 5.1. Clipping together the M charts one obtains the desired
assertion.
˚
Remark 5.45. We proved a little bit more than stated. The u-wavelet basis ˆjl in
Theorem 5.43 comes from Theorem 3.23. Then one has by (3.104) that

jl .g/ D 2j n=2 .g; ˆjl /: (5.237)

This can be transferred to the above cellular compact C 1 manifold as

jl .f / D 2j n=2 .f; ‰lj / ; (5.238)

˚
where ‰lj is a suitable u-wavelet system according to Definition 5.5 (i).

5.3.4 Wavelet bases in C 1 domains and cellular domains

We return to the problems discussed at the beginning of Section 5.3.3 asking for wavelet
bases in bounded C 1 domains not covered by Theorem 5.38. The assertions obtained
there hint how to proceed.
Proposition 5.46. Theorem 5.38 can be extended to spaces Apq
s
./ with

1 2 1
1  p < 1; 1<s < ; s 6D ; (5.239)
p p p
q as there, in .arbitrary/ bounded C 1 domains in Rn , n  3.
Proof. As mentioned in Step 1 of the proof of Theorem 5.27 there is nothing to prove
if s < 1=p. Let 1=p < s < 2=p. As in the proof of Theorem 5.38 one needs now
s 1
a wavelet basis for the boundary space Bpq p . / with D @. It follows from
Proposition 5.42 that is an .n  1/-dimensional cellular manifold. Then one obtains
the desired wavelet basis from Corollary 5.44.
Theorem 5.47. Theorem 5.38 with
1 m
1  p < 1; s>  1; s 62 N0 for m D 1; : : : ; n, (5.240)
p p
instead of (5.215), remains valid for any bounded C 1 domain in Rn , n  3, according
to Definition 3.4 (iii).
 5.3 Wavelet bases: higher dimensions 173

Remark 5.48. Theorem 5.35 covers the dimensions one and two. In case of s < 2=p
and n  3 we have Proposition 5.46. Its proof is based on the observation that D @
is a cellular manifold and that it is sufficient to construct wavelet bases inside each cell
separately. Boundary values do not play any role. If s > 2=p then the situation
is different. We outline now how to proceed and return in detail to this problem in
Section 6.1 below. By Proposition 5.34 and its proof we must find wavelet bases in
s 1 k
Bpq p . / for k 2 N0 with s  1
p
k >0

on the cellular C 1 manifold D @. This reduces the problem to find wavelet bases
for
s
Bpq .Q/; 1  p < 1; 0 < q  1; s > 2=p; (5.241)
where Q is a cube in Rn (switching from n  1 to n), say,

Q D fx 2 Rn W 0 < xl < 1g (5.242)

as a proto-type of a polyhedron. The worst case (for our task) is s > n=p. Then
f 2 Bpqs
.Q/ has boundary values in the corner points of Q. Along the edges, say,
s n1
I1 D fx 2 Rn W 0 < x1 < 1; x 0 D 0g one can construct wavelet bases for Bpq p .I1 /
as in Step 2 of the proof of Theorem 5.35 if s  n1
p
 p1 62 N0 . This gives a wavelet
basis on all edges of, say, a square Q in (5.242) with n D 2. The wavelet-friendly
extension operator as used in Theorem 5.14 fits pretty well in this scheme. Hence there
s n2
is a good chance to obtain wavelet bases in Bpq p .Q/ for the square Q where one
has again to exclude spaces with s  pl 2 N0 where l D n; n  1. This requires a closer
look at what happens near the corner points. Now one can proceed in this way shifting
wavelet bases on 2-dimensional faces of a 3-dimensional cube to this cube. Then one
can prove the above assertions by iteration. It requires some care what happens in
vertices, edges, faces etc. This is a somewhat tricky business. Details are shifted to
Section 6.1. These arguments apply equally to cellular manifolds and cellular domains.
Hence Theorem 5.47 can be complemented as follows.
Theorem 5.49. Theorem 5.38 with (5.240) in place of (5.215) remains valid for any
cellular domain in Rn , n  2, according to Definition 5.40 (i).
Remark 5.50. As mentioned above, details are shifted to Section 6.1. Both Theo-
s
rems 5.35, 5.38 and Theorems 5.47, 5.49 apply to spaces Apq ./ satisfying at least

1 1
1  p < 1; s>  1; s 62 N0 ; (5.243)
p p
and 0 < q < 1 for the B-spaces, 1  q < 1 for the F -spaces. The considerations
in Section 6.2.2 show that the restriction s  p1 62 N0 is natural. By Theorem 5.35
there are no further restrictions if  is a bounded interval .n D 1/ or a bounded
planar C 1 domain .n D 2/. If the bounded planar domain is only cellular (for
174 5 Spaces on smooth domains and manifolds

example the square Q in (5.242), n D 2) then Theorem 5.49 and (5.240) also exclude
s  p2 2 N0 . In higher dimensions, n  3, Theorems 5.47, 5.49 exclude s  m p
2 N0
for m D 1; : : : ; n. Hence Theorem 5.38 is a little bit better and as described at the
beginning of Section 5.3.1 for the torus of revolution (or tube) M3 in R3 the situation
is much better. But we do not know whether there is really a difference between
smooth domains and cellular domains or whether even the topology of D @ plays
a role. But it is clear that the method induction by dimension as used in the proof of
Theorem 5.38 and also in connection with Theorems 5.47, 5.49 does not work in these
exceptional cases. Let, for example,
˚
 D Q D x D .x1 ; x2 / 2 R2 ; 0 < x1 < 1; 0 < x2 < 1 (5.244)

be the unit square in the plane R2 . Then H 1 .Q/ D W21 .Q/ D B2;2
1
.Q/ is an excep-
S4
tional space. Let D @ D lD1 Il , where Il are the four sides of Q,

I1 D f.t; 0/ W 0  t  1g; I2 D f.0; t / W 0  t  1g; (5.245)

and similarly I3 ; I4 . Then the trace space tr  H 1 .Q/ is the collection of tuples of the
form g D .g1 ; g2 ; g3 ; g4 / with

gl 2 H 1=2 .Il /; l D 1; : : : ; 4; (5.246)

such that in addition

Z 1=2
jg1 .t /  g2 .t /j2
dt C CC < 1 (5.247)
0 t
where C C C indicates similar integrals at the other 3 corner points. We refer to
[Gri85], Theorem 1.5.2.3, 1.5.2.4, pp. 43, 47, and [Gri92], Theorem 1.4.6, p. 17, where
s
one finds also corresponding assertions for some spaces Bpq in polygonal domains
in R . The coupling (5.247) of the boundary values at different edges does not fit in
2

our approach. This applies also to s  mp

2 N0 excluded in (5.240).

5.4 Wavelet frames, revisited

5.4.1 Wavelet frames in Lipschitz domains

For spaces Apqs

./ having boundary values at D @ we constructed in Theorem 5.27
oscillating wavelet frames adapted to the underlying bounded C 1 domains . This
paved the way to obtain afterwards at least in some cases wavelet bases. If one does
not care about having oscillations near the boundary one may ask what happens if one
restricts wavelet bases on Rn according to Theorem 1.20 to domains  hoping that one
s
obtains at least frames. As before we use Apq ./ with A 2 fB; F g and correspondingly
apq .Z / with a 2 fb; f g according to Definitions 2.1 and 5.23. Equivalences  must
s 
 5.4 Wavelet frames, revisited 175
˚
be understood as in (3.4), (3.5). As explained in Remark 5.11 a system ˆjl is called
a frame if it is stable (with respect to a natural sequence space) and optimal. Let p
and pq be the same numbers as in (1.32).
Theorem 5.51. Let  be a bounded Lipschitz domain in Rn with n  2 according
˚
to Definition 3.4 (iii). For any u 2 N there is a u-wavelet system ˆjl according to
Definition 5.25 (i) with the following properties.
(i) Let 0 < p  1, 0 < q  1, s 2 R, and

u > max.s; p  s/: (5.248)

Then f 2 D 0 ./ is an element of Bpq

s
./ if, and only if, it can be represented as

1 X
X Nj
f D jl 2j n=2 ˆjl ;  2 bpq
s
.Z /; (5.249)
j D0 lD1

unconditional convergence being in Bpq

s
./ if p < 1, q < 1, and in Bpq
s"
./ for
any " > 0 if p D 1 and/or q D 1. Furthermore,

kf jBpq
s
./k  inf k jbpq
s
.Z /k; (5.250)

where the infimum is taken over all admissible representations (5.249) .equivalent
quasi-norms/. Any f 2 Bpqs
./ can be represented as

1 X
X Nj
f D jl .f / 2j n=2 ˆjl (5.251)
j D0 lD1

where jl .  / 2 Bpq

s
./0 are linear and continuous functionals on Bpq
s
./ and

kf jBpq
s
./k  k.f / jbpq
s
.Z /k (5.252)

.u-wavelet frame/.
(ii) Let 0 < p < 1, 0 < q  1, s 2 R, and

u > max.s; pq  s/: (5.253)

Then f 2 D 0 ./ is an element of Fpq

s
./ if, and only if, it can be represented as

1 X
X Nj
f D jl 2j n=2 ˆjl ;  2 fpq
s
.Z /; (5.254)
j D0 lD1

unconditional convergence being in Fpq

s
./ if q < 1 and in Fpq
s"
./ for any " > 0
if q D 1. Furthermore,

kf jFpq
s
./k  inf k jfpq
s
.Z /k (5.255)
176 5 Spaces on smooth domains and manifolds

where the infimum is taken over all admissible representations (5.254). Any f 2
s
Fpq ./ can be represented as (5.251) where jl .  / 2 Fpq
s
./0 are linear and contin-
uous functionals on Fpq
s
./ and

kf jFpq
s
./k  k.f / jfpq
s
.Z /k (5.256)

.u-wavelet frame/.
˚ j
Proof. Let ‰G;m be the same orthonormal wavelet bases in L2 .Rn / as used in Theo-
j
rem 1.20 based on (1.87)–(1.92). If the support of ‰G;m has a non-empty intersection
with  then its restriction to  is denoted as ˆjl ,

ˆjl .x/ D ‰G;m

j
.x/; x l D l.G; m/;
x 2 ; (5.257)

suitably re-numbered. Lipschitz domains are regular domains as described in Re-

mark 4.29 with references to [ET96], Section 2.5, and [TrW96]. Then
n
2j.s p / 2j n=2 ˆjl D aj l (5.258)

are -.s; p/-atoms. If f 2 D 0 ./ is represented by (5.249) or (5.254) then f 2

s
Apq ./ and
kf jApq
s
./k  c k japq
s
.Z /k: (5.259)
We prove the converse. As mentioned in Corollary 4.12 (i) with a reference to [Ry99]
there is a universal linear and bounded extension operator

ext W Apq
s s
./ ,! Apq .Rn /: (5.260)

With f 2 Apq
s
./ one has by Theorem 1.20 that
X  
j n=2 j j
ext f D j;G
m 2 m D2
‰G;m ; j;G j n=2
ext f; ‰G;m : (5.261)
j;G;m

The restriction to  gives the desired result.

Remark 5.52. The proof relies on the existence of (interior and boundary) atoms and
extension operators. But this applies also to other cases. It will be the subject of the
next Section 5.4.2. We add a discussion and compare the above theorem with other
j
assertions obtained so far. If supp ‰G;m   in (5.257) then the resulting wavelets
ˆjl preserve all the nice properties of the Rn -wavelets. The situation is less favourable
for the wavelets with supports intersecting D @. They need not to be linearly
independent any longer. The multiresolution property is inherited by  from Rn and
there is a temptation to apply the standard orthonormalisation procedure to the boundary
wavelets. But this might be rather unstable. There is no control about the resulting
constants and also the desired localisation cannot be guaranteed. Furthermore, although
the proof of the extension operator used in (5.261) is constructive (via local means)
 5.4 Wavelet frames, revisited 177

it is rather involved. In other words, to calculate for given f 2 Apq

s
./ the optimal
j
coefficients l .f / seems to be rather tricky. It might be a better choice, also from
the point of numerical calculations, to rely on the constrained wavelet expansions in
Lipschitz domains according to Theorem 4.23. In Theorem 5.27 we described wavelet
frames for spaces Apq s
./ in bounded C 1 domains under additional restrictions for
the parameters p, q, s. The related boundary wavelets are oscillating in contrast to
the boundary wavelets in the above Theorem 5.51. This might be of some interest
in applications but it is not the main point of Theorem 5.27. One obtains by the
same arguments Corollary 5.30 what cannot be expected by possible modifications
of Theorem 5.51 and its proof. Furthermore, Theorem 5.27 paves the way to obtain
wavelet bases at least in some of these spaces.

5.4.2 Wavelet frames in ."; ı/-domains

Let  be a bounded ."; ı/-domain according to Definition 3.1 (i). Then it follows from
Proposition 3.6 (i) that  is a bounded I -thick domain with j@j D 0. Furthermore, 
is interior regular according to (4.95). But these observations ensure that the proof of
s s
Theorem 5.51 can be carried over to the following situation. Let Apq ./, apq .Z /, the
use of the equivalence , and the numbers p pq be as in the preceding Section 5.4.1.
Theorem 5.53. Let  be a bounded ."; ı/-domain in Rn with n˚  2 according to
Definition 3.1 (i). For any u 2 N there is a u-wavelet system ˆjl according to
Definition 5.25 (i) with the following properties.
(i) Let 0 < p  1, 0 < q  1, and u > s > p . Then f 2 D 0 ./ is an
element of Bpqs
./ if, and only if, it can be represented as (5.249). One has (5.250).
Any f 2 Bpq ./ can be represented by (5.251) with (5.252) .u-wavelet frame/.
s

(ii) Let 0 < p < 1, 0 < q  1, and u > s > pq . Then f 2 D 0 ./ is an
element of Fpqs
./ if, and only if, it can be represented by (5.254). One has (5.255).
Any f 2 Fpq ./ can be represented by (5.251) with (5.256) .u-wavelet frame/.
s

Proof. As remarked above  is interior regular according to (4.95). Then it follows

from (4.99), (4.101), based on (4.97), (4.98), by the same arguments as in the proof of
Theorem 5.51 that (5.249), (5.254) are atomic decompositions. One obtains (5.259).
Again by the above references  is also a bounded I -thick domain with j@j D 0.
Then it follows from Theorem 4.4 that there are linear and bounded extension operators
(4.16), (4.17). Now one can argue as in the proof of Theorem 5.51.
Remark 5.54. One can extend part (ii) to Lp ./ D Fp;2
0
./ with 1 < p < 1. But
this is covered in a better way by Theorem 2.36.
Chapter 6
Complements

6.1 Spaces on cellular domains

6.1.1 Riesz bases

It is the main aim of Section 6.1 to prove Theorem 5.49 and, as a consequence, Theo-
rem 5.47. But first we describe the background and rephrase for this purpose some of
the previous notation and assertions (also to make Section 6.1 to some extent indepen-
dently readable).
Definition 6.1. Let  be a bounded Lipschitz domain in Rn with n  2 according to
Definition 3.4 (iii) or a bounded interval in R if n D 1. Let
˚
Z D xlj 2  W j 2 N0 I l D 1; : : : ; Nj ; (6.1)
typically with Nj  2j n , such that for some c1 > 0,

jxlj  xlj0 j  c1 2j ; j 2 N0 ; l 6D l 0 : (6.2)

 
For some c2 > 0 let j l be the characteristic function of the ball B xlj ; c2 2j  Rn
.centred at xlj and of radius c2 2j /. Let s 2 R, 0 < p  1, 0 < q  1. Then
s
bpq .Z / is the collection of all sequences
˚
 D jl 2 C W j 2 N0 I l D 1; : : : ; Nj (6.3)
such that
X
1 X
Nj q=p 1=q
 
 jb s .Z / D n
2j.s p /q jjl jp < 1; (6.4)
pq
j D0 lD1

s
and fpq .Z / is the collection of all sequences (6.3) such that

  X1 X 1=q 
Nj

 jf s .Z / D 
 2jsq jjl j l .  /jq

jLp ./ < 1 (6.5)
pq
j D0 lD1

.obviously modified if p D 1 and/or q D 1/.

Remark 6.2. This coincides essentially with Definition 5.23 (inserted here for sake of
completeness). As mentioned in Remark 5.24 with a reference to [T06], Section 1.5.3,
pp. 18–19, for any c20 > 0 one can replace j l in (6.5) by the characteristic function of

B.xlj ; c20 2j / or of B.xlj ; c20 2j / \ 

 6.1 Spaces on cellular domains 179

(equivalent quasi-norms). If no specification is required we write

s
apq .Z / with a 2 fb; f g: (6.6)

The spaces C u ./ have the same meaning as in Definition 5.17.

Definition 6.3. Let  be a bounded Lipschitz domain in Rn with n  2 according to
Definition 3.4 (iii) or a bounded interval in R if n D 1. Let D @ and let Z be as
in (6.1), (6.2). Let u 2 N.
(i) Then ˚
ˆ D ˆjl W j 2 N0 I l D 1; : : : ; Nj  C u ./ (6.7)
x if for some c3 > 0 and c4 > 0,
is called a u-wavelet system in 
 
x
supp ˆjl  B xlj ; c3 2j \ ; j 2 N0 I l D 1; : : : ; Nj ; (6.8)

and ˇ ˛ j ˇ
ˇD ˆ .x/ˇ  c4 2j n2 Cj j˛j ; j 2 N0 I l D 1; : : : ; Nj ; x 2 ; (6.9)
l
for all ˛ 2 Nn0 with 0  j˛j  u.
(ii) The above u-wavelet system is called oscillating if there are positive numbers
c5 ; c6 ; c7 with c6 < c7 such that

dist.B.xl0 ; c3 /; /  c6 ; l D 1; : : : ; N0 I (6.10)

and
ˇZ ˇ
ˇ ˇ
ˇ .x/ ˆjl .x/ dx ˇˇ
n
 c5 2j 2 j u k jC u ./k; 2 C u ./; (6.11)
ˇ


for all ˆjl with j 2 N and

dist.B.xlj ; c3 2j /; / 62 .c6 2j ; c7 2j /: (6.12)

(iii) An oscillating u-wavelet system according to part (ii) is called interior if, in
addition,

dist.B.xlj ; c3 2j /; /  c6 2j ; j 2 N0 I l D 1; : : : ; Nj : (6.13)

Remark 6.4. Parts (i) and (ii) coincide essentially with Definition 5.25 (again inserted
here for sake of completeness). The oscillation (6.11) is not required for the terms with
j D 0 and also not for the wavelets ˆjl with j 2 N and

dist.xlj ; /  dist.supp ˆjl ; /  2j ; (6.14)

roughly speaking. One may ask for a weaker version restricting (6.11) to all ˆjl with

dist.supp ˆjl ; /  c6 2j ; j 2 NI l D 1; : : : ; Nj : (6.15)

180 6 Complements

Boundary wavelets with \ supp ˆjl 6D ; would no longer assumed to oscillate. But
(almost) all of our constructions so far and also in the sequel produce the above sharper
version. The only exceptions are the u-wavelet systems in Section 5.4. But they are
somewhat outside of our main interests. One may consult Remark 5.52. Similarly one
may call a u-wavelet system interior if one has

supp ˆjl  ; j 2 N0 I l D 1; : : : ; Nj ; (6.16)

instead of the stronger assumption (6.13). But all our constructions of u-wavelet bases
or u-wavelet frames with (6.16) satisfy automatically (6.13).
We recalled at the beginning of Section 5.2.1 what is meant by a Riesz basis in
L2 . /. The related sequence space is `2 . Furthermore we refer to Section 1.2.2
for a definition of (unconditional) bases in complex quasi-Banach spaces. There is
s
the following natural generalisation of Riesz bases. Let apq .Z / be as in (6.6) and
Definition 6.1.
Definition 6.5. Let  be a bounded Lipschitz domain in Rn with n  2 according
to Definition 3.4 (iii) or a bounded interval on R if n D 1. Let 0 < p; q < 1
and s 2 R. Let AOpq s
./ be either Apq s
./ or AQpq
s
./ according to Definition 2.1.
˚
An Œoscillating finteriorg u-wavelet system ˆ D ˆjl  AOpq s
./ as introduced in
Definition 6.3 (i)[(ii)]f(iii)g is called an Œoscillating finteriorg u-Riesz basis for AOpq
s
./
if it has the following properties:
1. An element f 2 D 0 ./ belongs to AOpq
s
./ if, and only if, it can be represented
as
X1 X Nj
f D jl 2j n=2 ˆjl ;  2 apq
s
.Z /; (6.17)
j D0 lD1

unconditional convergence being in AOpq

s
./.
2. The representation (6.17) is unique,
1 X
X Nj
f D jl .f / 2j n=2 ˆjl ; (6.18)
j D0 lD1

where jl .  / 2 AOpq

s
./0 are linear and continuous functionals on AOpq
s
./.
3. Furthermore,
˚
f 7! jl .f / is an isomorphic map of AOpq
s s
./ onto apq .Z /. (6.19)

Remark 6.6. It is the main aim of this Section 6.1 to have a closer look at wavelet
bases in spaces Apqs
./ where  is a cellular domain and p < 1, q < 1. As
will be recalled below a cellular domain is a special bounded Lipschitz domain. This
 6.1 Spaces on cellular domains 181

may explain our restriction in the above definition to spaces with p < 1, q < 1 in
bounded Lipschitz domains. It is quite obvious that one can extend the above definition
of u-Riesz bases to spaces AOpqs
./ with p < 1, q < 1 in arbitrary domains  in Rn ,
Definition 2.1, or in diverse types of thick domains, Definition 3.1. Furthermore,
parallel to [oscillating]finteriorg u-Riesz bases one can introduce [oscillating]finteriorg
u-Riesz frames. Then one can also admit q D 1 and, for the B-spaces, p D 1. This
can be done in a rather obvious way, where Remark 5.11 suggests to incorporate in a
corresponding definition both stability and optimality. Furthermore there are natural
counterparts of u-Riesz bases and u-Riesz frames for spaces Apq s
. / on compact C 1
manifolds according to Definitions 5.1, 5.5. We stick at the above definition illustrated
by the following examples.
Theorem 6.7. (i) Let  be a bounded Lipschitz domain in Rn according to Defini-
tion 3.4 (iii). Then any space Apq
s
./, AQpq
s
./ covered by Definition 3.11 with p < 1,
q < 1 has for u 2 N with

u > max.s; p  s/ for B-spacesI u > max.s; pq  s/ for F -spaces; (6.20)

an interior u-Riesz basis.

(ii) Let  be a bounded C 1 domain in the plane R2 according to Definition 3.4 (iii).
Then any space Apq s
./ with (5.190) has for any u 2 N with u > s an oscillating
u-Riesz basis.
(iii) Let  be a bounded C 1 domain in Rn with n  3 according to Defini-
tion 3.4 (iii) such that each connected boundary component is diffeomorphic to the
sphere Sn1 . Then any space Apq s
./ with (5.215) and 0 < q < 1 for B-spaces,
1  q < 1 for F -spaces has for any u 2 N with u > s an oscillating u-Riesz basis.
Proof. These are special cases and reformulations of Theorems 3.13, 3.23, 5.35, 5.38.
In the related formulation one can replace r < u in Theorem 5.35 by s < u.
Remark 6.8. We rephrased the previous assertions to provide a better understanding
of what follows. It is one of the main aims of this Section 6.1 to get rid of the additional
assumption in part (iii) of the above theorem that the connected boundary components
of @ are diffeomorphic to Sn1 . In case of cellular domains (as recalled below) we
have so far interior u-Riesz bases for the spaces covered by Theorem 5.43. In what
follows it is our main goal to prove Theorems 5.47 and 5.49. But first we have a closer
s
look at some peculiarities for spaces of type Apq ./ in cellular domains also for its
own sake.

6.1.2 Basic properties

In Definition 5.17 we introduced the spaces C 1 ./ in domains  in Rn . Recall that

a one-to-one map from a bounded domain  in Rn onto a bounded domain ! in Rn ,

W  3 x 7! y D .x/ 2 !; (6.21)
182 6 Complements

is called a diffeomorphic map if

l 2 C 1 ./ and . 1
/l 2 C 1 .!/; l D 1; : : : ; n; (6.22)
1 1
for the components l of and . /l of its inverse ,
1 1
B D id in  and B D id in !:

Let again
Q D fx 2 Rn W 0 < xl < 1g (6.23)
be the unit cube in R . We explained at the beginning of Section 5.3.3 what is meant
n

by a polyhedron with Q in (6.23) as a proto-type. A bounded domain  in Rn is said

to be diffeomorphic to a bounded domain ! in Rn if there is a diffeomorphic map of
a neighbourhood of  x onto a neighbourhood of !
x with ! D ./. If is a set in Rn
ı ı
then is the largest open set in R with
n
 (the interior of ).
Definition 6.9. A domain  in Rn , n  2, is said to be cellular if it is a bounded
Lipschitz domain according to Definition 3.4 (iii) which can be represented as
S
L ı
D l ; with l \ l 0 D ; if l 6D l 0 ; (6.24)
lD1

such that each l is diffeomorphic to a polyhedron.

Remark 6.10. This coincides essentially with Definition 5.40 (i) inserted here for sake
of completeness. Since l .l / is a domain it follows that l is also a domain.
Proposition 6.11. A bounded C 1 domain in Rn according to Definition 3.4 (iii), n  2,
is cellular.
Proof. This follows by direct arguments. It has also been mentioned in Remark 5.41.

