470 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO.

2, FEBRUARY 2003

Wavelet Theory Demystified

Michael Unser, Fellow, IEEE, and Thierry Blu, Member, IEEE

Abstract—In this paper, we revisit wavelet theory starting from

the representation of a scaling function as the convolution of a
B-spline (the regular part of it) and a distribution (the irregular
or residual part). This formulation leads to some new insights on
wavelets and makes it possible to rederive the main results of the
classical theory—including some new extensions for fractional
orders—in a self-contained, accessible fashion. In particular, we
prove that the B-spline component is entirely responsible for five
key wavelet properties: order of approximation, reproduction
of polynomials, vanishing moments, multiscale differentiation
property, and smoothness (regularity) of the basis functions.
We also investigate the interaction of wavelets with differential
operators giving explicit time domain formulas for the fractional
derivatives of the basis functions. This allows us to specify a Fig. 1. Block diagram of the two-channel filterbank that forms the basis for
the wavelet transform algorithm. From a purely digital signal processing point
corresponding dual wavelet basis and helps us understand why
of view, the only constraints on the choice of the filters are the two perfect
the wavelet transform provides a stable characterization of the reconstruction equations.
derivatives of a signal. Additional results include a new peeling
theory of smoothness, leading to the extended notion of wavelet
differentiability in the -sense and a sharper theorem stating per se and their fundamental properties. To gain a deeper
that smoothness implies order. understanding, one needs to consider the continuous-time in-
Index Terms—Approximation order, Besov spaces, Hölder terpretation of the transform that is favored by mathematicians
smoothness, multiscale differentiation, splines, vanishing mo- [4]–[6]. This representation involves continuously-defined
ments, wavelets. basis functions (wavelets) that are rescaled and shifted versions
of each other; it also gives rise to a filterbank implementation
that is equivalent to the block diagram in Fig. 1—Mallat’s
I. INTRODUCTION
fast wavelet algorithm [4], [7]. However, one of the important

T HE mechanics of the wavelet transform are usually well

understood by engineers working in signal processing. In
essence, the system boils down to a two channel filterbank as
points of the continuous-time formulation is that the filters
cannot be completely arbitrary. A key constraint is that the
lowpass filter —also called the refinement filter—must
shown in Fig. 1 [1], [2]. In the decomposition (or analysis) step, be factorizable as 2 , which is an
the digital signal is split into two half-size sequences expression that involves some number of “regularity” factors
and by filtering it with a conjugate pair of ( ) as well as a stable residual term satisfying
lowpass and highpass filters ( and , respec- the lowpass constraint . A nontrivial factor is
tively) and down-sampling the results thereafter. The signal is generally necessary for obtaining orthogonal or biorthogonal
then reconstructed (synthesis step) by up-sampling, filtering, wavelets that can be implemented by means of FIR filters [8],
and summation of the components, as shown on the right-hand [9]. The presence of the regularity term is essential
side of the block diagram; note that the analysis and synthesis for theoretical reasons. It is responsible for a number of key
procedures are flow-graph transposes of each other. A standard wavelet properties such as order of approximation, vanishing
analysis shows that this kind of two channel decomposition moments, reproduction of polynomials, and smoothness of the
is one-to-one and reversible provided that the -transforms basis functions. If one excepts the vanishing moments, these
of the filters satisfy the perfect reconstruction (PR) equations are aspects of wavelets that are often not so well understood by
also given in Fig. 1 [1], [3]. In the tree-structured wavelet signal processing practitioneers, mainly because the connection
transform, the decomposition step is further iterated on the with the digital filterbank interpretation is not obvious. This
lowpass component . is rather unfortunate as many consider these properties as the
While the block diagram in Fig. 1 constitutes a valid descrip- very core of wavelet theory (cf. [6, Ch. 7]); they are almost
tion of the algorithm—different transforms simply correspond unavoidable if one wants to give precise explanations as to why
to different sets of PR filters—it tells us little about wavelets wavelets work so well for approximating piecewise-smooth
signals and why they are such a nice tool for characterizing
singularites [10].
Manuscript received December 12, 2001; revised September 24, 2002. The
associate editor coordinating the review of this paper and approving it for pub- The purpose of this paper is to demystify some of these con-
lication was Dr. Helmut Boelcskei. cepts by linking them to a factorization theorem involving the
The authors are with the Biomedical Imaging Group, Swiss Fed- convolution between a B-spline (the regular part of the wavelet)
eral Institute of Technology Lausanne (EPFL), Switzerland (e-mail:
[email protected]). and a distribution (the irregular or residual part). Indeed, we will
Digital Object Identifier 10.1109/TSP.2002.807000 show that it is the B-spline part—and nothing else—that brings
1053-587X/03$17.00 © 2003 IEEE
UNSER AND BLU: WAVELET THEORY DEMYSTIFIED 471

in all the desired mathematical properties. Although the effect A. Continuous Function Spaces and Notations
of the regularity factors is well understood by mathematicians The continuous-time version of the wavelet transform applies
working with wavelets, we are not aware of any deliberate effort to functions , that are square integrable. The space
to explain these properties from the perspective of B-splines. of those functions is denoted by ; it is the Hilbert space that
While it could be argued that this is essentially a matter of inter- corresponds to the inner product
pretation (regularity factors are equivalent to B-splines), we be-
lieve that the present formulation makes the whole matter more
transparent and accessible. Our only prerequisite here is to have (1)
a complete understanding of the properties of B-splines, which
is much easier than for other wavelets, since these are the only where the integral is taken in Lebesgue’s sense. The energy of
ones that have explicit formulas. We then use relatively simple a function is given by its squared -norm:
manipulations to show that these properties carry over to all ; thus, the notation is equivalent to the state-
scaling functions through the convolution relation. ment that is finite. More generally, one defines the
spaces for (where stands for Lebesgue) as the
A. Scope and Organization of the Paper set of functions whose -norm

This paper is written for researchers in signal processing who

have a basic knowledge of wavelets and filterbanks and who (2)
would like to improve their understanding of the more theoret-
ical aspects of wavelets that are usually confined to the math- is finite. These are Banach spaces; there is no corresponding
ematical literature. The paper is largely self-contained but as- inner product, except, of course, for .
sumes some familiarity with Mallat’s multiresolution formula- The Fourier transform of is denoted by . If ,
tion of the wavelet transform that rests on the fundamental no- it is given by . This definition can be ex-
tion of scaling function [4]. From a theoretical point of view, the tended for functions , as well as for generalized func-
scaling function is more important than the wavelet because it tions , where stands for Schwarz’ space of tempered
provides the elementary building blocks for the decomposition; distributions on [12].
it is responsible for the key properties of the transform—and this In wavelet theory, one usually considers some wavelet func-
is precisely what this paper is all about. tion that generates a basis of . What is meant here is
The presentation is organized as follows. In Section II, we that every function can be represented in a unique and
recall the main definitions and mathematical concepts that are stable fashion using the expansion
required for the formulation of the wavelet transform. We
also spend some time describing the fractional B-splines, which (3)
play a central role in our formulation. In Section III, we red-
erive—and often extend—the key theorems of wavelet theory
starting from the B-spline factorization of the scaling function. where the basis functions 2 2 are
Finally, in Section IV, we investigate the smoothness issue and obtained by dilation (index ) and translation (index ) of the
the interaction of wavelets with fractional differential operators. wavelet . The wavelet coefficients are obtained by
An important aspect of our formulation is that it allows for non- forming the (double infinite) sequence of inner products
integer orders , which are not covered by traditional wavelet
theory. This brings in two novelties: (4)

