Chapter 1

An Overview

1.1 Brief Review on Fourier Series and Fourier Trans-

form
1.1.1 Fourier series
Let L2 (0, 2π) denote the collection of all measurable (complex-valued) functions f defined
on the interval (0, 2π) with ż 2π
ˇf (x)ˇ2 dx ă 8.
ˇ ˇ
0
For the reader who is not familiar with the basic Lebesgue theory, the sacrifice is very
minimal by assuming that f is a piecewise continuous function. It will always be assumed
that functions in L2 (0, 2π) are extended periodically to the real line R = (´8, 8), namely:
f (x) ´ f (x ´ 2π) for all x. Hence, the collection L2 (0, 2π) is often called the space of 2π-
periodic square-integrable functions. That L2 (0, 2π) is a vector space can be verified very
easily. Any f in L2 (0, 2π) has a Fourier series representation:
8
ÿ
f (x) = cn einx , (1.1.1)
n=´8

where the constants cn , called the Fourier coefficients of f , are defined by

1 2π
ż
cn = f (x)e´inx dx. (1.1.2)
2π 0
The convergence of the series in (1.1.1) is in L2 (0, 2π), meaning that
ż 2π ˇ N
ÿ ˇ2
ˇ ˇ
inx ˇ
lim ˇf (x) ´ c n e ˇ dx = 0.
N,M Ñ8 0
ˇ
n=´M

There are two distinct features in the Fourier series representation (1.1.1).

1. Any f P L2 (0, 2π) is decomposed into a sum of infinitely many mutually orthogonal
components gn (x) ” cn einx , where orthogonality means that

xgn , gm yL2 (0,2π) = 0 for all m ‰ n (1.1.3)

1
 with the “inner product” in (1.1.3) being defined by

1 2π
ż
xgn , gm yL2 (0,2π) = gn (x)gm (x)dx. (1.1.4)
2π 0
That (1.1.3) holds is a consequence of the important, yet simple fact that

en P L2 (0, 2π) ˇ en (x) ” einx , n P Z

␣ ˇ (
(1.1.5)

is an orthonormal (o.n.) “basis” of L2 (0, 2π).

2. The orthonormal basis ten u8

n=´8 used in the Fourier series representation (1.1.1) is
generated by “dilation” of a single function

e(x) = eix ; (1.1.6)

that is, en (x) = e(nx) for all integers n. This will be called integral dilation. The
theory of the Fourier series shows that “every 2π-periodic square-integrable function
is generated by a “superposition” of integral dilations of the basic function eix .

Definition 1.1. An orthonormal set ten u8 n=1 in a Hilbert space (H, x¨, ¨y) is said to be
complete or is called an orthonormal basis if for every v P H, we have
8
ÿ
v= xv, en y en ,
n=1

where, with } ¨ } denote the norm induced by the inner product, the equality above is
understood in the sense that
› n
ÿ ›
lim ›v ´ xv, ek y ek › = 0.
› ›
nÑ8
k=1

Theorem 1.2. Let (H, x¨, ¨y) be a Hilbert space, and ten u8
n=1 be an orthonormal set in H.
The following three statements are equivalent:

1. ten u8
n=1 is complete;

2. xv, en y = 0 for all n P N implies that v = 0;

8 ˇ
ˇxv, en yˇ2 .
ˇ
3. for every v P H, }v}2 ” xv, vy =
ř
n=1
8
ÿ
4. for every u, v P H, xu, vy = xu, en y xv, en y.
n=1

From the o.n. property of ten u8

n=´8 , the Fourier series representation (1.1.1) also satisfies
the so-called Parseval identity:
ż 2π 8
1 ˇf (x)ˇ2 dx =
ˇ ˇ ÿ
|cn |2 . (1.1.7)
2π 0 n=´8
 Let ℓ2 denote the space of all square-summable bi-infinite sequences; that is, tcn u8
n=´8 P
2
ℓ if and only if
8
ÿ
|cn |2 ă 8.
n=´8
Hence, if the square-root of the quantity on the left of (1.1.7) is used as the “norm” for the
measurement of functions in L2 (0, 2π), and similarly, the square-root of the quantity on the
right of (1.1.7) is used as the norm for ℓ2 , then the function space L2 (0, 2π) and the sequence
space ℓ2 are “isometric” to each other. Returning to the above mentioned observation on the
Fourier series representation (1.1.1), we can also say that every 2π-periodic square-integrable
function is an ℓ2 -linear combination of integral dilations of the basic function e(x) = elx .
We emphasize again that the basic function

e(x) = eix = cos x + i sin x

which is a “sinusoidal wave”, is the only function required to generate all 2π-periodic square-
summable functions. For any integer n with large absolute value, the wave en (x) = e(nx)
has high “frequency”, and for n with small absolute value, the wave en has low frequency.
So, every function in L2 (0, 2π) is composed of waves with various frequencies.

In general, we can consider L2 (0, L), the collection of square-integrable (complex-valued)

functions defined on (0, L), and extend f P L2 (0, L) periodically (with period L) to the real
line satisfying f (x) = f (x ´ L) for all x P R. In other words, L2 (0, L) can be viewed as the
collection of all L-periodic complex-valued functions defined on R satsifying
żL
ˇf (x)ˇ2 dx ă 8.
ˇ ˇ
0
2
The inner production of L (0, L) is given by
żL
1
xf, gyL2 (0,L) = f (x)g(x)dx
L 0

with induced norm } ¨ }L2 (0,L) given by

(1 ż Lˇ )1
ˇf (x)ˇ2 dx 2 .
ˇ
}f }L2 (0,L) =
L 0
Any f P L2 (0, L) has the Fourier series representation
8
2πinx
ÿ
f (x) = cn e L , (1.1.8)
n=´8
żL
1 2πinx
where the Fourier coefficients tcn u8
n=´8 are given by cn = f (x)e´ L dx, and (1.1.8)
L 0
is understood in the sense
› N
ÿ ›
lim ›f ´ cn d L e n › = 0, (1.1.9)
› ›
N,M Ñ8 2π L2 (0,L)
n=´M

where for a constant c ą 0, dc is the dilation operator given by

(dc f )(x) = f (x/c).

1.1.2 Fourier transform
Next we consider functions defined on R without periodicity.

Definition 1.3. The space Lp (R), where 1 ď p ă 8, is a normed space that consists of all
complex-valued measurable functions satisfying
! ˇż )
L (R) = f : R Ñ C ˇ |f (t)|p dt ă 8
p ˇ
R

which is equipped with norm } ¨ }Lp (R) given by

(ż ˇ ˇ )1
}f }Lp (R) = ˇf (t)ˇp dt p ,
R

and the space L (R) consists of all complex-valued (essentially) bounded measurable func-
8

tions equipped with norm

ˇ ˇ
}f }L8 (R) = ess sup ˇf (x)ˇ .
xPR
p
Note that (L (R), } ¨ }Lp (R) ) is a Banach space; that is, a complete normed space.

Definition 1.4. For all f P L1 (R), the Fourier transform of f , denoted by F f or fp, is a
function defined by
ż
(F f )(ω) = fp(ω) = f (t)e´itω dt @ ω P R. (1.1.10)
R

Let C0 (R) denote the space of all bounded/continuous functions on R which decay at
infinity; that is, f P C0 (R) if and only if f (t) Ñ 0 as t Ñ 8. C0 (R) is a normed space
equipped the sup-norm defined by
ˇ ˇ
}f } = sup ˇf (t)ˇ .
tPR

Theorem 1.5. F : L1 (R) Ñ C0 (R) is a bounded linear map.

Proof. To show this theorem, we need to establish the following properties:

1. F is linear;

2. there exists a constant C such that }F [f ]}L8 (R) ď C}f }L1 (R) ;

3. F [f ] is continuous and F [f ](ω) Ñ 0 as |ω| Ñ 8.

The linearity of F is trivial. Moreover, if f P L1 (R),

ż
|fp(ω)| ď |f (t)|dt = }f }L1 (R) ă 8
R

which shows that }F [f ]}L8 (R) ď }f }L1 (R) . To show that fp is continuous, we in fact show
that fp is uniformly continuous as follows. Let ε ą 0 be given. Note that the Dominated
Convergence Theorem implies that
ż
ˇ ˇ
lim |f (t)|ˇe´it∆w ´ 1ˇ dt = 0;
∆wÑ0 R
thus there exists δ ą 0 such that
ż
ˇ ˇ
|f (t)|ˇe´it∆w ´ 1ˇ dt ă ε whenever |∆w| ă δ .
R

Therefore, if ω1 , ω2 P R satisfying |ω1 ´ ω2 | ă δ, we have

ż
ˇ ˇ ˇ ˇ
ˇfp(ω1 ) ´ fp(ω2 )ˇ ď |f (t)|ˇe´it(ω1 ´ω2 ) ´ 1ˇ dt ă ε
R

π
that shows that fp is uniformly continuous on R. Finally, since e´itω = ´e´iω(t+ ω ) ,
ż (
π ) ´iωt
ż ż
π
´itω ´iω(t+ ω )
f (ω) = f (t)e
p dt = ´ f (t)e dt = ´ f t ´ e dt.
R R R ω

Therefore,
ˇż ż ( ) ˇ
ˇ ˇ 1ˇ ´iωt π ´iωt
ˇ
ˇfp(ω)ˇ = ˇ f (t)e dt ´ f t ´ e dt ˇ
2ˇ R R ω ˇ
ˇż [ ( )] ( π )ˇˇ
ˇ ż ˇ
1 ˇˇ π ´iωt ˇ
ˇ 1 ˇ
= ˇ f (t) ´ f t ´ e dtˇ ď ˇf (t) ´ f t ´ ˇ dt
2 R ω 2 R ω

which converges to 0 as |ω| Ñ 8. ˝

ˇ ˇ
Remark 1.6. The result that lim ˇfp(ω)ˇ = 0 for f P L1 (R) is often called the Riemann-
|ω|Ñ8
Lebesgue Lemma.

Before proceeding, we define the following useful operators, called the translation, mod-
ulation and (scaled) dilation operators, respectively: for f P L1 (R),
1
(x)
(Tc f )(t) = f (t ´ c), (Mc f )(t) = eict f (t), (Dc f )(t) = a f . (1.1.11)
|c| c

In particular, D´1 is called the parity/reflection operator.

Theorem 1.7. Let f P L1 (R). Then

1. (Shifting) F [Tc f ](ω) = e´icω fp(ω).

( )
2. (Scaling) F [Dc f ](ω) = D 1 fp (ω).
c

[ ] [ ]
3. (Conjugation) F D´1 f (ω) = F D´1f (ω) = fp(ω).

4. (Modulation) F [Mc f ](ω) = (Tc fp)(ω).

1 t´b
( )
1
Example 1.8. Let ψ P L (R), and ψa,b (t) = a ψ . Using the translation and
|a| a
dilation operators we have ψa,b (t) = (Tb Da ψ)(t); thus

a,b (ω) = F [Tb Da ψ](ω) = e F [Da ψ](ω) = e´ibω D 1 ψ(ω)

´ibω
a
ψ
y p = |a|e´ibω ψ(aω).
p
a
Definition 1.9. Let f, g be complex-valued function defined on R. The convolution of f
and g, denoted by f ˙ g, is the function defined by
ż
(f ˙ g)(t) = f (τ )g(t ´ τ )dτ
R

whenever the integral makes sense.

Note that by the change of variable formula, f ˙ g = g ˙ f whenever the convolution

makes sense.
1 1
Remark 1.10. The convolution f ˙ g makes sense if f P Lp (R), g P Lq (R) with + ě 1.
p q
1 1 1
In fact, for f P Lp (R) and g P Lq (R), then f ˙ g P Lr (R) with + = 1 + and one has
p q r
Young’s inequality
}f ˙ g}Lr (R) ď }f }Lp (R) }g}Lq (R) .

Theorem 1.11. If f, g P L1 (R), then F [f ˙ g](ω) = fp(ω)p

g (ω).

Theorem 1.12. Let g P L1 (R) and define

ż ( )
h(t) = g(ω)eitω dω = gp(´t) .
R

If h P L1 (R), then for all f P L1 (R),

ż
(f ˙ h)(t) = g(ω)fp(ω)eiωt dω .
R

Definition 1.13. A summability kernel on R is a family tKλ uλą0 of continuous functions

with the following properties:
ż
(i) Kλ (x)dx = 1 for all λ ą 0;
R

(ii) there exists M ą 0 such that

ż
ˇ ˇ
ˇKλ (x)ˇ dx ď M @ λ ą 0;
R

ż ˇ ˇ
(iii) lim ˇKλ (x)ˇ dx = 0 for all δ ą 0.
λÑ8 |x|ąδ

A simple construction of a summability on R is as follows. Suppose F is a continuous

Lebesgue integrable function so that
ż
F (x)dx = 1.
R

Then, we set
Kλ (x) = λF (λx) for all λ ą 0 and x P R. (1.1.12)
Evidently, it follows that
ż ż ż
Kλ (x)dx = λF (λx)dx = F (x)dx = 1,
ż R żR Rż
ˇ ˇ ˇ ˇ ˇ ˇ
ˇKλ (x)ˇ dx = λˇF (λx)ˇ dx = ˇF (x)ˇ dx = }F }L1 (R)
R R R

and for δ ą 0,
ż ż ż
ˇ ˇ ˇ ˇ ˇ ˇ
ˇKλ (x)ˇ dx = λˇF (λx)ˇ dx = ˇF (x)ˇ dx Ñ 0 as λ Ñ 8.
|x|ąδ |x|ąδ |x|ąλδ

Therefore, the family tKλ uλą0 defined by (1.1.12) is a summability kernel on R.

Example 1.14. 1. The family tKλ uλą0 defined by Kλ (x) = λF (λx), where

1 sin2 (x/2)
F (x) =
2π (x/2)2
is called the Fejér kernel.

2. The family tKλ uλą0 defined by Kλ (x) = λG(λx), where

1
G(x) = ? exp(´x2 )
π
is called the Gaussian kernel.
Both the Fejér kernel and the Guassian kernel are summability kernels.

Theorem 1.15. Let tKλ uλą0 be a summability kernel on R. If f P L1 (R), then

lim }f ˙ Kλ ´ f }L1 (R) = 0.

λÑ8

Moreover, if in addition that f is essentially bounded and is continuous at c, then

lim (f ˙ Kλ )(c) = f (c).

λÑ8

Proof. Let f P L1 (R) be given, and M ą 0 be such that

ż
ˇ ˇ
ˇKλ (x)ˇ dx ď M .
R

Let ε ą 0 be given. By the definition of summability kernels and the convolution,

ż ż
(f ˙ Kλ )(x) ´ f (x) = f (x ´ y)Kλ (y)dy ´ f (x) Kλ (y)dy
żR R
[ ]
= Kλ (y) f (x ´ y) ´ f (x) dy . (1.1.13)
R
ż ˇ ˇ
1. Since lim ˇf (x) ´ f (x ´ y)ˇ dx = 0, there exists δ ą 0 such that if |y| ă δ,
yÑ0 R
ż
ˇf (x) ´ f (x ´ y)ˇ dx ă ε .
ˇ ˇ
R 2M
 By the Tonelli Theorem,
ż ż ż
ˇ ˇ ˇ ˇˇ ˇ
ˇ(f ˙ Kλ )(x) ´ f (x)ˇ dx ď ˇKλ (y)ˇˇf (x ´ y) ´ f (x)ˇ dydx
R ż ż R R
ˇ ˇˇ ˇ
= ˇKλ (y)ˇˇf (x ´ y) ´ f (x)ˇ dxdy
R R
(ż ż )ˇ ż
ˇ ˇ ˇ
= + ˇKλ (y)ˇ ˇf (x ´ y) ´ f (x)ˇ dxdy
|y|ăδ |y|ěδ R
ż ż
ε ˇ ˇ ˇ ˇ
ď ˇKλ (y) dy + 2}f }L1 (R)
ˇ ˇKλ (y)ˇ dy
2M |y|ăδ |y|ěδ
ż
ε ˇ ˇ
ď + 2}f }L1 (R) ˇKλ (y)ˇ dy .
2 |y|ěδ

Therefore, by the properties of summability kernels,

ż
ˇ ˇ ε
lim sup ˇ(f ˙ Kλ )(x) ´ f (x)ˇ dx ď ă ε.
λÑ8 R 2
Since ε ą 0 is given arbitrarily, we conclude that
ż
ˇ ˇ
lim sup ˇ(f ˙ Kλ )(x) ´ f (x)ˇ dx = 0
λÑ8 R

which shows that lim }f ˙ Kλ ´ f }L1 (R) = 0.

λÑ8

2. Now suppose in addition that f is continuous at c. Then there exists δ ą 0 such that

ˇf (a ´ y) ´ f (a)ˇ ă ε .
ˇ ˇ
2M
Therefore, (1.1.13) implies that
ż
ˇ ˇ ˇ ˇˇ ˇ
ˇ(f ˙ Kλ )(c) ´ f (c)ˇ ď ˇKλ (y)ˇˇf (c ´ y) ´ f (c)ˇ dy
R
(ż ż )ˇ ˇˇ ˇ
= + ˇKλ (y)ˇˇf (c ´ y) ´ f (c)ˇ dy
|y|ăδ |y|ěδ
ż ż
ε ˇ ˇ ˇ ˇ
ď ˇKλ (y) dy + 2}f }L8 (R)
ˇ ˇKλ (y)ˇ dy .
2M |y|ăδ |y|ěδ

ˇ ˇ
The same argument as in Part 1 shows that lim ˇ(f ˙ Kλ )(c) ´ f (c)ˇ = 0. ˝
λÑ8

Let Λ be the function defined by

1
#
(1 ´ |x|) if |x| ď 1,
Λ(x) = 2π
0 if |x| ą 1.
Then
ż1 ż1
1 ´ixω 1
Λ(ω)
p = (1 ´ |x|)e dx = (1 ´ x) cos(xω)dx
2π ´1 π 0
1 1 ´ cos ω 1 sin2 (ω/2)
= ¨ = = F (ω);
π ω2 2π (ω/2)2
thus the Fejér kernel is given by
? ?
Kλ (x) = λF (λx) = λD 1 Λ(x)
p = F [ λDλ Λ](x).
λ

By Theorem 1.12 (with h being the Fejér kernel), the fact that Λ
p is even implies that

1 λ ( |ω| ) p
ż ? ż
iωt
(f ˙ Kλ )(t) = λ(Dλ Λ)(ω)f (ω)e dω =
p 1´ f (ω)eiωt dω .
R 2π ´λ λ

Using the above identity, Theorem 1.15 and the Dominated Convergence Theorem show
the following

Theorem 1.16 (Fourier Inversion Formula). Let f P L1 (R). If fp P L1 (R), then

ż
1
f (t) = fp(ω)eiωt dω for a.a. t P R. (1.1.14)
2π R
In particular, the identity above holds for t at which f is continuous.

Remark 1.17. Note that in order to show (1.1.14) holds for continuities of f , the bound-
edness of f is required. Nevertheless, since fp P L1 (R), (1.1.14) shows that f is almost
everywhere equal to a continuous function that decays at infinity; thus f is essentially
bounded.

Remark 1.18. The integral operator

1
ż ( 1
ż )
f ÞÑ M‚ (ω)f (ω)dω or f (t) = Mt (ω)f (ω)dω ,
2π R 2π R

where M‚ is the Modulation operator defined in (1.1.11), is called the inverse Fourier trans-
form and is usually denoted by F ´1 or q. In other words,
ż ż
1 1
F [f ](t) = f (t) =
´1 q f (ω)Mt (ω)dω = f (ω)eiωt dω .
2π R 2π R
1 1
Note that F ´1 = D´1 F = F D´1 . Similar to Theorem 1.5, F ´1 : L1 (R) Ñ C0 (R) is
2π 2π
a bounded linear map and lim fq(t) = 0.
|t|Ñ8

Next we consider the Fourier transform of square-integrable functions. First we note that
there exists square-integrable function which is not integrable. For example, the function
" ´3/4
x if x ą 1,
f (x) =
0 otherwise

belongs to L2 (R) but not L1 (R). It is żnot possible to find the Fourier transform for this f
using Definition 1.4 since the integral f (t)eitω dt does not exist. In other words, when we
R
talk about Fourier transform of functions that are not integrable, we indeed try to extend
the domain of the original Fourier transform F .
 Before proceeding, we introduce the inner product x¨, ¨yL2 (R) used in L2 (R):
ż
xf, gyL2 (R) = f (x)g(x)dx @ f, g P L2 (R),
R

1
and the induced norm is indeed the L2 -norm; that is, }f }L2 (R) = xf, f yL22 (R) . The space
(L2 (R), x¨, ¨yL2 (R) ) is a Hilbert space.
Since L1 (R) X L2 (R) is dense in L2 (R) (for example, if f P L2 (R), then fn ” f 1[´n,n] is
2
square-integrable and tfn u8
n=1 converges to f in L (R)), it is natural to define the Fourier
transform of a square-integrable function as the limit of the Fourier transform of a sequence
in L1 (R) X L2 (R) whose L2 -limit is that function (if it exists). To talk about whether the
limit of such a sequence exists, we need the following

Lemma 1.19. Suppose that f P L2 (R) and f vanishes outside a bounded interval. Then
fp P L2 (R) and }fp}2L2 (R) = 2π}f }2L2 (R) .

Proof. Note that since f P L2 (R) and f vanishes outside a bounded interval, Cauchy-
Schwarz inequality shows that f P L1 (R) so fp P C0 (R) is well-defined.
Suppose f vanishes outside [´R, R] for some R ą 0. Let c = R/π and define
? ´ixt
g(t) = (M´x D 1 f )(t) = ce f (ct),
c

where x P R is arbitrarily given. Then g vanishes outside [´π, π] and Theorem 1.7 shows
that gp(ω) = (T´x Dc fp)(ω) = (Dc fp)(ω + x). On the other hand, the Parseval identity (1.1.7)
implies that
8 ˇ żπ 8 ˇ ż
1 2
ÿ ˇ1 ´int ˇ
ˇ2 ÿ ˇ1
ˇ2
}g}L2 (R) = g(t)e dtˇ = g(t)e´int dxˇ
ˇ
2π 2π ´π 2π R
ˇ ˇ
n=´8 n=´8
8 8
1 ÿ ˇˇ ˇ2 1 ÿ ˇˇ ˇ2
= 2 gp(n) = 2
ˇ (Dc fp)(n + x)ˇ .
4π n=´8 4π n=´8

Integrating the identity above in x from 0 to 1, we obtain that

8 ż 8 ż
1 ÿ 1 ˇˇ ˇ2 1 ÿ n+1 ˇˇ ˇ2
}g}2L2 (R) = (Dc f )(n + x) dx =
p ˇ (Dc fp)(x)ˇ dx
2π n=´8 0 2π n=´8 n
ż
1 ˇ(Dc fp)(x)ˇ2 dx = 1 }Dc fp}2 2 = 1 }fp}2 2 .
ˇ ˇ
= L (R) L (R)
2π R 2π 2π

The lemma is then concluded since }f }L2 (R) = }g}L2 (R) . ˝

The collection of square-integrable functions vanishing outside a bounded interval is

denoted by L2c (R); that is,

L2c (R) = f : R Ñ C ˇ f P L2 (R) and f vanishes outside [´R, R] for some R ą 0 .

␣ ˇ (
Since the support of a function f (P L2c (R)), denoted by supp(f ), is defined as the closure
of the collection of points at which f does not vanish; that is,
(␣
supp(f ) = cl x P R ˇ f (x) ‰ 0 .
ˇ ()

Functions in L2c (R) are also called square-integrable functions with compact support.

Lemma 1.20. Let f P L2 (R). There exists a (unique) function F P L2 (R) such that if
tfn u8 Ď L2 (R) converges to f in L2 (R), then tfpn u8 converges to F in L2 (R).
n=1 c n=1

Proof. Let f P L2 (R), and tfn u8 2 2

n=1 Ď Lc (R) that converges to f in L (R). Then Lemma
1.19 shows that
?
}fpn ´ fx
m }L2 (R) = 2π}fn ´ fm }L2 (R) Ñ 0 as n, m Ñ 8.

2 2
In other words, tfpn u8
n=1 is a Cauchy sequence in L (R); thus the completeness of L (R)
2
shows that tfpn u8
n=1 converges to some function F in L (R).
Next we show that such a function F is independent of the sequence tfn u8
n=1 that is used
2
to approach f . Suppose that there is another sequence tgn un=1 Ď Lc (R) that also converges
8

to f in L2 (R). Then the argument above shows that tgpn u8

n=1 converges to some function G
in L2 (R). On the other hand, the sequence thn un=1
8
defined by
"
fn if n is odd,
hn =
gn if n is even

xn u8 converges to some function H P L2 (R). Never-

also converges to f in L2 (R); thus th n=1
theless, since the odd terms of th
xn u8 is a subsequence of tfpn u8 and the even terms of
n=1 n=1
xn u8 is a subsequence of tgpn u8 , we must have F = H = G.
th ˝
n=1 n=1

The lemma above induces the following

Definition 1.21. Let f P L2 (R). The Fourier transform of f , still denoted by F [F ] and fp,
is the L2 -limit of the Fourier transform of (any) sequences tfn u8 2
n=1 in Lc (R) that converges
to f in L2 (R). In other words,

F [f ] = fp ” lim fpn whenever tfn u8 1 2

n=1 Ď L (R) X L (R) and lim fn = f ,
nÑ8 nÑ8

where the two limits above are all in the L2 -sense.

Theorem 1.22 (Plancherel’s identity).

1 @p D
xf, gyL2 (R) = f , gp 2 @ f, g P L2 (R). (1.1.15)
2π L (R)

Proof. First we note that

1 p2
}f }2L2 (R) = }f }L2 (R) @ f P L2 (R) (1.1.16)
2π
 2 2
since by choosing tfn u8 n=1 Ď Lc (R) with L -limit f , by Lemma 1.19 we have
1 p 2 1 p2
}f }2L2 (R) = lim }fn }2L2 (R) = lim }fn }L2 (R) = }f }L2 (R) ,
nÑ8 nÑ8 2π 2π
2
where the last identity follows from the fact that tfpn u8 n=1 converges to f in L (R). Identity
p
(1.1.15) then follows from the polarization identity
1[ ]
xf, gyL2 (R) = }f + g}2L2 (R) ´ }f ´ g}2L2 (R) + i}f + ig}2L2 (R) ´ i}f ´ ig}2L2 (R) . ˝
4
Theorem 1.23. Let f, g P L2 (R).
@ D @ D
1. f, gp 2 = fp, g 2 .
L (R) L (R)

1 1
2. If g = fp, then f = gp. In other words, the operator ? p is the inverse of itself.
2π 2π
Remark 1.24. The Plancherel identity often refers to the following identity
1 p2
}f }2L2 (R) = }f }L2 (R) @ f P L2 (R). (1.1.17)
2π
In this lecture the Plancherel identity of the form (1.1.16) will be used extensively. Never-
theless, the Plancherel identity of the form (1.1.17) can be applied to show that
[ ÿ 8 ] ÿ8
F cn ϕn = xn @ tcn unPZ P ℓ2 and tϕn unPZ is orthonormal in L2 (R). (1.1.18)
cn ϕ
n=´8 n=´8

In other words, under the conditions stated above the Fourier transform F commutes with
infinite sums. To see why (1.1.18) is true, we first note that the Plancherel identity (1.1.15)
ϕ
! x )
shows that ? n is an orthonormal set in L2 (R). By setting
2π nPZ k
ÿ
sk = cn ϕn ,
n=´k

the Plancherel identity (1.1.17) implies that

ÿ
}spk ´ spℓ }2L2 (R) = 2π}sk ´ sℓ }2L2 (R) = 2π |cn |2 Ñ 0 as k, ℓ Ñ 8. (1.1.19)
mintk,ℓuă|n|ďmaxtk,ℓu

This shows that tspk ukPN is a Cauchy sequence in L2 (R); thus tspk ukPN converges in L2 (R).
k
xn , so the fact that tcn unPZ P ℓ2 and that ?ϕn
! x )
Nevertheless, spk = is an or-
ř
cn ϕ
n=´k 2π nPZ
2
thonormal basis in L (R) implies that
8
ÿ
lim spk = cn ϕ
xn and the convergence is in L2 (R).
kÑ8
n=´8

On the other hand, part 1 of Theorem 1.23 for all functions ϕ P L2 (R) we have
B ÿ 8 F A E B ÿ 8 F
cn ϕ
xn , ϕ = lim xspk , ϕyL2 (R) = lim sk , ϕ
p = cn ϕn , ϕ
p
kÑ8 kÑ8 L2 (R)
n=´8 L2 (R) n=´8 L2 (R)
B [ ÿ
8 ] F
= F cn ϕn , ϕ .
n=´8 L2 (R)

Since this identity holds for all ϕ P L2 (R), we conclude (1.1.18).