In (3.84) we recalled what is meant by a pointwise multiplier for Apq s

.Rn / with a
reference to [RuS96] for explanations. We are interested in characteristic functions 
of domains  as pointwise multipliers.
Proposition 6.12. Let  be a cellular domain in Rn according to Definition 6.9. Let
s
Apq .Rn / and Apq
s
./ be the spaces introduced in Definitions 1.1, 2.1.
(i) Then  is a pointwise multiplier in Apq
s
.Rn / if, and only if,
   
1 1 1
0 < p  1; 0 < q  1; max n 1 ; 1 <s < ; (6.25)
p p p
.p < 1 for the F -spaces/.
(ii) Let 0 < p  1 and 0 < q  1 .with p < 1 for the F -spaces/. Let  2 R,
k 2 N and s D  C k. Then
˚
s
Apq ./ D f 2 Apq

./ W D ˛ f 2 Apq

./; j˛j  k (6.26)
 6.1 Spaces on cellular domains 183

and X
kf jApq
s
./k  kD ˛ f jApq

./k (6.27)
j˛jk

.equivalent quasi-norms/.

Proof. The assertion (5.222), (5.223) about characteristic functions of half-spaces as

s
pointwise multipliers in Apq .Rn / carries over to cubes, polyhedrons, and cellular do-
mains. There one finds also the necessary references. Part (ii) is a special case of
Proposition 4.21.

Proposition 6.13. Let  be a cellular domain in Rn according to Definition 6.9.

(i) Let p, q, s be as in (6.25) with p < 1 for the F -spaces. Then
s
Apq ./ D AQpq
s
./ D AQpq
s x
./ (6.28)

for the corresponding spaces from Definition 2.1 and the interpretation given in Re-
mark 2.2.
(ii) Let
   
1 1 1
0 < p < 1; 0 < q < 1; max n 1 ; 1 <s < : (6.29)
p p p
Then ı
s
Apq ./ D AQpq
s
./ D Apq
s
./ (6.30)
for the corresponding spaces from Definitions 2.1, 5.17 (i).

Proof. Step 1. According to Remark 2.2 the second equality in (6.28) is justified if
˚
g 2 Apq
s
.Rn / W supp g  @ D f0g: (6.31)

Recall that 1 
s
Apq ./ ,! L1 ./ if s > p D n p
1 C
: (6.32)
Since j@j D 0 one obtains (6.31) in these cases. It remains to prove (6.31) for the
admitted spaces with s  0. It is sufficient to deal with
s
Bpp .Rn /; 1 < p < 1; 1
p
 1 < s < 0: (6.33)

We use duality,
s
Bpp .Rn /0 D Bps n
0 p 0 .R /;
1
p
C 1
p0
D 1; 0 < s < 1
p0
; (6.34)

[T83], Section 2.11.2. Let ' 2 S.Rn /. Then it follows from Proposition 6.12 (i) that
both ' and '.1   / belong to Bps n
0 p 0 .R /. On the other hand, the translation

g.  / 7! gh .  / D g.  C h/ with h 2 Rn ; (6.35)
184 6 Complements

is continuous in Bps n n
0 p 0 .R /. To justify this claim we recall that D.R / is dense in
s s
n
Bp0 p0 .R /. If 2 D.R / approximates g 2 Bp0 p0 .R / then h approximates gh
n n

(uniformly in h) and the claimed continuity of the translation operator follows from
kf  fh jBps n
0 p 0 .R /k

 c kf  jBps
0 p 0 .R /k C c k
n
 h jBps
0 p 0 .R /k C c k.f 
n
/h jBps n
0 p 0 .R /k

 c 0 kf  jBps n 0
0 p 0 .R /k C c k  h jBps n
0 p 0 .R /k:
(6.36)
Then one has that D.Rn n @/ is dense in Bps
0 p 0 .R /. Let now g 2 Bpp .R / with
n s n

supp g  @ and let ' 2 S.Rn / be approximated in Bps n

0 p 0 .R / by '
k
2 D.Rn n @/.
Now one obtains that
g.'/ D lim g.' k / D 0 for any ' 2 D.Rn /: (6.37)
k!1

Hence g D 0. We obtain (6.31). This justifies the second equality in (6.28). The first
equality follows now from Proposition 6.12 (i).
Step 2. If p < 1, q < 1 then one obtains from the standard localisation (multipli-
cation with a smooth resolution of unity) and the continuity of the translation operator
as indicated above that D./ is dense in AQpq
s
./. This proves (6.30).
Remark 6.14. Assertions of type (6.28), (6.30) under the indicated conditions for the
parameters p, q, s have some history. As far as bounded C 1 domains are concerned
one may consult [T01], Sections 5.3–5.6, pp. 44–50, where also the above duality
argument comes from. In case of bounded Lipschitz domains we have Proposition 5.19
with the restriction q  min.p; 1/ for the F -spaces based on the references given in
Remark 5.20. It is not known whether one has a full counterpart of the pointwise
multiplier assertion in Proposition 6.12 (i) and, as a consequence, of Proposition 6.13
for bounded Lipschitz domains. As far as pointwise multipliers are concerned one may
consult [FrJ90], Corollary 13.6, and [Tri02]. This may support even at this level (before
discussing boundary values of derivatives) to specify bounded Lipschitz domains to
cellular domains which are bounded by definition. In Proposition 3.19 we described a
counterpart for more general domains.
Proposition 6.15. Let  be a cellular domain in Rn according to Definition 6.9. Let
k 2 N and
   
1 1 1
0 < p < 1; 0 < q < 1; max n 1 ; 1 < < : (6.38)
p p p
Then ı
AQpq
s
./ D Apq
s
./ with s D  C k (6.39)
.equivalent quasi-norms/ for the corresponding spaces in Definitions 2.1, 5.17.
Proof. If f 2 AQpq
s
./ then it follows from Proposition 6.12 (ii) and its Rn -counterpart
that
D ˛ f 2 AQpq 
./ D Apq

./; j˛j  k; (6.40)
 6.1 Spaces on cellular domains 185

(equivalent quasi-norms). Now one obtains (6.39) from (6.30) and the translation
argument in connection with (6.36) applied to f 2 AQpq
s
./ extended by zero to Rn
and its derivatives.
Remark 6.16. Assertions of this type for Sobolev spaces and classical Besov spaces as
mentioned in Remark 1.2 (with  in place of Rn ) in bounded C 1 domains are known
since a long time and may be found in [T78], Section 4.3.2. The extension of these
s
assertions to other spaces Apq , especially with p < 1, is a somewhat tricky game. One
may consult [T06], Section 1.11.6, the above Remark 5.20 and the references given
there.

6.1.3 A model case: traces and extension

At the end we wish to extend Theorem 6.7 (iii) to cellular domains without any ad-
ditional assumptions about the boundary. This will be done by the same method as
in Chapter 5 based on traces, wavelet-friendly extensions and related decompositions
as in Theorems 5.14, 5.21. But now one has to care for faces and edges of several
dimensions which complicates the situation. We deal first with a model case.
Let n 2 N and l 2 N with l < n. Let Rn D Rl  Rnl with x D .y; z/ 2 Rn ,

y D .y1 ; : : : ; yl / 2 Rl ; z D .z1 ; : : : ; znl / 2 Rnl ; (6.41)

where Rl is identified with the hyper-plane x D .y; 0/ in Rn . Let Ql be the unit cube
˚
Ql D x D .y; z/ 2 Rn W z D 0; 0 < ym < 1I m D 1; : : : ; l (6.42)

in this hyper-plane. Let

˚
Qln D x D .y; z/ 2 Rn ; y 2 Ql ; z 2 Rnl (6.43)

be the related cylindrical domain in Rn . Let tr l be the trace operator

tr l W f .x/ 7! f .y; 0/; f 2 Apq

s
.Rn /; (6.44)

on Rl or Ql (if exists) defined by obvious modification of (5.5) and the related com-
ments in Section 5.1.1. Let

Nnl D f˛ D .˛1 ; : : : ; ˛n / 2 Nn0 W ˛1 D    D ˛l D 0g: (6.45)

Then
@j˛j f
D˛ f D ˛ ; ˛ 2 Nnl ; (6.46)
@z1 lC1 : : : @zn˛n
are the derivatives perpendicular to Ql . Let
 
1
pl D l 1 ; 0 < p  1; l 2 N; (6.47)
p C
186 6 Complements

s
be as in (4.15), now indicating l. Let Apq .Rn / be the spaces as introduced in (1.95),
Definition 1.1, and Bpq 
.Ql / be the spaces from Definition 2.1 in the context of Rl .
We need an extension of the trace assertion (5.59)–(5.63) replacing the .n  1/-
dimensional manifold by Ql . The classical case, covering the Sobolev spaces
Hps .Rn / D Fp;2
s
.Rn /, 1 < p < 1, and the Besov spaces Bpq s
.Rn /, 1 < p < 1,
1  q  1 may be found in [T78], Section 2.9.4, 4.7.2, where we described also
s
the rich history of this problem. The full assertion for all relevant spaces Bpq .Rn /,
s n
Fpq .R / can be obtained by iteration of [T83], Theorem 3.3.3, p. 200, or nowadays by
atomic or wavelet arguments. For our later purpose p  1 would be sufficient. But as
in Section 5.1.3 traces and wavelet-friendly extensions are of interest for its own sake
and a complete description is desirable.
Proposition 6.17. Let l 2 N, n 2 N with l < n, and r 2 N0 . Let 0 < p  1
(p < 1 for the F -spaces), 0 < q  1, and
nl
s>rC C pl : (6.48)
p
Let Ql be as in (6.42). Let tr rl ,
˚
tr rl W f 7! tr l D ˛ f W ˛ 2 Nnl ; j˛j  r : (6.49)

Then
Y s nl
p j˛j
tr rl W Bpq
s
.Rn / ,! Bpq .Ql / (6.50)
˛2Nnl
j˛jr

and
Y s nl
p j˛j
tr rl W Fpq
s
.Rn / ,! Bpp .Ql / (6.51)
˛2Nn
l
j˛jr

(continuous embedding).
Remark 6.18. We refer to the above comments. But one can prove these embeddings
directly restricting atomic expansions of f 2 Apqs
.Rn / to Rl . No moment conditions
are needed in the counterparts of (5.64)–(5.67).
Crucial for what follows is the modification of the wavelet-friendly extension op-
erator as constructed in (5.75) and used in Theorem 5.14. Let C u .Ql / with u 2 N be
as in Definition 5.17 (ii) and let
˚ j
ˆm .y/ W j 2 N0 I m D 1; : : : ; Nj  C u .Ql / (6.52)

with Nj  2j l be an interior orthonormal u-wavelet basis in L2 .Ql / according to Theo-

rem 2.33 and Definitions 2.31, 2.4 (with Rl in place of Rn ). We apply Theorem 3.13
to
Bzpq

.Ql /; 0 < p  1; 0 < q  1; pl <  < u: (6.53)
 6.1 Spaces on cellular domains 187

Then one has for g 2 Bzpq


.Ql / the wavelet expansion

1 X
X Nj
gD jm .g/ 2j l=2 ˆjm (6.54)
j D0 mD1

with Z
jm .g/ D 2j l=2 g.y/ ˆjm .y/ dy (6.55)
Ql

and the isomorphic map

˚
g 7! jm .g/ of Bzpq
 
.Ql / onto bpq .ZQl /: (6.56)
˚

Here bpq .ZQl / comes from (6.4). If p < 1, q < 1 then ˆjm in (6.52) is an interior
u-Riesz basis in Bzpq

.Ql / according to Definition 6.5. We refer to Theorem 3.13 where
we clarified the convergence of (6.54) also in the cases where p D 1 and/or q D 1.
Next we modify the construction (5.75) of an extension operator such that it can be
applied to the right-hand sides of (6.50), (6.51). Let
˚
 2 D.Rnl /; supp   z 2 Rnl W jzj  2 ; .z/ D 1 if jzj  1; (6.57)

with z 2 Rnl as in (6.41). One may assume that  has sufficiently many moment
conditions, Z
.z/ z ˇ dz D 0 if jˇj  L: (6.58)
Rnl

Let ˆjm be the wavelets according to (6.52) and

j j˛j ˛
m .x/ D 2
ˆj;˛ z .2j z/ 2.nl/j=2 ˆjm .y/; ˛ 2 Nnl ; (6.59)

according to (6.41) and (6.45) be the counterpart of (5.76). Then

˚
g D Extr;u
l
g˛ W ˛ 2 Nnl ; j˛j  r .x/
1 X
XX Nj
1 j (6.60)
D m .g˛ / 2j j˛j 2j n=2 ˆj;˛
m .x/
˛Š
j˛jr j D0
mD1

is the counterpart of (5.75) where fg˛ g  L1 .Ql / and jm .g˛ / as in (6.55) (with g˛ in
place of g). Let AQpq
s
.Qln / be the spaces as introduced in Definition 2.1 where Qln is
the cylindrical domain (6.43).
˚
Theorem 6.19. Let ˆjm with u 2 N be the interior orthonormal u-wavelet basis in
L2 .Ql / according to (6.52). Let  be as in (6.57), (6.58) (with L 2 N0 sufficiently
large in dependence on q for the F -spaces). Let r 2 N0 and let Extr;u l
be given by
(6.60) with (6.55). Then
˚
Extr;u
l
W g˛ W ˛ 2 Nnl ; j˛j  r 7! g (6.61)
188 6 Complements

is an extension operator
Y nl
s p j˛j
Extr;u
l
W Bzpq .Ql / ,! Bzpq
s
.Qln /;
˛2Nn
l (6.62)
j˛jr

0 < p  1; 0 < q  1; u>s>rC nl

p
C pl ;

and
Y nl
s p j˛j
Extr;u
l
W Bzpp .Ql / ,! Fzpq
s
.Qln /;
˛2Nnl (6.63)
j˛jr

0 < p < 1; 0 < q  1; u>s>rC nl

p
C pl ;

with
Y nl
s p j˛j
tr rl B Ext r;u
l
D id; identity in Bzpq .Ql /: (6.64)
˛2Nn
l
j˛jr

Proof. Since s > pl and, for p < 1,

nl n
s> C pl D  l > pn ; (6.65)
p p

it follows from Definition 2.1 and Remark 2.2 that one can identify Bzpq 
.Ql / in
(6.62), (6.63) with Bzpq .Ql /, considered as a subspace of Bpq .R /, and AQpq .Qln / with
  l s

AQpq
s
.Qln / as a subspace of Apq
s
.Rn /. Then the case l D n  1 is essentially covered by
Theorem 5.14 and its proof. There we dealt with a bounded C 1 domain  in place of
Qln and its boundary D @ in place of Qn1 . Since we replaced Bpq 
. / now by
z 
Bpq .Qn1 / we avoid any difficulty which may be caused by the boundary @Qn1 of
Qn1 . But otherwise one can follow the arguments in the proof of Theorem 5.14 for all
l 2 N with l < n. In particular the functions ˆj;˛m in (6.59) are the direct counterpart
of (5.76) (not normalised atoms). By (6.65) no moment conditions are needed in case
of the B-spaces. If "  q < 1 for some " with 0 < " < 1 then one may choose
L 2 N with L  n 1"  1 in (6.58) in case of the F -spaces.
Remark 6.20. It comes out that Extr;u
l
is a common extension operator for given r; u
and 0 < "  1 with "  q in (6.63) in place of 0 < q.

6.1.4 A model case: approximation, density, decomposition

Let Ql and Qln be as in (6.42), (6.43). By Definition 5.17 (i) the completion of D.Qln /
ı
s
in Apq .Qln / is denoted by Apq
s
.Qln /. One can replace  in Proposition 6.15 by Qln .
 6.1 Spaces on cellular domains 189

Then one obtains that ı

AQpq
s
.Qln / D Apq
s
.Qln / (6.66)
(equivalent quasi-norms) if
1
1  p < 1; 0 < q < 1; 0<s 62 N: (6.67)
p
If p < 1 then one has to exclude some .p; s/-regions in order to obtain (6.66) (recall
that n  2). This is a first hint that it might be reasonable to restrict the further
considerations to p  1. But we add a second more substantial argument. For
0 < p < 1, 0 < q < 1 and the largest admitted r 2 N0 in (6.48) one may
ask under which circumstances
  ˚
D Qln n Ql is dense in f 2 AQpqs
.Qln / W tr rl f D 0 : (6.68)
Of course, Qln n Ql is considered as a set in Rn . As for the classical cases we refer to
[T78], Section 2.9.4, p. 223, complemented by [T83], Section 3.4.3, p. 210. We begin
with a preparation using the same notation as in (6.41)–(6.43) and in Proposition 6.17.
Proposition 6.21. Let l 2 N, n 2 N with l < n and r 2 N0 . Let
nl
0 < p < 1; 0 < q < 1; s>rC C pl : (6.69)
p
Let u 2 N with u > s. Then
˚
g 2 C u .Rn / W supp g  fy 2 Ql g  fjzj < 1g; tr rl g D 0 (6.70)
is dense in ˚
f 2 AQpq
s
.Qln / W tr rl f D 0 : (6.71)
Proof. By the same decomposition and translation arguments as in connection with
(6.36) now with h D .hl ; 0/ parallel to Rl it follows that it is sufficient to approximate
f 2 Apq
s
.Rn /; supp f  Qln ; tr rl f D 0; (6.72)
(hence vanishing near @Qln ). For given " > 0 one can approximate f in Apq
s
.Rn / by
f" 2 S.R / with
n

supp f"  Qln and k tr rl f" j tr rl Apq

s
.Rn /k  "; (6.73)
where tr rl Apq
s
.Rn / stands for the trace spaces according to (6.50), (6.51). Let
˚ r
f " D Extr;uC1
l
tr l f" (6.74)

with Extr;uC1
l
as in (6.60) and Theorem 6.19. By the properties of Ext r;uC1
l
as men-
tioned in Remark 6.20 one obtains that
f " 2 C u .Rn /; supp f "  Qln ; kf " jApq
s
.Rn /k  c"; (6.75)
for some c > 0 which is independent of ". Then f"  f is the desired approximation.
"
190 6 Complements

Remark 6.22. We return to the question (6.68). By Proposition 6.21 it is sufficient to

approximate the function g from (6.70). Let
˚
Sj D x D .y; z/ W y 2 Ql ; 2j 1  jzj  2j C1 ; j 2 N0 ; (6.76)
and let
1 X
X Nj
1D 'j m .x/; Nj  2j l ; x D .y; z/; y 2 Ql ; 0 < jzj < 1; (6.77)
j D0 mD1

be a resolution of unity by suitable C 1 functions 'j m with

supp 'j m  Bj m  Sj ; j 2 N0 ; 1  m  Nj ; (6.78)
where Bj m is a ball of radius 2j . Let g be as in (6.70) (in particular g vanishes near
the faces @Ql of Ql as an Rl -set). Then

X Nj
1 X X
n
gD j m 2j.rC1/ 2j.s p / 'j m g D j m aj m (6.79)
j D0 mD1 j;m

s
is an atomic decomposition of g in Bpq .Rn / according to Theorem 1.7 with p, q, s as
in (6.69) (not to speak about immaterial constants). By (6.65) no moment conditions
are needed. But we used tr rl g D 0 according to (6.49) for g 2 C u .Rn / and related
Taylor expansions in z-directions, where the factor 2j.rC1/ comes from. We have
n
j m  2j.rC1sC p / : (6.80)
Let gJ with J 2 N be given by (6.79) with j  J in place of j  0. Then it follows
by Theorem 1.7 that
X
1  X
Nj q=p 1=q
kgJ jBpq
s
.Rn /k c jj m j p

j DJ mD1
X
1
n l
1=q (6.81)
 c0 2jp.rC1sC p  p /q=p
j DJ

 c 2J ı
00
if ı D r C 1  s C nl
p
> 0:
s
Then ggJ approximates g in Bpq .Rn /. These functions vanish near Ql . An additional
mollification gives (6.68) for Bzpq .Qln /. The corresponding assertion for Fzpq
s s
.Qln /
follows from
Bzp;min.p;q/
s
.Qln / ,! Fzpq
s
.Qln / (6.82)
and Proposition 6.21. The largest admitted r 2 N0 in (6.69) is given by
nl
r Ds  pl  " for some " with 0 < "  1. (6.83)
p
 6.1 Spaces on cellular domains 191

If p  1 then
nl
ı D 1  " > 0 if, and only if, s  62 N: (6.84)
p
If p < 1 then (6.81) requires
 
1
ı D 1  "  pl D 1  "  l  1 > 0: (6.85)
p
Although there are some p < 1, 0 < " < 1 with (6.85) the situation is not very
satisfactory. Together with (6.66), (6.67) and the comments afterwards it is a second
good reason to restrict what follows to p  1.
After these discussions we are in a similar situation as in Corollary 5.16 and Theo-
rem 5.21, now with Qln in place of the bounded C 1 domain  and Ql in place of
D @. We describe briefly the (more or less technical) changes. First we remark
that one has by Proposition 6.17 for the same parameters as there the traces
Y nl
s p j˛j
tr rl W Bzpq
s
.Qln / ,! Bzpq .Ql / (6.86)
˛2Nnl
j˛jr

and
Y nl
s p j˛j
tr rl W Fzpq
s
.Qln / ,! Bzpp .Ql /: (6.87)
˛2Nn
l
j˛jr

In the same way as in (5.99), (5.100) one obtains now from Theorem 6.19 for the same
parameters p; q; s (and r; u) as there that

Prl D Extlr;u B tr rl W AQpq

s
.Qln / ,! AQpq
s
.Qln / (6.88)

is a projection. Let Prl AQpq

s
.Qln / be its range. Then Trl D id  Prl is also a projection
and ˚
Trl AQpq
s
.Qln / D f 2 AQpq s
.Qln / W tr rl f D 0 (6.89)
is the counterpart of (5.113). As in (5.101) one has the isomorphic maps
Y s nl
p j˛j
‡
Extr;u
l
Bzpq .Ql / D Prl Bzpq
s
.Qln /;
˛2Nn
l ;j˛jr
Y (6.90)
s nl j˛j
Extr;u
l
Bzpp p .Ql / D Prl Fzpq
s
.Qln /:
˛2Nn
l ;j˛jr

Now one can clip together the above considerations. Recall that Ql and related spaces
refer to Rl , whereas Qln and also Qln n Ql and related spaces are considered in Rn .
Then one has the following counterpart of Theorem 5.21.
192 6 Complements

Theorem 6.23. Let l 2 N and n 2 N with l < n. Let

nl 1
1  p < 1; 0 < q < 1; 0<s 62 N; s 62 N: (6.91)
p p
Let Ql and Qln be as in (6.42), (6.43). Then
ı
AQpq
s
.Qln / D Apq
s
.Qln /: (6.92)

Let r D s  nl
p
2 N0 be the largest integer smaller than s  nl
p
. For s < u 2 N let
r;u
Extl be the extension operator according to (6.60), (6.55) and Theorem 6.19. Then
Y nl
p j˛j
ı ı s
s
Bpq .Qln / D Bpq
s
.Qln n Ql /  Ext r;u
l
Bzpq .Ql / (6.93)
˛2Nnl ;
j˛jr

and
Y nl
p j˛j
ı ı s
s
Fpq .Qln / D Fpq
s
.Qln n Ql /  Ext r;u
l
Bzpp .Ql / (6.94)
˛2Nnl ;
j˛jr

(complemented subspaces).
ı
Proof. First we remark that (6.92) follows from (6.66), (6.67). Then Apqs
./ can be
decomposed by (6.90) and (6.89) into the second factors on the right-hand sides of
(6.93), (6.94) and the space on the right-hand side of (6.89). By Remark 6.22 we have
(6.68). This completes the proof of the theorem.

Corollary 6.24. Theorem 5.21 remains valid if one replaces (5.115) by

1
1  p < 1; 0 < q < 1; 1 < s  62 N0 ; (6.95)
p
both for B-spaces and F -spaces.
Proof. This follows from Proposition 6.13 (ii) (instead of Proposition 5.19) and the
proof of the above theorem with l D n  1.

Remark 6.25. The restriction q  1 for the F -spaces in Theorem 5.21 comes from
the reference to [T83] in its proof which can now be improved by the above arguments.

6.1.5 Cubes and polyhedrons: traces and extensions

Let

Q D fx 2 Rn W x D .x1 ; : : : ; xn /; 0 < xm < 1I m D 1; : : : ; ng (6.96)

 6.1 Spaces on cellular domains 193

be the unit cube in Rn where 2  n 2 N. We wish to extend Theorem 6.23 from Qln
to Q. This is mainly a technical matter, but it requires some care. It is based on the
crucial observation that corresponding extension operators of type (6.59), (6.60) from
l-dimensional faces of Q of type Ql into Q do not interfere with the boundary data
at all other faces (edges, vertices) of Q. This can be extended to bounded polyhedral
domains in Rn without essential changes in a more or less obvious way. One may
consider Q as a prototype of a polyhedron in Rn . The boundary D @Q of Q can be
represented as
[
n1
0
D l; l \ l 0 D ; if l 6D l ; (6.97)
lD0
where l collects all l-dimensional faces (edges, vertices). Then l consists of finitely
many l-dimensional cubes of type Ql as in (6.42) (open as a subset of Rl ). We wish to
apply Theorem 6.23 where we have now faces of dimensions l D 0; : : : ; n  1. This
suggests the restriction
1 k
1  p < 1; 0 < q < 1; s> ; s 62 N0 for k D 1; : : : ; n: (6.98)
p p
The trace (6.44) is now given by the restriction
tr l W f .x/ 7! f j l ; f 2 Apq
s
.Rn /; (6.99)
of f to l . If 2 l the D˛ f denotes the obvious counterpart of (6.46) where only
derivatives are admitted referring to directions perpendicular in Rn to l in . One
replaces (6.49) by
˚
tr rl W f 7! tr l D˛ f W j˛j  r ; l D 0; : : : ; n  1; (6.100)
for traces on l. Let 1  p < 1 and let be either
n n  l0 1
0<s 62 N or 0<s < for some l0 D 1; : : : ; n  1. (6.101)
p p p
If s > n=p then we put l0 D 0. Recall that Œa is the largest integer smaller than or
equal to a 2 R. Let
 
rN D r l0 ; : : : ; r n1 with r l D s  nl
p
; l0  l  n  1: (6.102)
Then
˚
tr rN W f 7! tr l D˛ f W 2 l; j˛j  r l I l0  l  n  1 (6.103)
is the appropriate modification of (6.49).
Proposition 6.26. Let Q be the cube (6.96) with the boundary D @Q according to
(6.97). Let
1 k
1  p < 1; 0 < q < 1; s> ; s 62 N0 for k D 1; : : : ; n: (6.104)
p p
194 6 Complements

Let l0 D 0 if s > n=p and otherwise l0 D 1; : : : ; n  1, such that

n  l0 1
0<s < : (6.105)
p p

Let trprN be as in (6.102), (6.103). Then

Y
n1 Y s nl
p j˛j
tr rN W Bpq
s
.Rn / ,! Bpq . l/ (6.106)
lDl0 j˛jr l

and
Y
n1 Y s nl
p j˛j
tr rN W Fpq
s
.Rn / ,! Bpp . l /: (6.107)
lDl0 j˛jr l

One can replace Bpq

s
.Rn / by Bpq
s
.Q/ and Fpq
s
.Rn / by Fpq
s
.Q/.