1) the extension of the theory for fractional wavelets such as

those introduced in [11]; where is the biorthogonal basis of
2) a new “peeling” theory of smoothness which general- such that . The
izes an interpretation of integer orders of differentiability
biorthogonal basis is also generated by a single template
given by Strang [6].
—the so-called biorthogonal wavelet [8]. In practice, the
The motivation here is to come up with a more intuitive—and inner products are never evaluated literally as in (4). They are
perhaps even more general—understanding of the concept of computed much more efficiently using Mallat’s algorithm,
wavelet smoothness. Another important goal of this paper is to which is based on a hierarchical application of the filterbank
clarify the notion that the wavelet representation is stable with decomposition in Fig. 1 [4]. The wavelet transform specifies
respect to (fractional) differentiation. the choice of a particular set of perfect reconstruction filters,
and vice versa [cf. (6) and (8)].

B. Scaling Functions
II. SCALING FUNCTIONS AND WAVELETS IN
The continuous-time interpretation of the wavelet decompo-
This section presents a brief review of the formulation and sition algorithm in Fig. 1 is based on the fundamental concept of
interpretation of the wavelet transform. It also contains a short scaling function [4]. The scaling function is either given
primer on fractional B-splines. explicitly—as is the case with the B-splines—or it is derived
472 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 2, FEBRUARY 2003

Fig. 2. Illustration of the key properties for the Haar case '(x) =  (x) (B-spline of degree zero). (a) Orthogonality. (b) Two-scale relation. (c) Wavelet
relation. (d) Partition of unity.

indirectly from the refinement filter as solution of the Condition i) ensures that generates a stable basis for the basic
two-scale relation [9]. The difficulty with the latter approach is function space
that not all filters do generate valid scaling functions [13], [14].
The mathematical requirements are the following.
Definition: is an admissible scaling function of if
and only if it satisfies the following three conditions: In other words, there is an equivalence between the -norm of
i) Riesz basis criterion; there exist two constants and the functions in and the -norm of their coefficients
such that

(5)
where and are the Riesz bounds of . The basis is or-
thonormal if and only if , as is precisely the case
ii) Refinability
with the example in Fig. 2(a).
The two-scale relation (6), which is equivalent to
(6) , is the key to the multiresolution structure of the transform.
It allows us to define the coarse-to-fine sequence of embedded
subspaces
iii) Partition of unity such that 2 .
The last, more technical partition of unity condition is neces-
(7) sary and sufficient (under mild conditions) for the approxima-
tion error to vanish as the wavelet size approaches zero [15]. It
ensures that the multiresolution decomposition is dense in .
These three conditions are satisfied by the Haar function We can prove that the above specification of a valid scaling
—or B-spline of degree zero—as illustrated in Fig. 2. function is equivalent to the axiomatic definition of a multires-
UNSER AND BLU: WAVELET THEORY DEMYSTIFIED 473

olution analysis of given by Mallat [7]. Hence, we have the

guarantee that there exists some corresponding wavelet
that generates a Riesz basis of . The wavelet is expressed as
a linear combination of shifted versions of the scaling function:

(8)

where the weights correspond to the impulse response of

the wavelet synthesis filter in Fig. 1. The dual scaling function
and wavelet are defined in a similar fashion using the dual
versions of (6) and (8), respectively; these involve the analysis
filters and in Fig. 1.

C. Fractional B-Splines Fig. 3. Fractional B-splines  (x) for   0. The polynomial ones ( 
integer) are represented using thicker lines.
The simplest example of a scaling function is the B-spline of
degree 0 (cf. Fig. 2). This function can be constructed by taking
the difference of two step functions shifted by one. This yields where ; is
the following mathematical representation: Euler’s Gamma function, which interpolates the factorial, i.e.,
for integer. If we specialize these formulae
for the more standard integer case ( ), then (10) yields the
(9) classical finite difference formula for the polynomial B-splines
(cf. [17]).
The fractional B-splines are shown in Fig. 3; they provide a
where denotes the one-sided power function
progressive transition between the polynomial ones ( integer)
; is the causal1 finite difference (convolution)
displayed in thicker lines. These functions have a number of in-
operator whose impulse response is . By
teresting properties that are briefly summarized here—for more
convolving with itself times, one generates the classical
details, see [11].
B-splines which are piecewise polynomials of degree [16].
These are also valid scaling functions simply because the con- • They generate valid Riesz bases for . In par-
volution of two scaling functions is a scaling function as well. It ticular, this means that they are square integrable, i.e.,
is still possible to go one step further by considering fractional , .
convolution products that yield the fractional B-splines [11]: • They belong to for all whenever , and for
all if .
• They satisfy the convolution relation:
(10) . This comes as a direct consequence of their
definition.
The right-hand side of this formula is easy to understand: It is • They reproduce polynomials of degree , where
the ( )th power of the Fourier transform in (9). The corre- denotes the ceiling of . In particular, they satisfy the
sponding definition of the fractional power of a complex number partition of unity for .
is with . The time-do- • They satisfy the two-scale relation. Their refinement filter
main formula on the left-hand side can be obtained by inverse is the generalized binomial. This is established easily by
Fourier transformation (cf. [11]). The key operator here is the expressing the two-scale relation in the Fourier domain
fractional causal finite difference , which is best defined and evaluating the following ratio:
in the Fourier domain

(11)

To obtain its time domain representation, we expand

(13)
using Newton’s generalized binomial the-
orem and perform an inverse Fourier transform
In fact, the fractional B-splines share all the properties of the
polynomial ones (cf. [18]) with two notable exceptions: They
are not strictly positive nor compactly supported unless is in-
teger. For noninteger, they decay like when .
with (12) Their lack of compact support is also apparent if we look at their
refinement filter 2 , which is infin-
1the subscript “+” is used to signify that the corresponding function or con-
volution operator is causal; this is to distinguish them from the centered version itely supported whenever is noninteger, although it decays like
of these operators which are often used in spline theory. , .
474 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 2, FEBRUARY 2003