Remark 1.25. There are ways to define the Fourier transform for more general “functions”
(termed tempered distributions). The key idea is to look at relationships between functions
and their Fourier transforms (such as the Parseval identity) and define the Fourier transform
in a way that preserves these relationships. For example, by Theorem 1.23, for all f, g P
L2 (R) we have
ż ż
f (x)p
g (x)dx = fp(x)g(x)dx (1.1.20)
R R

Therefore, the Fourier transform of a general “function” f must satisfy (some variant of)
equation (1.1.20) whenever g is a function whose Fourier transform gp is well-defined and the
left-hand side of (1.1.20) makes sense. For instance, the function f (x) = eitx (with t P R) is
neither integrable nor square-integrable, but the integral on the left-hand side:
ż ż
f (x)p
g (x)dx ” gp(x)eitx dx
R R

is well-defined and equal (almost everywhere) to 2πg(t), provided that g P L1 (R) and
gp P L1 (R). In this case, the Fourier transform fp of f is defined as the “function” that
satisfies:
ż
2πg(t) = fp(x)g(x)dx
R

for any g satisfying certain conditions. However, it is important to note that no actual
function satisfies this identity in the traditional sense. Eventually, the right-hand side
integral must also be interpreted symbolically, no longer representing the usual Lebesgue
integral. In other words, to define the Fourier transform of more general “functions,” we:

1. Generalize the concept of the integral of the product of two functions by treating the
integral
ż
f (x)g(x)dx
R

as a bilinear form of the pair (f, g). This bilinear form is identical to the integral of
f g whenever f g P L1 (R).

2. The Fourier transform of a general function f is defined as something that satisfies

equation (1.1.20) whenever that identity makes sense.

This is essentially what the theory of tempered distributions is about, though we will not
delve into it further here.
The bottom line is: everything we derived for the Fourier transform of f with f P
L (R) Y L2 (R) is treated as identities that the Fourier transform of generalized functions
1

must satisfy.
1.2 Poisson Summation Formula
Theorem 1.26. If f P L1 (R), then the series
8
ÿ
f (t + 2nπ) (1.2.1)
n=´8

converges absolutely for almost all t in (0, 2π) and its sum belongs to L1 (0, 2π) and is 2π-
1 p
periodic. If tan unPZ denotes the Fourier coefficient of the series, then an = f (n) for all
2π
n P Z.
Proof. By the Monotone Convergence Theorem,
ż 2π ÿ8 8 ż 2π
ÿ 8 ż 2(n+1)π
ÿ
ˇ ˇ ˇ ˇ ˇ ˇ
ˇf (t + 2nπ)ˇ dt = ˇf (t + 2nπ)ˇ dt = ˇf (t)ˇ dt
0 n=´8 0 2nπ
żn=´8 n=´8
ˇ ˇ
= ˇf (t)ˇ dt = }f }L1 (R) ă 8;
R
8 8
thus the series |f (¨ + 2nπ)| belongs to L1 (0, 2π) which shows that
ř ř
f (t + 2nπ)
n=´8 n=´8
converges in L1 (0, 2π) and also converges absolutely for a.a. t P (0, 2π).
By the definition of the Fourier coefficients,
8 8 ż
1 2π ÿ 1 ÿ 2π
ż
´imt
am = f (t + 2nπ)e dt = f (t + 2nπ)e´imt dt
2π 0 n=´8 2π n=´8 0
8
1 ÿ 2(n+1)π
ż ż
´imt 1 1 p
= f (t)e dt = f (t)e´imt dt = f (m)
2π n=´8 2nπ 2π R 2π
which concludes the theorem. ˝
8
For f P L1 (R), define F (t) = f (t + 2nπ). Then Theorem 1.26 shows that F P
ř
n=´8
L1 (0, 2π) and is 2π-periodic. If the Fourier series of F converges pointwise to F , then
8 8
ÿ ÿ 1 p
f (t + 2nπ) = F (t) = f (m)eimt . (1.2.2)
n=´8 m=´8
2π
In particular, letting t = 0 in (1.2.2) we obtain
8 8
ÿ 1 ÿ p
f (2nπ) = f (n).
n=´8
2π n=´8
This identity is called the Poisson summation formula. The following theorem provides a
condition that (1.2.2) holds (if f is continuous).
Theorem 1.27. If there exists δ ą 0 such that

@ x P R,
ˇ ˇ ˇ ˇ
ˇf (x)ˇ + ˇfp(x)ˇ ď C(1 + |x|)´1´δ

then
8 8
ÿ 1 ÿ p
fp(x + 2nπ) = f (n)einx @ x P [0, 2π]. (1.2.3)
q
n=´8
2π n=´8

The proof of this theorem is left as an exercise.

1.3 Shannon Sampling Theorem
An analog signal is a piecewise continuous function of time defined on R, with the exception
of perhaps a countable number of jump discontinuities. Almost all analog signals of interest
in engineering have finite energy. By this we mean that the signal is square-integrable. The
norm of a signal f : R Ñ R defined by
(ż ˇ ˇ )1
}f }L2 (R) = ˇf (t)ˇ2 dt 2 (1.3.1)
R

represents the square root of the total energy content of the signal f . The spectrum of a
signal f is represented by its Fourier transform fp, where the variable of fp, usually denoted
ω
by ω, is called the frequency. The frequency is measured by ν = in terms of Hertz.
2π
A signal f is called band-limited if its Fourier transform has a compact support; that is,

fp(ω) = 0 whenever |ω| ą ω0 (1.3.2)

for some ω0 ą 0. If ω0 ą 0 is the smallest value for which (1.3.2) holds, then it is called
the bandwidth of the signal. Even if an analog signal f is not band-limited, we can reduce
it to a band-limited signal by what is called an ideal low-pass filtering. To reduce f to a
band-limited signal fω0 with bandwidth less than or equal to ω0 , we consider
"
fp(ω) if |ω| ď ω0 ,
fω0 (ω) =
x (1.3.3)
0 if |ω| ą ω0

and we find the low-pass filter function fω0 by the inverse Fourier transform
1 ω0 itω p
ż ż
1 itω x
fω0 (t) = e fω0 (ω)dω = e f (ω)dω .
2π R 2π ´ω0

Define the gate/window function

"
1 if |ω| ď ω0 ,
Πω0 (ω) = 1[´ω0 ,ω0 ] (ω) =
0 if |ω| ą ω0 .

Then Πω0 P L1 (R) X L2 (R) whose inverse Fourier transform is given by

sin(ω0 t)
ż
1
Π}ω0 (t) = Πω0 (ω)eiωt dω = .
2π R πt
Therefore, by Theorem 1.12 we obtain that
sin(ω0 (τ ´ t))
ż ż
1 itω p
fω0 (t) = Πω0 (ω)e f (ω)dω = (f ˙ Πω0 )(t) =
} f (τ )dτ .
2π R R π(t ´ τ )
This gives the sampling integral representation of a band-limited signal fω0 . Thus, fω0 (t) can
sin ω0 (t ´ ¨)
be interpreted as the weighted average of f with the Fourier kernel as weight.
π(t ´ ¨)
In the field of digital signal processing, the sampling theorem is a fundamental bridge
between continuous-time signals (often called “analog signals”) and discrete-time signals
(often called “digital signals”). It establishes a sufficient condition for a sample rate（取
樣頻率）that permits a discrete sequence of samples to capture all the information from a
continuous-time signal of finite bandwidth. To be more precise, Shannon’s version of the
theorem states that “if an analog signal contains no frequencies higher than B hertz, it is
1
completely determined by giving its ordinates at a series of points spaced seconds apart.”
2B
Theorem 1.28. Let f : R Ñ R be a continuous integrable function. If supp(fp) Ď [´ω0 , ω0 ],
! ( ))8
πk
then f is fully determined by the sequence f , and
ω0 k=´8
ÿ8 ( πk )
f (t) = f sinc(ω0 t ´ kπ) @ t P R, (1.3.4)
k=´8
ω0
where the sinc function is given by
& sin x if x ‰ 0,
$

sinc(x) = x (1.3.5)
% 1 if x = 0.

Proof. Let f P L1 (R) X C (R; R) such that supp(fp) Ď [´B, B]. Then fp P L1 (R) (since it
must be bounded by Theorem 1.5); thus the Fourier inversion formula implies that
ż ż ω0
1
f (t) = ixω
fp(ω)e dω = fp(ω)eixω dω @ t P R.
2π R ´ω0

In particular,
( kπ ) 1
ż ω0
ikπω
f = fp(ω)e ω0
dω .
ω0 2π ´ω0
! ( ))8
π ´kπ
Treating fp as a function defined on [´ω0 , ω0 ], the identity above shows that f
ω0 ω0 k=´8
is the Fourier coefficients of fp.
Note that the boundedness of fp implies that fp P L2 (´ω0 , ω0 ). Therefore, if g P
L2 (´ω0 , ω0 ), the Parseval identity, together with the polarization identity, implies that
π ( ´kπ )
ż ω0 8
1 ÿ
f (ω)g(ω) dω =
p f gpk , (1.3.6)
2ω0 ´ω0 k=´8
ω0 ω0
ż ω0
1 ´ ikπx
where gpk = g(t)e ω0 dx.
2ω0 ´ω0

For each t P R, the Fourier coefficients of the function g(ω) = e´itω is given by
ż ω0 ż ω0
1 ikπω
´itω ´ ω0 1 (ω t+kπ)
´i 0ω ω
gpk = e e dω = e 0 dω = sinc(ω0 t + kπ);
2ω0 ´ω0 2ω0 ´ω0
thus the Fourier inversion formula and (1.3.6) imply that if x P R,
8
( ´kπ )
ż ż ω0
1 itω 1 ÿ
f (t) = fp(ω)e dω = fp(ω)g(ω) dx = f sinc(ω0 t + kπ)
2π R 2π ´ω0 k=´8
ω0
8
ÿ ( kπ )
= f sinc(ω0 t ´ kπ).
ω0
k=´8
! ( ))8
kπ
The identity above shows that f is fully determined by the sequence f . ˝
ω0 k=´8
Remark 1.29. Equation (1.3.4) is called the Whittaker–Shannon interpolation for-
mula.

Remark 1.30. The reconstruction formula (1.3.4) can be obtained formally as follows. In
the proof of the sampling theorem, since we have obtained that the Fourier coefficient of fp
!
π
( ´kπ ))8
is f , so we have
ω0 ω0 k=´8

π ( ´kπ ) ikπω π ( kπ ) ´ ikπω

ÿ8 ÿ8
fp(ω) = f e ω0
= f e ω0 .
k=´8
ω0 ω0 k=´8
ω0 ω0

Taking the inverse Fourier transform,

π ( kπ ) ´ ikπω ( kπ ) ikπω
ż 8 ż ÿ 8
1 ÿ
itω 1 ´
f (t) = f e ω0
e dω = f e ω0 eitω dω ;
2π R k=´8 ω0 ω0 2ω0 R k=´8 ω0

thus if we can switch the order of the infinite sum and the integration, we then immediately
obtain the reconstruction formula

1 ÿ ( kπ ) ( kπ )
8 ż 8
´ ikπω
ÿ
f (t) = f e ω0 eitω dω = f sinc(ω0 t ´ kπ).
2ω0 k=´8 ω0 R k=´8
ω0

In fact, this kind of “switch of order of the infinite sum and the integral” is usually valid
when the integral is due to the Fourier transform/inverse transform. In particular, later we
will take the Fourier transform of series of the form
8
ÿ 8
ÿ
cn ϕ(x ´ n) = cn (Tn ϕ)(x)
n=´8 n=´8

8
where ϕ P L2 (R) and tcn u8 2
n=´8 P ℓ ; that is, |cn |2 ă 8, and directly switch the order
ř
n=´8
of summation and the Fourier transform as
[ ÿ 8 ] 8
ÿ
F cn (Tn ϕ) = cn F [Tn ϕ].
n=´8 n=´8

A rigorous proof of the identity above will be ignored.

1.4 Time-Frequency Analysis

In signal processing, time-frequency analysis comprises those techniques that study a signal
in both the time and frequency domains simultaneously, using various time-frequency repre-
sentations. The mathematical motivation for this study is that functions and their transform
representation are tightly connected, and they can be understood better by studying them
jointly, as a two-dimensional object, rather than separately. A simple example is that the
4-fold periodicity of the Fourier transform - and the fact that two-fold Fourier transform
reverses direction - can be interpreted by considering the Fourier transform as a 90° rota-
tion in the associated time-frequency plane: 4 such rotations yield the identity, and 2 such
rotations simply reverse direction (reflection through the origin).
The practical motivation for time-frequency analysis is that classical Fourier analysis is
quite inadequate for most applications. In the first place, the Fourier analysis assumes that
signals are infinite in time or periodic, while many signals in practice are of short duration,
and change substantially over their duration. For example, traditional musical instruments
do not produce infinite duration sinusoids, but instead begin with an attack, then gradually
decay. Moreover, to extract the spectral information fp(ω) from the analog signal f , the
Fourier transform takes an infinite amount of time, using both past and future information
of the signal just to evaluate the spectrum at a single frequency ω. Besides, the formula
ż
(F f )(ω) = f (ω) = f (t)e´itω dt
p @ω P R (1.1.10)
R

does not even reflect frequencies that evolve with time. What is really needed is for one to
be able to determine the time intervals that yield the spectral information on any desirable
range of frequencies (or frequency band). In addition, since the frequency of a signal is
directly proportional to the length of its cycle, it follows that for high-frequency spectral
information, the time-interval should be relatively small to give better accuracy, and for low-
frequency spectral information, the time-interval should be relatively wide to give complete
information. In other words, it is important to have a flexible time-frequency window that
automatically narrows at high “center-frequency” and widens at low “center-frequency”. All
these issues are poorly represented by traditional methods, which motivates time-frequency
analysis.
In this lecture, concerning the time-frequency analysis we are going to study the following
subjects:

1. The Gabor Transform/Short-Time Fourier Transform (STFT): For a given

f P L2 (R), we consider the following integral
ż
G[f ](t, ω) = f (τ )gt,ω (τ )dτ ,
R

where gt,ω P L2 (R) is a time-localization window function taking the form

gt,ω (τ ) = ϕ(τ ´ t)eiωτ = (Mω Tt ϕ)(τ ) (1.4.1)

for some ϕ P L2 (R) and is used for extracting local information from a Fourier trans-
form of a signal f .
We are going to study the basic properties of the Gabor transform and derive the
inversion formula
[ ]
ż ż ż
1
f (t) = e G[f ](b, ω)ϕ(t ´ b)dωdb = F ´1 G[f ](b, ¨) (t)(Tb ϕ)(x)db.
iωt
2π R R R
 2. The Wigner-Ville Distribution (WVD) and Transform (WVT): The cross
Wigner-Ville Distribution of f, g P L2 (R) is the integral
ż (
τ) ( τ ) ´iωτ
Wf,g (t, ω) = f t + g t´ e dτ . (1.4.2)
R 2 2
At the first glance it looks very similar to the STFT since with the substitution of
τ
variable x = t + ,
2
ż ż
´2i(x´t)ω 2itω
Wf,g (t, ω) = 2 f (x)g(x ´ 2t)e dx = 2e f (x)g(x ´ 2t)e´2ixω dx
R R

and the last integral is indeed a STFT of f with window function g (at the point
(2t, 2ω)). When the windows function is the signal itself, it is called the Wigner-Ville
Transform of the signal. In other words, the Wigner-Ville Transform of f P L2 (R) is
Wf,f (which is also denoted by Wf in the textbook). We note that WVT is a nonlinear
map of the input f .
The WVD is a great method to perform time-frequency analysis; however, the study
of the Wigner-Ville transform is beyond the scope of this course, so we will only talk
about this breifly.

3. The Wavelet Transform: A given function ψ P L2 (R) satisfying the “admissibility”

condition ż 2
|ψ(ω)|
p
Cψ = dω ă 8,
R |ω|
is called a basic wavelet. The “integral wavelet transform” relative to the “basic
wavelet” ψ, denoted by Wψ (do not confuse with the Wigner-Ville transform), is an
integral operator on L2 (R) defined by
ż (t ´ b) ż
´ 12
Wψ [f ](a, b) = |a| f (x)ψ dt = f (t)(Tb Da ψ)(t)dt.
R a R

We are going to study the basic properties of the wavelet transform and derive the
inversion formula
ż ż
1 dbda
f (t) = Wψ [f ](a, b)(Tb Da ψ)(t) . (1.4.3)
Cψ R R a2

1.5 The Wavelet Series and Frames

For the purpose of localization, the basis wavelet is usually chosen to have compact support
or decay very fast, similar to the case of Fourier transform v.s. Fourier series, we might be
able to reconstruct f using only values of the wavelet transform at discrete points. Through-
out the lecture, any function ψ P L2 (R) is associated with a bi-infinite sequence of functions
tψj,k uj,kPZ defined by
ψj,k (x) = 2j/2 ψ(2j x ´ k) @ j, k P Z. (1.5.1)
We note that the index j measures the width of the window/support, and the parameter k
is used to represent translation of the window to cover the whole time domain.

Definition 1.31. A function ψ P L2 (R) is called an orthogonal wavelet (or o.n. wavelet),
if the family tψj,k uj,kPZ , as defined in (1.5.1), is an orthonormal basis of L2 (R); that is,

xψj,k , ψℓ,m yL2 (R) = δjℓ δkm @ j, k, ℓ, m P Z (1.5.2)

and every function f P L2 (R) can be written as

8
ÿ
f (x) = cj,k ψj,k (x), (1.5.3)
j,k=´8

where cj,k = xf, ψj,k yL2 (R) and the convergence of the series in (1.5.3) is in L2 (R); that is,
› N2
ÿ N1
ÿ
›
› ›
lim ›f ´ cj,k ψj,k ›› = 0.
M1 ,N1 ,M2 ,N2 Ñ8 ›
j=´M2 k=´M1 L2 (R)

Any arbitrarily given ψ P L2 (R) is most likely not an o.n. wavelet. For a given function
ψ P L2 (R), the first few questions we would like to answer are

1. Is the linear span of tψj,k uj,kPZ dense in L2 (R)?

2. If so, is there a effect way to express a function f P L2 (R) in terms of “linear combi-
nations” of ψj,k ’s?

3. Is the expression of a function f P L2 (R) as a “linear combinations” of ψj,k ’s unique?

The abstraction of the aforementioned questions leads to the concept of frames and frame
operators defined below.

Definition 1.32 (Frame). A sequence txn u in a separable Hilbert space (H, x¨, ¨y) (not
necessarily a basis of H) is called a frame if there exist two numbers A and B with
0 ă A ď B ă 8 such that

ˇxx, xn yˇ2 ď B}x}2 .

ÿˇ ˇ
A}x}2 ď (1.5.4)
n

The numbers A and B are called the frame bounds.

Definition 1.33 (Frame Operator). To each frame txn u there corresponds an operator T ,
called the frame operator, from H into itself defined by
ÿ
Tx = xx, xn y xn @x P H. (1.5.5)
n

For a given frame txn u of a Hilbert space, it can be shown that the frame operator is
bounded, and eventually answer (partially) the questions above by the following
Theorem 1.34. Suppose txn u8 n=1 is a frame on a separable Hilbert space with frame bounds
A and B, and T is the corresponding frame operator. Then,

(a) T is invertible and B ´1 I ď T ´1 ď A´1 I.

(b) tT ´1 xn u8
n=1 is a frame, called the dual frame of txn un=1 , with frame bounds B
8 ´1
and
A´1 .

(c) Every x P H can be expressed in the form

8
ÿ 8
ÿ
x= xx, T ´1 xn y xn = xx, xn y T ´1 xn . (1.5.6)
n=1 n=1

Later we will prove the following theorem which provides a sufficient condition of a given
ψ to generate a frame tψm,n um,n in L2 (R).

Theorem 1.35. Let ϕ P L2 (R), and a0 ą 1. If

8
ÿ ˇ m ˇ2
(i) there exist A, B ą 0 such that A ď ϕ(a0 ω)ˇ ď B for all 1 ď ω ď a0 , and
ˇp
m=´8
8
ÿ ˇ m ˇˇ m ˇ
(ii) sup ϕ(a0 ω + x)ˇ ď C(1 + |x|)´(1+δ) for some constants C and δ ą 0,
ϕ(a0 )ˇˇp
ˇp
ωPR m=´8

then there exists rb ą 0 such that the family tϕm,n um,nPZ given by

´m/2
ϕm,n = a0 ϕ(a´m
0 x ´ nb0 ) = (Dam
0
Tnb0 ϕ)(x)

forms a frame in L2 (R) for any b0 P (0, rb).

Suppose that a function ψ P L2 (R) generates a frame Ψ = tψj,k uj,kPZ in L2 (R). Even
though in theory we have a representation formula (1.5.6) using the frame operator T
associated with Ψ, it is still not practical enough since it requires to compute the dual
frame tT ´1 ψm,n um,nPZ . We hope, just like how we obtain Ψ, that the dual frame is a frame
simply generated by a function ψr P L2 (R); that is,

T ´1 ψm,n = ψrm,n @ m, n P Z.

However, this requires a subtle design of the function ψ.

Definition 1.36. A function ψ P L2 (R) is called an R-function if tψj,k uj,kPZ , as defined in

(1.5.1), is a Riesz basis of L2 (R), in the sense that the linear span of tψj,k uj,kPZ is dense in
L2 (R) and that there exist positive constants A and B, with 0 ă A ď B ă 8, such that
› 8 ›2
› ›2 › ÿ › › ›2
A tcj,k uj,kPZ ℓ2 ď ›
› › › cj,k ψj,k ›› ď B ›tcj,k uj,kPZ ›ℓ2 (1.5.7)
2L (R)
j,k=´8
for all doubly bi-infinite square-summable sequences tcj,k uj,kPZ ; that is,
8
›tcj,k uj,kPZ ›22 ”
› › ÿ
ℓ
|cj,k |2 ă 8.
j,k=´8

An R-function ψ P L2 (R) is called an R-wavelet (or wavelet), if there exists a function

ψr P L2 (R), such that tψj,k uj,kPZ and tψrj,k uj,kPZ , as defined in (1.5.1), are dual bases of L2 (R).
If ψ is an R-wavelet, then ψr is called a dual wavelet corresponding to ψ.

Remark 1.37. A collection of functions tϕn unPZ in L2 (R) is said to have Riesz bounds A
and B, where 0 ă A ď B ă 8, if
›2
› 8
› ÿ
› ›2 › › ›2
A tcn unPZ ℓ2 ď ›
› › › cn ϕn ›› ď B ›tcn unPZ ›ℓ2 @ tcn unPZ P ℓ2 . (1.5.8)
n=´8 L2 (R)

An orthonormal system in L2 (R) indeed has Riesz bounds 1 and 1. Similar to (1.1.18), the
argument used to establish (1.1.18) can be applied to show that
[ ÿ
8 ] 8
ÿ
F cn ϕn = cn ϕ
xn @ tcn unPZ P ℓ2 and tϕn unPZ satisfying (1.5.8). (1.5.9)
n=´8 n=´8

The key difference in the argument is that one has to modify (1.1.19) using (1.5.8):
ÿ
}spk ´ spℓ }2L2 (R) = 2π}sk ´ sℓ }2L2 (R) ď 2πB |cn |2 Ñ 0 as k, ℓ Ñ 8.
mintk,ℓuă|n|ďmaxtk,ℓu

We also remark that using (1.5.9), if tϕn unPZ Ď L2 (R) has Riesz bounds A and B, then
␣ (
ϕ
xn
nPZ
has Riesz bounds 2πA and 2πB because of the Plancherel identity.

Now, the question is how do we construct such kind of wavelet?

1.6 Multi-resolution Analysis and Construction of O.N.

Wavelets
A wavelet ψ in L2 (R) is called a semi-orthogonal wavelet (or wavelet) if the family tψj,k uj,kPZ
it generates satisfies
xψj,k , ψℓ,m y = 0 if j ‰ ℓ, j, k, ℓ, m P Z. (1.6.1)

Suppose that ψ is a semi-orthogonal wavelet and consider the family tψj,k uj,kPZ it generates.
For each j P Z, let Wj denote the closure of the linear span of tψj,k ukPZ , namely:
(␣
ψj,k ˇ k P Z .
ˇ ()
Wj ” closure}¨}2 (1.6.2)

Then it is clear that Wj K Wℓ if j ‰ ℓ, meaning that

xgj , gℓ y = 0 if j ‰ ℓ, gj P Wj and gℓ P Wℓ . (1.6.3)

Moreover, by the fact that the linear span of tψj,k uj,kPZ is dense in L2 (R), L2 (R) can be
decomposed as a direct sum of the spaces Wj :
à
L2 (R) = Wj = ¨ ¨ ¨ ‘ W´1 ‘ W0 ‘ W1 ‘ ¨ ¨ ¨ , (1.6.4)
jPZ

in the sense that every function f P L2 (R) has a unique decomposition:

8
ÿ
f (x) = gj (x) = ¨ ¨ ¨ + g´1 (x) + g0 (x) + g1 (x) + ¨ ¨ ¨ , (1.6.5)
j=´8

8
where gj = xf, ψrj,k y ψj,k P Wj for all j P Z.
ř
k=´8
For each j P Z, let us consider the closed subspaces
j´1
à
Vj ” Wℓ = ¨ ¨ ¨ ‘ Wj´2 ‘ Wj´1 (1.6.6)
ℓ=´8

of L2 (R). These subspaces clearly have the following properties:

(10 ) ¨ ¨ ¨ Ĺ V´1 Ĺ V0 Ĺ V1 Ĺ ¨ ¨ ¨ or Vj Ĺ Vj+1 for all j P Z;

(Ť )
(20 ) closure}¨}2 Vj = L2 (R);
jPZ

(30 )
Ş
Vj = t0u;
jPZ

(40 ) Vj+1 = Vj ‘ Wj for all j P Z; and

(50 ) For all j P Z, f P Vj if and only if d1/2 f P Vj+1 , where for λ ą 0 the dilation operator
dλ is given by (dλ f )(x) = f (λ´1 x).

Hence, in contrast to the subspaces Wj which satisfy

Wj X Wℓ = t0u if j ‰ ℓ,

the sequence of subspaces Vj is nested, as described by (10 ), and has the property that every
function f in L2 (R) can be approximated as closely as desirable by its projections Pj f in Vj ,
as described by (20 ). But on the other hand, by decreasing j, the projections Pj f could have
arbitrarily small energy, as guaranteed by (30 ). What is not described by (10 )-(30 ) is the
most important intrinsic property of these spaces which is that more and more “variations”
of Pj f are removed as j Ñ ´8. In fact, these variations are peeled off, level by level in
decreasing order of the “rate of variations” (better known as “frequency bands”) and stored
in the complementary subspaces Wj as in (40 ). This process can be made very efficient by
an application of the property (50 ).
In fact, if the reference subspace V0 , say, is generated by a single function ϕ P L2 (R) in
the sense that
(␣
ϕ0,k ˇ k P Z ,
ˇ ()
V0 ” closure}¨}2 (1.6.7)
where
ϕj,k (x) = 2j/2 ϕ(2j x ´ k), (1.6.8)

then all the subspaces Vj are also generated by the same ϕ (just as the subspaces Wj are
generated by as ψ in (1.6.2)), namely:
(␣
ϕj,k ˇ k P Z @ j P Z.
ˇ ()
Vj ” closure}¨}2 (1.6.9)

Hence, the “peeling-off” process from Vj to Wj´1 , Wj´2 , ¨ ¨ ¨ , Wj´ℓ can be accomplished
efficiently.

