Classical Wavelet Theory

Jonathan R. Partington, University of Leeds, School of Mathematics

April 29, 2010

http://www.maths.leeds.ac.uk/∼ pmt6jrp/gradcourse0910.html

LECTURE 1
Books:
Ingrid Daubechies, Ten lectures on wavelets, SIAM.
C.K. Chui, An introduction to wavelets, Academic Press.
Gerald Kaiser, A friendly guide to wavelets, Birkhäuser.

1 Introduction
Fourier series and transforms

For suitable functions f defined on [0, 1] we can form a Fourier series

∞
X
f (t) ∼ ck e2πikt
k=−∞
∞
X
= a0 + (aj cos 2πjt + bj sin 2πjt),
j=1

and Z 1
ck = f (t)e−2πikt dt = fˆ(k), say.
0
Likewise for suitable functions f on the real line we have
Z ∞
f (t) = fˆ(w)e2πiwt dw, where
−∞
Z ∞
ˆ
f(w) = f (t)e−2πiwt dt.
−∞

These are useful formulae, but not localised – if you want to examine f near
t = t0 , say, you still need all the information in fˆ.

1
Windowed Fourier transform (Gabor, 1946)

Let g be a function which is ‘well-localised’ – i.e., big near 0, say, and with g(t)
small for large t.
Examples:
2 /2
• g(t) = e−t ;

• g(t) = χ[−1,1] (t);

(
0 for |t| ≥ 2,
• g(t) a bump function, g ∈ C ∞ (R) and g(t) =
1 for |t| ≤ 1.

Let Z ∞
win ′
T f (w, t ) = f (t)g(t − t′ )e−2πiwt dt.
−∞

This contains information about f near t′ and at frequency 2πω.

Signal analysts can use a discrete form of this, say t′ = mt0 and w = kw0 for
m, k ∈ Z, to get Fourier coefficients of a “slice” of f , i.e.,
Z ∞
win
Tm,k (f ) = f (t)g(t − mt0 )e−2πikw0 t dt.
−∞

Wavelet transforms

Let ψ be a well-localised smooth function satisfying the technical condition

Z ∞
|ψ̂(w)|2
dw < ∞.
−∞ |w|

In particular, we want Z ∞
ψ(t) dt = ψ̂(0) = 0.
−∞

For example, the Mexican hat function,

ψ(t) = (1 − t2 ) exp(−t2 /2),

which is basically the second derivative of the Gaussian function. Define

Z ∞  
wav −1/2 t−b
(T f )(a, b) = |a| f (t)ψ dt = hf, ψ a,b i,
−∞ a

where  
a,b −1/2 t−b
ψ (t) = |a| ψ .
a

2
The number a 6= 0 controls the resolution (scaling) and b the position of the
function.
Then in fact there is an inversion formula
Z ∞Z ∞
−1 da db
f = Cψ ψ a,b hf, ψ a,b i 2 ,
−∞ −∞ a
where Z ∞
|ψ̂(w)|2
Cψ = 2π dw.
−∞ |w|
Again, to reconstruct f we only need a discrete version. Let
−n/2
ψn,m (t) = a0 ψ(a−n
0 t − mb0 )

for suitable a0 , b0 , e.g., a0 = 2 and b0 = 1. Then let

wav
Tn,m (f ) = hf, ψn,m i.

Then for suitable a0 , b0 (the ones above are OK for the Mexican hat), we can find
‘dual’ functions ψ̃n,m such that
∞
X ∞
X
f= hf, ψn,m iψ̃n,m .
n=−∞ m=−∞

Also, there exist A, B > 0 such that

∞
X ∞
X
2
Akf k ≤ |hf, ψn,m i|2 ≤ Bkf k2 ,
n=−∞ m=−∞

a condition which makes (ψn,m ) a frame. Sometimes we can do better (e.g. Haar
wavelets, Littlewood–Paley wavelets) and (ψn,m ) form an orthonormal basis, i.e.,
ψ̃n,m = ψn,m and A = B = 1.

Multi-resolution approximations

Illustrated by a crude but simple example. We work in L2 (R), square-integrable

functions on R.
Let k ∈ Z and let
V − k be the subspace of all functions that are constant on all
intervals 2mk , m+1
2k
for m ∈ Z. So

. . . ⊂ V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ . . . .
T S
Also Vn = {0} and Vn is dense in L2 (R).
We also note that f (t) ∈ Vn ⇐⇒ f (2t) ∈ Vn+1 (for each n), and f (t) ∈ V0 ⇐⇒
f (t − m) ∈ V0 (for each m).

3
In this case we also have that the function g(t) = χ(0,1) (t) has the property that
(g(t − m))m∈Z is an orthonormal basis for V0 .
From this we will later construct “wavelets” and find a function ψ that gives an
orthonormal basis of L2 (R) of the form

ψn,m (t) = 2−n/2 ψ(2n t − m)

such that (ψn,m )n≤N,m∈Z is an o.n.b. for VN .

Applications

Quantum mechanics (Heisenberg group), Signal processing, Image processing,

Numerical analysis.

LECTURE 2

2 Lp spaces and Fourier transforms

Wavelets give a technique for approximating and reconstructing functions, usually
on R.

2.1 Definition
Let 1 ≤ p < ∞. We define the space Lp (R) as the space of (measurable) functions
f : R → C satisfying
Z ∞ 1/p
p
kf kp := |f (t)| dt < ∞.
−∞

Similarly we define Lp (a, b) for an interval (a, b) or [a, b] ⊂ R. In the above we

regard two functions f and g as equal if f (t) = g(t) except on a set of measure 0
(i.e., almost everywhere (a.e.)). With this convention Lp (R) is a normed vector
space, and complete in the metric d(f, g) = kf − gkp , hence a Banach space.
The space L2 (R) is an inner-product space with the inner product being
Z ∞
hf, gi = f (t)g(t) dt,
−∞

linear in the first argument, anti-linear in the second; indeed it is a complete

inner-product space, or Hilbert space.
We write L∞ (R) for the space of essentially bounded measurable functions, with
norm
kf k∞ := inf{K ≥ 0 : |f (t)| ≤ K a.e.}.

4
Again we regard two functions f, g as equal if f (t) = g(t) a.e. Then L∞ (R), and
more generally L∞ (a, b), is a Banach space.

Fourier series

We work on the interval [0, 1]; other intervals can be treated by a change of
variables.
Let
en (t) = exp(2πint),
for each n ∈ Z.

2.2 Theorem
(en )n∈Z is an orthonormal basis of L2 (0, 1), that is, it satisfies:
(
1 if n = m,
1. hen , em i = δnm =
0 for n 6= m.
P
2. Every function f ∈ L2 (0, 1) has an expansion f = ∞ n=−∞ cn en , converging
in L2 norm, i.e.,
N
X
kf − cn en k2 → 0 as N → ∞.
n=−N

P∞
Moreover kf k2 = n=−∞ |cn |2 (Parseval’s formula), and
Z 1
cn = hf, en i = f (t)e−2πint dt
0

is the nth Fourier coefficient of f .

Proof: 1. Easy (exercise). P

2. More tricky. Suppose that cn = hf, en i and write fN = N n=−N cn en .
We may verify that hf − fN , fN i = 0, and so kf k = kf − fN k2 + kfN k2 , so that
2

N
X
2 2
kf k ≥ kfN k = |cn |2 .
n=−N
P∞
Hence N =−∞ |cN |2 ≤ kf k2 , and
X
kfN − fM k2 = |cn |2 , if N < M,
N <|n|≤M

5
from which we deduce that (fN ) is a Cauchy sequence, and so converges to g,
say. The hard part is to show that f = g, i.e., (en ) is a complete orthonormal
basis, and then
X∞
2 2
kf k = lim kfN k = |cn |2 .
n=−∞

The proof is omitted (see, e.g., N. Young’s An introduction to Hilbert space).

2.3 Corollary (Riesz–Fischer theorem)

P
A Fourier series ∞
P
2
n=−∞ cn en converges to a function in L (0, 1) if and only if
∞ 2
n=−∞ |cn | < ∞.

2.4 Remarks
In fact fN → f a.e. when f ∈ L2 , but this is much harder (Carleson, 1962). If
f only lies in L1 , then it is possible for its Fourier series to diverge everywhere
(Kolmogorov, 1926). Even for continuous periodic functions the Fourier series
may diverge at some points, and certainly need not converge uniformly.

We have a norm preserving invertible operator F

L2 (0, 1) → ℓ2 (Z)
f 7→ (fˆ(n))n := (hf, en i)n .

Fourier transforms.

Fourier series represent functions on [0, T ] in terms of a discrete sequence of

nodes exp 2πint/T – Fourier transforms represent functions on R in terms of
nodes exp 2πiωt, where ω ∈ R.

2.5 Proposition
Suppose that f ∈ L1 (R). Then the integral
Z ∞
f (t) exp(−2πiωt) dt
−∞

converges absolutely for all ω ∈ R, and defines a function which we write as fˆ(ω),
the Fourier transform of f . Moreover kfˆk∞ ≤ kf k1 .

LECTURE 3

2.6 Theorem

6
Let f ∈ L1 (R) ∩ L2 (R). Then the function fˆ exists and satisfies kfk
ˆ 2 = kf k2 ;
thus the Fourier transform can be extended continuously to the whole of L2 (R),
and satisfies the following:

kfkˆ 2 = kf k2 Plancherel identity;

ˆ
fˆ(t) = f (−t), that is,
Z ∞
f (t) = fˆ(w) exp(2πiwt) dw a.e.,
−∞

if f is sufficiently well-behaved that fˆ ∈ L1 (R) ∩ L2 (R).