D. Differentiation and Smoothness A. Order of Approximation

Our primary motivation for considering fractional splines in- A fundamental idea in wavelet theory is the concept of mul-
stead of conventional ones is that the enlarged family happens tiresolution analysis. There, one usually considers a sequence
to be closed under fractional differentiation. This will prove ex- of embedded subspaces with a dyadic scale progression, i.e.,
tremely useful for understanding and characterizing the smooth- 2, . Specifically, the approximation space at scale
ness properties of scaling functions and wavelets. 2 (or at resolution level ) is defined as
Our fractional derivative operator is defined in the Fourier
domain
(16)
(14)

It coincides with the usual derivative operator when is integer. Because of the two-scale relation, the subspaces have the fol-
Applying the operator to the fractional B-splines yields the fol- lowing inclusion property: . Given some input func-
lowing explicit differentiation formula tion , one considers its approximation in . The best
approximation in the least-squares sense (minimum -norm)
(15) is given by the orthogonal projection into

where is the fractional finite difference operator defined

by (12). Thus, the fractional derivative of a B-splines happens (17)
to be spline with a corresponding decrease of the degree. The
derivation of this result is rather straightforward if we work in Because of the Riesz basis property, the approximation of
the Fourier domain in is always well-defined and unique (Projection theorem).
A necessary condition for being able to construct wavelet bases
of is that the approximation improves as the scale gets
closer to zero. In other words, we want to be able to approxi-
mate any -function as closely as desired by selecting a scale
(or sampling step) that is sufficiently small. It is therefore crucial
to understand and characterize the rate of decay of the approxi-
Again, this is an extension of the well-known differentiation mation error as a function of the scale. This leads to the notion
formula for the polynomial B-splines. of order of approximation, which plays a fundamental role in
Thus, if we differentiate times, we end up with wavelet theory [6], [21], [22].
a function that is piecewise constant and Definition: A scaling function has order of approximation
bounded (because for ). The boundedness if and only if
of the derivative is an indication that is -Hölder continuous.
In fact, is infinitely differentiable everywhere except at (18)
the knots (integers), where it has discontinuities of order (e.g.,
). Hölder smoothness is a pointwise measure of conti-
where is a constant that may depend on but not on ; the
nuity that is often used to characterize wavelets [19]. In this con-
condition (Sobolev space of order ) ensures that
text, the Sobolev smoothness, which is a more global measure
, which is the -norm of the th derivative of , is
of differentiability in the -sense, is also important [6], [14],
finite so that the bound is well defined.
[20]. The critical Sobolev exponent for the fractional B-splines
is . In other words, has derivatives This definition means that for smooth signals, the -approx-
imation error should decay like the th power of the scale. All
in the -sense, i.e., for .
popular families (Daubechies, Battle-Lemarié, etc.) satisfy the
order constraint (29) and are indexed by their order , which is
III. WAVELET THEORY REVISITED always a positive integer. The least sophisticated transform—
Our goal in this section is to reformulate wavelet theory using Haar with —has a first order of approximation
a nontraditional point of view. Our argumentation is entirely ( ), which is enough for the approximation error to vanish
based on a B-spline factorization theorem (Section III-B), which as . However, one usually prefers to work with higher
is intimately related to the crucial notion of approximation order. order approximations which result in much faster convergence;
We will use this representation to derive the most important typically, (e.g., the 9/7 Daubechies and cubic spline
wavelet properties using relatively straightforward manipula- wavelets). While traditional wavelet theory requires to be in-
tions. In other words, we will show that the order of approxima- teger, our definition is also valid for noninteger orders
tion of the transform (Section III-A), the reproduction of poly- and applicable to the recent fractional extensions of wavelets
nomials property (Section III-C), the vanishing moments of the [11].
wavelet (Section III-D), and the multiscale differentiation prop- The way in which the order of approximation translates into
erty of the wavelet (Section III-E) can all be attributed to the filter design constraints is made explicit by the following the-
B-spline that lies hidden within. orem.
UNSER AND BLU: WAVELET THEORY DEMYSTIFIED 475

Fig. 4. Plot of k  (x 0 k ) for various values of . (a)  = 0 (piecewise constant). (b)  = 0:5 (fractional spline). (c)  = 1 (linear spline). (d)
 = 3 (cubic spline). This illustrates the property that the B-splines are able to reproduce a first order polynomial as soon as  > 0 (i.e., cases b, c, and d, but not
a).

Theorem 1: A valid scaling function has a th-order of where is the Fourier transform of a fractional B-spline
approximation if and only if its refinement filter can be as given by (10), and is a true function of bounded
factorized as on every closed interval. Because of our assumptions on , this
corresponds to a well-defined convolution product in the time
(19) domain

with and
where is stable, i.e., . (20)
The restricted version of this theorem for compactly sup- It is therefore always possible to express a scaling function as the
ported—or equivalently FIR—with integer is a standard convolution between a B-spline and a distribution. What The-
result in wavelet theory (cf. [6]). The important point here is that orem 2 also tells us is that the B-spline part is entirely respon-
the present version holds for any real with minimal re- sible for the approximation order of the transform. We will now
striction on (Riesz basis). The proof is technical and will be use the convolution relation (20) to show that the B-spline part
published elsewhere [23]. brings in three other very useful properties.

B. B-Spline Factorization Theorem C. Reproduction of Polynomials

The idea here is to interpret Theorem 1 in terms of B-splines. We have already mentioned that the B-splines of degree
If we were to consider the first term of the factorization alone, have the ability to reproduce polynomials of degree lesser or
then the refinement filter would generate a fractional B-spline equal to . Practically, this means that we can gen-
of degree [cf. (13)]. The filter alone, on the erate all polynomials by taking suitable linear combination of
other hand, would generate some distribution , which B-splines. In particular, one can construct the following series
is typically not a true function (in fact, we will see later on that of polynomials for :
it cannot have any Sobolev regularity at all). The only thing we
can say about is that it is lowpass (because ). Thus, with (21)
because of the convolutional nature of the two-scale relation, we
obtain the following corollary, which constitutes the main result
on which we will build our wavelet theory. which form a basis of the polynomials of degree because
Theorem 2 (B-Spline Factorization): is a valid scaling . The ability of B-splines to reproduce a straight line
function of order if and only if its Fourier transform can ( ) is illustrated in Fig. 4.
be factorized as The fact that (21) holds for the conventional case ,
with integer, is not too surprising because the B-splines are
themselves piecewise polynomial of degree . The noninteger
476 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 2, FEBRUARY 2003

case [cf. Fig. 4(b)] is less intuitive but is only truly relevant for
fractional wavelets, which exhibit a few exotic properties [11].
The main point of this section is that the polynomial repro-
duction property is preserved through the convolution relation
(20).
Proposition 3: Let be any distribution such that
and , . Then,
reproduces the polynomials of degree
lesser or equal to .
When is integer, Proposition 3 is a B-spline reformulation
of a standard result in wavelet theory [5], [6]. What is novel here
is the extension for noninteger, which is nontrivial.
Proof: We assume that the residual filter has sufficient
(inverse polynomial) decay for the moments of to be
bounded up to order . In other words, we want to be
times differentiable at the origin. This mild technical require-
ment—which is much weaker than compact support—ensures
that the convolutions and manipulations below are well-defined
in the distributional sense.
We start by showing that the convolution between and the
monomial produces a polynomial of degree

with Fig. 5. Illustration of the reproduction of polynomials of degree 0 and 1 using