Proof. This follows immediately from Proposition 6.17 and the above considerations.

Remark 6.27. One can extend the above assertion to q D 1 in (6.104), p D 1

for the B-spaces, and also to p < 1 with appropriately adapted restrictions for s as
in Proposition 6.17. But our main concern is the counterpart of Theorem 6.23 for the
above cube Q. For this purpose one has to clip together the extension operators Extr;u
l
as used in Theorems 6.19 and 6.23 with a reference to (6.60). Recall that l consists
of l-dimensional (open as subsets of Rl ) disjoint cubes of type Ql in (6.42),
[
l D l n k; l D 1; : : : ; n  1; (6.108)
k<l

(corner points if l D 0). In particular, D. l / is the disjoint union of sets of type D.Ql /.
ı
In the same way one has to interprete Bpq 
. l / and Bzpq

. l /. With p, q, s as in (6.104)
one obtains by Proposition 6.15 that

ı s nl j˛j s nl

p j˛j
Bpq p
. l / D Bzpq . l /; l D 1; : : : ; n  1; j˛j  r l : (6.109)

Let Extr;u
l be the obvious generalisation of (6.60) from Ql to l and let

N ˚ r l ;u
Extr;u
 D Ext l W l0  l  n  1 (6.110)

with rN and l0 as in (6.102). If s > n=p then l D l0 D 0 is admitted. These are the
corner points of Q. Then one has to modify (6.59), (6.60) as in Section 5.2.3 where
we dealt with the one-dimensional case and traces in points.
 6.1 Spaces on cellular domains 195

Theorem 6.28. Let Q be the cube in (6.96) with the boundary D @Q according to
(6.97). Let
1 k
1  p < 1; 0 < q < 1; s> ; s 62 N0 for k D 1; : : : ; n: (6.111)
p p
N
Let s < u 2 N. Let l0 and rN be as in (6.101), (6.102) and Proposition 6.26. Let Ext r;u

be given by (6.110). Then
N ˚ l
Ext r;u
 W g˛ W l0  l  n  1; j˛j  r
l
7! g (6.112)

is an extension operator,

Y
n1 Y nl
s p j˛j
N
Extr;u
 W Bzpq s
. l / ,! Bpq .Q/ (6.113)
lDl0 j˛jr l
and
Y
n1 Y nl
s p j˛j
N
Extr;u
 W Bzpp s
. l / ,! Fpq .Q/: (6.114)
lDl0 j˛jr l

Let tr rN be the trace operator according to (6.103). Then

Y
n1 Y nl
s p j˛j
tr rN B N
Ext r;u
 D id; identity in Bzpq . l /: (6.115)
lDl0 j˛jr l

Furthermore,
Y
n1 Y nl
s p j˛j
s
Bpq .Q/ D Bzpq
s N
.Q/  Ext r;u
 Bzpq . l/ (6.116)
lDl0 j˛jr l

and
Y
n1 Y nl
s p j˛j
s
Fpq .Q/ D Fzpq
s N
.Q/  Ext r;u
 Bzpp . l/ (6.117)
lDl0 j˛jr l

(complemented subspaces).
Proof. We use Theorem 6.23 and an induction by dimension l  l0 . Let l0  1. Then
one has by Proposition 6.13 that
nl0 nl0
s s
Bpq p
. l0 / D Bzpq p
. l0 /: (6.118)

Now one obtains by the above considerations with r l0 D 0 that

nl0
l0 ;u s
Extr l W Bpq p
. l0 /
s
,! Bpq .Q/: (6.119)
0
196 6 Complements

If l0 D 0 then one has to modify the above arguments according to Section 5.2.3 where
s
we dealt with the one-dimensional case. This decomposes Bpq .Q/ as in (6.89), (6.90),

˚ l l0 s
nl0
s
Bpq .Q/ D f 2 Bpq
s
.Q/ W tr rl0 f D 0  Extr l ;u Bpq p . l0 / (6.120)
0 0

into complemented subspaces. Let l D l0 C 1. The trace space of the first factor on
s nl
the right-hand side of (6.120) is now Bzpq p . l /. But this observation ensures that
one can apply the decomposition technique in connection with Theorem 6.23. Then
one obtains the factor with l D l0 C 1 on the right-hand side of (6.116). It remains a
l
complemented subspace now with tr rl f D 0 for l D l0 and l D l0 C 1. We remark
again that according to the construction (6.60) the extensions at different values of l do
not interfere. Iteration finally ends up with (6.116), (6.117) where one has to use the
counterpart of (6.92) or Proposition 6.15 with  D Q both for B-spaces and F -spaces.
This proves also (6.113)–(6.115).

Remark 6.29. Beyond the somewhat cumbersome technicalities the proof relies on
the following two remarkable facts:
1. With p, q, s as in (6.111) the boundary data of f 2 Apq s
.Q/ on different faces
0
of l , but also for different l and l 0 with l 6D l are totally decoupled.
2. The wavelet-friendly extension operator for a face of l respects this observation
and does not interfere with the boundary data of other faces of l , and with the
boundary data of faces l 0 with l 6D l 0 .

6.1.6 Cubes and polyhedrons: Riesz bases

Theorem 6.28 brings us in the same position as in Chapter 5 where we asked for wavelet
s
bases in spaces Apq ./ in case of bounded C 1 domains . We rephrased in Theo-
rem 6.7 some distinguished assertions obtained there now in terms of [oscillating]fin-
teriorg u-Riesz bases introduced in Definition 6.5. Let apqs
.Z / with a 2 fb; f g be
the sequence spaces according to (6.6) and Definition 6.1.
Theorem 6.30. Let  D Q be the cube (6.96) in Rn with n  2. Let Apq
s
./ be the
spaces according to Definition 2.1 (i) with
1 k
1  p < 1; s> ; s 62 N0 for k D 1; : : : ; n; (6.121)
p p

0 < q < 1 for B-spaces, 1  q < 1 for F -spaces. Then Apq

s
./ has for any u 2 N
with u > s an oscillating u-Riesz basis.
Proof. By Theorem 6.7 (i) all spaces on the right-hand sides of (6.116), (6.117) have
interior u-Riesz bases (where spaces on l are considered in Rl ). This is the point where
 6.1 Spaces on cellular domains 197

one has to strengthen 0 < q < 1 in (6.111) for F -spaces by 1  q < 1 (ensuring
pq D 0 in (3.45)). We described in Proposition 5.34 and Theorem 5.27 for bounded
C 1 domains  how wavelet bases for AQpq s
./ can be complemented by transferred
s
wavelet bases for trace spaces such that one obtains wavelet bases for Apq ./. This
applies also to the above situation now based on (6.116), (6.117), resulting in the above
theorem.

Corollary 6.31. Theorem 6.30 remains valid for bounded polyhedrons  in Rn .

Proof. The proof of Theorem 6.28 is based on the extension operators (6.59), (6.60),
clipped together by (6.110). One does not need that the respective faces intersecting
each other are orthogonal. It is sufficient to know that they intersect with non-zero
angles. Similarly one can replace the perpendicular derivatives in (6.103) by oblique
derivatives. This does not influence the arguments and results in the above assertion.

6.1.7 Cellular domains: Riesz bases

After all these preparations we reach now the main goal of Section 6.1, the proof of
Theorems 5.47 and 5.49. For sake of completeness we repeat in addition Theorem 5.43
rephrased in terms of u-Riesz bases.

Theorem 6.32. Let  be a cellular domain in Rn with n  2 according to Definition 6.9.

Let Apqs
./ be the spaces as introduced in Definition 2.1 (i).
(i) Then Apq
s
./ with (6.121), 0 < q < 1 for B-spaces, 1  q < 1 for F -spaces,
has for any u 2 N with u > s an oscillating u-Riesz basis according to Definition 6.5.
(ii) Let
 
1 n
0 < p < 1; 0 < q < 1; 1 < s < min ; : (6.122)
p n1

Then Apq
s
./ has for any sufficiently large u 2 N,

u > u.s; p/ 2 N for B-spaces; u > u.s; p; q/ 2 N for F -spaces;

an interior u-Riesz basis according to Definition 6.5.

Proof. Part (ii) is covered by Theorem 5.43 and the above reformulations. As for
part (i) we first remark that diffeomorphic maps of cubes onto corresponding domains
generate isomorphic maps both for related spaces and for wavelet bases. The extension
operators used in Theorem 6.28 transfer boundary spaces (and wavelets) from l to Rn ,
in particular to all adjacent cells. Hence the constructions resulting in Theorem 6.28
apply simultaneously to all cells (diffeomorphic images of cubes and, more general,
polyhedrons). Then one obtains the theorem as in Theorem 6.30 and Corollary 6.31.
198 6 Complements

Theorem 6.33. Let n  2. Theorem 6.32 remains valid for any bounded C 1 domain
in Rn according to Definition 3.4 (iii).
Proof. This follows from Theorem 6.32 and Proposition 6.11 (i).
Remark 6.34. In Definition 6.5 we introduced [oscillating]finteriorg u-Riesz bases
and mentioned in Remark 6.6 that there is an obvious counterpart in terms of u-Riesz
s
frames. According to Theorem 5.51 one has u-Riesz frames for all spaces Apq ./
in bounded Lipschitz domains  without any exceptional values p and s. Under the
restrictions of Theorem 5.27, in particular s  p1 62 N0 , the admitted spaces Apq
s
./
have oscillating u-wavelet frames. The step from frames to bases requires by our
method additional assumptions of type (6.121). The restriction s  p1 62 N0 is natural.
We discuss this point in Sections 6.2.2–6.2.4 below. The situation might be different
in case of the additional restrictions s  pk 62 N0 with 2  k 2 N. For some bounded
C 1 domains  in Rn with n  3 we have the restrictions (5.215). If  is a cube or a
cellular domain then one has in Theorems 6.30, 6.32 even the restrictions (6.121). We
discussed this situation in Section 5.3.1. Let

Hps ./ D Fp;2

s
./; 1 < p < 1; s 2 R; (6.123)

be the usual Sobolev spaces (the restrictions of the spaces Hps .Rn / in (1.17)–(1.19)
to ). Let M3 be the torus of revolution in R3 as in (5.197) and let B3 be the unit ball
in R3 as in (5.198). Then it follows from Theorems 5.38, 6.32, 6.33 and Remark 5.36
that one has oscillating u-Riesz bases in

Hps .M3 / if 1 < p < 1, 0 < s  1

p
62 N; (6.124)

in
Hps .B3 / if 1 < p < 1, 0 < s  1
p
62 N, s  2
p
62 N0 , (6.125)
and in

Hps ./ if 1 < p < 1, 0 < s  1

p
62 N, s  2
p
62 N0 , s  3
p
62 N0 , (6.126)

for cellular (in particular bounded C 1 ) domains  in R3 . But it is unlikely that

the additional exceptional values in (6.125) or in (6.126) are naturally related to the
different topologies of @M3 , @B3 or @. The situation is especially curious if p D 2.
Then it is unclear (by our method) whether the very classical Sobolev spaces

W2k .B3 /; k 2 N; (6.127)

have oscillating u-Riesz bases.

 6.2 Existence and non-existence of wavelet frames and bases 199

6.2 Existence and non-existence of wavelet frames and bases

s
6.2.1 The role of duality, the spaces Bpq .Rn /

Recall that fbj gj1D1  B in a separable complex quasi-Banach space B is called a

basis if any b 2 B can be uniquely represented as
1
X
bD j bj ; j 2 C (convergence in B): (6.128)
j D1

Of course j .b/ is linear in b. A basis is called a Schauder basis if j .  / 2 B 0 are

linear and continuous functionals on B. It is one of the very first observations of
Banach space theory by Banach himself, [Ban32], that any basis in a Banach space is a
Schauder basis, an early application of the Closed Graph Theorem. Recent proofs may
be found in [AlK06], Theorem 1.1.3, p. 3, and [Woj91], Corollary, p. 38. However this
observation can be extended to quasi-Banach spaces.
Proposition 6.35. Any basis in a separable complex quasi-Banach space B is a
Schauder basis.
Proof. For any quasi-metric % on a set X there is a number "0 with 0 < "0  1 such that
%" for " with 0 < "  "0 is equivalent to a metric. A proof may be found in [Hei01],
Proposition 14.5. One may also consult [T06], Section 1.17.4. As a consequence, any
quasi-Banach is a Fréchet space (also called an F -space, but should not to be mixed
s
with our occasional abbreviation of Fpq -spaces as F -spaces) for which the Closed
Graph Theorem remains to be valid, [Woj91], pp. 3–4. One may also consult [Yos80],
I.9, II.6, pp. 52, 79. Now one can carry over the proof in [AlK06], Theorem 1.1.3,
pp. 3, 4, from Banach spaces to quasi-Banach spaces.
Remark 6.36. Inserting b D bk in (6.128) the uniqueness of the representation ensures
the existence of a dual system fbj0 gj1D1  B 0 with
(
0
  1 if j D k;
bj bk D ıj k D (6.129)
0 if j 6D k.
Hence the existence of a basis in a quasi-Banach space B requires that it has a suf-
ficiently rich dual B 0 . This is always the case for all spaces Apq
s
.Rn / and Apqs
./
considered in the preceding chapters. But it is not the case, as we shall see, for other
function spaces of interest. Even worse, one has the same negative outcome if one
asks for frames instead of bases. Similarly as in Remark 5.11 we call fbj gj1D1  B
a (stable) frame if there is a quasi-Banach space ƒ of fj 2 C W j 2 Ng (sequence
space) with the following properties:
1. (Stability) Any b 2 B can be represented as
1
X
bD j bj ;  D fj g 2 ƒ (convergence in B); (6.130)
j D1
200 6 Complements

with
kb jBk  inf k jƒk (6.131)
where the infimum is taken over all representations (6.130).
2. (Optimality) There is a sequence  D fj g  B 0 with
1
X
bD j .b/ bj ; k.b/ jƒk  kb jBk; (6.132)
j D1

convergence in B.
Let m
h
f be the differences in Rn according to (1.21). Let
˚
s
Bpq .Rn / D f 2 Lp .Rn / W kf jBpq s
.Rn /km < 1 (6.133)

where
0 < p < 1; 0 < q < 1; 0 < s < m 2 N; (6.134)
and
Z 1=q
dh
kf jBpq
s
.Rn /km D kf jLp .Rn /k C jhjsq km
h f jLp .R /k
n q
:
Rn jhjn
(6.135)
We dealt in [T06], Chapter 9, with these spaces, where one finds also references to the
original papers. This will not be repeated here. One can replace the integration over
h 2 Rn in the second term on the right-hand side of (6.135) by jhj  1. Then one
obtains (1.23), and by (1.25) that
 
Bpqs
.Rn / D Bpq
s
.Rn /; 0 < p; q < 1; s > p D n p1  1 C ; (6.136)

(appropriately interpreted as subspace of S 0 .Rn / and of Lloc n

1 .R /, justified by the em-
bedding Bpq .R /  L1 .R //. By Theorem 1.20 all spaces covered by (6.136) have
s n loc n

wavelet bases. If 0 < s < p then the spaces Bpq s

.Rn / and Bpq
s
.Rn / do not coincide
0
any longer (they are even not comparable). As usual, B is the dual of the quasi-Banach
space B.
Theorem 6.37. Let
 
1
0 < p < 1; 0 < q < 1; 0 < s < p D n 1 : (6.137)
p
Then
s
Bpq .Rn /0 D f0g (6.138)
for the dual space of Bpq
s
.Rn /. Furthermore,

Lp .Rn /0 D f0g: (6.139)

 6.2 Existence and non-existence of wavelet frames and bases 201
 
Proof. Let ' 2 D.Rn / and 'j .x/ D ' 2j x where j 2 N. Then one obtains by
(6.135) that
k'j jBpq
s
.Rn /km
Z 1=q
j n n
j.s p / sq dh
D2 p k' jLp .R /k C 2
n
jhj km
h' jLp .R /k
n q
(6.140)
Rn jhjn
n
 2j.s p / k' jBpq
s
.Rn /km :
We assume that ' 2 D.Rn /, say, with '  0, generates a resolution of unity,
X
1D '.x  k/; x 2 Rn : (6.141)
k2Zn

Let g 2 Bpq
s
.Rn /0 and 2 D.Rn /. Then
X
jg. /j  jg. '.2j  k//j
jkjK
X
 kgk k '.2j  k/ jBpq
s
.Rn /km (6.142)
jkjK
n n
 c 2j n 2j.s p / D c 2j.sCn p / ;
where we used K  2j n , (6.140) and that is a pointwise multiplier in Bpq
s
.Rn /. This
can be checked by direct calculation or by real interpolation between Lp .Rn / and the

spaces Bpq .Rn / with  > p as indicated in [T01], pp. 373–74. The constant c in
(6.142) depends on g and , but not on j . Since s < pn  n one obtains by j ! 1
that g. / D 0. Again by real interpolation as indicated above or by direct calculation
it follows that D.Rn / is dense in Bpq
s
.Rn /. Hence g D 0. This proves (6.138). One
obtains (6.139) by the same arguments.
Corollary 6.38. Neither the spaces Lp .Rn / nor the spaces Bpq
s
.Rn / with (6.137) have
bases or frames with (6.130)–(6.132).
Proof. This follows from Theorem 6.37, Proposition 6.35 and Remark 6.36.
Remark 6.39. In [T06], Chapter 9, we dealt also with spaces Fpq
s
.Rn / and obtained
under some restrictions of the parameters p and q,
s
Bp;min.p;q/ .Rn / ,! Fpq
s
.Rn / ,! Bp;max.p;q/
s
.Rn / (6.143)
s
similarly as for the spaces Bpq .Rn / and Fpq
s
.Rn /. Then one obtains from (6.138) and
(6.143) that
Fpqs
.Rn /0 D f0g: (6.144)
In particular these spaces do not have bases or frames with (6.130)–(6.132). However
one has a remarkable substitute. There are constructive building blocks, called quarks,
both for Bpq
s
.Rn / and Fpq
s
.Rn / with expansions of type (6.130), (6.131). We refer for
details and further information to [T06], Section 9.2, and the recent papers [HaS08],
[Sch08], [Sch09].
202 6 Complements

6.2.2 The non-existence of Riesz frames in exceptional spaces

Let  be a bounded Lipschitz domain in Rn with n  2 according to Definition 3.4 (iii)

or a bounded interval in R if n D 1. Let
˚
ˆ D ˆjl W j 2 N0 I l D 1; : : : ; Nj  C u ./ (6.145)
be an [oscillating]finteriorg u-wavelet system as introduced in Definition 6.3. In Defi-
nition 6.5 we said what is meant by related [oscillating]finteriorg u-Riesz bases. Let
ı
AOpq
s s
./ be either Apq ./ or AQpq
s s
./ according to Definition 2.1, or Apq ./ as in Defi-
s 
nition 5.17 (i). Let apq .Z / be as in (6.6) and Definition 6.1. In good agreement with
the comments in Remarks 6.6 and 6.36 we call the [oscillating]finteriorg u-wavelet
system ˆ in (6.145) an [oscillating]finteriorg u-Riesz frame for
AOpq
s
./ with 0 < p < 1; 0 < q < 1; s 2 R;
if it has the following properties:
1. An element f 2 D 0 ./ belongs to AOpq
s
./ if, and only if, it can be represented
as
X1 X Nj
f D jl 2j n=2 ˆjl ;  2 apq
s
.Z /; (6.146)
j D0 lD1

unconditional convergence being in AOpq

s
./, and

kf jAOpq
s
./k  inf k japq
s
.Z /k; (6.147)
where the infimum is taken over all admissible representations (6.146).
2. Any f 2 AOpq
s
./ can be represented as
1 X
X Nj
f D jl .f / 2j n=2 ˆjl (6.148)
j D0 lD1

where jl .  / 2 AOpq

s
./0 are linear and continuous functionals on AOpq
s
./ and

kf jAOpq
s
./k  k.f / japq
s
.Z /k: (6.149)

This is the direct counterpart of Definition 6.5. Theorem 5.27 ensures the existence of
s
oscillating u-Riesz frames in some spaces Apq ./ excluding the exceptional spaces
with s  p 2 N0 . In this Section 6.2.2 and in the following Section 6.2.3 we have a
1

closer look at these exceptional spaces.

A bounded Lipschitz domain  is E-thick. This follows from Definitions 3.1, 3.4
and Proposition 3.8. Then one has by Theorem 3.13 that the spaces
(
p ; B-spaces;
AQpq ./; 0 < p; q < 1; s >
s
(6.150)
pq ; F -spaces;
 6.2 Existence and non-existence of wavelet frames and bases 203

have interior u-Riesz bases according to Definition 6.5 and hence also interior u-Riesz
s s
frames with u > s. The isomorphic map onto apq .Z /, now identified with apq .Z /,
in Theorem 3.13 and a subsequent mollification show that D./ is dense in the spaces
AQpq
s
./ with (6.150). (Of course, extending Definition 6.5 and the above definition to
u-Riesz frames one has the same assertions for E-thick domains.)
Proposition 6.40. Let  be a bounded Lipschitz domain in Rn with n  2 according
to Definition 3.4 (iii) or a bounded interval in R if n D 1. Let
(
p ; B-spaces;
0 < p; q < 1; s > (6.151)
pq ; F -spaces:
ı
Let s < u 2 N. Then Apq
s
./ has an interior u-Riesz frame if, and only if,
ı
s
Apq ./ D AQpq
s
./: (6.152)
ı
Proof. By the above considerations it follows from (6.152) that Apq
s
./ has an interior
ı
s
u-Riesz basis and hence an interior u-Riesz frame. Conversely, if Apq ./ has an
interior u-Riesz frame then any f 2 D./ can be represented by (6.148), (6.149).
Extended to Rn by zero outside of  one has an atomic decomposition of f 2 D.Rn /
s
in Apq .Rn / according to Theorem 1.7 (no moment conditions are needed). Hence,

kf jAQpq
s
./k D kf jApq
s
.Rn /k  c kf jApq
s
./k: (6.153)
ı
s
The converse of (6.153) follows from (2.5). Since D./ is dense both in Apq ./ and
Qs
Apq ./ one obtains (6.152) from

kf jAQpq
s
./k  kf jApq
s
./k; f 2 D./: (6.154)

Remark 6.41. For cellular domains  (and hence for bounded C 1 domains) we have
so far Propositions 6.13, 6.15, ensuring (6.152) under the given restrictions for the
parameters p, q, s. According to Remark 5.20 and the references given there one has
for bounded C 1 domains ,
ı
s
Apq ./ D AQpq
s
./ if 0 < p; q < 1; s > p ; s  1
p
62 N0 : (6.155)
ı
Hence by the remarks after (6.150) the spaces Apqs
./ with (6.151) and s  1
p
62 N0
have interior u-Riesz frames (even interior u-Riesz bases).
Theorem 6.42. Let  be a bounded C 1 domain in Rn with n  2 according to
ı
Definition 3.4 (iii) or a bounded interval in R if n D 1. Let Apq
s
./ be the spaces as
introduced in Definition 5.17 (i). Let
1
0 < p; q < 1; s D r 2 N0 and s < u 2 N: (6.156)
p
204 6 Complements

Then neither ı
s
Bpq ./ with min.1; p/ < q < 1; s > p ; (6.157)
nor ı
s
Fpq ./ with p > 1; 0 < q < 1; s > q : (6.158)
have an interior u-Riesz frame.
Proof. The case r D 0 follows from (5.122), (5.126) and Proposition 6.40. The
extension of this assertion to r 2 N is covered by (5.128).
Remark 6.43. The above theorem makes clear that the restrictions for s in Theo-
rems 5.21, 5.27 are natural.