(Proof sketched in lectures.)

2.7 Corollary
There is a linear operator F : L2 (R) → L2 (R) such that
Z ∞
(F f )(w) = f (t) exp(−2πiwt) dt a.e.
−∞

for f ∈ L1 (R) ∩ L2 (R). Also F is norm-preserving and surjective (unitary), and

hf, gi = hF f, F gi

for f, g ∈ L2 (R).

Proof: Use the polarization identity

1 1
hf, gi = (kf + gk2 − kf − gk2 ) + (kf + igk2 − kf − igk2 ).
4 4i
If the operator F preserves norms, then it preserves inner products (i.e., replacing
f and g by F f and F g doesn’t change the RHS, and hence doesn’t change the
LHS).


3 The Haar wavelet

This is the prototype for construction of wavelets, which here will mean or-
thonormal bases of functions related in a certain particular way. At present, we
use discontinuous (step) functions; later we will construct smoother wavelets.

Let V0 denote the subspace of L2 (R) consisting of all functions f that are constant
on every interval of the form (n, n + 1), where n is an integer.

7
3.1 Lemma
Every function f ∈ V0 has the form
∞
X ∞
X
f (t) = ak φ(t − k), with kf k22 = |ak |2 ,
k=−∞ k=−∞

where (
1 if 0 < t < 1,
φ(t) = χ(0,1) (t) =
0 otherwise.
Indeed, ak is the value of f on (k, k + 1). Thus the functions (φk )k∈Z , where
φk (t) = φ(t − k), form an orthonormal basis of V0 .

Proof: Simple verification.



3.2 Theorem
For j ∈ Z let
Vj denote the space of functions f constant on all intervals of the
form 2kj , k+1
2j
, with k ∈ Z. Then

. . . V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ . . . .
T S
Also j∈Z Vj = {0} and j∈Z Vj = L2 (R).
A function f lies in Vj if and only if the function t 7→ f (2−j t) lies in V0 , and Vj
has an orthonormal basis consisting of the functions

φjk (t) = 2j/2φ(2j t − k), k ∈ Z.

m


Proof:
Since every interval 2j+1 , m+1
2j+1
is contained in an interval of the form
k k+1
2j
, 2j , a function in Vj is automatically in Vj+1 .
T
If f ∈ j∈Z Vj , then f is constant on (0, 2j ) for each j. If this value is c then
Z 2j
2
kf k ≥ |c|2 dt = 2j |c|2
0

for each j, and if kf k < ∞, this means that c = 0 and f = 0 a.e. on (0, ∞).
Similarly for (−∞, 0). So f = 0 a.e.
S
Now to show that j∈Z Vj is dense, take f ∈ L2 (R) and ǫ > 0. By standard
results on L2 functions we have that ∃N ∈ N and a continuous function g such
that g(t) = 0 for |t| ≥ N and kf − gk2 < ǫ/2.

8
Now g is uniformly
√ continuous on [−N, N], so ∃j > 0 such that g doesn’t vary
by more than ǫ/ 8N on any interval (k/2j , (k + 1)/2j ). Hence there is a step
function h ∈ Vj such that
Z N 1/2
ǫ2
kg − hk2 ≤ dt = ǫ/2.
−N 8N
S
Thus kf − hk2 < ǫ by the triangle inequality, and so j∈Z Vj is dense.


If f ∈ Vj then f (t) is constant for t ∈ 2kj , k+1
2j
, for all k ∈ Z.
−j
This happens if and only if f (2 t) is constant for t ∈ (k, k + 1), for all k, i.e.,
f (2−j t) ∈ V0 .
Finally, any function in Vj has the form
∞
X
−j
f (2 t) = ak φ(t − k),
k=−∞

by Lemma 3.1, and so

∞
X
f (t) = ak φ(2j t − k).
k=−∞

Now
hφ(2j t − k), φ(2j t − ℓ)i = 2−j δkℓ ,
and so clearly (2j/2 φ(2j t − k))k∈Z is an orthonormal basis of Vj .


LECTURE 4
Now (Vj ) provides a multi-resolution approximation of L2 (R), the ‘resolution’
being 2−j at the jth level. The best approximation fj ∈ Vj to f ∈ L2 (R) is given
by
∞
X
fj = PVj f = hf, φjk iφjk ,
k=−∞

where PVj is the orthogonal projection onto Vj . Also kPVj f − f k → 0 as j → ∞.

We now want to use the orthonormal bases for the Vj to get a grand orthonormal
basis for L2 (R).
To do this we introduce the Haar wavelet ψ given by

ψ(t) = φ(2t) − φ(2t − 1) = χ(0,1/2) (t) − χ(1/2,1) (t).

Clearly ψ ∈ V1 \ V0 .

9
3.3 Lemma
The functions t 7→ ψ(t − k), k ∈ Z, form an orthonormal basis of a space W0 such
that V1 has the orthogonal decomposition V1 = V0 ⊕ W0 .

Proof: Clearly hψ(t − k), ψ(t − ℓ)i = δkl for all k, ℓ ∈ Z.

Also hψ(t − k), φ(t − ℓ)i = 0 for all k, ℓ ∈ Z, since

ψ(t − k)φ(t − k) = ψ(t − k) for each k, and

ψ(t − k)φ(t − ℓ) = 0 for k 6= ℓ.

Hence the collection of vectors S = {φ(t−k), ψ(t−ℓ) : k, ℓ ∈ Z} is an orthonormal

set. In fact it is an orthonormal basis of V1 : they all lie in V1 and
φ(t) + ψ(t) = 2φ(2t), while φ(t) − ψ(t) = φ(2t − 1), and so
1
φ(2t − 2k) = (φ(t − k) + ψ(t − k)), and
2
1
φ(2t − 2k − 1) = (φ(t − k) − ψ(t − k)), k ∈ Z.
2
Hence the linear span of S includes all the basis vectors of V1 , which is enough
to show that it is an orthonormal basis of V1 .

3.4 Lemma
For each j ∈ Z, Vj+1 = Vj ⊕ Wj , where Wj has the orthonormal basis
(2j/2 ψ(2j t − k))k∈Z .

Proof: A similar calculation using the fact that f (t) ∈ Vj ⇐⇒ f (2−j t) ∈ V0 .



3.5 Theorem
L2 (R) = . . . ⊕ W−2 ⊕ W−1 ⊕ W0 ⊕ W1 ⊕ W2 ⊕ . . . as an orthogonal direct sum;
that is, the functions

ψjk (t) = 2j/2 ψ(2j t − k), j, k ∈ Z

form an orthonormal basis of L2 (R).

Proof: Clearly, if i < j and f ∈ Wi , g ∈ Wj , then f ∈ Vi+1 ⊂ Vj and so

hf, gi = 0, since Vj ⊥ Wj .
This implies that if we combine the orthonormal bases from all the Wj we get a
sequence that is still orthonormal – but does it give a basis for L2 (R)?

LECTURE 5

10
Any closed subspace V of L2 (R) has an orthogonal complement V ⊥ , i.e.,
L2 (R) = V ⊕ V ⊥ , where

V ⊥ = {f ∈ L2 (R) : hf, gi = 0 for all g ∈ V }.

We define the orthogonal projection PV : L2 (R) → V by PV (v + w) = v, where

v ∈ V and w ∈ V ⊥ . It is the closest point in V to v + w.
Now,

L2 (R) = Vn+1 ⊕ Vn+1

⊥

⊥
= Vn ⊕ Wn ⊕ Vn+1 ,
so

I = PVn + PWn + PVn+1

⊥

= PVn + PWn + (I − PVn+1 ),

where I is the identity mapping. Thus

PWn = PVn+1 − PVn .

S
Now PVn f → f as n → ∞,Tfor all f ∈ L2 (R), since n Vn = L2 (R). Also,
PVn f → 0 as n → −∞, since n Vn = {0}.
Hence
N
X
PW−N ⊕...⊕W0 ⊕...WN f = PW n f
n=−N
N
X
= (PVn+1 − PVn )f
n=−N
= (PVN+1 − PV−N )f → f − 0 = f

as N → ∞. This implies that

L2 (R) = . . . ⊕ W−2 ⊕ W−1 ⊕ W0 ⊕ W1 ⊕ W2 ⊕ . . . .

3.6 Corollary
Every f ∈ L2 (R) has a multi-resolution expansion
∞
X ∞
X
f= hf, ψjk iψjk ,
j=−∞ k=−∞

11
converging in L2 (R) norm.
R∞
WARNING: Note that −∞ ψjk = 0 for all j and k. Hence if f ∈ L1 (R) ∩ L2 (R)
R∞ 2
and −∞ f 6= 0 (e.g. e−t /2 ), then the finite sums in Cor. 3.6 converge in L2 norm
but not in L1 norm, since
Z ∞ ∞
X ∞
X Z ∞
(f − cjk ψjk ) = f (constant) 6= 0,
−∞ j=−∞ k=−∞ −∞

whatever (cjk ) we take.

Later we construct similar wavelet bases in L2 (R) consisting of smoother func-

tions.

4 Band-limited functions and the sampling the-

orem
A function f ∈ L2 (R) is said to be time-limited, if there is a T > 0 such that
f (t) = 0 a.e. for all |t| ≥ T . It is said to be band-limited, if there is a b > 0 such
that fˆ(w) = 0 a.e. for all |w| > b.

4.1 Definition
Let b > 0. The Paley–Wiener space, P W (b), is the space of all functions
f ∈ L2 (R) such that fˆ(w) = 0 for all |w| > b. Also written P W (−b, b).