Definition 1.38. An MRA consists of a sequence tVm | m P Zu of embedded closed sub-

spaces of L2 (R) that satisfy the following conditions:

(i) ¨ ¨ ¨ Ď V´2 Ď V´1 Ď V0 Ď V1 Ď V2 Ď ¨ ¨ ¨ Ď Vm Ď Vm+1 Ď ¨ ¨ ¨ ;

8 ( Ť
8 )
(ii) Vm is dense in L2 (R); that is, closure}¨}2 Vm = L2 (R).
Ť
m=´8 m=´8

8
(iii)
Ş
Vm = t0u.
m=8

(iv) f P Vm if and only if d1/2 f P Vm+1 for all m P Z;

(v) there exists a function ϕ P V0 such that ϕ0,n = Tn ϕ ˇ n P Z is an orthonormal basis

␣ ˇ (

for V0 ; that is,

8
ˇxf, ϕn yˇ2
ÿ ˇ ˇ
}f }2L2 (R) = @ f P V0 .
n=´8

The function ϕ is called the scaling function or father wavelet. If tVm umPZ is a multi-
resolution of L2 (R) and if V0 is the closed subspace generated by the integer translates of a
single function ϕ, then we say that ϕ generates the MRA.

Remark 1.39. To define scaling function, sometimes condition (v) is relaxed by assuming
that Tn ϕ ˇ n P Z is a Riesz basis for V0 . Nevertheless, in most of the applications we look
␣ ˇ (

for ϕ such that tTn ϕ | n P Zu is an orthonormal basis of V0 , so we simply use Definition 1.38
(which is the one used in the textbook) for the scaling functions.

Theorem 1.40 (Orthonormalization Process). If ϕ P L2 (R) and if tTn ϕ | n P Zu is a Riesz

basis of V0 , then Tn ϕr ˇ n P Z is an orthonormal basis of V0 with
␣ ˇ (

ϕ(ω)
p
ϕ(ω) =d . (1.6.10)
p
r
8 ˇ
ř ˇ2
ϕ(ω + 2kπ)ˇ
ˇp
k=´8
Theorem 1.41. If tVn unPZ is an MRA with the scaling function ϕ, then there is an orthog-
onal wavelet ψ given by
8
ÿ
ψ(x) = (´1)´n´1 c´n´1 (D1/2 Tn ϕ)(x), (1.6.11)
n=´8

where the coefficients cn are given by

? ż
cn = xϕ, ϕ1,n y = 2 ϕ(x)ϕ(2x ´ n)dx. (1.6.12)
R

We will also give some examples of constructing orthogonal wavelet based on the theo-
rems above.
Chapter 6

The Wavelet Transforms and Their

Basic Properties

6.2 Continuous Wavelet Transforms and Examples

Definition 6.1 (Wavelet). A wavelet is a function ψ P L2 (R) which satisfies the admissibil-
ity condition
ż ˇˇ p ˇˇ2
ψ(ω)
Cψ ” dω ă 8, (6.2.1)
R |ω|
where ψp is the Fourier transform of ψ.

If ψ P L2 (R), then ψa,b P L2 (R) for all a and b since

1 ˇˇ ( t ´ b )ˇˇ
ż ˇ ˇ2 ż
2
ˇ ˇ2
}ψa,b }L2 (R) = ψ dt = ˇψ(x)ˇ dx = }ψ}2L2 (R) ă 8.
R |a|
ˇ a ˇ
R

The Fourier transform of ψa,b is given by

( ) a
ψ
y a,b (ω) = M ´b D 1 ψ
p (ω) = |a|e´ibω ψ(aω).
p (6.2.2)
a

Definition 6.2 (Continuous Wavelet Transform). If ψ P L2 (R), and ψa,b is given by

1 (t ´ b)
ψa,b (t) = a ψ = (Tb Da ψ)(t), (6.2.3)
|a| a

then the integral transformation Wψ defined on L2 (R) by

ż
Wψ [f ](a, b) = xf, ψa,b yL2 (R) = f (t)ψa,b (t)dt (6.2.4)
R

is called a continuous wavelet transform of f (relative to the wavelet ψ).

Using the Plancherel identity of the Fourier transform, it also follows from (6.2.4) that
ż a
1 @ p yD 1 ibω
Wψ [f ](a, b) = xf, ψa,b yL2 (R) = f , ψa,b 2 = |a|fp(ω)ψ(aω)e
p dω
2π L (R) 2π R
ż
1 ibω
= (fpD 1 ψ)(ω)e
p dω .
2π R a

26
This means that for a fixed a Wψ [f ](a, ¨) is the inverse Fourier transform of the function
fpD 1 ψp so the Fourier inversion formula shows that
a

[ ]
ż
F Wψ [f ](a, ¨) (ω) =
a
Wψ [f ](a, t)e´itω dt = (fpD 1 ψ)(ω)
p = |a|fp(ω)ψ(aω).
p (6.2.5)
a
R

Example 6.3 (The Haar Wavelet). The Haar wavelet (Haar 1910) is one of the classic
examples. It is defined by $
1
’
’ 1 if 0 ď t ă ,
’
& 2
ψ(t) = 1 (6.2.6)
´1 if ď t ă 1,
’
’
’ 2
0 otherwise.
%

The Haar wavelet has compact support. It is obvious that

ż ż
ˇψ(t)ˇ2 dt = 1.
ˇ ˇ
ψ(t)dt = 0,
R R

This wavelet is very well localized in the time domain, but it is not continuous. Its Fourier
transform ψp is calculated as follows: for ω ‰ 0 we have

i ( ´iω )
1 ż1
e´iωt ˇˇt= 12 e´iωt ˇˇt=1
ż
2
´iωt ´iωt ´iω
ψ(ω)
p = e dt ´ e dt = ´ = 2e 2 ´1´e
ω
ˇ ˇ
0 1 ´iω t=0 ´iω t= 12
( iω )( iω
2
´i ( ´iω )2 ´i ´iω )2
= 1´e 2 = exp ´ e4 ´e 4
ω ω 2
( iω ) sin2 (ω/4)
= i exp ´ (6.2.7)
2 ω/4

and for ω = 0 we have ψ(0)

p = 0; thus

ż ˇˇ p ˇˇ2
ψ(ω)
ż
ω
dω = 16 |ω|´3 sin4 dω ă 8.
R ω R 4

Both ψ and ψp are plotted in Figure 6.1.

ˇ ˇ
ψ(t) ˇψ(ω)
p ˇ

1 ‚ 1

1
t ‚ ‚ ‚ ‚ ‚ ‚ ω
0.5 ´12π´8π´4π 4π 8π 12π

´1

Figure 6.1: The Haar wavelet and its Fourier transform

Theorem 6.4. If ψ is a wavelet and ϕ is a bounded integrable function, then the convolution
function ψ ˙ ϕ is a wavelet.
Proof. By Young’s inequality,

}ψ ˙ ϕ}L2 (R) ď }ϕ}L1 (R) }ψ}L2 (R) ă 8,

so ψ ˙ ϕ P L2 (R). Moreover, by the fact that }ϕ}

p L8 (R) ď }ϕ}L1 (R) ,
ż ˇˇ{ ˇ2 ż ˇˇ p ˇˇ2 ˇˇp ˇˇ2
ψ ˙ ϕ(ω)ˇ ψ(ω) ϕ(ω)
dω = dω ď }ϕ}2L1 (R) Cψ ă 8.
R |ω| R |ω|
Thus, the convolution function ψ ˙ ϕ is a wavelet. ˝
Example 6.5. This example illustrates how to generate other wavelets by using Theorem
6.4. For example, if we take the Haar wavelet and convolute it with the following function
"
1 if 0 ď t ď 1,
ϕ(t) = 1[0,1] (t) =
0 otherwise,
we obtain a simple wavelet, as shown in Figure 6.2.
(ψ ˙ ϕ)(t)

1
2
3
2
1
t
2
´ 12

Figure 6.2: The wavelet ψ ˙ ϕ

Example 6.6. The convolution of the Haar wavelet with ϕ(t) = exp(´t2 ) generates a
smooth wavelet, as shown in Figure 6.3.

(ψ ˙ ϕ)(t)

0.2

0.1

t
´4 ´2 2 4
´0.1

´0.2

Figure 6.3: The wavelet ψ ˙ ϕ

Example 6.7 (The Mexican Hat Wavelet). The Mexican hat wavelet is defined by the
second derivative of a Gaussian function as
( t2 ) d2 ( t2 )
2
ψ(t) = (1 ´ t ) exp ´ = ´ 2 exp ´ = ψ1,0 (t),
2 dt 2 (6.2.8)
? ( ω2 )
2
ψ(ω)
p =ψy 1,0 (ω) = 2π ω exp ´ .
2
The Mexican hat wavelet ψ1,0 and its Fourier transform are shown in Figure 6.4(a)(b). This
wavelet has excellent localization in time and frequency domains and clearly satisfies the
admissibility condition.

(a) ψ(t) (b) ψ(ω)

p

? ?
´ 3 3
t ? ? ω
´ 2 2

Figure 6.4: (a) The Mexican hat wavelet ψ1,0 and (b) its Fourier transform ψ
y 1,0

Two other wavelets, ψ 3 ,´2 and ψ 1 ,?2 , from the mother wavelet (6.2.8) can be obtained.
2 4
These three wavelets, ψ1,0 , ψ 3 ,´2 , and ψ 1 ,?2 , are shown in Figure 6.5(i), (ii), and (iii),
2 4
respectively.
ψa,b (t)

1.5

1 (ii) (i) (iii)

0.5

0 t
´0.5

´1
´8 ´6 ´4 ´2 0 2

Figure 6.5: Three wavelets ψ1,0 , ψ 3 ,´2 , ψ 1 ,?2

2 4

1
Remark 6.8. If in addition ψ P L (R) (which is usually the case if ψ has rapid decay),
then its Fourier transform ψp is bounded continuous. Since ψp is continuous, Cψ can be finite
ż
only if ψ(0)
p = 0 or, equivalently, ψ(t)dt = 0. This means that ψ must be an oscillatory
R
function with zero mean.

6.3 Basic Properties of Wavelet Transforms

The following theorem gives several properties of continuous wavelet transforms.
Theorem 6.9. If ψ and ϕ are wavelets and f , g are functions which belong to L2 (R), then
(i) (Linearity) For any scalars α and β,

Wψ [αf + βg] = αWψ [f ] + βWψ [g].

 (ii) (Translation) With Tc denoting the translation operator defined by (Tc f )(t) = f (t ´ c),

Wψ [Tc f ](a, b) = Wψ [f ](a, b ´ c) and WTc ψ [f ](a, b) = Wψ [f ](a, b + ca).

(iii) (Dilation) For c ą 0, with Dc denoting the (scaled) dilation operator defined by
1
(t)
(Dc f )(t) = ? f ,
c c
1 (a b) 1
Wψ [Dc f ](a, b) = ? Wψ [f ] , and WDc ψ [f ](a, b) = ? Wψ [f ](ac, b).
c c c c

(iv) (Symmetry) For any a ‰ 0,

( b)
Wψ [f ](a, b) = Wf [ψ] 1, ´ .
a
(v) (Parity) With P denoting the parity operator defined by (P f )(t) = f (´t) (that is,
P = D´1 ),
WP ψ [P f ](a, b) = Wψ [f ](a, ´b).

(vi) (Anti-linearity) For any scalars α, β,

Wαψ+βϕ [f ] = αWψ [f ] + βWϕ [g].

Proofs of the above properties are straightforward and are left as exercises.

Theorem 6.10 (Parseval’s Formula for Wavelet Transforms). If ψ P L2 (R) is a wavelet

and Wψ [f ] is the wavelet transform of f (relative to ψ) defined by (6.2.4), then, for any
functions f , g P L2 (R), we obtain
ż ż
dbda
Wψ [f ](a, b)Wψ [g](a, b) 2 = Cψ xf, gyL2 (R) , (6.3.1)
R R a
where Cψ is defined by (6.2.1).

Proof from the textbook. By Parseval’s relation (3.4.37) for the Fourier transforms, we have
1 @ p yD 1 @p { D
Wψ [f ](a, b) = xf, ψa,b yL2 (R) = f , ψa,b 2 = f , Tb Da ψ 2
2π L (R)
ż 2π L (R)

1 p 1 1
fp(ω)|a| 2 eibω ψ(aω)dω
D
= xf , M´b D 1 ψp 2 = p , (6.3.2)
2π a L (R) 2π R
and substituting g for f in the identity above,
ż
1 1
Wψ [g](a, b) = gp(σ)|a| 2 e´ibσ ψ(aσ)dσ
p . (6.3.3)
2π R
Substituting (6.3.2) and (6.3.3) in the left-hand side of (6.3.1) gives
ż ż
dbda
Wψ [f ](a, b)Wψ [g](a, b) 2
R R a
ż ż ż ż
1 1 1 dbda
= 2
p g (σ)|a| 2 eib(ω´σ) ψ(aσ)dσdω
fp(ω)|a| 2 ψ(aω)p p
(2π) R R R R a2
ż ż ż ż
1 1 p p ib(ω´σ)
= 2
f (ω)ψ(aω)pg (σ)ψ(aσ)e
p dσdωdbda. (6.3.4)
(2π) R R R R |a|
Note that
ż (ż ) ż ż
1 1
f (t) = F ´1
[fp](t) = f (s)e ´isω
e iωt
dsdω = f (s)eiω(t´s) dsdω ;
2π R R 2π R R

thus interchanging the order of integration from dωdb to dbdω in (6.3.4) we obtain that
ż ż
dbda
Wψ [f ](a, b)Wψ [g](a, b) 2
R R a
ż ż ( ż ż )
1 1 p p 1 ib(ω´σ)
= f (ω)ψ(aω) gp(σ)ψ(aσ)e
p dωdb dσda
2π R R |a| 2π R R
ż ż ż ż ˇ ˇ2
1 1 p p 1 ˇψ(aω)
p ˇ
= f (ω)ψ(aω)p g (ω)ψ(aω)dσda
p = g (ω)
fp(ω)p dωda.
2π R R |a| 2π R R |a|

A further interchanging the order of integration and putting aω = x show that

(ż ˇp )
ˇψ(aω)ˇ2
ż ż ż ˇ
dbda 1
Wψ [f ](a, b)Wψ [g](a, b) 2 = f (ω)p
p g (ω) da dω
R R a 2π R R |a|
ż ( ż ˇ p ˇ2 ) ż
1 ˇψ(x)ˇ Cψ Cψ @ p D
= f (ω)p
p g (ω) dx dω = fp(ω)pg (ω)dω = f , gp 2 ,
2π R R |x| 2π R 2π L (R)

and the Plancherel identity is used to conclude the theorem. ˝

Remark 6.11. The proof above is not rigorous since the first interchange of the order of
integration (from dωdb to dbdω) cannot be valid since the integrand is indeed not integrable.
The second interchange of the order of integration (from dωda to dadω) is true due to the
fact that fp, gp and the admissibility condition.

Proof from another book. Define

1 1
F (ω) = |a|´ 2 F [Wψ [f ](a, ¨)](ω) and G(ω) = |a|´ 2 F [Wψ [g](a, ¨)](ω).

Using (6.2.5),
F (ω) = fp(ω)ψ(aω),
p G(ω) = gp(ω)ψ(aω).
p (6.3.5)
Applying the Plancherel identity and using (6.2.2), we find that
[ ]
ż
@ D
Wψ [f ](a, b)Wψ [g](a, b) db = Wψ [f ](a, ¨), Wψ [g](a, ¨) 2
R L (R)

1@ D |a|
= F [Wψ [f ](a, ¨)], F [Wψ [g](a, ¨) = xF, Gy.
2π 2π
Hence, by substituting (6.3.5) into the above expression, integrating with respect to da/a2
on R, and recalling the definition of Cψ from (6.2.1), we obtain
ż [ż
] ] da
ż [ ]
[ 1
Wψ [f ](a, b)Wψ [g](a, b) db 2 = xF, Gy da
R R a R 2π|a|
ż [ż 2 ] ż ż 2
1 |ψ(aω)|
p 1 |ψ(aω)|
p
= f (ω)p
p g (ω) dx da = fp(ω)p g (ω) dadx
2π R R |a| 2π R R |a|
1
ż ( ż |ψ(y)|
p 2 ) Cψ p
= f (ω)p
p g (ω) dy dx = xf , gpyL2 (R) = Cψ xf, gyL2 (R) .
2π R R |y| 2π
Note that the switch of the order of the integration in the second line is due to the Fubini
Theorem which requires that the integrand is integrable; nevertheless, the integrability of
the integrand can be shown using the Tonelli Theorem as long as the admissibility condition
(6.2.1) is satisfied. This completes the proof of the theorem. ˝

Remark 6.12. There is still one particular problem in the proof above: In order to apply
the Plancherel identity it is required that Wψ [f ](a, ¨) and Wψ [g](a, ¨) P L2 (R) for a ‰ 0.
However, if Wψ [f ](a, ¨) P L2 (R), then F [Wψ [f ](a, ¨)] P L2 (R) but if this might not be true
since fp, gp, ψp P L2 (R) only guarantees that F [Wψ [f ](a, ¨)], F [Wψ [g](a, ¨)] P L1 (R).

Proof. We first prove that (6.3.1) holds for all f, g P L1 (R)XL2 (R). Let f, g P L1 (R)XL2 (R),
and as in the previous proof we define
1 1
F (ω) = |a|´ 2 F [Wψ [f ](a, ¨)](ω) and G(ω) = |a|´ 2 F [Wψ [g](a, ¨)](ω).

Then for a ‰ 0, F, G P L2 (R); thus Wψ [f ](a, ¨), Wψ [g](a, ¨) P L2 (R). Therefore, the previous
proof goes through and we obtain that (6.3.1) holds for f, g P L1 (R) X L2 (R). In particular,
ż ż ˇ
ˇ1
ˇ2
ˇ Wψ [f ](a, b)ˇ dbda = Cψ }f }2L2 (R) (6.3.6)
ˇ
R R |a|

Now let f P L2 (R). Choose tfn u8 1 2

n=1 Ď L (R) X L (R) such that tfn un=1 converges to
8

f in L2 (R). Then tfn u8 2

n=1 is a Cauchy sequence in L (R) and the identity above, together
with the linearity of the wavelet transform, shows that
ż ż ˇ
ˇ1 1 ˇ2
ˇ Wψ [fn ](a, b) ´ Wψ [fm ](a, b)ˇ dbda = Cψ }fn ´ fm }2L2 (R) Ñ 0 as n, m Ñ 8.
ˇ
R R |a| |a|
This shows that the sequence thn u8
n=1 defined by

1 1
hn (a, b) = Wψ [fn ](a, b) = xfn , ψa,b yL2 (R)
|a| |a|
is a Cauchy sequence in L2 (R2 ); thus thn u8 2 2
n=1 converges to some function in L (R ). On the
other hand, the fact that fn Ñ f in L2 (R) shows that thn u8
n=1 converges a.e. to the function
1 2 2
h(a, b) = Wψ [f ](a, b). Therefore, thn u8
n=1 converges to g in L (R ) and this shows that
|a|
ż ż ˇ ż ż ˇ
ˇ1 ˇ1
ˇ2 ˇ2
2 2
ˇ Wψ [f ](a, b)ˇ dbda = }h}L2 (R) = lim }hn }L2 (R) = lim ˇ Wψ [fn ](a, b)ˇ dbda
ˇ ˇ
R R |a| nÑ8 nÑ8 R R |a|
2 2
= Cψ lim }fn }L2 (R) = Cψ }f }L2 (R) .
nÑ8

This establishes that (6.3.6) holds for all f P L2 (R), and we completes the proof of the
theorem using the polarization identity. ˝

Theorem 6.13 (Inversion Formula). If f P L2 (R), then f can be reconstructed by the

formula ż ż
1 dbda
f (t) = Wψ [f ](a, b)ψa,b (t) 2 , (6.3.7)
Cψ R R a
where the equality holds almost everywhere.
Proof from the textbook. For any g P L2 (R), we have, from Theorem 6.10, that
ż ż
dbda
Cψ xf, gyL2 (R) = Wψ [f ](a, b)Wψ [g](a, b) 2
R R a
ż ż (ż ) dbda
= Wψ [f ](a, b) g(t)ψa,b (t) dt
a2
żR (Rż ż R
)
dbda
= Wψ [f ](a, b)ψa,b (t) 2 g(t) dt.
R R R a
Since g is an arbitrary element of L2 (R), the inversion formula (6.3.7) follows. ˝

Remark 6.14. The problem in this proof again is that the interchange of the order of
integration cannot be guaranteed. Nevertheless, it is possible to show the validity of the
inversion formula for f P L1 (R)XL2 (R) with fp P L1 (R). However, even if this case is proved,
we cannot prove the general result by the density argument since the convergence
ż ż ż ż
dbda dbda
lim Wψ [fn ](a, b)ψa,b (t) 2 = Wψ [f ](a, b)ψa,b (t) 2 .
nÑ8 R R a R R a
cannot be guaranteed.

Proof. For fixed positive constants ε, A, B satisfying ε ă A, we define an operator S(ε, A, B)

on L2 (R) by
ż ż
1 dbda
[S(ε, A, B)f ](t) = Wψ [f ](a, b)ψa,b (t) f P L2 (R).
Cψ εă|a|ăA |b|ăB a2

Our goal is to show that S(ε, A, B)f Ñ f in L2 (R) as ε Ñ 0+ and A, B Ñ 8.

To see the L2 convergence, we note that
› › ˇ@ D ˇˇ
›S(ε, A, B)f ´ f › 2 = sup ˇ
S(ε, A, B)f ´ f, g
L (R)
ˇ 2
ˇ
}g}L2 (R) =1 L (R)

For g P L2 (R),
ż ż ż
@ D 1 dbda
S(ε, A, B)f, g 2 = Wψ [f ](a, b)ψa,b (t)g(t) dt. (6.3.8)
L (R) Cψ R εă|a|ăA |b|ăB a2
We first show that the integral above is absolutely convergent. By the Tonelli Theorem,
ż ż ż
ˇWψ [f ](a, b)ˇˇψa,b (t)ˇˇg(t)ˇ dbda dt
ˇ ˇˇ ˇˇ ˇ
R εă|a|ăA |b|ăB a2
ż ż (ż )
ˇ ˇ ˇ ˇˇ ˇ dbda
= ˇWψ [f ](a, b) ˇ ˇψa,b (t) g(t) dt
ˇ ˇ ˇ
εă|a|ăA |b|ăB R a2
[ż ż (ż )12 ]
ˇ ˇ ˇ ˇ2 dbda
= ˇWψ [f ](a, b)ˇ ˇψa,b (t)ˇ dt }g}L2 (R)
εă|a|ăA |b|ăB R a2
(ż ż ˇ1 )12 (ż ż ż )1
ˇ2 ˇ ˇ2 dbda 2
= ˇ Wψ [f ](a, b)ˇ dbda ψa,b (t) dt 2 }g}L2 (R)
ˇ ˇ ˇ ˇ
εă|a|ăA |b|ăB |a| εă|a|ăA |b|ăB R a
(ż ż ż )1
a ˇ ˇ2 dbda 2
= Cψ }f }L2 (R) }g}L2 (R) ˇψa,b (t)ˇ dt
εă|a|ăA |b|ăB R a2
and further computation shows that
ż ż ż ż ż
ˇ ˇ2 dbda 2 1
ˇψa,b (t)ˇ dt = }ψ} 2 (R) dbda ă 8.
L
εă|a|ăA |b|ăB R a2 εă|a|ăA |b|ăB |a|
2

Therefore, the integral on the RHS of (6.3.8) is absolutely convergent, so the Fubini Theorem
implies that
ż ż ż
@ D 1 dbda
S(ε, A, B)f, g 2 = Wψ [f ](a, b)ψa,b (t)g(t) 2 dt
L (R) Cψ R εă|a|ăA |b|ăB a
ż ż ż
1 dbda
= Wψ [f ](a, b) ψa,b (t)g(t)dt 2
Cψ εă|a|ăA |b|ăB R a
ż ż
1 dbda
= Wψ [f ](a, b)Wψ [g](a, b) 2 ;
Cψ εă|a|ăA |b|ăB a

thus (6.3.1) implies that

› › ˇ@ D ˇˇ
›S(ε, A, B)f ´ f › 2 = sup ˇ S(ε, A, B)f ´ f, g 2 ˇ
ˇ
L (R) L (R)
}g}L2 (R) =1
ˇ ż ˇ
ˇ 1 d(a, b) ˇˇ
= sup ˇ Wψ [f ](a, b)Wψ [g](a, b)
}g}L2 (R) =1
ˇ Cψ
t(a,b)|εă|a|ăA,|b|ăBuA a2 ˇ
ż
1 ˇ1 ˇ2
ď ˇ Wψ [f ](a, b)ˇ d(a, b).
a ˇ ˇ
Cψ t(a,b)|εă|a|ăA,|b|ăBuA |a|

1
The fact that the function (a, b) ÞÑ Wψ [f ](a, b) belongs to L2 (R2 ) shows that
|a|
ż ˇ1 ˇ2
lim ˇ Wψ [f ](a, b)ˇ d(a, b) = 0.
ˇ ˇ
εÑ0 ,A,BÑ8 t(a,b)|εă|a|ăA,|b|ăBuA |a|
+

› ›
This shows that +
lim ›S(ε, A, B)f ´ f › 2 = 0 and the proof is complete.
L (R)
˝
εÑ0 ,A,BÑ8

Remark 6.15. The proof above indeed shows that the RHS integral of (6.3.7) is obtained
by ż ż
dbda
lim Wψ [f ](a, b)ψa,b (t) .
+
εÑ0 ,BÑ8 εă|a| |b|ăB a2
On the other hand, for f P L1 (R)XL2 (R) with fp P L1 (R), the RHS integral of (6.3.7) means
the usual Lebesgue integral.

6.4 The Discrete Wavelet Transforms

It has been stated in the last section that the continuous wavelet transform (6.2.4) is a two-
parameter representation of a function. In many applications, especially in signal processing,
data are represented by a finite number of values, so it is important and often useful to
consider discrete versions of the continuous wavelet transform (6.2.4). Our goal in this
section is to answer the fundamental question whether we can reconstruct f from discrete
values of its wavelet transform Wψ [f ]. In particular, we would like to reconstruct f using
the discrete values of Wψ [f ] at a = am m
0 and b = nb0 a0 ; that is,
ż
m m ´m
(Wψ [f ])(a0 , nb0 a0 ) = a0 2
f (t)ψ(a´m
0 t ´ nb0 )dt,
R

where a0 ‰ 0, b0 are some given and fixed constants, and m, n are integers. Define
´m/2
ψm,n (x) = a0 ψ(a´m
0 x ´ nb0 ) = (Dam
0
Tnb0 ψ)(x), (6.4.1)

where we abuse the use of notation here and do not confuse with (6.2.3). Using (6.4.1), we
have
(Wψ [f ])(am m
0 , nb0 a0 ) = xf, ψm,n yL2 (R) .