6.2.3 Reinforced spaces

Let  be a bounded C 1 domain in Rn . Then one has u-Riesz frames (D u-wavelet

s
frames) in the spaces Apq ./ covered by Theorem 5.27. This excludes spaces with
s  p 2 N0 . Theorem 6.42 shows that the situation in these exceptional spaces is
1

rather peculiar. Let

d.x/ D dist.x; / D inffjx  j W 2 g; x 2 ; (6.159)

be the distance to the boundary D @. Then one has by (5.122), (5.126) and [T01],
Theorem 5.10. p. 54, for 1 < p < 1, 1  q < 1,
ı ˚
Fpq1=p
./ D Fpq1=p
./ 6D Fzpq
1=p
./ D f 2 Fpq 1=p
./ W d 1=p f 2 Lp ./ :
(6.160)
1=p 1=p
Although D./ is dense in Fpq ./ one obtains by Theorem 6.42 that Fpq ./ has
no interior u-Riesz frame or interior u-Riesz basis. The Hardy inequality
Z
dx
jf .x/jp  c kf jFzpq
1=p
./kp ; f 2 Fzpq1=p
./; (6.161)
 d.x/
is a special case of (5.124). The theory of envelopes as developed in [T01], Chapter II,
1=p
and in [Har07] produces the following sharp Hardy inequality for the spaces Fpq ./.
Let again 1 < p < 1, 1  q < 1, and let ~.t/ with t  0 be a positive monotonically
decreasing function. Then one has for some c > 0,
Z ˇ   ˇ
ˇ ~ d.x/ f .x/ ˇp dx  p
ˇ
ˇ
ˇ
ˇ  c f jFpq1=p
./ ; f 2 Fpq 1=p
./; (6.162)
 1 C j log d.x/j d.x/
if, and only if, ~ is bounded. We refer to [T01], Corollary 16.7, Remark 16.8, p. 242.
Hence one obtains the sharp Hardy inequality
Z ˇ ˇp
ˇ f .x/ ˇ dx  p
ˇ
ˇ
ˇ
ˇ  c f jFpq
1=p
./ ; f 2 Fpq 1=p
./: (6.163)
 1 C j log d.x/j d.x/
 6.2 Existence and non-existence of wavelet frames and bases 205

The sharp Hardy inequalities (6.161), (6.163) show the difference between these two
spaces. In contrast to this observation one has for 1  p; q < 1, 0 < s < 1=p, that
Z
z dx
Fpq ./ D Fpq ./;
s s
jf .x/jp sp  c kf jFpq
s
./kp (6.164)
 d .x/
for some c > 0 and all f 2 Fpq s
./. These are special cases of Proposition 5.19 and
s
(5.124). This suggests to reinforce the spaces Fpq ./ with s  p1 2 N0 as follows.
Let  D . / with 2 D @ be the outer normal at the boundary of the above
bounded C 1 domain  in Rn (bounded interval if n D 1), naturally extended to

" D fx 2 ; d.x/ < "g; with " > 0 sufficiently small. (6.165)

Definition 6.44. Let  be a bounded C 1 domain in Rn with n  2 according to

Definition 3.4 (iii) or a bounded interval in R if n D 1. Let 1  p; q < 1 and
s > p1  1. Then
˚ s
Fpq ./ if s  1
p
62 N0 ;
s;rinf
Fpq ./ D ˚ @r f
f 2 Fpqs
./ W d 1=p @ r 2 Lp ." / if s  1
p
D r 2 N0 .
(6.166)
s;rinf
Remark 6.45. If s  1
p
D r 2 N0 then Fpq ./ is normed by
Z ˇ r ˇ 
ˇ @ f .x/ ˇp dx 1=p
kf jFpq
s;rinf
./k D kf jFpq
s
./k C ˇ ˇ : (6.167)
ˇ r ˇ
" @ d.x/
r
If r D 0 then @@ fr D f and one obtains by [T01], Theorem 5.10, p. 54, incorporating
p  1 in the right-and side of (6.160),
1=p;rinf
Fpq ./ D Fzpq
1=p
./; 1  p; q < 1: (6.168)

Now one can extend the F -part of Theorem 5.27 to the above spaces with s D r C p1 ,
where we now use the notation of an oscillating u-Riesz frame as introduced at the
beginning of Section 6.2.2.
Theorem 6.46. Let  be a bounded C 1 domain in Rn with n  2 according to
Definition 3.4 (iii) or a bounded interval in R if n D 1. Let r 2 N0 and u 2 N with
r < u. Then there is a common oscillating u-Riesz frame according to (6.146)–(6.149)
s;rinf
for all spaces Fpq ./ with
1 1
1  p < 1; 1  q < 1; 1<sr  : (6.169)
p p

Proof. Step 1. In case of s  p1 62 N0 the above assertion is covered by Theorem 5.27.

s;rinf
It remains to deal with the exceptional spaces Fpq ./ where s D 1
p
C r. Let r D 0
206 6 Complements

and u 2 N. Then one obtains the desired assertion (as in Step 1 of the proof of
Theorem 5.27) from Theorem 3.23 and (6.168).
Step 2. Let r 2 N. We wish to show that the common u-Riesz frame for the
s;rinf
F -spaces with (5.142) is also a u-Riesz frame for Fpq ./ with s D p1 C r. We
begin with a preparation. We claim that
˚
Fzpq
s
./ D f 2 Fpq s;rinf
./ W tr r1
 f D0 ; (6.170)

where tr r1
 f has the same meaning as in (5.59) with r  1 in place of r. Let f 2
Fzpq
s
./. Since D./ is dense in Fzpq
s
./ and

Fzpq
s s
./ ,! Fpq ./ (6.171)

one obtains by (5.62), (5.63) that tr r1

 f D 0. By Definition 2.1 (ii) one has

D ˛ f 2 Fzpq
1=p
./ if j˛j D r and f 2 Fzpq
s
./: (6.172)

By (6.166) and (6.160) it follows that f 2 Fzpqs

./ with s D p1 C r is an element of
the right-hand side of (6.170). We prove the converse. Let n D 1, hence  D I is a
bounded interval, say, I D .0; 1/. It is sufficient to look what happens at x D 0. If
f is an element of the spaces on the right-hand side of (6.170), say with f .x/ D 0 if
x > 1=2, then
(
f .l/ .0/ D 0 if l D 0; : : : ; r  1;
f 2 Fpqs
.I /; R 1 .l/ (6.173)
0 jf .x/j x < 1 if l D r:
p dx

We extend f by zero to .1; 0/. This is also an extension of the distributional

derivatives f .l/ with l D 0; : : : ; r. By the last equality in (6.160) (which applies also
to p D 1, [T01], p. 54) it follows from (6.173) that f 2 Fzpq s
.I /. This proves (6.170)
in case of n D 1. Let n  2. Localisation and diffeomorphic maps show that we may
assume that f is supported near the origin and that a piece of the hyper-plane fx1 D 0g
is part of the boundary , in particular @@ D @x@1 . Then one obtains the desired assertion
from the one-dimensional case and the Fubini property for the F -spaces according to
[T01], Theorem 4.4, p. 36.
Step 3. After this preparation we extend now the construction in Theorem 5.27 for
s;rinf
p, q, s with (5.142) to Fpq ./ with s D p1 C r where r 2 N. Let Ext r1;u  be the
same extension operator as in (5.152). Then one has by Theorem 5.14
Y
r1
extr1;u
 D re B Ext r1;u
 W rk
Bpp s
. / ,! Fpq ./: (6.174)
kD0

Let g be as in (5.152) with the building blocks ˆj;k

l
. / in (5.153). From the explicit
j;k
form of ˆl . / and (5.72) it follows that
@r j;k
ˆ . / D 0 near I k D 0; : : : ; r  1: (6.175)
@ nr l
 6.2 Existence and non-existence of wavelet frames and bases 207

@r 1=p
In particular, @nr
g is an atomic decomposition in Fpq .Rn /. Hence

@r g
2 Fzpq
1=p
./: (6.176)
@ r
Using once more the right-hand side of (6.160) (again extended to p D 1) one obtains
by (6.167) and (6.174) that

Y
r1
extr1;u
 W rk
Bpp s;rinf
. / ,! Fpq ./: (6.177)
kD0

This is the decisive observation. Then (5.62) can be strengthened by

Y
r1
tr r1

s;rinf
Fpq ./ D rk
Bpp . /: (6.178)
kD0

But now one is in the same position as in Corollary 5.16 and (5.113) (with r  1 in
place of r). In particular, one obtains by (6.170) that

Qr1;u Fpq
s;rinf
./ D Fzpq
s
./: (6.179)

As a counterpart of (5.117) one has the decomposition

Y
r1
s;rinf
Fpq ./ D Fzpq
s
./  ext r1;u

rk
Bpp . /: (6.180)
kD0

Then one obtains the above theorem by the same arguments as in the proof of Theo-
rem 5.27.
Remark 6.47. We removed the restrictions s  p1 62 N0 in Theorem 5.27 for the
F -spaces at the expense of the reinforced spaces in (6.166). Recall that this covers in
particular the reinforced Sobolev spaces
s;rinf
Hps;rinf ./ D Fp;2 ./; 1 < p < 1; s  1
p
2 N0 : (6.181)

One may ask for u-Riesz bases according to Definition 6.5 naturally extended to
s;rinf
Fpq ./. One can again rely on the obviously modified Proposition 5.34.
Corollary 6.48. Let  be either a bounded interval on R .n D 1/ or a bounded planar
C 1 domain in R2 .n D 2/. Let r 2 N0 and u 2 N with r < u. Then there is a
s;rinf
common oscillating u-Riesz basis for all spaces Fpq ./ with
1 1
1  p < 1; 1  q < 1; 1<sr  : (6.182)
p p
Proof. One can use the same arguments as in the proof of Theorem 5.35 based on
(6.180).
208 6 Complements

6.2.4 A proposal

In Section 5.3 and also in Sections 6.1.6, 6.1.7 we dealt with u-Riesz bases for spaces
in higher dimensions based on induction by dimensions. This produces additional
exceptional values of s as in Theorems 5.38, 6.30, 6.32. We discussed these effects in
Section 5.3.1 and in Remark 6.34. The question arises of whether one can remove these
restrictions by reinforcing the original spaces in a similar way as in (6.167). Now d.x/
may be the distance to some edges, cutting faces or cutting lines of lower dimensions.
A corresponding theory has not yet been worked out. But we illustrate these admittedly
cryptical comments in a somewhat sketchy way by an example. Let

H 1 .Rn / D W21 .Rn / D F2;2

1
.Rn / (6.183)

be the distinguished classical Sobolev space in Rn , normed by

X
kf jH 1 .Rn /k D kD ˛ f jL2 .Rn /k: (6.184)
j˛j1

The space H 1 .R2 / in the plane R2 is exceptional in the context of the theory of envelopes
as developed in [T01], Chapter II, [Har07]. Similarly as in (6.162), (6.163) one has
the following sharp Hardy inequality. Let, as there, ~.t / with t  0 be a positive
monotonically decreasing function. Then one has for some c > 0,
Z ˇ ˇ
ˇ ~.jxj/ f .x/ ˇ2 dx
ˇ ˇ ˇ ˇˇ
ˇ  c kf jH 1 .R2 /k2 (6.185)
ˇ ˇ
jxj1 ˇ 1 C log jxj ˇ jxj
2

if, and only if, ~ is bounded. This is a special case of [T01], Theorem 16.2, p. 237.
On the other hand, the related refined localisation space H 1;rloc .R2 n f0g/ according to
Corollary 2.20 can be equivalently normed by
Z 1=2
dx
kf jH 1;rloc .R2 n f0g/k D kf jH 1 .R2 /k C jf .x/j2 : (6.186)
R2 jxj2
The situation is similar as in (6.162)–(6.164). Let n  2. The unit cube and the unit
ball are denoted by

Qn D fx 2 Rn W 0 < xr < 1g; Bn D fx 2 Rn W jxj < 1g: (6.187)

As in (6.97) we decompose the boundary of the unit square in R2 as

@Q2 D D 1 [ 0; 1 \ 0 D ;; (6.188)

where 0 collects the four corner points and 1 the connecting line segments. Let
temporarily Q D Q2 and let

d0 .x/ D dist.x; 0 /; x 2 Q: (6.189)

 6.2 Existence and non-existence of wavelet frames and bases 209

Then one has the sharp Hardy inequality

Z ˇ ˇ2
ˇ f .x/ ˇ dx
ˇ ˇ
ˇ 1 C j log d .x/j ˇ d 2 .x/  c kf jH .Q/k
1 2
(6.190)
Q 0 0

for the classical space H 1 .Q/, whereas the restriction

H 1 .Q; 0/ D reQ H 1;rloc .R2 n 0/ (6.191)

of the refined localisation space H 1;rloc .R2 n

can be normed by 0/
 X Z 1=2
dx
kf jH .Q; 0 /k D
1
kD f jL2 .Q/k C
˛ 2
jf .x/j2
: (6.192)
Q d0 .x/2
j˛j1

Asking for u-Riesz bases as in Theorem 6.30 it turns out that the desired decoupling
(6.117) for H 1 .Q/ D F2;21
.Q/ does not work. We discussed this point in Remark 5.50
with references to [Gri85], [Gri92]. On the other hand, according to Theorem 2.38 the
space H 1;rloc .R2 n 0 / has u-Riesz bases. By the specific construction one obtains that

tr 1 H 1 .Q; 0/
z 1=2 .
DH 1 /: (6.193)

This results in the decomposition

H 1 .Q; 0/
z 1 .Q/  Ext  H
DH z 1=2 . 1/ (6.194)

as a counterpart of (6.117). Both factors have u-Riesz bases which can be clipped
together as in Theorem 6.30. Then one obtains u-Riesz bases in H 1 .Q; 0 /.
In other words, if one reinforces H 1 .Q/ (satisfying the sharp Hardy inequality
(6.190)) by H 1 .Q; 0 / with (6.192) then one obtains the desired total decou-
pling of the boundary spaces and, as a consequence, u-Riesz bases.
These considerations can be extended to n dimensions, n  3. Let n2 be the .n  2/-
dimensional faces of Qn and let

dn2 .x/ D dist.x; n2 /; x 2 Qn ; (6.195)

be the counterpart of (6.189). Integrating (6.190) with respect to the remaining n  2

dimensions one obtains the sharp Hardy inequality
Z ˇ ˇ2
ˇ f .x/ ˇ dx
ˇ ˇ  c kf jH 1 .Qn /k2 (6.196)
ˇ ˇ
Qn 1 C j log dn2 .x/j
2
dn2 .x/

whereas the corresponding reinforced spaces H 1 .Qn ; n2 / are normed by

 X Z 1=2
dx
kf jH .Q ;
1 n
n2 /k D kD f jL2 .Q /k C
˛ n
jf .x/j 2
2 2
:
Q n dn2 .x/
j˛j1
(6.197)
210 6 Complements

Otherwise one is in the same position as in case of n D 2. One obtains the decompo-
sition
H 1 .Qn ; n2 / D H z 1=2 . n1 /
z 1 .Qn /  Ext  H (6.198)
and u-Riesz bases. This can be extended to other situations. One can replace the cube
Qn by the above unit ball Bn and the faces n2 by the equator

Sn2 D fx 2 Rn W jxj D 1; xn D 0g; n  3; (6.199)

of the sphere @Bn . The reinforced space H 1 .Bn ; Sn2 / is normed by (6.197) with Bn in
place of Qn and dn2 .x/ as the distance of x 2 Bn to Sn2 . There is hardly any doubt
that these sketchy comments can be applied to further spaces and other domains.
One takes sharp Hardy inequalities as they originate from the theory of envelopes
as a guide and complements the .quasi/-norms of .classical/ function spaces by
reinforcing terms of the above type, paying special attention to the distance of
.more or less/ natural cutting faces .lines, edges/. Employing the theory of the
s;rloc
refined localisation spaces Fpq .wavelet bases/ one tries to derive u-Riesz
bases for these reinforced spaces.

6.3 Greedy bases

6.3.1 Definitions and basic assertions

First we recall some notation and well-known assertions. A set fel g1

lD1
in a separa-
ble complex quasi-Banach space B is called a basis if any b 2 B can be uniquely
represented as
1
X
bD l el ; l 2 C (convergence in B): (6.200)
lD1

Of course l .b/ is linear in b. According to Proposition 6.35 any basis is a Schauder

basis. This means that l .  / 2 B 0 are linear and continuous functionals. A basis
fel g1
lD1
is called an unconditional basis if for any rearrangement  of N (one-to-one
map of N onto itself) fe .l/ g1
lD1
is again a basis,
1
X
bD  .l/ e .l/ (convergence in B); (6.201)
lD1

for any b 2 B with (6.200). Standard bases of separable sequence spaces as considered
s
in this book are always unconditional. One may think about the spaces apq .Z /
according to Definition 6.1 with p < 1, q < 1. But one should be aware that `2 and
any other Banach space with a basis has also a basis which is not unconditional. This
 6.3 Greedy bases 211

goes back to [PeS64] and may also be found in [AlK06], Section 9.5, pp. 235–40. We
always assume that fel g1
lD1
is normalised, hence kel jBk  1, which means
0 < inf kel jBk  sup kel jBk < 1: (6.202)
l l

(Recall that fel g1

lD1
is called normed if kel jBk D 1.) For a given basis fel g1
lD1
let
BN be the collection of all b 2 B with at most N 2 N non-vanishing coefficients
l D l .b/ in (6.200). Then
sN .b/ D inffkb  bN jBk W bN 2 BN g; N 2 N; (6.203)
is called the best N -term approximation error (with respect to the given basis fel g1
lD1
).
Let the coefficients l .b/ in (6.200) be ordered by magnitude,
jl1 .b/j  jl2 .b/j      jlr .b/j     : (6.204)
Then
X
N
GN .b/ D lk .b/ elk ; N 2 N; (6.205)
kD1
is the greedy algorithm. One may ask of whether this distinguished (non-linear) ap-
proximation is comparable with sN .b/ according to (6.203).
Definition 6.49. Let fel g1 lD1
be a normalised basis in a separable complex quasi-
Banach space.
(i) Then fel g1
lD1
is said to be greedy if there is a constant C . 1/ such that for any
b 2 B and any N 2 N,
kb  GN .b/ jBk  C sN .b/: (6.206)
(ii) Then fel g1
lD1
is said to be democratic if there is a constant D . 1/ such that
for any K tuplets
m1 < m2 <    < mK and l1 < l2 <    < lK (6.207)
of natural numbers,
XK  XK 
   
 emk jB   D  elk jB : (6.208)
kD1 kD1

Remark 6.50. Here we are interested only in the standard bases of complex quasi-
Banach sequence spaces of type apq s
.Z / and apq s
.Z / with a 2 fb; f g according
to Definitions 2.6 and 6.1 (or its R -counterpart in Definition 1.3). This simplifies
n

the situation. If j 2 N0 has the same meaning as in Definitions 2.6, 6.1 then the
normalised standard basis in these sequence spaces is given by
 n 
el D 0; : : : ; 0; 2j.s p / ; 0; : : : ; l 2 N; (6.209)
appropriately numbered, l D l.j; r/.
212 6 Complements

Theorem 6.51. (i) Let fel g1 lD1

be a normalised basis in a (separable infinite-dimen-
sional) complex Banach space. Then fel g1 lD1
is greedy if, and only if, it is unconditional
and democratic.
(ii) Let fel g1
lD1
be a normalised unconditional basis in a (separable infinite-dimen-
sional) complex quasi-Banach space. Then fel g1 lD1
is greedy if, and only if, it is
democratic.
Proof. Part (i) for real Banach spaces is due to Konyagin and Temlyakov, [KoT99].
This proof may also be found in [AlK06], Theorem 9.6.3, pp. 242–44. The arguments
given there apply also to complex Banach spaces. Presumably this can be extended to
quasi-Banach spaces. But it is not immediately clear whether a greedy basis in a quasi-
Banach space is also unconditional. This is the reason why we assumed in part (ii) that
fel g1
lD1
is unconditional. Otherwise the proof in [AlK06] does not use whether k  k
is a norm or only a quasi-norm (in a complex quasi-Banach space). Then one obtains
part (ii).
Remark 6.52. At the first glance one is a little bit surprised that the positive word
democratic is so closely connected with the somewhat negative word greedy. But this
reflects simply human life as can be seen from the following quotation.
Sir Winston Churchill (in the House of Commons, Nov. 11, 1947, perhaps react-
ing to corruption and other types of greediness in UK): Indeed, it has been said
that democracy is the worst form of government except all those other forms
that have been tried from time to time.

6.3.2 Greedy Riesz bases

s
The bases in spaces of type Apq ./ on bounded Lipschitz domains  considered in
the preceding sections are u-Riesz bases as introduced in Definition 6.5. Then one has
s
the isomorphic map (6.19) onto the sequence spaces apq .Z /. But for the question of
s
whether the standard basis of apq .Z / is greedy or not it does not matter whether one
deals with these sequence spaces or with the spaces apqs
.Z /, a 2 fb; f g, according to
Definition 2.6 incorporating now  D R . The standard basis (6.209) in apq
n s
.Z / with
p < 1, q < 1, is unconditional. Democracy and as a consequence of Theorem 6.51
also greediness of unconditional bases are invariant under isomorphic maps. Hence
the question of whether the u-Riesz bases (u-wavelet bases) in diverse types of spaces
s s
Bpq and Fpq are greedy can be reduced to the problem of whether the standard basis
s s
(6.209) is democratic in the sequence spaces, say, bpq .Z / of fpq .Z / according to
Definition 2.6. Recall that
s
bpp .Z / D fpp
s
.Z /; 0 < p < 1: (6.210)
Proposition 6.53. Let s 2 R, 0 < p < 1, 0 < q < 1.
(i) The standard basis (6.209) is greedy in any fpqs
.Z /.
(ii) The standard basis (6.209) is greedy in bpq .Z / if, and only if, p D q.
s
 6.3 Greedy bases 213

Proof. Step 1. We prove (i). By Theorem 6.51 (ii) and the above comments it is
sufficient to check that the unconditional basis (6.209) is democratic. Let el1 ; : : : ; elK
be K elements (6.209) with lk D lk .jk ; rk / for elk . For fixed j 2 N0 we may assume
that the characteristic functions jr in (2.38) have pairwise disjoint supports. We used
this possibility in this exposition several times with a reference to [T06], Section 1.5.3,
pp. 18–19. Hence for fixed x 2  one obtains that

X
K
n X
K
j2jk .s p / 2jk s jk ;rk .x/jq D ık 2jk nq=p ; x 2 ; (6.211)
kD1 kD1

with ık D 0 or ık D 1. Let jr1 < jr2 <    < jK 0 be the jk s contributing to (6.211)
with ık D 1. Then the right-hand side of (6.211) is equivalent to 2jK 0 nq=p . Taking
the q1 -power it follows that the integrand in (2.38) is equivalent to a function which is
independent of q. One has

XK p  XK p
   
 elk jfpq
s
.Z /   elk jfpp
s
.Z /
kD1 kD1
Z (6.212)
X
K
ˇ ˇp
jk n ˇ
D 2 jk ;rk .x/ˇ dx  K:

kD1

Hence, fel g in (6.209) is a unconditional democratic basis in fpq s

.Z /. Then it follows
from Theorem 6.51 (ii) that the basis fel g is greedy.
Step 2. By (6.210) and part (i) it follows that fel g is greedy in bpp
s
.Z /. It remains
to prove that bpq .Z / with p 6D q is not greedy. First we choose elk with k D 1; : : : ; K
s

such that they have the same j in (6.209). Then one obtains from (2.37) that

XK 
 
 elk jbpq
s
.Z /  K 1=p : (6.213)
kD1

Choosing for elk different levels of j , say jk D k, then one obtains that

XK 
 
 elk jbpq
s
.Z /  K 1=q : (6.214)
kD1

This shows that the basis fel g in (6.209) is not democratic and hence not greedy.

All u-wavelet bases (D u-Riesz bases) constructed in this book in spaces of type
s
Apq with A 2 fB; F g generate isomorphic maps onto corresponding sequence spaces
s
apq with a 2 fb; f g. Then one can apply the above Proposition 6.53. Recall that
s
Fpp D Bpps s
and similarly fpp D bpp
s
. Saying if, and only if, A D F , includes Bpp
s
,
but excludes Bpq with p 6D q.
s
214 6 Complements

Theorem 6.54. (i) Let s 2 R, 0 < p < 1, 0 < q < 1. Then the wavelet bases for
s
Apq .Rn / according to Theorem 1.20 are greedy if, and only if, A D F .
(ii) Let  be an arbitrary domain in Rn with  6D Rn . The wavelet bases for
s;rloc
Fpq ./ with 0 < p < 1, 0 < q < 1, s > pq , according to Theorem 2.38 are
greedy.
(iii) Let  be an arbitrary domain in Rn with  6D Rn . The wavelet bases for
Lp ./ with 1 < p < 1 according to Theorem 2.36 are greedy.
(iv) Let  be a bounded Lipschitz domain in Rn with n  2 according to Defi-
nition 3.4 (iii) or a bounded interval on R if n D 1. Let Axpq s
./ be the spaces in
(3.45), (3.46) with p < 1, q < 1. Then the wavelet bases for Axpq s
./ according to
Theorem 3.23 are greedy if, and only if, A D F .
(v) Let  be either a bounded interval on R .n D 1/ or a bounded planar C 1
domain in R2 .n D 2/. Then the u-wavelet bases for the spaces Apq s
./ according to
Theorem 5.35 are greedy if, and only if, A D F .
(vi) Let  be the special bounded C 1 domain in Rn with n  3 considered
in Theorem 5.38. Then the corresponding u-wavelet bases for the spaces Apq s
./
covered are greedy if, and only if, A D F .
(vii) Let  be a cellular domain in Rn with n  2 according to Definition 6.9
.which includes cubes and bounded C 1 domains/. The u-Riesz bases for the spaces
s
Apq ./ covered by Theorem 6.32 are greedy if, and only if, A D F .

Proof. In all cases one has isomorphic maps of the indicated u-Riesz bases, u-wavelet
s
bases, onto spaces of type apq .Z / and their appropriate modifications. Application
of Proposition 6.53 gives the desired result.
˚
Remark 6.55. Let  be an arbitrary domain in Rn with  6D Rn and let ˆjr be the
orthonormal u-wavelet basis in L2 ./ according to Theorem 2.36 where any smooth-
˚
ness u 2 N is admitted. By this theorem and part (iii) of the above theorem, ˆjr is
also a greedy basis in Lp ./ where 1 < p < 1. At least for Lp -spaces it is reasonable
to ask for greedy bases of smoothness u D 0. In Section 2.5.1 we recalled in (2.163)
the classical Haar system
˚ j
H D HG;m W j 2 N0 ; G 2 G j ; m 2 Zn (6.215)

in Rn and introduced in (2.183) its -adapted counterpart H  .

Theorem 6.56. Let 1 < p < 1 and n 2 N.

