See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/231222178

An Introduction to Wavelets

Book in Mathematics of Computation · January 1992

DOI: 10.2307/2153134

CITATIONS READS

3,749 10,508

1 author:

Charles K. Chui
Stanford University
508 PUBLICATIONS 22,865 CITATIONS

SEE PROFILE

All content following this page was uploaded by Charles K. Chui on 12 December 2016.

The user has requested enhancement of the downloaded file.

 An Introduction to Wavelets
Lectures delivered at Nicholas Copernicus University, Toruń

March 15th & 22nd, 2000

Richard M. Timoney
School of Mathematics, Trinity College, Dublin 2, Ireland.
Email: [email protected]

March 22, 2000

 Abstract

We introduce wavelets as a particular way of choosing bases in function spaces.

The concept of a multiresolution analysis provides a setting for constructing
certain wavelets and where practical algorithms have been developed. The con-
structions and the algorithms depend on a sequence of coefficients.
The term ‘wavelet’ is a relatively new term and most of the ideas are new, but
there is a very active interest in using wavelet techniques in applications. We will
mention some of these uses.
1 Introduction: A Review of Bases
In this section, we review briefly some concepts of basis and also recall the notion
of a Fourier series.

1.1 Vector spaces and algebraic bases

The most familiar context for bases is in linear algebra. If V is a vector space over
any field K, then a basis for V is a subset B ⊂ V such that every vector v ∈ V
can be expressed uniquely as a finitely nonzero sum
X
v= vb b (vb ∈ K∀b ∈ B, {b ∈ B : vb 6= 0} finite)
b∈B

Every vector space has a basis. We will refer to this type of basis as an alge-
braic basis when we need to make a distinction with other types of bases to be
introduced below.
Choosing a basis B for an n-dimensional vector space V and an ordering
B = {b1 , b2 , . . . , bn } for the basis is equivalent to choosing a linear isomorphism

: Kn → B
n
X
(α1 , α2 , . . . , αn ) 7→ αj bj
j=1

The fact that a given vector space will have many different bases (unless it is
{0}), can be viewed as an advantage because it often allows us to choose a basis
that adapts to the problem at hand. A simple example is the ability (in good cases)
to choose a basis of eigenvectors for a given linear transformation T : V → V .
Though bases always exist in theory, there are many infinite-dimensional cases
of interest where one cannot write down any basis explicitly.

1.2 Finite-dimensional inner product spaces and orthonormal

bases
Notation 1.2.1 From now on the symbol K is reserved for a field which can only
be either the reals R or the complex field C.

When given a finite-dimensional vector space V over K equipped with an

inner product h·, ·i: V ×V → K, it is very convenient to work with an orthonormal
basis B = {e1 , e2 , . . . , en } (satisfying hej , ek i = 0 for j 6= k and hej , ej i = 1 for
all j (1 ≤ j, k ≤ n)).

1
Wavelets 2

Then every vector can be expressed in terms of the basis in a computable way
n
X
v= hv, ej iej (v ∈ V ).
j=1

(Our notation is that inner products are linear in the first variable, and conju-
gate linear in the second variable in the complex case.)
An (ordered) orthonormal basis for a finite-dimensional inner product space
(V, h·, ·i) gives us an inner-product preserving linear isomorphism from Kn with
the standard (euclidean) inner product to V .
For hermitian linear operators T : V → V on a finite-dimensional inner product
space (hT v, wi = hv, T wi for all v, w ∈ V ) we can always find an orthonormal
basis of eigenvectors.

1.3 Banach spaces and Schauder bases

Recall that a Banach space consists of a vector space X over K equipped with
a norm k · kX so that X is complete in that norm (every Cauchy sequence in X
converges to a limit in X). Convergence in X is taken with respect to the metric
(or distance) associated with the norm d(x1 , x2 ) = kx1 − x2 kX .
With a norm we can contemplate infinite linear combinations, when we define
infinite sums via limits, as in
∞
X n
X
xn = lim xj
n→∞
n=1 j=1

Definition 1.3.1 A Schauder basis for a Banach space X is a sequence (en )∞n=1
of vectors in X such that every x ∈ X can be expressed uniquely as an ‘infinite
linear combination’
∞
X
x= x n en (with xn ∈ K∀n)
n=1

Examples 1.3.2 (i) The classical sequence spaces c0 and `p (1 ≤ p < ∞) have
as bases the ‘standard basis’ (en )∞
n=1 , where

en = (δnj )∞
j=1

(δnj = 0 if n 6= j while δnn = 1 for all n).

Wavelets 3

`p (1 ≤ p < ∞) is the space of all sequences (αj )∞

j=1 (αj ∈ K∀j) such that

∞
!1/p
X
(αj )∞
j=1 p
= |αj |p < ∞.
j=1

c0 consists of all sequences (αj )∞

j=1 such that limj→∞ αj = 0 with the supre-
mum norm
(αj )∞
j=1 ∞ = sup |αj |
1≤j<∞

(ii) The function spaces Lp [0, 1] and Lp (Rn ) are defined (somewhat) similarly
to `p except that they are (almost everywhere equivalence classes of) mea-
surable functions f on the domain which have |f |p integrable with respect
to Lebesgue measure.

Bases are not such a convenient idea for general Banach spaces. Only sepa-
rable Banach spaces can have a basis and many separable Banach spaces fail to
have a basis.
In general the order of summation in the infinite linear combination is impor-
tant, and this inhibits one from considering a more general kind of basis where
uncountably many basis vectors could be allowed. To remove dependence on the
order, one can consider unconditional bases, but even fewer Banach spaces have
these than have Schauder bases.
As in the vector space case, things are considerably simpler if one assumes that
the there is an inner product. However, in the vector space case inner products can
always be chosen (not necessarily in a useful and natural way) but in the Banach
space case the existence of a compatible inner product is a severe restriction.

1.4 Hilbert spaces and orthonormal bases

Recall that a Hilbert space is an inner product space (H, h, ·, ·iHp
) which is com-
plete (a Banach space) is the associated norm given by kxkH = hx, xiH .
Common examples are L2 [0, 1] and L2 (Rn ) with the inner product given by
Z
hf, gi = f ḡ

(integrals with respect to Lebesgue measure).