The discrete wavelet transform represents a function by a countable set of wavelet coeffi-
cients, which correspond to points on a two dimensional grid or lattice of discrete points in
the scale-time domain indexed by m and n.
The answer to the fundamental question is positive if the wavelets form a complete
system in L2 (R). The problem is whether there exists another function g P L2 (R) such that

xf, ψm,n yL2 (R) = xg, ψm,n yL2 (R)

for all m, n P Z implies f = g. In practice, the evaluation/measurement of xf, ψm,n yL2 (R)
might not be very accurate, so the best we can hope is that f and g are “close” if the two
␣ ( ␣ (
sequences xf, ψm,n yL2 (R) m,nPZ and xg, ψm,n yL2 (R) m,nPZ are “close”. This property can be
guaranteed if there exists an A ą 0 independent of f , such that
8
ˇxf, ψm,n y 2 ˇ2
ÿ ˇ ˇ
A}f }2L2 (R) ď L (R)
m,n=´8

since the inequality above implies that

8
ˇxf, ψm,n y 2 ´ xg, ψm,n y 2 ˇ2 .
ÿ ˇ ˇ
A}f ´ g}2L2 (R) ď L (R) L (R)
m,n=´8

␣ (
On the other hand, we also want two sequences xf, ψm,n yL2 (R) m,nPZ to be “close” if f and
g are “close”. This will be guaranteed if there exists a B ą 0 independent of f such that
8
ˇxf, ψm,n y 2 ˇ2 ď B}f }2 2
ÿ ˇ ˇ
L (R) L (R)
m,n=´8

since the inequality above implies that

8
ˇxf, ψm,n y 2 ´ xg, ψm,n y 2 ˇ2 ď B}f ´ g}2 2 .
ÿ ˇ ˇ
L (R) L (R) L (R)
m,n=´8

These two requirements are best studied in terms of the so-called frames.
6.4.1 Frames and frame operators
In the following, when the inner product of a Hilbert space is specified, } ¨ } is used to denote
the induced norm of the inner product.

Definition 6.16 (Frames). A collection of countably many vectors txn u in a Hilbert space
(H, x¨, ¨y) is called a frame if these exist constants A, B ą 0 such that

ˇxx, xn yˇ2 ď B}x}2

ÿˇ ˇ
A}x}2 ď @x P H. (6.4.2)
n

The constants A and B are called frame bounds, and a frame satisfying (6.4.2) is called a
frame with frame bounds A and B. If A = B, then the frame is called tight. The frame is
called exact if no proper subset of txn u is also a frame.

The following example shows that tightness and exactness are not related.

Example 6.17. If ten u is an orthonormal basis of H, then

(i) te1 , e1 , e2 , e2 , e3 , e3 , ¨ ¨ ¨ u is a tight frame with frame bounds A = B = 2, but it is not

exact.
?
(ii) t 2e1 , e2 , e3 , ¨ ¨ ¨ u is an exact frame but not tight since the frame bounds are easily
seen as A = 1 and B = 2.
e2 e2 e3 e3 e3
! )
(iii) e1 , ? , ? , ? , ? , ? , ¨ ¨ ¨ is a tight frame with the frame bound A = B = 1 but
2 2 3 3 3
not an orthonormal basis.
e e
! )
(iv) e1 , 2 , 3 , ¨ ¨ ¨ is a complete orthogonal sequence but is not a frame.
2 3
Theorem 6.18. Let txn u be a collection of countably many vectors in a Hilbert space
(H, x¨, ¨y). Then the following two statements are equivalent.

(a) The operator T x = xx, xn y xn is a bounded linear operator on H with AI ď T ď BI,

ř
n
where I is the identity operator on H.

(b) txn u is a frame with frame bounds A and B.

Proof. Before proceeding, we recall that the relation AI ď T ď BI means

xAIx, xy ď xT x, xy ď xBIx, xy @x P H, (6.4.3)

and note that if the series xx, xn y xn converges (if it is an infinite sum) for some particular
ř
n
x P H, then
Aÿ E
ÿA E ÿ
xx, xn y xn , y = xx, xn y xn , y = xx, xn yxxn , yy. (6.4.4)
n n n
“(a) ñ (b)” Suppose that (a) holds. Since T is defined on H, the series xx, xn y xn converges
ř
n
to T x for all x P H; thus (6.4.4) implies that
ˇxx, xn yˇ2 .
ÿ ÿˇ ˇ
xT x, xy = xx, xn yxxn , xy =
n n

Using (6.4.3), we conclude that txn un is a frame with frame bounds A and B. This
shows that (a) implies (b).

“(b) ñ (a)” We next prove that (b) implies (a). Suppose (b) holds. First we claim that
T x = xx, xn y xn converges for all x P H. To see this, it suffices to show the case
ř
n
that txn u = txn u8
n=1 is an infinite sequence. Recall that in any Hilbert space H the
norm of any element x P H is given by
ˇ ˇ
}x} = sup ˇxx, yyˇ .
}y}=1

N
For a fixed x P H, we consider TN x = xx, xn y xn . For 0 ď M ă N , we have, by the
ř
n=1
Cauchy-Schwarz inequality,
ˇ ˇ2
› ›2 ˇ ˇ2 ˇ ÿ ˇ
›T x ´ T x› = sup ˇxT x ´ T x, yyˇ = sup ˇ xx, x n yxx n , yyˇ
N M N M ˇ ˇ
}y}=1 }y}=1 M +1ďnďN
( ˇ )( ˇ)
ˇxx, xn yˇ2 ˇxxn , yyˇ2
ÿ ˇ ÿ ˇ
ď sup
}y}=1 M +1ďnďN M +1ďnďN
( ˇ)
ˇxx, xn yˇ2 B}y}2
ÿ ˇ
ď sup
}y}=1 M +1ďnďN
( ˇ)
ˇxx, xn yˇ2 Ñ 0
ÿ ˇ
=B as M, N Ñ 8.
M +1ďnďN

Thus, tTN xu8N =1 is a Cauchy sequence in H and hence it is convergent as N Ñ 8.

Therefore, T x ” xx, xn y xn converges for all x P H, and (6.4.4) implies that
ř
n
ÿ
xT x, yy = xx, xn yxxn , yy.
n

Following the preceding argument we obtain that

ˇ ˇ2 (ÿˇ ˇ2 )
2
ˇ ˇ ˇ ÿ ˇ
2
}T x} = sup xT x, yy = sup ˇ xx, xn yxxn , yyˇ ď B
ˇ ˇ ˇ ˇ ˇxx, xn yˇ ď B 2 }x}2
}y}=1 }y}=1 n n

which implies that }T } ď B. This shows the boundedness of T . The relation AI ď

řˇ ˇ2
T ď BI follows from that xT x, xy = ˇxx, xn yˇ and the frame txn u has frame bounds
n
A and B. ˝

Definition 6.19 (Frame Operator). To each frame txn u in a Hilbert space there corresponds
an operator T , called the frame operator, from H into itself defined by
ÿ
Tx = xx, xn y xn @x P H. (6.4.5)
n
 We remark that the frame operator associated with a frame is self-adjoint because of
(6.4.4):
Aÿ E ÿA E ÿ
xT x, yy = xx, xn y xn , y = xx, xn y xn , y = xx, xn yxxn , yy
n n n
ÿ ÿ
= xx, xn yxxn , yy = xy, xn yxxn , xy = xT y, xy = xx, T yy.
n n

Theorem 6.20. Suppose txn u is a frame on a Hilbert space (H, x¨, ¨y) with frame bounds A
and B, and T is the corresponding frame operator. Then,

(a) T is invertible and B ´1 I ď T ´1 ď A´1 I. Furthermore, T ´1 is a positive operator and

hence it is self-adjoint.

(b) tT ´1 xn u is a frame with frame bounds B ´1 and A´1 .

(c) Every x P H can be expressed in the form

ÿ ÿ
x= xx, T ´1 xn y xn = xx, xn y T ´1 xn . (6.4.6)
n n

Proof. (a) Since the frame operator T satisfies the relation

AI ď T ď BI;

it follows that ( A)
I ´ B ´1 T ď I ´ B ´1 AI = 1 ´ I
B
and hence
› ( )
›I ´ B ´1 T › ď 1 ´ A }I} ă 1.
›
B
Thus, B ´1 T is invertible and consequently so is T . In view of the fact that

xT ´1 x, xy = xT ´1 x, T T ´1 xy ě AxT ´1 x, T ´1 xy = A}T ´1 x}2 ‰ 0 whenever x ‰ 0,

we conclude that T ´1 is a positive operator and hence it is self-adjoint. Finally, since

1
T is a bounded positive operator, it is possible to define T 2 (using the spectral de-
1
composition in functional analysis) and T 2 is also positive definite (hence self-adjoint)
and invertible. Therefore, by AI ď T ď BI, the fact that
1 1 1 1
A´1 I ´ T ´1 = A´1 T ´ 2 (T ´ AI)T ´ 2 and T ´1 ´ B ´1 I = B ´1 T ´ 2 (BI ´ T )T ´ 2

shows that
B ´1 I ď T ´1 ď A´1 I.

(b) Since T ´1 is self-adjoint, we have

ÿ (ÿ ) ( )
´1 ´1 ´1
xx, T xn y T xn = T xT x, xn y xn = T ´1 T (T ´1 x) = T ´1 x.
´1
(6.4.7)
n n
 This gives
Aÿ E ÿ
xT ´1 x, xy = xx, T ´1 xn y T ´1 xn , x = xx, T ´1 xn yxT ´1 xn , xy.
n n

Hence,
ˇxx, T ´1 xn yˇ2 .
ÿ ÿˇ ˇ
xT ´1 x, xy = xx, T ´1 xn yxx, T ´1 xn y =
n n

Using the result from (a); that is, B ´1 I ď T ď A´1 I; it turns out that

B ´1 xIx, xy ď xT ´1 x, xy ď A´1 xIx, xy

and hence
B ´1 }x}2 ď xT ´1 x, xy ď A´1 }x}2 .

By Theorem 6.18 this shows that tT ´1 xn u is a frame with frame bounds B ´1 and A´1 .

(c) We replace x by T ´1 x in (6.4.5) to derive

ÿ ÿ
x= xT ´1 x, xn y xn = xx, T ´1 xn y xn .
n n

Similarly, replacing x by T x in (6.4.7) gives

ÿ ÿ
x= xT x, T ´1 xn y T ´1 xn = xx, xn y T ´1 xn .
n n

This completes the proof. ˝

Definition 6.21. Let H be a separable Hilbert space, txn u be a frame in H, and T be the
corresponding frame operator of frame txn u. The frame tT ´1 xn u is called the dual frame
of txn u.

By writing T ´1 xn as x
rn , according to formula (6.4.6), the reconstruction formula for x
has the form
ÿ ÿ
x= xx, x
rn y xn = xx, xn yr
xn .
n n

It is easy to verify that the dual frame of tr

xn u is the original frame txn u.

Theorem 6.22. Suppose txn u is a frame in a separable Hilbert space H with frame bounds
A and B. If there exists a sequence of scalars tcn u such that x = cn xn , then
ř
n
ÿ ÿ ÿ
|cn |2 = |an |2 + |an ´ cn |2 ,
n n n

where an = xx, T ´1 xn y so that x = an x n .

ř
n
Proof. Note that xxn , T ´1 xy = xT ´1 xn , xy = an . Substituting x = an xn in the first term
ř
n
in the inner product xx, T ´1 xy gives
Aÿ E ÿ ÿ
xx, T ´1 xy = an xn , T ´1 x = an xxn , T ´1 xy = |an |2 .
n n n

Similarly, substituting x = cn xn into the first term in the inner product xx, T ´1 xy yields
ř
n
Aÿ E ÿ ÿ
´1 ´1
xx, T xy = cn xn , T x = cn xxn , T ´1 xy = c n an .
n n n

Consequently, ÿ ÿ ÿ
|an |2 = c n an = c n an . (6.4.8)
n n n

Finally, we obtain, by using (6.4.8),

ÿ ÿ ÿ ÿ( ) ÿ
|an |2 + |an ´ cn |2 = |an |2 + |an |2 ´ an cn ´ cn an + |cn |2 = |cn |2 .
n n n n n

This completes the proof. ˝

Theorem 6.23. Let txn u be a frame in a Hilbert space (H, x¨, ¨y). If txn u is exact, then
txn u and its dual frame tT ´1 xn u is biorthonormal; that is,

xxm , T ´1 xn y = δmn @ m, n P Z. (6.4.9)

Proof. Suppose that txn u is a frame with frame bounds A and B. Let I denote the index
set, and m P I be a fixed index. Since txn u is an exact frame,
ˇxx, xn yˇ2 ď B
ÿˇ ˇ
Aď @ x P H with }x} = 1
n

but the fact that txn unPIztmu is not a frame implies that for each k P N, there exists yk P H
with }yk } = 1 such that
ˇxyk , xn yˇ2 .
ÿ ˇ ˇ
1ąk
n‰m
ˇ ˇ2
Adding εˇxyk , xm yˇ on both sides of the inequality above, we find that
ˇ2
ˇxyk , xn yˇ2 + ε ˇxyk , xn yˇ2
ˇ ÿ ˇ ˇ ÿˇ ˇ
1 + εˇxyk , xm yˇ ą (k ´ ε)
n‰m n
ˇxyk , xn yˇ2 + Aε.
ÿ ˇ ˇ
ě (k ´ ε) (6.4.10)
n‰m

Letting ε = 1/A in (6.4.10) and applying the Cauchy-Schwartz inequality we obtain that

ˇxyk , xn yˇ2 ă ˇxyk , xm yˇ2 ď }xm }2 ;

ÿ ˇ ˇ ˇ ˇ
(Ak ´ 1)
n‰m

ˇxyk , xn yˇ2 = 0. On the other hand, (6.4.6) implies that

ř ˇ ˇ
thus lim
kÑ8 n‰m
ÿ
xyk , xm y = xyk , xn yxT ´1 xn , xm y
n
so that ( ) ÿ
1 ´ xT ´1 xm , xm y xyk , xm y = xyk , xn yxT ´1 xn , xm y.
n‰m

By Theorem 6.20 and the Cauchy-Schwartz inequality we have

ˇ ( ÿ ˇ ˇ ) 1 ( ÿ ˇ ´1 ˇ )1
ˇxyk , xn yˇ2 2 ˇxT xn , xm yˇ2 2
ˇ ˇˇ
ˇ1 ´ xT ´1 xm , xm yˇˇxyk , xm yˇ ď
n‰m n‰m
1 ( ÿ ˇˇ ˇ2 ) 21
ď? xyk , xn yˇ }xm } Ñ 0 as k Ñ 8.
A n‰m
ˇ ˇ2 1
Letting ε = k in (6.4.10) shows that ˇxyk , xm yˇ ě A ´ , so the inequality above shows that
k
´1
xT xm , xm y = 1. (6.4.11)

The rest of (6.4.9); that is, xxm , T ´1 xn y = 0 for n ‰ m, follows from the identity
ˇ2 ÿ ˇ
ˇxxm , T ´1 xn yˇ2
ÿ ˇ ˇ
xxm , T ´1 xm y = xxm , T ´1 xn yxxn , T ´1 xm y = ˇxxm , T ´1 xm yˇ +
n n‰m

and the identity (6.4.11). ˝

6.4.2 A sufficient condition for a function generating a frame

As pointed out above, we want the family of functions tψm,n um,nPZ to form a frame in L2 (R).
Obviously, the double indexing of the functions is irrelevant. The following theorem gives
fairly general sufficient conditions for a sequence tψm,n um,nPZ to constitute a frame in L2 (R).

Theorem 6.24. Let ψ P L2 (R), and a0 ą 1. If

8
p m ω)ˇ2 ď B for all 1 ď ω ď a0 , and
ÿ ˇ ˇ
(i) there exist A, B ą 0 such that A ď ˇψ(a 0
m=´8
8
ÿ ˇ ˇˇ ˇ
(ii) sup p m )ˇˇψ(a
ˇψ(a p m ω + x)ˇ ď C(1 + |x|)´(1+δ) for some constants C and δ ą 0,
0 0
ωPR m=´8

then there exists rb ą 0 such that the family tψm,n um,nPZ given by
´m/2
ψm,n = a0 ψ(a´m
0 x ´ nb0 ) = (Dam
0
Tnb0 ψ)(x)

forms a frame in L2 (R) for any b0 P (0, rb).

Proof. Before proceeding, we remark that (i) is equivalent to that there exist A, B ą 0 such
that
8
p m ω)ˇ2 ď B
ÿ
@ ω P Rzt0u
ˇ ˇ
Aď ˇψ(a 0 (6.4.12)
m=´8

due to the fact that a0 ą 1. Suppose f P L2 (R). By the Plancherel identity (1.1.15),
8 8
ˇxf, ψm,n y 2 ˇ2 = ˇxf, Dam Tnb0 ψy 2 ˇ2
ÿ ˇ ˇ ÿ ˇ ˇ
L (R) 0 L (R)
m,n=´8 m,n=´8
8 8 ˇż ˇ2
ÿ 1 ˇˇ p ˇ2 1 ÿ ˇ m/2 ib a m nω ˇ
p m ω)dω ˇ .
= ˇxf , Da´m M´nb0 ψyL2 (R) ˇ = ˇ fp(ω)a0 e 0 0 ψ(a
p ˇ
0
m,n=´8
2π 0 2π m,n=´8
ˇ
R
ˇ
 ż
Since, for any s ą 0, the integral g(t)dt can be written as
R
8 żs
ÿ
g(t + ℓs)dt
ℓ=´8 0

2π
provided that g P L1 (R), by taking s = , we obtain
b0 a m
0

8
ˇxf, ψm,n y 2 ˇ2
ÿ ˇ ˇ
L (R)
m,n=´8
8 ˇ 8 żs ˇ2
1 ÿ ˇ ÿ ˇ
= am ˇ
0 ˇ fp(ω + ℓs)e 2πinω/s p m (ω
ψ(a 0 + ℓs))dω ˇˇ
2π m,n=´8 ℓ=´8 0
8 ˇż s ( ÿ 8 ) ˇ2
1 ÿ
mˇ
ˇ 2πinω/s m
ˇ
= a0 ˇ e f (ω + ℓs)ψ(a0 (ω + ℓs)) dω ˇˇ
p p
2π m,n=´8 0 ℓ=´8
8 8 ˇ żs ( ÿ8 ) ˇ2
1 ÿ
m 2
ÿ ˇ 1 2πinω/s m
ˇ
ˇ , (6.4.13)
= a0 s ˇ e fp(ω + ℓs) ψ(a
p
0 (ω + ℓs)) dω
2π m=´8 n=´8
ˇs
0 ℓ=´8
ˇ

where we have used (a version of) Theorem 1.26 to conclude that the series
8
ÿ
Fm (ω) ” p m (ω + ℓs))
fp(ω + ℓs)ψ(a 0
ℓ=´8

converges in L1 (0, s) so we can switch the order of infinite sum and the integration. Next
we show that Fm P L2 (0, s) by showing that the series
8
ÿ ˇ ˇˇ ˇ
Gm (ω) ” p m (ω + ℓs))ˇ
ˇfp(ω + ℓs)ˇˇψ(a 0
ℓ=´8

converges in L2 (0, s). To see this, we apply the Monotone Convergence Theorem and find
żs żs( ÿ
8 )
ˇGm (ω)ˇ2 dω =
ˇ ˇ ˇ ˇˇ m
ˇ
ˇfp(ω + ℓs)ˇˇψ(a
p
0 (ω + ℓs))
ˇ Gm (ω)dω
0 0 ℓ=´8
8 żs
ÿ ˇ ˇˇ ˇ
= p m (ω + ℓs))ˇGm (ω)dω
ˇfp(ω + ℓs)ˇˇψ(a 0
ℓ=´8 0
ÿ8 ż (ℓ+1)s
ˇ ˇˇ ˇ
= p m ω)ˇGm (ω ´ ℓs)dω
ˇfp(ω)ˇˇψ(a 0
ℓs
żℓ=´8
ˇ ˇˇ ˇ
p m ω)ˇGm (ω)dω ,
= ˇfp(ω)ˇˇψ(a 0
R

where the s-periodicity of Gm is used to conclude the last equality. Summing over m P Z,
we obtain that
8 żs 8 8 ż
ÿ ˇ ˇ2 ÿ ÿ ˇ ˇˇ ˇˇ ˇˇ ˇ
ˇGm (ω)ˇ dω = ˇfp(ω)ˇˇψ(a p m (ω + ℓs))ˇ dω .
p m ω)ˇˇfp(ω + ℓs)ˇˇψ(a
0 0
m=´8 0 ℓ=´8 m=´8 R
By applying Cauchy-Schwarz inequality twice,
8
ÿ ż
ˇ ˇˇ ˇˇ ˇˇ ˇ
p m ω)ˇˇfp(ω + ℓs)ˇˇψ(a
ˇfp(ω)ˇˇψ(a p m (ω + ℓs))ˇ dω
0 0
m=´8 R
8 (ż )12
ÿ ˇ ˇ2 ˇ m
ˇˇ m
ˇ
ď ˇf (ω) ψ(a0 ω) ψ(a0 (ω + ℓs)) dω ˆ
p ˇ ˇ p ˇˇ p ˇ
m=´8 R
(ż )12
ˇfp(ω + ℓs)ˇ2 ˇψ(a
ˇ ˇˇ ˇˇ ˇ
ˆ p m ω)ˇˇψ(a
p m (ω + ℓs))ˇ dω
0 0
R
( ÿ
8 ż )12
ˇ ˇ2 ˇ m
ˇˇ m
ˇ
ď ˇf (ω) ψ(a0 ω) ψ(a0 (ω + ℓs)) dω ˆ
p ˇ ˇ p ˇˇ p ˇ
m=´8 R
( ÿ
8 ż )12
ˇ ˇ2 ˇ m
ˇˇ m
ˇ
ˆ ˇfp(ω + ℓs)ˇ ˇψ(a
p
0 ω) ψ(a0 (ω + ℓs)) dω
ˇˇ p ˇ ,
m=´8 R

and the Monotone Convergence Theorem again allow us to switch the order of infinite sum
and the integration so that
8
ÿ ż
ˇ ˇˇ ˇˇ ˇˇ ˇ
p m ω)ˇˇfp(ω + ℓs)ˇˇψ(a
ˇfp(ω)ˇˇψ(a p m (ω + ℓs))ˇ dω
0 0
m=´8 R
(ż 8 )2 1

ˇfp(ω)ˇ2
ˇ ˇ ÿ ˇ ˇˇ ˇ
ď p m ω)ˇˇψ(a
ˇψ(a p m (ω + ℓs))ˇ dω ˆ
0 0
R m=´8
(ż 8 )2 1

ˇfp(ω)ˇ2
ˇ ˇ ÿ ˇ ˇˇ ˇ
ˆ p m ω)ˇˇψ(a
ˇψ(a p m (ω ´ ℓs))ˇ dω .
0 0
R m=´8
8 ˇ ˇˇ ˇ
Define β(ξ) = sup
ř p m ω)ˇˇψ(a
ˇψ(a p m ω + ξ)ˇ. The inequality above leads to that
0 0
ωPR m=´8
8
ÿ ż
ˇ ˇˇ ˇˇ ˇˇ ˇ
p m ω)ˇˇfp(ω + ℓs)ˇˇψ(a
ˇfp(ω)ˇˇψ(a p m (ω + ℓs))ˇ dω
0 0
m=´8 R
(ż )12 ( ż )12
ˇfp(ω)ˇ2 β(am ˇfp(ω)ˇ2 β(´am
ˇ ˇ ˇ ˇ ˇ
ď 0 ℓs)dω 0 ℓs) dω
ˇ
R R
[ [ ( 2πℓ ) ( 2πℓ )] 12
] 12
ď }f }2L2 (R) β(am = m
β
0 ℓs)β(´a0 ℓs) β ´ . }f }2L2 (R) (6.4.14)
b0 b0
8 [ ( 2πℓ ) ( 2πℓ )] 21
Since condition (ii) implies that ă 8, we find that
ř
β β ´
ℓ=´8 b0 b0
8 [ (
2πℓ ) ( 2πℓ )] 12
8 żs
ˇGm (ω)ˇ2 dω ď }f }2 2
ÿ ˇ ˇ ÿ
L (R) β β ´ ă 8.
m=´8 0 ℓ=´8
b 0 b 0

Therefore, Gm P L2 (0, s) for all m P Z; thus Fm P L2 (0, s) for all m P Z as well.

We next repeat the steps above (but use the Domninated Convergence Theorem instead
of the Monotone Convergence Theorem) to show that tψm,n um,nPZ is a frame. By the Parseval
identity,
8 ˇ żs ( ÿ8 ) ˇ2 żs
ˇ1 1 ˇFm (ω)ˇ2 dω
ÿ ˇ ˇ ˇ
2πinω/s m
ˇ
ˇs e fp(ω + ℓs)ψ(a
p
0 (ω + ℓs)) dω ˇ =
ˇ
n=´8 0 ℓ=´8
s 0
so (6.4.13) and the Dominated Convergence Theorem implies that
8 8 żs 8 żs
ˇxf, ψm,n y 2 ˇ2 = 1 1 ÿ
ˇ2
ˇFm (ω)ˇ2 dω
ÿ ˇ ˇ ÿ ˇ ˇ ˇ
m
a 0 s ˇ F m (ω) ˇ dω =
m,n=´8
L (R)
2π m=´8 0 b0 m=´8 0
8 ż s
1 ÿ p m (ω + ℓs))Fm (ω)dω
= fp(ω + ℓs)ψ(a 0
b0 m,ℓ=´8 0
8 ż (ℓ+1)s
1 ÿ p m ω)Fm (ω ´ ℓs)dω .
= fp(ω)ψ(a 0
b0 m,ℓ=´8 ℓs

Since Fm is s-periodic, the identity above implies that

8
ˇxf, ψm,n y 2 ˇ2
ÿ ˇ ˇ
L (R)
m,n=´8
8 ż (ℓ+1)s 8 ż
1 ÿ m 1 ÿ p m ω)Fm (ω)dω
= f (ω)ψ(a0 ω)Fm (ω)dω =
p p fp(ω)ψ(a 0
b0 m,ℓ=´8 ℓs b0 m=´8 R
8 ż
1 ÿ p m ω)fp(ω + ℓs)ψ(ap m (ω + ℓs))dω
= fp(ω)ψ(a 0 0
b0 m,ℓ=´8 R
8 ż ÿ ż
1 ÿ ˇ ˇ2 ˇ m
ˇ2 1 p m ω)fp(ω + ℓs)ψ(a
p m (ω + ℓs))dω
= ˇfp(ω)ˇ ˇψ(ap
0 ω) dω +
ˇ fp(ω)ψ(a 0 0
b0 m=´8 R b0 m,ℓPZ R
ℓ‰0
ż 8 ÿ ż
1 ˇf (ω)ˇ2
ˇ ˇ ÿ ˇ
p m ω)ˇ2 dω + 1
ˇ
p m ω)fp(ω + ℓs)ψ(a
p m (ω + ℓs))dω .
= ˇψ(a 0 fp(ω)ψ(a 0 0
b0 R m=´8
b 0 m,ℓPZ R
ℓ‰0

Therefore, using (6.4.14) we obtain that

ˇ 8 ż 8 ˇ
ˇ ÿ ˇ
ˇxf, ψm,n y 2 ˇ ´
ˇ2 1 ˇ ˇ2 ÿ
ˇf (ω)ˇ
ˇ
ˇψ(a m
ˇ2 ˇ
0 ω) dω ˇ
ˇ p ˇ ˇ
ˇ
m,n=´8
L (R)
b 0 R m=´8
ż
1 ÿ ˇ ˇˇ
ˇfp(ω)ˇˇψ(a
ˇˇ ˇˇ
p m ω)ˇˇfp(ω + ℓs)ˇˇψ(a
ˇ
p m (ω + ℓs))ˇ dω
ď 0 0
b0 m,ℓPZ R
ℓ‰0

1 ÿ [ ( 2πℓ ) ( 2πℓ )] 12 2 ÿ8 [ (
2πℓ ) ( 2πℓ )] 12
2 2
ď }f }L2 (R) β β ´ = }f }L2 (R) β β ´ .
b0 ℓPZ
b0 b0 b0 ℓ=1
b0 b0
ℓ‰0

Define
( 8 [ ( )
2πℓ ) ( 2πℓ )] 12
8
1 ÿ ˇ m
ˇ2 ÿ
A(b0 ) = inf ˇψ(a0 ω) ´ 2
p ˇ β β ´ ,
b0 ωPRzt0u
m=´8 ℓ=1
b0 b0
( 8 8 [ ( ) ( 2πℓ )] 12 )
1 ÿ ˇ m
ˇ2 ÿ 2πℓ
B(b0 ) = sup ˇψ(a
p
0 ω) + 2
ˇ β β ´ .
b0 ωPRzt0u m=´8 ℓ=1
b0 b0

Then
8
ˇxf, ψm,n y 2 ˇ2 ď B(b0 )}f }2 2 ,
ÿ ˇ ˇ
A(b0 )}f }2L2 (R) ď L (R) L (R)
m,n=´8
so using (6.4.12) the theorem is concluded provided we show that
8 [ (
ÿ 2πℓ ) ( 2πℓ )] 21
lim β β ´ = 0. (6.4.15)
b0 Ñ0
ℓ=1
b 0 b 0

Nevertheless, condition (ii) implies that β(ξ) ď C(1 + |ξ|)´1´δ ; thus

8 [ (
2πℓ ) ( 2πℓ )] 21
8 (
2πℓ )´1´δ
ż8(
ÿ ÿ 2πx )´1´δ Cb0
β β ´ ďC 1+ ďC 1+ dx = ;
ℓ=1
b0 b0 ℓ=1
b0 0 b0 2πδ

thus (6.4.15) is established. This completes the proof. ˝

Remark 6.25. Having established Theorem 6.24, we choose b0 in the form a´N 0 for some
´N /2
N P N such that b0 P (0, b). Let ϕ = ψN,0 or ϕ(x) = a0
r ´N
ψ(a0 x). Then for this choice of
b0 , we have
´m ( ´m ´N
) ´ m´N ´N ( ´N ´m+N )
(Dam
0
T nb 0 ψ)(x) = a 0
2
ψ a 0 x ´ na 0 = a 0
2
a 0 ψ a0 (a0
2
x ´ n)
´ m´N
= a0 2
ϕ(a´m+N
0 x ´ n) = ϕm´N,n (x),

´m/2 ␣ (
where ϕm,n (x) ” a0 ϕ(a´m
0 x ´ n). By the fact that Dam
0
Tnb0 ψ m,nPZ is a frame, there
exist A, B ą 0 such that

ˇxf, ϕm´N,n y 2 ˇ2 ď B}f }2 2 .