Note. fˆ can be regarded as lying in L2 (−b, b), and also

Z b Z b 1/2 Z b 1/2
|fˆ(w)| dw ≤ ˆ
|f(w)| 2
dw 2
1 dw < ∞,
−b −b −b

by the Cauchy–Schwarz inequality. So fˆ ∈ L1 (−b, b). This actually implies that

f is continuous and that f (t) → 0 as t → ±∞.

4.2 Theorem
If f ∈ P W (b), then f is the restriction to R of a function analytic on the whole
of C (which we also call f ), satisfying

|f (z)| ≤ kfˆk1 e2πb|y|

for z = x + iy ∈ C.

12
Proof: Define
Z b
f (z) = ˆ
f(w)e2πiwz
dw,
−b
and
Z b
g(z) = ˆ
(2πiw)f(w)e2πiwz
dw.
−b

Clearly,
Z b
|f (z)| ≤ |fˆ(w)||e2πiwz | dw
−b

≤ kfˆkL1 (−b,b) ke2πiwz kL∞ (−b,b) ≤ kfˆk1 e2πb|y| .

Thus the integral for f converges absolutely, and similarly for g. Also,
Z v Z vZ b
g(z) dz = ˆ
(2πiw)f(w)e 2πiwz
dw dz,
0 0 −b

and, changing the order of integration and using Fubini’s theorem, we get that
this is Z b


fˆ(w) e2πiwv − 1 dw = f (v) − f (0),
−b
′
and hence f (v) = g(v), i.e., f is analytic on C.

Functions satisfying such an inequality are said to be of exponential type. There
is a converse, which we don’t prove:

if f is analytic on C such that f|R lies in L2 (R) and, for some A > 0, we have
|f (z)| ≤ Ae2πb|y| for all z = x + iy ∈ C, then f ∈ P W (b).

4.3 Corollary
A function cannot be both time-limited and band-limited, unless it is identically
zero.

Proof: If f is analytic on C and f (t) = 0 for all real t with |t| > T , then f ≡ 0
by the principle of isolated zeroes.


LECTURE 6
We’re now working towards the sampling theorem for band-limited functions.

13
4.4 Lemma
Let f ∈ P W (b). Then

1 X  m  mπiw/b 1 X  n  −nπiw/b
∞ ∞
ˆ
f (w) = f − e = f e , (1)
2b n=−∞ 2b 2b n=−∞ 2b

with convergence in the L2 (−b, b) norm (and hence in L1 also).

Proof: Since fˆ ∈ L2 (−b, b) it has an expansion in terms of the orthonormal

basis functions em (w) = √12b emπiw/b (cf. Theorem 2.2), namely
∞
X
fˆ = hfˆ, em iem ,
m=−∞

and Z b
1 1  m
hfˆ, em i = √ fˆ(w)e−mφiw/b dw = √ f − .
2b −b 2b 2b
The other identity is obtained on writing n = −m.


4.5 Theorem (Whittaker–Kotel’nikov–Shannon)

n
Let f ∈ P W (b), and write tn = 2b
for n ∈ Z. Then
∞
X sin 2πb(t − tn )
f (t) = f (tn ) ,
n=−∞
2πb(t − tn )

converging in L2 (R) (i.e., P W (b)) norm, as well as uniformly. (As usual, we give
sin x
x
the value 1 at x = 0.)

Proof: Since the Fourier transform preserves L2 norms, we can invert (1) to
get
Z
1 X  n  b −nπiw/b 2πitw
∞
f (t) = f e e dw
2b n=−∞ 2b −b
∞ Z b
1 X
= f (tn ) e−2πitn w e2πitw dw,
2b n=−∞ −b
∞  b
1 X e2πi(t−tn )w
= f (tn ) ,
2b n=−∞ 2πi(t − tn ) −b

14
which gives the result since

e2πi(t−tn )b − e−2πi(t−tn )b = 2i sin 2πb(t − tn ).

Also, L1 convergence in (1) implies uniform convergence of the series for f .


For s ∈ R, we shall write ks for the function given by
sin 2πb(t − s)
ks (t) = for t ∈ R,
π(t − s)
with the value 2b at t = s. Then

k̂s (w) = e−2πisw χ[−b,b] (w).

4.6 Theorem
n √1 kt ,
Let tn = 2b for n ∈ Z. Then the functions 2b n
for n ∈ Z, form an orthonormal
basis for P W (b). Moreover,

f (s) = hf, ks i for f ∈ P W (b).

Proof: As in Lemma 4.4 we see that the functions

1 1
√ e−nπiw/b = √ k̂tn , n ∈ Z,
2b 2b
form an orthonormal basis for L2 (−b, b). Now use the fact that Fourier transforms
preserve norms and inner products. Also,
Z b
ˆ k̂s i =
hf, ks i = hf, fˆ(w)e2πisw dw = f (s).
−b


Definition. A reproducing kernel Hilbert space is a Hilbert space H consisting
of functions on a set S, such that for each s ∈ S there is a ks ∈ H such that
hf, ks i = f (s) for all f ∈ H.

Examples.
sin 2πb(t−s)
1. P W (b) with ks (t) = π(t−s)
.
P∞
2. The Hardy space H 2 of all power series f (z) = n=0 an z
n
in the disc
D = {z ∈ C : |z| < 1}, with
∞
X
2
kf k = |an |2 < ∞,
n=0

15
 and inner product
∞
X ∞
X ∞
X
h an z n , bn z n i = an bn .
n=0 n=0 n=0

1
For w ∈ D we take kw (z) = 1−wz
, the Cauchy–Szegö kernel.

3. The finite-dimensional Hilbert space Cn ; let S = {1, 2, . . . , n}, so that for

a = (a1 , . . . , an ) we have

a(j) = aj = ha, ej i

for j ∈ S, with ej the jth standard basis vector (all coordinates 0, except
that the jth is 1).

4.7 Proposition
Let H be a reproducing kernel Hilbert space with r.k. ks ∈ H for s ∈ S. Then
for s, t ∈ S we have ks (t) = kt (s). Also ks (s) > 0 unless f (s) = 0 for all s ∈ H.

Proof: We have
ks (t) = hks , kt i = hkt , ks i = kt (s).
Also ks (s) = hks , ks i = kks k2 , and ks 6= 0 unless f (s) = hf, ks i = 0 for all f ∈ H.

Note in P W (b) we have ks (t) = kt (s) = k0 (t − s) (real and symmetric).

LECTURE 7
Remark. The sampling theorem gives a way of reconstructing f from values
n
f (tn ), with tn = 2b , but is not robust – small errors in measurements can eas-
ily add up. Oversampling helps avoid this problem: since P W (b) ⊂ P W (c) for
n
c > b, we can reconstruct f from values f ( 2c ) if we wish.

Construction of a multi-resolution approximation and Littlewood–Paley

wavelets

1
We take b = 2
from now on, and write V0 = P W (1/2). Then V0 has an orthonor-
mal basis
sin π(t − k)
t 7→ φ(t − k) = , (k ∈ Z),
π(t − k)
sin πt
where φ(t) = k0 (t) = πt
.

4.8 Theorem

16
For j ∈ Z, let Vj = P W (2j−1). We have

. . . V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ . . . .
S T
Also j∈Z Vj = L2 (R) and j∈Z Vj = {0}. Moreover, for each j, we have
f (t) ∈ V0 ⇐⇒ f (2j t) ∈ Vj . Finally, Vj has an orthonormal basis consisting
of the functions 2j/2 φ(2j t − k), with k ∈ Z.

Proof: Clearly, the Vj are increasing with j. If f ∈ L2 (R), then fˆ ∈ L2 (R), so

given ǫ > 0 we can pick j such that
Z
ˆ
|f(w)| 2
dw < ǫ2 .
|w|>2j−1

Then define g ∈ L2 (R) by

ĝ(w) = fˆ(w)χ[−2j−1 ,2j−1 ] (w),

so that g ∈ Vj and kf − gk = kfˆ − ĝk < ǫ. So the union of the Vj is dense.

T
Also if f ∈ Vj , then fˆ = 0 a.e. for |w| > 2j−1 for each j; letting j → −∞ we
have fˆ = 0 a.e., so f = 0 a.e.

Now write fj (t) = f (2j t), so that

Z ∞
fˆj (w) = fj (t)e−2πiwt dt
Z−∞∞
j ds
= f (s)e−2πiws/2 j , with s = 2j t,
−∞ 2
1
= j fˆ(w/2j ),
2
so f ∈ P W (b) ⇐⇒ fj ∈ P W (2j b).

Finally, Vj has o.n.b.

k
1 sin 2πb(t − 2b )
√ k
= 2j/2 φ(2j t − k), (k ∈ Z),
2b π(t − 2b )
sin πt
where b = 2j−1 and φ(t) = πt
. 
Next, as for the Haar wavelet, we seek Wj such that Vj+1 = Vj ⊕Wj (orthogonal),
and a function ψ such that ψ(t − k), for k ∈ Z form an o.n.b. for W0 (cf. Lemma
3.3).

17
In the transform space, we want ψ̂ supported on [−1, 1] and orthogonal to all
functions fˆ in L2 (− 12 , 12 ). So we define
(
1 if 12 ≤ |w| ≤ 1,
ψ̂(w) =
0 otherwise.

This means that

1
φ(t) = (sin 2πt − sin πt) (exercise).
πt
4.9 Lemma
The functions ψk (t) = ψ(t − k) form an orthonormal basis of a space W0 such
that V0 ⊕ W0 = V1 (an orthogonal direct sum).

Proof:
Z ∞
ψ̂k (w) = ψ(t − k)e−2πiwt dt
Z−∞
∞
= ψ(x)e−2πiw(x+k) dx, with x = t − k,
−∞
−2πiwk
= e ψ̂(w).

