You are on page 1of 36

CHAPTER 5

SCALING FUNCTIONS IN WAVELET ANALYSIS

INTRODUCTION
Chui [5] has introduced new approach of harmonic analysis. Square integral
function with inner product and norm specification can be regarded as a Hilbert
space. By designing a collection of orthonormal bases function every function can be
expressed as double series of these bases functions. We brief an account of why,
what and how are wavelets introduced in Section I. A function called two - scale
function is explained as multi - resolution analysis by Hilbert space of square
integrable function. This property alongwith a functional equation satisfying two -
scale relation is studied in section II of this chapter. Another approach of wavelet
through wavelet matrices and two special wavelets form the contents of the last
section of this chapter.

Section I

5.1 PRELIMINARY RESULTS

Definition 5.1.1 : Fourier Series

Let L2 (0, 2%) denote the collection of all measurable functions f defined on

the interval (0, 2%) with

Jo* |f(x)|2dx<oo

by assuming that f is a piecewise continuous function. It will always be assumed that


functions in L2 (0, 2%) are extended periodically to the real line

n : = (-00,00),

85
namely: f(x) = f(x - 2k) for all x. Hence; the collection L2 (0, 2%) is often called the

space of 2n - periodic square - integrable functions. That L2 (0, 2k) is a vector space
can be verified very easily. Any f in L2 (0, 27t) has a Fourier series representation:

f(x)= £cnein\ (1)

where the constants cn, called the Fourier coefficients of f, are defined by
2%
1 Jf(x)e ““dx.
Cn= —

The convergence of the series in (1) is in L2 (0, 2k), meaning that

2n N
lim
f(x)~ 2>ne‘
M,N —> co i n=-M
dx = 0.

Thus every 2k - periodic square-integrable function is generated by a


‘superposition” of integral dilations of the basic functions w(x) = elx.

Remark 5.1.2
From the orthogonal property of {wn}, the Fourier series representation (1)
also satisfies the so-called Parseval Identity:

y'jflf(x)|2dx= £|cn |2.


2-k. „=-«

Let l denote the space of all square - summable bi-infinite sequences; that is,
{cn} e E1 if and only if
00

ZlCn I2 <00'
n=~oo

Definition 5.1.3 [4]


If a function \\f e L (R) has unit length, then all of the functions \j/j, k, defined by
Vj, k (x) = 2J/2 vg (2J x - k), j, k e Z, (2)

86
also have unit length; that is,

llVj, klb = IHb = 1, j, keZ,

Let the Kronccker symbol

8 .= flforj = k;
j,k [0 for j * k,
defined on Z x Z,

Then, A function vjy e L2 (R) is called an orthogonal wavelet (or o.n. wavelet), if the

family {vj/j k}, as defined in (2), is an orthonormal basis of L2 (R); that is,

< ¥j, k, ¥4 m > = 8j, I • 8k, m, j, M,me Z,

and every fe L (R) can be written as

f(x)= Zcjk'TjkCx), (3)


j,k=-co

where the convergence of the series in (3) is in L2(R), namely:

lim N,

M,,N„M2,N2 ->co f" j=-M2


I k=-M,
S cj,kVFik = 0.

The wavelet coefficients c,-;k are given by

cj,k=<f, ¥j,k>- (4)


That is, if we define an integral transform on L (R) by

(W„f) (b, a) :=|a| -1/1 £ fIdx, f e L2 (R),


V a /

then the wavelet coefficients in(3) and (4) become

C|.k = (Wvf)
(k O
2J 2J

87
Definition 5.1.4 Universal Scaling Functions

We shall say that a function x —»q>(x) is a universal scaling function if it is a

scaling function for every rank. This means that for every m > 1, there exists a

scaling vector a°(m) of rank m such that

<p(x) = Xak(m)q>(mx-k)
keZ

As we shall see, there are two fundamental examples of such functions, the

sine function and the Haar scaling function, and these turn out to be Fourier

transforms of each other. The wavelet functions which are paired with these universal

scaling functions will depend on the choice of wavelet matrix.