ÿ ˇ ˇ
A}f }2L2 (R) ď L (R) L (R)
m,nPZ

ř ˇˇ ˇ2 ř ˇˇ ˇ2
Since xf, ϕm´N,n yL2 (R) ˇ = xf, ϕm,n yL2 (R) ˇ , we conclude that tϕm,n um,nPZ is also a
m,nPZ m,nPZ
frame.

6.4.3 Riesz basis

A frame might have redundant vectors which makes the expression of a vector as the linear
combination of vectors in a frame not unique. What we really want is a basis. The last part
of this sub-section contributes to the Riesz basis.

Definition 6.26. Let H be a Hilbert space. A collection of vectors txn unPI is called a Riesz
basis of H if there exist a bounded linear bijection T : H Ñ H and an orthonormal basis
ten unPI of H such that T en = xn for every n P I.

Theorem 6.27. Let txn u be a collection of countably many vectors in a separable Hilbert
space (H, x¨, ¨y). The following two statements are equivalent:

1. txn u is a Riesz basis of H.

2. The linear span of txn u is dense in H, and there exist A, B ą 0 such that
ÿ ›ÿ ›2 ÿ ÿ
› ›
A |cn | ď ›› cn xn ›› ď B |cn |2 whenever
2
|cn |2 ă 8. (6.4.16)
n n n n
Proof. W.L.O.G. we assume that H is infinite dimensional, and we write txn u as txn u8
n=1 .

“ñ” Since txn u8

n=1 is a Riesz basis of H, there exist a bounded invertible linear map T

n=1 of H satisfying T en = xn for all n P N. By the

and an orthonormal basis ten u8
boundedness and invertibility of T , there exist A, B ą 0 such that

m}x} ď }T x} ď M }x} @x P H.
8
Note that the lower bound is due to the open mapping theorem. Since x = xx, en y en ,
ř
n=1
the Parseval identity implies that
8 › ( 8 )››2 8
ÿ › ÿ ÿ
2 2
m |cn | ď ›T
› cn en ›› ď M 2 |cn |2 @ tcn u8 2
n=1 P ℓ .
n=1 n=1 n=1
8
By the boundedness of T and the convergence of cn en in H for any given tcn u8
ř
n=1 P
n=1
ℓ2 , we conclude that
› ( 8 )›› › ( n )›› › n › › 8 ›
› ÿ › ÿ ›ÿ › ›ÿ ›
›T cn en ›› = lim ››T ck ek ›› = lim ›› ck xk ›› = ›› cn xn ›› .
› nÑ8 nÑ8
n=1 k=1 k=1 n=1

Therefore, (6.4.16) holds for A = m2 and B = M 2 .

For the denseness of the linear span of txn u, we note that for each x P H, T ´1 x P H
can be expressed as 8
ÿ @ ´1 D
´1
T x= T x, en en ;
n=1

thus the boundedness of T shows that

(ÿ
8
@ ´1 D ) ÿ 8
@ ´1 D 8
ÿ @ ´1 D
x=T T x, en en = T x, en T en = T x, en xn .
n=1 n=1 n=1

This shows that the linear span of txn u is dense in H.

“ð” Suppose that there exist A, B ą 0 such that (6.4.16) holds. Let ten u8
n=1 be an
orthonormal basis of H, and define T on (a subset of) H by
8
ÿ
Tx = xx, en y xn whenever the RHS makes sense.
n=1
8
Using (6.4.16), we find that xx, en y xn converges for all x P H so that T : H Ñ H
ř
n=1
is well-defined and by (6.4.16) again we have
8 › 8 ›2 8
ˇ2 › ÿ
ˇxx, en yˇ2
ÿ ˇ › ÿ ˇ ˇ
A ˇxx, en y ď ›
ˇ › xx, en y xn › ď B
› @x P H.
n=1 n=1 n=1

By the Parseval identity, the inequality above is equivalent to that

A}x}2 ď }T x}2 ď B}x}2 @x P H. (6.4.17)

 Therefore, T is bounded and injective. Next we show that T is surjective by showing
that the image of T is dense in H since the image of T must be closed due to (6.4.17).
Suppose that the image of T is not dense in H. Then there exists non-zero z P H
such that xz, T xy = 0 for all x P H. In particular,

xz, xn y = xz, T en y = 0 @ n P N.

However, since the linear span of txn u is dense in H, the statement above implies that
z = 0 which is a contradiction. Therefore, we establish the existence of a bounded
surjective linear map T : H Ñ H with the property that T en = xn for all n P N; thus
n=1 is a Riesz basis of H.
txn u8 ˝

Theorem 6.28. Let txn u be a collection of countably many vectors in a Hilbert space
(H, x¨, ¨y). Then txn u is an exact frame in H if and only if txn u be a Riesz basis of H.

Proof. Let txn u be a collection of countably many vectors.

“ñ” Suppose that txn u is an exact frame with frame bounds A and B, and T is the
corresponding frame operator. Note that by Theorem 6.23 txn u and tT ´1 xn u are
bi-orthonormal; that is, xxm , T ´1 xn y = δmn .
We first show that for every x P H there exists a unique tcn u such that x = cn xn .
ř
n
The existence of tcn u is clear since txn u is a frame so that Theorem 6.20 shows that
ÿ
x= xx, T ´1 xn y xn @x P H.
n

For the uniqueness, we have to show that cm has to agree with xx, T ´1 xm y for all
m P N. To see this, we note that
Aÿ E ÿ ÿ
xx, T ´1 xm y = cn xn , T ´1 xm = cn xxn , T ´1 xm y = cn δnm = cm .
n n n

Therefore, txn u is a basis.

Next we show that
›2
A2 ÿ B2 ÿ
›
2
›ÿ ›
|cn | ď › cn xn ›› ď
› |cn |2 @ tcn u P ℓ2 . (6.4.18)
B n n
A n

Nevertheless, the fact that txn u is a frame with frame bounds A and B as well as that
T is invertible implies that

ˇxT ´1 x, xn yˇ2 ď B}T ´1 x}2

ÿˇ ˇ
A}T ´1 x}2 ď @x P H. (6.4.19)
n

Since B ´1 I ď T ´1 ď A´1 I and T ´1 is self-adjoint,

1 1 1 1 1 1

2
}x}2 ď xx, T ´1 xy ď }T ´ 2 x}2 ď xT ´ 2 x, T ´1 T ´ 2 xy = }T ´1 x}2
B B B
 1
and similarly }T ´1 x}2 ď }x}2 . Therefore, (6.4.19) shows that
A2
A 2
ÿˇ
ˇxx, T ´1 xn yˇ2 ď B }x}2
ˇ
2
}x} ď @x P H,
B n
A2
or equivalently,
A2 ÿ ˇˇ ˇ2 B 2 ÿ ˇˇ ˇ2
xx, T ´1 xn yˇ ď }x}2 ď xx, T ´1 xn yˇ @x P H.
B n A n

Inequality (6.4.18) then follows from the fact that x = cn xn if and only if cn =
ř
n
xx, T ´1 xn y for all n P N.

“ð” Suppose that txn u is a Riesz basis of H. Then, there exists an orthonormal basis ten u
and a bounded linear bijection T : H Ñ H such that T en = xn for all n. For x P H,
we have
ˇxx, xn yˇ2 = ˇxx, T en yˇ2 = ˇxT ˚ x, en yˇ2 = }T ˚ x}2 ,
ÿˇ ˇ ÿˇ ˇ ÿˇ ˇ
n n n
where T ˚ is the adjoint of T . On the other hand, the fact that T is a bounded linear
bijection shows that (T ˚ )´1 exists and is bounded. Moreover, we have
› ˚ ´1 ›´1
›(T ) › }x} ď }T ˚ x} ď }T ˚ }}x};
› ›´2
thus the collection txn u is a frame (with frame bounds ›(T ˚ )´1 › and }T ˚ }2 ). The
collection txn u is obviously an exact frame because it ceases to be a basis whenever
any element is deleted from the collection.
This completes the proof. ˝

6.5 Orthonormal Wavelets

Since the discovery of wavelets, orthonormal wavelets with good time-frequency localization
are found to play an important role in wavelet theory and have a great variety of applications.
In general, the theory of wavelets begins with a single function ψ P L2 (R), and a family of
functions tψm,n um,nPZ is generated from this single function ψ by the operation of binary
dilations (that is, dilation by 2m ) and dyadic translation of n2´m so that
( ( n ))
ψm,n (x) = 2m/2 ψ 2m x ´ m = 2m/2 ψ(2m x ´ n), (6.5.1)
2
where the factor 2m/2 is introduced to ensure orthonormality so that }ψm,n }L2 (R) = }ψ}L2 (R)
for all m, n P Z.
A situation of interest in applications is to deal with an orthonormal family tψm,n um,nPZ ;
that is, ż
xψm,n , ψk,ℓ yL2 (R) = ψm,n (t)ψk,ℓ (t)dt = δmk δnℓ , (6.5.2)
R
where m, n, k, ℓ P Z.
To show how the inner products behave in this formalism, we prove the following lemma.
Lemma 6.29. If ψ and ϕ P L2 (R), then
xψm,k , ϕm,ℓ yL2 (R) = xψn,k , ϕn,ℓ yL2 (R) (6.5.3)
for all m, n, k, ℓ P Z.
Proof. By the substitution of variable x = 2n´m t, we have
ż ż
xψm,k , ψm,ℓ yL2 (R) = 2 ψ(2 x ´ k)ψ(2 x ´ ℓ)dx = 2n ψ(2n t ´ k)ψ(2n t ´ ℓ)dt
m m m
R R

= xψn,k , ψn,ℓ yL2 (R) . ˝

Definition 6.30 (Orthonormal Wavelet). A wavelet ψ P L2 (R) is called orthonormal if the

family of functions tψm,n um,nPZ defined by (6.5.1) is an orthonormal basis in L2 (R).

As in the classical Fourier series, the wavelet series for a function f P L2 (R) based on a
given orthonormal wavelet ψ is given by
8
ÿ
f (x) = cm,n ψm,n (x), (6.5.4)
m,n=´8

where the wavelet coefficients cm,n are given by

cm,n = xf, ψm,n yL2 (R) (6.5.5)

and the double wavelet series (6.5.4) converges to the function f in the L2 -norm.

Example 6.31 (Discrete Haar Wavelet). The simplest example of an orthonormal wavelet
is the classic Haar wavelet (6.2.6). To prove this fact, we first show that tψm,n um,nPZ is an
orthonormal set. With ψ defined by (6.2.6) and ψm,n defined by (6.5.1), we have
ż
xψm,n , ψk,ℓ yL2 (R) = 2m/2 ψ(2m x ´ n) ¨ 2k/2 ψ(2k x ´ ℓ)dx
R ż
(2m x´n=t) k/2 ´m/2
= 2 2 ψ(t)ψ(2k´m (t + n) ´ ℓ)dt (6.5.6)
R

1. For m = k, this result gives

ż
xψm,n , ψk,ℓ yL2 (R) = ψ(t)ψ(t + n ´ ℓ)dt = δ0(n´ℓ) = δnℓ ,
R

where ψ(t) ‰ 0 in 0 ď t ď 1 and ψ(t ´ (ℓ ´ n)) ‰ 0 in ℓ ´ n ď t ă 1 + ℓ ´ n, and these

intervals are disjoint from each other unless n = ℓ.

2. We now consider the case m ‰ k. In view of symmetry, it suffices to consider the case
m ă k. Putting r = k ´ m ą 0 in (6.5.6), we can complete the proof by showing that,
for k ‰ m, ż
r/2
xψm,n , ψk,ℓ yL2 (R) = 2 ψ(t)ψ(2r t + s)dt,
R

where s = 2 n ´ ℓ P Z. In view of the definition of the Haar wavelet ψ, we must prove

r

that ż 1 ż1
2
r
ψ(2 t + s)dt ´ ψ(2r t + s)dt = 0.
1
0 2
 Invoking a simple change of variables 2r t + s = x, with a = s + 2r´1 and b = s + 2r
we find that
ż1 ż1 ża żb
2
r r
ψ(2 t + s)dt ´ ψ(2 t + s)dt = ψ(x)dx ´ ψ(x)dx = 0,
1
0 2
s a

where we have used the fact that |a ´ s| = |b ´ a| = 2r´1 ě 1 and the integral of ψ on
an interval with length not less than 1 is zero to conclude the last equality.

This completes the proof that the Haar wavelet ψ is an orthonormal set.
Next we show that tψm,n um,nPZ is indeed a basis in L2 (R). Using (6.5.1) the discrete
Haar wavelet is defined by
$
n n + 1/2
’ 2m/2 if m
ďtă ,
2m
’
’
& 2
ψm,n (t) = 2m/2 ψ(2m t ´ n) = ´2m/2 if
n + 1/2 n+1
ďtď m ,
’
’
’ 2m 2
0 otherwise.
%

Since tψm,n um,nPZ is an orthonormal set, any function f P L2 (R) can be expanded in the
wavelet series in the form
8
ÿ
f= xf, ψm,n yL2 (R) ψm,n , (6.5.7)
m,n=´8

as long as we can show that

8
ˇxf, ψm,n y 2 ˇ2
ÿ ˇ ˇ
}f }2L2 (R) = L (R)
@ f P L2 (R). (6.5.8)
m,n=´8

To prove this, it suffices to show that (6.5.7)/(6.5.8) holds for the function
1 if 0 ď t ă 1,
"
f (t) = 1[0,1) (t) =
0 otherwise.

since this will show that (6.5.7)/(6.5.8) also holds for a collection of characteristic functions
␣ (
1[ nm , n+1
m ) m,nPZ
whose linear span is dense in L2 (R). Evidently,
2 2

m
xf, ψm,n yL2 (R) = 0 for m ě 0 or n ‰ 0 and xf, ψm,0 yL2 (R) = 2 2 if m ă 0.

Consequently,
8 ´1 8
ˇxf, ψm,n y 2 ˇ2 = ˇxf, ψm,0 y 2 ˇ2 =
ÿ ˇ ˇ ÿ ˇ ˇ ÿ
L (R) L (R)
2´m = 1 = }f }2L2 (R) .
m,n=´8 m=´8 m=1

This verifies (6.5.8).

Example 6.32 (The Discrete Shannon Wavelet). The Shannon function ψ whose Fourier
transform satisfies
ψ(ω)
p = 1I (ω), (6.5.9)
where I = [´2π, ´π) Y [π, 2π) is called the Shannon wavelet. This wavelet ψ can directly
be obtained from the inverse Fourier transform of ψp so that
ż [ ż ´π ż 2π ]
1 iωt p 1 iω iω
ψ(t) = e ψ(ω)dω = e dω + e dω
2π R 2π ´2π π
1( ) sin(πt/2) 3πt
= sin 2πt ´ sin πt = cos (6.5.10)
πt πt/2 2

Both ψ and ψp are shown in Figure 6.6.

ψ(t) ψ(ω)
p

‚ ‚ ‚ ‚ ‚ ‚ t ω
´3 ´2 ´1 1 2 3 ´2π ´π π 2π

Figure 6.6: The Shannon wavelet and its Fourier transform

We define ψm,n (t) = 2´m/2 ψ(2´m t ´ n) = (D2m Tn ψ)(t) whose Fourier transform is given
by
m
#
2 2 exp(´iωn2m ) if 2m ω P I,
ψ
z m,n (ω) = (D2´m M´n ψ)(ω) =
p (6.5.11)
0 otherwise.

Evidently, supp(ψ zm,n ) X supp(ψk,ℓ ) = H for m ‰ k. Hence, by the Plancherel identity

y
(1.1.15), it turns out that, for m ‰ k,
1 z y
xψm,n , ψk,ℓ yL2 (R) = xψm,n , ψk,ℓ yL2 (R) = 0. (6.5.12)
2π
For m = k, again by the Plancherel identity (1.1.15) we have
ż
1 z y 1
xψm,n , ψk,ℓ yL2 (R) = xψm,n , ψm,ℓ yL2 (R) = ψ
z m,n (ω)ψm,ℓ (ω)dω
y
2π 2π R
ż
1
= (D2m M´n ψ)(ω)(D
p 2m M´ℓ ψ)(ω)dω
p
2π R
( )ˇ
ż
1 ´m p ´m ω)ˇ2 dω
ˇ
= ¨2 exp ´ iω2´m (n ´ ℓ) ˇψ(2
2π R
( )ˇ
ż
1 p ˇ2 dσ
ˇ
= exp ´ i(n ´ ℓ)σ ˇψ(σ)
2π R
1 ( 2π
żπ )
( )
ż
= ´ exp ´ i(n ´ ℓ)σ dω = δnℓ . (6.5.13)
2π ´2π ´π

This shows that tψm,n um,nPZ is an orthonormal set in L2 (R).

␣ ˇ (
Next we show that tψm,n um,nPZ is an orthonormal basis in L2 (R). Let Im = ω ˇ 2m ω P I .
Note that for a fixed m P Z, ψ
␣ ( 2
m,n nPZ is an orthogonal basis of L (Im ). To see this, we first
z
␣ (
note that L2 (´2´m π, 2´m π) has a Fourier basis d2´m en nPZ , where en (x) = einx as defined in
Section 1.1. Since every f P L2 (Im ) corresponds to a unique function g P L2 (´2´m π, 2´m π)
satisfying #
f (x + 2´m π) if 0 ď x ă 2´m π,
g(x) =
f (x ´ 2´m π) if ´2´m π ď x ă 0,
8 @ D
and g = g, d2´m en 2 ´m m d2´m en in L2 (´2´m π, 2´m π), the fact that
ř
n=´8 L (´2 π,2 π)

d2´m en (x ˘ 2´m π) = (´1)n d2´m en (x)

8 @ D
shows that f = f, d2´m en 2 d2´m en in L2 (Im ), where
ř
n=´8 L (Im )

2m´1
ż ż
@ D 1 mx
f, d2´m en 2 = f (x)d2´m en (x)dx = f (x)e´in2 dx.
L (Im ) |Im | Im π Im

m,´n for all m, n P Z, we conclude that ψm,n nPZ is an orthogonal

␣ (
Since d2´m en = 2´m/2 ψ
{ z
basis of L2 (Im ) and we have
8
ÿ 8
ÿ
@ f P L2 (Im ).
´m
@ D ´m
@ D
f =2 f, ψ
{ m,´n
2
ψ
{ m,´n = 2 f, ψ
z m,n
2
ψ
z m,n
L (Im ) L (Im )
n=´8 n=´8

Now, every f P L2 (R) can be expressed as

[ ÿ ] [ ÿ ]
[ ] 8 8 8
ÿ
f =F ´1
F [f ] = F ´1
f 1Im = F
p ´1
2 ´m
xf 1Im , ψm,n yL2 (Im ) ψm,n
p z z
m=´8 m=´8 n=´8
[ ÿ
8 8 ] [ ÿ
8 8 ]
ÿ1 pz ÿ
=F ´1
xf , ψm,n yL2 (R) ψ
z m,n = F
´1
xf, ψm,n yL2 (R) ψ
z m,n
m=´8 n=´8
2π m=´8 n=´8
8
ÿ 8
ÿ
= xf, ψm,n yL2 (R) ψm,n .
m=´8 n=´8

This shows that the linear span of tψm,n um,nPZ is dense in L2 (R); thus tψm,n um,nPZ is an
orthonormal basis in L2 (R).
Chapter 7

Multi-resolution Analysis and

Construction of Wavelets

Throughout the chapter, for simplicity we use x¨, ¨y instead of x¨, ¨yL2 (R) .

7.2 Definition of MRA and Examples

Definition 7.1. An MRA consists of a sequence tVm | m P Zu of embedded closed subspaces
of L2 (R) that satisfy the following conditions:
(i) ¨ ¨ ¨ Ď V´2 Ď V´1 Ď V0 Ď V1 Ď V2 Ď ¨ ¨ ¨ ; that is, Vm Ď Vm+1 for all m P Z;
8 ( Ť
8 )
2
(ii) Vm is dense in L (R); that is, closure}¨}2 Vm = L2 (R).
Ť
m=´8 m=´8
8
(iii)
Ş
Vm = t0u.
m=8

(iv) f P Vm if and only if d1/2 f P Vm+1 for all m P Z;

(v) there exists a function ϕ P V0 such that ϕ0,n = Tn ϕ ˇ n P Z is an orthonormal basis

␣ ˇ (

for V0 ; that is, 8

ˇxf, ϕn yˇ2
ÿ ˇ ˇ
}f }2 2 =
L (R) @ f P V0 .
n=´8

The function ϕ is called the scaling function or father wavelet. If tVm umPZ is a multi-
resolution of L2 (R) and if V0 is the closed subspace generated by the integer translates of a
single function ϕ, then we say that ϕ generates the MRA.

Sometimes, condition (v) is relaxed by assuming that Tn ϕ ˇ n P Z is a Riesz basis for

␣ ˇ (
2
n=´8 P ℓ such that
V0 ; that is, for every f P V0 , there exists a unique sequence tcn u8
8
ÿ 8
ÿ
f (x) = cn (Tn ϕ)(x) = cn ϕ(x ´ n).
n=´8 n=´8

with convergence in L2 (R) and there exist two positive constants A and B independent of
f P V0 such that
8
ÿ 8
ÿ
A |cn |2 ď }f }2L2 (R) ď B |cn |2 ,
n=´8 n=´8

53
where 0 ă A ď B ă 8. In this case, we have an MRA with a Riesz basis.
Note that condition (v) implies that Tn ϕ ˇ n P Z is a Riesz basis for V0 with frame
␣ ˇ (

bounds A = B = 1.
Since ϕ0,n = Tn ϕ P V0 for all n P Z. Further, if n P Z, it follows from (iv) that
m
ϕm,n (x) = (D2´m Tn ϕ)(x) = 2 2 ϕ(2m x ´ n), mPZ (7.2.1)

is an orthonormal basis for Vm .

Consequences of Definition 7.1.

1. A repeated application of condition (iv) implies that f P Vm if and only if d2´k P Vm+k
for all m, k P Z. In other words, f P Vm if and only if d2m f P V0 for all m P Z.
This shows that functions in Vm are obtained from those in V0 through a scaling 2´m .
If the scale m = 0 is associated with V0 , then the scale 2´m is associated with Vm .
Thus, subspaces Vm are just scaled versions of the central space V0 which is invariant
under translation by integers; that is, Tn V0 = V0 for all n P Z.

2. It follows from Definition 7.1 that an MRA is completely determined by the scaling
function ϕ, but not conversely. For a given ϕ P V0 , we first define
" 8
ÿ ÿ8 ˇ *
ˇ 8 2
V0 = f = cn ϕ0,n = cn Tn ϕ ˇ tcn un=´8 P ℓ .
ˇ
n=´8 n=´8

Condition (v) implies that V0 has an orthonormal basis tϕ0,n unPZ . Then, V0 consists of
8 8
all functions f = cn Tn ϕ with finite energy }f }2L2 (R) = |cn |2 ă 8. Similarly,
ř ř
n=´8 n=´8
the space Vm has the orthonormal basis tϕm,n unPZ given by (7.2.1) so that fm is given
by
8
ÿ
fm (x) = cmn ϕm,n (x) (7.2.2)
n=´8

with the finite energy

8
ÿ
}fm }2L2 (R) = |cmn |2 ă 8.
n=´8

Thus, fm represents a typical function in the space Vm . It builds in self-invariance and

scale invariance through the basis tϕm,n unPZ .

3. Conditions (ii) and (iii) can be expressed in terms of the orthogonal projections Pm
onto Vm ; that is, for all f P L2 (R),

lim Pm f = 0 and lim Pm f = f . (7.2.3)

mÑ´8 mÑ8

The projection Pm f can be considered as an approximation of f at the scale 2´m .

Therefore, the successive approximations of a given function f are defined as the
 orthogonal projections Pm onto the space Vm :
8
ÿ
Pm f = xf, ϕm,n y ϕm,n , (7.2.4)
n=´8

where ϕm,n given by (7.2.1) is an orthonormal basis for Vm .

4. Since V0 Ď V1 , the scaling function ϕ that leads to a basis for V0 also belongs to V1 .
Since ϕ P V1 and tϕ1,n unPZ is an orthonormal basis for V1 , ϕ can be expressed in the
form
8
ÿ ? ÿ8
ϕ(x) = cn ϕ1,n (x) = 2 cn ϕ(2x ´ n), (7.2.5)
n=´8 n=´8

where
8
ÿ
cn = xϕ, ϕ1,n y and |cn |2 = 1.
n=´8

Equation (7.2.5) is called the dilation equation. It involves both x and 2x and is often
referred to as the two-scale equation or refinement equation because it displays ϕ in
the refined space V1 . The space V1 has the finer scale 2´1 and it contains ϕ which has
scale 1.

From the preceding facts MRA is described so that we can specify

(a) the subspaces Vm ,

(b) the scaling function ϕ,

n=´8 in the dilation equation (7.2.5).