Now
Z Z !
−1/2 1
hψ̂k , ψ̂ℓ i = + e−2πiwk e2πiwℓ dw
−1 1/2
(
1 if k = ℓ,
=
0 if k 6= ℓ.

Hence the (ψ̂k ), and so the (ψk ), are orthonormal.

Also, hφk , ψℓ i = 0 for all k and ℓ since hφ̂k , ψ̂ℓ i = 0, where as usual φk (t) = φ(t−k).

LECTURE 8
Finally, since
1 1 1 1
L2 (−1, 1) = L2 (− , ) ⊕ L2 ((−1, − ) ∪ ( , 1)),
2 2 2 2

it is sufficient to verify that (ψ̂k ) is an o.n.b. for L2 ((−1, − 12 ) ∪ ( 12 , 1)) (we have
already seen that they are orthonormal).
Note that (ek (w))k∈Z := (e2πiwk )k∈Z is an o.n.b. for L2 (0, 1).

18
Given γ ∈ L2 ((−1, − 12 ) ∪ ( 21 , 1)), define e
γ ∈ L2 (0, 1) by
(
γ(w − 1) if 0 < w < 21 ,
γe(w) =
γ(w) if 21 < w < 1.
Given ǫ > 0 there is an N and a−N , . . . , aN ∈ C such that
Z 1  N
X
2

 −2πiwk 
e
γ (w) − ak e  dw < ǫ2 .
0  
k=−N

This equals
Z Z ! 2
−1/2 1  N
X 
 −2πiwk 
+ γ(w) − ak e  dw,
−1 1/2  
k=−N
PN
and so kγ − −N ak ψ̂k k2 < ǫ, as required. 

4.10 Theorem
The functions ψjk (t) = 2j/2 ψ(2j t − k), (j, k ∈ Z), form an orthonormal basis
for L2 (R), the Littlewood–Paley wavelets.

Proof: We have done most of this. Note that

L2 (R) = . . . ⊕ W−1 ⊕ W0 ⊕ W1 ⊕ . . .
by Theorem 4.8, where Wj has orthonormal basis (2j/2 ψ(2j t − k)), for k ∈ Z: just
repeat the proof of Theorem 3.5. 

5 Riesz bases and frames

5.1 Definition
A sequence (φk ) is a Hilbert space H is called a Riesz basis for H if
(i) H is the closed linear span of the φk ;
(ii) there are constants A, B > 0 such that
X  X 2 X 

A |ak |2 ≤ ak φk ≤ B |ak |2

for all sequences (ak ) ∈ ℓ2 . A similar definition can be made in finite-dimensional

spaces, taking finite sums only, but we are mostly concerned with the infinite-
dimensional case.

Examples.
1. Any orthonormal basis (φk ), and then we can take A = B = 1.
2. In Cn , any basis {φ1 , . . . , φn }.

19
5.2 Theorem
The sequence (φk ) is a Riesz basis for H if and only if there is a bounded invert-
ible linear operator U : H → H such that (Uφk ) is an orthonormal basis for H.

Proof: If (Uφk ) is an orthonormal basis, then

X X 
ak φk = U −1 ak U(φk ) , and so
X X X 1/2

ak φk ≤ kU −1 k ak U(φk ) = kU −1 k |ak |2 .

Also,
X X

ak (Uφk ) ≤ kUk ak φk , and so
X X 1/2

ak φk ≤ kUk−1 |ak |2 .

Conversely, if (φk ) is a Riesz basis, then let (ψk ) be any o.n.b. for H. Define
U : H → H by X X
U ak φk = ak ψk
for finite sums. Then
X X
−1
ak ψk ≤ A ak φk , and
X X

ak φk ≤ B ak ψk ,

so that U and U −1 are bounded on finite sums. Since the closed linear span of
the (φk ) and (ψk ) both equal H, we can extend U to an isomorphism from H to
H. 

5.3 Definition
A sequence (φk ) in a Hilbert space H is called a frame, if there are constants
A, B > 0 such that
X
Akφk2 ≤ |hφ, φk i|2 ≤ Bkφk2 ∀φ ∈ H. (2)

Examples.
1. Orthonormal bases, with A = B = 1.
2. Riesz bases, since
X X X
|hφ, φk i|2 = |hφ, U −1 Uφk i|2 = |h(U −1 )∗ φ, ψk i|2 = k(U −1 )∗ φk2 ,

which can be bounded above by k(U −1 )∗ k2 kφk2 , and below by kU ∗ k−2 kφk2 .

20
However, frames needn’t be linearly independent. If (φk ) is a frame, then adding
in the zero vector, or repeating each vector a variable number ≤ N times, does
not stop it being a frame. Indeed, any finite set spanning Cn is a frame in Cn
(exercise).

We call the numbers hφ, φk i the frame coefficients of φ and would like to use
them to reconstruct φ.

Example 1. (Daubechies–Morlet) Take

2 2
ψ(x) = √ π −1/4 (1 − x2 )e−x /2 ,
3
the normalized Mexican hat function. Then the functions

ψjk (x) = 2j/2 ψ(2j x − k), (j, k ∈ Z),

form a frame for L2 (R) with B/A ≈ 1.116. Used in vision analysis, and process-
ing seismic data.

Example 2. (Landau) In P W (b) any set of reproducing kernels ktj , with

1
δ < tj+1 − tj < 2b (“uniformly discrete”) forms a frame.

A frame is tight, if A = B, e.g., any orthonormal basis. There are other examples;
e.g., in C2 , the set
( √ √ )
1 3 1 3
{φ1 , φ2 , φ3 } = (1, 0), (− , ), (− , − )
2 2 2 2

satisfies
3
X 3
|hφ, φk i|2 = kφk2
k=1
2
for all φ ∈ C2 .

LECTURE 9
From now on we are looking at countably infinite frames in a separable Hilbert
space.

5.4 Theorem
A sequence (φk ) in a Hilbert space H is a frame for H if and only if there is a
surjective bounded linear operator T : ℓ2 → H such that
X
T ((ak )) = ak φk for all (ak ) ∈ ℓ2 .

21
Proof: Suppose that T exists, and consider T ∗ : H → ℓ2 . Since
X X
hT ∗ ψ, (ak )i = hψ, T ((ak ))i = hψ, ak φk i = ak hψ, φk i,

we have (T ∗ ψ)k = hψ, φk i.

Hence, X
|hψ, φk i|2 = kT ∗ ψk2 ≤ kT ∗ k2 kψk2 ,
i.e., the upper bound in (2).
Also, T is onto, so by Banach’s open mapping theorem ∃C > 0 such that for each
ψ ∈ H there is an (ak ) such that T ((ak )) = ψ, and k(ak )k ≤ Ckψk.
Hence,
hT ∗ ψ, (ak )i = hψ, T ((ak ))i = kψk2 ≥ k(ak )k kψk/C,
so that kT ∗ ψk ≥ kψk/C, and hence
X
|hψ, φk i|2 ≥ kψk2 /C 2 .

Thus (φk ) is a frame.

Conversely, if we have a frame (φk ), let U : H → ℓ2 be defined by (Uψ)k = hψ, φk i,

and take T : ℓ2 → H to equal U ∗ . Now the image of T is H (check) and
X
hT ((ak )), ψi = h(ak ), Uψi = ak hφk , ψi,
P
so T ((ak )) = ak φk . 
Recovery. We want a formula recovering ψ from its frame coefficients hψ, φk i.
Consider the operator S = T T ∗ : H → H. Note that S ∗ = (T T ∗ )∗ = T ∗∗ T ∗ = S,
i.e., S is self-adjoint.
Also X
Sψ = T (hψ, φk i) = hψ, φk iφk ,
and X
hSψ, ψi = hT T ∗ψ, ψi = hT ∗ ψ, T ∗ ψi = |hψ, φk i|2 ≥ Akψk2 ,
so that kSψk ≥ Akψk, and S is invertible.
Now we have
X X
ψ = S −1 Sψ = S −1 hψ, φk iφk = hψ, φk iφek ,

where φek = S −1 φk .

5.5 Theorem

22
The sequence (φek ) = (S −1 φk ) forms a frame in H, called the dual frame to (φk ).
Moreover, the dual frame to (φek ) is (φk ) again.

Proof: We have to calculate

X X
|hψ, φek i|2 = |hS −1 ψ, φk i|2 = kT ∗ S −1 ψk2 = hT ∗ S −1 ψ, T ∗ S −1 ψi
= hS −1 ψ, T T ∗S −1 ψi = hS −1 ψ, ψi.

Since S is self-adjoint and invertible, we can say that ∃A′ , B ′ > 0 such that

A′ kψk2 ≤ |hS −1 ψ, ψi| ≤ B ′ kψk2 ,

which implies that (φek ) forms a frame.

e
So why is (φek ) the same as (φek )? Well, if Te is the frame operator corresponding
to (φek ), then we have

(Te∗ ψ)k = hψ, φek i = hψ, S −1 φk i = hS −1 ψ, φk i,

so that Te∗ = T ∗ S −1 , i.e., Te = S −1 T and

Se = TeTe∗ = S −1 T T ∗ S −1 = S −1 SS −1 = S −1 .
e
Thus φek = Se−1 φek = S φek = φk . 