Definition 5.1.5 Lawton Matrix


The Lawton Matrix is defined by
L := (Ly) = ^y2H, i, j = 0, ±1,........... ±(2g-2).This is a (4g-3) x (4g-3) matrix.

Definition 5.1.6 Operator G


Let H be an Hilbert Space, and (en) a Riesz basis of H. Let G be the operator
defined on H by
G(f) = X(f\e„)en.

Definition 5.1.7 Cardinal B - Splines

Cardinal B - Slines is given by Nm (x) = l2-m+1 Nm(2x-k)


k=0

Result 5.1.8
For any function <j) e L (R), the following statements are equivalent,

i). {<(> (x - k): ke Z} is an orthonormal family in the sense that

<())(., - k), <t> C,-0> = S\ A MeZ


88
I 7 •• i
ii). The fourier transform <{> of <j) satisfies — fe~1JX | <|>(x) n dx = 8:(), j e Z.
2n -00
■*

X A
iii). The identity j 4»(x + Ink) |2 = 1 holds for almost all x.
k=^»

5.1.9 Orthonormal wavelets, Analysis of operators and Applications to


Numerical Analysis [3]
The term wavelet bases tends now to more and more things. An orthonormal
wavelet basis ofL2(9?) is of the form v|/j k = 2j/2v|/(2-ix-k) where j and k belong to

Z, and \\) is a smooth and well localized function. We can suppose that

| vj/a)(x) |< Cexp(-y | x |)

for a< N and a positive y. The decomposition of a function/in this basis is then local
at high frequencies (that is for large j's), since
f=Zcj,knk
fk
where

ci,k = jfnk

and the function vgj k has a numerical support essentially of size 2'J, so that the
wavelet coefficients for j large yield a local information on / around k2'7. An
extremely important feature of this decomposition is that it is performed on functions

that have cancellation. Actually, for £ < N,

|x^vj/(x)dx = 0

Let us sketch the proof when £ = 0. We choose a € R such that vp (a)? 0.


Let j > 0 and k = [2ja].

0=| vjy(x)vt)j k (x) dx = J vj/(a)i|/ j k (x) dx + O (J| x - a | |v|/j k (x) | dx)

Using the localization estimates of vji, we get


89
J| x — a 1 lM/j>k (x) | dx) < C 2'3j/2
so that

j y(a)¥j,k (x)dx (= 2'J/2 V(a)J V) * C 2_3j/2


Choosing j large, we get /vjy = 0. To prove that moments of higher order vanish, the

same proof works, but one has to subtract the Taylor expansion of vj/ in a, hence the
link between the smoothness of the wavelet and the number of vanishing moments.
This last property implies that <\i j k is well localized near the frequencies of order of

magnitude 2J and - 2J.

Theorem 5.1.10
Let 0 < a< 1. A function f is uniformly Ca if

|f(x)-f(y)l<C|x-y

Then, f is uniformly Ca if and only if

Proof
First, suppose that/is uniformly C“, Then
I cj k | = ||f(x)v)/j k(x)dx|
= ljf(x)-f(k2 J))Vj,k(x)dxl

Conversely, if | cj k | < C 2

90
-j+i
Let J be such that 2"/<|x-y|<2'J+1. Using the mean value theorem,

-+a J
Vjsk(x)-¥j;k(y)I^ZZC2 aj2j |x-y|Sup
II
jsj k
2 U
j<J k
(l+|x-k2’J |)2 ’(1+1y-k2"j |)2

The sum in k is bounded by a constant independant of j, and the sum in j is bounded


by

C|x-y|<C|x-y|“ Furthermore

£ £ 2"WI I Yjjc W - Vji (y) I * 112 U:+“ J


l¥j,k OOMVj,k (y)l
j>J k j>J k

2-aj < C2“aJ < C | x - y


< 2CXI-
“jT(l+|y-k2“j |)2

Lemma 5.1.11
The operator G is a self- adjoint positive definite operator, if (hn), defined by
h„= G‘,,2(en),

forms an orthonormal basis of H.

Proof
The operator G is self-adjoint positive because
<G(f)|g> = X(f7e11><g/en}.

