Construction of Wavelets Through Walsh Functions: Yu. A. Farkov Pammy Manchanda Abul Hasan Siddiqi

Industrial and Applied Mathematics
Yu. A. Farkov
Pammy Manchanda
Abul Hasan Siddiqi
Construction
of Wavelets
Through Walsh
Functions
Editor-in-chief
Abul Hasan Siddiqi, Sharda University, Greater Noida, India
Editorial Board
Zafer Aslan, Istanbul Aydin University, Istanbul, Turkey
Martin Brokate, Technical University, Munich, Germany
N.K. Gupta, Indian Institute of Technology Delhi, New Delhi, India
Akhtar A. Khan, Rochester Institute of Technology, Rochester, USA
René Pierre Lozi, University of Nice Sophia-Antipolis, Nice, France
Pammy Manchanda, Guru Nanak Dev University, Amritsar, India
Zuhair Nashed, University of Central Florida, Orlando, USA
Govindan Rangarajan, Indian Institute of Science, Bengaluru, India
Katepalli R. Sreenivasan, NYU Tandon School of Engineering, Brooklyn, USA
The Industrial and Applied Mathematics series publishes high-quality research-level
monographs, lecture notes and contributed volumes focusing on areas where
mathematics is used in a fundamental way, such as industrial mathematics,
bio-mathematics, financial mathematics, applied statistics, operations research and
computer science.
More information about this series at http://www.springer.com/series/13577

Yu. A. Farkov Pammy Manchanda
• •
Abul Hasan Siddiqi
Construction of Wavelets
Through Walsh Functions
123
Yu. A. Farkov Pammy Manchanda
Department of Applied Information Department of Mathematics
Technology, School of Public Policy Guru Nanak Dev University
Russian Presidential Academy of National Amritsar, Punjab, India
Economy and Public Administration
(RANEPA)
Moscow, Russia
Abul Hasan Siddiqi

School of Basic Sciences and Research
and Centre for Advanced Research in
Applied Mathematics and Physics
Sharda University
Greater Noida, Uttar Pradesh, India
ISSN 2364-6837 ISSN 2364-6845 (electronic)

ISBN 978-981-13-6369-6 ISBN 978-981-13-6370-2 (eBook)
https://doi.org/10.1007/978-981-13-6370-2
Library of Congress Control Number: 2019930359
© Springer Nature Singapore Pte Ltd. 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface
The book is essentially devoted to the construction of wavelets by Walsh functions

and their transforms. This embodies the contemporary research on wavelets defined
on positive half line, wavelets on Vilenkin groups, Haar–Vilenkin wavelets, studied
by the authors of this book and their collaborators. More precisely, this book deals
with wavelet theory on Vilenkin groups, particularly on the Cantor group, which
can also be interpreted as wavelet theory on the positive half line. This theory is in
the framework of Walsh analysis (dyadic analysis in the case of the Cantor group),
which has been actively studied for the last four decades. The foundation of these
developments is discussed in books authored by Schipp, Wade, and Simon;
Golubov, Efimov, and Skvortsov; Maqusi; Siddiqi; and Beauchamp. Recent
research work not included in these books constitute important ingredients of the
present volume. These themes are an approximation of special cases of matrix
transform of Walsh–Fourier series in dyadic homogeneous Banach spaces and
Hardy spaces, Haar–Vilenkin wavelets and their properties, orthogonal and
biorthogonal wavelets and frames, nonuniform multiresolution analysis on the
positive half line, and vector-valued wavelets on the positive half line.
The book contains nine chapters and one appendix. Chapter 1 is devoted to the
basic results of Walsh analysis and wavelets on the real line. An exhaustive
background for Chaps. 1–3 can be found in Schipp et al. [1]. It has been observed
that Walsh functions were used for the transposition of conductors in open wire
lines as early as the late 1800s, and a complete system of Walsh functions seems to
have been found around 1900. Rademacher functions were introduced to the
mathematical world in 1922 by Rademacher. His widow told an audience of
mathematicians that Rademacher knew the concept of Walsh functions, but he
abandoned the idea as advised by his academic mentors. The concept of wavelets
was introduced by the French Geophysicist, Jean Morlet, in 1982. Its literal
meaning is “small wave”. Alex Grossmann, French Theoretical Physicist, studied
inverse formula for the wavelet transforms. In 1984, the collaboration of Morlet and
Grossmann yielded a detailed mathematical study of the continuous wavelet
transforms and their various applications, of course, without realizing that similar
results had already been obtained 20–50 years earlier by Calderon, Littlewood,
v
vi Preface
Paley, and Franklin. In 1985, Yves Meyer, recipient of the Abel Prize of 2017,
accidentally found the then existing literature of wavelets. With the knowledge
of the Calderon–Zygmund operators and the Littlewood–Paley theory, he was able
to lay the mathematical foundation of wavelet theory. The first major achievement
is due to Daubechies, Grossman, and Meyer for constructing a painless
non-orthogonal expansion in 1986.
Chapter 2 deals with the properties of Walsh–Fourier coefficients, the convergence
of Walsh–Fourier series, and an approximation of functions belonging to different
spaces by special classes of matrix transform of Walsh–Fourier series. The historical
note on most of the themes of this chapter is well documented in the monograph of
Schipp, Wade, and Simon. Here, we discuss approximation by matrix transform of
Walsh–Fourier series. After the result of Yano for arithmetic and Cesáro means, there
were no significant results on this theme till 1992. During the visit of Schipp in
January 1992 and Moricz in 1994 to Aligarh Muslim University, India, the problem
of generalizing the results of Yano, Jastrebova, and Skvorcov on the rate of
approximation by Cesáro means was given top priority. Moricz and Siddiqi published
an interesting paper generalizing all earlier results (see Theorems 2.24–2.26). Fridli,
Manchanda, and Siddiqi studied approximation by Norlund means in dyadic
homogeneous Banach spaces and Hardy spaces and extended the earlier results of
Moricz and Siddiqi (see Theorems 2.27–2.29).
Chapter 3 presents Haar systems and its generalization, Haar–Fourier series, the
relation between Haar and Walsh functions, and applications of Haar wavelets to
initial and boundary value problems. Alfred Haar was a Ph.D. student of David
Hilbert at Göttingen University. Haar function is the simplest example of mother
wavelet. Haar wavelets and its generalizations were studied in detail by Devore
et al. [2] and Dubeau et al. [3]. In a recent monograph, Lepik and Hein [4] have
studied the applications of Haar wavelets to initial and boundary problems.
Chapter 4 is devoted to the construction of dyadic wavelets and frames through
Walsh functions. This chapter is based on the research work of Farkov [5–12].
In Sect. 4.1 of this chapter, preliminary results needed for the subsequent chapters
are explained. Section 4.2 of this chapter presents the orthogonal wavelets and
multiresolution (MRA). Section 4.3 of this chapter is devoted to orthogonal wavelets
with compact support on the positive half line. Section 4.4 introduces the estimates
of the smoothness of the scaling functions. Section 4.5 presents the approximation
properties of dyadic wavelets.
Chapter 5 is devoted to orthogonal and periodic wavelets on Vilenkin groups.
This chapter is essentially based on the research work of Farkov and his collaborators.
As we know, Walsh functions are identified with characters of the Cantor dyadic
group. This fact was recognized by Gelfand as back as in 1940. Vilenkin group was
introduced in 1947. In a series of papers published after 2005, Farkov and his
research collaborators have studied orthogonal and periodic wavelets [6, 13–19].
The results of Chap. 4 and 5 may be extended for results given in [20]. Section 5.1
deals with multiresolution analysis on Vilenkin groups. Section 5.2 presents com-
pactly supported orthogonal p-wavelets. Section 5.3 introduces the periodic wavelets
Preface vii
on Vilenkin groups. Section 5.4 introduces periodic wavelets related to the Vilenkin–
Christenson transform. Section 5.5 presents applications to the coding of fractal
functions.
Chapter 6 is devoted to Haar–Vilenkin wavelets and their properties. Haar–Vilenkin
system was introduced around 1947. The concept of Haar–Vilenkin wavelets was
introduced and studied by Manchanda, Meenakshi, and Siddiqi [21, 22].
Chapter 7 is devoted to biorthogonal wavelets and frames on Vilenkin groups.
This chapter is based mainly on the research papers of Farkov published between
2008 and 2012; [14, 23, 24]. Section 7.1 introduces biorthogonal wavelets on the
positive half line. Section 7.2 presents biorthogonal wavelets on Vilenkin groups.
Parseval frames on Vilenkin groups are also presented in this chapter. This chapter
is concluded by applications of biorthogonal dyadic wavelets.
Chapter 8 deals with wavelets associated with nonuniform multiresolution analysis
on the positive half line. This chapter is based on the research of Manchanda,
Meenakshi, Siddiqi, and Vikram published during 2012 and 2017 [25–28].
Chapter 9 presents the concept of orthogonal vector-valued wavelets on the
positive half line. Section 9.4 presents vector-valued wavelets and wavelet packets
associated with the nonuniform multiresolution analysis. The study of the concept
of orthogonal vector-valued wavelets on the real line was initiated by Xia and Suter
(1996). This concept on the positive half line was initiated by Farkov in 2005 [16].
The chapter contains the results obtained by Manchanda, Meenakshi, Siddiqi, and
Vikram in the recent past [29–32].
The appendix contains topics on the basic results of functional analysis, Vilenkin
systems, Vilenkin–Pontryagin class of functions, and real and Fourier analysis
including pointwise, uniform, and absolute convergence.
The book will serve as a reference material for research in Fourier analysis.
The book, however, could also be used for a special course for graduate and
advanced undergraduate students of mathematics and engineering. We take this
opportunity to express our gratitude to our research groups. We also like to express
our sincere gratitude to the chancellor of Sharda University, P. K. Gupta, for
extending full support during the preparation of this book. We also take this
opportunity to thank Pooja and Mamta for meticulously preparing the final
manuscript. We take this opportunity to thank the World Scientific Publishing
Company for using results published in International Journal of Wavelets,
Multiresolution and Information Processing.
Moscow, Russia Yu. A. Farkov

Amritsar, India Pammy Manchanda
Greater Noida, India Abul Hasan Siddiqi
viii Preface
References
1. Schipp, F., Wade, W. R., & Simon, P. (1990). Walsh series. Adam Hilger.
2. Pfander, G. E., & Benedetto J. J. (2000). Periodic wavelet transforms and periodicity
detection. SIAM Journal of Applied Mathematics, 62(4), 1329–1368.
3. Dubeau, F., Elmejdani, S., & Krantini, R. (2004). Nonuniform haar wavelets. Applied
Mathematics and Computation, 159, 675–693.
4. Lepik, U., & Hein, H. (2014). Haar wavelets with applications. Berlin: Springer.
5. Farkov, Yu. A., & Protasov V. Yu. (2006). Dyadic wavelets and refinable functions on a half
line. Matematicheskii Sbornik, 197(10), 129–160. English translation, Sbornik: Mathematics,
197, 1529–1558.
6. Farkov, Yu. A. (2005). Orthogonal wavelets with compact support on locally compact abelian
groups. Izv. Ross. Akad. Nauk Ser. Mat. 69(3), 193–220. English translation, Izvestia:
Mathematics, 69(3), 623–650.
7. Farkov, Yu. A. (2009). On wavelets related to Walsh series. Journal of Approximation
Theory, 161, 259–279.
8. Farkov, Yu. A. (2010). Wavelets and frames based on Walsh-Dirichlet type kernels.
Communications in Mathematics and Applications, 1, 27–46.
9. Farkov, Yu. A., & Stroganov, S. A. (2011). The use of discrete dyadic wavelets in image
processing. Russian Mathematics (Iz. Yuz), 55(7), 47–55. Original Russian text published in
Izvestiya. Uchebnykh Zavednic. Mathematica, 7, 57–66.
10. Farkov, Yu. A. (2012). Periodic wavelets in Walsh analysis. Communication in Mathematics
and Applications, 3(3), 223–242.
11. Farkov, Yu. A., & Borisov, M. E. (2012). Periodic dyadic wavelets and coding of fractal
functions. Russian Mathematics (Izvestiya Vuz. Matematika), 56(9), 46–56.
12. Farkov, Yu. A. (2014). Wavelet expansions on the Cantor group. Mathematical Notes, 96(6),
996–1007.
13. Farkov, Yu. A. (2007). Orthogonal wavelets on direct products of cyclic groups.
Matematicheskie Zametki, 82(6), 934–952. English translation, Mathematical Notes, 82(6),
843–859.
14. Farkov, Yu. A. (2008). Multiresolution analysis and wavelets on Vilenkin groups. Facta
Univers. (Nis) ser.: Elec. Engineering, 21(3), 309–325.
15. Farkov, Yu. A. (2009). Biorthogonal wavelets on Vilenkin groups. Tr. Mat. Inst. Steklova,
265(1), 110–124. English translation, Proceedings of the Steklov Institute of Mathematics,
265(1), 101–114.
16. Farkov, Yu. A. (2005). Orthogonal p-wavelets on R+. In Proceedings of International
Conference on Wavelets and Spilnes, St. Petersburg, Russia, July 3–8 (p. 426). St. Petersberg:
St. Petersberg University Press.
17. Farkov, Yu. A. (2011). Discrete wavelets and the Vilenkin Chrestensen transform.
Mathematical Notes, 89(6), 871–884.
18. Farkov, Yu. A. (2011). Periodic wavelets on the p-adic Vilenkin group. P-Adic Numbers,
Ultrametric Analysis, and Applications, 3(4), 281–287.
19. Farkov, Yu. A., & Rodionov, E. A. (2011). Algorithms for wavelet construction on Vilenkin
groups. P-Adic Numbers, Ultrametric Analysis, and Applications, 3(3), 181–195.
20. Krivoshein, A. V., Protasov, V. Yu., & Skopina, M. A. (2016). Multivariate wavelet frames.
Berlin: Springer.
21. Manchanda, P., Meenakshi, & Siddiqi, A. H. (2008). Haar-Vilenkin wavelet. The Aligarh
Bulletin of Mathematics, 27(1), 59–73.
22. Manchanda, P., & Meenakshi. (2009). New classes of wavelets. In A. H. Siddiqi,
A. K. Gupta, & M. Brokate (Eds.), Proceedings of Conference on Modeling of Engineering
and Technological Problems, New York (vol. 1146, pp. 253–271).
Preface ix
23. Farkov, Yu. A., Maksimov A. Yu., & Stroganov, S. A. (2011). On biorthogonal wavelets
related to the Walsh functions. International Journal of Wavelets, Multiresolution and
Information Processing, 9(3), 485–499.
24. Farkov, Yu. A. (2012). Examples of frames on the Cantor dyadic group. Journal of
Mathematical Sciences, 187(1), 22–34.
25. Manchanda, P., & Sharma, V. (2013). Wavelet packets with nonuniform multiresolution
analysis on positive half-line. Asian-European Journal of Mathematics, 6(1), 1350007.
26. Meenakshi, Manchanda, P., & Siddiqi, A. H. (2012). Wavelets associated with nonuniform
multiresolution analysis on positive half line. International Journal of Wavelets,
Multiresolution and Information Processing, 10(2), 1250018, 1–27.
27. Sharma, V., & Manchanda, P. (2015a). Nonuniform wave packet frames in L2(R). Indian
Journal of Industrial and Applied Mathematics, 6(2), 139–152.
28. Sharma, V., & Manchanda, P. (2015b). Nonuniform wavelet frames in L2(R). Asian European
Journal of Mathematics, World Scientific, 8(2), 1550034(1–15).
29. Manchanda, P., & Sharma, V. (2012). Orthogonal vector valued wavelets on R+.
International Journal of Pure and Applied Mathematics, 75(4), 493–510.
30. Manchanda, P., Meenakshi, & Siddiqi, A. H. (2014). Construction of vector-valued
nonuniform wavelets and wavelet packets. Special Issue of Indian Journal of Industrial
and Applied Mathematics.
31. Meenakshi, Manchanda, P., & Siddiqi, A. H. (2014). Wavelets associated with vector valued
non-uniform multiresolution analysis. Applicable Analysis: An International Journal, 93(1),
84–104.
32. Meenakshi & Manchanda, P. (2017). Vector-valued nonuniform wavelet packets.
Numerical Functional Analysis and Optimization. Published online: https://doi.org/
10.1080/01630563.2017.1355814 (Taylor and Francis).
Contents
1 Introduction to Walsh Analysis and Wavelets . . . . . . . . . . . . . . . . . 1

1.1 Walsh Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Walsh–Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Haar Functions and Its Relationship with Walsh Functions . . . . . . 8
1.4 Walsh-Type Wavelet Packets . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Wavelet Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5.1 Continuous Wavelet Transform . . . . . . . . . . . . . . . . . . . . 12
1.5.2 Discrete Wavelet System . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5.3 Multiresolution Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6 Wavelets with Compact Support . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2 Walsh–Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1 Walsh–Fourier Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.1 Estimation of Walsh–Fourier Coefficients . . . . . . . . . . . . . 28
2.1.2 Transformation of Walsh–Fourier Coefficients . . . . . . . . . . 31
2.2 Convergence of Walsh–Fourier Series . . . . . . . . . . . . . . . . . . . . . 34
2.2.1 Summability in Homogeneous Banach Spaces . . . . . . . . . . 39
2.3 Approximation by Transforms of Walsh–Fourier Series . . . . . . . . 40
2.3.1 Approximation by Césaro Means of Walsh–Fourier
Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 40
2.3.2 Approximation by Nörlund Means of Walsh–Fourier
Series in Lp Spaces . . . . . . . . . . . . . . . . . . . . . . . . . .... 45
2.3.3 Approximation by Nörlund Means in Dyadic
Homogeneous Banach Spaces and Hardy Spaces . . . . . . . . 57
2.4 Applications to Signal and Image Processing . . . . . . . . . . . . . . . . 70
2.4.1 Image Representation and Transmission . . . . . . . . . . . . . . 70
2.4.2 Data Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.4.3 Quantization of Walsh Coefficients . . . . . . . . . . . . . . . . . . 74
xi
xii Contents
2.4.4 Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

2.4.5 ECG Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.4.6 EEG Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.4.7 Speech Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2.4.8 Pattern Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3 Haar–Fourier Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.1 Haar System and Its Generalization . . . . . . . . . . . . . . . . . . . . . . . 79
3.2 Haar Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.3 Haar System as Basis in Function Spaces . . . . . . . . . . . . . . . . . . 86
3.4 Non-uniform Haar Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.5 Generalized Haar Wavelets and Frames . . . . . . . . . . . . . . . . . . . . 93
3.6 Applications of Haar Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.6.1 Applications to Solutions of Initial and Boundary
Value Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.6.2 Applications to Solutions of Integral Equations . . . . . . . . . 96
3.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4 Construction of Dyadic Wavelets and Frames Through
Walsh Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.1 Preliminary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.2 Orthogonal Wavelets and MRA in L2 ðR þ Þ . . . . . . . . . . . . . . . . . 103
4.3 Orthogonal Wavelets with Compact Support on R þ . . . . . . . . . . . 114
4.4 Estimates of the Smoothness of the Scaling Functions . . . . . . . . . 133
4.5 Approximation Properties of Dyadic Wavelets . . . . . . . . . . . . . . . 158
4.6 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
5 Orthogonal and Periodic Wavelets on Vilenkin Groups . . . . . . . . . . 171
5.1 Multiresolution Analysis on Vilenkin Groups . . . . . . . . . . . . . . . . 171
5.2 Compactly Supported Orthogonal p-Wavelets . . . . . . . . . . . . . . . . 194
5.3 Periodic Wavelets on Vilenkin Groups . . . . . . . . . . . . . . . . . . . . . 205
5.4 Periodic Wavelets Related to the Vilenkin–Christenson
Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
5.5 Application to the Coding of Fractal Functions . . . . . . . . . . . . . . 226
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
6 Haar–Vilenkin Wavelet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
6.2 Haar–Vilenkin Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
6.2.1 Haar–Vilenkin Mother Wavelet . . . . . . . . . . . . . . . . . . . . 236
6.3 Approximation by Haar–Vilenkin Wavelets . . . . . . . . . . . . . . . . . 243
Contents xiii
6.4 Covergence Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246

6.5 Haar–Vilenkin Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
7 Construction Biorthogonal Wavelets and Frames . . . . . . . . . . . . . . . 253
7.1 Biorthogonal Wavelets on R þ . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
7.2 Biorthogonal Wavelets on Vilenkin Groups . . . . . . . . . . . . . . . . . 263
7.3 Construction of Biorthogonal Wavelets on The Vilenkin
Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
7.4 Frames on Vilenkin Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
7.5 Application to Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . 277
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
8 Wavelets Associated with Nonuniform Multiresolution Analysis
on Positive Half Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
8.2 Nonuniform Multiresolution Analysis on Positive Half Line . . . . . 285
8.2.1 Construction of Nonuniform Multiresolution Analysis
on Positive Half Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
8.2.2 The Analogue of Cohen’s Condition . . . . . . . . . . . . . . . . . 297
8.3 Nonuniform Wavelet Frames in L2 ðRÞ . . . . . . . . . . . . . . . . . . . . . 303
8.3.1 Necessary and Sufficient Condition . . . . . . . . . . . . . . . . . . 307
8.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
9 Orthogonal Vector-Valued Wavelets on R þ . . . . . . . . . . . . . . . . . . . 317
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
9.2 Vector-Valued Multiresolution Analysis . . . . . . . . . . . . . . . . . . . . 318
9.3 Vector-Valued Multiresolution p-Analysis on R þ . . . . . . . . . . . . 321
9.4 The Existence of Orthogonal Vector-Valued Wavelets on R þ . . . 326
9.5 Vector-Valued Nonuniform Multiresolution Analysis . . . . . . . . . . 334
9.5.1 Construction of Vector-Valued Nonuniform
Multiresolution Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 343
9.6 Vector-Valued Nonuniform Wavelet Packets and Its
Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
9.6.1 Vector-Valued Nonuniform Wavelet Bases . . . . . . . . . . . . 355
9.7 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
Notational Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
About the Authors
Yu. A. Farkov is a distinguished scientist and professor at the Russian Presidential

Academy of National Economy and Public Administration (RANEPA), Moscow,
Russia. Earlier, he worked as a professor and head of the Department of
Mathematics at the Russian State Geological Prospecting University, Moscow,
Russia, from 1997 to 2014; a professor and head of the Department of Higher
Mathematics at Dubna International University, Dubna, Russia, from 1996 to 2000;
an associate professor at the Department of Higher Mathematics and Mathematical
Modelling, Moscow State Geological Prospecting Academy, from 1988 to 1997;
and a lecturer at the Department of Higher Mathematics, Moscow State University
of Mechanical Engineering, Environmental and Chemical Engineering Institute,
Moscow, from 1981 to 1988. In 2013, Prof. Farkov received his DSc-Doctor
Fiz.- Mat. Nauk from the People’s Friendship University of Russia on “Optimal
Methods of Approximation of Function by Generalized Polynomials and
Wavelets”. In 1981, he received his Ph.D. degree in Mathematics from Moscow
Region Pedagogical Institute, on topic “Investigations of Asymptotic and
Approximation Properties of Faber–Erokhin Basis Functions”. From 1977 to 1980,
he did his postgraduate studies at the Moscow Electro-Technical Engineering
Institute; and in 1975, he received his BSc degree in Mathematics from Uralsk
Pedagogical Institute, Kazakhstan. Professor Farkov is on the editorial boards of the
journals American Journal of Computational Mathematics, Communications in
Mathematics and Applications and International Journal of Education. He is a
member of the Moscow Mathematical Society and American Mathematical Society.
His research interests include mathematical analysis, wavelet theory, dyadic
analysis and approximation theory.
Pammy Manchanda is a senior professor of mathematics at Guru Nanak Dev

University, Amritsar, India. She has attended and delivered talks and chaired sessions
at reputed academic conferences and workshops across the world, including ICIAM
(1999–2015) and ICM since 2002. She was invited twice to the Industrial
Mathematics Group of Professor Helmut Neunzert, Kaiserslautern University,
xv
xvi About the Authors
Germany, and visited the International Centre for Theoretical Physics (a UNESCO
institution) at Trieste, Italy, many times to carry out her research activities. She was
the joint secretary of the Indian Society of Industrial and Applied Mathematics
(ISIAM) from 1999 until 2016 and after that, she is the secretary of the society.
She has been actively engaged in organizing international conferences by the society.
She is the managing editor of the Indian Journal of Industrial and Applied
Mathematics (by ISIAM) and a member of the editorial board of the Springer’s book
series, Industrial and Applied Mathematics. Professor Manchanda has published 44
research papers in several international journals of repute, edited 2 proceedings of
international conferences of ISIAM and co-authored 3 books.
Abul Hasan Siddiqi is a distinguished scientist and Adjunct Professor at the

School of Basic Sciences and Research and Coordinator at the Centre for Applied
Mathematics and Physics at Sharda University, Greater Noida, India. He was also a
visiting consultant at the International Centre for Theoretical Physics (ICTP),
Trieste, Italy; Sultan Qaboos University, Muscat, Oman; MIMOS, Kuala Lumpur,
Malaysia; and professor at several reputed universities including Aligarh Muslim
University (Aligarh, India) and King Fahd University of Petroleum & Minerals
(Dhahran, Saudi Arabia). He has a long association with ICTP (a UNESCO
institution) in several capacities: short-time visitor, long-duration visitor, regular
associate, guests of the director and senior associate. He was awarded the German
Academic Exchange Fellowship thrice to carry out mathematical research in
Germany. He has published more than 100 research papers jointly with his research
collaborators, 13 books and edited proceedings of 17 international conferences, as
well as supervised 29 Ph.D. scholars. He is the founder secretary and the elected
president of the Indian Society of Industrial and Applied Mathematics (ISIAM),
which celebrated its silver jubilee year in January 2016. He is the editor-in-chief
of the Indian Journal of Industrial and Applied Mathematics (published by ISIAM)
and the Springer’s book series, Industrial and Applied Mathematics.
Introduction
The trigonometric Fourier series has played a very significant role in solving the
problems of science and technology. The concept of non-trigonometric Fourier
series such as Haar–Fourier series and Walsh–Fourier series was introduced by
Haar [1] and Walsh [2], respectively; Kaczmarz, Steinhaus, and Paley studied some
aspects of Walsh system between 1929 and 1931. Nowadays, Paley’s modification,
which is defined as the product of Rademacher functions, is known as Walsh
functions [3]. A major breakthrough came when N. J. Fine submitted his Ph.D.
dissertation to Pennsylvania University in 1946, which was published subsequently
in the Transactions of the American Mathematical Society in 1949. Fine [4]
introduced the concept of dyadic group G and proved that Walsh functions are its
characters. Haar introduced, in 1909, a class of functions closely related to Walsh
functions in response to a question posed by David Hilbert, namely “Does there
exist an orthonormal system such that the Fourier series of every continuous
function converges uniformly?” He showed that the Fourier series with respect to
the orthonormal system introduced by him, now known as the Haar system, of
continuous functions converges uniformly on [0,1]. Walsh [2] observed that Haar
and Walsh systems are Hadamard transforms of each other. The Walsh and Haar
systems perform all usual applications of orthonormal systems, for example, data
transmission, multiplexing, filtering, image enhancement, and pattern recognition.
An elegant presentation of these developments is given in Schipp et al. [5]; also,
see Beauchamp [6], Siddiqi [7], Maqusi [8], and Golubov et al. [9]. In the early
eighties, wavelet analysis was introduced and developed by Morlet, Daubechies,
Mallat, Meyer, and Coifman et al. [10–15]. Also, see Walnut [16], Christensen [17],
Siddiqi [18], and Krivoshein et al. [19].
Wavelet analysis is a refinement of Fourier analysis. A natural question arises:
What is the relation between Haar Fourier analysis, Walsh–Fourier analysis, and
wavelet analysis? It is known that Haar function is the simplest example of wavelet
and Walsh function is the simplest example of wavelet packet. This book focuses
on the construction of wavelets by Walsh functions and their transforms.
An interpretation of Walsh functions as characters of the Cantor dyadic group
was proposed by I. M. Gelfand, and Vilenkin [20] defined a broad class of locally
xvii
xviii Introduction
compact abelian groups (now called Vilenkin groups), containing the Cantor group
as a special case. Fine [4] independently examined Walsh functions as characters
of the Cantor dyadic group. Recall that for a given p [ 2, the Vilenkin group Gp is
defined as the weak direct product of a countable number of the cyclic groups of
order p equipped with the discrete topology (for p ¼ 2, we have the locally compact
Cantor group). A detailed bibliography of wavelets on locally compact abelian
groups is given in [21, 22], where the group Gp is interpreted as an additive group
of the field Fp ðtÞ of formal Laurent series.
According to [21], the specifics of constructions of wavelets on Vilenkin groups
and on the additive group of the field of p-adic numbers Qp are determined by the
fact that these groups contain compact subgroups. On the real line R, a classical
example of wavelet bases is the Haar system (see, for example, [23]). Haar wavelets
for the group Qp were constructed in [24], and on the groups Gp , in [25, 26]
(for p ¼ 2, see [27, 28]). An analysis of a general construction of Haar wavelets is
given in a recent paper [29]. In the space L2 ðQp Þ, any orthogonal wavelet basis
consisting of bandlimited functions is a modification of the Haar system [30].
As distinct from the group Qp , on Vilenkin groups, one may construct orthogonal
compactly supported wavelets that are substantially different from Haar wavelets
(see examples in [25–27]). We also note that the Fourier transform of the Haar
scaling function / on the group Gp coincides with this function, / ^ ¼ /, and hence,
Haar wavelets on Gp coincide with Kotelnikov–Shannon wavelets (for more details
on this phenomenon, see [21]).
As usual, by Z, Z þ , N, and C, we shall denote, respectively, the sets of integers,
positive integers, natural and complex numbers. The group Gp consists of the
sequences x ¼ ðxj Þ for which xj 2 f0; 1; . . .; p 1g for any j 2 Z, where only a
finite number of xj with negative indexes may be nonzero. For any nonzero
sequence x ¼ ðxj Þ from Gp , there exists a unique integer number k ¼ kðxÞ such
that xk 6¼ 0 and xj ¼ 0 for all j\k. The group operation on Gp is defined as the
coordinatewise addition modulo p, for all p,
ðzj Þ ¼ ðxj Þ þ ðyj Þ , zj ¼ xj þ yj ðmod pÞ for all j 2 Z
and the topology is defined by the complete system of neighborhoods of the origin,
Ul ¼ fðxj Þ 2 Gp : xj ¼ 0 for j lg; ; 2 Z
Let R þ :¼ ½0; 1Þ. With the help of the mapping k : Gp ! R þ ,

X
kðxÞ ¼ xj pj ; x ¼ ðxj Þ 2 Gp ;
j2Z
we interpret in the standard way the group Gp on the half line R þ so that to the
addition operation on Gp there corresponds the operation to be defined below,
and to the Haar measure on Gp there corresponds the Lebesgue measure on R þ
Introduction xix
[5, 9, 31, 32]. The integral and fractional parts of a number x are denoted by [x] and
x, respectively. For any m 2 Z, we denote by hmi ¼ p the remainder of the division
of m by p. Further, for any x 2 R þ , we set
xj ¼ h½p j xip ; xj ¼ h½p1j xip ; j 2 N: ð1Þ
These numbers are the digits of the p-adic expansion

X X
x¼ xj pj1 þ xj pj
j\0 j[0
(for a p-adic rational x, we get an expansion with a finite number of nonzero terms).
It is easily seen that
X
1 X
1
½x ¼ xj pj1 ; fxg xj pj ;
j¼1 j¼1
and moreover, for any x 6¼ 0, there exists a number k ¼ kðxÞ such that xk 6¼ 0 and
xj ¼ 0 for all j [ k. The p-adic addition operation p on R þ is defined by
X X
x p y ¼ hxj þ yj ip pj1 þ hxj þ yj ip pj ; ð2Þ
j\0 j[0
where xj , yj are calculated by (1). As usual, we denote by ⊖p the inverse operation

of p (for p ¼ 2, these operations coincide). Let ¼ 2 . A function w is called an
orthogonal wavelet in L2 ðR þ Þ if the functions
wj;k :¼ 2j=2 wð2 j x kÞ; j 2 Z; k 2 Z þ
form an orthonormal basis for L2 ðR þ Þ. In general, a function w is called a wavelet

in L2 ðR þ Þ if the following condition is satisfied
Z
^ 2 dx
0\ jwðxÞj \1
Rþ x
where w^ is the Walsh–Fourier transform of a function w (for the definition and some
properties of this transform, see Chap. 2). The continuous wavelet transform of a
function f 2 L2 ðR þ Þ with an analyzing wavelet w is defined by
Z
1=2
Ww f ða; bÞ ¼ a f ðxÞwððx bÞ=aÞdx; a [ 0; b 2 R þ
Rþ
An analogue of Grossmann–Morlet’s formula for this transform is proved in

[33]; for the transform Ww ðf Þ with f 2 L2 ðGp Þ, see [5, 34]. First examples of
xx Introduction
orthogonal compactly supported wavelets on the group G2 were constructed by

Lang [27, 35, 36] by using the group analogues of well-known theorems of Mallat
[14] and Cohen [37] on multiresolution analysis (MRA). Lang wavelets on the half
line R þ are constructed by using a function w 2 L2 ðR þ Þ satisfying the scaling
equations of the form
2X
n
1
/ðxÞ ¼ ck /ð2x kÞ; x 2 R þ ð3Þ
k¼0
A complete description of the orthogonal wavelets and scaling functions on R þ ,

whose masks are Walsh polynomials, was obtained in [38]; in [25, 38–42], these
results were extended and algorithms for the construction of biorthogonal Parseval
wavelets and frames for the group Gp were presented. Examples of step scaling
functions on the group Gp are given in [43–46]. For any p 2, necessary and
sufficient conditions for which solutions / of the scaling equation
pX
n
1
/ðxÞ ¼ ck /ðpx kÞ; x 2 R þ
k¼0
generate an MRA in L2 ðR þ Þ are provided in [47, 48]; these results are replicated in
Chap. 8 of [49]. A complete description of MRA wavelets in L2 -spaces on local
fields of positive characteristic (including the field Fp ððtÞÞ and, hence, the group
Gp ) is obtained in [50]; in connection with these results, see [43–46, 51, 52].
References
1. Haar, A. (2010). Zur Theorie Der Orthogonal Funktionen Systeme. Mathematische Annalen,
69, 331–371.
2. Walsh, J. L. (1923). A closed set of normal orthogonal functions. American Journal of
Mathematics, 45, 5–24.
3. Paley, R. E. A. C. (1932). A remarkable system of orthogonal functions. Proceedings of the
London Mathematical Society, 34, 241–279.
4. Fine, N. J. (1949). On the Walsh functions. Transactions of the American Mathematical
Society, 65(3), 372–414.
5. Schipp, F., Wade, W. R., & Simon, P. (1990). Walsh series. Adam Hilger.
6. Beauchamp, K. G. (1975). Walsh functions and their applications. New York: Academic
Press.
7. Siddiqi, A. H. (1978). Walsh function. Aligarh: AMU.
8. Maqusi, M. (1981). Applied Walsh analysis. London, Philadelphia, Rheine: Heyden.
9. Golubov, B. I., Efimov, A. V., & Skvortsov, V. A. (2008). Walsh series and transforms.
Moscow: Urss. English translation of 1st ed., 1991, Dordrecht: Kluwer.
Introduction xxi
10. Coifman, R. R., Meyer, Y., Quake, S., & Wickerhauser, M. V. (1993). Signal processing and
compression with wavelet packets progress. In Wavelets analysis and applications (pp. 77–93).
Gif-Sur-Yvette: Frontiers.
11. Daubechies, I. (1988). Orthonormal bases of compactly supported wavelets. Communications
on Pure and Applied Mathematics, 41, 909–996.
12. Daubechies, I. (1992). Ten lectures on wavelets. Philadelphia: SIAM.
13. Mallat, S. (1989). Multiresolution approximations and wavelets. Transactions of the
American Mathematical Society, 315, 69–88.
14. Mallat, S. (1999). A wavelet tour of signal processing. New York, London: Academic Press.
15. Meyer, Y. (1992). Wavelets and operators. Cambridge: Cambridge University Press.
16. Walnut, D. F. (2002). Wavelet analysis. Basel: Birkhauser.
17. Christensen, O. (2003). An introduction to frames and Riesz bases. Basel: Birkhaeuser.
18. Siddiqi, A. H. (2018). Functional analysis with applications. Berlin: Springer Nature.
19. Krivoshein, A. V., Protasov, V. Yu., & Skopina, M. A. (2016). Multivariate wavelet frames.
Berlin: Springer.
20. Vilenkin, N. Ya. (1947). On a class of complete orthonormal systems. Izvestiya Rossiiskoi
Akademii Nauk. Seriya Matematicheskaya, 11, 363–400. English translation, 1963, American
Mathematical Society Translations, 28 (Ser. 2), 1–35.
21. Benedetto, J. J., & Benedetto, R. L. (2004). A wavelet theory for local fields and related
groups. The Journal of Geometric Analysis, 14, 423–456.
22. Benedetto, J. J., & Benedetto R. L. (2009). The construction of wavelet sets, wavelets and
multiscale analysis. In J. Cohen et al. (Ed.), Theory and Applications. Selected Papers Based
on the Presentations at the International Conference on Wavelets: Twenty years of Wavelets,
De Paul University, Chicago, IL, USA, May 15–17 (pp. 17–56). New York: Springer (Alied
and Numerical Harmonic Analysis).
23. Novikov, I. Ya., Protasov, V. Yu., & Skopina, M. A. (2011). Wavelet theory (Moscow, 2006).
Providence: AMS.
24. Kozyrev, S. V. (2002). Wavelet analysis as a p-adic spectral analysis. Izvestiya: Mathematics,
66, 367–376.
25. Farkov, Yu. A. (2007). Orthogonal wavelets on direct products of cyclic groups.
Matematicheskie Zametki, 82(6), 934–952. English translation, Mathematical Notes, 82(6),
843–859.
26. Farkov, Yu. A. (2008). Multiresolution analysis and wavelets on Vilenkin groups. Facta
27. Lang, W. C. (1996). Orthogonal wavelets on the cantor dyadic group. SIAM Journal on
Mathematical Analysis, 27(1), 305–312.
28. Sendov, B. I. (1997). Multiresolution analysis of functions defined on the dyadic topological
group. East Journal on Approximations, 3(2), 225–239.
29. Novikov, I. Ya., & Skopina, M. A. (2012). Why are haar bases in various structures the same.
Mathematical Notes, 91(6), 895–898.
30. Evdokimov, S., & Skopina, M. (2015). On orthogonal p-adic wavelet bases. Journal of
Mathematical Analysis and Applications, 424(2), 952–965.
31. Agaev, G. H., Vilemkin, N. Ya., Dzhafarli, G. M., & Rubinstein, A. I. (1981). Multiplicative
systems of functions and analysis on 0 dimensional groups. ELM, Baku (in Russian).
32. Edwards, R. E. (1982). Fourier series: A modern introduction (2). Berlin: Springer.
33. Farkov, Yu. A. (2008). Walsh function and the continuous wavelet transform. In Proceedings
on V International Symposium on Fourier Series and Their Applications, Rostov na, Donu,
Tsvvr (pp. 27–32).
34. Lukashenko, T. P. (1994). Wavelets on topological groups. Izv. Ross. Akad. Nauk Ser. Mat.,
58(3), 88–102. English translation, 1995, Izvestiya: Mathematics, 44(3), 515–529.
35. Lang, W. C. (1998a). Fractal Multiwavelets related to the Cantor dyadic group. International
Journal of Mathematics and Mathematical Sciences, 21, 307–317.
xxii Introduction
36. Lang, W. C. (1998b). Wavelet analysis on the Cantor dyadic group. Houston Journal of
Mathematics, 24, 533–544.
37. Cohen, A. (1990). Ondelettes analysis multiresolutions et filtres miroir en quadrature. Annales
de l'Institut Henri Poincaré C, Analyse non linéaire, 7, 439–459.
38. Farkov, Yu. A., & Protasov V. Yu. (2006). Dyadic wavelets and refinable functions on a half
line. Matematicheskii Sbornik, 197(10), 129–160. English translation, Sbornik: Mathematics,
197, 1529–1558.
39. Farkov, Yu. A. (2005). Orthogonal wavelets with compact support on locally compact abelian
groups. Izv. Ross. Akad. Nauk Ser. Mat. 69(3), 193–220. English translation, Izvestia:
Mathematics, 69(3), 623–650.
40. Farkov, Yu. A. (2007). Biorthogonal dyadic wavelets on ℝ+. Russian Mathematical Surveys,
62, 1198. Translation from Usp. Mat. Nauk, 62(6), 189–190.
41. Farkov, Yu. A. (2009). Biorthogonal wavelets on Vilenkin groups. Tr. Mat. Inst. Steklova,
265(1), 110–124. English translation, Proceedings of the Steklov Institute of Mathematics,
265(1), 101–114.
42. Farkov, Yu. A., Lebedeva, E. A., & Skopina, M. A. (2015). Wavelet frames on Vilenkin
groups and their approximation properties. International Journal of Wavelets, Multiresolution
and Information Processing, 13(5). https://doi.org/10.1142/5021.
43. Berdnikov, G. S., & Lukomskii, S. F. (2014). N-valid trees in wavelet theory on Vilenkin
groups. http://arxiv.Org/abs/1412.309v1.
44. Lukomskii, S., & Vodolazov, A. (2014). Non-Haar MRA on local fields of positive charac-
teristics. http://arxiv.org/abs/1303.5635v/.
45. Lukomskii, S. F. (2014). Step refinable functions and orthogonal MRA on P-adic Vilenkin
groups. Journal of Fourier Analysis and Applications, 20(1), 42–65.
46. Lukomskii, S. F. (2015). Riesz multiresolution analysis on zero-dimensional groups.
Izvestiya: Mathematics, 79(1), 145–176.
47. Farkov, Yu. A. (2005). Orthogonal p-wavelets on R+. In Proceedings of International
Conference on Wavelets and Spilnes, St. Petersburg, Russia, July 3–8 (p. 426). St. Petersberg:
St. Petersberg University Press.
48. Farkov, Yu. A. (2009). On wavelets related to Walsh series. Journal of Approximation
Theory, 161, 259–279.
49. Debnath, L., & Shah, F. A. (2015). Wavelet transforms and their applications (2nd ed.). New York:
Birkhauser/Springer.
50. Behera, B., & Jahan, Q. (2012). Wavelet packets and wavelet frame packets on local fields of
positive characteristic. Journal of Mathematical Analysis and Applications, 395, 1–14.
51. Behera, B., & Jahan, Q. (2015). Characterization of wavelets and MRA wavelets on local
fields of positive characteristic. Collectanea Mathematica, 66(1), 33–53.
52. Berdnikov G., Kruss, Iu., & Lukomskii, S. (2017). On orthogonal systems of shifts of scaling
function on local fields of positive characteristic. http://arxiv.org/1503.08600.
Chapter 1
Introduction to Walsh Analysis
and Wavelets
1.1 Walsh Functions
The trigonometric Fourier series has played a very significant role in solving problems
of science and technology. The concept of non-trigonometric Fourier series such as
Haar–Fourier series and Walsh–Fourier series were introduced by Haar [1] and Walsh
[2], respectively; Kaczmarz, Steinhaus, and Paley studied some aspects of Walsh
system between 1929 and 1931. Nowadays, Paley’s modification, which is defined
as the product of Rademacher functions is known as the Walsh function [3]. A major
breakthrough came in N. J. Fine’s dissertation of Ph.D. submitted to Pennsylvania
University in 1946, which was published subsequently in Transactions of American
Mathematical Society in 1949. Fine [4] introduced the concept of dyadic group G
and proved that Walsh functions are its characters. Haar introduced in 1909 a class of
functions closely related to Walsh functions in response to a question posed by David
Hilbert, namely “Does there exist an orthonormal system such that the Fourier series
of every continuous function converges uniformly?” He showed that the Fourier
series with respect to the orthonormal system, introduced by him, now known as
the Haar system, of continuous functions converge uniformly on [0, 1]. Walsh [2]
observed that Haar and Walsh systems are Hadamard transforms of each other. The
Walsh and Haar systems perform all usual applications of orthonormal systems, for
example, data transmission, multiplexing, filtering, image enhancement, and pattern
recognition. An elegant presentation of these developments are given in Schipp,
Wade, Simon [5]; also see Beauchamp [6], Siddiqi [7], Maqusi [8], and Golubov,
Efimov, and Skvortsov [9]. In the early 80s, wavelet analysis was introduced and
developed by Morlet et al. [10–17]. Also see Walnut [14], Christensen [16], Siddiqi
[7, 18].
In this chapter, we briefly discuss Walsh function, its integration besides dyadic
group, properties of Walsh–Fourier transform, Haar function and its relationship with
Walsh function, Walsh-type wavelet packet, and wavelet analysis.
As usual, Z, Z+ , N, and C denote, respectively, the sets of integers, positive inte-
gers, natural numbers, and complex numbers.
© Springer Nature Singapore Pte Ltd. 2019 1
Yu. A. Farkov et al., Construction of Wavelets Through Walsh
Functions, Industrial and Applied Mathematics,
https://doi.org/10.1007/978-981-13-6370-2_1
2 1 Introduction to Walsh Analysis and Wavelets
The Walsh system {wn : n ∈ Z+ } on R+ is defined as follows:

k
w0 (x) = 1, wn (x) = (w1 (2 j x))ν j , n ∈ N,
j=0
x ∈ R+ , where R+ = [0, ∞),

where the ν j are the coefficients of the decomposition

k
n= νj2j, ν j ∈ {0, 1}, νk = 1, k = k(n),
j=0
and the function w1 (x) is defined on [0, 1] by the formula

1, x ∈ [0, 1/2)
w1 (x) =
−1, x ∈ [1/2, 1),
and is extended to R+ by periodicity: w1 (x + 1) = w1 (x) for all x ∈ R+ .

Walsh polynomials are finite linear combinations of the Walsh functions. An
arbitrary Walsh polynomial of order n can be written in the following form:

n
w(x) = c j w j (x)
j=0
where c j are complex coefficients. For more information about the properties of
Walsh polynomials and their role in the dyadic harmonic analysis see for example,
[5, 7].
We shall denote the integer and the fractional parts of a number x ∈ R+ , by [x]
and {x}, respectively.
For x ∈ R+ and j ∈ N, we define the number x j , x− j ∈ {0, 1} as follows:
x j = [2 j x](mod 2), x− j = [21− j x](mod 2). (1.1)
These are the digits of the binary expansion

x= x j 2− j−1 + x j 2− j .
j<0 j=0
(For dyadic x, we obtain an expansion with finitely many nonzero terms.)

For fixed x, y ∈ R+ , we set

x⊕y= |x j − y j |2− j−1 + |x j − y j |2− j ,
j<0 j>0
1.1 Walsh Functions 3
where x j , y j are defined in (1.1). By definition x y = x ⊕ y (because x ⊕ x = 0).

The binary operation ⊕ identifies R+ with the group G 2 and is useful in the study of
dyadic Hardy classes and for the construction of algorithms in signal processing (see
[5, 19]). We point out that this identification associates Haar measure in G 2 with the
Lebesgue measure on R+ and the characters of the group G 2 with generalized Walsh
functions.
A function f : R+ → C is said to be W -continuous (or continuous (W )) at a
point x ∈ R+ if
sup | f (x ⊕ h) − f (x)| → 0 as n → ∞.
0≤h<1/2n
A function f is W -continuous if it is W -continuous at each point of R+ . A continuous

(in the usual sense) function on R+ is also W -continuous. Walsh polynomials are
W -continuous. C W denotes set of all 1-continuous functions. (see for example, [5],
Sect. 1.3).
Theorem 1.1 Every continuous function is W -continuous but the converse is not
true.
Proof Suppose f (x) is a continuous function and let
w(δ, f ) = sup | f (x1 ) − f (x2 )|

|x1 −x2 |≤δ
If 0 ≤ h = |x1 − x2 | ≤ δ < 1, then by choosing x1 = x ⊕ h and x2 = x and keeping

in mind |(x ⊕ h) − x| ≤ h, if 0 ≤ x < 1 and 0 ≤ h < 1, see also [19]. From here,
it follows sup | f (x ⊕ h) − f (x)| → 0 as n → ∞. Hence f (x) is W -continuous.
It may be observed that the Walsh function is not continuous but it is
W -continuous:
sup | f (x ⊕ h) − f (x)| → 0 as n→∞

0≤h<1/2n
We give below integral of Walsh function. For derivative of Walsh function, see [5,
pp. 38–48].
Integration of Walsh Functions:
In this section, we shall discuss the properties of the integrals
x
Jk (x) = ψk (t)dt, k = 0, 1, 2, 3, . . .
0
Since ψ0 (t) = 1, it follows that

x
J0 (x) = ψ0 (t)dt
0
= x.
Theorem 1.2
∞

−(n+2) −r
Jk (x) = 2 ψk (x) − 2 ψ2+k
n+r
(x) ,
r =1
where k ≥ 1, k = 2n + k , 0 ≤ k < 2n , n = 0, 1, 2, . . .
∞

Proof Let x − [x] = 2−n xn , then
n=1
∞
1 1 −n
x − [x] = − 2 φn (x) (1.2)
2 4 n=0
because
1 − φn (x)
xn+1 = , n = 0, 1, 2, . . .
2
and by the fact that φn (x) takes the value 1 and −1 and xn+1 is 1 or 0, and that
∞
∞
−n 1 −n
2 xn = 2 xn+1
n=1
2 n=0
∞
∞
1 1 1 −n
= 1+ − 2 φn (x)
4 n=1
2n 4 n=0
∞
1 1 −n
= − 2 φn (x)
2 4 n=0
Let I ( p, n) denote the interval p. 2−n ≤ x < ( p + 1)2−n . Then ψk (x) is constant on
each interval I ( p, n + 1) and ψk (x) is constant on each interval I ( p, n). It follows
from the definition of Walsh functions that
ψk (x) = ψ2n (x)ψk (x)
and
Jk (x) = 2−n ψk (x)J1 (2n x) = ψk (x)J2n (x) (1.3)
Now

1 1
J1 (x) = + ψ1 (x) x − [x] −
2 2
From (1.3) with k = 2n , k = 0, we have

1.1 Walsh Functions 5
∞

−(n+2) −r
J2n (x) = 2 1− 2 ψ2r +1 (2 x)n
r =1
∞

= 2−(n+2) 1 − 2−r ψ2n+r +2n (x)
r =1
In general, therefore,
∞

−(n+2) −r
Jk (x) = 2 ψk (x) − 2 ψ2n+r +2n +k (x)
r =1
∞

−(n+2) −r
=2 ψk (x) − 2 ψ2n+r +k (x) .
r =1
This completes the proof.

It is also clear that Walsh system is closed under the finite products, every Walsh
function is piecewise constant with finitely many jump discontinuities on [0, 1),
and takes values +1 and −1. It may be noted that the Walsh system is complete
and orthonormal system. Every function f (x) which is periodic with period 1 and
Lebesgue integrable f (x) on [0, 1) can be expanded in a Walsh–Fourier series.
∞

f (x) ∼ cn wn (x),
n=0
1
where cn = 0 f (x)wn (x)d x, n = 0, 1, 2.....
Many properties of Walsh–Fourier series are analogous to trigonometric Fourier
series, for example, the Dirichlet kernels for the Walsh system, and trigonometric
system have the same order of growth. The theory of Walsh–Fourier series is quite
rich, see for example, [2, 5, 7, 9]. It has been shown in [5] that Walsh analysis is a
special case of the harmonic analysis on compact abelian groups. Some basic results
and applications of Walsh systems are given in Chap. 2.
1.2 Walsh–Fourier Transform
For x, ω ∈ R+ , we set
∞

χ (x, ω) = (−1)σ (x,ω) , where σ (x, ω) = (x j ω− j + x− j ω j )
j=1
and x j and ω j are defined by (1.1). Observe that for each positive integer n, we have
χ (x, 2−n l) = χ (2−n , l) = wl (2−n x) for all x ∈ [0, p n−1 ), l ∈ Z.
The Walsh–Fourier transform of a function f ∈ L 1 (R+ ) is defined as follows:

f (ω) = f (x)χ (x, ω) d x.
R+
If f ∈ L 2 (R+ ) and
a
Ja f (ω) = f (x)χ (x, ω) d x (a > 0) (1.4)
0
then f˜ is defined as limit of Ja f in L 2 (R+ ) as a → ∞.

The properties of the Walsh–Fourier transformation are quite similar to the prop-
erties of the classical Fourier transformations (see for example, [9, Chap. 6], [5,
Chap. 9]). We list some of them below:
Theorem 1.3 The following properties hold:

f is a W -continuous function and fˆ(ω) → 0 as ω → ∞;
(a) if f ∈ L 2 (R+ ), then
(b) the inversion formula

f (x) =
f (ω)χ (x, ω)dω; (1.5)
R+
holds for each x ∈ R+ , provided that both f and

f belongs to L 1 (R+ ) and f
is W -continuous.
(c) if f ∈ L 2 (R+ ), then
f ∈ L 2 (R+ ) and

f L 2 (R+ ) = f L 2 (R+ )
In what follows En is the space of dyadic entire functions of order n, that is, of
functions defined in R+ and constant on all intervals of range n. For f ∈ En , one has
∞

f (x) = f (2−n k)χl (n) (x), x ∈ R+ .
k
k=0
Clearly, each Walsh polynomial of order 2n − 1 belongs to En . The set E of dyadic

entire function on R+ is the union of all the spaces En . The refinable functions in
Examples 1 and 3, of [20] belong to E .
Theorem 1.4 The following properties hold: (a) if supp f ⊂ [0, 2n ], then f ∈ (E),
provided that bothf and fˆ belong to L 1 (R+ ) and f is W -continuous;
(b) if f ∈ L 1 (R) (E)n , then supp f ⊂ [0, 2n ];
(c) if f ∈ L 1 (R) and supp f ⊂ [0, 2n ], then
f ∈ En ; in a similar way, if g ∈ L 1 (R)
and supp g ⊂ [0, 2n ], then the inverse Walsh–Fourier transform of the function g
belongs to E ;
1.2 Walsh–Fourier Transform 7
(d) if fˆ ∈ En , then supp f ⊂ [0, 2n ], provided that f ∈ L 1 (R) and f is

W -continuous.
Proofs of all these properties can be found in [9, Sect. 6.2].
Proof of Theorem 1.2.1 (b) Let ε > 0, fix t ∈ R + , and choose by hypothesis an
integer n > 0 such that
| f (t ⊕ u) − f (t)| ≤ ε
for all u ∈ R + which satisfy |u| < 2−n . Clearly,

2n
fˆ(y)ψt (y)dy − f (t) = |(S2n f )(t) − f (t)|

0
2−n
≤ 2n | f (t + u) − f (t)|du
0
=ε
Therefore, 2n
lim fˆ(y)ψt (y)dy = f (t).
n→∞ 0
∞ ∞

But fˆ(y)ψt (y)dy ≤ | fˆ(y)|dy → 0 as n → ∞, since f is integrable.
2n 2n
This proves (b) part of Theorem 1.3.
Relationship between Fourier and Walsh–Fourier Transform: fˆ = Walsh–
Fourier transform of f is defined above and Fourier transform of f is
∞
F(ω) = f (x)e−iωx d x (1.6)
−∞
By the inverse of Walsh–Fourier transform

∞
f (x) = fˆχ (x, ω)dω (1.7)
0
and by the inverse of Fourier transform

∞
f (x) = F(ω)eiωx dω (1.8)
0
Substituting f (x) from (1.7) in (1.6), we get

∞ ∞
F(ω) = ˆ
f (τ )χ (x, τ )dτ e−iωx d x (1.9)
0 0
Interchanging order of integration of (1.9) yields

∞
F(ω) = fˆ(τ )E(τ, −ω)dω,
0
where E(τ, −ω) is defined as either the Walsh–Fourier transform of e−iωx or equiv-
alently as the Fourier transform of Walsh function χ (x, ω) (see Sect. 1.2). On the
other hand, Walsh–Fourier transform of f can be expressed in terms of its Fourier
transform.
1.3 Haar Functions and Its Relationship with Walsh

Functions
Let In (x) = I ( p, n), where
I ( p, n) = [ p2−n , ( p + 1)2−n ], 0 ≤ p < 2n , n, p ∈ N (1.10)
I ( p, n) is called a dyadic interval. For each x ∈ [0, 1), In (x) is called the dyadic
interval of length 2−n . Let l0 denote the collection of sequences a = {an }n∈N such
that an ∈ R. For each 0 < p < ∞, let l p denote the set of all sequences a ∈ l0 (space
of real sequences) such that
∞ 1/ p

a p = |an | p is finite. (1.11)
n=0
Let l∞ denote the collection of a ∈ l0 such that
a∞ = supn∈N |an | is finite. (1.12)
It may be recalled that l p , 1 ≤ p ≤ ∞ is a Banach space and for {ak }, {bk } ∈ l0

(Abel’s transformation)
⎛ ⎞
∞

n−1 k
n
ak bk = ⎝ ⎠
a j (bk − bk+1 ) + bn aj. (1.13)
k=0 k=0 j=0 j=0
For each n ∈ N set

j
ak(n) = 2−n/2 wk ,
j
2n
for 0 ≤ j, k < 2n , j, k ∈ N .
A matrix of form
A(n) = (ak(n)
j
)2n−1
j,k=0 , n ∈ N (1.14)
1.3 Haar Functions and Its Relationship with Walsh Functions 9
is called a Hadamard–Paley matrix.

Since w j (k/2n ) = wk ( j/2n ), by (properties of Walsh function)), each Hadamard–
Paley matrix A(x) is real and symmetric 2n × 2n matrix and furthermore, it is orthog-
onal matrix:
2
n
−1 2
n
−1

1 i
ak(n) a (n)
ji = 2−n wk wj
i=0
i
i=0
2n 2n
2
n
−1 1
= wk w j = wk w j .
i=0 I (i,n) 0
The Hadamard transform H is defined on l0 as follows:

Given b = {bk }k∈N ∈ l0 , define H b = bk−1 H bk , for all k ∈ N , H b ∈ l0 , H b0 =
b0 ; For 0 ≤ k < 2n , n ∈ N set
2
n
−1
H b2n +k = ak(n)
j
b2n + j (1.15)
j=0
Since the Hadamard–Paley matrices are symmetric and orthogonal, it is clear that
H takes l0 onto l0 and satisfies H H (b) = b. that is, if a is a Hadamard transform of
b then b is Hadamard transform of a. The concept of Hadamard transform can be
extended to function sequences like
⎛ ⎞
2
n
−1 2
n
−1 2
n
−1
b2n +k H ( f 2n +k ) = b2n +k ⎝ ak(n)
j
. f 2n + j ⎠
k=0 k=0 j=0
⎛ ⎞
2
n
−1 2
n
−1
= ⎝ ak(n) b2n +k ⎠ f 2n + j
j
j=0 j=0
2
n
−1
= H (b2n + j ) f 2n + j
j=0
for any function sequence { f k } and {bk } ∈ l0 . In particular,
2
n
−1 2
n
−1
bk H ( f k ) = H (bk ) f k (1.16)
k=2n−1 k=2n−1
Property (1.16) is called Hadamard transform self adjoint. We define below the
Haar system and indicate that the Walsh and Haar systems are Hadamard transforms
of each other, see [5, pp. 22–23].
The Haar system {h n } is defined as follows:

Set h 0 = 1. For n, k ∈ N with 0 ≤ k < 2n define h n on [0, 1] by
⎧ n/2
⎨ 2 , if x ∈ I (2k, n + 1)
h 2n +k (x) = −2n/2 , if x ∈ I (2k + 1, n + 1)
⎩
0, other wise
h 2n +k (x + 1) = h 2n +k (x).
Each Haar function is continuous from the right and the Haar system {h n } is orthonor-
mal on [0, 1). The following relation expresses Haar functions in terms of Walsh
functions.

n−1
2
1
h 2n +k = h kn (x) = √ [wm (k/2n ) = wm+1 (k/2n )]wk (x).
2 2n−1 k=0
1.4 Walsh-Type Wavelet Packets
A wavelet packet is a generalization of wavelets. Walsh function is an example

of wavelet packet. Walsh-type wavelet packet is a generalization of Walsh func-
tion and is a special case of wavelet packet. We mention here some beautiful
results of the Ph.D. thesis of Morten and Nielsen submitted to Washington Uni-
versity, Saint Louis, Missouri, May 1999 and approved by the committee chaired by
Prof. M. Victor Wickerhauser.
Definition 1.1 (Nonstationary Wavelet Packets) Let (ϕ, ψ) be the scaling function
( p) ( p)
and wavelet associated with a multiresolution analysis, and let (F0 , F1 ), p ∈ N
be a family of bounded operators on l2 (N ) of the form

(Fε( p) a)k = an−2k h ( p) , ε = 0, 1
n∈N

(Fε( p) a)k = an h ε( p) (n − 2k), ε = 0, 1
n∈Z
( p) ( p) ( p)
with h 1 = (−1)n h 0 (1 − n) a real value sequence in l1 (N ) such that each (F0 ,
( p)
F1 ) is a pair of conjugate quadrature filters (CQFs), see below. Let {h n } ∈ l1 (N ))
be a real valued sequence, and let gk = (−1)k h 1−k for k ∈ Z . Define the operators
H, G : l2 (N ) → l2 (N ) by
1.4 Walsh-Type Wavelet Packets 11

(H a)k = an h n−2k
n∈N

(Ga)k = an gn−2k
n∈N
The filters H and G are called a pair of CQF if
2H H ∗ = 2GG ∗ = 1,
H I = 1, where I = (1, 1, 1, ...1)
H ∗G + G∗ H = 1
H G∗ = G H ∗ = 0
of the form
( p) p ( p)
with h 1 (n) = (−1)n h 0 (1 − n) a real-valued sequence in l 1 (N ) such that (F0 ,
( p)
F1 ) is a pair of CQFs. H ∗ and G ∗ are defined as

(Ha∗ )k = a j h k−2 j
j∈Z

(G a∗ )k = a j gk−2 j
j∈Z
We define the family of nonstationary wavelet packets {wn }∞

n=0 recursively by
letting w0 = ϕ, w1 = ψ, and then for n ∈ N
( p)
w2n (x) = 2 h 0 (q)wn (2x − q) (1.17)
q∈Z
( p)
w2n+1 (x) = 2 h 1 (q)wn (2x − q), wher e 2 p ≤ n < 2 p+1 (1.18)
q∈Z
Definition 1.2 (Basic Stationary wavelet packets). Let (ϕ, ψ) be the scaling function
and wavelet associated with a multiresolution analysis, with associated CQFs {h n }
p
and {gn }. The functions {wn }n generated by Definition (1.1) by letting {h 0 } = {h n }
( p)
and {h 1 } = {gn } for all p ∈ N are called basic stationary wavelet packets.
Definition 1.3 The CQFs given by h 0 = h 1 = 21 , h k = 0 otherwise, and gk =

(−1)k h 1−k are called the Haar filters.
Definition 1.4 The Walsh system {wn }∞

n=0 is defined recursively on [0, 1) by w0 (x) =
χ[0,1] (x) and
w2n+ε (x) = wn (2x) + (−1)ε wn (2x − 1), ε = 0, 1, n = 0, 1.......

Definition 1.5 (Walsh type wavelets packets). Let {wn }n≥0,k∈Z be a family of non-
( p)
stationary wavelet packets constructed by using a family {h n }∞ p=1 of finite filters
( p)
in Definition 1.1. If there exists a constant J ∈ N such that h n is the Haar filter
for every p ≥ J and w1 has compact support then {wn }n≥0 is called a family of
Walsh-type wavelet packets.
It may be observed that Walsh-type wavelet packets resemble Walsh functions.
Interesting properties of Walsh-type wavelet packet expansions are proved by
M. Nielson [21].
Theorem 1.5 The Walsh-type wavelet packet expansion of any f ∈ L p (R), 1 <
p < ∞, converges a.e..
1.5 Wavelet Analysis
Wavelets meaning small waves were introduced in the early 80s. Wavelet theory deal-
ing with various aspects of wavelets is the outcome of a multidisciplinary endeavor
that brought together mathematicians, physicists, and engineers. Wavelet theory is a
refinement of Fourier analysis which enables to simplify the description of a cum-
bersome function in terms of a small number of coefficients. For a detailed account
of wavelet analysis and applications, we refer to Daubechies [10], Christensen [16],
Frazier [15], Mallat [22] and Siddiqi [18]. We present here a few basic results.
1.5.1 Continuous Wavelet Transform
Definition 1.6 A function ψ ∈ L 2 (−∞, ∞) is called a wavelet if it has zero aver-

age, that is, ∞
ψ(t)dt = 0 (1.19)
−∞
and ∞
|ψ̂(ω)|2
Cψ = 2π dω < ∞ (1.20)
−∞ |ω|
where ψ̂(ω) is the Fourier transform of ψ(t).

Remark 1.1 Very often, ψ is called wavelet without (1.20). Condition (1.20) is essen-
tial for proving inversion result. Many researchers take the following condition in
place of (1.20):
ψ(x) is continuous and has exponential decay

[ψ(x) ≤ Me−c|x| for some constants c and M.] (1.21)
1.5 Wavelet Analysis 13
Remark 1.2 If ψ ∈ L 2 (−∞, ∞) ∩ L 1 (−∞, ∞) then (1.20) implies (1.19). Equa-

tion (1.20) is called the wavelet admissibility condition.
Lemma 1.1 Let φ be a nonzero n-times, n ≥ 1, differentiable function such that
φ (n) ∈ L 2 (−∞, ∞). Then ψ(x) = φ n (x) is a wavelet.
Proof From the property of the Fourier transform (see for example, [18])
|ψ̃(ω)| = |ω|k |φ̃(ω)|.
Then
∞
|ψ̂(ω)|2
Cψ = dω
−∞ |ω|
∞
|ω|2k |φ̂(ω)|2
= dω
−∞ |ω|
−1
|ω|2k |φ̂(ω)|2
= |ω|2k−1 |φ̂(ω)|2 dω + dω
1 |ω|>1 |ω|
≤ 2π(φ2L 2 + φ (k) 2L 2 ) <∞
It is also clear that ∞ ∞

ψ(x)d x = φ (n) (x)d x = 0.
−∞ −∞
This proves that ψ is a wavelet.

Example 1.1 (Haar Wavelet) Let
⎧
⎨ 1, 0 ≤ x < 1/2
ψ(x) = −1, 1/2 ≤ x < 1
⎩
0, otherwise
It is clear that
∞ 0 1/2 1 ∞
ψ(x)d x = ψ(x)d x + ψ(x)d x + ψ(x)d x + ψ(x)d x
−∞ −∞ 0 1/2 1
1/2 1
= 1d x − 1d x
0 1/2
=0
and ψ(x) has compact support [0, 1].

We can verify that
1 (sin( ω4 ))2 −i(ω−π)/2

ψ̂(ω) = √ e
(2π ) ω/4
and
∞ ∞
|ψ̂(ω)|2 8 | sin ω/4|4
dω = dω < ∞
−∞ |ω π −∞ |ω|3
Function ψ(x) was introduced by Hungarian mathematician Alfred Haar in 1910

which is the simplest example of wavelet.
Example 1.2 (Mexican hat Wavelet) The function defined by
ψ(x) = (1 − x 2 )e−x /2
2
is a wavelet by Lemma 1.1.
d 2 −x 2 /2
= (1 − x 2 )e−x /2 .
2
ψ(x) = e
dx2
It is called Mexican hat wavelet.
It may be observed that the set of nonzero wavelets with compact support is a
dense subset of L 2 (−∞, ∞) (see Theorem 12.2 of [18]).
Definition 1.7 (Continuous Wavelet Transform) The continuous wavelet transform
Tψ of a function f ∈ L 2 (−∞, ∞) with respect to this wavelet ψ is defined as
∞

−1/2 t −b
Tψ f (a, b) = |a| f (t)ψ̄ dt (1.22)
−∞ a
where a ∈ (−∞, ∞) 0, b ∈ (−∞, ∞) and ψ̄ denotes complex conjugate of ψ(x).

If ψ is a real function then ψ̄ = ψ. If we consider ψa,b (t) a family of real functions
defined by

−1/2 t −b
ψa,b (t) = |a| ψ , a > 0, b ∈ R (1.23)
a
where ψ is a fixed function, often called mother wavelet, then
Tψ f (a, b) =< f, ψa,b >= inner product of f and ψa,b .
Remark 1.3 The following properties can be verified keeping in mind the properties
of inner product.
(a) Tψ (α f + βg)(a, b) = αTψ f (a, b) + βTψ g(a, b)
for any α, β ∈ R.
(b) Tψ (Sc f )(a, b) = Tψ f (a, b − c),
operator defined by Sc f (t) = S(t − c).
where Sc is a translation
(c) Tψ (Dc f )(a, b) = √1c Tψ f (a/c, b/c), where c is a positive number and Dc is
the dilation operator defined by Dc f (t) = (1/c) f (t/c).
(d) Tψ φ(a, b) = Tφ ψ(1/a, −b/a), a = 0, where ψ and φ are two wavelets.
(e) T Aψ A f (a, b) = Tψ f (a − b), where A is defined by
Aψ(t) = ψ(−t)
(f) TSc ψ( f (a, b)) = Tψ f (a, b + ca)

(g) (TDc ψ f )(a, b) = √1(c) (Tψ f )(ac, b), c > 0
Remark 1.4 Tψ f (a, b) is a function with scale (frequency) a and the spatial (time) b.
The plane defined by the variables (a, b) is called the scale-space or time–frequency
plane. It measures the variation of f in a neighborhood of b . For a compactly
supported wavelet, the value of Tψ f (a, b) depends upon the value of f in a neigh-
borhood of b of size proportional to the scale a. At small scales, Tψ f (a, b) provides
localized information such as localized regularity of f (x).
Few important results like Parseval’s formula, isometry, and inverse formula for
wavelet transformation are mentioned below whose proofs can be found in [10, 18].
1.5.2 Discrete Wavelet System
We discretize continuous wavelet transforms by choosing a = 2− j , b = k2− j , j, k ∈

N.
Definition 1.8 A function ψ ∈ L 2 (−∞, ∞) is a wavelet if the family of functions

ψ j,k (t) defined by
j
ψ j,k (t) = 2 2 ψ(2 j t − k) (1.24)
is an orthonormal basis in L 2 (R), that is,

∞

ψ j,k (t)ψ j ,k (t)dt = 1 i f j = j , k = k
−∞
= 0 other wise
be defined by the condition We say that the functions ψ j,k j, k ∈ N , form the
wavelet system associated to the function ψ. It may be observed that the wavelet
admissibility condition (1.5.2) is a necessary condition for ψ j,k (t) to be a wavelet
system.
Definition 1.9 Wavelet coefficients of a function f ∈ L 2 (−∞, ∞), denoted by d j,k ,

are defined as the inner product of f with ψ j,k (t), that is
∞
d j,k =< f, ψ j,k (t) >= f (t)ψ j,k (t)dt (1.25)
−∞
The series
< f, ψ j,k (t) > ψ j,k (t) (1.26)
j∈N k∈N
is called wavelet series of f .

The expression
< f, ψ j,k (t) > ψ j,k (t)
j∈N k∈N
is called the wavelet representation of f . A characterization of Lipschitz α class,

0 < α < 1 in terms of the wavelet coefficients is given below:
Theorem 1.6 f ∈ Li p α, that is, there is a constant K > 0 such that | f (x) −
f (y)| ≤ k|x − y|α , 0 < α < 1, if and if only
| d j,k |≤ K 2−( 2 +α)

1
(1.27)
For proof we refer [12, 14, 18].
1.5.3 Multiresolution Analysis
Multiresolution Analysis introduced by Mallat and Meyer in 1989 provides a general

method to construct wavelet orthonormal basis. It is considered as the most important
ingredient of wavelet theory.
Definition 1.10 (Multiresolution Analysis-[13]) A Multiresolution Analysis (MRA)

is a sequence {V j } of closed subspaces of L 2 (−∞, ∞) such that
(i) ..... ⊂ V−1 ⊂ V0 ⊂ V1 ......

(ii) V = L 2 (−∞, ∞)
j j∈Z
(iii) V j j∈Z = 0
(iv) f (x) ∈ V j if and only if f (2− j x) ∈ V0 ,
(v) f (x) ∈ V0 if and only if f (x − m) ∈ V0 for all m ∈ N and
• there exists a function φ ∈ V0 called scaling function, such that the system
{ψ(t − m)m∈N } is an orthonormal basis in V0 .
Construction of a wavelet from MRA:

For a given MRA {V j } in L 2 (R) with scaling function φ, a wavelet is obtained in the
following manner:
Let the subspace W j of L 2 (R) be defined by the condition
V j ⊕ W j = V j+1 , V j ⊥W j ∀ j.
j
Since 2 2 J j , where J j is the dyadic dilation operator defined as J j ( f )(x) = f (2 j x),
we have
J j (V1 ) = V j+1 . Thus,
V j+1 = J j (V0 ⊕ W0 ) = J j (V0 ) ⊕ J j (W0 )

= V j ⊕ J j (W0 )
This gives
W j = J j (W0 ) f or all j ∈ Z .
By conditions (i) to (iii) we obtain an orthogonal decomposition

L 2 (R) = W j = W1 ⊕ W2 ⊕ W3 ⊕ ... ⊕ Wn ⊕ ...
j∈N
We can find a function ψ ∈ W0 such that {ψ(t − m)m∈Z } is an orthonormal basis in

W0 . Such ψ is a wavelet and it is called wavelet associated with the MRA {V j } then

ψ(x) = an (−1)n φ(2x + n + 1), (1.28)
n∈Z
where ∞ x
an = φ φ(x − n), (1.29)
−∞ 2
is a wavelet.
1.6 Wavelets with Compact Support
Daubechies Wavelets
Daubechies (see for details [10, 12] and Pollen in [17]) has constructed, for an
arbitrary integer N, an orthonormal basis for L 2 (R) of the form
2 j/2 ψ(2 j x − k), j, k ∈ Z
having the following properties: The support of ψ N is contained in [−N + 1, N ]. To

emphasize this point, ψ is denoted by ψ N .
∞ ∞ ∞
ψ N (x) d x = xψ N (x) d x = . . . = x N ψ N (x) d x = 0 (1.30)
−∞ −∞ −∞
ψ N (x) has γ N continuous derivatives, where the
positive constant γ is about 1/5 (1.31)
Theorem 1.7 (Daubechies) There exists a constant K such that for each N =
2, 3, . . . , there exists an MRA with the scaling function φ and an associated wavelet
ψ such that
1. φ(x) and ψ(x) belong to C N .
2. φ(x) and ψ(x) are compactly supported and both suppφ and suppψ(x) are
contained
∞ in [−K N , K N ]. ∞
∞
3. −∞ ψ N (x) d x = −∞ xψ N (x) d x = . . . = −∞ x N ψ N (x) d x = 0.
We refer to [18] for a proof of the theorem. Here, we present a construction of the
Daubechies scaling function and wavelet on [0, 3] due to Pollen(for details, see [17]).
Theorem 1.8 The function ψ defined as
ψ(x) = −bφ(2x) + (1 − α)φ(2x − 1) − (1 − b)φ(2x − 2) + aφ(2x − 3)

(1.32)
satisfies the following conditions:
supp ψ(x) ⊂ [0, 3]

∞
1k=0
ψ(x)ψ(x − k) d x = (1.33)
−∞ 0 k = 0.
∞
ψ(x − k)φ(x) d x = 0 for all k ∈ Z . (1.34)
−∞
Thus, {2 j/2 ψ(2 j t − k)} j∈Z ,k∈Z is an orthonormal basis in L 2 (R).
Lemma 1.2 ([17], p. 90) For every x ∈ D, we have

ψ(x − k) = 1 (1.35)
k∈Z
√
3− 3
+ k ψ(x − k) = x (1.36)
k∈Z
2
Lemma 1.3 If x ∈ D and 0 ≤ x ≤ 1, then

√
1+ 3
2φ(x) + φ(x + 1) = x + (1.37)
2
1.6 Wavelets with Compact Support 19
√
3− 3
2φ(x + 2) + φ(x + 1) = x + (1.38)
2
√
−1 + 3
φ(x) − φ(x − 2) = x + (1.39)
2
Lemma 1.4 For 0 ≤ x ≤ 1 and x ∈ D, the following relations hold:

0+x
φ = aφ(x)
2

√
1+x 2+ 3
φ = bφ(x) + ax +
2 4

√
2+x 3
φ = aφ(1 + x) + bx +
2 4

3+x 1
φ = aφ(1 + x) − ax +
2 4

√
4+x 3−2 3
φ = aφ(2 + x) − bx +
2 4

5+x
φ = bφ(2 + x)
2
N
Lemma 1.5 Suppose that m(ξ ) = k=M ak e−ikξ is a trigonometric polynomial
such that
|m(ξ )|2 + |m(ξ + π )|2 = 1 for all ξ ∈ R (1.40)
m(0) = 1 (1.41)
π π
m(ξ ) = 0 for ξ ∈ − , . (1.42)
2 2
Then the infinite product
∞

θ (ξ ) = m(2− j ξ ) (1.43)
j=1
converges almost uniformly. The function θ (ξ ) is thus continuous. Moreover, it

belongs to L 2 (R). The function φ given by φ̂ = √12π θ (ξ ) has the support contained
in [M, N ] and is a scaling function of an MRA. In particular, {φ(x − k)} is an ONB
in L 2 (R). The function ψ(x) defined by

N
ψ(x) = 2 (−1)k a¯k φ(2x + k + 1) (1.44)
k=M
M−N −1 N −M−1
is a compactly supported wavelet with suppψ ⊂ 2
, 2
.
Lemma 1.6 Let D = ∪ j∈Z D j where D j = {k2− j /k ∈ Z } (D is a ring, that is, sums,
difference and product of elements of D are also in D. It is a dense subset of R). Then,
there exists a unique function φ : D → R having the following properties:
φ(x) = aφ(2x) + (1 − b)φ(2x − 1) + (1 − a)φ(2x − 2) + bφ(2x − 3) (1.45)

φ(k) = 1 (1.46)
k∈Z
φ(d) = 0 if d < 0 or d > 3 (1.47)

√ √
where a = 1+ 3
4
, b= 1− 3
4
. (It is clear that 1
2
< a < 1 and − 41 < b < 0).
Lemma 1.7 The function φ defined in Lemma 1.6 extends to a continuous function
on R which we also denote by φ: This continuous function φ has the following
properties: ∞
φ(x) d x = 1 (1.48)
−∞
and ∞
1k=0
φ(x)φ(x − k) d x = (1.49)
−∞ 0 k = 0.
In other words, φ is a scaling function.

Proof of Lemma 1.7 Let K is a nonlinear operator acting on the space of functions
on R. Let us define K ( f ) for x ∈ [0, 1] by the following set of conditions:

0+x
K( f ) = a f (x)
2

√
1+x 2+ 3
K( f ) = b f (x) + ax +
2 4

√
2+x 3
K( f ) = a f (1 + x) + bx +
2 4

3+x 1
K( f ) = b f (1 + x) − ax +
2 4

√
4+x 3−2 3
K( f ) = a f (2 + x) − bx +
2 4

5+x
K( f ) = b f (2 + x)
2
K ( f )(y) = 0 for y ∈
/ [0, 3].
This definition gives two values of K ( f ) at the points 0, 1/2, 1, 3/2, 2, 5/2, 3. us
denote by φ j , j = 0, 1, 2, . . ., the continuous, piecewise linear functions are which
on D j equals φ. The function K (φ j ) is well defined at each point and infact, K (φ j ) =
φ j+1 . Let x ∈ [0, 3] and j > 0, then we immediately get from the definition of K ( f )
φ j+1 (x) − φ j (x) = K (φ j )(x) − K (φ j−1 )(x)

= η(φ j (y) − φ j−1 (y)) (1.50)
where η = a or η = b and y ∈ R is a point depending on x. Since K ( f )(x) = 0 for

x∈
/ [0, 3] and max(|a|, |b|) = a, from (1.50), we get
φ j+1 − φ j ∞ ≤ aφ j − φ j−1 ∞
so by induction, we get
φ j+1 − φ j ∞ ≤ aφ1 − φ0 ∞ .
Since φ1 − φ0 is finite, the sequence {φ j } converges uniformly to a continuous

function which is denoted by φ. This proves the first part of the theorem.
We know that [17, p. 92, 4.6], holds for x ∈ R. Since
supp φ ⊂ [0, 3] for each
x ∈ R, there are atmost three nonzero terms in the k∈Z φ(x − k). For a positive
integer M, let
M
FM (x) = φ(x − k)
k=−M
From (1.48), we conclude that
|FM (x)| ≤ C for some constant C
and
FM (x) = 1 i f |x| ≤ M − 3
= 0 if |x| ≥ M + 3
Thus, for every integer M

∞
2(M − 3) − 12C ≤ FM (x) d x ≤ 2(M − 3) + 12C (1.51)
−∞
From the definition of FM , we also conclude that

∞ ∞
FM (x) d x = (2M + 1) φ(x) d x (1.52)
−∞ −∞
Since (1.51) and (1.52) hold for every positive integer M, making M → ∞, we
obtain (1.48). To prove (1.49), let
∞
Lk = φ(x)φ(x − k) d x (1.53)
−∞
Since suppφ ⊂ [0, 3], we find that
L k = 0 for |k| ≥ 3 (1.54)
It is clear that
L k = L −k . (1.55)
By a change of variable, we see that, for any l, m, n ∈ Z

∞
1
φ(2x − m)φ(2x − 2l − n) d x = L 2l+n−m (1.56)
−∞ 2
Substituting value of φ(x) given by (1.52) into (1.53) for l = 0, 1, 2 and applying
(1.54) and (1.55), we obtain the following equations:
(a(1 − a) + b(1 − b))L 0 = (1 − ab)L 1 + (a(1 − a) + b(1 − b))L 2

2L 1 = (a(1 − a) + b(1 − b))L 0 + (1 − b)L 1
+((1 − b)2 + (1 − a)2 + b2 )L 2
2L 2 = abL 1 + (a(1 − a) + b(1 − b))L 2
From given values of a and b (see Lemma 1.6), we have
a(1 − a) + b(1 − b) = 0. (1.57)
Thus, the above system of equations becomes
0 = (1 − ab)L 1
2L 1 = (1 − ab)L 1 + ((1 − b)2 + (1 − a)2 + b2 )L 2
2L 2 = abL 1
This implies that L 1 = L 2 = 0. Thus L k = 0 for k = 0. By (1.48) and (1.49), we

can compute
∞ ∞
1= φ(x) d x = φ(x). φ(x − k) d x
−∞ −∞ k∈Z

= L k = L 0.
k∈Z
This proves (1.49).

Proof of Theorem 1.7 Suppψ(x) ⊂ [0, 3] follows immediately from (1.32) and the
fact that suppφ ⊂ [0, 3].
To obtain (1.33), substitute ψ(x) given by (1.32) into left-hand side of (1.33). We
obtain (1.33) using (1.56), (1.48), (1.57) and values of a and b. To obtain (1.34),
we proceed similarly but substitute both ψ(x) given by (1.32) and φ(x) given by
(1.45) into left-hand side of (1.34). It follows directly from (1.33) and (1.34) that
2 2j ψ(2 j t − k) j∈Z ,k∈Z is orthogonal.
1.7 Exercises
1.1 Show that the characteristic function of this interval [0, x] is W -continuous for
x ∈ (0, 1) if and if x ∈ Q (set of rational numbers).
1.2 Let f = χ[0,1/2] . Show that f is W -continuous but f (|x|) is not continuous on
G (Dyadic group).
1.3 Show that the map x → |x| is continuous from G to [0, 1] if the metric d(x, y) =
|x
!− y| is used on [0, 1]. Examine whether it is continuous if the dyadic metric
is used instead.
1.4 Show that Walsh system is orthonormal.
1.5 Prove that Walsh–Fourier transform is W -continuous and f (ω) → 0 as ω →
∞.
1.6 Find the integral of Walsh polynomial of degree 2.
1.7 Let h n (x) denote Haar system and let wn (x) denote Walsh system. What is the
relation between h n (x) and wn (x)?
1.8 Discuss relationship between Fourier and Walsh–Fourier transform.
1.9 Explain the concept of Walsh-type wavelet packet.
1.10 Draw the graph of the oldest known mother wavelet, the Haar wavelet dates
back to 1910 ⎧
⎨ 1 if 0 ≤ t < 1/2
ψ(t) = −1 if 1/2 ≤ t < 1
⎩
0, otherwise
1.11 Write down the scaling function φ associated with Haar wavelet.
1.12 Draw the graph of ψ(22 t) and ψ(2t), where ψ(t) and φ(t) denotes respectively
Haar wavelet and Haar scaling function.
1.13 Draw the graph of ψ(t − 3), ψ(t + 3), φ(t − 4) and φ(24 t).
1.14 Let ψ(t) be the Haar mother wavelet. Then show that ψ j,k (t) = 2 j/2 ψ(2 j t − k)
is an orthonormal system in L 2 (R).
1.15 Let f ∈ L 2 (R) and Pr ( f ) = j<r k∈Z f, ψ j,k ψ j,k , Haar wavelet expan-
sion. Prove that limr →∞ Pr ( f ) = f, that is, Pr ( f ) converges to f ∈ L 2 (R).
1.16 Let ψ(t) and φ(t) be, respectively, Haar mother wavelet and Haar scaling
function then show that ψ(t) = φ(2t) − φ(2t − 1).
1.17 Let for a function f defined on R Hölder’s condition with exponent α, 0 ≤ α <
1 is satisfied that is, there exists a constant C such that
| f (x) − f (y)| ≤ C|x − y|α for all x, y ∈ R.

∞
Show that f, ψ j,k = −∞ f (t)ψ j,k (t)dt exists for all j, k ∈ Z and satisfies
| f, ψ j,k | ≤ C.2− j (α+1/2) ,
where {ψ j,k } is system of Haar wavelet.

1.18 Let φ = χ[ 0, 1], V0 = { f ∈ L 2 (R)/ f is constant on [k, k + 1], ∀ k ∈ Z },
D j (V0 ) = V j . Prove that {V j } is a multiresolution analysis for L 2 (R).
1.19 Prove Theorem 1.6.
1.20 Let wavelet ψ(t) be given by (1.10), then prove that
∞
xψ(x)d x = 0
0
References
1. Haar, A. (2010). Zur theorie der orthogonal funktionen systeme. Mathematische Annalen, 69,
331–371.
2. Walsh, J. L. (1923). A closed set of normal orthogonal functions. American Journal of Math-
ematics, 45, 5–24.
3. Paley, R. E. A. C. (1932). A remarkable system of orthogonal functions. Proceedings of the
London Mathematical Society, 34, 241–279.
4. Fine, N. J. (1949). On the Walsh functions. Transactions of the American Mathematical Society,
65(3), 372–414.
5. Schipp, F., Wade, W. R., & Simon, P. (1990). Walsh series. New York: Adam Hilger.
6. Beauchamp, K. G. (1975). Walsh functions and their applications. New York: Academic Press.
7. Siddiqi, A. H. (1978). Walsh Function. Aligarh: AMU.
8. Maqusi, M. (1981). Applied Walsh analysis. London: Heyden.
9. Golubov, B. I., Efimov, A. V., & V. A. Skvortsov (2008) Walsh series and transforms (English
Translation Of 1st ed.). Moscow: Urss; Dordrecht: Kluwer (1991).
References 25
11. Meyer, Y. (1992). Wavelets and operators. Cambridge: Cambridge University Press.
13. Mallat, S. (1989). Multiresolution approximations and wavelets. Transactions of the American
Mathematical Society, 315, 69–88.
15. Frazier, M. W. (1999). An introduction to wavelets through linear algebra. New York: Springer.
17. Wojtaszyk, P. (1997). A mathematical introduction to wavelets (Vol. 37)., London mathematical
society Cambridge Cambridge: Cambridge University Press.
19. Fridli, S., Manchanda, P., & Siddiqi, A. H. (2008). Approximation by Walsh Norlund means.
Acta Scientiarum Mathmaticarum (Szeged), 74, 593–608.
20. Rodionov, E. A., & Farkov, Y. A. (2009). Estimates of the smoothness of Dyadic orthogonal
wavelets of Daubechies type. Mathematical Notes, 86(3), 407–421.
21. Nielsen, M. (2000). Walsh type wavelet packet expansion. Applied and Computational Har-
monic Analysis, 9, 265–285.
22. Mallat, S. (1999). A wavelet tour of signal processing. New York: Academic Press.
23. Coifman, R. R., Meyer, Y., Quake, S., & Wickerhauser, M. V. (1993). Signal processing and
compression with wavelet packets. Progress in wavelets analysis and applications (pp. 77–93).
Gif-Sur-Yvette: Frontieres.
24. Manchanda, P., & Meenakshi. (2009). New classes of wavelets. In A.H., Siddiqi, A.K. Gupta,
& M. Brokate (Eds.) Conference Proceedings Modeling Of Engineering and Technological
Problems (vol. 1146, pp. 253–271). New York
25. Malozemov, V. N., & Masharskii, S. M. (2002). Generalized wavelet bases related to the discrete
Vilenkin-Chrestenson transform. St. Petersburg Mathematical Journal, 13, 75–106.
Chapter 2
Walsh–Fourier Series
In this chapter, we study the growth and decay of Walsh–Fourier coefficients for
classes of functions, conditions under which Walsh–Fourier series converges point-
wise and converges absolutely, summability of Walsh–Fourier series and degree
(order of approximation) for different classes of functions and different means of
Walsh–Fourier series.
2.1 Walsh–Fourier Coefficients
Let f ∈ L 1 (0, 1) then Walsh–Fourier coefficient of f is defined as

1
ck = fˆ(k) = f (x)wk (x) d x,
0
or
fˆ(x) = f ψn dμ, f ∈ L 1 (G),
G
∞ the Walsh–Fourier coefficients of f ∈ L 2 (0, 1) or f ∈ L 1 (G).

is called
k=0 ck ψk (x) is called the Fourier series of f ∈ L 1 (0, 1) and similarly we can
define Walsh–Fourier series on dyadic group G.
We present below a theorem which does not have an analogue in the Fourier
analysis.

Theorem 2.1 Let ∞ k=0 ck ψk (x) be the Walsh–Fourier series of the function f (x)
and h be any fixed number, then {ck ψk (h)} is a sequence of Walsh–Fourier coefficients
of f (x +̇h).

https://doi.org/10.1007/978-981-13-6370-2_2
28 2 Walsh–Fourier Series
Proof We have ψn (x) = ψn (h)ψn (x +̇h) at most everywhere and so

1 1
ψn (h) f (x +̇h) d x = ψn (h)ψn (x +̇h) f (x +̇h) d x
0 0
1
= ψn (h) ψn (x +̇h) f (x +̇h) d x
0
1
= ψn (h) ψn (x) f (x) d x
0
= ψn (h)cn
by Theorem 1.4.1 of Siddiqi [1]. Thus Walsh–Fourier coefficients of f (x +̇h) are

cn ψn (h), n = 0, 1, 2, 3, . . ..
Theorem 2.2 (Riemann–Lebesgue Theorem) If f ∈ L 1 then fˆ(n) → 0 as n → ∞.

This tells us that the Walsh–Fourier coefficients of Lebesgue integrable function tend
to zero as n becomes large.
Proof For given ε > 0 there is a g ∈ P (it denotes the set of Walsh polynomials)
such that f − g L 1 < ε, by the fact that Walsh polynomials are dense in L 1 .
1 1
f wn = ( f − g)wn for large n,
0 0
as ĝ(n) = 0 for large n.

Therefore,
lim sup| fˆ(n)| ≤ f − g L 1 < ε
n→∞
implying the desired result.
It may be observed that for f ∈ L 2 and Walsh–Fourier coefficients ck ’s of f we

have 1
∞
ck =
2
| f (x)|2 d x.
k=0 0
This is known as Parseval’s formula.
2.1.1 Estimation of Walsh–Fourier Coefficients
We present below the results concerning estimates of the growth of Walsh–Fourier

coefficients of functions from various classes such as L 1 , L p , continuous, functions
of bounded variation, and absolutely continuous. For functions defined on I = [0, 1]
or dyadic intervals [0, 2n ] which are W -continuous (continuous W ),
2.1 Walsh–Fourier Coefficients 29
w( f, I ) = sup{| f (x +̇y) − f (x)|/x ∈ I, 0 ≤ y ≤ 1}
is called the local moduli of continuity.

For 1 ≤ p < ∞ and each f ∈ L p
1/ p
1
w( p) ( f, I ) = sup 0≤y<|I | | f (x +̇y) − f (x)| p d x .
|I | I
The local modulus of continuity w( f, I ) is a measure of the oscillation of f on I . A

function f is said to be p-bounded fluctuation for some 1 ≤ p < ∞ if
2n −1 1/ p

sup n∈P |w( f, I (k, n))| p
< ∞.
k=0
For a bounded fluctuation function (1-bounded fluctuation)f, the total fluctuation is

denoted by Fl( f ) and it is defined as
2n −1

Fl( f ) = sup n∈P |w( f, I (k, n))| 2
.
k=0
Theorem 2.3 For f ∈ L 1
1 1
| fˆ(k)| ≤ w ( f, 1/k). (2.1)
2
Proof For fixed k ∈ P, choose n ∈ N such that 2n ≤ k ≤ 2n+1 . It has been proved
that
f − τ2−(n+1) fˆ (k) = fˆ(k)(1 − wk 2−(n+1) ),
τ2 fˆ(k) = wk (y) fˆ(k), k ∈ N
for details, see Schipp, Wade and Simon [2]. It follows that
1 ˆ
fˆ(k) = f − τ2−(n+1) f (k),

2
keeping in mind wk (2−(n+υ) ) = −1. Therefore,


1 1
ˆ
| f (k)| = ( f (x) − f (x +̇2 −(n+1)
))wk (x) d x
2 0
1
≤ w(1) ( f, 2−(n+1) )
2
1
≤ w(1) ( f, 1/k).
2
Remark 2.1 Theorem 2.3 can be extended for L p and C W , that is,
(i) | fˆ(k)| ≤ 21 w p ( f, 1/k), k ∈ P and f ∈ L p , 1 < p < ∞.
(ii) | fˆ(k)| ≤ 21 w( f, 1/k), k ∈ P and f ∈ C W .
Theorem 2.4 Let f be of bounded fluctuation, then
Fl( f )
| fˆ(k)| ≤ , k ∈ P. (2.2)
k
1 1
Proof Since 0 f (x +̇y) d x = 0 f (x) d x, that is, Lebesgue measure is invariant
with respect to +̇, we have
1
2 fˆ(k) = f (x +̇2−(n+1) )wk (x +̇2−(n+1) ) d x
0
1
+ f (x)wk (x) d x for any integer n, k ∈ N .
0
Therefore, for 2n ≤ k < 2n+1 , we get

1
2 fˆ(k) ≤ | f (x +̇2−(n+1) ) − f (x)| d x
0
−(n+1)
2
= | f (x +̇2−(n+1) ) − f (x)| d x.
p=0 I ( p,n)
It follows from the definition of local moduli of continuity that
2
n
−1
2| fˆ(k)| ≤ 2−n |w( f, I ( p, n))| ≤ 2−n Fl( f ),
p=0
as |I ( p, n)| = 2−n for 0 ≤ p < 2n .

The desired result follows from the above inequality.
Remark 2.2 There exists a function of bounded variation whose Walsh–Fourier coef-
ficients are not o(1/k) as k → 0, that is, (k fˆ(k)) does not tend to 0, as k → ∞.
Remark 2.3 Let F be absolutely continuous on [0, 1] and let F̂(k) = 0(1/k) as
k → ∞. Then F is constant on [0, 1].

Theorem 2.5 Let ∞ k=0 ck wk (x) be Walsh–Fourier
x series of f (x) and ∞k=0 bk wk (x)
be Walsh–Fourier series of F(x) = 0 f (t) dt. Then for fixed k ≥ 0 and n → ∞,
b2n +k = −2−(n+2) ck + o(2−n ). (2.3)
See proof [1, 2].
2.1.2 Transformation of Walsh–Fourier Coefficients
Theorem 2.6 Let f (x) be a real valued, periodic function of mean value zero on
[0, 1]. Furthermore, let F(x) = 0 f (t) dt and ∞
x
k=0 bk wk be the Walsh–Fourier
series of F(x). Then, the arithmetic mean of the sequence {k|bk |} → 0.

Proof Let ∞ k=0 ck ψk (x) be the Walsh–Fourier series of f (x). Then by Theorem
2.5
b2n +k = −2−(n+2) ck + o(2−n ), for
fixed k ≥ 0 and n → ∞. From it follows that k|bk | ≤ |ck | + 0(1), where k = 2v +

k , k < 2v .
Therefore
∞
1 1
k|bk | ≤ |ck |
n + 1 k=0 n + 1 0≤k≤n,k ≤T
1
+ |ck | + o(1)
n + 1 0≤k≤n,k >T
= S1 + S2 + o(1).
Given ε > 0, choose T so that ck < ε for k > T . Then S2 < ε. For S1 write k =
2v + k , where 0 ≤ v ≤ r , 0 ≤ k ≤ T and r is defined by 2r ≤ n ≤ 2r +1 . Then
1
r r
1 + log2n
S1 ≤ |ck | ≤ M = o(1).
n + 1 v=0 k =0 n+1
n
Therefore, 1
n+1 k=0 k|bk | = o(1) which completes the proof of Theorem 2.6.
Theorem 2.7 Let A = (am k ) be a regular matrix f ∈ BV [0, 1] and Bk (x) denote
the sequence {kck ψk (x)} where ck are Walsh–Fourier coefficients of f . Then for
every x ∈ [0, 1]
∞

lim am,k Bk (x) = 0 (2.4)
m→∞
k=0
if and only if
lim amk k Jk (x) = 0, (2.5)
m→∞
k=0
x
where Jk (x) = 0 wk (t) dt, k = 0, 1, 2, . . ., in every 0 < δ < t ≤ 1.
Proof It can be seen that

∞
∞

amk Bk (x) = amk kck wk (x)
k=0 k=0
∞ 1
= amk k f (t)wk (x)ψk (t) dt
k=0 0
∞ 1
= amk k f (t)ψk (x +̇t) dt, by Theorem1.4.1 [82], also
k=0 0
∞ 1
= amk k f (x +̇t)ψk (t) dt
k=0 0
∞
= amk k[ f (x +̇t)Jk (t)]10 −
k=0
∞ 1
amk k Jk (t) d f (x +̇t), integration by parts,
k=0 0
∞
1
=− amk k Jk (t) d f (x +̇t),
0 k=0

as ∞ k=0 amk k f (x +̇t)Jk (t) = 0 by virtue that Jk (1) and Jk (0) are zero.
Thus 1
∞
∞
| amk Bk (x) = − amk k Jk (t)d f (x +̇t).
k=0 0 k=0

Let K m (t) = ∞ k=0 amk k Jk (t)
and φx (t) = f (x +̇t). In order to prove the theorem we
have to show that if limm→∞ ∞ k=0 amk k Jk (x) = 0 holds then for every f ∈ BV [0, 1]
and for every x ∈ [0, 1],
1
lim K m (t) dφx (t) = 0 (2.6)
n→∞ 0
and conversely. Condition (2.6) is equivalent to the following condition:

1
lim K m (t) dφx (t) = 0 for every f ∈ BV [0, 1] (2.7)
m→∞ 0
for every x ∈ [0, 1] and for 0 < δ < 1. For f ∈ BV [0, 1] and x ∈ [0, 1], given any
ε > 0 there exists a δ > 0 such that
δ
ε
|dφx (t)| < . (2.8)
0 2M
By virtue of the regularity condition, we have

m ∞

|K m (t)| ≤ |amk ||Jm (t)k| ≤ M |amk | ≤ M (2.9)
k=0 k=0
as |k Jm (t)| ≤ M by Theorem 1.6.1 [1] so that

1 t ε
K m (t)dφx (t) − K m (t)dφx (t) < .
2
0 δ
Thus (2.6) and (2.7) are equivalent.

By a well theorem of weak convergence [3] of sequences in Banach space of all
continuous functions defined on a finite closed interval, it follows that (2.7) holds if
and only if
(i) |K m (t)| ≤ M for all m and t in [δ, 1] and
(ii) (2.5) holds.
Since (i) holds by virtue of (2.9), it follows that (2.7) holds if and only if (2.5) holds.
This completes the proof.
Corollary 2.1 For f ∈ BV [0, 1] and Walsh–Fourier coefficients ck of f , the arith-

metic mean of {kck } → 0 as k → ∞ if and only if the arithmetic mean of {k Jk (t)} →
0 as k → ∞ for 0 < δ ≤ t < 1.
Proof Choose x = 0 and
amk = 1/m, k < m

= 0, k ≥ m
in Theorem 2.7 we get Corollary 2.1.
We conclude the section by the following remark.

Remark 2.4 Let c1 , c2 , c3 , . . . be Walsh–Fourier coefficients of a function belonging
certain spaces such as L 1 , L p , C[a, b], C[G]. A natural question is whether Cn =
to
1 n
n k=0 ck will be also Walsh–Fourier coefficients of the same class of functions.
There exists a vast literature on this theme in the Fourier analysis, but there are only
few results on this problem in Walsh–Fourier

analysis, see Siddiqi [1] for details.
A typical known result is as follows: If ∞ 0 ck wk (x) is a Walsh–Fourier series of
a function f (x) ∈ L p (0, 1), 1 < p < 2, then ∞ k=0 C 2k wk (x) is the Walsh–Fourier
series of a function F(x) ∈ L q (0, 1), where 1p + q1 = 1.
2.2 Convergence of Walsh–Fourier Series
To revise notion of ordinary, uniform, almost absolute and L p , 1 ≤ p < ∞ norm,

convergence is introduced for sequences in Appendix. In this section,
we study such
concepts for partial sums of Walsh–Fourier series, namely, Sn (x) = n−1k=0 ck wk (x),
1
where ck = 0 f (x)wk (x) d x or ck = G f (x)wk (x) d x. Let Dn denote the Walsh–
Dirichlet kernel of order n, namely, D0 = 0 and

n−1
Dn (x) = wk (x), n ∈ P. (2.10)
k=0
It can be checked that

(a)
Sn f = f ∗ Dn f for f ∈ L 1 (G) and n ∈ N , (2.11)
where * denotes the convolution.

(b)
∞

Dn = wn n k (D2k+1 − D2k ) (2.12)
k=0
(c)
∞

Dn = wn n k ρk D2k (2.13)
k=0
(d) If 0 < n < 2m then
1
m−1
1
Dn = (D2m − wn ) − (γe wn )ρk D2k see [2, 5, 6]. (2.14)
2 2 k=0 k
Remark 2.5 The general problem of finding conditions under which it converges
in different senses, namely, convergence pointwise, uniform convergence, almost
convergence, absolute convergence, and L p norm convergence is quite challenging
task. Conditions (2.11) to (2.14) play very crucial rule in study of these problems.
We cite references [1, 2] for detailed study of these problems. We state and prove
below few typical results.
2.2 Convergence of Walsh–Fourier Series 35
Theorem 2.8 (i) S2n f → f uniformly on G as n → ∞ for all f ∈ C(G).

(ii) S2n f → f uniformly on [0, 1] as n → ∞ for f ∈ C W .

