Notes

Chapter 1.
The standard reference for trigonometric series is Zygmund [9]. Other
books on Fourier series that were helpful to us in the preparation of Chapters 1
and 2 are Bari [1], Helson [5], Katznelson [6], and Stein and Weiss [7].
The simplest orthogonal wavelet is the Haar function studied by A. Haar
in [55]. The notion of integral wavelet transform (IWT), (W^f)(b, a), was first
introduced by Grossmann and Morlet [54], although the techniques which are
based on the use of translations and dilations can be traced back to Calderón
[30] in the study of singular integral operators. The basic wavelet used to
formulate the IWT is also called a "mother wavelet" in the wavelet literature.
The formula for the recovery of any / G L 2 (R) from its IWT (W^/)(6,a),
a,b G R, can be found in [54].
The importance of the semi-discrete IWT, (W^/)(6, a), where b G R and
a = 2 - · 7 , j G Ζζ, in image compression was first pointed out by S. Mallat, and
the notion of dyadic wavelets was also introduced by Mallat in his joint work
with W.L. Hwang [60] and with S. Zhong [61, 62]. The stability condition on
dyadic wavelets studied in Mallat and Hwang [60] can be viewed as a gener­
alization of the Littlewood-Paley Identity. Inequalities of this type for frames
and wavelets were studied in Daubechies [50] and Chui and Shi [37], and in
particular, the identity was used by Chui and Shi [38] to characterize wavelets.
A general reference on signal processing is Oppenheim and Schäfer [25],
and a mathematical analysis of the subject is given in Chui and Chen [24].
Other references that deal with certain specialized but related topics in signal
and image processing are Auslander, Kailath, and Mitter [22], and Rosenfeld
[26].
That an 7^-function is not necessarily an 7^-wavelet (or wavelet) was
already observed by Y. Meyer in the first volume in [21] and discussed by
Daubechies in some detail in [50], where the results of Tchamitchian [68, 69]
were used. The proof in this chapter was given by Chui and Shi in [37], fol­
lowing the ideas of Daubechies and Meyer.
The notion of multiresolution analysis (MRA) was first introduced by
Meyer [63] and Mallat [58], and further developed by Mallat in [57, 59]. It was
also Mallat [57-59] who constructed the wavelet decomposition and reconstruc­
tion algorithms using the MRA spaces. The presentation of these algorithms in
this chapter follows [40, 43] in that the normalization constant \/2 is absorbed
by the basis functions to facilitate implementations. We remark that there is
245
246 Notes

some similarity between an MRA and the Laplacian pyramid algorithm due to
Burt and Adelson [29].
The most comprehensive work on cardinal splines is [14] by Schoenberg,
who was also responsible for the development of this subject. It was Meyer [63]
and Lemarié [56] who noted that cardinal spline spaces give rise to MRA, al­
though approximation theorists have also been considering subdivision schemes.
(See Chui [11] and the references therein.) Corresponding to the m t h order B-
spline ATm, the m t h order B-wavelet tpm was introduced by Chui and Wang
[43]. The dual ipm of V>m was also constructed in [43] in terms of the m t h
derivative of the shifted and scaled fundamental spline of order 2m introduced
by Chui and Wang in [40].
A predictor-corrector algorithm can be devised to implement an IIR filter
as an ARMA filter with poles lying both inside and outside the unit circle.
Such algorithms can be optimized by incorporating them with noise processes
(see Chui and Chen [23]). This procedure should apply to spline-wavelet de­
compositions without truncations.

Chapter 2.
There are many good references on Fourier transforms in the literature.
The ones that were helpful to us in the preparation of Chapter 2 are Goldberg
[4] and Titchmarsh [8].
Since functions in LP(R) or L p (0,2π) are considered to be "equivalence
classes" of functions, we are allowed to change the functions on sets of mea­
sure zero. In particular, in the statements of pointwise convergence, we always
mean convergence to some representative of the equivalence class under con­
sideration.
Again, the books [1, 5, 6, 7, 9] on harmonic analysis are good sources
for further reading. This subject has a very rich history. We only include a
very brief discussion on the development of pointwise convergence here. As
early as 1876, du Bois-Reymond already showed the existence of a 27r-periodic
continuous function whose Fourier series diverges at some point. From this
result, it is not difficult to prove the existence of a 27r-periodic continuous
function whose Fourier series is divergent on a dense subset of the real line R.
In 1923, Kolmogorov extended this observation and showed that the Fourier
series of some 27r-periodic continuous function may diverge almost everywhere.
Three years later, he even extended this result and proved the existence of some
/ G L 1 (0,2π) with an a.e. divergent Fourier series. On the other hand, in 1966,
Carleson proved that if / G L 2 (0,2π), then its Fourier series converges a.e. This
very deep result was extended by Hunt in 1967 to every / G L p (0,2π), p > 1.
A somewhat extensive discussion of the Poisson Summation Formula is
given in this chapter because several variations of this formula are used through­
out this book.
Notes 247