(c) the coefficient tcn u8

The real importance of an MRA lies in the simple fact that it enables us to construct an
orthonormal wavelet for L2 (R). In order to prove this statement, we first assume that
m=´8 is an MRA. Since Vm Ď Vm+1 , we define Wm as the orthogonal complement of
tVm u8
Vm in Vm+1 for every m P Z, so that we have
( )
Vm+1 = Vm ‘ Wm = Vm´1 ‘ Wm´1 ‘ Wm = ¨ ¨ ¨ = V0 ‘ W0 ‘ W1 ‘ ¨ ¨ ¨ ‘ Wm
(àm )
= V0 ‘ Wk (7.2.6)
k=0

and Vn K Wm for n ă m.
8
Since Vm is dense in L2 (R), we may take the limit as m Ñ 8 to obtain
Ť
m=´8
(à
8 )
V0 ‘ Wm = L2 (R).
m=0

Similarly, we may go in the other direction to write

V0 = V´1 ‘ W´1 = (V´2 ‘ W´2 ) ‘ W´1 = ¨ ¨ ¨ = V´m ‘ W´m ‘ ¨ ¨ ¨ ‘ W´1 .

 8
We may again take the limit as m Ñ 8. Since Vm = t0u, it follows that lim V´m =
Ş
m=´8 mÑ8
t0u. Consequently, it turns out that
8
à
Wm = L2 (R). (7.2.7)
m=´8

Finally, the difference between the two successive approximations Pm f and Pm+1 f is
given by the orthogonal projection Qm f of f onto the orthogonal complement Wm of Vm in
Vm+1 so that
Qm f = Pm+1 f ´ Pm f .
It follows from conditions (i)-(v) in Definition 7.1 that the spaces Wm are also scaled versions
of W0 and, for f P L2 (R),

f P Wm if and only if d 2m f P W 0 @m P Z (7.2.8)

since

f P Wm ô f P Vm+1 and f K Vm ô d2m f P V1 and d2m f K V0 ô d2m f P W0 .

Moreover, Wm ’s are mutually orthogonal spaces generating all L2 (R); that is,
à
Wm K Wk if m ‰ k and Wm = L2 (R), (7.2.9)
mPZ

and are translation-invariant for the discrete translations n P Z; that is,

f P W0 if and only if T n f P W0 ,

where the translation-invariant is due to the following equivalence:

8
L2
ÿ
2
f P W0 ô f P V1 and f K V0 ô f = ck ϕ1,k for some tck u8
k=´8 P ℓ and f K V0
k=´8
8
L2
ÿ
2
ô Tn f = ck´2n ϕ1,k for some tck u8
k=´8 P ℓ and Tn f K V0
k=´8

ô Tn f P V1 and Tn f K V0 ô Tn f P W0 .

Moreover, it can be shown that there exists a function ψ P W0 such that ψ0,n = Tn ψ
constitutes an orthonormal basis for W0 . It follows from (7.2.8) that

ψm,n (x) = 2m/2 ψ(2m x ´ n) for n P Z (7.2.10)

constitute an orthonormal basis for Wm . Thus, the family tψm,n um,nPZ represents an or-
thonormal basis of wavelets for L2 (R).

Example 7.2 (Characteristic Function and Piecewise Constant Function). We assume that
ϕ = 1[0,1] is the characteristic function of the interval [0, 1]. Define spaces Vm by
" ÿ 8 ˇ *
ˇ 8 2
Vm = ck ϕm,k ˇˇ tck uk=´8 P ℓ ,
k=´8
where
ϕm,n (x) = 2m/2 ϕ(2m x ´ n).
The spaces Vm satisfy all the conditions of Definition 7.1, and so, tVm umPZ is an MRA.
[ n n + 1]
The space Vm consists of functions in L2 (R) which are constant on intervals m , m ,
2 2
where n P Z. Obviously, Vm Ď Vm+1 because any function that is constant on intervals of
length 2´m is automatically constant on intervals of half that length. The space V0 contains
all functions f in L2 (R) that are constant on n ď x ă n + 1. The function d1/2 f in V1 is
n n+1
then constant on ďxă . Intervals of length 2´m are usually referred to as dyadic
2 2
intervals. A sample function in spaces Vm is shown in Figure 7.1.

Figure 7.1: Piecewise constant functions in V´1 ; V0 and V1

As we shall see later, this MRA is related to the classic Haar wavelet.

7.3 Properties of Scaling Functions and Orthonormal

Wavelet Bases
Lemma 7.3. Let f, g P L2 (R). Then the function
8
ÿ 8
ÿ
[T2kπ (fpgp)](¨) ” fp(¨ + 2kπ)p
g (¨ + 2kπ)
k=´8 k=´8

belongs to L1 (0, 2π), and for all n P Z,

[ 8 ]
1 2π inω ÿ p
ż
xf, Tn gy = e g (ω + 2kπ) dω .
f (ω + 2kπ)p (7.3.1)
2π 0 k=´8

␣ ( 8
In other words, xf, T´n gy nPZ is the Fourier coefficient of
ř
[T2kπ (fpgp)].
k=´8
Proof. Let f, g P L2 (R) and n P Z be given. Then fp, gp P L2 (R) so that fpgp P L1 (R); thus
8
Theorem 1.26 shows that the series [T2kπ (fpgp)] converges in L1 (0, 2π). Therefore, the
ř
k=´8
2
fact that Tn g P L (R) and the Plancherel identity (1.1.15) show that
8 ż
1 ÿ 2(k+1)π p
ż
1 py 1 inω
xf, Tn gy = xf , Tn gy = f (ω)e gp(ω)dω =
p g (ω)einω dω
f (ω)p
2π 2π R 2π k=´8 2kπ
8 ż
1 ÿ 2π p
= f (ω + 2kπ)p g (ω + 2kπ)einω dω
2π k=´8 0
[ 8 ]
1 2π inω ÿ p
ż
= e f (ω + 2kπ)p g (ω + 2kπ) dω . ˝
2π 0 k=´8

Corollary 7.4. Let f, ϕ P L2 (R), and ϕ0,n ” Tn ϕ ˇ n P Z be an orthonormal sysmtem.

␣ ˇ (

Then xf, ϕ0,n y = 0 for all n P Z (this can be expressed as f K V0 , where V0 is the closure of
the linear span of the orthonormal system) if and only if
8
ÿ
fp(ω + 2kπ)ϕ(ω
p + 2kπ) = 0 for a.a. ω P R. (7.3.2)
k=´8

Proof. Let ϕ be the role of g in Lemma 7.3. Since ϕ0,n = Tn ϕ, (7.3.1) shows that
ż [ 8 ]
1 2π ÿ p p + 2kπ) einω dω .
xf, ϕ0,n y = f (ω + 2kπ)ϕ(ω
2π 0 k=´8

Consequently, it follows from the completeness of einω ˇ n P Z (which holds for functions in
␣ ˇ (

L1 (0, 2π) as well since the Cesàro mean of the Fourier series of f converges to f in L1 (0, 2π)
if f P L1 (0, 2π)) that xf, ϕ0,n y = 0 for all n P Z if and only if (7.3.2) holds. ˝

Theorem 7.5. For any function ϕ P L2 (R), the following conditions are equivalent.

(a) The system ϕ0,n ” Tn ϕ ˇ n P Z is orthonormal.

␣ ˇ (

8 ˇ ˇ2
(b) ϕ(ω + 2kπ)ˇ = 1 for a.a. ω P R.
ř ˇp
k=´8

Proof. Letting f = g = ϕ in Lemma 7.3, we find that

8
1 2π i(m´n)ω ÿ ˇˇp
ż
ˇ2
xϕ0,n , ϕ0,m y = xϕ0,0 , ϕ0,m´n y = e ϕ(ω + 2kπ)ˇ dω .
2π 0 k=´8
ż 2π
1
By the fact that ei(m´n)ω dω = δnm , the identity above implies that
2π 0
ż 2π [ ÿ
8 ]
ˇ2
@ n, m P Z.
i(m´n)ω
ˇ
e ϕ(ω + 2kπ)ˇ ´ 1 dω = xϕ0,n , ϕ0,m y ´ δmn
ˇp
0 k=´8

8 ˇ ˇ2
Thus, xϕ0,n , ϕ0,m y = δnm if and only if ϕ(ω + 2kπ)ˇ = 1 almost everywhere.
ř
ˇp ˝
k=´8
 ␣ ˇ
Theorem 7.6. For any two functions ϕ, ψ P L2 (R), the sets of functions ϕ0,n ” Tn ϕ ˇ n P
Z and ψ0,m ” Tm ψ ˇ m P Z are bi-orthogonal; that is,
( ␣ ˇ (

xϕ0,n , ψ0,m y = 0 @ n, m P Z

if and only if
8
ÿ
ϕ(ω
p + 2kπ)ψ(ω
p + 2kπ) = 0 for a.a. ω P R.
k=´8

Proof. Letting f = ϕ and g = ψ in Lemma 7.3, we find that

ż 2π [ ÿ
8 ]
1 i(m´n)ω
xϕ0,n , ψ0,m y = xϕ, ψ0,m´n y = e ϕ(ω
p + 2kπ)ψ(ω
p + 2kπ) dω .
2π 0 k=´8

Thus, the same reason for proving Theorem 7.5, we conclude that

xϕ0,n , ψ0,m y = 0 @ n, m P Z

if and only if
8
ÿ
ϕ(ω
p + 2kπ)ψ(ω
p + 2kπ) = 0 for a.a. ω P R. ˝
k=´8

A somewhat weaker property than the property of orthonormality in the previous theo-
rem is the “Riesz (or unconditional) condition”, which we study in the following.

Theorem 7.7. For any function ϕ P L2 (R) and constants 0 ă A ď B ă 8, the following
two statements are equivalent:
␣ (
(i) Tk ϕ = ϕ(¨ ´ k) kPZ satisfies the Riesz condition with Riesz bounds A and B; that is,
› 8 ›2
› ÿ ›
A}tck ukPZ }2ℓ2 ď›
› ck ϕ(¨ ´ k)›› ď B}tck ukPZ }2ℓ2 @ tck ukPZ P ℓ2 . (7.3.3)
2 L (R)
k=´8

(ii) The Fourier transform ϕp of ϕ satisfies

8
ÿ ˇ2
for a.a. x P R.
ˇ
Aď ϕ(x + 2kπ)ˇ ď B
ˇp (7.3.4)
k=´8

Proof. Let ϕ P L2 (R) and 0 ă A ď B ă 8 be given. For each n P Z, let en (x) = e´inx . Since
each sequence in ℓ2 corresponds to a unique function in L2 (0, 2π) and vice versa, (7.3.3) is
equivalent to that
› 8 ›2
› ÿ ›
A}C}2L2 (0,2π) ď ›› xC, ek yL2 (0,2π) Tk ϕ ›› ď B}C}2L2 (0,2π) @ C P L2 (0, 2π), (7.3.5)
k=´8 L2 (R)
ż 2π
1
here we recall that xf, gyL2 (0,2π) = f (x)g(x)dx. Moreover, for every function C P
2π 0
 8
L2 (0, 2π), C(ω) = xC, ek yL2 (0,2π) e´ikω for almost all ω P R; thus the fact that T
ř
k ϕ(ω) =
y
k=´8
e´ikω ϕ(ω)
p shows that
8
ÿ
C(ω)ϕ(ω)
p = xC, ek yL2 (0,2π) T
y k ϕ(ω) for a.a. ω P R
k=´8

so that the monotone convergence theorem further implies that

ˇ2
ˇ 8 8 ż 2(k+1)π
ż ˇ ÿ ż
ˇ2
p ˇ2 dω
ˇ ˇ ÿ ˇ ˇ
ˇ xC, e k yL2 (0,2π)
T
y k ϕ(ω) ˇ dω = ˇC(ω) ϕ(ω)
p ˇ dω = ˇC(ω) ϕ(ω)
R k=´8 R
ˇ ˇ
k=´8 2kπ
8 ż 2π
p + 2kπ)ˇ2 dt
ÿ ˇ ˇ
= ˇC(t)ϕ(t
k=´8 0
ż 2π 8
ˇC(t)ˇ2
ˇ ˇ ÿ ˇ ˇ2
= ϕ(t + 2kπ)ˇ dt.
ˇp (7.3.6)
0 k=´8

8
“(i) ñ (ii)” Let C P L2 (0, 2π) be given. Note that (7.3.5) implies that
ř
xC, ek yL2 (0,2π) Tk ϕ
k=´8
converges in L2 (R). By the Plancherel identity (1.1.16),
› 8 ›2 › 8 ›2
› ÿ › › ÿ ›
› xC, ek yL2 (0,2π) Tk ϕ ›› = lim ›› xC, ek yL2 (0,2π) Tk ϕ ››
› nÑ8
k=´8 L2 (R) k=´8 L2 (R)
› n ›2 › 8 ›2
1 ›› ÿ › 1 ›› ÿ ›
= lim xC, ek yL2 (0,2π) Tk ϕ ›
y › = xC, ek yL2 (0,2π) Tk ϕ ››
y ; (7.3.7)
nÑ8 2π › 2π k=´8
›
k=´n L2 (R) L2 (R)

and (7.3.6) further shows that

› 8 ›2 ż 2π 8
› ÿ › 1 ˇC(t)ˇ2
ˇ ˇ ÿ ˇ ˇ2
› xC, ek yL2 (0,2π) Tk ϕ ›› = ϕ(t + 2kπ)ˇ dt.
ˇp
›
k=´8 2 L (R) 2π 0 k=´8

Therefore, condition (7.3.5) and the Parseval identity show that

ż 2π 8
1 ˇC(t)ˇ2
ˇ ˇ ÿ ˇ ˇ2
A}C}2L2 (0,2π) ď ϕ(t + 2kπ)ˇ dt ď B}C}2L2 (0,2π)
ˇp @ C P L2 (0, 2π).
2π 0 k=´8

1
Let tgn u8
n=1 be an approximation of the identity. Replacing |C|2 by Tx gn in the
2π
inequality above and passing to the limit as n Ñ 8, we conclude (7.3.4).

“(ii) ñ (i)” Let C P L2 (0, 2π). Then (7.3.4) and (7.3.6) imply that
ˇ2
ˇ 8
ż ˇ ÿ
1 ˇ
A}C}2L2 (0,2π) ď ˇ xC, e k y 2 T
y k ϕ(ω) ˇ dω ď B}C}2 2
L (0,2π) .
2π R k=´8
ˇ L (0,2π) ˇ

8
2
This shows the series k ϕ converges in L (R). The desired inequality
ř
xC, ek yL2 (0,2π) T
y
k=´8

(7.3.5) then follows from (7.3.7). ˝

 We next proceed to the construction of a mother wavelet by introducing a generating
function in L2 (0, 2π). Before proceeding, we first establish the following

p f P L2 (0, 2π) such that

Lemma 7.8. For every f P V1 , there exists m
(ω ) (ω )
f (ω) = m
p pf ϕp . (7.3.8)
2 2
Indeed, m
p f is given by
8
1 ÿ
m
p f (ω) = ? xf, ϕ1,n y e´inω . (7.3.9)
2 n=´8

Proof. Let cn = xf, ϕ1,n y. Since f P V1 ,

8
ÿ 8
ÿ
f (x) = cn ϕ1,n (x) = cn (D1/2 Tn ϕ)(x),
n=´8 n=´8
8
where |cn |2 = }f }2L2 (R) ă 8. Using (1.1.18), the Fourier transform of the identity above
ř
n=´8
gives
8
( ) 1 ÿ
8 (ω )
´ inω
ÿ
fp(ω) = cn D2 M´n ϕ (ω) =
p ? cn e 2 ϕ
p . ˝
n=´8 2 n=´8 2

p P L2 (0, 2π) in the

The mother wavelet ψ can be generated by the generating function m
following lemma.

Lemma 7.9. The Fourier transform of the scaling function ϕ satisfies the following condi-
tions:
8
ÿ ˇ2
for a.a. ω P R
ˇ
ϕ(ω + 2kπ)ˇ = 1
ˇp (7.3.10)
k=´8 (ω ) (ω )
ϕ(ω) = m
p p ϕp (7.3.11)
2 2
where
8
1 ÿ
m(ω)
p ”m
p ϕ (ω) = ? xϕ, ϕ1,n y e´inω (7.3.12)
2 n=´8
is a 2π-periodic function and satisfies the so-called the orthogonality condition
ˇ2 ˇ ˇ2
for a.a. ω P R.
ˇ
ˇm(ω)
p p + π)ˇ = 1
ˇ + ˇm(ω (7.3.13)

Remark 7.10. The Fourier transform ϕp of the scaling function ϕ satisfies the functional
equation (7.3.11). The function m p is called the generating function of the MRA. This
function is often called the discrete Fourier transform of the sequence tcn u ” txϕ, ϕ1,n yu. In
signal processing, m
p is called the transfer function of a discrete filter with impulse response
tcn u or the low-pass filter associated with the scaling function ϕ.

Proof. Condition (7.3.10) follows from Theorem 7.5, and (7.3.11) follows from Lemma 7.8
(with f being the scaling function ϕ in the lemma).
 To verify the orthogonality condition (7.3.13), we substitute (7.3.11) in (7.3.10) so that
condition (7.3.10) becomes
8 8 ˇ ( )ˇ2 ˇ ( ω )ˇ2
ÿ ˇ2 ÿ ω
for a.a. ω P R.
ˇ
1= ϕ(ω + 2kπ)ˇ = m + kπ ϕ + kπ
ˇp ˇ ˇ ˇp ˇ
p
2 2
ˇ ˇˇ ˇ
k=´8 k=´8

This is true for almost all ω P R and hence, replacing ω by 2ω gives

8
ÿ ˇ2 ˇ ˇ2
for a.a. ω P R.
ˇ
1= ˇm(ω
p + kπ)ˇ ˇpϕ(ω + kπ)ˇ
k=´8

We now split the above infinite sum over k into even and odd integers and use the 2π-periodic
p to obtain that for almost all ω P R,
property of the function m
8 8
ÿ ˇ ˇ2 ˇ ˇ2 ÿ ˇ ˇ2 ˇ ˇ2
1= ˇm(ω
p + 2kπ)ˇ ˇpϕ(ω + 2kπ)ˇ + ˇm(ω
p + (2k + 1)π)ˇ ˇpϕ(ω + (2k + 1)π)ˇ
k=´8 k=´8
8 8
ÿ ˇ ˇ2 ˇ ˇ2 ÿ ˇ ˇ2 ˇ ˇ2
= ˇm(ω)
p ˇ ϕ(ω + 2kπ) +
ˇp ˇ ˇm(ω
p + π)ˇ ˇpϕ(ω + (2k + 1)π)ˇ
k=´8 k=´8
8 8
ˇ ˇ2 ÿ ˇ ˇ2 ˇ ˇ2 ÿ ˇ ˇ2
= ˇm(ω)
p ˇ ϕ(ω + 2kπ) + m(ω
ˇp ˇ ˇ p + π) ˇ ϕ(ω + π + 2kπ)ˇ .
ˇp (7.3.14)
k=´8 k=´8

Using (7.3.10),
8 8
ÿ ˇ2 ÿ ˇ2
for a.a. ω P R;
ˇ ˇ
ϕ(ω + 2kπ) =
ˇp ˇ ϕ(ω + π + 2kπ)ˇ = 1
ˇp
k=´8 k=´8

thus (7.3.14) leads to the desired condition (7.3.13). ˝

The following lemma is useful for reducing the computation in the follow up theorems.

Lemma 7.11. Let tVm umPZ be an MRA with the scaling function ϕ. Then for all f, g P V1 ,
8
ÿ ( )( ω ) ( )( ω )
fp(ω + 2kπ)p
g (ω + 2kπ) = mp fm
pg + mp fm
pg +π for a.a. ω P R, (7.3.15)
k=´8
2 2

where m
p f and m
p g are functions satisfying
(ω ) (ω ) (ω ) (ω )
fp(ω) = mpf ϕp , gp(ω) = m
pg ϕp .
2 2 2 2
Proof. Using (7.3.8) and (7.3.11) (as well as the case with g replacing f ),
8
ÿ
fp(ω + 2kπ)p
g (ω + 2kπ)
k=´8
8
ÿ (ω ) (ω ) (ω ) (ω )
= m
pf + kπ ϕp + kπ mpg + kπ ϕp + kπ
k=´8
2 2 2 2
ÿ8 (ω ) (ω )ˇ ( ω )ˇ2
= m + kπ m + kπ ˇϕp + kπ ˇ .
ˇ ˇ
pf pg
k=´8
2 2 2
Splitting the sum into even and odd integers k, by the 2π-periodicity of m
p f and m
p we obtain
the for almost all ω P R,
8
ÿ
fp(ω + 2kπ)p
g (ω + 2kπ)
k=´8
8
ÿ (ω ) (ω )ˇ ( ω )ˇ2
= m + 2kπ m + 2kπ ˇϕ + 2kπ ˇ
ˇp ˇ
pf pg
k=´8
2 2 2
ÿ8 (ω ) (ω )ˇ ( ω )ˇ2
+ m + (2k + 1)π m + (2k + 1)π ˇϕp + (2k + 1)π ˇ
ˇ ˇ
pf pg
k=´8
2 2 2
(ω ) (ω ) ÿ 8 ˇ ( )ˇ2
ˇp ω
=m m ˇϕ + 2kπ ˇ
ˇ
pf pg
2 2 k=´8 2
(ω ) (ω ) ÿ 8 ˇ ( )ˇ2
ˇp ω
+m +π m +π ˇϕ + (2k + 1)π ˇ .
ˇ
pf pg
2 2 k=´8
2

The desired identity (7.3.15) then follows from Theorem 7.5. ˝

We next prove the following major technical lemma.

Lemma 7.12. Let tVn unPZ be an MRA with the scaling function ϕ, and m
p be the associated
generating function given by (7.3.12). Then the Fourier transform of any function f P W0
can be expressed in the form
[ ( )] ( iω ) ( ω ) (ω )
fp(ω) = νp(ω)d2 M1 ϕT p (ω) = νp(ω) exp
p ´π m m
p + π ϕp , (7.3.16)
2 2 2

where νp is a 2π-periodic function satisfying

1 2π ˇˇ
ż
ˇ2
νp(ω)ˇ dω = }f }2L2 (R) , (7.3.17)
2π 0
[ ( )]
and the factor d2 M1 ϕT p is independent of f .
p ´π m

Proof. Since f P W0 , it follows from V1 = V0 ‘ W0 that f P V1 and is orthogonal to V0 . By

Lemma 7.8, (ω ) (ω )
f (ω) = m
p pf ϕp . (7.3.18)
2 2
where, with cn denoting xf, ϕ1,n y, the function m
p f is given by
8
1 ÿ
m
p f (ω) = ? cn e´inω .
2 n=´8

p f is a 2π-periodic function which belongs to L2 (0, 2π). Moreover, since f K V0 ,

Evidently, m
by Corollary 7.4 and Lemma 7.11, we have
(ω ) (ω ) (ω ) (ω )
mpf m
p +mpf +π m p +π =0 for a.a. ω P R.
2 2 2 2
Replacing ω by 2ω in the identity above gives

0=m
p f (ω)m(ω)
p +m
p f (ω + π)m(ω
p + π) for a.a. ω P R,

or, equivalently,
ˇ ˇ
ˇ m
p f (ω) m(ω
p + π) ˇ
ˇ=0 for a.a. ω P R.
ˇ ˇ
ˇ
ˇ ´m
p f (ω + π) m(ω)
p ˇ

This can be interpreted as the linear dependence of two vectors

( ) ( )
m
p f (ω), ´m
p f (ω + π) and m(ω
p + π), m(ω)
p
( )
for almost all ω P R. Since (7.3.13) implies that the vector m(ω
p + π), m(ω)
p is not a zero
vector for all ω P R, there exists a function λ,
p depending on f , such that
( ) ( )
m
p f (ω), ´m
p f (ω + π) = λ(ω)
p m(ω
p + π), m(ω)
p for a.a. ω P R. (7.3.19)

Using (7.3.19), the 2π-periodicity of m

p and m
p f implies that λ
p is also 2π-periodic. Further-
more,
λ(ω)
p + λ(ω
p + π) = 0 for a.a. ω P R.
Thus, the function M´1 λ p (or the function y = e´iω λ(ω))
p is π-periodic, so there exists a
2π-periodic function νp defined by

λ(ω)
p = eiω νp(2ω). (7.3.20)

A simple combination of (7.3.18), (7.3.19), and (7.3.20) gives the desired representation
(7.3.16).
ˇ ˇ
Finally, by (7.3.13) the π-periodicity of ˇp λˇ implies that
ˇ2 (ˇ ˇ2 )
ż 2π ż 2π ż 2π
ˇ ˇ2 ˇ ˇ2 ˇ ˇ2 ˇ
ˇνp(ω)ˇ dω = 2 λ(ω)ˇ dω = 2
ˇp λ(ω)ˇ ˇm(ω)
ˇp p ˇ + ˇm(ω
p + π)ˇ dω
0
[ πż π π
ż 2π ]
ˇ ˇ2 ˇ ˇ2 ˇ ˇ2 ˇ ˇ2
=2 λ(ω + π)ˇ ˇm(ω
ˇp p + π)ˇ dω + λ(ω)ˇ ˇm(ω
ˇp p + π)ˇ dω
0 π
ż 2π
ˇ ˇ2 ˇ ˇ2
=2 λ(ω)ˇ ˇm(ω
ˇp p + π)ˇ dω .
0

Using (7.3.19) and the Parseval identity,

ż 2π ż 2π 8
ˇνp(ω)ˇ2 dω = 2
ˇ ˇ ˇ ˇ2 ÿ
ˇmp f (ω)ˇ dω = 2π |cn |2 = 2π}f }2L2 (R) ă 8.
0 0 n=´8

This completes the proof of Lemma 7.12. ˝

Now, we return to the main problem of constructing a mother wavelet ψ(x). Suppose
that there is a function ψ such that tψ0,n unPZ is a basis for the space W0 . Then, every
function f P W0 has a series representation
8
ÿ 8
ÿ
f (x) = hn ψ0,n (x) = hn ψ(x ´ n), (7.3.21)
n=´8 n=´8
where hn = xf, ψ0,n y satisfies
8
ÿ
}f }2L2 (R) = |hn |2 ă 8.
n=´8

Using (1.1.18), the application of the Fourier transform to (7.3.21) gives

( ÿ
8 )
fp(ω) = hn e´inω ψ(ω)
p =p
h(ω)ψ(ω),
p (7.3.22)
n=´8

where the function p

h is
8
ÿ
h(ω) =
p hn e´inω ,
n=´8

and it is a square integrable and 2π-periodic function in [0, 2π], and the Parseval identity
implies that
8
1 2π ˇˇp ˇˇ2
ÿ ż
2 2
}f }L2 (R) = |hn | = h(ω) dω (7.3.23)
n=´8
2π 0
When (7.3.22) and (7.3.23) are compared with (7.3.16) and (7.3.17), by picking up the terms
independent of f we see that one possible choice of ψp should be
( iω ) ( ω ) (ω ) (ω ) (ω )
ψ(ω)
p = exp m
p + π ϕp ”m
pψ ϕp , (7.3.24)
2 2 2 2 2
where the function m
p ψ is given by

p + π)eiω = m
p ψ (ω) = m(ω
m p ϕ (ω + π)eiω .