Theorem
∞ 2.9 Let ck → 0 and ∞ k=0 |ck − ck−1 | < ∞ then the Walsh–Fourier series
c w
k=0 k k (x) converges uniformly in δ ≤ u < 1, δ > 0.
Theorem 2.10 Let ( f (x) − c)/(u − x0 ) be absolutely integrable in interval |u −

x0 | < δ for some δ > 0. Then the Walsh–Fourier series of f (x) converges to c at the
point x0 .
Theorem 2.11 If f (x) is continuous and if its modulus of continuity satisfies

w(δ, f ) = o(log δ −1 )−1 as δ → 0, then its Walsh–Fourier series converges to f (x)
uniformly.
Theorem 2.12 Let f (x) ∈ Li pα, α > 21 . Then the Walsh series of f (x) converges
absolutely.
Theorem 2.13 Let f (x) be a continuous function of period 1. Let

n+1
2 p

u n (x) = p −1 f (x)+̇ n+1 − f (x +̇(2 p + 1))/2n+1 .
p=1
2
Then
lim u n (x) = 0 uniformly in x

n→∞
implies that
lim Sk ( f, x) = f (x) uniformly in x.
k→∞
Theorem 2.14 Let f (x) satisfy

1 1
| f (x + t) − f (x − t)| p
d x dt < ∞, 2 ≥ p ≥ 1.
0 o t
Then, the Walsh–Fourier series of f (x) converges almost everywhere.
Theorem 2.15 If f (t) is a p-periodic function defined on [0, 1] and

f (x) = ∞ n=0 cn wn (t),
∞
n=0 |cn | < ∞, then f 1(t) = ∞n=0 dn wn (t), where
∞
n=0 |dn | < ∞ provided inf t | f (t)| > 0.
Theorem 2.16 If 1 < p < ∞ and f ∈ L p then
||Sn f − f || L p → 0, as n → ∞.
Theorem 2.17 For f ∈ L p , p > 1, Sm f converges to f almost everywhere.

Theorem 2.18 For f ∈ L 1 and satisfying
1
w(1) ( f, δ) = o(log )−1 as δ → 0,
δ
||Sn f − f || L 1 → 0 as n → ∞.
We prove here Theorems 2.9–2.12 and 2.16, for the proof of other theorems, we
refer to [1, 2]. It may be remarked that short proof of Theorem 2.15 is available
using results from Banach Algebra. In the proof of Theorem 2.9, we require the
following lemma.
Lemma 2.1 For all u such that 0 < u < 1, |Dk (u)| < 2/u.
Proof Suppose that 2−n ≤ u < 2−n+1 , and write k in the form k = p2n + q, 0 ≤ q ≤
2n . We have wr (u)ws (u) = ws+r (u) provided that r and s have dyadic expansions
without any exponents common to both, and
p−1 2 −1

n q−1
k−1
Dk (u) = wr (u) = wr 2n +1 (u) + w p2n +1 (u)
r =0 r =0 i=0 i=1
p−1 2 −1

n q−1
= w2n (u)wi (u) + w p.2n (u)wi (u)
r =0 i=0 i=0

p−1
= D2n (u) wr (2n u) + w p (2n u)Dq (u)
r =0
= D2n (u)D p (2n u) + w p (2n u)Dq (u).
Since D2n (u) contains 1 + w2n−1 (u) = 1 + w1 (2n−1 u) as a factor, and 1

2
≤ 2n−1
u < 1, D2n (u) = 0, hence
|Dk (u)| = |Dq (u)| ≤ q ≤ 2n < 2/u.
This proves the lemma.
Proof of Theorem 2.9 We have

∞

n−1
ck wk (u) = (ck − ck+1 )Dk (u) + ak Dn (u)
k=0 k=0
by Abel’s transformation. By Lemma 2.1 and the fact that ck → 0 as k → ∞,

∞
k=0 ck wk (x) converges uniformly in δ ≤ u < 1.
Proof of Theorem 2.10 We have

1
sk (x0 , f ) − c = ( f (u) − c)Dk (x0 + u) du
0
= ( f (u) − c)k(x0 + u) du + ( f (u) − c)Dk (x0 + u) du.
|u−x0 |<δ |u−x0 |≥δ
By Lemma 2.1 and

2 2
|Dk (x0 +̇u)| < ≤ .
x0 +̇u |u − x0 |
We have

f (u) − c
|sk (x0 , f ) − c| ≤ 2
u − x du
|u−x0 |<δ 0

+ ( f (u) − c)Dk (x0 +̇u) du .
|u−x0 |>δ
Given ε > 0, we can choose δ small enough so that the first integral does not exceed
δ/2; with δ thus chosen and fixed, the second integral can be made less than ε/2 for
k > k0 , (δ, ε) = k0 (ε). This completes the proof of the theorem.
Proof of Theorem 2.11 By Theorem 2.8, s2n (x) → f (x) uniformly. It is sufficient
to consider for any given k = 2n + k , 0 ≤ k < 2n , the difference
1
s2n +k (x) − s2n (x) = w2n (x)w2n (u)Dk (x +̇u) f (u) du. (2.15)
0
We know that for p < 2n , w p (2−(n+1) ) = 1 so that
Dk (z +̇2−(n+1) ) = Dk (z); also w2n (u +̇2−(n+1) ) = −w2n (u).
By invariance of integration, we get

1
s2n +k (x) − s2n (x) = − w2n (x)w2n (u)Dk (x +̇u) f (u +̇2−(n+1) ) du (2.16)
0
By adding (2.15) and (2.16), we get

1
2(s2n +k (x) − s2n (x)) = w2n (x +̇u)Dk (x +̇u){ f (u) − f (u +̇2−(n+1) )} du
0
or
1
−(n+1)
2|s 2n +k − s (x)| ≤ max0≤u≤1 | f (u) − f (u +̇2
2n )| |Dk (x +̇u)| du
0
≤ w(2−(n+1) , f )L k
≤ w(2−n , f )log(2n )
= 0(1).
This proves the theorem.

∞
Proof of Theorem 2.12 ∞ By Theorem 2.1, k=0 ck wk (h)wk (x) is Walsh–Fourier
series of f (x +̇h) if u=0 ck wk (x) is Walsh–Fourier series of f (x). By Parseval’s
formula 1 ∞

[ f (x +̇h) − f (x)]2 d x = ck2 (1 − wk (h))2 .
0 u=0
Set h = 2−(n+1) . Then wk (2−(n+1) ) = −1 for 2n ≤ k < 2n+1 . Hence
−1
2n+1 1
ck2 ≤ [ f (x +̇2−(n+1) ) − f (x)]2 d x.
k=2n 0
Since | f (x +̇h) − f (x)| ≤ w(δ, f ), we have
−1
2n+1
ck2 ≤ [w(2−(n+1) , f )]2 ≤ A2−nα ,
k=2n
where A is constant. By Schwarz’s inequality

⎛ ⎞1/2 ⎛ ⎞1/2
−1
2n+1 −1
2n+1 −1
2n+1
|ck | ≤ ⎝ ck2 ⎠ ⎝ 12 ⎠
k=2n k=2n k=2n
≤ A1/2 2−n(α−1)/2 .
Since α > 1/2, the right-hand side of the above inequality is the nth term of a con-
vergent series, which completes the proof.
Proof of Theorem 2.16 For ε > 0, choose a Walsh polynomial P such that
f − P L p < ε. Since Sn P = P for n large, it is clear that
f − Sn f L p ≤ f − P L p + §n P − Sn f L p .
By Corollary 6 of Sect. 3.3, [2] gives the desired result.

2.2.1 Summability in Homogeneous Banach Spaces
Definition 2.1 Let X be a Banach space with norm .. It is called a dyadic homo-
geneous Banach space if P ⊆ X ⊆ L 1 (G) hold, where P denotes the set of Walsh
polynomials.
f L 1 ≤ f X , ∀ f ∈ X. (2.17)
For Tx f (t) = f (x +̇t), t ∈ [0, 1).
Tx f = f X , x ∈ [0, 1). (2.18)
P is dense in X , that is,

P̄ = X.
L p (1 ≤ p < ∞), H (dyadic Hardy space), and C(G) are homogeneous spaces, for
details, see [2, 4].
Definition 2.2 A sequence {Pn , n ∈ N } of functions in C(G) is called an approxi-

mate identity (for dyadic convolutions) if
Pn = O(1) , as n → ∞. (2.19)

Pn dμ = 1, n ∈ N (2.20)
G

lim |Pn | dμ = 0 for eachk ∈ N . (2.21)
n→∞ G\Ik (o)
n
{D2n , n ∈ N } and {K n , n ∈ N }, where K n = 1
n k=1 Dk , n ∈ N are examples of
approximate identity.
The following theorem

shows that in any homogeneous Banach space S2 f and
n
σn f , where σn f = n1 nk=1 sk f , converge to f in norm as n → ∞.

Theorem 2.19 Let {Pn , n ∈ N } be an approximate identity and X be a homogeneous
Banach space. Then
lim Pn ∗ f − f = 0 for all f ∈ X.

n→∞
For proof of this theorem, we refer to [2, p. 158].

2.3 Approximation by Transforms of Walsh–Fourier Series
Let A = (amn ), m = 1, 2, 3, . . . , n = 1, 2, 3, . . . be an infinite matrix of real num-

bers. Then a transform of a sequence {sn }, y = {yn } is the sequence defined by the
equation
∞
ym = amn xn , m = 1, 2, 3, . . . .
n=1
n−1
Let sn f = k=0 ck wk (x), where ck is the kth Walsh–Fourier coefficient of f ∈ L 2 ,
then
∞

tm ( f, x) = amn sn ( f, x)
n=1
is called the matrix mean of Walsh–Fourier series. The general problem is to

evaluate tm f − f X for different X and special cases of tm . For X = Li p α and
{tm }(C, 1) mean, namely,
1
amn = for n < m
m
= 0 for n ≥ m.

β β +n β
For amn = An = for n ≥ 1 and A0 = 1, β = −1, −2, . . ..
n
For X = Li p(α, p), Li p α, Li p j (t) and (C, 1) and (C, β) means the degree of
approximation, that is, evaluation of tm f − f X was carried out between 1947 and
1963. A comprehensive account of this development is presented in [1]. In Sect. 2.3.1,
approximation by Cesàro mean of f ∈ Li p α is presented. Case of Nörlund means
of Walsh–Fourier series in L p space is studied in Sect. 2.3.2. The same problem in
homogeneous Banach spaces is presented [4] in Sect. 2.3.3.
2.3.1 Approximation by Césaro Means of Walsh–Fourier

Series
Fejer kernel or kernel for (C, 1) summability for Walsh–Fourier series is defined as
1
k
K k (x, u) = Dr (x, u).
k r =1
This can be written as

2.3 Approximation by Transforms of Walsh–Fourier Series 41
1
k
K k (x, u) = K k (x +̇u) = Dr (x +̇u)
k r =1
(x +̇u = dyadic rational) by setting K k (u) = K x (0, u).

(C, 1) mean of order k is defined as
1
σk ( f, x) = K k (x +̇u) f (u) du.
0
Walsh–Fourier series of f (x) is said to uniformly summable by (C, 1) mean to f (x)

if σn (x, f ), converges uniformly to f (x). For any nonnegative integer n and any real
α(> −1), we define
(α + 1)(α + 2) . . . (α + n)
Aαn = , Aα0 = 1.
n!
The kernel for (C, α) summability is defined by
1
n−1
K n(α) (x) = Aαn−k−1 wk (x), n = 1, 2, 3, . . . .
Aαn−1 k=0
Lemma 2.2 (i) For n ≥ 0, 0 ≤ k ≤ 2n ,
(2n + k )K 2n +k (u) = 2n K 2n (u) + k D2n (u) + w2n (u)k K k (u).
(ii) K 2n (u) ≥ 0 for n ≥ 0.
Proof (i) By definition and by the relation
Dn (t) = D2n (t) + w2n (t)Dk (t),
where k = 2n + k , with 0 ≤ k < 2n , we have

n
2 k
(2n + k )K 2n +k (u) = Dr (u) + D2n +q (u)
r =1 q=1

k
= 2 K 2n (u) +
n
{D2n (u) + w2n (u)Dq (u)}
q=1

k

= 2 K 2n (u) + k D2n (u) + w2n (u)
n
Dq (u)
q=1
= 2n K 2n (u) + K D2n (u) + w2n (u)k K k (u).

(ii) Take k = 2n in part (i), then
2n+1 K 2n+1 (u) = (1 + w2n (u))2n K 2n (u) + 2n D2n (u).
Since D2n (u) ≥ 0, 1 + w2n (u) ≥ 0, K 1 (u) = 1. By the principle of induction,

the result holds for all n.
The failure of the estimate of Fejér kernel has created great difficulties in proving the
analogue of Fejér approximation theorem. For several years, it was an open problem.
We present the proof of this result by Yano [1, 2], and a generalization of Yano’s
result by Siddiqi and Gupta [1].
Theorem 2.20 If f (x) ∈ Li pα, 0 < α < 1, then σn (x; f ) − f (x) = O(n −α ),
where σn (x, f ) denote the arithmetic mean of the Walsh–Fourier series of f (x).
We require the following lemmas in the proof of this theorem.
Lemma 2.3 Let K n (t) be the Fejér kernel for the Walsh functions and let I p(n) denote
the interval p.2−n ≤ t < ( p + 1)/2−n , 0 ≤ p < 2n . Then for n ≥ 2;
2n + 1
(i) K 2n (t) = (t ∈ I0(n) ),
2
(ii) K 2n (t) = 2n−r −2 (t ∈ I2(n),r
n
=0,1,...,n−1
),
(iii) K 2 (t) = 0 elsewhere in (0, 1).
n (2.22)
Lemma 2.4 Under the assumption of Theorem 2.1
σ2n (x, f ) − f (x) = O(2−αn ). (2.23)
Proof of Lemma 2.3(i) This lemma holds clearly for n = 2. Suppose (2.22) holds
for n ≥ 2, then
1 + ψ2n (t) 1
K 2n+1 (t) = K 2n (t) + D2n (t). (2.24)
2 2
By using this relation, we show that (2.22) holds for n + 1. We know that K 2n+1 (t)
is constant on I p(n+1) , 0 ≤ p < 2n+1 , on I0(n+1) and I1(n+1) , D2n (t) = 2n , on I2m
(n+1)
ψ2n
(n+1) (n+1)
(t) = 1, and on I2m+1 , w2n (t) = −1. Hence, on I0 , K 2n+1 (t) = K 2n (t) + 2 n−1
=
(2(n−1) /2 + 2n+1 ) = (2n+1 + 1)/2. So that (i) is true for (n + 1). On I1(n+1) , K 2n+1
(t) = 2n−1 , so that (ii) is true for (n + 1), with r = 0. For 2−n ≤ t < 1, D2n (t) =
0, hence if t ∈ I2(n+1)
r ⊂ I2nr −1 for 1 ≤ r < n + 1, we have K 2n+1 (t) = K 2n (t) =
2n−(r −1)−2 = 2(n+1)−r −2 , and (ii) is true for 0 ≤ r < n + 1. Now (iii) follows from
the fact that the integral of K 2(n+1) (t) over (0, 1) is 1, that K 2(n+1) (t) is nonnegative
and that the integral over the interval specified in (i) and (ii) (with n replaced by
n + 1) is
2n+1 + 1 1
2−(n+1) { + 2n−1 + · · · + 1 + } = 1.
2 2
This completes the proof of the lemma.
Proof of Lemma 2.4 We have

1
σ2n (x; f ) − f (x) = [ f (x +̇t) − f (x)]K 2n (t) dt.
0
Then by Lemma 2.1

−n
2n + 1 2
σ2n (x; f ) − f (x) = [ f (x +̇t) − f (x)] dt
2 0

n−1 2−n+i+2−n
+ 2 n−i−2
[ f (x + t) − f (x)] dt.
i=0 2−n+ j
2−n
n−1 −n+2−n
σ2n (x, f ) − f (x) = 2n o(t α ) dt + 2n−i−2 o(t α ) dt
0 i=0 2−n+i

n−1
= o(2−αn ) + 2n−i−2 o(2α(−n+i+1) )2−n
i=0

n−1
= o(2−αn ) + o(2αn ) 2−(1−α)i
i=0
= o(2−αn ).
This completes the proof of the lemma.
Proof of Theorem: Let n = 2n 1 + 2n 2 + . . . + 2nr , (n 1 > n 2 > n 3 . . . > n r ≥ 0),

n (0) = n, n 1 = n − 2n 1 , n (i) = n (i−1) − 2ni , where i = 2, 3, . . . , r − 1, and n (r ) = 0.
Then by Lemma 2.2, we get
1
σn (x, f ) − f (x) = [ f (x +̇t) − f (x)]K n (t) dt
0

1 ni 1
r
(i−1)
= 2 [ f (x +̇t) − f (x)]ψn−n (t)K 2ni (t) dt
n i=1 0

1 (i) 1
r
(i−1)
+ n [ f (x +̇t) − f (x)]ψk−k (x)D2in (t) dt
n i=1 0
= Pn + Q n , (2.25)
say. By Lemma 2.3 and the fact |ψn (x)| = 1, we get


1 ni 1
r
|Pn | ≤ 2 | f (x +̇t) − f (x)|K 2ni (t) dt
n i=1 0
r n
1 1 1
= O 2ni 2−αni = O 2(1−α)i (2.26)

n i=1
n i=1
1
= O(2(1−α)n 1 ) = O(n −α ).
n
On the other hand, by the property of D2ni (t) stated in the proof of Lemma 2.3
1 2−n

[ f (x +̇t) − f (x)]D2ni (t)| dt| ≤ 2ni O(t α ) dt = O(2−αni ). (2.27)

0 0
By definition n (i) < 2ni . Therefore

r r
1 1
|Q n | ≤ O n (i) 2−αni = O 2(1−α)ni = O(n α−x ). (2.28)
n i=1
n i=1
By (2.25), (2.26), and (2.27), the theorem is proved. Theorem 2.20 can be generalized
in the following manner.
(β)
Theorem 2.21 If f (x) ∈ Li p α, then for any β > α, σn (x, f ) − f (x) = O(n −α ),
(β)
where σn (x, f ) denotes the (C, β) mean of the Walsh–Fourier series of f (x).
A function f (x) is said to belong Li p j (t) class if the following condition is satisfied:

| f (x + t) − f (x)|
f = sup t>0,x < ∞,
f (t)
where j (t) is a positive and nondecreasing function defined on the interval (0, 1).
By taking j (t) = t α , we get Li pα class. This class was first introduced by A. H.
Siddiqi and further studied by M. Izumi and S. Izumi, see [1] for references.
Theorem
1 2.22 If f (x) ∈ Li p j (t), then σn (x, f ) − f (x) = O( j (1/n)) where
−2
t j (u)u du ≤ A j (t)t −1 , as t → 0 and j (t) is subadditive.
Theorem 2.23 If f (x) ∈ Li p j (t) then for 0 < β < 1

2N
σn(β) (x, f ) − f (x) = O n −β
j (1/t)t β−1
dt ,
0
1
where t j (u)u −2 du ≤ At −1 j (t) as t → 0 and N > n, 2 N −1 < n < 2 N , j (t) is
subadditive.
2.3.2 Approximation by Nörlund Means of Walsh–Fourier

Series in L p Spaces
This section is devoted to the study of approximation by Nörlund means for Walsh–
Fourier series of a function in L p and in particular, in Li p(α, p) over the unit interval
(0, 1) where α > 0 and 1 ≤ p ≤ ∞. In case p = ∞, by L p , we mean C W , the collec-
tion of uniformly W -continuous functions over (0, 1). The discussion in this section
is based on paper of Moricz and Siddiqi [5]. As special cases, we obtain the earlier
results by Yano [12, 13], Jastrebova [14] and Skvorcov [10] on the rate of approxi-
mation by Cesáro means. We denote by Pn the collection of Walsh polynomials of
order less than n, that is, function of the form

n−1
P(x) = ck wk (x),
k=0

where n ≥ 1 and {ck } is any sequence of real(or complex) numbers. Denote by m
the finite σ -algebra generated by the collection of dyadic intervals of the form
Im (k) = [k2−m , (k + 1)2−m ), k = 0, 1, . . . , 2m − 1,

where m ≥ 0. It is not difficult to see that the collection of m -measurable functions
on I coincides with P2m , m ≥ 0. We will study approximation by means of Walsh
polynomials in the norm L p = L p (I ), 1 ≤ p < ∞, and C W = C W (I ). We remind
the reader that C W is the collection of functions f : I → R that are uniformly contin-
uous from the dyadic topology of I to the usual topology of R, or in short, uniformly
W -continuous. The dyadic topology is generated by the union m for m = 0, 1, . . ..
As is known (see, e.g., [2, p. 9]), a function belongs to C W if and only if it is
continuous at every dyadic irrational of I , is continuous from right on I , and has a
finite limit from the left on (0, 1], all these in the usual topology. Hence, it follows
immediately that if the periodic extension of a function f from I to R with period 1
is classically continuous, then f is also uniformly W -continuous on I. The converse
statement is not true. For example, the Walsh functions wk belong to C W , but they
are not classically continuous for k ≥ 1.
For the sake of brevity in notation, we agree to write L ∞ instead of C W and set
1 1/ p
f p = | f (x)| p d x , 1 ≤ p < ∞,
0
f ∞ = sup{| f (x)| : x ∈ I }.
After these preliminaries, the best approximation of a function f ∈ L p , 1 ≤ p ≤ ∞,

by polynomials in Pn is defined by
E n ( f, L p ) = in f P∈P n f − P p .
Since Pn is a finite-dimensional subspace of L p for any 1 ≤ p ≤ ∞, this infimum

is attained.
From the results of [2, pp. 142 and 156–158], it follows that L p is the closure of
the Walsh polynomials when using the norm ., 1 ≤ p ≤ ∞. In particular, C W is
the uniform closure of the Walsh polynomials.
Next, define the modulus of continuity in L p , 1 ≤ p ≤ ∞ of a function f ∈ L p
by
w p ( f, δ) = sup |t|<δ τt f − f p δ > 0,
where τt means the dyadic translation of t:
τt f (x) = f (x +̇t), x, t ∈ I.
Finally, for each α > 0, Lipschitz classes in L p are defined by
Li p(α, p) = { f ∈ L p : w p ( f, δ) = O(δ α )asδ → 0}.
Unlike the classical case, Li p(α, p) is not trivial when α > 1. For example, the
function f = w0 + w1 belongs to Li p(α, p) for all α > 0 since
w p ( f, δ) = 0 when 0 < δ < 2−1 .
Main Results Given a function f ∈ L 1 , its Walsh–Fourier series is defined by

∞
1
ck wk (x), where ak = f (t)wk (t). (2.29)
k=0 0
The nth partial sums of series in (2.29) are

n−1
sn ( f, x) = ck wk (x), n ≥ 1.
k=0
As is well known, 1
sn ( f, x) = f (x +̇t)Dn (t) dt,
0
where

n−1
Dn (t) = wk (t), n ≥ 1,
k=0
is the Walsh–Dirichlet kernel of order n. Let {qk : k ≥ 0} be a sequence of nonneg-

ative numbers. The Nörlund means for series (2.29) are defined by
1
n
tn ( f, x) = qn−k sk ( f, x),
Q n k=1
where

n−1
Qn = qk , n ≥ 1.
k=0
We always assume that q0 > 0 and
lim Q n = ∞. (2.30)
n→∞
In this case, the summability method generated by {qk } is regular if and only if
qn−1
lim = 0. (2.31)
n→∞ Qn
As to this notion and result, we refer the reader [[2, 4, 5] and references therein].
We note that in the particular case when qk = 1 for all k, these tn ( f, x) are the first
arithmetic or (C, 1)-means. More generally, when

β β +k β
qk = A k = for k ≥ 1 and q0 = A0 = 1,
k
where β = −1, −2, . . . , the tn ( f, x) are the (C, β)-means for series (2.29). The
representation
1
tn ( f, x) = f (x +̇t)L n (t) dt (2.32)
0
plays a central role in the sequel, where
1
n
L n (t) = qn−k Dk (t), n ≥ 1, (2.33)
Q n k=1
is so-called Nörlund kernel.

Theorem 2.24 Let f ∈ L p , 1 ≤ p ≤ ∞, let n = 2m + k, 1 ≤ k ≤ 2m , m ≥ 1, and
let {qk : k ≥ 0} be a sequence of nonnegative numbers such that
n γ −1 γ
n−1
γ q = O(1) for some 1 < γ ≤ 2. (2.34)

Q n k=0 k
If {qk } is nondecreasing, then
5 j
m−1
tn ( f ) − f p ≤ 2 qn−2 j w p ( f, 2− j ) + O{w p ( f, 2−m )}, (2.35)
2Q n j=0
while if {qk } is nonincreasing, then
5
m−1
tn ( f ) − f p ≤ (Q n−2 j +1 − Q n−2 j+1 +1 )w p ( f, 2− j ) + O{w p ( f, 2−m ))}.
2Q n j=0
(2.36)
Clearly, condition (2.34) implies (2.30) and (2.31).
We note that if {qk } is nondecreasing, in sign qk ↑, then
nqn−1
= O(1) (2.37)
Qn
is a sufficient condition for (2.34). In particular, (2.37) is satisfied if
qk k β or (logk)β for some β > 0.
Here and in sequel, qk rk means that the two sequences {qk } and {rk } have the
same order of magnitude; that is, there exist two positive constants C1 and C2 such
that
C1rk ≤ qk ≤ C2 rk for all k large enough.
If {qk } is nonincreasing, in sign qk ↓, then condition (2.34) is satisfied if, for example,
(i) qk k −β for some 0 < β < 1, or

(ii) qk (logk)−β for some 0 < β. (2.38)
Namely, it is enough to choose 1 < γ < min(2, β −1 ) in case (i), and γ = 2 in case
(ii).
Theorem 2.25 Let {qk : k ≥ 0} be a sequence of nonnegative numbers such that in
case qk ↑ condition (2.37) is satisfied, while in case qk ↓ condition (2.38) is satisfied.
If f ∈ Li p(α, p) for some α > 0 and 1 ≤ p ≤ ∞, then
⎧
⎨ O(n −α ) if 0 < α < 1,
tn ( f ) − f p = O(n −1log n) if α = 1, (2.39)
⎩
O(n −1 ) if α > 1.
Now we make a historical comments. The rate of convergence of (C, β)-means for
functions in Li p(α, p) was first studied by Yano [13] in the cases when 0 < α < 1,
β > α and 1 ≤ p ≤ ∞, and then by Jastrebova [14] in the case when α = β = 1 and
p = ∞. Later on, Skvorkov [10] showed that these estimates hold for 0 < β ≤ α as
well, and also studied the cases when α = 1, β > 0 and 1 ≤ p ≤ ∞. In their proofs,
the above authors rely heavily on the specific properties of the binomial coefficients
β
Ak .
In 1963, Watari [11] proved that a function f ∈ L p belongs to Li p(α, p) for some
α > 0 and 1 ≤ p ≤ ∞ if and only if
E n ( f, L p ) = O(n −α ).
Thus for 0 < α < 1 the rate of approximation to functions f in Li p(α, p) by tn ( f ) is

as good as the best approximation. For complete references, we refer to [1, 2, 4–6].
Auxiliary Results
Yano proved that the Walsh–Fejér kernel
1
n n−1
k
K n (t) = Dk (t) = (1 − )wk (t), n ≥ 1,
n k=1 k=0
n
is a quasi-positive, and K 2m (t) is even positive. These facts are formulated in the
following.
Lemma 2.5 Let m ≥ 0 and n ≥ 1; then K 2m (t) ≥ 0 for all t ∈ I ,
1 1
|K n (t)| dt ≤ 2 and K 2m (t) dt = 1.
0 0
A Sidon-type inequality proved by Schipp and Moricz implies that the Nörlund kernel
L n (t) is also quasi-positive. More exactly, C = [O(1)]1/γ , 2γ /(γ − 1) in the next
lemma, where O(1) is from (2.34).
Lemma 2.6 If condition (2.34) is satisfied, then there exists a constant C such that
1
|L n (t)| dt ≤ C, n ≥ 1.
0
Now, we give a specific representation of L n (t), interesting in itself.

Lemma 2.7 Let n = 2m + k, 1 ≤ k ≤ 2m , and m ≥ 1; then

m−1 2
j
−1
Q n L n (t) = − r j (t)w2 j −1 (t) i(qn−2 j+1 +i − qn−2 j+1 +i+1 )K i (t)
j=0 i=1

m−1
− r j (t)w2 j −1 (t)2 j qn−2 j K 2 j (t)
j=0

m−1
+ (Q n−2 j +1 − Q n−2 j+1 +1 )D2 j+1 (t)
j=0
+Q k+1 D2m (t) + Q k rm (t)L k (t). (2.40)
Proof We have
2
m
−1 2
m
+k
Q n L n (t) = qn−i Di (t) + qn−2m D2m (t) + qn−i Di (t)
i=1 i=2m +1
2
m−1 j
−1
= qn−2 j −i (D2 j +i (t) − D2 j+1 (t))
j=0 i=0
⎛ ⎞

m−1 2
j
−1
+ ⎝ qn−2 j −i ⎠ D2 j+1 (t)
j=0 i=0

k
+qn−2m D2m (t) + qn−2m −i D2m +i (t). (2.41)
i=1
As is well known(see,e.g., [2, p. 46]),
D2m +i (t) = D2m (t) + rm (t)Di (t), 1 ≤ i ≤ 2m . (2.42)
Furthermore, by (1.1) of [5], it is not difficult to see that
w2 j−1−l (t) = w22 j −1 (t)wl (t), 0 ≤ l < 2 j .
Hence, we deduce that
2
j
−1 −i−1
2 j
D2 j+1 (t) − D2 j +i (t) = r j (t) wl (t) = r j (t) w2 j −1−l (t)
l=i l=0
= r j (t)w2 j −1 (t)D2 j −i (t), 0 ≤ i < 2 j . (2.43)
Substituting (2.42) and (2.43) into (2.41) yields

m−1 2
j
−1
Q n L n (t) = − r j (t)w2 j −1 (t) qn−2 j +1 D2 j −i (t)
j=0 i=0

m−1
+ (Q n−2 j +1 − Q n−2 j+1 +1 )D2 j+1 (t)
j=0
+Q k+1 D2m (t) + Q k rm (t)L k (t). (2.44)
Performing a summation by part gives
2
j
−1 2
j
−1
qn−2 j −1 (t)D2 j −i (t) = i K i (t)(qn−2 j+1 +i − qn−2 j+1 +i+1 ) + 2 j K 2 j (t)qn−2 j .
i=0 i=1
Substituting this into (2.44) results in (2.40).

Lemma 2.8 If g ∈ P2m , f ∈ L p , where m ≥ 0 and 1 ≤ p ≤ ∞, then for 1 ≤ p <
∞
p 1/ p
1 1

rm (t)g(t)[ f (x +̇t) − f (x)] dt d x

0 0
1
−1 −m
≤ 2 w p ( f, 2 ) |g(t)| dt, (2.45)
0
while for p = ∞
1
sup | rm (t)g(t)[ f (x +̇t) − f (x)] dt : x ∈ I
0
1
−1 −m
≤ 2 w∞ ( f, 2 ) |g(t)| dt. (2.46)
0
Proof Since g ∈ P2m , it takes a constant value, say gm (k) on each dyadic interval
Im (k), where 0 ≤ k < 2m . We observe that if t ∈ Im (k) then t + 2−m−1 ∈ Im (k).
We will prove (2.45). By Minkowski’s inequality in the usual and in the generalized
form, we obtain that
p 1/ p
1 1
rm (t)g(t)[ f (x +̇t) − f (x)] dt d x

0 0
m −1 p 1/ p
1 2
−m−1
= gm (k) [ f (x +̇t) − f (x +̇t +̇2 )] dt d x
0 k=0 Im+1 (2k)
2
m
−1 1 p 1/ p
−m−1
≤ |gm (k)| | f (x +̇t) − f (x +̇t +̇2 )| dt dx
k=0 0 Im+1 (2k)
2
m
−1 1 1/ p
−m−1
≤ |gm (k)| | f (x +̇t) − f (x +̇t +̇2 )| d x
p
dt
k=0 0
2
m
−1
≤ |gm (k)|2−m−1 w p ( f, 2−m ).
k=0
This is equivalent to (2.45). Inequality to (2.46) can be proved analogously.

Proofs of Theorem 2.24 We carry out the proof of Theorem 2.21 for 1 ≤ p < ∞.
The proof for p = ∞ is similar and even simpler. By (2.32) and (2.40) and the usual
Minkowski inequality, we may write that
p 1/ p
1 1
Q n tn ( f ) − f p =
Q n L n (t)[ f (x +̇t) − f (x)] dt d x

0 0
p 1/ p

m−1 1 1
≤ r j (t)g j (t)[ f (x +̇t) − f (x)] dt d x

j=0 0 0
p 1/ p

m−1 1 1
+ r j (t)g j (t)[ f (x +̇t) − f (x)] dt d x

j=0 0 0

m−1
+ (Q n−2 j +1 − Q n−2 j+1 +1 )
j=0
p 1/ p
1 1
×
D2 j+1 (t)[ f (x +̇t) − f (x)] dt d x

0 0
p 1/ p
1 1
+Q k+1
D2m (t)[ f (x +̇t) − f (x)] dt d x

0 0
p 1/ p
1 1
+Q k rm (t)L k (t)[ f (x +̇t) − f (x)] dt d x

0 0
A1n + A2n + A3n + A4n + A5n , (2.47)
say, where
2
j
−1
g j (t) = w2 j −1 (t) i(qn−2 j+1 +i − qn−2 j+1 +i+1 )K i (t),
i=1
h j (t) = w2 j −1 (t)2 j qn−2 j qn−2 j K 2 j (t), 0 ≤ j < m.
Applying Lemma 2.5, in the case when qk ↑ we get that
1 2
j
−1
|g j (t)| dt ≤ i|qn−2 j+1 +i − qn−2 j+1 +i+1 |
0 i=1

j
2
= 2(2 qn−2 j − j
qn−2 j+1 +i ) ≤ 2 j+1 qn−2 j ,
i=1
while in the case when qk ↓

⎛ ⎞
1
2 j
|g j (t)| dt ≤ 2 ⎝ qn−2 j+1 +i − 2 j qn−2 j ⎠

0 i=1
≤ 2(Q n−2 j+1 +i − Q n−2 j+1 +i+1 ).
Thus by Lemma 2.8 in the case qk ↑

m−1
A1n ≤ 2 j qn−2 j w p ( f, 2− j ), (2.48)
j=0
while in the case qk ↓

m−1
A1n ≤ (Q n−2 j+1 +i − Q n−2 j+1 +i+1 )w p ( f, 2− j ). (2.49)
j=0
By virtue of Lemmas 2.5 and 2.8 again, we again that

m−1
A2n ≤ 2−1 2 j qn−2 j w p ( f − 2− j ). (2.50)
j=0
Obviously, in the case qk ↓
2 j qn−2 j ≤ Q n−2 j+1 +i − Q n−2 j+1 +i+1 . (2.51)
Since
2m if t ∈ [0, 2−m ),
D2m (t) =
0 if t ∈ [2−m , 1),
(see,e.g., [2, p. 7]), by the generalized Minkowski inequality, we find that

m−1
A3n ≤ (Q n−2 j+1 +i − Q n−2 j+1 +i+1 )
j=0
1 1 1/ p
× D2 j+1 | f (x + t) − f (x)| p d x dt
0 0

m−1
≤ (Q n−2 j+1 +i − Q n−2 j+1 +i+1 )w p ( f, 2− j ), (2.52)
j=0
A4n ≤ Q k+1 w( f, 2−m ). (2.53)
Clearly, in the case qk ↑

Q n−2 j+1 +i − Q n−2 j+1 +i+1 ≤ 2 j qn−2 j . (2.54)
Finally, by Lemmas 2.6 and 2.8, in a similar way to the above, we deduce that
1
−1 −m
A5n ≤ 2 Q k w p ( f, 2 ) |L k (t)| ≤ C Q n w p ( f, 2−m ). (2.55)
0
Combining (2.57)–(2.55) yields (2.53) in the case qk ↑ and (2.54) in the case qk ↓.
Proof of Theorem 2.25 Case(a).qk ↑. We have
n − 2 j ≥ 2m−1 for 0 ≤ j ≤ m − 1.
Consequently, for such j s
2 j qn−2 j (n − 2 j + 1)qn−2 j Q n−2 j +1 2j

= ≤ C2 j−m+1 ,
Qn Q n−2 j +1 Qn n − 2 j + 1
where C equals O(1) from (2.37). Since f ∈ Li p(α, p), from (2.35), it follows that
O(1) j
m−1
tn ( f ) − f p = 2 qn−2 j 2− jα + O(2−mα )
Q n j=0

m
= O(1)2−m 2 j−α
j=0
⎧
⎨ O(2−mα ) if 0 < α < 1,
= O(m2−m ) if α = 1,
⎩
O(2−m ) if α > 1.
This is equivalent to (2.39).

Case (b).qk ↓. For example, we consider case (i) in (2.38). Then Q n n 1−β . This
time we have
n − 2 j+1 ≥ 2m−1 for 0 ≤ j ≤ m − 2.
Since f ∈ Li p(α, p), from (2.36) it follows that
5 j
m−2
tn ( f ) − f p ≤ 2 qn−2 j +1 w p ( f, 2− j )
2Q n j=0
5
+ w p ( f, 2−m ) + O{w p ( f, 2−m )}
2
O(1) j
m−2
= 2 qn−2 j +1 2− jα + O(2−mα )
Q n j=0
O(1)2−mβ j (1−α)
m−2
= 2 + O(2−mα )
n 1−β j=0
⎧ −1 m(1−α)
⎨ O(n 2 ) if 0 < α < 1,
= O(n −1 m) if α = 1,
⎩
O(n −1 ) if α > 1.
Clearly, this is equivalent to (2.39).

Case (ii) in (2.38) can be proved analogously.
Concluding Remarks and Problems
(A) We have seen that (2.34) is satisfied when qk = (k + 1)β for some β > −1, and
Theorems 2.24 and 2.25 apply. If qk increases faster than a positive power of k,
the relation (2.34) is no longer true in general. But the case, for example, when qk
grows exponentially is not interesting, since the condition (2.31) of regularity is not
satisfied. On the other hand, the case when β = −1 is of special interest.
Problem 1. Find substitutes of (2.36) and (2.39) when qk = (k + 1)−1 . In this case,
the tn ( f ) are called the logarithmic means for series (2.29).
(B) It is also of interest that Theorems 2.21 and 2.22 remain valid when
qk k β φ(k), (2.56)
where β > −1 and φ(k) is a positive and monotone(nondecreasing or nonincreasing)

function in k, slowly varying in the sense that
φ(2k)
lim = 1.
k→∞ φ(k)
It is not difficult to check that in this case
Q n n 1+β φ(n).
(C) Now, we turn to the so-called saturation problem concerning the Nörlund means
tn ( f ). We begin with the observation that the rate of approximation by tn ( f ) to
functions in Li p(α, p) cannot be improved too much as α increases beyond 1. Indeed,
the following is true.
Theorem 2.26 If {qk } is a sequence of nonnegative numbers such that
lim in f q2m −1 > 0, (2.57)

m→∞
and if for some f ∈ L p , 1 ≤ p ≤ ∞,
t2m ( f ) − f p = O(Q −1
2m ) as m → ∞, (2.58)
then f is constant. We note that condition (2.34) is correctly satisfied if qk ↑ or qk ↓

and lim qk > 0.
Proof Since by the definition
E 2m ( f, L p ) ≤ t2m ( f ) − f p ,
and by a theorem of Watari [1963], Moricz and Siddiqi [5] obtained
s2m ( f ) − f p ≤ 2E 2m ( f, L p ),
and it follows from (2.58) that
s2m ( f ) − f p = O(Q −1
2m ) as m → ∞. (2.59)
A simple computation gives that
2
m
−1
Q 2m {s2m ( f, x) − t2m ( f, x)} = (Q 2m − Q 2m −k )ak wk (x).
k=1
Now (2.58) and (2.59) imply that
2
m
−1
lim (Q 2m − Q 2m −k )ak wk (x) p = 0.
m→∞
k=1
Since .1 ≤ . p , for any p ≥ 1 it follows that
lim |(Q 2m − Q 2m − j )a j |
m→∞
2m −1
1

= lim w j (x) (Q 2m − Q 2m −k )ak wk (x) d x
m→∞ 0
k=1
2
m
−1
≤ lim (Q 2m − Q 2m −k )ak wk (x)1 = 0.
m→∞
k=1
Hence, by (2.57), we conclude that a j = 0 for all j ≥ 1. Therefore, f = a0 is con-

stant.
In particular case when qk = 1 for all k, the tn ( f ) are the (C, 1)-means for series
(2.29) defined by
1
n
σn ( f, x) = sk ( f, x), n ≥ 1,
n k=1
and Theorem 2.23 is known (see, e.g., [2, p. 191]). It says that if for some f ∈ L p ,
1 ≤ p ≤ ∞,
σ2m ( f ) − f p = O(2−m ) as m → ∞,
then f is necessarily constant.
Problem 2. How can one characterize those functions f ∈ L p such that
σn ( f ) − f p = O(n −1 )for some 1 ≤ p ≤ ∞? (2.60)
We conjecture that (2.60) holds if and only if

∞
∞

2m w p ( f, 2−m ) < ∞, or equivalently w p (k −1 ) < ∞.
m=0 k=1
The “if” part can be proved in the same manner as in the case when w p ( f, δ) = O(δ α )
for some α > 1 (cf. [2, p. 190]). The proof(or disproof) of the “only if” part is a
problem.
2.3.3 Approximation by Nörlund Means in Dyadic

Homogeneous Banach Spaces and Hardy Spaces
We have introduced Dyadic Homogeneous spaces in Sect. 2.2.3. In the previous

section (Sect. 2.3.2), we have discussed the approximation properties of Walsh–
Nörlund means in L p spaces. As indicated, this study embraced earlier results
in the area by Yano, Jastrebova, and Skvorcov on the rate of approximation by
Césaro means. The main objective of this section is to present extension of results of
Sect. 2.3.2 which have been studied by Fridli, Manchanda and Siddiqi [4]. We present
here approximation by Walsh–Nörlund means in dyadic homogeneous Banach
spaces, and to dyadic Hardy spaces H p ; p < 1. Let us list some basic but useful
properties of homogeneous Banach spaces. The first one is about the norm of con-
volution operators. If h ∈ L 1 and f ∈ X , then
h ∗ f X ≤ h1 f X , (2.61)
1
where ∗ stands for dyadic convolution (h ∗ f )(x) = 0 h(x +̇t) f (t) dt.
The modulus of continuity in X is defined by
w( f, δ, X ) = sup μ<δ f − τ f X ( f ∈ X, 0 < δ ≤ 1).
Since
D2n = 2n χ [0, 2−n ) (n ∈ N), (2.62)
the 2n th partial sums plays a special role in Walsh analysis. For instance, (2.62)
implies S2n f X ≤ f X . Moreover, they enjoy the following approximation prop-
erties ([2]):
1
w( f, 2−n , X ) ≤ f − S2n f X ≤ w( f, 2−n , X ),
2
1
w( f, 2−n , X ) ≤ E 2n ( f, X ) ≤ w( f, 2−n , X ) ( f ∈ X, n ∈ N). (2.63)
2
E n ( f, X ) stands for the best approximation of f in X by Walsh polynomials of
order not greater than n. We note that for the L p case and for dyadically continuous
functions the inequalities were originally proved by Watari in 1963.
Let the σ −algebra generated by the dyadic intervals of length 2−n be denoted by
An . A sequence f = ( f n ) of integrable functions is a dyadic martingale if
(I) f n is An measurable,
(II) S2n f n+1 = f n (n ∈ N).
It follows from the definition of martingales that fˆn (k) = fˆm (k) for any k < 2n , 2m ..
Therefore, the definition fˆ(k) = fˆn (k)(k ∈ N, k < 2n ) makes sense. This means that
the concept of Walsh–Fourier partial sums and Walsh–Fourier series can be extended
to martingales.
In order to define the dyadic Hardy spaces H p (0 < p < ∞), we introduce the
maximal function of martingale:
f ∗ = sup n∈N | f n |.
Then H p is the collection of dyadic martingales f for which f ∗ is in L p and the

norm is defined as
1 1/ p
∗ p
f Hp = (f ) (0 < p < ∞).
0
We note that for p > 1 the Hardy space H p coincides with the Lebesgue space L p ,
H 1 is a homogeneous Banach spaces which is a proper subspace of L 1 , and for p < 1
the H p spaces are homogeneous Banach spaces.
In several cases, it is useful to have other characterizations of the H p norm. One
of them relies on the quadratic variation of martingales.
For a martingale f , let the
quadratic variation of f be defined as Q f = ( ∞ n=0 |dn f | )
2 1/2
, where dn f = f n+1 −
f n . Then f ∈ H p if and only if Q f ∈ L p (0 < p < ∞), and f H p ≈ Q f L p (see
[7]).
The concept of modulus of continuity and best approximation in H p can be defined
in the same way as in homogeneous Banach spaces. Moreover, by using quadratic
variation, it is easy to see that for any f ∈ H p , and n ∈ N we have
f − S2n f H p ≈ Q( f − S2n f ) L p = min Q( f − p) L p ≈ E 2n ( f, H p ).

p∈P 2n
(2.64)
The Lebesgue measure of a measurable set A will be denoted by |A|, and throughout
the paper C will stand for an absolute positive constant not necessarily the same in
different occurrences.
Results We will Norölund means when the generating sequence (qk ) is nonnegative,
and monotone. The Norölund means of the Walsh–Fourier series of a martingale f
is defined by
1
n
tn f = qn−k f (n ∈ N),
Q n k=1

where Q n = n−1 k=0 qk . We may suppose that q0 > 0. This implies Q n > 0, n ∈ N,
as well.
The following theorem is a generalization of the result of Moricz and Siddiqi in
[5] on the approximation properties of Nörlund means in L p spaces
Theorem 2.27 Let X be a homogeneous Banach space, f ∈ X , and n, m ∈ N, with
2m ≤ n < 2m+1 .
(i) If (qk ) is a monotonically increasing sequence of positive real numbers, then
⎛ ⎛ ⎞ ⎞
m−1 2l+1 −1
5 1 ⎝ ⎝
tn f − f X ≤ qn− j ⎠ ω( f, 2−l , X ) + qn− j ω( f, 2−m , X )⎠ .
2 Q n l=0 l j=2 m
j=2
(ii) If (qk ) is a monotonically decreasing sequence of nonnegative real numbers, then

⎛ ⎛ ⎞

m−1
2l+1 −1
5 1 ⎝ ⎝
tn f − f X ≤ qn− j ⎠ ω( f, 2−l , X )+
2 Q n l=0
j=2l
⎞

n
(n − 2 m
)q
qn− j (1 + log+ n )ω( f, 2−m , X )⎠ .
n− j
+C
m s=2 m qn−s
j=2
We note that the statements of Theorem 2.27 can be stated in several equivalent
forms. For instance, as a consequence of (2.63), we receive an equivalent form of
the3 above theorem if ω( f, 2−l , X ) is replaced by E 2 l( f, X )(l ∈ N).
Another version relies on a property of the modulus itself. However, the dyadic
modulus of continuity has properties different from the classical one. It is easy to see
that similarly to the classical case the inequality ω( f, 2−l , X ) ≤ 2ω( f, 1/j, X ) (2l ≤
j ≤ 2l+1 ). Then, for example, the right side of statement (i) is equivalent to
5 1
n
qn− j ω( f, 1/j, X ).
2 Q n j=1
Remark 2.6 As it is easy to see that in the proof, in case the Walsh system will be
a basis in X the statement in (i) of Theorem 2.27 holds without the monotonicity
of (qk ). So the statements of Theorem 2.27 are of interest only if the Walsh system
is not a Schauder basis X . Examples of homogenous Banach spaces in which the
Walsh system is not a basis include L 1 , H 1 , the dyadic VMO space, the space of
dyadically continuous functions, and nonreflexive Orlicz spaces.
In our second theorem, we are concerned about the approximation properties of
Nörlund means in H p (0 < p < 1) spaces. It shows that, even though these spaces
are not homogeneous, the results received are similar to those in Theorem 2.27.
Theorem 2.28 Let (qk ) be a sequence of nonnegative real numbers.

(i) If (qk ) is monotonically increasing, and 0 < p < 1 then
⎛ ⎛ ⎞p

m−1
2l+1 −1
1 ⎝ ⎝ qn− j ⎠ E 2l ( f, H p ) +
p
tn f − f H p ≤ C p
Q n l=0
j=2l
⎛ ⎞p ⎞1/ p

n
⎝ qn− j ⎠ E 2m ( f, H p )⎠
p
j=2m
( f ∈ H p , n, m ∈ N, 2m ≤ n < 2m+1 ).
(ii) Suppose that (qk ) is monotonically decreasing and there exists r > 1 such that
n r −1 r
n−1
q = O(1) (n ∈ N). (2.65)
Q rn j=0 n
If p > r/(2r − 1) then

⎛ ⎛ ⎞p ⎞1/ p
m−1 2l+1 −1
1 ⎝ ⎝
qn− j ⎠ E 2l ( f, H p )⎠ + O(E 2m , H p ))
p p
tn f − f H p ≤ Cp
Q n l=0 l
j=2
( f ∈ H p , n, m ∈ N, 2m ≤ n < 2m+1 ).
Auxiliary results
In the proof of (ii) of Theorem 2.27, we will need a Sidon-type inequality.
Theorem A [4] Let n ∈ N, and e j ( j = 1, . . . , n) be real numbers. Then

n
n
|c j |

cj Dj ≤ C +

|c j | 1 + log −1 n . (2.66)

j=1 j=1
n s=1 |cs |
1
A multiplier theorem will be used in the proof of (ii) of Theorem 2.28. It shows
that certain Hörmander–Mihlin-type multipliers are bounded on H p . For any real
sequence, ϕ = (ϕk ) and martingale Tψ f is given by
φk f (k) = ϕk fˆ(k)

T (k ∈ N).
Let the difference sequence Δϕ be defined as Δϕk = ϕk − ϕk+1 (k ≥ 0). Then the
following theorem holds.
Theorem B [4] Let 1 < r ≤ 2, p > r/(2r − 1), and (ϕ j ) be a real sequence. If (ϕ j )
is bounded and satisfies
⎛ ⎞1/r

2l+1 −1
|ϕ j | ⎠
r
2l ⎝ ≤C (l ∈ N), (2.67)
2l
k=2l
then the Walsh multiplier operator Tϕ is bounded on H p .

Lemma 2.9 Let g ∈ P2m (m ∈ N) and f ∈ X . Then
1
τm g ∗ f X ≤ g1 ω( f, 2−m , X ).
2
Proof Multiplying g by τm results in shifting the spectrum by 2m . Hence, ( τm g)(k) =

/ {2m , . . . , 2m+1 − 1}. Since (τm
0 whenever k ∈ g ∗ f )(k) = (τm g)(k) fˆ(k), we have
that
τm g ∗ f = τm g ∗ (S2m+1 f − S2m f ).
Thus it follows from (2.61) that
τm g ∗ f X ≤ g1 S2m+1 f = S2m f X .
By the approximation properties of the S2n f (n ∈ N) partial sum noted in introduction,

we conclude
1
τm g ∗ f X ≤ g1 ω( f, 2−m , X ).
2
In the proofs of our Theorems 2.27 and 2.28, we will apply summation by parts. As a
result of it, Dirichlet kernels will be replaced by Fejér kernels. The nth Walsh–Fejér
kernel K n is defined as
1
n
Kn = Dk (n ∈ N).
n j=1
The L p norm (0<p<1) of linear combinations of Fejér kernels can be estimated as

follows.
Lemma 2.10 Let 1
2
< p < 1 and μ j ∈ R( j = 1, . . . , 2l − 1, l ∈ N). Then
2
l
−1 2
l
−1
|μ j K j | ≤ C p (2 ) l 1−1/ p
|μ j |.
j=1 j=1
Proof It is known [2] that
1 k
l−1 l−1
|K j | ≤ 2 (D2i + τ2−k D2i ) (2l−1 ≤ j ≤ 2l ).
2l k=0 i=k
Since the terms on the right side are positive, the estimate holds for indices j < 2l−1
as well. l−1 l−1 k l−1
Set X 1 = l−1k=0 2
k
i=k D2i and X 2 = k=0 2 i=k τ2−k D2i . Then it follows
from p > 1 that
⎛ ⎞p ⎛ ⎞p
2
l
−1 2
l
−1 p
⎝ 1
|μ j K j |⎠ ≤ ⎝ |μ j |⎠
p p
(X 1 + X 2 ).
j=1 j=1
2l
1 p
For the proof of the lemma, we only need to show that 0 X i ≤ C p (2l )2 p−1 . By
(2.62) and change the summation, we have
l−1 p
p

i
l−1
X1 = (2 )χ[0,2−i )
i
2 k
≤ 2p (22i ) p χ[0,2−i ) .
i=0 k=0 i=0
Hence, p > 1
2
implies
1
l−1
p
X1 ≤ 2p (2i )2 p−1 ≤ C p (2l )2 p−1 .
0 i=0
We show that the same estimate holds for X 2 . Let us start by noting that
l−1 m

l−1
l−1
τ2−k D2i = 2 χ[2−k ,2−k +2−l+1 ) +
i
2i χ[2−k +2−m−1 ,2−k +2−m )
i=k i=k m=k i=k

l−2
≤ 2l χ[2−k ,2−k +2−l+1 ) + 2m+1 χ[2−k +2−m−1 ,2−k +2−m ) .
m=k
This, in particular, means that the support of the various terms with respect to k in
X 2 are pairwise disjoint. Therefore, we have
l−1 p
1
p

l−1
l−1
l−1
X2 = (2 )
k p
τ2 D2
−k i ≤2 p
(2 )
k p
(2m ) p−1 .
0 k=0 i=k k=0 m=k
1 p
Then 1
2
< p < 1 implies 0 X 2 ≤ C p (2l )2 p−1 .
H p is not homogeneous Banach space for 0 < p < 1, and the norm of convolution
operators cannot be estimated by (2.61) within H p . Let Pn be the set of linear
combinations of wk s, k = 0, . . . , n − 1. In case of Walsh polynomials, the following
lemma will serve as a substitute for (2.61) in H p .
Lemma 2.11 Let h, g ∈ P2l (i ∈ N). Then h ∗ g p ≤ (2l )1/ p−1 h p g p .
Proof Set xk = k2−l , and t j = j2−l (k, j = 0, . . . , 2l − 1). Since h and g are Walsh
polynomial of order less than 2l , they are constants on dyadic intervals of length 2−l .
Consequently,
p
p 2
l
−1 2
l
−1
1 1 1
h ∗ g pp = h(x + t)g(t)dt d x = l 1 h(x + t )g(t )
l
2 k=0 2 k=0
k j j
0 0
2 −1 2l −1
1 1 p
l
= l |h(xk + t j )| p |g(t j )| p
2 k=0 2l j=0
p+1 2
l
−1 2
l
−1
1
= |g(t j )| p
|h(xk + t j )| p = (2l )1− p g pp h pp .
2l j=0 k=0
Proofs
Proof of Theorem 2.27. Let n = 2m + k(n, m ∈ N, 0 ≤ k < 2m ), and f ε X In case

(qk ) is increasing we will use the following decomposition:
1
n
f − tn f = f − qn− j S j f
Q n j=1
⎛ ⎞
m−1 2l+1 −1
1 ⎝ n
= qn− j ( f − S j f ) + qn− j ( f − S j f )⎠
Q n l=0 l j=2 m
j=2
⎛ ⎞
m−1 2 −1
1 ⎝
l+1
n
= qn− j ( f − S2l f ) + qn− j ( f − S2m f )⎠ (2.68)
Q n l=0 l j=2 m
j=2
⎛ ⎞
m−1 2l+1 −1
1 ⎝ n
− qn− j (S j f − S2l f ) + qn− j (S j f − S2m f )⎠
Q n l=0 l j=2 +1
m
j=2
= A1 + A2 .
By (2.63), we have
⎛ ⎞
m−1 2l+1 −1
1 1 ⎝ n
A1 X ≤ qn− j ω( f, 2−l , X ) + qn− j ω( f, 2−m , X )⎠ .
2 Q n l=0 l j=2 m
j=2
Let us consider A2 . We can write S2l + j f − S2l f = (τl D j ) ∗ dl f ( j = 1, . . . , 2l , l ∈

N). Then for the first term in A2 , we have
⎛ ⎞

2l+1 −1 2
l
−1
qn− j (S j f − S2l f ) = ⎝τl qn−2l − j D j ⎠ ∗ dl f. (2.69)
j=2l +1 j=1
In the same way, we obtain for the second term that

⎛ ⎞
2
m
+k
k
qn− j (S j f − S2m f ) = ⎝τm qn−2m − j D j ⎠ ∗ dm f. (2.70)
j=2m +1 j=1
Let us use summation by parts to obtain
2
l
−1 2
l
−2
qn−2l − j D j = jqn −2l − j K j + (2l − 1)qn−2l+1 +1 K 2l −1 . (2.71)
j=1 j=1
Recall q j = q j − q j+1 , and K j is the jth ( j ∈ N) Walsh–Fejér kernel. Since

K 1 ≤ 2 (see [2]), we have
⎛l ⎞
2
−1 2−2
l

qn−2l − j D j ⎝ jqn −2l − j K j + (2l − 1)qn−2l+1 +1 ⎠
≤2
j=1 j=1
1
2
l
−1
=2 qn−2l − j . (2.72)
j=1
In a similar way, we have

⎛ ⎞

k k−1 k
q D ≤ 2 ⎝ jq + kq ⎠ = 2 qn−2m − j .
n−2 − j j
m n−2 − j
m 0
j=1 j=1 j=1
1
Then, it follows that (2.61)

⎛ ⎞
m−1 2l+1 −1
2 ⎝ n
A X ≤ qn− j dl f X + qn− j dm f X ⎠ .
Q n l=0 j=2 +1
l j=2 +1 m
Since dl f = S2l+1 ( f − S2l f ), we have by (2.61) and (2.63) that dl f X ≤ ω

( f, 2−l , X ). Consequently,
⎛ ⎞
m−1 2l+1 −1
2 ⎝ n
A X ≤ qn− j ω( f, 2−l , X ) + qn− j ω( f, 2−l , X )⎠ .
Q n l=0 j=2m +1
l j=2 +1
The proof of part (i) of the theorem is completed.

In case (qk ) is depressing, we slightly modify the decomposition in (2.68) by
using S2l+1 f instead of S2l f . Then, we will have the following three terms:
f − tn f
1
n
= f − qn− j S j f
Q n j=1
⎛ ⎞
m−1 2l+1 −1
1 ⎝ n
= qn− j ( f − S2l+1 f )⎠ + qn− j ( f − S2m f ) + (2.73)
Q n l=0 j=2m
j=2l
⎛ ⎞
m−1 2l+1 −1
1 ⎝ 1
n
+ qn− j (S2l+1 f − S j f )⎠ − qn− j (S j f − S2m f )
Q n l=0 l
Q n j=2m
j=2
= B1 + B2 + B3 . (2.74)
In the same way as for A1 , we have

⎛ ⎞
m−1 2l+1 −1
1 1 ⎝ n
B1 ≤ qn− j ω( f, 2−(l+1) , X ) + qn− j ω( f, 2−(m) , X )⎠ .
2 Q n l=0 l j=2 m
j=2
In order to estimate B2 X , let us consider the corresponding kernel

2l+1 −1
qn− j (D2l+1 − D j f ).
j=2l
By use summation of Walsh functions, it is easy to see that w2n −1 ws = w2n −1−s holds
for any 0 ≤ s < 2v , v ∈ N. Hence
D2l+1 − D j f = w2l+1 −1 − D2l+1 − j .
This implies

2l+1 −1
2l+1 −1
qn− j (S2l+1 f − S j f ) X = qn− j (w2l+1 −1 − D2l+1 − j ) ∗ dl f
j=2l j=2l

l
2
≤ qn−2l+1 + j D j 1 dl f X .
j=1
In the last sum, the coefficients qn−2l+1 + j decrease by j. Therefore, similarly to (2.72),
we can obtain
2l
2l
qn−2l+1 + j D j 1 ≤ 2 qn−2l+1 + j .
j=1 j=1
Consequently,
m−1 2 −1
2
l+1
B2 ≤ qn− j ω( f, 2−l , X ).

Q n l=0 lj=2
Recall that n = 2m + k, 0 ≤ k < 2m . Then, for the last term B3 , we can write

n
1 1 k

B3 X = q (S f − S f ) = (w qn−2m − j D j ) ∗ dm f
Qn Qn
n− j j 2 m 2 m
j=2m j=1
X X

1
k
≤ qn−2m − j D j

Q n j=1 dm f X .

1
At the point, we will use the Sidon-type inequality in Theorem A to obtain

1
n
(n − 2m )qn− j
B3 X = C qn− j 1 + log n ω( f, 2−m , X ).
Q n j=2m +1 q
s=2 +1 n−s
m
Summarizing the estimate for B1 X (i = 1, 2, 3), we conclude

⎛ ⎛ ⎞
m−1 2l+1 −1
1 ⎝ 5 ⎝
f − tn X ≤ qn− j ω( f, 2−l , X )⎠ +
Q n 2 l=0 l j=2
⎞
n
(n − 2m )qn− j
+ C qn− j 1 + log n ω( f, 2−m , X )⎠ .
m s=2 m +1 qn−s
j=2
Proof of Theorem 2.28 Let us first consider the case when (q j ) is decreasing. Since
the maximal function operator is sublinear and p < 1, we have a decomposition
similar to (2.68)
⎛⎛ ⎞∗ ⎞ p
1 n
⎝⎝ f − qn− j S j f ⎠ ⎠ ≤ A∗1, p + A∗2, p ,
Q n j=1
where
⎛ ⎛ ⎞p ⎛ ⎞p ⎞
m−1 2l+1 −1
1 ⎝ ⎝ n

A∗1, p qn− j ⎠ ( f − S2l f )∗ ) + ⎝ qn− j ⎠ ( f − S2m f )∗ ) ⎠ ,
p p
= p
Q n l=0 l j=2m
j=2
and
⎛ ⎛⎛ ⎞∗ ⎞ p

m−1
2l+1 −1
1 ⎝ ⎝⎝
A∗2, p = p qn− j (S j f − S2l f )⎠ ⎠ +
Q n l=0
j=2 +1
l
⎛⎛ ⎞∗ ⎞ p ⎞
n
+ ⎝⎝ qn− j (S j f − S2m f )⎠ ⎠ ⎠ .
j=2m +1
It follows from (2.64) that

⎛ ⎛ ⎞p ⎛ ⎞p ⎞
1
m−1
2l+1 −1
n
1 ⎝ ⎝
A∗1, p ≤ qn− j ⎠ E 2l ( f, H p ) + ⎝ qn− j ⎠ E 2m ( f, H p )⎠ .
p p
p
0 Q n l=0 j=2m
j=2l
In connection with A∗2, p , let us first note that

⎛ ⎞∗
−1 2l+1 −1

2l+1

⎝ ⎠
qn− j (S j f − S2l f ) = qn− j (S j f − S2l f ) .
j=2l +1 j=2l +1
Then by (2.69),(2.70), and (2.71), we obtain

⎛⎛ ⎞ p
2 −2
l
1
m−1

A∗2, p = p⎝ ⎝τl jqn−2l − j K j + τl (2l − 1)qn−2l+1 +1 K 2l −1 ⎠ ∗ dl f +
n l=0 j=1
⎛ ⎞ p⎞

k−1

+ ⎝τm jqn−2m − j K j + τm kq0 K k ⎠ ∗ dm f ⎠ .
j=1
It follows from (2.64) and Lemmas 2.9, 2.10 that

m−1
1
1
A∗2, p ≤ p (2l )1− p
0 Qn l=0
2
l
−2
× jqn−2l − j K j + (2l − 1)qn−2l+1 +1 K 2l −1 pp dl f pp +
j=1
⎞

k−1
+ (2m )1− p jqn−2m − j K j + τm kq0 K k pp dl f pp ⎠
j=1
⎛
⎛ ⎞p
m−1 2l −2
1 ⎝ ⎝
| jqn−2l − j | + (2l − 1)qn−2l+1 +1 ⎠ E 2l ( f, H p )+
p
≤ Cp p
Q n l=0 j=1
⎛ ⎞p ⎞

k−1
+ ⎝ j| jqn−2m − j | + kq0 ⎠ E 2m ( f, H p )⎠
p
j=1
⎛ ⎛ ⎞p ⎛ ⎞p ⎞

m−1
2l+1 −1
n
1 ⎝ ⎝ qn− j ⎠ E 2l ( f, H p ) + ⎝ qn− j ⎠ E 2m ( f, H p )⎠ ,
p p
= Cp p
Q n l=0 j=2m
j=2l
which is the desired estimate.

Finally, let us prove part (ii), that is, when (qk ) is decreasing and condition (2.65)
holds for the sequence. Then, similarly to (2.73), we have
⎛ ⎛ l+1 ⎞p
2 −1
1
m−1
p
( f − t n f )∗ ≤ p ⎝ ⎝ qn− j ⎠ ( f − S2l+1 f )∗
p
Q n l=0
j=2l
⎛ ⎞p ⎞

n

+ ⎝ qn− j ⎠ ( f − S2m f )∗ ⎠
p
(2.75)
j=2m
⎛ ⎞p

2l+1 −1
1
m−1
p
+ p
⎝ qn− j ⎠ (S2l+1 f − S j f )∗ +
Qn l=0 j=2l
⎛ ⎞p
1 ⎝
n
p
+ p qn− j ⎠ (S j f − S2m f )∗
Q n j=2m
∗ ∗ ∗
= B1, p + B2, p + B3, p . (2.76)
Following the reasoning used for Ai ∗ and for Bi (i = 1, 2) in the proof part (ii)
of Theorem (2.27), we receive that the proper estimates for B1∗ and B2∗ . Let us
∗
start the case of B3, p by noting the spectrum nj=2m qn− j (S j f − S2m f ) is part of
{2m , . . . , n − 1}. Recall that n = 2m + k. Set
k− j−1
1
qi , if s = 2m+1 ± j, j = 0, . . . , k − 1;
ϕs = Qn i=0
0, otherwise.
Then, it is easy to see that

n
qn−1
k−1
1
k− j−1
τm+1 (S j f − S2m f ) = qi fˆ(2m + j)w2m+1 + j
j=2m
Qn j=0
Q n i=0
= Tϕ (τm+1 (dm f )).
Hence
1 1
∗ Tϕ (τm+1 (dm f )) p = Tϕ (τm+1 (dm f )) p p .
B3, p = H
0 0
We will show that condition (2.67) in Theorem B holds for ϕ. Since

qk−i−1
, if u = 2m+1 ± i, i = 0, . . . , k − 1;
|uϕ | = Qn
0, otherwise,
we have that there are only two dyadic blocks for which the left side of (2.67) is not
0. For those two blocks, we obtain
⎛ ⎞1/r ⎛ ⎞1/r
j=2m+2 −1
| j ϕ|r ⎠ k−1
q r
n 1−1/r r
n−1
2m+1 ⎝ m+1 1−1/r ⎝ j ⎠
= (2 ) ≤ 2 q ,
2m+1 j=0
Q rn Q n j=0 j
j=2m+1
(2.77)
and
⎛ ⎞1/r ⎛ ⎞1/r
j=2m+1 −1
| j ϕ|r ⎠ k−1
q r 1−1/r
n−1
2m ⎝ = (2 m 1−1/r ⎝
)
j ⎠ ≤n q rj . (2.78)
j=2m
2m j=0
Q r
n Q n j=0
Then (2.78), (2.77), and our assumption (2.65) on the q j s implies that condition (2.67)
holds for ϕ (We note that in (2.65) we may suppose 1 < r ≤ 2). Consequently,
Tϕ (τm+1 (dm f )) H p ≤ C(τm+1 (dm f )) H p = Cdm f p

≤ C f − S2m f H p ≤ C p E 2m ( f, H p ).
This means that 1

∗ p
p ≤ C p E 2m ( f, H ).
p
B3,
0
This proves Theorem 2.28(ii).

2.4 Applications to Signal and Image Processing
A signal can be treated as a function of one variable while an image can be represented
by a function of two variables. A digital signal is nothing but a discrete function or a
sequence, and a digital image can be represented by a matrix. The problem of digital
signal processing is in general deals with the representation of signals by sequences
of numbers or symbols, and the processing by a digital computer of these numbers.
The goal of such processing is to find certain characteristic parameters of the signal,
or it may be to obtain a simpler form of handling the signal as in the case of digital
transmission. Walsh–Fourier series are ideally suited for this purpose. Communica-
tion, storing, improving, or refining images are challenging problems. Applications
of Walsh–Fourier analysis may provide superior results in many situations.
2.4.1 Image Representation and Transmission
Let f (t) be a periodic signal with period 1. f (t) can be written as f (t) =
∞
n=0 cn wn (t), where 1
cn = f (t)wn (t) dt
0
By Parseval’s formula for Walsh analysis, we have

1 ∞

f 2 (t) dt = cn2 .
0 n=0
In view of these results, the signal energy is distributed in the Walsh domain over the
Walsh coefficients. Therefore, instead of transmitting the signal in the spatial domain
(for example, time domain), we may encode it by its Walsh–Fourier coefficients and
transmit these coefficients instead. In case of rapid convergence of the Walsh series,
many higher order coefficients are of low magnitude and become possible to transmit
fewer coefficients without sacrificing a big loss in the total signal energy. A data
reduction in Walsh coefficients entails bandwidth reduction in the sequency Walsh
domain.
Let us consider images f (x, y) and g(x, y) in the spatial domain and a Walsh
function and a Walsh function of D = {(x, y)/0 ≤ x < P, 0 ≤ y < ∞} two vari-
ables.
In the Walsh or sequency domain, a representation of the image made possible by
the application of a double Walsh transform, that is,
∞ ∞
F(σ, α) = f (x, y)wσ (x)wα (y) d x d y. (2.79)
0 0
2.4 Applications to Signal and Image Processing 71
By inverse Walsh transform, we have

∞ ∞
f (x, y) = F(σ, α)wσ (x)wα (y) dσ dα. (2.80)
0 0
Analogue of Parseval’s formula for two-dimensional case is

∞ ∞ ∞ ∞
(| f (x, y)|)2 d x d y = (|F(σ, α)|2 ) dσ dα. (2.81)
0 0 0 0
This indicates conservation of energy under double Walsh transform.

Let f (x, y) be an image defined over a finite domain R = {(x, y)/0 ≤ x < a, 0 ≤
y < b} and
[ f (x, y)]2 d x d y < ∞. (2.82)
R
It is possible to expand f (x, y) in the form of double Walsh–Fourier series

∞
∞
f (x, y) = amn wm (x/a)wn (y/b), (2.83)
m=0 n=0
where x, y ∈ R and
a b
1
amn = f (x, y)wm (x/a)wn (y/b) d x d y. (2.84)
ab 0 0
Thus, it is possible to encode a continuous image by its corresponding Walsh–Fourier

coefficients.
The application of digital image transmission is more profitable than analogue or

analytic methods. There are three main problems associated with image representa-
tion:
(a) Image digitization and coding,
(b) Image enhancement and restoration, and
(c) Image segmentation and description.
Image Digitization and Coding This comprise conversion of continuous image
(Continuous function of two variables) into a discrete form, using double Walsh–
Fourier series. Coding the digitized picture is a succeeding scheme which aims at
efficient and a more simplified representation of the digitized data. Digitization often
leads to data compression. Once the image has been sampled and quantized within
a tolerable quantized error, the result is a discrete image made of say N × N = N 2
samples.
The sampling theorem is a very fundamental concept in signal processing with
many applications. It is a notion linking continuous and discrete signals. We present
below a Dyadic Sampling Theorem without proof, for details, we refer to Butzer and
Splettstöber in [2].
Walsh–Fourier Analysis x ∈ R+ = [0, ∞),
∞

x= xi 2−i , xi ∈ {0, 1}, N (x) ∈ Z = {0, ±1, ±2, . . .}
i=−N (x)
D+ = {x ∈ R+ | x = p/2q , p ∈ P = {0, 1, 2, . . .}, q ∈ Z } dyadiic rational numbers.
If x ∈ D+ , select finite expansion.

x∈ / D+ , unique representation.
ψ(y, x), x, y ∈ R+ -generalized Walsh functions, y-sequency.
r
J (x, r ) = 0 wx,s ds-Walsh–Dirichlet kernel.
r = 2n ,
n
2 0 ≤ x < 2−n , 1 x ∈ [2−n s, 2−n (s + 1)),
J (x, 2 ) =
n
, J (1, 2 x ⊕ s) =
n
0 other wise 0 other wise.
Theorem 2.29 Dyadic Sampling Theorem

If f and its Walsh–Fourier spectrum S j (w) belong to L 1 (R+ ),
d
f is continuous on R+ ,
f is continuous from right on D+ .
S f (w) = 0 for w ≥ 2n , n ∈ Z , Z—the set of integers, i.e.,
f is sequency limited, then
∞

f (x) = f (s/2n )J (1, 2n x ⊕ s), x ∈ R+ . (2.85)
s=0
(b) Image Enhancement and Restoration This process deals essentially with
improvement in blurred and noisy images. Our main objective is to minimize the
expression
[ f (x, y) − f γ (x, y)]2 d x d y, (2.86)
R
where f (x, y) and f γ (x, y) denote, respectively, the original and reconstructed
images. These functions can be represented by Walsh–Fourier series.
(c) Image Segmentation and Description Image segmentation is devoted to division
of given image into segments for measurements and studies on a particular segment,
or classification or the description of the image in terms of these segments or parts.
Different series with orthonormal systems have used to study this theme. However,
Walsh–Fourier series gives better results in many cases. Figure 2.1 presents a block
diagram of a basic transform image transmission system.
Fig. 2.1 Diagram of a basic image transmission system
In Fig. 2.1, the original is first sampled to get a two-dimensional array of elements,
also called pixel or pels. This is followed by two-dimensional Walsh transform.
Thresholding the transformed image data results in the rejection of those transform
coefficients that fall below the threshold. The remaining Walsh transform components
are then quantized and coded. At the receiver, an inverse process takes place. The
received data is first decoded and then inverse transformed to give two-dimensional
data from which an image is reconstructed.
The digital image is usually represented by an m × n matrix. Elements of such
an image are called pixels, pels, or simply points.
The digitized image can be represented by a two-dimensional discrete Walsh
transform, that is, an image point denoted by g(n, m) yields discrete Walsh transform
N −1
N −1
G(x, s) = g(n, m)wn (r )wm (s). (2.87)
n=0 m=0
The inverse two-dimensional Walsh is given by
N −1 N −1
1
g(n, m) = q(r, s)wn (r )wm (s).
N 2 r =0 s=0
Conservation of energy between the spatial and sequency domain exists via Parseval-
type formula
N −1
N −1 N −1 N −1
1 2
g (n, m) = 2
2
G (r, s). (2.88)
n=0 m=0
N n=0 m=0
2.4.2 Data Compression
Data compression is devoted to reduction of data size according to appropriate cri-

teria. In transform image transmission, data compression is applied to the transform
points. A decision criterion must be applied for discarding certain transformed points.
A popular criterion is thresholding. Hence, the transmitted data are specified by

f (n, m) | f (n, m)| ≥ threshold
G (n, m) =
0 | f (n, m)| < threshold.
According to this thresholding scheme, the normalized mean-squared error arising

in the reconstruction of the image from the compressed data is given by

m [ f (n, m) − G (n, m)]
2
r eemn = n
2 . (2.89)
n m f (n, m)
2.4.3 Quantization of Walsh Coefficients
The quantized image is operated on by a coder which converts quantized trans-

formed image samples into appropriate digital code words, one code word for each
transformed point; for more details, we refer to [8].
2.4.4 Signal Processing
Signal process deals with the representation of signals by sequence of numbers or

symbols and processing of these numbers. The goal may be to estimate characteristic
parameters of the signal, or it may be to yield a simpler form of treating a signal.
Signal processing via Walsh–Fourier series have been used to signals like speech
and electrocardiogram (ECG). For details, we refer to [8] and references therein.
2.4.5 ECG Analysis
An electrocardiogram (ECG) heart wave represents a graphic recording of the electri-

cal activity associated with the heartbeat. By applying electrodes to various positions
on the human body and connecting those electrodes to an electrocardiographic appa-
ratus (monitor or recorder), an ECG signal is obtained. By properly sampling an
ECG graph, a discrete signal is obtained.
Let f T = [ f 0 , f (1), . . . , f N −1 ] be such a signal vector, where T denotes the
vector or matrix transposition.
Using Walsh function system {wn (x)}, we have
N −1

F(r ) = Fr = f n wr (n), r = 0, 1, 2, . . . , N − 1. (2.90)
n=0
By inverse discrete Walsh transform, we obtain

N −1
1
f (n) = F(r )wr (n), n = 0, 1, 2, 3, . . . (2.91)
n r =0
For detailed ECG analysis using Walsh analysis, we cite [8, pp. 227–230]. It has
been noted that there are several advantages of using Walsh transform.
2.4.6 EEG Analysis
Brain is an extremely complex system. The cerebral cortex of the human brain
contains roughly 15−33 billion neurons, perhaps more, depending on gender and
age, linked with up to 10,000 synaptic connections each. Each cubic millimeter of
cerebral cortex contains roughly one billion synapses. These neurons communicate
with one another by means of long protoplasmic fibers called axons, which carry
trains of signal pulses called action potentials to distant parts of the brain or body and
target them to specific recipient cells. Methods of observation such as EEG recording
and functional brain imaging tell us that brain operations are highly organized, while
single-unit recording can resolve the activity of single neurons, but how individual
cells give rise to complex operations is unknown.
Electroencephalography (EEG) is the recording of electrical activity along the
scalp produced by the firing of neurons within the brain. In clinical contexts, EEG
refers to the recording of the brain’s spontaneous electrical activity over a short
period of time, usually 20–40 mins, as recorded from multiple electrodes placed
on the scalp. In neurology, the main diagnostic application of EEG is in the case of
epilepsy, as epileptic activity can create clear abnormalities on a standard EEG study.
A secondary clinical use of EEG is in the diagnosis of coma, encephalopathies, and
brain death. EEG used to be a first-line method for the diagnosis of tumors, stroke, and
other focal brain disorders, but this use has decreased with the advent of anatomical
imaging techniques such as MRI and CT.
A routine clinical EEG recording typically lasts 20–30 mins (plus preparation
time) and usually involves recording from scalp electrodes. Routine EEG is typically
used in the following clinical circumstances:
to distinguish epileptic seizures from other types of spells, such as psychogenic
seizures, syncope (fainting), sub-cortical movement disorders, and migraine vari-
ants;
to differentiate “organic” encephalopathy or delirium from primary psychiatric
syndromes such as catatonia;
to serve as an adjunct test of brain death;
to prognosticate, in certain instances, in patients with coma; and
to determine whether to wean anti-epileptic medications. Both effects are inde-
pendent and additive.
Walsh–Fourier methods applied in Sect. 2.4.5 could be applied for EEG analysis.
For an interesting application of Walsh–Fourier analysis to EEG, we cite references
62–65 in [8].
2.4.7 Speech Processing
Man communicates through speech, and the theme of transmission of speech from a
transmitter to a listener has been investigated using orthonormal systems including
Walsh systems [8]. The use of Walsh functions to serve as a basis for the transform
processing aspect yields several advantages, for example, bandwidth reduction for
transmitted speech, efficient speech synthesis, and automatic speech or word recog-
nition. For applying the Walsh transform, the number of speech in a window is often
of the form N = 2n , for n positive integer. This yields fast computational algorithms.
If a signal is sampled at a frequency fs, then considering N sampled values for a sam-
pling window, the sampling interval is T = 1fs , and the analysis window has a time
duration of N T . The frequency resolution of this window is given by N1T .
Out of the N block of transformed samples, we may select the most significant M
Walsh components of speech synthesis. The fundamental theorem of speech synthesis
is to maximize the efficiency of voice transmission while preserving reasonable
distortion levels for voice characteristics . Now let f (t) be a speech signal. Sampling
the signal properly, and choosing a block of N samples, we got a set of sampled
values, { f (nT )/n = 0, 1, 2, . . . , N − 1}, where T signifies the sampling interval.
The discrete Walsh transform of this block of speech samples can be specified by
N −1

F(k) = f (n)wn (k). (2.92)
n=0
As in the image transmission problem, one can encode the speech by its Walsh
transform coefficients and transmit these coefficients instead of the original speech
samples.
2.4.8 Pattern Recognition
Pattern space P comprises a large amount of data (digital or analogue signals). One
may be interested in finding special features and the pattern and combining correlated
data (redundancy) to obtain bandwidth reduction. The resulting data gives a feature
space F. Significant information of F is used to form a classification space C, in
which the original pattern will have been properly indicated. The main objective of
applying discrete Walsh transform in pattern recognition is to reduce computation
time for the transformed data from the pattern space P. A more general approach
2.5 Exercises 77
in pattern recognition is to carry out a discrete Walsh transform of the pattern and
its model, and to cross-correlate the two transformed sets of values to determine the
degree of recognition, rather than to attempt cross-correlation of the original signal.
Use of discrete Walsh Transform is advantageous for computation. For a detailed
discussion on pattern recognition, we cite Kamath [9].
2.5 Exercises

2.1 Let f be continuous on [0,1] and fˆ(k) = o k(log1 k)α as k → ∞ for some α > 1,
then show that f is constant. Is this result true for α = 1?
2.2 How can one characterize those functions f ∈ L p such that
σn ( f ) − f p = o(n −1 )
for some 1 ≤ p ≤ ∞?
2.3 Let f ∈ L p , and that
2
n
−1
sup w(2) ( f, I (k, n)) < ∞.
n∈N k=0
Prove that fˆ(k) = o( k1 ) as k → ∞.

2.4 Prove that Sn ( f ) → f uniformly as n → ∞ for f ∈ C W which is of bounded
variation.
2.5 Let f (x) = χ[0, 13 ] (x). Show that Sn f ( 13 ) not converge as n → ∞.
2.6 Let f ∈ C W . Then show that σn f → f uniformly as n → ∞, where σn f is
arithmetic mean of Walsh–Fourier series of f .
2.7 Write an essay on application of Walsh analysis to ECG and EEG analyses.
References

2. Schipp, F., Wade, W.R., & Simon, P. (1990) Walsh series. Bristol: Adam Hilger
Acta Scientiarum Mathematicarum (Szeged), 74, 593–608.
5. Moricz, F., & Siddiqi, A. H. (1992). Approximation by Norlund means of Walsh-Fourier.
Journal of Approximation Theory, 70(3–6), 375–389.
6. Golubov, B. I., Efimov, A. V., & Skvortsov, V. A. (2008). Walsh series and transforms (English
Translation of 1st ed.). Moscow: Urss; Dordrecht: Kluwer (1991).
7. Malozemov, V. N., & Masharskii, S. M. (2002). Generalized wavelet bases related to the
Discrete Vilenkin-Chrestenson transform. St. Petersburg Mathematical Journal, 13, 75–106.
8. Maqusi, M. (1981). Applied Walsh analysis. London: Heyden.

9. Kamath, C. (2009). Scientific data mining: A practical perspective. PhiladelPhia: SIAM.
10. Skvorcov, V. A. (1981). Certain estimates of approximation of function by Cesaro means of
Walsh-Fourier series. Matematicheskie Zametki, 539–547. [Russian]
11. Watari, Sh. (1963). Best approximation by Walsh polynomials. Tôhoku Mathematical Journal,
15, 1–5.
12. Yano, Sh. (1951). On Walsh series. Tohoku Mathematical Journal, 3, 223–242.
13. Yano, Sh. (1951). On approximation by Walsh function. Proceedings of the American Mathe-
matical Society, 2, 962–967.
14. Jastrebova, M.A. (1968). On the approximation of function satisfying a Lipshitz condition by
the arithmetic means of their Walsh Fourier Series. American Mathematical Society Transla-
tions: Series 2, 77, 149–162.
Chapter 3
Haar–Fourier Analysis
3.1 Haar System and Its Generalization
It is well known that the history of Walsh series began with Haar’s(Hungarian Mathe-
matician Alfred Haar) dissertation of 1909 Zur Theories Orthogonal Function system
in which Haar system was introduced. Supervisor of Haar, David Hilbert at Göttingen
university asked him to find an orthonormal system on the interval whose Fourier
series of continuous functions converges uniformly. He constructed the system which
is under discussion in this chapter, now known as Haar system provided answer to
the problem posed by Hilbert.
Haar system {h k }∞
k=0 is defined as
h 0 (x) = 1, h 1 (x) = χ[0,1/2) (x) − χ(1/2,1] (x) (3.1)
In general writing k1 = 2n + k, 0 ≤ k < 2n where n is the largest power of 2 which

is less than or equal to k, one defines
2 1
h k (x) = h kn (x) = 2n/2 , for 2k − < x < 2k − n+1
2n+1 2
1 2k
= −2n/2 , for 2k − n+1 < x < n+1
2 2
= (2n /4)1/2 , for x = k/2n ,
= 0, otherwise.
The Haar orthonormal sequences of functions labeled by two indices are designated
by
n (x)} = {h 0 (x), h 1 (x), h 2 (x), . . . , h k (x), . . . , h k
{h m 1 1 1 2k−1
(x), . . . , }
where h 0 (x), h 11 (x), . . ., h 2k−1

k ,. . . are given by
h 0 (x) = 1, 0 ≤ x ≤ 1.
https://doi.org/10.1007/978-981-13-6370-2_3
80 3 Haar–Fourier Analysis

1, 0 ≤ x < 1/2
h 11 (x) =
−1, 1/2 < x ≤ 1
⎧√
⎨ √ 2, 0 ≤ x < 1/4
h 12 (x) = − 2, 1/4 < x < 1/2
⎩
0, 1/2 < x ≤ 1
⎧
⎨ 0,
√ 0 ≤ x < 1/2
h 22 (x) = 2 1/2 < x < 3/4 (3.2)
⎩ √
− 2, 3/4 < x ≤ 1.
and in general
⎧√
⎪
⎪ 2n−1 ,
√
m−1
≤x < m−1/2
⎪
⎪
2n−1 2n−1
⎪
⎪ − 2n−1 , m−1/2
< x < 2n−1
m
⎨ 2n−1
...
n (x) =
hm (3.3)
⎪
⎪ 0, 0 < x < m−1
⎪
⎪ 2n−1
⎪
⎪ 0, m
<x ≤1
⎩ 2n−1
n = 1, 2, 3, . . . ; m = 1, 2, . . . , 2n−1 .
At a point of discontinuity, a Haar function may be defined by a mean value around

n (x 0 ) by
that point. Hence, if x0 is such a point, we define h m
n (x 0+ ) + h n (x 0− )
hm m
n (x 0 ) =
hm . (3.4)
2
where h mn (x 0+ ) and h n (x 0− ) denote the right-hand and left-hand limits of the function
m
at x0 respectively. Figure 3.1 shows the first eight Haar functions. Outside the unit
interval, Haar functions are considered periodic with unity period.
The previous expression for Haar function reveal a different characteristic from
corresponding √expression.√
{0, ±1, ± 2, ±2, ±2 2, ±4, . . .}. On the other hand, the ordering of these func-
tions by some parameter like sequency is not easily apparent from their expressions.
But according to (8.3) we can still put Haar functions in some suitable sequencing
order by referring to the indices n and m in h m n (x) by degree and order, respectively.
Disregarding h 0 (x), then the degree n defines a subset of Haar functions having the
same number of zero-crossings with a least interval of constancy 1/2n . The order
m now defines the position of the particular Haar function in such a subset. In this
manner the degree n may be considered to provide a form of frequency or sequency
ordering.
3.1 Haar System and Its Generalization 81
Fig. 3.1 The first eight Haar functions
n is [0, T ], then√
If the domain of definition of h m the normalized Haar function
on this interval would be altered in magnitude to 2n−1 /T . Using this interval,
orthogonality of the normalized Haar functions {h m n (x)} may be expressed as
T
1n=k
n (x)h k (x) d x =
hm m
(3.5)
0 0 n = k
3.2 Haar Fourier Series
Haar functions form a complete orthonormal basis over a finite interval. Conse-
quently, it becomes possible to expand suitable functions in terms of this set. Such a
suitable function may be a Lebesgue or a square-integrable function on the indicated
interval. Hence if f (x) is square-integrable on [0, 1], we may write a Haar series
expansion representation [1] for it, as
∞
2 n−1
f (x) = a0 + n (x),
anm h m (3.6)
n=1 m=1
where the Haar expansion coefficients are given by

1
a0 = f (x) d x (3.7)
0
and
1
amn = anm = f (x)h m
n d x, n = 1, 2, 3, . . . ; m = 1, 2, . . . , 2
n−1
. (3.8)
0
The convergence problem may now be a conveniently studied by considering partial

sums of the form

N
2n−1
S N (x) = a0 + n (x).
anm h m (3.9)
n=1 m=1
Clearly S N (x) contains 2 N terms, while S N +1 (x) contains 2 N more terms than S N (x).
For a continuous function, the sequence of partial sums {S N (x) : N = 1, 2, 3, . . .}
converges uniformly at continuity points. This convergence is still true even when x
is a dyadic rational, that is, x = l/2r , for l, r = nonnegative integers such that l ≤ 2r .
For a function with discontinuities at nondyadic rational points, we have pointwise
convergence but uniform convergence may not hold any longer. The convergence
feature provides yet another difference between Haar and Walsh functions. In this
respect, Haar functions provide a superior performance since a Walsh series expan-
sion of some continuous function may actually diverge at certain points. This cannot
occur for Haar expansions.
A truncated (finite) series expansion of f (x) containing the first 2 N terms is then
given by S N (x). The function S N (x) is a step function with 2 N equal length(not
height!) steps. On each step, the value of S N (x) is simply the mean or average value
of f (x) over the step length. Function approximation by S N (x) provides the best
step function approximation in accuracy of a Haar partial sum is given in terms of a
first derivative f (x), by
3.2 Haar Fourier Series 83
|S N (x) − f (x)| ≤ max| f (x)|/2 N , (3.10)
where x lies in a subinterval of [0, 1], and max| f (x)| denotes the maximum absolute
value of f (x) in the selected subinterval. For large N, (8.10) becomes at any point
|S N (x) − f (x)| < | f (x)|/2 N +1 . (3.11)
Furthermore, a Haar expansion coefficient bears a simple relationship to the mean

value of f (x) over subintervals of [0, 1]. The relation is given by
anm = 2−(n+1)/2 { f [(2m − 2)/2n , (2m − 1)/2n ]

− f [(2m − 1)/2n , 2m/2n ]} (3.12)
where f (a, b) denotes the mean value of f (x) in the interval (a, b) that is for b > a
f (a, b) = [ f (b) − f (a)]/(b − a). (3.13)
From (3.12), it is clear that anm is proportional to the difference of two adjacent steps of
Sn (x), namely the steps on either side of x = (2m − 1)/2n . This mean-value property
of the coefficient {anm } is responsible for the modest computational requirements of
Haar transforms.
Relations Between Haar and Walsh Functions
Relations between Haar and Walsh functions may be developed by expanding one
set in terms of the other, and vice versa. Kremer, see for Example [1] developed a
convenient method for establishing conversion relations between Haar and Walsh
functions. This method utilizes the block orthogonal functions as an intermediate
step.
Block orthogonal function may be defined on a unit interval by

1 2rn ≤ x < r2+1
b2n (r, x) = n
(3.14)
0 otherwise
where r = 0, 1, 2, . . . and n = 0, 1, . . . , 2r − 1. In terms of Walsh functions, these

block functions can be expressed by
b1 (0, x) = wal(0, x) = 1. (3.15)
2 −1
1
n
b2n (r, x) = n wr (k/2n )wk (x). (3.16)

2 k=0
Now the Haar functions may be expressed in terms of block functions by
h 0 (x) = b1 (0, x) (3.17)

and √
n (x) =
hm 2n−1 [b2n (m, x) − b2n (m + 1, x)]. (3.18)
Comparison of (3.15) and (3.17), and (3.16) and (3.18), respectively, gives
φ0 (x) = w0 (x) = 1 (3.19)
and

n−1
2
1
φnm (x) = √ [wm (k/2n ) − wm+1 (k/2n )]wk (x). (3.20)
2 2n−1 k=0
Relation (3.20) express Haar functions in terms of Walsh functions. The converse
may be established in a similar way.
Discrete Haar Transform
As in the case the discrete Walsh transform, a discrete Haar transform (DHT) may
be developed for the representation of discrete signals.
Thus for a discrete signal sequence { f (i) : i = 0, 1, 2, . . . , N − 1} a correspond-
ing discrete Haar transform sequence may be defined by
N −1
2i
F(n) = f (i)φnm (i), n = 0, 1, . . . N − 1. (3.21)

i=0 m=1
where φnm (i) denotes the nth value of the Haar function φnm (x) in the i-th subinterval.
In this case, the unit interval is divided into N subintervals of length 1/N each. The
inverse discrete transform is given by
N −1 2
1
n
f (i) = F(n)φnm (i), i = 0, 1, . . . , N − 1. (3.22)

N n=0 m=1
The sequence of functions h m n (i) : i, n = 0, 1, . . . , N − 1 obviously defines a set of

discrete Haar functions. If the values of these discrete functions are arranged properly
in a matrix, we refer to this matrix as a Haar matrix. This matrix constitutes the kernel
of the discrete Haar transformations.
In matrix form, (3.21) and (3.22) can then be written as
F = Hf (3.23)
and
f = H −1 F, (3.24)
where the matrices H and H −1 denote direct and inverse Haar matrices of order N .
3.2 Haar Fourier Series 85
Extension of the one-dimensional discrete Haar transform to a two-dimensional

case is rather straightforward. Assume a two-dimensional discrete signal specified
by { f (i, j) : i, j = 0, 1, 2, . . . , N − 1}. Its discrete Haar transform components are
then specified by
N −1
N −1
2
n
2 r
F(n, r ) = f (i, j)h m

n (i)h r ( j), n, r = 0, 1, 2, . . . , N − 1.
1 m2
i=0 j=0 m 1 =1 m 2 =1
(3.25)
N −1
N −1
2n 2r
1
f (i, j) = 2 f (i, j)h m
n (i)h r ( j), n, r = 0, 1, 2, . . . , N − 1.
1 m2
N n=0 r =0 m =1 m =1
1 2
(3.26)
Applications of Haar Functions. To a large extent the utility of Walsh functions is
based on two primary aspects. The first concerns the ease with which these functions
may be generated on digital computers. The second aspect lies in the simplicity of
implementing fast computational techniques based on these functions. Both features
rely on the fact that a Walsh function attains a constant value of either +1 and −1 on
each subinterval of length, say 1/2n , where the unit interval of operation is divided
into 2n equal subintervals.
Ignoring scaling constants, a Haar function assumes one of three possible values
on each such subinterval. The three values are +1, −1 and 0. Thus, generation of Haar
functions, their binary representations, and other involved operations are not likely
to be as convenient as with Walsh functions. However, certain other features of Haar
functions make them attractive for certain applications. For instance, the convergence
aspect of a Haar series expansion provides desirable approximations, particularly in
some cases of discontinuous signals. Another feature is noted if we consider the Haar
expansion of function f (x). The resulting first two Haar coefficients are effected by
all values of f (x). But the rest of the coefficients are affected by only small sections
of f (x). Thus the first two Haar functions help to represent f (x) globally, and are
thus called global functions. The remaining Haar functions describe f (x) on local
sections, and are thus called local functions. In this sense, Walsh functions are all
global functions. The Haar function feature of localization may be very desirable if
interest is in describing certain local portions of a s signal as in television images
where the center of a picture is the significant part.
Practical applications of Haar functions and discrete Haar transforms have been
made in a number of other areas. In data coding, for instance, a signal may be encoded
by its Haar coefficients. Information transmission is then based on these coefficients
instead of the original time domain signal. In this case, the rapid convergence aspect
may be put to advantage in achieving good bandwidth reduction. In addition, certain
immunity to channel errors results.
Multiplexing with Haar carries is another area of application. In general, Haar
multiplexing provides an operation with features between time division and sequency
division multiplexing. Cross-talk problems may be less severe, and signal recovery
at the demultiplexing and may be performed rather easily and accurately.
Further application is made in some areas of pattern recognition. The property

of a Haar coefficient being easily related to the mean value of the function is made
useful in studying certain aspects of pattern recognition such as edge detection.
Applications of discrete Haar transforms have been made in digital signal process-
ing of one-dimensional and two-dimensional data. For instance, Haar transforms are
found useful in the image transmission of slowly varying images, or images which
are rather constant over large areas.
3.3 Haar System as Basis in Function Spaces
In this section we introduce C W , L p , 1 < p < ∞, H p , BMO and VMO spaces

indicating that Haar system is basis in these spaces. We will prove that Haar system
is a complete orthonormal system of L 2 [0, 1].
Let C W represent the set of functions g : [0, 1) → R which are continuous at
every dyadic irrational, continuous from the right on [0, 1) and have a finite limit
from the left on (0, 1], all this in the usual topology. Let L p denote the set of all
almost everywhere finite, Lebesgue measurable functions from [0, 1) into [−∞, ∞]
1
such that ( 0 | f | p d x)1/ p < ∞.
1
f p = ( 0 | f | p d x)1/ p is a norm and so L p , 1 ≤ p < ∞, is a Banach space.
Dyadic Hardy spaces H p , 1 < p < ∞, BMO and VMO are discussed in detail in
[2] and these spaces are special cases of Homogeneous Banach spaces discussed in
[2, 3], see also Chap. 2. Let (X, .) be a Banach space. {h n } is called a (Schauder)
basis in X if to each x ∈ X there exists a unique sequence {xk } such that
∞

x= xk h k (3.27)
k=0
in X , that is, the sequence of partial sums

n−1
sn (x) = xk h k
k=0
converges to x in the norm of X as n → ∞.
Remark 3.1 (i) The Haar system {h n } is a basis in L 1 . Infact, it is a basis in L p for
1 < p < ∞.
(ii) The Haar system {h n } is a basis in the dyadic Hardy space H p .
(iii) The Haar system {h n } is a basis in C W .
(iv) The shifted Haar system {h n } is a basis in VMO.
(v) The famous problem posed by Stefan Banach whether every separable Banach
space has a basis is known as the “Basis problem”. It has been shown [2,
3.3 Haar System as Basis in Function Spaces 87
Sect. 5.6] that answer is no for the dyadic Hardy spaces as there is a dyadic
Hardy space(separable) having no basis.
Remark 3.2 (i) Let ⎧
⎨ 1 0 ≤ t < 1/2,
ψ(t) = −1 1/2 ≤ t < 1,
⎩
0, otherwise
Haar mother wavelet and

1 0≤t <1
φ(t) =
0 otherwise
ψn,k (t) = 2n/2 ψ(2n t − k), t ∈ R.
This function is supported on In,k = [k2−n , (k + 1)2−n ), that is, it vanishes out-
side this interval. It has integral 0 and norm 1 in L 2 (R), that is,

ψn,k (t) dt = 0, ψn,k dt = 2
(ψn,k (t))2 dt = 1.
R L 2 (R) R
The Haar functions are orthogonal, that is,

ψn 1 ,k1 (t)ψn 2 ,k2 dt = δn 1 ,n 2 δk1 ,k2 ,
R
where δi, j denotes the Kronecker delta.

(ii) The Haar system on R, {ψn,k (t); n ∈ Z , k ∈ Z }, is complete in L 2 (R). The
Haar system is an orthonormal basis in L 2 (R).
Proposition 3.1 Haar functions in L 2 [0, 1] constitute an orthonormal basis
(Schauder basis).
Proof Let ψ = χ[0,1/2] − χ[1/2,1] , ψ j,k (t) = 2 j/2 ψ(2 j t − k). Let
A = {( j, k) : j ≥ 0, k = 0, 1, 2, . . . , 2 j − 1}.
We can prove that A = {χ[0,1] } ∪ {ψ j,k : ( j, k) ∈ A} is an orthonormal system in

L 2 [0, 1] using the fact that each of them is supported on [2− j k, 2− j (k + 1)], and
each different pairs i, j either has disjoint support or contained in each other support.
Let us show that A is complete. For let g ∈ L 2 [0, 1] with < g, ψi, j >= 0 and
< g, χ[0,1] >= 0 for all (i, j) ∈ A.
We show that g = 0 a.e. Let
I lj,k = [2− j , 2− j k + 2− j−1 ], I rj,k = [2− j k + 2− j−1 , 2− j (k + 1)]

l r
then ψi, j = 2− j (χ Ii,l j − χ Ii,r j ), so Ii, j ψ= Ii, j ψ for all (i, j) ∈ A.

ψ d x = 0.
[0,1]
Hence ψ = 0 a.e.
Alternative Approach. To show that every continuous function on [0, 1] is a uniform

limit of linear combination of Haar functions. Suppose f is continuous, n is a positive
integer and let
g= < f, ψ j,k > ψ j,k .
j≤n,k
By construction of piecewise constant function: it is constant on subintervals of the

form [2−n−1 k, 2−n−1 (k + 1)), k = 0, 1, . . . , 2n+1 . We claim that on each subinter-
val, the value of g is equal to the average of f on that subinterval: this implies uniform
convergence. Thanks to the uniform continuity of f .
The proof is by induction. The base case is
1 1 1
ψ= < f, χ[0,1] > χ[0,1] = f.
0 0 0
which holds because all functions except χ[0,1] have zero mean on [0, 1]. Once it has
been established that on some dyadic interval I j,k the averages of f and g are equal,
consider its halves I− and I+ : then
1 1
g− g = 2− j/2 ψ j,k g = 2− j/2 ψ j,k f = f − f
I+ I− 0 0 I+ I−
which together with

g+ g= g= f = f + f.
I+ I− I I I+ I−

imply that I+ g= I+ f and I− g= I− f.
3.4 Non-uniform Haar Wavelets
In the first place we discuss uniform Haar wavelets or classical wavelets.

Let I j,k = [2− j k, 2− j (k + 1)) be a dyadic interval.
A dyadic step function with scaling j is a function which is constant on each
interval I j,k (with j fixed).
Haar scaling function of order j are given by φ j,k (x) = 2 j/2 φ(2 j x − k), where
3.4 Non-uniform Haar Wavelets 89

10≤x <1
φ(x) =
0 otherwise
φ j,k (x) can be written as

2 j/2 x ∈ I j,k
φ j,k (x) =
0 otherwise
Let V j = {Set of all dyadic step functions of scale j which belong to L 2 (R)}.
V j is called the Haar approximation space of order j.
Remark 3.3 (i) V j ⊂ V j+1 for all j,

(ii) ∩∞
j=−∞ V j = {0}.
(iii) ∪∞
j=−∞ V j is dense in L 2 (R).
(iv) For any j ∈ Z the set of functions {φ j,k }∞
k=−∞ is an orthonormal basis for V j .
Definition 3.1 The Haar system (of wavelets) is the family
ψ j,k (x) = 2 j/2 ψ(2 j x − k), j, k ∈ Z,
where ψ is the Haar wavelet:

⎧
⎨ 1 0 ≤ x < 1/2,
ψ(x) = −1 1/2 ≤ x < 1,
⎩
0 otherwise
Remark 3.4 The set of functions {ψ j,k }∞ j,k=−∞ is an orthonormal basis in L 2 (R).