Since the (en) are a Riesz basis,


CiE(f/e„>2 <||I(f/en>enf <C2X(f/en>2

so that
C,(G(f)/f><||G(f)|2£C2(G(f)/f);

hence G is positive definite and


C,Id <G <C2Id

where Cl and C2 are positive.


91
Let us prove that the (h„) forms an orthonormal basis of H. Since G is positive
definite, G'12 can be defined (as the only positive operator T such that T2 = G'1), and
it is an isomorphism on H. Hence, the G‘l/2(en) are a basis of H. We first check that
Vx s H,X(G-',2(en)/x)G-|'2(e1,) = x (5)

We have
G(y) = X(en/y)en

so that
Gl,2(Gl,2(y)) = X{G_l/2(en)/G,,2(y)) e„

and G1/2(y) = X{G-1,2(e„)/G1,2(y))G-1,2(e„)

Hence Equation (5) if we take x = G1/2(y) (which is possible since G12 is an


isomorphism). The G4/2(en} are a basis of H. If we take X = G“1/2(em) in Equation (5),
we- obtain

2{G-1'2(en)/G"1,2(en,))G-,,2(en) + (||G','2(em)||2 -l)G-,,2(em) = 0.


n^m

Since the (en) are a basis of H, this implies


{G-|,2(en)/G-|,2(em)) = 6nim

and the (hn) are an orthonormal basis of H.

5.1.12 Wavelets on arbitrary domains

We shall construct wavelets which are an orthonormal basis of L2(Q) and are

bases of the Sobolev spaces Hq(Q) .


Consider the space V) of functions that are C2m‘2, vanish outside Q, and are
polynomials of degree 2m - 1 in each variable in the cubes
K2'J+2’j[0,l]n, with keZn.

92
Let om(x1-m,......*n-m) = £*........ * 4(2m times), where is the characteristic

function of [0, l]n. The Fourier transform of am is

. 2m \2m
Sin /2 Sin 12
. Si/2 . . £„/2

These functions are the classical B-splines in dimension n. Let


Oj,k = 2n j/2 a ( 2J x - k),

and let Aj be the set of points k2“J such that supp (cij k )cQ. First, we have

Proposition 5.1.13
The (cr f k) j , are a Riesz basis of Vj.

The equivalence of norms in the definition of Riesz basis holds because it holds

on Rn, if we take all the cjjk, keZn (this is easily checked using the Fourier

transform). Let us sketch the proof of the completeness. If f is a function compactly

supported in Vj, its support has a comer. One easily checks that, on the cube of size

2 J located in that comer, f coincides, up to a multiplicative constant, with a function

Cj k. We substract this function and we can make the support of f shrink until

ultimately it has vanished.

In order to obtain an orthonormal basis of Vj,we use the orthonormalization

procedure. Since the Gram matrix of the Oj;k has exponential decay, Using the

following Lemma.

Lemma 5.1.14
Let G be a positive matrix in Ay such that

CiId<G<C2 Id.
Then 3 y >0 such that G '2 e AY where y' depends only on y, C, C', Q and C2.

We obtain the following localization for the functions q>j,k

| aa(pJ;k (x) |< C2ja2nj/2 exp (-y2J | x - k2~J |)

93
for | a |< 2m - 2, and a positive y.

Let us now describe how to obtain the wavelets. First we remark that a function
in Vj is determined by its values on the k2~J. More precisely, there exist Ci and C2
positive such that
\ 1/ 2

Ci ||f||2< Zifwi <C2 ||f||2.


j

This holds because it is equivalent to the fact that the positive self adjoint matrix

Gk>/ = om (k - i?)is invertible. As before we first check it fork € Zn, where it is

straightforward using the Fourier transform, and we then use the fact that, if

C,Id<G <C2Id,

this property is stable under restriction, with the same constants. Thus the constants
in the equivalence are independant of j. Let (Ljk)=G_1(cjjk). The LJik are called

cardinal splines because of the following property which makes them useful for
interpolation

LJ,k«’2'I) = 8M

The orthonormal basis of W, is obtained by projecting on Wj the functions Lj k

such that k 2 JeAj+1\A j and then by orthonormalizing the set thus obtained. Here

again
| 3aq>j>k(x) |< C2ja2nj/2 exp (-y2j | x - k2-j |)

for|a|<2m-2, and a positive y. This estimate holds because all the matrices of the

transforms performed on the bases in order to obtain the vpj k have entries

exponentially decreasing away from the diagonal, so that we can apply several times
Lemma 5.1.14.