5.6 Theorem
We have
2 B−A
I − S ≤ < 1, (3)
B+A B+A
and so ∞  n
−1 2 X 2
S = I− S . (4)
B + A n=0 B+A
Proof: Let C denote (A + B)/2. Then

h(S − CI)ψ, ψi ≤ hSψ, ψi − Ckψk2

= hT ψ, T ψi − Ckψk2
B−A
≤ (B − C)kψk2 = kψk2 .
2
But S − CI is self-adjoint, and hence kS − CIk ≤ B−A
2
, or,P
equivalently (3).
Now write U = I − S/C, so that kUk < 1, and (I − U)−1 = ∞ n
n=0 U , which gives
(4). 

23

Hence, if I − B+A
2
S is small, we have a rapidly converging series for S −1 . Then
we can use this, since
X X
ψ= hψ, φk i(S −1 φk ) = S −1 hψ, φk iφk ,

giving an inversion formula to recover ψ from its frame coefficients. Here’s an

estimate for the error.

LECTURE 10

5.7 Corollary
PN
and φek =
2 (N ) 2
Let R = I − B+A
S, B+A n=0 Rn φk for k = 1, 2, . . .. Then
X  B − A N +1

hψ, φk iφek ≤
(N )
ψ − kψk.
B+A

Proof: ∞
2 X n
φek = R φk = φek + RN +1 φek .
(N )
B + A n=0
Hence X X
(N )
ψ − hψ, φk iφek = hψ, φk iRN +1 φek = kRN +1 ψk.

But kRk ≤ (B − A)/(B + A), by Theorem 5.6, and hence the result.

6 Windowed Fourier transforms

Let g(t) be a well-localised function. We’ll assume that g ∈ L2 (R), and w.l.o.g.
kgk
R 2 = 21. Usually g is real and non-negative. “Well-localised” means that
|t|>T
|g| is small for some small T > 0.

Examples. q (i) g(t) = π −1/4 exp(−t2 /2).

(
3
2
(1 − |t|) if |t| ≤ 1,
(ii) g(t) =
0 otherwise.
Let f ∈ L2 (R), and write ft (u) = f (u)g(u − t). Then we define
Z ∞
e ˆ
f (w, t) = ft (w) = e−2πiwu f (u)g(u − t) du.
−∞

This describes the behaviour of f near time t at frequency w.

6.1 Theorem (Time-frequency duality)

24
 eˆ
fe(w, t) = e−2πiwt f(−t, w), where the windowed transform of fˆ is formed using
the window ĝ.

Proof: We have
e t) = hf, gw,ti = hf,
f(w, ˆ ĝw,ti,

where gw,t(u) = e2πiwu g(u − t).

Now,
Z ∞
ĝw,t(ν) = e−2πiνu e2πiwu g(u − t) du
Z−∞
∞
= e−2πiν(x+t) e2πiw(x+t) g(x) dx, with x = u − t,
−∞
−2πiνt 2πiwt
= e e ĝ(ν − w).

So Z ∞
fe(w, t) = e−2πiwt fˆ(ν)e2πiνt ĝ(ν − w) dν,
−∞

as required. 
−πt2
Hence if g and ĝ are both well-localised (e.g. g(t) = 21/4 e and ĝ(w) =
1/4 −πw 2
2 e ), then the windowed Fourier transform gives localised behaviour of
both f and fˆ simultaneously. But there is a limitation as to how well we can do.

6.2 Theorem (Heisenberg’s inequality)

For every f ∈ L2 (R) we have
Z ∞ 1/2 Z ∞ 1/2
2 2 2 ˆ 2 1
t |f (t)| dt w |f(w)| dw ≥ kf k22 .
−∞ −∞ 4π

Comments. In probability and statistics, if f has L2 norm equal to 1, then |f |2

ˆ 2 are probability density functions whose standard deviations satisfy
and |f|
1
σf σfˆ ≥ √ .
4π
In quantum mechanics the uncertainties in measured position and momentum
satisfy
(∆x)(∆p) ≥ ~/2,
where ~ is Planck’s constant.

We get equality if f (and hence fˆ) is a Gaussian function.

25
More generally, for all real t0 and w0 we have, for g ∈ L2 (R), that
Z ∞ 1/2 Z ∞ 1/2
2 2 2 2 1
(t − t0 ) |g(t)| dt (w − w0 ) |ĝ(w)| dw ≥ kgk22.
−∞ −∞ 4π

This can be deduced from the theorem by putting

f (t) = g(t + t0 )e−2πiw0 t ,

and

ˆ
f(w) = g(w + w0 )e2πiwt0 e2πiw0 t0 .

LECTURE 11
Proof of Theorem 6.2: Without loss of generality we may assume that f is
ˆ
C 1 and that |f (t)| = O(1/t2) at ±∞ and |f(w)| = O(1/w 2) at ±∞. This is a
dense set of functions, and the general case follows by approximation.
Now
Z ∞
2
kf k2 = 1.f f
−∞
h i∞ Z ∞
= tf (t)f (t) − t[f ′ (t)f (t) + f (t)f ′ (t)] dt,
−∞ −∞

on integrating by parts. The first term vanishes, and so

Z ∞
2
kf k2 ≤ 2 |tf (t)| |f ′(t)| dt
−∞
≤ 2ktf (t)k2 kf ′ k2 ,

using Cauchy–Schwarz. Now

Z ∞ Z ∞
′ −2πiwt
 
−2πiwt ∞
f (t)e dt = f (t)e −∞
+ f (t)2πiwe−2πiwt dt,
−∞ −∞

ˆ
again integrating by parts, so that kf ′k2 = kF f ′k2 = 2πkw f(w)k 2.
Hence
ˆ
kf k22 ≤ 4πktf (t)k2kw f(w)k 2.


The inversion formula.
We write C for kgk2.

6.3 Theorem

26
 Z ∞Z ∞
1
f (u) = 2 fe(w, t)g(u − t)e2πiwu dw dt,
C −∞ −∞
assuming that the integrals converge absolutely.

Proof: Recall that

Z ∞
ft (u) = f (u)g(u − t) = fe(w, t)e2πiwu dw.
−∞

Hence
Z ∞ Z ∞ Z ∞
2
f (u)|g(u − t)| dt = e t)g(u − t)e2πiwu dw dt.
f(w, (5)
−∞ −∞ −∞

Now the LHS of (5) is C 2 f (u), and hence the result.


Remark 1. By Plancherel’s formula,
Z ∞Z ∞
kfe(w, t)kL2(R2 ) = |ft (u)|2 du dt
Z−∞ −∞
∞ Z ∞
= |f (u)|2|g(u − t)|2 du dt = kf k22 kgk22,
−∞ −∞

so the windowed Fourier transform gives a linear operator

W : L2 (R) → L2 (R2 )
(which is not surjective), satisfying kW f k2 = Ckf k2.

Remark 2. The W.F.T. can be discretized – compute fe(mw0 , nt0 ) for m, n ∈ Z,

where w0 and t0 are fixed positive real numbers. Then
e
f(mw 0 , nt0 ) = hf, gmn i,

where gmn (u) = e2πimw0 u g(u − nt0), so by the results of Section 5, we get a stable
inversion formula if and only if the (gmn ) form a frame.

It turns out to be necessary, but not sufficient in general, that w0 t0 < 1 (so sam-
pling rapidly enough); for the Gaussian window this is also sufficient (Lyubarskii,
1989).

7 The wavelet transform

We saw in Sections 3 and 4 that for certain functions ψ ∈ L2 (R) the family
ψjk (t) = 2j/2 ψ(2j t − k) with j, k ∈ Z forms an o.n.b. of L2 (R). Here j con-
trols scaling (frequency) and k controls position (time). We now generalize this,
starting with a continuous version of it.

27
7.1 Definitions
Let ψ ∈ L2 (R) satisfy the admissibility condition
Z ∞
|ψ̂(w)|2
Cψ := dw < ∞. (6)
−∞ |w|
We write  
x,y −1/2 t−y
ψ (t) = |x| ψ ,
x
for x, y ∈ R with x 6= 0.
For example, take ψ(t) to be the Haar wavelet, so that

(1 − e−iπwt )2
ψ̂(w) = = O(w) as w → 0.
2πiw
√
Then ψ x,y (for x > 0 at least) takes values ±1/ x on the interval between y and
x + y.

Now define
◦
f (x, y) = hf, ψ x,y i
Z ∞  
−1/2 t−y
= f (t)|x| ψ ,
−∞ x
the Wavelet Transform of f .
Note that if ψ̂ is continuous, then by (6) it must satisfy ψ̂(0) = 0, i.e.,
Z ∞
ψ(t) dt = 0.
−∞

So ψ cannot be a positive function like the Gaussian, but it could be the Mexican
hat function.

LECTURE 12

7.2 Theorem
For all f, g ∈ L2 (R), we have
Z ∞Z ∞ ◦
◦ dx dy
f (x, y)g (x, y) 2 = Cψ hf, gi.
−∞ −∞ x
Proof: (Fuller calculations given in lectures.) The LHS above equals
Z ∞
dx dy
A := hf, ψ x,y i hψ x,y , gi 2 .
−∞ x

28
Now
d
ψ b
x,y (w) = |x|1/2 e−2πiwy ψ(xw),

so Z ∞ Z ∞
b ψbx,y i hψbx,y , b dx dy dx
A= hf, gi = hF̌x , Ǧx i ,
−∞ |x| −∞ |x|

where F̌x (y) denotes the inverse Fourier transform of Fx (w) = fb(w)ψ(xw).
b Thus
Z ∞ Z ∞
dx dy b
b ψ(xw)b b dx dw
A = hFx , Gx i = f(w) g (w)ψ(xw)
−∞ |x| −∞ |w|
= Cψ hfb, b
g i = Cψ hf, gi.