(i) The classical Haar system in (6.215) is a greedy basis in Lp .Rn /.
(ii) Let  be an arbitrary domain in Rn with  6D Rn . Then the Haar system H 
according to (2.183) is a greedy basis in Lp ./.

Proof. This follows immediately from the famous Paley–Littlewood assertion in Theo-
rem 2.40 (i) for the spaces Lp .Rn /, its -counterpart in Theorem 2.44 and Proposi-
tion 6.53 (adapted to the above situation).
 6.4 Dichotomy: traces versus density 215

Remark 6.57. If one has a basis in a space of type Apq s

which is isomorphic to the
s
standard basis of a related sequence space apq then one obtains by Proposition 6.53
a greedy basis if, and only if, a D f . Theorems 6.54 and 6.56 are examples. But
s
this applies also to the weighted spaces Apq .Rn ; w/ according to Theorem 1.26 and
s n
the periodic spaces Apq .T / in Theorem 1.37. One can extend these assertions to
s;˛
anisotropic spaces Apq .Rn /, where corresponding wavelet expansions may be found
in [T06], Section 5.2.2, pp. 252–54.

Remark 6.58. Greedy bases came up at the end of the 1990s by the work of Temlyakov,
where [KoT99] is a nowadays often quoted paper. We refer to [AlK06], Section 9.6,
and the survey [Woj03] for further information and references. The considerations are
mostly restricted to real Banach spaces. But it is also mentioned that the assertions can
be extended to complex Banach spaces and, more generally, to complex quasi-Banach
spaces, as we did in Theorem 6.51. Here we are mainly interested in the standard
s
basis (6.209) in spaces of type apq .Z /. Then everything can be done directly. But
assertions of this type for sequence spaces of f -type and b-type are not new. We refer
in this context also to [GaH04], Theorem 2.1, dealing with anisotropic (not necessarily
diagonal) sequence spaces. The first observation about the greediness of Haar bases in
Lp -spaces with 1 < p < 1 in Rn and on cubes goes back to [Tem98]. The search for
greedy bases in L1 , Lorentz spaces Lp;q , Orlicz spaces Lˆ and other rearrangement
invariant spaces seems to be a somewhat tricky business. It is subject to several papers.
We mention [GHM07], [Woj06] and the references within. Greedy bases in so-called
˛-modulation spaces with brushlets in place of wavelets have been studied in [BoN06a],
[BoN06b].

6.4 Dichotomy: traces versus density

6.4.1 Preliminaries

Traces of spaces Apqs

.Rn / with 2  n 2 N, s > 0 and 1  p; q  1 on hyper-planes
R with n > d 2 N or on boundaries
d
D @ of smooth domains  have been
considered since a long time. They played also a role in this exposition. One may
consult Sections 5.1.1, 5.1.3 and also Section 6.1. There is always the question about
the alternative (dichotomy)

• either D.Rn n / is dense in Apq

s
.Rn /,
s
• or Apq .Rn / has a trace on as in (5.5).

One has to say what is meant by traces. If D @ is the boundary of a bounded C 1

domain furnished naturally with the .n1/-dimensional Hausdorff measure D Hn1
then the question (5.5) is reasonable, especially if s; p; q are restricted as in (5.3).
However if D Hn1 C ı, where ı is the Dirac measure at 0 2 , then the situation
216 6 Complements

s
is different. For pointwise traces of Apq .Rn / one needs s  n=p, whereas the density
requires s  1=p. There is a gap and the dichotomy we are asking for cannot be
expected. This suggests to restrict the above question to isotropic Radon measures.
We assume that the reader is familiar with basic measure and integration theory.
Short descriptions of what is needed may be found in [Mat95], pp. 7–13, [Fal85],
pp. 1–6, [T97], pp. 1–2, or [T06], Section 1.12.2, pp. 80–81. With exception of the
Lebesgue measure in Rd , considered as a hyper-plane in Rn , we always assume that
is a Radon measure in Rn with

0 < .Rn / < 1 and supp D compact: (6.216)

Recall that there is an one-to-one relation between these Radon measures and the
tempered distribution T ,
Z
T W ' 7! '.x/ .dx/; ' 2 S.Rn /; (6.217)
Rn

generated by . This justifies to identify with T and to write 2 S 0 .Rn /. It is

essentially the famous Riesz representation of continuous functionals on C0 .Rn / (the
collection of all continuous functions on Rn tending to zero at infinity). Details may be
found in [T06], Proposition 1.123, pp. 80–81. There is an interesting recent extension
to unbounded tempered Radon measures, [Kab08]. But we stick here at (6.216). Let
Lr . ; / with 0 < r < 1 be the usual complex quasi-Banach spaces, quasi-normed
by
Z 1=r Z 1=r
kg jLr . ; /k D jg. /jr .d / D jg. /jr .d / : (6.218)
Rn 

If r  1 then g 2 Lr . ; / generates as in (6.217) with the complex Radon measure

g in place of a tempered distribution and one has again an one-to-one relation
between
g 2 Lr . ; / and g 2 S 0 .Rn /: (6.219)
However if 0 < r < 1 then this is no longer the case. Just this point is responsible for
some curious effects which will be discussed later on. By the above comments it is
reasonable to assume that the measure with (6.216) is isotropic. As before a ball in
Rn centred at x 2 Rn and of radius % is denoted by B.x; %/. A non-negative function
h on the unit interval Œ0; 1 is called strictly increasing if h.t1 / > h.t2 / for t1 > t2 .
Then a Radon measure with (6.216) is called isotropic if there is a continuous strictly
increasing function h on the interval Œ0; 1 with h.0/ D 0 and
 
B. ; %/  h.%/ where 2 and 0 < % < 1; (6.220)

(the equivalence constants are independent of and %). We refer to [T06], Sec-
tion 1.15.1, pp. 95–97, for details and, in particular, for a criterion which functions
h generate a measure with (6.220). It goes back to [Bri03], Theorem 2.7. One
 6.4 Dichotomy: traces versus density 217

may also consult [Bri04]. It might be of interest to study dichotomy in the context of
measures with (6.216), (6.220). But we restrict ourselves to the most distinguished
isotropic measures. These are the so-called d -sets as supports of measures with
(6.216) and  
B. ; %/  %d ; 2 ; 0 < % < 1; 0 < d < n: (6.221)
It is well known that for given and given d a measure with (6.216), (6.221) is
equivalent to the restriction Hd of the Hausdorff measure H d in Rn to . A short
proof may be found in [T97], Theorem 3.4, pp. 5–6. This justifies to speak simply
about d -sets and to simplify in this case Lr . ; / by Lr . / and (6.218) by
Z 1=r
kg jLr . /k D jg. /jr .d / ;  Hd : (6.222)


This coincides also with our previous notation in (5.4), (5.18). If Lr . / or, more
general, Lr . ; / with (6.216), (6.218) is chosen as a target space for some f 2
s
Apq .Rn / one may ask to which extent the trace (if exists) depends on the chosen
source and target spaces. Roughly speaking, it should depend only on f . This is also
largely the case and will be justified below. But for this purpose it is desirable to have
a constructive procedure for the well-known assertion that D.Rn / and S.Rn / are dense
s
in Apq .Rn / with p < 1, q < 1. The spaces Apq s
.Rn / have the same meaning as in
(1.95) and Definition 1.1.
Proposition 6.59. There are linear operators
Ij;l W S 0 .Rn / ,! S.Rn /; j 2 N; l 2 N; (6.223)
such that for all f 2 Apq
s
.Rn / with s 2 R, 0 < p < 1, 0 < q < 1, and " > 0,
kf  Ij;l f jApq
s
.Rn /k  " if j  j" .f; Apq
s
/; l  l" .f; j; Apq
s
/: (6.224)
Proof. Let '0 be the same standard function as in (1.5) and let f 2 Apq
s
.Rn / with
p < 1, q < 1. Then one has
 _
fj D '0 .2j  /fO ! f in Apq s
.Rn / if j ! 1 (6.225)
as a consequence of Definition 1.1, Fourier multiplier assertions and, in case of F -
spaces, Lebesgue’s dominated convergence theorem, [Mall95], p. 37. Recall that fj is
an entire analytic function. Let 2 S.Rn / with .0/ D 1 such that y has compact
support. Then
˚
fjl D .2l  /fj 2 S.Rn / and supp fyjl  y 2 Rn W jyj  c 2j (6.226)

for some c > 0, all j 2 N and all l 2 N. Here fjl 2 S.Rn / follows from the Paley–
Wiener–Schwartz theorem, [T83], p. 13, whereas the second assertion in (6.226) is a
consequence of the convolution

fyjl D 2nl y .2l  / ? fyj :

218 6 Complements

One obtains again by Lebesgue’s dominated convergence theorem that

 
kfjl  fj jApq
s
.Rn /k  cj k 1  .2l  / fj jLp .Rn /k ! 0 (6.227)

and hence
fjl ! fj in Apq
s
.Rn / if l ! 1: (6.228)
With Ij;l f D fjl one obtains (6.224).
Remark 6.60. It is known since a long time that S.Rn / and also D.Rn / are dense in
s
Apq .Rn / with p < 1, q < 1, [T83], Theorem, p. 48. But it will be helpful for us to
have the explicit universal approximating sequence fjl of f .

6.4.2 Traces

Although we are mainly interested in d -sets it is reasonable to deal first with Radon
measures in Rn according to (6.216) and the related spaces Lr . ; / quasi-normed
s
by (6.218). Let Apq .Rn / with A 2 fB; F g be the spaces introduced in Definition 1.1.
Definition 6.61. Let be a Radon measure in Rn with (6.216). Let 0 < r < 1,

0 < p < 1; 0 < q < 1; s 2 R: (6.229)

Let for some c > 0,

k' jLr . ; /k  c k' jApq

s
.Rn /k for all ' 2 S.Rn /: (6.230)

Then the trace operator tr ,

tr W Apq
s
.Rn / ,! Lr . ; /; (6.231)

is the completion of the pointwise trace .tr '/. / D '. / with ' 2 S.Rn / and 2 .
Remark 6.62. Since S.Rn / is dense in Apq s
.Rn / it follows by standard arguments that
this definition makes sense and that the outcome

tr f 2 Lr . ; /; f 2 Apq
s
.Rn /; (6.232)

is independent of the chosen approximating sequence

'j ! f in Apq
s
.Rn /; 'j 2 S.Rn /; (6.233)

on the source side. On the target side one has

tr 'j ! tr f in Lr . ; /: (6.234)

By standard arguments of measure theory, see e.g. [Mall95], Theorem 5.2.7, p. 23,
Lemma 9.3.3, p. 51, it follows that there is a subsequence of f'j g which converges
 6.4 Dichotomy: traces versus density 219

-a.e. to tr f . Identifying this subsequence with f'j g we may assume in addition to

(6.234) that
'j . / D .tr 'j /. / ! .tr f /. / -a.e. on : (6.235)
One obtains that tr f is -a.e. independent of r as long as one has (6.230). This
applies in particular to 0 < rQ  r with (6.230) as a consequence of . / < 1 and
Hölder’s inequality. There is a similar situation on the source side assuming that one
has (6.230) for Aps11 q1 .Rn / and Aps22 q2 .Rn / instead of Apqs
.Rn / on the right-hand side
and that
f 2 Aps11 q1 .Rn / \ Aps22 q2 .Rn /: (6.236)
Then one can rely on the distinguished approximation Ij;l f of f according to Propo-
sition 6.59 where one may assume that it applies to both spaces in (6.236). Hence one
has a common approximating sequence, and tr f is -a.e. the same for both spaces
in (6.236). If (6.230) holds both for Lr1 . 1 ; 1 / and Lr2 . 2 ; 2 / then it is also valid
for Lr . ; / with D 1 C 2 and r D min.r1 ; r2 /. By the above considerations,
tr 1 and tr 2 are restrictions of tr . In other words,
for individual elements f the traces are independent of the source spaces and
of the target spaces as long as one has (6.230) and whenever comparison makes
sense.
In (5.6) we dealt with the more specific trace spaces Lr . /, r  1. Then one has (5.8).
But this requires the necessary restriction
 
1
s
Apq .Rn /  Lloc n
1 .R /; hence s  p D n 1 ; (6.237)
p C

which does not fit in our concept. As will be seen, the above definitions apply even
to some singular distributions, for example the ı-distribution, for which (5.8) does not
make any sense.

At the beginning of Section 6.4.1 we have given a first rough description of what
is meant by dichotomy. Now it is clear that traces must be understood according to
Definition 6.61. One has the following almost obvious observation. If is a compact
set in Rn or the hyper-plane Rd in Rn then we abbreviate now

D D D.Rn n /: (6.238)

Proposition 6.63. Let Apq s

.Rn / and Lr . ; / be as in Definition 6.61. If D is dense
in Apq .R / then there is no c > 0 with (6.230).
s n

Proof. We assume that there is a constant c > 0 with (6.230). We approximate a

function ' which is identically 1 near by D -functions. Then one has by Remark 6.62
that tr ' D 0 -a.e. This contradicts . / > 0.

As said at the beginning of Section 6.4.1 it does not make much sense to deal
with the dichotomy problem in the context of general Radon measures according to
220 6 Complements

(6.216). We concentrate on d -sets with (6.221) where D Hd is the restriction

of the Hausdorff measure H d in Rn to . Then we simplify Lr . ; / by Lr . / with
(6.222). Similarly we put
tr  D tr ;  Hd ; D supp ; (6.239)
in this case, in agreement with our previous notation in (5.5), (5.6). As far as traces on
d -sets are concerned one has the following assertion.
Proposition 6.64. Let be a compact d -set in Rn with 0 < d < n furnished with the
Radon measure D Hd satisfying (6.221). Then
nd
tr  Bpqp .Rn / D Lp . / if 0 < p < 1; 0 < q  min.1; p/; (6.240)
and
nd
tr  Fpqp .Rn / D Lp . / if 0 < p  1; 0 < q < 1: (6.241)
Remark 6.65. Of course, (6.240), (6.241) means that the linear and bounded trace
operator tr  D tr according to Definition 6.61 exists and that it is a map onto Lp . / D
Lp . ; /. These assertions have a substantial history. First we remark that can be
replaced by Rd with n > d 2 N and the d -dimensional Lebesgue measure, interpreted
as d -dimensional hyper-plane in Rn . Then one has in obvious notation,
nd
tr Bpqp .Rn / D Lp .Rd / if 0 < p < 1; 0 < q  min.1; p/; (6.242)
and
nd
tr Fpqp .Rn / D Lp .Rd / if 0 < p  1; 0 < q < 1: (6.243)
The first proof of (6.242) for 1  p < 1 (and d D n1) goes back to [Pee75], [Gol79],
[BuG79]. This has been extended to 0 < p < 1 in [FrJ85]. The F -counterpart,
hence (6.243) (with d D n  1), is due to [Tri89], [FrJ90]. Further details may be
found in [T92], Section 4.4.3, pp. 220–21. The step from Rd to d -sets requires new
technical instruments, especially atomic decompositions. Proposition 6.64 coincides
with [T06], Proposition 1.172. But it goes back to [T97], Corollary 18.12, p. 142,
and the later observation that any d -set with d < n is porous (called ball condition
in [T97]) according to Definition 3.4 (i). We refer to [T01], Remark 9.19, pp. 140–41,
and also to the above Section 3.2.4. However we wish to mention that one finds some
assertions of this type in an earlier paper by Yu. V. Netrusov, [Net90], Corollary, p. 193.

6.4.3 Dichotomy

The above discussions suggest that the dichotomy traces versus density requires iso-
tropic measures with (6.220). We deal here exclusively with d -sets and the related
s
target spaces Lr . / according to (6.221), (6.222). The spaces Apq .Rn / have the same
meaning as above, (1.95) and Definition 1.1. Let D and tr  be as in (6.238) and
(6.239).
 6.4 Dichotomy: traces versus density 221

Definition 6.66. Let n 2 N and 0 < p < 1. Let

˚ s
Ap .Rn / D Apq .Rn / W 0 < q < 1; s 2 R : (6.244)

Let 0 < d < n and let D supp be a compact d -set in Rn with D Hd . Let
 2 R. Then
 
D Ap .Rn /; Lp . / D .; u/ with 0 < u < 1; (6.245)
˚
is called the dichotomy of Ap .Rn /; Lp . / if
(
s > ; 0 < q < 1;
tr  W s
Apq .Rn / ,! Lp . / exists for (6.246)
s D ; 0 < q  u;

and (
s s D ; u < q < 1;
D is dense in Apq .Rn / for (6.247)
s < ; 0 < q < 1:
Furthermore,
 
D Ap .Rn /; Lp . / D .; 0/ (6.248)
means that )
tr  exists for s > , 0 < q < 1;
(6.249)
D is dense in Apq s
.Rn / for s   , 0 < q < 1;

and
 
D Ap .Rn /; Lp . / D .; 1/ (6.250)
means that )
tr  exists for s   , 0 < q < 1;
(6.251)
D is dense in Apq s
.Rn / for s < , 0 < q < 1:

Remark 6.67. Recall that D.Rn / is dense in Apq

s
.Rn / with p < 1, q < 1. Further-
more for fixed A D B or A D F and 0 < p < 1 one has the continuous embeddings
s1
Apq 1
.Rn / ,! Apq
s2
2
.Rn / for s2 < s1 and 0 < q1 ; q2 < 1 (6.252)
and
s
Apq1
.Rn / ,! Apq
s
2
.Rn / for s 2 R and 0 < q1  q2 < 1. (6.253)

Together with Proposition 6.63 it follows that the above definition makes sense. If
one deals with more general isotropic measures (6.220) and with more general source
spaces of B-type and F -type then it might be reasonable to replace q  u in (6.246)
by q < u and to leave open (at least in the definition) what happens at the breaking
point q D u. But in our case one has the additional information that the sharp breaking
point .; u/ is on the trace side.
222 6 Complements

Theorem 6.68. Let n 2 N, 0 < d < n and 0 < p < 1. Let be a compact d -set in
Rn and D Hd . Then
˚ nd

  p
;1 if p > 1;
D Bp .Rn /; Lp . / D  nd  (6.254)
p
;p if p  1;

and ˚ nd

  p
;0 if p > 1;
D Fp .Rn /; Lp . / D  nd  (6.255)
p
;1 if p  1:

Proof. Step 1. In the Steps 3 and 5 below we prove that

nd
D is dense in Bpqp .Rn / if 0 < p < 1; q > min.1; p/: (6.256)

Then (6.254) follows from (6.240) and the comments in Remark 6.67. Since D.Rn / is
s
dense in Fpq .Rn /, q < 1, one obtains by (6.256) that

s
D is dense in Fpq .Rn /; s< nd
p
; 0 < q < 1: (6.257)

Then (6.255) with p  1 is a consequence of (6.241). By (6.240) it follows that tr 

s
exists for all spaces Fpq .Rn / with s > nd
p
. Hence it remains to prove that
nd
D is dense in Fpqp .Rn / if 1 < p < 1; 0 < q < 1: (6.258)

In other words, the theorem follows from (6.256) (subject to Steps 3 and 5) and (6.258)
(subject to Step 4). ˚ 1
Step 2. We begin with a preparation. We wish to construct a sequence ' J J D1 
D.Rn / with
' J .x/ D 1 in an open neighbourhood of (6.259)
(depending on J ) and
nd
' J ! 0 in Bpqp .Rn / if J ! 1; p  1; q > 1: (6.260)

For given j 2 N we cover a neighbourhood of with balls Bj;m centred at and of

radius 2j , where m D 1; : : : ; Mj with Mj  2jd such that there is a resolution of
unity,
Mj
X
'j;m .x/ D 1 near ; 0  'j;m 2 D.Bj;m /; (6.261)
mD1

with the usual properties,

jD  'j;m .x/j  c 2j jj ; 2 Nn0 : (6.262)

 6.4 Dichotomy: traces versus density 223

For 2  J 2 N let J 0 2 N be such that

0 C1
JX
rj D 1 with rj D j 1 if J  j  J 0 ; 0 < rJ 0 C1  .J 0 C 1/1 : (6.263)
j DJ

Then
0 C1
JX Mj
d X jd
' .x/ D
J
rj 2j p 2 p 'j;m .x/; x 2 Rn ; (6.264)
j DJ mD1
s
is an atomic decomposition in Bpq .Rn / according to Theorem 1.7 and Definition 1.5 for
s D ndp
, p  1, q > 1, with L D 0 (no moment conditions). We used s  pn D  dp .
One obtains by Theorem 1.7 that

 q 0 C1
JX X
Mj q=p 1
X
 J nd
 dq
' jBpqp .Rn /  c rjq 2j p 1  c0 j q  J 1q : (6.265)
j DJ mD1 j DJ

This proves (6.259), (6.260).

Step 3. We prove (6.256) for p > 1, q > 1. It is sufficient to approximate
f 2 D.Rn / in Bpq s
.Rn / with s D nd p
by functions f J 2 D . Let ' J be the
functions according to (6.259), (6.260) and let

f D fJ C f J with fJ D ' J f and f J D .1  ' J /f 2 D : (6.266)

nd
By the pointwise multiplier theorem in [T92], Corollary, p. 205, one has for % > p
,
some c > 0, all f 2 D.Rn / and all ' J that

kfJ jBpq
s
.Rn /k  c kf jC % .Rn /k  k' J jBpq
s
.Rn /k ! 0 (6.267)

if J ! 1 where we used (6.260). This proves (6.256) for p > 1, q > 1.

Step 4. We prove (6.258). As mentioned at the end of Remark 6.65 the compact
d -set with d < n is porous. Then one can apply [T06], (9.90), Definition 9.18,
s
pp. 392–93, to the atomic decomposition (6.264) in Fpq .Rn / with s D nd
p
, p > 1,
q  1 (where one may choose L D 0 in Theorem 1.7, no moment conditions) with the
outcome
lim k ' J jFpq
s
.Rn /k D lim k' J jBpp
s
.Rn /k D 0: (6.268)
J !1 J !1

For this type of argument one may also consult [T01], pp. 142–43. If 0 < q < 1 then it
s
may happen that (6.264) is no longer an atomic decomposition in Fpq .Rn / according
to Theorem 1.7 since one may need moment conditions. Because is porous one
can complement 'j;m outside of in an appropriate way such that one has the needed
moment conditions. Details may be found in [T97], p. 143, with a reference to [TrW96].
After this modification one has (6.268) for all s D nd
p
, p > 1, 0 < q < 1. Then one
obtains (6.258) in the same way as in Step 3.
224 6 Complements

Step 5. We prove (6.256) for p < q. This covers in particular the remaining cases
with p  1. We begin with a preparation, covering , say, with . / D 1, for given
L 2 N by d -sets l such that
0 0
[
L X
L
1
D l; . l/  l ; . l /  . / D 1; (6.269)
lDL lDL

where L0 2 N with L0 > L is chosen appropriately. This can be done as follows. For
given l 2 N and appropriately chosen large k 2 N (in dependence on l) one finds
 l 1 2kd balls centred at , of radius  2k and having pairwise distance of at least
 2k , such that the intersection of with the union of these balls is a sub-d -set l of
with . l /  l 1 . Now one can start for given L 2 N with L D L and applies
afterwards the above procedure to n L and l D L C 1. Iteration gives the desired
decomposition. This can be done in such a way that there are functions l 2 D.Rn /,
l  0,

0
X
L

l. / D 1 if 2 ; l  supp l  fy 2 Rn W dist.y; l/ < "l g (6.270)

lDL

for some "l > 0. Let for given l 2 N between L and L0 and appropriately chosen
j.l/ 2 N,
Mj.l/
X
'j.l/;m .x/ D 1 near ; 0  'j.l/;m 2 D.Bj.l/;m /; (6.271)
mD1

as in (6.261), (6.262) with Mj.l/  2j.l/d . Let

j.L/ <    < j.l/ < j.l C 1/ <    < j.L0 / (6.272)

and in analogy to (6.264),

0 Mj.l/
X
L X
 j.l/d j.l/d
' .x/ D
L
l .x/ 2
p 2 p 'j.l/;m .x/; x 2 Rn : (6.273)
lDL mD1

First we assume that nd

p
D s > p . Then one does not need moment conditions in
Theorem 1.7, and (6.273) with large j.l/ is an atomic decomposition which can be
written as
0
Mj.l/
L0
X j.l/d X
' L .x/ D 2 p 'Qj.l/;m .x/; x 2 Rn ; (6.274)
lDL mD1

with
0
Mj.l/  . l / 2j.l/d  l 1 2j.l/d ; (6.275)
 6.4 Dichotomy: traces versus density 225

counting only non-vanishing terms, where the equivalence constants are independent
of L. We have ' L .x/ D 1 near . Then one obtains by Theorem 1.7 (i) for q > p that
0
 q L0
X  MX
j.l/  1
X
 L nd
n  j.l/ dq
q=p q
' jBpq .R /  c
p
2 p 1  c0 l q=p  L1 p :
lDL mD1 lDL
(6.276)
This is the counterpart of (6.265). Now one can argue in the same way as in Step 3. This
proves (6.256) provided that one does not need moment conditions in Theorem 1.7.
But otherwise one can rely on the same arguments and references as in Step 4. This
proves (6.256) also in the remaining cases.
According to (6.242), (6.243) there is a perfect counterpart of Proposition 6.64 with
the hyper-plane Rd , n > d 2 N, in place of . Similarly one can replace the target
space Lp . / in Definition 6.66 by Lp .Rd /. Then one obtains the following assertion.
Corollary 6.69. Let n > d 2 N. Then
˚ nd

  p
;1 if p > 1;
D Bp .Rn /; Lp .Rd / D  nd  (6.277)
p
;p if p  1;

and ˚ nd

  p
;0 if p > 1;
D Fp .Rn /; Lp .Rd / D  nd  (6.278)
p
;1 if p  1:

Proof. This is a by-product of the proof of Theorem 6.68. On the one hand one has
(6.242), (6.243). On the other hand the approximation of g 2 D.Rn / by functions
from D D D.Rn n / now with D Rd is a local matter and covered by the above
arguments.
Remark 6.70. In Remark 6.65 one finds relevant references as far as traces are con-
cerned. Density and trace assertions for the Sobolev spaces Hps .Rn / in (1.17) with
s > 0, 1 < p < 1, and for the classical Besov spaces in (1.23), (1.24) with s > 0,
1 < p < 1, 1  q  1, are known since the late 1960s and early 1970s. Corre-
sponding formulations and detailed references may be found in [T78], Section 2.9.4,
pp. 223–26, with [Tri73] as our own contribution at this time. Density problems and
related approximations for Sobolev spaces Hps .Rn / with respect to arbitrary sets
have been considered in [AdH96], Chapters 9, 10, based on capacity arguments and
the so-called spectral synthesis. This has been extended in [HeN07] to general spaces
of B-type and F -type. We followed here essentially [Tri07d].
226 6 Complements

6.4.4 Negative smoothness

For any 0 < p < 1 and 0 < s < n=p there is a number d with 0 < d < n and
s D ndp
. Hence the corresponding spaces maybe breaking points in Theorem 6.68.
If s < 0 then the situation is different.
Proposition 6.71. Let be a compact set in Rn with Lebesgue measure j j D 0. Then
D D D.Rn n / is dense in all spaces
s
Apq .Rn / with s < 0; 0 < p < 1; 0 < q < 1: (6.279)

Proof. Obviously, D is dense in Lp .Rn / with 1 < p < 1. Since D.Rn / is dense in
s
Apq .Rn / it is sufficient to approximate f 2 D.Rn / in Apq
s
.Rn / by functions belonging
to D . Using

Lp .Rn / ,! Apq
s
.Rn / if s < 0; 1 < p < 1; 0 < q < 1; (6.280)

one obtains the desired assertion from the density of D in Lp .Rn /. For any ball K in
Rn (and s 2 R, 0 < q < 1) there is a number cK > 0 such that

kf jAps 1 ;q .Rn /k  cK kf jAps 2 ;q .Rn /k; 0 < p1  p2 < 1; (6.281)

for all f 2 Aps 2 ;q .Rn / with supp f  K. This well-known assertion follows from
s
Hölder’s inequality and characterisations of the spaces Apq .Rn / in terms of local means,
[T92], Sections 2.4.6, 2.5.3, pp. 122, 138. Then the density assertion for the spaces in
(6.280) can be extended to all spaces in (6.279).