Definition 3.2 (i) For f ∈ L 2 (R), P j ( f ) = ∞ k=−∞ < f, φ j,k > φ j,k , where φ j,k
are the Haar scaling functions of order j. P j is called the approximation operator
of order j.
(ii) The Haar detail operator Q j of order j is defined as
∞

Q j( f ) = < f, ψ j,k > ψ j,k .
k=−∞
For more details, see Walnut [4].
Non-uniform Haar wavelets. Now we present investigation of Dubeau et al. [5] on

a frame work for non-uniform wavelets based on non-uniform partition of R. In this
paper Haar scaling function is used to define the associated basic non-uniform Haar
wavelet. Non-uniform multiresolution associated to these functions.
Basic Haar Scaling Function and Non-uniform Haar Wavelet
The characteristic function of the interval [0, 1[, noted χ[0,1[ will be considered as
the basic Haar scaling function ϕ

1 i f x ∈ [0, 1[,
ϕ(x) = (3.28)
0 elsewher e.
For any α1 , α2 ∈ [0, 1] such that α1 < α2 we have

x − α1
ϕ = χ[α1 ,α2 [ (x), (3.29)
α2 − α1
hence, for any α ∈]0, 1[, we obtain

x
x −α
ϕ(x) = χ[0,α[ (x) + χ[α,1[ (x) = ϕ +ϕ (3.30)
α 1−α
We define the basic non-uniform Haar wavelet, noted ψα , by
ψα (x) = −(1 − α)χ[0,α[ (x) + αχ[α,1[ (x)

x
x −α
= −(1 − α)ϕ + αϕ (3.31)
α 1−α
Proposition 3.2 The functions ϕ and ψα have the following properties:

∞
ϕ(x)d x = 1, (3.32)
−∞
∞
ψ(x)d x = 0 (3.33)
−∞
and for any n ∈ N ∞

|ϕ(x)|n d x = 1, (3.34)
−∞
∞
|ψα (x)|n d x = α(1 − α)[(1 − α)n−1 + α n−1 ] (3.35)
−∞
x x−α
We can solve for ϕ α
and ϕ 1−α
and obtain
x
ϕ = αϕ(x) − ψα (x). (3.36)
α

x −α
ϕ = (1 − α)ϕ(x) + ψα (x) (3.37)
1−α
Now for any A, B ∈ R we have from (3.36) and (3.37)

Fig. 3.2 Graph of φ(x) and ψα (x)
x
x −α
Aϕ + Bϕ = [Aα + B(1 − α)]ϕ(x) + [B − A]ψα (x), (3.38)
α 1−α
and from (3.30) and (3.31)

x
x −α
Aϕ(x) + Bψα (x) = [A − B(1 − α)]ϕ + [A + Bα]ϕ . (3.39)
α 1−α
Remark 3.5 In (3.38) the value Aα + B(1 − α) is the average of the left-hand side
expansion on [0, 1[. The value B − A is the step size of the discontinuity at x = α.
In (3.39), the value B of the step size at x = α is splitted in two parts −B(1 − α) on
[0, α[ and Bα on [α, 1[, and added to the average A on [0, 1[.
Non-uniform Multiresolution Analysis

The multiresolution analysis is based on a family {m }m∈Z of partitions of R such
that the partition {m+ } is finer than {m } for any m ∈ Z. To be more precise,
{m } = {xk(m) }k∈Z (m ∈ Z) are such that for any k ∈ Z (Fig. 3.2).
xk(m) = x2k
(m+1) (m+1)
< x2k+1 (m+1)
< x2k+2 (m)
= xk+1 (3.40)
and
lim xk(m) = −∞, lim xk(m) = ∞.
k→−∞ k→∞
For each m ∈ Z and k ∈ Z we set

x − xk(m)
ϕk(m) (x) =ϕ (m)
= χ[x (m) ,x (m) [ (x) (3.41)
xk+1 − xk(m) k k+1
and, for α = αk(m) ∈]0, 1[ defined by
(m+1) (m+1) (m+1)

x2k+1 − x2k x2k+1 − xk(m+1)
αk(m) = (m+1) (m+1)
= (m)
(3.42)
x2k+2 − x2k xk+1 − xk(m)
And using (3.31), we set


x − xk(m)
ψk(m) (x) = ψ (m) (3.43)
αk(m) (m)
xk+1 − xk(m)
= −(1 − αk(m) )χ[x (m+1) ,x (m+1) [ (x) + αk(m) χ[x (m+1) ,x (m+1) [ (x).
2k 2k+1 2k+1 2k+2
In the next proposition, properties of ϕ and ψα are extended to ϕk(m) and ψk(m) .
Proposition 3.3 The functions ϕk(m) and ψk(m) are such that
∞
ϕk(m) (x)d x = xk+1
(m)
− xk(m) , (3.44)
−∞
∞
ψk(m) (x)d x = 0, (3.45)
−∞
and for any n ∈ N

∞
|ϕk(m) (x)|n d x = xk+1
(m)
− xk(m) , (3.46)
−∞
∞
|ψk(m) (x)|d x = αk(m) (1 − αk(m) ) (1 − αk(m) )n−1 + (αk(m) )n−1 xk+1
(m)
− xk(m)
−∞
(3.47)
Moreover, orthogonality conditions and useful integrals are given in the proposition
4, 5 [5] with proof. We present these results in Remark 3.6.
Remark 3.6 (a) For any m, n ∈ Z and k, l ∈ Z we have

∞
(i) ϕk(m) (x)ψl(n) (x)d x = 0 (3.48)
−∞
∞
(ii) ϕk(m) (x)ϕl(n) (x)d x = δklmn xk+1
(m)
− xk(m) , (3.49)
−∞
∞
ψl(m) (x)ψl(n) (x)d x = δklmn αk(m) 1 − αk(m) xk+1
(m)
− xk(m)
−∞
where
1 i f m = n and k = 1,
δklmn =
0 elsewher e
(b) The functions ϕk(m−1) , ψk(m−1) , ϕ2k

(m) (m)
, ϕ2k+1 are related as follows
ϕk(m−1) (x) = ϕ2k

(m) (m)
(x) + ϕ2k+1 (x) (3.50)
φk(m−1) (x) = −(1 − αk(m−1) )φ2k

(m)
(x) + αk(m) φ2k+1
(m)
(x) (3.51)
and
(m)
φ2k (x) = αk(m−1) φk(m−1) (x) − ψk(m−1) (x), (3.52)
(m)
φ2k+1 (x) = (1 − αk(m−1) φk(m−1) (x)) + ψk(m−1) (x). (3.53)
Let us define the vector spaces
Vm = Lin{φk(m) /k ∈ Z}
and
Wm = Lin{ψk(m) /k ∈ Z}
for any m ∈ Z. Remark 3.6 yields
Vm = Vm−1 ⊕ Wm−1 . (3.54)
{Vm } is a multiresolution.
3.5 Generalized Haar Wavelets and Frames
Definition 3.3 A generalized Haar wavelet of degree M is a wavelet with the prop-
erty that there exists M ∈ R and si ∈ R such that ψ/[si , si+1 ] = ci ∈ C, Msi ∈ Z
for all i ∈ Z.
Definition 3.4 A family of functions {φi }i∈I in a Hilbert space H is a frame if there
exists A > 0 and B < ∞ such that for all f ∈ H

A f 2H ≤ | < f, ψi > |2 ≤ B f 2H (3.55)
i∈I
The result given in Remark 3.7 and its generalization are proved in G. E. Pfander [6].
Remark 3.7 For generalized Haar wavelet ψ ∈ l2 (Z) the following statements are
equivalent
(i) The family {ψn,m }m∈Z+ , n ∈ Z is a frame for l2 (Z).
(ii) There exists A > 0 and B < ∞ such
that A −s≤ ψ s (r ) ≤ B for almost all r ∈ T
+
where ψ : T → R ∪ {∞}, r → m∈Z+ m |dm (r )ψ̂(mr )2 a.e.
s

m−1
dm (r ) = e−2πilr .
l=0
P. E. Pfander and J. J. Benedetto [6] have studied applications of generalized Haar

wavelet to predict occurrence of seizure to an epileptic patient. The theory of pe-
riodic wavelet transforms presented in [6] was developed to deal with the problem
of epileptic seizure prediction. The main result is the characterization of wavelets
having time and scale periodic transforms. In reality it has been proved that such
wavelets are nothing but generalized Haar wavelets plus a logarithmic term. In the
cited paper an algorithm for periodicity detection based on the periodicity of wavelet
transforms defined by generalized Haar wavelets and implemented by wavelet aver-
aging methods. The algorithm detects periodicities imbedded in significant noise.
3.6 Applications of Haar Wavelets
Applications of Haar functions are discussed in Sect. 3.2. Haar wavelets have sim-
ilar applications but we explain here applications to initial value, boundary value
problems and integral equations. This section is based on Lepik and Hein [7] which
contains important results of 19 reputed research papers.
3.6.1 Applications to Solutions of Initial and Boundary Value

Problems
Problems Consider the nth order linear differential equation

n
Av (x)y (v) (x) = f (x), x ∈ [α, β] (3.56)
v=0
with the initial conditions
y (v) (A) = y0(v) , v = 0, 1, . . . , n − 1. (3.57)
Here Av (x), f (x) are prescribed functions, y0(v) given constants. The Haar wavelet
solution is sought in the form

2M
y (n) (x) = ai h i (x), (3.58)
i=1
where ai are the wavelet coefficients.

By integrating (3.58) n-v times we obtain

2M
y (v) (x) = ai pn−v,i (x) + Z v (x), (3.59)
i=1
3.6 Applications of Haar Wavelets 95
where

n−v−1
1
Z v (x) = (x − A)α y0(v+σ ) . (3.60)
σ =0
σ!
We shall satisfy (3.56), (3.58)–(3.60) in the collocation points x(l). If we substitute

x → x(l) and replace (3.58)–(3.59) into Eq. (3.56) we get a system of linear equa-
tions for calculating the wavelet coefficient ai . After solving this system, the wanted
function y = y(x) is calculated from (3.59) by assuming v = 0.
It is essential to estimate the exactness of the achieved results. For this purpose
we introduce the following error estimates:
(i) If the exact solution y = yex (x) is known, we define the error estimates as

y(xl )
δex
= max l − 1 (local estimate) (3.61)
y (x )
ex l
and
σex = y − yex /(2M) (global estimate). (3.62)
(ii) If the exact solution is unknown, we define the error vector

n
ε(l) = Av (xl )y (v) (xl ) − f (xl ). (3.63)
v=0
The smaller the s(l)/(2M) the more exact the solution is (ε = 0 in the case
of exact solution).
If ε(l)/(2M) is not small enough, we have to go to a higher level of resolution
J and repeat the calculations.
Boundary Value Problems. This topic is explained nicely in [7]. We mention here
main steps. Let us consider (3.56). Let us assume that μ < m are given the initial
conditions specified at x = A, but the remaining n − μ conditions are prescribed in
some interval point x1 < B or at the boundary x = B. Furthermore assume boundary
condition y (ν) (x) = y (ν) , where y∗(ν) is a given number. It follows from (3.59) and
(3.60)
Σi=1
2M
ai pn−v,i (x∗ ) + Z v (x∗) = y∗(v)

n−v−1
1
Z v (x∗) = (x∗ − A)σ y0v+σ
σ =0
σ!
3.6.2 Applications to Solutions of Integral Equations
The integral equations can be classified in the following way:

(i) Fredholm equation
B
u(x) = K [x, t, u(t)] dt + f (x), (x, t) ∈ [A, B]. (3.64)
A
(ii) Volterra Equation

x
u(x) = K [x, t, u(t)] dt + f (x), (x, t) ∈ [A, B]. (3.65)
A
(iii) Integro-differential equation

x
αu (x) + βu(x) = K [x, t, u(t), u (t)] dt + f (x), for u (A) = u 0 , x ∈ [A, B].
A
(3.66)
Here α, β, A, B denote prescribed constants, f and K are given functions, the quan-
tity K is called the kernel. The problem is to find the unknown function u(x).
If f (x) = 0 the integral equation is called first kind equation. If f (x) = 0 we
have the second kind equation.
If the kernel has the form
K [x, t, u(t)] = A(x, t)u(t). (3.67)
the Eqs. (3.64)–(3.66) are linear integral equations, if (3.67) does not hold, these
equations are nonlinear. In the case
K [x, t, u(t)] = A(x, t)(x − t)−γ , where γ ∈ (0, 1) (3.68)
the corresponding integral equation is weakly singular.

Fredholm Integral Equation. Consider this linear Fredholm integral equation
B
u(x) = K (x, t)u(t) dt + f (x), x ∈ [A, B]. (3.69)
A
The solution is sought in the form

2M
u(x) = ai h i (x), (3.70)
i=1
where ai are the wavelet coefficients and h i (x) is ( ith) computed Haar wavelet
according to (2.1) of [7]. Substituting (3.70) into (3.69) we obtain
3.6 Applications of Haar Wavelets 97

2M
2M
ai h i (x) − ai G i (x) = f (x), (3.71)
i=1 i=1
where B
G i (x) = K (x, t)h i (t) dt. (3.72)
A
(i) Collocation method. Satisfying (3.71) only at the collocation points [7], we get
a system of linear equations

2M
ai [h i (xl ) − G i (xl )] = f (xl ), l = 1, 2, . . . , 2M. (3.73)
i=1
The matrix form of this system is
a(H − G) = F (3.74)
where G il = G i (xl ), Fl = f (xl ).

(ii) Galerkin Method. For realizing this approach each term of (3.71) is multiplied
by h l (x) and the result is integrated over x ∈ [A, B]. Due to the orthogonality
condition (2.4) [7] we obtain
2M B
al
− ai Γil = f (x)h l (x) d x. (3.75)
ml i=1 A
Here l = m + k + 1, m = 2 j , j = 0, 1, . . . , J , k = 0, 1, . . . , m − 1 and
B
Γil = G i (x)h l (x) d x. (3.76)
A
3.7 Exercises
3.1 Write down a Haar series expansion of the following function

3.2 Find expansion of the function given in Exercise 3.1 with respect to non-uniform
wavelet given by (3.31).
3.3 What is the relationship between Haar and non-uniform wavelet?
1
3.4 Sketch graph of the Haar Fourier series of f (x) = x 2 for x ∈ (0, 1).
3.5 Discuss the evaluation of integrals
β
f (x)h i (x)d x
α
where h i (x), i = 0, 1, 2, ..., m − 1 are Haar functions.

3.6 Using Haar wavelets solve the boundary problems:
d3 y d2 y dy
x2 3
+ 6x 2
+6 = 6, x ∈ [4, 8]
dx dx dx
for y(4) = y(8) = 0, , y (3) = 0.

3.7 Discuss the numerical results obtained by the Haar wavelet of the equation
t
u(x) − Π λ cos Π (x + t)u(t)dt = 0
0
References
1. Maqusi, M. (1981). Applied Walsh analysis. London, Philadelphia, Rheine: Heyden.

2. Schipp, F., Wade, W. R., & Simon, P. (1990). Walsh series. Bristol: Adam Hilger.
Acta Scientiarum Mathematicarum (Szeged), 74, 593–608.
5. Dubeau, F., Elmejdani, S., & Krantini, R. (2004). Nonuniform Haar wavelets. Applied Mathe-
matics and Computation, 159, 675–693.
6. Pfander, G. E., & Benedetto, J. J. (2000). Periodic wavelet transforms and periodicity detection.
SIAM Journal of Applied Mathematics, 62(4), 1329–1368.
7. Lepik, U., & Hein, H. (2014). Haar wavelets with applications. Berlin: Springer.
Chapter 4
Construction of Dyadic Wavelets and
Frames Through Walsh Functions
4.1 Preliminary
Walsh system of functions {wl : l ∈ Z+ } on half line R+ = [0, ∞) is determined by

the equations

k
w0 (x) ≡ 1, wl (x) = (w1 (2 j x))ν j , l ∈ N, x ∈ R+ ,
j=0
where the numbers k and ν j are taken from the binary expansion

k
l= ν j 2 j , ν j ∈ {0, 1}, νk = 1, k = k(l),
j=0
but w1 (x)-function given on [0, 1) by the formula

1, if x ∈ [0, 21 ),
w1 (x) =
−1, if x ∈ [ 21 , 1),
and continued on R+ such that w1 (x + 1) = w1 (x) for all x ∈ R+ . The integer and
fractional parts of number x ∈ R+ are denoted by [x] and {x}, respectively. For each
x ∈ R+ and any j ∈ N, the numbers x j , x− j ∈ {0, 1} are determined as follows
x j = [2 j x](mod 2), x− j = [21− j x](mod 2). (4.1)
These numbers are digits of binary expansion

x= x j 2− j−1 + x j 2− j
j<0 j>0
https://doi.org/10.1007/978-981-13-6370-2_4
100 4 Construction of Dyadic Wavelets and Frames Through Walsh Functions
(in case of dyadic rational x, we get disintegration with finite number of nonzero
terms). For x, y ∈ R+ , we get

x⊕y= |x j − y j |2− j−1 + |x j − y j |2− j ,
j<0 j>0
where x j , y j are calculated by the formula (4.1). From definition, we assume x y =

x ⊕ y (since x ⊕ x = 0). It is well known that
wm (x)wn (x) = wm⊕n (x), m, n ∈ Z+ ,
but if the sum x ⊕ y is double irradiational, then
wn (x ⊕ y) = wn (x)wn (y), n ∈ Z+ , (4.2)
(in such a way that in case of fixed y, the Eq. (4.2) takes place for all x from R+ , in
except countable set).
The Walsh functions are 1-periodic, these are examined in the interval Δ = [0, 1).
For a given integer n, the numerical intervals
Ik(n) = [k2−n , (k + 1)2−n ], k ∈ Z+
are called dyadic interval of rank n. It is easy to see that
2
n
−1
Ik(n) ∩ Il(n) = φ and k = l, Δ = Ik(n) .
k=0
As Walsh functions wl (x), 0 ≤ l ≤ 2n − 1, are either equal to 1 or −1 in each of the

dyadic interval of rank n and wl (x) = 1 at x ∈ I0(n) . Any 1-periodic function w(x),
constant in each dyadic interval Ik(n) , 0 ≤ k ≤ 2n − 1 is representable in the form
2
n
−1
w(x) = cl wl (x), (4.3)
l=0
i.e., is the Walsh polynomial of order not more than 2n − 1. Such a representation
of the function w(x) is unique. The coefficients cl in the formula (4.3) coincide with
the values bk = w(k2−n ), 0 ≤ k ≤ 2n − 1, discrete Walsh transformation:
2 −1
1
n
cl = n bk wk (l2−n ), 0 ≤ l ≤ 2n − 1. (4.4)
2 k=0
Theorem 4.1 The Walsh system {wl : l ∈ Z+ } is the basis in L p (Δ) for 1 < p < ∞
(orthonormal basis in L 2 (Δ)) and not the basis in L 1 (Δ).
4.1 Preliminary 101
The function f : R+ → C is called W -continuous at the point x ∈ R+ , if
sup | f (x ⊕ h) − f (x)| → 0 as n → ∞.
0≤h< 21n
Each Walsh function W -continuous at each point of R+ . Any function f , continuous

in the classical sense at the point x, is W -continuous at this point.
Let E ⊂ R+ . If the function f is W -continuous at each point of the set E, then it is
called W -continuous on E. If the function f : R+ → C continuous at each irrational
point of R+ and has a finite limit to the left at each point of R+ \{0}, then the function
f is W -continuous on R+ .
The function f is called uniformly W -continuous on E, if for any ε > 0 there
exists the number n ∈ Z+ such that the inequality | f (x) − f (y)| < ε follows from
the conditions x ∈ E and x ⊕ y < 2−n . If the function f : R+ → C is constant in
the dyadic interval of rank n, i.e.,
f (x) = ck , where x ∈ I0(n) ,
where {ck } certain numerical sequence, then f is uniformly W -continuous on R+ .

For any x, y ∈ R+ , we write
∞

χ (x, y) = (−1)σ (x,y) , σ (x, y) = x j y− j + x− j y j ,
j=1
where the numbers x j , y j are calculated by the formulae (4.1). The function χ (x, y)
possesses the properties as given below:
(1) χ (y, x) = χ (x, y), χ (x, 0) = 1;
(2) χ (x, y) = χ (x, [y])χ ([x], y);
(3) χ (x, k) = wk (x) = wk ({x}) and any k ∈ Z+ ;
(4) if x, y, z ∈ R+ and x ⊕ y are dyadic irrational, then χ (x, z)χ (y, z) = χ (x ⊕
y, z);
(5) in case of fixed y ∈ [0, 2n ), where n ∈ Z+ , the function χ (x, y) is constant for
x on each dyadic interval I j(n+1) = [ 2n+1
j
, (2j+1)
n+1 ], j ∈ Z+ ;
(6) for any k ∈ Z+ , we have

k+1
wk (x), 0 ≤ x < 1,
χ (x, y)dy =
k 0, x ≥ 1.
The Fourier–Walsh transform of function f ∈ L 1 (R+ ) ∩ L 2 (R+ ) is defined by the

formula
ˆ
f (ω) = f (x)χ (x, ω)d x, ω ∈ R+ ,
R+
and valid for all functions from L 2 (R+ ). Let us remember that
1p
f L p (R+ ) = | f (t)| p , 1 ≤ p < ∞.
R+
Proposition 4.1 The following properties hold:

(a) if f ∈ L 1 (R+ ), then fˆ is W -continuous function on R+ and fˆ(ω) → 0 when
ω → ∞;
(b) if f, fˆ ∈ L 1 (R+ ), and f are the W -continuous function, then at each point
x ∈ R+ there takes place the inversion formula

f (x) = fˆ(ω)χ (x, ω)dω;
R+
(c) if f ∈ L 2 (R+ ), then fˆ ∈ L 2 (R+ ) and fˆ L 2 (R+ ) = f L 2 (R+ ) .

The following criterion of orthonormality of the system of shifts of the function from
L 2 (R+ ).
Proposition 4.2 Let f ∈ L 2 (R+ ). For orthonormality of the system { f (· ⊕ k) : k ∈
Z+ } in L 2 (R+ ), it is necessary and sufficient that

| fˆ(ω ⊕ l)|2 = 1, for almost all ω ∈ R+ .
l∈Z+
Let f k ∈ L 2 (R+ ), k ∈ Z+ . The family { f k } is Riesz system in L 2 (R+ ), if there exist

the positive constants A and B such that for any set of coefficients a = {ak } ∈ l 2 is
fulfilled the inequality

A a l 2 ≤ ak f k
≤ B a l 2 .
k∈Z+
By A and B, it is usually their best values. Then, the maximum value of A is the
least Riesz constant and the minimum value of B is the upper Riesz constant. Accord-
ing to Parseval’s theorem, the constants A and B are equal to 1 for the orthonormal
system { f k }.
Proposition 4.3 Let { f k } be a Riesz system in L 2 (R+ ) with the constants A and B.
Then

∞
(a) { f k } is Riesz basis in space V := { f = ∞ k=0 ck f k : k=0 |ck | };
2
(b) V = span{ f k : k ∈ Z+ };
(c) for any f ∈ L 2 (R+ ), the following inequality is fulfilled
∞

A f L 2 (R+ ) ≤ f, f k ≤ f L 2 (R+ ) ;
k=0
4.1 Preliminary 103
The Proposition 4.3 follows from the well-known property of Riesz system in Hilbert
spaces (see, e.g., Novikov et al. [1, Theorem 1.1.2]).
Proposition 4.4 Let f ∈ L 2 (R+ ). For the system { f (· ⊕ k) : k ∈ Z+ } in L 2 (R+ ) to

be Riesz system in L 2 (R+ ) with the constant A and B, it is necessary and sufficient
that
A≤ | fˆ(ω ⊕ l)|2 ≤ B, ω ∈ R+ .
l∈Z+
Let En (R+ ) be the set of functions given on R+ and constant on dyadic intervals
of rank n. Clearly, each Walsh polynomial of order 2n − 1 to the set En (R+ ). The
characteristic function of the set M ⊂ R+ is denoted by 1 M . If f ∈ En (R+ ), then
∞

f (x) = f (2−n k)1 I (n) (x), x ∈ R+ .
k
k=0
The set
E (R+ )
is defined as the collection of all sets En (R+ ). Let us indicate by Ec (R+ ), the set of
functions from E (R+ ) having compact support.
Proposition 4.5 The following properties are true:

(a) each function f ∈ E (R+ ) is uniformly W -continuous on R+ ;
(b) the sets E (R+ ) and Ec (R+ ) are compact in the spaces L p (R+ ), 1 ≤ p ≤ ∞;
(c) if f ∈ L 1 (R+ ) ∩ En , then m 0 supp f ⊂ [0, 2n ];
(d) if f ∈ L 1 (R+ ) and supp f ⊂ [0, 2n ], then fˆ ∈ En .
The last two properties show that the dyadic entire functions on R+ are analogous to
the usual entire functions on R+ (compare with the Paley-Wiener theorem on Fourier
transform of entire functions).
4.2 Orthogonal Wavelets and MRA in L 2 (R+ )
As on the real line R+ , orthogonal wavelets on positive half line R+ are constructed
using the notion of multiresolution analysis. For an arbitrary function f ∈ R+ , we
set
f j,k (x) := 2 j/2 ψ j,k (2 j x ⊕ k), j ∈ Z, k ∈ Z+ .
Definition 4.1 An orthogonal wavelet in L 2 (R+ ) is a function ψ so that the system

{ψ j,k : j ∈ Z, k ∈ Z+ .} is an orthonormal basis in L 2 (R+ ). In other words, ψ is an
orthogonal wavelet in L 2 (R+ ) if the system {ψ j,k } is orthonormal and every function
f ∈ L 2 (R+ ) can be expanded in a Fourier series in this system:

f = f, ψ j,k ψ j,k .
j,k
The simplest example of an orthogonal wavelet in L 2 (R+ ) is the Haar wavelet

⎧
⎪
⎨+1, if x ∈ [0, 1/2)
ψ H (x) = −1, if x ∈ [1/2, 1)
⎪
⎩
0, otherwise.
Definition 4.2 MRA in L 2 (R+ ) A multiresolution analysis in L 2 (R+ ) is a family

of closed subspaces V j ⊂ L 2 (R+ ) satisfying the following conditions:
(i) V ⊂ V j+1 for all j ∈ Z;
j
(ii) V j = L 2 (R+ ) and V j = {0};
(iii) f (·) ∈ V j if and only if f ∈ ψ(2· j ) ∈ V0 ;
(iv) there exists a function φ ∈ L 2 (R+ ) so that the system {φ(· ⊕ k) : k ∈ Z} is an
orthonormal basis in V0 .
According to (i), the family {V j } is a sequence of embedded subspaces. Let W j be
the orthogonal complement of V j in V j+1 , that is,
V j+1 = V j ⊕ W j+1 , j ∈ Z. (4.5)
From properties (i) and (ii), it follows the equalities
L 2 (R+ ) = ⊕ j∈Z W j = V0 ⊕ (⊕ j≥0 W j ). (4.6)
Property (iii) allows us to reproduce the entire family {V j } from one subspace V0 .
From properties (iii) and (iv), it follows that the system of functions
√
φ1,k (t) = 2φ(2t ⊕ k), k ∈ Z+
is an orthonormal basis of the subspace V1 . Since φ ∈ V0 ⊂ V1 function φ can be

expanded in a Fourier series as per this system as shown below:

φ= φ, φ1,k φi,k . (4.7)
k∈Z+
Definition 4.3 A scaling function in L 2 (R+ ) is a function φ from L 2 (R+ ) such that

φ(t) = ck φ(2t ⊕ k), (4.8)
k∈Z+
where {ck } is some sequence from l 2 .

4.2 Orthogonal Wavelets and MRA in L 2 (R+ ) 105
The function φ from condition (iv) of Definition 4.2 is called the scaling function of
the multiresolution analysis {V j }. According to (4.7), for this function of Eq. (4.8)
is satisfied with constants
√
ck = 2 φ, φi,k = 2 φ(t)φ(2t ⊕ k)dt, k ∈ Z+ . (4.9)
R+
From (4.8) and (4.9), by virtue of the orthonormality of the system {φ(· ⊕ k) : k ∈
Z+ }, by means of Parseval’s identity, the equality is derived as

¯ = 2δ0,l and, in particular,
ck ck⊕2l |ck |2 = 2. (4.10)
k∈Z+ k∈Z+
For an arbitrary sequence {ck } from l 2 , equality (4.8) can be considered as a functional
equation with respect to φ. This equation is called the scaling function for φ. We say
that a function φ generates an MRA in L 2 (R+ ), if, first, the system {φ(· ⊕ k) : k ∈
Z+ } is orthonormal in L 2 (R+ ) and, second, the family of subspaces
V j = span{φ j,k : k ∈ Z+ }, j ∈ Z+ (4.11)
satisfies the conditions

V j ⊂ V j+1 , V j = L 2 (R) V j = {0}.
Note that if the function φ generates MRA, then equality (4.7) is true for it, and hence
φ is a scaling function in L 2 (R+ ). The function
1
m 0 (ω) = ck wk (ω) (4.12)
2 k∈Z
+
is called the mask of the scaling equation (4.8) (or its solution φ). By the Parseval’s
theorem, from (4.10) and (4.12), we have
1
1 1
|m 0 (ω)|2 dω = |ck |2 = . (4.13)
0 4 2
k∈Z+
Applying the Walsh–Fourier transform, we can write Eq. (4.8) in the form
ω ω
φ̂(ω) = m̂ 0 φ̂ , ω ∈ R+ . (4.14)
2 2
Proposition 4.6 Suppose that φ ∈ L 2 (R+ ) satisfies the scaling equation (4.8) and
φ̂(0) = 0. If the system {φ(· ⊕ k) : k ∈ Z+ } is orthonormal in L 2 (R+ ), then
m 0 (0) = 1, and|m 0 (ω)|2 + |m 0 (ω ⊕ 1/2)|2 = 1, a.e. ω ∈ R+ . (4.15)
Proof The first equality in (4.15) follows from (4.14). According to Proposition 4.2,
the orthonormality of the system {φ(· ⊕ k) : k ∈ Z+ } is equivalent to the condition

|φ̂(ω + l)|2 = 1, almost everywhere ω ∈ R+ . (4.16)
l∈Z+
By virtue of the 1-periodicity of the mask m 0 , the second equality in (4.15) is obtained
after substituting (4.14) into (4.16) and grouping the summands with even and odd
indices. The proposition is proved.
Recall that for integer m, we denote by Em (R+ ), the collection of all functions f
on R+ which are constant on [2−m s, (s + 1)2−m ) (cf. Proposition 4.5). Further, we
set

E˜m (R+ ) := { f : f is W -continuous and fˆ ∈ E (R+ )} and E˜ (R+ ) := E˜m (R+ ).
m
It is known (see Golubov et al. [2, §6.2 and §10.5]) that

(1) E˜ (R+ ) is dense in L p (R+ ) for 1 ≤ p ≤ ∞;
(2) if f ∈ L 1 (R+ ) ∩ E˜m (R+ ), then supp f ∈ [0, 2m ].
Proposition 4.7 Let {V j } be a family of subspaces defined by (4.1) with a fixed

function ϕ ∈ L 2 (R+ ) φ ∈ L 2 (R+ ) with the help of formula (4.11). If the system
{φ(· ⊕ k) : k ∈ Z+ } is orthonormal basis of the subspace V0 , then V j = {0}.

Proof Let P j be the orthogonal projection of L 2 (R+ ) to V j , let f ∈ V j . Given an
ε > 0, we choose u ∈ L 2 (R+ ) ∩ E˜m (R+ ) such that f − u < ε. Then
f − P j u ≤ P j ( f − u) ≤ f − u < ε,
and so
f ≤ P j u + ε for every j ∈ Z. (4.17)
Now choose R > 0 so that suppu ⊂ [0, R). Then

R
(P j u, φ j,k ) = (u, φ j,k ) = 2 j/2 u(x)φ(2 j x ⊕ k)d x.
0
Hence, by Cauchy–Schwarz inequality

R
P j u = 2
|(P j (u), φ j,k )| ≤ u 2 2
2 j
|φ(2 j x ⊕ k)|2 d x.
k∈Z+ k∈Z+ 0
Therefore, if j is chosen enough so that Rp j < 1, then

P j u ≤ u
2 2
|φ(x)| d x = u
2 2
1 SR, j (x)|φ(x)|2 d x, (4.18)
SR j R+

where S R j := k∈Z+ {y ⊕ k : y ∈ [0, 2 j R)}. It is easy to check that
lim 1 SR j = 0 for all x ∈

/ Z+ .
j→−∞
Thus, by the dominated convergence theorem from (4.18), we get
lim P j u = 0.
j→−∞

In view of (4.17), this implies that f < ε, and thus V j = {0}.
Theorem 4.2 Let the solution φ of the scaling equation (4.8) generates MRA in
L 2 (R+ ) and satisfies the condition φ̂(0) = 1. If the function ψ is defined by the
formula
ψ(x) = (−1)k ck⊕l φ(2x ⊕ k), x ∈ R+ . (4.19)
k∈Z+
then, ψ is an orthogonal wavelet in L 2 (R+ ).
Proof According to (4.6), for the function ψ to be an orthogonal wavelet, it is suf-

ficient that the system {ψ0,k : k ∈ Z+ } to be an orthonormal basis in W0 . Applying
the Walsh–Fourier transform to (4.19), we obtain
ψ̂(ω) = m 1 (ω/2) · φ̂(ω/2),

where m 1 (ω) = 21 k∈Z+ (−1)k ck⊕l wk (ω) = −w1 (ω)m 0 (ω ⊕ 21 ). From this and
from (4.15), we see that the matrix

m 0 (ω) m 0 (ω ⊕ 21 )
(4.20)
m 1 (ω) m 1 (ω ⊕ 21 )
is unitary for almost all ω ∈ R+ . Further, using (4.15) and (4.16), we have

|ψ̂(ω ⊕ l)|2 = |ψ̂(ω ⊕ 2l)|2 + |ψ̂(ω ⊕ 2l ⊕ 1)|2
l∈Z+ l∈Z+ l∈Z+
ω 2 ω 2 ω
1 2

= m 0 + φ̂ ⊕ l + m 0 ⊕
2 l∈Z+
2 2 2

ω
1 2
+
φ̂ 2 ⊕ l ⊕ 2
l∈Z+

ω 1 2 ω 2
= m 0 ⊕ + m 0 .
2 2 2
Hence, by Proposition 4.2, the system {ψ0,k : k ∈ Z+ } is orthonormal.

Besides,
φ(t ⊕ k)ψ(t)dt = 0,
R+
In fact,

φ(t ⊕ k)ψ(t)dt = φ̂(ω)ψ̂(ω)χ (k, ω)dω
R+ R+
1
= φ̂(ω ⊕ l)ψ̂(ω ⊕ l) χ (k, ω)dω
0 l
= 0.

φ̂(ω ⊕ 2l)ψ̂(ω ⊕ 2l) + φ̂(ω ⊕ 2l ⊕ 1)ψ̂(ω ⊕ 2l ⊕ 1)
l l
ω ω
ω 2

= m0 m1 φ̂ ⊕l
2 2 2
l∈Z+

ω 1 ω 1 ω 1 2
+ m0 ⊕ m1 ⊕ φ̂ ⊕ l ⊕
2 2 2 2 l∈Z 2 2
+
ω ω
ω 1 ω 1
= m0 m1 + m0 ⊕ m1 ⊕
2 2 2 2 2 2
= 0.
Thus, ψ ∈ W0 . We now prove that W0 = span{ψ0,k : k ∈ Z+ }. By virtue of relations
V1 = V0 ⊕ W0 and V1 = span{φ0,l : l ∈ Z+ }.
it follows from the fact that for each l ∈ Z+


φ(2t ⊕ l) = αk φ(t ⊕ k) + βk ψ(t ⊕ k), (4.21)
k∈Z+ k∈Z+
where αk , βk are the Fourier coefficients of the

function φ(2t ⊕ l) with respect to
the orthonormal system {φ(t ⊕ k) : k ∈ Z+ } {ψ(t ⊕ k) : k ∈ Z+ }. According to
Parseval’s theorem, the expansion (4.21) holds the place in that and only in that case,
when

|αk | + |βk | =
2 2
|φ(2t ⊕ l)|2 . (4.22)
k∈Z+ R+
Obviously, the right side in (4.22) is equal to 21 . Comparing (4.9) with expression

αk = φ(2t ⊕ l)φ(t ⊕ l)dt = φ(2x ⊕ 2k ⊕ l)φ(x)d x,
R+ R+
1 c2k⊕l
in view of (4.12), we note that αk = 0 m 0 (ω)w2k⊕l (ω)dω = 2
. Similarly, for the
coefficients βk , we have
1
βk = m 1 (ω)w2k⊕l (ω)dω
0

1
=− m0 ω ⊕ w1 (ω)w2k⊕l (ω)dω
R+ 2
1
1
=− m 0 (ω)w2k⊕l⊕1 ω + dω
0 2
1
= (−1) l
m 0 (ω)w2k⊕l⊕1 (ω)dω
0
(−1)l
= c2k⊕l⊕1 .
2
Consequently,
1 1
|αk |2 + |βk |2 = |ck |2 = .
k∈Z
4 k∈Z
2
+ +
Therefore, (4.22) is true and the theorem is proved.
Under the conditions of the Theorem 4.2 for each fixed j, the system {ψ j,k : k ∈ Z+ }
is orthonormal basis of space W j . Orthogonal projectors P j : L 2 (R+ ) → V j and
Q j : L 2 (R+ ) → W j for each j ∈ Z are determined by the formulae

Pj f = a jk φ jk , a jk = f, φ jk
k
and

Qj f = d jk ψ jk , d jk = f, ψ jk .
k
For any f ∈ L 2 (R+ ) in accordance with (4.5), we have
P j+1 f = P j f + Q j f, P j+1 f 2 = P j f 2 + Q j f 2 , j ∈ Z. (4.23)
Sometimes it is said that Q j f contains “the details” necessary for the transition from
the j −th level of the approximation of the function f to the more accurate ( j + 1)
level. Accordingly, the subspaces {V j } (and the coefficients {a jk }) are said to be
approximating, and the subspaces {W j } (and coefficients {d jk }) are detailed. For a
fixed j and any s ∈ N from equalities (4.23), we obtain
P j f = P j−1 + Q j−1 = P j−s f + Q j−s f + · · · + Q j−1 f. (4.24)
From the relations

V j = L 2 (R+ ) and V j = {0}.
It follows that for each f ∈ L 2 (R+ ), we have the equalities
lim f − P j f = 0, and lim P j f = 0.

j→+∞ j→−∞
Thus, as j increases, the approximation error f ≈ P j f decreases to zero, and if

j → −∞, then the projections of P j f tend to the zero element of the space L 2 (R+ ).
Theorem 4.3 Let f ∈ L 2 (R+ ) and let h k = φ, φ1,k , gk = ψ, φ1,k , k ∈ Z+ ,
where the scaling function φ satisfies the conditions of the Theorem 4.2 and the
wavelet ψ is given by the formula 4.19. If the coefficients a jk of the expansion

Pj f = a jk φ jk
k∈Z+
are known, then the coefficients of the expansions of

P j−1 f = a j−1,k φ j−1,k , Q j f = d jk ψ j−1,k
k∈Z+ k∈Z+
are calculated from formulas

a j−1,k = h l⊕2k a jl , d j−1,k = gl⊕2k a jl . (4.25)
l l
Conversely, if the coefficients a j−1,k and d j−1,k are known, then the coefficients a jl
are recovered from the formula

a jl = h l⊕2k a j−1,l + gl⊕2k d j−1,k . (4.26)
k
Proof According to (4.9) and (4.19), for the coefficients in formulas (4.25) and (4.26)
there occur equality
ck
h k = √ , gk = (−1)k h k⊕l .
2
Taking into account (4.7), we have expansions

√
φ(t) = 2 h l φ(2t ⊕ l) (4.27)
l∈Z+
and √
ψ(t) = 2 gl φ(2t ⊕ l). (4.28)
l∈Z+
Using (4.27), for any j ∈ Z, k ∈ Z+ , we have

2( j−1)/2 φ(2( j−1) t ⊕ k) = 2 j/2 φ(2 j t ⊕ (2k ⊕ l))
l∈Z
and, consequently
φ j−1,l = h l φ j,2k⊕l ,
l∈Z
That is,
φ j−1,l = h l⊕2k φ jl . (4.29)
l∈Z
Similarly from (4.28), we obtain the equality

ψ j−1,k = gl⊕2k φ jl . (4.30)
l∈Z
According to (4.29), we have

a j−1,k = f, φ j−1,k = h l⊕2k f, φ j,l = h l⊕2k a jl .
l∈Z l∈Z
Similarly from (4.30), we deduce

d j−1,k = f, φ j−1,k = gl⊕2k f, φ jl = gl⊕2k a jl .
l∈Z l∈Z
Thus, the formulas (4.25) are proved. We now prove (4.26). According to (4.29) and
(4.30), we have
h l⊕2k = φ j−1,k φ jk , gl⊕2k ψ j−1,k , φ jk . (4.31)
Using the fact that the difference f − P j f is orthogonal to the subspace V j , and
applying (4.23), we have
a jl = f, φ jl = P j f, φ jl = P j−1 f, φ jl + Q j−1 f, φ jl .
Taking into account (4.31), hence, by the definition of the operators P j and Q j , we
obtain

a jl = a j−1,k φ j−1,k φ jl + ψ j−1,k φ jl
k k

= a j−1,k h l⊕2k + d j−1,k gl⊕2k ,
k k
i.e., the formula (4.26) is valid. The theorem is proved.
Further, we shall consider the case when the scaling function φ has a support on the
segment. It is easy to see that then in Eq. (4.8) the coefficients ck can be non-zero only
when 0 ≤ k ≤ 2n − 1. Indeed, let us assume that for the function φ from Definition
4.3, the condition supp φ ⊂ [0, 2n − 1] is fulfilled. According to (4.8), we have
2
n
−1 ∞

φ(x) = S1 (x) + S2 (x) = ck φ(2x ⊕ k) + ck φ(2x ⊕ k).
k=0 k=2n
If x ≥ 2n−1 and k < 2n , then 2x ⊕ k ≥ 2n , and hence φ(x) = S1 (x) = 0. If x ≥

2n−1 , then S2 (x) = 0 (since 2x ⊕ k ≥ 2n for all k ≥ 2n ). Thus, if supp φ ⊂ [0, 2n−1 ],
then φ(x) = S1 (x) for all x ∈ R+ and Eq. (4.8) can be written in the form of
2
n
−1
φ(x) = ck φ(2x ⊕ k), x ∈ R+ . (4.32)
k=0
The mask of this scaling equation is the Walsh polynomial
2 −1
1
n
m 0 (ω) = ck wk (ω). (4.33)

2 k=0
We denote by L 2c (R+ ) the class of all functions from L 2 (R+ ) with compact supports.
The following theorem contains the properties that solution of Eq. (4.32) possesses
in the general case.
Theorem 4.4 If the scaling equation (4.32) has a solution φ ∈ L 2c (R+ ) such that
φ̂(0) = 1, then
2
n
−1
ck = 2, suppφ ⊂ [0, 2n−1 ]. (4.34)
k=0
This solution is unique and is given by the formula

∞

φ̂(ω) = m 0 (2− j ω) (4.35)
j=0
and possess the following properties:

(1) φ̂(r

) = 0, for all r ∈ N;
(2) k∈Z+ φ(x ⊕ k) = 1 for x ∈ R+ .
Proof Let φ be a solution of Eq. (4.32), belonging to the space L 2c (R+ ) and such,
that φ̂(0) = 1. From the equality φ̂(ω) = φ̂( ω2 )m 0 ( ω2 ) at ω = 0. We obtain m 0 (0) =

n −1
1, and hence 2k=0 ck = 2. Further, let j be the largest integer so that function
φ on the section [ j − 1, j] does not turn into zero on a set of positive measure.
Let us assume that j ≥ 2n−1 + 1. Take an arbitrary binary irrational x ∈ [ j − 1, j].
If [2x] = l(mod 2n ), where l ∈ {1, 2, 3, . . . , 2n − 1}, then 2x > 2n + l and for any
k ∈ {0, 1, 2, 3, . . . , 2n − 1}
2x ⊕ k ≥ 2x − l ≥ 2n ,
since {2x} > 0, then 2x ⊕ k > 2n . Hence, from (4.32) it follows that j ≥ 2n + 1.
Indeed, otherwise φ(2x ⊕ k) = 0 for almost all x ∈ [ j − 1, j] for any k.It means,
by (4.32) and φ(x) = 0 almost everywhere on [ j − 1, j], which contradicts the
election of j. Further, since x ≥ j − 1 and 2x is not an integer, for any k ∈
{0, 1, 2, 3, . . . , 2n − 1}
2x ⊕ k ≥ 2x − (2n − 1) > (2 j − 2) − (2n − 1) ≥ j.
(The inequality j ≥ 2n + 1 is used). Hence, as above, we obtain that φ(x) = 0 almost

everywhere on [ j − 1, j]. Thus, j ≤ 2n−1 and therefore supp φ ⊂ [0, 2n−1 ]. Thus,
the relations (4.34) are proved. Further, since φ belongs to L 2c (R+ ), then it belongs
to L 1 (R+ ). Thus, from (4.34) by Proposition 4.5 it follows that φ̂ ∈ En−1 . Using the
condition φ̂(0) = 1, we get that φ̂(ω) = 1 for all ω ∈ [0, 21−n ). On the other hand,
m 0 (ω) = 1 for ω ∈ [0, 21−n ). Therefore, for any ω ∈ [0, 2r ) (r is a natural number),
we have
+n
r ∞
φ̂(ω) = φ̂(2−r −n ω) m 0 (2−k ω) = m 0 (2−k ω),
k=1 k=1
which proves (4.35) and uniqueness φ. Note that for any r ∈ N


j−1
φ̂(r ) = φ̂(r ) m 0 (2s r ) = φ̂(2 j r ) → 0.
s=1
when j → ∞ (since φ̂ ∈ L 1 (R+ ) and m 0 (2s r )) by virtue of the equality m 0 (0) = 1

and the periodicity of the polynomial (4.33). It follows that φ̂(r ) = 0. Applying now
the Poisson summation formula, we get
∞
∞

φ(x ⊕ k) = φ̂(r )wr (x),
k=0 r =0
(the equality is understood almost everywhere

with respect to Lebesgue measure)
and using the equality φ̂(r ) = δ0r , we obtain ∞
k=0 φ(x ⊕ k) = φ̂(0)w0 (x) = 1. The
theorem is proved.
4.3 Orthogonal Wavelets with Compact Support on R+
For every natural number n, we shall characterize all the solutions φ of the scaling
equation
2
n
−1
φ(x) = φ(2x ⊕ k), x ∈ R+ (4.36)
k=0
generating multiresolution analysis in L 2 (R+ ) and we indicate the algorithm for

constructing corresponding orthogonal wavelets. The mask of the scaling equation
(4.36) is the Walsh polynomial
2 −1
1
n
m 0 (ω) = ck wk (ω), (4.37)

2 k=0
and the Eq. (4.36) in terms of the Fourier–Walsh transform is written in the form
ω ω
φ̂(ω) = m 0 φ̂ , ω ∈ R+ . (4.38)
2 2
By Theorem 4.4, the scaling equation (4.36), under rather general conditions, has
a unique solution represented by (4.35). It follows from the properties of Walsh
functions that the mask (4.33) assumes constant values at each of the intervals
Il(n) = [l2−n , (l + 1)2−n ], 0 ≤ l ≤ 2−n .
The coefficients of the scaling equation (4.36) are calculated from the values bl =
m 0 (l2−n ), 0 ≤ l ≤ 2−n using the discrete Walsh transform
4.3 Orthogonal Wavelets with Compact Support on R+ 115
2
n
−1
1
ck = bl wl (k2−n ), 0 ≤ k ≤ 2n − 1. (4.39)
2n−1 l=0
Thus, the choice of mask values m 0 on dyadic intervals of rank n completely

determines the coefficients of Eq. (4.36), to which satisfies the corresponding
function φ.
Let us assume that φ ∈ L 2c (R+ ), φ̂(0) = 0, satisfies the scaling equation (4.36).
Then, it follows from Propositions 4.2 and 4.5 that the orthonormality of the system
{φ(· ⊕ k) : k ∈ Z+ } is equivalent to the condition

|φ̂(ω ⊕ l)|2 = 1, for all ω ∈ R+ , (4.40)
l∈Z+
and by Proposition 4.6, in this case, we have

1 2
m 0 (0) = 1 and |m 0 (ω)|2 + m 0 ω ⊕ = 1, for all ω ∈ R+ . (4.41)
2
Obviously, condition (4.41) means that for the values bl = m 0 (l2−n ) the following
equalities are fulfilled
b0 = 1, |bl |2 + |bl+2n−1 |2 = 1, 0 ≤ l ≤ 2n−1 − 1. (4.42)
According to Theorem 4.2, if φ generates MRA in L 2 (R+ ), then corresponding

orthogonal wavelet ψ is determined by the formula
2
n
−1
ψ(x) = (−1)k c̄k⊕l φ(2x ⊕ k), x ∈ R+ . (4.43)
k=0
Example 4.1 If n = 1 and c0 = c1 = 1, then the solution of Eq. (4.36) is the Haar
function φ = 1[0,1) (recall that 1 E is the characteristic function of the set E). In this
case, the mask is determined by the formula

1, ω ∈ [0, 1/2),
m 0 (ω) =
0, ω ∈ [1/2, 1).
and the wavelet ψ coincides with the Haar wavelet which is given below:
⎧
⎪
⎨1, x ∈ [0, 1/2),
ψ H (x) = −1, x ∈ [1/2, 1),
⎪
⎩
0, x ∈ R+ \[0, 1).
The corresponding system of wavelets {ψ jk } is the classical Haar basis.
Example 4.2 The Lang scaling function is obtained if for n = 2, we set

⎧
⎪
⎪1, ω ∈ [0, 1/4),
⎪
⎨a, ω ∈ [1/4, 1/2),
m(ω) =
⎪0,
⎪ ω ∈ [1/2, 3/4),
⎪
⎩
b, ω ∈ [3/4, 1),

where 0 < |a| < 1, |b| < 1 − |a|2 . In this case, the function φ satisfies the equation

3
φ(x) = φ(2x ⊕ k), (4.44)
k=0
whose coefficients are calculated according to (4.39), from formulas
1+a+b 1+a−b 1−a−b 1−a+b

c0 = , c1 = , c2 = , c3 = . (4.45)
2 2 2 2
This function generates MRA in L 2 (R+ ) and is represented by a Walsh series:
1 x ∞ x
φ(x) = 1[0,1) 1+a b j w2 j+1 −1 , x ∈ R+ . (4.46)
2 2 j=0
2
When n = 2 from (4.43), we obtain
ψ(x) = c̄1 φ(2x) − c̄0 φ(2x ⊕ 1) + c̄3 φ(2x ⊕ 2) − c̄4 φ(2x ⊕ 3).
In the case a = 1, b = 0, the Haar case is obtained (Example 4.1). If a = 0, b = 1,

then, the solution of Eq. (4.44) is the function φ(x) = ( 21 )1[0,1) ( x2 ) and the system
{φ(· ⊕ k) : k ∈ Z+ } is linearly dependent (it is easy to see that φ(x) = φ(x ⊕ 1)).
From the expansion (4.44) for the function g(x) := φ(2x) when 0 < |b| < 1, we
easily deduce Lang [3] a fractal property:

1+a−b
+ bg(2x), 0 ≤ x < 21
g(x) = 2
1−a+b
2
− bg(2x − 1), 21 ≤ x < 1.
Hence, it follows that
φ(0) = 1+a−b
2(1−b)
, φ − (1) = 1+a+b
2(1+b)
,
φ(1) = 1−a−b
2(1−b)
, φ − (2) = 1−a+b
2(1+b)
.
Here, in what follows, the left-hand limit of the function φ(x) is denoted at the point
x0 through φ − (x0 ).
It turns out that conditions (4.41) are sufficient for the fact that solution of φ of Eq.
(4.36) belonged to the class L 2 (R+ ).
Theorem 4.5 If the mask m 0 of the scaling equation (4.36) satisfies the conditions
(4.41), then the solution of φ of this equation belongs to L 2 (R+ ) and the following
inequality is correct
φ L 2 (R+ ) ≤ |φ̂(0)|. (4.47)
Proof Define the function φ̂(ω) by the formula (4.35) (recall that for each ω in this
product only a finite number of nonzero factors). Let us prove that it belongs to
L 2 (R+ ). Then, its inverse Walsh–Fourier transform will also belong to L 2 (R+ ) and
will obviously be the solution of Eq. (4.36). We have
∞

|φ̂(ω)|2 = |m 0 (2−k ω)|2 .
k=1
Since |m 0 (ω)| ≤ 1 for all ω then for any j, we have

j
|φ̂(ω)| ≤ |m 0 (2−k ω)|2 , ω ∈ R+ .
k=1
Consequently,

2j 2j j 1 j

|φ̂(ω)|2 dω ≤ |m 0 (2−k ω)|dω = 2 j m 0 (2k ω)2 dω. (4.48)
0 0 k=1 0 k=1
The function |m 0 (ω)|2 has period 1 and is piecewise constant with step 2−n . Conse-
quently, it is a Walsh polynomial of order 2n − 1:
2
n
−1
|m 0 (ω)|2 = ak wk (ω).
k=0
δ0k
From the conditions on the mask, it follows that a2k = 2
. Thus,
2 −1 n−1
1
|m 0 (ω)| = + 2
a2s+1 w2s+1 (ω).
2 s=0
Then

j−1
k 2
m 0 2 ω = as0 · · · as j−1 wr (ω),
k=1 s0 +2s1 +···+2 j−1 s j−1
where for each i either si = 0 or si is an odd number. Therefore, if r = 0, then

s0 = · · · = s j−1 = 0. Therefore,

j−1
k 2
m 0 2 ω = 2− j + br wr (ω).
k=1 r ≥1
Since 1
wr (ω)dω = 0, for all r ≥ 1,
0
then

j−1
k 2
m 0 2 ω = 2 − j ,
k=1
and from (4.48), we have

2j
|φ̂(ω)|2 dω ≤ 1.
0
Hence, when j → ∞, we obtain (4.47). The theorem is proved.

We now find out that under what conditions on the mask m 0 the solution φ of
the scaling equation (4.36) generates the orthonormal system {φ(· ⊕ k) : k ∈ Z+ }.
For this purpose, we first establish a criterion for the linear independence of the
system of entire shifts of an arbitrary function f from L 2c (R+ ). Recall that the family
{ f (· ⊕ k) : k ∈ Z+ } is a Riesz system in L 2 (R+ ) if there exist positive constants A
and B such that
⎛
⎞ 21 ⎛ ⎞ 21

A⎝ |ak |2 ⎠ ≤
a k f (· ⊕ k)
≤ B⎝ |ak |2 ⎠ , (4.49)
k∈Z+ k∈Z+ k∈Z+
L q (R+ )
for each {ak } ∈ l 2 . We can say that the function g : R+ → C has a periodic zero at
the point x ∈ R+ . If f (x ⊕ k) = 0 for all k ∈ Z+ .
Theorem 4.6 For an arbitrary function f ∈ L 2c (R+ ), the following properties are
equivalent:
(a) the system { f (· ⊕ k) : k ∈ Z+ } is a Riesz system in L 2 (R+ );
(b) the system { f (· ⊕ k) : k ∈ Z+ } is linearly independent;
(c) the Walsh–Fourier transform of f does not have periodic zeros.
Proof The implication (a) ⇒ (b) follows from the Proposition 4.3. Our next claim
is that f ∈ L 1 (R+ ) since f has compact support and f ∈ L 2 (R+ ). Let us choose a
positive integer n such that supp f ⊂ [0, 2n−1 ]. Then, by Proposition 4.5 we have
fˆ ∈ En−1 (R+ ). Besides, if k > 2n−1 , then f (x ⊕ k) = 0 for almost all x ∈ [0, 2n−1 ].
Therefore, the linear independence of the system { f (· ⊕ k) : k ∈ Z+ } in L 2 (R+ ) is
equivalent to that for the finite system { f (· ⊕ k) : k = 0, 1, . . . , 2n−1 − 1}. Further,
if some vector (a0 , a1 , . . . , a2n−1 −1 ) satisfies conditions
−1
2n−1
ak f (· ⊕ k) = 0, and |a0 | + · · · + |a2n−1 −1 | > 0, (4.50)
k=0
then using the Walsh–Fourier transform, we obtain
−1
2n−1
fˆ(ω) ak wk (ω) = 0, for almost everywhere ω ∈ R+ .
k=0

n−1
The Walsh polynomial w(ω) = 2k=0 −1 ak wk (ω) is not identically equal to zero.
Hence, among Is(n−1) , 0 ≤ s ≤ 2n−1 − 1, there exists an interval (denote it by I ) for
which w(I ⊕ k) = 0, k ∈ Z+ . Since fˆ ∈ En−1 (R+ ), it follows that (4.50) holds if
and only if there exists a dyadic interval I of range n − 1 such that f (I ⊕ k) = 0
for all k ∈ Z+ . Thus, (b) ⇔ (c). It remains to prove that (c) ⇒ (a). Suppose that fˆ
does not have periodic zeros. Then

F(ω) := | fˆ(ω ⊕ k)|2 , ω ∈ R+ ,
k∈Z+
is positive and 1-periodic function. Moreover, since fˆ ∈ En−1 (R+ ) we see that F
is constant on each Is(n−1) , 0 ≤ s ≤ p n−1 − 1. The application of Proposition 4.4
completes the proof of the theorem.
Remark 4.1 In the proof of Theorem 4.6, it was established that if the function
f ∈ L 2 (R+ ) has compact support and the system of integer translates of this function
is linearly dependent, then on the interval [0, 1) there exists a dyadic interval I , whose
all points are periodic zeros for fˆ. Moreover, if supp f ⊂ [0, 2n−1 ], then the interval
I has rank n − 1 and any periodic zero ω0 ∈ [0, 1) for f is located in such interval
I . For an arbitrary M ⊂ [0, 1), we set

1 1 1
T M := M + M ,
2 2 2
where α + β M := {α + βx : x ∈ M}. For the mask m 0 of the scaling equation

(4.36), we introduce notations N (m 0 ) := {ω ∈ [0, 1) : m 0 (ω) = 0}. Suppose that
the set M ⊂ [0, 1) is representable as a union of dyadic intervals of rank n − 1 or
m 0 ) if
coincides with one of these intervals. A set M is called blocking set (for mask
it does not contain the interval [0, 2−n+1 ) and has the property T M ⊂ M N (m 0 )
(that is, the mask m 0 turns into zero at each point of the set T M\M). Obviously,
each mask can have only a finite number of blocking sets.
( j)
If the dyadic interval Ik = [2− j k, 2− j (k + 1)] lies on [0, 1), then we sometimes
( j)
write Ik = Id1 d2 ···d j where 0ḋ1 ḋ2 · · · d j = 2− j k. If M is the blocking set for the
mask m 0 , then for any dyadic interval Id1 d2 ···dn−1 ⊂ M each of the intervals I0d1 d2 ···d j
and I1d1 d2 ···d j is located either in M, or in N (m 0 ). It is easy to see that for the mask
from Example 4.2 the interval [ 21 , 1), in the case, a = 0 is a blocking set.
Theorem 4.7 Let φ be the solution of the scaling equation (4.36) such that φ ∈
L 2c (R+ ) and φ̂(0) = 1. The system {φ(· ⊕ k) : k ∈ Z+ } is linearly dependent if and
only if the mask m 0 has a blocking set.
Proof By Theorem 4.4, we have suppφ ⊂ [0, 2n−1 ], and therefore φ̂ ∈ En−1 . If the
system {φ(· ⊕ k) : k ∈ Z+ } is linearly dependent, then the set of all periodic zeros
of the function φ̂ on [0, 1) is the blocking set for m 0 . Indeed, let
M0 = {ω ∈ [0, 1) : φ̂(ω + k) = 0, for almost all k ∈ Z+ }.
According to Remark 4.1, the set M0 is the union of dyadic intervals of rank n − 1.
Since φ̂(0) = 1,then M0 does not contain [0, 2−n+1 ). In addition, if ω ∈ M0 , then
according to (4.38) we have

ω k ω k
m0 + φ + = 0, for all k ∈ Z+ ,
2 2 2 2
and, consequently, the numbers ω2 , ω2 + 21 belong to either M0 or N (m 0 ). Conversely,

if the mask m 0 has a blocking set M, then each ω ∈ M has the property: φ̂(ω + k) =
0, k ∈ Z+ , hence, φ̂ has periodic zeros and (by Theorem 4.6) the system {φ(· ⊕ k) :
k ∈ Z+ } is linearly dependent. In fact, presuppose that there is ω ∈ M such that
φ̂(ω + k) = 0 for some k. Then, we take so high j that 2− j (ω + k) < 21−n and for
each r ∈ {0, 1, . . . , j} denote the fractional part of the number 2−r (ω + k) through
xr . Obviously, x0 = ω and x j = 2− j (ω + k). We have,

j

j
φ̂(ω + k) = φ̂(2− j (ω + k)) m 0 (2−r (ω + k)) = φ̂(x j ) m 0 (xr ). (4.51)
r =1 r =1
It is easy to see that if xr ∈ M, then xr +1 ∈ T M and, therefore, xr +1 belongs to either

N (m 0 ), or M. We also note that xr ∈ / N (m 0 ) (otherwise, from (4.51) we would have
the equality φ̂(ω + k) = 0). Thus, if xr ∈ M, then xr +1 ∈ Mr and since x0 = ω ∈ M,
then xr ∈ M for all 1 ≤ r ≤ j. However, this is impossible since x j = M. Indeed,
x j = 2− j (ω + k) < 21−n , but M does not contain points of the interval [0, 21−n ).
This completes the proof of the theorem.
The question of the existence of a blocking set by Theorem 4.7 leads to the verification
of some to the verification of some combinatorial fact, which can be verified, at least
theoretically, in finite time by simple brute force. In practice, however, for large n
n−1
this search can take considerable time to perform the order of 22 operations. In
connection with this, we formulate the following two corollaries from Theorem 4.7.
Let us call a number ω symmetric zero of the mask m 0 , if m 0 (ω) = m 0 (ω + 21 ) = 0.
Corollary 4.1 If the mask m 0 of Eq. (4.36) has a symmetric zero, then it has a
blocking set.
Indeed, if ω = 0 · d1 · d2 · · · dn · · · is a symmetric zero, then the dyadic interval Id2 ···dn

is a blocking set.
Corollary 4.2 If m 0 ( 21 − 21n ) = 0, then the system of entire shift of the solution φ of
the scaling equation (4.36) is linearly dependent.
In fact, if m 0 ( 21 − 21n ) = 0, then the interval [1 − 21−n , 1) is a blocking set.

According to the following proposition and Theorem 4.7, the condition m 0 ( 21 ) is
necessary so that the mask of the scaling equation (4.36) does not have blocking sets.
Proposition 4.8 If the system of entire shifts L 2 of the solution φ of the scaling
equation (4.36) is linearly independent, then
−1
2n−1 −1
2n−1
1
m0 = 0, c2k = c2k+1 = 1. (4.52)
2 k=0 k=0
Proof Suppose that m 0 ( 21 ) = 0. Using property 1 of Theorem 4.4, for any k ∈ Z+ ,

we have

1 1 1 1
0 = φ̂(2k + 1) = m 0 k + φ̂ k + = m0 φ̂ k + .
2 2 2 2
And, hence, φ̂(k + 21 ) = 0. Thus, the number 21 is the periodic zero of the function φ̂
which contradicts Theorem 4.6. Therefore, m 0 ( 21 ) = 0 and, by virtue of (4.34), the
equalities (4.52) are true. The proposition is proved.
L 2c (R+ ) and φ̂(0) = 1. The system {φ(· ⊕ k) : k ∈ Z+ } is an orthonormal family if
and only if the mask m 0 satisfies condition (4.41) and does not have blocking sets.
Proof We set
F(ω) := |φ̂(ω ⊕ l)|2 .
l∈Z+
As noted above (see (4.40)), the orthonormality in L 2 (R+ ) of the system {φ(· ⊕
k) : k ∈ Z+ } is equivalent to the condition F(ω) ≡ 1. Moreover, as in the proof of
Proposition 4.6, we have

2 ω 1 2 ω 1

F(ω) = |m 0 (ω) F + |m 0 ω ⊕ F ⊕ . (4.53)
2 2 2 2
1. Necessity. The absence of blocking sets follows from Theorem 4.7 and (4.41)
from the condition F(ω) ≡ 1 by (4.53), (4.41) is obtained.
2. Adequacy. We set
δ = in f F(ω)|ω ∈ [0, 1).
By Theorem 4.8, we have F(0) = 1 and, hence, δ ≤ 1: The function F is

nonnegative, has Period 1, and is constant on dyadic intervals of rank n − 1:
Therefore, either δ > 0; or F converts into zero at one of these intervals (and
then has a periodic zero). The latter is impossible by Theorem 4.6 since the
system {φ(· ⊕ k) : k ∈ Z+ } is linearly independent. Thus, 0 < δ ≤ 1. We set
Mδ = {F(ω) = δ : ω ∈ [0, 1)}. If 0 < δ < 1, then it follows from (4.41) and
(4.53) that for any ω ∈ Mδ numbers ω2 and ω2 + 21 belong to either Mδ , or N (m 0 ).
Hence, Mδ is a blocking set, which contradicts the condition. Thus, F(ω) ≥ 1
for all ω ∈ [0, 1). From this equation
1 l+1
F(ω)dω = |φ̂(ω)|2 dω = φ̂ 2L 2 (R+ ) .
0 l∈Z+ l
By Theorem 4.4, we obtain

1
F(ω)dω = 1.
0
Repeating the inequality F(ω) ≥ 1 and using the fact that the function F is
constant on dyadic intervals of rank n − 1, we conclude that F(ω) ≡ 1 The
theorem is proved.
We recall that the family {[2− j )} is a fundamental system of neighborhoods of zero of

the dyadic topology on R+ . The set E on R+ that is compact with respect to the dyadic
topology is said to be W -compact. It is easy to see that the union of a finite number
of dyadic intervals is W compact. Suppose that the set E is W compact. The set E
is said to be congruent [0, 1) as per module Z+ (the notation: E ≡ [0; 1)(modZ+ ))
if the Lebesgue measure of E is 1 and for every x ∈ [0, 1), there exists k ∈ Z+ , such
that x ⊕ k ∈ E. We say that the Walsh polynomial m 0 satisfies the modified Cohen
condition if there exists a W − compact set, E ⊂ R+ congruent [0, 1), modulo Z+ ,
containing a neighborhood of zero and such that
inf inf |m 0 (2− j ω)| > 0.

j∈Z ω∈E
The following theorem is an analogue of Cohen’s well-known theorem.

Theorem 4.9 Let the Walsh polynomial
1
2n−1
m 0 (ω) = ck wk (ω) (4.54)
2 k=0
satisfies the conditions

1 2 1
2
m 0 (0) = 1 |m 0 (ω)| + m 0 ω + = 1 for all ω ∈ [0, ), (4.55)
2 2
and the function φ ∈ L 2 (R+ ) is defined with the help of the formula
∞

φ̂(ω) = m 0 (2− j ω). (4.56)
j=1
The system {φ(· ⊕ k) : k ∈ Z+ } is an orthonormal family in L 2 (R+ ) then and only

then, when polynomial m 0 is satisfied the modified condition of Coehen.
Proof Note that φ belongs to class L 2 (R+ ) by Theorem 4.5. According to (4.56),
we have equality of φ̂(ω) = φ̂( ω2 )m 0 (( ω2 )) equivalent (4.36). It means that the given
function φ satisfies scaling equation (4.36), mask of which is the polynomial (4.54).
For each ω, all the multipliers of the work (4.56), starting with some numbers, are
equal to 1 as a mask m 0 is equal to 1 on I0(n) as 2− j ω → 0 when j → ∞. From here it
follows that φ̂ is W -continuous function (noted in the Proposition 4.1 W -continuity
of the function φ̂ follows from the Theorem 4.4). Let the system {φ(· ⊕ k) : k ∈ Z+ }
be orthonormal. As per (4.40), for every ω ∈ [0, 1), there exists a number lω such
that,
lω
1
|φ̂(ω ⊕ l)|2 > .
l=0
2
Since φ̂ is a W -continuous function, for every ω ∈ [0, 1), such a dyadic interval Iω
can be found, that
lω
1
|φ̂(t ⊕ l)|2 ≥ .
l=0
4
For every t ∈ Iω . By the W compactness of the interval [0, 1) from the covering
{Iω : ω ∈ [0, 1)}, a finite sub-covering {Iω1 , . . . , IωL } can be distinguished. Assuming
that l0 = max{Iω1 , . . . , IωL }, then the inequality

l0
1
|φ̂(ω ⊕ l)|2 ≥ (4.57)
l=0
4
is true for all ω ∈ [0, 1). Let c0 = √1

2 l0 +1
. According to (4.57), for each ω ∈ [0, 1), a
number l ∈ {0, 1, . . . , l0 } can be found such that |φ̂(ω ⊕ l)| ≥ c0 . As φ̂(0) = 1 and
function φ̂ is W -continuous, the set
S0 := {ω ∈ [0, 1) : |φ̂(ω)| ≥ c0 }
contains a neighborhood of zero. Let us assume
S1 := {ω ∈ [0, 1)\S0 : |φ̂(ω ⊕ 1)| ≥ c0 }

S2 := {ω ∈ [0, 1)\(S0 ∪ S1 ) : |φ̂(ω ⊕ 2)| ≥ c0 }
·······································

l0
St0 := ω ∈ [0, 1)\ Sl : |φ̂(ω ⊕ l0 )| ≥ c0 .
l=0
0
The set E 0 = ll=0 (Sl ⊕ l) contains neighborhood of zero, which is W -continuous
and congruent [0, 1), as per modulo Z+ . Using the fact that the polynomial m 0 (ω)
is equal to 1 in neighborhood of zero, we choose j0 in such a way that,
m 0 (2− j ω) = 1 for all j > j0 , ω ∈ E 0 . (4.58)
In view of (4.56)

j0
|φ̂(ω)| = |m 0 (2− j ω)| · |φ̂(2− j0 ω)|, (4.59)
j=1
where |φ̂(ω)| ≥ c0 for ω ∈ E 0 . As |m 0 (ω)| ≤ 1. For any ω, it follows from (4.58)

and (4.59) that

j0
|m 0 (2− j ω)| ≥ |m 0 (2−l ω)| ≥ c0 > 0, for all 1 ≤ j ≤ j0 , ω ∈ E 0 . (4.60)
l=1
From (4.58) and (4.60), we acquire, that
inf inf |m 0 (2− j ω)| > 0. (4.61)

j∈N ω∈E 0
Conversely, we assume that the polynomial m 0 (ω) in formula (4.56) satisfies condi-
tions (4.61) and (4.56). We show that the system {φ(· ⊕ k) : k ∈ Z+ } is an orthonor-
mal family in L 2 (R+ ). For every l ∈ N, we assume

l ω ω
μ[l] (ω) = m0 1E , ω ∈ L 2 (R+ ),
j=1
2j 2j
where E is the set from (4.61). As per the condition, E contains a neighborhood of
zero and m 0 (ω) = 1 for all ω ∈ [0, 2−n ). Therefore, from (4.56) it follows that
lim μ[l] (ω) = φ̂(ω), ω ∈ R+ . (4.62)

l→∞
Also by (4.61) and (4.55), there exists a number j0 such that m 0 ( 2ωj ) = 1, j >
j0 , ω ∈ E. Thus,
j0
φ̂(ω) = m 0 (2− j ω), ω ∈ E.
j=1
By (4.61), there is a constant c1 > 0 such that, |m 0 ( 2ωj )| ≥ c1 , j ∈ N, ω ∈ E. And so
−j
c1 0 |φ̂(ω)| ≥ 1 E (ω), ω ∈ R+ .
Therefore,
l ω ω l ω ω
− j0
|μ[l] (ω)| = m 0 j
1 E l
≤ c1 m 0 j
φ̂ l ,
j=1
2 2 j=1
2 2
which by (4.56) yields

−j
|μ[l] (ω)| ≤ c1 0 |φ̂(ω)| for all l ∈ N, ω ∈ R+ . (4.63)
Now, for l ∈ N, we define

Il (s) := |μ[l] (ω)|2 χ (s, ω)dω, s ∈ Z+ .
R+
Setting El := {ω ∈ R+ : 2−l ω ∈ E} and ξ = p −l ω, we have

l
l−1
ω
Il (s) = |m 0 ( j )|2 χ (s, ω)dω = 2l |m 0 (ξ )| 2
|m 0 (2 j ξ )|2 χ (s, 2l ξ )dξ.
El j=1
2 E j=1
(4.64)
Where the last integrand is 1-periodic. Using the assumption E ≡ [0, 1)(modZ+ )
and changing the variable we get from (4.64).

l l−1
1
ξ i
Il (s) = 2l−1 |m 0 ( + )|2 |m 0 (2 j−1 ξ )|2 χ (s, 2l−1 ξ )dξ.
0 0
2 2 j=1
Therefore, in view of (4.55),

1
l−2
Il (s) = 2l−1 |m 0 (2 j ξ )|2 χ (s, 2l−1 ξ )dξ
0 j=0
which by (4.64) becomes

Il (s) = Il−1 (s).
Since
1 1
Il (s) = p |m 0 (ξ )|2 χ (s, pξ )dξ = χ (s, ξ )dξ = δ0,s .
0 0
We get
Il (s) = δ0,s l ∈ N, s ∈ Z+ . (4.65)
In particular, for all l ∈ N

Il (0) = |μ[l] (ω)|2 dω = 1.
R+
By (4.62) and Fatou’s lemma, we then obtain

|φ̂(ω)|2 dω ≤ 1.
R+
Thus, from (4.62) and (4.63) by Lebesgue’s dominated convergence theorem it fol-
lows that
|φ̂(ω)|2 χ (s, ω)dω = lim Il (s).
R+ l→∞
Hence, by (4.65) and

φ(x)φ(x ⊕ k)d x = |φ̂(ω)|2 χ (k, ω)dω, k ∈ Z+
R+ R+
we have
φ(x)x ⊕ sd x = δ0,s , s ∈ Z+ .
R+
The theorem is proved.
The necessary and sufficient conditions, under which the solutions of the scaling
equation (4.36) generate MRA, are contained in the following theorem.
L 2c (R+ ) and φ̂(0) = 1. Let us suppose that the mask m 0 of this solution is φ which
satisfies the conditions (4.41). Then, the following three statements are equivalent:
(a) The function φ generates MRA in L 2c (R+ ).
(b) The mask m 0 does not have blocking sets.
(c) The mask m 0 satisfies the modified Cohen condition.
Proof Implications (a) ⇒ (b) and (a) ⇒ (c) follow from Theorems 4.8 and 4.9. To
prove the converse implications, let us suppose that for m 0 any of the conditions (b) or
(c) is completed. From Theorems 4.8 and 4.9, it follows that the system {φ(· ⊕ k) :
k ∈ Z+ } is orthonormalized.
We define {V j } by the formula (4.11). According to
Proposition 4.6, we have V j = {0}. The embedding V0 ⊂ V1 follows from the fact
that φ satisfies Eq. (4.36). Taking (4.11) into account, we therefore have V j ⊂ V j+1
for all j ∈ Z. It remains to prove that

V j = L 2 (R+ ).
Or, equivalently,
⊥
Vj = {0}. (4.66)

Let f ∈ ( V j )⊥ . For the given ε > 0, we choose a dyadic-integer function u ∈
L 1 (R+ ) ∩ L 2 (R+ ) such that f − u < ε (here and further through · signifies
the norm of space L 2 (R+ )). Then, for any j ∈ Z+ for any orthogonal projection P j f
of the function f on V j , we have
P j f 2 = (P j f, P j f ) = ( f, P j f ) = 0
and consequently,
P j u = P j ( f − u) ≤ f − u < ε. (4.67)
Next, take a natural j so large that supp û ⊂ [0, 2 j ) and 2− j ω ∈ [0, 2−n+1 ) for all ω ∈
supp û and then put g(ω) = supp û φ̂(2− j ω) (such j exists in virtue of Proposition
j
4.5). Since the system {2− 2 χ (2− j k, ·)}∞
k=0 is orthonormal basis in L [0, 2 ], we have
2 j
2j
−j
|ck (g)| = 2
2
|g(ω)|2 dω, (4.68)
k∈Z+ 0
2j
where ck (g) = 2− j 0 g(ω)χ (2− j k, ω)dω. Noting that

φ(2 j x ⊕ k)χ (x, ω)d x = 2− j φ̂(2− j ω)χ (2− j k, ω).
R+
By the Plancherel formula, we obtain

2j
−j
2 2 (u, φ j,k ) = 2− j g(ω)χ (2− j k, ω)dω.
0
Therefore, in view of (4.68),

2j
P j u =
2
|(u, φ j,k )| = 2
|û(ω)φ̂(2− j ω)|2 dω. (4.69)
k∈Z+ 0
Since m(ω) = 1 on I0(n) and since 2− j ω ∈ [0, 2−n+1 ) for ω ∈ supp û from formula
∞

φ̂(ω) = m 0 (2− j ω)
j=1
it follows that φ̂(2− j ω) = 1 for all ω ∈ supp û. Considering that supp û ⊂ [0, 2 j ),
from (4.67) and (4.68) by the Parseval theorem, we obtain
ε > P j u = û = u .
It means f < ε + u < 2ε and therefore it is true (4.66). The theorem is proved.
Applying Theorems 4.2 and 4.10, we obtain the following algorithm for constructing
orthogonal wavelets in L 2 (R+ ).
Algorithm A0
• Step-1 Select the numbers b0 , b1 , . . . , b2n −1 , satisfying the condition
b0 = 1, |bl |2 + |bl+2n−1 |2 = 1, 0 ≤ l ≤ 2n−1 − 1. (4.70)
• Step-2 Calculate the coefficients ck by formulas
2
n
−1
1
ck = bl wl (k2−n ), 0 ≤ k ≤ 2n − 1.
2n−1 l=0
• Step-3 For the Walsh polynomial
2 −1
1
n
m 0 (ω) = ck wk (ω), ω ∈ R+ ,
2 k=0
verify the feasibility of conditions (b) or (c) of Theorem 4.10.

• Step-4 Find the function φ ∈ L 2 (R+ ) such that
∞

φ̂(ω) = m 0 (2− j ω), ω ∈ R+ . (4.71)
j=1
• Step-5 Determine the function ψ by the formula

2
n
−1
ψ(x) = (−1)k c̄k⊕1 φ(2x ⊕ k), x ∈ R+ . (4.72)
k=0
Step-2 of the algorithm A0 is realized using the fast Walsh transform. Step-3 will
be implemented if the numbers bl in Step-1 are chosen so that bl = 0 for 0 ≤ l ≤
2n−1 − 1. In this case, m 0 (ω) = 0 on [0, 21 ) and condition (c) of Theorem 4.10, that
is, the modified Cohen condition, is satisfied with E = [0, 1). Note that if there are
blocking sets for the polynomial m 0 found at Step-2, then the function ψ defined
by the algorithm A0 generates the Parseval Frame in L 2 (R+ ) (for more details, see
Chap. 5).
Let us explain how to implement Step-4. For the values of bs selected in Step-1,
we set
γ (i 1 , i 2 , . . . , i n ) = bs ,
if
s = i 1 20 + i 2 21 + · · · + i n 2n−1 , i j ∈ {0, 1},
and the numerical sequence {dl : l ∈ N} as follows. Let

k
l= μ j 2 j , μk = 1, k = k(l) ∈ Z+ , (4.73)
j=0
then
dl = γ (μ0 , 0, 0, . . . , 0, 0), i f k(l) = 0;

dl = γ (μ1 , 0, 0, . . . , 0, 0)γ (μ0 , μ1 , 0, . . . , 0, 0), i f k(l) = 1;
···········································
dl = γ (μk , 0, 0, . . . , 0, 0)γ (μk−1 , μk , 0, . . . , 0, 0) · · · γ (μ0 , μ1 , μ2 , . . . , μn−2 , μn−1 ).
If k = k(l) ≥ n − 1. In the last product, the indices of each factor, starting from
the second one, are obtained by “shifting” the indices of the previous factor by one
position to the right and adding one new digit to the vacant first place from the binary
expansion of the number l. We set

M = M0 {1, 2, . . . , 2n−1 − 1},
where M0 is the set of integers l ≥ 2n−1 , for which in the above binary expansion
among the sets {μ j , μ j+1 , μ j+2 , . . . , μ j+n−1 }, there is no set (0, 0, . . . , 0, 1). Step-4
is justified by the following proposition.
Proposition 4.9 Suppose that the polynomial m 0 and the numbers dl are defined by
the values of bs under the condition (4.70) as indicated above. If the function φ is
given by its Fourier transform according to formula (4.71), then it belongs to the
class L 2 (R+ ) and ⎧

⎪
⎨1, ω ∈ [0, 2n−1 ),
1
φ̂(ω) = dl , ω ∈ [ 2n−1
l
, (l+1) ), l ∈ M, (4.74)
⎪
⎩
2n−1
(l+1)
0, ω ∈ [ 2n−1 , 2n−1 ), l ∈
l
/ M.
Proof By Parseval’s theorem and Theorem 4.5, we obtain that it belongs to the class
L 2 (R+ ). If ω ∈ I0(n−1) , then 2− j ω ∈ I0(n) for all j ∈ N. But m 0 (ω) ≡ 1 on I0(n) . This
and (4.71) imply the first equality in (4.74). We take l ∈ N with the expansion (4.73)
and find the smallest number j0 such that 2 j0 > l + 1. Then, for any ω ∈ I0(n−1) for
j ≤ j0 , we have 2 j0 ω ∈ I0(n) , and hence m 0 (2− j ω) = 1. Hence, according to (4.71),
we obtain
j0 −1

φ̂(ω) = m 0 (2− j ω), ω ∈ Il(n−1) . (4.75)
j=1
In the case, l = μ0 (that is, when k(l) = 0), in formula (4.75) there will be j0 =
2, 2−1 ω ∈ Iμ(n)
0
and 2− j ω ∈ I0(n) . Therefore, if l ∈ {1, 2, . . . , p − 1} and ω ∈ I0(n−1) ,
then
φ̂(ω) = b1 = γ (μ0 , 0, 0, . . . , 0, 0) = dl .
Similarly, in the case k(l) = 1, we have
m 0 (2−1 ω) = γ (μ1 , 0, 0, . . . , 0, 0), m 0 (2−2 ω) = γ (μ0 , μ1 , 0, . . . , 0, 0), . . . ,
and, hence, φ̂(ω) = dl . In the general case, from formula (4.73) for any j ∈ N, it
follows that
l 1 μ0 μ j−2
= + · · · + + μ j−1 + μ j 2 + · · · + μk 2 k−n
.
2n= j−1 2n 2 j−1 2
(n+ j−1)
In addition, if ω ∈ Il(n−1) then 2− j ω ∈ Il . If l ∈
/ M then among the factors in
(4.71) there is a zero (since by virtue of (4.70)). If, however, l ∈ M then in formula
(4.71) there will be
2−1 ω ∈ Iμ(n)
k
, 2−2 ω ∈ Iμ(n)
k−1 +2μk
, . . . , 2 j0 ω ∈ Iν(n)
0
,
where ν0 = μ j0 −1 + μ j0 2 + · · · + μk 2k−n . Hence, taking into account the periodicity

of the polynomial m 0 , formula (4.74) follows. The proposition is proved.
In connection with step-5, we note that under the conditions of Theorem 4.10 the
orthogonal wavelet with respect to the scaling function φ is not uniquely determined.
Namely, if for the Walsh polynomial
2 −1
1 (1)
n
m 1 (ω) = c wk (ω)
2 k=0 k
matrix m (ω) m (ω ⊕ 1 )
0 0 2
m 1 (ω) m 1 (ω ⊕ 21 )
is unitary for all ω ∈ R+ , then the burst ψ can be given by formula
2
n
−1
ψ(x) = ck(1) φ(2x ⊕ k), x ∈ R+ .
k=0
We write formula (4.74) in the form

φ̂(ω) = 1[0, 1
) (ω) + dl 1[0, 1
) (ω ⊕ 2−n+1 l). (4.76)
2n−1 2n−1
l∈M
Because
1
2n−1
−n+1
1[0, 1
) (ω ⊕ 2 l)χ (x, ω)dμ(ω) = χ (x, 2−n+1 l) χ (x, ω)dμ(ω)
2n−1
R+ 0
= 2−n+1 1[0,1) (2 −n+1

x)χ (2−n+1 x, l)
= 2−n+1 1[0,1) (2−n+1 x)wl (2−n+1 x),
the formal application of the inverse Fourier transform to (4.76) leads to the decom-
position

1 x x
φ(x) = 1[0,1) n−1 1+ dl wl n−1 , x ∈ R+ . (4.77)
2n−1 2 l∈M
2
We denote C A(n)
0 , the class of all scaling functions φ defined by steps 1–4 of algorithm
A0 for which the series (4.77) converges absolutely and uniformly on the interval
[0, 2n−1 ). Every function φ ∈ C A(n) 0 satisfies a scalable equation of the form (4.36)
since by (4.71), we have φ̂(ω) = m 0 ( ω2 )φ̂( ω2 ). The Lang’s scaling function from
Example 4.2 belongs to the class C A(2) 0 (R+ ) since for n = 2 the expansion (4.77)
coincides with (4.46). Any binary-integer scalable function belongs to the class
C A(2)
0 (R+ ) (in this case, the expansion (4.77) contains a finite number of nonzero
terms). The scaling function φ from the following example does not belong to the
class C A(3)
0 (R+ ).
Example 4.3 Let n = 3 and the function φ is determined by the algorithm A0 with
the parameters
1
b0 = 1, b4 = 0, bl = √ , for all the remaining l.
2
Grouping the coefficients of the series (4.77) with respect to the number of factors
γ (μ j , μ j+1 , μ j+2 ) included in the definition dl , l ∈ M, we note that the number of
coefficients in the k-th group is equal to Fk+1 from the Fibonacci sequence
F0 = 0, F1 = F2 = 1, Fk+1 = Fk + Fk−1 , k = 2, 3, . . . .
(k+1)
In addition, dl ≥ 2 2 for 2k ≤ l < 2k + 1. Hence, since
√
Fk+2 5+1
lim = ,
k→∞ Fk+1 2
on the basis of Dalember’s theorem, it is easy to conclude that for the chosen values
of the parameters bl , the series (4.77) diverges at the point x = 0. Therefore, the
series (4.77) does not converge absolutely.
The following example shows that for any n ≥ 3 there are Walsh series of the form
(4.77) with a finite number of nonzero summands.
Example 4.4 For an arbitrary n ≥ 3, we set

⎧
⎪
⎪ 1, ω ∈ [0, 21 − 2n−1
1
) ∪ [ 21 − 21n , 21 ),
⎪
⎨α, ω ∈ [ 21 − 2n−1
1
, 21 − 21n ),
m 0 (ω) =
⎪
⎪ 0, ω ∈ [ 2 , 2 − 2n ) ∪ [1 − 21n , 1),
1 1 1
⎪
⎩
β, ω ∈ [1 − 2n−11
, 1 − 21n ),

where 0 ≤ |α| < 1, |β| = 1 − |α|2 . For this mask with the help of (4.77), we obtain
1 x −3 x
2n−1 x
φ(x) = 1
n−1 [0,1) n−1
1 + wl n−1 + αw2n−1 − 2 n−1
2 2 l=1
2 2
x x
+ w2n−1 −1 n−1 + βw2n −2 n .
2 2
This function generates MRA in L 2 (R+ ) since condition (c) of Theorem 4.10 is
satisfied for
1 1 1 1
E = [0, 1 − ) ∪ [1 − , 1) ∪ [2 − ,2 − ).
2n−2 2n−1 2n−2 2n−1
The corresponding orthogonal wavelet ψ is found from formula (4.72).
The following exercise shows that the Lang surge ψ from Example 4.2 can be inter-
preted as a multiwavelet on R+ .
Example 4.5 Let φ, ψ be defined s in Example 4.2 in such a way that

φ(x) = c0 φ(2x) + c1 φ(2x ⊕ 1) + c2 φ(2x ⊕ 2) + c3 φ(2x ⊕ 3),

ψ(x) = c̄1 φ(2x) − c̄0 φ(2x ⊕ 1) + c̄3 φ(2x ⊕ 2) − c̄4 φ(2x ⊕ 3),
where
1+a+b 1+a−b 1−a−b 1−a+b
c0 = , c1 = , c2 = , c3 = .
2 2 2 2
It is easy prove that the vector functions

φ(x) ψ(x)
φ(x) = , ψ(x) =
φ(x ⊕ l) ψ(x ⊕ 1)
satisfy the equalities

x x
φ( ) = P0 φ(x) + P2 φ(x ⊕ 2), ψ( ) = Q 0 φ(x) + Q 2 φ(x ⊕ 2),
2 2
where
1 c0 c1 1 c2 c3
P0 = √ , P2 = √
2 c2 c3 2 c0 c1

1 c1 −c0 1 c3 −c2
Q0 = √ , Q2 = √ .
2 c3 −c2 2 c1 −c0
The dyadic derivative of the function f ∈ L 1 (R+ ) at the point x ∈ R+ is defined by

the equality
f [1] (x) = lim dl f (x),
l→∞
where

l−1
dl f (x) = 2 j−1 ( f (x) − f (x ⊕ 2− j−1 )).
j=−l+1
For any integer r ≥ 2, we put
f [r ] (x) = ( f [r −1] (x))[1] .
4.4 Estimates of the Smoothness of the Scaling Functions
In this section, we obtain estimates for the smoothness of the functions φ ∈ L 2c (R+ ),
satisfying a scaling equation of the form
2
n
−1
φ(x) = ck φ(2x ⊕ k), x ∈ R+ , (4.78)
k=0
and belonging to the class C A(n)

0 (R+ ). In this case, for the mask
2 −1
1
n
m 0 (ω) = ck wk (ω), (4.79)

2 k=0
the conditions are fulfilled

1 2
m 0 (0) = 1, |m 0 (ω)|2 + m 0 ω + = 1, for all ω ∈ R+ ,
2
and the coefficients of Eq. (4.78) are determined from the values bl = m 0 ( 2ln ), l =
0, 1, . . . , 2n , using the formula
2
n
−1
1
ck = bl wl (k2−n ), 0 ≤ k ≤ 2n − 1.
2n−1 l=0
We assume that the values of the parameters bl are real, although all estimates
of smoothness given below are also preserved in the complex case. Since φ ∈
C A(n)
0 (R+ ) the expansion takes place

1 x x
φ(x) = n−1 1[0,1) n−1 1+ dl wl n−1 , x ∈ R+ , (4.80)
2 2 l∈M
2
where the coefficients are defined as in Proposition 4.9 and the series converges
absolutely and uniformly on [0, 2n−11
).
Recall that a dyadic metric on a positive half line is introduced by means of the
distance ρ(x, y) = x ⊕ y, x, y ∈ R+ . Functions that are continuous with respect
to this metric are said to be binary continuous or W -continuous. It is easy to show
that the dyadic metric on the half line majorizes the Euclidean metric. Therefore, the
topology generated by this metric is stronger than the Euclidean one. A sequence that
converges in dyadic topology converges also in Euclidean topology, and functions
that are continuous in the usual sense are also W -continuous. The Walsh functions
wk (x) are dyadic continuous at each point x ∈ R+ .
The dyadic modulus of continuity of a scaling function φ, satisfying Eq. (4.78) is
defined by
ω(φ, δ) := sup{|φ(x ⊕ y)| − φ(x) : x, y ∈ R+ , |y| < δ}, δ > 0.
If the function φ is such that ω(φ, 2− j ) ≤ C2α j , j ∈ N for some α > 0, then there
exists see [5] the constant C(φ, α), such that
4.4 Estimates of the Smoothness of the Scaling Functions 135
ω(φ, δ) ≤ C(φ, α)δ α . (4.81)
We denote through αφ the upper bound of the set of all values α > 0, for which
inequality (4.81) is satisfied. By Theorem 4.4 for a given function φ, we have
suppφ ⊂ [0, 2n−1 ]. Therefore, an estimate of the quality αφ reduced to a study of a
sequence
ω j (φ) := sup{|φ(x) − φ(y)| : x, y ∈ [0, 2n−1 ), x ⊕ y ∈ [0, 2− j )} j ≥ 1 − n.
This can be seen from formula (4.72) that the smoothness of the orthogonal wavelet
ψ in L 2 (R+ ), corresponding to the scaling function φ coincides with αφ . We set
N = 2n−1 . By Proposition 4.8, we have
N −1
N −1

c2k = c2k+1 = 1. (4.82)
k=0 k=0
Recall that the left-side limit of function f at point a is denoted by f − (a), that is
f − (a) := lim f (x).

x→a −0
From formula (4.80) and easily verified equations
N −1
N −1

k k
wm = wm− = 0, m ∈ N,
k=0
2n−1 k=0
2n−1
it follows that
N −1
N −1

φ(k) = φ − (k + 1) = 1. (4.83)
k=0 k=0
For i, j, k ∈ {1, 2, . . . , 2N − 1} the equality k = i ⊕2 j by definition means that
ks = i s + js (mod 2) for s ∈ {0, 1, . . . , n − 1},
where i s , js , and ks are the digits of binary expansions

n−1
n−1
n−1
i= i s 2s , j = js 2s , k = k s 2s .
s=0 s=0 s=0
We define (N × N ) the matrices T0 and T1 by the formulas
(T0 )i, j = c2(i−1)⊕2 ( j−1) and (T1 )i, j = c(2i−1)⊕2 ( j−1) , (4.84)
where i, j ∈ {1, 2, . . . , N }. In particular, for n = 2, we have

c0 c1 c1 c0
T0 = , T1 = ,
c2 c3 c3 c2
where the coefficients ck are computed as in Example 4.2.

Example 4.6 Suppose that n = 3 and b0 = 1, b1 = a, b2 = b, b3 = c, b4 =
0, b5 = α, b6 = β, b7 = γ . Then, for the coefficients ck , we have formulas
1
c0 = (1 + a + b + c + α + β + γ ),
16
1
c1 = (1 + a + b + c − α − β − γ ),
16
1
c2 = (1 + a − b − c + α − β − γ ),
16
1
c3 = (1 + a − b − c − α + β + γ ),
16
1
c4 = (1 − a + b − c − α + β − γ ),
16
1
c5 = (1 − a + b − c + α − β + γ ),
16
1
c6 = (1 − a − b + c − α − β + γ ),
16
1
c7 = (1 − a − b + c + α + β − γ ).
16
In the case, the matrices (4.84) have the form
⎛ ⎞ ⎛ ⎞
c0 c1 c2 c3 c1 c0 c3 c2
⎜c2 c3 c0 c1 ⎟ ⎜c3 c2 c1 c0 ⎟
T0 = ⎜
⎝c4
⎟,T = ⎜ ⎟.
c5 c6 c7 ⎠ 1 ⎝c5 c4 c7 c6 ⎠
c6 c7 c4 c5 c7 c6 c5 c4
Note that for a = 0 and c = 0, the blocking sets for the mask
1
7
m 0 (ω) = ck wk (ω)
2 k=0
are the intervals [ 41 , 1) and [ 43 , 1), respectively. If a and c are nonzero, then the mask
m 0 satisfies the modified Cohen condition (with interval [0, 1) for b = 0 and with the
set [0, 21 ) ∪ [ 43 , 1) ∪ [ 23 , 74 ) for b = 0). Let e1 = (1, 1, . . . , 1) be an N-dimensional
vector with all components equal to 1. According to (4.82), we have
e1 T0 = e1 T1 = e1 . (4.85)
For any two N -dimensional vectors v = {v1 , . . . , v N } and w = {w1 , . . . , w N }, we set

N
√
v · w :=
t
v j w j , and v := v · vt ,
j=1
where vt and wt are the columns vector obtained from the row vectors v and w by
transposition.
We denote through E 1 the space of N -dimensional column vectors orthogonal to the
vector e1 :
E 1 := {u = (u 1 , . . . , u N )t : u 1 + · · · + u N = 0}.
For an arbitrary real (N × N ) matrix M, we set
Mu
M := sup{ : u ∈ R+ , uc = 0}
u
and
Mu
M| E1 := sup{ : u ∈ E 1 , u = 0}.
u
It is well known that the value M coincides with the square root of the largest
eigenvalue of the matrix M t M. The following Proposition holds:
Proposition 4.10 Suppose that the function φ satisfies Eq. (4.78) and belongs to the
class C A(n)
0 (R+ ). Suppose that the elements (N × N ) matrices T0 , T1 have the form
(4.84). If for all m ∈ N
max{ Td1 Td2 · · · Tdm | E1 : d j ∈ {0, 1}, 1 ≤ j ≤ m} ≤ Cq m , (4.86)
where 0 < q < 1 and C > 0, then for any integers j ≥ n the inequality
ω j (φ) ≤ Cq j . (4.87)
Proof For any x ∈ [0, 1) and {0, 1, . . . , N − 1}, we set
φ0 (x ⊕ k) := (1 − x)φ(k) + xφ − (k). (4.88)
The sequence of vector functions v j (x) for j ∈ Z+ , x ∈ [0, 1) we define the equal-
ities
v0 (x) := (φ0 (x), φ0 (x ⊕ 1), . . . , φ0 (x ⊕ (N − 1)))t ,

T0 v j (2x), if x ∈ [0, 21 ),
v j+1 (x) :=
T1 v j (2x ⊕ 1), if x ∈ [ 21 , 1).
For each x ∈ [0, 1) having double expansion

∞

x= x j 2 j+1 , x j ∈ {0, 1}.
j=0
Set d j (x) = x j (if the number x is dyadic rational, then among the digits d j (x) only
a finite number of nonzero ones). Let

2x, if x ∈ [0, 21 ),
τ x :=
2x ⊕ 1, if x ∈ [ 21 , 1).
Then
v j+1 (x) = Td1 (x) v j (τ x),
and, consequently,
v j (x) = Td1 (x) Td2 (x) · · · Td j (x) v0 (τ j x). (4.89)
From (4.83) and (4.88), we find
N −1

e1 · v0 (x) = φ0 (x ⊕ k)
k=0
N −1
N −1

= (1 − x) φ(k) + x φ − (x)
k=0 k=0
= 1.
From this and (4.85) and (4.89), we obtain
e1 · v j (x) = e1 Td1 (x) Td2 (x) · · · Td j (x) v0 (τ j x) = e1 · v0 (τ j x) = 1.
Consequently, for each l ∈ Z+ , we have
e1 · (v j+1 (x) − vl (x)) = 0.
Thus, for all x ∈ [0, 1) and for all j, l ∈ Z+
e1 · v j (x) = 1 and v j+1 (x) − vl (x) ∈ E 1 (4.90)
and according to (4.89),
v j+1 (x) − v j (x) = Td1 (x) Td2 (x) · · · Td j (x) [vl (τ j x) − v0 (τ j x)]. (4.91)
For l = 1 from (4.86), (4.90), and (4.91), we have

v j+1 (x) − v j (x) ≤ Cq j sup v1 (y) − v0 (y) .

y∈[0,1)
Therefore,

v j (x) ≤ v0 (x) + vl (x) − vl−1 (x)
l=1
≤ sup v0 (x) + C(1 − q)−1 sup v1 (y) − v0 (y) .
y∈[0,1) y∈[0,1)
Thus, the sequence {v j (·)} is uniformly bounded on [0, 1):
sup{ v j (x) |x ∈ [0, 1), j ∈ Z+ } < ∞. (4.92)
As above, for any x ∈ [0, 1) and l ∈ N in view of (4.86), (4.90), and (4.91), we have
v j+1 (x) − v j (x) ≤ Cq j sup v1 (y) − v0 (y) .

y∈[0,1)
From this and (4.92) follow the inequality
sup v j+1 (x) − v j (x) ≤ Cq j , (4.93)

y∈[0,1)
where C does not depend on l. Hence, the sequence {v j (·)} is fundamental in the
space
[C([0, 1))] N = C([0, 1)) × · · · × C([0, 1)).
By (4.93), for the limit vector-valued function ṽ(·), we have
sup ṽ(x) − v j (x) ≤ Cq j . (4.94)

y∈[0,1)
Let
v(x) := (φ(x), φ(x ⊕ h [1] ), . . . , φ(x ⊕ h [N −1] ))t .
Then, for x ∈ [0, 1) v(x) = Td1 (x) v(τ x) and assuming j → ∞ in equality v j+1 (x) =
Td1 (x) v j (τ x). We conclude that ṽ(x) = v(x). From this and (4.94) for all j ∈ Z+ , we
have
sup φ(x) − φ j (x) ≤ Cq j . (4.95)
y∈[0,1)
We fix an integer j ≥ n and choose in [0, N ) of the numbers

∞
∞

x= x j 2 j+1 and y = y j 2 j+1 .
j=2−n j=2−n
So that xi = yi for 2 − n ≤ i ≤ j and x j+1 = y j+1 . It is seen that both the numbers
x and y belong to the interval [ 2mj , (m+1)
2j
), where

j

j
m= xi 2 j−i = yi 2 j−i .
i=2−n i=2−n
According to (4.95), we have

m
|φ(x) − φ(y)| ≤ |φ(x) − φ j (x)| + |φ j (x) − φ j ( )|
2j
m
+ |φ j () − φ j (y)| + |φ j (y) − φ(y)|
2 j m m

≤ 2Cq j + φ j (x) − φ j + φ j (y) − φ j . (4.96)
2j 2j
Suppose that x ∈ [0, 1) and x0 = x−1 = · · · = x2−n = 0. Then, m = x1 2 j−1 +

x2 2 j−2 + · · · + x j 20 and, similarly to (4.91)
m τ jm
v j (x) − v j ( j
) = Td1 (x) Td2 (x) · · · Td j (x) [v0 (τ j x) − v0 ( j )]. (4.97)
2 2
Taking into account that v0 (τ j ·) are uniformly bounded on [0, 1), and also that
m m

φ j (x) − φ j ≤ v j (x) − v j ( )
2j 2j
and
Td1 (x) Td2 (x) · · · Td j (x) ≤ Cq j
from (4.97), we obtain the inequality

m

φ j (x) − φ j ≤ Cq j . (4.98)
2j
Let now x ∈ [1, N ). We set x = x ⊕ k and m = m − k, where k = x0 2 j +

x−1 2 j+1 + · · · + x2−n 2 j+n−2 . Then x ∈ [0, 1) and
τ j m
v j (x ) − v j (m /2 j ) = Td1 (x ) Td2 (x ) · · · Td j (x ) [v0 (τ j x ) − v0 ( )].
2j

Because |φ j (x) − φ j ( 2mj )| ≤ v j (x ) − v j ( m2 j ) , we again come at the estimate
(4.98). The last term in (4.98) is estimated similarly. Therefore, the inequality (4.87)
is true and Proposition 4.10 is proved.
We note that the vector-valued function v(x) is related to the matrices T0 and T1 by
the formula

T0 v(2x), if x ∈ [0, 21 ),
v(x) = (4.99)
T1 v(2x ⊕ 1), if x ∈ [ 21 , 1).
From this formula for x = 0, it follows that the vector
v(0) = (φ(0), φ(1), . . . , φ(N − 1))t
is an eigenvector of the matrix T0 corresponding to the eigenvalue 1. Formula (4.99)

is useful in calculating the values of the function φ. For example, for n = 2, the
values φ(0) and φ(1) given in Example 4.2 are obtained from (4.83) and (4.99).
Example 4.7 Let n = 2, 0 < |b| < 1, and let the function φ be given by (4.46). Then,
we have the estimate
ω j (φ) ≤ C|b| j , j ∈ N. (4.100)
Indeed, in this case E 1 = {v ∈ R2 |v1 + v2 = 0} = {te10 |t ∈ R} and T0 e10 = be10 =

−be10 , where e10 = (−1, 1)t . Therefore, inequality (4.100) follows from Proposition
4.10. Because φ(0) − φ(2− j ) = 1−bab j
, then the estimate (4.100) is exact in order.
Example 4.8 Let φ ∈ C A(3)

0 (R+ ) and the coefficients ck be defined as in Example
4.6. Suppose that a = 1 and |γ | < 1. Then, using (4.83) and (4.99), we obtain the
equations
1+c−γ c(1 − β) b
φ(0) = − + ,
2(1 − γ ) 4(1 − γ ) 4
1−c−γ c(1 − β) b
φ(1) = + − ,
2(1 − γ ) 4(1 − γ ) 4
b c(1 − β)
φ(2) = − ,
2 2(1 − γ )
b c(1 − β)
φ(3) = − + .
2 2(1 − γ )
We show that
ω j (φ) ≤ C|γ | j , j ∈ N. (4.101)
For n = 3, the basis of the space E 1 is formed by the vectors

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
1 1 1
⎜−1⎟ ⎜−1⎟ ⎜1⎟
e1 = ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎝ 1 ⎠ , e2 = ⎝−1⎠ , e3 = ⎝−1⎠ . (4.102)
−1 1 −1
Under the condition a = 1, we have

T0 e1 = T1 e1 = 0,
T0 e2 = −T1 e2 = βe1 + γ e2 ,
T0 e3 = T1 e3 = be1 + ce2 .
Using these equalities, for an arbitrary vector ν = ν1 e1 + ν2 e2 + ν3 e3 . Space E 1 , we

obtain
T02 w = −T1 T0 w = (ν2 γ + ν3 c)T0 e2 , T12 w = −T0 T1 w = (ν2 γ − ν3 c)T0 e2 ,

(4.103)
and, in particular, T02 e2 = T12 e2 = −T0 T1 e2 = −T1 T0 e2 = γ T0 e2 . Therefore, for any
d1 , d2 , . . . , dm ∈ {0, 1}, the following equality is correct:
Td1 · Td2 · · · Tdm e2 = ±γ m−1 T0 e2 .
Taking into account (4.103), we obtain
Td1 · Td2 · · · Tdm | E1 ≤ C|γ |m ,
and estimate (4.101) follows from Proposition 4.10. From formula (4.80) with a =
1, b = 0, |γ | < 1 for all j ∈ N, we derive the equality

1 cγ j+3
φ(0) − φ = .
2j 4(1 − γ )
Thus, the estimate (4.101) is exact in order. The estimate is also preserved for γ = 0
since, in this case,
1 x
φ(x) = 1[0,1) (y)(1 + w1 (y) + bw2 (y) + w3 (y) + βw6 (y)), y =
4 4
(See Example 4.4).
We recall that the joint spectral radius of two complex matrices A0 and A1 of size
N × N is determined by the formula
1
ρ̂(A0 , A1 ) := lim max{ Ad1 Ad2 · · · Adk k : d j ∈ {0, 1}, 1 ≤ j ≤ k},
k→∞
where · is an arbitrary norm in C N ×N (a bibliographic reference on the joint

spectral radius is given in the comments on Chap. 7 in see [1]). Obviously, if A0 = A1 ,
then value ρ̂(A0 , A1 ) coincides with the spectral radius ρ(A0 ). The following two
propositions take place:
Proposition 4.11 For any (d × d) matrices A0 , A1 and for any q ≤ 0 the following
conditions are equivalent:
(1) there exists constants C > 0, p ≤ 0 such that for all m ∈ N, we have
max{ Ad1 · · · Adm : d j ∈ {0, 1}, 1 ≤ j ≤ m} ≤ Cm p q m ;
(2) ρ̂(A0 , A1 ) ≤ q.
Moreover, if any of these properties are fulfilled, then we can always take p ≤ d − 1,
and if the matrices are irreducible (do not have nontrivial real common eigensub-
spaces), then p = 0.
Proposition 4.12 Let the function φ and the matrices T0 , T1 be the same as in
Proposition 4.10, and let ρ̂ = ρ̂(L 0 , L 1 ) be the joint spectral radius of the linear
operators L 0 L 1 given on R N by the matrices T0 , T1 and bounded on the subspaces
E 1 = {u = (u 1 , . . . , u N )t : u 1 + · · · + u N = 0}.
Then
1
αφ = log2 ( ). (4.104)
ρ̂
Analogues of Propositions 4.11 and 4.12 for wavelets on the line R have been proved
by V. Yu. Protasov [4], and for dyadic wavelets on the half line R+ , the proofs
are completely analogous. Formula (4.104) allows us to accurately find the Hölder
exponent; however, its practical use is complicated by the need to calculate the joint
spectral radius. It follows from this formula that the function φ is W -continuous,
then, and only then ρ̂ < 1. In the general case, if for the matrices T0 and T1 , the
inequality is fulfilled
max{ Td1 · · · Tdm | E1 : d j ∈ {0, 1}, 1 ≤ j ≤ m} ≤ Cm p q m , m ∈ N,
where 0 ≤ q < 1, p ≥ 0 then ρ̂(L 0 , L 1 ) ≤ q.