In other words, we choose ψ so that for every f P W0 we have p h = νp. The function mp ψ is
called the filter conjugate to m
p and hence, m
p and m
p ψ are called conjugate quadratic filters
in signal processing.
In the following, we show that the function ψ whose Fourier transform is given by (7.3.24)
is indeed an orthonormal wavelet. We start with the following

Lemma 7.13. Let tVm umPZ be an MRA with the scaling function ϕ and its associated
generating function
8
1 ÿ
m(ω)
p =? cn e´inω , cn = xϕ, ϕ1,n y.
2 n=´8

Then the Fourier transform of a function ψ is given by

( iω ) ( ω ) (ω )
ψ(ω) = exp
p m
p + π ϕp (7.3.25)
2 2 2
if and only if ψ P V1 takes the form
8
ÿ
ψ(x) = (´1)´n´1 c´n´1 ϕ1,n (x). (7.3.26)
n=´8
Proof. By the fact that tcn unPZ P ℓ2 , (1.1.18) implies that
[ ÿ8 ] 8
ÿ
F (´1)´n´1 c´n´1 ϕ1,n = (´1)´n´1 c´n´1 ϕ
y 1,n .
n=´8 n=´8

Therefore, by the injectivity of the Fourier transform and the fact that tϕ1,n unPZ is an
orthonormal basis for V1 , it suffices to show that
8
ÿ ( iω ) ( ω ) (ω )
(´1)´n´1 c´n´1 ϕ
y 1,n (ω) = exp m
p + π ϕp .
n=´8
2 2 2
Nevertheless, the fact that ϕ1,n = D1/2 Tn ϕ implies that
8
ÿ 8
ÿ
(´1)´n´1 c´n´1 ϕ
y 1,n (ω) = c´n´1 (´1)´n´1 (D2 M´n ϕ)(ω)
p
n=´8 n=´8

1
8 (ω ) 1 ÿ
8 (ω )
´i(n+1)π ´in ω inπ+i(n+1) ω
ÿ
=? c´n´1 e e 2 ϕ
p =? cn e 2 ϕ
p
2 n=´8 2 2 n=´8 2
( iω ) 1 ÿ 8
ω
(ω ) ( iω ) ( ω ) (ω )
= exp ? cn e´in( 2 +π) ϕp = exp m
p + π ϕp ,
2 2 n=´8 2 2 2 2
so the lemma is concluded. ˝

Lemma 7.14. Let tVm umPZ be an MRA with the scaling function ϕ and its associated
generating function m,
p and ψ be the function whose Fourier transform is given by
( iω ) ( ω ) (ω )
ψ(ω)
p = exp m
p + π ϕp . (7.3.25)
2 2 2
Then the system ψ0,n ˇ n P Z is an orthonormal system in W0 .
␣ ˇ (

Proof. By Lemma 7.13, ψ P V1 , so it suffices to show that tψ0,n unPZ is an orthonormal

system and ψ K V0 . Since ψp is given by (7.3.25), by setting

m p + π)eiω
p ψ (ω) = m(ω
(ω ) (ω )
we have ψ(ω)
p =m
pψ ϕp . Letting f = g = ψ in (7.3.15), we obtain
2 2
8 ˇ ( ω )ˇ2 ˇ ( ω )ˇ2 ˇ ( ω )ˇ2 ˇ ( ω )ˇ2
p + 2kπ)ˇ2 = ˇˇm
ÿ ˇ ˇ
ˇψ(ω pψ ˇ + ˇm
ˇ ˇ
pψ + π ˇ = ˇm
ˇ ˇ
p ˇ + ˇm
ˇ ˇ
p +π ˇ ,
ˇ
k=´8
2 2 2 2

and the orthogonality condition (7.3.13) further shows that

8
p + 2kπ)ˇ2 = 1
ÿ
for a.a. ω P R.
ˇ ˇ
ˇψ(ω
k=´8

By Theorem 7.5, tψ0,n unPZ is an orthonormal system. Moreover, letting f = ψ and g = ϕ

in (7.3.15), we obtain
8
ÿ ( )( ω ) ( )( ω )
ψ(ω
p + 2kπ)ϕ(ω
p + 2kπ) = m p ψm
p + m p ψmp +π
k=´8
2 2
(ω ) ω (ω ) (ω ) ω (ω )
i2 i( 2 +π)
=mp +π e ¨mp +mp + 2π e ¨m
p +π ,
2 2 2 2
and the 2π-periodicity of m
p shows that
8 [ ] (ω ) (ω )
p + 2kπ) = ei ω2 + ei( ω2 +π) m
ÿ
ψ(ω
p + 2kπ)ϕ(ω p m
p + π = 0.
k=´8
2 2

By Corollary 7.4, ψ K V0 for all n P Z. The translation invariant property then shows that
tψ0,n unPZ K V0 ; thus tψ0,n unPZ is an orthonormal system in W0 . ˝

Lemma 7.15. Let ϑ P L2 (R) be such that tϑ0,n unPZ is an orthonormal system in L2 (R). If
f P L2 (R) and fp = νpϑp for some 2π-periodic function νp P L2 (0, 2π), then
( )
f P closure}¨}L2 (R) span(tϑ0,n unPZ ) ;

8
2
that is, there exists thn u8
n=´8 P ℓ such that f =
ř
hn ϑ0,n .
n=´8
In particular, if ψ is the function given in Lemma 7.14, then tψ0,n unPZ is an orthonormal
basis of W0 .

Proof. Suppose that f P L2 (R) satisfies fp = νpϑp for some 2π-periodic function νp P L2 (0, 2π).
Since νp P L2 (0, 2π), there exists thn unPZ P ℓ2 such that
8
ÿ
νp(ω) = hn e´inω
n=´8

and the convergence is in L2 (0, 2π). Therefore,

8
ÿ 8
ÿ 8
ÿ 8
ÿ
fp(ω) = hn e´inω ϑ(ω)
p = hn (M´n ϑ)(ω)
p = hn Ty
n ϑ(ω) = hn ϑy
0,n (ω).
n=´8 n=´8 n=´8 n=´8
! ϑy )
Using (1.1.18) (with F ´1
and q replacing F and p, respectively), the fact that ?0,n
2π nPZ
is an orthonormal system (which is a direct consequence of the Plancherel identity (1.1.15))
and thn unPZ P ℓ2 imply that
8
ÿ
f= hn ϑ0,n .
n=´8
( )
This shows that f P closure}¨}L2 (R) span(tϑ0,n unPZ ) .
Next we establish that tψ0,n unPZ is an orthonormal basis of W0 . By Lemma 7.14, it
( )
suffices to show that W0 Ď closure}¨}L2 (R) span(tϑ0,n unPZ ) . Nevertheless, by Lemma 7.12
every f P W0 corresponds to a 2π-periodic function νp P L2 (0, 2π) such that fp = νpψ;
p thus
( )
the argument above then shows that W0 Ď closure}¨}L2 (R) span(tψ0,n unPZ ) . ˝

Remark 7.16. The key element to establish Lemma 7.15 is (1.1.18). Due to its similar
version (1.5.9), one can relaxed the condition that “tϑ0,n unPZ is an orthonormal system in
L2 (R)” to that “tϑ0,n unPZ has Riesz bounds A and B for some positive A and B”.

The combination of Lemma 7.13, 7.14 and 7.15 leads to the main theorem of this section.
Theorem 7.17. If tVn unPZ is an MRA with the scaling function ϕ, then there is a mother
wavelet ψ given by
? ÿ8 8
ÿ
´n´1
ψ(x) = 2 (´1) c´n´1 ϕ(2x ´ n) = (´1)´n´1 c´n´1 ϕ1,n (x), (7.3.26)
n=´8 n=´8

where the coefficients cn are given by

? ż
cn = xϕ, ϕ1,n y = 2 ϕ(x)ϕ(2x ´ n)dx.
R

That is, the system ψm,n ˇ n P Z is an orthonormal basis for Wm .

␣ ˇ (

Example 7.18 (The Shannon Wavelet). We consider the Fourier transform ϕp of a scaling
function ϕ defined by ϕ(ω)
p = 1[´π,π] (ω) so that
& sin πx if x ‰ 0,
$
żπ
1 ixω
ϕ(x) = e dω = sinc(πx) = πx
2π ´π % 1 if x = 0.
This is also known as the Shannon sampling function. Clearly, the Shannon scaling function
does not have finite support. However, its Fourier transform has a finite support (band-
limited) in the frequency domain and has good frequency localization. Evidently, the system

ϕ0,k (x) = ϕ(x ´ k) = sinc(π(x ´ k)) kPZ

is orthonormal because
1 @ { pD
xϕ0,k , ϕ0,ℓ yL2 (R) = xϕ0,k´ℓ , ϕ0,0 yL2 (R) = xTk´ℓ ϕ, ϕyL2 (R) = Tk´ℓ ϕ, ϕ 2
żπ 2π L (R)

1
= e´i(k´ℓ)ω dω = δkℓ .
2π ´π
In general, we define, for m = 0,
" ÿ8 ˇ 8 *
ˇ ÿ 2
V0 = ck sinc(π(x ´ k)) ˇˇ |ck | ă 8
k=´8 k=´8

and, for other m ‰ 0, m P Z,

" ÿ 8 ˇ 8 *
m/2 m
ˇ ÿ 2
Vm = ck 2 sinc(π(2 x ´ k)) ˇ
ˇ |ck | ă 8 .
k=´8 k=´8

It is easy to check that all conditions of Definition 7.1 are satisfied. We next find out the
coefficients ck defined by
@ D 1 p {D
ck = xϕ, ϕ1,k yL2 (R) = ϕ, D 1 Tk ϕ 2 = xϕ, D 1 Tk ϕ 2
2 L (R) 2π 2 L (R)

1
żπ
1
żπ
ikω
(ω ) 1
żπ
ikω
´ 2 p
= (D2 M´k ϕ)(ω) dω = ?
p e ϕ dω = ? e´ 2 dω
2π ´π 2 2π ´π 2 2 2π ´π
1
$
1 ? if k = 0,
$ ’
? if k = 0,
’
’
& ’
& 2
2
= ? =
? 0 if k is even and k ‰ 0,
2 kπ
sin if k ‰ 0.
’ ’
% 2 (´1) k´1
% ’
if k is odd.
’
kπ 2 2
kπ
Consequently, we can use the formula (7.3.26) to find the Shannon mother wavelet
? ÿ8
ψ(x) = 2 (´1)n´1 c´n´1 ϕ(2x ´ n)
n=´8
[ ]
? 1 ÿ
n´1
= 2 ? ϕ(2x + 1) + (´1) c´n´1 ϕ(2x ´ n)
2 nPZ,n‰´1
[ 8 ]
? 1 ÿ
2ℓ´1
= 2 ? ϕ(2x + 1) + (´1) c´2ℓ´1 ϕ(2x ´ 2ℓ)
2 ℓ=´8
( ) 8
1 ÿ 2(´1)ℓ
= sinc π(2x + 1) ´ sinc(2π(x ´ ℓ)).
π ℓ=´8 (2ℓ + 1)

By Theorem 7.17, the system tψm,n | m, n P Zu is an orthonormal basis in L2 (R). It is known

as the Shannon system.
Theorem 7.19. Let ϕ be a scaling function for an MRA, and ψ be the mother wavelet given
by Theorem 7.17. Then a function ϑ P W0 is an orthonormal wavelet for L2 (R) if and only
if
ϑ(ω)
p = νp(ω)ψ(ω)
p (7.3.27)
ˇ ˇ
for some 2π-periodic function νp such that ˇνp(ω)ˇ = 1.
Proof. First we recall Lemma 7.13 that the mother wavelet ψ satisfies
( iω ) ( ω ) (ω )
ψ(ω)
p = exp m
p + π ϕp . (7.3.25)
2 2 2
Suppose that g P W0 satisfying gp = νpψp for some 2π-periodic νp P L2 (0, 2π). By setting

m p + π)eiω ,
p g (ω) = νp(2ω)m(ω
(ω ) (ω )
we have gp(ω) = m
pg ϕp . By Lemma 7.11, the 2π-periodicity of m
p and νp implies that
2 2
for almost all ω P R,
8 (ω ) (ω ) ˇ ( ω )ˇ2 ˇ ( ω )ˇ2
ˇgp(ω + 2kπ)ˇ2 = (m
ÿ ˇ ˇ
p gm
p g) + (m
p gmp g) + π = ˇm ˇ + ˇm +π ˇ
ˇ ˇ ˇ ˇ
pg pg
k=´8
2 2 2 2
( ω )ˇ ( ω )ˇ
ˇ2 ˇ ˇ2
ˇ ˇ
ˇ2 ˇ ˇ
ˇ2 ˇ
= ˇνp(ω)ˇ ˇm p + π ˇ + ˇνp(ω + 2π)ˇ ˇm p + 2π ˇ
[ 2 )ˇ2 ˇ ( ω )ˇ2 ]
2
ˇ ˇ2 ˇˇ ( ω
= ˇνp(ω)ˇ ˇm + π ˇ + ˇm ˇ ,
ˇ ˇ ˇ
p p
2 2
and the orthogonality condition (7.3.13) further implies that
8
ˇgp(ω + 2kπ)ˇ2 = ˇνp(ω)ˇ2
ÿ
for a.a. ω P R.
ˇ ˇ ˇ ˇ
(7.3.28)
k=´8

“ñ” Suppose that ϑ P W0 is an orthonormal wavelet. By Lemma 7.12, there must be a

2π-periodic function νp P L2 (0, 2π) such that (7.3.27) holds. Letting g = ϑ in (7.3.28),
Theorem 7.5 implies that
8
ÿ ˇ2 ˇ ˇ2
for a.a. ω P R.
ˇ
1= ϑ(ω + 2kπ)ˇ = ˇνp(ω)ˇ
ˇp
k=´8
 ˇ ˇ
“ð” Suppose that for some νp satisfying ˇνpˇ = 1 a.e. the function ϑ satisfies (7.3.27). Lemma
7.15 shows that ϑ P W0 Ď V1 ; thus letting g = ϑ in (7.3.28) shows that
8
ÿ ˇ2 ˇ ˇ2
for a.a. ω P R.
ˇ
ϑ(ω + 2kπ)ˇ = ˇνp(ω)ˇ = 1
ˇp
k=´8

Theorem 7.5 then implies that tϑ0,n unPZ is an orthonormal system in W0 .

Next we show that tϑ0,n unPZ is indeed an orthonormal basis of W0 . Let f P W0 be
p P L2 (0, 2π) satisfying
given. By Lemma 7.12 there exists a 2π-periodic function µ
µp(ω) p
fp(ω) = µ
p(ω)ψ(ω)
p = ϑ(ω).
νp(ω)
µ
Since is 2π periodic and belongs to L2 (0, 2π), Lemma 7.15 implies that
p
νp
( )
f P closure}¨}L2 (R) span(tϑ0,n unPZ ) ;
( )
thus W Ď closure}¨}L2 (R) span(tϑ0,n unPZ ) . ˝

If the scaling function ϕ of an MRA is not an orthonormal basis of V0 but rather is a

Riesz basis, we can use the following orthonormalization process to generate an orthonormal
basis.

Theorem 7.20 (Orthonormalization Process). Let ϕ P L2 (R) be such that tϕ0,n unPZ is a
Riesz basis of V0 ; that is, the linear span of tϕ0,n unPZ is dense in V0 and (by Theorem 7.7)
there exists two constants A, B ą 0 such that
8
ÿ ˇ ˇ2
0ăAď ϕ(ω + 2kπ)ˇ ď B ă 8.
ˇp (7.3.29)
k=´8
␣ (
Then ϕr0,n nPZ is an orthonormal basis of V0 with

ϕ(ω)
p
ϕ(ω) =d . (7.3.30)
p
r
8 ˇ
ř ˇ2
ϕ(ω + 2kπ)ˇ
ˇp
k=´8

Proof. It follows from (7.3.30) that

8 ˇ
ÿ ˇ2
ˇϕ(ω
r + 2kπ)ˇˇ = 1.
ˇp
k=´8

Thus, the function ϕr satisfies condition (b) of Theorem 7.5, and this shows that tϕr0,n unPZ is
an orthonormal set in L2 (R).
Next we show that tϕr0,n unPZ Ď V0 . We consider a 2π-periodic function νp defined by
1
νp(ω) = d .
8 ˇ
ř ˇ2
ϕ(ω + 2kπ)ˇ
ˇp
k=´8
Using (7.3.29), νp P L2 (0, 2π). Since ϕr = νpϕ,
p Lemma 7.15 and Remark 7.16 together imply
p

that
( )
ϕr P closure}¨} span(tϕ0,n unPZ ) .
L2 (R)

The fact that tϕ0,n unPZ is a Riesz basis of V0 shows that ϕr P V0 ; thus tϕr0,n unPZ Ď V0 .
1r 1
On the other hand, we also have ϕp = ϕ. Since is 2π-periodic and belongs to L2 (0, 2π)
p
νp νp
(due to (7.3.29)), Lemma 7.15 shows that
( )
ϕ P closure}¨}L2 (R) span(tϕr0,n unPZ ) .

This further implies that

( )
tϕ0,n unPZ Ď closure}¨}L2 (R) span(tϕr0,n unPZ ) . (7.3.31)
Since tϕ0,n unPZ is a Riesz basis, we have
( )
L2 (R) = closure}¨}L2 (R) span(tϕ0,n unPZ ) ;
( )
thus (7.3.31) shows that L2 (R) = closure}¨}L2 (R) span(tϕr0,n unPZ ) . ˝

Finally, we provide a sufficient condition for a function being a qualified scaling function.

Theorem 7.21. Let ϕ be a bounded function with compact support, ϕ(0)

p = 1, and tϕ0,n unPZ
is an orthonormal system in L2 (R). If it holds the two-scale equation
8
ÿ
ϕ(x) = xϕ, ϕ1,n yϕ1,n (x), (7.2.5)
n=´8
( ( ))
then Vm defined Vm = closure}¨}L2 (R) span tϕm,n unPZ forms an MRA tVm umPZ .

Proof. W.L.O.G, we assume that supp(ϕ) Ď [0, L] for some L P N. First we note that
ϕ P L2 (R) since it is continuous with compact support. Due to the two-scale equation
(7.2.5), we have

Vm Ď Vm+1 @m P Z and f P Vm ô d1/2 f P Vm+1 .

Therefore, it suffices to show (iii) and (iv) in the Definition of MRA; that is,
( ď
8 ) 8
č
closure}¨}L2 (R) Vm = L2 (R) and Vm = t0u.
m=´8 m=´8
( 8
)
We first focus on showing that closure}¨}L2 (R) = L2 (R). Let Pm : L2 (R) Ñ Vm
Ť
Vm
m=´8
be the orthogonal projection defined by
8
ÿ
Pm f = xf, ϕm,ℓ yϕm,ℓ .
ℓ=´8

By the nature of the orthogonality,

}f }2L2 (R) = }f ´ Pm f }2L2 (R) + }Pm f }2L2 (R) .

 › ›
The identity above implies that to show that lim ›Pm f ´ f ›L2 (R) = 0, we only need to that
mÑ8
lim }Pm f }L2 (R) = }f }L2 (R) .
mÑ8
Claim: For all finite sequence thk unk=1 Ď C and ´8 ă a1 ă b1 ď a2 ă b2 ď ¨ ¨ ¨ ď ak ă bk ď
¨ ¨ ¨ ď an ă bn ă 8, ( ď )
n
ÿ 8
hk 1[ak ,bk ) P closure}¨}L2 (R) Vm .
k=1 m=´8

Proof of Claim: We first prove the case that n = 1 and h1 = 1. Let a = a1 and b = b1 . For
each m P Z,
8 (ż b
ÿ )
Pm 1[a,b) = ϕm,ℓ (x)dx ϕm,ℓ .
ℓ=´8 a

Therefore, the orthonormality of tϕm,ℓ uℓPZ shows that for m " 1,

ˇ2 8 ˇ ż 2m b´ℓ
8 ˇż b
ˇ2
ˇ
› ›2 ÿ ˇ ˇ ´m
ÿ ˇ
›Pm 1[a,b) › 2 = ˇ ϕm,ℓ (x)dxˇ = 2 ˇ ϕ(x)dxˇˇ
L (R) ˇ ˇ ˇ
ℓ=´8 a ℓ=´8 2m a´ℓ
( [2mÿ
a]´L m a]
[2ÿ [2mÿ
b]´L m b]
[2ÿ 8
ÿ )ˇ ż 2m b´ℓ ˇ2
´m
ˇ ˇ
=2 + + + + ˇ
ˇ m ϕ(x)dxˇˇ .
ℓ=´8 ℓ=[2m a]´L+1 ℓ=[2m a]+1 ℓ=[2m b]´L+1 ℓ=[2m b]+1 2 a´ℓ

1. For ℓ ď [2m a] ´ L or ℓ ě [2m b] + 1, we have

[ ]
2m a ´ ℓ, 2m b ´ ℓ X [0, L] = H.

Therefore,
( [2mÿ
a]´L 8
ÿ )ˇ ż 2m b´ℓ ˇ2
ˇ ˇ
+ ˇ
ˇ m ϕ(x)dx ˇ = 0.
ˇ
ℓ=´8 ℓ=[2m b]+1 2 a´ℓ

ˇ ż 2m b´ℓ ˇ
ˇ ˇ
2. Since ˇˇ ϕ(x)dxˇˇ ď }ϕ}L1 (R) , we have
m2 a´ℓ

( m a]
[2ÿ m b]
[2ÿ )ˇ ż 2m b´ℓ ˇ2
ˇ ˇ
+ ˇ
ˇ m ϕ(x)dxˇˇ ď 2L}ϕ}2L1 (R) ;
ℓ=[2m a]´L+1 ℓ=[2m b]´L+1 2 a´ℓ

thus
( [2m
ÿ a]´1 m b]
[2ÿ )ˇ ż 2m b´ℓ ˇ2
´m
ˇ ˇ
lim 2 + ˇ ϕ(x)dxˇˇ = 0.
mÑ8 ˇ m
2 a´ℓ
ℓ=[2m a]´L ℓ=[2m b]´L+1

3. For [2m a] + 1 ď ℓ ď [2m b] ´ L, we have

2m a ´ ℓ ď 2m a ´ [2m a] ´ 1 ă 0 and 2m b ´ ℓ ě 2m b ´ [2m b] + L ě L;

thus the fact that supp(ϕ) Ď [0, L] implies that

[2mÿ
b]´L ˇ ż 2m b´ℓ ˇ2 [2m b]´L ˇ ż ˇ2 [2m b]´L
ˇ ˇ ÿ ˇ ˇ ÿ ˇ ˇ2
ˇ ϕ(x)dx ˇ = ˇ ϕ(x)dxˇ = ϕ(0)
ˇp ˇ
R
ˇ m ˇ ˇ ˇ
ℓ=[2m a]+1 2 a´ℓ ℓ=[2m a] ℓ=[2m a]
m m
= [2 b] ´ [2 a] ´ L.
 Therefore,
[2mÿ
b]´L ˇ ż 2m b´ℓ
( )
ˇ2
ˇ ˇ
lim 2 ´m ˇ ϕ(x)dx ˇ = lim 2´m [2m b] ´ [2m a] ´ L
mÑ8 ˇ m
2 a´ℓ
ˇ mÑ8
ℓ=[2m a]
[ ]
´m m m m m m m
= lim 2 2 b ´ 2 a + ([2 b] ´ 2 b) + (2 a ´ [2 a]) ´ L = b ´ a.
mÑ8

› ›
The discussion above shows that lim ›Pm 1[a,b) ›L2 (R) = b ´ a = }1[a,b) }2L2 (R) .
mÑ8
n
Now if s = hk 1[ak ,bk ) , where thk unk=1 Ď C and ´8 ă a1 ă b1 ď a2 ă b2 ď ¨ ¨ ¨ ď an ă
ř
k=1
bn ă 8, the linearity of Pm implies that
› › › n
›ÿ ( )›› n
ÿ ›
›
›
›
›Pm s ´ s› 2 = › hk Pm 1[ak ,bk ) ´ 1[ak ,bk ) ›› ď |hk |››Pm 1[ak ,bk ) ´ 1[ak ,bk ) ››
L (R) ›
k=1 L2 (R) k=1 L2 (R)

which converges to 0 as m Ñ 8 (because it is a finite sum). This concludes the claim.

Let f P L2 (R) be given, and ε ą 0 be given. There exists finite sequence thk unk=1 Ď C
and 2n real numbers satisfying ´8 ă a1 ă b1 ď a2 ă b2 ď ¨ ¨ ¨ ď an ă bn ă 8 such that the
8
step function s = hk 1[ak ,bk ) satisfies
ř
k=´8
ε
}f ´ s}L2 (R) ă .
3
The orthogonality of Pm further shows that

›Pm f ´ Pm s› 2 ď }f ´ s}L2 (R) ă ε .

› ›
L (R) 3
By Claim 1, there exists N ą 0 such that

›Pm s ´ s› 2 ă ε
› ›
L (R)
whenever m ě N .
3
Therefore, if m ě N ,
› › › ›
}f ´ Pm f }L2 (R) ď }f ´ s}L2 (R) + ›Pm s ´ s›L2 (R) + ›Pm f ´ Pm s›L2 (R) ă ε,
( 8
)
and this shows that f P closure}¨}L2 (R)
Ť
Vm .
m=´8
8
Next we show that Vm = t0u or equivalently,
Ş
m=´8

lim }Pm f }L2 (R) = 0 @ f P L2 (R).

mÑ´8

ε
Let f P L2 (R) and ε ą 0 be given. Choose g P Cc (R) be such that }f ´ g}L2 (R) ă . Suppose
2
that supp(g) Ď [´R, R] and N P N satisfies 2N ě R. Then the fact that supp(ϕm,n ) Ď
[2´m n, 2´m (n + L)] implies that if m ď ´N ,
żR
xg, ϕm,n y = g(x)ϕm,n (x) dx = 0 whenever n ď ´L ´ 1 or n ě 1.
´R
The discussion above shows that if m ď ´N ,
8
ÿ 0
ÿ
Pm g = xg, ϕm,n yϕm,n = xg, ϕm,n yϕm,n .
n=´8 n=´L

Therefore, using the estimate

ˇż ˇ
ˇxg, ϕm,n yˇ = ˇ g(x)2 m2 ϕ(2m x ´ n) dxˇ ď 2 m2 }ϕ}L8 (R) }g}L1 (R) ,
ˇ ˇ ˇ ˇ
R
ˇ ˇ

we have for m ď ´N ,
0
ˇxg, ϕm,n yˇ2 ď 2m (L + 1)}ϕ}2L8 (R) }g}2 1 .
ÿ ˇ ˇ
}Pm g}2L2 (R) = L (R)
n=´L

Therefore, by choosing N even larger, we have for m ď ´N ,

›
}Pm f }L2 (R) ď ›Pm (f ´ g)}L2 (R) + }Pm g}L2 (R) ď }f ´ g}L2 (R) + }Pm g}L2 (R) ă ε

which concludes that lim }Pm f }L2 (R) = 0 for all f P L2 (R). ˝
mÑ0

7.4 Construction of Orthonormal Wavelets

We now use the properties of scaling functions and filters for constructing orthonormal
wavelets.

Example 7.22 (The Haar Wavelet). Example 7.2 shows that spaces of piecewise constant
functions constitute an MRA with the scaling function ϕ = 1[0,1) . Moreover, ϕ satisfies the
dilation equation
? ÿ8
ϕ(x) = 2 cn ϕ(2x ´ n), (7.4.1)
n=´8

where the coefficients cn are given by

? ż
cn = xϕ, ϕ1,n yL2 (R) = 2 ϕ(x)ϕ(2x ´ n) dx. (7.4.2)
R

Evaluating this integral with ϕ = 1[0,1) gives cn as follows:

1
c0 = c1 = ? and cn = 0 for n ‰ 0, 1.
2
Consequently, the dilation equation becomes

ϕ(x) = ϕ(2x) + ϕ(2x ´ 1). (7.4.3)

This means that ϕ is a linear combination of the even and odd translates of d1/2 ϕ and
satisfies a very simple two-scale relation (7.4.3), as shown in Figure 7.2.
 Figure 7.2: Two-scale relation of ϕ(x) = ϕ(2x) + ϕ(2x ´ 1)

Thus, the Haar mother wavelet is obtained from (7.3.26) as a simple two-scale relation
ψ(x) = ϕ1,´1 (x) ´ ϕ1,´2 (x) = ϕ(2x + 1) ´ ϕ(2x + 2) = 1[´ 1 ,0) (x) ´ 1[´1,´ 1 ) (x)
2 2
$ [ 1 )
’
’ 1 if x P ´ , 0 ,
’ 2
[ 1)
&
= ´1 if x P ´1, ´ , (7.4.4)
’
’
’ 2
0 otherwise.
%

This two-scale relation (7.4.4) of ψ is represented in Figure 7.3.