Remark. We can interpret this as saying that
Z ∞Z ∞ ◦
−1 ψ x,y
f = Cψ f (x, y) 2 dx dy,
−∞ −∞ x
where the integral is taken in a weak sense (i.e., take the inner product of both
sides with any g ∈ L2 (R).
In fact, it converges in norm as a vector-valued integral, too, in the sense that
Z Z
◦ ψ x,y
f − C −1
f (x, y) dx dy →0
ψ
x2
|y|≤c a≤|x|≤b

as a → 0, and b, c → ∞. Proof omitted.

7.3 Corollary
For f ∈ L2 (R), Z Z
∞ ∞ ◦ dx dy
kf k22 = Cψ−1 | f (x, y)|2 ,
−∞ −∞ x2
2
i.e., the wavelet transform maps L (R) isometrically into a closed subspace H of
L2 (R2 , dµ), where the measure on R2 is given by dµ = Cdxψdy
x2
.
Further, H is a reproducing kernel Hilbert space with kernel
k(s,t) (x, y) = hψ s,t , ψ x,y i.
Proof: The first part follows directly from Theorem 7.2, on putting f = g.
For the reproducing kernel properties, note that
◦ ◦ ◦
f (s, t) = hf, ψ s,t i = hf, ψ s,t iL2 (R2 ,dµ)
Z ∞Z ∞ ◦
−1
◦ dx dy
= Cψ f (x, y)ψ s,t(x, y) 2
−∞ −∞ x
◦
= hf , k(s,t) iL2 (R2 ,dµ) ,

29
where ◦
s,t
k(s,t) (x, y) =ψ (x, y) = hψ s,t , ψ x,y i.

Remark. This transform has found applications in the analysis of radar and
sonar data; there are also connections with quantum mechanics and group rep-
resentations.
It is considered superior to the windowed Fourier transform (WFT) because the
WFT cannot deal efficiently with several scalings/resolutions at once (the “win-
dow” is a fixed size). So the wavelet transform tends to be more efficient.

Discretising the wavelet transform.

For simplicity, assume ψ is an even function, so that ψ x,y = ψ −x,y and so we only
need to consider x > 0.

Fix a > 1 and b > 0. For j, k ∈ Z set x = a−j and y = ka−j b. Write

ψj,k (t) = ψ x,y (t) = aj/2 ψ(aj t − kb),

and consider the products hf, ψj,k i.

By the results of Section 5 we know that we can recover f in a robust manner if
and only if the (ψj,k ) form a frame.
It is necessary for ψ to be admissible, i.e., (6) holds. If this is true and ψ and ψb
are “well-behaved” (a complicated condition to do with decay at infinity), then
for a, b sufficiently small we do get a frame. Often, indeed, a = 2 and b = 1 will
work (Daubechies, very technical). To go further, i.e., to get orthonormal bases,
would be nicer still, and we do this in the next section.

LECTURE 13

8 Multi-resolution analysis (MRA)

By a multi-resolution analysis, we mean a chain of subspaces of L2 (R),

. . . V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ . . . .
S T
such that j∈Z Vj = L2 (R) and j∈Z Vj = {0}.
We also ask that for j, k ∈ Z

• f ∈ V0 ⇐⇒ the function t 7→ f (2j t) lies in Vj ;

• f ∈ V0 ⇐⇒ the function t 7→ f (t − k) lies in V0 ; and

30
 • there is a function φ ∈ V0 such that (φk )k∈Z form an o.n.b. of V0 , where
φk (t) = φ(t − k).
Then for j ∈ Z the space Vj has an orthonormal basis consisting of the functions

φjk (t) = 2j/2φ(2j t − k), k ∈ Z.

Writing Pj for the orthogonal projection onto Vj , we have that Pj f → f as

j → ∞, and Pj f → 0 as j → −∞.

We are seeking a “wavelet” ψ ∈ V1 such that

ψjk (t) = 2j/2ψ(2j t − k), j, k ∈ Z. (7)

form an o.n.b. of L2 (R).

As in Sections 3 and 4, we write Vj ⊕ Wj = Vj+1 , an orthogonal direct sum, i.e.,

Wj = {f ∈ Vj+1 : hf, gi = 0 ∀g ∈ Vj }
= {f ∈ Vj+1 : hf, φjk i = 0 ∀k ∈ Z},

since (φjk )k∈Z is an o.n.b. of Vj .

By the same arguments as before, once we have a MRA we can deduce that

L2 (R) = . . . ⊕ W−2 ⊕ W−1 ⊕ W0 ⊕ W1 ⊕ W2 ⊕ . . . ,

and also f ∈ W0 if and only if t 7→ f (2j t) lies in Wj .

Finally, we seek a ψ ∈ W0 such that (ψk )k∈Z form an o.n.b. for W0 , where ψk (t) =
ψ(t − k). Then we can conclude that the family (ψjk )j,k∈Z, as in (7), forms an
o.n.b. for L2 (R).

8.1 Example (Meyer wavelets)

Fix an integer d ≥ 0 and let θ : [0, 1] → R be a non-negative C d function satisfying
(
0 if u ∈ [0, 31 ],
θ(u) =
π/2 if u ∈ [ 32 , 1],

and
π
θ(u) + θ(1 − u) =
2
for all u ∈ [0, 1]. The way we shall do this is to let Pk be an odd polynomial of
(m)
degree 2d + 1 satisfying Pk (1) = 1 and Pk = 0 for m = 1, 2, . . . , d (no condition
here if d = 0).
Thus we can take P0 (x) = x, and P1 (x) = 23 x − 21 x3 .

31
Define  
π 1 2
θ(u) = (1 + Pk (6u − 3)) for θ ∈ , .
4 3 3
Finally we define the “father wavelet” by means of its Fourier transform, taking
b
φ(w) = cos(θ(|w|)) on [−1, 1]
and 0 outside that interval. Clearly φb ∈ C d and has compact support. This
implies that φ is smooth and rapidly-decreasing, in that td φ(t) lies in L2 (R).

8.2 Proposition
For any φ ∈ L2 (R) the orthonormality property
hφk , φℓ i = δk,ℓ
is equivalent to the condition that
∞
X
K(w) := b + k)|2 = 1
|φ(w a.e.
k=−∞

Proof: Note that K, if finite, is 1-periodic. Now

Z ∞ X∞ Z k+1
b 2
kφk2 = b 2
|φ(w)| dw = b
|φ(w)| 2
dw
−∞ k=−∞ k
∞
X Z 1 Z 1
= b + k)|2 dw =
|φ(w K(w) dw,
k=−∞ 0 0

(exchanging the integral sum is O.K. by the monotone convergence theorem, for
example). So K ∈ L1 (0, 1) and
Z ∞
hφk , φℓ i = hφbk , φbℓ i = b
φ(w)e−2πikw b
φ(w)e2πiℓw dw
−∞
Z ∞ Z 1
= b
|φ(w)| 2 −2πi(k−ℓ)w
e dw = K(w)e−2πi(k−ℓ)w dw,
−∞ 0

which is the (k − ℓ)th Fourier coefficient of the 1-periodic function K. We have

orthonormality if and only if this coefficient is δk,ℓ ; this is equivalent to the con-
dition that K = 1 a.e.

P∞ b + k)|2 = 1; this is obvious
Note that for the Meyer wavelet, we have k=−∞ |φ(w
1 2
for w ∈ [0, 3 ] ∪ [ 3 , 1], as all terms are 0 except for one which is 1. Also, for
w ∈ [ 31 , 23 ] we get
K(w) = cos2 θ(w) + cos2 θ(1 − w)
π 
= cos2 θ(w) + cos2 − θ(w) = cos2 θ(w) + sin2 θ(w) = 1.
2
32
LECTURE 14
We can also identify V0 , or at least its transform Vb0 .
X X
fb ∈ Vb0 ⇐⇒ fb = ak φbk , with |ak |2 < ∞
X
⇐⇒ fb(w) = b
ak e−2πikw φ(w) (L2 convergence)
⇐⇒ fb(w) = φ(w)g(w),
b

where g is 1-periodic and g|[0,1] ∈ L2 (0, 1).

Likewise, fb ∈ Vbj ⇐⇒ f = φbj b

gj , where gj has period 2j and is locally L2 .
T b S b 2
Using this, we can see for the Meyer MRA that j Vj = {0} and j Vj = L (R).

We next need to construct ψ, and we work towards a general construction of this.

8.3 Proposition
2
Let (Vn ) and φ give a MRA
P of2 L (R). Then
2
(i) ∃(hk )k ∈ ℓ (Z) with |hk | = 1 such that
X
φ= hk φ1,k .
P∞ 2πikt
(ii) There is a 1-periodic function h(t) = k=−∞ hk e (convergence in L2 ),
such that khkL2 (0,1) = 1 and
∞
X
hk hk+2ℓ = δℓ,0 , (ℓ ∈ Z).
k=−∞
P
(iii) A function gk φ1,k lies in W0 if and only if (gk ) ∈ ℓ2 and
∞
X
gk hk+2ℓ = 0, (ℓ ∈ Z).
k=−∞

Proof: (i) Since φ ∈ V0 ⊂ V1 we may express φ in terms of the orthonormal

basis (φ1,k )k∈Z of V1 .
(ii) The existence and properties of h come straight from Parseval’s identity. Also
X
φ(t) = hk 21/2 φ(2t − k)

and X X
φ(t + ℓ) = hj 21/2 φ(2t + 2ℓ − j) = hn+2ℓ 21/2 φ(2t − n),
n

33
with n = j − 2ℓ. Now the result follows since
X
hφ−ℓ , φi = hn+2ℓ hn = δℓ,0 .
n
P
(iii) A function χ = gk φ1,k lies in W0 if and only if hφ−ℓ , χi = 0 for all ℓ ∈ Z;
the calculation is now similar to (ii) above.