It follows from Proposition 4.12 and Example 4.7 that for the Lang scaling function
the equality αφ = log2 ( |b|
1
) is correct (so that αφ → ∞ for b → 0 and αφ → 0 for
b → 1). We indicate a few more cases when the value αφ can be calculated exactly.
Let n = 3 and the coefficients of the scaling equation (4.78) are defined as in
Example 4.6 with respect to the parameters b0 = 1, b1 = a, b2 = b, b3 = c, b4 =
0, b5 = α, b6 = β, b7 = γ , where |a|2 + |α|2 = |b|2 + |β|2 = |c|2 + |γ |2 = 1. In
this case, the matrices of the linear operators L 0 , L 1 defined on R4 by the matrices
T0 and T1 and bounded on the subspaces E 1 , we have the following form in the basis
(4.102): ⎛ ⎞ ⎛ ⎞
0βb 0 −β b
A0 = ⎝ 0 γ c ⎠ , A1 = ⎝ 0 −γ c ⎠ (4.105)
α 00 −α 0 0
and in this case ρ̂(L 0 , L 1 ) = ρ̂(A0 , A1 ).

Theorem 4.11 Suppose that φ be the solution of (4.78) defined above for n = 3 and
let ρ̂ be the joint spectral radius of the matrices (4.105). Then
⎧√
⎪
⎨ |α|,√
3
if b = 0, |c| = 1, 0 ≤ |α| < 1,
ρ̂ = max{ |α|, |γ |}, if |b| = 1, 0 ≤ |α| < 1, 0 ≤ |γ | < 1,
⎪
⎩
|γ |, if |a| = 1, 0 ≤ |γ | < 1.
and in this case αφ = log2 ( ρ̂1 ).
Proof Suppose that β = c = 1. Then, for the basis vectors (4.102), we have
T0 e1 = −T1 e1 = αe3 , T0 e2 = −T1 e2 = e1 , T0 e3 = T1 e3 = e2 .
We denote through mod (m, 3) the remainder when dividing m by 3. By induction,

we verify
T0m e1 = α [ 3 ] e3−mod (m+2,3) , m ∈ N,
m+2
where [x] is the integer part of the number x. Similarly,
T0m e2 = α [ 3 ] e3−mod (m+4,3) , T0m e3 = α [ 3 ] e3−mod (m+3,3) , m ∈ N.

m+1 m
Hence, we obtain that

m
T0m e j ≤ C|α| 3 , j = 1, 2, 3.
Moreover, since T0 e j = ±T1 e j , similar estimates exist for the vectors Td1 · Td2 · · ·
Tdm e j . Therefore, for all m ∈ N
m
Td1 · Td2 · · · Tdm | E1 ≤ C|α| 3 , d1 , d2 , . . . , dm ∈ {0, 1},
√
and the estimate ρ̂ ≤ 3 |α| is valid (see Propositions 4.10 and 4.11). We prove the
converse inequality. Since b = γ = 0, by the formula (4.80), we obtain the expansion
1 "
φ(x) = 1[0,1) (y) 1 + α w1 (y) + w3 (y) + w6 (y) + αw13 (y) + αw27 (y) + αw54 (y)
4 #
+ α 2 w109 (y) + α 2 w219 (y) + α 2 w438 (y) + α 3 w877 (y) + · · · ,
where y = x4 . The absolute convergence of this series follows from the condition
|α| < 1. It is also seen that
1 3a
φ(0) = 1+ .
4 1−α
In addition, if s = 3k, then

1 1 " s−3 s s+3

#
φ( ) = 1 + α 3 + 3α + · · · + 3α 3 − α3 − α 3 + ···
2s 4
1
1 + α(3 − 4α 3 )(1 − α)−1 .
s
=
4
Further, if s = 3k + 1, then
1 1 " s−4 s−1 s+2 s+5

#
φ( ) = 1 + a 3 + 3α + · · · + 3α 3 + α 3 − α 3 − α 3 ···
2s 4
1
1 + a{(3(1 − α 3 ) − α 3 )(1 − α)−1 + α 3 } ,
s−1 s+2 s−1
=
4
and if s = 3k + 2, then
1 1 " s−2 s+1 s+4

#
φ( ) = 1 + a 3 + 3α + · · · + 3α 3 − α 3 − α 3 + ···
2s 4
1
1 + a{(3 − 4α 3 )(1 − α)−1 } .
s+1
=
4
From the above expansions, we obtain

φ(0) − φ 1 = C|α| 3s , s ∈ N,
2
s
√
and, therefore, ρ̂ ≥ 3 |α|. Cases β = −1, c = 1, β = 1, c = −1, and β = c = √ −1
are considered similarly. Therefore, if b = 0, |c| = 1, 0 ≤ |α| < 1, then ρ̂ = 3 |α|.
Let now b = 1. Then
T0 e1 = −T1 e1 = αe3 , T0 e2 = −T1 e2 = γ e2 , T0 e3 = T1 e3 = e1 + ce2 .
Therefore, for an even m
T0m e1 = α 2 (e1 + ce2 ) + (γ 2 α 2 −1 + γ 4 α 2 −2 + · · · + γ m−2 α)ce2 ,

m m m
T0m e2 = γ m e2 ,
T0m e3 = α 2 e3 + (γ α 2 −1 + γ 3 α 2 −2 + · · · + γ m−1 α)ce2
m m m
and
m
T0m e1 ≤ C |α| 2 + |γ |2 |α| 2 −1 + |γ |4 |α| 2 −2 + · · · + |γ |m−2 |α|
m m

Cm|γ |m , |α| ≤ |γ |2 ,
≤ m
Cm|γ | 2 , |α| ≥ |γ |2 ,
T0m em ≤ C|γ |m ,

Cm|γ |m , |α| ≤ |γ |2 ,
T0 em ≤
m
m
Cm|α| 2 , |α| ≥ |γ |2 .
Further, for odd m, we have

m+1
m−1 m−3

T0m e1 = α 2 e3 + γ α 2 + γ 3 α 2 + · · · + γ m−2 α ce2 ,
T0m e2 = γ m e2 ,
m−1
m−3 m−5

T0m e3 = α 2 (e1 + ce2 ) + γ 2 α 2 + γ 4 α 2 + · · · + |γ |m−2 α ce2
and, as above,
m+1 m−1 m−3

T0m e1 ≤ C |α| 2 + |γ ||α| 2 + |γ |3 |α| 2 + · · · + |γ |m−2 |α|

Cm|γ |m , |α| ≤ |γ |2 ,
≤ m
Cm|γ | 2 , |α| ≥ |γ |2 ,
T0m em ≤ C|γ |m ,

Cm|γ |m , |α| ≤ |γ |2 ,
T0 em ≤
m
m
Cm|α| 2 , |α| ≥ |γ |2 .
Similar estimate also occurs for the vectors Td1 Td2 · · · Tdm e j . Therefore,
1 1
Td1 Td2 · · · Tdm | E1 m ≤ Cm m max{ |α|, |γ |}, m ∈ N,
√
and, consequently, ρ̂ ≤ max{ |α|, |γ |}. For the proof of the inverse inequality, we
use the expansion

4φ(x) = 1 + α w1 (y) + w2 (y) + cw3 (y) + αw5 (y) + γ cw7 (y) + αw10 (y)
+ αcw11 (y) + γ 2 cw15 (y) + α 2 w21 (y) + αγ cw23 (y) + γ 3 cw31 (y)
+ α 2 w42 (y) + α 2 cw43 (y) + αγ 2 cw47 (y) + γ 4 cw63 (y) + α 3 w85 (y)
+ α 2 γ cw87 (y) + αγ 3 cw95 (y) + γ 5 cw127 (y) + α 3 w170 (y) + α 3 cw171 (y)

+ α 2 γ 2 cw175 (y) + αγ 4 cw191 (y) + γ 6 cw255 (y) + α 4 w341 (y) + · · · ,
(4.106)
where y = x
4
and x ∈ [0, 4). Substitute in (4.106) the value of x = 0, we find
1 α c
φ(0) = 1+ 2+ .
4 1−α 1−γ
Further, it follows from (4.106) that if s is even, then
1 1 2a " s
s s−2
c(γ s − 1 ) #
φ( s ) = 1+ 1 − α 2 − c α 2 + γ 2 α 2 + · · · + αγ s−2 − 2 ,
2 4 1−α 1−γ
and if s is odd, then
1 1 2a " s+1
s−1 s−3
c(γ s − 1 ) #
φ( )= 1+ 1 − α 2 − cγ α 2 + γ 2 α 2 + · · · + αγ s−3 − 2
.
2 s 4 1−α 1−γ
Therefore, s

φ(0) − φ 1 = Cs max{ |α|, |γ |} , s ∈ N,
2s
√ √
and, ρ̂ ≥ max{ |α|, |γ |}, Thus, ρ̂ = max{ |α|, |γ |}.The case b = −1 is treated
similarly. Derivation of necessary estimates for the case |a| = 1, 0 ≤ |γ | < 1 is
contained in Example 4.8. The equality αφ = log2 ( ρ̂1 ) follows from Proposition 4.12.
The theorem is proved.
For n = 4, the mask of Eq. (4.78) is written in the form
1
15
m 0 (ω) = ck wk (ω). (4.107)
2 k=0
Here, the coefficients ck are defined as in the algorithm A0 with respect to the pa-
rameters
b0 = 1, bl = βl , b8 = 0, bl+8 = γl , 1 ≤ l ≤ 7, (4.108)
where |βl |2 + |γl |2 = 1. We give a complete list of sets of parameter values, for
which the mask (4.107) has blocking sets, and for each of these sets we indicate one
blocking set as given below:
(1) β1 = 0, M = [ 18 , 1),
(2) β6 = 0, M = [ 78 , 1),
(3) β2 = β3 = 0, M = [ 41 , 21 ) ∪ [ 58 , 1),
(4) β2 = γ5 = 0, M = [ 41 , 1) ∪ [ 58 , 34 ),
(5) β3 = β5 = 0, M = [ 83 , 21 ) ∪ [ 85 , 1).
Also note that for n = 4 vectors
e1 = (1, −1, 1, −1, 1, −1, 1, −1)t , e2 = (1, 1, −1, −1, −1, −1, 1, 1)t ,
e3 = (1, 1, 1, 1, −1, −1, −1, −1)t , e4 = (1, −1, 1, −1, −1, 1, −1, 1)t ,
e5 = (1, 1, −1, −1, 1, 1, −1, −1)t , e6 = (1, −1, −1, 1, 1, −1, −1, 1)t ,
e7 = (1, −1, −1, 1, 1, −1, 1, −1)t , (4.109)
for a basis of E 1 . We recall that T0 and T1 are matrices, determined by formulas

(4.84) under the condition (4.108). Let us introduce the matrices
⎛ ⎞ ⎛ ⎞
0 −β7 0 0 β4 γ4 0 0 −β7 0 0 β4 −γ4 0
⎜0 0 β3 γ3 0 0 0⎟ ⎜0 0 β3 −γ3 0 0 0⎟
⎜ ⎟ ⎜ ⎟
⎜γ1 0 0 0 0 0 0⎟ ⎜γ1 0 0 0 0 0 0⎟
⎜ ⎟ ⎜ ⎟
B0 = ⎜
⎜0 β7 0 0 β5 γ5 0⎟ ⎜
⎟ B1 = ⎜ 0 β7 0 0 β5 −γ5 0⎟⎟
⎜0 0 β2 γ2 0 0 0⎟ ⎜0 0 β2 −γ2 0 0 0⎟
⎜ ⎟ ⎜ ⎟
⎝ 0 β6 − β7 0 0 0 0 0⎠ ⎝ 0 β6 − β7 0 0 0 0 0⎠
0 2β7 0 0 0 0 0 0 2β7 0 0 0 0 0
and denote by Γ matrix of size 7 × 7, for which the first six columns are zero, and the
last column coincides with the vector (γ4 − γ7 , −γ3 , γ1 , γ5 + γ7 , −γ2 , γ6 − γ7 , γ7 )t .
Then, the linear transformations y = T0 x, y = T1 x, x ∈ E 1 have in the basis
(4.109) of matrices
1 1
A0 = B0 + Γ, A1 = B1 − Γ,
2 2
respectively. By Proposition 4.12, we have ρ̂(A0 , A1 ) = 2−αφ . We indicate three

cases when the values of αφ is calculated exactly.
Theorem 4.12 Let the function φ ∈ C A(4) 0 (R+ ) satisfy Eq. (4.78), in which n = 4
and the coefficients are determined from the parameters (4.108). Then
(1) αφ = ( 21 ) log2 ( |γ12 | ), if γ2 = 0, γ1 = γ5 = γ7 = 0;
(2) αφ = ( 21 ) log2 ( |β51γ2 | ), if β5 γ2 = 0, γ1 = γ3 = γ7 = 0;
(3) αφ = log2 ( |γ17 | ), if γ7 = 0,
and one of the following conditions is fulfilled (a) γ1 = γ2 = γ3 = 0, (b) γ1 = γ2 =
γ5 = 0, (c) β5 = γ1 = γ3 = 0.
Proof Since φ ∈ C A(4)

0 (R+ ), the following expansion takes place

1 x x
φ(x) = 1[0,1) 1+ dl wl , x ∈ R+ , (4.110)
8 8 l∈M
8
where the coefficients dl are defined as in Proposition 4.9.

10 : Let β1 = β5 = β7 = 1 and γ2 = 0. Then (4.110) is written as

8φ(x) = 1[0,1) (y) 1 + w1 (y) + β2 w2 (y) + β3 w3 (y) + β2 β4 w4 (y) + β2 w5 (y)
+ β3 β6 w6 (y) + β3 w7 (y) + β2 γ2 w10 (y) + β2 γ3 w11 (y) + β3 β6 γ4 w12 (y)
+ β3 γ6 w14 (y) + β2 β4 γ2 w20 (y) + β2 γ2 w21 (y) + β2 β6 γ3 w22 (y) + β2 γ3 w23 (y)
+ β3 γ4 γ6 w28 (y) + β2 γ22 w42 (y) + β2 γ2 γ3 w43 (y) + β2 β6 γ3 γ4 w44 (y)
+ β2 γ3 γ6 w46 (y) + β2 β4 γ22 w84 (y) + β2 γ22 w85 (y) + β2 β6 γ2 γ3 w86 (y)

+ β2 γ2 γ3 w87 (y) + · · · ,
where y = x8 . We put
G 1 := 2 + 2β2 + 2β3 + β2 β4 + β3 β6 + 2β2 γ2 + 2β2 γ3 + β3 β6 γ4

+ β3 γ6 + β2 β4 γ3 + β2 β6 γ3 + β3 γ4 γ6 .
Then
d2 2
8φ(0) = G 1 + 2γ2 + 2γ2 γ3 + β6 γ3 γ4 + γ3 γ6
1 − γ2

+ β4 γ22 + β6 γ2 γ3 + γ3 γ4 γ6 .
Moreover, for any odd s ≥ 5
1 β2 2 s−2
8φ( ) = G 1 + (γ + γ γ
2 3 + β γ γ
6 3 4 + γ γ
3 6 ) 1 − 2γ 2
2s 1 − γ2 2 2

+ (γ22 + β4 γ22 + β6 γ2 γ3 + γ2 γ3 + γ3 γ4 γ6 ,
and for any given s ≥ 6
1 β2 2
8φ( ) = G 1 + (γ2 + γ2 γ3 + β6 γ3 γ4 + γ3 γ6 )
2s 1 − γ2
s−4
+ (γ22 + β4 γ22 + β6 γ2 γ3 + γ2 γ3 + γ3 γ4 γ6 ) 1 − 2γ2 2 .
s
Therefore, |φ(0) − φ( 21s )| = C|γ2 | 2 , and hence αφ ≤ 21 log2 ( |γ12 | ). To prove the con-
verse inequality, we note that for the vectors (4.109) and for any i, j, k, l ∈ {0, 1}.
Ti e1 = Ti T j Tk Tl e2 = 0, Ti e4 = ±(γ3 e2 + γ2 e5 ) Ti T j = ±Ti e4 .
m
Therefore, Td1 Td2 · · · Tdm ei ≤ C|γ2 | 2 for i = 1, 2, 4, 5. Further, for all i, j, k, l ∈
{0, 1}, the equalities take place
Ti T j e6 = 0,
Ti e3 = β3 e2 + β2 e5 ,
Ti T j Tk Tl e4 = ±γ22 (β4 e1 + e4 ) + γ3 (Ti T j Tk e2 ± γ2 Ti e2 ),
1
Ti T j Tk e7 = ± Ti T j e4 .
2
Hence, for any d1 , d2 , . . . , dm ∈ {0, 1} and m ∈ N, we obtain the estimate
m
Td1 Td2 · · · Tdm | E1 ≤ C|γ2 | 2 ,
from which, Proposition 4.12, we deduce αφ ≥ 21 log2 ( |γ12 | ). Thus, if β1 = β5 = β7 =

1 and γ2 = 0, then αφ = 21 log2 ( |γ12 | ).
20 Let β1 = β3 = β7 = 1 and β5 γ2 = 0. Then, according to (4.110)

8φ(x) = 1[0,1) (y) 1 + w1 (y) + β2 w2 (y) + w3 (y) + β2 d4 w4 (y) + β2 β5 w5 (y)
+ β6 w6 (y) + w7 (y) + β2 β5 γ2 w10 (y) + β6 γ4 w12 (y) + β6 γ5 w13 (y) + γ6 w14 (y)
+ β2 β4 β5 γ2 w20 (y) + β2 β52 γ2 w21 (y) + β6 γ2 γ5 w26 (y) + γ4 γ6 w28 (y) + γ5 γ6 w29 (y)
+ β2 β52 γ22 w42 (y) + β4 β6 γ2 γ5 w52 (y) + β5 β6 γ2 γ5 w53 (y) + γ2 γ5 γ6 w58 (y)

+ β2 β4 β52 γ22 w84 (y) + · · · ,
where y = x8 . We set
G 2 := 4 + β2 + β2 β4 + β2 β5 + β6 + β2 β5 γ2 + β6 γ4 + β6 γ5 + γ6
+ β2 β4 β5 γ2 + β2 β52 γ2 + β6 γ2 γ5 + γ4 γ6 + γ5 γ6 .
Then
1 γ2
2γ
8φ( ) = G2 + β2 β52 γ2 + β2 γ4 β5 2 + β4 β6 γ5 + β5 β6 γ5
2 s 1 − β5 γ2

+ γ5 γ6 + β5 β6 γ2 γ5 + β4 γ5 γ6 + β5 γ5 γ6 .
For any odd s ≥ 5
1 γ2 s−3 s−3
8φ( ) = G2 + (β2 β52 γ2 + γ5 γ6 + β5 β6 γ2 γ5 )(1 − 2(β5 γ2 ) 2 + 2(β5 γ2 ) 2 )
2s 1 − β5 γ2
s−1
+ (β4 β6 γ5 + β5 β6 γ5 + β5 γ5 γ6 + β5 γ5 γ6 )(1 − 2(β5 γ2 ) 2 + β2 β4 β52 γ2 + β2 β52 γ2 ),
For any even s ≥ 6


1 γ2 s−2
8φ = G2 + β2 β52 γ2 + (β4 β6 γ5 + β5 β6 γ5 ) 1 − 2(β5 γ2 ) 2
2s 1 − β5 γ2
s−4

+ (β4 β2 β52 γ2 + β2 β52 γ2 + β4 γ5 γ6 + d5 γ5 γ6 ) 1 − 2(β5 γ2 ) 2
s−2 s
s−4

+ γ5 γ6 1 − 2(β5 γ2 ) 2 + 2(β5 γ2 ) 2 + β5 β6 γ2 γ5 1 − 2(β5 γ2 ) 2
s−2

+ 2(β5 γ2 ) 2 ) .
s
Therefore, |φ(0) − φ( 21s )| = C|β5 γ2 | 2 and αφ ≤ 1
2
log2 ( |β51γ2 | ). We now notice that
for vectors (4.109) and any i, j ∈ {0, 1}
Ti e1 = 0, Ti T j e5 = ±γ2 β5 e5 , Ti e4 = ±γ2 e5 .
m
Hence, for i = 1, 4, 5 estimate Td1 Td2 · · · Tdm ei ≤ C|β5 γ2 | 2 is correct. Moreover,
for all i, j, k, l ∈ {0, 1} equalities take place
Ti T j Tk Tl e2 = ±γ2 β5 Ti e4 ± (β6 − 1)γ2 γ5 Ti e5 + Ti T j Tk e7 ,

Ti e3 = e2 + β2 e5 ,
Ti T j e6 = ±γ5 Ti e4 ,
1
Ti T j Tk e7 = ± (±γ2 γ5 Ti e5 + (±γ5 γ6 − β5 )Ti e4 ).
2
Using these equalities, for any m ∈ N, we obtain estimate
m
Td1 Td2 · · · Tdm | E1 ≤ C|β5 γ2 | 2 .
Consequently, according to Proposition 4.12 the equality αφ ≥ 1

2
log2 ( |β51γ2 | ).
30 Let β1 = β2 = β3 = 1 and γ7 = 0. Then

8φ(x) = 1[0,1) (y) 1 + w1 (y) + w2 (y) + w3 (y) + β4 w4 (y) + β5 w5 (y) + β6 w6 (y)
+ β7 w7 (y) + β6 γ4 w12 (y) + β6 γ5 w13 (y) + β7 γ6 w14 (y) + β7 γ7 w15 (y)
+ β7 γ4 γ6 w28 (y) + β7 γ5 γ6 w29 (y) + β7 γ72 w31 (y) + β7 γ4 γ6 γ7 w60 (y)
+ β7 γ5 γ6 γ7 w61 (y) + β7 γ6 γ72 w62 (y) + β7 γ73 w63 (y) + β7 γ4 γ6 γ72 w124 (y)

+ β7 γ5 γ6 γ72 w125 (y) + · · · ,
G 3 := 4 + β4 + β5 + β6 + β6 γ4 + β6 γ5 .
Then
β7
8φ(0) = G 3 + (1 + γ6 (1 + γ4 + γ5 )).
1 − γ7
Moreover, for any integer s ≥ 2
1 β7
8φ( ) = G3 + (1 − 2γ7s + (1 − 2γ7s−2 )(β6 γ5 + γ4 γ6 ) + γ6 (1 − 2γ7s−1 )).
2s 1 − γ7
Therefore, |φ(0) − φ( 21s )| = C|γ7 |s . For all i, j, k ∈ {0, 1} equalities are valid
Ti T j Tk e1 = Ti T j Tk e4 = Ti T j Tk e5 = Ti T j Tk e6 = 0,
Ti T j Tk e2 = ±β7 γ7 Ti T j e7 ,
Ti T j Tk e3 = (β6 − β7 )(γ4 e1 + γ5 e4 ) + 2β7 Ti e7 ,
1
Ti T j Tk e7 = ±γ7 Ti T j e7 , Ti T j e7 = ± (±(γ6 − γ7 )(γ4 e1 + γ5 e4 ) + 2γ7 Ti e7 ).
2
Consequently, for any m ∈ N, we get
Td1 Td2 · · · Tdm | E1 ≤ C|γ7 |m .
Hence, as above we conclude αφ ≥ log2 ( |γ17 | ). Thus, if β1 = β2 = β3 = 1 and γ7 = 0,

then αφ = log2 ( |γ17 | ).
40 Let β1 = β2 = β5 = 1 and γ7 = 0. Then

8φ(x) = 1[0,1) (y) 1 + w1 (y) + w2 (y) + β3 w3 (y) + β4 w4 (y) + w5 (y) + β3 β6 w6 (y)
+ β3 β7 w7 (y) + γ3 w11 (y) + β3 β6 γ4 w12 (y) + β3 β7 γ6 w14 (y) + β3 β7 γ7 w15 (y)
+ β6 γ3 w22 (y) + β7 γ3 w23 (y) + β3 β7 γ4 γ6 w28 (y) + β3 β7 γ6 γ7 w30 (y)
+ β3 β7 γ72 w31 (y) + β6 γ3 γ4 w44 (y) + β7 γ3 γ6 w46 (y)

+ β7 γ3 γ7 w47 (y) + · · · ,
where y = x8 . Let
G 4 := 4 + β3 + β4 + β3 β6 + β3 β7 + γ3 + β3 β6 γ6 + β3 β7 γ7
+ β6 γ3 + β7 γ3 + β3 β7 γ4 γ6 + β3 β7 γ72 + β6 γ3 γ4
+ β7 γ3 γ6 + β7 γ3 γ7 + β3 β7 γ4 γ6 γ7 + β3 β7 γ6 γ72 .
Then
β7
8φ(0) = G 4 + (β3 γ72 + γ3 γ4 γ6 + γ3 γ6 γ7 + γ3 γ72 + d3 γ4 γ6 γ72 + β3 γ6 γ72 ).
1 − γ7
For any integer s ≥ 5

β7
8φ(B −s h [1] ) = G 4 + β3 γ72 (1 − 2γ7s−3 ) + γ3 (γ4 γ6 + γ6 γ7 + γ72 )
1 − γ7

(1 − 2γ7s−4 + 2γ7s−3 − 2γ7s−2 ) + d3 γ6 γ72 (γ4 + γ7 )(1 − 2γ7s−4 ) .
From here it is seen that

1
|φ(0) − φ( )| = C|γ7 |s ,
2s
and hence αφ ≤ log2 ( |γ17 | ). Further as above, for any i, j, k ∈ {0, 1}
Ti e1 = Ti T j e6 = 0,
Ti T j Tk e2 = ±β7 γ3 Ti e2 + 2β7 Ti T j e7 ,
Ti T j Tk e3 = β3 Ti T j e2 ± γ3 e2 , Ti e4 = Ti T j e5 = ±γ3 e2 ,
1
Ti T j e7 = ± (γ3 Ti e2 + γ7 Ti e4 ± (γ6 − γ7 )γ4 e1 + γ7 Ti e7 ).
2
Consequently, for any m ∈ N the estimate is correct and according to the Proposition
4.12, αφ ≥ log2 ( |γ17 | ). Thus, if β1 = β2 = β5 = 1 and γ7 = 0 then αφ = log2 ( |γ17 | ).
50 Let β1 = β3 = β5 = 1 and γ7 = 0. Then

8φ(x) = 1[0,1) (y) 1 + w1 (y) + β2 w2 (y) + w3 (y) + β2 β4 w4 (y) + β6 w6 (y) + β7 w7 (y)
+ β6 γ4 w12 (y) + β6 w13 (y) + β7 γ6 w14 (y) + β7 γ7 w15 (y) + β6 γ2 w26 (y)
+ β7 γ4 γ6 w28 (y) + β7 γ6 w29 (y) + β7 γ6 γ7 w30 (y) + β7 γ72 w31 (y)
+ β4 d6 γ2 w52 (y) + β7 γ2 γ6 w58 (y) + β7 γ4 γ6 γ7 w60 (y)

+ β7 γ6 γ7 w61 (y) + β7 γ6 γ72 w62 (y) + · · · ,
G 5 := 3 + β2 + β2 β4 + 2β6 + β7 + β6 γ4 + 2β7 γ6 + β7 γ7
+ β6 γ2 + β7 γ4 γ6 + β7 γ6 γ7 + 2β7 γ72 + β4 β6 γ2
+ β2 γ2 γ6 + β7 γ4 γ6 γ7 + β7 γ6 γ72 .
Then
β7
8φ(0) = G 5 + (γ 3 + β4 γ2 γ6 + γ2 γ6 γ7 + γ4 γ6 γ72 + γ6 γ72 + γ6 γ73 ).
1 − γ7 7
For any s ≥ 5, we have

β7 3
8φ(B −s h [1] ) = G 5 + γ7 1 − γ7s−3 + γ6 (β4 γ2 + γ2 γ7
1 − γ7

+ γ4 γ72 + γ72 + γ73 ) 1 − 2γ7s−4
and estimate |φ(0) − φ( 21s )| = C|γ7 |s is correct. Hence, αφ ≤ log2 ( |γ17 | ). Further,
Ti e1 = Ti T j Tk e4 = Ti T j e5 = Ti T j Tk Tl e6 = 0,
Ti e2 = −β7 e1 + β7 e4 + (β6 − β7 )e6 + 2β7 ,
Ti e3 = e2 + β2 e6 ,
1
Ti e7 = ± ((γ4 − γ7 )e1 − (γ7 + 1)e4 − γ2 e5 + (γ6 − γ7 )e6 + γ7 e7 ).
2
Therefore, for any m ∈ N inequality is valid.
Td1 Td1 · · · Tdm | E1 ≤ C|γ7 |m .
Using Proposition 4.12, from this we derive the estimate αφ ≥ log2 ( |γ17 | ). Thus, if
γ5 = β1 = β3 = 1 and γ7 = 0 then αφ = log2 ( |γ17 | ). The theorem is proved.
Recall that a binary-integer function of order r is a function f , defined on R+ and

a constant on dyadic intervals of rank r . The space of such functions is denoted
by
Er = Er (R+ ), and the set of all dyadic-integer functions is denoted by E = r Er .
Walsh polynomial of order 2r − 1 belongs to the set Er . In the notations of Theorem
4.9, we indicate three cases, when ρ̂(A0 , A1 ) = 0 and the solution of Eq. (4.78)
for n = 4 is a binary-integer function. Suppose that x ∈ [0, 8) and y = x8 . Then,
assuming
W (y) := w1 (y) + β2 w2 (y) + β3 w3 (y) + β2 β4 w4 (y) + β3 β6 w6 (y) + β3 β7 w7 (y).
We obtain form (4.110), the following three expansions:

1. If γ1 = γ2 = γ3 = γ7 = 0, then

8φ(x) = 1 + β1 W (y) + β2 β5 w5 (y) + β3 β6 γ4 w12 (y) + β3 β6 γ5 w13 (y)

+ β3 β7 γ6 w14 (y) + β3 β7 γ4 γ6 w28 (y) + β3 β7 γ5 γ6 w29 (y) .
2. If γ1 = γ2 = γ5 = γ7 = 0, then

8φ(x) = 1 + β1 W (y) + β2 β5 w5 (y) + β2 β5 γ3 w11 (y) + β3 β6 γ4 w12 (y)
+ β3 β7 γ6 w14 (y) + β2 β5 β6 γ3 w22 (y) + β2 β5 β7 γ4 γ6 w28 (y)
+ β2 β5 β6 γ3 γ4 w44 (y) + β2 β5 β7 γ3 γ6 w46 (y)

+ β2 β5 β7 γ3 γ4 γ6 w92 (y) .
3. If β5 = γ1 = γ3 = γ7 = 0, then

8φ(x) = 1 + β1 W (y) + β3 β6 γ4 w12 (y) + β3 β7 γ6 w14 (y) + d3 d6 γ2 γ5 w23 (y)
+ β3 β7 γ4 γ6 w28 (y) + β3 β7 γ5 γ6 w29 (y) + β3 β4 β6 γ2 γ5 w52 (y)

+ β3 β7 γ2 γ6 w58 (y) + β3 β4 β7 γ2 γ5 γ6 w116 (y) .
In Example 4.4, for an arbitrary n ≥ 3, there is a dyadic entire function of Eq. (4.78),
generating MRA in L 2 (R+ ). The following theorem gives the conditions for the
arrangement of the zeros of the mask of the scaling equation (4.78), for which its
solution φ for a given natural number n is a dyadic entire function. Binary-rational
numbers will sometimes be written in the binary system (for example, 0.01 = 41 ).
Theorem 4.13 Let φ be a finite L 2 solution of a scaling equation (4.78) and m 0 is

its mask. Suppose that φ̂(0) = 1. The function φ is not a dyadic entire if and only
if there exists a finite sequence {dk }k=1
N
, N ≤ j ≤ 2n−1 + n − 1, dk ∈ {0, 1}}, such
that
(a) d1 = · · · = dn−1 = 0, dn = 1;
(b) There exists an integer j, n − 1 ≤ j ≤ N − 1, such, that d j−s = d N −s for s =
0, . . . , n − 2;
(c) m 0 (0, ·dk+1 · · · dk+n ) = 0 for k = 0, . . . , N − n.
Proof Let {dk }k=1

N
be a sequence, satisfying conditions (a) − (c), and let
β = 0 · d1 · · · d j−n+1 (d j−n+2 · · · d N −n+1 )
is a binary-rational number with period (d j−n+2 · · · d N −n+1 ). Since m 0 ∈ En and

m 0 (0) = 1, then for any r ∈ N.
∞

φ̂(2r β) = m 0 (2r −k β)
k=1
∞

= m 0 (d1 · · · dr −k , dr −k+1 · · · dr −k+n · · · )
k=1
∞
= m 0 (0 · dr −k+1 · · · dr −k+n ) = 0.
k=1
(Assume di = 0 for i ≤ 0, use the periodicity m 0 and the fact that each piece of
length n of a binary entry of number β coincides with some piece of length n of
the sequence {dk }k=1
N
.). Hence, the Walsh–Fourier transform carrier φ̂ is not compact
and, therefore, by Proposition 4.5, the function φ is not binary integer. Conversely,
suppose that a finite L 2 -solution φ of Eq. (4.78) is not a dyadic entire function. Then,
φ ∈ L 1 (R+ ) and by Proposition 4.5 the support supp φ̂ is not compact. Let us take
n−1
the number ω > 22 such that φ̂(ω) = 0. The binary expansion of this number has
the form

l ∞

ω= ω− j−1 2 j + ω j 2 j , ω j ∈ {0, 1}, ω−l−1 = 1, l > 2n−1 .
j=0 j=1
It can be seen from 4.5 that m 0 (2−s ω) for all s ∈ N. Hence, as above using the fact
that m 0 ∈ En , we get that at points
γl = 0 · 0 · · · 0 · ω−l−1 , γl−1 = 0 · 0 · · · ω−l ω−l−1 , γ0 = 0 · ω−n · · · ω−2 ω−1
the mask m 0 does not vanish. In the binary notation of each binary-rational number
γs (s = 0, . . . , l) there are exactly n significant digits. Since l > 2n−1 among the
numbers γ0 , . . . , γl there are two, for which the first n − 1 digits coincide. Let r be
the smallest number such that for some k ∈ N, the numbers γr and γr −k have the
first n − 1 digits. Then, the sequence 0, . . . , 0, ω−l−1 , . . . , ω−r is required and the
theorem is proved.
Corollary 4.3 If the mask m 0 of the scaling equation (4.78) is equal to zero on the
interval [2−r , 2−r +1 ) for some r ∈ N, then the solution φ of this equation is a dyadic
entire function.
Proof Since on the intervals [k, k + 2−n ) the mask m 0 is equal to 1, then 1 ≤ r ≤ n.
Then, m 0 (ω) = 0 for any number ω, the binary record of which begins with the
digits 0.0...01 (before the unit of r zeroes). It immediately follows that for this mask
there is no sequence {dk }k=1
N
satisfying the conditions (a) − −(c) of Theorem 4.13.
In particular, if the mask is equal to zero in the interval [ 21 , 1), then the solution φ of
Eq. (4.78) is a binary integer. Note also that if

1, ω ∈ [0, 2−r ),
m 0 (ω) =
0, ω ∈ [2−r , 2−r +1 ),
where r ∈ N, r ≤ n, then irrespective of the values taken by the mask m 0 on the

remaining dyadic intervals, we have φ(x) = 21−r 1[0,2r −1 ) (x). For small n, the ver-
ification using Theorem 4.13, whether the solution of the scaling equation (4.78)
is binary integer, reduces to a search of all sequences of zeros and units of length
2n−1 − n + 1.
Example 4.9 For the case n = 3, the following step functions are solutions of Eq.
(4.78):
(1) φ(x) = ( 41 )1[0,1) ( x4 ) (b1 = 0),

(2) φ(x) = ( 41 )1[0,1) ( x4 )(1 + b1 w1 ( x4 )) (b2 = b3 = 0),
(3) φ(x) = ( 41 )1[0,1) ( x4 )(1 + b1 w1 ( x4 ) + b1 b3 w3 ( x4 )) (b2 = b6 = b7 = 0),
(4) φ(x) = ( 41 )1[0,1) ( x4 )(1 + b1 w1 ( x4 ) + b1 b2 w2 ( x4 )) (b3 = b4 = b5 = 0),
(5) φ(x) = ( 41 )1[0,1) ( x4 )(1 + b1 w1 ( x4 ) + b1 b2 w2 ( x4 ) + b1 b3 w3 ( x4 ) + b1 b3 b6 w6 ( x4 ))
(b4 = b5 = b7 = 0).
It is well known that for orthogonal wavelets with compact support on a line the
smoothness cannot be infinite, it is always bounded by the length of the support of
the scaling function. The exact analogue of this statement for dyadic wavelets is not
true, if only because dyadic scaling functions can be dyadic entire finite functions.
However, it turns out that for dyadic scaling functions there is an alternative, either
the function is dyadic entire or it has finite smoothness. More precisely, by Theorem
4.14, if the solution φ of Eq. (4.78) is not binary integer, then its smoothness is finite.
Moreover, the Holder exponent αφ is estimated above with the help of the smallest
nonzero value of the mask module, and this estimate is unimprovable. Since the mask
module |m 0 (ω)| is a piecewise constant function with step 2−n , it can take not more
than 2n different values. One of these values is equal to zero, because of the modified
Strang–Fix condition (Exercise 4.2.19), the mask must convert to zero on the period.
Let am be the least positive value of the module of the mask m 0 , i.e.,
am = min{|m 0 (ω)| : ω ∈ R+ , m 0 (ω) = 0}.
Since the mask has period 1 and is constant at dyadic intervals of rank n. Then, the
value of am coincides with the minimal positive number from the set {|m 0 (2−n k)|, k =
0, . . . , 2n−1 }.
Theorem 4.14 For any function φ ∈ L 2c (R+ ), that is a solution of Eq. (4.78) and
satisfying the condition φ̂(0) = 1, the following alternative takes place:
(1) either φ is binary integer;
(2) either am < 1 and αφ ≤ − log2 am .
Proof Suppose that φ is not a binary-integer function. Then, by Proposition 4.5, the
support of the function φ̂ is not compact. It means that there exists arbitrarily large ω
for which φ̂(ω) = 0. Having determined the natural number k from the inequalities
2k ≤ ω ≤ 2k+1 , and using the fact that m 0 (2− j ω) = 1 for all j ≥ n + k + 1, we
obtain
∞
n+k
φ̂(ω) = m 0 (2− j ω) = m 0 (2− j ω). (4.111)
j=1 j=1
Since φ̂(ω) then for each j = 1, . . . , n + k value of |m 0 (2− j ω)| is not equal to zero,
and therefore not less than am . Having substituted the inequality |m 0 (2− j ω)| ≥ am
in (4.111), we obtain
|φ̂(ω)| ≥ (am )n+k . (4.112)
If we now assume that am ≥ 1, then, by (4.112), there exists arbitrarily large ω

for which |φ̂(ω)|2 ≥ 1. But, then, since the function |φ̂(ω)|2 is constant on binary
intervals of rank n − 1, it is not summable on the half line R+ , and hence φ̂(ω) ∈ /
L 2 (R+ ). Consequently, φ ∈/ L 2 (R+ ) which contradicts the condition. Thus, am < 1.
Substituting in (4.112), the inequality k ≤ log2 ω, we have |φ̂(ω)| ≥ ω−(alogm 2) am . At
n
the same time, according to Shipp Wade Simon [5, Exercise 9.12], for any ω > 0
the inequality |φ̂(ω)| ≤ 21 ω(φ, ω2 ) is satisfied. Consequently, there exists a constant
C > 0 such that |φ̂(ω)| ≤ ωCαφ for all ω ≤ 1. Therefore, αφ ≤ − log2 am . The theorem
is proved.
Remark 4.2 There exist infinitely smooth functions in R+ that are not dyadic integer.
However, as from Theorem 4.14, this is impossible for scaling functions. The estimate
αφ ≤ − log2 am is unimprovable for scaling functions. For example, for the Lang
function (Example 4.2), it is achieved. In contrast to the scaling functions on the line
R where for each n there exists only one scaling function of maximum smoothness
(namely, a cardinal B-spline of order n − 1), the smoothness of the solutions of
Eq. (4.78) for a given n is unbounded. However, it is limited for non-dyadic entire
functions for a given value of am . This estimate is attained for the data n and am on
the set of finite scaling dyadic functions.
Thus, in constructing dyadic wavelets, we can check for a given mask m 0 in finite
time whether the wavelets are dyadic entire and if not, then obtain a boundary from
above for smoothness. Another useful consequence of Theorem 4.14 concerns the
case where the mask module can take only two values: 0 and 1. This includes, in
particular, the case when the mask is a characteristic function of the set obtained by
combining some family of dyadic intervals of rank n.
Corollary 4.4 If the mask module of the scaling equation (4.78) assumes only the
values 0 and 1, then any L 2 solution of this equation is dyadic entire.
Theorem 4.10 provides an algorithm for the finite verification of the scaling function
for belonging to the class of dyadic entire function. For masks whose module takes
only two values, according to Corollary 4.4, this algorithm yields a criterion for the
solvability of the scaling equation in L 2 (R+ ).
In conclusion of the section, applying Theorem 4.10, we obtain
Corollary 4.5 If the function φ generates an MRA in L 2 (R+ ), and the mask module
of the corresponding scaling equation takes only the values 0 and 1, then this function
and the dyadic wavelets generated by it are binary integers.
4.5 Approximation Properties of Dyadic Wavelets
In this section, we describe the approximation properties of dyadic wavelets in the

Sobolev and Hölder spaces on the half line R+ . We first consider approximations of
functions that are smooth in the usual term, i.e., in the Euclidean metric on R+ , and
4.5 Approximation Properties of Dyadic Wavelets 159
then the approximations of W -smooth functions are studied. Since we study bursts
with a compact support, it is quite natural to restrict ourselves to approximating
the finite functions f . It will be established that (in contrast to classical wavelets
on the straight line) all systems of dyadic wavelets on R+ have the same orders of
approximation. Namely, in Sobolev spaces W1l , the order of approximation by any
system of dyadic wavelets is equal to 1, while in the class of W -smooth functions the
order of approximation is +∞. Therefore, to compare the approximation properties,
a fine gage—the defect of the approximation is introduced and an explicit formula
is obtained for it, which has an interesting analogy with the formula for the order of
approximation by classical wavelets. It is known (see, for example, Novikov et al. [1],
Theorems 3.2.3 and 3.4.16]) that the order of approximation of a system of wavelets
with compact support on a line is equal to the multiplicity of the zero of the mask
(trigonometric polynomial) m(ξ ) at point ξ = 21 . For dyadic wavelets, the mask is a
Walsh polynomial, i.e., periodic function, piecewise constant on dyadic intervals of
the given order n. It is shown that the approximation error is determined by the largest
integer k, for which the mask vanishes on the half-interval [ 21 , 21 + 2k−n ). This result
allows us to characterize dyadic wavelets with the best approximation properties.
Next, we investigate the relationship of the approximation properties of wavelets
with their smoothness (or, something is the same, with the smoothness of the scaling
function). It is proved that the wavelets with the smallest error of approximation have
a maximum smoothness, i.e., binary-integer functions.
Recall that for a positive integer l, the Sobolev space W2l (R+ ) of all functions
f , for which the derivative f (l−1) is absolutely continuous and f (l) ∈ L 2 (R+ ). For
l = 0, we set W20 (R+ ) = L 2 (R+ ). The space of algebraic polynomials of order not
above r is denoted by Pr . We will assume that the polynomials are defined on R+ .
In particular, P0 is the space of identity constants on R+ .
For any function φ ∈ L 2c (R+ ), the family of subspaces {V j } is defined as
V j := span{φ j,k : k ∈ Z+ }, j ∈ Z,
where φ j,k = φ(2 j x ⊕ k), x ∈ R+ . We note that V j

= Ṽ j ∩ L 2c (R+ ), where Ṽ j is
the set of all possible linear combinations of the form k∈Z+ ak φ(2 j x ⊕ k), ak ∈ C.
(The sums are finite since φ has a compact support.) The distance from an arbitrary
function f ∈ L 2 (R+ ) to the subspace V j is a quantity given by
dist ( f, V j ) := inf f − g .
g∈V j
It is clear that the rate of decrease of dist ( f, V j ) depends on the approximated

function f , in particular, on its smoothness (by virtue of the inverse theorems of
approximation theory). Therefore, it is natural to put the problem for certain classes
of functions. For approximations in the class of continuously differentiable functions
C l (R), an analogous problem was first investigated by I. Schoenberg [Schoenberq]
for the spaces V j , generated by the functions φ(2 j − k), k ∈ Z, where φ is an arbi-
trary continuous finite function that does not necessarily satisfy the scaling equation.
I. Schonberg [10] established the connection between the maximum degree of alge-
braic polynomials generated by integer-valued shifts of the function f , and the order
of decreasing distance from an arbitrary function f ∈ C l (R) to the spaces V j . In
terms of the Fourier transform of the function, this connection was found explicitly
by G. Strang and G. Fix in the monograph [6]. For the Strang–Fix conditions in the
theory of wavelets, see, for example, Sect. 3.3. Books (Novikov et al. [1]).
In the case, when {ψ j,k } is the system of wavelets defined by the algorithm A0
of Sect. 4.2 and {V j } corresponding to the MRA, for any function f ∈ L 2 (R+ ) the
dist ( f, V j ) = f − P j f is true, where P j f is the orthogonal projection of f to
V j . The expansion of the projection P j f over the system {ψ j,k } has the form

Pj f = cr,k ψr,k , cr,k = f, ψr,k .
k∈Z,r < j
Therefore, the approximation properties of the

wavelet system also determine the
rate of decrease of the “tail” of the expansion k∈Z,r ≥ j |cr,k |2 or, in fact, the rate of
decrease of the coefficients c j,k for j → ∞. The smoothness index of the function
f ∈ L 2 (R+ ) is denoted by S f and is defined as
S f =:= k + sup{S : f (k) (· + h) − f (k) (·) ≤ Ch S },
where k is the largest integer for which f ∈ W2k (R+ ).

For an arbitrary function φ ∈ L 2c (R+ ), the order of approximation νφ is the exact
upper bound of the set of numbers γ ≥ 0 such that
dist ( f, V j ) ≤ C( f, γ )2− jγ , j ∈ Z, (4.113)
for any function f ∈ L 2 (R+ ), for which S f > γ . It is seen that the order of approx-
imation νφ will be larger if the distance from any sufficiently smooth finite function
f to the space V j is equal to O(2− j γ ) for j → ∞. According to the following
proposition, the order of approximation per unit exceeds the largest degree of the
space of polynomials, which the function φ generates with its entire shifts.
Proposition 4.13 For any function φ ∈ L 2c (R+ ), we have

1 + max{r : Pr ⊂ Ṽ0 }, if P0 ⊂ Ṽ0 ,
νφ =
0, if P0 Ṽ0 .
Proof Let supp φ ⊂ [0, 2n−1 ), where n ∈ N. Let us assume that Pr ⊂ Ṽ0 . Then,
at the interval [0, 2n−1 ) any degree polynomial not higher than n is generated by
functions φ(· ⊕ k), k = 0, 1, . . . , 2n−1 . From here it follows that V j contains all
functions g of the following form: on a given dyadic interval of length 2n−1− j , the
function g coincides with a given polynomial of degree not higher than r , and outside
of this interval is equal to zero. Hence, the space V j approximates an arbitrary finite
function f ∈ W2r +1 (R+ ) not worse than a space of splines of order r with nodes on
a uniform grid with step h = 2n−1− j . It is well known that the distance from f to
this space of splines does not exceed C( f, r )h r +1 . Consequently,
dist ( f, V j ) ≤ [2(n−1)(r +1) C( f, r )]2− j (r +1) , j ∈ Z.
In this way, if Pr ⊂ Ṽ0 , then νφ ≥ r + 1. It is assumed now that P0 Ṽ0 . Let us

take the smallest q ∈ Z+ for which x q ∈ / Ṽ0 . By assumption, q ≤ r . The linear span
of functions φ(· ⊕ k), k = 0, 1, . . . , 2n−1 − 1 does not contain the restriction x q to
the interval [0, 2n−1 ). Let a = dist (x q , L n−1 ) in the space L 2 [0, 2n−1 ). By virtue of
finite dimension of space L n−1 we have a > 0. Hence, the distance from the function
x q , bounded to any interval till the space V0 is not less than a. We apply the conversion
f (·) → 2− jq f (2 j ·), which reduces the L 2 -norm in 2 j (q+ 2 ) , transforms the space V0
1
into V j , and leaves the function x q in place. We had obtained that the distance from
the function x q bounded to any dyadic interval Is(k) of order k = n − 1 − j, till the
space V j is not less than 2 j (q+ 2 ) . Hence, for any function g ∈ V j , we have
1
2n−1
|x q − g(x)|2 ≥ 2− j (2q+1) a 2 .
0
Adding this inequality over all 2 j intervals of order n − 1 − j, joining the interval
[0, 2n−1 ), we obtain
2n−1
|x q − g(x)|2 ≥ 2 j · 2− j (2q+1) a 2 = 2−2q j a 2 ≥ 2−2r j a 2 .
0
In this way, the distance from the function x q which is bounded on the interval
[0, 2n−1 ) till the space V j is not less than 2−r j a. Now taking an arbitrary function
f ∈ W2r +1 (R+ ) that coincides with the polynomial x q on the interval [0, 2n−1 ), we
find that dist ( f, V j ) ≥ 2−r j a. In this way, if Pr Ṽ0 , then νφ ≤ r . The proposition
is proved.
Corollary 4.6 For any function φ ∈ L 2c (R+ ), the order of approximation νφ is finite
and is a nonnegative integer.
We show that the order of approximation of an arbitrary dyadic scaling function with
compact support is equal to 1. For this, we apply the results to a scaling function φ
satisfying the equation
2
n
−1
φ(x) = ck φ(2x ⊕ k), x ∈ R+ . (4.114)
k=0
We recall that the mask of this scaling equation is the following Walsh polynomial
2 −1
1
n
m 0 (ω) = ck wk (ω). (4.115)

2 k=0
Theorem 4.15 For any scaling function φ ∈ L 2c (R+ ), satisfying Eq. (4.114), the
equality νφ = 1 is true.
Proof Since the carrier of the L 2 function φ is determined with an accuracy till the set
of measure zero, as per Theorem 4.4, we have suppφ ⊂ [0, 2n−1 ) and the partition
of unity property is fulfilled:

φ(x ⊕ k) = 1, f or a.e. x. (4.116)
k∈Z+
Consequently, P0 ⊂ Ṽ0 and with the help of the proposition 1, we obtain νφ ≥

1. Further, if the function f (x) = x belongs to the space Ṽ0 , then the coefficients
a0 , a1 , . . . , a2n−1 −1 exist for which
−1
2n−1
ak φ(x ⊕ k) = x f or a.e. x.
k=0
Substituting x ⊕ k for x in this equation and adding over all l = 0, 1, . . . , 2n−1 − 1,

we obtain
−1 2 −1
2n−1 −1
n−1
2n−1
ak φ((x ⊕ l) ⊕ k) = (x ⊕ l).
l=0 k=0 l=0
On the other hand, with the help of (4.116), we have
−1 2 −1
2n−1 −1 −1 −1
n−1
2n−1 2n−1 2n−1
ak φ((x ⊕ l) ⊕ k) = ak φ((x ⊕ l) ⊕ k) = ak .
l=0 k=0 k=0 l=0 k=0
Consequently,
−1
2n−1 −1
2n−1
ak = (x ⊕ l) f or a.e. x. (4.117)
k=0 l=0
Note, that if x ∈ [0, 1) then x ⊕ l = x + l. Therefore, for x ∈ [0, 1) Eq. (4.117) takes
the form
−1
2n−1 −1
2n−1
ak = (x + l) = 2n−1 x + 2n−2 (2n−1 − 1),
k=0 k=0
which is impossible in view of the arbitrariness of x ∈ [0, 1). From here it follows
/ Ṽ0 . In this way, Ṽ0 does not hold the set P1 . As per Proposition
that f (x) = x ∈
4.13, we obtain νφ ≤ 1, which completes the proof of the theorem.
It is clear from Theorem 4.15 that for approximation function of the Sobolev class
W2k (R+ ), all the wavelets determined by the algorithm A0 of Sect. 4.2 are equivalent.
Partially, this phenomenon can be explained by the fact that dyadic wavelets are
equal as well as the classical Walsh system is intended for approximation of not
smooth (in the general sense) but dyadic-smooth functions. The remaining part of the
present section is devoted to such approximations. For a scaling function φ ∈ L 2c (R+ )
satisfying Eq. (4.114), the dyadic module of continuity is defined in Sect. 4.3. In this
section, for each function f ∈ L 2 (R+ ), we set
ω2 ( f ; h) := sup{ f (· ⊕ x) − f (·) : 0 ≤ x < δ}, h > 0,
and we will estimate the smoothness of function f by H0̈lder indicator
α f := sup{α ≥ 0 : ω2 ( f ; h) ≤ Ch α }.
We note that, in the given case, the Holder indicator of function, which is different
from the constant can be greater than 1. The binary order of approximation by a
function φ ∈ L 2c (R+ ) is named as the exact upper bound νφ of numbers γ ≥ 0, such
that (4.113) is fulfilled for all f ∈ L 2c (R+ ) for which α f > γ .
Proposition 4.14 For any function φ ∈ L 2c (R+ ), there is an alternative, either νφ =
0, or νφ = +∞. In this case, νφ = +∞ if and only if

φ(x ⊕ k) = c, (4.118)
k∈Z+
for some constants c = 0.

Proof We choose a positive integer n from the condition supp φ ⊂ [0, 2n−1 ). If

n−1
(4.118) is satisfied, then 2k=0 −1 φ(x ⊕ k) = c, for x ∈ [0, 2n−1 ). Consequently,
1[0,2n−1 ) ∈ V0 from where 1[0,2n−1 ) ∈ Vn−1 . Hence, the space Ṽn−1 contains all dyadic
entire function of order 0. It follows from here that E0 ⊂ Ṽn−1 , and therefore Er ⊂
Ṽr +n−1 for any r ∈ Z+ . It is known [Golubov et al., Sect. 10.5.2] that for an arbitrary
function f ∈ L 1 (R+ ) ∩ L 2 (R+ ) we have the inequalities
1
ω2 ( f ; 2−r ) ≤ dist ( f, Er ) ≤ ω2 ( f ; 2−r ). (4.119)
2
Consequently, dist ( f ; Ṽr +n−1 ) ≤ dist ( f, Er ) ≤ ω2 ( f ; 2−r ). Assuming j=r +n−1

and noting that dist ( f ; Ṽr +n−1 ) = dist ( f ; Vr +n−1 ), such that f is finite, we have
dist ( f ; V j ) ≤ ω2 ( f ; 2− j+n−1 ). (4.120)

If α f ≥ γ , then ω2 ( f ; 2− j+n−1 ) ≤ C( f )2γ (− j+n−1) , from where
dist ( f ; V j ) ≤ 2(n−1)γ C( f )2− jγ .
This inequality is true for any γ > 0 and f ∈ lc2 (R+ ), under the condition that α f >
γ . By virtue of arbitrary γ , we obtain νφ = +∞. Now assume that the function φ
does not have the property (4.118). Then, for any coefficients ak , 0 ≤ k ≤ 2n−1 − 1,
we have
−1
2n−1

1[0,2n−1 ) (·) − φ(· ⊕ k) > 0.

k=0
In fact, if for some coefficients {ak } the equality is completed
−1
2n−1
φ(x ⊕ k) = 1[0,2n−1 −1) (x),
k=0
then, by applying the Walsh–Fourier conversion, we obtain a(ω)φ̂(ω) = 1[0,2n−1 −1) (ω),

n−1
where a(ω) = 2k=0 −1 wk (ω). Consequently, a(0)φ̂(0) = 1 and a(k)φ̂(k) = 0 for
all k ∈ N. It follows from the first equality that φ̂(0) = 0 and a(0) = 0. Then, for
any k ∈ N we have a(k) = a(0) = 0. Now by applying the second
equality, we get
φ̂(k) = 0 for all k ∈ N. Putting a 1-periodic function Φ(x) := k∈Z+ φ(x ⊕ k), in
the Walsh–Fourier series, we obtain

Φ(x) = φ̂(k)wk (x) = φ̂(0)w0 (x) = φ̂(0), f or x ∈ R+ .
k∈Z+
Consequently, the condition (4.118) is completed with constant c = φ̂(0), which

contradicts the proposition. Therefore, the continuous function

−1
2n−1

{ak }2k=0 −1
n−1

→ 1[0,2n−1 ) (·) − φ(· ⊕ k)

k=0
takes only positive values, from which (by finite dimensionality) it follows

−1
2n−1

inf 1[0,2 )
n−1 (·) − φ(· ⊕ k) =: q > 0.

{ak }
k=0
In this way, dist (1[0,2n−1 ) ; V0 ) = q > 0. Then, for any binary interval Is(k) of order
k = n − 1 − j, we have
j
dist (1 Is(k) , V0 ) ≥ 2− 2 q.
Consequently, for any function g ∈ V j , we have

|g(x) − 1|2 ≥ 2− j q 2 .
Is(k)
Adding these inequalities over all 2 j intervals Is(k) , constituting the interval [0, 2n−1 ),
we obtain q − 1[0,2n−1 ) > q. Thus, dist (1[0,2n−1 ) ; V j ) ≥ q. Thus, the characteristic
function of the interval [0, 2n−1 ) cannot be approximated with any accuracy by the
spaces V j . But since the Hölder indicator α of the function 1[0,2n−1 ) is equal to +∞,
we have νφ = 0, which completes the proof.
The following theorem is the direct consequence of Theorem 4.4 and Proposition
4.14.
Theorem 4.16 For any scaling function φ ∈ L 2c (R+ ), satisfying Eq. (4.114), is true
to the equation νφ = +∞.
Thus, the order of approximation by any system of dyadic wavelets is equal to +∞.
Therefore, the dyadic wavelets come closer to the W -smooth functions better than
any classical system of wavelets. On the other hand, Theorem 4.16 shows that the
binary order of approximation νφ has a very rough characteristic. A more detailed
analysis proposes the constant C( f, γ ) in inequality (4.113) and its growth depending
on γ . Taking into account the estimate (4.118), we introduce the following notion.
Let for the function νφ = +∞, the equation φ ∈ L 2c (R+ ) is fulfilled. By the defect
of approximation bφ , function φ is named the smallest integer number b, such that
for all f ∈ L 2c (R+ ) the following inequality takes place
dist ( f, V j ) ≥ ω2 ( f ; 2− j+b ). (4.121)
The correctness of the definition follows from Proposition 4.12, but first we will
prove the following lemma.
Lemma 4.1 For any l ∈ Z+ , the following properties of the function φ ∈ L 2c (R+ )
are equivalent:
(a) 1[0,2l ) ∈ Ṽ0 ;
(b) E−1 ⊂ Ṽ0 ;
(c) The function φ̂ at every interval [k, k + 2−l ) is identically equal to zero and it
does not vanish at the interval [0, 2−l ).
Proof The equivalence of the properties (a) and (b) are evident. We will prove that
the property (a) is equal to (c). Let supp φ ⊂ [0, 2n−1 ). If
−1
2n−1
φ(x ⊕ k) = 1[0,2l ) (x),
k=0
then, applying the conversion of Walsh–Fourier, we obtain
a(ω)φ̂(ω) = 1[0,2l ) (ω), (4.122)

n−1
where a(ω) = 2k=0 −1 wk (ω). It follows that the function a(ω) is not changed to
zero at the interval [0, 2−l ), and this means that, at all intervals [k, k + 2−l ), k ∈ N.
That is why, the function φ̂(ω) at these intervals is equal to zero. Conversely, if (c)
holds, then, putting
1
, ω ∈ [0, 2−l ),
a(ω) = φ̂(ω)
0, ω∈/ [0, 2−l ),
and continuing with the periodicity on R+ , we obtain (14), which implies (a). The
lemma is proved.
Proposition 4.15 For any function φ ∈ L 2 (R+ ), satisfying (4.118), the defect of
approximation bφ is equal to the minimal number l for which conditions (a)–(c) of
Lemma 4.1 are satisfied.
Proof If E−1 ⊂ Ṽ0 for a given l, i.e., satisfying the condition (b), then E j−1 ⊂ Ṽ j for
any j. Now, applying (4.119) in any function f ∈ L 2c (R+ ), we obtain dist ( f, V j ) ≥
ω2 ( f ; 2− j+1 ). Consequently, for b = l the inequality (4.121) is satisfied. Conversely,
if property (a) is not satisfied for the number l − 1, that is, 1[0,2l−1 ) ∈
/ Ṽ0 , then, as it
is in the proof of Proposition 4.14, we conclude that dist ( f 0 , Vl−1 ) ≥ q > 0, where
f 0 = 1[0,1) . Since ω2 ( f 0 ; 2−l+1+b ) for any b < l − 1, then the inequality (4.121) is
not completed when j = l − 1. Therefore b ≥ l − 1. The proposition is proved.
Corollary 4.7 For any function φ ∈ L 2c (R+ ), which satisfies the condition (4.118),
we have 0 ≤ b ≤ n − 1, where n is the smallest natural number for which the
suppφ ⊂ [0, 2n−1 ).
It can be seen from Proposition 4.12 that for a scaling function φ, satisfying Eq.
(4.114), the value of bφ is entirely determined by the set of zeros of the function φ̂,
and hence by the zero set of the mask m 0 (ω), given by formula (4.115). A scaling
function φ is named as stable, if the system of entire shifts of this function forms a
Riesz basis in L 2 (R+ ), as per Theorem 4.6, this is equivalent to linear independence
in the shift system {φ(· ⊕ k) : k ∈ Z+ } in L 2 (R+ ).
Lemma 4.2 A stable scaling function φ ∈ L 2c (R+ ) has the properties (a)–(c) of
Lemma 4.1 if and only if, its mask is equal to zero on the interval [ 21 , 21 + 2l+1
1
).
Proof If m 0 ( 21 + ω2 ) = 0 for all ω ∈ [0, 2−l ), then from (4.114) for any odd k ∈ N,
we obtain

k ω k ω
φ̂(k + ω) = m 0 + φ̂ +
2 2 2 2

1 ω k ω
= m0 + φ̂ + = 0.
2 2 2 2
If the number k is even, then let us assume that its form k = 2s r , where r is odd.
Then, for any ω ∈ [0, 2−l ), we obtain

s
φ̂(k + ω) = φ̂(2s r + ω) = φ̂(r + 2−s ω) m 0 (2−i k + 2−i ω) = 0.
i=1
Since r is odd, then φ̂(r + 2−s ω) = 0. Thus, the function φ̂ is equal to zero on all
intervals [k, k + 2−l ), k ∈ N. Consequently, on the interval [0, 2−l ) it does appear in
zero, otherwise would have a periodic zero and, as per Theorem 4.6, would had been
unstable, which contradicts the assumption. In this way, property (c) of Lemma 4.1
is satisfied. Conversely, if this property is satisfied, then for any ω ∈ [0, 2−l ) and for
any k ∈ Z+ , we have

1 ω 1 ω
0 = φ̂(2k + 1 + ω) = m 0 k + + φ̂ k + +
2 2 2 2

1 ω 1 ω
= m0 + φ̂ k + + .
2 2 2 2
If there exists ω ∈ [0, 2−l ), for which m 0 ( 21 + ω2 ) = 0, then φ̂(k + 21 + ω2 ) = 0 For

any k ∈ Z+ . Consequently, the function φ̂ has periodic zero, and this means φ is
unstable. Therefore, m 0 ( 21 + ω2 ) = 0 for all ω ∈ [0, 2−l ). The lemma is proved.
By Proposition 4.15 and Lemma 4.2, we obtain
Theorem 4.17 The defect of approximation bφ by the stable scaling function φ ∈
L 2c (R+ ) (and the corresponding dyadic wavelet system) is equal to the smallest
integer r such that the mask m 0 (ω) is zero on the interval [ 21 , 21 + 2−r −1 ).
Thus, for dyadic wavelets, the binary order of approximation is one and the same,
the defect of approximation is determined by the length of the binary interval at the
starting point 21 , at which the mask turns into zero. For example, for the scaling Lang
function from Example 4.2, we have bφ = 1.
Corollary 4.8 Minimal defect of approximation is equal to zero and is achieved,
when the mask is zero identically on the interval [ 21 , 1).
For any n, there are dyadic orthogonal wavelets, which are generated by scaling func-
tions of order n and have a minimal approximation defect, i.e., the best approximate
properties on the class of W -continuous functions. A necessary and sufficient condi-
tion for this is the equality |m 0 (ω)| = 1 for all ω ∈ [0, 21 ). In order to construct such a
wavelet ψ in the algorithm A0 of Sect. 4.2, it is sufficient to choose for 0 ≤ l ≤ 2n−1 .
Indeed, according to Theorem 4.10 and Corollary 4.8, the following proposition is
true:
Proposition 4.16 A scaling function φ, satisfying Eq. (4.114), generates a system
of dyadic wavelets with a minimal approximation defect, if and only if, when

1, ω ∈ [0, 21 ),
|m 0 (ω)| =
0, ω ∈ [ 21 , 1).
In particular, the Haar splash has a minimal approximation error. It corresponds to the
condition n = 1. For every n ≥ 2, there are many systems of wavelets with minimal
defect. As a mask m 0 (ω), one can take an arbitrary 1-periodic binary-integer function
from En , which take the values ±1 on the interval [0, 21 ) (if the mask is complex
valued, then the values of the mask on [0, 21 ) can be arbitrary complex values with
Module 1) and equal to zero on [ 21 , 1).
Proposition 4.17 A stable scaling function φ has a minimal defect of approximation

if and only if, when it belongs to E0 .
Proof We can confine ourselves to the case when φ̂(0) = 1. If bφ = 0, then, by

Corollary 4.8, the mask m 0 (ω) is equal to zero for all ω ∈ [ 21 , 1). Then, by taking any
natural number j for any ω > 1, for which ω ∈ [2 j−1 , 2 j ), we obtain m 0 (2− j ω) = 0.
But by Theorem 4.4, we have
∞

φ̂(ω) = m 0 (2−s ω),
s=1
and also this means φ̂(ω) = 0 for all ω > 1. Consequently, φ ∈ E0 . Conversely, if
φ ∈ E0 , then φ̂(ω) = 0 for every ω > 1. Let us suppose that bφ > 0. This means that
m 0 (ω0 ) = 0 for some ω0 ∈ [ 21 , 1). Then, 0 = φ̂(2ω0 ) = m 0 (ω0 )φ̂(ω0 ), from where
φ̂(ω0 ) = 0. But since φ̂(ω0 + k) = 0 for all k ≥ 1, we obtain that ω0 -periodic zero
of the function φ̂, consequently, φ is unstable. The proposition is proved.
Thus, dyadic wavelets with minimal defect approximations are binary integers. Con-
versely, generally speaking, it is incorrect as shown in Proposition 4.18.
Proposition 4.18 For any n, there exists a scaling equation of order n, the solution
of which generates MRA, is a dyadic entire function, but it has the maximum possible
defect of approximation bφ = n − 1.
Proof For a given n, we take an arbitrary mask m 0 (ω), which is zero on the intervals
[2−n , 21−n ) and [ 21 , 21 + 2−n ), equal to 1 on the intervals [0, 2−n ) and [ 21 + 2−n , 21 +
21−n ), and at the remaining points of the interval [0, 1) does not vanish and satisfies
the condition
1 2
2
|m 0 (ω)| + m 0 ω + = 1.
2
By Theorem 4.10, the corresponding scaling function φ generates an MRA. As in

the proof of Proposition 4.17, we establish that φ ∈ En−1 . Nevertheless, bφ = n − 1
according to Theorem 4.17. The proposition is proved
For classic wavelets on R, the smoothness is a sufficient condition for good approxi-
mation properties (see, for example, De Boor et al. [7, 8], Jia [9]). In a dyadic case, as
from Proposition 4.18, there is no direct connection between the smoothness of the
wavelets and the approximation properties. Dyadic entire wavelets (whose smooth-
ness equals ∞) can have the greatest approximation error. At the same time, by
Proposition 4.17, all scaling functions belonging to E0 have optimal approximation
properties, that is, the smallest approximation error. In particular, Haar approxima-
tions are optimal.
In connection to this, we note that in some applied problems, when the approxi-
mation “signal” f by its projections P j f , the coefficients of Eq. (4.114) (and thus
the scaling function φ) at each level j are selected separately with the help of the
chosen criterion of optimization for concrete task, and such “adaptive” approach, as
a rule, leads to approximations, which are different from Haar approximations (for
more details on the adaptive approach, see Sects. 5.3, 5.5, and 7.5).
4.6 Exercise
4.1 If the mask m 0 does not have symmetric zeros, it does not turn into zero at the
point ω = 21 − 21n and is nonzero on the interval [0, 41 ), then the system of entire
shifts of the solution φ of the scaling equation (4.36) is linearly independent.
4.2 Prove that for a given integer L ≥ 0 and a given function φ ∈ L 2c (R+ ), φ̂(0) = 1,
the following properties are equivalent:
(i) φ̂ [r ] (n) = 0 for all r = 0, 1, . . . , L; n ∈ N (modified Strang–Fix condi-
tions);

k∈Z+ (x ⊕ k) φ(x ⊕ k) = 1 for each r = 0, 1, . . . , L and x ∈ R+ .

r
(ii)
For scaling functions on a real line, the Strang–Fix conditions are proved, for
example, in Novikov et al. [1, § 3.2].
4.3 Prove that if a finite L 2 solution φ of scaling equation (4.78) is a dyadic integer,
n−1
then supp φ̂ ⊂ [0, 22 ].
References
1. Novikov, I. Ya., Protasov, V. Yu., & Skopina, M. A. (2011). Wavelet theory. Providence: AMS
(Moscow, 2006).
2. Golubov, B. I., Efimov, A. V., & Skvortsov, V. A. (2008). Walsh series and transforms. URSS,
Moscow; English Transl. of 1st ed. Dordrecht: Kluwer (1991).
3. Lang, W. C. (1998). Fractal multiwavelets related to the Cantor dyadic group. International
4. Protasov, V. Yu. (2007). Approximation by Dyadic wavelets. Matematicheskii Sbornik, 198(11),
135–152; English Transl. Sbornik: Mathematics, 1665–1681.
5. Schipp, F., Wade, W. R. & Simon, P. (1990). Walsh series. Adam Hilger.
6. Strang, G., & Fix, J. (1973). An analysis of the finite element method. Englewood Clifs, N.J.:
Prentice Hall (second edition in 2008).
7. De Boor, C., DeVore, R., & Ron, A. (1991). Approximation from shift-invariant subspaces of
L 2 (Rd ), CMS-TSR University of Wisconsin-Madison, 92-2.
8. De Boor, C., DeVore, R., & Ron, A. (1992). The structure of finitely generated shift-invariant
spaces in L 2 (Rd ), CMS-TSR University of Wisconsin-Madison, 92-8.
9. Jia, R. Q. (1993). A Berstein type inequality associated with wavelet decomposition. Construc-
tion Aroximation, 9, 299–318.
10. Schoenberg, I. J. (1946). Contribution to the problem of approximation of equidistant data by
analytic functions, Part A and B. Quarterly of Applied Mathematics, 4(2), pp. 45–99, 112–141.
Chapter 5
Orthogonal and Periodic Wavelets on
Vilenkin Groups
5.1 Multiresolution Analysis on Vilenkin Groups
As noted in Chap. 1, the Walsh function can be identified with characters of the
Cantor dyadic group. This fact was first recognized by Gelfand in 1940s, who offered
to Vilenkin study series with respect to characters of a large class of abelian groups
which includes the Cantor group as special case see Vilenkin [1], Fine [2], Agaev
et al. [3]. For wavelets on Vilenkin groups most of the results relate to the locally
compact group G p , which is defined by a fixed integer p ≥ 2. The group G p have
a standard interpretation on R+ . Since the case p = 2 corresponds to the Cantor
group C , all the results on wavelets on R+ presented in Chap. 4 can be rewritten
for wavelets on C . In this section, necessary and sufficient conditions are given for
refinable functions to generate an MRA in the space L 2 (G p ). The partition of unity
property, the linear independence, the stability, and the orthogonality of “integer
shifts” of refinable functions in L 2 (G p ) are also considered.
Let us recall that the group G p consists of sequences x = x j , where x j ∈
{0, 1, 2, . . . p − 1} for j ∈ Z and with at most finite number of negative j such
that x j = 0. The zero sequence is denoted by θ . For any x = θ , there exists a unique
k = k(x) such that xk = 0 and x j = 0 for j < k. The group operation on G p is
defined as the coordinatewise addition modulo p:
(z j ) = (x j ) ⊕ (y j ) ⇔ z j = x j + y j (modp) f or j ∈ Z
and topology on G p is introduced via the complete system of neighborhoods of zero
Ul = {(x j ) ∈ G p : x j = 0 for all j ≤ l}, l ∈ Z .
The equality z = x y means that z ⊕ y = x. For p = 2, we have x ⊕ y = x y.

In this case, the group G 2 coincides with the locally compact Cantor group C and
the subgroup U0 is isomorphic to the compact Cantor group C; i.e., the topological

https://doi.org/10.1007/978-981-13-6370-2_5
172 5 Orthogonal and Periodic Wavelets on Vilenkin Groups
Cartesian product of a countable set of cyclic groups with discrete topology. It is

well known that C is a perfect nowhere dense totally disconnected metrizable space,
and therefore U0 is homeomorphic to the Cantor ternary set.
For any p ≥ 2, we let G = G p and U = U0 . One can show that G is self-dual.
The duality pairing on G takes x, w ∈ G to
⎛ ⎞
2πi
χ (x, w) = ex p ⎝ x j w1− j ⎠ .
p j∈Z
The Lebesgue spaces L q (G), 1 ≤ q ≤ ∞, are defined with the help of the Haar
measure μ on Borel subsets of G normalized by μ(U ) = 1. Denote by (·, ·) and
·
the inner product and the norm in L 2 (G), respectively.

The Fourier transform of a function f ∈ L 1 (G) ∩ L 2 (G),

fˆ(w) = f (x)χ (x, w)dμ(x), w ∈ G,
G
admits a standard extension to the space L 2 (G). By C0 (G), we denote the set of all
continuous complex-valued functions g on G such that, for any ε > 0 there exists a
compact set E ⊂ G (depending on g and ε) such that |g(x)| < ε for all x ∈ G \ E.
The following properties are well known (see, e.g.,):
1) If f ∈ L (G), then fˆ ∈ C0 (G).
2) The Fourier operator maps L 2 (G) into itself linearly, continuously, and one to
one.
3) If f, g ∈ L 2 (G), then ( f, g) = ( fˆ, ĝ) (Plancherel’s equality).
Take a discrete subgroup H = {(x j ) ∈ G : x j = 0 f or j > 0} and define an auto-
morphism A ∈ Aut G by the formula (Ax) = x j+1 . It is easy to see that the quotient
H
group A(h) contains p elements. We define a map λ : G → R+ by

λ(x) = x j p − j , x = (x j ) ∈ G.
j∈Z
The image of H under λ is the set of nonnegative integers: λ(H ) = Z+ . For every α ∈
Z+ , let h [α] denote the element of H such that λ(h [α] ) = α. Notice that χ (Ax, w) =
χ (x, Aw) for all x, w ∈ G.
The Walsh function for the group G can be defined by
wα (x) = χ (x, w[α] ), α ∈ Z+ , x ∈ G.
These functions are continuous on G and satisfy the orthogonality relations

wα (x)wβ (x)dμ(x) = δα,β , α, β ∈ Z+ ,
U
5.1 Multiresolution Analysis on Vilenkin Groups 173
where δα,β is the Kronecker delta. Also, it is well known that the system {wα } is
complete in L 2 (U ).
For any positive integer n, let εn (G) denote the collection of all functions defined
on G and constant on the sets
Un,s := A−n (h[s]) ⊕ A−n (U ), s ∈ Z+ .
The elements of the set ε(G) := ∪n εn (G) will be called p-Adic entire functions on
G (Golubov et al. [4, § 6.2]). Note that the sets Un,s , 0 ≤ s ≤ p n − 1, are cosets
of the subgroup A−n (U ) in the group U . It is easy to check that the functions wα ,
0 ≤ α ≤ p n − 1, belong to the class ε(G).
As above, the characteristic function of a set E is denoted by 1 E . Any function
f ∈ εn (G) can be expressed as the series
∞

f (x) = f n,s 1Un,s , (5.1)
s=0
where f n,s are the values of f on U n, s (i.e., f n,s = f (A−n (h [ s])))). Besides, note
that the proof of the equalities

wβ (x)dμ(x) = 0, 0 ≤ α ≤ pl+1 − 1, pl ≤ β ≤ pl+1 − 1 , l ∈ Z+ (5.2)
Ul,α
is similar to that of relations (1.5.18) from [4]. Using 5.1 and 5.2, we prove the
following.
Proposition 5.1 The following properties hold:
(a) If f ∈ L 1 (G) ∪ ξn (G), then suppose fˆ ⊂ U−n .
(b) If f ∈ L 1 (G) and supp fˆ ⊂ U−n , then fˆ ∈ ξn (G)
Proof Suppose that f ∈ L 1 (G) ∩ ξn (G). Then, for any w ∈ G, from 5.1, we have
∞

fˆ(w) = f (t)χ (t,¯ w)dμ(t) = f n,α χ (t,¯ w)dμ(x).
G α=0 Un
Setting t = A−n (h [α] ) ⊕ x and noting that A−n (U ) = Un , we obtain

χ (t, w)dμ(t) = ξ(A−n (h [α] ), w) χ (x, w)dμ(x).
Un,α Un
For an arbitrary element w = (w j ) from G \ U−n , we set s = min{ j : w j = 0}. We

can easily see that s ≤ −n. Let l = −s. Then, for any element x = (x j ) from Un we
have

2πi
χ (x, w) = ex p (xn+1 w−n + xn+2 w−n−1 + · · · + xl+1 w−l ) = wβ (x),
p
where β = w−n p n + w−n−1 p n−1 + · · · + w−l pl . Further, splitting Un into the sets
Ul,α , and using 5.2 we obtain
pl−n −1

χ (x, w)dμ(x) = wβ (x)dμ(x) = 0.
Un α=0 Ul,α
Thus, fˆ(w) = 0 for all w ∈ G \ U−n and assertion (a) is proved.

Now suppose that f ∈ L 1 (G) and supp f ⊂ U−n . Then
pn −1

fˆ(w) = f (t)χ (t, w)dμ(t) = f (t)χ (t, w)dμ(t).
U−n s=0 U0,s
Hence, for
w = (. . . , 0, 0, w−l , w−l+1 , . . . , w0 , w1 , w2 , . . .).
Substituting t = h [s] ⊕ x, we obtain
pn −1

fˆ(w) = cs,β ws∗ (w), (5.3)
s=0
where cs,β = U f (h [ s]) ⊕ xwβ (x)dμ(x), β=β(w) = w−l pl + w−l+1 pl−1 + · · · +

w−l p + w0 . Since ws (w) and β(w) are constant on the sets Un,s , s ∈ Z+ , it follows
from 5.3 that fˆ ∈ ξn G. Proposition 5.1 is proved.
A function f ∈ L 2 (G) is said to be stable if there exist positive constants A0 and B0

such that
21

21
∞

∞

∞
A0 |aα | 2
≤ aα f · h [α] ≤ B0 |aα | 2
,

α=0 α=0 α=0
for each sequence {αn } ∈ l 2 . In other words, a function f is stable in L 2 (G) if

functions f (· h), h ∈ H , form a Riesz system in L 2 (G). Note also, that a function
f is stable in L 2 (G) with constants A0 and B0 if and only if

A0 ≤ | fˆ(w H )|2 ≤ B0 for a.e ∈ w ∈ G. (5.4)
h∈H
The proof of this fact is quite similar to that of Theorem 1.1.7 in [5]. We say that
a function g : G → C has a periodic zero at a point w ∈ G if g(w ⊕ h) = 0 for all
h ∈ H. The set of all compactly supported functions in L 2 (G) will be denoted by

L 2c (G).
Proposition 5.2 For any f ∈ L 2c (G), the following properties are equivalent:
(a) f is stable in L 2 (G).
(b) { f (· h)|h ∈ H } is a linearly independent system.
(c) fˆ does not have periodic zeros.
Proof The implication (a) ⇒ (b) follows from the well-known property of the Riesz
system (see, e.g., Novikov et al. [5], Theorem 1.1.2). Our next claim is that f ∈
L 1 (G), since f has compact support and f ∈ L 2 (G). Let us choose a positive integer
n such that supp f ⊂ U1−n . By Proposition 5.1, then f ∈ ξn−1 (G). Besides, if λ(h) >
p n−1
μ{supp f (· h) ∩ U1−n } = 0.
Therefore, the linearly independence of the system and { f (· h)|h ∈ H } is equiv-

alent to that of the finite system { f (· h [α] )|α = 0, 1, 2, . . . p n−1 − 1}. Further, if
some vector (a0 , . . . , a pn−1 −1 ) satisfies conditions
pn−1 −1

aα f (· α) = 0 and |a0 | + · · · + |a2n−1 −1 | > 0, (5.5)
α=0
then using the Fourier transform we obtain
pn−1 −1

fˆ(w) aα wα (w) = 0.
α=0
The Walsh polynomial

pn−1 −1

W (w) = aα wα (w)
α=0
is not identically equal to zero; hence among Un−1,s , 0 ≤ s ≤ p n−1 − 1 there exists a
set (denote it by X ) for which W (X ⊕ h) = 0, h ∈ H . Since fˆ ∈ ξn−1 (G), it follows
that 5.5 holds if and only if there exists a set X = Un−1,s , X ⊂ U , such that fˆ(X ⊕
h) = 0 for all h ∈ H . Thus (b) ⇒ (c). It remains to prove that (c) ⇒ (a). Suppose
that fˆ does not have periodic zeros. Then

F(w) = | fˆ(w h)|2
h∈H
is positive and H -periodic function. Moreover, since fˆ ∈ ξn−1 (G), we see that F is
constant on each Un−1,s , 0 ≤ s ≤ p n−1 − 1. Hence, (5.4) is satisfied and Proposition
5.2 is established.
Proposition 5.3 Let φ ∈ L 2 (G) and the system {φ(· h) : h ∈ H } be Riesz basis
in the closed subspace V of space L 2 (G) to belong to V , it is necessary and sufficient
that there exists H -periodic function m f ∈ L 2 (U ) such that
fˆ(w) = m f (w)φ̂(w) w ∈ G. (5.6)
Besides, the coefficient of decomposition f for the system {φ(· h) : h ∈ H } coin-

cide with the corresponding Walsh coefficient of function m f .
Proof According to the known properties of Riesz system (see, for example, Theorem
1.1.2 in [5]) the subspace V consists of function of the type

f (x) = ch φ(x H ) = ch [α] φ x h [α] , x ∈ G, (5.7)
h∈H α∈Z+

where h∈H |ch |2 < ∞. Using the Fourier equation, we get f ∈ L 1 (G), since f
has compact support and f ∈ L 2 (G). Let us choose a positive integer n such that
supp f ⊂ U1−n . By Proposition 5.1, then f ∈ ξn−1 (G). Besides, if λ(h) > p n−1 ,

fˆ(w) = ch χ (h, w)φ̂(w) = ch [α] wα (w)φ̂(w). (5.8)
h∈H α∈Z+
It is evident from here that 5.6 is fulfilled if we assume

m f (w) = wα (w), w ∈ U. (5.9)
α∈Z+
Conversely if 5.6 is satisfied and m ∈ L 2 (U ), then m f may be written in the form

5.9. From here follows 5.8 and that means even 5.7. The proposition is proved.
We say that a function φ ∈ L 2c (U ) is a refinable function, if it satisfies an equation

of the type
pn −1

φ(x) = p aα φ Ax h [α] . (5.10)
α=0
The functional equation 5.10 is called the refinement equation. The Walsh polyno-
mials on G, i.e., finite linear combination of the characters, play the same role as the
trigonometric polynomials in the real setting. So, we have

φ̂(w) = m 0 A−1 w φ̂ A−1 w , (5.11)
where the Walsh polynomial

pn −1

m 0 (w) = p aα wα (w) (5.12)
α=0
is called the mask of Eq. 5.10 (or the mask of its solution φ). The coefficients of
Eq. 5.10 are related to the values bs of m 0 on Un,s by means of the direct and the
inverse Vilenkin–Chrestenson transforms:
p −1 n
1
aα = n bs wα (A−n h [s] ), 0 ≤ αp n − 1, (5.13)
p s=0
pn −1

bs = aα wα (A−n h [s] ), 0 ≤ s ≤ p n − 1. (5.14)
s=0
These transforms can be realized by the fast algorithms (see, for instance, [Schipp
et al. [18]]). Thus, any choice of the values of m 0 on Un,s defines also the coefficient
of Eq. 5.10.
Now, we prove the following analogue of Theorem 4.4
Theorem 5.1 Let ϕ ∈ L 2c (G) satisfy Eq. 5.10. If ϕ̂(θ ) = 1, then
pn −1

aα = 1 and suppϕ ⊂ U1−n .
α=0
The solution is unique, and is given by the formula

ϕ̂(w) = m 0 (A− j w), (5.15)
and possesses the following properties:

ϕ̂(h) = 0 for all h ∈ H \{θ };
1.
2. h∈H ϕ(x ⊕ h) = 1 for almost every x ∈ G (the partition of unity property).
Proof As wα (θ ) = φ(θ ) = 1, after submission in 5.11 and 5.12 the values of w = θ

we get the equation
pn −1

aα = 1.
α=0
Let s be the highest integer such that
μ{x ∈ U0,s−1 |φ(x) = 0} > 0,
where U0,s−1 = h [s−1] ⊕ U. Let us assume that s ≥ p n−1 + 1. Then the arbitrary
element x from U0,s−1 has the form
x = (. . . , 0, 0, x−k , x−k+1 , . . . , x−1 , x0 , x1 , . . .), x j ∈ {0, 1, . . . , p − 1}, (5.16)
where
x−k p k + x−k+1 p k+1 + · · · + x−1 p + x0 = s − 1, x−k = 0, k ≥ n − 1. (5.17)
Let us select this element such that the number λ(x) is not p rational (then the average
of the component x j with positive indices is an infinite number other than zero). For
any α ∈ {0, 1, . . . , p n − 1}, element y (α) = Ax h [α] is the form

(α) (α) (α)
y (α)
= . . . , 0, 0, y−k−1 , y−k , . . . , y−1 , y0(α) , y1(α) , y2(α) , . . . , k + 1 ≥ n,
(α)
where y−k−1 = x−k = 0 and average of the number y αj , j ≥ 0 is other than zero.
Consequently,
λ(Ax)h [α] > p n , x ∈ U0,s−1 . (5.18)
If s ≤ p n then from (5.18), we get that φ(Ax)h [α] ) = 0 for x ∈ U0,s−1 .

Then, on the strength of (5.10), φ(x) = 0 for any x ∈ U0,s−1 , which contradicts
selection s. Hence s ≥ p n + 1. By using this inequality for any α ∈ {0, 1, 2, . . .
p n − 1} from 5.16 and 5.17, we get
λ(Ax h [α] ) > p(s − 1) − ( p n − 1) ≥ 2(s − 1) − (s − 2) = s.
From here just as above, it follows that φ(x) = 0 for x ∈ U0,s−1 . Hence, s ≤ p n − 1
and φ ⊂ U1−n .
Let us prove that φ̂ can be represented by the Formula (5.16). As φ̂ is finite and
belongs to L 2 (G), then it belongs also to L 1 (G). As suppφ ⊂ u 1−n , then as per
Proposition 5.1, we have φ̂ ∈ ξn−1 (G). On the strength of the condition φ̂(θ ) = 1,
we get that φ̂(w) = 1 for all w ∈ Un−1 . On the other hand, m 0 (w) = 1, w ∈ Un−1 .
This means that for any natural number l

l+n ∞

φ̂(w) = φ̂(A−l−n w) m 0 (A− j w) = m 0 (A− j w), w ∈ U−l .
j=1 j=1
Hence the equality is true and the solution φ is unique. As φ ∈ L 1 (G) then φ̂ ∈
C0 (G). Besides m 0 (As h) = 1 on the strength of equality m 0 (θ ) = 1n 0 . Hence for
any h ∈ H \ {θ }

j−1
φ̂(h) = φ̂(h) m 0 (As h) = φ̂(A j h) → 0
s=0
when j → +∞. It follows from here that

φ̂(h) = 0, h ∈ H \ {θ }. (5.19)
Using now the Poisson summation formula:

φ(x ⊕ h) = φ̂(h)χ (x, h),
h∈H h∈H
and using 5.19 and the condition φ̂(θ ) = 1, we get

φ(x ⊕ h) = 1 ∀x ∈ G.
h∈H
Theorem 5.1 is proved.
Proposition 5.4 Let φ ∈ L 2 (G). Then {φ(· h) : h ∈ H } is an orthonormal system

in L 2 (G) if and only if

|φ̂(w h)|2 = 1 f ora.e.w ∈ G.
h∈H
Proof The function

φ(w) = |φ̂(w h)|2
h∈H
is H -periodic: φ(w ⊕ h) = φ(w) for all h ∈ H. Furthermore, it has a finite L 1 -norm

on U , because

0≤ φ(w)dμ(w) = |φ̂(w h)|2 dμ(w)
U U h∈H

= |φ̂(w)| dμ(w) =
2
|φ̂(w)|2 dμ(w) < +∞.
h∈H U ⊕h G
Denote by {φ̂(h)} the Fourier coefficients of function φ with respect to the system
{χ (h, ·)}. For any h ∈ H by a change of variable η = w h, we have

φ̂(h) = φ(w)χ (h, w)dμ(w) = χ (h, w) |φ̂(w)|2 dμ(w)
U U h∈H

= χ (h, η)|η̂| dμ(η) =
2
|φ̂(w)|2 χ (h, w)dμ(w).
h∈H U ⊕h G
Applying Plancherel’s equality, we get

φ(x h)φ(x)dμ(x) = φ̂(h), h ∈ H.
G
To complete the proof, we note that 3.1 is equivalent to φ̂(h) = δθ,h for h ∈ H.
For each l ∈ {0, 1, 2, . . . p − 1} let us denote by δl the sequence w = (w j ) such
that w1 = l and w j = 0 for j = 1 (in particular, δ0 = θ .) Note that δl = A−1 wl ∈ U,
λ(δl ) = pl and wα (δl ) = ex p( 2παl
p
) for all l ∈ {0, 1, 2, . . . p − 1}.
Proposition 5.5 Suppose that φ ∈ L 2 (G) satisfies 5.10 and the translates {φ(·
h) : h ∈ H } are orthonormal in L 2 (G). Then

p−1
|m 0 (w ⊕ δl )|2 = 1 f orall ∈ w ∈ G, (5.20)
l=0
where m 0 is the mask of Eq. 5.10.

Proof For l ∈ {0, 1, 2, . . . p − 1}, we put
Hl = {h ∈ H : A−1 h δl ∈ H }
such that h 0 = l for h = (h j ) ∈ Hl . Since φ̂ is continuous on G, by Proposition 5.4

we obtain
|φ̂(A−1 w ⊕ A−1 h)|2 = 1, l ∈ {0, 1, 2, . . . p − 1},
h∈Hl
for all w ∈ G. It follows from this and 5.11 that

2
1= m 0 (A−1 (w ⊕ h))2
φ̂(A−1 (w ⊕ h))
h∈H
2 2
−1
p−1
= −1
m(A w ⊕ δl ) φ̂(A h) = |m(A−1 w ⊕ δl )|2 .
h∈H h∈Hl l=0
Thus, 5.20 holds. The proof is complete.

We recall that a collection of closed subspaces V j ⊂ L 2 (G), j ∈ Z is called a mul-
tiresolution analysis (MRA) in L 2 (G) if the following hold:
(i) V j ⊂ V j+1 for all j ∈ Z,
(ii) ∪V j = L 2 (G) and ∩V j = {0},
(iii) f (·) ∈ V j ⇔ f (A·) ∈ V j+1 for all j ∈ Z,
(iv) f (·) ∈ V0 ⇒ f (· h) ∈ V0 for all h ∈ H,
(v) there is a function φ ∈ L 2 (G) such that the system {φ(· h) : h ∈ H } is an
orthonormal basis of V0 .
The function φ in condition (v) is called a scaling function in L 2 (G).
For arbitrary φ ∈ L 2 (G) we set
j
φ j,h (x) = p 2 φ(A j x h), j ∈ Z, h ∈ H,
and
V j = span{φ j,h : h ∈ H }, j ∈ Z. (5.21)
We say that a function φ generates an MRA in L 2 (G) if the system {φ(· h) : h ∈ H }

is orthonormal in L 2 (G) and, in addition, the family of subspaces V j is an MRA in
L 2 (G). If a function φ generates an MRA in L 2 (G), then it is a scaling function in
L 2 (G). In this case the system {φ j,h \ h ∈ H } is an orthonormal basis of V j for every
j ∈ Z and one can define orthonormal wavelets ψ1 , ψ2 , ψ3 , . . . ψ p−1 in such a way
that the functions
j
ψl, j,h = p 2 ψl (A j x h), 1 ≤ l ≤ p − 1, j ∈ Z, h ∈ H,
form an orthonormal basis of L 2 (G) (see Sect. 5.2)
Proposition 5.6 If the system, {φ(· h) : h ∈ H } is orthonormal basis in V0 , then

∩V j = {0}.
Proof From Eq. 5.20, it follows that at each j ∈ Z the system {φ j,h \ h ∈ H } is the
orthonormal basis of space V j . Hence, the orthogonal projector P j : L 2 (G) → V j
as per the formula
Pj f = ( f, φ j,h )φ j,h , f ∈ L 2 (G).
h∈H
Let us assume that f ∈ ∩V j and fix ε > 0. The C0 (G) denotes the set of continuous
on G function with compact suppport on L 2 (G). Let us select f 0 ∈ C0 (G) such that
f − f 0
< ε.

f − P j f 0
≤
P j ( f − f 0 )
≤
f − f 0
< ε
and means

f
≤
P j f 0
+ ε (5.22)
for each j ∈ Z. If supp f 0 ⊂ Ul,n

j
(P j f 0 , φ j,h ) = ( f 0 , φ j,h ) = p 2 f o (x)φ(A j x h)dμ(x),
Ul
where the number l depends on f 0 . Using Cauchy–Bunyakovsky inequality, from

here we get

P j f 0
=
2
|(P j f 0 , φ j,h )| ≤
f 0
2 2
p j
|φ(A j x h)|2 dμ(x).
h∈H h∈H Ul
For j < l, we have


pj |φ(A j x h)|2 dμ(x) = |φ(x)|2 dμ(x),
h∈H Ul Sl, j
where
Sl, j = {y h|y ∈ Ul− j }.
h∈H
Consequently,

P j f 0
≤
f 0
2 2
1 Sl, j (x)|φ(x)|2 dμ(x). (5.23)
G
It is easy to see that

lim 1 Sl, j (x) = 0 x ∈
/ H.
j→−∞
According to Lebesgue theorem, we get from 5.23 that
lim
P j u
= 0.
j→−∞
Taking into consideration 5.22, we conclude that

f
≤ ε and hence ∩V j = {0}.
Proposition 5.7 Let the mask m 0 of the measuring Eq. 5.10 meet the conditions

p−1
m(θ ) = 1 u |m(w⊕)δl |2 = 1 w ∈ G. (5.24)
l=0
Then the function φ given by the equation

∞

φ̂(w) = m(A− j w), (5.25)
j=1
is the L 2 -solution of Eq. 5.10 when

φ
≤ 1.
Proof Flow convergence of the product in 5.24 follows from the fact that the mask
m 0 is equal to 1 for the set Un,0 (such that for any w ∈ G only a finite number of
the factors in 5.24 may be different from 1). Let us represent the right-hand part of
Eq. 5.25 by g(w). It follows from 5.24 that |m 0 (w)| ≤ 1 for all w ∈ G. Hence for
any s ∈ N, we have
s
|g(w)|2 ≤ |m 0 (A− j w)|2
j=1
and consequently,

s
s−1
|g(w)|2 dμ(w) ≤ |m 0 (A− j w)|2 dμ(w) = 2s |m 0 (A j w)|2 dμ(w).
U−l U−l j=1 U j=0
(5.26)
From the equation,
pn −1

m 0 (w) = aα wα (w), wα (w)wβ (w) = wαβ (w),
α=0
is seen that
pn −1

|m 0 (w)|2 = cα wα (w), (5.27)
α=0
where the coefficients cα are calculated using acα . Let us substitute 5.27 in the part
Equation 5.24 and using the fact that α is smaller than p, then

p−1
wα (δl ) = p,
l=0
and for the remaining α this sum is equal to 0. As a result, we get c0 = 1

p
and cα = 0
for nonzero α, multiple p. It means, the following equation is true
p −1 p−1 n−1
1
|m 0 (w)| = +
2
c pα+l w pα+l (w).
p α=0 l=1
From here, we get

s−1 σ (s)

|m 0 (A j w)|2 = p −s + bγ wγ (w), σ (s) ≤ sp n−1 ( p − 1),
j=0 γ =1
where each of the coefficients bγ is equal to the product of certain coefficients

c pα+l , l = 1, 2, . . . p − 1. Taking into consideration that

wγ (w)dμ(w) = 0, γ ∈ N,
U
we have

s−1
|m 0 (A j w)|2 dμ(w) = p −s .
U j=0
Substituting in 5.26, for any arbitrary l ∈ N, we derive the equation


|g(w)|2 dμ(w) ≤ 1
U−l
and consequently,
|g(w)|2 dμ(w) ≤ 1. (5.28)
G
Now let φ ∈ L 2 (G) and φ̂ = g. Then, from 5.25, it follows that
φ̂(w) = m 0 (A−1 w)φ̂(A−1 w)
and this means φ satisfies Eq. 5.10. Besides, as per Parseval’s identity, we get from
5.28, that
φ
≤ 1. Proposition 5.7 is proved.
A compact subset E of G is said to be congruent to U modulo H , if μ(E) = 1
and for each w ∈ E, there is an element h ∈ H such that w ⊕ h ∈ U. Let m 0 be the
mask of Eq. 5.10. We say that m 0 satisfies the modified Cohen condition, if there
exists a compact subset E of G containing a neighborhood of the zero element such
that:
(1) the set E congruent to U modulo H ;
(2) the inequality
inf inf |m 0 (A− j w)| > 0 (5.29)
j∈N w∈E
is true.
Since E is compact, we note that if m 0 (θ ) = 1 than there exists a number j0
such that m 0 (A− j w) = 1 for all j > j0 , w ∈ E. Therefore, 5.29 holds if the poly-
nomial m 0 (w) does not vanish on the sets A−1 (E), . . . , A− j0 (E). Moreover, we can
choose j0 ≤ p n , because m 0 is completely defined by the values bs , 0 ≤ s ≤ p n − 1.
Proposition 5.8 Let the mask m 0 of Eq. 5.10 satisfy the conditions

p−1
m 0 (θ ) = 1 u |m(w⊕)δl |2 = 1 for w ∈ G. (5.30)
l=0
Suppose that the function φ ∈ L 2 (G) is determined with the help of formula
∞

φ̂(w) = m 0 (A− j w), w ∈ G. (5.31)
j=1
The system {φ(· h) : h ∈ H } is orthonormal in L 2 (G) then and only then when
the full m 0 satisfies the modified condition of Cohen.
Proof From 5.30 and 5.31, with the help of Proposition 5.5 we derive that the given
function φ is the solution of Eq. 5.10 and satisfies the condition of the Theorem 5.1.
Let us assume that the system {φ(· h) : h ∈ H } is orthonormal in L 2 (G). Let us

examine
φ(w) = |φ̂(w h)|2 . (5.32)
h∈H
Evidently, the function φ is nonnegative and possess the property H -periodicity.