Actually, though these wavelets are not the same as in the case Q = Rn, they are

94
almost the same; that is, numerically, only the wavelets that are close to the boundary

are modified. Thus, we can essentially keep the fast decomposition algorithms, with

only small modifications of the values of the associated filters near the boundary.

More precisely, there exists classical wavelets \j/j k on Rn such that

(-+a)j
l|da(v|/j,k -viJj)k)||00<C2 2 exp(-y2J d(k2~J,<3f2)).

We have that the wavelets constructed are bases of the Hq(Q) spaces and that the

following characterization holds


>2js
f II"2h5(Q) -II 'j.k
Section II

5.2 Scaling Functions in Wavelet Analysis

If we assume that <)> e V,, then for a suitable a e i2(Z), we will have

«t) = 5>j*(2t-j) (6)

such an equation is a two-scale relation. It is an example of a “Functional Equation”.

Lemma 5.2.1 [32]

Let <{> be an element of L (R) such that {4>0j : j e Z }is orthonormal and such

that the two-scale equation (6) is valid. Then for all n and m in Z,

I[a„.2j am_2j + (-1 )n+ma,_„+2j a,.mt2j] = 28„ (7)


j

Proof
If n+m is odd, the expression on the left in equation (7) becomes
S an-2j am-2j S ^l-n+2j ^l-m+2j
j j

In the first sum, let j=-i. In the second sum, let j = i + (n+m-l)/2. The result is
S **n+2i am+2i — S ain+2i an+2i — ® — ^nm
i i

If n+m is even, the expression on the left on equation (7) is

95
S an-2j am-2j al-n+2j al-m+2j
J j

In the first term, let j = -i. In the second sum let j — i + (m+n) 12, we get
S an+2i am+2i 2 am+2i+l an+2i+l — an+j am+j — S ai a
i+m-n
i i j

Since (<|)lk : k e Z}is orthonormal, the parseval equation applies to give

Yj ai ai+m-n = (X ai ^li’ X ai+m-n <Pli (8)


> \ > J

By the two-scale equation, we have


I I
Za1(f.li(t) = Sai22(}.(2t-i) = 22(|)(t)

Similarly,
\
2 _ 4>ii (t)=I ai+m_n 2 2 <|>(2t - i)
i l
1
= Za; 2z<t>(2t-j + m-n)
j

,2v
l
X/T (n ~ m)x •,
= 2Z Zaj <j)(2 (t-------—)-j)
i 2

_1_
= 22<t> fn J®
n-m
0.

Finally, the right side of the equation (8) becomes,

22a>o,0,22<I> n-m =25


nm

96
Lemma 5.2.2
Let 0 be an element of L2(R) such that {0Oj: j e Z }is orthonormal. Assume

that 0 satisfies the two-scale equation (6). Define ¥ by the equation


v(t)=X(-l)kai-k<t'(2t-k)
k
_1
then 4>ln =2‘2X[a„_2j4>0j +(-!)" a,.„t2j >Poj] (9)
j

Proof
By the two-scale equation d>(t) = ^ai<J>(2t - i)
i
Therefore, 0(t-j) = ZM>(2(t-j)-i)
i
_i 1
= £ 2 2aj220 (2t-2j-i)
i

In other terms 0Oj = = ^ 2 2 ai<J>1>2j+i ■ A similar equation holds for v|/, namely
i

VOj =S(-1)'2 2al-i^>l,2j+i


i

The expression on the right in equation (9) can now be written as


II
2^Z[a„-2jX 2'2ai<l.u+jj+(-l)”a1_n+2jX 2'2(-l)la1_iO,_i+2j]
I
j ' >
changing the summation index i to m by the equation m = i +2j, obtaining
2*'X X[a„-2ja„-2j +(-0"+ma1-n+2ja,_mt2j]®ln,
m j

According to the preceding lemma, the sum over j is 28nm. Hence the entire
expression reduces to 0ln.