The Haar wavelet revisited

We have
1 1
φ = √ φ1,0 + √ φ1,1 ,
2 2
and so h(t) = √12 (1 + e2πit ).
In this case we defined
1 1
ψ = √ φ1,0 − √ φ1,1 ,
2 2
so our two sequences are
1 1
(hj ) = (. . . , 0, 0, √ , √ , 0, 0, . . .),
2 2
1 1
(gj ) = (. . . , 0, 0, √ , − √ , 0, 0, . . .), (8)
2 2
so that (gj ) is orthogonal to (hj ) and all its translates by an even number of steps.

LECTURE 15
Back to the general case

8.4 Proposition
Let h be as above, and consider it as a function in L2 (0, 1). then
(i)
Z 1
1
|h(t)|2 e4πiℓt dt = δ0,ℓ , (ℓ ∈ Z); and |h(t)|2 + |h(t + )|2 = 2 a.e.
0 2
P
(ii) a function g ∈ L2 (0, 1) corresponds to gk φ1,k in W0 if and only if
Z 1
h(t)g(t)e4πiℓt dt = 0 (ℓ ∈ Z).
0

Equivalently,
1 1
h(t)g(t) + h(t + )g(t + ) = 0 a.e.
2 2
34
We may equivalently write g(t) = λ(t)h(t + 12 ), with λ a 1-periodic function such
that λ(t) + λ(t + 12 ) = 0 a.e., or, equivalently

1
g(t) = e2πit µ(2t)h(t + ),
2
where µ is 1-periodic.

Proof: (i) Note that

Z 1 X X
|h(t)|2 e4πiℓt dt = h hk e2πi(k+2ℓ)t , hk e2πikt i
0 k k
X
= hk hk−2ℓ = δℓ,0 ,
k

and so |h(t)|2 − 1 is orthogonal to e4πiℓt for each ℓ ∈ Z, i.e., the even-numbered

Fourier coefficients are all zero. Hence h(t)e2πit has all odd-numbered coefficients
zero, so is 21 -periodic, that is,

1 1
(|h(t)|2 − 1)e2πit = (|h(t + )|2 − 1)e2πi(t+ 2 ) ,
2
or |h(t)|2 − 1 = −(|h(t + 21 )|2 − 1), as required.
(ii) Similarly, h(t)g(t)e2πit is 12 -periodic, i.e.,

1 1 1
h(t)g(t)e2πit = h(t + )g(t + )e2πi(t+ 2 ) .
2 2
Now if h(t + 21 ) = 0 then h(t) 6= 0, so g(t) = 0, and we can write g(t) =
λ(t)h(t + 12 ) a.e., with λ 1-periodic.
Now
1 1 1 1 1
h(t)g(t) + h(t + )g(t + ) = h(t)λ(t)h(t + ) + h(t + )λ(t + )h(t) = 0,
2 2 2 2 2
and so we have λ(t) + λ(t + 21 ) = 0 (if h(t + 21 ) = 0, then we can define λ(t) to
ensure this).
This means we can write λ(t) = e2πit µ(2t) with µ 1-periodic.

Example. Take µ ≡ 1. So g(t) = h(t + 12 ).
Thus X X 1
gn e2πint = hk e−2πik(t+ 2 ) e2πit ,
n k
n
or gn = (−1) h1−n . This fits in with the example (8) above.

35
8.5 Lemma
With φ and h as above we have

b 1 b
φ(w) = √ h(w/2)φ(w/2).
2
P
Letting ψ = gn φ1,n where gn = (−1)n h1−n , so that g(w) = e2πiw h(w + 1/2),
we have

b 1 b 1 b
ψ(w) = √ g(w/2)φ(w/2) = √ eπiw h((w + 1)/2)φ(w/2).
2 2
P
Proof: With φ = hk φ1,k we have
Z ∞√
b
φ1,k (w) = 2φ(2t − k)e−2πiwt dt
−∞
Z ∞
1
= √ φ(x)e−2πiw(x+k)/2 dx (x = 2t − k)
2 −∞
1 b −πiwk
= √ φ(w/2)e .
2
b b P b
So φ(w) = √12 φ(w/2) k hk e
−πiwk
with L2 convergence, which equals √12 h(w/2)φ(w/2).
b
Similarly ψ(w) b
= √12 g(w/2)φ(w/2).

Another example. √
For the Littlewood–Paley wavelet we had φb = χ(−1/2,1/2) , and h(w) = 2φ(2w)/
b b
φ(w),
which is 1-periodic.
b
We get that ψ(w) = eπiw χ(−1,−1/2)∪(1/2,1) (w). By choosing µ differently we can
obtain the ψ we obtained before.

8.6 Theorem
In the MRA situation of this chapter, let ν(w) be 1-periodic and satisfy |ν(w)| = 1
a.e. Define ψ by
b 1 b
ψ(w) = √ ν(w)h((w + 1)/2)φ(w/2).
2
Then (ψk )k∈Z , given by ψk (t) = ψ(t − k), generate an o.n.b. for W0 ; hence
(ψj,k )j,k∈Z, given by ψj,k (t) = 2j/2 ψ(2j − k), form an orthonormal basis for L2 (R).

LECTURE 16

36
9 Spline wavelets and compactly-supported wavelets
9.1 Definition
Suppose that a0 < a1 < . . . < aN are real and that n ≥ 1. A spline func-
tion of degree n, with knots at a0 , . . . , aN is a function f : [a0 , aN ] → R such that
f ∈ C n−1 [a0 , aN ] and f|[aj ,aj+1 ] is a polynomial of degree n for each 0 ≤ j ≤ N −1.

So for n = 1 these functions are piecewise linear, for n = 2 piecewise quadratic

with continuous derivative, and so on.
We also allow n = 0, when we simply have a step function that can only jump at
the knots.
These are of use in approximation and interpolation schemes. Certain basis func-
tions are known as B-splines. In particular for n = 0 we see an obvious link with
the space V0 used for constructing Haar wavelets.

To correspond to the case n = 1 we define

(
1 − |x| for 0 ≤ |x| ≤ 1,
φ(x) =
0 otherwise.

Now if f is a degree-1 spline with knots in Z, it can be written in terms of φk ,

where as usual φk (t) = φ(t − k).
The Battle–Lemarié construction produces wavelets using this φ, and similar
basis functions.

9.2 Proposition
With φ as above we have:
(i)
 2
b sin πw
φ(w) = ;
πw
(ii)
1 1
φ(x) = φ(2x + 1) + φ(2x) + φ(2x − 1),
2 2
i.e., P
φ = hk φ1,k , where h−1 = h1 = 2√1 2 , and h0 = √12 , and the rest 0; as usual
φj,k (t) = 2j/2 φ(2j t − k).
(iii)
X∞
K(w) := b + k)|2 = 2 + 1 cos 2πw.
|φ(w
3 3
k=−∞

37
Proof: We leave parts (i) and (ii) as exercises. For (iii), note that K is a
period-1 function whose mth Fourier coefficient is hφ, φmi, as in the proof of
(8.2).
Now Z 1
2
hφ, φ0i = 2 (1 − x)2 dx = ,
0 3
and Z 1
1
hφ, φ1 i = hφ, φ−1i = (1 − x)x dx = ,
0 6
with all other inner products zero, so
2 1 2πiw 1 −2πiw 2 1
K(w) = + e + e = + cos 2πw.
3 6 6 3 3

In general, it is possible to construct similar smoother φ that is a finite linear
combination on translates, and with
 n+1
b −iπrw sin πw
φ(w) = e ,
πw

where r = 0 for N odd and r = 1 for N even.

Problem: in general φ is not a “father wavelet”, as the translates of φ are not
orthonormal, except in the case n = 0.

9.3 Theorem (the orthonormalization theorem)

Let φ ∈ L2 (R) be such that for some constants A, B > 0 the function
∞
X
K(w) := b + k)|2
|φ(w
k=−∞

satisfies 0 < A ≤ K(w) ≤ B for almost all w ∈ R. Then the family (φk )k∈Z
forms a Riesz basis for its closed linear span V0 , and an orthonormal basis can be
obtained by taking the functions (φ♯k )k∈Z defined by

φb♯ (w) = φ(w)/K(w)

b 1/2
,

and φ♯k (t) = φ♯ (t − k).

38
Proof: We calculate
2 2 2
X X X
b −2πilw b
cl φ l = cl φ l = cl e φ(w)

l 2 l 2 l
Z ∞ X  
2
 −2πilw  b
=  cl e  |φ(w)|2 dw
−∞  
l
∞ Z k+1 X
 2
X   2
 −2πilw   b 
=  cl e  φ(w) dw
 
k=−∞ k l
 
X∞ Z 1 X 2  2
 −2πilw   b 
=  cl e  φ(w + k) dw
 
k=−∞ 0 l

Z 1 X 
2
 −2πilw 
=  cl e  K(w) dw.
0  l


Thus 2
X X X
2
A |cl | ≤ cl φ l ≤ B |cl |2 ,

l l 2 l

and it’s a Riesz basis.

Now, repeating the calculation with φ♯ we get an orthonormal sequence (i.e., the
same with A = B = 1), since
∞
X
♯
K (w) := |φb♯ (w + k)|2 = 1.
k=−∞

We should remark that (φk ) and (φ♯k ) have the same closed linear span, since

V0 := {f ∈ L2 (R) : fb = ν φ,
b ν ∈ L2 (0, 1), extended 1-periodically}
= {f ∈ L2 (R) : fb = ν1 φb♯ , ν1 ∈ L2 (0, 1), extended 1-periodically},

with ν1 = Kν. See also (8.6).