Every Hilbert space H has an orthonormal basis in a sense that involves con-
vergence. In the Hilbert space case an orthonormal basis may be defined as a
maximal orthonormal subset B ⊂ H — a set consisting of unit norm pairwise
Wavelets 4

orthogonal elements of H with the property that it is not a proper subset of any
other such set.
Given any orthonormal basis B for H, we can write every x ∈ H as
X
x= hx, biH b
b∈B

in the sense that there are at most a countable number of nonzero terms in the
summation and for any enumeration

{b ∈ B : hx, biH 6= 0} = {b1 , b2 , . . .}

of the nonzero terms, we have

n
X
x = lim hx, bj iH bj .
n→∞
j=1

(No limit is needed if there are only a finite number of nonzero terms.)
Moreover, we have a convenient representation of the norm on H in terms of
the basis coefficients sX
kxkH = |hx, biH |2
b∈B

1.5 Fourier series

From now on, we will typically consider C-valued function spaces.

Example 1.5.1 For H = L2 [0, 1] there is a very simple orthonormal basis that is
so frequently used that it is almost the standard basis. It is {en : n ∈ Z} where

en (t) = exp(2πint)

From the general theory, we know that every f ∈ L2 [0, 1] can be expressed as
a sum X
f= hf, en ien
n∈Z
2
with convergence of the sum in L -norm. This is normally known as the Fourier
series of the function f and we often write
Z 1
fˆ(n) = hf, en i = f (t) exp(−2πint) dt.
0
Wavelets 5

It is also common to write

n
X
Sn f (t) = fˆ(j) exp(2πijt)
j=−n

for certain (symmetrical) partial sums of the series and the general Hilbert space
theory tells us that for all f ∈ L2 [0, 1] we have

lim kf − Sn f k2 = 0
n→∞

but it is a much deeper result due to Carleson that we also have almost everywhere
convergence of Sn f (t) to f (t).

For f ∈ Lp [0, 1] and 1 < p < ∞ one also knows that

lim kf − Sn f kp = 0
n→∞

so that the exponentials en also form a Schauder basis for Lp [0, 1].
They do not form a basis for L1 [0, 1] or for C[0, 1] = the continuous functions
f : [0, 1] → K (with the supremum norm). What we can say however is that for
f ∈ L1 [0, 1], there is enough information in the Fourier coefficients (fˆ(n))n∈Z to
completely determine f , but it is difficult to determine whether a given sequence
(an )n∈Z is the sequence of Fourier coefficients of some (unknown) function f ∈
L1 [0, 1].

Remark 1.5.2 Why do we use Fourier series?

One may justify the choice of the Fourier series example of an orthonormal
basis for L2 [0, 1] on the basis that it has proved its value over time, but one may
also argue that the complex exponentials en (t) = exp(2πint) are ‘eigenvectors’
of the differentiation operator dtd .
Perhaps there is a slight problem because the operator is not globally defined
: L [0, 1] → L2 [0, 1] and there are other eigenvectors eλ (t) = exp(2πiλt) with
2

λ ∈ C.
However, one may argue that the right context is to deal with is that of periodic
functions (having 1 as period). For this, we extend all functions f ∈ L2 [0, 1]
periodically to R by discarding the value at 1 and extending from [0, 1) to R using
def
period 1 (f (x) = f (x − [x]) with [x] the greatest integer ≤ x). At least the terms
of the Fourier series are naturally periodic in this way and the en are the 1-periodic
eigenvectors of differentiation.
In the context of periodic functions, one can argue that we are actually dealing
with functions on the unit circle of the complex plane T = {exp(2πit) : t ∈
Wavelets 6

R} = {exp(2πit) : t ∈ [0, 1)} and in this case the complex exponentials en may
be viewed as the irreducible unitary representations of T.
There is a generalisation of the Fourier theory to L2 (G), for G a locally com-
pact abelian group. We define the L2 space with respect to Haar measure on G
(normalised to give G measure 1 in the case G is compact) and then we have a
‘Fourier series’ representation of every f ∈ L2 (G) where the series is indexed by
the set Ĝ of irreducible unitary representations of G in place of Z.

2 Fourier Transform
We will continue in this section to consider C-valued function spaces. However, at
some places it will be convenient to assume that a given f ∈ L2 to be represented
is actually R-valued. Most of the remarks where this assumption is invoked can
be adapted by linearity or other means to the case of general complex valued f .
We progress to consider Fourier series where the period is not 1 and from
there, by a limiting argument, we arrive at the Fourier transform on R. We then
deal with some of the limitations of the Fourier transform as a preparation for
motivating the notion of a wavelet later.

2.1 Fourier series on general intervals

We can transfer the theory of Fourier series on the unit interval [0, 1] to any other
interval [a, b] (a < b ∈ R) by a simple change of variables. We have an isometric
(and inner product preserving) map

: L2 [0, 1] → L2 [a, b]
  
1 s−a
f 7→ s 7→ √ f
b−a b−a

and so a way to transfer Fourier series to L2 [a, b].

Specialising to the case [a, b] = [−T, T ] we have that for each g ∈ L2 [−T, T ],

(−1)n
  
X 2πins
g(s) = ĝ(n) √ exp
n∈Z
2T 2T

with
T
(−1)n
 
−2πins
Z
ĝ(n) = √ g(s) exp ds.
2T −T 2T
Wavelets 7

2.2 The Fourier transform

Combining these last two together, we can say
X 1 Z T 
−2πint
  
2πins

g(s) = g(t) exp dt exp
n∈Z
2T −T 2T 2T

for g ∈ L2 [−T, T ].
If we take g ∈ L2 (R) with compact support then we have
X 1 Z ∞ 
−2πint
  
2πins

g(s) = g(t) exp dt exp
n∈Z
2T −∞ 2T 2T

for all T large.