97
Theorem 5.2.3 (Lawton) [10]
A necessary and sufficient condition that the wavelet system W(A) associated
to the wavelet matrix A be an orthonormal system is that 1 is a nondegenerate
eigenvalue of the matrix L.

Before we discuss the proof of this theorem, let us look at the three examples.
Example 5.2.4
For the case of Haar, the matrix is 1 x 1 and
L = Loo= ^Yoo =^.2= 1,

as desired, since the Haar wavelet system is clearly orthonormal.


Example 5.2.5
For the case of g = 2, we have the 5 x 5 generic Lawton matrix
/ = -2,...,2,j = -2,...,2.
(0 y3 0 0 0^
l 2 Yi 0 y3 0
L=~ 0 Yi 2 Yj 0
0 Ys 0 Yi 2
0 0 0 Y3

Substituting the values of Yj for AE , we find


f 1
0-000
2
10 0-0
2
L(E) 0 0 10 0
0-001
2
0 0 0 -0
V 2 j

with eigen values {-1,--,--, 1, 1}, and indeed, 1 is a repeated eigenvalue.


2 2

98
Example 5.2.6
For D2, we find

_ J_
0 0 0
16
_9_
1 0
16 16
_9 _9
L(D2) = 0 1
16 16
_ 9 9
0 0
16 8
_J_
0 0 0
16
with eigenvalues {1, 0.0582127, 0.156643, 0.398996, 1.07365}, and since 1 is a
nondegenerate eigenvalue in this case, we can use the above Theorem 5.2.3 to
conclude that the Daubechies system W(D2) is an orthonormal system. This method
works well for computable examples of Daubechies wavelet systems.

Proof of sufficient condition

12g—1
Let A
52g-l

be a wavelet matrix of rank 2, and let (p be the associated scaling function


satisfying

<P00= Zak<P(2x-h) (10)


h=0

Define p; := j <p(x)(p(x + i)dx, i e Z, which is a measure of the orthonormality of

W(A). We want to show that pi=5i0 under the hypothesis that 1 is a nondegenerate

eigenvalue of the Lawton matrix L.

By using (10), we can express the vector pin terms of itself. Namely,

r|j = J (p(x)q>(x +i) dx

99
^ a na(| (p(2x - k)tp(2x + 2i - ^)dx
kt

Z ~Zakak+2i-j
Tli
j V2 k

Then, the (4g - 3)-dimensional vector r| satisfies the equation


tl=Lr!,

that is, r| is an eigenvector of L with eigenvalue 1. Let 8 be the vector defined by


5i = 5i0

Then, we see that


L8 = 8
as well, since

(LS)i-X oZakak+2i-j
8jo
;,
k \.2‘k

Zakak+2i

and now using the quadratic condition Xak+mfak+m£ = m8s,s8w for a wavelet
k
matrix, we see that
(L6)i = 8i0

that is, LS = 8, as desired. Since we assume that 1 is a nondegenerate eigenvalue, it


follows that r| = C8, for some constant C.

Now, we need to show that C = 1. To do this, we use the Fourier transform to


compute r|j. Namely, recall that(p is real,

% = J <P(* Mx + *)& = j <pfe)<P-L


by Parseval's theorem for the Fourier transform. But

thus

100
■n< = Je 2m¥ I <p(£) I2 d£
which by a sequence of changes of variable (divide the real axis up into the union of
intervals of length 2 n) yields
2k ( A
cs,„ = e-2"« Ei<p(5+2*k)r
0 Vkez J
which implies that the periodic function

sa
£|tp(i; + 27tk)|2
kez

has the same Fourier coefficients as the constant C, and hence

XM5 + 2rf)|2=C
ie z
Recall that
Fo(OI2+|F„(5 + 7t)|2=l

and F0 (0) = 1, and hence F0 (tc) = 1, Now Sharathlar Unlvartlty


Library

A <»
<pfi)=nF„«/2") T421
n=l

by construction, so we can compute

(p(27ik) = cp(2.2€ (2r + 1)tc),

for some t >0,reZ. We find

q*27ck) = flF0 (2(+l~n (2m + 1)tc)F0 ((2r + 1)ti). J((2r +1)71),


11=1

and thus, since F0(£) is periodic of period 2 n, we see that

(p(27tk) = 0, k* 0, k e Z.