Daubechies scaling functions of order  = 2.
The leading coefficient is one because
. We then use this result to evaluate the proof that is more accessible to an engineering audience is avail-
sum able in [26]. Note that Theorem 4—the converse part of Propo-
sition 3—is only applicable in the conventional framework:
compactly supported, and . We have already encoun-
tered some counter examples with the fractional B-splines (i.e.,
using (20) and (21) ), which are not compactly supported.
In addition, note that the results in this section do not require
the two-scale relation. As such, they are also applicable outside
the wavelet framework—for instance, in the context of interpo-
lation [27].
In addition, here, the leading coefficient is one implying that the
sequence of these polynomials for is a basis for D. Vanishing Moments
the polynomials of degree . A wavelet transform is said to have vanishing mo-
This polynomial reproduction property is illustrated in Fig. 5 ments if the analysis wavelet is such that ,
for the Daubechies scaling function of order 2. Although no one . These vanishing moments are undoubtedly one of
will question the fact that the linear B-splines reproduce the con- the better known and most useful properties of wavelets [10]. In
stant and the ramp, it is much less obvious that the fractal-like particular, they give the transform the ability to kill all polyno-
Daubechies functions are endowed with the same property. mials of degree [5], [21]. This property translates into a sparse
For sake of completeness, we mention the existence of a con- representation of piecewise smooth signals because the wavelet
verse implication which goes back to the Strang–Fix theory of coefficients will be essentially zero over all regions where the
approximation. signal is well approximated by a polynomial, e.g., the first few
Theorem 4: If a function reproduces the polynomials of terms of its Taylor series. This will produce streams of zero co-
degree and is compactly supported, then . efficients that can be coded with very few bits [28], [29]. The
This result was first conjectured by Strang in [24] and was vanishing moments also allow the characterization of singular-
later proven by Ron in a more abstract framework [25]. A recent ities based on the decay of the coefficients across scale—the
UNSER AND BLU: WAVELET THEORY DEMYSTIFIED 477

so-called persistence across scale property of the wavelet trans- Proof: Let us denote by and the determinant
form [10], [30], [31]. The study of this decay plays an important of the matrices and in (22), respectively. Then,
role in the analysis of fractal and multifractal signals [32]. clearly, . By inverting , we get
The vanishing moments are nothing but an indirect manifes-
tation of the ability of the scaling function to reproduce polyno- (24)
mials.
Proposition 5: If the scaling function reproduces the As a result, imposing (19) yields (23), with
polynomials of degree , then the analysis wavelet has 2 , which is bounded on the unit circle by the
vanishing moments. stability hypothesis. Conversely, since ,
This is a standard result in wavelet theory (cf. [5], [6]) that is imposing (24) with the assumption that is stable implies
rederived here for sake of completeness. (19) with 2 , which is stable as well.
Proof: The polynomial reproduction property is equiv- Finally, because for ,
alent to saying that the polynomials are in the linear span we note that the factorization (23) with stable is equivalent
of . Since the wavelet is perpendicular to to the condition .
by construction (cf. the general biorthogonal To get a better feel for this result, we consider (23) and take
wavelet formulation of Cohen-Feauveau-Daubechies [8]), it is the limit to obtain the asymptotic form of the filter as .
therefore perpendicular to the polynomials; in particular, the Using the fact that and assuming that
monomials with . is continuous, we obtain
Thus, by combining Propositions 3 and 5, we can claim that
as (25)
it is the B-spline component once again that is entirely respon-
sible for the vanishing moments of . The above argument can where 2 because
also be applied to the synthesis side of the transform. In other (stability) and ( is of maximum
words, if the dual scaling function can be factorized as order ).
—meaning that is of order —then the synthesis The transfer function of the analysis wavelet is obtained by
wavelet will have exactly vanishing moments taking the Fourier transform of (8) with and :
(general biorthogonal case). When the analysis and synthesis
spaces are identical (orthogonal and semi-orthogonal cases), the (26)
number of vanishing moments are the same on both sides. Its low-frequency response obviously depends on the behavior
of near the origin.
E. Multiscale Differentiator Theorem 7: Let and be two valid biorthogonal scaling
functions with and continuous at . Then, is
We will now show that another consequence of the B-spline of order (i.e., ) if and only if
factor is that the analysis wavelet essentially behaves like
a th-order differentiator. The proof of this property rests al-
most entirely on the perfect reconstruction property of the fil-
terbank in Fig. 1. In the sequel, we will assume that the four The proof is given in the Appendix. If we now assume that
filters , , , and are stable in the sense is continuous as well, we can easily obtain the asymp-
that their Fourier transforms are bounded. We also recall that the totic version of the result by plugging (25) into (26) and making
perfect reconstruction property has an equivalent -formulation use of the property . This yields
that is expressed by four biorthogonality relations between the as
various filters pairs, which can be written in matrix form, cf.
[33] with (27)

The practical implication of Theorem 7 is that the wavelet

(22) transform acts as a multiscale differentiator. Specifically, the
wavelet coefficients will correspond to the samples of the th
derivative of a smoothed version of the input signal :

These are strong constraints that have a direct implication on the

form of the wavelet filter . where the smoothing kernel is defined by its frequency response
Proposition 6: Under the constraint of a stable perfect recon-
; it is necessarily lowpass and bounded
struction filterbank, the order condition (19) is equivalent to the
because of (27) and Theorem 7. This kernel essentially limits the
following factorization of the wavelet analysis filter:
bandwidth of the signal being analyzed with two practical ben-
efits. First, it regularizes the differentiation process by reducing
(23) its noise amplification effect, and second, it attenuates the signal
components above the Nyquist frequency so that the differenti-
where the filter is stable, i.e., . Another ated signal is well represented by its samples (or wavelet coef-
equivalent formulation is . ficients).
478 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 2, FEBRUARY 2003