If we now define a function F(g) by
Z ∞
F(g)(ξ) = g(t) exp(−2πiξt) dt
−∞

we have  
X 1  n  2πins
g(s) = F(g) exp
n∈Z
2T 2T 2T
By treating this summation as a Riemann sum for an integral and taking the
limit as T → ∞ we can justify
Z ∞
g(s) = F(g)(ξ) exp(2πiξs) dξ (1)
−∞

for g ∈ L2 (R) compactly supported. In fact, the formula (1) holds for all g ∈
L2 (R).
The map F is called the Fourier transform on R and it can be proved (Parse-
val’s theorem) that F: L2 (R) → L2 (R) is an isometric isomorphism. The formula
(1) is the Fourier inversion formula, which exhibits the inverse transform as being
almost of the same form as F (in fact it is the adjoint of F).
In the Fourier series case, functions on [0, T ), or periodic functions with pe-
riod T , are exhibited as superpositions of exponentials t 7→ exp(2πnt/T ) (with
periods a multiple of T ). For the infinite line, we no longer have this granularity of
the periods and we need almost all possible periods 1/ξ for the Fourier inversion
formula. We can argue that the continuous range of periods used implies that the
summation in Fourier series becomes an integral in the case of R. But we can still
think of every f ∈ L2 (R) as being given by a ‘superposition’ of exponentials.
Wavelets 8

We can view these exponentials as being all the bounded eigenfunctions of

the derivative operator on R and this gives a clue to important applications of the
Fourier transform in differential equations.
We can also view the Fourier transform from the group-theoretical point of
view. Then the dual group of R (that is, the space of irreducible unitary represen-
tations) is again R — if we identify ξ ∈ R with the one dimensional representation
of R given by its matrix as t 7→ exp(2πiξt).
For future reference note that t 7→ exp(2πiξt) is periodic with period 1/|ξ|.
We can say it repeats |ξ| times when t increases by 1, and this justifies saying that
it has frequency |ξ| (measured in cycles or repetitions per unit of t).

2.3 Shannon-Nyquist sampling

For applications to digital communication, it is important to consider sending a
signal over a channel which is limited in frequency range. One can communicate
over such a channel only signals f (t) with the property that the Fourier transform
Ff is supported inside the range of the channel. By a simple phase change of
the signal (multiplying by a suitable complex exponential exp(2πiξ0 t)), we may
assume that the channel can carry frequencies in the range [−W/2, W/2].
Now, if we take g = Ff , we can reconstruct g from its Fourier series
  
X 1 2πinξ
g(ξ) = ĝ(n) √ exp (|ξ| < W )
n∈Z
W W

where1
W/2  
−2πins
Z
1
ĝ(n) = √ g(s) exp ds
W −W/2 W
Z W/2  
1 −2πins
= √ F(f )(s) exp ds
W −W/2 W
Z ∞  
1 −2πins
= √ F(f )(s) exp ds
W −∞ W
√
(using the fact that F(f )(s) = 0 for |s| > W/2). This should be (1/ W )f (−n/W )
by the Fourier inversion formula. However the Fourier inversion formula is a
formula in L2 (R) and so cannot be applied pointwise in general. It turns out
(see below) that band-limited functions are automatically continuous, and then
1
Strictly speaking we should have some factors (−1)n in ĝ(n) and also on the exponentials
above, but we have canceled these signs.
Wavelets 9

we can justify
√ this pointwise application of the inversion formula. We can show
ĝ(n) = (1/ W )f (−n/W ). Thus the Fourier series formula says
X  −n   1 
2πinξ

F(f )(ξ) = f exp (|ξ| < W )
n∈Z
W W W

and using the Fourier inversion formula (plus the assumption that f is band-
limited) we can show the following result.

Theorem 2.3.1 If f ∈ L2 (R) is band-limited so that F(f ) is supported in [−W/2, W/2],

then f is completely determined by its values f (n/W ) (n ∈ Z) and in fact
∞
X sin(π(n − W t))
f (t) = f (n/W )
n=−∞
π(n − W t)

This can be interpreted to mean that if a channel is band-limited to a frequency

band of total width W (or limited by |ξ| < W/2), then we cannot transmit a
continuous function along the channel, but only one value every time interval
1/W .

2.4 Compact support

We can be tempted from looking at band-limited functions (compact support of
the Fourier transform) to consider compactly supported functions, but we cannot
have both compact support for f and F(f ) simultaneously. This is a consequence
of the following theorem.

Theorem 2.4.1 (Paley-Wiener) If f ∈ L2 (R) has compact support then its Fourier
transform F(f )(ξ) (ξ ∈ R) extends to be an analytic function ζ 7→ F(f )(ζ): C →
C (or an entire function).
In fact this entire function must be of exponential type, that is it must satisfy

|F(f )(ζ)| ≤ AeB|ζ| (ζ ∈ C)

for some constants A, B ≥ 0.

Moreover, the (restrictions to R of) entire functions of exponential type are
exactly the Fourier transforms of compactly supported functions in L2 (R).

A similar result applies to the inverse Fourier transform, so that the band-
limited functions are those that are restrictions to R of entire functions of expo-
nential type.
Wavelets 10

As an entire function cannot be zero on any interval of positive length, unless it

is identically zero, it follows that f and F(f ) cannot both be compactly supported.
A quantitative form of this fact can be stated for functions which are somewhat
localised in time and have Fourier transforms that are somewhatRlocalised as well.
For f ∈ L2 (R) normalised to have unit norm (kf k2 = 1 or R |f (t)|2 dt = 1)
we can treat |f (t)|2 as a probability density function on R. Then we can try to
compute the mean µ and variance σ 2 of this probability distribution.
Z Z
2
µ= t|f (t)| dt, σ = (t − µ)2 |f (t)|2 dt
2
R R

We will certainly be able to do this for compactly supported f ∈ L2 (R), but

we will also succeed for more general (normalised) f ∈ L2 (R).
As kF(f )k2 = kf k2 = 1, we can also contemplate the mean µ̂ and variance
(σ̂)2 of |F(f )|2 .

Theorem 2.4.2 (Heisenberg uncertainty principle) If we have f ∈ L2 (R) and

also f 0 (t), f 00 (t), tf (t), t2 f (t) ∈ L2 (R) and kf k2 = 1, then
1
σσ̂ ≥
4π
We omit the proof of this, though it requires only some simple properties of
the Fourier transform, the Cauchy-Schwarz inequality and integration by parts.
See [1, Appendix F] for a proof.