Using this as well as that <p(0) = 1 m j v|/jk(x)v|/jf(x)dx = 8^8^, we find that C = 1 as


desired.

101
Theorem 5.2.7
Let <|> be a scaling function governed by a finite two - scale relation

<Kx) = Xpk<K2x-k)> Po. Pn*o (ii)


k=0

as in «*) = I Pk*«2x - k), p*, p* * 0


k=0 *

with N = N«(, and = p*. Suppose that {<(>(.,-k): k g Z) is in orthonormal family that

constitutes a partition of unity and <j> is skew-symmetric in the sense that

<Ka + x) = <Ka-x), xeR (12)

for some aeR. Then <|) must be the first order cardinal B - Spline.

Proof
As Usual, let
p(z) = ^Xpkzk
1 k=0
i\v

and z = e 2 . Then (11) is equivalent to

,w. (13)
4>(w) = p(z)4)(—), wgR

on the other hand, assertion (12) is equivalent to

(j>(w)eiaw = <j)(w)eiaw , weR


(14)

Hence from (13) and (14), we have

«7>

for |z|=l. Since p(z) is a polynomial of degree N with non-zero leading coefficient
and nonzero constant term, it follows that 2a = N, So that,
p(z) = z~Np(z), | Z |= 1 (15)

102
Let us now consider the hypothesis that (<K,,-k): k e Z) is an orthonormal family. By

Result (5.1.7), this hypothesis is equivalent to

<J)(w + 27t:k) 1 (16)


z
k=-oo

so that an application of (13) yields


=

|p(z)|2 + |p(-z)|2 = 1, H=1 (17)

so by substituting (15) into (17), we have

(p(z))2 + (- 1)N (p(-z))2 = zN’ I z 1= 1 (18)

For <|> to be a function, it is necessary that N > 0.

Hence from (18) and the hypothesis PN * 0, we see that N must be an odd integer.

Now by setting

Pe(z)="ZP2kZk
1 k

Po(z) = ^ZP2k+lzk
1 k

we can write,

P(z) = Pc(z2) + zp0(z2) (19)

and the identity (17) yields

|Pe(z)|2 +|Po(z)|2 =^> M=1 (20)

Hence by applying (15) and (19), we obtain

Pe(z2) + zp0(z2) = p(z)


= zNp(z)

= zN[pe(z2) + zPo(z2)]

= zNPe(z2) + zN"1Po(z2X lzl=l


Since N is odd, equating the odd and even parts give,

103
pe(z2) = zN''p0(z2)
p0(z2) = zN'1p^?), z 1= 1

So that |pe(z)|2 = |p0(z)r z =1


Applying this identity to (20) gives rise to

|Pe(z)|2 =|Po(z)| 2 1
Z =1

This is not possible unless pe and p0 are monomials or


p(z) = ^(pe +pNzN), N is odd

Since (<«.,-k): k e Z) is a partition of unity, we have p(l) = 1 and p(-l) = 0, so that

1 + ZN
P(z)

and hence
iwN ^

1+e
<Kw)=n
k=l

-iwN

iwN
consequently by
j2m-l
Cot x
k=^o (x + 7ik)2m (2m -1)! dx 2m-l

We obtain

2„ . 2/WNx „
4Sin2(——)
I (|)(w + 2n k)
k
N2 k=-oo (w + 2k k)2

Sin2(^)
2
N2Sin2(—)
2

104
1 *i" Z
It follows form (16) that N=l, that is p0(z) = p(z) = -—or <j> is the first order cardinal

B - Spline Nj.