In the classical case where is an integer, there is an with a new subscript notation that makes the order of the func-
equivalence between (27) and the vanishing moment property. tions explicit, e.g., . We then use this relation to compute
This can be shown by determining the Taylor series of the th derivative of for :
around : , where
is proportional to the magnitude of the first
nonvanishing moment. Using (15) with and and noting that
In the fractional case, there is no Taylor series of order and, , which is consistent with (10), we get
therefore, no direct equivalence with the vanishing moments (30)
property. The unique property of this kind of wavelets is that
they give access to fractional orders of differentiation, which This explicit time-domain differentiation formula is known for
can be very useful in some applications [34], in particular, when integer, but its extension for arbitrary is new to the best
dealing with fractal-like signals. of our knowledge. This result indicates that the function
generates a basis for representing the fractional derivatives of
IV. WAVELET DIFFERENTIABILITY AND INTEGRABILITY the scaling function and of the wavelet; indeed, by linearity, we
have that . The question
One of the primary reasons for the mathematical successes that then arises is the following: How far can we differentiate,
of wavelets is their stability with respect to differentiation and or equivalently, how much B-spline can we “peel off” before
integration. In the conclusion of his classical monograph, Meyer the so-called residue in (29) really blows up? What we
writes: “everything takes place as if the wavelets were mean by “blowing up” can be made mathematically precise by
eigenvectors of the differential operator , with corresponding requiring that some -norm of the residue remains finite.
eigenvalue ” [35]. We will now use our B-spline formalism This particular interpretation turns out to be intimately related
to shed some light on this important aspect of wavelet theory. to the concept of smoothness in a general -sense and leads to
As starting point, we use the B-spline factorization the following smoothness characterization theorem.
(20) together with the B-spline convolution property Theorem 8: If with ,
( ) to manipulate the basic scaling then , i.e., has derivatives in the sense.
function biorthogonality relation as follows: Proof: First, we rewrite (30) as

(28) where the coefficients of the finite difference operator are given
by (12). We also recall that , which implies
where denotes the anticausal B-spline of de- that the sequence is absolutely summable (i. e.,
gree . This formula suggests that the new pair of functions for ). We then use the above formula in conjunction with
( , ) should also generate a valid biorthogonal basis. In Minkowsky’s inequality to obtain the following bound for the
the sequel, we will show that these scaling functions play a cru- -norm of the derivative
cial theoretical role for they provide the building blocks for the
fractional integrals and fractional derivatives of the analysis and
synthesis wavelets, respectively. For this purpose, we will first
investigate the extent to which (or equivalently ) is differ- which proves the desired results.
entiable and propose a peeling formulation of smoothness that Theorem 8 provides an explicit link between the smoothness
provides some new insights on the various notions of wavelet properties of wavelets and the B-spline factorization. It is also
regularity. We will also prove that the B-spline component is en- interesting because it yields a general and coherent approach
tirely responsible for the smoothness of the basis functions. Fi- to the concept of smoothness, i.e., fractional differentiability in
nally, in Section IV-C, we will specify the biorthogonal wavelet the -sense. For , the present definition of smoothness is
basis that is associated with the fractional differentiation oper- equivalent to the widely used Sobolev regularity [36]. Another
ator. interesting case is because it penalizes the worst case
(max norm); this is very close to Hölder regularity, even though
A. Smoothness: Peeling Theory the latter is a measure of continuity rather than of differentia-
In this subsection, we characterize the fractional derivatives bility [19], [37].
of the scaling function and show that the presence of the Our peeling theory of smoothness is illustrated in Fig. 6 for
B-spline component is absolutely necessary for these to be the case of Daubechies’ scaling function of order 2. It is clear
well-defined. The argument is entirely based on the convolu- from the graph that the residue becomes rougher as a
tion property of B-splines (cf. Section II-C). Specifically, we higher order B-spline component gets pulled out. These various
rewrite the B-spline factorization (20) in terms of the function plots were obtained by running a Fourier version [38] of the cas-
that are already encountered in (28): cade algorithm with ten levels of iteration on the residual factor
of the refinement filter corresponding to in (29). Past the
where limiting case ( ) [37], is no longer bounded, and it
(29) does not make much sense to attempt to represent it graphically.
UNSER AND BLU: WAVELET THEORY DEMYSTIFIED 479

Fig. 6. Residual factor ' (x) in (29) as a function of r for Daubechies’ scaling function of order = 2. The critical value r = 0:55 corresponds to the
limiting case where ' (x) is bounded (r derivatives in L -sense). It also yields the Hölder exponent.

Note that there are computational techniques for determining follows from the Sobolev inequality (for compactly supported
this critical Hölder exponent (cf. Section IV-B). functions) , where is the Hölder
To get some further insight on our notion of smoothness, exponent of . In fact, our bound is sharp, as demonstrated by
we now consider the example of the fractional B-splines. The the example below.
borderline cases here are the B-splines of negative degree; these The B-splines of order are very regular, but they fall short of
can be shown to belong to the following functional spaces: the maximum possible Sobolev smoothness by 1/2. Yet we can
perturb the B-splines to achieve the maximum possible smooth-
for and ness by taking in (19), where
is small but nonzero. Since we are dealing with a
For the fractional splines, we have , so that we can two-tap residual, the Hölder (alias ) and Sobolev (alias )
write (29) as , where is the exponents can be computed explicitly using the technique out-
corresponding residue. Applying Theorem 8, we conclude that lined in [19] and [37]. Specifically, we find that
has up to derivatives in the -sense with
the strict equality only being reached for . Indeed, we
have already mentioned that is bounded ( derivatives
in ), which is consistent with the fact that is -Hölder for
continuous.
There is also a converse part to Theorem 8, albeit only within
the restricted framework of wavelets.
Theorem 9: If is a valid scaling function such that as
, then with .
The proof is more technical and is given in the Appendix . Thus, by letting tend to 0 , we are able to saturate the above
Since the order of the B-spline factor in Theorem 9 is inequalities such that , which proves that the
, the above statement is equivalent to saying that the order bound in Theorem 9 is sharp. If we now push it a little further
of a valid scaling function is necessarily greater or equal to by letting , we increase the B-spline factor by one which
its critical Sobolev exponent . In other words, we have makes us jump to , while the Hölder continuity
that , . This is a new re- remains at , but at the same time, the order is also
sult that extends a classical theorem in wavelet theory stating increased by one when compared with the previous degenerate
that -continuity ( integer) implies some minimum approx- case.
imation order: [5]. Note that our re- The functions that saturate the inequality (i.e., ) are
sult is slightly more generous (i.e., it yields more order); this the only ones for which the residual in (29) can be in . Indeed,
480 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 2, FEBRUARY 2003