2.5 Windowed Fourier transforms

A drawback of both Fourier series and Fourier transforms is that they destroy local
information. Both allow reconstruction of functions in L2 (and Lp for certain p,
though the reconstruction is not possible by exactly the same elegant inversion
formula). But it is part of the nature of the Fourier transform that the whole
transform is needed to recover the function.
A simple example is provided by the characteristic functions of intervals [0, a).
By a simple computation the Fourier transform is

1 − exp (−2πiaξ)
F(χ[0,a) )(ξ) =
2πiξ
and we can see that a small shift in a affects the Fourier transform at each ξ.
In particular a relatively small localised change in the function requires a re-
calculation of the Fourier transform/series and the whole transform is likely to be
Wavelets 11

altered by such a change. Similarly a small local change on the frequency side (a
small change of the Fourier transform) will normally affect the whole function.
One may analyse a signal or a sound wave such as a piece of music played over
a certain time by taking its Fourier transform. First we must consider the signal
(or sound wave) to extend over infinite time (perhaps by extending it with zero
backwards in time to −∞ and forwards in time to +∞). One may consider (and
this is typically done) the absolute value of the Fourier transform of the signal at
frequency ±ξ to be a measure of the amount of the total amount of frequency |ξ|
present in the signal. (Assuming that the signal is a real one, we should combine
F(f )(ξ) exp(2πiξt) + F(f )(−ξ) exp(−2πiξt) to get A(ξ) cos(2πξ(t − t0 )) with
amplitude A(ξ) = 2|F(f )(ξ)| and phase t0 .)
For example in the case f (t) = χ[0,a) (t), we get
sin(πaξ) a 1
A(ξ) = 2 , t0 = − .
πξ 2 4ξ
The phase can be thought of loosely as representing a time shift in frequency
ξ, but it is not easy to interpret it directly as related to a beginning time as one
must consider the cancellation between all the terms to recover the signal. Thus
the Fourier transform of the sound wave from a piece of music must contain all
the notes as the original sound wave can be recovered from it. However, there is
no simple way to detect from the Fourier transform which notes were played at
which times.
In the L2 setting (or even L1 ) we cannot handle a sound wave that consists of
a single note played with constant intensity for all time, but if we considered a
suitable generalisation (distributions) where such signals could be handled, then
the Fourier transformed signal would be concentrated at that one frequency. In
this way we can justify the interpretation of the amplitude as the ‘amount’ of a
given frequency present in a signal.
In an effort to also get a hold of local information, the windowed Fourier
transform takes the Fourier transform of many localised versions of the original
signal. We take a particular fixed function g of compact support, such as g(t) =
χ[−1,1] or a smoother version. Then we translate g by arbitrary amounts to get
t 7→ g(t − a) and take the Fourier transform of f (t)g(t − a). We get a transform
for each translation amount a
Z
WF (f )(a, ξ) = F(f (t)g(t − a))(ξ) = f (t)g(t − a) exp (−2πiξt) dt
R

The definition of WF depends on the choice of g and g is known as the win-

dow function. One may find strange effects caused by discontinuities or lack of
smoothness of g and so choosing a smooth compactly supported window function
is desirable as a rule.
Wavelets 12

The resulting WF does give local information about the function f , at the
expense of introducing a ‘position’ parameter a. For a fixed a we must be able
to find from WF(a, ·) all the information about f in the support of t 7→ g(t − a)
(which is the support of g translated by a).
However, the information is still not very easy to decode. We can say that
WF(f )(a, ξ) represents the amount of frequency ξ present in the graph of f in-
tersection the window around a, though there will be an effect from the shape of
the graph of g also.
We can argue that WF must be limited in its ability to localise because of
the Heisenberg uncertainty principle. Applying the window to get f (t)g(t − a)
certainly gives us a windowed signal with relatively small standard deviation σ,
but we do not always get good control on σ̂, the standard deviation on the trans-
form side. At the best, we could generate window functions where σ̂ would be
proportional to the reciprocal of the width of the window.
One might argue that for a relatively high frequency ξ where the corresponding
period 1/ξ is smaller than the window, we are in a reasonably good position to
detect whether there is a frequency ξ component inside the window. For relatively
small frequency ξ, and large period, we do not really have a big enough window
on the graph of f to say that we are detecting any variation in the graph at that
frequency.
Of course, when we take the Fourier transform, we do see something at all fre-
quencies, including small frequencies, but we are really detecting features caused
by the zero-extension of the graph of f (t)g(t − a) past the window.
At high frequencies, small periods, we have room for several oscillations in-
side the window and we are now back to a similar position to that we had with the
Fourier transform F(f )(ξ) where we are not able to say where within the window
the high frequency change in f happens.
If we managed to optimise σ̂ as roughly proportional to the reciprocal of the
width of the window, we would be able to do quite a good job of pin-pointing the
position of reasonably low frequency components of f , but for higher frequencies
we could not expect to pin down the location except within a multiple of the
period, a high multiple in the case of high frequencies.

3 Wavelets via Frames

We continue the ideas introduced in studying the windowed Fourier transform to
get one approach to wavelets, perhaps the most general approach. Later we will
take a more restrictive setting where there are better computational algorithms
available.
Wavelets 13

3.1 Continuous wavelet transform

If we take the discussion above a step further, it suggests that if we want to know
where we can find a component of frequency ξ in the graph of f , we should not
expect to be able to answer unless we can consider a section (or window) of the
graph of length comparable to the period 1/ξ.
We then take a window such as

sin(2πt) −1/2 < t < 1/2
ψ(t) =
0 t ≥ 1/2 or t ≤ −1/2

but we could
2
R take any ψ. Normally we assume that ψ is compactly supported, ψ ∈
L (R) and R ψ(t) dt = 0. We should think of the graph of ψ as a single ‘cycle’
of a more or less periodic wave, but we will often want additional properties such
as smoothness of ψ which will make this only approximately correct.
Instead of translating the wavelet around as we did for windows, we also scale
the wavelet to give stretched versions of the original wavelet ψ with the same basic
shape but at different scale or frequency.
We define
 
1 t−b
ψa,b (t) = √ ψ (a > 0, b ∈ R)
a a
and this will give us a single ‘wavelet’ with support stretched to a of its previous
√ frequency 1/a times the original frequency in some sense. (The factor
length,
1/ a is not really essential, but it is there to preserve L2 norms — kψa,b k2 =
kψk2 .)
Instead of imposing a further frequency on this, we just consider its inner
product with a given f .