Figure 7.3: Two-scale relation of ψ(x) = ϕ(2x) + ϕ(2x ´ 1)

Alternatively, the Haar wavelet can be obtained from the Fourier transform of the scaling
function ϕ = 1[0,1) so that
( iω ) (ω )
ϕ(ω) = 1[0,1) (ω) = exp ´
p z sinc
2 2
( iω ) (ω ) ( iω ) (ω )
= exp ´ cos exp ´ sinc
( ω ) 4( ω ) 4 4 4
=mp ϕp , (7.4.5)
2 2
where the associated filter m
p and its complex conjugate are given by
( iω ) (ω ) 1
m(ω)
p = exp ´ cos = (1 + e´iω ), (7.4.6a)
2 2 2
( iω ) (ω ) 1
m(ω)
p = exp cos = (1 + eiω ). (7.4.6b)
2 2 2
Thus, the Haar wavelet can be obtained form (7.3.24) or (7.3.27) and is given by
( iω ) ( ω ) (ω )
ψ(ω) = νp(ω) exp
p m
p + π ϕp
2 2 2
( iω ) 1 ( iω )
(ω )
= νp(ω) ¨ exp ¨ 1´e 2 ¨ϕp ,
2 2 2
where νp(ω) = ´ exp(´iω) is chosen to find the exact result (7.4.4) since using this νp, we
obtain
( iω ) 1 ( iω )
(ω ) 1 (ω ) 1 ( iω ) ( ω )
ψ(ω)
p = ´ exp ´ ¨ 1 ´ e 2 ¨ ϕp = ϕp ´ exp ´ ϕp
2 2 2 2 2 2 2 2
1 [ ] 1 [ ]
= ? (D2 ϕ)(ω)
p ´ (M´ 1 D2 ϕ)(ω)
p =? D z1 ϕ(ω) ´ T{1 D 1 ϕ(ω)
2 2
2 2 2 2

so that the inverse Fourier transform gives the exact result (7.4.4) as
1 [ ]
ψ(x) = ? (D 1 ϕ)(x) ´ (T 1 D 1 ϕ)(x) = ϕ(2x) ´ ϕ(2x ´ 1).
2 2 2 2

On the other hand, using (7.3.24) also gives the Haar wavelet as
( iω ) ( ω ) (ω ) ( iω ) 1 ( iω )
( iω ) (ω )
ψ(ω)
p = exp m
p + π ϕp = exp ¨ 1 ´ e 2 ¨ exp ´ sinc
2 2 2 2 2 4 4
( iω ) 1 ( ´iω iω )
(ω ) ( iω ) ω (ω )
= exp ¨ e 4 ´ e 4 ¨ sinc = ´ exp ¨ sin ¨ sinc
[ 2 2 4 2 4 4
( iω ) sin2 (ω/4) ]( )
= i exp ´ ´exp(iω) . (7.4.7)
2 ω/4

This corresponds to the same Fourier transform (6.2.7) of the Haar wavelet (7.4.4) except for
the factor ´ exp(iω). This means that this factor induces a translation of the Haar wavelet
to the left by one unit. Thus, we have chosen νp(ω) = ´ exp(´iω) in (7.3.27) to find the
same value (7.4.4) for the classic Haar wavelet.

Example 7.23 (Cardinal B-splines and Spline Wavelets). When we talk about “cardinal
splines”, we mean “polynomial spline functions with equally spaced simple knots”. We first
consider the set Z of all integers as the “knot sequence”. Let πn denotes the collection of all
algebraic polynomials of degree at most n, and C n (R) denote the collection of all functions
f such that f , f 1 , ¨ ¨ ¨ , f (n) are continuous everywhere, with the understanding that C ´1 (R)
is the space of piecewise continuous functions.
For each positive integer m, the space Sm of cardinal splines of order m and with knot
sequence Z is the collection of all functions f P C m´2 (R) such that the restrictions of f to
any interval [k, k + 1), k P Z, are in πm´1 ; that is,

fæ[k,k+1) P πm´1 , k P Z.

The space S1 of piecewise constant functions is easy to understand. The most convenient
basis to use is B1 (x ´ k) ˇ k P Z , where B1 = 1[0,1) is the scaling function given in the
␣ ˇ (

previous example. To give a basis of Sm , m ě 2, let us first consider the space Sm,N
consisting of the restrictions of functions f P Sm to the interval [´N, N ], where N is a
positive integer. In other words, we may consider Sm,N as the subspace of functions f P Sm
such that the restrictions

fæ(´8,´N +1) and fæ[N ´1,8)

of f are polynomials in πm´1 . This subspace is easy to characterize. Indeed, for an arbitrary
function f in Sm,N , by setting pm,j = fæ[j,j+1) P πm´1 , j = ´N , ¨ ¨ ¨ , N ´ 1, we have, in view
of the fact that f P C m´2 ,
( (ℓ) (ℓ) )
pm,j ´ pm,j´1 (j) = 0, ℓ = 0, 1, ¨ ¨ ¨ , m ´ 2; m ě 2.

That is, by considering the “jumps” of f (m´1) at the knot sequence Z, namely:
(m´1) (m´1) (m´1) (m´1)
cj = pm,j (j + ) ´ pm,j´1 (j ´ ) ” lim+ pm,j (x) ´ lim0 pm,j´1 (x)
xÑj xÑj
[ (m´1) ]
= lim+ f (j + ε) ´ f (m´1) (j ´ ε) , (7.4.8)
εÑ0

the adjacent polynomial pieces of f are related by the identity

cj
pm,j (x) = pm,j´1 (x) + (x ´ j)m´1 . (7.4.9)
(m ´ 1)!
Hence, by introducing the notation
#
x+ = maxt0, xu,
(7.4.10)
xm´1
+ = (x+ )m´1 , m ě 2,

it follows from (7.4.9), that for all x P [´N, N ],

N ´1
ÿ cj
f (x) = fæ[´N,´N +1) (x) + (x ´ j)m´1
+ . (7.4.11)
j=´N +1
(m ´ 1)!

This holds for every f P Sm,N , with the constants cj given by (7.4.8). Consequently, the
collection
1, ¨ ¨ ¨ , xm´1 , (x + N ´ 1)m´1 , ¨ ¨ ¨ , (x ´ N + 1)m´1
␣ (
+ + (7.4.12)

of m + 2N ´ 1 functions is a basis of Sm,N . This collection consists of both monomials

and “truncated powers”. Since we restrict our attention to the interval [´N, N ], it is also
possible to replace the monomials 1, ¨ ¨ ¨ , xm´1 in (7.4.12) by the truncated powers:

(x + N + m ´ 1)m´1
+ , ¨ ¨ ¨ , (x + N )m´1
+ . (7.4.13)

That is, the following set of truncated powers, which are generated by using integer translates
of a single function xm´1
+ , is also a basis of Sm,N :
ˇ
(x ´ k)m´1
␣ (
+
ˇ k = ´N ´ m + 1, ¨ ¨ ¨ , N ´ 1 . (7.4.14)

This basis is more attractive than the basis in (7.4.12) for the following reasons. Firstly,
each function (x ´ j)m´1
+ vanishes to the left of j; secondly all the basis functions in (7.4.14)
are generated by a single function xm´1
+ which is independent of N . Moreover, since
8
ď
Sm = Sm,N ,
N =1
it follows that the basis in (7.4.14) can also be extended to be a “basis” T of the infinite
dimensional space Sm , simply by taking the union of the bases in (7.4.14); that is, we have

T = (x ´ k)m´1 ˇk P Z .
␣ ˇ (
+ (7.4.15)

Since we are mainly concerned with the Hilbert space L2 (R), we are especially interested
in cardinal splines that are in L2 (R). Unfortunately, there is not a single function in T
m´1
that qualifies to be a function in L2 (R), and in fact, each (x ´ k)+ grows to infinity fairly
rapdily as x Ñ +8. To create L2 (R) functions from TN , we must tame the polynomial
growth of (x ´ k)m´1
+ .Define the difference operator ∆ recursively by
#
(∆f )(x) = f (x) ´ f (x ´ 1),
( ) (7.4.16)
(∆n f )(x) = ∆n´1 (∆f ) (x), n = 2, 3, ¨ ¨ ¨ .

Observe that just like the mth order differential operator, the mth order difference of any
polynomial of degree m ´ 1 or less is identically zero, that is,

∆m f = 0, f P πm´1 . (7.4.17)

m=1 : the function M1 ” 1[0,1) , and

This motivates the definition of the sequence tMm u8

1
Mm (x) ” ∆m xm´1
+ @ m ě 2. (7.4.18)
(m ´ 1)!

y y y

x x x

Figure 7.4: The graph of functions Mm , m = 1, 2, 3, from left to right.

It is clear from the definition that Mm is a linear combination of the basis functions in
(7.4.15). In fact, it is easy to verify by induction that if ℓ P N, then for all m P N,
m ( )
ÿ m
∆m xℓ+ = (´1) j
(x ´ j)ℓ+ ,
j
j=0

so we indeed have
m ( )
1 ÿ
k m m´1
Mm (x) = (´1) (x ´ k)+ . (7.4.19)
(m ´ 1)! k=0 k

From (7.4.17), it follows that Mm (x) = 0 for all x ě m. Since Mm (x) clearly vanishes
for x ă 0, we have supp Mm Ď [0, m]. By working a little harder, we can even conclude that

supp(Mm ) = [0, m]. (7.4.20)

So, Mm is certainly in L2 (R). However, is the collection

B ” Tk Mm kPZ
␣ (
(7.4.21)

of integer-translates of Mm a “basis” of Sm ? Let us again return to Sm,N which, according

to (7.4.12) or (7.4.14), has dimension m + 2N ´ 1. Now, by using the support property
(7.4.20), each function in the collection
␣ (N ´1
Tk Mm k=´N ´m+1 (7.4.22)

is non-trivial (at least one function in this collection is non-zero) on the interval [´N, N ]
and Mm (x ´ k) vanishes identically on [´N, N ] for k ă ´N ´ m + 1 or k ą N ´ 1. Since it
can be shown that (7.4.22) is a linearly independent set, we have obtained another basis of
Sm,N . So, analogous to (7.4.15), if we take the union of the bases in (7.4.22), N = 1, 2, ¨ ¨ ¨ ,
we arrive at B in (7.4.21). One advantage of B over T in (7.4.15) is that we can now talk
about a spline series
8
ÿ
f (x) = ck Mm (x ´ k) (7.4.23)
k=´8

without worrying too much about convergence. Indeed, for each fixed x P R, since Mm has
compact support, all except for a finite number of terms in the infinite series (7.4.23) are
zero.
As mentioned earlier, we are mainly interested in those cardinal splines that belong to
L2 (R), namely: Sm X L2 (R). Let V0m denote its L2 (R)-closure. That is, V0m is the smallest
closed subspace of L2 (R) that contains Sm X L2 (R). Since Mm has compact support, we see
that B Ď V0m . We can even show that B is a Riesz basis of V0m , so a scaling function ϕ for
V0m can be constructed using the orthonormalization process described by Theorem .
So far, we have only considered cardinal splines with knot sequence Z. More generally,
j
we will also consider the spaces Sm of cardinal splines with knot sequences 2´j Z, j P Z. Since
a spline function with knot sequence 2´j1 Z is also a spline function with knot sequence 2´j2 Z
whenever j1 ă j2 , we have a (doubly-infinite) nested sequence
´1 0 1
¨ ¨ ¨ Ď Sm Ď Sm Ď Sm Ď ¨¨¨

0
of cardinal spline spaces, with Sm = Sm . Analogous to the definition of V0m , we will let Vjm
denote the L2 (R)-closure of Smj
X L2 (R). Hence, we have a nested sequence
m
¨ ¨ ¨ Ď V´1 Ď V0m Ď V1m Ď ¨ ¨ ¨ (7.4.24)

of closed cardinal spline subspaces of L2 (R). It will be clear that this nested sequence of
subspaces satisfies: $ (ď )
m
’
’
& closure }¨}2 V j = L2 (R),
jPZ
č (7.4.25)
’
’
% Vjm = t0u.
jPZ
Furthermore, it is clear that once we have shown that B is a Riesz basis of V0m , then for any
j P Z, the collection
2 Mm (2j x ´ k) ˇ k P Z
␣ j/2 ˇ (
(7.4.26)
is also a Riesz basis of Vjm with the same Riesz bounds as those of B.
The cardinal B-splines (basis splines) consist of functions in C m´1 (R) with equally
[
spaced integer knots that coincide with polynomials of degree n on the intervals 2´m k, 2´m (k+
]
1) . These B-splines of order n with compact support generate a linear space V0 in L2 (R).
This leads to an MRA tVm | m P Zu by defining f P Vm if and only if d1/2 f P Vm+1 .
The cardinal B-splines Bm of order m are defined by the following convolution product

B1 = 1[0,1] and Bm = B 1 ˙ B1 ˙ ¨ ¨ ¨ ˙ B1 = B1 ˙ Bm´1

looooooooooomooooooooooon n ě 2, (7.4.27)
there are m B1 ’s

where m factors are involved in the convolution product. Obviously,

ż ż1 żx
Bm (x) = Bn´1 (x ´ t)B1 (t) dt = Bm´1 (x ´ t) dt = Bm´1 (t) dt. (7.4.28)
R 0 x´1

Using the formula (7.4.28), we can obtain the explicit representation of splines B2 , B3 , and
B4 as follows: żx żx
B2 (x) = B1 (t) dt = 1[0,1] (t) dt.
x´1 x´1
Evidently, it turns out that

B2 (x) = 0 if x ď 0 or x ě 2,
żx
B2 (x) = dt = x if 0 ď x ď 1,
0
ż1
B2 (x) = dt = 2 ´ x if 1 ď x ď 2.
x´1

Or, equivalently,
B2 (x) = x1[0,1] (x) + (2 ´ x)1[1,2] (x). (7.4.29)
Similarly, using żx
B3 (x) = B2 (t) dt,
x´1
we find the explicit expression of B3 :

B3 (x) = 0 if x ď 0 or x ě 3,
żx żx
x2
B3 (x) = B2 (t) dt = t dt = if 0 ď x ď 1,
0 2
(ż 1 żx) 0 ż1 żx
B3 (x) = + B2 (t) dt = t dt + (2 ´ t) dt if 1 ď x ď 2,
x´1 1 x´1 1
1[ ] 1 1
= 1 ´ (x ´ 1)2 + 2x ´ 2 ´ (x2 ´ 1) = (6x ´ 2x2 ´ 3)
ż2 2 ż2 2 2
´1 ˇt=2
2ˇ 1
B3 (x) = B2 (t) dt = (2 ´ t) dt = (2 ´ t) ˇ = (3 ´ x)2 if 2 ď x ď 3.
x´1 x´1 2 t=x´1 2
Or, equivalently,

x2 1 1
B3 (x) = 1[0,1] (x) + (6x ´ 2x2 ´ 3)1[1,2] (x) + (3 ´ x)2 1[2,3] (x). (7.4.30)
2 2 2
żx
Finally, B4 (x) = B3 (t) dt we have
x´1

B4 (x) = 0 if x ď 0 or x ě 4,
żx
1
B4 (x) = B3 (t) dt = x3 if 0 ď x ď 1,
6
ż0x
1
B4 (x) = B3 (t) dt = (´3x3 + 12x2 ´ 12x + 4) if 1 ď x ď 2,
6
żx´1
x
1
B4 (x) = B3 (t) dt = (3x3 ´ 24x2 + 60x ´ 44) if 2 ď x ď 3,
x´1 6
ż3
1
B4 (x) = B3 (t) dt = (4 ´ x)3 if 3 ď x ď 4.
x´1 6

Or, equivalently,
1 1
B4 (x) = x3 1[0,1) (x) + (´3x3 + 12x2 ´ 12x + 4)1[1,2) (x)
6 6 (7.4.31)
1 1
+ (3x3 ´ 24x2 + 60x ´ 44)1[2,3) (x) + (4 ´ x)3 1[3,4] (x).
6 6
In general, we have the following

Theorem 7.24. The mth order cardinal B-spline Bm satisfies the following properties:

(i) For every f P C (R),

ż ż1 ż1
f (x)Bm (x)dx = ¨¨¨ f (x1 + x2 + ¨ ¨ ¨ + xm )dx1 ¨ ¨ ¨ dxm . (7.4.32)
R 0 0

(ii) For every g P C m (R),

ż m ( )
(m)
ÿ
m´k m
g (x)Bm (x)dx = (´1) g(k). (7.4.33)
R k
k=0

(iii) Bm (x) = Mm (x) for all x P R; that is,

m ( )
1 ÿ
k m
Bm (x) = (´1) (x ´ k)m´1
+ .
(m ´ 1)! k=0 k

(iv) supp(Bm ) = [0, m].

(v) Bm (x) ą 0 for all x P (0, m).

8
(vi) Bm (x ´ k) = 1 for all x P R.
ř
k=´8
(vii) Bm1 (x) = (∆Bm´1 )(x) ” Bm´1 (x) ´ Bm´1 (x ´ 1) for all x P R.

(viii) The cardinal B-splines Bm and Bm´1 are related by the identity:

x m´x
Bm (x) = Bm´1 (x) + Bm´1 (x ´ 1) @ x P R. (7.4.34)
m´1 m´1

(ix) Bm is symmetric with respect to the center of its support, namely:

(m ) (m )
Bm + x = Bm ´x @ x P R.
2 2

Proof. (i) Assertion (7.4.32) certainly holds for m = 1. Suppose it also holds for m ´ 1 (for
some m ě 2), then by the definition of Bm in (7.4.28) and this induction hypothesis,
we have
ż ż ż1 ż1ż
f (t)Bm (t)dt = f (t) Bm´1 (t ´ xm )dxm dt = f (t)Bm´1 (t ´ xm )dtdxm
R
żR1ż 0 0 R

= f (t + xm )Bm´1 (t)dtdxm
0 R
ż 1(ż 1 ż1 )
= ¨ ¨ ¨ f (x1 + ¨ ¨ ¨ xm´1 + xm )dx1 ¨ ¨ ¨ dxm´1 dxm .
0 0 0

It follows from induction that (7.4.32) holds for all m.

(ii) Assertion (7.4.33) holds for m = 1. Suppose that it also holds for m ´ 1 (for some
m ě 2). Then by part (i) we obtain that

ż ż1 ż1
(m)
g (t)Bm (t)dt = ¨¨¨ g (m) (x1 + ¨ ¨ ¨ + xm )dxm ¨ ¨ ¨ dx1
R
ż01 ż01
= ¨ ¨ ¨ g (m´1) (x1 + ¨ ¨ ¨ + xm´1 + 1)dxm´1 ¨ ¨ ¨ dx1
0
ż 1 0ż 1
´ ¨ ¨ ¨ g (m´1) (x1 + ¨ ¨ ¨ + xm´1 )dxm´1 ¨ ¨ ¨ dx1
ż 0 0 ż
(m´1)
= g (t + 1)Bm´1 (t)dt ´ g (m´1) (t)Bm´1 (t)dt
R R
m´1 ) ( m´1 ( )
m´1´k m ´ 1 m´1
ÿ ÿ
m´1´k
= (´1) g(k + 1) ´ (´1) g(k).
k k
k=0 k=0

Since
m´1 ( ) m ( )
ÿ
m´1´k m´1 ÿ
m´k m´1
(´1) g(k + 1) = (´1) g(k),
k k´1
k=0 k=1
 we find that
m´1 ( ) m´1 ( )
ÿ m´1 ÿ m´1
(´1)m´1´k g(k + 1) ´ (´1)m´1´k g(k)
k k
k=0 k=0
m ( ) m´1 ( )
m´k m ´ 1 m´1
ÿ ÿ
= (´1) g(k) ´ (´1)m´1´k g(k)
k´1 k
k=1 k=0
m´1 ( ) m´1 ( )
m´k m ´ 1 m´k m ´ 1
ÿ ÿ
m
= g(m) + (´1) g(k) + (´1) g(0) + (´1) g(k)
k´1 k
k=1 k=1
m´1 [( ) ( )]
m
ÿ
m´k m´1 m´1
= g(m) + (´1) g(0) + (´1) + g(k)
k´1 k
k=1
m´1 ( ) m ( )
m
ÿ
m´k m ÿ
m´k m
= g(m) + (´1) g(0) + (´1) g(k) = (´1) g(k).
k k
k=1 k=0

It follows from induction that (7.4.33) holds for all m P N.

(iii) Clearly B1 (x) = M1 (x) = 1[0,1) (x) for all x P R. Assume that Bm´1 (x) = Mm´1 (x) for
all x P R (for some m ě 2). Then
ż1 ż1
Bm (x) = Bm´1 (x ´ t)dt = Mm´1 (x ´ t)dt
0 0
m´1 ( )ż 1
1 ÿ
k m´1
= (´1) (x ´ t ´ k)m´1
+ dt.
(m ´ 1)! k=0
k 0

Since
$ ż1
’
’
’ (x ´ t ´ k)m´1 dt if x ´ k ě 1,
ż1 & ż 0
’
(x ´ t ´ k)m´1 dt = x´k
0
+
’
’ (x ´ t ´ k)m´1 dt if 0 ď x ´ k ă 1,
% 0
’
’
0 if x ´ k ă 0,
we have
1[ ]
ż1
(x ´ t ´ k)m´1
+ dt = (x ´ k)m m
+ ´ (x ´ k ´ 1)+ .
0 m
Therefore,
m´1 ( )ż 1
1 ÿ
k m´1
Bm (x) = (´1) (x ´ t ´ k)m´1
+ dt
(m ´ 1)! k=0 k 0
( )
m´1 1 [ ]
m´1
1 ÿ
= (´1)k (x ´ k)m+ ´ (x ´ k ´ 1)m+
(m ´ 1)! k=0 k m
[ m´1 ( ) m ( ) ]
1 ÿ k m´1 m
ÿ
k m´1 m
= (´1) (x ´ k)+ + (´1) (x ´ k)+
m! k=0 k
k=1
k´1
m ( )
1 ÿ m
= (´1)k (x ´ k)m
+ = Mm (x).
m! k=1 k

By induction, Bm (x) = Mm (x) for all x P R and m P N.

(vii) Using (7.4.28) again, the Fundamental Theorem of Calculus shows that
d x
ż
1
Bm (x) = Bm´1 (t)dt = ´Bm´1 (x ´ 1) + Bm´1 (x) = (∆Bm´1 )(x).
dx x´1

(viii) To verify the identity in (viii), we use the definition of Mm in (7.4.18) instead. Of
course, we have already shown in (iii) that Bm = Mm . The idea is to represent as the
product of a monomial and a truncated power, namely:

xm´1
+ = x ¨ xm´2
+

and then apply the following “Leibniz Rule” for differences:

n ( )
[ ] ÿ n
∆n (f g) (x) = (∆k f )(x)(∆n´k g)(x ´ k). (7.4.35)
k
k=0

This identity for differences can be easily established by induction. It is almost exactly
the same as the Leibniz Rule for derivatives. Now, if we set f (x) = x and g(x) = xm´2
+
k
in (7.4.35) and recall that ∆ f = 0 for k ě 2 from (7.4.17), we then have
1 1 [ ]
Bm (x) = Mm (x) = ∆m xm´1
+ = x∆ m m´2
x + + m∆ m´1
(x ´ 1) m´2
+
(m ´ 1)! (m ´ 1)!
1 [ ( ) ]
= x ∆m´1 xm´2
+ ´ ∆m´1 (x ´ 1)+
m´2
+ m∆m´1 (x ´ 1)m´2
+
(m ´ 1)!
x m´x
= Bm´1 (x) + Bm´1 (x ´ 1)
m´1 m´1
Assertions (iv), (v), (vi) and (ix) can also be easily derived by induction, using the
definition of Bm in (7.4.28). This completes the proof of Theorem 7.24. ˝

In order to obtain the two-scale relation for the B-splines of order n, we apply the Fourier
transform of (7.4.27) so that
( ) ( )
x1 (ω) = exp ´ iω sinc ω .
B (7.4.36)
2 2
( iω )
Using (7.4.5) and (7.4.6a) we can also express (7.4.36) in terms of z = exp ´ as
2
(ω )
x1 (ω) = 1 (1 + z)B
B x1 . (7.4.37)
2 2
Application of the convolution theorem of the Fourier transform to (7.4.27) gives
[ ] ( 1 + z )n [ ( ω )]n ( 1 + z )n ( ω )
B
xn (ω) = B x1 (ω) n = B
x1 (ω)Bzn´1 (ω) = Bx1 = B
xn
(ω ) (ω ) 2 2 2 2
=M xn B
xn ,
2 2
where the associated filter M
xn is given by
(ω ) ( 1 + z )n n ( ) ( ikω )
1( ´ iω
)n 1 ÿ n
M
xn = = n 1+e 2 = n exp ´
2 2 2 2 k=0 k 2
1 ÿ
8 ( ikω )
”? cn,k exp ´ , (7.4.38)
2 k=´8 2
where the coefficients cn,k are given by
$ ? ( )
& 2 n if 0 ď k ď n,
cn,k = 2n k (7.4.39)
otherwise.
%
0
Therefore, the spline function in the time domain is
? ÿ8 n ( )
1´n n
ÿ
Bn (x) = 2 cn,k Bn (2x ´ k) = 2 Bn (2x ´ k). (7.4.40)
k
k=´8 k=0

This may be referred to as the two-scale relation for the B-splines of order n.
We next show that the cardinal B-spline basis

B = Tk Bm ˇ k P Z
␣ ˇ (
(7.4.41)

is a Riesz (or unconditional) basis of V0m in the sense of Definition 6.26 (or Theorem 6.27).
By Theorem 7.7, this is equivalent to investigating the existence of lower and upper bounds
A, B in (7.3.4). From (7.4.27), we see that so that Bxm = Bx1 m , so that
ˇ2 ˇ 1 ´ e´iω ˇ2m ˇ sin(ω/2) ˇ2m
ˇ ˇ ˇ ˇ
ˇ
ˇBxm (ω)ˇ = ˇ
ˇ iω ˇ = ˇ ω/2 ˇ .
ˇ ˇ ˇ

Hence, replacing ω by 2ω, we have

8 8 8
ÿ ˇ
xm (2ω + 2kπ)ˇ2 =
ˇ ÿ sin2m (ω + kπ) 2m
ÿ 1
ˇB
2m
= sin ω . (7.4.42)
k=´8 k=´8
(ω + kπ) k=´8
(ω + kπ)2m
By the residue theorem, for ω ‰ kπ for all k P Z,
[ ÿ 8 ]
π cot(πz) cot(πz)
¿
1 d ˇˇ
lim dz = 2πi + ˇ ,
N Ñ8 C (ω + zπ)2 (ω + kπ) 2 dz z=´π/ω π
N k=´8
( )
where CN is the square contour with four corners ˘(N + 0.5), ˘(N + 0.5) . The choice of
CN leads to that
π cot(πz)
¿
lim dz = 0,
N Ñ8 C (ω + zπ)2
N

so we conclude that
8
ÿ 1 d ˇˇ cot(πz) 2
␣ ˇ
ˇk P Z .
(
= ´ = csc ω whenever ω R kπ
(ω + kπ)2 dz z=´π/ω π
ˇ
k=´8

Taking another 2m ´ 2 derivative w.r.t. ω yields that

8
ÿ 1 1 d2m´2
2m
= 2m´2
csc2 ω . (7.4.43)
k=´8
(ω + kπ) (2m ´ 1)! dx

Therefore, substituting (7.4.43) into (7.4.42), we obtain

8 2m 2m´2
xm (2ω + 2kπ)ˇ2 = sin ω d
ÿ ˇ ˇ
ˇB
2m´2
csc2 ω . (7.4.44)
k=´8
(2m ´ 1)! dx

This shows that the first order B-spline B1 defined by (7.4.27) is a scaling function that
generates the classic Haar wavelet.