From Proposition 5.1, it follows that the function φ (and also φ̂) is constant on
the sets Un−1,s . 0 ≤ s ≤ p n−1 − 1, merging of which coincides with U . Now, as in
Proposition 5.5 we get that φ(w) = 1 for all w ∈ U . Let us construct the compact
set E ⊂ G, which congruent to U modulo H , for which is fulfilled Eq. 5.29.
For the arbitrarily fixed w(0) ∈ U , let us select the neighborhood
Uk = {(w j ) ∈ G : w j = 0 j ≤ k},
such that,
1
|φ̂(w(0) h)|2 ≥ ,
h ∗ ∈Vk
2
where Vk = Uk ∩ H. The set Vk is finite, but the function φ̂ is continuous. Hence,

the neighborhood (w(0) ) such that for all ξ ∈ (w(0) ) fulfilled the inequality
1
|φ̂(ξ h)|2 ≥ .
h∈Vk
4
Let us find out such neighborhood (w) for all w ∈ U. On the strength of com-
pactness of the subgroup U , from the covering/mask {(w) : w ∈ U }, it is pos-
sible to separate the finite covering/mask {(w(α) )}sα=1 . In the structure for each
α ∈ {1, 2, . . . s}, there exists a number k(α) such that for all ξ ∈ (w(α) ) there is
inequality
1
|φ̂(ξ h)|2 ≥ ,
h∈V
4
k
s
where Vk(α) = Uk(α) ∩ H . Suppose V = α=1 Vk(α) . Then the inequality
1
|φ̂(w h)|2 ≥ (5.33)
h∈V
4
is true for all w ∈ U. Let c0 = 1

√
2 γ
, where γ is the number of elements of the set
V . On the strength of 5.33, for each w ∈ U the inequality |φ̂(w h)| ≥ c0 is met at
certain h ∈ V. Let us note that the set
S0 = {w ∈ U : |φ̂(w)| ≥ c0 }
contains the neighborhood of zero element θ on the strength of the continuous func-
tion φ̂ and the equation φ̂(θ ) = 1. From the elements h 1 , . . . h γ forming the set V ,
let us determine by induction method the sets
S1 = {w ∈ U \ S0 : |φ̂(w h 1 )| ≥ c0 },
S2 = {w ∈ U \ (S0 ∪ S1 ) : |φ̂(w h 2 )| ≥ c0 },
γ −1

Sγ = w ∈ U \ Sl : |φ̂(w h γ −1 )| ≥ c0 .
l=1
γ
The set E = l=0 is the neighborhood of the zero element of the group G and the
congruent to U modulo H . Let us select the number j0 such that m 0 (A− j w) = 1 for
all j ≥ j0 and all w ∈ E. According to 5.29, for w ∈ E we get

j0
|φ̂(w)| = |m 0 (A− j w)||φ̂(A− j0 w)|,
j=1
where |φ̂(w)| ≥ cα . It is seen from here that for the constructed set E Condition 5.29
is true.
Conversely, let us assume that there exists a compact set E congruence to U as per
modulo H containing the neighborhood of zero element and such that the 5.29 is
fulfilled. Let us prove that the system {φ(· h) : h ∈ H } is orthonormal in L 2 (G).
For arbitrary natural k, let us assume

k
[k]
μ = m 0 (A− j w)1 E (A−k w), w ∈ G.
j=1
From Condition 5.30, it is seen that m 0 (w) = 1 for all w ∈ Un,0 . Hence, it follows
from 5.31 that
lim μ[k] = m 0 (w), w ∈ G. (5.34)
k→∞
Taking into consideration that A−n w ∈ Un,0 for w ∈ U , we have also

n−1
m 0 (w) = m(B − j w), w ∈ U.
j=1
On the strength of 5.29, there exists the constant c1 > 0 such that
|m 0 (A− j w)| ≥ c1 for j ∈ N, w ∈ E
and this means c11−n |m(w)| ≥ 1 E (w) for w ∈ G. From here, we get
[k] k
k

μ (w) = m 0 (A− j w)1 E (A−k w) ≤ c1−n m 0 (A− j w) m 0 (A−k w)
1
j=1 j=1
and taking into consideration 5.31

[k]
μ (w) ≤ c1−n |m 0 (w)|, k ∈ N, w ∈ G. (5.35)
1
Now let
[k] 2
Ik (s) = μ (w) ws (w)dμ(w) k ∈ N, s ∈ Z+ .
G
Assuming ζ = A−k w, we find out

k
Ik (s) = |m 0 (A− j w)|2 ws (w)dμ(w)
Uk,0 j=1

k−1
= p k
|m 0 (ζ )| 2
|m 0 (A j ζ )|2 ws∗ (Ak ζ )dμ(ζ ). (5.36)
U j=1
p−1
As U = l=0 (A−1 (U ) ⊕ δl ), then
p−1
k−1
Ik (s) = p k−1 |m(A−1 w ⊕ δl )|2 |m(A j−1 w)|2 ws (Ak−1 w)dμ(w)
U j=1
and on the strength of 5.30, it follows

k−2
Ik (s) = p k−1 |m 0 (A j w)|2 ws (Ak−1 w)dμ(w).
U j=0
From here and from 5.36, it follows that
Ik (s) = Ik−1 (s).
Similarly, when k = 1

I1 (s) = p |m(w)|2 ws (Aw)dμ(w) = dμ(w),
U U
where the last integral is equal to δ0,s on the strength of the system {wα } ∈ L 2 (U ).
It means
Ik (s) = δ0,s k ∈ N, s ∈ Z+ . (5.37)
In particular, for all k ∈ N

[k] 2
Ik (0) = μ (w) dμ(w) = 1.
G
Using Lebesgue’s theorem on limited convergence, and from 5.34, 5.35, and 5.37
we find out that

|φ̂(w)|2 ws (w)dμ(w) = lim Ik (s) = δ0,s .
G k→∞
From here as per Plancheral’s formula, it follows that the system {φ(· ⊕ h) : h ∈ H }
is orthonormal in L 2 (G). Proposition 5.7 is proved.
Let M ⊂ U and let

p−1
Tp M = {A−1 w[l] + A−1 (w) : w ∈ M}.
l=0
The set M is said to be blocked (for the mask m 0 ) if it coincides with some union of
sets
Un−1,s = A1−n (w[s] ) ⊕ A1−n (U ), 0 ≤ s ≤ p n−1 − 1, (5.38)
which does not contain the set Un−1,0 and satisfies the condition T p M ⊂ M ∪ Z (m 0 ),
where Z (m 0 ) is the set of all zeros of m 0 on U . It is clear that each mask can have
only a finite number of blocked sets.
Proposition 5.9 Let φ ∈ L 2c (G) be a solution of Eq. 5.10 such that φ̂(θ ) = 1. Then
φ is not stable if and only if its mask m 0 possess a blocked set.
Proof Using Theorem 5.1 and the Proposition 5.1, we have suppφ ⊂ U1−n φ̂ ∈
ξn−1 (G). Let us assume that the function φ is unstable. While proving Proposi-
tion 5.2, it was established that if the unstable function f ∈ L 2 (G) has a carrier in
U1−n , then there exists as set X = Un−1,s such that all the points of the set X are
periodic zeros of the Fourier transform fˆ (and any periodic zero w ∈ U for fˆ is in
the same set X ).
Hence the set
M0 = w ∈ U : φ̂(w ⊕ h) = 0
is represented in the form of convergence of some from the sets Un−1,s , 0 ≥ s ≥

p n−1 − 1. As φ̂(θ ) = 1, then M0 does not contain Un−1,s . Besides, if w ∈ M0 , then
as per Formula 5.11

m 0 A−1 w ⊕ A−1 h φ̂ A−1 w ⊕ A−1 h = 0 for all h ∈ H
and consequently, the elements A−1 w ⊕ A−1 w[l] , l = 0, 1, 2, . . . p − 1 belongs either

to M0 or Z (m 0 ). Thus, if φ is unstable, then the set M0 is blocking set for m o . Con-
versely, let the mask m 0 has a blocking set M. Let us show that then each element from
M is a periodic zero for φ̂ (and, consequently, as per Proposition 5.2 the function φ is
unstable). Let us propose that there exists w ∈ M and h ∈ H also that φ̂(w ⊕ h) = 0.
Let us select the natural number j, for which A− j (w ⊕ h) ∈ Un−1 and then for each
r ∈ {0, 1, 2, . . .}, let us find out u r ∈ U, h r ∈ H such that A−r (w ⊕ h) = u r ⊕ h r .
Further for each r ∈ {0, 1, 2, . . . j − 1}, let us take l ∈ {0, 1, 2, . . . p − 1} such that
A−1 h r = A−1 w[l] ⊕ wr ,
where wr ∈ H . Then
u r +1 ⊕ h r +1 = A−r −1 (w ⊕ h) = A−1 (u r ⊕ h r ) = (A−1 u r ⊕ A−1 w[l] ) ⊕ wr
and consequently,
u r +1 = w A−1 u r ⊕ A−1 w[l] , l = l(r ) ∈ {0, 1, 2, 3 . . . , p − 1}.
It is evident from here that if u r ∈ M, then u r +1 ∈ T p M. Besides, it follows from the

equation

j

j
φ̂(w ⊕ h) = φ̂(A−1 (w ⊕ h)) m 0 (A−r (w ⊕ h)) = φ̂(u j ) m 0 (u r )
r =1 r =1
that all u r ∈
/ Z (m 0 ). Thus, if u r ∈ M then u r +1 ∈ M, u 0 = w ∈ M, it follows from
here that all u r ∈ M. this contradicts the fact that u j = A− j (w ⊕ h) ∈ Un−1 and the
set M does not contain Un−1 = Un−1,s . Proposition 5.9 is proved.
Proposition 5.10 Let φ be the finite L 2 solution of Eq. 5.10 such that θ̂ = 1. The
system {φ(· h) : h ∈ H } is orthonormal in L 2 (G) then and only then when the
mask m of Eq. 5.10 does not have blocking sets and satisfies the condition

p−1
|m 0 (w ⊕ δl )|2 = 1 w ∈ G. (5.39)
l=0
Proof If the system {φ(· h) : h ∈ H } is orthonormal in L 2 (G), then the condition

5.39 is fulfilled (see Proposition 5.5 and the absence of blocking sets follows from
Propositions 5.2 and 5.3. Conversely, let mask m 0 does not have blocking sets and
Condition 5.39 is fulfilled. As in 5.2, we suppose

φ(w) = |φ̂(w h)|2 . (5.40)
h∈H
As the function φ is not negative and possess the property H -periodicity, on the
strength of proposition 5.4 it is sufficient to verify that φ(w) = 1. Let
a = in f {φ(w)|w ∈ U }.
From Theorem 5.1 and Proposition 5.1, it follows that the function φ just as φ̂ is
constant on the sets Un−1,s then the function φ̂ has a periodic zero and this means
φ̂ is unstable. According to Proposition 5.2 and 5.3, this contradicts the absence of
blocking sets of mask m 0 . This means that the number α is positive. Besides, taking
into consideration Theorem 5.1, we have φ(θ ) = 1. Thus 0 ≤ a ≤ 1. We notice that
from 5.11 and 5.40 follows the equation

p−1

φ(w) = m 0 (A−1 w δl )2 φ(A−1 w δl ). (5.41)
l=0
Now let Ma = {φ(w) = a|w ∈ U }. If 0 < a < 1, then from 5.39 and 5.41, it follows
that for any w ∈ Ma the elements belonging to A−1 w δl , l = 0, 1, 2, . . . p − 1 are
the blocking set, which contradicts the condition. Thus φ(w) ≥ 1 for all w ∈ U .
From here and from the equation

ˆ 2 dμ(w) = ˆ 2
φ(w)dμ(w) = |φ(w)| φ(w) dμ(w) =
φ
2
U h∈H U ⊕h G
as per Proposition 5.7, we get that

φ(w)dμ(w) = 1
U
Repeatedly using the inequality and using the fact that function φ is constant in sets
Un−1,s 0 ≤ s ≤ p n−1 − 1, we conclude that φ(w) ≡ 1. Proposition 5.8 is proved.
Theorem 5.2 Suppose that the refinement Eq. (5.10) possesses a solution ϕ ∈
L 2c (G) such that ϕ̂(θ ) = 1 and the corresponding mask m 0 satisfies the condition
(5.30). Then the following are equivalents:
(a) The function ϕ generates an MRA in L 2 (G).
(b) The mask m 0 satisfies the modified Cohen’s condition.
(c) The mask m 0 has no blocked sets.
Proof According to Proposition 5.9, we have ∩V j = {0}. Embedding V − j ⊂ V j+1
from the fact that satisfies Eq. 5.10. Let us show that
∪V j = L 2 (G). (5.42)
Let f ∈ (∪V j )⊥ and ε > 0. Let us select u ∈ L 2 (G) such that and û ∈ C0 (G) with
fˆ − û
< ε. For the orthogonal projection P j f , the function f in the subspace V j
we have

P j f
2 = (P j f, P j f ) = ( f, P j f ) = 0
and it means,

P j u
=
P j ( f − u)
≤
f − u
=
fˆ − û
< ε. (5.43)
Let us fix the number j ∈ N such that suppû ⊂ U− j , A− j w ∈ Un−1 for all w ∈ suppû.
j
Using the fact that the system { p − 2 wα (A− j )} is orthonormal and full in L 2 (U− j ) for
the function Γ (w) = n̂(w)φ̂(A− j w), we have

p− j |Γ (w)|2 dμ(w) = |cα (Γ )|2 , (5.44)
Uj α∈Z+
where
j
cα (Γ ) = p − 2 Γ (w)Wα∗ (B − j w)dμ(w).
U− j
Noting that

φ(A j x h [α] )χ (x, w)dμ(x) = p − j φ̂(A− j w)wα (A− j w)
G
we get
j
p − 2 (u, , φ j,h ) = p − j Γ (w)wα (A− j w)dμ(w).
U− j
From here and from (5.42), we find out

2

P j u
2 = |(u, φ j,h )|2 = |ŵ|2 φ̂(A− j w) dμ(w). (5.45)
h∈H U− j
According to the condition m 0 (w) = Un,0 the number j is selected such that A− j w ∈
Un+1 for w ∈ suppû. From here and from (5.31), it is seen that φ̂(A− j w) = 1 for all
w ∈ suppû. Taking into consideration that suppû ⊂ U− j from 5.43 and 5.45, we get
that
ε >
P j u
=
û
=
u
.
Consequently,

f
< ε +
u
< 2ε.
Thus, ⊥
∪ Vj = {0},
and hence 5.42 is true. Thus the implications (b) ⇒ (a) and (c) ⇒ (a) are true. The
inverse implications follow Proposition 5.8 and 5.10 and Theorem 5.2 is proved.
Taking into consideration the H -periodicity of mask m 0 , we note that Condition (30)
is equivalent to the equations
b0 = 1, |bl |2 + |bl+ pn−1 |2 + · · · + |bl+( p−1) pn−1 |2 = 1, 0 ≤ l ≤ p n−1 − 1, (5.46)
Examples of the function φ satisfying Condition 5.10 and generating MRA in L 2 (G)
are given below.
Example 5.1 If a0 = . . . a p−1 = 1p and all aα = 0 for α ≥ p then function φ = 1Un−1

is the solution of Eq. 5.10 in particular when n = 1, Haar function satisfies Eq. 5.10;
(φ = 1U compare with Example 4.1).
Example 5.2 Let p = n = 2 and b0 = 1, b1 = a, b2 = 0, b3 = b, where |a|2 +

|b|2 = 1. Let us propose
(1 + a + b) (1 + a − b) (1 − a − b) (1 − a + b)
a0 = , a1 = , a2 = , a3 = .
4 4 4 4
When α = 0, the modified Cohen condition is fulfilled on the set E = U and the
corresponding solution φ regenerates MRA in L ( G). In particular, when a = 1 and
a = −1, we get, respectively, Haar function : φ(x) = 1U (x) and mixed Haar func-
tion : φ(x) = 1U (x h [1] ). If 0 < |a| < 1, then the solution φ is determined as in
Example 4.2 by decompositon
⎛ ⎞
∞
1 −1 −1
φ(x) = 1U A x ⎝1 + a b w2 j+1 −1 A x ⎠ , x ∈ G.
j
2 j=0
When a = 0, the set U1,1 is blocked, the function φ is determined by the formula
{φ(x) = φ(x h [1] )} and the system {φ(· h) : h ∈ H } is linearly dependent (since
φ(x) = 21 1U (A−1 x)).
Example 5.3 Let p = 2, n = 3 and b0 = 1, b1 = a, b2 = b, b3 = c, b4 = 0, b5 = α,

b6 = β, b7 = γ , where
|a|2 + |α|2 = |b|2 + |β|2 = |c|2 + |γ |2 = 1.
The coefficients a0 , a1 , . . . a7 of Eq. 5.10 are determined simultaneously by the val-

ues b0 , b1 , . . . b7 and the expression of these coefficients through the parameters
a, b, c, α, β, γ is obtained with the help of 5.13 and has the form
1
a0 = (1 + a + b + c + α + β + γ ),
8
1
a1 = (1 + a + b + c − α − β − γ ),
8
1
a2 = (1 + a − b − c + α − β − γ ),
8
1
a3 = (1 + a − b − c − α + β + γ ),
8
1
a4 = (1 − a + b − c − α + β − γ ),
8
1
a5 = (1 − a + b − c + α − β + γ ),
8
1
a6 = (1 − a − b + c − α − β + γ ),
8
1
a7 = (1 − a − b + c + α + β − γ ).
8
Blocking sets for mask

7
m(w) = aα wα∗ (w).
α=0
If a = 0 or c = 0, then the system {φ(· h)|h ∈ H } is linearly dependent. If, how-

ever, a and c are nonzero, then the function regenerates MRA in L 2 (G). Modi-
fied Cohen condition is fulfilled for E = U when abc = 0, and in case a = 0,
b = 0, c = 0 it takes place for E = B(U3,0 ∪ U3,1 ∪ U3,3 ∪ U3,6 ). In particular, when
a = c = 1, we get
1
φ(x) = 1U (y)(1 + w1 (y) + bw2 (y) + w3 (y) + βw6 (y)), y = A−2 x,
4
(compare with 4.36).
Example 5.4 Let us assume that Eq. 5.46 is fulfilled for certain numbers bs , 0 ≤ s ≤
P n − 1. Using Formula (5.13), let us find out the coefficient of mask
pn −1

m 0 (w) = aα wα (w).
α=0
In the sets Un,s , 0 ≤ s ≤ p n − 1 and the value bs . If it is additionally known that

b j = 0 for j ∈ {1, 2, . . . p n−1 − 1}, then Eq. 5.10 with the coefficients found out aα
has the solution φ generating KMA in L 2 (G) (modified Cohen condition is fulfilled
for E = U ).
5.2 Compactly Supported Orthogonal p-Wavelets
As in the previous section, we denote by G the Vilenkin group G p . Let us show that
for each function φ ∈ L 2c (G) satisfying the equation
pn −1

φ(x) = p aα φ(Ax h [α] ) (5.47)
α=0
and generating KMA {V j } ∈ L 2 (G). We may construct the function ψ (1) , . . . ψ ( p−1)
such that the functions
j (ν) j
ψ (ν)
j,h = ψ (A x h), 1 ≤ ν ≤ p − 1, j ∈ Z, h ∈ H,
2
form orthogonal basis in L 2 (G). Such function ψ (1) , . . . ψ ( p−1) will be called orthog-
onal p-wavelets.
Proposition 5.11 Let f ∈ L 2 (G) and function φ generate KMA V j ∈ L 2 (G). For
the function f to belong to V j , it is necessary and sufficient that there exists h-periodic
function m f ∈ L 2 (U ) such that
fˆ(w) = m f (A− j w)φ̂(A− j w) w ∈ G.
This proposition is proved in a way similar to Proposition 5.3. As above, the

orthogonal addition V j in V j+1 will be denoted as W j . In the following variant, the
principle of unitary prolongation is used in finding the algorithm of the structure of
orthogonal p-wavelets along with Theorem 5.2.
Theorem 5.3 Let the scaling function φ ∈ L 2c (G) satisfy Eq. 5.47 and generate an
MRA {V j }. Let us assume that
p−1
M = {m ν (w ⊕ δ)k)}ν,k=0 (5.48)
where m 0 is the mask of Eq. 5.47 and m 1 , m 2 , . . . m p−1 , are certain H -periodic func-
tions from L 2 (U ) is unitary for almost all w ∈ U . Then the functions ψ (ν) determined
by the equations
ψ̂ (ν) (w) = m ν (A−1 w)φ̂(A−1 w), ν = 1, 2, . . . p − 1, (5.49)
belong to W0 and posses orthogonality properties
(ψ (ν) )(· h [α] ), ψ (χ) (·, h [β] ) = δν,χ δα,β ν, χ ∈ {1, 2, . . . p − 1}, α, β ∈ Z+ .
Proof As per Proposition 5.11, under Condition of 5.49, we have ψ (ν) ∈ V1 . Further
as per Proposition 5.4, the equations
5.2 Compactly Supported Orthogonal p-Wavelets 195
(ν)
ψ (·), ψ (ν) (· h) = δθ,h , h ∈ H, ν = 1, 2, . . . p − 1, (5.50)
take place then and only then when ν = 1, . . . p − 1 fulfill the conditions

|ψ̂ (ν) (Aw ⊕ h)|2 = 1, w ∈ G. (5.51)
h∈H
Using 5.49 and 5.51 in the form

|m ν (w ⊕ A−1 h)φ̂(w ⊕ A−1 h)|2 = 1 (5.52)
h∈H
for w ∈ G. Assuming h = w[α] , wk,s = δk⊕w[s] where α = ps + k, δk = A−1 w[k] ,

s ∈ Z+ , 0 ≤ k ≤ p − 1 and using H -periodicity of masks, let us transform the left
part of 5.52 as follows:

m ν w ⊕ A−1 w[α] φ̂ w ⊕ A−1 w[α] m ν w ⊕ A−1 w[α] φ̂ w ⊕ A−1 w[α]
α∈Z+

p−1

= m nu (w ⊕ δk )m ν (w ⊕ δk ) φ̂(w ⊕ wks )φ̂(w ⊕ wks ).
k=0 s∈Z+
From the unitary nature of matrix 5.48 and Proposition 5.4, it follows that the last
expression is equal to zero for almost all w ∈ G. Thus, Eq. 5.51 and, it also means,
5.50 are proved. The equations

φ(· h), ψ (ν) (·) = 0, h ∈ H, ν = 1, . . . , p − 1,
and (ν)
ψ (· h), ψ (χ) (·) = 0, h ∈ H, ν = χ ,
are established similarly. Hence, ψ (ν) ∈ W0 for ν = 1, 2 . . . , p − 1 and the theorem

is proved.
From Theorems 5.1 and 5.2, we get the following algorithm for making the orthogonal
p-wavelets.
Algorithm A
bs , 0 ≤ s ≤ p n − 1
• Step 1 Select the numbers bs , 0 ≤ s ≤ p n − 1, such that the following equation is

satisfied
b0 = 1, |bl |2 + |bl+ pn−1 |2 + · · · + |bl+(n+1) pn−1 |2 = 1, 0 ≤ l ≤ p n−1 − 1. (5.53)

• Step 2 Calculate the coefficients aα , 0 ≤ α ≤ p n − 1 by 5.13 and to verify that the

mask
pn −1

m 0 (w) = aα wα (w)
α=0
satisfies the condition (b) or (c) of Theorem 5.2.

• Step 3 Define the function φ ∈ L 2 (G) such that
∞

φ̂(w) = m 0 (A− j w), w ∈ G. (5.54)
j=1
• Step 4 Find
pn −1

m ν (w) = aα(ν) wα (w), 1 ≤ ν ≤ p − 1,
α=0
p−1
such that the matrix (m ν (w ⊕ δk ))ν,k=0 is unitary.
• Step 5 Define p-wavelets ψ (1) , . . . , ψ ( p−1) by the formula
ψ̂ (ν) (x) = m ν (A−1 w)φ̂(A−1 w), 1 ≤ ν ≤ p − 1.
Using the Fourier’s inverse transform, we have
pn −1

(ν)
ψ (x) = p aα(ν) φ(Ax h [α] ), 1 ≤ ν ≤ p − 1. (5.55)
α=0
We notice that when p = 2, Algorithm A is equivalent to Algorithm A0 examined

in Chap. 4.
Let us show how to calculate the values of function φ̂(w) given by Formula 5.54
by the numbers selected in Step 1. For each 0 ≤ s ≤ p n − 1, let us assume
γ (i 1 , . . . i n ) = bs ,
if
s = i 1 p 0 + i 2 p 1 + · · · + i n p n−1 , i j ∈ {0, 1, . . . p − 1},
and determine numerical sequence {dl : l ∈ N} as follows. Let

k
l= μ j p j , μ j ∈ {0, 1, 2, . . . p − 1}, μk = 0, k = k(l) ∈ Z+ , (5.56)
j=0
then
dl = γ (μ0 , 0, . . . , 0, 0), k(l) = 0;
dl = γ (μ0 , 0, . . . , 0, 0)γ (μ0 , 0, . . . , 0, 0), k(l) = 1;
.................................................
dl = γ (μk , 0, . . . , 0, 0)γ (μk−1 , 0, . . . , 0, 0) . . . γ (μ0 , μ1 , μ2 , . . . μn−2 , μn−1 ), k(l) = 1;
if k = k(l) ≥ n − 1. Further, let us denote by M0 the set of all natural l ≥ p n−1 in

which the p-decomposition 5.56, among the collections (μ j , μ j+1 , . . . μ j+n−1 ) there
is not a single one from the collections
(0, 0, . . . , 0, 1), (0, 0, . . . 0, 2), . . . , (0, 0, . . . 0, p − 1)
and assume M = {1, 2, . . . p n−1 − 1} ∪ M0 . We notice that when n = 2 the set M

has the form
⎧ ⎫
⎨ k ⎬
M= μ j p j : μ j ∈ {1, 2, . . . p − 1}, k ∈ Z+ .
⎩ ⎭
j=0
There takes place the next generalization of the Proposition 4.8.
Proposition 5.12 Let m 0 and dl be determined by bs under Condition 5.53 as shown

above. If the function φ is given by Formula 5.54 then
⎧
⎪
⎨1, w ∈ Un−1,0
φ̂(w) = dl , w ∈ Un−1,l l ∈ M (5.57)
⎪
⎩
0, w ∈ Un−1,l l ∈/ M.
Proof If w ∈ Un−1,0 then A− j w ∈ Un,0 for all j ∈ N but m 0 (w) = 1 on Un,0 . From
here and from 5.54 follows the first equation in 5.57. Further, according to 5.53
b pn−1 = · · · = b( p−1) pn−1 = 0. (5.58)
Let us take l ∈ N with Decomposition 5.56 and find out the least number j0 such that
p j−1 > l + 1. Then, for any w ∈ Un−1,l when j ≥ j 0 we have A− j w ∈ Un,0 and it
means m 0 (A− j w) = 1. From here, according to 5.54 we have
j0 −1

φ̂(w) = m 0 (A− j w), w ∈ Un−1,l . (5.59)
j=1
When l = μ0 (that is, when k(l) = 0), in Formula 5.59 we will have j0 = 2, A−1 w ∈
Un,μ0 and A− j w ∈ Un,0 . Consequently, if l ∈ {1, 2, . . . p − 1} and w ∈ Un−1,l , then
φ̂(w) = b1 = γ (μ0 , 0, . . . 0, 0) = dl .
Similarly, when k(l) = l, we have

m 0 (A−1 w) = γ (μ1 , 0, 0 . . . , 0, 0) m 0 A−2 w = γ (μ0 , μ1 , 0 . . . 0, 0), . . . ,
and this means that φ̂(w) = dl . In general case, from Formula 5.56 for any j ∈ N it
follows that

l 1 μ0 μ j−1
= + · · · + + μ j−1 + μ j p + · · · + μk p k−n
.
p n+ j−1 p n p j−1 p
Besides, if w ∈ Un−1,l , then A− j w ∈ Un+ j−1,l . If l ∈

/ M, then on the strength of 5.58,
then among the factors in 5.59 there is zero. If l ∈ M, then Formula 5.59 will have
A−1 w ∈ Un,μk A−2 w ∈ Un,μk−1 + pμk ,... , A− j0 w ∈ Uμ,ν0 ,
where
ν0 = μ j0 −1 + μ j0 p + · · · μk p k−n .
From here taking into consideration H -periodicity of the polynomial m 0 follows the
formula 5.57. Proposition 5.12 is proved.
Let us show how to carry out Step 4 of the Algorithm A. When p = 2, it is possible
to take just as in the theorem
m 1 (w) = −w1 (w)m 0 (w ⊕ δ1 ).
For p ≥ 3, let us introduce the notation
bs(ν) = m ν (A−n h [s] ), ν = 0, 1, . . . p − 1, s = 0, 1 . . . p n − 1.
According to 5.53, we have
b0(0) = |bl(0) |2 + |bl+

(0) 2 (0)
pn−1 | + · · · + |bl+( p−1) pn−1 | = 1, l = 1, 2, . . . p
2 n−1
− 1.
For carrying out Step 4, it is necessary to find out the coefficients
b0(ν) , b1(ν) , . . . , b(ν)

pn −1 . ν = 1, 2, . . . , p − 1, (5.60)
such that the matrices

⎛ ⎞
bl(0) bl+
(0) (0)
pn−1 . . . . . . bl+( p−1) pn−1
⎜ (1) (1) ⎟
⎜ b (1)
bl+ pn−1 . . . . . . bl+( ⎟
⎜ l p−1) pn−1 ⎟
⎜ ⎟ (5.61)
⎜ ... ... ... ... ... ⎟
⎝ ⎠
( p−1) ( p−1) ( p−1)
bl bl+ pn−1 . . . . . . bl+( p−1) pn−1
will be unitary. For this to happen, it is necessary to use the following proposition
based on Householder transformation (see Novikov et al. [5, §2.6])
r
Proposition 5.13 Let ck ∈ C, k = 1, 2 . . . r , c = 1, k=0 |ck | = 1, and let ck0 =
2
c c
ck , c0 j = c j0 1−c
1−c00
00
, ck j = δk j − 1−c
k0 j0
00
, j, k = 1, . . . r. Then the matrix is unitary.
In addition to Proposition 5.13, let us show that the method of unitary prolongation
of the matrix proposed in [6]. the unitary nature of the matrix 5.61 is of equal force
due to the fact that, for each l, 0 ≤ l ≤ p n−1 − 1, is satisfied the equations:
bl(0) bl(ν) + bl+

(0) (ν) (0) (ν)
pn−1 bl+ pn−1 + · · · + bl+( p−1) pn−1 bl+( p−1) pn−1 = 0, ν = 1, 2, . . . p − 1
(5.62)
(μ) (μ) (μ)
bl(ν) bl + bl+ (ν)
b
pn−1 l+ pn−1 + · · · + b (ν)
b
l+( p−1) pn−1 l+( p−1) pn−1 = 0, ν, μ = 1, 2 . . . p−1
(5.63)
These equations form a system p( p+1) 2
− 1 unknown constant with respect to p( p −
1). Let us determine the matrices
⎛ ⎞ ⎛ ⎞
1 c2
c1
0 ... 0 0 0 0 ... 0
⎜ −c2 1 0 ... ⎟
0⎟ ⎜0⎜ 0 0 ... 0⎟
|c1 | ⎜ c1 ⎟
R0 = ! ⎜ 0 0 0⎟ ⎜ ... 0⎟
⎜ 0 ... ⎟+⎜ 0 0 1 ⎟
|c1 |2 + |c2 |2 ⎝ . . . . . . ... ... ... ⎠ ⎝. . . ... ... ... . . .⎠
0 0 0 ... 0 0 0 0 ... 1
⎛ ⎞
⎛ ⎞ 0 0 0 0 0 ... 0
1 0 c3
... 0 ⎜0 ... 0⎟
c1(1) ⎜ 1 0 0 0 ⎟
⎜ 0 0⎟
|c1(1) | ⎜ 0 0 ... ⎟ ⎜
⎜0 0 1 0 0 ... 0⎟ ⎟
⎜ −c3 0⎟
R1 = " ⎜ (1) 0 1 ... ⎟+⎜ 0 0 0 1 0 ... 0⎟
⎜c ⎟ ⎜
⎜0
⎟
(1) 2
|c1 | + |c3 | ⎝ . 1. .
2
... ... ... ... ⎠ ⎜ 0 0 0 1 ... 0⎟ ⎟
⎝. . . ... ... ... ... ... . . .⎠
0 0 0 ... 0
0 0 0 0 0 ... 1
.........................................................................................
⎛ cp ⎞ ⎛ ⎞
1 0 0 . . . ( p−2) 0 0 0 ... 0 0
⎜ 0 0 0 ... 0 ⎟ ⎜ 0⎟
c1
0 1 0 ... 0
( p−2) ⎜ ⎟ ⎜ ⎟
|c1 | ⎜ 0 0 0 ... 0 ⎟ ⎜ ⎜ 0 0 1 ... 0 0⎟ ⎟
R p−2 =" ⎜ ⎟+
( p−2) 2 ⎜ ... ... ... ... ... ⎟ ⎜ ⎜ ... ... ... ... ... . . .⎟
⎟
|c1 | + |c p | ⎝
2 ⎠ ⎝ 0 0 0 ...
−c p 1 0⎠
0 0 . . . 1
( p−2)
c1 0 0 0 ... 0 0
where c1 = b0(0) , c2 = b(0) pn−1 ,…c p = b"

(0)
( p−1) p(n−1)
! |c(1) | ( p−2) |c
( p−3)
|
c1(1) = |cc11 | |c1 |2 + |c2 |2 , c1(2) = 1(1) |c1(1) | + |c3 |2 . . . c1 = 1( p−3)
" c1 c1
( p−3)
|c1 | + |c p |2
Let us now examine the unitary transformation C P with the matrix R = R p−1 R p−2 . . .
R0 . This transformation converts the plane
c1 z 1 + c2 z 2 + · · · + c p z p = 0
into plane 21 = 1. Let us select in C P the orthogonal system {a1 , a2 , . . . a p−1 } con-
sisting of vectors of the type
a1 = (0, a12 , a13 , . . . a1 p )t

a2 = (0, a22 , a23 , . . . a2 p )t
........................................
a p−1 = (0, a( p−1)2 , a( p−1)3 , . . . a( p−1) p )t
and find out part of the coefficients 5.60 from the equations
(ν) (ν) t
b0 , b pn−1 , . . . , b((0)
p−1) pn−1 = R −1 aν ν = 1, . . . p − 1
Using this algorithm successively for l = 1, . . . , p n−1 − 1 and assuming, respec-

tively,
c1 = bl(0) , c2 = bl+
(0) (0)
pn−1 , . . . c p = bl+( p−1) pn−1 ,
we find out all the remaining coefficients 5.60, which also permit in constructing
p-spurs ψ (1) , . . . , ψ ( p−1)
Example 5.5 Let p = 3, n = 2. According to Step 1 of the algorithm A, let us select
the numbers a, b, c, α, β, γ such that
|a|2 + |b|2 + |c|2 = |α|2 + |β|2 + |γ |2 = 1,
and then propose
b0 = 1, b1 = a, b2 = α, b3 = 0, b4 = b, b5 = β, b6 = 0, b7 = c, b8 = γ .
Then for l = 1, we have

⎛ |a|b
⎞
|a| a
ac
|a|
⎜ |a|b ⎟
|a|
⎜ ⎟
R = ⎜− a √1−|c|2 √1−|c|2 0 ⎟
⎝ ! ⎠
− √ ac 2 − √ bc 2 1 − |c|2
1−|c| 1−|c|
Further, for the numerical values
a = 0.900000, b = 0.435889, c = 0.000000,
α = 0.900000, β = 0.2064742, γ = 0.383886
and vectors
(0, 1, 0)t , l = 0, l = 2
a1 =
(0, 0, 1)t , l = 1

(0, 0, 1)t , l = 0, l = 2
a2 =
(0, 1, 0)t , l = 1,
Using the method shown above, we get b0(1) = 0.000000, b1(1) = 0.000000, b2(1) =
−0.223607, b3(1) = 1.000000, b4(1) = 0.00000, b5(1) = 0.974679, b6(1) = 0.000000,
b7(1) = 1, b8(1) = 0.000000, b0(2) = 0.000000, b1(2) = −0.435890, b2(2) = −0.374166,
b3(2) = 0.000000, b4(2) = 0.900000, b5(2) = −0.085840, b6(2) = 1.000000, b7(2) =
0.000000, b8(1) 0.923380.
According to 5.55, we find out from here
∞

ψ (ν) (x) = 3 aαν φ Ax h [α] , ν = 1, 2
α=0
where aαν is calculated from the values found out with the help of discrete transfor-
mation 5.13.
From Formula 5.55, it is seen that the calculation of values of ψ (ν) (x) it is sufficient
to be able to calculate the values φ(x). For this, the decomposition of function φ(x)
in the Walsh–Fourier series may be useful (see Remark of Chap. 4). According to
5.57, we have
φ̂(w) = 1Un−1 (w) + dl 1Un−1 (w A1−n h [l] ). (5.64)
l∈M
As

(w A1−n h [l] )χ (x, w)dμ(w) = χ (x, A1−n h [l] ) χ (x, w)dμ(w)
G Un−1
= p 1−n 1U (A1−n x)χ (A1−n x, h [l] )

= p 1−n 1U (A1−n x)wl (A1−n x).
The formal application of Fourier inverse transformation to Eq. 5.64 leads to the
decomposition

1
φ(x) = 1U (A1−n x)(1 + dl wl (A1−n x)), x ∈ G. (5.65)
p n−1
l∈M
Example 4.3 shows that the Fourier–Walsh series does not converge absolutely for
certain scaling functions φ. Sufficient condition of absolute and uniform of series
5.64 are formulated in analogy with the Lemma 7.3. The series 5.64 gets transformed
into the final sum in that and only in that case when the function φ belongs to ε(G).
The following criterion takes place.
Proposition 5.14 Let φ ∈ L C2 (G) be a solution of Eq. (5.47) such that φ(θ ) = 1.
Then φ does not belong to ε(G) if and only if there exists a finite sequence
{dk }k=1
V
, N ≤ p n−1 + n − 1, dk ∈ {0, 1, · · · , . p − 1}, such that
(i) d1 = · · · = dn−1 = 0, dn = 0,
(ii) there exists an integer j, n − 1 ≤ j ≤ N − 1, such that d j−s = d N −n for s =
0, · · · , n − 2,
(iii) m 0 (ω(0) )m 0 (ω(1) ) · · · m 0 (ω(N −n) ) = 0, where m 0 is the mask of Eq. (5.47) and
ω(k) ∈ U are determined by the sequence {dk }k=1 N
such that

N
λ(ω(k) ) = dν+k p −v , k = 0, 1, . . . , N − n.
ν=1
The proof of this proposition is essentially the same as in the case p = 2 (see Theo-
rem 4.13).
Example 5.6 Let p = 3, n = 2. Just as in Example 5.5, let us propose
b0 = 1, b1 = a, b2 = α, b3 = 0, b4 = b, b6 = 0, b7 = c, b8 = γ ,
where
|a|2 + |b|2 + |c|2 = |α|2 + |β|2 + |γ |2 = 1.
Then in conformity with step 2 of Algorithm A the coefficients of the mask m 0 has
the form
1
a0 = (1 + a + b + c + α + β + Γ ),
9
1
a1 = (1 + a + α + (b + β)ε32 + (c + γ )ε3 ),
9
1
a2 = (1 + a + α + (b + β)ε3 + (c + γ )ε32 ),
9
1
a3 = (1 + (a + b + c)ε32 + (α + β + γ )ε3 ),
9
1
a4 = (1 + c + β + (a + γ )ε32 + (b + α)ε3 ),
9
1
a5 = (1 + b + γ + (a + β)ε32 + (c + α)ε3 ),
9
1
a6 = (1 + (a + b + c)ε3 + (α + β + γ )ε32 ),
9
1
a7 = (1 + b + γ + (a + β)ε3 + (c + α)ε32 ),
9
1
a8 = (1 + c + β + (a + γ )ε3 + (b + α)ε32 ),
9
where ε3 = ex p(2πi/3). The blocking sets of this mask are as follows:
1. U1,1 when a = c = 0,
when α = β = 0,
2. U1,2
3. U1,2 U1,2 when a = α = 0.
It is seen from here that the scaling function φ formed from the given parameters
regenerate MRA in L 2 (G) in the following cases:
1. a = 0, α = 0,
2. a = 0, α = 0, c = 0,
3. α = 0, a = 0, β = 0.
In each of these cases, the summation in 5.65 is carried out as per the set
⎧ ⎫
⎨k ⎬
M= μ j 3 j : μ j ∈ {1, 2}, k ∈ Z+
⎩ ⎭
j=0
= {1, 2, 4, 5, 7, 8, 13, 14, 16, 17, 22, 23, 25, · · · }.
Let us give the values of certain coefficients d1 = a, d2 = 0, d1 = ab, d5 = aβ,

d7 = αc, d8 = αγ , d13 = ab2 , d14 = abβ, d16 = aβcβ, d17 = aβγ , d22 = abc,
d23 = αβc, d25 = αcγ , d26 = αγ 2 , d40 = ab3 , d41 = ab2 β, d43 = abβc,
d44 = abβγ , d49 = abβγ , d50 = aβ c, d52 = aβcγ , d53 = aβγ 2 ,
2
d67 = ab2 c,
d68 = αbβc, d70 = αβc , d71 = αβcγ , d76 = abcγ , d77 = αβcγ , . . .
2
With help of Proposition 5.14, we find that the series in (5.65) get transformed
into the final sums (and the function φ is step function) in the following two cases:
1. If |α| = 1, |a|2 + |c|2 = 1, b = β = γ = 0, then

φ(x) = (1/3)1v (A−1 x) 1 + aw1 A−1 x + αw2 A−1 x + αcw7 A−1 x .
(5.66)
2. If |a| = 1, |α|2 + |β|2 = 1, b = c = γ = 0, then

φ(x) = (1/3)1v A−1 x 1 + aw1 A−1 x + αw2 A−1 x + aβw5 A−1 x .
(5.67)
According to Proposition 5.6, if in Algorithm A the selected vectors, all the
values of bs with indices (b0 , b1 , . . . , b pn −1 < p n−1 are zero, then the corresponding
mask m 0 satisfies the modified condition of Cohen and consequently the function φ
determined by formula (54) generates MRA. If there is zero among the components
bs , s < p n−1 , then the mask m 0 the conditions (b) or (c) from Theorem 5.2 are verified
in Step 2. In particular, it is necessary to clarify if the mask m 0 has a blocked set.
In case of small p and n, this task is solved by a simple method as in Example 5.6,
but when there is increase in the dimension the computing complexity in the search
for blocked sets increases sharply. In recent works by Lukomskii [7, 8], Berdnikov
and Lukomskii [9] on class of step scaling functions for the search of similar masks
N -valid trees is used. In particular, there is the following proposition (see Lukomskii
[7, 8]).
Proposition 5.15 Let l ∈ {1, 2, . . . , p − 2} where p > 2 is a simple number. Let

us propose that El(0) and El(1) form a division of set {0, 1, . . . , p − 1} such that
El(0) = { j1 , j2 , . . . , jl } and El(1) . For fixed j0 ∈ El(1) , j0 = 0 when n = 2 let us select
in (5.49) numbers such that bs , s = s1 + s2 p, s1 , s2 ∈ {0, 1, . . . , p − 1}, such that
1. b0 = 1,
2. |bs | = 1 for all s ∈ El(1) 0,
3. |b j1 + j0 p | = |b j2 + j1 p | = · · · = |b jl + jl−1 p | = 1.
4. bs = 0 in the rest of the cases.
Further let us determine the coefficients of the mask m 0 by Formula (5.13). Then the
function φ given by the formula (5.50) belonging to the class εl (G) and generates
KMA in L 2 (G).
Let us assume that the scaling function φ and mask m 0 are determined as in
Proposition 5.15. Let us select the mask
p2 −1

m̃ 0 (ω) = ãα wα (ω),
α=0
values bs of which in the sets in Un,s satisfy the condition
b̃0 = 1, |b̃k |2 + |b̃k+ p |2 + · · · + |b̃k+( p−1) p |2 = 1, 0 ≤ k ≤ p − 1
while m̃ 0 (ω) = 0 for all values of ω such that |m 0 (ω)| = 1. Then as in step 3 of
Algorithm A, let us determine the function φ̃ by the formula
∞
ˆ
φ̃(ω) = m˜0 (A− j ω).
j=1
This function φ̃ generates KMA in L 2 (G) (in fact, if the condition b of Theorem 5.2 is
met for m 0 with compact set E, then this is true even for the mask m̃ 0 of the function
φ̃). We also notice that when p = 3, Proposition 5.15 leads to two conditions
1) b0 = 1, |b1 | = |b5 | = 1, bs = 0 for the remaining s(E 1(0) = {2}, E 1(1) =
{0, 1}, j0 = 1);
5.3 Periodic Wavelets on Vilenkin Groups
As above, we write G = G p , N = p n , and ε p = exp(2πi/ p). We recall that the

automorphism A is defined by (Ax) j = x j+1 where x = (x j ) ∈ G. For 0 ≤ k ≤
N − 1, we let xn, k := A−n h [k] and Uk(n) := xn, k + A−n (U ). It is easily seen that

N −1
Uk(n) ∩ Ul(n) = ∅ for k = l, Uk(n) = U.
k=0
Further, we shall use the notation
wl,(n)k := wl (xn, k ) for 0 ≤ l, k ≤ N − 1.
Notice that
wl,(n)k = wk,
(n) −sq (n+1)
l = ε p w pk+s, N q+l , 0 ≤ s, q ≤ p − 1, (5.68)
N −1
N −1

wi,(n)l wi,(n)k = wl,(n)j wk,
(n)
j = N δl, k , 0 ≤ l, k ≤ N − 1. (5.69)
i=0 j=0
A finite sum
N −1

D N (x) := w j (x), x ∈ G,
j=0
is called the Walsh–Dirichlet kernel of order N . It is well known that

#
N , x ∈ U0(n) ,
D N (x) = (5.70)
0, x ∈ U \ U0(n) .
Let us introduce the following spaces

(j)
Vn := span{1, w1 (x), . . . , wN−1 (x)}, Wn := span{wjN (x), wjN+1 (x), . . . , w(j+1)N−1 (x)},
( p−1)
respectively. Note that the orthogonal direct sum of Vn , Wn(1) , . . . , Wn coincides
$ $ ( p−1) $
with Vn+1 , that is, for Wn := Wn(1) · · · Wn , we have Vn Wn = Vn+1 . The
( j)
spaces Vn and Wn will be called the approximation spaces and wavelet spaces ,
respectively.
We can use the discrete Vilenkin–Chrestenson transform to recover v ∈ Vn from
the values v(xn, l ), 0 ≤ l ≤ N − 1. Indeed, if
N −1

v(x) = ck wk (x), x ∈ U, (5.71)
k=0
then
N −1
1
ck = v(xn, l )wl,(n)k , 0 ≤ k ≤ N − 1. (5.72)
N l=0
Suppose that a = (a0 , a1 , . . . , a N −1 ), where ak = 0, 0 ≤ k ≤ N − 1. Then we set
N −1
1
Φ Na (x) := ak wk (x), φn, k (x) := Φ Na (x xn, k ), 0 ≤ k ≤ N − 1, x ∈ G.
N
k=0
Proposition 5.16 Let v ∈ Vn . Assume that
N −1
(n)
αn, k = αn, k (v) := al−1 cl wl,k , 0 ≤ k ≤ N − 1, (5.73)
l=0
where cl are defined as in (5.72). Then
N −1

v(x) = αn, k φn, k (x). (5.74)
k=0
Proof According to the orthogonal relations (5.73), we get
N −1

wl,(n)k φn, k (x) = al wl (x), 0 ≤ l ≤ N − 1,
k=0
and, in view of (5.71), (5.72), and (5.73),
N −1
N −1 N −1

v(x) = al−1 cl wl,(n)j φn, j (x) = αn, k φn, k (x).
l=0 j=0 k=0
Therefore, the expansion in (5.78) is valid for any v ∈ Vn . The proposition is proved.
%n,k are defined by
Remark 5.1 Suppose that φ
5.3 Periodic Wavelets on Vilenkin Groups 207
N −1

%n,0 (x) =
φ a −1 % %
j w j (x), φn,k (x) = φn,0 (x ⊕ x n, k ), k = 1, . . . , N − 1.
j=0
Then {φ%n,k }k=0

N −1 N −1
is a dual shift basis for {φn,k }k=0 . Indeed, using (5.75) and (5.77),
for any v ∈ Vn we have
& '
%n,k ) :=
(v, φ %n,k (x) d x =
v(x), φ %n,0 (x ⊕ xn, k ) d x
cl wl (x) φ
U U l
& '&

'
(n)
= cl wl (x) a l−1 wl,k wl (x) d x = αn, k (v),
U l l
where the last equality follows from the orthogonality of {wk : k ∈ Z+ }.
Let b = (b0 , b1 , . . . , b pN −1 ), where bk = 0 for all 0 ≤ k ≤ pN − 1. In particular,

we can choose
# #
ak/ p if k is divisible by p , ak if k ≤ N − 1 ,
bk = or bk =
1 if k is not divisible by p 1 if 0 ≤ k ≤ pN − 1 .
Then we set
pN −1
1
Φ pN
b
(x) := bk wk (x), φn+1, k (x) := Φ pN
b
(x xn+1, k ), 0 ≤ k ≤ pN − 1, x ∈ G,
pN
k=0
and define
( j)

p−1
ψn, k (x) := ε−p js φn+1, pk+s (x), 0 ≤ k ≤ N − 1, 1 ≤ j ≤ p − 1.
s=0
N −1 ( j)
Let us show that, for each j, the system {ψn,k }k=0 is a bases for the corresponding
( j)
wavelet space Wn .
( j)
Proposition 5.17 Suppose that w ∈ Wn for some j ∈ {1, . . . , p − 1}. Then
N −1
( j)
w(x) = βn,k ψn,k (x), (5.75)
k=0
where with the notations as in (5.77),
βn,k = βn,k (w) = αn+1, pk (w), 0 ≤ k ≤ N − 1. (5.76)

( j) ( j)
Proof Let w ∈ Wn where j ∈ {1, . . . , p − 1}. Then, since Wn ⊂ Vn+1 , as in
Proposition 2.1 we have
( j+1)N
−1
pN −1
p−1 N −1
w(x) = cl wl (x) = αn+1, k (w)φn+1, k (x) = αn+1, pk+s (w)φn+1, pk+s (x), (5.77)
l= j N k=0 s=0 k=0
where
−1 pN −1
1
N
(n+1) (n+1)
αn+1, pk+s (w) = b−1
j N +l c j N +l w j N +l, pk+s , c j N +l = w(xn+1, l )wl, j N +l .
pN
l=0 l=0
Here, in view of (5.72), w(n+1) (n+1) js

j N +l, pk+s = ε p w j N +l, pk , and hence
− js
αn+1, pk+s (w) = ε p αn+1, pk (w), 0 ≤ k ≤ N − 1, 0 ≤ s ≤ p − 1,
which by (5.77) and (5.78) yields (5.75). The proposition is proved.
Let α = 0. For the case where

# #
α if k = 0 or k = N − 1 , α if k = 0 or k = pN − 1 ,
ak = bk =
1 otherwise, 1 otherwise
(5.78)
the assertions which are equivalent to Propositions 5.16 and 5.17 can be found in
Farkov [10]. Notice that the value α = 1 corresponds to the Haar case (so, we use
α = 1 in the sequel).
For each l ∈ {0, 1, . . . , N − 1} with p-ary expansion

n−1
l= ν j p j , ν j ∈ {0, 1, . . . , p − 1},
j=0
n−1
we let γ (l) := j=0 ν j . In Case (5.82), we have the following equalities

p−1
(1 − α) −γ (k)
φn, k (x) = φn+1, pk+s (x) − εp w N −1 (x), (5.79)
s=0
N
& N −1
'
1 − α γ (ν)−γ (k) 1 j ( p−s) ( j)
p−1
1
φn+1, pk+s (x) = φn,k (x) + εp φn,ν (x) + εp ψn,k (x),
p αN p
ν=0 j=1
(5.80)
where 1 ≤ k ≤ N − 1, 0 ≤ s ≤ p − 1. Note also, that w N −1 (x) can be expressed as
N −1 N −1
1 γ (s) p−1
w N −1 (x) = ε φn, s (x) = γn+1, pk+s φn+1, pk+s (x), (5.81)
α s=0 p k=0 s=0
where γn+1, pk+s := w(n+1)

N −1, pk+s .
For any functions f n ∈ Vn and gn ∈ Wn , we write
N −1

p−1 N −1
( j)
f n (x) = Cn,k φn,k (x), gn (x) = gn( j) (x), gn( j) (x) = Dn,k ψn,k (x),
k=0 j=0 k=0
(5.82)
where the coefficient sequences
( j)
Cn = {Cn,k }, D(nj) = {Dn,k }, 1 ≤ j ≤ p − 1, (5.83)
uniquely determine f n and gn , respectively. Let us describe the algorithms, in terms

of the coefficient sequences (5.87), for decomposing f n+1 ∈ Vn+1 as the orthogonal
( j) ( j) ( j)
sum of f n ∈ Vn and gn ∈ Wn , and for reconstructing f n+1 from f n and gn .
As a consequence of (5.80), we observe that
N −1

p−1
( j)
φn+1, pk+s (x) = A(n)
pk+s,ν φn,ν (x) + B (n)
pk+s, j ψn,k (x), (5.84)
ν=0 j=1
where
⎧
⎨ 1/ p + (1 − α)/(αpN ), ν = k,
A(n)
pk+s,ν = B (n) −1 j ( p−s)
pk+s, j = p ε p .
⎩ γ (ν)−γ (k)
εp (1 − α)/(αpN ), ν = k,
Since f n + gn = f n+1 , it follows from (5.82) and (5.84) that
N −1
p−1 N −1
p−1 N −1

( j) ( j)
Cn,ν φn,ν (x) + Dn,ν ψn,ν (x) = Cn+1, pk+s φn+1, pk+s (x)
ν=0 j=1 ν=0 s=0 k=0
⎧ ⎫
⎨N −1
p−1 ⎬
( j)
= Cn+1, pk+s A(n)
pk+s,ν φn,ν (x) + B (n)
pk+s, j ψn,k (x)
⎩ ⎭
s,k ν=0 j=1

p−1
( j)
= Cn+1, pk+s A(n)
pk+s,ν φn,ν (x) + Cn+1, pk+s B (n)
pk+s, j ψn,k (x).
ν s,k j=1 s,k
This implies that

Cn,ν = A(n)
pk+s,ν C n+1, pk+s ,
( j)
Dn,ν = B (n)
pk+s, j C n+1, pk+s . (5.85)
s,k s,k
Now, using (5.79) and (5.80), we obtain
N −1
p−1
φn,ν (x) = Q (n)
pk+s,ν φn+1, pk+s (x),
k=0 s=0
where ⎧ γ (k)
⎨ 1 − ε p (1 − α)γn+1, pk+s /N , k = ν,
Q (n)
pk+s,ν =
⎩ γ (k)
−ε p (1 − α)γn+1, pk+s /N , k = ν.
Therefore, we have

Cn+1, pk+s φn+1, pk+s (x) = Cn,ν Q (n)
pk+s,ν φn+1, pk+s (x)
k,s ν k,s
p−1 N −1
p−1
( j)

+ Dn,k ε−p js φn+1, pk+s (x)
j=1 k=0 s=0
⎧ ⎫
⎨ ( j)
⎬
= Q (n)
pk+s,ν C n,ν + ε−p js Dn,k φn+1, pk+s (x)
⎩ ⎭
k,s ν j
and so ( j)
Cn+1, pk+s = Q (n)
pk+s,ν C n,ν + ε−p js Dn,k . (5.86)
ν j
We remark that the decomposition and reconstruction algorithms based on For-

mulas (5.85) and (5.86) have more simple structure than the similar algorithms con-
structed in Chui and Mhaskar [11] for the case of trigonometric wavelets.
To conclude this section, let us consider the case where p = 2, N = 2n , and
#
ak , 0 ≤ k ≤ N − 1,
bk = (5.87)
a N −k , N ≤ k ≤ 2N − 1;
recall here that all ak = 0. Then, for any k ∈ {0, 1, . . . , N − 1},
φn,k (x) = φn+1,2k (x) + φn+1,2k+1 (x), ψn,k (x) = φn+1,2k (x) − φn+1,2k+1 (x),
and thus
1 1
φn+1,2k (x) = [φn,k (x) + ψn,k (x)], φn+1,2k+1 (x) = [φn,k (x) − ψn,k (x)].
2 2
So, under Condition (5.87), instead of (5.85) and (5.86) we obtain the classical Haar
discrete transforms.
5.4 Periodic Wavelets Related to the Vilenkin–Christenson

Transform
Let us denote by k p the remainder from the division of the integer k by the natural
number p, and let [ a ] be the integer part of a number a. For any a ∈ R+ , the digits
of the p -adic expansion
∞
∞

ν−1
a= a−ν p + aν p −ν (5.88)
ν=1 ν=1
are defined by a−ν = [ p 1−ν a ] p , aν = [ p ν a ] p (so, the finite representation

for a p -adic rational a is taken). We can easily see that, for each a ∈ R+ there exists
a natural number μ such that a−ν = 0 for all ν > μ as well as that the first sum in
(5.88) is equal to [ a ]. The representation (5.92) induces the operation of addition
modulo p (or p -adic addition) on R+ as follows:
∞
∞

a ⊕ p b := a−ν + b−ν p p ν−1 + aν + bν p p −ν , a, b ∈ R+ .
ν=1 ν=1
As usual, the equality c = a p b means that c ⊕ p b = a.

For N = p n , we set Z N = {0, 1, . . . , N − 1}. Suppose that the space C N con-
sists of complex sequences x = (. . . , x(−1), x(0), x(1), x(2), . . . ), such that x( j +
N ) = x( j) for all j ∈ Z. An arbitrary sequence x from C N is given if the values of
x( j) are given for j ∈ Z N ; therefore, the element x is often identified with the vector
(x(0), x(1), . . . , x(N − 1)). The space C N is equipped with the following natural
inner product:
N −1

x, y := x( j)y( j).
j=0
For an arbitrary j ∈ Z N , let j ∗ denote the nonnegative integer defined by the

condition j ⊕ p j ∗ = 0. For p = 2, we have j ∗ = j, and, for p > 2, the number
j ∗ is p -adic opposite to j. For each x ∈ C N we denote by % x the vector from C N
such that %
x ( j) = x( j ∗ ) for all j ∈ Z . Further, for k, j ∈ Z , we set {k, j} :=
n N N p
ν=1 k ν−n−1 j−ν , where

n
n
k= k−ν p ν−1 , j= j−ν p ν−1 , k−ν , j−ν ∈ {0, 1, . . . , p − 1}.
ν=1 ν=1
The Vilenkin–Chrestenson functions w0(N ) , w1(N ) , . . . , w(N )

N −1 for the space C N are
{k,
defined by the equalities wk(N ) ( j) = ε p p and wk(N ) (l) = wk(N ) (l + N ), where k, j ∈
j}
Z N , l ∈ Z. For n ≥ 2 and p = 2, the Vilenkin–Chrestenson functions coincide with

Walsh functions and, in the case n = 1 and p ≥ 2, they are exponential functions:
( p) kj
wk ( j) = ε p , k, j ∈ {0, 1, . . . , p − 1} .
The functions w0(N ) , w1(N ) , . . . , w(N )
N −1 constitute an orthogonal basis in C N and
wk(N )
2 = N for all k ∈ Z N . To an arbitrary vector x from C N the Vilenkin–
Chrestenson transform by the definition) assigns the sequence ( x of the Fourier
coefficients of x in the system w0(N ) , w1(N ) , . . . , w(N )
N −1 :
N −1
1
(
x (k) := x( j)wk(N ) ( j), k ∈ Z N .
N j=0
For all x, y ∈ C N , we define the p -convolution x ∗ y by the formula
N −1

(x ∗ y)(k) := x(k p j)y( j), k ∈ Z N .
j=0
By a unit N -periodic impulse we mean the vector δ N from C N defined by the

equality #
1, if j is divisible by N ,
δ N ( j) :=
0, if j is not divisible by N .
The system of shifts {δ N (· p k) | k ∈ Z N } is an orthonormal basis in C N and
N −1

x( j) = (x ∗ δ N )( j) = x(k)δ N ( j p k), j ∈ ZN ,
k=0
for all x ∈ C N . For each k ∈ Z N the p -adic shift operator Tk : C N → C N is defined

as
(Tk x)( j) := x( j p k), x ∈ C N , j ∈ Z N .
It follows from the definitions that, for all x, y ∈ C N , the following relations hold:
x, y = N ( y , x
x,( ∗y = N(
x(y,
(T (N )
k x)(l) = wk (l) (
)
x (l), (%x )(k) = (
x (k),
(5.89)
y, Tk x = y ∗ %
x (k), Tk x, Tl y = x, Tl p k y, k, l ∈ Z N . (5.90)
For ν = 0, 1, . . . , n, we set Nν = N / p ν and Δν = p ν−1 . The operators D :

C N → C N1 and U : C N1 → C N given by the formulas
(Dx)( j) := x( pj), j = 0, 1, . . . , N1 − 1,
5.4 Periodic Wavelets Related to the Vilenkin–Christenson Transform 213
and #
y( j/ p) if j is divisible by p ,
(U y)( j) :=
0 if j is not divisible by p ,
where x ∈ C N and y ∈ C N1 are called the thickening sampling operator and the
thinning sampling operator, respectively. Note that D(U y) = y for all y ∈ C N1 . Fur-
ther, suppose that D 1 = D, U 1 = U and, for ν = 2, . . . , n, we define the operators
D ν : C N → C Nν and U ν : C Nν → C N by the formulas
#
ν ν ν y( j/ p ν ) if j is divisible by p ν ,
(D x)( j) := x( p j), (U y)( j) :=
0 if j is not divisible by p ν ,
where x ∈ C N and y ∈ C Nν For any y ∈ C Nν , the following relation holds: U ν y(l) =

−ν
p ( y(l), l ∈ Z N , where, on the left-hand side, the Vilenkin–Chrestenson transform
is taken in C N , while, on the right-hand side, it is taken in C Nν . In addition,
U ν (x ∗ y) = U ν (x) ∗ U ν (y) for all x, y ∈ C Nν . (5.91)
Following the approach from Frazier [12], we introduce the concept of the wavelet
basis of the first stage in C N :
Definition 5.1 Suppose that u 0 , u 1 , . . . , u p−1 ∈ C N . If the system
N1 −1 N1 −1 N1 −1
B(u 0 , u 1 , . . . , u p−1 ) = {T p k u 0 }k=0 ∪ {T p k u 1 }k=0 ∪ · · · ∪ {T p k u p−1 }k=0
is an orthonormal basis in C N , then B(u 0 , u 1 , . . . , u p−1 ) is called the wavelet basis

of the first stage in C N generated by the collection of vectors u 0 , u 1 , . . . , u p−1 .
The following theorem characterizes all collections of vectors generating wavelet

bases of the first stage in C N .
Theorem 5.4 The collection of vectors u 0 , u 1 , . . . , u p−1 generates a wavelet basis

of the first stage in C N if and only if the matrix
⎛ ⎞
(
u 0 (l) (
u 1 (l) ... (
u p−1 (l)
⎜(u 0 (l + N1 ) (
u 1 (l + N1 ) ... (
u p−1 (l + N1 ) ⎟
N ⎜
⎜(
⎟
⎟
√ ⎜ u 0 (l + 2N1 ) (
u 1 (l + 2N1 ) ... (
u p−1 (l + 2N1 ) ⎟ (5.92)
p⎜ .. .. .. ⎟
⎝ . . ... . ⎠
u 0 (l + ( p − 1)N1 ) (
( u 1 (l + ( p − 1)N1 ) . . . (
u p−1 (l + ( p − 1)N1 )
is unitary for l = 0, 1, . . . , N1 − 1.
Using Theorem 5.4, for each 1 ≤ m ≤ n we define the following procedure for the
construction of a wavelet basis of the first stage in C N .
Step 1. Choose complex numbers bl , 0 ≤ l ≤ p m − 1, satisfying the condition

p−1
|bl+kp m−1 |2 = 1, l = 0, 1, . . . , p m−1 − 1. (5.93)
k=0
Step 2. Calculate a0 , . . . , a pm −1 by the formulas
pm −1
( pm )
−m+1/2
aj = p bl wl ( j), j = 0, 1, . . . , p m − 1.
l=0
Step 3. Define a vector u 0 ∈ C N , for which

#
a j , 0 ≤ j ≤ p m − 1,
u 0 ( j) = (5.94)
0, p m ≤ j ≤ p n − 1.
Step 4. Find vectors u 1 , . . . , u p−1 ∈ C N such that, for all l = 0, 1, . . . , N1 − 1,

Matrix (5.92) is unitary.
For p = 2, the last step of this procedure is carried out simply by the formula
u 1 ( j) = (−1) j u 0 (1 ⊕ 2 j), j ∈ ZN . (5.95)
One of algorithms for the realization of step 4 in the case p > 2 is based on the
Haushölder transform [5, § 2.6] and can be described by the formulas
u 0 (l)
1 −( u 0 (l + j N1 )(
( u 0 (l + k N1 )
u k (l) = (
( u 0 (l + k N1 ) u k (l + j N1 ) = δk j −
, ( ,
1 −(
u 0 (l) 1 −(
u 0 (l)
(5.96)
where δk j is the Kronecker delta, k, j = 1, 2, . . . , p − 1 and l = 0, 1, . . . , N1 − 1.
Example 5.7 Suppose that N > p . Take m = 1 and b0 = 1, b1 = · · · = b p−1 = 0.

Then the system B(u 0 , u 1 , . . . , u p−1 ) is generated by the vectors

u μ = p −1/2 1, ε pμ , ε2μ ( p−1)μ
p , . . . , εp , 0, 0, . . . , 0 , μ = 0, 1, . . . , p − 1.
In particular, for p = 2, we have the Haar basis of the first stage in C N :

√ √ √ √
u 0 = (1/ 2, 1/ 2, 0, 0, . . . , 0), u 1 = (1/ 2, −1/ 2, 0, 0, . . . , 0).
The following example is obtained by modifying the orthogonal wavelets con-

structed in Example 5.14; it corresponds to the case m = p = 2, b0 = 1, b1 = a,
b2 = 0, b3 = b.
Example 5.8 Suppose that a and b are complex numbers such that |a|2 + |b|2 = 1.
Suppose that p = 2 and N ≥ 4, , and the vectors u 0 , u 1 ∈ C N are given by the
equalities
1+a+b 1+a−b 1−a−b 1−a+b

u 0 (0) = √ , u 0 (1) = √ , u 0 (2) = √ , u 0 (3) = √ ,
2 2 2 2 2 2 2 2
1+a−b 1+a+b 1−a+b 1−a−b

u 1 (0) = √ , u 1 (1) = − √ , u 1 (2) = √ , u 1 (3) = − √ ,
2 2 2 2 2 2 2 2
under the condition that u 0 ( j) = u 1 ( j) = 0 for 4 ≤ j ≤ N − 1. Then the vectors

u 0 , u 1 generate a wavelet basis of the first stage in C N . Note that, for a = 1, b = 0,
the resulting wavelet basis B(u 0 , u 1 ) coincides with the Haar wavelet basis of the
first stage described in Example 5.7.
The following two examples are similar to Examples 5.3 and 5.4, respectively.
Example 5.9 Suppose that p = 2, n > 3, and m = 3. We set
1
( b0 , b1 , . . . , b7 ) = ( 1, a, b, c, 0, α, β, γ ),
2
where |a|2 + |α|2 = |b|2 + |β|2 = |c|2 + |γ |2 = 1. Then, by relation (5.94), we

have
1 1
u 0 (0) = √ (1 + a + b + c + α + β + γ ), u 0 (1) = √ (1 + a + b + c − α − β − γ ),
4 2 4 2
1 1
u 0 (2) = √ (1 + a − b − c + α − β − γ ), u 0 (3) = √ (1 + a − b − c − α + β + γ ),
4 2 4 2
1 1
u 0 (4) = √ (1 − a + b − c − α + β − γ ), u 0 (5) = √ (1 − a + b − c + α − β + γ ),
4 2 4 2
1 1
u 0 (6) = √ (1 − a − b + c − α − β + γ ), u 0 (7) = √ (1 − a − b + c + α + β − γ ).
4 2 4 2
Further, we set u 1 ( j) = u 0 ( j) = 0 for 8 ≤ j ≤ 2n − 1 , and we choose the other

components of the vector u 1 so that relation (5.95) are valid, i.e.,
u 1 (0) = u 0 (1), u 1 (1) = −u 0 (0), u 1 (2) = u 0 (3), u 1 (3) = −u 0 (2),
u 1 (4) = u 0 (5), u 1 (5) = −u 0 (4), u 1 (6) = u 0 (7), u 1 (7) = −u 0 (6).
The resulting pair u 0 , u 1 generates a wavelet basis of the first stage in C N .

Example 5.10 Suppose that p = 3, n > 2, m = 2 and
1
( b0 , b1 , . . . , b8 ) = √ ( 1, a, α, 0, b, β, 0, c, γ ),
3
where |a|2 + |b|2 + |c|2 = |α|2 + |β|2 + |γ |2 = 1. Then, using (5.93) and (5.94),
we obtain
1
u 0 (0) = √ (1 + a + b + c + α + β + γ ),
3 3
1
u 0 (1) = √ 1 + a + α + (b + β)ε32 + (c + γ )ε3 ,
3 3
1
u 0 (2) = √ 1 + a + α + (b + β)ε3 + (c + γ )ε32 ,
3 3
1
u 0 (3) = √ 1 + (a + b + c)ε32 + (α + β + γ )ε3 ,
3 3
1
u 0 (4) = √ 1 + c + β + (a + γ )ε32 + (b + α)ε3 ,
3 3
1
u 0 (5) = √ 1 + b + γ + (a + β)ε32 + (c + α)ε3 ,
3 3
1
u 0 (6) = √ 1 + (a + b + c)ε3 + (α + β + γ )ε32 ,
3 3
1
u 0 (7) = √ 1 + b + γ + (a + β)ε3 + (c + α)ε32 ,
3 3
1
u 0 (8) = √ 1 + c + β + (a + γ )ε3 + (b + α)ε32 ,
3 3
where ε3 = exp(2πi/3). We set u 0 ( j) = u 1 ( j) = u 2 ( j) = 0 for 9 ≤ j ≤ 3n − 1 and

use formulas (5.96) to define the other components of the vectors u 1 , u 2 ∈ C N so
that the matrix ⎛ ⎞
(
u (l) (
u 1 (l) (
u 2 (l)
9 ⎝ 0
√ u 0 (l + 3) (
( u 2 (l + 3) ⎠
u 1 (l + 3) (
3 ( u 0 (l + 6) (
u 1 (l + 6) (
u 2 (l + 6)
is unitary for l = 0, 1, 2. The resulting collection of the vectors u 0 , u 1 , u 2 generates

a wavelet basis of the first stage in C N .
Definition 5.2 Suppose that m ∈ N, m ≤ n. By a sequence of orthogonal wavelet

filters of the mth stage we mean a sequence of vectors
u (1) (1) (1) (m) (m) (m)

0 , u 1 , . . . , u p−1 , . . . , u 0 , u 1 , . . . , u p−1 ,
such that u (ν)

μ ∈ C Nν−1 for ν = 1, 2, . . . , m , μ = 0, 1, . . . , p − 1 and the matrices
⎛ (ν) (ν) (ν)
⎞
u 0 (l)
( (
u 1 (l) ... (
u p−1 (l)
⎜ ⎟
⎜ (ν) (ν) (ν) ⎟
⎜(u (l + Nν ) (
u 1 (l + Nν ) ... (
u p−1 (l + Nν ) ⎟
⎜ 0 ⎟
N ⎜ (ν) ⎟
A(ν) (l) := √ ⎜ (
u (l + 2Nν ) (
(ν)
u 1 (l + 2Nν ) ... (
(ν)
u p−1 (l + 2Nν ) ⎟
p⎜ 0 ⎟
⎜ ⎟
⎜ ... ... ... ... ⎟
⎝ ⎠
(ν) (ν) (ν)
(
u 0 (l + ( p − 1)Nν ) (
u 1 (l + ( p − 1)Nν ) ... (
u p−1 (l + ( p − 1)Nν )
are unitary for ν = 1, 2, . . . , m , l = 0, 1, . . . , Nν − 1.

Theorem 5.5 Suppose that the collection of vectors u 0 , u 1 , . . . , u p−1 generates a
wavelet basis of the first stage in C N . For a given m ∈ N, m ≤ n, set
Δ
ν −1
u (1) (ν) −1
μ ( j) = u μ ( j), u μ ( j) = Δν u (1)
μ ( j + k Nν−1 ), j ∈ Z Nν−1 , (5.97)
k=0
where ν = 2, . . . , m, μ = 0, 1, . . . , p − 1. Then the vectors
u (1) (1) (1) (m) (m) (m)

0 , u 1 , . . . , u p−1 , . . . , u 0 , u 1 , . . . , u p−1 ,
constitute a sequence of orthogonal wavelet filters of the mth stage.

Thus, from a given vector u 0 ∈ C N , defined by (5.93) and (5.94) we can, first,
find a wavelet basis of the first stage u 0 , u 1 , . . . , u p−1 , using (5.95) or (5.96), and
then, using (5.97) obtain the sequence of orthogonal wavelet filters of the mth stage.
Denote by ⊕ the direct sum of the subspaces of the space C N . By the theorem that
follows, from any sequence of orthogonal wavelet filters of the mth stage we can
construct an orthonormal wavelet basis in C N .
Theorem 5.6 Suppose that a sequence of orthogonal wavelet filters of the mth stage
is given in the space C N :
u (1) (1) (1) (m) (m) (m)

0 , u 1 , . . . , u p−1 , . . . , u 0 , u 1 , . . . , u p−1 .
Let φ (1) = u (1) (1) (1)

0 , ψμ = u μ , μ = 1, . . . , p − 1, and define φ
(ν)
, ψμ(ν) for ν =
2, . . . , m, μ = 1, . . . , p − 1 by the formulas
φ (ν) = φ (ν−1) ∗ U ν−1 u (ν) (ν)

0 , ψμ = φ
(ν−1)
∗ U ν−1 u (ν)
μ .
Further, for ν = 1, . . . , m, μ = 1, . . . , p − 1, we set

(μ)
φ−ν,k = T pν k φ (ν) , ψ−ν,k = T pν k ψμ(ν) , k = 0, 1, . . . , Nν − 1,
and define the subspaces

Nν −1 (μ) Nν −1 (μ)
V−ν = span{φ−ν,k }k=0 , W−ν = span{ψ−ν,k }k=0 ,
(1) ( p−1)
W−ν = W−ν ⊕ · · · ⊕ W−ν .
Then the following expansion holds:
C N = W−1 ⊕ W−2 ⊕ · · · ⊕ W−m ⊕ V−m (5.98)
and, for each ν = 1, 2, . . . , m the following properties are valid :

(a) V−ν = V−ν−1 ⊕ W−ν−1 ;
Nν −1
(b) {φ−ν,k }k=0 is an orthonormal basis in V−ν ;
(1) Nν −1 ( p−1) Nν −1
(c) {ψ−ν,k }k=0 ∪ · · · ∪ {ψ−ν,k }k=0 is an orthonormal basis in W−ν .
This theorem justifies the method of constructing subspaces V−1 , . . . , V−n in C N

with the following properties:
(i) V−ν−1 ⊂ V−ν for all ν ∈ {1, 2, . . . n} ;
(ii) for each ν ∈ {1, 2, . . . n}, there exists a vector φ (ν) ∈ V−ν such that the system
Nν −1
{T pν k φ (ν) }k=0 is an orthonormal basis in V−ν ;
(iii) for each 1 ≤ m ≤ n, relation (5.98) is valid;
(iv) for each ν ∈ {1, 2, . . . n} there exist vectors ψ1(ν) , . . . , ψ p−1
(ν)
∈ W−ν such that the
p−1 Nν −1
system μ=1 {T pν k ψμ(ν) }k=0 is an orthonormal basis in W−ν .
According to the terminology used in the theory of multiresolution analysis and
the sequence {φ (ν) }nν=1 in property (ii), it is natural to call a scaling sequence in C N .
In particular, for p = 2, n = 3 by Theorem 5.6 we obtain three orthonormal wavelet
bases in C8 :
{ψ−1,k }3k=0 ∪ {φ−1,k }3k=0 (m = 1),
{ψ−1,k }3k=0 ∪ {ψ−2,k }1k=0 ∪ {φ−2,k }1k=0 (m = 2),
{ψ−1,k }3k=0 ∪ {ψ−2,k }1k=0 ∪ {ψ−3,0 } ∪ {φ−3,0 } (m = 3).
In the Haar case (see Example 5.7), these bases consist of the vectors
1 1
φ−1,0 = √ (1, 1, 0, 0, 0, 0, 0, 0), ψ−1,0 = √ (1, −1, 0, 0, 0, 0, 0, 0),
2 2
1 1
φ−1,1 = √ (0, 0, 1, 1, 0, 0, 0, 0), ψ−1,1 = √ (0, 0, 1, −1, 0, 0, 0, 0),
2 2
1 1
φ−1,2 = √ (0, 0, 0, 0, 1, 1, 0, 0), ψ−1,2 = √ (0, 0, 0, 0, 1, −1, 0, 0),
2 2
1 1
φ−1,3 = √ (0, 0, 0, 0, 0, 0, 1, 1), ψ−1,3 = √ (0, 0, 0, 0, 0, 0, 1, −1),
2 2
1 1
φ−2,0 = (1, 1, 1, 1, 0, 0, 0, 0), ψ−2,0 = (1, 1, −1, −1, 0, 0, 0, 0),
2 2
1 1
φ−2,1 = (0, 0, 0, 0, 1, 1, 1, 1), ψ−2,1 = (0, 0, 0, 0, 1, 1, −1, −1),
2 2
1 1
φ−3,0 = √ (1, 1, 1, 1, 1, 1, 1, 1), ψ−3,0 = √ (1, 1, 1, 1, −1, −1, −1, −1).
2 2 2 2
In the general case, the orthogonal projections P−ν : C N → V−ν and Q −ν :

C N → W−ν act by the formulas

N ν −1 p−1 Nν −1
(μ) (μ)
P−ν x = x, φ−ν,k φ−ν,k , Q −ν x = x, ψ−ν,k ψ−ν,k , x ∈ C N .
k=0 μ=1 k=0
(5.99)
Suppose that I is the identity operator on C N . Setting P0 = I , V0 = C N and using
Theorem 5.6 for any x ∈ C N , we obtain the equalities
ν

x = P−ν x + Q −k x, P−ν+1 x = P−ν x + Q −ν x, ν = 1, 2, . . . , n.
k=1
An arbitrary vector x from C N can be regarded as the input signal a0 = x and,

for ν = 1, 2, . . . , m, we can set
u (ν)
aν = D(aν−1 ∗ % 0 ), dν
(μ)
u μ(ν) ), μ = 1, . . . , p − 1.
= D(aν−1 ∗ % (5.100)
We can easily see that the components of the vectors aν and dν(μ) are the coefficients of
the expansions (5.100) for a chosen x. The application of formulas (5.105) constitutes
the phase of the analysis of the signal x and yields the collection of vectors
d1(1) , . . . , d p−1
(1)
, . . . , d1(m) , . . . , d p−1
(m)
, am . (5.101)
The inverse passage from the collection (5.101) to the original vector x constitutes
the reconstruction phase and is defined by the formulas

p−1
aν−1 = u (ν)
0 ∗ U aν + u (ν) (ν)
μ ∗ U dμ , ν = m, m − 1, . . . , 1. (5.102)
μ=1
Therefore, the formulas (5.100) and (5.102) specify the direct and inverse
discrete wavelet transforms associated with the sequence of wavelet filters u (1)
0 ,
(1) (1) (m) (m) (m)
u 1 , . . . , u p−1 , . . . , u 0 , u 1 , . . . , u p−1 .
It is easily seen that

N −1
Uk(n) ∩ Ul(n) = ∅ for k = l, Uk(n) = U.
k=0
Further, we shall use the notation
wl,(n)k := wl (xn, k ) for 0 ≤ l, k ≤ N − 1.
Notice that
wl,(n)k = wk,
(n) −sq (n+1)
l = ε p w pk+s, N q+l , 0 ≤ s, q ≤ p − 1, (5.103)
N −1
N −1

wi,(n)l wi,(n)k = wl,(n)j wk,
(n)
j = N δl, k , 0 ≤ l, k ≤ N − 1. (5.104)
i=0 j=0
A finite sum
N −1

D N (x) := w j (x), x ∈ G,
j=0
is called the Walsh–Dirichlet kernel of order N . It is well known that

#
N , x ∈ U0(n) ,
D N (x) = (5.105)
0, x ∈ U \ U0(n) .
( j)
Vn := span{1, w1 (x), . . . , w N −1 (x)}, Wn := span{w j N (x), w j N +1 (x), . . . , w( j+1)N −1 (x)},
( p−1)
respectively. Note that the orthogonal direct sum of Vn , Wn(1) , . . . , Wn coincides
$ $ ( p−1) $
with Vn+1 , that is, for Wn := Wn(1) · · · Wn , we have Vn Wn = Vn+1 . The
( j)
spaces Vn and Wn will be called the approximation spaces and wavelet spaces ,
respectively.
We can use the discrete Vilenkin–Chrestenson transform to recover v ∈ Vn from
the values v(xn, l ), 0 ≤ l ≤ N − 1. Indeed, if
N −1

v(x) = ck wk (x), x ∈ U, (5.106)
k=0
then
N −1
1
ck = v(xn, l )wl,(n)k , 0 ≤ k ≤ N − 1. (5.107)
N l=0
Suppose that a = (a0 , a1 , . . . , a N −1 ), where ak = 0, 0 ≤ k ≤ N − 1. Then we set
N −1
1
Φ Na (x) := ak wk (x), φn, k (x) := Φ Na (x xn, k ), 0 ≤ k ≤ N − 1, x ∈ G.
N
k=0
Proposition 5.18 Let u ∈ C N , ν ∈ {0, 1, . . . , n}. The system is orthonormal then

Nν −1
only then when {T p ν k u}k=0 CN
pν
|(
u (l)|2 + |( u (l + ( p ν − 1)Nν )|2 =
u (l + Nν )|2 + · · · + |( (5.108)
N2
for all l ∈ Z Nν .

Proof According to (5.89), (T (N )
p ν k u)(l) = w p ν k (l) (
u (l). Using Parseval’s equation,
we get
ν −1

N −1 p ν −1
N
(N ) (N )
u, T p ν k u =N u (l)|2 w p ν k (l) = N
|( u (l + μNν )|2 wk
|( ( p ν (l + μNν ))
l=0 μ=0 l=0
ν
Nν −1 p −1
1 N2
= w(N
ν
)
(l) |(
u (l + μNν )|2 . (5.109)
Nν l=0 p k p ν μ=0
Thus the derivative Nν u, T p ν k u coincides with the inverse Vilenkin–Chrestenson

transform in C Nν of sequence
ν
p −1
N2
Φν (l) := ν |(
u (l + μNν )|2 , l ∈ Z Nν .
p μ=0
It is easy to see that

N ν −1
N ν −1
w(N )
p ν k (l) = wl(Nν ) (k) = Nν δ Nν (k), k ∈ Z Nν .
l=0 l=0
Consequently, the equation

#
1 k = 0,
u, T pνk u =
0 k = 1, . . . , Nν − 1,
is fulfilled in that and only in that case when
p ν −1
pν
|(
u (l + μNν )|2 = , l ∈ Z Nν .
μ=0
N2
Besides, using (5.90) for any k, l ∈ Z Nν we have
T p ν k u, T p ν l u = u, T p ν (l p k) u.
Nν −1
It means, the orthonormality of the system {T p ν k u}k=0 is equivalent to the condition
(5.108). The proposition 5.18 is proved.
Proof of Theorem 5.4. For μ = 0, 1, . . . , pi − 1, λ = 0, 1, 2, . . . , N1 , we have
N1 −1 N1 −1
1 1
p−1
(N1 ) ( pN )
û (2)
μ (l) = (2)
u ( j)wl ( j) = u (1) ( j + k N1 )w pl 1 ( j)
N1 j=0 μ pN1 j=0 k=0 μ
N1 −1
1 1
p−1
(N1 ) (N1 )
u (1)
μ ( j + k N1 )w pl ( j) = j = 0 N −1 u (1)
μ ( j)w pl ( j).
N1 j=0 k=0 N
Using the induction for any ν ∈ {2, . . . , m}, we get
ˆ
u (ν) (1)
μ (l) = û μ ( p
ν−1
l), μ = 0, 1, . . . , p − 1, l ∈ Z Nν−1 .
Hence for any ν ∈ {1, 2, . . . , m} true is the equation A(ν) (l) = A(1) ( p ν−1l), where,
on the strength of Theorem , matrix A(1) ( p ν−1l) is unitary for all l ∈ Z N . Theorem 5.4
is proved.
Proof of Theorem 5.5. According to Theorem 5.4, the systems
N1 −1 N1 −1
{T p k u 0 }k=0 , . . . , {T p k u p−1 }k=0
are orthonormal in C N then only then when the column vectors

⎛ ⎞ ⎛ ⎞
(u 0 (l) (u p−1 (l)
⎜ u 0 (l + N1 )
( ⎟ ⎜ (u p−1 (l + N1 ) ⎟
N ⎜
⎜
⎟ N ⎜ ⎟
√ ⎜ ( u 0 (l + 2N1 ) ⎟ ⎟, ...,
⎜
√ ⎜ ( u p−1 (l + 2N1 ) ⎟ ⎟ (5.110)
p⎝ ... ⎠ p⎝ ... ⎠
u 0 (l + ( p − 1)N1 )
( u p−1 (l + ( p − 1)N1 )
(
have length equal to 1. Besides , as per Proposition 5.4.2, the condition
T p k u μ , T p j u ν = 0, μ = ν, k, j ∈ Z N1 ,
is uniform on the paired orthogonality of vectors. It is to be noted that the number

of elements of the system (5.110). B(u 0 , u 1 , . . . , u p−1 ) for the given set of vectors
coincides with the dimension of space C N .
For μ = 0, 1, . . . , p − 1, l = 0, 1, . . . , N1 − 1,we have
N1 −1 N1 −1
1 1
p−1
(N1 ) ( pN )
u μ(2) (l) =
( u (2) ( j)w ( j) = u (1) ( j + k N1 )w p l 1 ( j)
N1 j=0 μ l
pN1 j=0 k=0 μ
N1 −1 N −1
1 1 (1)
p−1
(N )
= u (1) ( j + k N )w ( j) = u ( j)w(N )
u μ(1) ( p l).
p l ( j) = (
N j=0 k=0 μ N j=0 μ
1 pl
Using the induction for any ν ∈ {2, . . . , m} we get
u μ(ν) (l) = (
( u μ(1) ( p ν−1l), μ = 0, 1, . . . , p − 1, l ∈ Z Nν−1 .
Hence for any ν ∈ {1, 2, . . . , m} true is the equation where A(ν) (l) = A(1) ( p ν−1l),
on the strength of Theorem 5.5 the matrix A(1) ( p ν−1l) is unitary for all l ∈ Z Nν−1 .
Theorem 5.5 is proved.
Proposition 5.19 Let f ∈ C N u u 0 , u 1 , . . . , u p−1 ∈ C Nν−1 , where 1 ≤ ν ≤ n. Assum-
Nν−1 −1
ing that the system {T pν−1 k f }k=0 is orthonormal in C N ,and the system
Nν −1 Nν −1
{T pk u 0 }k=0 ∪ · · · ∪ {T pk u p−1 }k=0 is orthonormal basis of space C Nν−1 . Let us
assume
g = f ∗ U ν−1 u 0 , h 1 = f ∗ U ν−1 u 1 , . . . , h p−1 = f ∗ U ν−1 u p−1 . (5.111)
Then the system

Nν −1 Nν −1 Nν −1
{T pν k g}k=0 ∪ {T pν k h 1 }k=0 ∪ · · · ∪ {T pν k h p−1 }k=0 (5.112)
is orthonormal in C N .
Proof For the arbitrary k ∈ {0, 1, . . . , Nν−1 − 1}, on the strength (5.90), (5.91), and
(5.111), we have
g ( p ν k) = ( f ∗ U ν−1 u 0 ∗ %
g, T pν k g = g ∗ % f ∗ U ν−1%
u 0 )( p ν k)
=(f ∗ %
f ) ∗ (U ν−1 (u 0 ∗ %
u 0 ))( p ν k)
or as per the convolution definition,
N −1

g, T pν k g = (f ∗ %
f )( p ν k p j) U ν−1 (u 0 ∗ %
u 0 ) ( j).
j=0
As
#
u 0 )( j/ p ν−1 ),
(u 0 ∗ % j if j is divisible by p ν−1 ,
U ν−1 (u 0 ∗ %
u 0 )( j) =
0, j if j is not divsible by p ν−1 ,
then,
Nν−1 −1

g, T pν k g = (f ∗ %
f )( p ν k p p ν−1 m)(u 0 ∗ %
u 0 )(m).
m=0
N −1
Taking into consideration the orthonormality of the system {T pν−1 k f }k=0
ν−1
, for the
arbitrary m ∈ Z Nν−1 we have
#
1, if m = pk,
f ∗ %
f ( p ν−1 ( pk p m)) = f, T pν−1 ( pk p m) f =
0, if m = pk.
Hence
#
1, if k = 0,
g, T pν k g = (u 0 ∗ %
u 0 )( pk) = u 0 , T pk u 0 =
0, if k = 1, 2, . . . , Nν − 1.
For each μ ∈ {1, . . . , p − 1}

#
1, if k = 0,
h μ , T pν k h μ = g, T pν k h μ =
0, if k = 1, 2, . . . , Nν − 1.
Consequently, the system (5.108) is orthonormal in C N . Proposition 5.19 is proved.

Proof of Theorem 5.6. The equation C N = W−1 ⊕ V−1 and the properties (b), (c)
for ν = 1 follow from the fact that, as per Theorem 5.4 the system
N1 −1 (1) N1 −1 N1 −1 ( p−1)
{φ−1,k }k=0 ∪ {ψ−1,k }k=0 ∪ · · · ∪ {ψ−1,k }k=0
is orthonormal basis in C N . As per the condition,

(μ)
φ−2,k = T p 2 k φ (2) , ψ−2,k = T p 2 k ψμ( 2) , k = 0, 1, . . . , N 2 − 1,
where
φ ( 2) = φ (1) ∗ U (u (02) ), ψμ( 2) = ψ (1) ∗ U (u (μ2) ), μ = 1, . . . , p − 1.
Using Proposition 5.19, we notice that the system
N2 −1 (1) N2 −1 N2 −1 ( p−1)
{φ−2,k }k=0 ∪ {ψ−2,k }k=0 ∪ · · · ∪ {ψ−2,k }k=0
is orthonormal basis of space V1 . From here we get the property (a) for ν = 1.
Now let ν ∈ {2, 3, . . . , m}. Using Proposition 5.18 for f = ψ (ν−1) , u 0 = u (ν)
0 ,
. . . , u p−1 = u (ν)
p−1 , we get that the system
Nν −1
Nν −1 (1) Nν −1 ( p−1)
{φ−ν,k }k=0 ∪ {ψ−ν,k }k=0 ∪ · · · ∪ ψ−ν,k (5.113)
k=0
is orthonormal in C N . Hence
* + * +
(μ) (μ ) (μ)
φ−ν,k , ψ−ν,l = ψ−ν,k , ψ−ν,l = 0
for k, l ∈ Z Nν , μ, μ ∈ {1, . . . , p − 1}, μ = μ . (5.114)
Further, for arbitrary k ∈ Z Nν , j ∈ Z N we have
φ−ν,k ( j) = φ (ν) ( j p p ν k) = (φ (ν−1) ∗ U ν−1 u (ν) ν

0 )( j p p k)
N −1

= φ (ν−1) ( j p p ν k p m) U ν−1 u (ν)
0 (m)
m=0
Nν−1 −1

= φ (ν−1) ( j p p ν k p p ν−1 m) u (ν)
0 (m)
m=0
Nν−1 −1

= u (ν)
0 (m)T pν−1 ( pk⊕ p m) φ
(ν−1)
( j).
m=0
It means, for any k ∈ Z Nν
Nν−1 −1

φ−ν,k = u (ν)
0 (m)φ−ν+1, pk⊕ p m (5.115)
m=0
Similarly, for all k ∈ Z Nν , μ ∈ {1, . . . , p − 1} we have
Nν−1 −1
(μ)

ψ−ν,k = u (ν)
μ (m)φ−ν+1, pk⊕ p m . (5.116)
m=0
From Formulas (5.114)–(5.116) follow the relations
V−ν ⊥ W−ν , V−ν ⊂ V−ν+1 , W−ν ⊂ V−ν+1
and as dim V−ν + dim W−ν = dim V−ν+1 , V−ν+1 is the direct sum of subspaces V−ν
and W−ν :
V−ν+1 = V−ν ⊕ W−ν . (5.117)
From here and from the orthogonality of 5.110 follow the properties (a)–(c) for each
ν. For the decomposition of the system 5.98, let us show that the system
p−1 N1 −1 p−1 N2 −1 p−1 Nm −1

(μ) (μ)
(μ) Nm −1
ψ−1,k ∪ ψ−2,k ∪ ··· ∪ ψ−m,k ∪ {φ−m,k }k=0
k=0 k=0 k=0
μ=1 μ=1 μ=1
(5.118)
is orthonormal in C N . From the orthonormality of the system 5.110 follow the equal-
Nν −1 (μ) Nν −1
ity 5.111 and orthonormality of each of the system {φ−m, k }k=0 {ψ−ν, k }k=0 . for
(μ) (μ) (μ)
ν < ν , let us examine the elements ψ−ν, k ψ−ν , l . Then ψ−ν , k ∈ W−ν and as per
(μ)
proof, ψ−ν, k ∈ W−ν , where W−ν ⊂ V−ν+1 ⊂ · · · ⊂ V−ν . Taking into consideration
(μ) Nν −1 (μ) Nν −1
(5.117),we get from here that the sets {ψ−ν, k }k=0 and {ψ−ν , l }l=0 are orthonor-
Nm −1 (μ) Nν −1
mal. Similarly, the set {φ−m, k }k=0 is orthonormal to any from the sets {ψ−ν, k }k=0 ,
1 ≤ ν ≤ m. It has to be noted that the number of elements of the sets (5.118) coincides
with the dimension of the space C N . Theorem 5.6 is proved.
5.5 Application to the Coding of Fractal Functions
It is well known that a quick discrete wavelet transform breaks down the signals into
low-frequency(approximating) and high-frequency(detailing) components with their
subsequent incomplete sample; the inverse transformation regenerates the signal(see
for example, [13, § 7.3]).The results of the calculation experiments given below show
that the discrete wavelet transformation associated with wavelets from Sects. 5.2, 5.3,
and 5.4 may be used for coding of fractal functions. In this section, the p -adic
operations ⊕ p and p on R+ will be denoted by ⊕ and , respectively.The notations
N , Nν , and ε p are understood as before.
In Sect. 5.4. for any complex vector b = (b0 , b1 , . . . , b N −1 ), satisfying the condi-
tion
|bl |2 + |bl+N1 |2 + · · · + |bl+( p−1)N1 |2 = 1, 0 l N1 − 1, (5.119)
is constructed an orthonormal basis in space of N -periodic complex sequences C N .

Let us denote by G( p, n) the set of all vectors b, satisfying Condition 5.116. Using
any method of unitary expansion of matrix, for any arbitrary fixed vector b(0) =
(b0(0) , b1(0) , . . . , b(0) (s)
N −1 ) from G( p, n) , let us find out the numbers bk , 0 k N − 1,
1 s p − 1, such that the matrices
⎛ ⎞
bl(0) (0)
bl+N (0)
. . . bl+( p−1)N1
⎜ (1) (1)
1
(1) ⎟
⎜ b bl+N . . . bl+( p−1)N1 ⎟
Ml = ⎜ l 1
⎟ , 0 l N1 − 1,
⎝ ... ... ... ... ⎠
( p−1) ( p−1) ( p−1)
bl bl+N1 . . . bl+( p−1)N1
are unitary. Further, with the help of Vilenkin–Christenson transformation, let us

calculate the coefficients
N −1
1 (s) (N )
ck(s) = b w ( j), 0 k N − 1, 0 s p − 1, (5.120)
N1 j=0 j k
5.5 Application to the Coding of Fractal Functions 227
and propose bk(s) = ck(s) = 0 k N .

The orthogonal discrete wavelet transform O( p, n), is associated with vector b(0) ,
and is defined by the formulae
(0) (1)
(1) ( p−1)
( p−1)
a j−1,k = cl pk a j,l , d j−1,k = cl pk a j,l , . . . , d j−1,k = cl pk a j,l ,
l∈Z+ l∈Z+ l∈Z+
(5.121)
where the coefficients ck(0) of the low-frequency filter and the coefficients ck(1) , . . . ,
( p−1)
ck of high-frequency filter are determined by Formula 5.118 (compare with [7,
Sect. 7.3.2]). The dimension of vector a j with the components a j,l is proposed by the
multiple number of the inlet canals p. The transformation 5.519 translates vector a j
( p−1)
into approximating vectors a j−1 and detailing vectors d (1) j−1 , . . . , d j−1 , the dimen-
sions of which are p times less than the dimensions of the vector a j . The inversion
formula of this transformation is
(0) (1) (1) ( p−1) ( p−1)
a j,l = cl pk a j−1,k + cl pk d j−1,k + · · · + cl pk d j−1,k . (5.122)
l∈Z+
√ √
The discrete transformation O(2, 1), associated with the vector b(0) = (1/ 2, 1/ 2),
coincides with classic Haar transform. Here the formulae (5.118) is written in the
form
a j,2k + a j,2k+1 a j,2k − a j,2k+1
a j−1,k = √ , d j−1,k = √ .
2 2
For arbitrary p and n the discrete Haar transform corresponds to the values b0(0) =
√ √
b1(0) = · · · = b(0)
p−1 = 1/ p, bk(0) = 0 when k p; then ck(s) = (1/ p)
exp(2πiks/ p) when 0 k, s p − 1. In particular p = 3 we have
√ √ √
(0) (0) (0) (1) (2) 3 (1) (2) 3 (1) (2) 3 2
c0 = c1 = c2 = c0 = c0 = , c1 = c2 = ε3 , c2 = c1 = ε ,
3 3 3 3
(5.123)
where ε3 = exp(2πi/3).
The discrete transform O(2, 2), associated with vector b(0) = (1, a, 0, b), 0 < a
1, a 2 + b2 = 1. was suited by Lang [14–16]. For this transformation the nonzero
coefficients in (5.118) and (5.119) are determined by the formulae
1+a+b 1+a−b
c0(0) = c1(1) = √ , c1(0) = −c0(1) = √ ,
2 2 2 2
1−a−b 1−a+b
c2(0) = c3(1) = √ , c3(0) = −c2(1) = √ .
2 2 2 2
While processing the discrete signal, the selected discrete wavelet transform is
( p−1)
used iteratively, after the first step the vectors d (1)
j−1 , . . . , d j−1 remain in memory,
but the vectors a j−1 changes to approximating vector a j−2 and detailing vectors
( p−1)
d (1)
j−2 , . . . , d j−2 etc. The dimensions of the vectors decrease by p times in each
step. As a result, after j0 steps, we get the vectors
( p−1) ( p−1)
a j− j0 , d (1) (1)
j−1 , . . . , d j−1 ; . . . ; d j− j0 , . . . , d j− j0 .
The vector a j is regenerated by inverse transformation of these vectors. The freedom

in selecting the vector b(0) helps in adapting the transformation O( p, n) to the signal
being processed by root mean square, entropy or other criteria.
In a recent work, it is shown that for coding certain fractural signals the discrete
wavelet transform O( p, n) is preferred in comparison with discrete Haar transform
and zone encoding method. This result is given as an example of coding values
generalized by Weierstrass function.
∞
k
πi x
Va,b (x) = a k eb , 0 < a 1, b 1/a.
k=1
As the source of four sources of signals we selected the values of function Va,b (X )
in 24 nodes of uniform division of interval (0, 1) for the pair of indices a and b given
in table .
(N ) (N )
Let TN = (wl,k ) be the matrix composed of numbers wl,k = wk (l/N ) , 0
l, k N − 1 . . As is usual, we designate the matrix complexly joined with the
matrix TN∗ TN . The direct and inverse discrete, multiplicative transforms in the
space C N , are written in the form
x = N −1 TN∗ x, x = TN (
( x, x ∈ CN .
Method of zone coding Z ( p) is used in [4, § 11.3], consists in using direct dis-
crete multiplicative transform to vector x with subsequent zeroing out vector ( x
of arbitrary p n − p n−1 components: ( x ( p n−1 ), (
x ( p n−1 + 1), . . ., (
x ( p n − 1). Then
the vector %x is obtained after application of inverse direct multiplicative √ transform.
The error in regeneration is evaluated by the values δ = ||x − % x ||2 / N . In calcula-
tion experiments, the orthogonal discrete spike transformation O(3, 1)O(3, 2) were
compared in Rodionov [17] not only with the method of zone coding Z (3), but also
with discrete Haar transform H (3), determined by the coefficients 5.120. The results
of the calculations are given in Table 5.1.
Let us now assume that the spaces Vn and Wn are determined as in § 5.3 when
p = 2N = 2n :
Vn = span{1, w1 (x), . . . , w N −1 (x)},
Wn = span{w N (x), w N +1 (x), . . . , w2N −1 (x)},

N −1 N −1
and basis systems {φn,k }k=0 , {ψn,k }k=0 are given by the parameter α under the
condition (5.78). We notice that for the decomposition coefficients
Table 5.1 Root mean square compression errors of functions Va,b with the help of sound coding
method, discrete Haar transform and orthogonal spike transformations
ł Z (3) H (3) O(3, 1) O(3, 2)
V0.6,9 0.3774 0.2835 0.2068 0.1310
V0.8,6 0.6660 0.5393 0.4826 0.4629
V0.8,9 0.6572 0.4599 0.3448 0.2934
V0.9,4 1.3786 0.9706 0.8763 0.8269

2N −1
w N −1 (x) = γn+1,k φn+1,k (x).
k=0
The recurrent formulae
γ2,0 = γ2,1 = 1, γ2,2 = γ2,3 = −1,
γn+1,k = γn,k , γn+1,N +k = −γn,k , n ≥ 2, 0 ≤ k ≤ N − 1.
According to (5.82) and (5.83), for any function f n ∈ Vn gn ∈ Wn we have the equal-
ities
N −1
N −1

f n (x) = Cn,k φn,k (x), gn (x) = Dn,k ψn,k (x), (5.124)
k=0 k=0
and sequence of coefficients
Cn = {Cn,k }, Dn = {Dn,k } (5.125)
uniquely defined by f n and gn , respectively. When p = 2 from (5.85), (5.86) we get

(n)
(n)
Cn,k = Al,k Cn+1,l , Dn,k = Bl,k Cn+1,l . (5.126)
l l
(n) (n)
Cn+1,l = Pl,k Cn,k + Q l,k Dn,k . (5.127)
k
In these formulae
⎧
⎨ 1/2 + (1 − α)/(2α N ), k = [l/2],
(n)
Al,k =
⎩
(1 − α)(−1)γ (k)−γ ([l/2]) /(2α N ), k = [l/2],
⎧
⎨ (−1)l /2, k = [l/2],
(n)
Bl,k =
⎩
0, k = [l/2],
and ⎧
⎨ 1 + γn+1,k (1 − α)(−1)γ (l)+1 /N , k = [l/2],
(n)
Pl,k =
⎩
γn+1,k (1 − α)(−1)γ (l)+1 /N , k = [l/2].
⎧
⎨ (−1)l , k = [l/2],
(n)
Q l,k =
⎩
0, k = [l/2].
For any function f n ∈ Vn and any n 0 ∈ {1, 2, . . . , n − 1} with the help of For-
muals (5.121), (5.123) after substitution of n by n − 1, n − 2, . . . , n − n 0 we get the
decomposition

n0
f n = f n−n 0 + gn− j
j=1
and, correspondingly the vector Cn gets transformed into a set of vectors Cn−n 0 ,
Dn−1 , . . . , Dn−n 0 . Thus is direct discrete diadic periodic spur transformation with
parameter α (let us call it Wα ). The corresponding inverse transformation Wα−1 is
based on Formulas (5.122), (5.124) and helps in regenerating the function f n and
vectors Cn as per the set of vectors Cn−n 0 , Dn−1 , . . . , Dn−n 0 .
The method of coding consisting of the following steps will be called the
Method A.
1. To use direct discrete wavelet transformation as input mass.
2. Select from the wavelet coefficients a certain number (for example, 10 or 1%)
having high modulus, the remaining coefficients equal to zero.
3. Use inverse wavelet transform for the obtained coefficients file.
4. Calculate from the regenerated file %

Cn , the average of root mean square deviation
,
-
-1 m
δ(Cn , Cn ) := .
% %n,k ),
(Cn,k − C
m k=1
where m is the dimension of vectors Cn %

Cn .
Method B is different from the Method A by the fact that here the step 2 is replaced
by uniform quantization of wavelet coefficients
Let us remember that uniform quantization of arbitrary vector x = (x1 , . . . , xm )
takes place by the formula
#
0, |x j | < Δ,
xj =
Δ([x j /Δ] + 0.5sign(x j )), |x j | ≥ Δ,
where Δ -step in quantization, in Method C given below, we assume

Δ = Δt = max x j − min x j /t, t ∈ N.
1≤ j≤m 1≤ j≤m
If the vector x represents the alphabet of the spatial source without storage/memory,
then the entropy of this source identified by the formula

m
H (x) = − p j log2 ( p j ),
j=1
where p j − probability with which the source yields the value x j . In practice, the
probability is substituted by relative frequency.
For arbitrary vectors
x = (x1 , . . . , xm ), y = (y1 , . . . , yk ), . . . , z = (z 1 , . . . , z s )
the vector
(x1 , . . . , xm , y1 , . . . , yk , . . . , z 1 , . . . , z s ).
is denoted by (x, y, . . . , z)
Method C is the coding method consisting of the following five steps.
1. To select parameters n 0 ∈ N, α ∈ {0.01, 0.02, . . . , 0.99, 1}, t ∈ N p ∈
{1, 2, . . . , 100}.
2. To transform the given vector Cn with the help of mapping Wα into a set of vectors
Cn−n 0 , Dn−1 , . . . , Dn−n 0 .
3. To quantize uniformly the vector
(Cn−n 0 , Dn−1 , . . . , Dn−n 0 )
with the step Δt .