Definition 3.1.1 The continuous wavelet transform of f ∈ L2 (R) corresponding

to a choice of ‘wavelet’ ψ is
 
t−b
Z
1
W (f )(a, b) = hf, ψa,b i = f (t) √ ψ dt (a > 0, b ∈ R).
R a a
The key to the continuous wavelet transform is that we can choose ψ so that
W (f )(a, b) contains enough information to reconstruct the function f . The sim-
plest example where this is possible is called the Haar Wavelet

 1 0 ≤ t < 1/2
ψ(t) = −1 1/2 ≤ t < 1
0 t ≥ 1 or t < 0

Wavelets 14

Theorem 3.1.2 (Inverse wavelet transform) If ψ ∈ L1 (R)∩L2 (R) is real-valued

and satisfies the admissibility condition
Z ∞
|F(ψ)(ξ)|2
cψ = dξ < ∞
0 ξ

then for f ∈ L2 (R)

Z 1/2
√ 2da
kf k2 = cψ |W (f )(a, b)| 2 db
(a,b)∈(0,∞)×R a

For f ∈ Lp (R) (1 < p < ∞)

Z
1 da
f (t) = W (f )(a, b)ψa,b (t) 2 db
cψ (a,b)∈(0,∞)×R a

(if the integral is interpreted in a distributional sense).

We will not try to prove this result.

Notice that the Haar wavelet has Fourier transform
(1 − exp(−πiξ))2
F(ψ)(ξ) =
2πiξ
and it is admissible.
It is easy to see that a compactly supported C ∞ function ψ with R ψ(t) dt = 0
R

is ‘admissible’ because the Fourier transform of any such function will be analytic,
have a zero at ξ = 0 and decay faster than any power of 1/|ξ| as |ξ| → ∞.

3.2 Discretisation of the CWT

In the case of the Haar wavelet at least, it is easy to see that there is a great
redundancy of information in W (f )(a, b). Among the ψa,b , those with a = 1
contain an orthonormal set in L2 (R):

{ψ1,m : m ∈ Z} = {χ[m,m+1/2) − χ[m+1/2,m+1) : m ∈ Z}

This is clearly not an orthonormal basis because no function can be expressed in

terms of these unless it is constant on the intervals with endpoints at adjacent half
integers.
However, the orignal work of Haar showed that

{ψ2n ,2n m : n, m ∈ Z}
Wavelets 15

does form an orthonormal basis for L2 (R). Thus we can express every f ∈ L2 (R)
as X X
f= hf, ψ2n ,2n m i ψ2n ,2n m = W (f )(2n , 2n m)ψ2n ,2n m
n,m∈Z n,m∈Z

from the general theory of orthonormal bases in Hilbert spaces.

This has been known and used by various authors since Haar invented it, but
it has not had so many applications because the basis functions are not continuous
and not well suited to many applications.
What is new in the wavelet context is that we can use other admissible ψ
which have good behaviour in many ways. We can even find much more regular
(for example continuous or continuously differentiable) ψ so that
{ψ2n ,2n m : n, m ∈ Z}
forms an orthonormal basis for L2 (R).
For some purposes, we might be happy with less, a discrete set of points
(an , bm ) which are sufficient to recover f from W (f )(an , bm ).

3.3 Frames
There are a number of ways to generalise the concept of an orthonormal basis in
a separable Hilbert space.
One generalisation, known as a Riesz basis, allows for a sequence of vectors
xn ∈ H with the property that they form aP Schauder basis for H and so that for
any convergent infinite linear combination n αn xn (αn ∈ K for all n) we have
2 2
X X X
2
A αn xn ≤ |αn | ≤ B αn xn
n H n n H

for some fixed A, B > 0.

Every Riesz basis for H can be got by applying an invertible continuous linear
map T : H → H to some orthonormal basis of H and so we can view a Riesz basis
as an orthonormal basis which has been somewhat distorted.
The more general concept of a frame was first introduced by Duffin and Scha-
effer in 1952.
Definition 3.3.1 A sequence (xn )∞n=1 of elements of a Hilbert space H is called a
frame for H if there are constants A, B > 0 so that
∞
X
Akxk2H ≤ |hx, xn i|2 ≤ Bkxk2H
n=1

holds for all x ∈ H.

Wavelets 16

Examples 3.3.2 (i) A wavelet frame for L2 (R) is a frame of the type

{a−n/2 ψ(a−n t − mb)}n,m∈Z

for a fixed ψ ∈ L2 (R) and some a > 0, b 6= 0.

(ii) Relatively uninteresting examples of frames can be generated from orthonor-

mal bases e1 , e2 , . . . by listing each basis vector a finite number of times
in the frame. For example e1 , e1 , e2 , e2 , e3 , e3 , . . . (where each is repeated
twice) is a frame with A = B = 2.
A frame with A = B is called a tight frame and a frame which ceases to
be a frame on removal of one vector from the sequence is called an exact
frame. The frame generated from an orthonormal basis by
1 1
√ e1 , √ e1 , e2 , e3 , e4 , . . .
2 2
is tight with A = B = 1 but is not exact.

To make a connection between frames and bases, we introduce the coefficient

operator U : H → `2 associated with a frame x1 , x2 , . . . in H given by

U x = (hx, xn i)∞
n=1 ∈ `
2

What we know is that

Akxk2H ≤ kU xk22 = hU x, U xi = hU ∗ U x, xiH ≤ Bkxk2H

and this implies that U ∗ U : H → H is a positive bounded linear operator on H

(in the sense of positive definiteness that hU ∗ U x, xiH ≥ 0 for all x ∈ H). Also
U ∗ U satisfies the inequalities2

A idH ≤ U ∗ U ≤ B idH

in the sense that the differences U ∗ U − A idH and B idH − U ∗ U are positive
operators. Moreover we can compute
∞
X
U ∗U x = hx, xn ixn
n=1

We call U ∗ U the frame operator and denote it by S.