if would have some residual Sobolev smoothness, it would be (30) for fractional differentiation, which is specific to scaling
possible, by Theorem 9, to factor out a larger order B-spline that (or refinable) functions.
would contradict the assumption that is of maximum order
. For all other cases, is not in and is typically only a C. Biorthogonal Wavelet Basis for Fractional Differentiation
distribution; for instance, in the case of The stability of wavelets with respect to differentiation is a
the B-splines. We can therefore safely state that in (20) has prerequisite for using them to characterize smoothness spaces,
no Sobolev regularity at all, which also means that there is no in particular, for proving that they provide unconditional bases
smoothness possible without the B-spline part of the wavelet. for Sobolev, Hölder, and Besov spaces [35]. Although these
While smoothness implies approximation order, there is gen- functional aspects of wavelet theory are beyond the scope of
erally no such implication in the reverse direction [39]. In other this paper, we want to make an interesting connection by iden-
words, the presence of the B-spline factor is not enough to guar- tifying and characterizing the biorthogonal wavelet basis asso-
antee that the scaling function is in , not to mention its differ- ciated with the fractional differentiation operator.
entiability. Indeed, one can conceive of very irregular wavelets The idea is simple and is based on the following manipulation
where the distributional part is so rough that it counterbal- of the wavelet biorthogonality relation:
ances the inherent smoothness of the B-spline factor. This is the
reason why smoothness is relatively difficult to control when (31)
applying conventional filter design procedures; it is a property
that is usually checked a posteriori. where denotes the anticausal fractional in-
tegration operator; it is the adjoint of , which
B. Determining Wavelet Smoothness is the inverse of the fractional differentiation operator . Based
on the results in Section IV-A, we obtain the explicit form of the
The peeling theory provides an interpretation of wavelet
“derivative” wavelet
smoothness that is appealing intuitively. However, we should
not be fooled by the apparent simplicity of the argument.
Wavelet smoothness is and remains one of the most difficult
theoretical aspects of wavelet theory. While Theorem 8 pro-
vides a mathematical criterion for testing differentiability in the where is defined by (28), and 2 .
-sense, it does not give a numerical method for determining We now close the loop by showing that the biorthogonal partner2
the critical exponents for a given filter . As far as we of in (29), that is, the scaling function
know, an exact computation is only possible for determining the
Sobolev index (i.e., -smoothness) of compactly supported
basis functions. The preferred method is based on the determi-
provides the complementary basis for expanding , which
nation of the spectral radius of the so-called transition operator;
is the -order fractional integral of the analysis wavelet. This
in practice, this amounts to computing the maximum eigenvalue
manipulation is performed in the Fourier domain starting from
of a reduced transition matrix associated with the residual filter
(26)
[6], [20]. Various techniques have also been proposed
to estimate the Hölder exponent (alias -smoothness) [37],
[19]. These methods are not exact anymore (except when is
symmetrical) but can provide tight upper and lower bounds.
Determining fractional orders of smoothness in norms
other than is more challenging mathematically, especially
since there is no single universal definition of fractional
differentiation in the time domain. Mathematicians have turned
the difficulty by testing the appartenance of the function to The key here is the result of Proposition 6, which allows us
some “smoothness” spaces. Villemoes [20] proposed to study to extract the relevant B-spline factor as long as . After
wavelet smoothness using Besov spaces, which leads to the inverse Fourier transformation, we end up with a stable, explicit
determination of a critical Besov exponent that is qualitatively representation
similar to the notion of -smoothness that we have considered
here. His approach is based on the fact that wavelets provide
unconditional bases for Besov spaces [35]. More recently,
Micchelli and Sauer have proposed to extend the Hölder where 2 . Indeed, we can invoke
notion of continuity to -space by introducing what they Young’s convolution inequality (cf. [42])
call generalized Lipschitz spaces [40]. Their formulation is
general but also rather involved—about 90 pages of dense
mathematics. By contrast, our approach to measuring wavelet
smoothness is much closer to a classical -Sobolev analysis, which proves that (since
except that the traditional method for is restricted to , ). Hence, we have established that the new wavelet
integer orders of differentiation. What makes the Sobolev 2This is a suggestive denomination that we are happy to borrow from
technique applicable here is our explicit time-domain formula Vaidyanathan [41].
UNSER AND BLU: WAVELET THEORY DEMYSTIFIED 481

Fig. 7. Summary of wavelet theory: main properties and equivalences.

pair ( , ) generates a biorthogonal basis that is in V. CONCLUSION

provided that and . Note that it is essential In this paper, we proposed a new spline-based interpretation
here to work with wavelets because the fractional integral of a and derivation of some fundamental wavelet properties. Our
scaling function is generally not in (since ). argumentation entirely rests on the representation of a scaling
Finally, by using the rescaling property function as the convolution between a B-spline and a distribu-
, we obtain the differentiation formulas for
tion: . The B-spline constitutes the regular part
the basis functions: of the wavelet, and its presence is necessary (and sufficient) for
the transform to have approximation order (cf. Theorem
and (32) 2). In particular, this implies that one cannot build a multireso-
lution wavelet basis of without any B-spline factor. We have
also shown that the key properties of B-splines (smoothness,
which are the “eigen-relations” to which Meyer was alluding, differentiability, reproduction of polynomial, etc.) get carried
even though he did not write them down explicitly. Of course, over to the scaling function in an essentially mechanical fashion
the qualifying statement is not rigorously correct—the im- through the convolution relation. These results and the various
portant point is that there are basis functions with the same equivalences that have been established are summarized in
wavelet structure on both sides of the identity. The practical Fig. 7. The advantage of approaching the problem from this end
relevance of these “differential” wavelets is that they give us a is that these basic properties are relatively easy to understand
direct way of gauging the fractional derivative of a signal based for the B-splines because of their simple analytical form in
on its wavelet coefficients in the original basis. Specifically, both the time and frequency domain. This particular point of
we have that 2 , view leads to the conclusion that the primary mathematical
which implies that the original wavelet coefficients divided by wavelet properties—order of approximation, reproduction
2 are the coefficients of in the modified wavelet basis of polynomials, vanishing moments, and smoothness of the
{ }. This provides a strong hint as to why the Sobolev basis function—are entirely due to the B-spline component.
norm of a signal can be measured from the -norm of its Another interesting property that has been explained from the
rescaled wavelet coefficients. The Besov case is analogous B-spline perspective is the stability of the wavelet basis with
with -norms being used instead; the argument there is more respect to fractional differentiation. Our key observation is
involved and relies on some Riesz-type -norm equivalences that there is a direct correspondence between the process of
[35]. In addition, note that the wavelets that have just been moving a B-spline factor from one side to another in a pair of
specified are fractional ones, which, in themseves, may serve biorthogonal scaling functions and the exchange of fractional
as an a posteriori justification for our extended formulation. integrals/derivatives on their wavelet counterparts.
482 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 2, FEBRUARY 2003