For the Battle–Lemarié example, with n = 1 we have
 2
b sin πw 1
φ♯ (w) = 2 1 .
πw ( 3 + 3 cos 2πw)1/2

Now φ♯ is still a spline, but no longer compactly supported. It does have exponen-
tial decay, however. V0 consists of all continuous L2 functions that are piecewise

39
linear on every [n, n + 1]. We could now work out ψ as in Section 8, but it gets
very complicated.

LECTURE 17
Compactly-supported wavelets
For good localization we would like wavelets which decrease rapidly as |t| → ∞.
The following result links the behaviour of f and fb.

9.4 Proposition
(i) Suppose that f, f ′ , . . . , f (n) are all L1 (R) functions (so that in fact f, f ′ , . . . , f (n−1)
are all absolutely continuous as well). Then fb(w) = O(|w|−n) as |w| → ∞.
b
(ii) Suppose that f(w) = O(|w|−n) as |w| → ∞. Then f is C n−2 .
(iii) Parts (i) and (ii) hold if we swap the roles of f and fb.

Proof: (i)
Z ∞
b
f(w) = f (t)e−2πiwt dt
−∞
Z ∞
1
= f ′ (t)e−2πiwt dt
2πiw −∞
= ... Z ∞
1
= f (n) (t)e−2πiwt dt,
(2πiw)n −∞

integrating by parts repeatedly. Then since f (n) ∈ L1 , the final integral is

bounded independently of w so |f (w)| ≤ kf (n) k1 /(2π|w|)n.
(ii) Z ∞
f (t) = b
f(w)e 2πiwt
dw.
−∞

We may differentiate under the integral sign as everything is absolutely conver-

gent, and so Z ∞
(m)
f (t) = fb(w)(2πiw)me2πiwt dw,
−∞

at least for 1 ≤ m ≤ n − 2. Moreover, fb(w)(2πiw)n−2 lies in L1 (R), and so f (n−2)

is continuous.
b
(iii) Easy, since fb(t) = f (−t).


9.5 Corollary

40
If f is C ∞ with all derivatives in L1 (R), then fb decreases more rapidly than any
power of w. Likewise, if fb decreases rapidly, then f ∈ C ∞ .

Properties of particular wavelets

Wavelet Properties of ψ(t) Properties of ψ̂(w)

Haar Compact support, discontinuous O(1/w), C ∞
Littlewood–Paley O(1/t), C ∞ Compact support, discontinuous
∞
Meyer Rapidly-decreasing, C Compact support, can be C ∞
k
Battle–Lemarié Rapidly-decreasing, C O(1/w k ), C ∞
Daubechies Compact support, C k O(1/w k ), C ∞

Here k is arbitrarily large, but finite.

What is not possible is for both ψ and ψb to have compact support, since a function
f and its transform fb cannot both be time-limited.
Also we can’t have ψ being C ∞ with exponential decay if the (ψj,k ) form an or-
thonormal basis. We can if they form a frame (e.g., Mexican hat).

Construction of Daubechies wavelets (sketch)

Recall that we had, in Section 8, the notation

X
φ = hk φ1,k ,
X
h(t) = hk e2πikt ,
b 1 b
φ(w) = √ h(w/2)φ(w/2),
2
X
ψ = gk φ1,k , with
gk = h1−k (−1)k ,
|h(t)|2 + |h(t + 1/2)|2 = 2 (orthonormality of (φk )).

9.6 Proposition
Suppose that φ is a compactly-supported function, φ(0)b = 0, that φ generates a
MRA (Vk ), and
√ (φk ) is an o.n.b. for V0 . Then h(t) is a trigonometric polynomial
in t, h(0) = 2, and ψ is also compactly-supported.

Proof: Since φ and φ1,k are compactly-supported, it’s clear that only finitely-
many φ1,k have supports meeting supp φ, so hk = 0 for all but finitely-many
k. √
b
Also φ(w) b
= √12 h(w/2)φ(w/2); put w = 0 and we see that h(0) = 2.
P
Finally, gk = h1−k (−1)k 6= 0 for only finitely-many k and so ψ = gk φ1,k is
compactly-supported as each φ1,k is. 

41
LECTURE 18

9.7 Proposition
The scaling function h determines φ (up to a constant) by the identity
Y∞  
b b 1 m
φ(w) = φ(0) √ h(w/2 ) .
m=1
2
√
Proof: Since h(0) = 2 and h is a trigonometric polynomial, we have
w
h(w/2m ) = h(0) + m h′ (ξ),
2
for some 0 < ξ < w/2m . Thus there is a constant C > 0 such that
 
 1 
 √ h(w/2 ) − 1 ≤ Cw/2m
m
 2 

for all |w| ≤ 1. Q P

Now an infinite
P product a m “converges absolutely” when | log am | < ∞, or
equivalently (1 − |am |) < ∞. We apply this idea with am = √12 h(w/2m).
So
b 1 b
φ(w) = √ h(w/2)φ(w/2)
2
= .". . #
Y n
1 b
= √ h(w/2m) φ(w/2 n
) for n = 1, 2, . . . .
m=1
2
b
Letting n → ∞ and noting that φ(w/2 n b
) → φ(0), we have the result. 

9.8 Construction
√
We take h to be given by h(w) = 2P (e2πiw ), with P a polynomial of degree
2N − 1; here N ≥ 1 is a fixed integer.
If we have |P (z)|2 + √
|P (−z)|2 = 1 whenever |z| √
= 1, then, with z = exp(2πiw),
we will have h(w) = 2P (z) and h(w + 1/2) = 2P (−z), so we will obtain
|h(w)|2 + |h(w + 1/2)|2 = 2,
which is one of the necessary conditions.
We shall choose  N
1+z
P (z) = W (z),
2
with deg W = N − 1. Thus P (−1) = 0 and we will take P (1) = 1 by making
W (1) = 1.
This property leads to smoothness of φ (details omitted).

42
9.9 Proposition
For N = 1 the construction produces P (z) = (1+z)/2, and |P (z)|2 +|P (−z)|2 = 1
for |z| = 1.
Then h(w) = √12 (1 + e2πiw ), the scaling function for the Haar wavelet.

Proof: W is degree 0, hence constant, so W = W (1) = 1. It is easy to verify

that
 
1 + z 
 
 2  = | cos πw|, and
 
1 − z 
 
 2  = | sin πw|,

for z = exp(2πiw). Thus |P (z)|2 + |P (−z)|2 = 1, and the remainder of the

calculation is straightforward. 

9.10 Theorem (Strichartz, 1993)

Let C(z) = cos πw = (z 1/2 + z 1/2 )/2 and S(z) = sin πw = (z 1/2 − z 1/2 )/2i. Then
2N
X −1  
2N − 1
1= C 4N −2−2k S 2k ,
k=0
k

and if |P (z)|2 equals the sum of the first N terms, then P satisfies the conditions
of Construction 9.8.

Proof: The first identity is just what we get on expanding 1 = (C 2 + S 2 )2N −1

by the binomial theorem.
Now if
N
X −1  
2 2N − 1
|P (z)| = C 4N −2−2k S 2k ,
k=0
k
which is non-negative since all its terms are non-negative, then the transformation
z 7→ −z corresponds to w 7→ w + 1/2 (modulo 1); thus C(−z) = cos π(w + 1/2) =
−S(z) and S(−z) = sin π(w + 1/2) = C(z).
Thus, with ℓ = 2N − 1 − k below, we get
N
X −1  
2 2N − 1 4N −2−2k 2k
|P (−z)| = S C
k=0
k
2N
X −1  
2N − 1 2ℓ 4N −2−2ℓ
= S C ,
ℓ=N
ℓ

43
and indeed
2N
X −1  
2 2 2N − 1
|P (z)| + |P (−z)| = C 4N −2−2k S 2k = 1.
k=0
k


Note that z = −1 is when w = 1/2 and C = 0, and |P (z)|2 has a zero of degree
4N − 2 − 2(N − 1) = 2N there, so P will have a zero of degree N.

Example. N = 2.

N
X −1  
2 2N − 1
|P (z)| = C 4N −2−2k S 2k
k
k=0
X1  
3
= C 6−2k S 2k
k
k=0
= C + 3C 4 S 2
6
 
 1 + z 4 2

=   (C + 3S 2 ).
2 
We now need to do a spectral factorization, choose W a polynomial of degree 1
with W (1) = 1 such that

|W (e2πiw )|2 = C 2 + 3S 2 = 1 + 2S 2 = 2 − cos 2πw,

or
z+z
|W (z)|2 = 2 −
2
when |z| = 1.
We try W (z) = a + bz with a, b real, so that |W (z)|2 = (a + bz)(a + bz), and
hence 2 = a2 + b2 and − 21 = ab. Also a + b = W (1) = 1.
The solution to these equations is
 2
1 1+z h √ √ i
P (z) = (1 ± 3) + (1 ∓ 3)z .
2 2
9.11 Final details
P determines h, and h determines φb and hence φ, by Prop. 9.7. Using the coef-
ficients of h we can now obtain ψ too.

Details that we have omitted.

44
1. φ is compactly-supported – the idea is to show that the infinite product
defining φb produces an entire function lying in a suitable Paley–Wiener
space, and hence φ itself has compact support contained in [0, 2N − 1].

2. The continuity and differentiability properties of φ and ψ are also to do

b
with the properties of φ.

3. We need to do a spectral factorization to obtain the higher-order wavelets.

This requires the Fejér–Riesz theorem which says they exist. Can be done
algebraically for N ≤ 3, and numerically/iteratively beyond that.

THE END

45