6.4.5 Curiosities

Any Radon measure with (6.216) can be interpreted as a tempered distribution

2 S 0 .Rn /. This applies also to g 2 Lr . ; / with r  1 interpreted as in (6.219).
However if r < 1 then the situation is totally different and the spaces S 0 .Rn / and
Lr . ; / have nothing in common. On the other hand, Definition 6.61 makes sense
and we have Proposition 6.64. But this has some curious consequences. Let be a
compact d -set in Rn with 0 < d < n furnished with the Hausdorff measure D Hd .
Then it follows from Proposition 6.64 that the trace
 
tr  W Apq
s
.Rn / ,! Lp . /; 0 < p < 1; nd p
< s < n p1  1 (6.282)

makes sense, although there are singular distributions which are elements of these
s n. 1 1/
spaces Apq .Rn /; for example, the ı-distribution belongs to Bp;1 p
.Rn / for all
0 < p  1. What does this mean for the trace of these distributions as elements
of Lp . /? In this context it is also reasonable to ask how to understand the traces,
 
tr W Apq
s
.Rn / ,! Lp .Rm /; 0 < p < 1; nm p
< s < n p1  1 ; (6.283)
 6.4 Dichotomy: traces versus density 227

on hyper-planes Rm  Rn with n  m 2 N admitting n D m, where (6.242), (6.243)

with d D m may be considered as limiting cases. According to [T06], pp. 96–97,
Proposition 7.32, and its proof, pp. 315–16 , one has for a compact d 0 -set in Rn with
0
0  d 0 < n and its Hausdorff measure Hd ,

0 .nd 0 /. p
1
1/
D Hd 2 Bp1 .Rn /; j j D 0; 0 < p  1; (6.284)

admitting d 0 D 0 for the ı-distribution D ı with D f0g. In particular one finds in

all spaces with (6.282) and (6.283) singular distributions f with j supp f j D 0.

Proposition 6.72. (i) If s > 0, 0 < p < 1, 0 < q < 1, then

tr W Apq
s
.Rn / ,! Lp .Rn / (6.285)
1 
exists. If, in addition, 0 < p < 1, 0 < s  n p
 1 , then

tr f D 0 in Lp .Rn / for any f 2 Apq

s
.Rn / with j supp f j D 0. (6.286)

0
(ii) Let 0 < p < 1. Let be a compact d -set, D Hd , and be a compact
0
d -set, 0 D Hd0 with
0

0  d 0 < d < n; d  d 0 > p.n  d 0 /: (6.287)

nd
0
Then 2 Bppp .Rn /,
0 0
tr  2 Lp . / exists and tr  D0 (6.288)

in Lp . /.

Proof. Step 1. We prove part (i). The case p  1 is well known. Let 0 < p  1. Then
one has by Definition 1.1 that
X1 p
 
kf jLp .Rn /kp D  .'j fO/_ jLp .Rn /
j D0
1
(6.289)
X
 k.'j fO/_ jLp .Rn /kp  c kf jBpq
s
.Rn /kp :
j D0

One obtains

k' jLp .Rn /k  c k' jApq

s
.Rn /k; s > 0; 0 < p < 1; 0 < q < 1; (6.290)

for all ' 2 S.Rn / as requested in Definition 6.61. This proves (6.285). Let, in addition,

f 2 Apq
s
.Rn / with D supp f compact, j j D 0; (6.291)
228 6 Complements

1 
be non-trivial (hence 0 < p < 1, 0 < s  n p
 1 ) and let

" D fy 2 Rn W dist.y; / < "g (6.292)

s
for " > 0. Then f can be approximated in Apq .Rn / by D.Rn /-functions with support
in " (multiplying an approximating sequence of D.Rn /-functions with a suitable
function from D. " /). This proves tr f  " for any " > 0. Hence tr f  and
tr f D 0 in Lp .Rn /. This can be extended to all f covered by (6.286).
Step 2. We prove part (ii). First we remark that (6.287) and (6.284) ensure that
 
1 0 nd 0
nd
.n  d / 1 > and 2 Bppp .Rn /: (6.293)
p p
0
Then one obtains by Proposition 6.64 that tr  2 Lp . / exists. Let
Mj
[   0
0 0 0 j
 j D B j;m ; 2 ; Mj  2jd ; (6.294)
mD1

0 j 0
where B. j;m ; 2 / are open balls centred at and of radius 2j with j 2 N. One
nd
0
can approximate 0 in Bpp .Rn / by D.Rn /-functions with support in
p
j (similarly as
indicated above). However,
0 0
. j/  c 2jd 2jd ! 0 if j ! 1: (6.295)
0
This proves tr  D 0 in Lp . /.
Remark 6.73. The above proposition shows that (tempered) distributions in Rn on the
one hand and measurable functions on the other hand are rather different worlds and
that great care is required when crossing the border. For the ı-distribution one has
(6.284) with D ı and d 0 D 0. Then one obtains by (6.286) that

tr ı D 0 in Lp .Rn / with 0 < p < 1; (6.296)

what looks somewhat curious.

6.4.6 Pointwise evaluation

In Definition 6.66 we excluded d D 0 and d D n. Whereas d D n does not fit in the

above scheme (with exception of a few curiosities discussed in Section 6.4.5) the case
d D 0 makes sense especially when consists of finitely many points in Rn ,
˚
D x l 2 Rn W l D 1; : : : ; L ; L 2 N; n 2 N: (6.297)

The trace problem reduces to the question of whether the embedding

id W Apq
s
.Rn / ,! C.Rn / (6.298)
 6.4 Dichotomy: traces versus density 229

is continuous. As before, C.Rn / is the space of all complex-valued continuous bounded

functions in Rn , normed by

kf jC.Rn /k D sup jf .x/j; (6.299)

x2Rn

whereas Apq s
.Rn / with A 2 fB; F g has the same meaning as in Definition 1.1. For the
embedding (6.298) one has the following final answer,
(
s n n 0 < p  1; 0 < q  1; s > n=p;
Bpq .R / ,! C.R / if, and only if,
0 < p  1; 0 < q  1; s D n=p;
(6.300)
and
(
s 0 < p < 1; 0 < q  1; s > n=p;
Fpq .Rn / ,! C.Rn / if, and only if,
0 < p  1; 0 < q  1; s D n=p:
(6.301)
This is well known. One may consult [T01], Section 11, and [Har07], Section 7.2, where
one finds also the necessary references. We only mention that the sharp assertion in the
limiting case s D n=p goes back to [SiT95]. This is the counterpart of Proposition 6.64.
The difference between (6.240) and (6.300) is also reflected by what follows. The
counterpart of Definition 6.66 reads now as follows. As before we put D D D.Rn n /.

Definition 6.74. Let n 2 N and 0 < p < 1. Let

˚ s
Ap .Rn / D Apq .Rn / W 0 < q < 1; s 2 R : (6.302)

Let be as in (6.297). Let  2 R. Then

 
D Ap .Rn /; D .; u/ with 0 < u < 1 (6.303)

is called the dichotomy of fAp .Rn /; g if

(
s > ; 0 < q < 1;
id W s
Apq .Rn / ,! C.R /n
for (6.304)
s D ; 0 < q  u;

and (
s s D ; u < q < 1;
D is dense in Apq .Rn / for (6.305)
s < ; 0 < q < 1:
Furthermore,  
D Ap .Rn /; D .; 0/ (6.306)
means that )
id exists for s > , 0 < q < 1;
(6.307)
D is dense in Apqs
.Rn / for s   , 0 < q < 1;
230 6 Complements

and
 
D Ap .Rn /; D .; 1/ (6.308)
means that )
id exists for s   , 0 < q < 1;
(6.309)
D is dense in Apqs
.Rn / for s < , 0 < q < 1:

Remark 6.75. This is the direct counterpart of Definition 6.66. The justification of
the above definition is the same as in Remark 6.67. We are now not only interested in
a counterpart of Theorem 6.68 with the breaking smoothness s D n=p, but also in the
question of whether D is dense in
˚ PL
s;
Apq .Rn / D f 2 Apq
s
.Rn / W lD1 jf .x /j D 0
l
(6.310)
s
for spaces Apq .Rn / satisfying (6.300), (6.301). This cannot happen if one has the
embedding (6.298) with C 1 .Rn / in place of C.Rn /, because one has in this case the
same sharp criterion as in (6.300), (6.301) with pn C1 in place of pn , [T01], Theorem 11.4,
pp. 170–71, [Har07], Section 7.2. We do not bother about the limiting cases s D 1 C pn
s;
and restrict the question of whether D is dense in Apq .Rn / to the spaces (6.300),
(6.301) with s < 1 C p . Recall again that D D D.R n /.
n n

Theorem 6.76. Let n 2 N and

˚
D x l 2 Rn W l D 1; : : : ; L ; L 2 N: (6.311)

(i) Let 0 < p < 1. Then

  n 
D Bp .Rn /; D ;1 (6.312)
p

and ˚ n

  p
;0 if p > 1;
D Fp .Rn /; D n  (6.313)
p
;1 if p  1.
s;
(ii) Let 0 < p < 1, 0 < q < 1 and s < 1 C pn . Then D is dense in Apq .Rn /
with (6.300), (6.301).

Proof. Step 1. It is sufficient to deal with D f0g. We begin with a preparation and
n=p
approximate f 2 D.Rn / in Bpq .Rn / with 0 < p < 1, 1 < q < 1 by functions
f 2 D.R / with f .0/ D 0. Basically we rely on the same construction as in Step 2
J n J

of the proof of Theorem 6.68. For 2  J < 1 let J 0 2 N be such that

0 C1
JX
rj D 1 with rj D j 1 if J  j  J 0 and 0 < rJ 0 C1  .J 0 C 1/1 . (6.314)
j DJ
 6.4 Dichotomy: traces versus density 231

Let 0  2 D.Rn / with

.x/ > 0 if, and only if, jxj  1; .0/ D 1: (6.315)

Let
0 C1
JX
fJ .x/ D rj .2j x/ f .x/; x 2 Rn : (6.316)
j DJ

Then it follows from Theorem 1.7 and Definition 1.5 that

0 C1
JX
kfJ jBpq
n=p
.Rn /kq c j q  J 1q : (6.317)
j DJ

n=p
Since q > 1 one obtains that f J D f fJ approximates f in Bpq .Rn /. Furthermore
n=p
f J .0/ D 0. To justify a corresponding approximation in Fpq .Rn / with p > 1,
0 < q < 1, we recall that one has the embedding

Bpn=p
0u
0
.Rn / ,! Fpq
n=p
.Rn / ,! Bpn=p
1v
1
.Rn /; 0 < p0 < p < p1 < 1; (6.318)

if, and only if, 0 < u  p  v  1. This may be found in [ET96], p. 44, with the
reference to the original papers [Fra86], [Jaw77], [SiT95]. There is also a new proof
based on wavelets in [Vyb08b]. In particular one may choose u D p > 1. Then one
n=p
obtains the desired approximation in Fpq .Rn / with p > 1, 0 < q < 1, from (6.318)
n=p0
and the above approximation in Bp0 p .Rn /. Recall that D.Rn / is dense in all spaces
s
Apq .Rn / with p < 1, 0 < q < 1. Then it follows from the above considerations
and elementary embeddings that
˚
f 2 D.Rn / W f .0/ D 0 is dense in all spaces Apq s
.Rn /, p < 1, q < 1; (6.319)

which are not covered by (6.300), (6.301).

Step 2. We prove both parts of the theorem. Let f 2 Apq s
.Rn / with p < 1,
q < 1 in (6.300), (6.301) and f .0/ D 0. Then f can be approximated in Apq s
.Rn / by
functions f" 2 D.R / with f" .0/ ! 0 if " ! 0. This is the same situation as in the
n

proof of Proposition 6.21. One has similarly as there that the functions f 2 D.Rn /
s;
with f .0/ D 0 are dense in Apq .Rn / where D f0g: Then one obtains by Step 1 that
it is sufficient to approximate a function

f 2 D.Rn /; supp f  fx 2 Rn W jxj < 1g; f .0/ D 0; (6.320)

in
1 
s
Bpq .Rn /; 0 < p < 1; 0 < q < 1; n p
1 C
< s < 1 C pn ; (6.321)

by functions g 2 D.Rn / with 0 62 supp g. Let

˚
' 2 D.Rn /; supp '  x 2 Rn W 1
2
< jxj < 2 (6.322)
232 6 Complements

such that
1
X
'.2j x/ D 1 if 0 < jxj < 1: (6.323)
j D0

Then
1
X n n
f .x/ D 2j.s p /j 2j.s p / 2j '.2j x/ f .x/ (6.324)
j D0
s
is an atomic decomposition in Bpq .Rn / with (6.321). We refer again to Theorem 1.7
(no moment conditions are needed). The extra factor 2j comes from jf .x/j  c jxj.
Let f J be given by (6.324) with j  J in place of j 2 N0 . One has for given " > 0
that
X1
n
kf J jBpq
s
.Rn /kq  2j.s p 1/q  " (6.325)
j DJ

if J  J."/. Then g D f  f J
is the desired approximation.

6.4.7 A comment on sampling numbers

Let  be a bounded Lipschitz domain in Rn according to Definition 3.4 (iii) if n  2

or a bounded interval in R if n D 1. There is a more or less obvious counterpart
of Theorem 6.76 with  in place of Rn . But it seems to be reasonable to fix some
s
related notation. Recall that we introduced the spaces Apq ./ in Definition 2.1 (i) as
ı
s
the restriction of Apq .Rn / to  and the spaces Apq
s
./ in Definition 5.17 (i) as the
s
completion of D./ in Apq ./. Let C./ be the restriction of C.Rn / to . This
coincides with C./ D C 0 ./ in Definition 5.17 (ii). Then one has

id W Apq
s
./ ,! C./ (6.326)

for the same parameters as in (6.300), (6.301). With

˚
D x l 2  W l D 1; : : : ; L ; L 2 N; n 2 N; (6.327)

one has for the spaces in (6.326) the counterpart

˚ PL
s;
Apq ./ D f 2 Apq s
./ W lD1 jf .x /j D 0
l
(6.328)

of (6.310). Both parts of Theorem 6.76 with  in place of Rn and the restriction

D D D j of D D D.Rn n / (6.329)

in place of D remain valid. One can also replace there (in obvious notation)
ı ı
s
Apq s
./ by Apq ./; s;
Apq s;
./ by Apq ./; and D by D. n /; (6.330)

where now p D 1 and/or q D 1 are admitted.

 6.4 Dichotomy: traces versus density 233

Problems of sampling, tractability and optimal recovery of functions are very fash-
ionable subjects nowadays. They have been studied by many authors. As for the
background of this theory we refer to [TWW88], [Nov88], [NoW08]. Sampling num-
bers rely on standard information, hence methods which are based on function values.
They have a substantial history. Related references may be found in [NoT04], Re-
mark 24, p. 341, and [T06], Remark 4.42, p. 228, which will not be repeated here.
We do not deal systematically with sampling numbers. We are only interested in the
connection between sampling numbers and the -version of Theorem 6.76 described
above. But first we recall what is meant by (non-linear) sampling numbers in the con-
s
text of the above spaces Apq ./. We assume that the source space Aps11 q1 ./ satisfies
(6.326), hence
id W Aps11 q1 ./ ,! C./; (6.331)
which means that s1 ; p1 ; q1 are restricted by the right-hand sides of (6.300), (6.301).
Let
id W Aps11 q1 ./ ,! Aps22 q2 ./ be compact; (6.332)
which is equivalent to
 
1 1
s1 > s2 C n  ; 0 < q1 ; q2  1: (6.333)
p1 p2 C

Recall that aC D max.0; a/ where a 2 R. With as in (6.327) the information map,

NL W Aps11 q1 ./ 7! CL ; L 2 N; (6.334)

given by  
NL f D f .x 1 /; : : : ; f .x L / ; f 2 Aps11 q1 ./; (6.335)
makes sense. As usual, CL is the collection of all L-tuples of complex numbers. Let

SL D ˆL B NL with ˆL W CL 7! Aps22 q2 ./ (6.336)

be an arbitrary map (also called method or algorithm), hence

 
SL f D ˆL f .x 1 /; : : : ; f .x L / 2 Aps22 q2 ./; f 2 Aps11 q1 ./: (6.337)

One wishes to recover a given function f 2 Aps11 q1 ./ in Aps22 q2 ./ by asking for
optimally scattered sampling points and optimally chosen methods ˆL . With id as
in (6.332),
˚
gzL .id/ D inf sup kf  SL f jAps22 q2 ./k W kf jAps11 q1 ./k  1 (6.338)

is the Lth sampling number, where the infimum is taken over all L-tuples fx l gL
lD1

and all maps SL D ˆL B NL according to (6.334)–(6.337). This definition goes back
to [NoT04], Definition 17, p. 339, and may also be found in [T06], Definition 4.32,
p. 219. It can be (equivalently) reformulated as follows.
234 6 Complements

Definition 6.77. Let  be a bounded Lipschitz domain in Rn according to Defini-

˚ L
tion 3.4 (iii) if n  2 or a bounded interval in R if n D 1. Let D x l lD1  .
Let

s1 2 R; s2 2 R; and 0 < p1 ; p2  1 (< 1 for F -spaces), 0 < q1 , q2  1

(6.339)
with  
1 1
s1 > s2 C n  ; (6.340)
p1 p2 C
and let
id W Aps11 q1 ./ ,! Aps22 q2 ./: (6.341)
(i) Let, in addition,
Aps11 q1 ./ ,! C./ (6.342)
and ˚ PL
q1 ./ D f 2 Ap1 q1 ./ W
Aps11; lD1 jf .x /j D 0 :
s1 l
(6.343)
Then for k 2 N,
˚
gk .id/ D inf sup kf jAps22 q2 ./k W kf jAps11 q1 ./k  1; f 2 Aps11;
q1 ./
(6.344)
where the infimum is taken over all with card  k.
(ii) Then for k 2 N,
˚
g k .id/ D inf sup kf jAps22 q2 ./k W kf jAps11 q1 ./k  1; f 2 D (6.345)

where the infimum is taken over all with card  k.

Remark 6.78. As remarked in (6.332), (6.333) the embedding (6.341) makes sense.
It is compact, if and only if, one has (6.340). This justifies the definitions of gk .id/
and g k .id/. According to [NoT04], Proposition 19, p. 340, or [T06], Proposition 4.34,
p. 220, one has
gzk .id/  gk .id/; k 2 N; (6.346)
where the original (non-linear) sampling numbers gzk .id/ are given by (6.338). In
contrast to the sampling numbers gk .id/ in part (i) the corresponding numbers in
(6.345) make sense for any compact embedding (6.341). How they are related and are
the numbers g k .id/ of any use? We do not deal with this question systematically. We
have a closer look at a model case, replacing the target space Aps22 q2 ./ by L1 ./,
hence
id W Apq
s
./ ,! L1 ./: (6.347)
Characterisations for which A 2 fB; F g and s, p, q the embedding (6.347) is continu-
ous may be found in [T01], Theorem 11.2, pp. 168–69, and [Har07], pp. 111–12, with
a reference to [SiT95]. It follows that id in (6.347) is compact if, and only if,
 
0 < p  1 (p < 1 for F -spaces); 0 < q  1; s > p D n p1  1 C : (6.348)
 6.4 Dichotomy: traces versus density 235

Theorem 6.79. Let  be a bounded Lipschitz domain in Rn according to Defini-

tion 3.4 (iii) if n  2 or a bounded interval in R if n D 1. Let gk .id/ and g k .id/ be the
sampling numbers according to Definition 6.77 with respect to the compact embedding
(6.347) (or equivalently (6.348)) in place of (6.341).
(i) Let, in addition,
s
Apq ./ ,! C./ (6.349)
(or equivalently the right-hand sides of (6.300), (6.301)) and s < 1 C pn . Then
s 1
g k .id/  gk .id/  k  n C. p 1/C ; k 2 N: (6.350)

(ii) Let, in addition, (6.349) not be valid. Then

g k .id/  1; k 2 N: (6.351)

Proof. Step 1. We prove part (i). If p < 1, q < 1 then the first equivalence in
(6.350) follows from Theorem 6.76 (ii) transferred to  as indicated at the beginning
of this Section 6.4.7. If s > n=p then the second equivalence in (6.350) is a special
case of [NoT04], Theorem 23, p. 341, or [T06], Theorem 4.37, p. 224. But it is valid
for all spaces in (6.349). This is not immediately covered by [NoT04], [T06]. We add
a respective comment in Remark 6.80 below. It remains to extend these assertions to
spaces with p D 1 and/or q D 1. Let
ı
id0 W Apq
s
./ ,! L1 ./: (6.352)

The different behaviour of id in (6.347) and of id0 in (6.352) near the boundary of  is
immaterial. But otherwise one can now extend the above arguments to p D 1 and/or
q D 1 as indicated in (6.330). In particular, by (6.345)

g k .id/ D g k .id0 /  gk .id0 /; k 2 N: (6.353)

By the proof in [NoT04], [T06] one obtains that

s 1
gk .id0 /  k  n C. p 1/C  gk .id/; k 2 N: (6.354)

This proves (6.350) also if p D 1 and /or q D 1.

Step 2. We prove part (ii). If (6.349) is not valid then p < 1. If, in addition, q < 1
then (6.351) follows from (6.345) and the -version of Theorem 6.76 saying that D
s
is dense in Apq ./. Spaces with q D 1 can now be incorporated by monotonicity
arguments since one has already (6.351) for suitable smaller spaces.
Remark 6.80. We add a comment about the second equivalence in (6.350) for spaces
s
Apq ./ with (6.349) and s D n=p. First we remark that one can extend [T06],
Proposition 4.36(i), p. 222, to the spaces Fpq s
./ with 0 < p  1, 0 < q  1,
s D n=p, according to (6.301). As there one may assume that q  p. The arguments
s
rely on [T06], Corollary 4.13(ii), p. 202, which can be extended to Fpq with 0 < p  1,
236 6 Complements

q  p, s D n=p. Afterwards one can apply the arguments from the proof of [T06],
Theorem 4.37, p. 224, to these cases. This proves the second equivalence in (6.350)
s
for all spaces Fpq ./ with (6.349). By

n=p
Bp;q n
./ ,! B1;1 ./; 0 < p  1; 0 < q  1; (6.355)

one obtains a corresponding assertion for these B-spaces in the limiting situation s D
n 0
n=p in (6.300). Complex interpolation between B1;1 ./ and B1;1 ./ extends these
n=p
considerations to the spaces Bp;q ./ with 1 < p < 1, q  1.

Remark 6.81. It is not the aim of this Section 6.4.7 to study the behaviour of (non-
linear) sampling numbers gk for diverse source and target spaces for their own sake.
We wanted only to make clear how closely and also how naturally these numbers are
related to the pointwise dichotomy according to Theorem 6.76. This may justify that
we fixed the target space by L1 ./ in (6.347) and in Theorem 6.79. Although we used
in the above arguments occasionally the right-hand side of (6.350) there is little doubt
that these considerations can be extended to (6.332) with (6.331). Secondly we tried to
find a characterisation of gk which does not use (6.331) and which gives the possibility
to extend the definition of non-linear sampling numbers gk to source spaces which
do not necessarily satisfy (6.331). These are the sampling numbers g k according to
(6.345) with (6.350). Questions of this type have attracted some attention. We refer
to [Hei08]. But the outcome both in [Hei08] and also in the above Theorem 6.79 (ii)
is largely negative. But it should be said that this observation is the main motivation
in [Hei08] to step from deterministic approximations to randomised approximations
what changes the situation.