If these fundamental properties are to be attributed to the where we have defined 2 . In par-
B-spline part exclusively, then what is the purpose of the ticular, this implies that
distribution corresponding to the residual filter ? This
part is essential for imposing additional properties—such as
orthonormality—and, more important, to balance the local- We now prove that is integrable over [ ].
ization properties (size) of the analysis and synthesis filters in First, we observe that is integrable over any closed
Fig. 1. B-splines are optimal in terms of size and smoothness, subset of [ ] that does not include 0. This is because
but they are not orthogonal. To construct a pure spline wavelet is bounded over , which implies that is
transform, one needs to orthogonalize the B-splines [43], bounded by up to a multi-
[44] or to specify a dual pair of spline functions [45]–[47]; plicative constant. Now, because is in
in both cases, this is equivalent to selecting a distributional
part of the form , where is
an appropriate digital filter. However, the implementation of
these (semi-orthogonal) spline transforms requires IIR filters, is in and, thus, is in .
which is often considered a handicap. Early on, Daubechies Second, the continuity of —and of —at 0
and others [8], [9] have shown that the only way to construct implies that for any —here, we choose —we can
wavelets and basis functions that are compactly supported find such that for all , 2 .
on both sides (analysis and synthesis) is by careful selection Consequently
of the factor , eventually moving it to the analysis side,
which yields biorthogonal spline wavelets [8]. This explains
why all popular wavelet families (Daubechies, Coiflets, Then, we define , where 2
Cohen-Daubechies-Feauveau, 9/7, etc.) include nontrivial . are bounded quantities because the
distributional factors and/or , which we like to ’s are closed subsets of . We thus have
view as the price to pay for the additional but more classical 2 , i.e., 2 or, by summing up all the contri-
filter design constraints (e.g., FIR filterbank and orthogonality). butions

APPENDIX
PROOFS which is bounded, independently of . Letting tend to infinity,
A. Proof of Theorem 7 we conclude that is integrable in [ ] (Fatou’s the-
orem), in particular, in the neighborhood of 0. Thus, is
If is of order , then by Proposition 6. fully integrable in [ ].
This, together with the wavelet scaling (26) and the bounded- It is now a simple matter to state that the function
ness of (consequence of the Riesz condition), implies that is in . This is because
.
Conversely, if , then
. Now, using the assumption that in a
neighborhood of (consequence of the continuity of where is the same as defined by (33) and has just been
at since ) and the stability of the filter , shown to be integrable in [ ].
we can claim that . This is also equivalent to
(23), where is stable, as seen in the proof of Proposition REFERENCES
6. Thus, from Proposition 6, this is equivalent to the -order [1] P. P. Vaidyanathan, “Quadrature mirror filter banks, M-band extensions
property of . and perfect-reconstruction technique,” IEEE Acoust., Speech, Siganl
Process. Mag., vol. 4, pp. 4–20, July 1987.
[2] O. Rioul and M. Vetterli, “Wavelets and signal processing,” IEEE Signal
B. Proof of Theorem 9 Processing Mag., vol. 8, pp. 11–38, Oct. 1991.
[3] M. Vetterli, “A theory of multirate filter banks,” IEEE Trans. Acoust.
We require one (very mild) technical assumption, namely, Speech Signal Processing, vol. ASSP-35, pp. 356–372, Mar. 1987.
that the refinement filter be continuous at . To [4] S. G. Mallat, “A theory of multiresolution signal decomposition: The
prove Theorem 9, we define wavelet representation,” IEEE Trans. Pattern Anal. Machine Intell., vol.
11, pp. 674–693, July 1989.
[5] I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA: SIAM,
1992.
(33) [6] G. Strang and T. Nguyen, Wavelets and Filter Banks. Wellesley, MA:
Wellesley-Cambridge, 1996.
[7] S. G. Mallat, “Multiresolution approximations and wavelet orthogonal
bases of L (R),” Trans. Amer. Math. Soc., vol. 315, no. 1, pp. 69–87,
This is a true function (possibly taking the value for some 1989.
) because it is a sum of positive functions. We observe that [8] A. Cohen, I. Daubechies, and J. C. Feauveau, “Bi-orthogonal bases of
because of (6), satisfies the scaling relation compactly supported wavelets,” Commun. Pure Appl. Math., vol. 45, pp.
485–560, 1992.
[9] I. Daubechies, “Orthogonal bases of compactly supported wavelets,”
(34) Commun. Pure Appl. Math., vol. 41, pp. 909–996, 1988.
UNSER AND BLU: WAVELET THEORY DEMYSTIFIED 483