4. To apply the transform Wα−1 to the vector obtained in step 3 and as a result getting
of vector %
Cn .
5. To verify the condition
p
δ(Cn , %
Cn ) < · max |Cn,k |. (5.128)
100 0≤k≤N −1
If Condition (5.162) is not satisfied, then change t by t + 1 and repeat steps 3–5.
6. To calculate the value of parameter α, in which the entropy of the vector obtained
after quantization in step 3 is the minimum.
Table 5.2 Root mean square error when coding Weierstrass function by Method A
p (a, b) α Wα Haar Db4
85 (0.7,5) 0.555 0.6142 0.6223 0.7495
85 (0.9,3) 0.765 0.8983 0.9008 6.9543
85 (0.9,7) 0.95 1.5876 1.5882 3.0923
95 (0.7,5) 0.68 0.6448 0.6733 3.6479
95 (0.9,3) 0.36 1.0690 1.0927 3.9443
95 (0.9,7) 0.525 1.9078 1.9184 6.7063
Table 5.3 Root mean square error when coding Weierstrass function by Method B
t (a, b) α Wα Haar Db4
10 (0.7,5) 0.09 0.2389 0.2662 0.2773
10 (0.9,3) 0.06 0.4593 0.4754 0.4714
10 (0.9,7) 0.17 0.5243 0.5419 0.5376
100 (0.7,5) 0.77 0.0754 0.0799 0.0819
100 (0.9,3) 0.39 0.1308 0.1384 0.1437
100 (0.9,7) 0.73 0.1476 0.1537 0.1709
A similar method is employed even when using discrete Haar and Daubechies trans-
formations (Haar transformation is obtained when α = 1). We notice that in examples
given below the value of p shows also the percentage of zero spike coefficients when
the methods A and B are used. Weierstrass function with parameters a,b is determined
by the formula
∞

Wa,b (x) = a n cos(bn π x), 0 < a < 1, b ≥ 1/a.
n=1
The results obtained by the methods A, B, and C in steps of digitization 0.005 on

section [0,1.28] and pairs of parameters (a, b) = (0.7, 5), (0.9, 3), (0.9, 7). are given
in Tables 5.2, 5.3, and 5.4. It is clear that the periodic dynamic spur with optimum
values of parameter α give result in many cases better than the classic spurs.
Let us now use Method A for the fractal functions of Van-der-waerden V (x).
Riemann R(x),Takagi T (x), and Hankel G (x), determined as follows:
1. Let the function f 0 (x) have a period 1 and on section [0,1] defined by the
formula #
x, x ∈ [0, 1/2),
f 0 (x) =
1 − x, x ∈ [1/2, 1]
(the value of f 0 (x) is equal to distance between x and the nearest integer point). Then
Table 5.4 Entropy when coding Weieratrass function by Method C

p (a, b) α Wα Haar Db4
95 (0.7,5) 0.48 4.0248 4.0678 4.2059
95 (0.9,3) 0.45 3.9607 4.0120 4.0335
95 (0.9,7) 0.67 4.3553 4.5534 4.3612
90 (0.7,5) 0.52 2.7626 2.8200 3.5927
90 (0.9,3) 0.41 2.6586 2.6803 2.8044
90 (0.9,7) 0.51 3.3664 3.4892 3.2403
Table 5.5 Root mean square error in 95% zeroing

Function α Wα Haar Db4
Van-der-Waerden 0.7850 0.0605 0.2486 0.0594
Riemann 0.4400 0.1637 0.4509 0.1607
Takagi 0.8950 0.0992 0.4039 0.0991
Hankel 0.3950 0.1722 0.3006 0.1600
∞
f 0 (4n x)
V (x) = (5.129)
n=1
4n
2. Riemann functoin:
∞
sin(n 2 π x)
R(x) = . (5.130)
n=1
n2
3. Takagi function:
∞
t0 (2n−1 x)
T (x) = , (5.131)
n=1
2n
where t0 (x) = 2|x − [x + (1/2)]|.

4. Hankel function:
∞
/ 0
sin(π nx) 1
G (x) = sin . (5.132)
n=1
n2 sin(π nx)
Table 5.5 contains the results obtained by Method A with initial vectors Cn of
length 256 obtained from the calculation of the values of the functions (5.126)–
(5.128) on the section [0,1.28] with step 0.005 and function (5.129) on section [0,01,
1.29]. It is seen that periodic dyadic spurs with optimum values of parameters α
surpass Haar spur and give results approximately the same as Daubechies Db4.
References
1. Vilenkin, N. Y. (1947). On a class of complete orthonormal systems. Izv. Akad. Nauk Sssr, Ser.
Mat. (No. 11, pp. 363–400). English Translation, American Mathematical Society Translations,
28 (Series 2), 1–35 (1963).
65(3), 372–414.
3. Agaev, G. H., Vilemkin, N. Y., Dzhafarli, G. M., & Rubinstein, A. I. (1981). Multiplicative
systems of functions and analysis on 0 dimensional groups. Baku: ELM [In Russian].
Transl. Of 1st ed.). Moscow: Urss; Dordrecht: Kluwer (1991).
5. Novikov, I. Y., Protasov, V. Y., & Skopina, M. A. (2011). Wavelet theory (Moscow, 2006).
Providence: AMS.
6. Farkov, Y. A., & Rodionov, E. A. (2011). Algorithms for wavelet construction on Vilenkin
groups. P-Adic Numbers, Ultrametric Analysis and Applications, 3(3), 181–195.
7. Lukomskii, S. F. (2014). Step refinable functions and orthogonal MRA on P-Adic Vilenkin
groups. Journal of Fourier Analysis and Applications, 20(1), 42–65.
8. Lukomskii, S. F. (2015). Riesz multiresolution analysis on zero-dimensional groups. Izvestiya:
Mathematics, 79(1), 145–176.
9. Berdnikov, G. S., & Lukomskii, S. F. N-valid trees in wavelet theory on Vilenkin groups. http://
arxiv.Org/abs/1412.309v1.
10. Farkov, Y. A. (2011). Periodic wavelets on the p-Adic Vilenkin group. P-Adic Numbers, Ultra-
metric Analysis, and Applications, 3(4), 281–287.
11. Chui, C. K., & Mhaskar, H. N. (1993). On trigonometric wavelets. Construction Approximately,
9, 167–190.
12. Frazier, M. W. (1999). An Introduction to wavelets through linear algebra. New York: Springer.
13. Mallat, S. (1999). A wavelet tour of signal processing. New York, London: Academic Press.
14. Lang, W. C. (1996). Orthogonal wavelets on the cantor dyadic group. SIAM Journal on Math-
ematical Analysis, 27(1), 305–312.
Journal of Mathematics and Mathematical, 21, 307–317.
16. Lang, W. C. (1998). Wavelet analysis on the Cantor dyadic group. Houston Journal of Mathe-
matics, 24, 533–544.
17. Rodionov, E. A., & Farkov, Y. A. (2009). Estimates of the smoothness of Dyadic orthogonal
Chapter 6
Haar–Vilenkin Wavelet
6.1 Introduction
Haar System
The Haar system H = (Hn , n ∈ N) is defined as follows:
H0 = 1. For n, r ∈ N and 0 ≤ r < 2n the function Hn is defined on [0, 1) by
⎧ n
⎨ 2 2 x ∈ I (2r, n + 1)
n
H2n +r (x) = −2 2 x ∈ I (2r + 1, n + 1)
⎩
0 other wise
where
I (2r, n + 1) = [2r 2−(n+1) , (2r + 1)2−(n+1) )

2r 2r + 1
= n+1 , n+1 .
2 2
It can be extended to R by the periodicity of period 1. Each Haar function is contin-

uous from the right and the Haar system H is orthonormal on [0, 1).
Haar–Vilenkin System
The following system which is a generalization of Haar system is connected with the
name of Vilenkin. Very often it is termed as a generalized Haar system or a Haar-type
Vilenkin system (Schipp et al. [1]). Let m = (m k , k ∈ N) be a sequence of natural
numbers such that m k ≥ 2, N denotes the set of nonnegative integers. Let M0 = 1
and Mk = m k−1 Mk−1 , k ∈ P.
Let P denote the set of positive integers and let k ∈ P can be written as
k = Mn + r (m n − 1) + s − 1. (6.1)

https://doi.org/10.1007/978-981-13-6370-2_6
236 6 Haar–Vilenkin Wavelet
where n ∈ N, r = 0, 1, . . . , Mn − 1 and s = 1, 2, . . . , m n − 1. This expression is

unique for each k ∈ P. Let us write an arbitrary element t ∈ [0, 1) in the form
∞
tk
t= , (0 ≤ tk < m k ). (6.2)
k=0
Mk+1
It may be noted that there exist two such expressions (6.2), for so-called m-adic
rational numbers. In such cases we use the expression which contains only a finite
number of terms different from zero.
Define the function system (h n , n ∈ N) by h 0 = 1 and
√
Mn exp 2πist n r
≤ t < rM+1n
h k (t) = mn Mn . (6.3)
0 other wise
This system can be extended to R (the set of real numbers) by periodicity of period 1:
h k (t + 1) = h k (t), t ∈ [0, 1). It can be checked that {h k (t)} is a complete orthonormal
system in L 2 (R). It is clear that
2π istn
h k (t) = χ[ Mr , r +1 ] (t) Mn exp .
n Mn mn
Certain properties of this system have been recently studied.

In the recent years, various extensions and concepts related to Haar wavelet has
been studied. In this chapter, we study basic properties of Haar–Vilenkin wavelets
and Haar–Vilenkin scaling function [2, 3]. We prove that the system {ψa,b }, a, b ∈ Z

in
L (R), while the convergence properties of expansion of f
2
is an orthonormal basis
namely the series a∈Z b∈Z f, Dm an Tb h k (t)Dm an Tb h k (t) for arbitrary coefficients
(k fixed) are studied.
6.2 Haar–Vilenkin Wavelets
6.2.1 Haar–Vilenkin Mother Wavelet
The function h k (t) as defined in (6.3) can also be written as

⎧√
⎪
⎪ Mn r
≤ t < Mr n + M1n+1
⎪√
⎪
Mn
⎪ 2πis
⎨ √ Mn exp m n
r
Mn
+ M1n+1 ≤ t < Mr n + 2
Mn+1
h k (t) = Mn exp 4πis r
+ M2n+1 ≤ t < Mr n + 3
. (6.4)
⎪
⎪
mn Mn Mn+1
⎪
⎪ ···
⎪ √
⎩ M exp 2πis(m n −1) m n −1 r +1
n mn
r
Mn
+ Mn+1
≤t < Mn
6.2 Haar–Vilenkin Wavelets 237
It can be easily seen that h k (t) ∈ L 2 [0, 1) and th k (t) ∈ L 1 [0, 1) for k ∈ P and
∞ r
+M1 r
+M2
Mn n+1 Mn n+1 2πis
h k (t) dt = Mn dt + Mn exp dt + · · ·
−∞ r
Mn
r
Mn +M1 mn
n+1
r +1

Mn 2πis(m n − 1)
+··· + Mn exp dt
r
Mn + m n −1
Mn+1
mn
√
Mn 2πis 4πis 2πis(m n − 1)
= 1 + exp + exp + · · · + exp
Mn+1 mn mn mn
√
Mn 1 − exp2πis
=
Mn+1 1 − exp 2πis
mn
= 0.
Thus the function h k (t) is a mother wavelet for k ∈ P and for t ∈ [0, 1). The function
h k (t) is called a Haar–Vilenkin Wavelet.
Define
ψa,b (t) = m a/2
n h k (m n t − b)
a
(6.5)
The collection {ψa,b (t)}a,b∈Z is referred to as the Haar–Vilenkin

system.
ψa,b (t) is supported on the interval Ia,b where Ia,b = m a Mn + mba , mra+1
r
+ b
a ,
n n n Mn mn
a, b ∈ Z. a
The system ψa,b (t) can also be written as {m n2 h k (m an t − b)} = Dm an Tb h k (t).
Haar–Vilenkin Scaling Function

For k ∈ P and t ∈ [0, 1) as defined in (6.1) and (6.2) the Haar–Vilenkin scaling
function is defined as

pk (t) = Mn χ[ Mr , rM+1 )
√
n n
Mn , Mr n ≤ t < rM+1n
= (6.6)
0 other wise
Define
φa,b (t) = m a/2
n pk (m n t − b)
a
(6.7)
The collection {φa,b (t)}a,b∈Z is referred to as the system of Haar–Vilenkin scaling

functions. For a given a ∈ Z, the collection {φa,b (t)}b∈Z is referred to as the system
of scale a Haar–Vilenkin scaling functions.
φa,b (t) is supported on the interval Ia,b where a, b ∈ Z, where Ia,b =
r
m a Mn
+ mba , mra+1
Mn
+ mba
n n n n
For each a, b ∈ Z
1
φa,b (t) dt = φa,b (t) dt = m a/2
n Mn = m −a/2
n Mn−1/2 .
R Ia,b m an Mn
and

1
|φa,b (t)|2 dt = |φa,b (t)|2 dt = m an Mn = 1.
R Ia,b m an Mn
Remark 6.1 1. Haar system is a special case of Haar–Vilenkin system for m n = 2

for all n ∈ N
2. Given any a ∈ Z, the collection of scale a Haar–Vilenkin scaling functions is an
orthonormal system on R.
Remark 6.2 We have
φa,b (t) = m −1/2

n φa+1, r (mn −1) +m n b (t) + m −1/2
n φa+1, r (mn −1)+1 +m n b (t) + · · ·
Mn Mn
··· + m −1/2
n φa+1, (r +1)(mn −1) +m n b (t). (6.8)
Mn
and
2πis −1/2
ψa,b (t) = m −1/2
n φa+1, r (mn −1) +m n b (t) + exp m φa+1, r (mn −1)+1 +m n b (t) + · · ·
Mn mn n Mn
2πis(m n − 1) −1/2
· · · + exp m n φa+1, (r +1)(mn −1) +m n b (t). (6.9)
mn Mn
The Eqs. (6.8) and (6.9) shows the relationship between Haar–Vilenkin scaling
function and Haar–Vilenkin wavelet.
Remark 6.3 For k = 1, h k (t) is the well-known mother Haar wavelet.
Lemma 6.1 For a ∈ Z, let ga (x) be a scale a function which is constant on Ia,b , b ∈
Z. Then ga (x) can be written as
ga (x) = ra−1 (x) + ga−1 (x)
where ra−1 (x) has the form

ra−1 (x) = αa−1 ψa−1,b (x) (6.10)
b
for some coefficients {αa−1 (b)}b∈Z and ga−1 (x) is a function which is constant on
Ia−1,b .
Proof ga (x) is a function which is constant on Ia,b . Suppose ga (x) has the value
Ca (b) on the interval Ia,b . For each interval Ia−1,b , define the function ga−1 (x) which
is constant on Ia−1,b by

ga−1 (x) = m a−1
n Mn ga (t) dt
Ia−1,b

m a−1
n Mn r (m n − 1) r (m n − 1) + 1
= C a + m n b + C a + m n b + ···
m an Mn Mn Mn

(r + 1)(m n − 1)
· · · + Ca + mn b
Mn

1 r (m n − 1) r (m n − 1) + 1
= Ca + m n b + Ca + mn b + · · ·
mn Mn Mn

(r + 1)(m n − 1)
· · · + Ca + mn b
Mn
as
Ia−1,b = Ia, r (mn −1) +m n b ∪ Ia, r (mn −1)+1 +m n b ∪ · · · ∪ Ia, (r +1)(mn −1) +m n b .
Mn Mn Mn
In other words on Ia−1,b , ga−1 (x) takes the average value of ga (x).
Let ra−1 (x) = ga (x) − ga−1 (x). ga−1 (x) is a function which is constant on the
interval Ia−1,b , b ∈ Z and |Ia−1,b | = m a−11 M .
n n
Thus

ra−1 (x) d x = ga (x) d x − ga−1 (x) d x
Ia−1,b Ia−1,b Ia−1,b

= ga (x) d x + ga (x) d x + · · ·
I I
a, r (mMn −1) +m n b a, r (m nM−1)+1 +m n b
n
n
··· + ga (x) d x − ga−1 (x) d x

I a−1,b
a, (r +1)(m
Mn
n −1) +m n b

1 r (m n − 1) r (m n − 1) + 1
= C a + m n b + C a + m n b + ···
m an Mn Mn Mn

(r + 1)(m n − 1) 1 r (m n − 1)
+ Ca + mn b − a Ca + mn b +
Mn m n Mn Mn

r (m n − 1) + 1 (r + 1)(m n − 1)
Ca + m n b + · · · + Ca + mn b
Mn Mn
= 0.
Thus on Ia−1,b , ra−1 (x) must be a multiple of the Haar–Vilenkin function ψa−1,b (x)
and must have the form (6.10).
Theorem 6.1 Given any a ∈ Z, the collection {φa, Mb }b∈Z is an orthonormal system
n
on R.
Proof Since a ∈ Z is fixed and suppose b, b

∈ Z are given. Then

φ b = b
Ia, Mb ∩ Ia, b
= Ia, Mb b = b
n Mn
n
If b = b
, then the product φa, Mb (t)φa, b
(t) = 0 for all t, since the functions are
n Mn
supported on disjoint intervals.
Hence if b = b

φa, Mb , φa, b
= φa, Mb (t)φa, b
(t) dt = 0.
n Mn n Mn
R
If b = b
, then
φa, b , φa, b
= |φa,b (t)|2 dt = 1.
Mn Mn
R
Orthonormality of Haar–Vilenkin Wavelet

We prove in the following theorem that the Haar–Vilenkin wavelet system is an
orthonormal basis.
a
Theorem 6.2 The system {m n2 h k (m an t − b)} = {ψa,b }, a, b ∈ Z is an orthonormal
system in L 2 (R).
Proof √
Mn exp 2πist n r
≤ t < rM+1n
h k (t) = mn Mn
0 other wise
for k ∈ P and t ∈ [0, 1).

Thus h k (t) is supported on the interval [ Mr n , rM+1n ) and thus h k (m an t − b) is sup-
ported on the interval Ia,b , where

r b r +1 b
Ia,b = + a, a + a
m an Mn m n m n Mn mn
√
Mn exp 2πist n
+ mba ≤ t <
r r
+ b
h k (m an t − b) = mn m an Mnn m an Mn m an
0 other wise
This function can also be written as

⎧√
⎪ Mn A ≤ t < A + m a M1 n+1
⎪
⎪ √
⎪
n
⎪ 2πis
⎨ √ Mn exp m n A + m a M1 n+1 ≤ t < A + 2
m an Mn+1
n
h k (t) = Mn exp 4πis A + m a M2 n+1 ≤ t < A + 3
⎪
⎪
mn n m an Mn+1
⎪
⎪ · · ·
⎪
⎩√ n −1) m n −1
Mn exp 2πis(m
mn
A+ m an Mn+1
≤t < A+ 1
m an Mn
where A = r
m an Mn
+ b
m an
.
First we will show the orthonormality with a given scale. Let a ∈ Z be fixed and
suppose b, b
∈ Z are given Then

φ i f b = b
Ia,b ∩ I a,b
=
Ia,b i f b = b
If b = b
, then the product ψa,b (t)ψa,b
(t) = 0, ∀t since the functions are supported
on the disjoint intervals.
Hence if b = b
, then

< ψa,b , ψa,b
>= ψa,b (t)ψa,b
(t) dt = 0.
R
If b = b
, then

< ψa,b , ψa,b
> = ψa,b (t)ψa,b (t) dt = ψa,b (t)2 dt
I Ia,b
a,b
= m an Mn dt = 1.
Ia,b
Next we will show the orthonormality between the scales. Suppose a, a

∈ Z with
a = a
, say a > a
and let b, b
∈ Z. Then we have the possibilities
1. Ia
,b
∩ Ia,b = φ. In this case ψa,b (t)ψa
,b
(t) = 0∀t and

< ψa,b , ψa
,b
>= ψa,b (t)ψa
,b
(t) dt = 0.
Ia,b
2. If a > a
then either the intervals Ia,b and Ia
,b
are disjoint or Ia,b is contained in
the one of the m n subintervals

1
1
2
A , A + a
, A + a
, A + a
,...
m n Mn+1 m n Mn+1 m n Mn+1

mn − 1 1
. . . A
+ a
, A
+ a
m n Mn+1 m n Mn
where A
= r
m an
Mn
+ b
m an
In each case, we will get

< ψa,b , ψa
,b
>= ψa,b (t)ψa
,b
(t) dt = 0.
Ia,b
Thus ψa,b , a, b ∈ Z is an orthonormal system in L 2 (R).

In order to show that {ψa,b }a,b∈Z is an orthonormal basis in L 2 (R), let us consider
the two families of subspaces of L 2 (R).
S p = span{ψa,b }a< p,b∈Z (6.11)
L p = {Set of all functions which are constant on intervals I p,b for b ∈ Z} (6.12)
Both of these families have the following properties:
· · · ⊂ S−2 ⊂ S−1 ⊂ S0 ⊂ S1 ⊂ S2 ⊂ · · · (6.13)
f (t) ∈ S p ⇔ f (2t) ∈ S p+1 (6.14)
f (t) ∈ S0 ⇔ f (t + k) ∈ S0 for k ∈ Z (6.15)
In order to prove that {ψa,b } is an orthonormal basis in L 2 (R)it remains to prove that
L p = Sp∀ p ∈ Z
Lemma 6.2 For all p ∈ Z, we have L p = S p .

Proof From (6.14) above it suffices to show that L 0 = S0 .
Since each ψa,b for a < 0 is constant on any interval [u + Mr n , u + rM+1n ) we see that

S0 ⊂ L 0 . Also each function in L 0 can be written as u∈Z au χ[u+ Mr ,u+ rM+1 ) , Hence
n n
by (6.15) it suffices to show that χ[ Mr , rM+1 ) ∈ S0 .
n n
To show this, let us consider the series

n ψa,0 =
m a/2 m an h k (m an t).
a<0 a<0
a/2
Since m an h k (m an t)2 = m n and a < 0, this series is absolutely convergent in
L 2 (R).
One can easily see from the definition of h k (t) that

a/2
a<0 m n ψa,0 (t) = 0 for t ≤ Mn
r
and

√ √
a/2 r +1
a<0 m n ψa,0 (t) = a<0 m n Mn = m n −1 for Mn < t < Mn .
a Mn r
m vn r m vn r +1
For Mn
≤t < Mn
where v = 1, 2, 3, . . . one has

n ψa,0 (t) =
m a/2 m an h k (m an t).
a<0 a<0
If r = 0, then
∞
1 2π istn
n ψa,0 (t) =
m a/2 v
M n exp + m −a
n Mn
a<0
mn mn a=v+1
1 2π istn 1
= M n exp + Mn v .
m vn mn m n (m n − 1)
If r = 0, then
1 2π istn
n ψa,0 (t) =
m a/2 v
Mn exp .
a<0
mn mn
This shows that S0 = L 0 , so S p = L p for all p ∈ Z. It can be easily verified that

∪∞
p=−∞ L p is dense in L (R).
2
Thus the system {ψa,b }a,b∈Z is an orthonormal basis in L 2 (R).
Thus a function f ∈ L 2 (R) has a decomposition

f = f, ψa,b ψa,b .
a∈Z b∈Z
6.3 Approximation by Haar–Vilenkin Wavelets
Definition 6.1 For each a ∈ Z define the approximation operator Pa on the functions
f (x) ∈ L 2 (R) by
Pa f (x) = f, φa, Mb φa, Mb (x)
n n
b
Remark 6.4
1. For each a ∈ Z, define the approximation space Va by

Va = span φa, Mb
n b∈Z

Since φa, Mb : b ∈ Z is an orthonormal system on R. This implies that Pa f (x)
n
is a function in Va best approximating f (x) in L 2 -sense.

2.
φa, Mb (x) = m a/2
n Mn χ Ia, b (x)
n Mn
Thus ⎛ ⎞

f, φa, Mb φa, Mb (x) = m an Mn ⎝ f (t) dt ⎠ χ Ia, b (x)
n n Mn
Ia, b
Mn
In other words, on the interval Ia, Mb , Pa f (x) is the average value of f (x) on
n
Ia, Mb .
n
We can prove the following facts about the operator Pa :

Theorem 6.3 1. For each a ∈ Z, Pa is linear, given f (x), g(x) ∈ L 2 (R) and α, β ∈
C
Pa (α f + βg)(x) = α Pa ( f )(x) + β Pa (g)(x)
2. For each a ∈ Z, Pa is idempotent, that is given f (x) ∈ L 2 (R)
Pa (Pa f )(x) = Pa f (x)
3. Given integers a, a
with a ≤ a
and g(x) ∈ Va
Pa
g(x) = g(x)
4. Given a ∈ Z and f (x) ∈ L 2 (R)
Pa f 2 ≤ f 2
Proof 1.
⎡ ⎤

Pa (α f + βg)(x) = m an Mn ⎣ (α f (t) + βg(t)) dt ⎦ χ I (x)
a, Mb
Ia, b n
Mn

= αm an Mn f (t) dt.χ I (x) + βm an Mn g(t) dt.χ I (x)
a, Mb a, Mb
Ia, b n Ia, b n
Mn Mn
= α Pa f (x) + β Pa g(x).
2. If a ∈ Z and for f (x) ∈ L 2 (R)

Pa (Pa f )(x) = Pa f, φa, Mb φa, Mb
n n
b

= < f, φa, b
>, φa, Mb φa, Mb (x)
Mn n n
b b

= < f, φa, b
>< φa, b
, φa, Mb > φa, Mb (x)
Mn Mn n n
b b

= < f, φa, Mb > φa, Mb (x)
n n
b
= Pa f (x).
3. It follows from the fact that if g(x) ∈ Va , then Pa g(x) = g(x).

6.3 Approximation by Haar–Vilenkin Wavelets 245
4. Since { pa, Mb }b∈Z is an orthonormal system on R.

n
Hence
2

Pa f 22 = f, φa, Mb φa, Mb d x
R n n
b

= | f, φa, Mb |2
n
b
2

= m a/2 Mn f (t) dt
n
b Ia, b
M

n
By Cauchy–Schwarz’s Inequality
2 ⎛ ⎞⎛ ⎞

a/2
m Mn f (t) dt ≤ ⎝ m an Mn dt ⎠ ⎝ | f (t)|2 dt ⎠
n
Ia, b Ia, b Ia, b

Mn Mn Mn
= | f (t)|2 dt
Ia, b
Mn

Thus
Pa f 22 ≤ | f (t)|2 dt = | f (t)|2 dt = f 22 .
b Ia, b R
Mn
Theorem 6.4 Given f (x), C 0 on R
lim Pa f − f 2 = 0 (6.16)

a→∞
Proof Suppose that f (x) is supported in an interval of the form [−m nN , m nN ] for
some integer N . Then there exists an integer A and a function g(x) ∈ V A such that

f − g∞ = max x∈R | f (x) − g(x)| < .
m nN +3
If a ≥ A, then by Theorem 6.3, Pa g(x) = g(x) and by Minkowski inequality and

Theorem 6.3
Pa f − f 2 ≤ Pa f −a g2 + Pa g − g2 + f − g2

= Pa ( f − g)2 + g − f 2
≤ 2 f − g2 . (6.17)
where
m nN m nN
2 2 2 2
g − f 22 = |g(x) − f (x) d x ≤2
= <
−m nN −m nN m nN =3 m 3n m 2n

⇒ g − f 2 < .
mn
Combining this result with (6.17) proves (6.16).
6.4 Covergence Theorems
The number
Ca,b = f, ψa,b (6.18)
is called the a, bth wavelet coefficient and

f, ψa,b ψa,b (6.19)
a∈Z b∈Z
is called the wavelet series of f ∈ L 2 (R).

We investigate the convergence of this series for any f ∈ L p (R), 1 < p < ∞.
We prove the following theorem:
Theorem 6.5 If f ∈ L p (R) with 1 < p < ∞ and f is continuous function, then
lim Pp ( f ) = f (6.20)
p→∞
where
Pp ( f ) = f, ψa,b ψa,b
a< p b∈Z
i.e.
lim Pp ( f ) − f p → 0
p→∞
Furthermore for each p ∈ Z
lim Pa ( f ) + Q aμ ( f ) = Pa+1 ( f ) (6.21)

μ→∞
where
Q aμ (t) = < f, ψa,b > ψa,b
b≤μ
It may be observed that the convergence part may be obtained from the fact that Pp
are conditional expectations but we present here a direct proof.
6.4 Covergence Theorems 247
Proof For 1 < p < ∞, let us consider the families of subspaces of L p (R)
p
Sl = span{ψa,b }a<l,b∈Z
p
L l = {Set of all functions which are constant on intervalsIl,b for b ∈ Z}

where Ia,b = r
m an Mn
+ b
, r +1
m an m an Mn
+ b
m an
.
p p
The proof of Lemma 6.2 can be easily modified to prove that Sl = L l ∀l ∈ Z.
Since Pa ( f ) is an orthonormal projection onto Sa and Sa = L a , we can write a
different presentation of the operator Pa , namely we have

Pa ( f ) = m an Mn f (t) dt.χ Ia,b
b∈Z Ia,b
This equation is valid as the right-hand side of this equation defines an orthogonal
projection onto L a .
By Holder’s Inequality, we have
! !p "1/ p
! ! 1
Pa ( f ) p = m ap p ! !
f (t) dt !
n Mn !
b∈Z Ia,b m an Mn
"1/ p
mn
−ap/q
Mn
− p/q
= m ap p
n Mn | f (t)| dt p
a
b∈Z Ia,b m n Mn
"1/ p
# $ p−1− p/q
= m an Mn | f (t)| p dt
b∈Z Ia,b
∞ 1/ p
= | f (t)| dtp
−∞
If f ∈ C0 (R), then it is uniformly continuous. Then for given > 0 we can find N
such that a > N and for each b ∈ Z
sup{| f (x) − f (y)|x, y ∈ Ia,b } <
For a given a > N and each t ∈ R, we fix an integer b such that t ∈ Ia,b . Then

|Pa f (t) − f (t)| = m an Mn f (s) ds − f (t)
Ia,b

a

= m n M n ( f (s) − f (t)) ds
Ia,b
< .
This implies that

supt∈R |Pa f (t) − f (t)| → 0 as a → ∞.
This proves the first part of the theorem.
Since for a fixed a, {ψa,b }b∈Z have disjoint supports, we have

! ! ⎛ ⎞1/ p
! !
! !
! < f, ψa,b > ψa,b ! ⎝ | < f, ψa,b > | p ψa,b p ⎠
! ! =
!b≤μ ! b≤μ
p
⎛ ⎞1/ p
# $(1/2−1/ p)
=⎝ m an Mn ψa,b p ⎠
b≤μ
μ
This shows that limμ→∞ Q a ( f ) exists in the norm of the space. Clearly it is equal
to Pa+1 ( f ) − Pa ( f ).
6.5 Haar–Vilenkin Coefficients
Behavior of Haar–Vilenkin Coefficients Near Jump % Discontinuities

r +1
Suppose that f (x) is defined on interval Mn , Mn with a jump discontinuity at
r
& %
x0 ∈ Mr n , rM+1n and continuous at all other points in Mr n , rM+1n . We have to check
whether Haar–Vilenkin coefficients < f, ψa,b > such that x0 ∈ Ia,b behave differ-
ently than do the Haar–Vilenkin coefficients.
Let us assume that given function f (x) is C 2 on the intervals [0, x0 ] and [x0 , 1].
This means that both f
(x) and f
(x) exist, are continuous functions and hence are

bounded on these intervals. Fix integers a ≥ 0 and 0 ≤ b ≤ m an − 1 and let xa,b be
the mid point of the interval Ia,b , i.e.,
r +1/2
xa,b = m an Mn
+ b
m an
.
Case I: If x0 ∈ / Ia,b , then expanding f (x) about xa,b by Taylor’s formulae, it

follows that for all x ∈ Ia,b
f (x) = f (xa,b ) + f
(xa,b )(x − xa,b ) + f (ξa,b )(x − xa,b )2
2
where ξa,b is some point in Ia,b .
6.5 Haar–Vilenkin Coefficients 249
'
Since Ia,b ψa,b (x) d x = 0,

f, ψa,b = f (x)ψa,b (x) d x
Ia,b

= f (xa,b ) ψa,b (x) d x + f
(xa,b ) ψa,b (x)(x − xa,b ) d x
Ia,b Ia,b

1
+ ψa,b (x)(x − xa,b )2 f
(ξa,b ) d x
2 Ia,b
= αa,b (x) + βa,b (x)
where
αa,b (x) = f
(xa,b ) ψa,b (x)(x − xa,b ) d x
Ia,b
and
1
βa,b (x) = ψa,b (x)(x − xa,b )2 f
(ξa,b ) d x
2 Ia,b
Now

|αa,b (x)| ≤ | f (xa,b )| |ψa,b (x)||(x − xa,b )| d x

Ia,b

= | f (xa,b )|m n
a/2
Mn |(x − xa,b )| d x
Ia,b

1
= | f (xa,b )|m n a/2
Mn
2m 2a 2
n Mn
−3a/2 −3/2
mn Mn
= | f
(xa,b )| .
4
and

1
|βa,b (x)| = ψa,b (x)(x − xa,b ) f (ξa,b ) d x
2
2 Ia,b

1
≤ max x∈Ia,b | f
(x)| |ψa,b (x)|(x − xa,b )2 d x

2 Ia,b

1
= Mn m a/2
n max x∈Ia,b | f (x)| (x − xa,b )2 d x
2 Ia,b

1
1
= Mn m a/2
n max x∈Ia,b | f (x)|
6 4m 3an Mn
3
1 −5/2 −5a/2
= M mn max x∈Ia,b | f
(x)|.
24 n
−5a/2 −3a/2
If j is large, then m n will be very small compared with m n , so we conclude
that for the large values of j
1 −3a/2 −3/2
| f, ψa,b | ≈ m Mn | f (xa,b )|.

4 n
Case II: If x0 ∈ Ia,b , then it is contained in
% one of the m n subintervals of Ia,b .
b r +1/m n
Assume that x0 ∈ m a Mn + m a , m a Mn + m a .
r b
n n n n
Expanding f (x) in a Taylor series about x0 , we have

r b
f (x) = f (x0− )
+ f (ξ− )(x − x0 ), x ∈ + a , x0 , ξ− ∈ [x, x0 ]

m an Mn mn
and

r + 1/m n b
f (x) = f (x0+ )
+ f (ξ+ )(x − x0 ), x ∈ x0 , a
+ a , ξ+ ∈ [x0 , x].
m n Mn mn
Thus

f, ψa,b = f (x)ψa,b (x) d x
Ia,b
x0 A+ m a M1
f (x0− )ψa,b (x) d x f (x0+ )ψa,b (x) d x
n n+1
= +
A x0
A+ m a mMn
f (x0+ )ψa,b (x) d x + a,b
n n+1
+
A+ m a M1
n n+1

− + 1
= m a/2
n M n f (x 0 )(−A + x 0 ) + f (x 0 ) A+ − x0
m an Mn+1
A+ m a mMn
f (x0+ )
n n+1
+ ψa,b (x) d x + a,b
A+ m a M1
n n+1

1
= m a/2
n Mn f (x0− )(−A + x0 ) + f (x0+ ) A+ a − x0
m n Mn+1
+ 1
− Mn m a/2
n f (x 0 ) a + a,b
m n Mn+1
− +
= Mn m a/2
n (x 0 − A)[ f (x 0 ) − f (x 0 )] + a,b
where A = r
m an Mn
+ b
m an
and
6.5 Haar–Vilenkin Coefficients 251
x0 A+ a 1
m n Mn+1
a,b = f (ξ− )(x − x0 )ψa,b (x) d x + f (ξ+ )(x − x0 )ψa,b (x) d x.
A x0

|a,b | ≤ maxt∈Ia,b \{x0 } | f
(t)| |x − x0 ||ψa,b (x)| d x
Ia,b

= maxt∈Ia,b \{x0 } | f (t)| Mn m n a/2

|x − x0 | d x
Ia,b
−3/2 −3a/2
Mn m n
≤ maxt∈Ia,b \{x0 } | f
(t)| .
4
−3/2 −3a/2 −1/2 −a/2
If j is large, then Mn mn will be very small compared with Mn mn , so
for large values of j

r b
| < f, ψa,b > | ≈ m a/2
Mn x 0 − a − a | f (x0− ) − f (x0+ )|.
n
m n Mn mn

b
The quantity x0 − r
m an Mn
− can be small if x0 is close to m arMn +
m an
b
m an
and can
n
%
b r +1/m n
even be zero. We can expect that in middle of m a Mn + m a , m a Mn +
r b
so that
n n n man
b
x0 − m an Mn − m an ≈ 2m an Mn+1 .
r 1
Thus for the large values of j

1
| < f, ψa,b > | ≈ m a/2
n Mn | f (x0− ) − f (x0+ )|
2m an Mn+1
−a/2 1/2
m n Mn
= | f (x0− ) − f (x0+ )|.
2Mn+1
Comparing the two cases, we see that the decay of | < f, ψa,b > | for the large j is
considerably slower if x0 ∈ Ia,b than if x0 ∈
/ Ia,b .
The large coefficient in the Haar–Vilenkin expansion of the coefficient f(x) that
persist for all scales suggests the presence of jump discontinuity in the intervals Ia,b
corresponding to the large coefficient.
6.6 Exercises
6.1 Prove the statement made in Remark 6.1.

6.2 Prove that for a ∈ Z, the collection of scale a Haar scaling functions is an
orthonormal system on R.
6.3 The collections {φa, Mb , ψa, Mb , b ∈ Z} is an orthonormal system on R.
n n
p p
6.4 Prove that Sl = L l ∀l ∈ Z where the terms Sl and L l are same as in Lemma
6.2.
6.5 Calculate√the Haar–Vilenkin coefficients for the scale a = 4 for the function
f (x) = 4 on (0,1).
References
2. Manchanda, P., Meenakshi, & Siddiqi, A. H. (2008). Haar-Vilenkin wavelet. The Aligarh Bulletin
of Mathematics, 27(1), 59–73.
3. Manchanda, P., & Meenakshi. (2009). New classes of wavelets. In A. H., Siddiqi, A. K. Gupta, &
M. Brokate (Eds.) Conference Proceedings Modeling of Engineering and Technological Prob-
lems (vol. 1146, pp. 253–271). New York.
Chapter 7
Construction Biorthogonal Wavelets
and Frames
In this chapter, basic properties of biorthogonal wavelets on positive real lines and
Vilenkin groups, frames on Cantor group, Parseval frames on Vilenkin group, and
application of biorthogonal dyadic wavelets to image processing are presented and
these results are discussed in more detail [1–3].
7.1 Biorthogonal Wavelets on R+
Definition 7.1 A sequence {gk }∞ k=1 in a Hilbert space H is called biorthogonal to

another sequence { f k } in H if the following conditions are satisfied:
< gk , f j >= δk, j , (7.1)
where
0 if k = j,
δk, j =
1 if k = j.
It may be proved that for every basis { f k } of a Hilbert space, there exists a unique
basis {gk } such that { f k } and {gk } are biorthogonal and in H
∞

f = < f k , gk > f k , for all f ∈ H.
k=1
For proof see Christenson [4].

The main goal is to present an algorithm for computing biorthogonal compact
supported dyadic wavelets related to the Walsh functions on the positive half-line
R+ , see the example Farkov et al. [2].

https://doi.org/10.1007/978-981-13-6370-2_7
254 7 Construction Biorthogonal Wavelets and Frames
Let us recall that Walsh system {wn |n ∈ Z+ } on R+ is defined by

k
w0 (x) ≡ 1, wn (x) = (w1 (2 j x))v j , n ∈ N, x ∈ R+
j=0
where w1 is defined on [0,1) by the formula

1 x ∈ [0, 1/2),
w1 (x) =
−1 x ∈ [1/2, 1).
and is extended to R+ by periodicity: w1 (x + 1) = w1 (x) for all x ∈ R+ , while the

v j are the digits of the binary expansion of n:

k
n= vj2j, v j ∈ {0, 1}, vk = 1, k = k(n)/.
j=0
By a Walsh polynomial, we shall mean a finite linear combination of Walsh functions.

We shall denote the integer and the fractional parts of a number x by [x] and {x},
respectively. For x ∈ R+ and j ∈ N, we define x j , x− j ∈ {0, 1} as follows:

x = [x] + {x} = x j 2− j−1 + x j 2− j (7.2)
j<0 j>0
(for a dyadic rational x, we obtain an expansion with finitely many nonzero terms).
The dyadic addition on R+ is defined by the formula

x⊕y= |x j − y j |2− j−1 + |x j − y j |2− j
j<0 j>0
and plays a key role in the Walsh analysis and its applications [5, 6].
For x, y ∈ R+ , put
∞

χ (x, y) = (−1)t (x,y) where t (x, y) = (x j y− j + x− j y j )
j=1
and x j , y j are calculated as in (7.2). For every integer j, we have
χ (x, 2 j ) = χ (2 j x, ) = w1 (2 j , x), ∈ Z+ , x ∈ [0, 2− ).
Further if x, y, z ∈ R+ and x ⊕ y is dyadic irrational, then
χ (x, z)χ (y, z) = χ (x ⊕ y, z). (7.3)

7.1 Biorthogonal Wavelets on R+ 255
See for example Golubav et al. [5]. Thus, for fixed x and z, last equality holds for all
y ∈ R+ , except for countably many.
The inner product and the norm on L 2 (R+ ) will be denoted by ., . and . , respec-
tively. For each function f ∈ L 1 (R+ ) ∩ L 2 (R+ ), its Walsh–Fourier transform fˆ

fˆ(w) = f (x)χ (x, w)d x, w ∈ R+
R+
belongs to L 2 (R+ ). The Walsh–Fourier operator
F : L 1 (R+ ) ∩ L 2 (R+ ) → L 2 (R+ ), F f = fˆ
extends uniquely to the whole space L 2 (R+ ). The properties of the Walsh–Fourier
transform are quite similar to those of the classical Fourier transform. In particular,
if f, g ∈ L 2 (R+ ) then f, g= fˆ, ĝ. (Parseval’s equality).
We denote the support of a function f ∈ L 2 (R+ ) by supp f : it is defined as
the minimal (with respect to inclusion) closed set such that on its complement, f
vanishes almost everywhere. The set of functions in L 2 (R+ ) with compact support
is denoted L 2c (R+ ).
Definition 7.2 A function ϕ ∈ L 2c (R+ ) is called a refinable function, if it satisfies

an equation of the type
2
n
−1
ϕ(x) = ck ϕ(2x ⊕ k), x ∈ R+ , (7.4)
k=0
where ck are complex coefficients.
We denote the characteristic function of a set E ⊂ R+ by 1 E . If c0 = c1 = 1 and

ck = 0 for all k ≥ 2, then ϕ = 1[0,2n−1 0] is a solution of (7.4). In particular, when n = 1
we obtain the Haar function. Some other examples of refinable functions determining
orthogonal wavelets in L 2c (R+ ) are given in Lang [7–9], Farkov [10, 11], Farkov and
Protasov [12].
The functional equation (7.4) is known as the refinement equation. Applying
the Walsh–Fourier transform, we can write this equation as ϕ̂(ω) = m(ω/2)ϕ̂(ω/2)
where
2n −1
1
m(ω) = ck ωk (ω)
2 k=0
is a Walsh polynomial, which is called the mask of the refinable function ϕ.

By a dyadic interval of range n, we shall mean an interval of the form Is(n) =
[s2−n , (s + 1)2−n ), s ∈ R+ . It follows easily from definition that, for every 0 ≤
1 ≤ 2n − 1, the Walsh function ω1 (x) is piecewise constant: on each interval Is(n) , it
is either equal to 1 or to −1. Besides, ωl (x) = 1 for x ∈ [0, 2−n ) and the coefficients
of the refinement equation (7.4) are related to the values bs = m(s2−n ), 0≤s≤
2n − 1 by the discrete Walsh transform:
2 −1
1
n
ck = n bs ωs (k2−n ), 0 ≤ k ≤ 2n − 1 (7.5)
2 s=0
which can be realized by a fast algorithm, see Schipp et al. [6, p. 463]
A set M ⊂ [0, 1) is said to be blocking set for a mask m, if it is the union of dyadic
intervals of range n − 1 or coincides with some of these intervals, does not contain
the interval [0, 2−n+1 ), and such that each point of the set (M/2) ∪ (1/2 + M/2)
which is not contained in M is a zero for m. It is evident that m can have only finitely
many blocking set.
Theorem 7.1 Let ϕ be a function in L 2c (R+ ) satisfying the refinement equation (7.4)
with m as its mask, and let ϕ̂(0) = 1. Then
2
n
−1
ck = 2; supp ϕ ⊂ [0, 2n−1 ]
α=0
and
∞

ϕ̂(ω) = m(2− j ω).
j=1
Moreover, the following properties are true:

ϕ̂(k) = 0 for all k ∈ N (the modified Strang–Fix condition);
(i)
(ii) k∈Z+ ϕ(x ⊕ k) = 1 for almost every x ∈ R+ (the partition of unity property);
(iii) {ϕ(. ⊕ k)|k ∈ Z+ } is a linearly independent system if and only if m does not
have blocking sets.
Definition 7.3 We say that a function f ∈ L 2 (R+ ) is stable, if there exist positive
constants c1 and c2 such that
∞ 21 ∞
∞ 21

c1 |ak |2 ≤ || ak f (/ ⊕ k)|| ≤ c2 |ak |2 (7.6)
k=1 k=1 k=1
for every sequence {ak } from l2 . In other words, a function f is stable in L 2 (R+ ) if
the family { f (. ⊕ k)|k ∈ Z+ } is a Riesz system in L 2 (R+ ).
We say that a function g : R+ → C has periodic zero at a point ω ∈ R+ if g(ω ⊕

k) = 0 for all k ∈ Z+ . The following theorem characterizes compactly supported
stable functions in L 2 (R+ ).
Theorem 7.2 For any f ∈ L 2c (R+ ), the following properties are equivalent:
(i) f is stable in L 2 (R+ )
(ii) { f (. ⊕ k)|k ∈ Z+ } is a linearly independent system;
(iii) the Walsh–Fourier transform of f does not have periodic zeroes.
Theorems 7.1 and 7.2 were proved in Farkov and Protasov [12] (see also their
generalizations in Farkov [13].
Definition 7.4 A family of closed subspaces V j ⊂ L 2 (R+ ), j ∈ Z is called a mul-

tiresolution analysis (or, briefly, an MRA) in L 2 (R+ ) if the following holds:
(i) V j ⊂ V j+1 for j ∈ Z
(ii) ∪V j = L 2 (R+ ) and ∩V j = {0}
(iii) f (.) ∈ V j ⇔ f (2.) ∈ V j+1 for j ∈ Z
(iv) f (.) ∈ V0 ⇒ f (. ⊕ k) ∈ V0 for j ∈ Z+
(v) there exists a function ϕ ∈ L 2 (Z+ ) such that the system {ϕ(. ⊕ k)|k ∈ Z+ } is a
Riesz basis in V0 .
Given a function f ∈ L 2 (R+ ), we set
f j,k (x) = 2 j/2 f (2 j x ⊕ k), j ∈Z k ∈ Z+ .
Definition 7.5 We say that a function ϕ generates an MRA in f ∈ L 2 (R+ ) if, first,
the family {ϕ(. ⊕ k)|k ∈ (Z+ )} is a Riesz system in L 2 (R+ ) and, second, the closed
subspaces V j = span{ϕ j,k |k ∈ R} j ∈ Z form an MRA in f ∈ L 2 (R+ ).
For each refinable function ϕ generating an MRA in f ∈ L 2 (R+ ), we can construct an

orthogonal wavelet ψ such that {ψ j,k } is an orthonormal basis in L 2 (R+ ). Moreover,
according to [13, Theorem 2], for any positive integer n, there exist the coefficients
ck such that the corresponding refinement equation (7.4) has a solution ϕ ∈ L 2c (R+ )
which generates an MRA in L 2 (R+ ). In Sect. 7.2, we study the following problem:
given two masks
2 −1 −1
1
n
2
n
m(ω) = ck ωk (ω), m̃(ω) = c̃k ωk (ω). (7.7)

2 k=0 k=0
How can we find biorthogonal wavelet bases in L 2c (R+ )? Then, in Sect. 7.3, we
give some examples of masks with their application to maximizing the PSNR values
for several images. In these numerical experiments, we find biorthogonal dyadic
wavelets which have some advantages in comparison with the Haar, Daubechies,
and biorthogonal 9/7 wavelets. By analogy with Navikov et al. [14, Proposition
1.1.12], we have the following.
Lemma 7.1 Let ϕ, ϕ̃ ∈ L 2 (R+ ). The systems {ϕ(. ⊕ k)|k ∈ R) and {ϕ̃(. ⊕ k)|k ∈
R) are biorthonormal in L 2c (R+ ) if and only if
∞

ϕ̃(ω ⊕ l)ϕ̃(ω ⊕ l) = 1; for a.e.ω ∈ R+ .
k=0
Let εn be the space of functions defined on R+ and constant on all dyadic intervals
of range n.
Lemma 7.2 If f ∈ L 2 (R+ ) and supp f ⊂ [0, 2n ] then f˜ ∈ εn . In a similar way, if

g ∈ L 1 (R+ ) and supp g ⊂ [0, 2n ], then the inverse Walsh–Fourier transform of g
belongs to εn .
Now, let us represent each l ∈ N in the form of a binary expansion

k
l= μj2j, μ j ∈ {0, 1}, μk = 1, k = k(l) ∈ Z+ (7.8)
j=0
and denote by N0 (n) the set of all positive integers l ≥ 2n−1 for which the ordered sets
(μ j , μ j+1 . . . μ j+n−1 ) of the coefficients in (7.8) do not contain (0, 0, 0, 1). Further,
let
γ (i 1 , i 2 , . . . , i n ) = bs s = i 1 2i + i 2 21 + · · · + i n 2n−1 , i j ∈ {0, 1}
where bs are defined as in (7.5) then we write

cl [m] = γ (μ0 , 0, 0, . . . , 0, 0) i f k(l) = 0
cl [m] = γ (μ1 , 0, 0, . . . , 0, 0)γ (μ0 , μ1 , 0, 0, . . . , 0, 0) i f k(l) = 1
cl [m] = γ (μk , 0, 0, . . . , 0, 0)γ (μk−1 , 0, . . . , 0, 0)γ (μ0 , μ1 , . . . , μn−2 , μn−1 )
i f k = k(l) ≥ n − 1.
Note that in the last product, the subscripts of each factor starting with the second
are obtained by starting those of the preceding factor by one position to the right and
placing one new digit from the binary expansion (7.8) at the vacant first position. If
ϕ generate an MRA in L 2 (R+ ), then the following expansion takes place:

ϕ̃(ω) = 1[0,21−n ) (ω) + cl [m]1[0,21−n ) (ω ⊕ 21−n l), ω ∈ R+ (7.9)
l∈N(n)
where N(n) = {1, 2, . . . , 2n−1 − 1} ∪ N0 (n). The last expansion was obtained in
Farkov (for some partial cases see Lang).
Let r = 2n−1 recall that the joint spectral radius of r × r complex matrices A0
and A1 is defined as
p̃(A0 , A1 ) := lim max{||Ad1 Ad2 . . . Adk ||1/k } d j ∈ {0, 1}, 1 ≤ j ≤ k,

k→∞
where ||.|| is an arbitrary norm on Cr ×r . If A0 = A1 , then p̃(A0 , A1 ) coincides

with the spectral radius p(A0 ). The joint spectral radius of finite-dimensional linear
operators L 0 , L 1 is defined as the joint spectral radius of their matrics in an arbitrary
fixed basis of the corresponding linear space.
Given a refinement equation of the form (1.3), we set (r × r ) matrix To , T1 by
(T0 )i, j = c(2i−2)⊕( j−1) , (T1 )i, j = c(2i−1)⊕( j−1)
where i. j ∈ {1, 2, . . . , r }. Consider the subspace
V := {u = (u 1 , . . . , u r )t |u 1 + · · · + u r = 0}
and denote by L o , L 1 , the restrictions to V of the linear operators defined on the

whole space Cr by the matrices To , T1 , respectively.
Lemma 7.3 Let the mask m of refinement equation (7.4) satisfies m(0) − 1,
m(1/2) = 0 and p̃[m] = p̃(L 0 , L 1 , . . . , L p−1 ) < 1. Then the function ϕ defined by
(1.7) satisfies (1.3) and belongs to L 2 (R+ ). Moreover, we have
−1
2n−1 −1
2n−1
c2k = c2k−1 = 1
k=0 k=0
This lemma can be proved by analogy with the similar results in Farkov, Sects. 2
and 4; see also Farkav et al. [14, Sect. 1.6].
Construction of Biorthogonal Wavelets on R+
Let us define two scaling functions ϕ, ϕ̃ by the Walsh–Fourier transforms
∞
∞

ϕ̃(ω) = m(2− j ω) ϕ̃(ω) = m(2− j ω) (7.10)
j=1 j=1
where ϕ̃(ω) = ϕ̃(ω) = 1 and m, m̃ are as in (1.5). Suppose that
m(1/2) = m̃(1/2) = 0. (7.11)
We set m ∗ (ω) = m(ω)m̃(ω), N = max{n.ñ} and ck = c̃l = 0 for k ≥ 2n , l ≥ 2ñ .

Using the equality ωk (w)ωl (w) = ωk⊕l (w), we obtain
2 −1 2 −1
1 ∗ 1
N N
m ∗ (w) = c ωk (w); ck∗ (w) = cl c̃k⊕l (7.12)

2 k=0 k 2 l=0
Proposition 7.1 If the systems {ϕ(. ⊕ k)|k ∈ Z+ } {ϕ̃(. ⊕ k)|k ∈ Z+ } are biorthonor-
mal in L 2 (R+ ), then
m ∗ (w) + m ∗ (w + 1/2) = 1 f or all w ∈ [0, 1/2). (7.13)
Proof The Walsh–Fourier transforms of ϕ and ϕ̃ belong to ε N −1 (see Theorem 7.1

and Lemma 7.2) Then, by means of Lemma 7.1 for any ω ∈ [0, 1), we have
∞
∞

1= ϕ̃(w ⊕ l)ϕ̃(w ⊕ l) = m ∗ (w/2 ⊕ l/2)ϕ̃(w/2 ⊕ l/2)ϕ̃(w/2 ⊕ l/2).
l=0 l=0
Further, using periodicity of m ∗ , we obtain

∞
∞

1 = m ∗ (w/2) ϕ̃(w ⊕ l)ϕ̃(w ⊕ l) + m ∗ (w/2 + 1/2) m ∗ (w/2 ⊕ l/2)ϕ̃(w/2 ⊕ l/2)ϕ̃(w/2 ⊕ l/2)
l=0 l=0
where both sums equal to 1.

Let us define ϕ ∗ by the formula

ϕ ∗ (x) = ϕ(t ⊕ x)ϕ̃(t)dt.
R+
The Walsh–Fourier transform of ϕ ∗ can be written as ϕ̂ ∗ (w) = ϕ̂(w)ϕ̂(w) (see Gol-

ubov, etc.). Hence, ϕ̂ ∗ = m ∗ (w/2)ϕ̃(w/2). Using the inverse Walsh–Fourier trans-
form, we can deduce from (7.11) that ϕ ∗ is a scaling function and ϕ ∗ satisfies
2
N
−1
∗
ϕ (x) = ck∗ ϕ ∗ (2x ⊕ k) x ∈ R+ . (7.14)
k=0
Thus, m ∗ is the mask of ϕ ∗ .
Proposition 7.2 If one of the masks m, m̃.m ∗ has blocking set. Then, the systems
{ϕ(. ⊕ k)|k ∈ Z+ }, {ϕ̃(. ⊕ k)|k ∈ Z+ } are not biorthonormal in L 2 (R+ ).
Proof Assume that systems
{ϕ(. ⊕ k)|k ∈ Z+ }, {ϕ̃(. ⊕ k)|k ∈ Z+ } (7.15)
arc biorthonormal in L 2 (R+ ). It is well known that each biorthonormal system is

linearly independent. It follows from Theorem 7.1 that m and m̃ have no blocking
sets. Besides, in view of Lemmas 7.1 and 7.2, for any w ∈ [0, 1), we have Φ(w) = 1,
where Φ is defined by the formula
∞

Φ(w) = ϕ̃(w ⊕ l)ϕ̃(w ⊕ l) w ∈ R+ . (7.16)
l=0
Now, let m ∗ have a blocking set. Then, according to Theorem 7.2, {ϕ ∗ (. ⊕ k)|k ∈
Z+ } is a linearly dependent system. Let us show that there is a dyadic interval of a
range N − 1 such that all points of it are periodic zeros of ϕ̃ ∗ Indeed, if k > 2 N , then
support of ϕ ∗ (. ⊕ k) disjoint with [0, 2 N ]. Therefore, the system {ϕ ∗ (. ⊕ k)|k ∈ Z+ }
is linearly dependent only in the case when there exist numbers ao , . . . , a2 N −1 −1 such
that
−1 −1
2 N −1 2 N −1
∗
ak ϕ (. ⊕ k) = 0 and |ak | > 0.
k=0 k=0
Applying the Walsh–Fourier transform, we get

−1
2 N −1
∗
ϕ̃ (w) ak wk (w) = 0 f or a, e, w ∈ R+ .
k=0
N −1
The polynomial ω(w) = 2k=0 −1 ak wk (w) is not identically equal to zero. There-
fore, there is an dyadic interval Ic [0.1) with range N − 1 such that w(I + r ) = 0 for
all r ∈ Z+ . Since ϕ̃ ∗ ∈ ε N −1 we obtain that ϕ̃ ∗ (I + r ) = 0 for all r ∈ Z+ . That is,
on [0, 1), there is a dyadic interval of range N − 1, entirely consisting of periodic
zeros ϕ̃ ∗ But then Φ(w) = 0 on this interval. This contradiction completes the proof
of Proposition 7.2.
Remark 7.1 From identity
1
m ∗ (w) + m ∗ (w + 1/2) = ck c̃k⊕2l wk⊕2l (w)
2 k l
we see that (7.12) is equivalent to the following:

ck c̃k⊕2l = 2δ0,l l ∈ Z+ . (7.17)
k
Besides, it is possible to write (7.12) in the form
bl b̃l + b̃l+2 N −1 b̃l+ν2 N −1 = 1, 0 ≤ l ≤ 2 N −1 − 1, (7.18)
where bl = m(l2−N ), b̃l = m̃(l2−N ).
The family {[0, 2− j ]| j ∈ Z+ } forms a fundamental system of neighborhoods of zero

in the dyadic topology on R+ (e.g., Ref. [6]). A subset E of R+ that is compact in
the dyadic topology is said to be W -compact. It is easy to see that the union of finite
family of dyadic intervals is W -compact.
A W -compact set is said to be congruent to [0, 1) modulo Z+ . If its Lebesgue
measure is equal to 1 and for each x ∈ [0.1) there exists k ∈ Z+ such that x ⊕ k ∈ E+ .
Theorem 7.3 Let m ∗ satisfy the condition (7.13). Then, the systems {ϕ(. ⊕ k)|k ∈
Z+ }, {ϕ̃(. ⊕ k)|k ∈ Z+ } are biorthonormal in L 2 (R+ ) if and only if there exists a
W -compact set E congruent to [0, 1) modulo Z+ , containing a neighborhood of zero
such that
in f l∈N in f w∈E |m(2− j w)| > 0, in f l∈N in f w∈E |m̃(2− j w)| > 0. (7.19)
Note that the condition (7.19) is valid for E = [0.1) if m ∗ (w) = 0 for all w ∈
[0, 1/2).
Example 7.1 Let n = ñ = 2. Suppose that the masks of scaling functions ϕ, ϕ̃ are
given by ⎧
⎪
⎪ 1 w ∈ [0, 1/4),
⎪
⎨a w ∈ [1/4, 1/2),
m(w) =
⎪0 w ∈ [1/2, 3/4),
⎪
⎪
⎩
b w ∈ [3/4, 1),
⎧
⎪
⎪ 1 w ∈ [0, 1/4),
⎪
⎨ã w ∈ [1/4, 1/2),
m̃(w) =
⎪
⎪0 w ∈ [1/2, 3/4),
⎪
⎩
b̃ w ∈ [3/4, 1),
where a ã + bb̃ = 1. If a = 0 or ã = 0, then [1/2.1) is blocking set for m or m̃,

otherwise blocking sets for m, m̃, m ∗ do not exist. Now let |b| < 1 and b̃ < 1. Then
a ã = 1 and blocking sets are absent. Moreover, as above, the condition (7.19) is
satisfied for E = [0, 1). Besides, both functions belong to L 2 (R+ ). Really, from [1,
Example 4.3] and [15, Remark 3] we see that p̃|m| = |b|, p̃|m̃| = |b̃|, hence Lemma
7.3 is applicable. The following decompositions takes place:
⎛ ⎞
∞
ϕ(x) = (1/2)10,1 (x/2) ⎝1 + a b j w2 j+1 −1 (x/2)⎠
j=0
and the similar decomposition is true for ϕ̃. Thus, under the conditions a ã + bb̃ = 1,
|b| < 1, and b̃ < 1, the family of integer shifts of scaling functions ϕ, ϕ̃ forms an
biorthonormal system in L 2 (R+ ).
Example 7.2 Let n = 3.ñ = 2. Suppose m̃ is the same as in Example 7.1 and the
mask m is defined by
m(w) = 1 for w ∈ [0, 1/8), m(w) = a for w ∈ [1/8, 1/4)

m(w) = b for w ∈ [1/4, 3/8), m(w) = c for w ∈ [3/8, 1/2)
m(w) = 0 for w ∈ [1/2, 5/8), m(w) = α for w ∈ [5/8, 3/4)
m(w) = β for w ∈ [3/4, 7/8), m(w) = γ for w ∈ [7/8, 1)
7.2 Biorthogonal Wavelets on Vilenkin Groups
The foundations of the theory of Walsh series and transforms as a branch of modern
harmonic analysis are expounded in Chaps. 1–3, while the harmonics eikt are charac-
ters of the group of rotations of the circle and the Walsh functions are characters of the
Cantor dyadic group. Orthogonal wavelets and the corresponding refinable functions
representable as lacunary Walsh series were studied in [10–12, 14–16]. For p ≥ 2,
the Vilenkin group G can be defined as the weak direct product of a countable set of
cyclic groups of order p (in the case of p = 2, G is isomorphic to the Cantor group).
A peculiarity of the construction of wavelets on the Cantor and Vilenkin groups is
associated with the fact that these groups (as well as the additive group of the p-adic
number field) contain open compact subgroups (see [17]). In the present chapter, by
analogy with compactly supported biorthogonal wavelets on R+ , which are deter-
mined by appropriate chosen trigonometric polynomials (see, e.g., [18, Sect. 8.3.5;
34, Sect. 1.3]), we introduce biorthogonal wavelets and refinable functions whoso
masks are generalized Walsh polynomials. Note that for each p ≥ 2 the generalized
Walsh functions under consideration are characters of the corresponding Vilenkin
group.
Let G be a locally compact abelian group consisting of sequences of the form
x = (x j ) = (. . . , 0, 0, xk , xk+1 , xk+2 . . .),
where x j ∈ {1, 2, . . . , p − 1} for j ∈ Z and x j = 0 for j < k = k(x). The group

operation on G is denoted by and defined as coordinate-wise addition modulo p:
(x j ) ⊕ (y j ) ⇔ z j = x j + y j (mod p) f or j ∈ Z.
The topology on G is determined by the basis of neighborhoods of zero
Ul = {(x j ) ∈ G|x j = 0 f or j ≤ l} l ∈ Z
The group G is known as the Vilenkin group (see, e.g., [6, p. 511]). We denote
the inverse operation of ⊕ by (so that x x = θ , where θ is the zero sequence).
Define as a subgroup of G with the set of elements U0 .
The Lebesgue spaces L q (G) with 1 ≤ q ≤ ∞ are considered with respect to the
Haar measureμ defined on the Borel subsets of G and normalized by the condition
μ(U ) = 1. By (., .) and ||.||, we denote the inner product and the norm on L 2 (G).
The dual group of G is denoted by G ∗ and consists of sequences of the form
w = (w j ) = (. . . , 0, 0, wk , wk+1 , wk+2 . . .)
where w j ∈ {1, 2, . . . , p − 1} for j ∈ Z and w j = 0 for j < k = k(w). Addition and

subtraction, neighborhoods of zero {Ul∗ }, and the Haar measure μ∗ are defined for
G ∗ in the same way as for G. Each character of the group G can be represented as
⎛ ⎞
2πi
χ (x, w) = ex p ⎝ x j wl− j ⎠ , x∈G
p j∈Z
for some w ∈ G ∗ . Consider the discrete subgroup H = {((x j ) ∈ G|x j = 0 f or j >

U } in C and define an automorphism A ∈ Aut G ∗ as (Ax) j = x j+i . It is easy to see
that the quotient group H/A(H ) contains p elements, and the annihilator H − of H
consists of those sequences (w j ) ∈ G ∗ in which w j = 0 for j > 0.
Consider the mapping A : G → (0, +∞) defined by

λ(x) = x j p− j , x = (x j ) ∈ G.
j∈Z
The image of the subgroup H under λ is the set of nonnegative integers: λ(H ) =
Z+ . For each α ∈ Z+ , let h [α] denote the element of H such that λ(h [α] ) = α (in
particular, h [0] = θ ). A mapping λ∗ : G ∗ ⇒ [0, +∞], an automorphism B ∈ Ant G ∗ ,
a subgroup in U ∗ in G ∗ , and elements w[α] in H + are defined by analogy with λ, A, H ,
and h [α] , respectively. Note that χ (Ax, w) = χ (x, Bw) for all x ∈ G and w ∈ G ∗
(i.e., the automorphism B is conjugate to A).
Generalized Walsh functions for the group G can be defined as
Wα (x) = χ (x, w[α] ) α ∈ Z+ , x =∈ G.
These functions are continuous on G and satisfy the orthogonality relations

Wα (x)Wβ (x)dμ(x) = δα,β , α, β ∈ Z+
U
where δα,β denotes the Kronecker delta. It is known that tile system {Wα }is complete
in L 2 (U ). The corresponding system for the group G ∗ is defined by
Wα∗ (w) = χ (h [α],w ), α ∈ Z+ , w =∈ G ∗ .
The system {Wα8 } is orthonormal basis in L 2 (U ∗ ). For each function f ∈ L ! (G) ∩

L 2 (G), its Fourier transform f˜
7.2 Biorthogonal Wavelets on Vilenkin Groups 265

f˜(w) = f (x)χ (x, w)dμ(x), w =∈ G
G
belongs to the space L 2 (G). The Fourier operator
F : L 1 (G) ∩ L 2 (G) → L 2 (G), F f = f˜
admits a standard extension to the whole space L 2 (G). For any f.g ∈ L 2 (G), Parse-
val’s equality ( f, g) = ( f˜, g̃) holds.
We denote the support of a function f ∈ L 2 (G) by supp f ; it is defined am the
minimal (with respect to inclusion) closed set such that on its complement f vanishes
almost everywhere. The set of functions in L 2 (G) with compact support is denoted
by L 2c (G).
Definition 7.6 A function ϕ ∈ L 2c (G) is called a refinable function if it satisfies an

equation of the form
pn −1

ϕ(x) = p aα ϕ(Ax ⊕ h [α] ), x∈G (7.20)
α=0
where a0 are complex coefficients.
We denote the characteristic function of a set E ⊂ G by 1 E . If a0 = · · · = a p−1 =

1/ p and aα = 0 for all α ≥ p, then the function ϕ = 1Un−1 is a solution of equation
(1.1). The corresponding orthogonal wavelets in L 2 (G) have the form

p−1
ϕ (ν) (x) = ενα
p (Ax ⊕ h [α] ), ν = 1, . . . , p − 1
α=0
where ε p = ex p(2πi/ p) ([19] Theorem 2;15, Sect. 4). Some other examples of
refinable functions determining orthogonal wavelets in L 2 (G) are given in [12, 14].
The functional equation (7.20) is known as refinement equation. Applying the
Fourier transform, we can write this equation as
ϕ̃(w) = m(B−1 wϕ̃(B−1 w) (7.21)
where
pn −1

m(w) = aα Wα∗ (w) (7.22)
α=0
is a generalized Walsh polynomial, which is called the mask of the refinable function
ϕ̃. The sets
∗
Un,s = B −n (w[s] ) ⊕ B −n (U ∗ ) 0 ≤ s ≤ pn − 1 (7.23)
arc cosets of the subgroup D −n (U ∗) in the group U ∗ . Each function Wα∗ (.) with
0 ≤ α ≤ p n − 1 is constant on sets (7.22) and the coefficients of the refinement
∗
equation (7.20) are related to the values bs of mask (7.21) on the classes Un,s by the
direct and inverse Vilenkin–Chrestenson discrete transforms;
p −1 n
1
aα = n bs Wα∗ B −n (w[s] ) 0 ≤ α ≤ pn − 1 (7.24)
p s=0
pn −1

bs = aα Wα∗ B −n (w[s] ) 0 ≤ s ≤ pn − 1 (7.25)
α=0
Parseval’s formula for these transformations takes the form

pn −1 p −1 n
1
aα a˜α = n bα b˜α (7.26)
α=0
p s=0
where aα , bs and a˜α , b˜s are arbitrary sets of numbers satisfying equalities (7.23) and
(7.25). The algorithms for calculating the Vilenkin–Chrestenson discrete transforms
are similar to the classical fast Fourier transform algorithms (see, e.g., Schipp et al.
[6])
Theorem 7.4 If a function ϕ ∈ L 2 (G) satisfies Eq. (7.20) and ϕ̃(θ ) = 1, then
pn −1

aα = 1 and suppϕ ⊂ U1−n .
α=0
In the space L 2 (G), this solution of equation (7.20) is unique and is given by
∞

ϕ̃(w) = m(B − j w)
j=1
and satisfies the following conditions:

ϕ̃(h ∗ ) = 0 for all h ∗ ∈ H + \{0} (a modified Strang–Fix condition):
(1)
(2) h∈H ϕ(x ⊕ h) = 1 for a, e, x ∈ G (the partition of unity property).
Definition 7.7 We say that a function f ∈ L 2 (G) is stable if there exist positive
constants c1 and c2 such that
∞ 21 ∞
∞ 21

c1 |ak | 2
≤ || ak f (/ ⊕ k)|| ≤ c2 |ak | 2
.
k=1 k=1 k=1
For every sequence {aα } from l 2 in other words, a function f is stable in L 2 G if the
function f (. ⊕ h)|h ∈ H forms a Riesz system in L 2 (G).
We say that a function g : G → C has periodic zero at a point ω ∈ G ∗ if g(ω ⊕

h ∗ ) = 0 for all h ∗ ∈ H + . The following theorem characterizes compactly supported
stable functions in L 2 (G).
Theorem 7.5 For any f ∈ L 2c (G), the following properties are equivalent: (i) f is
stable in L 2 (G)
(ii) { f (. ⊕ h)|h ∈ H } is a linearly independent system;
(iii) The Walsh–Fourier transform of f does not have periodic zeroes.
Theorems 7.4 and 7.5 were proved in Farkov and Protasov [12] (see also their
generalizations in Farkov [13]).
Definition 7.8 A family of closed subspaces V j ⊂ L 2 (G), j ∈ Z is called a mul-

tiresolution analysis (or, briefly, an MRA) in L 2 (G) if the following hold:
(i) V j ⊂ V j+1 for j ∈ Z
(ii) ∪V j = L 2 (R+ ) and ∩V j = {0}
(iii) f (.) ∈ V j ⇔ f (2.) ∈ V j+1 for j ∈ Z
(iv) f (.) ∈ V0 ⇒ f (. ⊕ k) ∈ V0 for j ∈ Z+
(v) there exists a function ϕ ∈ L 2 (G) Such that the system {ϕ(. ⊕ h)|h ∈ H } is a
Riesz basis in V0 .
Given a function f ∈ L 2 (G), we set
f j,k (x) = p j/2 f (2 j x ⊕ k), j ∈Z h∈H
Definition 7.9 We say that a function ϕ generates an MRA in f ∈ L 2 (G) if, first,
the family {ϕ(. ⊕ h)|h ∈ (H ) is a Riesz system in L 2 (G) and, second, the closed
subspaces V j = span{ϕ j,h |h ∈ H } j ∈ Z form an MRA in f ∈ L 2 (G).
In [1, 14], for arbitrary p.n ∈ N with p ≥ 2, we found coefficient such that the
refinement equation (7.1) has a solution ϕ ∈ L 2 (G) in the form of a lacunary series
in generalized Walsh functions that generates an MRA in L 2 (G). Moreover, for each
refinable function ϕ generating an MRA in L 2 (G), we can construct orthogonal
wavelets ψ (1) , . . . ψ ( p−1) such that the function
ψ (ν)
j,k (x) = p
j/2 (ν)
ψ (A j x ⊕ h), 1≤ν ≤ p−1 j ∈Z h∈H
forms an orthogonal basis in L 2 (G).

Let us expand a refinable function in a lacunary Walsh series. Let l ∈ {0, 1, . . . ,
p − 1}, and let δ1 denote the sequence w = (w j ) in which w1 = 1 and w j = 0 for
j = 1 (in particular, δ0 = 0). It is easy to see that
{w ∈ U ∗ |χ (x, w) = 1 f or x ∈ A(H )} = {δ0 δ1 , . . . , δ p−1 }

i.e., the set of sequences δ is an annihilator of the subgroup A(H ) in H . Note that
Wα∗ = εα,h −n
p and δ1 = B wlpn−1 for 0 ≤ l ≤ p − 1.
Suppose that a function ϕ ∈ L 2 (G) is a solution of the refinement equation (1.1)
and its mask satisfies the conditions

p−1
m(0) = 1 and |m(w ⊕ δl )|2 = 1, w ∈ G∗.
l=0
Then, as shown in [20], we have

⎛ ⎞
1
ϕ(x) = 1U (A1−n x) ⎝1 + c1 [m]Wl (A1−n x)⎠ (7.27)
p n−1 l∈N ( p,n)
where N( p, n) and cl [m] are defined as follows. Let us represent each l ∈ N in the
form of a p-ary expansion

k
l= μj pj μ j ∈ {0, 1, . . . , p − 1}, μk = 0 k = k(l) ∈ Z+ (7.28)
j=0
and denote the set of all positive integers l ≥ p n−1 for which the ordered sets
(μ j , μ j+1 , . . . , μ j+n−1 ) of the coefficients of (1.9) do not contain
(0, 0, . . . , 0, 1), (0, 0, . . . , 0, 2), . . . (0, 0, . . . , 0, p − 1)
by N0 ( p, n). Then N( p, n) = {1, 2, . . . , p − 1} ∪ N0 ( p, n). Let
γ (i 1 , i 2 , . . . , i n ) = bs , s = i 1 p0 + i 2 p1 + · · · i)npn−1 , i j ∈ (0, 1, . . . , 0, p − 1)
where bs are defined as (1.6), then
cl [m] = γ (μ0 , 0, 0 . . . , 0, 0), if k(l) = 0
cl [m] = γ (μ1 , 0, 0 . . . , 0, 0)γ (μ0 , 0, 0 . . . , 0, 0), if k(l) = 1
........................................
cl [m] = γ (μk , 0, 0 . . . , 0, 0)γ (μk−1 , 0, 0 . . . , 0, 0), . . . , γ (μ0 , 0, 0 . . . , 0, 0) i f k(l) ≥ n − 1

Note that in the last product, the subspaces of each factor starting with the second
are obtained by shifting those of the preceding factor by one position to the right and
placing one new digit from the p-ary decomposition (1.9) at the vacant first position.
7.3 Construction of Biorthogonal Wavelets on The Vilenkin

Group
Suppose that we are given two refinable functions ϕ and ϕ̃ with masks
pn −1 pn −1

m(ω) = aα Wα∗ (ω), and m̃(ω) = ãα Wα∗ (ω), (7.29)
α=0 α=0
respectively. We are interested in the following questions:

(i) When do the H -shifts of the functions ϕ and ϕ̃ form a biorthonormal system
in L 2 (G)?
(ii) How can one construct biorthogonal bases in L 2 (G) from masks (2.1)?
Proposition 7.3 Let ϕ, ϕ̃ ∈ L 2 (G) the systems {ϕ(. ⊕ h)|h ∈ H } and {ϕ̃(. ⊕ h)|h ∈
H } be biorthonormal in L 2 (G) ,if and only if

ϕ̃(ω ⊕ h ∗ )ϕ̃ω ⊕ h ∗ = 1, f or a, e, ω ∈ G ∗ .
h ∗ ∈H +
Proposition 7.4 Let ϕ, ϕ̃ ∈ L 2 (G) be refinable functions with masks m and m̃,
respectively. If the {ϕ(. ⊕ h)|h ∈ H } and {ϕ̃(. ⊕ h)|h ∈ H } are biorthonormal in
L 2 (G), then

p−1
m(ω ⊕ δl )m̃ω ⊕ δl = 1, f or all ω ∈ G∗. (7.30)
l=0
Analogs of Propositions 7.3 and 7.4 were proved in [1, Sect. 3] and [13, Sect. 1.2]
Given masks (7.28), we set m ∗ (ω) = m(ωm̃ω) and N = max{n, ñ}. Condition
(7.29) is then written in the form

p−1
m ∗ (ω ⊕ δl ) = 1, ω ∈ G∗
l=0
and is equivalent to the equalities

p−1
(N ) (N )
bl+νp N −1 b̃l+νp N −1 , 0 ≤ l ≤ p N −1 − 1 (7.31)
ν=0
where b1(N ) = m(B −N ω[l] ) and b̃1(N ) = m̃(B −N ω[l] ) Let ⊕ p and p denote, respec-
tively. addition and subtraction of integers modulo p. The equality Wα∗ (ω)Wβ∗ (ω) =
∗
Wα⊕ pβ
(ω) implies
pn −1 pn −1

∗ ∗
m (ω) = aα ãβ Wα⊕ pβ
(ω)
α=0 β=0
setting aα = ãβ = 0 for α ≥ p n and β ≥ p n , we obtain
p N −1 p N −1

∗
m (ω) = aα Wα∗ (ω) aα∗ aγ ãγ ⊕ p α .
α=0 γ =0
Consider the function ϕ ∗ defined by

ϕ ∗ (x) = ϕ(t ⊕ x)ϕ̃(t)dμ(t).
G
The Fourier transform of this function is related to those of ϕ and ϕ̃ by ϕ̃ ∗ (ω) =

ϕ̃(ω)ϕ̃(ω) Moreover, ϕ̃ ∗ is a refinable function satisfying the equation
p N −1

∗
ϕ (x) = p aα∗ ϕ ∗ (Ax ⊕ α), x ∈ G.
α=0
Thus, the polynomial m ∗ is the mask of the function ϕ ∗ . For M ⊂ U ∗ , let

p−1
Sp M = {B −1 ω[l] + B −1 (ω)|ω ∈ M}.
t=0
Suppose that M either coincides with one of the sets ∪∗n−1,s,1 ≤ s ≤ p n−1 − 1, or
is the union of one of these sets. Such a set M is said to be blocking for a mask if
m(ω) = 0 for all ω ∈ S p M\M.
Proposition 7.5 If one of the masks m, m̃, m ∗ has a blocking set, then the systems
{ϕ(. ⊕ h)|h ∈ H } and {ϕ̃(. ⊕ h)|h ∈ H } are not biorthogonal in L 2 (G).
This proposition is proved by using Theorem 7.5 (the case of m = m̃ was consid-
ered in detail in [14]).
Let E be a compact set in G ∗ . The set E is said to be congruent to U ∗ modulo
H μ (E) = 1 and, for any ω ∈ E, there exists an h ∗ ∈ H + such that ω ⊕ h ∗ ∈ U ∗ .
+ ∗
The following analog of the well-known Cohen’s criterion (see, e.g., [17, Theorem
2.5.6] is valid.
7.3 Construction of Biorthogonal Wavelets on The Vilenkin Group 271
Theorem 7.6 If ϕ and ϕ̃ are refinable functions whose masks m and m̃ satisfy con-
dition (7.29) and ϕ̃(0) = ϕ̃(0) = 1, then the following conditions are equivalent:
(a) the systems {ϕ(. ⊕ h)|h ∈ H } and {ϕ̃(. ⊕ h)|h ∈ H } are biorthonormal in
L 2 (G).
(b) there exists a set E that is congruent to U ∗ modulo H + contains a neighbor-
hood of zero in G ∗ and is such that
inf inf |m(B − j ω)| > 0 boxand inf inf |m̃(B − j ω)| > 0. (7.32)
j∈N ω∈E j∈N ω∈E
For ϕ = ϕ̃, this theorem was proved in Sect. 5.1 (cf. Theorem 5.2). When constructing
biorthogonal wavelet systems, one of the main questions is whether the functions ϕ
and ϕ̃ belong to the class L 2 . In some cases, this question can be answered by using
the joint spectral radius of special linear operators defined for the corresponding
refinement equations (see, e.g., Theorem A,6.5 in [13]). Let r = p n−1 . Recall that
the joint spectral radius of r × r complex matrices A0 , A1 , . . . , A p−1 is defined as
p̃(A0 , A1 , . . . , A p−1 ) = lim max{||Ad1 , Ad2 , . . . , Adk ||1/k : d j ∈ {0, 1, . . . , p − 1}, 1 ≤ j ≤ k},
k⇒∞
where ||.|| is an arbitrary norm on C r ×r In the case of A0 = A1 = · · · = A p−1 ,

ϕ̃(A0 , A1 , . . . A p−1 ) coincides with the spectral radius p(A0 ). The joint spectral
radius of finite-dimensional linear operators L 0 , L 1 , . . . L p−1 is defined as the joint
spectral radius of their matrices in an arbitrary fixed basis of the corresponding linear
space.
Given a refinement equation of the form (7.19), we set cα = paα and define r × r
matrices T0 , T1 , . . . , T p−1 by
(T0 )i, j = c( pi− p)⊕ p ( j−1) , (T1 )i, j = c( pi− p+1)⊕ p ( j−1) , . . . , (T p−1 )i, j = c( pi−1)⊕ p ( j−1) .
for i, j ∈ {1, 2, . . . , r }. Consider the subspace
V = {u = (u 1 , . . . , u r )t |u 1 + · · · + u r = 0}
and denote by L 0 , L 1 , . . . , L p−1 the restrictions to V oi the linear operators defined

on the whole space by the matrices T0 , T1 , . . . , T p−1 , respectively.
Proposition 7.6 If the mask m of the refinement equation (7.19) satisfies the condi-
tions
m(0) = 1, m(δ1 ) = m(δ2 ) = · · · = m(δ p−1 ) = 0
and p̃[m] = p̃(L 0 , L 1 , . . . , L p−1 ) < 1, then the function ϕ defined by (7.26) satisfy
Eq. (7.19) and is continuous on G.
7.4 Frames on Vilenkin Group
In this section we discuss construction of frames on Vilenkin group based on [3, 25,
34, 38–41].
Let C be the Cantor dyadic group. The Walsh functions can be identified [1] with
characters of the group C . The family of compactly supported orthogonal MRA-
based wavelets on C was constructed by Lang [7]. Moreover, Lang [8] has found
wavelets that generate unconditional bases for the spaces L q (C ), 1 < q < ∞.
Let an integer p ≥ 2 be fixed. The Vilenkin group G = G p consists of sequences
x = (x j ), where x j ∈ {0, . . . , p − 1} for j ∈ Z and there exist at most a finite number
of negative j such that x j = 0 (the case p = 2 corresponds to the group C ). The
group operation on G is the coordinate-wise addition modulo p, while the sets
Ul := {(x j ) ∈ G | x j = 0 for j < l}, l ∈ Z,
form a complete system neighbourhoods of the zero element θ . Denote by μ the

Haar measure on G normalized so that μ(U ) = 1 where U := U0 .
One can show that G is self-dual. The duality pairing on G takes x, ω ∈ G to
⎛ ⎞
2πi
χ (x, ω) = exp ⎝ x j ω1− j ⎠ .
p j∈Z
The Fourier transform:

f (ω) = f (x)χ (x, ω) dμ(x), ω ∈ G.
G
We define a map λ : G → [0, +∞) by

λ(x) = x j p− j , x = (x j ) ∈ G.
j∈ Z
Take in G a discrete subgroup H = {(x j ) ∈ G | x j = 0 for j > 0}. The image of H

under λ is the set of non-negative integers: λ(H ) = Z+ . For every k ∈ Z+ , let h [k]
denote the element of H such that λ(h [k] ) = k. The Walsh functions on G are defined
by
Wk (x) = χ (x, h [k] ), x ∈ G, k ∈ Z+ .
Let A be the automorphism of G defined by (Ax) j = x j+1 for x = (x j ) ∈ G. The

sets
Un, s := A−n (h [s] ) ⊕ A−n (U ), 0 ≤ s ≤ p n − 1,
are cosets of the subgroup A−n (U ) in the group U . For every 0 ≤ k ≤ p n − 1 the
function Wk is constant on each set Un, s .
7.4 Frames on Vilenkin Group 273
For all integers p, n ≥ 2, necessary and sufficient conditions for the solutions of
the corresponding refinement equations with p n coefficients to generate an MRA
in L 2 (G p ) were put forward in [25]. Some of these solutions can be written as the
Walsh series:

ϕ(x) = (1/ p n−1 )1U (A1−n x)(1 + dl Wl (A1−n x)), x ∈ G,
l∈ M
where 1U is the characteristic function of U (cf. Proposition 7.7 below). Further

results (see the expository paper [34]) were concerned with the following topics:
• algorithms for constructing orthogonal and biorthogonal compactly supported

wavelets on G p ;
• algorithm for constructing an MRA with a Riesz basis for L 2 (G p );
• smoothness estimates of Lang’s wavelets;
• unconditional convergence of wavelet expansions in the dyadic Hardy space;
• approximation properties of dyadic wavelet expansions;
• periodic wavelets in Walsh analysis;
• discrete wavelets related to the Vilenkin–Chrestenson transform;
• nonstationary wavelets related to the Walsh functions;
• wavelets associated with non-uniform multiresolution analysis;
• wavelet packets related to the Walsh polynomials;
• applications for digital signal processing.
Below we describe the basic construction of MRA-based tight frames on the

Vilenkin/Cantor groups. The direct and the inverse Vilenkin–Chrestenson transforms:
p −1 n
1
aα = n bs Wα (A−n h [s] ), 0 ≤ α ≤ p n − 1, (7.33)
p s=0
pn −1

bs = aα Wα (A−n h [s] ), 0 ≤ s ≤ p n − 1. (7.34)
α=0
These discrete transforms can be realized by fast algorithms (e.g., [1]).

For any f ∈ L 2 (G) we let
f j,k (x) := p j/2 f (A j x h [k] ), j ∈ Z, k ∈ Z+ .
Given := {ψ (1) , . . . , ψ (r ) } ⊂ L 2 (G) with r ≥ p, we define the wavelet system

as
X () := {ψ (ν)
j,k : 1 ≤ ν ≤ r, j ∈ Z, k ∈ Z+ }.
The system X () is called a Parseval frame for L 2 (G) if

r
| f , ψ (ν)
j,k | = f
2 2
j∈Z k∈Z+ ν=1

for all f ∈ L 2 (G). This is equavelent to f = g∈X () f , gg for all f ∈ L 2 (G).
Using the unitary extension principle (e.g. [40]), the Parseval frames for L 2 (G)
can be constructed by the following algorithm:
Algorithm A.
• Step 1. Choose numbers bs , 0 ≤ s ≤ p n − 1, so that
b0 = 1, |bl |2 + |bl+ pn−1 |2 + · · · + |bl+( p−1) pn−1 |2 ≤ 1, 0 ≤ l ≤ p n−1 − 1.

(7.35)
• Step 2. Compute aα , 0 ≤ α ≤ p n − 1, by 7.33 and define
pn −1

m 0 (ω) = aα Wα (ω).
α=0
• Step 3. Define ϕ ∈ L 2 (G) such that

∞

ϕ (ω) = m 0 (A− j ω), ω ∈ G. (7.36)
j=1
• Step 4. Given r ≥ p find the Walsh polynomials
pn −1

m ν (ω) = aα(ν) Wα (ω), 1 ≤ ν ≤ r,
α=0
such that, for each ω ∈ G, the rows of the following matrix form an orthonormal
system:
⎛ ⎞
m 0 (ω) m 1 (ω) . . . m r (ω)
⎜ m 0 (ω ⊕ δ1 ) m 1 (ω ⊕ δ1 ) . . . m r (ω ⊕ δ1 ) ⎟
M(ω) = ⎜
⎝ ...
⎟.
⎠
... ... ...
m 0 (ω ⊕ δ p−1 ) m 1 (ω ⊕ δ p−1 ) . . . m r (ω ⊕ δ p−1 )
• Step 5. Define ψ (1) , . . . , ψ (r ) as follows:
pn −1

ψ (ν) (x) = p aα(ν) ϕ(Ax h [α] ), 1 ≤ ν ≤ r.
α=0
7.4 Frames on Vilenkin Group 275
Theorem 7.7 Suppose that ϕ ∈ L 2 (G),

ϕ (θ ) = 1, satisfies a refinement equation
pn −1

ϕ(x) = p aα ϕ(Ax h [α] ), x ∈ G,
α=0
or
p n −1

−1 −1
ϕ (ω) = m 0 (A ω)
ϕ (A ω), m 0 (ω) = aα Wα (ω).
α=0
Then ϕ generates a Parseval frame for L 2 (G) iff its mask m 0 satisfying

p−1
|m 0 (ω ⊕ δl )|2 ≤ 1, ω ∈ G, (7.37)
l=0
where δl ∈ G, λ(δl ) = l/ p, l = 0, . . . , p − 1.
Note: (7.37) can be written as (7.35) with bs = m 0 ((A−n h [s] )).

Assume that bs satisfy condition (7.35). For each 0 ≤ s ≤ p n − 1 we set
γ (i 1 , i 2 , ..., i n ) = bs ,
if
s = i 1 p 0 + i 2 p 1 + · · · + i n p n−1 , i j ∈ {0, 1, . . . , p − 1}.
Then for an integer l with the p -ary expansion

k
l= μ j p j , μ j ∈ {0, 1, . . . , p − 1}, μk = 0, k = k(l) ∈ Z+ , (7.38)
j=0
we define dl as follows
dl = γ (μ0 , 0, 0, . . . , 0, 0), if k(l) = 0;
dl = γ (μ1 , 0, 0, . . . , 0, 0)γ (μ0 , μ1 , 0, . . . , 0, 0), if k(l) = 1;
...................................................................
dl = γ (μk , 0, 0, . . . , 0, 0)γ (μk−1 , μk , 0, . . . , 0, 0) . . . γ (μ0 , μ1 , μ2 , . . . , μn−2 , μn−1 ),
if k = k(l) ≥ n − 1. Further, denote by M 0 the set of all positive integers l ≥ p n−1

whose p -ary expansion (7.38) contains no n-tuple (μ j , μ j+1 , . . . , μ j+n−1 )
coinciding with any of the n-tuples
(0, 0, . . . , 0, 1), (0, 0, . . . , 0, 2), . . . , (0, 0, . . . , 0, p − 1).
and set M = {1, 2, . . . , p n−1 − 1} ∪ M 0 . In particular, if n = 2 then

k
M= μ j p : μ j ∈ {1, 2, . . . , p − 1}, k ∈ Z+ .
j
j=0
Proposition 7.7 Suppose that m 0 and dl are defined as above. If

ϕ is given by
(Theorem 7.7), then
⎧
⎨ 1, ω ∈ Un−1, 0 ,

ϕ (ω) = dl , ω ∈ Un−1, l , l ∈ M,
⎩
0, ω ∈ Un−1, l , l ∈ / M.
Let us set
bs(ν) = m ν (A−n h [s] ), 0 ≤ s ≤ p n − 1.
Then, by Step 1, we have
b0(0) = 1, |bl(0) |2 + |bl+

(0) 2 (0)
pn−1 | + · · · + |bl+( p−1) pn−1 | ≤ 1, 0 ≤ l ≤ p
2 n−1
− 1.
As usual, let M ∗ denote the adjoint matrix of a matrix M . According to Step 4, for
each l, we must find
bl(1) , bl+
(1) (1) (r ) (r ) (r )
pn−1 , . . . , bl+( p−1) pn−1 , . . . , bl , bl+ pn−1 , . . . , bl+( p−1) pn−1 ,
such that the matrices

⎛ ⎞
bl(0) bl+
(0)
pn−1 ... (0)
bl+( p−1) pn−1
⎜ (1) (1) (1) ⎟
⎜ bl bl+ pn−1 ... bl+( p−1) pn−1 ⎟
Ml = ⎜ ⎟.
⎝ ... ... ... ... ⎠
bl(r ) bl+
(r )
pn−1 ... (r )
bl+( p−1) p n−1
satisfy the condition Ml Ml∗ = I , where I is the identity matrix. To find such matrices
can be used the methods given in [38].
Remark 7.2 1. If all inequalities in Step 1 become equalities (as for the orthogonal
wavelets case, cf. [34]), then Algorithm A can be applied to r = p − 1.
2. The Parseval frames on the group C related to the generalized Walsh-Dirichlet
kernel (which are not MRA-based wavelet frames) are defined in [3].
7.5 Application to Image Processing 277
7.5 Application to Image Processing
In image processing, there are several methods of finding the best wavelet for each
input image from within a class of wavelets of parameters [34]. The so-called
“feedback-based” approach to this problem can be broken down into three basic
steps (a) coding the image by encoding the wavelet coefficient of the input image (b)
decoding (reconstruction) the image and calculating the PSNR, and (c) updating the
parameters to achieve the best PNSR [9]. Let us recall that for given N × N input
image f (x, y) and output image g(x, y), the PNSR value is defined by
255N
PNSR = 20lg .
N
x,y=1 ( f (x, y) − g(x, y))1/2
As usual, the Daubechies wavelet of order N will be denoted through DN (see,

e.g., [2, 15, 17, 36]). In [2], the following method which we shall call Method A
was used:
(i) Apply a direct discrete wavelet transform to the wavelet coefficient if the input
image.
(ii) Allocate a certain percentage of the wavelet coefficient with the largest absolute
values (10%, 5%, 1%) and nullify the remaining coefficient.
(iii) Apply a reverse wavelet transform to the modified array of the wavelet coef-
ficient.
(iv) Calculate the PNSR value.
Using this method, it is shown in [17] that the wavelets have an advantage over
the Haar wavelet for the “lena” image. In the following, we compare the Haar,
Daubechies, and biorthogonal 9/7 wavelets for several images with the wavelets
from Examples 2.1 and 2.2 (marked DBW4 and DBW8, respectively) and also with
the orthogonal wavelets (marked DOW8) from the next example
Example 7.3 Let ϕ be the solution of equation (7.4) for n = 3 with the coefficients
1 1
c0 = (1 + a + b + c + α + β + γ ), c1 = (1 + a + b + c − α − β − γ )
4 4
1 1
c2 = (1 + a − b − c + α + β − γ ), c3 = (1 + a − b − c − α − β + γ )
4 4
1 1
c4 = (1 − a + b − c + α − β − γ ), c5 = (1 − a + b − c − α − β + γ )
4 4
Fig. 7.1 Reconstructed “lena” image for Haar (top left), D8 (top right), DOW8 (bottom left), and
DBW8 (bottom right) wavelets
1 1
c6 = (1 − a − b + c − α − β + γ ), c7 = (1 − a − b + c + α + β − γ ),
4 4
where |a|2 + |α|2 = |b|2 + |β|2 = |a|2 + |γ |2 = 1, a = 0, c = 0. The corre-

sponding wavelet (see [11, 37]) is defined by

ψ(t) = (−1)k c1⊕k ϕ(2t ⊕ k).
k∈Z+
Let us recall the definitions of the direct

√ and reverse biorthogonal dyadic
√ wavelet
transforms. Suppose that h k = ck / 2, gk = (−1)k h 1⊕k , h̃ k = c̃k / 2, g̃k =
(−1)k h̃ 1⊕k . Then, the approximating and detailing wavelet coefficients are calcu-
lated by the formulas
Fig. 7.2 Reconstructed “rentegen” image for Haar (top left), D8 (top right), DOW8 (bottom left),
and DBW8 (bottom right) wavelets

ak = h l⊕2k xl dk = g̃l⊕2k xl .
l l
The reconstruction formula is

xl = h̃ l⊕2k + gl⊕2k dk .
k
In numerical experiments, Method A and its two modified versions were tested on
three images: “lena”, “rentgen”, and “bird”. All images are grayscale and have 256 ×
256 size. To find the best set parameters, we used the Nelder–Mead simplex method
(see, e.g., Ref. [2]) with PSNR as the target function. As free parameters, we choose
a, b, α̃ for Example 2.1, b, c, β, γ for Example 2.2, and a, b, c for Example 3.1.
Fig. 7.3 Reconstructed “bird” image for Haar (top left), D8 (top right), and DBW8 (bottom)
wavelets
When we used Method A with feedback process, the best PSNR values for the dyadic
wavelets are the same as for the Haar wavelet. In Method B, we replace Step 2 of
Method A on the uniform quantization of wavelet coefficient tliat is, a line segment.
containing all wavelet coefficients. are divided into several segments of equal length,
and then the set of wavelet coefficients that are within at the subsegment replaces by
its center). In Step 2 of Method C, we used the corresponding algorithm from the
.JPEG2000 standard. So each wavelet coefficient ai, j was quantized by the formula
a
i, j
ai, j = sign(ai, j ) ,
Δ
where Δ is the length of the quantization interval. The most significant results were
obtained using Method B with Δ = 50; see Table 7.1 and the reconstructed images
in Figs. 7.1, 7.2, and 7.3. We see that for considered images, the introduced dyadic
Table 7.1 PSNR values for Method B with Δ = 50

Image Haar D4 D8 D9/7 DOW8 DBW4 DBW8
Lena 23,139304 23,115207 23,064597 23,094191 23,819995 27,648008 28,555064
Bird 32,326904 21,865437 21,586381 21,35232 32,326904 33,07666 33,450892
Rentgen 21,342584 21,20615 21,192967 21,161501 22,159873 27,819712 31,704022
wavelets have an advantage over the standard Haar, Daubechies, and biorthogonal
9/7 wavelets. For details, see [34, 35].
References
1. Farkov, Yu. A. (2008). Multiresolution analysis and wavelets on vilenkin groups. Facta
2. Farkov, Yu. A, Maksimov, A. Yu., & Stroganov, S. A. (2011). On biorthogonal wavelets related
to the Walsh functions. International Journal of Wavelets, Multiresolution and Information
Processing, 9(3), 485–499.
3. Farkov, Yu. A. (2012). Examples of frames on the Cantor dyadic group. Journal of Mathe-
matical Sciences, 187(1), 22–34.
Translation of 1st ed.). Moscow: Urss; Dordrecht: Kluwer (1991).
7. Lang, W. C. (1996). Orthogonal wavelets on the Cantor dyadic group. SIAM Journal on
Mathematical Analysis, 27(1), 305–312.
9. Lang, W. C. (1998). Wavelet analysis on the Cantor dyadic group. Houston Journal of Math-
ematics, 24, 533–544.
10. Farkov, Yu. A. (2005). Orthogonal p-wavelets on R + . In Proceedings of International Con-
ference Wavelets and Splines (St. Petersburg, Russia, July 3–8) (p. 426). St. Petersberg: St.
Petersberg University Press.
11. Farkov, Yu. A. (2005). Orthogonal Wavelets with compact support on locally compact abelian
groups. Izv. Ross. Akad. Nauk Ser. Mat. 69(3), 193–220. English translation, Izvestia: Math-
ematics 69(3), 623–650.
12. Farkov, Yu. A. & Protasov V. Yu. (2006). Dyadic wavelets and refinable functions on a half
line. Mat. Sbornik, 197(10), 129–160. English translation, Sbornik: Mathematics, 197, 1529–
1558.
13. Farkov, Yu. A. (2009). Biorthogonal wavelets on Vilenkin groups. Tr. Mat. Inst.
Steklova 265(1), 110–124. English translation, Proceedings of the Steklov Institute of Math-
ematics, 265(1), 101–114.
14. Novikov, I. Ya., Protasov, V. Yu., & Skopina, M. A. (2011). Wavelet theory (Moscow, 2006).
Providence: AMS.
16. Farkov, Yu. A., & Rodionov, E. A. (2009). Estimates of the smoothness of dyadic orthogonal
17. Welstead, S. (2000). Fractal and wavelet image compression techniques. Bellingham: SPIE
Optical Engineering Press.
18. Hereford, J., Roach, D. W., & Pigford, R. (2003). Image compression using parameterized
wavelet with feedback. Proceeding SPIE, 5102, 267–277
19. Kozyrev, S. V. (2002). Wavelet analysis as a p-adic spectral analysis. Izvestiya: Mathematics,
(66), 367–376.
20. Benedetto, J. J., & Benedetto, R. L. (2009). The construction of wavelet sets, wavelets and
multiscale analysis. In Theory and Applications. Selected papers based on the presentations
at the International conference on Wavelets: Twenty years of Wavelets. Cohen, J. et al. (ed.) De
Paul University, Chicago, IL, USA, May 15–17, New York: Springer (Alied And Numerical
Harmonic Analysis), (pp. 17–56).
21. Agaev, G. H., Vilemkin, N. Ya., Dzhafarli, G. M., & Rubinstein, A. I. (1981). Multiplicative
systems of functions and analysis on 0 dimensional groups. Baku: ELM. [In Russian].
22. Pontryagin, L. S. (1939). Topological groups. New Jersey: University Press Princiton.
23. Edwards, R. E. (1982). Fourier series: A modern introduction (Vol. 2). Berlin: Springer.
24. Benedetto, J. J., & Benedetto, R. L. (2004). A wavelet theory for local fields and related
groups. Journal of Geometric Analysis, 14, 423–456.
25. Farkov, Yu. A. (2007). Orthogonal wavelets on direct products of cyclic groups. Matematich-
eskie Zametki, 82(6), 934–952. English Translation, Mathematical Notes, 82(6), 843–859.
26. Protasov, V. Yu., & Farkov, Yu A. (2006). Dyadic wavelet and refinable functions on a half
line. Sboinik: Mathematics, 197(10), 1529–1558.
27. Chui, C. K., & Mhaskar, H. N. (1993). On trigonometric wavelets. Constructive Approxima-
tion, 9, 167–190.
28. Shah, F. A. (2012). Biorthogonal p-wavelet packets related to the Walsh polynomials. Journal
of Classical Analysis, 1(2), 135–146.
29. Farkov, Yu. A. (2010). Wavelets and frames based on Walsh-Dirichlet type kernels. Commu-
nications in Mathematics and Applications, 1, 27–46.
31. Farkov, Yu. A. (2009). On wavelets related to Walsh series. Journal of Approximation Theory,
161, 259–279.
32. Farkov, Yu. A, & Rodionov, E. A. (2012). Nonstationary wavelets related to the Walsh func-
tions. American Journal of Computational Mathematics, 2, 82–87.
33. Farkov, Yu. A, & Stroganov, S. A. (2011). The use of discrete dyadic wavelets in image
processing. Russian Mathematics (Iz. Yuz), 55(7), 47–55. Original Russian Text Published in
Izvestiya. Uchebnykh Zavednic, Mathematica, (7), 57–66.
34. Farkov, Yu. A. (2015). Construction of MRA- based wavelets and frames in Walsh analysis.
Poincare Journal of Analysis and Applications, 2015(2), Special Issue (IWWFA-II, Delhi),
13–36.
35. Farkov, Yu. A. (2014). Wavelet expansions on the Cantor group. Mathematical Notes, 96(6),
996–1007.
37. Farkov, Yu. A., & Rodionov, E. A. (2011). Algorithms for wavelet construction on Vilenkin
groups. P-Adic Numbers, Ultrametric Analysis, and Applications, 3(3), 181–195.
38. Farkov, Yu. A., Lebedeva, E. A., & Skopina, M. A. (2015). Wavelet frames on Vilenkin groups
and their approximation properties. International Journal of Wavelets, Multiresolution and
Information Processing, 13(5). https://doi.org/10.1142/5021.
39. Farkov, Yu. A. (2016). MRA-Based Wavelet Frames in Walsh Analysis, Conference on Har-
monic Analysis and Approximation Theory, (CRM, Bellaterria, Spain).
40. Dong, B. & Shen, Z. (2013). Framelets: MRA-based wavelet frames and applications, in
Mathematics in image processing / Hogkai Zhao editor.- (IAS/Park City mathematics series)
7-208.
41. Farkov, Yu. A. Wavelet frames related to Walsh functions. European Journal of Mathematics.
https://doi.org/10.1007/s40879-018-0220-6. (in press).
Chapter 8
Wavelets Associated with Nonuniform
Multiresolution Analysis on Positive Half
Line
8.1 Introduction
Gabardo and Nashed considered a generalization of the notion of multiresolution

analysis, which is called nonuniform multiresolution analysis (NUMRA) and is based
on the theory of spectral pairs. In this set up, the associated subspace V0 of L2 (R)
has, as an orthonormal basis, a collection of translates of the scaling function φ of
the form {φ(x − λ)}λ∈Λ where Λ = {0, r/N } + 2Z, N ≥ 1 is an integer and r is an
odd integer with 1 ≤ r ≤ 2N − 1 such that r and N are relatively prime and Z is
the set of all integers. The main results of Gabardo and Nashed deal with necessary
and sufficient condition for the existence of associated wavelets and extension of
Cohen’s theorem.
The objective of this chapter is to study analogue of Gobardo’s and Nashed’s
results concerning nonuniform MRA on L2 (R+ ). More precisely, we study nonuni-
form MRA on positive half line [1, 2]. Nonuniform wavelet frames are also discussed
[3].
Definition 8.1 Given an integer N ≥ 1 and an odd integer r with 1 ≤ r ≤ 2N − 1

such that r and N are relatively prime, an associated nonuniform multiresolution
analysis (abbreviated NUMRA) is a collection {Vj }j∈Z of closed subspaces of L2 (R)
satisfying the following properties:
(a) Vj ⊂ Vj+1 ∀ j ∈ Z,
(b) ∪j∈Z Vj is dense in L2 (R),
(c) ∩j∈Z Vj = {0},
(d) f (x) ∈ Vj if and only if f (2Nx) ∈ Vj+1 ,
(e) There exists a function φ ∈ V0 , called a scaling function such that the collection
{φ(x − λ)}λ∈Λ , where Λ = {0, r/N } + 2Z is a complete orthonormal system for
V0 .
It is worth noticing that when N = 1, one recovers from definition above the standard
definition of a one-dimensional MRA with dyadic dilation. When N > 1, the dilation

https://doi.org/10.1007/978-981-13-6370-2_8
284 8 Wavelets Associated with Nonuniform Multiresolution …
factor of 2N ensures that 2N Λ ⊂ 2Z ⊂ Λ. However, the existence of associated

wavelets with the dilation 2N and translation set Λ is no longer guaranteed as is the
case in the standard setting.
Given a NUMRA, we denote by Wm the orthogonal complement of Vm in Vm+1 ,
for any integer m. It is clear from (a), (b), and (c) of Definition 8.1 that
L2 (R) = ⊕m∈Z Wm .
Definition 8.2 A collection {ψk }k=1,2,...,2N −1 of functions in V1 will be called a set

of wavelets associated with a given NUMRA if the family of functions {ψk (x −
λ)}k=1,...,2N −1,λ∈Λ is an orthonormal system for W0 .
The following result proved in provides the simple necessary and sufficient conditions
for the existence of the associated set of wavelets:
Theorem 8.1 Consider a NUMRA with associated parameters N and r, as in Def-

inition 8.1, such that the corresponding space V0 has an orthonormal system of the
form {φ(x − λ)}λ∈Λ , where Λ = {0, r/N } + 2Z, φ̂ satisfies the scaling relation
φ̂(2N ξ ) = m0 (ξ )φ̂(ξ ),
where φ̂ denotes the Fourier transform of a function φ and m0 has the form
m0 (ξ ) = m10 (ξ ) + e−2πiξ r/N m20 (ξ ),
for some locally L2 , 1/2-periodic functions m10 and m20 . Define M0 by
M0 (ξ ) = |m10 (ξ )|2 + |m20 (ξ )|2 .
Then, each of the following conditions is necessary and sufficient for the existence
of associated wavelets ψ1 , . . . , ψ2N −1 :
(a) M
0Nis 1/4-periodic.