Theorem 5.2.8
Let pe W satisfying |p(z)|2 + |p(-z)|2 = 1, |z| = 1

N
and p(z) = ^£pkzk = (\ + z'l
S(z), where N is same positive integer and S e W
1 k v 4 y
satisfying S(l) - 1
For some N > 1 such that

Ilsk||k|E< oo

for some s > 0 and B < 2 N-l

S(z) = £Stzk
where the notation k
B: Max | S(z) |
!Z|=1

is used. Then infmte product

CO f i(0
------- 'N
g(©):=ripe 2k
k=! ^ y

converges to ge C(R)nL'(R) nL2(R) everywhere. Furthermore, the function

<(>eL2(R)with <)> = g is an orthogonal scaling function that generates an Multi-

Resolution Analysis of L2(R).

Section III
5.3 SPECIFIC TYPES OF WAVELETS

5.3.1 Multi - Resolution Analysis [16]


Let As be generated by the bases
S S

(Oks: 22(J>(2St-k);k e Z)and Ws by (vj/ks: 22 v|/(2st-k);keZ}.


105
In other words, any function xs(t) and ys(t) can be represented as the linear

combination of Oks(t) and T*ks(t) respectively. Observe that the function

and Ys_j(t) e Ws_,(t)are both derived from xs(t)eA. Therefore, we

should expect that the bases d\s.i of As_i and vFks_i of Ws.i should somehow be

related to the bases <J>k,s of As such a relationship with help in devising an algorithm

to obtain the functions xs.[ and ys.i from xsmore efficiently.

To achieve a multi-resolution analysis of a function as shown in figure 5.2. we

must have a finite-energy function <t>(t) eL ( R) called a scaling function, that

generates a nested sequence {Aj} namely


{0} <—........c A_ic Aoc Aic................—> L2

and satisfies a dilation (refinement) equation

4>(t) = IgoGO«at-k)
k
for some a>0 and coefficients {g0(k)}e f2. We will consider a = 2 which corresponds

to octave scales. Observe that the function <J>(t) is represented as a superposition of a

scaled and translated version of itself — hence the name scaling function. More

precisely, Ao is generated by {0(. - k): k € Z} and in general, As, by {Ok,s: k,s e Z}.

Consequently, we have the following two obvious results.

x(t) eAso x(2t) e As+1 (21)

x(t) e As o x(t + 2~s) e As (22)

There are many functions that generate nested sequences of subspaces. But the

properties (21) and (22) and the dilation equation are unique to MRA.

For each s, since As is a proper subspace of As+i there is some space left in

As, called Ws which when combined with As gives us As+1. This space {Ws} is called

the Wavelet subspace and is complementary to As in As+1, meaning that

As n Ws = {0}, s e Z (23)

Asews= As+] (24)

106
with the condition (23), the summation in (24) is referred to as a direct sum and the
decomposition in (24) as a direct-sum decomposition.
Subspaces (Ws) are generated by 'F(t)eL2 called the wavelet, in the same way as

(AJ is generated by <f>(t). In other words, any xs(t)eAs can be written as

*»(«) = Z»k,.«2‘t-k)
k

and any function ys(t) eWs can be written as

y,(0 = Zwk,,v(2’t-k)
k

for some coefficients {ak>s}keZ, {Wk s}keZ e £2

Since AS+1=WS©AS

= Ws0Ws,©A,1

= Ws ® Ws_, © Ws_2 ©.......


s-1
we have As = © We

•• •

Figure (5.1) splitting of MRA subspaces.

107
Observe that the {As} are nested while the {Ws} are mutually orthogonal.
Consequently, we have

As n Am = m>i

W<nWm = {0}, £*m

A^Wm = {0}, £<m

A Schematic representation of the hierarchical nature of As and Ws is shown in


Figure 5.1

24

108
16
- 16

Figure (5.2) multilevel representation of a function.

109
5.3.2 Wavelet Matrix [15]

5.3.2.1 Formulation of Principal Results and Examples

In this section we formulate the principal results concerning the use of wavelet
function for the harmonic analysis of functions in L2 (R) where L2 (R) denotes the

Hilbert space of square - integrable Lebesgue measurable functions defined on R,


equipped with the inner product

<f, g> := JRf(x)g(x)dx


for f, g e L2 (R).