Chapter 3.
The window Fourier transform was first introduced by Gabor [53] by using
a Gaussian function as the window function. This is why it is also called
the Gabor transform. In the engineering literature, this windowing process,
which is not restricted to the use of a Gaussian function, is called a short-time
Fourier transform (STFT) as well. See Daubechies [20, 50], Mallat [57], some
of the chapters in the edited volumes [16, 17, 18, 19] and the lists of references
therein. The reader is also referred to the same sources for material concerning
the use of the standard deviation to define the radius, and hence width, of a
window function as well as for discussions on the Uncertainty Principle. As
mentioned earlier, when we refer to a function in LP(1R), we actually mean a
representative of an equivalence class of functions which are identical except on
sets of measure zero. In particular, we always use a continuous representative
as a window function whenever it is available. See also Champeney [2] for
further discussions and references from the harmonic analysis point of view.
As mentioned in the notes for Chapter 1, the IWT was introduced by
Grossmann and Morlet [54], where the admissibility condition (3.3.1) on the ba­
sic wavelets (also called the mother wavelets) was imposed. The semi-discrete
version of the IWT has proved to be very useful in Mallat's work on image
compression, first using zero-crossings and later using wavelet maxima (or lo­
cal extrema) of the IWT on dyadic scale-levels. See the work of Mallat and
Zhong [61, 62] and Mallat and Hwang [60]. In this regard, the stability condi­
tion (3.4.6) is crucial to the inversion formula (3.4.14) introduced in Mallat and
Hwang [60]. The characterization of dyadic duals using the Littlewood-Paley
identity is given in Chui and Shi [38].
The notion of frames was introduced by Duffm and Schaeffer [51] and
studied in some detail by Daubechies [20, 50]. Example 3.18 is also attributed
to Daubechies. The Interior Mapping Principle used in our discussion on the
boundedness of the inverse map generated by the frame is standard in oper­
ator theory (see Vol. 1, page 57, of Dunford and Schwartz [3]). The stability
property (3.5.18) of a frame was derived in Chui and Shi [37]. We also remark
that Frazier and Jawerth [52] used dilation and translation in their work on
(^-transforms.
Semi-orthogonal wavelets were introduced independently by Auscher [27]
and Chui and Wang [40, 41, 43], while semi-orthogonal cardinal spline-wavelets
were first constructed by Chui and Wang in [40, 43]. The orthonormalization
procedure in (3.6.18) is due to Schweinler and Wigner [66] and is therefore
called the Schweinler-Wigner o.n. procedure.
Chapter 4.
Spline analysis is an established subject. The reader is referred to de Boor
[10], Nürnberger [13], Schoenberg [14], and Schumaker [15] for the univariate
theory, and to Chui [11] for a multivariable investigation. A spline function
with equally spaced (simple) knots is called a cardinal spline, and the reader is
referred to Schoenberg [14, 65] for further study. In particular, the structure
of Euler- Frobenius polynomials is documented in [14, 65].
248 Notes

The graphical display algorithm (Algorithm 4.7) introduced in Section 4.3

is an iteration of a standard subdivision scheme. It is presented here in order to
motivate the introduction of the (wavelet) reconstruction algorithm in Chapter
5. The cardinal B-spline B-net algorithm (Algorithm 4.10) is the one-variable
version of the box-spline graphical display algorithm developed in Chui and
Lai [36].
The notion of quasi-interpolation is due to de Boor and Fix (see de Boor
[10] and Schumaker [15]). The Neumann series approach was first introduced
by Chui and Diamond in [35], where Theorem 4.13 was established. The char­
acterization of quasi-interpolants presented here (see (4.5.35)-(4.5.36)) was
given by Chui and Diamond [33]. The theory of interpolation by cardinal
splines was developed by Schoenberg (see Schoenberg [14, 65]). The con­
struction of local interpolation formulas presented in Section 4.6 by taking
a "Boolean sum" of a completely local interpolation operator and a quasi-
interpolation operator was introduced by Chui and Diamond in [34]. A tutorial
study of this topic is given in Chui [32], and in more generality, in Chui [31].