Remark 6.82. Sampling numbers for embeddings of type (6.341), (6.342) in bounded
Lipschitz domains  with the target spaces Lp2 ./, 0 < p2  1, in place of
Aps22 q2 ./, have been considered in [NoT04]. This has been extended in [Tri05] to
target spaces Aps22 q2 ./ with s2 > p2 . One may also consult [T06], Section 4.3. In
[NoT04], [Tri05], [T06] one finds references to earlier papers. To some extent one can
replace bounded Lipschitz domains by arbitrary domains  with jj < 1, or E-thick
domains  according to Definition 3.1 (ii) with jj < 1. We return to this point in
Theorem 6.87 below with a reference to [Tri07b]. Sampling numbers for target spaces
of type Bp02 q2 ./ show some curious behaviour, different from Lp2 ./, [Vyb08a].
Finally we refer to [Vyb07a] where also target spaces Aps22 q2 ./ with s2 < 0 have
been considered. Some of these recent results have also been surveyed in [NoW08],
Section 4.2.4, where the non-linear sampling numbers gk are denoted as
 
gk D e wor k; ƒstd ; k 2 N: (6.356)

As far as this notation is concerned one may consult [NoW08], Sections 4.1.1, 4.14.
We only mention that ƒstd refers to pointwise evaluation, called standard information
in the context of theory developed there.
 6.5 Polynomial reproducing formulas 237

6.5 Polynomial reproducing formulas

6.5.1 Global reproducing formulas

Polynomial reproducing formulas play some role in numerical analysis. They can
be used to construct stable wavelet bases on intervals and on domains. We refer to
[Coh03], Section 2.12, especially pp. 125–26. We commented in Section 6.4.7 on
sampling numbers. Assertions as on the right-hand side of (6.350) are proved in
[NoT04] and [T06], Section 4.3.3, with the help of a local polynomial reproducing
formula due to H. Wendland, [Wen01]. The aim of the present Section 6.5 is rather
modest. We wish to make clear how one derives polynomial reproducing formulas
from wavelet expansions. We begin with corresponding assertions in Rn .
s
We use the weighted spaces Apq .Rn ; w/ as introduced in Definition 1.22, specified
by
Bpps
.Rn ; w /; 1  p  1; s > 0; w .x/ D .1 C jxj2 /=2 ; (6.357)
with 2 R as in (1.136). According to Theorem 1.26 any f 2 Bpps
.Rn ; w / can be
expanded by X
j n=2 j
f D j;G
m 2 ‰G;m ;  2 bpp
s
.w /; (6.358)
j;G;m

with Z
j
j;G
m D2 j n=2
f .x/ ‰G;m .x/ dx (6.359)
Rn
j
where ‰G;m are the real wavelets (1.91) based on (1.87), (1.88). All these wavelets
satisfy moment conditions of type (1.88) extended to Rn with exception of
Y
n
‰F ;m .x/ D 0
‰G;m .x/ D F .xr  mr /; m 2 Zn ; x 2 Rn ; (6.360)
rD1

with G D .F; : : : ; F /. Let f D P be a polynomial of degree N and let N < u 2

N in (1.88) and Theorem 1.26. If P 2 Bpp s
.Rn ; w / then one has the expansion
(6.358) where all coefficients j;G 0;G
m are zero with possible exception of m with G D
fF; : : : ; F g and m 2 Z , hence
n

X Z 
P .x/ D P .y/ ‰F ;m .y/ dy ‰F ;m .x/; x 2 Rn : (6.361)
m2Zn Rn

We justify this expansion.

Theorem 6.83. Let P be a polynomial of degree at most N in Rn . Let s > 0,
n
1  p  1; max.s; N / < u 2 N; < N  : (6.362)
p
Then P 2 Bpp s
.Rn ; w /, and one has (6.361), unconditional convergence being in
s
Bpp .Rn ; w /.
238 6 Complements

Proof. Since supp Py D f0g it follows from (1.117) that

P 2 Bpp
s
.Rn ; w / if, and only if, w P 2 Lp .Rn /: (6.363)

But this is ensured by (6.362). Furthermore we have by Theorem 1.26 and (1.119) that
 X 1=p
kP jBpp s
.Rn ; w /k  k.P / jbpp
s
.w /k  c .1 C jmj/pCNp (6.364)
m2Zn

with the usual modification if p D 1. This show that (6.361) converges uncondition-
ally for all p with 1  p  1, including p D 1.
R
Remark 6.84. Recall the well-known assertion R F .x/ dx D 1, [T06], p. 31. Then
one obtains by (6.361) with P .x/ D 1 that
X
‰F ;m .x/ D 1; x 2 Rn ; (6.365)
m2Zn

is a resolution of unity. The right-hand side of (6.361) makes sense for any, say,
continuous function f in Rn , hence
X Z 
.Uf /.x/ D f .y/ ‰F ;m .y/ dy ‰F ;m .x/; x 2 Rn : (6.366)
m2Zn Rn

Polynomials f D P of degree at most N are reproduced, what explains the notation

polynomial reproducing formula. Recall that
˚
‰F ;m 2 C u .Rn /; supp ‰F ;m  B.m; c/ D y 2 Rn W jy  mj < c ; (6.367)

for some c > 0. One can replace ‰F ;m .x/ by suitable functions Hm .x/ with

Hm 2 C 1 .Rn /; supp Hm  B.m; c 0 /; m 2 Zn ; (6.368)

for some c 0 > 0. Let ! be a compactly supported C 1 function satisfying

Z Z
!.y/ dy D 1; y ˛ !.y/ dy D 0; 0 < j˛j  N: (6.369)
Rn Rn

Mollification
Z Z
.! ? g/.x/ D g.y/ !.x  y/ dy D g.x  y/ !.y/ dy (6.370)
Rn Rn

of (6.361) results in
X Z 
P .x/ D P .y/ ‰F ;m .y/ dy Hm .x/ (6.371)
m2Zn Rn

with Hm D .! ?‰F ;m / 2 D.Rn /. Then one obtains a polynomial reproducing formula

(6.361) with Hm according to (6.368) in place of ‰F ;m .
 6.5 Polynomial reproducing formulas 239

6.5.2 Local reproducing formulas

One may ask whether one can replace the evaluation in terms of the scalar product
.f; ‰F ;m / in (6.366) by pointwise evaluation of f . This is just what one needs in
the context of sampling numbers. This can be done. We formulate the outcome and
indicate the connection with wavelet expansions.
Let  be an arbitrary domain in Rn with jj < 1. Let for  > 0 and D @,

 D fx 2  W dist.x; / > g: (6.372)

Then  is bounded. Let C loc ./ be the collection of all complex-valued continuous
functions in , hence

f 2 C loc ./ if, and only if, f j 2 C. /;  > 0; (6.373)

in the notation of Definition 5.17. As before, B.x; %/ stands for a ball in Rn centred at
x 2 Rn and of radius % > 0.
Theorem 6.85. Let  be an arbitrary domain in Rn with jj < 1. Let N 2 N. Then
there is a number 0 > 0 (depending on /, and, in dependence on N and 0 , numbers
a > 0; b > 0; c > 0; d > 0 with the following property. For any  with 0 <   0
one finds points x j 2  having pairwise distance of at least a , and real function
hj 2 D./ with

sup jhj .x/j  c; supp hj  B.x j ; b /  d  ; (6.374)

such that the mapping U ,

X
U f D f .x j / hj ; f 2 C loc ./; (6.375)
j

is polynomial reproducing in  ,

.U P /.x/ D P .x/; x 2  ; (6.376)

for any polynomial of degree at most N .

˚
Proof (Outline). Let ˆjr be an interior orthonormal u-wavelet basis in L2 ./ ac-
cording to Theorem 2.33 based on Definitions 2.31 and 2.4,
1 X
X Nj
 
f D f; ˆjr ˆjr ; f 2 L2 ./: (6.377)
j D0 rD1

One may choose L 2 N in (2.32)–(2.34), based on (2.27), such that 2L  . Let 
be a cut-off function with
 
 2 D =2 ;  .x/ D 1 if x 2  : (6.378)
240 6 Complements

Then (6.377) can be applied to f D  P , where P is a polynomial of degree at

most N . We choose N < u 2 N. Then the interior wavelets in (2.33), (2.27) satisfy
the moment conditions (1.88) extended to Rn . If, in addition, supp ˆjr   then
   
 P; ˆjr D P; ˆjr D 0: (6.379)

Furthermore if L 2 N is chosen sufficiently large in dependence on  then one may

assume that supp ˆjr   n =2 for the boundary wavelets in (2.34). Hence the
corresponding scalar products in (6.377) are zero. Now one obtains that

X
N0
 
P .x/ D 2Ln=2  P; ˆ0r 2Ln=2 ˆ0r .x/; x 2  : (6.380)
rD1
 
By (2.31), (2.26) one has supp ˆ0r  B x L;r ; c 0 2L taking L with 2L   into
account. There are points
˚ K  
x L;r;k kD1
 B x L;r ; c 00 2L

having pairwise distance of at least  2L , and constants ckL;r with jckL;r j  C for
some C > 0 and all admitted L; r; k such that
Z X
K
 
2Ln=2
P .y/ ˆ0r .y/  .y/ dy D ckL;r P x L;r;k (6.381)

kD1

for all polynomials of degree at most N . A detailed proof of this assertion may be
found in [Tri07b], pp. 484–86. Inserting (6.381) in (6.380) one obtains that
X
P .x/ D P .x j / hQj .x/; x 2  ; (6.382)
j

where the points x j 2  have the desired properties and hQj satisfies (6.374) with hQ in
place of h. Finally one applies the same mollification argument as in (6.370), (6.371),
assuming that ! is supported by a ball of radius   . This proves (6.376) with (6.375).

Remark 6.86. The proof of (6.381) uses in a decisive way that P is a polynomial of
degree at most N , but not the specific nature of ˆ0r (only its magnitude and support).
Details may be found in [Tri07b].

6.5.3 A further comment on sampling numbers

In Definition 6.77 we introduced for the compact embeddings (6.341) in bounded

Lipschitz domains  the (non-linear) sampling numbers gk .id/. We obtained (6.350)
 6.5 Polynomial reproducing formulas 241

for the modified embedding (6.347) with (6.349). In Section 4.3.4 we discussed to
which extent one can replace bounded Lipschitz domains by more general domains
in connection with entropy numbers and approximation numbers for corresponding
compact embeddings. One can ask the same question for the above sampling numbers
gk .id/. This was the main aim of [Tri07b]. We formulate the outcome as an application
of wavelet bases and polynomial reproducing formulas according to Theorem 6.85. In
what follows the definition of the sampling numbers gk .id/ in (6.344) is naturally
extended to arbitrary compact embeddings

id W G1 ./ ,! G2 ./ with G1 ./ ,! C./ (6.383)

by
˚
gk .id/ D inf sup kf jG2 ./k W kf jG1 ./k  1; f 2 G1 ./ ; (6.384)

k 2 N, where
˚ PL ˚ L
G1 ./ D f 2 G1 ./ W lD1 jf .x /j D 0
l
with D xl lD1
 : (6.385)

As before, the infimum in (6.384) is taken over all with card  k. Let pq be as
in (1.32).

Theorem 6.87. (i) Let u 2 N and let  be an arbitrary domain in Rn with jj < 1.
s;rloc
Let Fpq ./ with
 
n
0 < p  1; 0 < q  1; s > max ; pq (6.386)
p

.q D 1 if p D 1/ be the refined localisation spaces according to Definition 2.14.

Then
id W Fpq
s;rloc
./ ,! L t ./; 0 < t  1; (6.387)
is compact and
s 1 1
gk .id/  k  n C. p  t /C ; k 2 N: (6.388)
(ii) Let u 2 N and let  be an E-thick domain in Rn according to Definition 3.1 (ii)
with jj < 1 and let AQpq s
./ be the corresponding spaces as introduced in Defini-
tion 2.1 (ii) with
n
0 < p  1; 0 < q  1; s > ; (6.389)
p
(p < 1 for the F -spaces). Then

id W AQpq
s
./ ,! L t ./; 0 < t  1; (6.390)

is compact and
s 1 1
gk .id/  k  p C. p  t /C ; k 2 N: (6.391)
242 6 Complements

Remark 6.88. This coincides essentially with [Tri07b], Theorem 23, where one finds
also a proof. It is based on the polynomial reproducing mapping U in Theorem 6.85
which in turn relies on wavelet expansions. Part (ii) follows from part (i), Proposi-
tion 3.10 and some interpolation as far as the B-spaces are concerned. According
to Proposition 3.8 (i) bounded Lipschitz domains are thick, and hence E-thick. This
shows that (6.391) is in good agreement with (6.350). One may compare the above
assertions with corresponding results in Section 4.3.4. As there
the adequate assumption to obtain the desired behaviour of diverse types of
numbers for compact embeddings between function spaces on domains  is not
so much that  is bounded, but that jj < 1.
ı
Remark 6.89. Recall that Apq s
./ is the completion of D./ in Apq s
.Rn /, Defini-
tion 5.17. If a bounded domain  is not only E-thick, but Lipschitz, cellular or C 1
and if s is not an exceptional value, s  p1 62 N0 , then one can replace AQpq
s
./ in
ı
s
part (ii) of the above theorem by Apq ./. We refer to Proposition 5.19, Remark 5.20
and Propositions 6.13 (ii), 6.15.
Bibliography

The numbers following I at the end of each item refer to the page(s) on which the respective
work is cited.

[AdH96] D. R. Adams, L. I. Hedberg. Function spaces and potential theory. Grundlehren Math.
Wiss. 314, Springer-Verlag, Berlin, 1996. I 131, 225
[AlK06] F. Albiac, N. J. Kalton. Topics in Banach space theory. Grad. Texts in Math. 233,
Springer-Verlag, New York, 2006. I 15, 103, 199, 211, 212, 215
[Ban32] S. Banach. Théorie des opérations linéaires. Monogr. Mat. 1, Warszawa, 1932. I 199
[BeS88] C. Bennett, R. Sharpley. Interpolation of operators. Pure Appl. Math. 129, Academic
Press, Boston, 1988. I 61
[BeL76] J. Bergh, J. Löfström. Interpolation spaces, an introduction. Grundlehren Math. Wiss.
223, Springer-Verlag, Berlin, 1976. I 109
[BIN75] O. V. Besov, V. P. Il’in, S. M. Nikol’skij. Integral representations of functions and em-
bedding theorems. Nauka, Moskva, 1975, sec. ed., 1996 (in Russian); English translation:
Wiley, New York, 1978/79. I 118, 146
[BoN06a] L. Borup, M. Nielsen. Nonlinear approximation in ˛-modulation spaces. Math. Nachr.
279 (2006), 101–120. I 215
[BoN06b] L. Borup, M. Nielsen. Banach frames for multivariate ˛-modulation spaces. J. Math.
Anal. Appl. 321 (2006), 880–895. I 215
[Bri03] M. Bricchi. Complements and results on h-sets. In Function Spaces, Differential Oper-
ators and Nonlinear Analysis. Birkhäuser, Basel, 2003, 219–229. I 216
[Bri04] M. Bricchi. Tailored Besov spaces and h-sets. Math. Nachr. 263–264 (2004), 36–52.
I 217
[BrL75] Th. Bröcker, L. Lander. Differential germs and catastrophes. London Math. Soc. Lec-
ture Note Ser. 17, Cambridge University Press, Cambridge, 1975. I 133
[BuG79] V. I. Burenkov, M. L. Gol’dman. On the extension of functions of Lp (Russian). Trudy
Mat. Inst. Steklov 150 (1979), 31–51. I 220
[Cae98] A. M. Caetano. About approximation numbers in function spaces. J. Approximation
Theory 94 (1998), 383–395. I 124
[CLT07] A. M. Caetano, S. Lopes, H. Triebel. A homogeneity property for Besov spaces. J.
Funct. Spaces Appl. 5 (2007), 123–132. I 34, 35, 57
[Car06] A. Carvalho. Fractal geometry of Weierstrass-type functions. Fractals 17 (2009), 23–37.
I 91
[Chr84] M. Christ. The extension problem for certain function spaces involving fractional orders
of differentiability. Arkiv Mat. 22 (1984), 63–81. I 72
[Cie84] Z. Ciesielski. Spline bases in classical function spaces on compact C 1 manifolds, III.
In Constructive Theory of Functions, Proc. Intern. Conf. Varna. Publishing House Bulg.
Acad. Sciences, Sofia, 1984, 214–223. I 167, 168
244 Bibliography

[CiF83a] Z. Ciesielski, T. Figiel. Spline bases in classical function spaces on compact C 1

manifolds, I. Studia Math. 76 (1983), 1–58. I 167
[CiF83b] Z. Ciesielski, T. Figiel. Spline bases in classical function spaces on compact C 1
manifolds, II. Studia Math. 76 (1983), 95–136. I 167
[CFMM07] F. Cobos, L. M. Fernández-Cabrera, A. Manzano, A. Martinez. Logarithmic in-
terpolation spaces between quasi-Banach spaces. Z. Anal. Anwend. 26 (2007), 65–86.
I 65
[CFT04] F. Cobos, L. M. Fernández-Cabrera, H. Triebel. Abstract and concrete logarithmic
interpolation spaces. J. London Math. Soc. 70 (2004), 231–243. I 64
[Coh03] A. Cohen. Numerical analysis of wavelet methods. Stud. Math. Appl. 32, North-
Holland, Elsevier, Amsterdam, 2003. I 52, 168, 237
[CDV00] A. Cohen, W. Dahmen, R. DeVore. Multiscale decompositions on bounded domains.
Trans. Amer. Math. Soc. 352 (2000), 3651–3685. I 52
[CDV04] A. Cohen, W. Dahmen, R. DeVore. Adaptive wavelet techniques in numerical sim-
ulation. In Encyclopedia Computational Mechanics. Wiley, Chichester, 2004, 1–64.
I 52
[DNS06a] S. Dahlke, E. Novak, W. Sickel. Optimal approximation of elliptic problems by linear
and nonlinear mappings I. J. Complexity 22 (2006), 29–49. I 90
[DNS06b] S. Dahlke, E. Novak, W. Sickel. Optimal approximation of elliptic problems by linear
and nonlinear mappings II. J. Complexity 22 (2006), 549–603. I 90
[DNS07] S. Dahlke, E. Novak, W. Sickel. Optimal approximation of elliptic problems by linear
and nonlinear mappings III: Frames. J. Complexity 23 (2007), 614–648. I 90
[Dah97] W. Dahmen. Wavelet and multiscale methods for operator equations. Acta Numer. 6
(1997), 55–228. I 52, 168
[Dah01] W. Dahmen. Wavelet methods for PDEs – some recent developments. J. Comput Appl.
Math. 128 (2001), 133–185. I 52, 168
[DaS99] W. Dahmen, R. Schneider. Wavelets on manifolds I: construction and domain decom-
position. SIAM J. Math. Anal. 31 (1999), 184–230. I 168
[Dau92] I. Daubechies. Ten lectures on wavelets. CBMS-NSF Regional Conf. Series Appl.
Math., SIAM, Philadelphia, 1992. I 13, 14, 27
[DeS93] R. A. DeVore, R. C. Sharpley. Besov spaces on domains in Rd . Trans. Amer. Math.
Soc. 335 (1993), 843–864. I 118
[Dis03] S. Dispa. Intrinsic characterizations of Besov spaces on Lipschitz domains. Math. Nachr.
260 (2003), 21–33. I 118
[EdE04] D. E. Edmunds, W. D. Evans. Hardy operators, function spaces and embeddings.
Springer Monogr. Math., Springer-Verlag, Berlin, 2004. I 61, 77
[EdT89] D. E. Edmunds, H. Triebel. Entropy numbers and approximation numbers in function
spaces. Proc. London Math. Soc. 58 (1989), 137–152. I 124
[EdT92] D. E. Edmunds, H. Triebel. Entropy numbers and approximation numbers in function
spaces, II. Proc. London Math. Soc. 64 (1992), 153–169. I 124
[EdT95] D. E. Edmunds, H. Triebel. Logarithmic Sobolev spaces and their applications to
spectral theory. Proc. London Math. Soc. 71 (1995), 333–371. I 62, 64
 Bibliography 245

[ET96] D. E. Edmunds, H. Triebel. Function spaces, entropy numbers, differential opera-

tors. Cambridge Tracts in Math. 120, Cambridge University Press, Cambridge, 1996
(Reprinted 2008). I 18, 61, 64, 77, 119, 120, 124, 176, 231
[EdT98] D. E. Edmunds, H. Triebel. Logarithmic spaces and related trace problems. Funct.
Approx. Comment. Math. 26 (1998), 189–204. I 64
[Fal85] K. J. Falconer. The geometry of fractal sets. Cambridge Tracts in Math. 85, Cambridge
University Press, Cambridge, 1985. I 216
[FeS71] C. Fefferman, E. M. Stein. Some maximal inequalities. Amer. J. Math. 93 (1971),
107–115. I 10
s
[Fra86] J. Franke. On the spaces Fpq of Triebel-Lizorkin type: Pointwise multipliers and spaces
on domains. Math. Nachr. 125 (1986), 29–68. I 231
[FrJ85] M. Frazier, B. Jawerth. Decomposition of Besov spaces. Indiana Univ. Math. J. 34
(1985), 777–799. I 5, 220
[FrJ90] M. Frazier, B. Jawerth. A discrete transform and decompositions of distribution spaces.
J. Funct. Anal. 93 (1990), 34–170. I 5, 184, 220
[GaH04] G. Garrigós, E. Hernández. Sharp Jackson and Bernstein inequalities for N -term ap-
proximation in sequence spaces with applications. Indiana Univ. Math. J. 53 (2004),
1739–1762. I 215
[GHM07] G. Garrigós, E. Hernández, J. M. Martell. Wavelets, Orlicz spaces, and greedy bases.
Appl. Comput. Harmon. Anal. 24 (2008), 70–92. I 215
[Gol79] M. L. Gol’dman. On the extension of functions of Lp .Rn / in spaces with a larger
number of dimensions (Russian). Mat. Zametki 25 (1979), 513–520. I 220
[Gri85] P. Grisvard. Elliptic problems in nonsmooth domains. Monogr. Stud. Math. 24, Pitman,
Boston, 1985. I 174, 209
[Gri92] P. Grisvard. Singularities in boundary value problems. Rech. Math. Appl. 22, Masson,
Paris; Springer-Verlag, Berlin, 1992. I 174, 209
[Gro01] K. Gröchenig. Foundations of time-frequency analysis. Appl. Numer. Harmon. Anal.,
Birkhäuser, Boston, 2001. I 159
[Haar10] A. Haar. Zur Theorie der orthogonalen Funktionensysteme. Math. Ann. 69 (1910),
331–371. I 55
[HaS04] H. Harbrecht, R. Schneider. Biorthogonal wavelet bases for the boundary element
method. Math. Nachr. 269–270 (2004), 167–188. I 168
[Har07] D. D. Haroske. Envelopes and sharp embeddings of function spaces. CRC Research
Notes in Math. 437, Chapman & Hall, Boca Raton, 2007. I 61, 204, 208, 229, 230, 234
[HaS08] D. D. Haroske, C. Schneider. Besov spaces with positive smoothness on Rn , embed-
dings and growth envelopes. Jenaer Schriften Mathematik Informatik, Math/Inf/02/08,
Jena, 2008. I 201
[HaT94a] D. D. Haroske, H. Triebel. Entropy numbers in weighted function spaces and eigen-
value distributions of some degenerate pseudodifferential operators I. Math. Nachr. 167
(1994), 131–156. I 18
[HaT94b] D. D. Haroske, H. Triebel. Entropy numbers in weighted function spaces and eigen-
value distributions of some degenerate pseudodifferential operators II. Math. Nachr. 168
(1994), 109–137. I 18
246 Bibliography

[HaT05] D. D. Haroske, H. Triebel. Wavelet bases and entropy numbers in weighted function
spaces. Math. Nachr. 278 (2005), 108–132. I 17
[HaT08] D. D. Haroske, H. Triebel. Distributions, Sobolev spaces, elliptic equations. EMS
Textbooks in Math., European Math. Soc. Publishing House, Zürich, 2008. I 118, 147
[HeN07] L. I. Hedberg, Yu. Netrusov. An axiomatic approach to function spaces, spectral syn-
thesis, and Luzin approximation. Mem. Amer. Math. Soc. 188, 882 (2007). I 131, 225
[Hei01] J. Heinonen. Lectures on analysis on metric spaces. Universitext, Springer-Verlag, New
York, 2001. I 199
[Hei08] S. Heinrich. Randomized approximation of Sobolev embeddings. In Monte Carlo and
Quasi-Monte Carlo Methods 2006. Springer-Verlag, Berlin, 2008, 445–460. I 236
[Hel78] S. Helgason. Differential geometry, Lie groups, and symmetric spaces. Pure Appl. Math.
80, Academic Press, New York, 1978. I 133
[HoL05] J. A. Hogan, J. D. Lakey. Time-frequency and time-scale methods. Appl. Numer. Har-
mon. Anal., Birkhäuser, Boston, 2005. I 52
[JNP83] S. Janson, P. Nilsson, J. Peetre. Notes on Wolff’s note on interpolation spaces. Proc.
London Math. Soc. 48 (1983), 283–299. I 111
[Jaw77] B. Jawerth. Some observations on Besov and Lizorkin-Triebel spaces. Math. Scand. 40
(1977), 94–104. I 231
[JeK82] D. Jerison, C. Kenig. Boundary behavior of harmonic functions in non-tangentially
accessible domains. Adv. Math. 146 (1982), 80–147. I 77
[Jon81] P. W. Jones. Quasiconformal mappings and extendability of functions. Acta Math. 147
(1981), 71–88. I 71, 72, 74
[JoK07] A. Jouini, M. Kratou. Wavelet bases on a manifold. J. Funct. Anal. 248 (2007), 128–151.
I 168
[Kab08] M. Kabanava. Tempered Radon measures. Rev. Mat. Complutense 21 (2008), 553–564.
I 216
[Ken94] C. E. Kenig. Harmonic analysis techniques for second order elliptic boundary value
problems. CBMS Regional Conf. Series Math. 83, Amer. Math. Soc., Providence, 1994.
I 77
[Kon86] H. König. Eigenvalue distribution of compact operators. Oper. Theory Adv. Appl. 16,
Birkhäuser, Basel, 1986. I 127
[KoT99] S. V. Konyagin, V. N. Temlyakov. A remark on greedy approximation in Banach spaces.
East J. Approx. 5 (1999), 365–379. I 212, 215
[Kyr03] G. Kyriazis. Decomposition systems for function spaces. Studia Math. 157 (2003),
133–169. I 12, 14
[Mal99] S. Mallat. A wavelet tour of signal processing. Academic Press, San Diego, 1999 (sec.
ed.). I 13, 14, 27, 52
[Mall95] P. Malliavin. Integration and probability. Grad. Texts in Math. 157, Springer-Verlag,
New York, 1995. I 217, 218
[Mar37] J. Marcinkiewicz. Quelques theorèmes sur les séries orthogonales. Ann. Soc. Polon.
Math. 16 (1937), 84–96. I 56
 Bibliography 247