2
idH denotes the identity operator : H → H
Wavelets 17

The frame operator S is invertible and if x ∈ H we can show

∞
X ∞
X
x= hS −1 x, xn ixn = hx, xn iS −1 xn
n=1 n=1

With a frame we do not have to have linear independence in general, but we

can show an optimality property for the `2 norm of the coefficients in the above
canonical expansion of x in terms of xn (with coefficients hS −1 x, xn i). If x =
P ∞
n=1 αn xn for some scalars αn , then

∞
X ∞
X ∞
X ∞
X
|αn |2 = |hS −1 x, xn i|2 + |hS −1 x, xn i − αn |2 ≥ |hS −1 x, xn i|2
n=1 n=1 n=1 n=1

If we have a tight frame (A = B) then S = A idH and

∞
1X
x= hx, xn ixn
A n=1

By scaling the frame, we can arrive at one where A = 1. Every x ∈ H has a

representation in terms of such a frame which has many of the properties of an
expansion in terms of an orthonormal basis, except uniqueness.
For non-tight frames, the necessity to compute S −1 to arrive at a concrete
representation for x ∈ H is an obstacle to the practical use of such frames.

Proposition 3.3.3 Suppose ψ ∈ L1 (R) has the property that F(ψ)(ξ) has no
zeros for 1 < |ξ| < k for some k > 2. Then the set

{2n/2 ψ(2n t − 2n m) : n, m ∈ Z}

is a frame in L2 (R).
Wavelets 18

4 Wavelets via Multiresolution Analysis

In this section we take a more restrictive approach to wavelets, which leads to a
more effective approach in practice than the frame-based approach.

4.1 Multiresolution analysis on R

Looking at the Haar wavelet example, we have the discrete version where every
f ∈ L2 (R) can be expressed
X
f= hf, ψ2n ,2n m i ψ2n ,2n m
n,m∈Z

ψ2n ,2n m (t) = 2n/2 χ[2−n m,2−n (m+1/2) − χ[2−n (m+1/2),2−n (m+1)

For a fixed n we see functions constant on intervals of length 2−n−1 (and all the
functions have integral 0). Recall that we can have n positive and negative so that
we get short as well as long intervals.
If we take

Wn = span{ψ2−n ,2−n m : m ∈ Z} Vn = span{ψ2−k ,2−k m : k, m ∈ Z, k < n}

(we take the closed linear spans), then we can deduce from orthonormality and
the fact the the ψ2n ,2n m form a basis that
\
{0} = Vn ⊂ · · · ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ · · ·
n∈Z
[
⊂ Vn dense in L2 (R) (2)
n∈Z
f (t) ∈ Vn ⇐⇒ f (2t) ∈ Vn+1 (3)
Vn = Vn−1 ⊕ Wn
(orthogonal direct sum) (4)
 
functions constant on intervals
Vn = span
of length 2−n starting at 0
 

 functions constant on intervals 

of length 2−n−1 starting at 0
 
Wn = span

 and which average 0 on the intervals 

of length 2−n
 

See Figure 1 for a graphical view of this. In Figure 2, we try to show what
happens when we build up linear combinations of ψ2n ,2n m starting with n = 1
Wavelets 19

A selection of graphs of ψ1,m , A selection of graphs of ψ1/2,b ,

m = −2, 0, 2 b = −1, 0, 1, 2
Figure 1: Haar wavelets at different scales

Figure 2: Combining Haar wavelets at finer scales

(constant on intervals of length 1) and then adding some more terms (with smaller
coefficients) with n = 0, −1, −2. The additional terms affect the details of the
graph. In fact, we show successive approximations to 2(x − 1)χ[0,1) (x), first using
ψ2,0 , then adding a combination of ψ1,m , next adding a combination of ψ1/2,(1/2)m ,
etc.
We take the point of view that Vn is what one can see if one is restricted to
taking averages over (dyadic) blocks of length 1/2n (blocks [k/2n , (k + 1)/2n )).
A sequence Vn of subspaces on L2 (R) with the properties (2) and (3) is called
a multiresolution analysis of L2 (R).
In the example at hand, we have one additional key property, that all the Vn
Wavelets 20

are spanned by the integer translates of one function. For example

V0 is spanned by the translates φ(x − m) (m ∈ Z) (5)

of the function φ(t) = χ[0,1) (t). In fact these form an orthonormal basis for V0
(which is even more convenient). Then property (3) implies that φ2−n ,2−n m (t) =
2n/2 φ(2n t − m) (m ∈ Z) form an orthonormal spanning sequence for Vn (all
n ∈ Z).
Moreover, the orthogonal complement Wn−1 of Vn−1 inside Vn is also spanned
by the translates of one function, {ψ2−n+1 ,2−n+1 m : m ∈ Z}.
It turns out that, in general, we can deduce that Wn is spanned by the translates
of one function based on the fact that (2), (3), (4) and (5) hold. We have
√
φ(t) ∈ V1 = span{φ1/2,m/2 (t) = 2φ(2t − m) : m ∈ Z}
√
and since the 2φ(2t−m) are orthonormal, this means we must be able to express
X X
φ= hφ, φ1/2,m/2 iφ1/2,m/2 = cm φ1/2,m/2 .
m∈Z m∈Z

Indeed, in the particular Haar case we were considering, we have

1 1
φ(t) = φ(2t) + φ(2t − 1) = √ φ1/2,0 (t) + √ φ1/2,1/2 (t)
2 2
and notice that we also have
1 1
ψ(t) = √ φ1/2,0 (t) − √ φ1/2,1/2 (t)
2 2
It turns out in general that something similar happens.

Theorem 4.1.1 Suppose (Vn )n∈Z is a sequence of subspaces of L2 (R) which form
a multiresolution analysis (that is, they satisfy (2) and (3))) and suppose that (5)
holds for a function φ with orthonormal translates φ(t − m) (m ∈ Z).

(i) Then we must have

X
|cm |2 = 1 for cm = hφ, φ1/2,m/2 i
m∈Z

and X
cm cm−2` = 0 for 0 6= ` ∈ Z (6)
m∈Z
Wavelets 21

(ii) If we define X
ψ= (−1)m c1−m φ1/2,m/2
m∈Z

then ψ ∈ V1 and it is orthogonal to V0 .

Moreover the translates ψ1,m (m ∈ Z) form an orthonormal basis for the
orthogonal complement of V0 in V1 and the whole collection of translated
and scaled functions
{ψ2n ,2n m : n, m ∈ Z}
based on ψ is an orthonormal basis for L2 (R).
R
(iii) If we assume that φ ∈ L1 (R) ∩ L2 (R), R φ(t) dt 6= 0 and {m : cm 6= 0} is
finite, then we have
X √
cm = 2 (7)
m∈Z
F(φ)(ξ) = P (ξ/2)F(φ)(ξ/2) (8)
X
where P (ξ) = cm e2πimξ
m∈Z
∞
Y
F(φ)(ξ) = F(φ)(0) P (ξ/2k ) (9)
k=1

Proof.