[10] S. Mallat, A Wavelet Tour of Signal Processing. San Diego, CA: Aca- [38] T. Blu and M. Unser, “The fractional spline wavelet transform: Defi-
demic, 1998. nition and implementation,” in Proc. Int. Conf. Acoust., Speech, Signal
[11] M. Unser and T. Blu, “Fractional splines and wavelets,” SIAM Rev., vol. Process., Istanbul, Turkey, 2000, pp. 512–515.
42, no. 1, pp. 43–67, Mar. 2000. [39] A. Cohen, I. Daubechies, and A. Ron, “How smooth is the smoothest
[12] Y. Katznelson, An Introduction to Harmonic Analysis. New York: function in a given refinable space?,” Appl. Comput. Harmon. Anal.,
Dover, 1976. vol. 3, no. 1, pp. 87–90, 1996.
[13] I. Daubechies and J. C. Lagarias, “Two-scale difference-equations .1. [40] C. A. Micchelli and T. Sauer, “Regularity of multiwaveletes,” Adv.
Existence and global regularity of solutions,” SIAM J. Math. Anal., vol. Comput. Math., vol. 7, pp. 455–545, 1997.
22, no. 5, pp. 1388–1410, 1991. [41] P. P. Vaidyanathan and B. Vrcelj, “Biorthogonal partners and applica-
H
[14] G. Strang, “Eigenvalues of (# 2) and convergence of the cascade algo- tions,” IEEE Trans. Signal Processing, vol. 49, pp. 1013–1027, May
rithm,” IEEE Trans. Signal Processing, vol. 44, pp. 233–238, Feb. 1996. 2001.
[15] M. Unser, “Sampling—50 years after Shannon,” Proc. IEEE, vol. 88, [42] E. M. Stein and G. Weiss, Fourier Analysis on Euclidean
pp. 569–587, Apr. 2000. Spaces. Princeton, NJ: Princeton Univ. Press, 1971.
[16] I. J. Schoenberg, “Contribution to the problem of approximation of [43] G. Battle, “A block spin construction of ondelettes. Part I: Lemarié func-
equidistant data by analytic functions,” Quart. Appl. Math., vol. 4, pp. tions,” Commun. Math. Phys., vol. 110, pp. 601–615, 1987.
45–99, 1946. [44] P.-G. Lemarié, “Ondelettes à localization exponentielle,” J. Math. Pures
[17] M. Unser, “Splines: A perfect fit for signal and image processing,” IEEE et Appl., vol. 67, no. 3, pp. 227–236, 1988.
Signal Processing Mag., vol. 16, pp. 22–38, Nov. 1999. [45] C. K. Chui and J. Z. Wang, “On compactly supported spline wavelets
[18] C. de Boor, A Practical Guide to Splines. New York: Springer-Verlag, and a duality principle,” Trans. Amer. Math. Soc., vol. 330, no. 2, pp.
1978. 903–915, 1992.
[19] O. Rioul, “Simple regularity criteria for subdivision schemes,” SIAM J. [46] M. Unser, A. Aldroubi, and M. Eden, “On the asymptotic convergence
Math. Anal., vol. 23, pp. 1544–1576, Nov. 1992. of B-spline wavelets to Gabor functions,” IEEE Trans. Inform. Theory,
[20] L. F. Villemoes, “Wavelet analysis of refinement equations,” SIAM J. vol. 38, pp. 864–872, Mar. 1992.
Math. Anal., vol. 25, no. 5, pp. 1433–1460, 1994. [47] , “A family of polynomial spline wavelet transforms,” Signal
[21] G. Strang, “Wavelets and dilation equations: A brief introduction,” SIAM Process., vol. 30, no. 2, pp. 141–162, Jan. 1993.
Rev., vol. 31, pp. 614–627, 1989.
[22] M. Unser, “Approximation power of biorthogonal wavelet expansions,”
IEEE Trans. Signal Processing, vol. 44, pp. 519–527, Mar. 1996.
[23] T. Blu and M. Unser, “Wavelet regularity and fractional orders of ap-
Michael Unser (M’89–SM’94–F’99) received the
proximation,” SIAM J. Math. Anal., submitted for publication.
M.S. (summa cum laude) and Ph.D. degrees in elec-
[24] G. Strang and G. Fix, “A fourier analysis of the finite element variational
trical engineering in 1981 and 1984, respectively,
method,” in Constructive Aspects of Functional Analysis. Rome, Italy:
from the Swiss Federal Institute of Technology
Edizioni Cremonese, 1971, pp. 793–840.
(EPFL), Lausanne, Switzerland.
[25] A. Ron, “Factorization theorems for univariate splines on regular grids,”
From 1985 to 1997, he was with the Biomedical
Israel J. Math., vol. 70, no. 1, pp. 48–68, 1990.
Engineering and Instrumentation Program, National
[26] T. Blu, P. Thévenaz, and M. Unser, “Complete parametrization of piece-
Institutes of Health, Bethesda, MD. He is now Pro-
wise-polynomial interpolation kernels,” IEEE Trans. Image Processing,
fessor and Head of the Biomedical Imaging Group at
submitted for publication.
EPFL. His main research area is biomedical image
[27] P. Thévenaz, T. Blu, and M. Unser, “Interpolation revisited,” IEEE
processing. He has a strong interest in sampling the-
Trans. Med. Imag., vol. 19, pp. 739–758, July 2000.
ories, multiresolution algorithms, wavelets, and the use of splines for image
[28] J. Shapiro, “Embedded image coding using zerotrees of wavelet coef-
processing. He is the author of 100 published journal papers in these areas.
ficients,” IEEE Trans. Acoust., Speech, Signal Processing, vol. 41, pp.
Dr. Unser is Associate Editor-in-Chief for the IEEE TRANSACTIONS ON
3445–3462, Dec. 1993.
MEDICAL IMAGING. He is on the editorial boards of several other journals,
[29] A. Said and W. A. Pearlman, “A new fast and efficient image codec
including IEEE SIGNAL PROCESSING MAGAZINE, Signal Processing, IEEE
based on set partitioning in hierarchical trees,” IEEE Trans. Circuits
TRANSACTIONS ON IMAGE PROCESSING (from 1992 to 1995), and IEEE SIGNAL
Syst. Video Technol., vol. 6, pp. 243–250, June 1996.
PROCESSING LETTERS (from 1994 to 1998). He serves as regular chair for the
[30] S. Jaffard, “Pointwise smoothness, two-microlocalization and wavelet
SPIE Conference on Wavelets, which has been held annually since 1993. He
coefficients,” Publicacions Matemàtiques, vol. 35, pp. 155–168, 1991.
was general co-chair of the first IEEE International Symposium on Biomedical
[31] S. Mallat and W. L. Hwang, “Singularity detection and processing with
Imaging, Washington, DC, 2002. He received the 1995 Best Paper Award and
wavelets,” IEEE Trans. Inform. Theory, vol. 38, pp. 617–643, Mar. 1992.
the 2000 Magazine Award from the IEEE Signal Processing Society.
[32] A. Arneodo, F. Argoul, E. Bacry, J. Elezgaray, and M. J. F., Ondelettes,
Multifractales et Turbulence. Paris, France: Diderot, 1995.
[33] M. Vetterli and C. Herley, “Wavelets and filter banks: Theory and de-
sign,” IEEE Trans. Signal Processing, vol. 40, pp. 2207–2232, Sept.
1992. Thierry Blu (M’96) was born in Orléans, France, in 1964. He received the
[34] F. J. Herrmann, “Singularity characterization by monoscale analysis: “Diplôme d’ingénieur” from École Polytechnique, Paris, France, in 1986 and
Application to seismic imaging,” Appl. Comput. Harmon. Anal., vol. 11, from Télécom Paris (ENST), in 1988. In 1996, he received the Ph.D degree in
no. 1, pp. 64–88, July 2001. electrical engineering from ENST for a study on iterated rational filterbanks ap-
[35] Y. Meyer, Ondelettes et Opérateurs I : Ondelettes. Paris, France: Her- plied to wideband audio coding.
mann, 1990. He is currently with the Biomedical Imaging Group, Swiss Federal
[36] T. Eirola, “Sobolev characterization of solutions of dilation equations,” Institute of Technology (EPFL), Lausanne, Switzerland, on leave from
SIAM J. Math. Anal., vol. 23, pp. 1015–1030, 1992. France Télécom National Center for Telecommunications Studies (CNET),
[37] I. Daubechies and J. C. Lagarias, “Two-scale difference-equations .2. Issy-les-Moulineaux, France. His research interests include (multi)wavelets,
Local regularity, infinite products of matrices and fractals,” SIAM J. multiresolution analysis, multirate filterbanks, approximation and sampling
Math. Anal., vol. 23, no. 4, pp. 1031–1079, 1992. theory, mathematical imaging, etc.