−1
k=0 δk/2 ∗ j∈Z δjN ∗ |φ̂| = 1.
2
(b)
(c) For any odd integer m, we have

φ(x)φ(x − m/N ) dx = 0.
R
Given a 1-periodic trigonometric polynomial m0 satisfying the identities m0 (0) =

1 and |m0 (ξ + 1/2)|2 + |m0 (ξ )|2 = 1, Cohen’s condition provides a necessary and
sufficient condition for a function φ, defined via the Fourier transform by the infi-
∞
ξ
nite product φ̂(ξ ) = m0 k , to be the scaling function of some multiresolution
2
k=1
8.1 Introduction 285
analysis on R and in particular, for the sequence of translates of φ, {φ(x − n)}n∈Z ,

that generates the associated subspace V0 , to be orthonormal. Gabardo and Nashed
derive the analogue of Cohen’s condition for the generalized setting of NUMRA.
8.2 Nonuniform Multiresolution Analysis on Positive Half

Line
Definition 8.3 For an integer N > 1 and an odd integer r with 1 ≤ r ≤ 2N − 1 such
that r and N are relatively prime, an associated nonuniform multiresolution analysis
on positive half line is a sequence of closed subspaces Vj ⊂ L2 (R+ ), j ∈ Z such that
the following properties hold:
(i) Vj ⊂ Vj+1 ∀j ∈ Z,
(ii) ∪j∈Z is dense in L2 (R+ ),
(iii) ∩j∈Z Vj = {0},
(iv) f (.) ∈ Vj ⇔ f (N .) ∈ Vj+1 ∀j ∈ Z,
(v) There exists a function ϕ ∈ V0 such that {ϕ(x λ), λ ∈ Λ+ } where Λ+ =
{0, r/N } + Z+ , is a complete orthonormal system for V0 .
The function ϕ is called a scaling function in L2 (R+ ).
When N > 1, the dilation factor of N ensures that
N Λ+ ⊂ Z+ ⊂ Λ+ .
By the conditions (iv) and (v) of the Definition 8.3
ϕ1,λ (x) = N 1/2 ϕ(Nx λ), λ ∈ Λ+
constitute an orthonormal basis in V0 .

Since V0 ⊂ V1 , the function ϕ ∈ V1 and has the Fourier expansion

ϕ(x) = hλ (N )1/2 ϕ(Nx λ),
λ∈Λ+

where hλ = ϕ(x)ϕ1,λ (x) dx.
R+
This implies that

ϕ(x) = N aλ ϕ(Nx λ), |aλ |2 < ∞
λ∈Λ+ λ∈Λ+
where aλ = (N )−1/2 hλ .
Lemma 8.1 Consider a nonuniform multiresolution analysis on positive half line as

in Definition 8.3. Let ψ0 = ϕ and suppose ∃ N − 1 functions ψk , k = 1, 2, . . . , N −
1 in V1 such that the family of functions {ψk (x λ)}λ∈Λ+ ,k=0,1,...,N −1 forms an
orthonormal system in V1 . Then the system is complete in V1 .
ψk ∈k V 1 , k = 0, 1, . . . , N − 1 there exists the sequences {hλ }λ∈Λ+ sat-

k
Proof Since
isfying λ∈Λ+ |hλ | < ∞ such that
2

ψk (x) = hkλ N 1/2 ϕ(Nx λ).
λ∈Λ+
This implies that

ψk (x) = N aλk ϕ(Nx λ), |aλk |2 < ∞,
λ∈Λ+ λ∈Λ+
where aλk = (N )−1/2 hkλ .

On taking the Walsh–Fourier transform, we have

ξ ξ
ψ̃k (ξ ) = mk ϕ̃ , (8.1)
N N

where mk (ξ ) = λ∈λ+ aλk χ (λ, ξ ).
Since Λ+ = {0, r/N } + Z+ , we can write

r
mk (ξ ) = m1k (ξ ) + χ , ξ m2k (ξ ), k = 0, 1, . . . , N − 1 (8.2)
N
where m1k and m2k are locally L2 functions.

According to for λ ∈ Λ+ , where Λ+ = {0, r/N } + Z+ and by assumption we
have

ψk (x)ψl (x λ) dx = ψ̃k (ξ )ψ̃l (ξ )χ (λ, ξ ) d ξ
R+ R+
= δkl δ0λ ,
where δkl denotes the Kronecker delta.

Define

hkl (ξ ) = ψ̃k (ξ + Nj)ψ̃l (ξ + Nj), 0 ≤ k, l ≤ N − 1.
j∈Z+
If λ ∈ Z+ , we have
8.2 Nonuniform Multiresolution Analysis on Positive Half Line 287

ψk (x)ψl (x λ) dx = hkl (ξ )χ (λ, ξ ) d ξ
R+ [0,N ]
⎡ ⎤

N −1
= ⎣ hkl (ξ + p)⎦ χ (λ, ξ ) d ξ
[0,1] p=0
On taking λ = r
N
+ n where n ∈ Z+ , we obtain

r
ψk (x)ψl (x λ) dx = χ n + , ξ hkl (ξ ) d ξ
R+ [0,N ] N

= χ (n, ξ ) χ (r/N , ξ )hkl (ξ ) d ξ
[0,N ]
⎡ ⎤
−1
N
= χ (n, ξ ) χ (r/N , ξ ) ⎣ χ (r/N , p)hkl (ξ + p)⎦ d ξ.
[0,N ] p=0
By the orthonormality of the system {ψk (x λ)}λ∈Λ+ ,k=0,1,...,N −1 , we conclude that

N −1
hkl (ξ + p) = δkl (8.3)
p=0

N −1
and χ (r/N , p)hkl (ξ + p) = 0
p=0

N −1
i.e., α p hkl (ξ + p) = 0 (8.4)
p=0
p
where α = χ (r/N , 1), since χ (r/N , p) = χ (r/N , 1) for p = 0, 1, . . . , N − 1.
Now, we will express the conditions (8.3) and (8.4) in terms of mk as follows:

hkl (N ξ ) = ψ̃k (N ξ + Nj)ψ̃l (N ξ + Nj)
j∈Z+

= ψ̃k [N (ξ + j)]ψ̃l [N (ξ + j)]
j∈Z+

= mk (ξ + j)ml (ξ + j)|ϕ̃(ξ + j)|2
j∈Z+

= [m1k (ξ )m1l (ξ ) + m2k (ξ )m2l (ξ )] |ϕ̃(ξ + j)|2
j∈Z+
⎡ ⎤

+ ⎣m1k (ξ )m2l (ξ ) χ (r/N , ξ + j)|ϕ̃(ξ + j)|2 ⎦
j∈Z+
⎡ ⎤

+ ⎣m2k (ξ )m1l (ξ ) χ (r/N , ξ + j)|ϕ̃(ξ + j)|2 ⎦ .
j∈Z+
Therefore,

N −1
hkl (N ξ ) = [m1k (ξ )m1l (ξ ) + m2k (ξ )m2l (ξ )] h00 (ξ + j)
j=0
⎡ ⎤

N −1
+ ⎣m1k (ξ )m2l (ξ )χ (r/N , ξ ) χ (r/N , j)h00 (ξ + j)⎦
j=0
⎡ ⎤

N −1
+ ⎣m2k (ξ )m1l (ξ )χ (r/N , ξ ) χ (r/N , j)h00 (ξ + j)⎦
j=0
= m1k (ξ )m1l (ξ ) + m2k (ξ )m2l (ξ )
By using the last identity and the Eqs. (8.3) and (8.4), we obtain
−1

N
ξ +p 1 ξ +p 2 ξ +p 2 ξ +p
m1k ml + mk ml = δkl , (8.5)
p=0
N N N N
and
−1

N
ξ +p ξ + p ξ + p ξ + p
α p
m1k m1l + mk
2
m2l = 0, (8.6)
p=0
N N N N
0 ≤ k, l ≤ N − 1.
Both of these conditions together are equivalent to the orthonormality of the
system {ψk (x λ)}λ∈Λ+ ,k=0,1,...,N −1 .
The completeness of the system {ψk (x λ)}λ∈Λ+ ,k=0,1,...,N −1 in V1 is equivalent
1 x
to the completeness of the system ψk ( λ)}λ∈Λ+ ,k=0,1,...,N −1 in V0 . For a
N N
given
arbitrary function f ∈ V0 , by assumption ∃ a unique function m(ξ ) of the form
bλ χ (λ, ξ ), where λ∈Λ+ |bλ |2 < ∞ such that f˜ (ξ ) = m(ξ )ϕ̃(ξ ).
λ∈Λ+
Hence, in order to prove the claim, it is enough to show that the system of functions
S = {χ (N λ, ξ )mk (ξ )}λ∈Λ+ ,k=0,1,...,N −1

is complete in L2 [0, 1].

Let g ∈ L2 [0, 1], therefore, ∃ locally L2 functions g1 and g2 such that
g(ξ ) = g1 (ξ ) + χ (r/N , ξ )g2 (ξ ).
Assuming that g is orthogonal to all functions in S, we then have for any λ ∈ Λ+

and k ∈ {0, 1, . . . , N − 1} that

0= χ (ξ, N λ)mk (ξ )g(ξ ) d ξ
[0,1]

= χ (ξ, N λ)[m1k (ξ )g1 (ξ ) + m2k (ξ )g2 (ξ )] d ξ (8.7)
[0,1]
Taking λ = m where m ∈ Z+ and defining
wk (ξ ) = m1k (ξ )g1 (ξ ) + m2k (ξ )g2 (ξ ), k = 0, 1, . . . , N − 1
we obtain

0= χ (ξ, Nm)wk (ξ ) d ξ
[0,1]

N −1
= χ (ξ, Nm) wk (ξ + j/N ) d ξ.
[0, N1 ] j=0
Since this equality holds for all m ∈ Z+ , therefore

N −1
wk (ξ + j/N ) = 0 for a.e. ξ. (8.8)
j=0
Similarly, on taking λ = m + r
N
where m ∈ Z+ , we obtain

0= χ (ξ, Nm) χ (ξ, r)wk (ξ ) d ξ
[0,1]

N −1
= χ (ξ, Nm) χ (ξ, r) α j wk (ξ + j/N ) d ξ.
[0, N1 ] j=0
Hence, we deduce that

N −1
α j wk (ξ + j/N ) = 0 for a.e. ξ
j=0
which proves our claim.

If ψ0 , ψ1 , . . . , ψN −1 ∈ V1 are as in Lemma 8.1, one can obtain from them an

orthonormal basis for L2 (R+ ) by following the standard procedure for construction
of wavelets from a given MRA. It can be easily checked that for every m ∈ Z, the
collection {N m/2 ψk (N m x λ)}λ∈Λ+ ,k=0,1,...,N −1 is a complete orthonormal system
for Vm+1 .
Given a NUMRA on positive half line, we denote by Wm the orthogonal com-
plement of Vm in Vm+1 , m ∈ Z. It is clear from (i), (ii), and (iii) of Definition 8.3
that
L2 (R+ ) = ⊕m∈Z Wm
where ⊕ denotes the orthogonal direct sum with the inner product of L2 (R+ ).
From this, it follows immediately that the collection {N m/2 ψk (N m x λ)}
λ∈Λ+ ,m∈Z,k=1,2,...,N −1 forms a complete orthonormal system for L (R+ ).
2
Definition 8.4 A collection {ψk }k=1,2,...,N −1 of functions in V1 will be called a set

of wavelets associated with a given nonuniform multiresolution analysis on positive
half line if the family of functions {ψk (x λ)}k=1,2,...,N −1,λ∈Λ+ is an orthonormal
system for W0 .
The following theorem proves the necessary and sufficient condition for the existence
of associated set of wavelets to nonuniform multiresolution analysis on positive half
line.
Theorem 8.2 Consider a nonuniform multiresolution analysis on a positive half

line with associated parameters N and r, as in Definition 8.3, such that the corre-
sponding space V0 has an orthonormal system of the form {ϕ(x λ)}λ∈Λ+ where
Λ+ = {0, r/N } + Z+ , ϕ̃ satisfies
ϕ̃(ξ ) = m0 (ξ/N )ϕ̃(ξ/N ), (8.9)
and m0 has the form
m0 (ξ ) = m10 (ξ ) + χ (r/N , ξ )m20 (ξ ), (8.10)
for some locally L2 functions m10 and m20 . M0 is defined as
M0 (ξ ) = |m10 (ξ )|2 + |m20 (ξ )|2 . (8.11)
Then a necessary and sufficient condition for the existence of associated wavelets
ψ1 , ψ2 , . . . , ψN −1 is that M0 satisfies the identity
M0 (ξ + 1) = M0 (ξ ). (8.12)
Proof The orthonormality of the collection of functions {ϕ(x λ)}λ∈Λ+ which sat-
isfies (8.9) implies the following identities as shown in the proof of Lemma 8.1
−1

N
1
m (ξ + p/N )2 + m2 (ξ + p/N )2 = 1
0 0
p=0
and

N −1 2 2
α p m10 (ξ + p/N ) + m20 (ξ + p/N ) = 0,
p=0
where α = χ (r/N , 1). Similarly, if {ψk }k=1,2,...,N −1 is a set of wavelets associated

with the given nonuniform multiresolution analysis then it satisfies the relation (8.1)
and the orthonormality of the collection {ψk }k=0,1,...,N −1 in V1 is equivalent to the
identities
−1

N
1 ξ +p ξ + p ξ + p ξ + p
mk m1l + mk 2
m2l = δkl , (8.13)
p=0
N N N N
and
−1

N
ξ +p 1 ξ +p 2 ξ +p 2 ξ +p
α p
m1k ml + mk ml = 0, (8.14)
p=0
N N N N
0 ≤ k, l ≤ N − 1.
If ξ ∈ [0, 1/N ] is fixed and ak (j) = m1k (ξ + j/N ), bk (j) = m2k (ξ + j/N ) are vec-
tors in C N for j = 0, 1, . . . N − 1 where 0 ≤ k ≤ N − 1, the solvability of system of
Eqs. (8.13) and (8.14) is equivalent to
1
M0 (ξ + (j + N )/N ) = M0 (ξ + j/N ), ξ ∈ [0, ], j = 0, 1, . . . , N − 1,
N
which is equivalent to (8.12).
The main purpose in the next section is to construct nonuniform multiresolution anal-
ysis on positive half line starting from a Walsh polynomial m0 satisfying appropriate
conditions and finding suitable analogue of Cohen’s conditions.
8.2.1 Construction of Nonuniform Multiresolution Analysis

on Positive Half Line
Our goal in this section is to construct nonuniform multiresolution analysis on a

positive half line starting from a polynomial m0 of the form
m0 (ξ ) = m10 (ξ ) + χ (r/N , ξ )m20 (ξ ) (8.15)

where N > 1 is an integer and r is an odd integer with 1 ≤ r ≤ 2N − 1 such that r

and N are relatively prime and m10 (ξ ) and m20 (ξ ) are locally L2 Walsh polynomials.
The scaling function ϕ associated with given nonuniform multiresolution analysis
on positive half line should satisfy the scaling relation
ϕ̃(ξ ) = m0 (ξ/N )ϕ̃(ξ/N ). (8.16)
Define
M0 (ξ ) = |m10 (ξ )|2 + |m20 (ξ )|2
and suppose

N −1
M0 (ξ + p/N ) = 1 (8.17)
p=0

N −1
and α p M0 (ξ + p/N ) = 0 (8.18)
p=0
where α = χ (r/N , 1). It follows from (8.16) that

∞

ξ
ϕ̃(ξ ) = m0 . (8.19)
Nk
k=1
∞

ξ
Also assume that m0 (0) = 1 in order for the infinite product to converge
m0
Nk
k=1
pointwise. For an arbitrary function m0 of the form (8.15) the conditions (8.17) and
(8.18) imply that |m0 | ≤ 1 a.e. Since if |m0 (ξ )| > 1 for a fixed ξ , then |m10 (ξ )| +
|m20 (ξ )| > 1 and thus |M0 (ξ )| > 21 .
We obtain the inequalities

N −1
1
M0 (ξ + p/N ) <
p=1
2
and
N
−1 1
M0 (ξ + p/N ) = |M0 (ξ )| >

p=1 2
which yields a contradiction.
Theorem 8.3 Let m0 be a polynomial of the form (8.15), where m10 and m20 are
locally square integrable functions and M0 satisfy (8.17) and (8.18). Let ϕ be defined
by (8.19) and assume that the infinite product defining ϕ̃ converges a.e. on R+ . Then
the function ϕ ∈ L2 (R+ ).
Proof Consider the integrals

N
A1 = M0 (ξ/N ) d ξ
0
and
k−1
2
Nk
ξ
Ak = M0 m0 ξ d ξ for k ≥ 2
Nk N
j
0 j=1

1N −1
N
ξ p
A1 = M0 (ξ/N ) d ξ = M0 + dξ = 1
0 0 p=0
N N
and
k−1
2
Nk
ξ
Ak = M0 m0 ξ d ξ
Nk N
j
0 j=1
−1 2 k−2 2
N k−1 N
m0 ξ ξ p ξ
= + p M0 + m0 d ξ.
0 p=0
N k−1 N k N j=1 Nj
We find that
−1
N 2 2 2 N
−1
ξ ξ p ξ ξ ξ p
m0 + p M0 + = m10 + m2 M0 +
N k N k N N k−1 0
N k−1 N k N
p=0 p=0
N
−1
ξ ξ r ξ ξ p
+m10 m20 χ , α p M0 +
N k−1 N k−1 N N k−1
p=0
Nk N
N
−1
ξ ξ r ξ ξ p
+m10 m20 χ , α −p M0 +
N k−1 N k−1 N N k−1
p=0
Nk N
2 2
ξ ξ
= m10 + m2

N k−1 0
N k−1

ξ
= M0 .
N k−1
This shows that

k−2
2
N k−1
ξ
Ak = M0 m0 ξ d ξ = Ak−1 .
N k−1 N
j
0 j=1
It follows that, for all k

Ak = Ak−1 = Ak−2 = · · · = A1 = 1.
Hence
k−1
Nk
Nk ξ 2
|ϕ̃(ξ )| d ξ ≤
2 m0 dξ
Nj
0 0 j=1
k−1
2
k
ξ
m0 ξ d ξ
N
≤ 2M0
0 Nk j=1
N
j
= 2Ak = 2.
Since k is arbitrary, it follows that ϕ ∈ L2 (R+ ).
We will construct nonuniform multiresolution analysis on a positive half line from

a polynomial m0 of the form (8.15), which satisfies (8.17) and (8.18) and also the
condition m0 (0) = 1.
By Theorem 8.3, we obtain a compactly supported function ϕ ∈ L2 (R+ ) which
satisfies
ϕ̃(N ξ ) = m0 (ξ )ϕ̃(ξ ). (8.20)
Now, it is necessary to determine the orthonormality of the system of functions

{ϕ(x λ)}λ∈Λ+ in L2 (R+ ) where Λ+ = {0, r/N } + Z+ .
If the orthonormality condition is satisfied, we can define
V0 = span{ϕ(x λ)}λ∈Λ+
and Vj for j ∈ Z is defined as

x
f (x) ∈ Vj ⇔ f ∈ V0 (8.21)
Nj
so that (iv) and (v) of Definition 8.3 hold.
Also Eq. (8.20) implies that (i) also holds. The remaining two conditions (ii) and
(iii) follow from the results of the Theorems 8.4 and 8.5 which are analogues of
results in standard theory.
For an integer m, let εm (R+ ) denotes the collection of all functions f on R+ which
are constant on [sN −m , (s + 1)N −m ) for each s ∈ Z+ . Further we set,
ε˜m (R+ ) = {f : f is W -continuous and f˜ ∈ εm (R+ )}
and
ε(R+ ) = ∪∞ ∞
m=1 εm (R+ ), ε̃(R+ ) = ∪m=1 ε˜m (R+ ).
The following properties are true:

1. εm (R+ ) and ε˜m (R+ ) are dense in Lq (R+ ) for 1 ≤ q ≤ ∞.
2. If f ∈ L1 (R+ ) ∪ εm (R+ ), then supp f˜ ⊂ [0, N m ].
3. If f ∈ L1 (R+ ) ∪ ε˜m (R+ ), then supp f ⊂ [0, N m ].
For ϕ ∈ L2 (R+ ), we put
ϕj,λ (x) = N j/2 ϕ(N j x λ), j ∈ Z, λ ∈ Λ+ .
Let Pj be the orthogonal projection of L2 (R+ ) to Vj .
Theorem 8.4 Let Λ+ = {0, r/N } + Z+ . Suppose that ϕ ∈ L2 (R+ ) is such that the
collection {ϕ(x λ)}λ∈Λ+ is an orthonormal system in L2 (R+ ) with closed linear
span V0 and Vj is defined by (8.21) then ∩j∈Z Vj = {0}.
Proof Let f ∈ ∩j∈Z Vj . Given an ε > 0 and a continuous function u which is com-
pactly supported in some interval [0, R], R > 0 and satisfies f − u2 < ε. Then we
have
f − Pj u2 ≤ Pj (f − u)2 ≤ f − u2 < ε
so that
f 2 < Pj u2 + ε.
Using the fact that the collection {N j/2 ϕ(N j x λ)}λ∈Λ+ is an orthonormal bases for
Vj .

Pj u22 = < Pj u, ϕj,λ > 2
λ∈Λ+
R
2

= (N ) j u(x)ϕ(N j x λ) dx

λ∈Λ+ 0
R
≤ (N )j u2∞ R |ϕ(N j x λ)|2 dx
λ∈Λ+ 0
where u∞ denotes the supremum norm of u. If j is chosen small enough so that
RN j ≤ 1, then

Pj u22 ≤ u2∞ |ϕ(x)|2 dx
SR,j

= u2∞ ISr,j (x)|ϕ(x)|2 dx, (8.22)
R+
where SR,j = ∪λ∈Λ+ {y λ/y ∈ [0, RN j ]} and ISr,j denotes the characteristic function
of Sr,j .
It can be easily checked that
lim ISR,j (x) = 0 ∀x ∈

/ Λ+ .
j→−∞
Thus, from Eq. (8.22) by using the dominated convergence theorem, we get
lim Pj u2 = 0.

j→−∞
Therefore, we conclude that f 2 < ε and since ε is arbitrary f = 0 and thus

∩j∈Z Vj = {0}.
Theorem 8.5 Let ϕ ∈ L2 (R+ ) is such that the collection {ϕ(x λ)}λ∈Λ+ is an
orthonormal system in L2 (R+ ) with closed linear span V0 and Vj is defined by
(8.21) and assume that ϕ̃(ξ ) is bounded for all ξ and continuous near ξ = 0 with
|ϕ̃(0)| = 1, then ∪j∈Z Vj = L2 (R+ ).
⊥
Proof Let f ∈ ∪j∈Z Vj .
Given an ε > 0 we choose u ∈ L1 (R+ ) ∩ ε(R+ ) such that f − u2 < ε.
Then for any j ∈ Z+ , we have

Pj f 22 = Pj f , Pj f = f , Pj f = 0
and so
Pj u2 = Pj (f − u)2 ≤ f − u2 < ε.
Then we put g(ξ ) = ũ(ξ )ϕ̃(N −j ξ ) for some function g of the form

r
g(ξ ) = g1 (ξ ) + χ ξ, g2 (ξ )
N
where g1 and g2 are locally square integrable,
1/2-periodic functions.
If g(ξ ) has the expansion of the form λ∈Λ+ cλ χ (ξ, λ) on the set [0, 1], then
1
cλ = g(ξ )χ (ξ, λ) d ξ
0
= ũ(ξ )ϕ̃(N −j ξ )χ (ξ, λ) d ξ, λ ∈ Λ+ .
R+
If λ ∈ Z+ , we have

1 1
ξ k
2g1 (ξ )χ (ξ, λ) d ξ = ũ(ξ + k)ϕ̃ + χ (ξ, λ) d ξ
0 0 k∈Z Nj Nj
+
Therefore
1 ξ k
g1 (ξ ) = ũ(ξ + k)ϕ̃ + j .
2 Nj N
k∈Z+
On taking λ = r
N
+ m where m ∈ Z+ , we obtain

1 ξ k r
g2 (ξ ) = ũ(ξ + k)ϕ̃ + χ ξ + k, .
2 Nj Nj N
k∈Z+
Therefore
1 ξ k
g(ξ ) = ũ(ξ + k)ϕ̃ + j (1 + α k )
2 Nj N
k∈Z+
where α = χ (r/N , 1).

Since the collection {N j/2 ϕ(N j x λ)}λ∈Λ+ is an orthonormal basis for Vj , if in
addition ũ has compact support, for large values of j

Pj u22 = u, ϕj,λ 2 = |u(ξ )|2 |ϕ̃(N −j ξ )|2 d ξ.
λ∈Λ+ R+
By Lebesgue-dominated convergence theorem as j → ∞, the expression on R.H.S

converges to |ϕ̃(0)|2 ũ22 . Therefore,
ε > Pj u2 = ũ2 = u.
Consequently
f 2 < ε + u2 < 2ε.
Since ε is arbitrary, therefore f = 0.
8.2.2 The Analogue of Cohen’s Condition
On the basis of construction in the previous section, it is necessary to determine

the orthonormality of the system of functions {ϕ(x λ)}λ∈Λ+ . We are going to con-
sider the analogue of Cohen’s condition for nonuniform multiresolution analysis on
positive half line.
Theorem 8.6 Let m0 be a polynomial of the form (8.15), which satisfies m0 (0) = 1
together with the conditions (8.17) and (8.18). Let ϕ be defined by the formula (8.19)
and Λ+ = {0, r/N } + Z+ . Then the following are equivalent:
1. ∃ a W -compact set E such that 0 ∈ int(E), μ(E) = 1, E ≡ [0, 1](mod Z+ ) and
infj∈N infw∈E |m0 (N −j w)| > 0 (8.23)
2. The system {ϕ(x λ)}λ∈Λ+ is orthonormal in L2 (R+ ).
Proof We will start by proving (1) ⇒ (2)

For k ∈ N, define

k
ξ ξ
μk (ξ ) = m0 IE , ξ ∈ .R+
j=1
Nj Nj
Since 0 ∈ int (E)

μk → ϕ̃ pointwise as k → ∞. (8.24)
By Eq. (8.19), there is a constant c1 > 0 such that

m0 ξ ≥ c1 > 0 for k ∈ N, ξ ∈ E.
N
k
By (a) and (c), ∃ an integer k0 such that

ξ
m0 = 1 for k > k0 , ξ ∈ E.
Nk
Thus

k0
ξ
ϕ̃(ξ ) = m0 , ξ ∈E
Nk
k=1
and so
c1−k0 |ϕ(ξ )| ≥ IE (ξ ), ξ ∈ R+ . (8.25)
Therefore

k
ξ ξ
|μk (ξ )| = m0 IE
N j N j
j=1

k

≤ c1−k0 m0 ξ ϕ̃ ξ
N j N j
j=1
which yields
|μk (ξ )| ≤ c1−k0 |ϕ̃(ξ )|, k ∈ N, ξ ∈ R+ .
This implies
|μk (ξ )| ≤ c|ϕ̃(ξ )|, k ∈ N, ξ ∈ R+ .
Therefore, by Lebesgue DCT the sequence {μk } converges to ϕ̃ in L2 (R+ ) norm.

We will compute the integral

|μk (ξ )|2 χ (λ σ, ξ ) d ξ where λ, σ ∈ Λ+ .
R+
If k = 1, then
2

|μ1 (ξ )| χ (λ σ, ξ ) d ξ =
2 m0 ξ IE ξ χ (λ σ, ξ ) d ξ
N N
R+ R+

=N |m0 (ξ )|2 χ (λ σ, N ξ ) d ξ.
E
Using the assumption E ≡ [0, 1](mod Λ+ ), we get

|μ1 (ξ )|2 χ(λ σ, ξ ) d ξ = N |m0 (ξ )|2 χ(λ σ, N ξ ) d ξ
R+ [0,1]

=N M0 (ξ )χ(λ σ, N ξ ) d ξ
[0,1]
⎡ ⎤

p

N −1
=N ⎣ M0 ξ+ χ(λ σ, p)⎦ χ(λ σ, N ξ ) d ξ
[0,1/N ] N
p=0

=N χ(λ σ, N ξ ) d ξ
[0,1/N ]

= χ(λ σ, ξ ) d ξ = δλσ .
[0,1]
For k > 1
⎡ ⎤
k
2
ξ
|μk (ξ )|2 χ (λ σ, ξ ) d ξ = ⎣ m0 ⎦ IE ξ χ (λ σ, ξ ) d ξ.
Nj Nk
R+ R+ j=1
Setting Ek = {ξ ∈ R+ /N −k ξ ∈ E} and w = N −k ξ , we have
k
2

|μk (ξ )|2 χ (λ σ, ξ ) d ξ = m0 ξ χ (λ σ, ξ ) d ξ
N
j
R+ Ek j=1

k−1
=N k
|m0 (w)| 2
|m0 (N j w)|2 χ (λ σ, N k w) d w.
E j=1
Using the assumption E ≡ [0, 1](modZ+ ) and changing the variable, we get

1 k−1
|μk (ξ )|2 χ(λ σ, ξ ) d ξ = N k |m0 (N j ξ )|2 χ (λ σ, N k ξ ) d ξ
R+ 0 j=0
⎡ ⎤
1
k−1
=N k ⎣ |m0 (N ξ )| j 2⎦
M0 (ξ )χ(λ σ, N k ξ ) d ξ
0 j=1
⎡ ⎤⎡ ⎤

p ⎦

k−1
N −1
=N k ⎣ |m0 (N ξ )|
j 2⎦ ⎣
|m0 (N ξ + p)| M0 ξ +
2
[0,1/N ] N
j=2 p=0
χ(λ σ, N k ξ ) d ξ
where

N −1
N −1
|m0 (N ξ + p)|2 M0 (ξ + p) = |m10 (N ξ + p) + χ(r/N , N ξ + p)m20 (N ξ + p)|2 M0 (ξ + p/N )
p=0 p=0
−1
N
= |m10 (N ξ )|2 + |m20 (N ξ )|2 M0 (ξ + p/N )
p=0
⎡ ⎤

N −1
+ ⎣m10 (N ξ )m20 (N ξ )χ(r/N , N ξ ) χ(r/N , p)M0 (ξ + p/N )⎦
p=0
⎡ ⎤

N −1
+ ⎣m20 (N ξ )m10 (N ξ )χ(r/N , N ξ ) χ(r/N , p)M0 (ξ + p/N )⎦
p=0
= |m10 (N ξ )|2 + |m20 (N ξ )|2

= M0 (N ξ ).
Therefore
⎡ ⎤

k−1
|μk (ξ )| χ(λ σ, ξ ) d ξ = N
2 k ⎣ |m0 (N ξ )| j 2⎦
M0 (N ξ )χ(λ σ, N k ξ ) d ξ
R+ [0,1/N ] j=2
⎡ ⎤

k−1
= N k−1 ⎣ |m0 (N j−1 ξ )|2 ⎦ M0 (ξ )χ(λ σ, N k−1 ξ ) d ξ
[0,1] j=2

k−2
= N k−1 |m0 (N j ξ )|2 M0 (ξ )χ(λ σ, N k−1 ξ ) d ξ
[0,1] j=1

= |μk−1 (ξ )|2 χ(λ σ, ξ ) d ξ.
R+
Therefore, for any k ∈ N

|μk (ξ )|2 χ (λ σ, ξ ) d ξ = δλσ , λ, σ ∈ Λ+ .
R+
In particular, for k ∈ N
|μk (ξ )|2 d ξ = 1.
R+
Since μk → ϕ̃ pointwise as k → ∞ by using Fatou’s Lemma, we obtain

|ϕ̃(ξ )|2 d ξ ≤ 1.
R+
From Eqs. (8.24) and (8.25), by Lebesgue’s DCT it follows that

|ϕ̃(ξ )|2 χ (λ σ, ξ ) d ξ = lim |μk (ξ )|2 χ (λ σ, ξ ) d ξ.
R+ k→∞ R+
Therefore,
ϕ(x λ)ϕ(x σ ) = δλσ , λ, σ ∈ Λ+ .
R+
We will now prove the converse (2) ⇒ (1).

Since the system {ϕ(x λ)}λ∈Λ+ is orthonormal in L2 (R+ ), therefore we have the
identity

ξ
|ϕ̃(ξ + Nj)|2 = M0 (8.26)
j∈Z
N
+
which can be deduced as in proof of Theorem 1.2.1.

From Eq. (8.17), we obtain

N −1
p
M0 ξ + =1
p=0
N
and thus we can find, for each ξ ∈ [0, 1] an integer p(ξ ) ∈ {0, 1, 2, . . . , N − 1} such
that
ξ p(ξ ) 1
M0 + ≥ .
N N N
Therefore, by using (8.26) there exists l(ξ ) ∈ N such that

1
|ϕ̃(ξ + p(ξ ) + Nj)|2 > .
2N
j≤l(ξ )

Since ϕ̃ is continuous, the finite sum |ϕ̃(. + p(ξ ) + Nj)|2 is continuous. There-
j≤l(ξ )
fore there exists, for every ξ in [0, 1], a neighborhood {ζ : |ζ − ξ | ≤ Rξ } so that, for
all ζ in this neighborhood
1
|ϕ̃(ζ + p(ξ ) + Nj)|2 > .
2N
j≤l(ξ )
Since [0, 1] is compact, we can find a finite subset of the collection of intervals
{ζ : |ζ − ξ | ≤ Rξ } which covers [0, 1]. Let l0 be the maximum of the corresponding
numbers l(ξ ) associated with this finite covering. Then for all ζ ∈ [0, 1]
1
|ϕ̃(ζ + p(ξ ) + Nj)|2 > .
2N
j≤l0
It follows that for every ξ ∈ [0, 1] ∃p(ξ ) between 0 and l0 such that
|ϕ̃(ξ + p(ξ ) + Nk(ξ ))|2 > [2N (2l0 + 1)]−1 = C.
By continuity of ϕ̃, there exists for every ξ in [0, 1], a neighborhood {ζ : |ζ − ξ | ≤

r(ξ )} so that, for all ζ in this neighborhood
|ϕ̃(ζ + p(ξ ) + Nk(ξ ))|2 > C.
Let U (ξ ) ≡ {ζ ∈ [0, 1], |ζ − ξ | ≤ r(ξ )}.

Since ϕ̃(0) = 1, we can take p(0) = k(0) = 0.
Since [0, 1] is compact, we can find a finite collection {U (ξi )}M
i=0 which covers
[0, 1] with ξ0 = 0.
Define
S0 = U (ξ0 )
and
Sl = U (ξl )\ ∪l−1
j=0 U (ξj ) if 1 ≤ l ≤ M .
The Sl , 0 ≤ l ≤ M form a partition of [0, 1].

We define
E = ∪M l=0 (Sl + p(ξl ) + Nk(ξl )),
then E is clearly compact E ≡ [0, 1](mod Z+ ).

By construction, |ϕ̃(ξ )| ≥ C on E and E contains a neighborhood of 0.
Next we will show that E satisfies (8.23).
ξ
From the remark in, we need only to check that infξ ∈E |m0 j | > 0 for a finite
N
number j, 1 ≤ j ≤ j0 .
For ξ ∈ E, we have
⎛ ⎞
j0
ξ ξ
|ϕ̃(ξ )| = ⎝ m0 ⎠ ϕ̃ (8.27)
j=1
Nj N j0
is bounded below away from zero.

Since |ϕ̃| is also bounded, the first factor on the right-hand side of above equation
has, therefore, no zeros on the compact set E. Since finite product of a continuous
function is also continuous, therefore
j0

m0 ξ ≥ C1 > 0 for ξ ∈ E.
Nj
j=1
Since |m0 | ≤ 1, we therefore have, for any j, 1 ≤ j ≤ j0

j0

m0 ξ ≥ m0 ξ ≥ C1 > 0.
Nj Nj
j =1
This proves that (8.23) is satisfied.
8.3 Nonuniform Wavelet Frames in L2 (R)
We begin with some notations. Given an integer N ≥ 1 and an odd integer r with
1 ≤ r ≤ 2N − 1, r and N are relatively prime, we consider the translation set Λ as

rk
Λ = {0, r/N } + 2Z = + 2n : n ∈ Z, k = 0, 1 . (8.28)
N

Definition 8.5 If f ∈ L1 (R) L2 (R), then the Fourier transform of f is the function
fˆ defined by
fˆ (ξ ) = f (x)e−2πiξ x dx. (8.29)
R
Let
ψ ∈ L2 (R), ψj,λ (t) = (2N )j/2 ψ((2N )j t − λ), j ∈ Z, λ ∈ Λ, (8.30)
On taking Fourier transform, we have

−j
ψ̂j,λ (ξ ) = (2N )−j/2 ψ̂((2N )−j ξ )e−2πi(2N ) .λξ
(8.31)
The function system {ψj,λ (t)}j∈Z,λ∈Λ is called a nonuniform wavelet frame for L2 (R)
if there exist constants A and B, 0 < A ≤ B < ∞ such that

Af 2 ≤ |f , ψj,λ |2 ≤ Bf 2 (8.32)
j∈Z λ∈Λ
for all f ∈ L2 (R). We call A, B the lower frame bound and the upper frame bound for
the frame, respectively. If A = B, then the frame is called a tight frame. If A = B = 1,
then the frame is called a normalized tight frame.
First, we prove a lemma which will be used in the proofs of the main results.
Lemma 8.2 If f , ψ ∈ L2 (R), then
2N −1 !
1
|f , ψj,λ |2 = fˆ (ξ + (2N )j p/2)ψ̂((2N )−j ξ + p/2) G 1 (ξ )d ξ
4N R
λ∈Λ p=0

2N −1 !
1
+ fˆ (ξ + (2N )j p/2)ψ̂((2N )−j ξ + p/2)e−π irp/N G 2 (ξ )d ξ
4N R
p=0
(8.33)
where
G 1 (ξ ) = fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2) (8.34)
l∈Z
and
fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)eπi N l .
r
G 2 (ξ ) = (8.35)
l∈Z
Proof By (8.31), we have for any f ∈ L2 (R)

1 −j
f , ψj,λ = fˆ (ξ )ψ̂((2N )−j ξ )e2π i(2N ) λξ d ξ
(2N )j/2 R
1 21 (2N )j (l+1) −j
= fˆ (ξ )ψ̂((2N )−j ξ )e2π i(2N ) λξ d ξ
(2N ) j/2 1
(2N ) jl
l∈Z
2
1 1
2 (2N )
j
−j
= fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)eπ iλl e2π i(2N ) λξ d ξ
(2N )j/2 0
l∈Z
(2N )j " #
1 2 −j
= fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)eπ iλl e2π i(2N ) λξ d ξ.
(2N )j/2 0 l∈Z
Now
|f , ψj,λ |2 = |f , ψj,λ |2 + |f , ψj,λ |2 (8.36)
λ∈Λ λ∈2Z λ∈(r/N +2Z)
√ !

is an orthonormal basis for L2 0, (2N2 ) , Parseval’s
−j j
Since (2N )2j/2 e2πi(2N ) 2mξ
m∈Z
formula gives
8.3 Nonuniform Wavelet Frames in L2 (R) 305

|f , ψj,λ |2
λ∈2Z
2
(2N )j √
1 2
ˆ −j
2 2π i(2N )−j 2mξ
= f (ξ + (2N ) j
l/2) ψ̂((2N ) ξ + l/2) e d ξ
(2N )j/2
m∈Z 0
2
l∈Z

(2N ) j
2
1 2
= fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2) d ξ
2 0
l∈Z
(2N )j
1 2
= |G 1 (ξ )|2 d ξ.
2 0
Using Parseval’s formula again, we obtain

|f , ψj,λ |2
λ∈(r/N +2Z)

(2N )j √ 2π i(2N )−j (r/N +2m)ξ 2
1 2 2e
= fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)eπ irN /l d ξ
2 0 (2N )j/2
m∈Z l∈Z

(2N )j
2
1 2
= fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)eπ irN /l d ξ
2 0
l∈Z
(2N )j
1 2
= |G 2 (ξ )|2 d ξ.
2 0
Therefore (8.36) becomes

(2N )j (2N )j

1 2 2
|f , ψj,λ | = 2
|G 1 (ξ )| d ξ +
2
|G 2 (ξ )| d ξ
2
(8.37)
λ∈Λ
2 0 0
1
= [I + II ] (8.38)
2
$ (2N )j $ (2N )j
where I = 0 2 |G 1 (ξ )|2 d ξ and II = 0 2 |G 2 (ξ )|2 d ξ .
For 1 ≤ p ≤ 2N − 1, we have

1 ˆ 1 −j
1
G 1 ξ + (2N ) p =
j
f ξ + (2N ) (l + p) ψ̂ (2N ) ξ + (l + p)
j
2 2 2
l∈Z
= G 1 (ξ ) (8.39)
and

1 1 1
G 2 ξ + (2N )j p = fˆ ξ + (2N )j (l + p) ψ̂ (2N )−j ξ + (l + p) eπ irl/N
2 2 2
l∈Z
= e−π irp/N G 2 (ξ ). (8.40)
Since
(2N )j (2N )j+1 (2N )j+1
2 2 2
I= |G 1 (ξ )| d ξ =
2
|G 1 (ξ )| d ξ −
2
|G 1 (ξ )|2 d ξ.
(2N )j
0 0 2
Using (8.39), we get
(2N )j+1
2
|G 1 (ξ )|2 d ξ
(2N )j
2
(2N )j
2 (2N ) 2 (2N )
3 j 1 j+1
= |G 1 (ξ )|2 d ξ + |G 1 (ξ )|2 d ξ + · · · + |G 1 (ξ )|2 d ξ

2N −1
2 (2N ) (2N )j 2 (2N )
1 j j
= (2N − 1)I . (8.41)
Therefore
2 (2N )
1 j+1
1
I= |G 1 (ξ )|2 d ξ. (8.42)
2N 0
Proceeding as above, we obtain

2 (2N )
1 j+1
1
II = |G 2 (ξ )|2 d ξ. (8.43)
2N 0
The function G 1 (ξ ) can be expressed as
1

1

G 1 (ξ ) = fˆ ξ + (2N )j l ψ̂ (2N )−j ξ + l
2 2
l∈Z

1 1
= fˆ ξ + (2N )j 2Nm ψ̂ (2N )−j ξ + 2Nm
2 2
m∈Z

1 1
+ fˆ ξ + (2N )j (2Nm + 1) ψ̂ (2N )−j ξ + (2Nm + 1)
2 2
m∈Z
+
..
.
1

1

+ fˆ ξ + (2N )j (2Nm + 2N − 1) ψ̂ (2N )−j ξ + (2Nm + 2N − 1)
2 2
m∈Z
" #

2N −1
= ˆf ξ + 1 (2N )j+1 m + 1 (2N )j p ψ̂ (2N )−j ξ + 1 (2Nm + p) .
2 2 2
p=0 m∈Z
Therefore, we can also write G 2 (ξ ) as

⎧ ⎫

2N −1 ⎨ ⎬
1 1 1 π
G 2 (ξ ) = fˆ ξ + (2N ) m + (2N ) p ψ̂ (2N )−j ξ + (2Nm + p) e
j+1 j irp/N .
⎩ 2 2 2 ⎭
p=0 m∈Z
The functions G 1 (ξ ) and G 2 (ξ ) are 21 (2N )j+1 -periodic, which gives

"
2N −1 (2N )
j+1
1 2 1 1
I = fˆ ξ + (2N )j+1 m + (2N )j p
2N 0 2 2
p=0 m∈Z

1
× ψ̂ (2N )−j ξ + (2Nm + p) G 1 (ξ )d ξ
2
−1 " #
1 2 (2N ) (m+1)
2N 1 j+1
= ˆf ξ + 1 (2N )j p ψ̂ (2N )−j ξ + 1 p G 1 (ξ )d ξ

2N 1 2 2
p=0 m∈Z 2 (2N ) m
j+1
−1
" #
1
2N
1 1
= fˆ ξ + (2N )j p ψ̂ (2N )−j ξ + p G 1 (ξ )d ξ
2N R 2 2
p=0
(8.44)
and
2N −1
" #
1 ˆ 1 −j 1 −πirp/N
II = f ξ + (2N ) p ψ̂ (2N ) ξ + p e
j G 2 (ξ )d ξ
2N p=0 R 2 2
(8.45)
Combining all these results, we get

|f , ψj,λ |2
λ∈Λ
2N −1
" #
1 1 1
= fˆ ξ + (2N )j p ψ̂ (2N )−j ξ + p G 1 (ξ ) + e−πirp/N G 2 (ξ ) d ξ.
4N R 2 2
p=0
The proof is complete.
8.3.1 Necessary and Sufficient Condition
We prove a necessary condition for the system {ψj,λ (t)}j∈Z,λ∈Λ to be wavelet frame
for L2 (R).
Theorem 8.7 If {ψj,λ (t)}j∈Z,λ∈Λ is a wavelet frame for L2 (R) with bounds A and B,
then
A ≤ Sψ (ξ ) ≤ B a.e ξ ∈ R, (8.46)
where
Sψ (ξ ) = |ψ̂((2N )−j ξ )|2 . (8.47)
j∈Z
Proof By definition, we have

|f , ψj,λ |2 ≤ Bf 2
j∈Z λ∈Λ
for f ∈ L2 (R), so that for any M > 0, M ∈ Z, we write

M
|f , ψj,λ |2 ≤ Bf 2 .
j=−M λ∈Λ
Using Lemma 8.2, we get
M 2N −1 + !
1
fˆ (ξ + (2N )j p/2)ψ̂((2N )−j ξ + p/2)
4N j=−M p=0 R

× fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)d ξ
l∈Z
!
+ fˆ (ξ + (2N )j p/2)ψ̂((2N )−j ξ + p/2)e−πirp/N
R

× fˆ (ξ + (2N ) j
l/2)ψ̂((2N )−j ξ + l/2)e πirl/N
dξ ≤ Bfˆ 2 .
l∈Z
(8.48)
Let ξ0 ∈ R, given ε > 0, we consider

√1 , ξ ∈ (ξ0 − ε, ξ0 + ε)
fˆ (ξ ) = 2ε (8.49)
0, otherwise
For sufficiently small ε, j ≥ −M , the intervals (ξ0 − ε − (2N )j l/2, ξ0 + ε −

(2N )j l/2), l ∈ Z are mutually disjoint, we have
fˆ (ξ + (2N )j l/2)fˆ (ξ + (2N )j m/2) = 0, l = m; l, m ∈ Z. (8.50)

The inequality (8.48) becomes
M 2N −1
1 ξ0 +ε−(2N ) p/2 ˆ
j
|f (ξ + (2N )j p/2)|2 |ψ̂((2N )−j ξ + p/2)|2 d ξ ≤ Bfˆ 2 (8.51)

2N ξ0 −ε−(2N )j p/2
j=−M p=0
or
M 2N −1 ξ0 +ε−(2N )j p/2
1 1
|ψ̂((2N )−j ξ + p/2)|2 d ξ ≤ B. (8.52)
2N j=−M p=0 2ε ξ0 −ε−(2N )j p/2
By taking ε → 0 and M → ∞ consecutively, we have
2N −1
1
|ψ̂((2N )−j ξ )|2 d ξ ≤ B
2N j∈Z p=0

⇒ |ψ̂((2N )−j ξ )|2 d ξ ≤ B (8.53)
j∈Z
which is the right inequality of (8.46). To prove left inequality of (8.46) (2.19),
consider
|f , ψj,λ |2 = I1 + I2 , (8.54)
j∈Z λ∈Λ

where I1 = j>−M λ∈Λ |f , ψj,λ |2 and I2 = j≤−M λ∈Λ |f , ψj,λ |2 .
By the frame condition, I1 ≥ A − I2 . As we have shown that I1 → Sψ (ξ ), it
remains to prove that I2 → 0 as M → ∞.
Using (8.34), (8.35), and (8.38), we obtain

I2 = |f , ψj,λ |2
j≤M λ∈Λ
(2N ) 2
1
j

2

= fˆ (ξ + (2N ) l/2)ψ̂((2N ) ξ + l/2) d ξ
j −j
2 j≤−M 0
l∈Z
(2N ) 2
1
j

2
r
+ fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)eπi N l d ξ
2 j≤−M 0
l∈Z
(2N ) 2
1
j

2

≤ |fˆ (ξ + (2N ) l/2)ψ̂((2N ) ξ + l/2)| d ξ
j −j
2 j≤−M 0
l∈Z
(2N ) 2
1
j

2
rk
+ |fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)eπi N l | d ξ
2 j≤−M 0
l∈Z
2
(2N )j
2

= |fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)| d ξ
0
j≤−M l∈Z
" #
(2N )j
2
= ˆ −j
|f (ξ + (2N ) l/2)ψ̂((2N ) ξ + l/2)|
j
j≤−M 0 l∈Z
" #

× ˆ −j
|f (ξ + (2N ) m/2)ψ̂((2N ) ξ + m/2)| d ξ.
j
m∈Z

Since the function l∈Z |fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)| is 21 (2N )j -periodic, we
can continue with
(2N )j
2 (l+1)
I2 ≤ |fˆ (ξ )ψ̂((2N )−j ξ )| |fˆ (ξ + (2N )j m/2)ψ̂((2N )−j ξ + m/2)|d ξ
(2N )j
j≤−M l∈Z 2 l m∈Z

= |fˆ (ξ )ψ̂((2N )−j ξ )||fˆ (ξ + (2N )j m/2)ψ̂((2N )−j ξ + m/2)|d ξ.
j≤−M m∈Z R
By Cauchy–Schwarz inequality, we get
1/2 1/2
I2 ≤ |fˆ (ξ )ψ̂((2N )−j ξ )|2 d ξ |fˆ (ξ + (2N )j m/2)ψ̂((2N )−j ξ + m/2)|2 d ξ
j≤−M m∈Z R R
Let ξ0 ∈ R and fˆ be defined as above. If ξ + (2N )j m/2 ∈ (ξ0 − ε, ξ0 + ε), then

|(2N )j m/2| < ε for fixed j ≤ −M . By the hypothesis of fˆ , the number of summation
2ε
index m is bounded by (2N )j
. Thus
(2N )−j (ξ0 +ε)

2ε ˆ −j
I2 ≤ |f (ξ )ψ̂((2N ) ξ )| d ξ =
2
|ψ̂(ξ )|2 d ξ.
(2N )j j≤−M R j≤−M (2N )−j (ξ −ε)
0
(8.55)
Also for any ξo = 0, given η > 0, a positive integer M may be chosen so that
∞
|ψ̂(ξ )|2 d ξ < η. (8.56)
2(2N )M ξ0
1+2N
−1
It is clear that for 0 < ε < 2N2N +1 0
ξ , the intervals ((2N )−j (ξ0 − ε), (2N )−j (ξ0 +
ε)), j ∈ Z are mutually disjoint. It follows that
(2N )−j (ξ0 +ε)

I2 ≤ |ψ̂(ξ )|2 d ξ
j≤−M (2N )−j (ξ0 −ε)
∞
≤ |ψ̂(ξ )|2 d ξ
2(2N )M ξ0
1+2N
< η.
This completes the Theorem.

We establish a sufficient condition for the system {ψj,λ (t)}j∈Z,λ∈Λ to be wavelet frame
for L2 (R). For this, we prove the following lemma.
Lemma 8.3 Let f be in L2 (R) such that fˆ ∈ Cc (R). If ψ ∈ L2 (R) and ess supξ ∈R
−j
j∈Z |ψ̃((2N ) ξ )| < ∞ then
2

|f , ψj,λ | =2
|fˆ (ξ )|2 |ψ̂((2N )−j ξ )|2 d ξ + Rψ (f ) (8.57)
j∈R λ∈Λ R j∈Z
where Rψ (f ) = R0 + R1 + · · · + R2N −1 and for 0 ≤ p ≤ 2N − 1, Rp is given by

!
1
Rp = fˆ (ξ + (2N )j p/2)ψ̂((2N )−j ξ + p/2)
4N j∈Z R l=p
!
fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)(1 + eπi N (l−p) ) d ξ.
r
(8.58)
Proof By (8.33), (8.34), and (8.34), we have

|f , ψj,λ |2
j∈Z λ∈Λ

1
= 2 |fˆ (ξ )|2 |ψ̂((2N )−j ξ )|2 d ξ
4N R
j∈Z
!
1
fˆ (ξ )ψ̂((2N )−j ξ ) fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)(1 + eπ i N l )d ξ
r
+
4N
j∈Z l=0 R

1
+ 2 |fˆ (ξ + (2N )j /2)|2 |ψ̂((2N )−j ξ + 1/2)|2 d ξ
4N R
j∈Z
!
1
+ fˆ (ξ + (2N )j /2)ψ̂((2N )−j ξ + 1/2)
4N
j∈Z l=1 R
!
× fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)(1 + eπ i N (l−1) ) d ξ +
r
..
.
1
+ 2 |fˆ (ξ + (2N − 1)(2N )j /2)|2 |ψ̂((2N )−j ξ + (2N − 1)/2)|2 d ξ
4N R
j∈Z
!
1
+ fˆ (ξ + (2N − 1)(2N )j /2)ψ̂((2N )−j ξ + (2N − 1)/2)
4N R
j∈Z l=2N −1
!
× fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)(1 + eπ i N (l−2N +1) ) d ξ.
r
Since for 0 ≤ p ≤ 2N − 1,

|fˆ (ξ + (2N )j p)|2 |ψ((2N )−j ξ + p)|2 d ξ = |fˆ (ξ )|2 |ψ((2N )−j ξ )|2 d ξ.
j∈Z R j∈Z R
(8.59)
Therefore, we obtain

|f , ψj,λ |2 = |fˆ (ξ )|2 |ψ((2N )−j ξ )|2 d ξ + R0
j∈R λ∈Λ R j∈Z
R1 + · · · + R2N −1
which is the desired result.
We establish a sufficient condition for nonuniform wavelet frame for L2 (R).

Theorem 8.8 Let ψ ∈ L2 (R) be such that
⎧ ⎫
⎨ ⎬
A= inf |ψ̂((2N )−j ξ )|2 − |ψ̂((2N )−j ξ )ψ̂((2N )−j ξ + l/2)| > 0,
|ξ |∈[1,2N ] ⎩ ⎭
j∈Z j∈Z l=0
and
⎧ ⎫
⎨ ⎬
B= sup |ψ̂((2N )−j ξ )ψ̂((2N )−j ξ + l/2| < ∞. (8.60)
|ξ |∈[1,2N ] ⎩ j∈Z ⎭
l∈Z
Then {ψj,λ }j∈Z,λ∈Λ is a wavelet frame with bounds A and B.
Proof Let f ∈ L2 (R) such that fˆ is continuous and compactly supported. We estimate
Rψ (f ). By (8.58), we have

!
1
|Rp | =
fˆ (ξ + (2N )j p/2)ψ̂((2N )−j ξ + p/2)
4N j∈Z l=p R
!
r
× fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2)(1 + eπ i N (l−p) ) d ξ
!
1 ˆ
≤ f (ξ + (2N )j p/2)ψ̂((2N )−j ξ + p/2)
2N
j∈Z l=p R
!

× fˆ (ξ + (2N )j l/2)ψ̂((2N )−j ξ + l/2) d ξ.
Applying Cauchy–Schwarz’ inequality on the integral, we get

1/2
1
|Rp | ≤ |fˆ (ξ + (2N )j p/2)|2 |ψ̂((2N )−j ξ + p/2)ψ̂((2N )−j ξ + l/2)|d ξ
2N R
j∈Z l =p
1/2
× |fˆ (ξ + (2N )j l/2)|2 |ψ̂((2N )−j ξ + p/2)ψ̂((2N )−j ξ + l/2)|d ξ .
R
Again using Cauchy–Schwarz’ inequality on the summation over l, we obtain

⎧ ⎫1/2

1 ⎨ ˆ −j −j
⎬
|Rp | ≤ |f (ξ + (2N ) p/2)| |ψ̂((2N ) ξ + p/2)ψ̂((2N ) ξ + l/2)|d ξ
j 2
2N ⎩ R ⎭
j∈Z l=p
⎧ ⎫1/2
⎨ ⎬
× |fˆ (ξ + (2N )j l/2)|2 |ψ̂((2N )−j ξ + p/2)ψ̂((2N )−j ξ + l/2)|d ξ
⎩ R ⎭
l=p
1
= IJ .
2N
j∈Z
where
⎧ ⎫1/2
⎨ ⎬
I= |fˆ (ξ + (2N )j p/2)|2 |ψ̂((2N )−j ξ + p/2)ψ̂((2N )−j ξ + l/2)|d ξ ,
⎩ ⎭
l=p R
⎧ ⎫1/2
⎨ ⎬
J = |fˆ (ξ + (2N )j l/2)|2 |ψ̂((2N )−j ξ + p/2)ψ̂((2N )−j ξ + l/2)|d ξ .
⎩ R ⎭
l=p
On changing the variable ξ → ξ + 21 (2N )j (p − l) in J ,

⎧ ⎫1/2
⎨ ⎬
J = |fˆ (ξ + (2N )j p/2)|2 |ψ̂((2N )−j ξ + (2p − l)/2)ψ̂((2N )−j ξ + p/2)|d ξ
⎩ R ⎭
l=p
⎧ ⎫1/2
⎨ ⎬
= |fˆ (ξ + (2N )j p/2)|2 |ψ̂((2N )−j ξ + l /2)ψ̂((2N )−j ξ + p/2)|d ξ
⎩ ⎭
l =p R
= I.
We get

1
|Rp | ≤ |fˆ (ξ + (2N )j p/2)|2 |ψ̂((2N )−j ξ + l/2)ψ̂((2N )−j ξ + p/2)|d ξ
2N
j∈Z l=p R

1
= |fˆ (ξ )|2 |ψ̂((2N )−j ξ )ψ̂((2N )−j ξ + (l − p)/2)|d ξ
2N R
j∈Z l=p

1
= |fˆ (ξ )|2 |ψ̂((2N )−j ξ )ψ̂((2N )−j ξ + (l − p)/2)|d ξ.
2N R j∈Z l=p
Therefore, we can write
2N −1
1
|Rψ (f )| ≤ |fˆ (ξ )|2 |ψ̂((2N )−j ξ )ψ̂((2N )−j ξ + (l − p)/2)|d ξ.
2N p=0 R j∈Z l=p
(8.61)
From (8.57) and (8.61), we have

|f , ψj,λ |2 ≤ |fˆ (ξ )|2 |ψ̂((2N )−j ξ )|2 d ξ
j∈Z λ∈Λ R j∈Z
2N −1
1
+ |fˆ (ξ )|2 |ψ̂((2N )−j ξ )ψ̂((2N )−j ξ + (l − p)/2)|d ξ
2N
p=0 R j∈Z l =p
⎛ ⎞

= ˆ
|f (ξ )|2 ⎝ −j
|ψ̂((2N ) ξ )ψ̂((2N ) ξ + l/2|⎠ d ξ
−j (8.62)
R j∈Z l∈Z
and

|f , ψj,λ |2
j∈Z λ∈Λ
⎛ ⎞

≥ |fˆ (ξ )|2 ⎝ |ψ̂((2N )−j ξ )|2 − |ψ̂((2N )−j ξ )ψ̂((2N )−j ξ + l/2)|⎠ d ξ
R j∈Z j∈Z l=0
(8.63)
Taking supremum in (8.62) and infimum in (8.63), we get

Af ≤ |f , ψj,λ |2 ≤ Bf 2 .
j∈Z λ∈Λ
This completes the theorem.
8.4 Exercises
8.1 Prove Theorem 8.1.

8.2 State and prove Analogue of Cohen’s conditions for NUMRA.
8.3 Show that Collection {N m/2 ψk (N m x λ)}λ ∈ Λ+ , m ∈ Z, k = 1, 2, . . . , N − 1
as defined in Lemma 8.1 forms a complete orthonormal system for L2 (R+ ).
8.4 For parameters N and r a s in Definition 8.3 and for m10 (ξ ) = m11 (ξ ) = 1/2, ξ ∈
[0, 2). Show that the system of functions generates a NUMRA in L2 (R+ ).
8.4 Exercises 315
8.5 Give definition of nonuniform wavelet frames.

8.6 Prove that the system {ψj,λ }j∈Z,λ∈Λ forms a Bessel sequence for B given by
(8.60).
References
1. Sharma, V., & Manchanda, P. (2013). Wavelet packets associated with nonuniform multireso-
lution analysis on positive half-line. Asian-European Journal of Mathematics, 6(1), 1350007
(1–16).
2. Meenakshi, M., Manchanda, P., & Siddiqi, A. H. (2012). Wavelets associated with non uniform
multiresolution analysis on positive half line. International Journal of Wavelets, Multiresolution
and Information Processing 10(2), 1250018(1–27).
3. Sharma, V., & Manchanda, P. (2015). Nonuniform wavelet frames in L2 (R). Asian European
Journal of Mathematics, 8(2), 1550034(1–15).
Chapter 9
Orthogonal Vector-Valued Wavelets
on R+
9.1 Introduction
We have considered the notion of vector-valued multiresolution analysis (VMRA)

on positive half line R+ and studied associated vector-valued wavelets and wavelet
packets. Xia and Suter in 1996 generalized the concept of multiresolution analysis
(MRA) on R to vector-valued multiresolution analysis (VMRA) on R and stud-
ied associated vector-valued wavelets. Farkov [1] introduced MRA on R+ . In this
chapter, we have introduced vector-valued multiresolution analysis (VMRA) on R+ ,
where the associated subspace V0 of L 2 (R+ , C N ) has an orthonormal basis, a family
of translates of a vector-valued function Φ, i.e, {Φ(x l)}l∈Z+ , where Z+ is the set
of nonnegative integers. The necessary and sufficient condition for the existence of
associated vector-valued wavelets has been obtained and the construction of vector-
valued multiresolution analysis (VMRA) on R+ has been presented.
Lang [2–4] introduced the concept of compactly supported orthogonal wavelets
on the locally compact Cantor dyadic group C . These wavelets turn out to be certain
lacunary Walsh series on the R+ . Later on, Farkov [5] extended these results on
the wavelet analysis to the locally compact Abelian group G which is defined for an
integer p ≥ 2 and coincides with C when p = 2. Farkov [1] has presented the general
construction of all compactly supported orthogonal p-wavelets in L 2 (R+ ). He proved
necessary and sufficient conditions for scaling filters with p n many terms ( p, n ≥ 2)
to generate a p-MRA in L 2 (R+ ). Farkov’s approach is connected with Walsh–Fourier
theory. Thereafter, Protasov and Farkov [6] constructed dyadic compactly supported
wavelets in L 2 (R+ ). Farkov et al. [7] provided an algorithm for computing compactly
supported biorthogonal wavelets on positive half line.
Xia and Suter (1996) initiated the study of vector-valued multiresolution analy-
sis and orthogonal vector-valued wavelets on R. They unveiled that vector-valued
wavelets can be viewed as a class of generalized multiwavelets and multiwavelets
can be yielded from component functions in vector-valued wavelets. Associated
with VMRA, they have described vector-valued scaling functions, which leads to

https://doi.org/10.1007/978-981-13-6370-2_9
318 9 Orthogonal Vector-Valued Wavelets on R+
vector-valued wavelet functions. They carried out vector-valued wavelet transform

for vector-valued signals where correlation exists between their components.
In the present chapter, we discuss vector-valued MRA p-analysis on positive
half line and associated vector-valued wavelets. These results are analog of earlier
results cited above. Some results on vector-valued nonuniform MRA [8–10] are also
presented.
9.2 Vector-Valued Multiresolution Analysis
Let s be a constant and 2 ≤ s ∈ Z. By L 2 (R, Cs ) we denote the set of all vector-valued

functions f(t), i.e.,

L 2 (R, Cs ) = f(t) = ( f 1 (t), f 2 (t), . . . , f s (t))T : t ∈ R, f k (t) ∈ L 2 (R), k = 1, 2, . . . , s ,
where T means the transpose and Cs denote the s-dimensional complex Eucledian
space.
The space L 2 (R, Cs ) is called vector-valued function space. For f ∈ L 2 (R, Cs ),
f denotes the norm of the vector-valued function f and is defined as
s 1/2

f = | f k (t)|2 dt . (9.1)
k=1 R
For a vector-valued function f ∈ L 2 (R, Cs ) the integration of f(t) is defined as

T
f(t) dt = f 1 (t), f 2 (t), . . . , f s (t) .
R R R R
The Fourier transform of f(t) is defined by

f̂(ξ ) = f(t)e−2πitξ dt.
R
For any two vector-valued functions f, g ∈ L 2 (R, Cs ) the inner product < f, g > is
defined as
< f, g >= f(t)g∗ (t) dt, (9.2)
R
where * denotes the transpose and the complex conjugate.

The inner product in (9.2) is vector-valued, but still satisfies the property for an
inner product:
the linearity
< f, a1 g1 + a2 g2 >= a1 < f, g1 > +a2 < f, g2 >,
9.2 Vector-Valued Multiresolution Analysis 319
and the commutativity

< f, g >=< g, f >∗ .
A sequence {fk (t)} ∈ L 2 (R, Cs ) is said to be orthonormal if it satisfies
< fk (.), fn (.) >= δk,n Is , k, n ∈ Z, (9.3)
where δk,n denotes the Kronecker symbol such that δk,n = 1 when k = n and δk,n = 0
when k = n, Is denotes the identity matrix of order s × s.
Definition 9.1 A sequence {fk (t)} ∈ L 2 (R, Cs ), k ∈ Z is called an orthonormal

basis for L 2 (R, Cs ) if it satisfies (9.3) and moreover for any f ∈ L 2 (R, Cs ) there
exists a sequence of s × s constant matrices {Fk }k∈Z such that

f(t) = Fk fk (t), t ∈ R, (9.4)
k∈Z
where the multiplication Fk fk (t) for each fixed t is the s × 1 matrix multiplication,
and the convergence for infinite summation is as same as of the norm . defined by
(9.1) for the vector-valued function space.
Let {fk (t)}k∈Z is an orthonormal basis for L 2 (R, Cs ). Then the expansion (9.4)
for any f ∈ L 2 (R, Cs ) is unique and
Fk =< f, fk >, k ∈ Z. (9.5)
We also called the expansion (9.4), the Fourier expansion of f.

The corresponding Parseval equality is

< f, f >= Fk Fk∗ . (9.6)
k∈Z
From Eq. (9.6), it is clear that < f, f >= 0 if and only if f = 0 where 0 is the zero
vector.
Let φ(t) = (φ1 (t), φ2 (t), . . . , φs (t))T ∈ L 2 (R, Cs ) satisfy the following refine-
ment equation:
φ(t) = Pk φ(2t − k), (9.7)
k∈Z
where {Pk }k∈Z is a s × s constant matrix sequence.

Define a closed subspace V j ⊂ L 2 (R, Cs ) by
V j = clos L 2 (R,Cs ) (span{φ(2 j t − k) : k ∈ Z}), j ∈ Z. (9.8)
Vector-valued multiresolution analysis defined by Xia and Suter (1996) is as follows:

Definition 9.2 φ(t) defined by (9.7) generates a vector-valued multiresolution anal-

ysis {V j } j∈Z of L 2 (R, Cs ), if the sequence {V j } j∈Z defined in (9.8) satisfies the
following:
1. · · · ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ · · · ,
2. ∩ j∈Z V j = {0}, ∪ j∈Z V j is dense in L 2 (R, Cs ), where 0 is the zero vector of
L 2 (R, Cs ),
3. h(t) ∈ V0 if and only if h(2 j t) ∈ V j ∀ j ∈ Z,
4. there exists φ(t) ∈ V0 such that the sequence {φ(t − k), k ∈ Z} is an orthonormal
basis of V0 . The vector-valued function φ(t) is called a scaling function of the
vector-valued multiresolution analysis.
On taking the Fourier transform on both sides of (9.7), and assuming that φ̂(η) is
continuous at zero, we have
φ̂(η) = α(η/2)φ̂(η/2), η ∈ R, (9.9)
where
1
α(η) = Pk exp{−ikη}. (9.10)
2 k∈Z
Let W j , j ∈ Z denote the orthogonal complement of V j in V j+1 and there exists a

vector-valued function Ψ (t) ∈ L 2 (R, Cs ) such that the translations and dilations of
Ψ (t) form a Riesz basis of W j , i.e.,
W j = clos L 2 (R,Cs ) (span{Ψ (2 j t − k) : k ∈ Z}), j ∈ Z. (9.11)
Since Ψ (t) ∈ W0 ⊂ V1 , there exists a unique finitely supported sequence {Bk }k∈Z of
s × s constant matrices such that

Ψ (t) = Bk φ(2t − k). (9.12)
k∈Z
Let
1
β(η) = Bk .exp {−ikη}. (9.13)
2 k∈Z
Then the Eq. (9.12) becomes
Ψ̂ (η) = β(η/2)φ̂(η/2), η ∈ R. (9.14)
The following theorem proves the existence of orthogonal vector-valued wavelets:
Theorem 9.1 Let φ(t) defined in (9.7) is an orthogonal vector-valued scaling func-
tion. Assume that Ψ (t) ∈ L 2 (R, Cs ) and α(η) and β(η) are defined, respectively,
by (9.10) and (9.13). Then Ψ (t) is an orthogonal vector-valued wavelet function
9.2 Vector-Valued Multiresolution Analysis 321
associated with φ(t) if and only if
α(η)β(η)∗ + α(η + π )β(η + π )∗ = O, η ∈ R.
β(η)β(η)∗ + β(η + π )β(η + π )∗ = Is , η ∈ R.
The following theorem proved by Chen and Cheng (2007) present an algorithm
for the construction of compactly supported orthogonal vector-valued wavelets:
Theorem 9.2 Let φ(t) ∈ L 2 (R, Cs ) be a 3-coefficient compactly supported orthog-

onal vector-valued scaling functions satisfying the following refinement equation:
φ(t) = P0 φ(2t) + P1 φ(2t − 1) + P2 φ(2t − 2).
Assume that there exists an integer n, 0 ≤ n ≤ 2, such that the matrix A defined in
the following equation, is not only an invertible matrix but also a Hermitian matrix:
A2 = (2Is − Pn Pn∗ )−1 Pn Pn∗ .
Define
B j = A Pj , j = n,
B j = −A−1 P j , j = n, j, n ∈ {0, 1, 2}
and
Ψ (t) = B0 φ(2t) + B1 φ(2t − 1) + B2 φ(2t − 2).
Then Ψ (t) is an orthogonal vector-valued function associated with φ(t).
9.3 Vector-Valued Multiresolution p-Analysis on R+
Let N ≥ 2 be an integer and C N denotes N -dimensional complex Euclidean space,

IN and O represent N × N identity matrix and the zero matrix, respectively.
L 2 (R+ , C N ) represents the set of square integrable vector-valued functions f(t) on
positive half line, R+ i.e,
T
L 2 (R+ , C N ) = f(t) = f 1 (t), f 2 (t), , . . . , f N (t) ,
where t ∈ R+ , f v (t) ∈ L 2 (R+ ), v = 1, 2, . . . N and T denotes transpose.

The space L 2 (R+ , C N ) is called vector-valued function space.
For f ∈ L 2 (R+ , C N ), f L 2 (R+ ,CN ) is the norm of the function f, i.e,

N

f L 2 (R+ ,C N ) = | f v (t)|2 dt
v=1 R+
and integration of f is given by

T
f(t)dt = f 1 (t)dt, f 2 (t)dt, . . . , f N (t)dt .
R+ R+ R+ R+
The Walsh–Fourier transform of f(t) is defined by

T
f̃(w) = f(t)χ (k, w)dt = f˜1 (w), f˜2 (w), . . . , fÑ (w). , (9.15)
R+
where
f˜i (w) = f i (t)χ (t, w)dt, 1 ≤ i ≤ N.
R+
For any f ∈ L 2 (R+ , C N ), we have

f(t)f(t k)dt = f̃(w)f̃(w)∗ χ (k, w)dw (9.16)
R+ R+
For two vector-valued functions f, h ∈ L 2 (R+ , C N ), their symbol inner product is

defined by
f, h L 2 (R+ ,CN ) = f(t)h(t)∗ dt,
R+
where ‘*’ means complex conjugate and transpose. The product defined above is
matrix-valued (usually it is a scalar-valued) and it satisfies properties of an inner
product such as linearity and commutativity.
Definition 9.3 A sequence {fk (t)}k∈Z+ ⊂ U ⊆ L 2 (R+ , C N ) is called an orthonormal

set of U, if it satisfies
fk (t), fn (t) = δk,n I N , (9.17)
where δk,n is the Kronecker delta such that δk,n = 1 when k = n and δk,n = 0 when
k = n.
A vector-valued function f(t) ∈ U ⊆ L 2 (R+ , C N ) is said to be orthogonal in U
if its translations, i.e., {f(t k)}k∈Z+ satisfy
f(t k), f(t n) = δk,n I N , k, n ∈ Z+ . (9.18)

9.3 Vector-Valued Multiresolution p-Analysis on R+ 323
Definition 9.4 A sequence of vector-valued functions
{fk (t)}k∈Z+ ⊂ U ⊆ L 2 (R+ , C N ),
is called an orthonormal basis of U if it satisfies (9.17) and for any h(t) ∈ U, there
exists a unique sequence {Ak }k∈Z+ whose each element is an N × N constant matrix
such that
h(t) = Ak fk (t).
k∈Z+
The multiresolution analysis approach is one of the main approaches in the construc-
tion of wavelets. We introduce vector-valued multiresolution p-analysis on positive
half line and give the definition for associated orthogonal vector-valued wavelets [1,
11].
Definition 9.5 A vector-valued multiresolution p-analysis in L 2 (R+ , C ) is a

N
sequence of closed subspaces V j , j ∈ Z of L 2 (R+ , CN ) such that the following hold:

V j ⊂ V j+1 , j ∈ Z,
(a)
j V j is dense in L (R+ , C ) and j V j = {0}, where 0 is the zero vector of
2 N
(b)
+ N
L (R , C ),
2
(c) f(t) ∈ V j if and only if f( pt) ∈ V j+1 ,

(d) f(t) ∈ V0 ⇒ f(t ⊕ k) ∈ V0 for all k ∈ Z+ ,
(e) there exists a function Φ ∈ V0 called scaling function such that its translations
{Φk (t) = Φ(t k), k ∈ Z+ }, form an orthonormal basis for V0 .
Now Φ(t) ∈ V0 ⇒ Φ( pt) ∈ V1 , by (e),

δ0,k I N = Φ(t)Φ(t k)dt
R+

√ √
= pΦ( pt) pΦ( pt k)dt.
R+
√
So, Φ1,k (t) = { pΦ( pt k)}k∈Z+ form an orthonormal basis for V1 .
Therefore, the space V j is defined by
j
V j = clos L 2 (R+ ,CN ) (span{ p 2 Φ( p j t k)}, k ∈ Z+ ), j ∈ Z. (9.19)
Now Φ = (φ1 (t), φ2 (t), . . . , φ N (t))T ∈ V1 , we write

Φ(t) = Rk Φ( pt k) , t ∈ R+ , (9.20)
k∈Z+
where {Rk }k∈Z+ is a sequence of N × N constant matrices. If the sequence {Rk }k∈Z+
is finite, we say that Φ(t) is a compactly supported vector-valued function.
On taking Walsh–Fourier transform, we have
Φ̃(w) = R(w/ p)Φ̃(w/ p), w ∈ R+ , (9.21)
where
1
R(w) = Rk χ (k, w). (9.22)
p k∈Z
+
Noting that χ (k, w + l) = χ (k, w), k, l ∈ Z+ , so R(w) is 1-periodic function of w.

By (9.21), we have

w w w
Φ̃(w) = R R 2
Φ
p p p

w w
=R R . . . Φ̃(0)
p p2
∞
w
= R l
Φ̃(0).
l=1
p
We get
R(0) = I N or Rk = I N .
k∈Z+
For j ∈ Z, let W j denote the orthocomplement subspace of V j in V j+1 and there

exist p − 1 vector-valued functions Ψm (t) ∈ L 2 (R+ , C N ), m ∈ Λ, where Λ =
{1, 2, . . . , p − 1}, such that their translations and dilations form a Riesz basis of
W j , i.e,
j
W j = clos L 2 (R+ ,C N ) (span{ p 2 Ψm ( p j t k)}, m ∈ Λ, k ∈ Z+ ), j ∈ Z. (9.23)
For each m ∈ Λ, Ψm (t) ∈ W0 ⊂ V1 , there exist p − 1 finitely supported sequences

of N × N constant matrices {Sk(m) }k∈Z+ such that

Ψm (t) = Sk(m) Φ( pt k), m ∈ Λ, t ∈ R+ . (9.24)
k∈Z+
By taking Walsh–Fourier transform, the refinement equation (9.24) becomes
Ψ̃m (w) = S (m) (w/ p)Φ̃(w/ p), w ∈ R+ , m ∈ Λ, (9.25)
where
1 (m)
S (m) (w) = Sk χ (k, w). (9.26)
p
k∈Z+
9.3 Vector-Valued Multiresolution p-Analysis on R+ 325
If Φ(t) ∈ L 2 (R+ , C N ) is an orthogonal vector-valued scaling function, then it follows

from (9.18) that
Φ(t), Φ(t k) = δ0,k I N , k ∈ Z+ . (9.27)
Definition 9.6 The collection of functions Ψm (t) ∈ L 2 (R+ , C N ), m ∈ Λ are orthog-

onal vector-valued wavelet functions associated with the orthogonal vector-valued
scaling function Φ(t), if they satisfy
Ψm (t), Φ(t k) = O, m ∈ Λ, k ∈ Z+ (9.28)
and the family {Ψm (t k), m ∈ Λ}k∈Z+ is an orthonormal basis of the subspace W0 .
Therefore,
Ψm (t l), Ψn (t k) = δl,k δm,n I N , m, n ∈ Λ, l, k ∈ Z+ . (9.29)
The following lemma gives a characterization in the frequency domain of an orthog-

onal vector-valued function f(t).
Lemma 9.1 Let f(t) ∈ L 2 (R+ , C N ). Then f(t) is an orthogonal vector-valued func-
tion if and only if
f̃(w + l)f̃(w + l)∗ = I N , w ∈ R+ . (9.30)
l∈Z+
T
Proof Let f(t) = f 1 (t), f 2 (t), . . . , f N (t) ∈ L 2 (R+ , C N ) is an orthogonal func-
tion, then

δ0,k I N = f(t)f(t k)∗ dt,
R+
By identity (9.16),

δ0,k I N = f̃(w)f̃(w)∗ χ (k, w)dw
R+
l+1
= f̃(w)f̃(w)∗ χ (k, w)dw
l∈Z+ l
1
= f̃(w + l)f̃(w + l)∗ χ (k, w)dw
0 l∈Z
+

So, f(t) ∈ L 2 (R+ , C N ) is orthogonal ⇔ l∈Z+ f̃(w + l)f̃(w + l)∗ = I N .
Lemma 9.2 If Φ(t) ∈ L 2 (R+ , C N ), defined by (9.20), is an orthogonal vector-

valued scaling function, then for every k ∈ Z+ , we have

Ru (Ru⊕ pk )∗ = pδ0,k I N . (9.31)
u∈Z+
Proof On using (9.20) in the relation (9.27), we have
δ0,k I N = Φ(t k), Φ(t)

= Ru Φ( pt pk u) Φ( pt v)∗ Rv∗ dt
u∈Z+ R+ v∈Z+

= Ru Φ( pt pk u)Φ( pt v)Rv∗ dt
u∈Z+ v∈Z+ R+
1
= Ru Φ(t pk u), Φ(t v) Rv∗
p
u,v∈Z+
1
= Ru (Ru⊕ pk )∗ .
p u∈Z
+
9.4 The Existence of Orthogonal Vector-Valued Wavelets

on R+
In this section, we start with considering the existence of compactly supported orthog-
onal vector-valued wavelets on R+ . The necessary and sufficient condition for the
vector-valued wavelets associated with VMRA on R+ is derived as follows:
Theorem 9.3 Let Φ(t) ∈ L 2 (R+ , C N ) defined in (9.20) be an orthogonal vector-

valued scaling function. Suppose that Ψm (t) ∈ L 2 (R+ , C N ), m ∈ Λ and R(w) and
S (m) (w) are defined by (9.22) and (9.26), respectively. Then Ψm (t), m ∈ Λ are
orthogonal vector-valued wavelet functions associated with Φ(t) if and only if

p−1 ∗
w+l w+l
R S (m) = O, m ∈ Λ, w ∈ R+ , (9.32)
l=0
p p
and

p−1 ∗
(m) w+l (n) w+l
S S = δm,n I N , m, n ∈ Λ, w ∈ R+ . (9.33)
l=0
p p
Proof First, we prove the necessary part of the theorem.

By Lemma 9.1 and (9.28), we have
9.4 The Existence of Orthogonal Vector-Valued Wavelets on R+ 327
O = Φ(t), Ψm (t)

= Φ̃(w + l)Ψ̃m (w + l)∗
l∈Z+

w+l w+l w + l ∗ (m) w + l ∗
= R Φ̃ Φ̃ S
p p p p
l∈Z+
∗ ∗
w w w w
= R + n Φ̃ + n Φ̃ + n S (m) +n
p p p p
l= pn
∗
w 1 w 1 w 1
+ R + + n Φ̃ + + n Φ̃ + +n
p p p p p p
l= pn+1
∗
w 1
×S (m) + +n
p p
.
.
.
∗
w p−1 w p−1 w p−1
+ R + + n Φ̃ + + n Φ̃ + +n
p p p p p p
l= pn+( p−1)
∗
w p−1
×S (m) + +n
p p
On simplification, we get
⎛ ⎞
∗ ∗
w ⎝ w w w
O=R Φ̃ + n Φ̃ + n ⎠ S (m)
p p p p
l= pn
⎛ ⎞
∗
w 1 ⎝ w 1 w 1 w 1 ∗
+R + Φ̃ + + n Φ̃ + + n ⎠ S (m) +
p p p p p p p p
l= pn+1
.
.
.
.
⎛ ⎞
∗
w p−1 ⎝ w p−1 w p−1
+R + Φ̃ + + n Φ̃ + +n ⎠
p p p p p p
l= pn+ p−1

w p−1 ∗
×S (m) +
p p
p−1
w+l (m) w+l ∗
= R S .
p p
l=0
Again from (9.29) and Lemma 9.1, for m, n ∈ Λ, we get


δm,n I N = Ψ̃m (w + l)Ψ̃n (w + l)∗
l∈Z+

w+l w+l w + l ∗ (m) w + l ∗
= S (m) Φ̃ Φ̃ S
p p p p
l∈Z+
⎛ ⎞
∗ ∗
w ⎝ w w w
= S (m) Φ̃ + n Φ̃ + n ⎠ S (n)
p p p p
l= pn
⎛ ⎞
∗
(m) w 1 ⎝ w 1 w 1 w 1 ∗
+S + Φ̃ + + n Φ̃ + + n ⎠ S (n) +
p p p p p p p p
l= pn+1
.
.
.
.
⎛ ⎞
∗
w p − 1 w p − 1 w p − 1
+ S (m) + ⎝ Φ̃ + + n Φ̃ + +n ⎠
p p p p p p
l= pn+ p−1

(n) w p−1 ∗
×S +
p p
p−1
w+l w+l ∗
= S (m) S (n) .
p p
l=0
Conversely, suppose identities (9.32) and (9.33) hold. By similar steps performed
above, we have

w+l ∗
p−1
w+l
Φ̃(w + l)Ψ̃m (w + l)∗ = R S (m) = O,
l∈Z l=0
p p
+
and

p−1 ∗
∗ (m) w+l (n) w+l
Ψ̃m (w + l)Ψ̃n (w + l) = S S = δm,n I N .
l∈Z+ l=0
p p
Therefore,
l+1
Φ(t), Ψm (t k) = Φ̃(w)Ψ̃m (w)∗ χ (k, w)dw
l∈Z+ l
1
= Ψ̃ (w + l)Ψ̃m (w + l)∗ χ (k, w)dw
0 l∈Z
+
= O, m ∈ Λ, k ∈ Z+ .
and
1
Ψm (t), Ψn (t k) = Ψ̃m (w + l)Ψ̃n (w + l)∗ χ (k, w)dw
0 l∈Z
+
= δ0,k δm,n I N , m, n ∈ Λ, k ∈ Z+ .
Thus, Φ(t) and Ψm (t), m ∈ Λ are mutually orthogonal and {Ψm (t), m ∈ Λ} is a
finite family of orthogonal vector-valued functions. This proves the orthogonality of
{Ψm (t k), m ∈ Λ}k∈Z+ .
We now prove the completeness of {Ψm (t k), m ∈ Λ}k∈Z+ in W0 . For any
f ∈ W0 ⊂ V1 , there exists finitely supported sequence of N × N constant matrices
{Ak }k∈Z+ such that
f= Ak Φ( pt k). (9.34)
k∈Z+
Thus
f̃(w) = A (w/ p)Φ̃(w/ p), (9.35)
where
1
A (w) = Ak χ (k, w).
p
k∈Z+
Also for f ∈ W0 and f ∈

/ V0 means

f(t)Φ(t k)∗ dt = O, k ∈ Z+ . (9.36)
R+
This is equivalent to
f̃(w + l)Φ̃(w + l)∗ = O. (9.37)
l∈Z+
On using (9.21), (9.35) and Lemma 9.1, we have

p−1 ∗
w+l w+l
A R = O, w ∈ R+ . (9.38)
l=0
p p
Let
∗
A1 (w) = A wp , A w+1
p
, . . . , A w+pp−1 ,
and
∗
R1 (w) = R wp , R w+1
p
, . . . , R w+ p−1
p
.
For i = 1, 2, . . . , p − 1, we set
∗
Si (w) = S (i) wp , S (i) w+1
p
, . . . , S (i) w+ p−1
p
.
It is clear that for any w ∈ R+ , the identities (9.32) and (9.33) imply that the column
vectors in the pN × N matrix R1 (w) and column vectors in the pN × N Si (w) are
orthogonal and all these vectors form an orthonormal basis for the pN -dimensional
complex Euclidean space C pN . The identity (9.38) implies that the column vectors in
the pN × N matrix A1 (w) and column vectors in the pN × N R1 (w) are orthogonal.
Then, for each m ∈ Λ, there exist p − 1 matrices L (m) (w), m ∈ Λ whose all entries
are 1-periodic functions of w such that

A (w) = L (m) (w)S (m) (w), w ∈ R+ . (9.39)
m∈Λ
Therefore, by (9.35)

f̃(w) = L (m) (w/ p)S (m) (w/ p)Φ(w/ p)
m∈Λ

= L (m) (w/ p)Ψ̃m (w). (9.40)
m∈Λ
By the orthonormality of {Ψm (t k), m ∈ Λ}k∈Z+ , we get

l+1
∗
f̃( pw)f̃( pw) dw = L (m) (w)Ψ̃m ( pw)Ψ̃m ( pw)∗ L (m) (w)∗ dw
R+ l∈Z+ l m∈Λ
1
= L (m) (w)L (m) (w)∗ dw.
0 m∈Λ
This proves that L (m) (w) has Walsh–Fourier series expansion. Let constant N × N
matrices {Q (m)
k }k∈Z+ , m ∈ Λ be its Walsh–Fourier coefficients. Then

f= Q (m)
k Ψm (t k).
k∈Z+ m∈Λ
This proves the completeness of {Ψm (t k), m ∈ Λ}k∈Z+ in W0 .
By Walsh–Fourier analysis and (9.22), (9.26), identities (9.32) and (9.33) are
equivalent to, respectively,
(m) ∗
Rv (Sv⊕ pk ) = O, m ∈ Λ, k ∈ Z+ (9.41)
v∈Z+
(n)
Sv(m) (Sv⊕ ∗
pk ) = pδm,n δ0,k I N , m, n ∈ Λ, k ∈ Z+ . (9.42)
v∈Z+
Theorem 9.3 implies that a vector-valued multiresolution p-analysis in L 2 (R+ , C N )

provides vector-valued scaling function Φ, which leads to associated p-1 orthogonal
vector-valued wavelet functions Ψm (t), m ∈ Λ such that its dilations and translations
Ψ j,k,m (t) = p j/2 Ψm ( p j t k), j ∈ Z, k ∈ Z+ , m ∈ Λ
form an orthonormal basis for L 2 (R+ , C N ). The following theorem deals with the
construction of vector-valued scaling function Φ.
Theorem 9.4 Let

1
R(w) = Rk χ (k, w),
p k∈Z
+
be a matrix-valued scaling filter satisfying following conditions:

(a)

p−1 ∗
w+l w+l
R R = IN . (9.43)
l=0
p p
(b) There exists a constant C > 0 and integer M > 0, such that for w ∈ (0, p M )
M ∞
w w

R <C R . (9.44)
pl pl
l=1 l=1
If Φ ∈ L 2 (R+ , C N ) such that Φ̃(0)Φ̃(0)∗ = I N and its Walsh–Fourier transform can

be written as
∞
w
Φ̃(w) = R Φ̃(0), (9.45)
l=1
pl
then Φ(t) is a vector-valued scaling function for a vector valued multiresolution p-

analysis in L 2 (R+ , C N ). Thus, the corresponding family of vector-valued functions
j
{Ψ j,k,m (t) = p 2 Ψm ( p j t k), j ∈ Z, k ∈ Z+ , m ∈ Λ}
form an orthonormal basis for L 2 (R+ , C N ).

Proof It is sufficient to prove the orthonormality of Φ(t k), k ∈ Z+ . Now

Φ(t)Φ(t k)dt = Φ̃(w)Φ̃(w)∗ χ (k, w)dw. (9.46)
R+ R+
For an integer M > 0, let

M
w
μ M (w) = R Φ̃(0)1(0, p M ) (w), (9.47)
l=1
pl
where 1(0, p M ) (w) is the characteristic function of a subset (0, p M ) of R+ .

Now

μ M (w)μ M (w)∗ χ (k, w)dw
R+

pM
w w w
= R R . . . R Φ̃(0)
0 p p2 pM
∗ ∗ ∗
w w w
× Φ̃(0)∗ R R . . . R χ (k, w)dw
pM p M−1 p
1 M−1

M−1

∗
∗
=p M
R( p M−l
w) R(w)R(w) R( p M−l
w) χ (k, p M w)dw
0 l=1 l=1
M−1 M−1 ∗
1
p
∗
=p M
R( p M−l
w) χ (k, p M w)dw
0 l=1 l=1
M−1 M−1 ∗
2
p
∗
+p M
R( p M−l
w) χ (k, p M w)dw
1
p l=1 l=1
.
.
.
M−1 M−1 ∗
1
∗
+p M
R( p M−l
w) χ (k, p M w)dw
p−1
p l=1 l=1
M−1 p−1
1
p l l ∗
= pM R( p M−l w) R w+ R w+
0 l=1 l=0
p p
M−1 ∗

× R( p M−l w) χ (k, p M w)dw
l=1
.
Using (9.43) and Φ̃(0)Φ̃(0)∗ = I N , we have

μ M (w)μ M (w)∗ χ (k, w)dw
R+
M−1 M−1 ∗
p M−1 w w
= R R χ (k, w)dw
0 l=1
pl l=1
pl

= μ M−1 (w)μ M−1 (w)∗ χ (k, w)dw
R+
.
.
.
= μ1 (w)μ1 (w)∗ χ (k, w)dw
R+
1
= I N χ (k, w)dw
0
= δ0,k I N .
From (9.47), we get that μk (w) converges to Φ(w) pointwise. In view of (9.44), we
have
μ M (w)μ M (w)∗ − Φ̃(w)Φ̃(w)∗ ≤ (C + 1)Φ̃(w)Φ̃(w)∗ , w ∈ R+ .
Since all matrix norms are equivalent, there exists a constant C1 > 0 such that

μ M μ∗M − Φ̃ Φ̃ ∗ ≤ C1 μ M (w)μ M (w)∗ − Φ̃(w)Φ̃(w)∗ 2 dw.
R+
By the Dominated convergence theorem, we get
μ M μ∗M − Φ̃ Φ̃ ∗ → 0 as → ∞.
Therefore,

Φ̃(w)Φ̃(w)∗ χ (k, w)dw = lim μ M (w)μ M (w)∗ χ (k, w)dw
R+ M−→∞ R
+
= δ0,k I N .
This proves the orthonormality of Φ(t k), k ∈ Z+ .

9.5 Vector-Valued Nonuniform Multiresolution Analysis
Definition 9.7 Given integers N ≥ 1 and r odd with 1 ≤ r ≤ 2N − 1 such that r and
N are relatively prime, we say that ϕ ∈ L 2 (R, Cs ) generates a VNUMRA {V j } j∈Z
of L 2 (R, Cs ), if the sequence {V j } j∈Z satisfies:
(a) · · · ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ · · · ,
(b) ∪ j∈Z V j is dense in L 2 (R, Cs ),
(c) ∩ j∈Z V j = {0}, where 0 is the zero vector of L 2 (R, Cs ),
(d) ϕ(t) ∈ V j if and only if ϕ(2N t) ∈ V j+1 ∀ j ∈ Z.
(e) there exists ϕ(t) ∈ V0 such that the sequence {ϕ(t − λ), λ ∈ Λ} is an orthonor-
mal basis of V0 where Λ = {0, r/N } + 2Z. The vector-valued function ϕ(t) is
called a scaling function of the VNUMRA.
Note that when N = 1, one recovers from the above definition the definition of
vector-valued multiresolution analysis with dilation factor equal to 2.
Let ϕ(t) = (ϕ1 (t), ϕ2 (t), . . . , ϕs (t))T ∈ L 2 (R, Cs ) satisfy the following refine-
ment equation:
ϕ(t) = Aλ ϕ(2N t − λ) (9.48)
λ∈Λ
where {Aλ }λ∈Λ is s × s constant matrix sequence that has only finite number of
terms.
Define a closed subspace V j ∈ L 2 (R, Cs ) by

V j = clos L 2 (R,Cs ) span{ϕ (2N ) j t − λ , λ ∈ Λ} , j ∈ Z (9.49)
Given a VNUMRA let Wm denotes the orthogonal complement of Vm in Vm+1 ,

for any integer m. It is clear from the conditions (a), (b) and (c) of the Definition 9.7
that
L 2 (R, Cs ) = ⊕m∈Z Wm .
As is the case in the standard situation, the main purpose of VNUMRA is to construct
orthonormal basis of L 2 (R, Cs ) given by appropriate translates and dilates of a finite
collection of functions, called the associated wavelets.
Definition 9.8 A collection {Ψ k }k=1,2,...,2N −1 of functions in V1 will be called a set

of wavelets associated with a given VNUMRA if the family of functions {Ψ k (x −
λ)}k=1,2,...,2N −1, λ∈Λ is an orthonormal system of W0 .
On taking the Fourier transform on both sides of Eq. (9.48), we have
ϕ̂(2N η) = A(η)ϕ̂(η), η ∈ R (9.50)

9.5 Vector-Valued Nonuniform Multiresolution Analysis 335
where
1
A(η) = Aλ e−2πiλη . (9.51)
2N λ∈Λ
Since Λ = {0, r/N } + 2Z, we can write that
A(η) = A1λ + A2λ e−2πir/N η (9.52)
where {A1λ } and {A2λ } are s × s constant symmetric matrix sequences.

Then
η η
ϕ̂(η) = A ϕ̂
2N 2N
η η η
=A A A . . . ϕ̂(0)
2N (2N )2 (2N )3
∞
η
= A ϕ̂(0). (9.53)
k=1
(2N )k
Equation (9.53) implies that

A(0) = Is or A λ = Is , (9.54)
λ∈Λ
where Is denotes the identity matrix of order s × s.