P− m) and the second by

(i) The first part follows from orthonormality of φ(t
expanding hφ(t), φ(t − `)i = 0 in terms of φ = m∈Z cm φ1/2,m/2 .

(ii) We can express

X
ψ(t − `) = (−1)m c1−(m+2`) φ1/2,m/2 (t)
m∈Z

and so expand
X
hφ(t), ψ(t − `)i = (−1)m cm c1−(m+2`) = 0
m∈Z

(since the m and 1 − m − 2` terms have opposite signs).

The orthonormality of the ψ1,m follows by (i).
To show that they span the orthogonal complement of V0 inP V1 , suppose f ∈
V
P1 is perpendicular to V0 and to all ψ1,m . We can write f = m∈Z fm φ1/2,m/2 =
m∈Z hf, φ1/2,m/2 iφ1/2,m/2 and let us denote the sequence of coefficients of
Wavelets 22

f by F = (fm )m∈Z . Suppose we also denote by Φ` = (cm−2`
)m∈Z the

coefficients of φ(t − `) = φ1,` and by Ψk = (−1)m c1−(m+2k) m∈Z the
coefficients of ψ(t − k) = ψ1,k (t) in the same basis for V0 .
Consider the matrix M that has as its columns

(. . . , Φ1 , Ψ1 , Φ0 , Ψ0 , Φ−1 , Ψ−1 , Φ−2 , . . .).

These columns are orthonormal by the earlier parts of the proof. The matrix
M represents a linear operator on `2 (where we take Z as the index set for
the sequence space) that transforms the standard basis to the orthonormal
sequence given by the columns. When we compute M ∗ M we get the iden-
tity matrix. But M ∗ F = 0 by assumption and so what we want is M M ∗ to
be the identity.
If we break M into 2 × 2 blocks Mrs (r, s ∈ Z) we get
 
c2r+2s c2s−2r+1
Mrs =
c2r+2s+1 −c2s−2r
and if we work out M M ∗ we get the matrix made up of the blocks
X
Mr,j (Ms,j )∗
j
X  c2r+2j c2j−2r+1   c2s+2j c2s+2j+1 
=
c2r+2j+1 −c2j−2r c2j−2s+1 −c2j−2s
j
 P 
j (c2r+2j c2s+2j + c c
P 2j−2r+1 2j−2s+1
)
j (c2r+2j c2s+2j+1 − c2j−2r+1 c2j−2s )
 
=  P
 

 j (c2r+2j+1 c2s+2j − c2j−2r c2j−2s+1 ) 
P
j (c2r+2j+1 c2s+2j+1 + c2j−2r c2j−2s )
 
δrs 0
=
0 δrs
(The off-diagonal sums rearrange to 0 and the orthogonality relations (6)
show that the diagonal entries are 0 unless r = s, when they are 1 by (i).)
Hence M M ∗ is the identity, F = M M ∗ F = 0 and so f = 0. Thus the
collection {φ1,m : m ∈ Z} ∪ {ψ1,m : m ∈ Z} is a maximal orthonormal
subset (an orthonormal basis) of V0 .

P to φ, it follows that hψ1,` , φ1,m i = hψ1,`−m , φi = 0

Since ψ1,` is orthogonal
for all m. As φ2,0 = m∈Z cm φ1,m it follows that hψ1,` , φ2,0 i = 0.
Iterating these ideas, we can show that ψ2n ,2n m are all orthonormal. Since
T
n∈Z Vn = {0} we can show that any f ∈ Vk which is orthogonal to all
Wavelets 23

ψ2n ,2n m (n > −k) must be zeroSand {ψ2n ,2n m : n > −k} is an orthonormal
basis for Vn . Using density of n∈Z Vn in L2 (R) we can show that ψ2n ,2n m
form an orthonormal basis of L2 (R).
P
(iii) Integrating both sides of φ = m∈Z cm φ1/2,m/2 gives (7). Taking Fourier
transforms of both sides gives (8). Iterating (8) gives
n
Y
F(φ)(ξ) = F(φ)(ξ/2n ) P (ξ/2k )
k=1

and letting n → ∞ gives (9).

4.2 Daubechies wavelets

By Theorem 4.1.1, we can find wavelets if we can find finitely nonzero sequences
of coefficients cm that satisfy the conditions of part (i) of that theorem and (7).
There is another step required, to show that the infinite product in (9) converges to
the Fourier transform of a function in L1 (R) ∩ L2 (R) with nonzero integral. This
step usually works because of results which we will not explain and the resulting
φ (and ψ) will be compactly supported in most cases.
The function ψ is the basic wavelet (sometimes called the ‘mother wavelet’)
resulting from the multiresolution analysis and the φ is a generating function that
is sometimes called the ‘father function’, but also known as the scaling function.

Examples 4.2.1 (i) If we allow only two non-zero coefficients c0 and c1 , then
the conditions we have force us to have the Haar wavelet situation.
Allowing a possible nonzero c2 as well as c0 and c1 still reduces to the Haar
wavelet situation.

(ii) Daubechies showed that there are a family of choices of coefficients c0 , c1 , . . . , c2p−1
(p ≥ 1) that lead to continuous compactly-supported φ (and continuous
compactly-supported wavelets). As p increases the corresponding wavelet
becomes more differentiable.
For p = 2 she exhibited
√ √ √ √
1+ 3 3+ 3 3− 3 1− 3
c0 = √ , c1 = √ , c2 = √ , c3 = √
4 2 4 2 4 2 4 2
Wavelets 24

4.3 Higher dimensions

Generalisations to Rd often require the use of 2d − 1 wavelets and not just one. So
3 wavelets for R2 , for example.
The simplest approach is to take a wavelet ψ arising from a multiresolution of
L2 (R) and its associated scaling function φ. Then we consider the wavelet basis
of L2 (R2 ) given by