[Mat95] P. Mattila. Geometry of sets and measures in Euclidean spaces. Cambridge Stud. Adv.
Math. 44, Cambridge University Press, Cambridge, 1995. I 91, 133, 216
[Maz85] V. G. Maz’ya. Sobolev spaces. Springer Ser. Soviet Math., Springer-Verlag, Berlin,
1985. I 68, 72, 76
[MeM00] O. Mendez, M. Mitrea. The Banach envelopes of Besov and Triebel-Lizorkin spaces
and applications to partial differential equations. J. Fourier Anal. Appl. 6 (2000), 503–531.
I 109, 111
[Mey92] Y. Meyer. Wavelets and operators. Cambridge Stud. Adv. Math. 37, Cambrigde Uni-
versity Press, Cambridge, 1992. I 13, 14
[Miy90] A. Miyachi. H p spaces over open subsets of Rn . Studia Math. 95 (1990), 205–228.
I 108
[Miy93] A. Miyachi. Extension theorems for the function spaces of DeVore and Sharpley. Math.
Japonica 38 (1993), 1033–1049. I 72
[Miy98] A. Miyachi. On the extension properties of Triebel-Lizorkin spaces. Hokkaido Math.
J. 27 (1998), 273–301. I 72
[Net90] Yu. V. Netrusov. Sets of singularities of functions in spaces of Besov and Lizorkin-
Triebel type. Proc. Steklov Inst. Math. 187 (1990), 185–203. I 220
[Nik77] S. M. Nikol’skij. Approximation of functions of several variables and embedding the-
orems. Sec. ed., Nauka, Moskva, 1977, first ed., Nauka, Moskva, 1969 (in Russian);
English translation, Springer-Verlag, Berlin, 1975. I 118, 146
[Nov88] E. Novak. Deterministic and stochastic error bounds in numerical analysis. Lecture
Notes in Math. 1349, Springer-Verlag, Berlin, 1988. I 233
[NoT04] E. Novak, H. Triebel. Function spaces in Lipschitz domains and optimal rates of
convergence for sampling. Constr. Approx. 23 (2006), 325–350. I 233, 234, 235, 236,
237
[NoW08] E. Novak, H. Woźniakowski. Tractability of multivariate problems. Vol. I: Linear
information. EMS Tracts Math. 6, European Math. Soc. Publishing House, Zürich.
I 233, 236
[Pal32] R. E. A. C. Paley. A remarkable series of orthogonal functions I. Proc. London Math.
Soc. 34 (1932), 241–264. I 56
[Pee75] J. Peetre. The trace of Besov space – a limiting case. Technical Report, Lund, 1975.
I 220
[PeS64] A. Pełczyński, I. Singer. On non-equivalent bases and conditional bases in Banach
spaces. Studia Math. 25 (1964/1965), 5–25. I 211
[Pie80] A. Pietsch. Operator ideals. North-Holland Math. Library 20, North-Holland, Amster-
dam, 1980. I 127
[Pie87] A. Pietsch. Eigenvalues and s-numbers. Cambridge Stud. Adv. Math. 13, Cambridge
University Press, Cambridge, 1987. I 127
[Pie07] A. Pietsch. History of Banach spaces and linear operators. Birkhäuser, Boston, 2007.
I 127
[Rog06] L. G. Rogers. Degree-independent Sobolev extension on locally uniform domains. J.
Funct. Anal. 235 (2006), 619–665. I 71
248 Bibliography

[RuS96] T. Runst, W. Sickel. Sobolev spaces of fractional order, Nemytskij operators, and
nonlinear partial differential equations. De Gruyter Ser. Nonlinear Anal. Appl. 3, W. de
Gruyter, Berlin, 1996. I 86, 87, 169, 182
[Ry98] V. S. Rychkov. Intrinsic characterizations of distribution spaces on domains. Studia
Math. 127 (1998), 277–298. I 118
[Ry99] V. S. Rychkov. On restrictions and extensions of the Besov and Triebel-Lizorkin spaces
with respect to Lipschitz domains. J. London Math. Soc. 60 (1999), 237–257. I101, 108,
113, 118, 176
[Ry00] V. S. Rychkov. Linear extension operators for restrictions of function spaces to irregular
open sets. Studia Math. 140 (2000), 141–162. I 72
[ST87] H.-J. Schmeisser, H. Triebel. Topics in Fourier analysis and function spaces. Wiley,
Chichester, 1987. I 20, 21, 22
[Sch08] C. Schneider. On dilation operators in Besov spaces. Rev. Mat. Complutense 22 (2009),
111–128. I 89, 201
[Sch09] C. Schneider. Spaces of Sobolev type with positive smoothness on Rn , embeddings and
growth envelopes. J. Funct. Spaces Appl. (to appear). I 201
[See89] A. Seeger. A note on Triebel-Lizorkin spaces. In Approximation and Function Spaces,
Banach Center Publ. 22, PWN Polish Sci. Publ., Warsaw, 1989, 391–400. I72, 74, 108,
118
[Shv06] P. Shvartsman. Local approximations and intrinsic characterization of spaces of smooth
functions on regular subsets of Rn . Math. Nachr. 279 (2006), 1212–1241. I 108, 118
[SiT95] W. Sickel, H. Triebel. Hölder inequalities and sharp embeddings in function spaces of
s
Bpq s
and Fpq type. Z. Anal. Anwendungen 14 (1995), 105–140. I 229, 231, 234
[Skr97] L. Skrzypczak. Besov spaces on symmetric manifolds – the atomic decomposition.
Studia Math. 124 (1997), 215–238. I 139
[Skr98a] L. Skrzypczak. Mapping properties of pseudodifferential operators on manifolds with
bounded geometry. J. London Math. Soc. 57 (1998), 721–738. I 139
[Skr98b] L. Skrzypczak. Atomic decompositions on manifolds with bounded geometry. Forum
Math. 10 (1998), 19–38. I 134, 139
[Skr03] L. Skrzypczak. Heat extensions, optimal atomic decompositions and Sobolev embed-
dings in presence of symmetries on manifolds. Math. Z. 243 (2003), 745–773. I 139
[Skr08] L. Skrzypczak. Wavelet frames, Sobolev embeddings and negative spectrum of
Schrödinger operators on manifolds with bounded geometry. J. Fourier Anal. Appl. 14
(2008), 415–442. I 139
[Soa97] P. Soardi. Wavelet bases in rearrangement invariant function spaces. Proc. Amer. Math.
Soc. 125 (1997), 3669–3673. I 64
[Sob50] S. L. Sobolev. Einige Anwendungen der Funktionalanalysis auf Gleichungen der math-
ematischen Physik. Math. Lehrbücher 12, Akademie-Verlag, Berlin, 1964 (Translation
from the Russian, Leningrad, 1950). I 68
[Ste70] E. M. Stein. Singular integrals and differentiability properties of functions. Princeton
Math. Ser. 30, Princeton University Press, Princeton, 1970. I 30
 Bibliography 249

[Tem98] V. N. Temlyakov. The best m-term approximation and greedy algorithms. Adv. Comput.
Math. 8 (1998), 249–265. I 215
[Tor86] A. Torchinsky. Real-variable methods in harmonic analysis. Pure Appl. Math. 123,
Academic Press, San Diego, 1986. I 10
[TWW88] J. F. Traub, G. W. Wasilkowski, H. Woźniakowski. Information-based complexity.
Comput. Sci. Sci. Comput., Academic Press, Boston, 1988. I 233
[T78] H. Triebel. Interpolation theory, function spaces, differential operators. North-Holland
Math. Library 18, North-Holland, Amsterdam, 1978. (Sec. ed. Barth, Heidelberg, 1995)
vi, 40, 61, 97, 109, 111, 118, 146, 151, 185, 186, 189, 225
[T83] H. Triebel. Theory of function spaces. Monogr. Math. 78, Birkhäuser, Basel, 1983 v, 2,
7, 20, 22, 35, 58, 97, 110, 113, 114, 140, 141, 143, 145, 146, 149, 183, 186, 189, 192,
217, 218
[T92] H. Triebel. Theory of function spaces II. Monogr. Math. 84, Birkhäuser, Basel, 1992 v,
2, 3, 5, 35, 38, 101, 109, 116, 118, 139, 141, 220, 223, 226
[T97] H. Triebel. Fractals and spectra. Monogr. Math. 91, Birkhäuser, Basel, 1997 91, 125,
216, 217, 220, 223
[T01] H. Triebel. The structure of functions. Monogr. Math. 97, Birkhäuser, Basel, 2001 29,
33, 34, 35, 37, 39, 40, 73, 80, 86, 91, 125, 131, 148, 150, 184, 201, 204, 205, 206, 208,
220, 223, 229, 230, 234
[T06] H. Triebel. Theory of function spaces III. Monogr. Math. 100, Birkhäuser, Basel, 2006 v,
vi, 2, 3, 5, 7, 13, 14, 15, 16, 17, 18, 19, 29, 30, 31, 33, 35, 38, 40, 41, 42, 43, 44, 46, 48,
52, 57, 80, 82, 85, 86, 87, 89, 91, 101, 108, 109, 111, 118, 123, 124, 125, 128, 145, 148,
152, 160, 178, 185, 199, 200, 201, 213, 215, 216, 220, 223, 227, 233, 234, 235, 236, 237,
238
[Tri73] H. Triebel. Boundary values for Sobolev spaces with weights. Density of D./ in
1 
s
Wp; 0 ;:::;r
s
./ and in Hp; 0 ;:::;r
./ for s > 0 and r D Œs  p . Ann. Scuola Norm.
Sup. Pisa 27 (1973), 73–96. I 225
[Tri78] H. Triebel. Spaces of Besov-Hardy-Sobolev type. Teubner-Texte zur Mathematik, Teub-
ner, Leipzig, 1978. I 146
[Tri83] H. Triebel. Periodic spaces of Besov-Hardy-Sobolev type and related maximal inequal-
ities for trigonometrical polynomials. In Functions, Series, Operators, Budapest 1980.
Colloquia Math. Soc. Janos Bolyai 35, II. North-Holland, Amsterdam, New York, 1983,
1201–1209. I 20, 22
[Tri86] H. Triebel. Analysis and mathematical physics. Math. Appl. (East European Ser.) 24,
Reidel Publ. Comp., Dordrecht, 1986. I 133
s
[Tri89] H. Triebel. Atomic decompositions of Fpq spaces. Applications to exotic pseudodiffer-
ential and Fourier integral operators. Math. Nachr. 144 (1989), 189–222. I 220
[Tri92] H. Triebel. Higher analysis. Hochschulbücher für Math., Barth, Leipzig, 1992. I 48
[Tri02] H. Triebel. Function spaces in Lipschitz domains and on Lipschitz manifolds. Charac-
teristic functions as pointwise multipliers. Rev. Mat. Complutense 15 (2002), 475–524.
I 89, 109, 111, 148, 184
[Tri03] H. Triebel. Non-smooth atoms and pointwise multipliers in function spaces. Ann. Mat.
Pura Appl. 182 (2003), 457–486. I 86, 87, 89
250 Bibliography

[Tri04] H. Triebel. A note on wavelet bases in function spaces. In Orlicz Centenary Volume,
Banach Center Publ. 64, Polish Acad. Sci., Warszawa, 2004, 193–206. I 14, 16, 17
[Tri05] Sampling numbers and embedding constants. Trudy Mat. Inst. Steklov 248 (2005),
275–284. [Proc. Steklov Inst. Math. 248 (2005), 268–277]. I 236
[Tri07a] H. Triebel. Wavelets in function spaces on Lipschitz domains. Math. Nachr. 280 (2007),
1205–1218. I 31, 33, 41, 52
[Tri07b] H. Triebel. Wavelet para-bases and sampling numbers in function spaces on domains.
J. Complexity 23 (2007), 468–497. I240–242 12, 31, 33, 42, 52, 236, 240, 241, 242
[Tri07c] H. Triebel. Local means and wavelets in function spaces. In Function Spaces 8,
Be˛dlewo, 2006. Banach Center Publ. 79, Polish Acad. Sci., Warszawa, 2008, 215–234.
I 12, 17, 27
[Tri07d] H. Triebel. The dichotomy between traces on d -sets in Rn and the density of
D.R n / in function spaces. Acta Math. Sinica 24 (2008), 539–554. I 150, 225
n

[Tri08] H. Triebel. Wavelet bases in Lorentz and Zygmund spaces. Georgian Math. J. 15 (2008),
389–402. I 61, 62, 63, 64
[TrW96] H. Triebel, H. Winkelvoss. Intrinsic atomic characterizations of function spaces on
domains. Math. Z. 221 (1996), 647–673. I 77, 78, 119, 120, 176, 223
[Vyb07a] J. Vybíral. Sampling numbers and function spaces. J. Complexity 23 (2007), 773–792.
I 236
[Vyb07b] J. Vybíral. Widths of embeddings in function spaces. J. Complexity 24 (2008),
545–570. I 127
[Vyb08a] J. Vybíral. On dilation operators and sampling numbers. J. Funct. Spaces Appl. 6
(2008), 17–46. I 89, 236
[Vyb08b] J. Vybíral. A new proof of the Jawerth-Franke embedding. Rev. Mat. Complutense 21
(2008), 75–82. I 231
[Wen01] H. Wendland. Local polynomial reproduction and moving least squares approximation.
IMA J. Numer. Anal. 21 (2001), 285–300. I 237
[Woj91] P. Wojtaszczyk. Banach spaces for analysts. Cambridge Stud. Adv. Math. 25, Cam-
bridge University Press, Cambridge, 1991. I 199
[Woj97] P. Wojtaszczyk. A mathematical introduction to wavelets. London Math. Soc. Stud.
Texts 37, Cambridge University Press, Cambridge, 1997. I 13, 14, 27, 56
[Woj03] P. Wojtaszczyk. Greedy type bases in Banach spaces. In Constructive Theory of Func-
tions, Darba, Sofia, 2003, 136–155. I 215
[Woj06] P. Wojtaszczyk. Greediness of the Haar system in rearrangement invariant spaces. In
Approximation and Probability, Banach Center Publ. 72, Polish Acad. Sci., Warszawa,
2006, 385–395. I 215
[Wol82] T. H. Wolff. A note on interpolation spaces. In Harmonic Analysis, Lecture Notes in
Math. 908, Springer-Verlag, Berlin, 1982, 199–204. I 110, 111
[Yos80] K. Yosida. Functional analysis. Grundlehren Math. Wiss. 123, Springer-Verlag, Hei-
delberg, 1980 (sixth ed.). I 48, 199
Symbols

Sets bpq , 4
s
Ap .Rn /, 221 bpq , 14
s
B.x; %/, 31 bpq .Z /, 33
Bn , 163 s
bpq .Z /, 151, 178
C, 1 s
bpq .Z /, 136
H  , 59 s;per
bpq , 24
fF; M gn , 13
fF; M gn , 13 bNpq
s
,6
s
N, 1 bpq .w/, 18
x 31
N, fpq , 4
N0 , 1 fpqs
, 14
Nn0 , 1 fpqs
.Z /, 33
Nnl , 185 s
fpq .Z /, 151, 178
f .%/, 61 s
fpq .Z /, 136
x ı , 73
./ s;
c D Rn n , 101 fpq , 43
s;per
Pjn , 23 fpq , 24
Q, rQ, 4 fNpq
s
,6
Qj m , 4 fpqs
.w/, 18
Qe , 119  N 
R, 1 `q `p j , 98
 M 
Rn , 1 `q 2j ı `p j , 125
RnC , 114 .A0 ; A1 / ;q , 108
S  , S ;1 , S ;2 , 42 ŒA0 ; A1 , 109
Sn , 163 s
Apq .Rn /, 14
T n , 19 s;per
Apq .Rn /, 23
U , 34 s;per
UA , 123 Apq .Rn ; w /, 20
s
Z, 1 Apq ./, 28
s
Zn , 1 Apq . /, 133
Z , 31 s
Apq ./rN , 157
Z , 151, 178 ı
s
Z , 134 Apq ./, 147
‰ 1; , ‰ 2; , 42 AOpq
s
./, 180
AQpq
s
./, 28
Spaces
Qs x 28
Apq ./,
apq .Z /, 119
s
apq .Z /, 114, 115, 117 AQpq
s
.c /, 102
s
apq .w/, 18 Axpq
s
./, 80
252 Symbols

s;
Apq ./, 232 xps ./, 111
H
s
Bpq .Rn /, 2 Hps;rinf ./, 207
s;per
Bpq .Rn /, 22 H 1;rloc , 208
s
Bpq .T n /, 20, 21 H 1 .Q; 0 /, 209
Bpq .Rn ; w/, 17
s H 1 .Qn ; n2 /, 210
s;per idAQpq
s
./, 84
Bpq .Rn ; w /, 21
z s
idBpq ./, 84
Bpq .Rn /, 200
s
s
Bpq ./, 29, 30 idFzpqs
./, 84
z s
Bpq ./, 28
loc
L1 ./, 53
Lp .Rn /, 1
s x
Bzpq ./, 28
Lp ./, 28
x s
Bpq .U /, 34 Lp . /, 130, 133
Bxpq
s
./, 79, 111 Lr . /, 217
C./, 147 Lr . ; /, 216
C loc ./, 239 Lp;q ./, 61
C u .R/, 13 Lp .log L/a ./, 61
C u ./, 147 ƒp;q .Z /, 62
C 1 ./, 147 ƒp .log ƒ/a .Z /, 62
C u . /, 135 S.Rn /, 1
C ~ .Rn /, 35 S./, 111
C s .Rn /, 3 S 0 .Rn /, 1
C s .T n /, 21 S 0 ./, 120
0
D./, 28 S per .Rn /, 20
D , 219 Wpk .Rn /, 2
D 0 ./, 28 Wpk ./, 65
D 0 .T n /, 19, 20, 21 Wpk ./, 65
s
Fpq .Rn /, 2 Wpk .T n /, 21
Fpq .T n /, 21
s
Wpk;rloc ./, 39
s
Fpq .Rn ; w/, 18 zpk ./, 120
W
s;per
Fpq .Rn /, 22
s;per
Fpq .Rn ; w /, 21 Operators
Fpq
s
.Rn /, 201 ext, Ext, 65, 87, 101
Fpqs
./, 29 ext per , 20
z s ext u , 103, 107
Fpq ./, 28
ext " , 105, 106, 107
s x
Fzpq ./, 28 ext r;u
 , 143, 206
s;rinf
Fpq ./, 205 Extpk , 71
Fxpq
s
.U /, 34 Ext r;u , 187
l
Fxpq
s
./, 79 Ext r;u
 , 142, 143
s;rloc N
Fpq ./, 36 Ext r;u
 , 158, 194
s n
Hp .R /, 3 P r;u , 146
 Symbols 253

Qr;u , 148 ˆ , 154

tr l , 185, 193 ˆr , 155
tr  , 131, 220 ˆ , 154
tr , 218 ˆ;r , 154
tr rl , 186, 193 ˆjr , 32
tr r , 140 F ; M , 13
tr rN , 193 L L
F , M , 31
tr s;p
 , 131 ‰  , 154
re, 101 ‰F ;m , 237
gj, 28 j
‰G;m , 13, 41
I , 2 j;L
y F ', 1
', ‰G;m , 23, 31
j
' _ , F 1 ', 1 ‰G;m;per , 23
kT k, 123 j;per
‰G;m , 24
L.A; B/, 123
rank T , 123 Numbers, relations

Functions, functionals D.Ap .Rn /; /, 229

D.Ap .Rn /; Lp . //, 221
xˇ , 1
ak .T /, 123
hi, 3
ek .T /, 123
j m , 4, 6
gk .id/, 234
j.p/
m, 4 g k .id/, 234
jr , 32
hk .T /, 128
lh , 3
GN .b/, 211
lh; , 30 sN .b/, 211
M
d t;u f , 38 bC D max.b; 0/, 5
f  .t/, 61 ai  bi , 69
j
HG;m , 59 p , pq , 5, 79, 103
k.f /; kj m .f /, 7 dist. 1 ; 2 /, 30
M g, 10 ,! , 83
ı.x/, 120
Index

algorithm, greedy, 211 domain, John, 77

approximation number, 123 domain, Lipschitz, special, 72
approximation, N -term, 211 domain, Lipschitz, bounded, 72
approximation problem, inverse, 129 domain, plump, 77
atlas, 132 domain, snowflake, 75
atom, .s; p/, 4, domain, ."; ı/, 70, 71, 177
atom, .s; p/ , 134 domain, planar, 161
atom, -.s; p/, 119 domain, E-porous, 85
atom, .s; p/K;L;d , 5 domain, uniformly E-porous, 85
atom, .s; p/K;L;d;D , 134 domain, E-thick, 70
domain, I -thick, 70
basis, 15, 210 domain, thick, 70
basis, democratic, 211 duality, 97, 98, 99
basis, greedy, 211
basis, orthonormal, 13, 26 English Dictionary, Oxford, 117
basis, Riesz, 159 entropy number, 123
basis, Schauder, 199 entropy problem, inverse, 129
basis, standard, 211 evaluation, pointwise, 228
basis, unconditional, 15, 210 extension operator, 101,
extension operator, common, 103, 142
Churchill, Sir Winston, 212 extension operator, universal, 107
complemented subspace, 102 extension problem, 65, 70, 101
constraint, 115 extension, wavelet-friendly, 141, 186, 187
convergence, local, 15
formula, polynomial reproducing, 238
convergence, unconditional, 15
Fourier transform, 1
corkscrew property, 77
frame, 138
cube, complementing, 70
frame representation, 45
cube, reflected, 70
function, Lipschitz, 72
curve, Koch, 91
function, scaling, 13
curve, snowflake, 91
Haar system, 214
d -set, 85, 106, 217 Haar wavelet, 55
decomposition, optimal, 8, 10 homogeneity, 34, 92
decomposition, Whitney, 30, 41, 49
dense, 98, 99 inequality, Hardy, 150, 204, 208, 210
dichotomy, 215, 221, 229 information, standard, 236
diffeomorphism, 182 interpolation, complex, 88, 109
domain D open set interpolation, real, 88, 109
domain, C 1 , 72 interpolation theorem, Wolff, 110
domain, cellular, 73, 169, 182 intrinsic, 117
256 Index

kernel, 6 restriction, 28
Koch curve, 91 restriction operator, 101
resolution of unity, 1
lattice, approximate, 31
Lebesgue point, 131 set, porous, 72
lift, 3 set, uniformly porous, 85
local means, 6 space, Besov, classical, 3
localisation, refined, 36, 79 space, Hölder–Zygmund, 3
space, Hölder–Zygmund, periodic, 22
manifolds, C 1 , 132 space, Lorentz, 60
manifolds, cellular, 169 space, periodic, 19
maximal function, 10 space, refined localisation, 36
maximal inequality, 10 space, reinforced, 204, 205, 209
measure, Hausdorff, 220 space, Sobolev, 3
measure, isotropic, 85, 216 space, Sobolev, classical, 2
multi-index, 1 space, Sobolev, on domains, 65
multiplier, pointwise, 35, 36, 86 space, Sobolev, periodic, 21
space, Sobolev, intrinsic, 117
number, approximation, 123 space, weighted, 17
number, entropy, 123 space, Zygmund, 61
number, Gelfand, 127 stability, 138, 199
number, Kolmogorov, 127 subspace, complemented, 83, 102
number, s-, 127
number, sampling, 233, 234, trace, 140, 218
241
u-Riesz basis, 180
operator, compact, 123 u-Riesz basis, interior, 180
operator, linear bounded, 123 u-Riesz basis, oscillating, 180
operator, trace, 131 u-Riesz frame, 181, 202
optimality, 138, 200 u-Riesz frame, interior, 202
oscillation, 152 u-Riesz frame, oscillating, 202
u-wavelet basis, 49
Paley–Littlewood theorem, 2 u-wavelet system, 32, 135, 152, 179
para-basis, 43, 45 u-wavelet system, interior, 179
porosity, 85, 86 u-wavelet system, oscillating, 135, 152,
porous, 72 179
porous, E-, 85
porous, uniformly, 85 wavelet, 13
projection, 83, 102 wavelet, basic, 32
property, refined localisation, 80 wavelet, boundary, 32
wavelet, interior, 32
rearrangement, 61 wavelet expansion, constrained, 65, 67,
regular, 119 115
regular, exterior, 119 weight, 17
regular, interior, 119 widths, 127