Chapter 5.
The notion of multiresolution analysis (MRA) was first introduced by
Meyer [63] and Mallat [58], and further developed by Mallat in [57, 59]. The­
orems 5.5 and 5.6 are due to Cohen [44]. The result on the support of a
scaling function being given by the length of its two-scale sequence is due to
Daubechies [49], and the scaling function defined in (5.2.6) was also given in
[49]. The characterization of minimally supported scaling functions in terms
of the non-existence of symmetric zeros was given by Chui and Wang [41],
where the notion of generalized Euler-Frobenius (Laurent) polynomials was
introduced and their properties, such as those in Theorem 5.10, were stud­
ied. The fact that the m t h order B-spline is the only function that generates
the MRA {V™} and has a finite two-scale sequence was also proved in [41].
The presentation of the direct-sum decomposition in Section 5.3 seems to be
new, and the approach in Section 5.4 is a generalization of the work of Cohen,
Daubechies, and Feauveau in [46]. In particular, the criterion (5.4.11) in Theo­
rem 5.19 is new. The duality principle for spline-wavelets and, more generally,
semi-orthogonal wavelets, was introduced in Chui and Wang [43] and [41], re­
spectively, and the nonorthogonal (or biorthogonal) version is again found in
Cohen, Daubechies, and Feauveau [46].
As mentioned earlier, wavelet decomposition and reconstruction algorithms
were constructed in Mallat [58]. The formulation here follows [40, 43]. The
importance of linear phase in filtering and the relation between linear phase
and symmetry are well known in the engineering literature (see Oppenheim
and Schäfer [25]). Our discussion in Sections 5.5 and 5.6 is an extension of the
results of Chui and Wang in [41].
The general solution of (5.6.39) of the dual relation (5.6.17) was given
in Daubechies [49] for the orthonormal setting, and in Cohen, Daubechies,
and Feauveau [46], in general. Compactly supported wavelets with compactly
Notes 249

supported duals that meet the linear-phase requirement were also constructed
by Cohen, Daubechies, and Feauveau in [46]. The filter bank approach was
considered by Vetterli and Herley [70].

Chapter 6.
The interpolatory spline-wavelets given in Theorem 6.1 are due to Chui
and Wang [40], and the identification between the MRA subspaces V^2m'° of
Vi™ with the wavelet spaces W™ was introduced in Chui and Wang [39]. The
compactly supported cardinal spline-wavelets (or B-wavelets) given in (6.2.5),
together with their duals in (6.2.10), were introduced by Chui and Wang in
[43]. However, the presentation in Section 6.2 is quite different in that it makes
use of the more general result, Theorem 5.19, from Chapter 5. The Pascal
triangular algorithm (PTA) for computing B-wavelets and their derivatives
was introduced in Chui and Wang [42].
The presentation of properties of the Euler-Frobenius polynomials follows
Schoenberg [65], while the material on error analysis in spline-wavelet decom­
position is taken from Chui and Wang [42]. The most comprehensive reference
on total positivity is Karlin [12], and additional information on Pólya frequency
(PF) sequences can be found in Schoenberg [14]. Theorem 6.21 on certain lin­
ear PTA's producing PF sequences was proved in Chui and Wang [39], where
the notion of complete oscillation was first introduced and its relation to zero-
crossings explored.

Chapter 7.
The first nontrivial wavelet was constructed by Strömberg [67], using
spline functions. The Meyer wavelets [64], as studied in Example 7.3, are
o.n. wavelets with compactly supported Fourier transforms. The o.n. spline
wavelets given in Example 7.1 are usually called the Battle-Lemarié wavelets,
since they were constructed independently by Battle [28] and Lemarié [56] using
different methods. However, none of these o.n. wavelets has compact support.
Based on the structure of MRA, Daubechies [49] was the first to construct
compactly supported o.n. wavelets. Hence, her construction was based on the
identification of o.n. scaling functions as discussed in Section 7.2, although the
presentation there is somewhat different from that in [49]. The construction of
the Daubechies wavelets also depends on the Riesz Lemma as stated in Theo­
rem 7.17. Theorem 7.22 was also proved by Daubechies [49], while the result
stated in (7.3.28) was established by Cohen and Daubechies in [45].
Wavelet packets, also called "wave packets" by Coifman and Meyer, were
introduced in Coifman, Meyer, Quake, and Wickerhauser [47]. See also Coif­
man, Meyer, and Wickerhauser [48] for further information.