{ψ2ws 2 sw ww
n ,2n m : n ∈ Z, m ∈ Z } ∪ {ψ2n ,2n m } ∪ {ψ2n ,2n m }

where

ψ ws (x, y) = ψ(x)φ(y), ψ sw (x, y) = φ(x)ψ(y), ψ ww (x, y) = ψ(x)ψ(y)

are ‘tensor products’ of the wavelet ψ in one variable with the scaling function φ
or with ψ in the other variable.
This corresponds to a multiresolution analysis of L2 (R2 ) which respects the
dilations x 7→ 2x and the translations x 7→ x + b with b ∈ Z2 .
One can replace the dilation x 7→ 2x by x 7→ Ax where A is a matrix with
integer entries which is expansive (kAn xk → ∞ as n → ∞ for each 0 6= x ∈ R2 ).
The number of wavelets required is then | det(A)| − 1 and this can be 1, for
example if  
1 −1
A=
1 1
Although only 1 wavelet may be needed, there are some A where there is no
scaling function to generate a multiresolution analysis and there do not seem to be
effective algorithms with such approaches.

5 Applications and Concluding Remarks

There have been a vast range of practical applications of wavelets.
There are effective algorithms for dealing with orthonormal wavelets that arise
from a multiresolution analysis. All the calculations are done in terms of the
coefficients cm .
Practical applications normally rely on a finite-dimensional version of the
wavelet basis, similar to the way the Discrete Fourier Transform (DFT) works.
There is a Fast Wavelet Transform algorithm to rival (or even surpass) the Fast
Fourier Transform (FFT) algorithm for computing the DFT.
Wavelets are applied in removing high-frequency noise from signals and in
trying to identify strong features of signals. Strong features (such as steep slopes
or discontinuities) will normally result in significant magnitudes for the wavelet
Wavelets 25

coefficients at that point for many different scales. Similarly high frequency noise
will normally only produce significant wavelet coefficients at one scale at each
position.
The most spectacular uses are in image analysis and the big success story of
wavelets with respect to images was its adoption by the FBI as a standard for
compressing digital fingerprint images. An image (in one colour or grayscale)
may be regarded as a function of a position in the planar image, with the value
of the function being the brightness or intensity of the image at that point. Using
2-dimensional wavelets, one can express the function by its wavelet coefficients.
To compress the image, one discards some of the smaller coefficients and
stores or transmits only the ‘important’ ones. If this is done in a clever way and
if one uses a cleverly selected set of wavelet coefficients (cm ), then high compres-
sion rates can be obtained without much loss of quality.
Apart from the fingerprints, there is a new JPEG2000 standard that is intended
to supplant the JPEG version now used for storing many images on computers.
(JPEG is based on an algorithm that uses the Discrete Cosine Transform, a variant
of the DFT).
On a more mathematical front there are orthonormal wavelet bases that form
unconditional bases for Banach spaces of functions such as Lp (R) and Lp [0, 1]
(1 < p < ∞). See [6]. For the Fourier series case, we get only a conditional basis
for Lp [0, 1] (1 < p < ∞, p 6= 2).
Wavelets have certain apparent disadvantages compared to Fourier series and
transforms. The wavelets are not eigenfunctions for differentiation and they do
not behave especially well with respect to convolutions. However, some results
remain valid that seem to suggest that differentiation is ‘almost diagonal’ when
expressed in a wavelet basis. See the characterisations of Sobolev spaces in terms
of coefficients with respect to a wavelet basis ([6]).
The number of potential applications of wavelets in signal processing, image
analysis and compression, sound wave analysis, numerical solution of differen-
tial equations and noise reduction is very large and there is a vast literature on
wavelets. Not all potential applications seem to have succeeded, possibly because
the right approach or the right wavelet has not been found. Another explanation
is that wavelet-like techniques were already standard techniques in some fields
like recognition of shapes in digital images, even before the term ‘wavelet’ was
identified around 1988.
Current research on applications of wavelets is very active and, on the theoret-
ical front, there are various more complex variants of wavelets being investigated.
Also the n-dimensional versions of wavelets (associated with matrix dilations)
seem to pose some unresolved problems.
Wavelets 26

References
[1] Burke Hubbard, B., The world according to Wavelets: The stroy of a Math-
ematical Technique in the Making (2nd ed.), A. K. Peters, Natick, Mas-
sachusetts (1998).

[2] Chui, C., Wavelets: a mathematical tool for signal processing, SIAM,
Philadelphia, PA, 1997.

[3] Strang, G. and Nguyen, T., Wavelets and filter banks, Wellesley-Cambridge
Press, Wellesley, MA, 1996.

[4] G. Strang, Wavelet transforms versus Fourier transforms, Bull. Amer. Math.
Soc. 28 (1993) 288–305.

[5] B. Vidaković and P. Müller, Wavelets for kids, ftp://ftp.isds.duke.

edu/pub/Users/brani/papers/wav4kidsA.ps.Z

[6] Wojtaszczyk, P., A mathematical introduction to wavelets. London Mathe-

matical Society Student Texts, 37, Cambridge University Press, Cambridge,
1997.
Contents
1 Introduction: A Review of Bases 1
1.1 Vector spaces and algebraic bases . . . . . . . . . . . . . . . . . 1
1.2 Finite-dimensional inner product spaces and orthonormal bases . . 1
1.3 Banach spaces and Schauder bases . . . . . . . . . . . . . . . . . 2
1.4 Hilbert spaces and orthonormal bases . . . . . . . . . . . . . . . 3
1.5 Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Fourier Transform 6
2.1 Fourier series on general intervals . . . . . . . . . . . . . . . . . 6
2.2 The Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Shannon-Nyquist sampling . . . . . . . . . . . . . . . . . . . . . 8
2.4 Compact support . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Windowed Fourier transforms . . . . . . . . . . . . . . . . . . . 10

3 Wavelets via Frames 12

3.1 Continuous wavelet transform . . . . . . . . . . . . . . . . . . . 13
3.2 Discretisation of the CWT . . . . . . . . . . . . . . . . . . . . . 14
3.3 Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4 Wavelets via Multiresolution Analysis 18

4.1 Multiresolution analysis on R . . . . . . . . . . . . . . . . . . . . 18
4.2 Daubechies wavelets . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3 Higher dimensions . . . . . . . . . . . . . . . . . . . . . . . . . 24

5 Applications and Concluding Remarks 24

View publication stats