W j = clos L 2 (R,Cs ) span{Ψ k (2N ) j t − λ , λ ∈ Λ, k = 1, 2, . . . , 2N − 1} , j ∈ Z.
(9.55)
Since Ψ k (t) ∈ W0 ⊂ V1 there exists a uniquely supported sequence
{Bλ,k }λ∈Λ,k=1,2,...,2N −1 of s × s constant matrices such that

Ψ k (t) = Bλ,k ϕ(2N t − λ) (9.56)
λ∈Λ
On taking the Fourier transform on both sides of Eq. (9.56), we have
Ψˆ k (2N η) = Bk (η)ϕ̂(η), (9.57)
where
1
Bk (η) = Bλ,k e−2πiλη . (9.58)
2N λ∈Λ
Bk (η) = Bλ,k
1
+ Bλ,k
2
e−2πir/N η , (9.59)
where {Bλ,k
1
} and {Bλ,k
2
} are s × s constant symmetric matrix sequences.
Lemma 9.3 If ϕ(t) ∈ L 2 (R, Cs ) defined by Eq. (9.48) is an orthogonal vector-

valued scaling function then we have

Am A∗2N (λ−σ )+m = 2N δλ,σ Is ∀λ, σ ∈ Λ, (9.60)
m∈2Z
where δλ,σ denotes the Kronecker’s delta.
Proof Since ϕ ∈ L 2 (R, Cs ) defined by Eq. (9.48) are orthogonal vector-valued scal-
ing functions, therefore for λ, σ ∈ Λ, we have

δλ,σ Is = ϕ(x − λ)ϕ ∗ (x − σ ) d x
R

= Au ϕ(2N x − 2N λ − u) ϕ ∗ (2N x − 2N σ − v)∗ Av d x
u∈Λ R v∈Λ

= Au ϕ(2N x − 2N λ − u)ϕ ∗ (2N x − 2N σ − v)∗ d x Av
u∈Λ v∈Λ R

1
= Au ϕ(x − 2N λ − u)ϕ ∗ (x − 2N σ − v)∗ d x Av .
2N u∈Λ v∈Λ R
Taking u = 2m, v = 2n where m, n ∈ Z, we have

δλ,σ Is = ϕ(x − λ)ϕ ∗ (x − σ ) d x
R
1
= Au < ϕ(x − 2N λ − u), ϕ(x − 2N σ − v) > A∗v
2N u∈Λ v∈Λ
1
= A2m < ϕ(x − 2N λ − 2m), ϕ(x − 2N σ − 2n) > A∗2n
2N
m∈Z n∈Z
1
= Am < ϕ(x − 2N λ − m), ϕ(x − 2N σ − n) > A∗n
2N
m∈2Z n∈2Z
1
= Am A∗2N (λ−σ )+m .
2N
m∈2Z
Therefore,
Am A∗2N (λ−σ )+m = 2N δλ,σ Is ∀λ, σ ∈ Λ.
m∈2Z
Taking u = r
N
+ 2m, v = 2n where m, n ∈ Z, we have
1
δλ,σ = Au < ϕ(x − 2N λ − u), ϕ(x − 2N σ − v) > A∗v
2N u∈Λ v∈Λ
1 r
= A Nr +2m < ϕ(x − 2N λ − − 2m), ϕ(x − 2N σ − 2n) > A∗2n
2N m∈Z n∈Z N
1 r
= A r +m < ϕ(x − 2N λ − − m), ϕ(x − 2N σ − n) > A∗n
2N m∈2Z n∈2Z N N
1
= Am A∗2N (λ−σ )+m .
2N m∈2Z
Thus, in both the cases, we get

Am A∗2N (λ−σ )+m = 2N δλ,σ Is ∀λ, σ ∈ Λ, ∀λ, σ ∈ Λ.
m∈2Z
Lemma 9.4 Consider a VNUMRA as in Definition 9.7. Let Ψ 0 = ϕ, B0 (.) = A(.)

and suppose that there exists 2N-1 functions Ψ k , k = 1, 2, . . . , 2N − 1 in V1 . Then
the family of functions {Ψ k (x − λ)}λ∈Λ,k=0,1,2,...,2N −1 will form an orthonormal sys-
tem for V1 iff for k, l ∈ {0, 1, 2, . . . , 2N − 1}

2N −1
ξ p ξ p
Bk + Bl∗ + = δk,l Is . (9.61)
p=0
2N 4N 2N 4N
Proof First, we will prove the necessary condition.

By the orthonormality of Ψ k (t) ∈ L 2 (R, Cs ), k = 0, 1, 2, . . . , 2N − 1 (or the
orthonormality of VNUMRA V j ), we have in the time domain

< Ψ k (. − λ), Ψ l∗ (. − σ ) >= Ψ k (t − λ)Ψ l∗ (t − σ ) dt = δk,l δλ,σ Is where
R
where λ, σ ∈ Λ and k, l ∈ {0, 1, 2, . . . , 2N − 1}. Equivalently, in the frequency

domain, we have

∗
e−2πi(λ−σ )w Ψˆ k (w)Ψˆ l (w) dw = δk,l δλ,σ Is where λ, σ ∈ Λ, k, l ∈ {0, 1, 2, . . . , 2N − 1}.
R
Taking λ = 2m, σ = 2n where m, n ∈ Z, we have


∗
δk,l δm,n Is = e−2πi(2m−2n)ξ Ψˆ k (ξ )Ψˆ l (ξ ) dξ
R
∗
= e−4πi(m−n)ξ Ψˆ k (ξ )Ψˆ l (ξ ) dξ
R ∗
= e−4πi(m−n)ξ Ψˆ k (ξ + N j)Ψˆ l (ξ + N j) dξ,
[0,N ) j∈Z
∗
Let wk,l (ξ ) = j∈Z Ψˆ k (ξ + N j)Ψˆ l (ξ + N j).
Therefore,

δk,l δm,n Is = e−4πi(m−n)ξ wk,l (ξ ) dξ
[0,N )
⎡ ⎤

2N −1
p ⎦
= e−4πi(m−n)ξ ⎣ wk,l ξ + dξ,
[0,1/2) p=0
2
Therefore, we obtain

2N −1 p
wk,l ξ + = 2δk,l Is . (9.62)
p=0
2
Also on taking λ = r
N
+ 2m and σ = 2n where m, n ∈ Z, we have

∗
0= e−2πiξ.2(m−n) e−2πiξr/N Ψˆ k (ξ )Ψˆ l (ξ ) dξ
R ∗
e−4πiξ(m−n) e−2πiξ. N Ψˆ k (ξ + N j)Ψˆ l (ξ + N j) dξ
r
=
[0,N ) j∈Z

e−4πiξ(m−n) e−2πiξ. wk,l (ξ ) dξ
r
= N
[0,N )
⎡ ⎤

2N −1
p
e−4πiξ(m−n) e−2πi N ξ ⎣ e−πi p N wk,l (ξ + )⎦ dξ.
r r
=
[0, 21 ) p=0
2
We conclude that

2N −1
p
α p wk,l (ξ + ) = 0 where α = e−πir/N . (9.63)
p=0
2
Also we have

2N −1
j ˆ j ˆ∗ j
wk,l ξ + = Ψk ξ + Ψl ξ + .
j=0
2 j∈Z
2 2
Therefore, Eq. (9.62) reduces to

p ˆ ∗ p
Ψˆ k ξ + Ψl ξ + = 2δk,l Is . (9.64)
p∈Z
2 2
Also

j ∗ j
wk,l (2N ξ ) = Ψˆ k 2N ξ + Ψˆ l 2N ξ +
2 2
j∈Z
j

j

j

j

= Bk ξ + ϕ̂ ξ + ϕ̂ ∗ ξ + Bl∗ ξ +
2 2 2 2
j∈Z

= Bk (ξ + n N )ϕ̂(ξ + n N )ϕ̂ ∗ (ξ + n N )Bl∗ (ξ + n N )
j=n.2N

1 1 1 1
+ Bk ξ + n N + ϕ̂ ξ + n N + ϕ̂ ∗ ξ + n N + Bl∗ ξ + n N +
2 2 2 2
j=n.2N +1

2 2 2 2
2 2 2 2
j=n.2N +2

3 3 3 3
2 2 2 2
j=n.2N +3
+···

2N − 1 2N − 1 2N − 1
+ Bk ξ + n N + ϕ̂ ξ + n N + ϕ̂ ∗ ξ + n N +
2 2 2
j=n.2N +(2N −1)

2N − 1
Bl∗ ξ + n N +
2
⎡ ⎤

= Bk (ξ ) ⎣ ϕ̂(ξ + n N )ϕ̂ ∗ (ξ + n N )⎦ Bl∗ (ξ )
j=n.2N
⎡ ⎤

1 ⎣ 1 ∗ 1 ⎦ ∗ 1
+Bk ξ+ ϕ̂ ξ + n N + ϕ̂ ξ + n N + Bl ξ +
2 2 2 2
j=n.2N +1
⎡ ⎤

2 2 2 ⎦ ∗ 2
+Bk ξ + ⎣ ϕ̂ ξ + nN + ϕ̂ ∗ ξ + n N + Bl ξ +
2 2 2 2
j=n.2N +2
⎡ ⎤

3 3 3 ⎦ ∗ 3
+Bk ξ + ⎣ ϕ̂ ξ + nN + ϕ̂ ∗ ξ + n N + Bl ξ +
2 2 2 2
j=n.2N +3
+···
⎡ ⎤

2N − 1 ⎣ 2N − 1 ∗ 2N − 1 ⎦
+Bk ξ + Ψ̂ ξ + nN + Ψ̂ ξ + n N +
2 2 2
j=n.2N +(2N −1)

2N − 1
Bl∗ ξ +
2

1 1 2 2
= 2 Bk (ξ )Bl∗ (ξ ) + Bk ξ + Bl∗ ξ + + Bk ξ + Bl∗ ξ + + ···
2 2 2 2

2N − 1 2N − 1
· · · + Bk ξ + Bl∗ ξ +
2 2

2N −1
j j
=2 Bk ξ + Bl∗ ξ + .
2 2
j=0
Therefore, we have

p ˆ ∗ p

2N −1
ξ p ξ p
Ψˆ k ξ + Ψl ξ + =2 Bk + Bl∗ + .
2 2 p=0
2N 4N 2N 4N
p∈Z
By using (9.64), we conclude that

2N −1
ξ p ξ p
Bk + Bl∗ + = δk,l Is , i.e.,
p=0
2N 4N 2N 4N

ξ ξ ξ 1 ξ 1 ξ 2 ξ 2
Bk Bl∗ + Bk + Bl∗ + + Bk + Bl∗ + + ···
2N 2N 2N 4N 2N 4N 2N 4N 2N 4N

ξ 2N − 1 ξ 2N − 1
· · · + Bk + Bl∗ + = δk,l Is .
2N 4N 2N 4N
Now we will prove the sufficiency.

By Eq. (9.57), for ξ ∈ R

j ∗ j
Ψˆk ξ + Ψˆ l ξ +
2 2
j∈Z

ξ j ξ j ξ j ξ j
= Bk + ϕ̂ + ϕ̂ ∗ + Bl∗ +
2N 4N 2N 4N 2N 4N 2N 4N
j∈Z
⎡ ⎤

ξ n ⎣ ξ n ξ n ⎦ ∗ ξ n
= Bk + ϕ̂ + ϕ̂ ∗ + Bl +
2N 2 2N 2 2N 2 2N 2
j=n.2N
⎡ ⎤

ξ n 1 ⎣ ξ ξ n n1 1 ⎦
+Bk + + ϕ̂ + +
+ + ϕ̂ ∗
2N 2 4N 2N2N 2 4N
2 4N
j=n.2N

ξ n 1
Bl∗ + + + ···
2N 2 4N
⎡ ⎤

ξ n 2N − 1 ⎣ ξ n 2N − 1 ∗ ξ n 2N − 1 ⎦
+Bk + + ϕ̂ + + ϕ̂ + +
2N 2 4N 2N 2 4N 2N 2 4N
j=n.2N

ξ n 2N − 1
Bl∗ + +
2N 2 4N

ξ ∗ ξ ξ 1 ∗ ξ 1
= 2 Bk Bl + Bk + Bl + +
2N 2N 2N 4N 2N 4N

ξ 2 ξ 2 ξ 2N − 1 ξ 2N − 1
Bk + Bl∗ + + · · · + Bk + Bl∗ +
2N 4N 2N 4N 2N 4N 2N 4N
= 2δk,l Is .
It proves the orthonormality of {Ψ k (t − λ), λ ∈ Λ, k = 0, 1, 2, . . . , 2N − 1}.

Now, we have the following result on the existence of a vector-valued wavelet
function:
Theorem 9.5 Suppose {Ψ k (t − λ)}λ∈Λ, k=0,1,...,2N −1 is the system as defined in
Lemma 9.4 and orthonormal in V1 . Then this system is complete in W0 ≡ V1 V0 .
Proof Since the system {Ψ k (t − λ)}λ∈Λ, k=0,1,...,2N −1 is orthonormal in V1 . By
Lemma 9.4 we have

ξ ξ ξ 1 ξ 1 ξ 2 ξ 2
Bk Bl∗ + Bk + Bl∗ + + Bk + Bl∗ + + ···
2N 2N 2N 4N 2N 4N 2N 4N 2N 4N

ξ 2N − 1 ξ 2N − 1
· · · + Bk + Bl∗ + = δk,l Is .
2N 4N 2N 4N
We will now prove its completeness.

For fk ∈ W0 ∃ the constant matrices {Fλ,k } such that

fk (t) = Fλ,k ϕ(2N t − λ), k = 0, 1, . . . , 2N − 1.
λ∈Λ
Thus
ξ ξ
f̂k (ξ ) = Fk ϕ̂ , (9.65)
2N 2N
where
1
Fk (ξ ) = Fλ,k e−2πiλξ .
2N λ∈Λ
On the other hand, f ∈

/ V0 and f ∈ W0 implies

fk (t)ϕ ∗ (t − λ) dt = 0, λ ∈ Λ.
R
This condition is equivalent to

j j
f̂k ξ + ϕ̂ ∗ ξ + = 0 ξ ∈ R.
j∈Z
2 2
Therefore, from Eqs. (9.50) and (9.65) we have

ξ j ξ j ξ j ξ j
Fk + ϕ̂ + ϕ̂ ∗ + A∗ + = 0, ξ ∈ R.
2N 4N 2N 4N 2N 4N 2N 4N
j∈Z
As similar to the identity (9.61) in Lemma 9.4, we have

ξ ∗ ξ ξ 1 ∗ ξ 1
Fk A + Fk + A + + ···
2N 2N 2N 4N 2N 4N

ξ 2N − 1 ξ 2N − 1
· · · + Fk + A∗ + = 0. (9.66)
2N 4N 2N 4N
ξ ξ ξ ∗
Let Fk 2N = F , F ξ + 1 , . . . , Fk 2N + 2N4N−1 ,
ξ ξ k 2N ξ k 2N1 4N ξ ∗
A1 2N = A 2N , A 2N + 4N , . . . , A 2N + 2N4N−1 ,
ξ ξ ξ ξ
∗
Bk 2N = Bk 2N , Bk 2N + 4N 1
, . . . , Bk 2N + 2N4N−1 . Then the
Eq. (9.61) implies that for any ξ ∈ R, the column vectors in 2N s × s matrix
A1 (ξ ) and the column vectors in 2N s × s matrix Bk (ξ ) are orthogonal for k =
0, 1, . . . , 2N − 1 and these vectors form an orthogonal basis of 2N s dimensional
complex Eucledian space C2N s .
The identity (9.66) implies that the column vectors in 2N s × s matrix Fk (ξ ) and
the column vectors in A1 (ξ ) are orthogonal. Thus there exists an s × s matrix Pk (ξ )
such that
Fk (ξ ) = Pk (ξ )Bk (ξ ), ξ ∈ R, k = 0, 1, . . . , 2N − 1.
Therefore, from Eqs. (9.57) and (9.65), we have

ξ ξ
f̂k (ξ ) = Fk ϕ̂
2N 2N

ξ ξ ξ
= Fk Bk ϕ̂
2N 2N 2N

ξ
= Pk Ψ̂ k (ξ ).
2N
By using the orthonormality of {Ψ k (t − λ) : λ ∈ Λ}, we have

∗
f̂k (2N ξ )f̂k∗ (2N ξ ) dξ = Pk (ξ )Ψ̂ k (2N ξ )Ψ̂ k (2N ξ )Pk∗ (ξ ) dξ.
R R
Therefore, we have
1/2
f̂k (2N ξ )f̂k∗ (2N ξ ) dξ = 2 Pk (ξ )Pk∗ (ξ ) dξ.
R 0
This shows that Fk (ξ ) has the Fourier series expansion and let the constant s × s
matrices {Q λ,k }λ∈Λ,k=0,1,...,2N −1 be its Fourier coefficients. Therefore,

fk (t) = Q λ,k Ψ k (t − λ).
λ∈Λ
It proves the completeness of {Ψ k (t − λ)}λ∈Λ, k=0,1,...,2N −1 in W0 .
If Ψ 0 , Ψ 1 , . . . , Ψ 2N −1 ∈ V1 are as in Theorem 9.5, one can obtain from them an

orthonormal basis for L 2 (R, Cs ) by following the standard procedure for construc-
tion of wavelets from a given MRA. It can be easily checked that for every m ∈ Z,
the collection {(2N )m/2 Ψ k ((2N )m x − λ)}λ∈Λ,k=0,1,...,2N −1 is a complete orthonor-
mal system for Vm+1 .
Given a VNUMRA, since Wm the orthogonal complement of Vm in Vm+1 , m ∈ Z
and
L 2 (R, Cs ) = ⊕m∈Z Wm
where ⊕ denotes the orthogonal direct sum with the inner product of L 2 (R, Cs ).
From this, it follows immediately that the collection {(2N )m/2 Ψ k ((2N )m x −
λ)}λ∈Λ,m∈Z,k=0,1,2,...,2N −1 forms a complete orthonormal system for L 2 (R, Cs ). When
N = 1, we recover the usual construction of vector-valued wavelets from vector-
valued multiresolution analysis.
9.5.1 Construction of Vector-Valued Nonuniform

Multiresolution Analysis
We are going to construct VNUMRA starting from a matrix-valued function A(ξ )

which is of the form
A(ξ ) = A1λ + e−2πir/N ξ A2λ , (9.67)
where the integers r and N satisfy N ≥ 1, 1 ≤ r ≤ 2N − 1, r is odd, r and N are

relatively prime and A1λ and A2λ are s × s constant symmetric matrix sequences.
The scaling function ϕ associated with the given VNUMRA should satisfy the scaling
relation
ϕ̂(2N ξ ) = A(ξ )ϕ̂(ξ ), ξ ∈ R. (9.68)
Also assume that the following condition holds:

ξ ∗ ξ ξ 1 ∗ ξ 1
A A +A + A + + ···
2N 2N 2N 4N 2N 4N

ξ 2N − 1 ξ 2N − 1
··· + A + A∗ + = Is .
2N 4N 2N 4N
,i.e.,

2N −1
ξ j ξ j
A + A∗ + = Is . (9.69)
j=0
2N 4N 2N 4N
!
Also assume that A(0) = Is in order for the infinite product ∞ ξ
k=1 A (2N )k to con-
verge pointwise.
The remaining part of this section will be devoted to the construction of a
VNUMRA starting from a matrix function A(ξ ) of the form (9.67) that satisfies
(9.69) as well as A(0) = Is .
We obtain a matrix-valued function ϕ which satisfies the scaling relation
ϕ̂(2N ξ ) = A(ξ )ϕ̂(ξ ). (9.70)
At this point, it is necessary to determine whether the collection of functions

{ϕ(x − λ)}λ∈Λ where Λ = {0, r/N } + 2Z is an orthonormal system in L 2 (R, Cs ).
Theorem 9.6 If A(ξ ) is of the form (9.68) which satisfies A(0) = Is together with
the conditions (9.69), Aλ = A∗λ for λ ∈ Λ (or A(ξ ) = A∗ (ξ ) ∀ ξ ∈ R). There exists a
constant c > 0 and a compact set K ⊂ R that contains a neighborhood of the origin
and
ξ
A ≥ C ∀ξ ∈ K ∀ k ≥ 1 (9.71)
(2N ) k
then the solution ϕ(t) in the matrix dilation equation (9.48) is a vector-valued scal-
ing function for VNUMRA. Thus the corresponding ϕ j,k (t), j ∈ Z, λ ∈ Λ form an
orthonormal basis of L 2 (R, Cs ).
Proof In order to prove this proposition, we need only to prove the orthonormality
of {ϕ(t − λ) : λ ∈ Λ}.
Since Aλ = A∗λ for λ ∈ Λ, therefore, we have A∗ (ξ ) = A(−ξ ) and ϕ̂ ∗ (ξ ) =
ϕ̂(−ξ ) ∀ξ .
For any integer k ≥ 1, define
⎡ ⎤

k
⎣ ξ ⎦ ξ
μk (ξ ) = A χK .
j=1
(2N ) j (2N )k
Since 0 belongs to interior of K by assumption, therefore μk → ϕ̂ pointwise as

k → ∞.
There is also a constant B > 0 such that
|A(ξ ) − A(0)| ≤ B|ξ | ∀ ξ ∈ R
and thus |A(ξ )| ≥ 1 − B|ξ | ∀ ξ .

Since K is bounded, we can find an integer k0 ≥ 1 such that B
(2N )k
|ξ | < 1
2
if ξ ∈ K
and k > k0 .
There exists a constant C > 0 such that
χ K (ξ ) ≤ C|ϕ̂(ξ )| ∀ ξ ∈ R.
Hence
⎡ ⎤
k "
" " "
" ξ " " ξ "
|μk (ξ )| ≤ C ⎣ " " ⎦ " "
"A (2N )k " "ϕ̂ (2N )k "
j=1
= C ϕ̂(ξ ).
We will commute the result by induction on the integral

μ∗k (ξ )μk (ξ )e−2πiξ(λ−σ ) dξ, when λ, σ ∈ Λ.
R
If k = 1, then we have

ξ ξ ξ
μ∗1 (ξ )μ1 (ξ )e−2πiξ(λ−σ ) dξ = A∗ A χK e−2πiξ(λ−σ ) dξ
R R 2N 2N 2N

= 2N A∗ (ξ )A(ξ )e−2πi(2N )ξ(λ−σ ) dξ
K
1/2
= 4N A∗ (ξ )A(ξ )e−2πi(2N )ξ(λ−σ ) dξ
0
⎡ ⎤
1

2N −1
4N
⎣ 1 1
= 4N A∗ ξ + A ξ+ e−πi(λ−σ ) j ⎦
0 4N 4N
j=0
e−2πi(2N )ξ(λ−σ ) dξ,
If λ − σ ∈ 2Z then the expression in the brackets in the above integral is equal to Is

and thus

∗ −2πiξ(λ−σ )
μ1 (ξ )μ1 (ξ )e dξ = 4N Is e−4πi N ξ(λ−σ ) dξ
R [0,1/4N )

=2 Is e−2πiξ(λ−σ ) dξ
[0,1/2)
= δλ,σ Is .
On the other hand, if λ = 2m, σ = 2n + r

N
where m, n ∈ Z then the same expression
will vanish and the integral becomes

μ∗1 (ξ )μ1 (ξ )e−2πiξ(λ−σ ) dξ = 0.
R
Therefore, we conclude that

μ∗1 (ξ )μ1 (ξ )e−2πiξ(λ−σ ) dξ = δλ,σ Is . (9.72)
R
Now, if k ≥ 2, we have

μk (ξ )μ∗k (ξ )e−2πiξ(λ−σ ) dξ
R

ξ ξ ξ ∗ ξ ∗ ξ ∗ ξ
= A A . . . A A A . . . A
R 2N (2N )2 (2N )k (2N )k (2N )k−1 (2N )
e−2πiξ(λ−σ ) dξ

= (2N )k A((2N )k−1 ξ )A((2N )k−2 ξ ) . . . A(ξ )A∗ (ξ )A∗ ((2N )ξ ) . . . A∗ ((2N )k−1 ξ )
K
e−2πi(2N )
k ξ(λ−σ )
dξ

∗ ∗ ∗
= (2N ) k
A ((2N ) k−1
ξ )A ((2N ) k−2
ξ ) . . . A (ξ )A(ξ )A((2N )ξ ) . . . A((2N ) k−1
ξ)
K
e−2πi(2N )
k ξ(λ−σ )
dξ
⎡ ⎤∗ ⎡ ⎤

k−1
k−1
= (2N )k ⎣ A((2N ) j ξ )⎦ ⎣ A((2N ) j ξ )⎦ e−2πi(2N ) ξ(λ−σ ) dξ
k
K j=0 j=0
⎡ ⎤∗ ⎡ ⎤

k−1
k−1
= (2N ) k ⎣ A((2N ) ξ )⎦ A (ξ )A(ξ ) ⎣
j ∗
A((2N ) ξ )⎦ e−2πi(2N )
j k ξ(λ−σ )
dξ
K j=1 j=1
⎡ ⎤∗ ⎡ ⎤
1/2
k−1
k−1
= 2(2N ) k ⎣ A((2N ) ξ )⎦ A (ξ )A(ξ ) ⎣
j ∗
A((2N ) ξ )⎦ e−2πi(2N )
j k ξ(λ−σ )
dξ.
0 j=1 j=1
⎡ ⎤∗ ⎡ ⎤
p ∗ p p p ⎦
1

k−1
2N −1
4N
= 2(2N ) k ⎣ A((2N ) ξ )⎦ ⎣ j ∗
A 2N ξ + A ξ+ A ξ+ A 2N ξ +
0 2 2 2 2
j=2 p=0
⎡ ⎤

k−1
⎣ A((2N ) j ξ )⎦ e−2πi(2N ) ξ(λ−σ ) dξ
k
j=2
⎡ ⎤∗ ⎡ ⎤
1

k−1
k−1
4N
= 2(2N )k ⎣ A((2N ) j ξ )⎦ H(ξ ) ⎣ A((2N ) j ξ )⎦ e−2πi(2N )
k ξ(λ−σ )
dξ,
0 j=2 j=2
where
⎡ ⎤

2N −1 p ∗ p p p ⎦
H(ξ ) = ⎣ A∗ 2N ξ + A ξ+ A ξ+ A 2N ξ + .
p=0
2 2 2 2
Using (9.69), we have
−1
p p
2N
A1λ + e2πi N (2N ξ + 2 ) A2λ
∗ p ∗
A∗ ξ +
r
H(ξ ) = A ξ+
p=0
2 2

A1λ + e2πi N (2N ξ + 2 ) A2λ
r p
∗ ∗
= A1λ A1λ + A2λ A2λ
= A∗ (2N ξ )A(2N ξ ).
It follows that

μk (ξ )μ∗k (ξ )e−2πiξ(λ−σ ) dξ
R
⎡ ⎤∗ ⎡ ⎤

k−1
k−1
= (2N )k ⎣ A((2N ) j ξ )⎦ ⎣ A((2N ) j ξ )⎦ e−2πi(2N )k ξ(λ−σ ) dξ
K /2N j=1 j=1
⎡ ⎤∗ ⎡ ⎤

k−2
k−2
= (2N )k−1 ⎣ A((2N ) j ξ )⎦ ⎣ A((2N ) j ξ )⎦ e−2πi(2N )
k
ξ(λ−σ )
dξ
K j=0 j=0

= μk−1 (ξ )μ∗k−1 (ξ )e−2πiξ(λ−σ ) dξ,
R
which together with (9.72) shows that for any k ≥ 1,

μk (ξ )μ∗k (ξ )e−2πiξ(λ−σ ) dξ = δλ,σ λ, σ ∈ Λ.
R
On taking the limit k → ∞, we get

ϕ(x − λ)ϕ ∗ (x − σ ) d x = ϕ̂(ξ )ϕ̂ ∗ (ξ )e−2πiξ(λ−σ ) dξ = δλ,σ where λ, σ ∈ Λ,
R R
which proves the condition of orthonormality.
9.6 Vector-Valued Nonuniform Wavelet Packets

and Its Properties
We will define the vector-valued nonuniform wavelet packets(VNUWP) and inves-

tigate their properties. Let
(0) (k)
γ 0 (t) = φ(t), γ k (t) = Ψ k (t), Q λ = Aλ , Q λ = Bλ,k , λ ∈ Λ, k = 1, 2, . . . , 2N − 1.
Definition 9.9 The family of vector-valued nonuniform functions {γ 2N n+k (t), n ∈

Z+ , k = 0, 1, . . . , 2N − 1} is called a vector-valued nonuniform wavelet packet w.r.t
the orthogonal vector-valued scaling function γ 0 (t) where

γ 2N n+k (t) = Q (k)
λ γ (2N t − λ), k = 0, 1, . . . , 2N − 1.
n
(9.73)
λ∈Λ
By taking the Fourier transform both sides, we get
γ̂ 2N n+k (2N η) = Q (k) (η)γ̂ n (η), k = 0, 1, . . . , 2N − 1, (9.74)
where
Q (k) (η) = Q (k)
λ e
−2πiλη
, k = 0, 1, . . . , 2N − 1. (9.75)
λ∈Λ
Q (k) (η) = Rλ(k) + Sλ(k) e−2πir/N η (9.76)
where {Rλ(k) } and {Sλ(k) } are s × s constant symmetric matrix sequences.

Thus, we have
Q (0) (η) = A(η), Q (k) (η) = Bk (η). (9.77)
Then the Eq. (9.61) can be written as

1 1 ∗
Q (k) (η)Q (l) (η)∗ + Q (k) (η + )Q (l) (η + ) +
4N 4N
2 2 ∗ 2N − 1 (l) 2N − 1 ∗
Q (k) (η + )Q (l) (η + ) + · · · + Q (k) (η + )Q (η + )
4N 4N 4N 4N
= δk,l Is , η ∈ R, k, l ∈ 0, 1, 2, . . . , 2N − 1. (9.78)
The Eq. (9.63) can be written as

2N −1
α p wn (ξ + p/2) = 0 where α = e−πir/N , (9.79)
p=0

and wn (ξ ) = j∈Z γ̂ n (ξ + N j)γ̂ n (ξ + N j)∗ . Now we will investigate the proper-
ties of the vector-valued nonuniform wavelet packets.
Theorem 9.7 If {γ n (t), n ∈ Z+ } is a vector-valued nonuniform wavelet packet with

respect to orthogonal vector-valued nonuniform scaling function φ(t), then ∀ n ∈
Z+ , we have
< γ n (. − λ), γ n (. − σ ) >= δλ,σ Is , λ, σ ∈ Λ. (9.80)
Proof We will prove the result by induction on n.

If n = 0 then the Eq. (9.80) follows directly from hypothesis. # n $
Suppose 0 ≤ n < (2N )l for some integer l. Then, for some (2N )l−1 ≤ 2N <
# n $
2 where [x] denotes the greatest integer of x and order n = 2N 2N + k, k =
l
0, 1, 2, . . . , 2N − 1.
Therefore, by induction, we have
% [n] &
γ 2N (. − λ), γ [ 2N ] (. − σ ) = δλ,σ Is .
n
(9.81)
9.6 Vector-Valued Nonuniform Wavelet Packets and Its Properties 349
We obtain

% n &
γ (. − λ), γ n (. − σ ) = e−2πi(λ−σ )ξ γ̂ n (ξ )γ̂ n (ξ )∗ dξ
R

e−2πi(λ−σ )ξ γ̂ n (ξ + N j)γ̂ n (ξ + N j)∗ dξ,
[0,N ) j∈Z
Let
wn (ξ ) = γ̂ n (ξ + N j)γ̂ n (ξ + N j)∗ .
j∈Z
Therefore, by using (9.74) and (9.81), we obtain
wn (2N ξ )
n
= γ̂ (2N (ξ + j/2))γ̂ n (2N (ξ + j/2))∗
j∈Z
# n
$ # n
$
= Q (k) (ξ + j/2)γ̂ 2N ((ξ + j/2))γ̂ 2N ((ξ + j/2))∗ Q (k) (ξ + j/2)∗
j∈Z
# n
$ # n
$
= Q (k) (ξ + n N )γ̂ 2N (ξ + n N )γ̂ 2N (ξ + n N )∗ Q (k) (ξ + n N )∗
j=n.2N
# $ # $ ∗
1 n 1 n 1 1 ∗
+ Q (k) ξ + n N + γ̂ 2N ξ + n N + γ̂ 2N ξ + n N + Q (k) ξ + n N +
2 2 2 2
j=n.2N +1
# $ # $ ∗
2 n 2 n 2 2 ∗
+ Q (k) ξ + n N + γ̂ 2N ξ + n N + γ̂ 2N ξ + n N + Q (k) ξ + n N +
2 2 2 2
j=n.2N +2
...
# $
2N − 1 n 2N − 1
+ Q (k) ξ + n N + γ̂ 2N ξ + n N +
2 2
j=n.2N +(2N −1)

# n $ 2N − 1 ∗ (k) 2N − 1 ∗
γ̂ 2N ξ + n N + Q ξ + nN +
2 2
⎡ ⎤
# n
$ # n
$
= Q (k)
(ξ ) ⎣ γ̂ 2N (ξ + n N )γ̂ 2N (ξ + n N ) ∗⎦
Q (k) (ξ )∗
j=n.2N
⎡ ⎤
# n $ # $ ∗
(k) 1 ⎣ 1 n 1 ⎦ (k) 1 ∗
+Q ξ+ γ̂ 2N ξ + nN + γ̂ 2N ξ + nN + Q ξ+
2 2 2 2
j=n.2N +1
⎡ ⎤
# n $ # $
2 ⎣ 2 n 2 ∗ ⎦ (k) 2 ∗
+Q (k) ξ + γ̂ 2N ξ + n N + γ̂ 2N ξ + n N + Q ξ+
2 2 2 2
j=n.2N +2
+···
⎡
# $
(k) 2N − 1 ⎣ n 2N − 1
+Q ξ+ γ̂ 2N ξ + nN +
2 2
j=n.2N +(2N −1)
⎤
# n $
2N − 1 ∗ ⎦ (k)

2N − 1 ∗
γ̂ 2N ξ + n N + Q ξ+
2 2

1 1 ∗ 2 2 ∗
= 2 Q (k) (ξ )Q (k) (ξ )∗ + Q (k) ξ + Q (k) ξ + + Q (k) ξ + Q (k) ξ + + ···
2 2 2 2
∗
(2N − 1) (2N − 1)
· · · + Q (k) ξ + Q (k) ξ +
2 2

2N −1
=2 Q (k) (ξ + j/2)Q (k) (ξ + j/2)∗ .
j=0
Also

2N −1
wn (ξ + p/2) = γ̂ n (ξ + j/2)γ̂ n (ξ + j/2)∗
p=0 j∈Z

2N −1 ∗
ξ j ξ j
=2 Q (k) + Q (k) + .
j=0
2N 4N 2N 4N
If λ = 2m 1 , σ = 2m 2 , where m 1 , m 2 ∈ Z, by using (9.78) we have

% &
γ n (. − λ), γ n (. − σ )

= e−2πi(2m 1 −2m 2 )ξ γ̂ n (ξ )γ̂ n (ξ )∗ dξ
R

= e−4πi(m 1 −m 2 )ξ γ̂ n (ξ )γ̂ n (ξ )∗ dξ
R
n
= e−4πi(m 1 −m 2 )ξ γ̂ (ξ + N j)γ̂ n (ξ + N j)∗ dξ
[0,N ) j∈Z

= e−4πi(m 1 −m 2 )ξ wn (ξ ) dξ
[0,N )
⎡ ⎤

2N −1
= e−4πi(m 1 −m 2 )ξ ⎣ wn (ξ + p/2)⎦ dξ
[0,1/2) p=0

2N −1 ∗
ξ j ξ j
=2 e−4πi(m 1 −m 2 )ξ Q (k) + Q (k) + dξ
[0,1/2) j=0
2N 4N 2N 4N
= δm 1 ,m 2 Is
= δλ,σ Is .
When λ = 2m 1 , σ = 2m 2 + r/N where m 1 , m 2 ∈ Z, we obtain by using (9.79)
< γ n (. − λ), γ n (. − σ ) >


= e−4πi(m 1 −m 2 )ξ e−2πiξr/N wn (ξ ) dξ
[0,N )
⎛ ⎞

2N −1
= e−4πi(m 1 −m 2 )ξ e−2πiξr/N ⎝ e−πi pr/N wn (ξ + p/2)⎠ dξ
[0,1/2) p=0
= 0.
Theorem 9.8 For any n 1 , n 2 ∈ Z+ and λ, σ ∈ Λ, we have
< γ n 1 (. − λ), γ n 2 (. − σ ) >= δn 1 ,n 2 δλ,σ Is ,
where {γ n (t) : n ∈ Z+ } is VNUWP with respect to orthogonal vector-valued scaling

function φ(t).
Proof If n 1 = n 2 then the result follows by Theorem 9.5.

If n 1 = n 2 , without loss of generality we can assume that n 1 > n 2 .
Write 'n ( 'n (
1 2
n 1 = 2N + k, n 2 = 2N + l,
2N 2N
where k, l ∈
# n{0, # n.2.$. , 2N − 1}.
$ 1, 2,
Case (i) If 2N 1
= 2N , then k = l.

% n1 &
γ (. − λ), γ n 2 (. − σ ) = e−2πi(λ−σ )ξ γ̂ n 1 (ξ )γ̂ n 2 (ξ )∗ dξ
R

= e−2πi(λ−σ )ξ γ̂ n 1 (ξ + N j)γ̂ n 2 (ξ + N j)∗ dξ
[0,N ) j∈Z

= e−2πi(λ−σ )ξ v(ξ ) dξ,
[0,N )

where v(ξ ) = j∈Z γ̂ n 1 (ξ + N j)γ̂ n 2 (ξ + N j)∗ . Therefore,

v(2N ξ ) = γ̂ n 1 (2N (ξ + j/2))γ̂ n 2 (2N (ξ + j/2))∗
j∈Z
n1 n2
= Q (k) (ξ + j/2)γ̂ [ 2N ] (ξ + j/2)γ̂ [ 2N ] (ξ + j/2)∗ Q (l) (ξ + j/2)∗ .
j∈Z
On solving as same as in Theorem 9.5, we obtain

2N −1
v(2N ξ ) = 2 Q (k) (ξ + j/2)Q (l) (ξ + j/2)∗ .
j=0
Also

2N −1
v(ξ + p/2) = γ̂ n 1 (ξ + j/2)γ̂ n 2 (ξ + j/2)∗
p=0 j∈Z

2N −1 ∗
ξ j ξ j
=2 Q (k) + Q (l) + .
j=0
2N 4N 2N 4N
If λ = 2m 1 and σ = 2m 2 where m 1 , m 2 ∈ Z, we obtain

% n &
γ (. − λ), γ n (. − σ ) = e−4πi(m 1 −m 2 )ξ v(ξ ) dξ
[0,N )
⎡ ⎤

2N −1
= e−4πi(m 1 −m 2 )ξ ⎣ v(ξ + p/2)⎦ dξ
[0,1/2) p=0
−1
j ∗
2N
ξ j ξ
=2 e−4πi(m 1 −m 2 )ξ Q (k) + Q (l) + dξ
[0,1/2) 2N 4N 2N 4N
j=0
= δm 1 ,m 2 δk,l Is
= δλ,σ δk,l Is .
If λ = 2m 1 + r/N and σ = 2m 2 where m 1 , m 2 ∈ Z, we obtain by using (9.79)

% n &
γ (. − λ), γ n (. − σ ) = e−4πi(m 1 −m 2 )ξ e−2πiξr/N v(ξ ) dξ
[0,N )
⎡ ⎤

2N −1
= e −4πi(m 1 −m 2 )ξ −2πiξr/N
e ⎣ e −πi pr/N
v(ξ + p/2)⎦ dξ
[0,1/2) p=0
= 0.
'n ( 'n ( 'n ( 'n (

1 2 1 [n 1 /2N ] 2
Case (ii) If = then take = 2N + k1 and =
2N 2N 2N 2N 2N
[n 2 /2N ]
2N + l1 where k1 , l1 ∈ {0, 1, 2, . . . , 2N − 1}.
2N
'n ( 'n ( [n 1 /2N ]
1 2
Let = 2N p1 + k1 and = 2N q1 + l1 where p1 = and
2N 2N 2N
[n 2 /2N ]
q1 = .
2N
If p1 = q1 , then the result follows' pimmediately
( from Case (i). 'q (
1 1
If p1 = q1 , then take p1 = 2N + k2 = 2N p2 + k2 and q1 = 2N +
2N 2N
l2 = 2N q2 + l2 where k2 , l2 ∈ {0, 1, 2, . . . , 2N − 1}.
If p2 = q2 then the result follows from Case (i).
If p2 = q2 then apply the above procedure. After performing a finite num-

ber of steps, we have pm−1 = 2N pm + km and qm−1 = 2N qm + lm where km , lm ∈
{0, 1, 2, . . . , 2N − 1} and pm , qm ∈ {0, 1, 2, . . . , 2N − 1}.
Case I If pm = qm
Case II If pm = qm .
For Case I, the result follows from Case (i).
For Case II, we have
% n1 &
γ (. − λ), γ n 2 (. − σ )

= e−2πi(λ−σ )ξ γ̂ n 1 (ξ )γ̂ n 2 (ξ )∗ dξ
R

ξ n1 ξ n2 ξ ∗ (l1 ) ξ ∗
= e−2πi(λ−σ )ξ Q (k1 ) γ̂ [ 2N ] γ̂ [ 2N ] Q dξ
R 2N 2N 2N 2N

ξ ξ ξ ξ
= e−2πi(λ−σ )ξ Q (k1 ) Q (k2 ) . . . Q (km )
γ̂ pm
R 2N (2N )2 (2N )m (2N )m
∗ ∗ ∗ ∗
ξ (lm ) ξ (l2 ) ξ (l1 ) ξ
γ̂ qm Q . . . Q Q dξ
(2N )m (2N )m (2N )2 (2N )
)m *
ξ ξ
−2πi(λ−σ )ξ (kn ) pm
= e Q γ̂
R n=1
(2N )n (2N )m
∗ ) m *∗
qm ξ (ln ) ξ
γ̂ Q dξ
(2N )m n=1
(2N )n
) ⎡

m *
−2πi(λ−σ )ξ (kn ) ξ ⎣ pm ξ
= e Q γ̂ + Nj
[0,N ) n=1
(2N )n j∈Z
(2N )m
∗ )
m *∗
qm ξ (ln ) ξ
γ̂ + Nj Q dξ
(2N )m n=1
(2N )n
= 0,
which completes the proof.
Corollary 9.1 If {γ n (t), n ∈ Z+ } is a vector-valued nonuniform wavelet packet with

respect to orthogonal vector-valued nonuniform scaling function φ(t), then ∀ n ∈
Z+ , and k, l ∈ {0, 1, . . . , 2N − 1}, we have
< γ 2N n+k (. − λ), γ 2N n+l (. − σ ) >= δλ,σ δk,l Is , λ, σ ∈ Λ. (9.82)
Proof We have
< γ 2N n+k (. − λ), γ 2N n+l (. − σ ) >

= e−2πi(λ−σ )ξ γ̂ 2N n+k (ξ )γ̂ 2N n+l (ξ )∗ dξ
R
2N n+k
= e−2πi(λ−σ )ξ γ̂ (ξ + N j)γ̂ 2N n+l (ξ + N j)∗ dξ.
[0,N ) j∈Z
Let
vn (ξ ) = γ̂ 2N n+k (ξ + N j)γ̂ 2N n+l (ξ + N j)∗ .
j∈Z
Therefore, on solving the equation as same as in Theorem 9.5, we have

vn (2N ξ ) = γ̂ 2N n+k (2N (ξ + j/2)) γ̂ 2N n+l (2N (ξ + j/2))∗
j∈Z

= Q (k) (ξ + j/2) γ̂ n (ξ + j/2) γ̂ n (ξ + j/2)∗ Q (l) (ξ + j/2)∗
j∈Z

2N −1
=2 Q (k) (ξ + j/2) Q (l) (ξ + j/2)∗ .
j=0
Also

2N −1
vn (ξ + p/2) = γ̂ 2N n+k (ξ + j/2)γ̂ 2N n+l (ξ + j/2)∗ .
p=0 j∈Z
Therefore, we have

2N −1 ∗
ξ j ξ j
γ̂ 2N n+k (ξ + j/2)γ̂ 2N n+l (ξ + j/2)∗ = 2 Q (k) + Q (l) + .
2N 4N 2N 4N
j∈Z j=0
When λ = 2m 1 and σ = 2m 2 , where m 1 , m 2 ∈ Z, by using (9.78) we obtain
< γ 2N n+k (. − λ), γ 2N n+l (. − σ ) >

= e−2πi(λ−σ )ξ vn (ξ ) dξ
[0,N )

= e−4πi(m 1 −m 2 )ξ vn (ξ ) dξ
[0,N )
⎡ ⎤

2N −1
= e−4πi(m 1 −m 2 )ξ ⎣ vn (ξ + p/2)⎦ dξ
[0,N ) p=0

2N −1 ∗
ξ j ξ j
=2 e−4πi(m 1 −m 2 )ξ Q (k) + Q (l) + dξ
[0,1/2) j=0
2N 4N 2N 4N

=2 e−4πi(m 1 −m 2 )ξ δk,l Is dξ
[0,1/2)
= δm 1 ,m 2 δk,l Is = δλ,σ δk,l Is .
When λ = 2m 1 , σ = 2m 2 + r/N where m 1 , m 2 ∈ Z, we obtain by using (9.79)
< γ 2N n+k (. − λ), γ 2N n+l (. − σ ) >

= e−4πi(m 1 −m 2 )ξ e−2πiξr/N vn (ξ ) dξ
[0,N )
⎛ ⎞

2N −1
= e−4πi(m 1 −m 2 )ξ e−2πiξr/N ⎝ e−πi pr/N vn (ξ + p/2)⎠ dξ
[0,1/2) p=0
= 0,
9.6.1 Vector-Valued Nonuniform Wavelet Bases
In this subsection, we shall construct orthogonal vector-valued nonuniform wavelet

bases of L 2 (R, Cs ) by using orthogonal vector-valued nonuniform wavelet packets.
Define a dilation operator by (DF)(t) = F(2N t) for F ∈ L 2 (R, Cs ).
∀Ω ⊂ L 2 (R, Cs ) and ∀ n ∈ Z+ , denote DΩ = {D F : F ∈ Ω} and

Ωn = {F(t) : F(t) = Mλ γ n (t − λ), {Mλ }λ∈Λ ∈ l 2 (Λ)s×s }. (9.83)
λ∈Λ
Then Ω0 = V0 and Ω1 ⊕ Ω2 ⊕ Ω3 ⊕ · · · ⊕ Ω2N −1 = W0 , where ⊕ denotes the

orthogonal direct sum.
Lemma 9.5 If {γ n (t), n = 0, 1, 2, . . .} is vector-valued wavelet packet with respect
to the orthogonal vector-valued scaling function γ 0 (t), then for every n ∈ Z+ ,

2N −1
2N γ n (2N t − k) = (Q (l) ∗ 2N n+l
k−2N λ ) γ (t − λ), k ∈ Λ.
l=0 λ∈Λ
Proof Consider
2N −1
1 (l)
(Q )∗ γ 2N n+l (t − λ)
2N l=0 λ∈Λ k−2N λ
2N −1
1 (l) (l)
= (Q k−2N λ )∗ Q λ1 γ n (2N t − 2N λ − λ1 )
2N l=0 λ∈Λ λ ∈Λ 1
)2N −1 *
1 (l)
∗ (l)
= (Q k−2N λ ) (Q λ2 −2N λ ) γ n (2N t − λ2 )
2N λ ∈Λ l=0 λ∈Λ
2
1
= δk,λ2 Is γ n (2N t − λ2 )
2N λ ∈Λ
2
= γ n (2N t − k),

Lemma 9.6 For n ∈ Z+ , the space DΩn can be orthogonally decomposed into the
spaces Ω2N n+k , k = 0, 1, 2, . . . , 2N − 1, i.e.,
DΩn = Ω2n ⊕ Ω2n+1 ⊕ · · · ⊕ Ω2n+(2N −1) .
Proof By Eqs. (9.73) and (9.83), we have
Ω2n ⊕ Ω2n+1 ⊕ · · · ⊕ Ω2n+(2N −1) ⊂ Ωn ∀n ∈ Z+ .
But Ω2n , Ω2n+1 , . . . , Ω2n+(2N −1) are orthogonal to each other by Theorem 9.6. By
Lemma 9.5, we have

2N −1
2N γ (2N t − k) =
n
(Q (l) ∗ 2N n+l
k−2N λ ) γ (t − λ), k ∈ Λ.
l=0 λ∈Λ
Therefore, the basis of the space Ωn can be linearly represented by the basis of the
space Ω2n , Ω2n+1 , . . . , Ω2n+(2N −1) .
Thus
Ωn ⊂ Ω2n ⊕ Ω2n+1 ⊕ · · · ⊕ Ω2n+(2N −1) .
Therefore
DΩn = Ω2n ⊕ Ω2n+1 ⊕ · · · ⊕ Ω2n+(2N −1) ,
i.e.,
−1
DΩn = ⊕2N
k=0 Ω2n+k .

Let X, Y ⊂ R and a X = {ax : x ∈ X } for a ∈ R.
X + Y = {x + y : x ∈ X and y ∈ Y }, X − Y = {x − y : x ∈ X and y ∈ Y }.
For a fixed positive integer r denote

r
Ũr = (2N )i {0, 1, 2, . . . , 2N − 1} and Ur = Ũr \Ũr −1 .
i=0
Theorem 9.9 The set of vector-valued functions {γ n (. − λ), n ∈ Ur , λ ∈ Λ} forms

an orthonormal basis of Dr W0 . In particular, the collection {γ n (. − λ), n ∈ Z+ , λ ∈
Λ} constitute an orthonormal basis of L 2 (R, Cs ).
Proof By Lemma 9.6, we have
−1
DΩ0 = Ω0 ⊕ Ω1 ⊕ · · · ⊕ Ω2N −1 = ⊕2N
k=0 Ωk ,
i.e.,
DV0 = V0 ⊕ W0 .
By using induction and by using Theorem 9.6 and Lemma 9.6, it can be proved that
Dr V0 = ⊕k∈Ũr Ωk and Dr V0 ⊕ Dr W0 = Dr +1 V0 ,
i.e.,
Dr W0 = ⊕k∈Ur Ωk .
Therefore, the set {γ n (. − λ), n ∈ Ur , λ ∈ Λ} forms an orthonormal basis of Dr W0 .

Moreover,

L 2 (R, Cs ) = V0 ⊕ ⊕0≤r Dr W0

= Ω0 ⊕ ⊕0≤r ⊕k∈Ur Ωk = ⊕k∈Z+ Ωk .
Therefore, the set {γ n (. − λ), n ∈ Z+ , λ ∈ Λ} forms an orthonormal basis of

L 2 (R, Cs ).
Theorem 9.10 For each r ∈ Z+ \{0}, the family of vector-valued functions
{γ n ((2N ) j t − λ), n ∈ Ur , j ∈ Z, λ ∈ Λ} forms an orthonormal basis of L 2 (R, Cs ).
Proof By Theorem 9.7, {γ n (. − λ), n ∈ Ur , λ ∈ Λ} forms an orthonormal basis of
Dr W0 . Then for each j ∈ Z, {γ n ((2N ) j t. − λ), n ∈ Ur , λ ∈ Λ} forms an orthonor-
mal basis of Dr W0 . Hence, for each r ∈ Z+ \{0},
⊕ j∈Z D j Dr W0 = ⊕ j∈Z D j+r W0 = ⊕ j∈Z D j W0 .
Thus the family {γ n ((2N ) j t − λ), n ∈ Ur , j ∈ Z, λ ∈ Λ} forms an orthonormal

basis of L 2 (R, Cs ).
9.7 Exercise
• Prove Theorems 9.1 and 9.2.

• Prove the necessary and sufficient condition for the existence of associated
wavelets with VMRA on R+ .
• Prove that the system of wavelets Ψ0 , Ψ1 , . . . , Ψ2N −1 in V1 as defined in Theo-

rem 9.5 forms an orthonormal basis for L 2 (R, Cs ).
• Construct vector-valued nonuniform wavelets associated with φ(t) ∈ L 2 (R, C2 )
with supp φ(t) = [0, 2].
References
1. Farkov, Yu. A. (2005). Orthogonal p-wavelets on R + . In Proceedings of the International

Conference on Wavelets and Splines (p. 426). Saint Petersburg: St. Petersburg University Press.
(Saint Petersburg, Russia, July 3–8).
2. Lang, W. C. (1996). Orthogonal wavelets on the cantor dyadic group. SIAM Journal on Math-
ematical Analysis, 27(1), 305–312.
3. Lang, W. C. (1998). Fractal multiwavelets related to the cantor dyadic group. International
4. Lang, W. C. (1998). Wavelet analysis on the cantor dyadic group. Houston Journal of Mathe-
matics, 24, 533–544.
5. Farkov, Yu. A. (2005). Orthogonal Wavelets with compact support on locally compact abelian
groups. Izv. Ross. Akad. Nauk Ser. Mat. 69(3), 193–220. (English Translate, Izvestia: Mathe-
matics, 69(3), 623–650.)
6. Protasov, V. Yu., & Farkov, Yu. A. (2006). Dyadic Wavelet and refinable Functions on a half
line. Sboinik: Mathematics, 197(10), 1529–1558.
7. Farkov, Yu. A. Maksimov A. Yu. & Stroganov S. A. (2011). On biorthogonal wavelets related
to the walsh functions. International Journal of Wavelets Multiresolution and Information
Processing, 9(3), 485–499.
8. Manchanda, P., Meenakshi, M., & Siddiqi, A. H. (2014). Construction of vector-valued nonuni-
form wavelets and wavelet packets. Special Issue of Indian Journal of Industrial and Applied
Mathematics.
9. Meenakshi, M., Manchanda, P., & Siddiqi, A. H. (2014). Wavelets associated with vector valued
non-uniform multiresolution analysis. Applicable Analysis: An International Journal, 93(1),
84–104.
10. Meenakshi, M., & Manchanda, P. (2017). Vector-valued nonuniform wavelet packets. Numer-
ical Functional Analysis and Optimization, Taylor and Francis. Published Online, https://doi.
org/10.1080/01630563.2017.1355814
11. Manchanda, P., & Sharma V. (2012). Orthogonal vector valued wavelets on R+. International
Journal of Pure and Applied Mathematics, 75(4), 493–510.
12. Xia, X. G., & Suter, B. W. (1996). Vector-valued wavelets and vector filter banks. IEEE Trans-
actions on Signal Processing, 44(3), 508–518.
13. Chen, Q., & Cheng, Z. (2007). A study of compactly supported orthogonal vector-valued
wavelets and wavelet packets. Chaos, Solitons and Fractals, 31(4), 1024–1034.
Appendix
A.1 Basic Results of Functional Analysis
Definition A.1 1. Let X be a set, and d a real function on X × X , with the prop-
erties
(a) d (x, y) ≥ 0 ∀ x, y ∈ X , d (x, y) = 0 if and only if x = y.
(b) d (x, y) = d (y, x) ∀ x, y ∈ X (symmetry property).
(c) d (x, y) ≤ d (x, z) + d (z, y) (the triangle inequality).
Then d is called a metric, or a metric function on or over or in X . (X , d ) is called
a metric space.
2. Sr (x) = {y/d (x, y) < r} is called an open sphere with radius r and center x in
X.
3. Let A and B be subsets of a metric space (X , d ). Then
(a) d (A, B) = supa∈A inf b∈B d (a, b).
(b) The diameter of A, which is denoted by δ(A), is the number supa,b∈A d (a, b).
(c) S(A, ε) = {x/d (A, x) < ε is called the ε− neighborhood of A.
(d) A is bounded if δ(A) < ∞.
4. A sequence in a metric space is a function whose domain is the set of positive
integers and range is a subset of the metric space.
5. A sequence {An } of subsets of a metric space is called a decreasing sequence if
A1 ⊇ A2 ⊇ A3 ⊇ . . ..
Theorem A.1 1. The collection of open spheres in a metric space (X , d ) forms a

base for the metric topology.
2. Every metric space is normal and hence a Hausdorff topological space.
Definition A.2 1. A sequence {xn } in a metric space (X , d ) is said to be convergent

to an element x ∈ X if limn→∞ d (xn , x) = 0. {xn } is called a Cauchy sequence
if limm,n→∞ d (xm , xn ) = 0.
https://doi.org/10.1007/978-981-13-6370-2
360 Appendix
2. A metric space is called complete if every Cauchy sequence in it is convergent

to an element of this space.
3. Let (X , d1 ) and (Y , d2 ) be metric spaces, and f is a mapping of X into Y . f
is called continuous at a point x0 in X if either of the following equivalent
conditions is satisfied:
(a) For every ε > 0, there exists a δ > 0 such that d1 (x, x0 ) < δ implies
d2 (f (x), f (x0 )) < ε.
(b) For each open sphere Sε (f (x0 )) centered at f (x0 ), there exists an open sphere
Sδ (x0 ) centered at x0 such that f (Sδ (x0 )) ⊆ Sε (f (x0 )). f is called continuous
if it is continuous at each point of its domain.
4. In (3), it can be seen that the choice of δ depends not only on ε but also on the
point x0 . The concept of continuity, in which for each ε > 0, a δ > 0 can be
found which works uniformly over the entire metric space X , is called uniform
continuity.
Theorem A.2 1. If a convergent sequence in a metric space has infinitely many

distinct points, then its limit is a limit point of the set of points of the sequence.
2. Let X and Y be metric spaces and f is a mapping of X into Y . Then f is continuous
if and only if {xn } in X converges to x ∈ X implies that f (xn ) in Y converges to
f (x) ∈ Y .
3. A point x is in the closure of a set A in a metric space if and only if there is a
sequence {xn } of points of A converging to x.
Definition A.3 A norm on a vector space X denoted by ||.||is a function defined on

X satisfying the following conditions:
(i)
x
> 0;
x
= 0 if and only if x = 0;
(ii)
αx
= |α|
x
, α is an arbitrary scalar.
(iii)
x + y
≤
x
+
y
for all x, y ∈ X .
Every normed space (X ,
.
) is a metric space with d (x, y) =
x − y
. A complete
normed space is called a Banach space.
The vector space Rn of all n-tuples x = (x1 , x2 , . . . , xn ) of real numbers is a normed
space, infact Banach space with respect to norms

(i)
x
1 = nk=1 |xk |,
(ii)
x
2 = (nk=1 |xk |2 )1/2 ,
(iii)
x
3 = ( nk=1 |xk |p )1/p , where 1 ≤ p < ∞, and
(iv)
x
4 = max{|x1 |, |x2 |, . . . , |xk |}..
Spaces
∞ The set of all sequences x = {x1 , x2 , . . . , xn , . . .} of real numbers such that
n=1 |xn | p
< ∞, for 1 ≤ p < ∞ denoted by lp , in the case p = 2, it is called the
infinite-dimensional Euclidean space.
For a, b ∈ R and a < b, [a, b] = {x ∈ R/a ≤ x ≤ b} (a, b) = {x ∈ R/a < x <
b} and (a, b] = {x ∈ R/a < x ≤ b} are called, respectively the closed, open, semi-
closed (semi-open) intervals of R. R is a metric space with the metric d (x, y) =
Appendix 361
|x − y| and therefore a closed and bounded subset of R can be defined. As a special

case, a real- valued function f (x) on [a, b] is called bounded if there exists a real
number k such that |f (x)| ≤ k ∀ x ∈ [a, b]. It is said to be continuous at x0 ∈ [a, b]
if, for ε > 0, there exists δ > 0 such that |f (x) − f (x0 ))| < ε whenever |x − x0 | < δ.
It is called uniformly continuous on [a, b] if the choice of δ does not depend on
the point x0 in the definition of continuity. We can define a continuous function in a
similar fashion on an arbitrary subset of R. C[a, b] denotes the set of all continuous
real functions defined on [a, b]. It is called the space of continuous functions on
[a, b]. Similarly, C(T ), where T ⊆ R, called the space of continuous functions on
T . If X is an arbitrary topological space, then C(X ) is called the space of continuous
real functions on the topological space X .
p(x) = a0 + a1 x + . . . + an xn for [a, b], where a0 , a1 , . . . , an are real numbers,
is called a polynomial of degree n. It is a continuous real function. P[a, b] denotes
the set of all polynomials over [a, b].Pn [a, b] denotes the set of all polynomials of
degree less than or equal to n over [a, b]. Cn [0, π ] denotes the set of all functions on
[0, π ] of the form
f (x) = b1 cos x + b2 cos 2x + . . . + bn cos nx.
Let f (x) be a real-valued function defined on [a, b]. The limits
f (y) − f (x)
f− (x) = lim
y→x,y<x y−x
and
f (y) − f (x)
f+ (x) = lim
y→x,y>x y−x
are called left and right derivatives, respectively. The function f is said to be differen-
tiable if the right and left derivatives at x exist and their values are equal. This value is
called the first derivative of f at x and is denoted byf or f 1 or df
dx
. If the first derivative
exists at every point of [a, b], then it is said to be differentiable over [a, b]. If the first
derivative of f is also differentiable, then f is called twice differentiable. In general,
the nth derivative of the function f is the derivative of the (n − 1)th derivative. f
is called n times differentiable over [a, b] if nth derivative exists at every point of
[a, b]. C (n) [a, b] denotes the set of all functions which have continuous derivatives
upto and including the nth order over [a, b]. If derivatives of all order on [a, b] exist,
then f is called infinitely differentiable. C ∞ [a, b] denotes the class of all infinitely
differentiable functions over [a, b].
Let f (x) be a real-valued function
[a, b]. Let P : x0 ≤ x1 ≤ x2 ≤ . . . , ≤
defined on
xn = b be a partition of [a, b], ba (f ) = supP ni=0 |f (xi ) − f (xi−1 )| is a variation of

f (x) over [a, b]. f (x) is called a function of bounded variation on [a, b] if ba (x) < ∞.
BV [a, b] denotes the space of all functions of bounded variation over [a, b].
A real-valued function f (x) on [a, b] is called absolutely continuous on [a, b] if,
for ε > 0, there exists a δ > 0 such that for any collection {(ai , bi )}n1 of disjoint open
362 Appendix
∞
subintervals of [a, b], ∞ i=1 |f (bi ) − f (ai )| < ε holds whenever i=1 (bi − ai ) <
δ.AC[a, b] denotes the class of all absolutely continuous function on [a, b].
A real function f (x) defined on [a, b] is said to satisfy a Hölder condition of
exponent α over [a, b] or to be Hölder continuous or Lipschitz class if
|f (x) − f (y)|
sup <∞
x,y∈[a,b] |x − y|α
The class of Hölder continuous or Lipschitz class on [a, b] denoted by Ca [a, b] or

Lipα [a, b].
We write
|f (x)|
f (x) = O(g(x)) if there exists a constant K > 0 such that |g(x)| ≤ K ∀ x.
f (x)
f (x) = o(g(x)) if g(x) → 0.
These relationships are valid when x → ∞, for x → −∞ or x → x0 , where x0 is
some fixed number. If bn > 0, n = 0, 1, 2, . . . and abnn → 0 as n → ∞, then we write
an = o(bn ). If abnn is bounded, then we write an = O(bn ).
Remark A.1 1. a. Every function of the Lipschitz class over [a, b] belongs to
AC[a, b].
b. AC[a, b] ⊂ C[a, b]
c. AC[a, b] ⊂ BV [a, b]
2. All continuous differentiable functions over [a, b] are absolutely continuously
over [a, b].
Examples
(i) lp , 1 ≤ p < ∞ is Banach space with respect to norm
∞ 1/p

x
p = |xk | p
k=1
is a Banach space
(ii) If A is any measurable subset of Rn , in particular Rn itself, then Lp (A) with
1 ≤ p < ∞ of all (classes of equal a.e.) measurable functions such that
1/p

f
p = |f (x)|p dx <∞
A
is a Banach space
Lp (a, b), 1 ≤ p < ∞ is a Banach space. Often Lp (a, b) is written as Lp (a, b).
Definition A.4 Let the σ − algebra generated by the dyadic intervals of length 2−n be
denoted by An . A sequence f = (fn ) of integrable functions is a dyadic martingale if
Appendix 363
(i) fn in An is measurable,
(ii) S2n fn+1 = fn (n ∈ N)
It follows from the definition of martingales that fˆn (k) = fˆm (k) for any k < 2n , 2m .
Therefore the definition fˆ (k) = fˆn (k)(k ∈ N, k < 2n ) makes sense. This means that
the concept of Walsh–Fourier partial sums, and Walsh–Fourier series can be extended
to martingales.
In order to define dyadic Hardy spaces H p (0 < p < ∞) we introduce the maximal
function of a martingale:
f ∗ = supn∈N |fn |.
Then H p is the collection of dyadic martingales f for which f ∗ is in Lp and the norm
is defined by
1 1/p
∗ p

f
H p = (f ) (0 < p < ∞).
0
1 if 0 ≤ x < 1/2
γ0 (x) =
−1 if 1/2 ≤ x < 1
periodic by 1, and
γk (x) = γ0 (2k x) (0 ≤ x < 1, k ∈ N)
The Walsh
function can be represented as products of Rademacher functions. Namely,
if n = ∞ γ
k=0 k 2 k
(γ k = 0 or 1, n ∈ N) is the binary decomposition of n. Then the
nth Walsh function wn in the Paley enumeration is defined as
∞
wn = Πk=0 γknk .

Then Dn = n−1 k=0 wk is the nth Walsh-Dirichlet kernel. The Walsh–Fourier coeffi-
cients, partial sums, and series of an integrable function f are defined by

n−1 ∞

fˆ (k) = f , wk , Sn f = fˆ (k)wk , Sf = fˆ (k)wk
k=0 k=0
Intervals of the form [k2−n , (k + 1)2−n ) (k, n ∈ N, k < 2n ) are called dyadic inter-
vals. The dyadic expansion of an x ∈ [0, 1) is
∞

x= xk 2−(k+1) (xk = 0 or 1).
k=0
For the so-called dyadic rationals there are two expressions of this form. In this case
we take the one which terminates in 0s. The dyadic addition (+̇) is defined as follows:
364 Appendix
∞

x+̇y = xk 2−(k+1) (xk = 0 or 1).
k=0
Then by definition
wk (x+̇y) = wk (x)wk (y) (k ∈ N, 0 ≤ x, y < 1)
Let P denote the set of Walsh polynomials, and let τx f be the dyadic translate of f
by x ∈ [0, 1), i.e.,
τx f (t) = f (x+̇t) (t ∈ [0, 1)).
Definition A.5 The space X is called to be a (dyadic) homogeneous Banach space,

if
(i) P ⊂ X ⊂ L1 ,
(ii)
f
1 ≤
f
X ,
(iii) τx f ∈ X ,
τx f
X =
f
X (0 ≤ x < 1),
iv) P is dense in X.
Remark A.2
(i) There are several examples for homogeneous spaces such as the Lp (1 ≤ p < ∞)
spaces, the dyadic Hardy space H , dyadic V MO space, the space of dyadic
continuous functions CW .
(ii) If h ∈ L1 and f ∈ X then

h ∗ f
X ≤
h
1
f
X ,
where * stands for dyadic convolution

1
(h ∗ f )(x) = h(x+̇t)f (t) dt
0
(iii) The modulus of continuity in X is defined by
w(f , δ, x) = supμ<δ
f − τμ f
X (f ∈ X , 0 < δ ≤ 1).
Since
D2n = 2n χ[0,2−n ) (n ∈ N)
the 2n th partial sums plays a special role in Walsh analysis. This relation implies

S2n f
X ≤
f
X . Moreover, they enjoy the following approximation properties:
1
w(f , 2−n , X ) ≤
f − S2n f
X ≤ w(f , 2−n , X ),
2
Appendix 365
1
w(f , 2−n , X ) ≤ E2n (f , X ) ≤ w(f , 2−n , X ) (f ∈ X , n ∈ N).
2
En (f , X ) stands for the best approximation of f in X by Walsh polynomials of
order not greater than n.
(iv) We note that for p > 1, Hardy space H p coincides with the Lebesgue space Lp ,
H 1 is a homogeneous Banach space which is proper subspace of L1 , and for
p < 1 the H p spaces are not homogeneous Banach spaces.
(v) In several cases it is useful to have other characterizations of the H p norm.
One of them relies on the quadratic variation of martingales. For a martin-
gale f let the quadratic variation of f be defined as Qf = ( ∞ n=0 |dn f | )
2 1/2
,
where dn f = fn+1 − fn . Then f ∈ H if and only if Qf ∈ L (0 < p < ∞), and
p p

f
H p ≈
Qf
Lp [1].
(vi) The concept of modulus of continuity, the best approximation in H p can be
defined in the same way as in homogeneous Banach spaces. Moreover by using
quadratic variation it is easy to see that for any f ∈ H p , and n ∈ N we have

f − S2n f
H p ≈
Q(f − S2n f )
Lp = minp∈P2n
Q(f − p)
Lp ≈ E2n (f , H p ). (A.1)
Definition A.6 Let X be a vector space over the field of real or complex numbers.
A mapping denoted by < ., . > defined on X × X into the underlying field, is called
an inner product of any two elements x and y of X if the following conditions are
satisfied:
1. < x + x , y >=< x, y > + < x , y >
2. < αx, y >= α < x, y >, α belongs to the underlying field
3. < x, y > =< y, x >
4. < x, x >≥ 0, ∀x ∈ X ; and < x, x >= 0 if and only if x = 0.
If the inner product < ., . > is defined for every pair of elements (x, y), x and y ∈ X ,
then the vector space X together with the inner product < ., > is called an inner
product space or pre-Hilbert space, usually denoted by (X , < ., . >) or simply X .
Pre-Hilbert space is a normed space. It is called a Hilbert space if it is complete.
Rn , l2 , L2 (a, b) are Hilbert spaces.
Definition A.7 Two normed spaces (X ,

.
1 ) and (Y ,
.
2 ) are said to be isometric
and isomorphic to one another if there exists a one-to-one mapping T of X onto Y
such that
d (Tu1 , Tu2 ) =
Tu1 − Tu2
2 =
u1 − u2
1 = d (u1 , u2 ) or
Tx
2 =
x
1
and T is linear, that is,
T (x + y) = Tx + Ty, ∀x, y ∈ X
T (αx) = αTx, ∀x ∈ X , α ∈ R or C
366 Appendix
This definition means that there exists a mapping of X into Y which is one-one, onto,
norm-preserving and linear.
If such mapping exists, we write X = Y .
Definition A.8(a) Two normed spaces (X ,

.
1 ) and (X ,
.
2 ) are called topologi-
cally equivalent or two norms (
.
1 and
.
2 ) are called equivalent if there exist
constants k1 and k2 such that
k1
x
1 ≤
x
2 ≤ k2
x
1 .
(b) A normed space is called finite-dimensional if the underlying vector space is

finite-dimensional; otherwise it is infinite-dimensional.
Bounded and Unbounded Operators Definitions and Examples
Definition A.9 Let U and V be two normed spaces. Then

1. A mapping T from U into V is called an operator or a transformation. The value
of T at x ∈ U is denoted by T (x) or Tx.
2. T is called a linear operator or a linear transformation if the following conditions
are satisfied:
(a) T (x + y) = Tx + Ty, ∀ x, y ∈ U .
(b) T (αx) = αT (x), ∀ x ∈ U and real α.
3. The operator T is called bounded if there exists a real k > 0 such that
Tx
≤
k
x
∀ x ∈ U .
4. T is called continuous at a point x0 ∈ U if, given ε > 0, there exists a δ > 0,
depending upon ε and x0 such that
Tx − Tx0
< ε, whenever
x − x0
< ε. T
is called continuous on X if it is continuous at every point of X .
5. T is called uniformly continuous if, for ε > 0, there exists a δ > 0, independent of
x0 , such that for any x0 and x ∈ X with
x − x0
< δ, we have
Tx − Tx0
< ε.
6.
T
= sup {
Tx
/x = 0} is called the norm of the bounded operator T (For an

unbounded operator, the sup may not exist).
7. If Y = R, the normed space of real numbers, then T is called a functional and it
is usually denoted by F.
8. For the operator T , the set R = {Tx ∈ U/x ∈ V } and the set N = {x ∈ U/T (x)
= 0} are called the range and null spaces, respectively.
9. X ∗ = F : X → R, FlinearboundedonnlsX is called dual of X .
Banach–Alagolu Theorem
Theorem A.3 Suppose X is a normed space and X ∗ is its dual. Then the closed unit
sphere S1∗ = {f ∈ X ∗ /
f
≤ 1} is compact with respect to weak topology.
Appendix 367
Principle of Uniform Boundedness

Theorem A.4 Let X be a Banach space, Y a normed space and Ti a sequence of
bounded linear operators over X into Y such that {Ti (x)} is a bounded subset of Y
for all x ∈ X . Then {
Ti
} is a bounded subset of real numbers, i.e., {Ti } is a bounded
sequence in the normed space B[X , Y ].
A.2 Vilenkin Systems
Topological group is a group G together with a topology on G such that the group’s
binary operation and the group’s inverse function are continuous with respect to the
topology.
A topological group is a mathematical object with an algebraic structure and a
topological structure.
In mathematical terms a topological group G, is a topological space, which is a
also a group such that the group operation of product:
G × G −→ G : (x, y) −→ xy
and taking inverses

G −→ G : x −→ x−1
are continuous. Here G × G is viewed as a topological space with the product topol-
ogy.
A topological group G is called compact if G is a compact topological space.
G is called abelian topological group if G is abelian group.
The Vilenkin systems were introduced in 1947 by Vilenkin. They include as a
special case the Walsh system.
We below give a brief introduction to the Vilenkin systems and relate them to other
types of orthonormal systems which have come up in the course of our narrative.
Let m = (mk , k ∈ N) be a sequence of natural numbers such that mk ≥ 2 (k ∈ N).
We shall construct a compact abelian group G m for each such sequence m. Let
Zmk (k ∈ N) be the mk th discrete cyclic group, i.e., Zmk can be represented by the
set {0, 1, 2, . . . , mk − 1}, where the group operation is mod mk addition, and every
subset is open. Haar measure on Zmk can be generated by insisting that the measure
of singleton is 1/mk . The group G m is defined as the complete discrete product of
the compact groups Zmk (k ∈ N). Thus on G m we use coordinate-wise addition as
the group operation, the product topology and the product measure. Thus G m is a
compact abelian group.
The group G m is metrizable. Let M0 = 1 and Mk = mk−1 Mk−1 (k ∈ P). Define
the distance between the elements (xk , k ∈ N) ∈ G m and (yk , k ∈ N) ∈ G m by
368 Appendix
∞
|xk − yk |
ρ(x, y) = .
Mk+1
k=0
It is obvious that the topology induced by this metric coincides with that of G m . It is
easy to give a base for the neighborhoods of G m
I0 (x) = G m
In (x) = {y = (yi , i ∈ N ∈ G m : yi = xi for i = 0, 1, . . . , n − 1)}
for x ∈ G m , n ∈ P. Moreover, if 0 denotes the null element of G m , then the topology

induced by the sets In (0) (n ∈ N) (the base of neighborhoods of 0) coincides with the
topology of G m . Furthermore, it can be shown that the product measure introduced
in G m is the normalized Haar measure of G m .
Clearly, if mk = 2 (k ∈ N) then G m coincides with the dyadic group. Thus no
confusion will arise if we use some of the same notation introduced in the study of
the dyadic groups for these Vilenkin groups G m .
Accordingly, for each n ∈ N set
2π ixn
ρn (x) = exp
mn
for x = (xk , k ∈ N) ∈ G m , n ∈ N. Clearly, each ρn is acharacter of G m . Infact, every

complex-valued function defined on G m whose values depend only on one coordinate
the element x is continuous on G m in the topology of G m .
To enumerate finite products of the functions ρn (n ∈ N), write each n ∈ N
uniquely with the help of the sequence (Mn , n ∈ N) in the form
∞

n= nk Mk (0 ≤ nk < mk , nk ∈ N). (A.2)
k=0
As in the dyadic case, it can be shown that the set Gˆm consists of the functions
∞

ψn = ρknk (n ∈ N),
k=0
where the sequence (nk , k ∈ N) has been defined in A1 . The system Gˆm = (ψn , n ∈
N) is called a Vilenkin system.
A group operation ⊕ corresponding to that introduced earlier in the set N can be
defined as follows:
∞

n⊕j = (nk + jk (mod mk ))Mk
k=0
Appendix 369
where
∞
∞

n= nk Mk and j = jk Mk .
k=0 k=0
It is clear, then, that ψn⊕j = ψn ψj (n, j ∈ N) and Gˆm isomorphic to N.
A.3 Vilenkin-Pontrjagin Class of Functions
Let G be a compact Abelian group satisfying the second axiom of countability. It

is well known that we can associate with G a countable set of functions which are
continuous on G and satisfy the equation φ(x + y) = φ(x)φ(y), (Pontrjagin, [2]).
These functions form a complete orthonormal system on G. Such a class of functions
was introduced by Pontrjagin and its Fourier expansion was studied by Vilenkin [3].
This system of functions includes Walsh functions as a special case when the Group
G is a topological direct sum of cyclic groups of order two. For the details of this
class one can see [1–3] .
A.3.1 The Dyadic Group G
The dyadic group G is the set of all sequences x̄ = {xn }, xn = 0, 1 for n = 1, 2, 3, . . .,

the operation in G, denoted by +∗ , is the addition modulo 2 in each coordinate.
Corresponding to each element x̄ = {xn } of G, there is a real number,
x1 x2 x3
λ(x̄) = + 2 + 2 + ... (A.3)
2 2 3
lying in the closed interval [0, 1]. At the dyadic rationals x, 0 < x < 1, we have two
representations in the dyadic scale and therefore, the mapping λ is not one-one but
it is onto. If μ(x) is the inverse of λ, for all real x, we have,
λ(μ(x)) = x − [x], . . . (A.4)
the finite expansion in G being associated under μ with the dyadic rationals. Thus
μ(λx̄) = x̄, provided λ(x̄) is not a dyadic rational.
Let
χn (x̄) = χn ({x1 , x3 , . . . , xn , . . .)

1 xn = 0
=
−1 xn = 1
370 Appendix
It is clear that χn is a character of G. Fine [4] has established that
φn−1 (x) = χn (μ(x)), n ≥ 1.
All characters of G may be obtained by taking finite products of these basic characters.
Thus each Walsh function can be expressed in the form
ψn (x) = χj1 (μ(x))χj2 (μ(x)) . . . χjt (μ(x)),
and therefore Walsh functions are identified with the whole set of characters of G.
Notation: Henceforth λ(μ(y) +∗ μ(z)) would be abbreviated to y+̇z, where y and

z are reals.
There is a natural topology of G which is obtained by taking the set of points
{x1 , x2 , . . . , xn , dn+1 , . . .}
as neighborhoods, where x1 , x2 , . . . , xn are fixed and dn+1 vary independently. These

neighborhoods, denoted by N (x1 , x2 , . . . , xn ), form a basis for the topology of G.
Proposition A.1 λ defines a metric on G. This metric is continuous on G, and the

topology induced on G by this metric is equivalent to the original topology of G.
Proposition A.2 |λ(ȳ) − λ(z̄)| ≤ λ(ȳ +z̄), ȳ, z̄ ∈ G.
Proposition A.3 For any two real numbers x and h

(a) |(x+̇h) − (x − [x])| ≤ h − [h],
(b) |(x+̇h) − x| ≤ h if 0 ≤ x < 1 and 0 ≤ h < 1.
Theorem A.5 (a) ψn (y + z) = ψn (y)ψn (z) where y + z is not a dyadic rational,

its exception is a denumerable event for each fixed x.
(b) Let A is ameasurable set, x be a fixed element and
Tx (A) = {(x+̇y), y ∈ A}
then the Lebesgue measure of Tx (A) is equal to the Lebesgue measure of A.

(c) if f is integrable then for every fixed x,
1 1
f (x+̇y) dy = f (y) dy.
0 0
The following definition forms a basis for discussing the relationship between two
given classes of functions one on G and another one on [0, 1]:
Appendix 371
Definition A.10 For each real-valued function g(x) of period 1, there corresponds
a function ḡ(x̄) on G defined by

g(x) μ(x) = x̄ for some x ∈ [0, 1]

ḡ(x̄) = (A.5)
limȳ→x̄ supḡ(ȳ) μ(x) = x̄ for any x ∈ [0, 1]
where the lim sup is taken over those ȳ’s which correspond to the dyadic rationals.
Characteristic functions of neighborhoods of G are continuous on G because each

neighborhood is both open and closed. Finite linear combinations of such character-
istic functions are then continuous on G.
Classes of functions:
The discussion of the Fourier properties of the Walsh system may be proceeded from
two points of view
(a) Characters χn (x) of G and their properties,
Haar measure of G,
Various classes of functions on G,
(b) Functions ψn (x) on [0, 1],
Lebesgue measure on [0, 1].
Definition A.11 A function f (x) of period 1 is said to Continuous (W) at a point x,

0 ≤ x < 1 if for every ε > 0, there exists δ > 0 depending upon ε and x such that
|f (x+̇y) − f (x)| < ε
for all y for which 0 ≤ y < δ.
Theorem A.6 Every continuous function is Continuous (W) but converse is not true.
Theorem A.7 If f (x) is a continuous (W) on 0 ≤ x < 1 and if f (x − 0) exists and

is finite at each dyadic rational x, 0 ≤ x < 1, then f¯ (x̄), the G-extension of f (x), is
continuous on G. If f¯ (x̄) is continuous on G, then the extension f (x) = f¯ (μ(x)) is
continuous (W) on 0 ≤ x < 1 and f (x − 0) exists and is finite at each dyadic rational
x, 0 ≤ x < 1.
Definition A.12 A function f¯ (x̄) is said to belong to the class of Lipschitz α on G,

0 ≤ α ≤ 1 if there exists a constant C > 0 such that
|f¯ (x̄) − f¯ (ȳ)| < Cλ(x̄ +ȳ)α , for all x, y ∈ G.
Definition A.13 A function f (x) defined on [0, 1) is said to belong to the class
Lipschitz α(W). 0 < α ≤ 1 if there is a constant C > 0 such that
|f (x+̇y) − f (x)| < Cyα ,

372 Appendix
for 0 ≤ x < 1, 0 ≤ y < 1 and μ(x+̇y) = μ(x) +μ(y), that is, except when μ(x)
+μ(y) ends in a sequence of 1’s.
The following theorem of Morgenthaler, see for example Siddiqi [5], gives the rela-
tionship between Lipschtz α functions on G and Lipschitz α(W) functions on [0, 1):
Theorem A.8 If f¯ (x̄) is a Lipschitz function on G, 0 < α ≤ 1, then f (x) = f¯ (μ(x))

is a Lipschitz α(W) function on [0, 1), and f (x0 − 0) exists, is finite at each dyadic
rational x, 0 ≤ x < 1. If f (x) is a Lipschitz α(W) function on [0, 1), 0 ≤ α < 1 and
f (x0 − 0) exists, is finite at each dyadic rational x, 0 ≤ x < 1, then the G-extension
of f , f¯ (x̄) is a Lipschitz function on G.
Remark A.3
(i) If f (x) is a function of class Lipschitz α on [0, 1) then f (x) is also a Lipschitz
α(W) function.
(ii) If f (x) is a function of class Lipschitz α(W) on [0, 1), then f (x) is continuous
(W) on [0, 1).
(iii) The G-extension of the Walsh functions are Lipschitz 1 functions on G.
(iv) The Walsh functions ψn (x) are Lipschitz α(W) functions on [0, 1).
(v) If we define
w̄(δ, f¯ ) = max |f¯ (x̄) − f¯ (ȳ)|
for all x̄, ȳ ∈ G, with λ(x +y) < δ,
then w̄(δ, f¯ ) ≤ Cδ α if f¯ belongs to the class Lipschitz α on G.

(vi) If
WW (δ, f ) = max0≤x<1,0≤y<δ |f (x+̇y) − f (x)|
μ(x) +μ(y) not ending in 1’s, then WW (δ, f ) < Cδ α if f belongs to the class
Lipschitz α(W).
A.4 Introduction to Real and Fourier Analysis

A series ∞ n=1 an
of real numbers is said to be convergent if the sequence sn of
partial sums sn = ni=1 ai converges.

Absolute Convergence:
∞ A series ofreal numbers ∞ n=1 an is said to converge
absolutely if n=1 |an | is finite (or if ∞
n=1 |an | is convergent).
Absolute convergence is important for the study of infinite series because it is a
stronger notion than that of convergence.
Pointwise Convergence: Let D be a subset of R and let an be a sequence of real-
valued functions defined on D. Then an converges pointwise to a if given any x in D
Appendix 373
and given any ε > 0, there exists a natural number N = N (x, ε) such that
|fn (x) − f (x)| < ε for every n > N .
Note: The notation N = N (x, ε) means that the number N depends on the choice of
x and ε.
Uniform Convergence: Let D be a sequence of real-valued functions defined on D.
Then fn converges uniformly to f if given any ε > 0, there exists a natural number
N = N (ε) such that
|fn (x) − f (x)| < ε for every n > N and for every x in D.
Note: In the above definition the number N depends on ε but not on x.
It is clear that Uniform Convergence implies pointwise convergence. But the
converse neednot be true.
A series ∞ n=1 fn (x) is called pointwise convergent
if the sequence < sn > of
(x) = f1 (x) + f2 (x) + ... + fn (x) or ni=1 fi (x) is pointwise convergent.
partial sums sn
The series ∞ n=1 fn (x) is a called uniformly convergent if the sequence of partial
sums is uniformly convergent.
Convergence almost everywhere: A sequence of function fn n∈N is said to con-
verge almost everywhere on D to f if μ{x ∈ D|fn (x) does not converge to f (x)}=0
where μ is measure of the set namely there exists a set E with μ(E) = 0 such that
limn→∞ fn (x) = f (x) for all x ∈ D \ E.
In case of Lebsegue measure, the measure of an interval will be length of the
interval. See for example Appendix of [5]. Let f (t) be a periodic function with period
T and Lebesgue integrable continuous function having at most finite discontinuous
over (−T /2, T /2). Then the fourier series of f (t) is the trigonometric series
∞

1 2π k 2π k
A0 + Ak cos x + Bk sin x (A.6)
2 T T
k=1
where
T /2
2 2π kt
Ak = f (t) cos dt, k = 1, 2, 3, . . . (A.7)
T −T /2 T
T /2
1
A0 = f (t)dt (A.8)
T −T /2
T /2
2 2π kt
Bk = f (t) sin dt, k = 1, 2, 3, . . . (A.9)
T −T /2 T
and we write it as
∞

1 2π k 2π k
f ∼ A0 + Ak cos x + Bk sin x (A.10)
2 T T
k=1
Here, we take
374 Appendix
k
wk = , k = 0, 1, 2, 3, . . . (A.11)
T
Very often, we choose T = 2π, Ak and Bk are called cosine Fourier coefficient and
sine Fourier coefficient, respectively. The set of triplex (Ak , Bk , wk ) where Ak , Bk , wk
are given by Eqs. (A.7), (A.8), and (A.9), respectively, is called the Fourier series,
frequency content.
The complex form of the Fourier series of f (x) is
∞

Cn e2πint/T
−∞
where
An + iBn
Cn = , n>0
2
C0 = A0
An − iBn
Cn = , n>0
2
n
wn = , n = −2, −1, 0, 1, 2.
T
Let T = 2π and

n
1 n
Sn (f )(x) = Ck eikx = A0 + (Ak cos kx + Bk sin kx)
2
k=−n k=1
be the nth partial sum of the Fourier series of f . Then

2
1
Sn (f )(x) = π0 f (x − t)Dn (t)dt
π
2
1
= π0 f (t)Dn (x − t)dt (A.12)
π
where
1
n
sin n + 21 x
Dn (x) = + cos kx = (A.13)
2 2 sin 2x
k=1
is the “Dirichlet kernel”, and

S0 (f ) + S1 (f ) + · · · + Sn (f ) 1 2
σn (f ) = = π0 f (x − t)Kn (t)dt (A.14)
n+1 π
Appendix 375
where

D0 (x) + D1 (x) + · · · + Dn (x) 1 sin2 n+1 (x)
Kn (x) = = 2
(A.15)
n+1 n + 1 2 sin2 2x
is called the “Fejer Kernel”.

Theorem A.9 (Bessel’s Inequality)
∞

|Ck |2 ≤
f
2L2 (0,2π)
k=−∞
or
∞
1 2 2
A + (A + Bn2 ) ≤
f
2L2 (0,2π)
4 0 n=1 n
This also mean that {Ak } and {Bk } are elements of l2 .

Theorem A.10 (Reisz -Fisher Theorem) Let {Ck } ∈ l2 . Then there exists f ∈ L2 (−π, π )
such that {Ck } is the kth Fourier coefficient of f . Furthermore
∞

|Ck |2 ≤
f
2L2 (−π,π) .
k=−∞
For {Ak }, {Bk } belonging to l2 , there exists f ∈ L2 (0, 2π ) such that Ak , Bk are, respec-
tively, kth cosine and sine Fourier coefficient of f . Furthermore
∞
1 2 2 π
A + (A + Bk2 ) = |f |2 dt .
2 0 n=1 k −π
Theorem A.11 Let f ∈ L2 (−π, π ), then
lim
f − Sn (f )
L2 (−π,π) = 0.
n→∞
Theorem A.12 Let f ∈ C[0, 2π ] such that

2
w(f , t)
π0 dt < ∞, w(f , t) = sup |f (x + t) − f (x)|.
t t
Then the Fourier series of f converges uniformly to f , that is
lim
f − Sn (t)
L∞ (−π,π) = 0.
n→∞
If w(f , η) = 0(ηα ), then the condition of the theorem holds.

376 Appendix
Theorem A.13 If f is a function of bounded variation, then
f (x+ ) + f (x− )
Sn (f ) → as n → ∞
2
where
f (x+ ) = lim+ f (x + h)
h→0
f (x− ) = lim+ f (x − h)
h→0
exists at every x, a < x < b.

Theorem A.14 Let f ∈ L1 (R), then the series k=∞k=−∞ f (x + 2π k) converges almost
everywhere to a function λ(x) ∈ L1 (0, 2π ). Moreover, the Fourier coefficient ck of
1 ˆ
λ(x) are given by ck = 2π f (k)
lim f (k) (x) = 0 for k = 0, 1, . . . , n − 1

|x|→∞
then
F{f (n) } = (iω)n F(f ).
Definition A.14 Let f , g ∈ L1 (R) then the convolution of f and g is denoted by

f ∗ g and is defined by
∞
1
F(f ∗ g)(x) = √ f (x − u)g(u)du.
2π −∞
Theorem A.15 For f , g ∈ L1 (R), F(f ∗ g) = F(f )F(g) holds.
Theorem A.16 Let f be a continuous function on R vanishing outside a bounded

interval. Then f ∈ L2 (R) and

f
L2 (R) =
f
L2 (R) .
Definition A.15 (Fourier Transform in L2 (R)) Let f ∈ L2 (R) and {Φn

be a sequence
of continuous functions with compact support to f in L2 (R), that is,
f − φn
L2 (R) →
0. The Fourier transform of f is denied by
fˆ = lim Φn
n→∞
where the limit is with respect to the norm in L2 (R).

Theorem A.17 If f ∈ L2 (R), then (i) f , gL2 = fˆ , ĝ (Parseval’s formula)
L2
(ii)
f
L2 =
f
L2 (Plancherel formula)
Appendix 377
In physical problems the quantity

f
L2 is a measure of energy while
fˆ
L2 rep-
resents the power spectrum of f .
Theorem A.18 (1) Let f ∈ L2 (R). Then

π
1
fˆ (ω) = lim √ e−iωxdx
n→∞ 2π −π
(2) If f , g ∈ L2 (R), then

∞ ∞
f (x)ĝ(x)dx = fˆ (x)g(x)dx.
−∞ −∞
Theorem A.19 (Inversion of the Fourier Transform in L2 (R)) Let f ∈ L2 (R). Then
π
1
f (x) = lim √ eiωx fˆ (ω)d ω
n→∞ 2π −π
where convergence is with respect to the norm in L2 (R).

∞
Corollary A.1 (1) If f ∈ L1 (R) ∩ L2 (R), then fˆ (ω) = √12π −∞ e−iωx f (x)dx.
π
(2)
fˆ (ω) − √12π −π e−iωx f (x)dx
L2 (R) → 0 as n → ∞.
π
(3)
f (x) − √12π −π eiωx fˆ (ω)d ω
L2 (R) → 0 as n → ∞.
(4)
f
2 =
fˆ
2 .
L2 (R) L2 (R)
(5) The map f → fˆ is an isometry of L2 (R) onto L2 (R).
Example A.1 (1) If f (x) = (1 − x2 )e−x /2 = Second derivative of a Gaussian func-

2
tion, then
fˆ (ω) = ω2 e−ω /2 .
2
(2) If the Shannon function is defined by
sin 2π x − sin π x
f (x) =
πx
then
√1 , if π < |ω| < 2π

fˆ (ω) = 2π
0, otherwise
The results given in the appendix are based on references [1, 2, …]

378 Appendix
References
2. Pontrjagin, L. S. (1939). Topological groups. New Jercy: University Press Princiton.
3. Vilenkin, N. Y. (1947). On a class of complete orthonormal systems, Izv. Akad. Nauk Sssr Ser.
Mat. 11:363–400. English Translation, American Mathematical Society Translations 28(Series
2)(1963), 1–35.
65(3), 372–414.
6. Schoenenberg, I. J. (1946). Contribution to the problem of approximation of equidistant data
by analytic functions. Quarterly of Applied Mathematics, 4(1), 45–99.
7. Selesnick, I. W. (1998). Multiwavelet bases with extra approximation properties. IEEE Trans-
actions on Signal Processing, 46(11), 2898–2908.
8. Wade, W. R. (1995). A Walsh system for polar coordinates. Computers & Mathematics with
Applications, 30(3–6), 221–227.
9. Weide, B. W., Andrews, L. T., & Iannone, A. M. (1978). Real-line analysis of EEG using Walsh
transforms. Computers in Biology and Medicine, 8(4), 255–263.
10. Welstead, S. (2000). Fractal and wavelet image compression techniques. Washington: SPIE
Optical Engineering Press.
11. Siddiqi, A. H. (2018). Functional Analysis and Applications, Springer.
Notational Index
Symbols φ(x), 114

CW , H p , Lp , BMO and VMO, 86 ψa,b (t), 14
Dn (x), 34 ψj,k (t), 15
G p , 171 aα , 177
H ∗ , G ∗ , 11 bs , 177
H2n +r (x), 235 dj,k , 15
I (p, ω), 8 h2n +k (x), 10
Jk (x), 3, 4 m0 (ω), 114, 177, 193

f (ω), 6 pk (t), 237
Ja f (ω), 6 tn (f , x), 47
Kk (x, n), 40 W (δ, f ), 3
Kn (t), 49 Wp (f , δ), 46
(α) x ⊕ y, 3
Kn (x), 41
Lipj(t), 44
L (R+ , CN ), 326
2
Ml , 226
Ln (t), 49 MRA in L2 (G), 190
MRA, 16 MRA in L2 (R+ ), 103
Pa (f ), 247 NUMRA, 284, 285
Pj f , 110 PNSR, 277
Sn f , 34 VNUMRA, 334
Tψ f (a, b), 14 [x], xi
φ̂(ω), 123 x ⊕p y, xi
λ(x), 172 Z, Z+ , N , C, x

https://doi.org/10.1007/978-981-13-6370-2
Index
A EEG analysis, 75
Approximation by Walsh–Fourier series, 40
Approximation in dyadic homogeneous
Banach spaces, 57 F
Fejer Kernel, 375
Fourier transform, 376
B inversion, 377
Banach space, 360 Fractional functions, 226
Bessel’s inequality, 375
Biorthogonal wavelets, 253
construction, 259 G
Vilenkin group, 263, 269 Generalized Haar wavelets, 93
Bounded fluctuation, 29
H
C Haar filters, 11
Classes of functions, 371 Haar Fourier series, 82
Cohen’s criterion, 270 Haar System, 79, 86, 235
Convergence Haar–Vilenkin coefficients, 248
absolute, 372 Haar–Vilenkin scaling function, 237
almost everywhere, 373 Haar–Vilenkin wavelet, 235, 236
pointwise, 372 approximation, 243
uniform, 373 convergence, 246
orthonormality, 240
Haar wavelet, 13
D applications
Data compression, 73 boundary value, 95
Discrete Haar Transform (DHT), 84 initial value, 94
Discrete wavelet system, 15 integral equations, 96
Dyadic convolution, 364 Hadamard–Paley matrix, 9
Dyadic group G, 369 Hadamard transform self adjoint, 9
Dyadic wavelets
approximation properties, 158
I
Image digitization, 71
E Image enhancement, 72
ECG analysis, 74 Image transmission, 71

https://doi.org/10.1007/978-981-13-6370-2
382 Index
L V
Lipj(t), 44 Vector-valued MRA, 318
Lipschitz class, 362 Vector-valued MRA p-analysis, 321
Locally compact Cantor group, 171 Vector-valued Nonuniform MRA, 334
Vector-valued nonuniform wavelet packets,
347
M Vilenkin–Christenson transform
Mask of scaling equation, 112 periodic wavelets, 211
Mask of the refinable function, 265 Vilenkin group, 171
Modified Strang–Fix condition, 256 periodic wavelets, 205
Multiresolution Analysis (MRA), 16, 267 Vilenkin-Pontrjagin, 369
Vilenkin systems, 367
N
Non-uniform Haar wavelets, 88
Non-uniform MRA (NUMRA), 91, 283 W
Nörlund Kernel, 47 Walsh–Dirichlet Kernel, 220, 363
Walsh–Fourier coefficients, 27
transformation, 31
O Walsh–Fourier series
Orthogonal wavelets in L2 (R+ ), 103, 114 applications
image representation, 70
approximation in Lp spaces, 45
P convergence, 34
Pattern recognition, 76 Walsh–Fourier transform, 5, 101
p-wavelets, 194
relation with Fourier transform, 7
Walsh function, 363
integration, 3
R
relations with Haar function, 8, 83
Reisz -Fisher Theorem, 375
Riemann–Lebesgue Theorem, 28 Walsh polynomials, 2
Riesz system, 118 Walsh-type wavelet packets, 10
Wavelet coefficients, 16
Wavelet representation, 16
S Wavelets, 12
Scaling function, 104 compact wavelets, 17
σ − algebra, 362 construction, 16
Speech processing, 76 Wavelet series, 16
Wavelet transform
continuous, 12
U W -compact, 122
Uniformly W -continuous, 101 W -continuous, 6

Construction of Wavelets Through Walsh Functions: Yu. A. Farkov Pammy Manchanda Abul Hasan Siddiqi

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Construction of Wavelets Through Walsh Functions: Yu. A. Farkov Pammy Manchanda Abul Hasan Siddiqi

Uploaded by

Copyright:

Available Formats

Industrial and Applied Mathematics

More information about this series at http://www.springer.com/series/13577

Abul Hasan Siddiqi

Abul Hasan Siddiqi

ISSN 2364-6837 ISSN 2364-6845 (electronic)

Library of Congress Control Number: 2019930359

© Springer Nature Singapore Pte Ltd. 2019

The book is essentially devoted to the construction of wavelets by Walsh functions

Moscow, Russia Yu. A. Farkov

1 Introduction to Walsh Analysis and Wavelets . . . . . . . . . . . . . . . . . 1

2.4.4 Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

6.4 Covergence Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246

Yu. A. Farkov is a distinguished scientist and professor at the Russian Presidential

Pammy Manchanda is a senior professor of mathematics at Guru Nanak Dev

Abul Hasan Siddiqi is a distinguished scientist and Adjunct Professor at the

ðzj Þ ¼ ðxj Þ þ ðyj Þ , zj ¼ xj þ yj ðmod pÞ for all j 2 Z

Ul ¼ fðxj Þ 2 Gp : xj ¼ 0 for j lg; ; 2 Z

Let R þ :¼ ½0; 1Þ. With the help of the mapping k : Gp ! R þ ,

xj ¼ h½p j xip ; xj ¼ h½p1j xip ; j 2 N: ð1Þ

These numbers are the digits of the p-adic expansion

where xj , yj are calculated by (1). As usual, we denote by ⊖p the inverse operation

wj;k :¼ 2j=2 wð2 j x kÞ; j 2 Z; k 2 Z þ

form an orthonormal basis for L2 ðR þ Þ. In general, a function w is called a wavelet

An analogue of Grossmann–Morlet’s formula for this transform is proved in

orthogonal compactly supported wavelets on the group G2 were constructed by

A complete description of the orthogonal wavelets and scaling functions on R þ ,

1.1 Walsh Functions

The Walsh system {wn : n ∈ Z+ } on R+ is defined as follows:

x ∈ R+ , where R+ = [0, ∞),

and the function w1 (x) is defined on [0, 1] by the formula

and is extended to R+ by periodicity: w1 (x + 1) = w1 (x) for all x ∈ R+ .

x j = [2 j x](mod 2), x− j = [21− j x](mod 2). (1.1)

These are the digits of the binary expansion

(For dyadic x, we obtain an expansion with finitely many nonzero terms.)

where x j , y j are defined in (1.1). By definition x  y = x ⊕ y (because x ⊕ x = 0).

A function f is W -continuous if it is W -continuous at each point of R+ . A continuous

Proof Suppose f (x) is a continuous function and let

w(δ, f ) = sup | f (x1 ) − f (x2 )|

If 0 ≤ h = |x1 − x2 | ≤ δ < 1, then by choosing x1 = x ⊕ h and x2 = x and keeping

sup | f (x ⊕ h) − f (x)| → 0 as n→∞

Since ψ0 (t) = 1, it follows that

ψk (x) = ψ2n (x)ψk  (x)

From (1.3) with k = 2n , k  = 0, we have

This completes the proof.

1.2 Walsh–Fourier Transform

The Walsh–Fourier transform of a function f ∈ L 1 (R+ ) is defined as follows:

then f˜ is defined as limit of Ja f in L 2 (R+ ) as a → ∞.

Theorem 1.3 The following properties hold:

holds for each x ∈ R+ , provided that both f and

Clearly, each Walsh polynomial of order 2n − 1 belongs to En . The set E of dyadic

(d) if fˆ ∈ En , then supp f ⊂ [0, 2n ], provided that f ∈ L 1 (R) and f is

for all u ∈ R + which satisfy |u| < 2−n . Clearly,

By the inverse of Walsh–Fourier transform

and by the inverse of Fourier transform

Substituting f (x) from (1.7) in (1.6), we get

Interchanging order of integration of (1.9) yields

1.3 Haar Functions and Its Relationship with Walsh

Let In (x) = I ( p, n), where

I ( p, n) = [ p2−n , ( p + 1)2−n ], 0 ≤ p < 2n , n, p ∈ N (1.10)

Let l∞ denote the collection of a ∈ l0 such that

a ∞ = supn∈N |an | is finite. (1.12)

It may be recalled that l p , 1 ≤ p ≤ ∞ is a Banach space and for {ak }, {bk } ∈ l0

For each n ∈ N set

xj ¼ h½p j xip ; xj ¼ h½p1j xip ; j 2 N: ð1Þ

where x j , y j are defined in (1.1). By definition x y = x ⊕ y (because x ⊕ x = 0).

ψk (x) = ψ2n (x)ψk (x)

From (1.3) with k = 2n , k = 0, we have

a∞ = supn∈N |an | is finite. (1.12)

It is also clear that ∞ ∞