Let A e WM (m, g; C) be a wavelet matrix, and consider the functional


difference equation
mg-l
<p(x)= Z a[!<p(mx-k). (25)
k=0
This equation is called the scaling equation associated with the wavelet matrix

A = (ak). If tp e L2 (R) is a solution of this equation, then cp is called a scaling

function associated with A. If (p is a scaling function for the wavelet matrix A, then
we define the wavelet functions {vp1,....... vjA1} associated with the wavelet matrix A

and the scaling function cp by the formula


mg-l
vps(x):= £ akcp(mx-k). (26)
k=0
If we want to indicate the dependence of {cp, ip1,....... ,vpm-1} on the wavelet matrix A,

we write

<pM v'M.... ■v”'1 [A],


but normally this will be clear in the context, and the explicit dependence on A will
not be indicated except to avoid ambiguity. Daubechies proved the following
theorem for rank 2, and we generalize it here to the case of arbitrary rank m.

110
Theorem S.3.2.2
Let A e WM (m, g; C) be a wavelet matrix. Then, there exists a unique
cp e L2 (R) such that:

(a) cp satisfies (25);

(b) lR <p(x) dx = 1;

f m ^
(c) supp cp c o,(g-0 + 1 .
m -1

Proof
It will depend on Fourier transform arguments, but for practical computations,
we can define an iterative procedure which will readily approximate a solution on a
computer. Let
<p°:= X[0.1]

be the characteristic function of the interval [0, 1), and define

mg-l
<Py(x):= Z ak(Py (rnx-k).
k=o
Then, this sequence will be a converge to the solution cp(x):= lim cpY(x). The
y—>oo

functions v|/s(x) can be computed by using (26). Thus, one obtains, for any wavelet

matrix A, a corresponding scaling function and wavelet functions.

Example 5.3.2.3 The Haar Functions


f1
If H =
l-l lJ
is the canonical Haar wavelet matrix of rank 2, then the scaling and wavelet functions
cpand vp given by Theorem 5.3.2.2 can be determined in an elementary fashion: no

iteration scheme is necessary. Their graphs are shown in Figures 5.3 and 5.4.

Ill
1.5

1■ -.............
0.5

0 ————.................................................
-0.5 i
I
-1 i*
ti
»
.1-5i j ^
Figure 5.3. The Haar scaling function for rank 2

------------ ---------—j------- ------

-0.5 0.5 1.5 2

Figure 5.4. The Haar wavelet function for rank 2.

Example S.3.2.4 Wavelet Matrix Asillt


For each m>l, there is a wavelet matrix Asinc which is not compact. This
example is important not only as a concrete illustration of the general theory, but also
because of its special relation to signal processing applications. Moreover, the scaling

Sin Ti x
function associated with the wavelet matrix Aanc is the function, Sin (nx) =--------- , it
nx
is universal in the sense that it is independent of rank m.
The sequence defined by the coefficients of the Laurent’s series,

y -t—L+ ll{z2k+1 + z-(2k+,)}


i+-itS2k
112
is the first row a of non compact wavelet matrix of rank 2. The second row p is

defined by

b0 = 1 and b2k. 1 = - a2JM. The wavelet matrix Asinc is

ASin

vtv
More generally, we can define a wavelet matrix of rank m and infinite genus which

will depend on a choice of a Haar wavelet matrix of rank m

A ■
^sinc , 0 < r < m, -oo<k<oo
Vak )

Theorem S.3.2.5

There exists a solution h to

(|>(x) = <|)(2x)+<t>(2x-l) (27)

on [0,oo) which is continuous on (0, go), bounded on [0, oo), and square-integrable

on [0, oo).

Proof

It is obvious that a continuous solution h to

h(t) = h(2t) + h(2t + l), te[0,oo) (28)

on (0, oo) is determined by its values on D n (0,co) (by density) and, moreover, the

values of h on D n [0, oo) are determined by its values on Z n [0, oo). Thus, it

suffices to find a solution to (28) on Z n [0, oo). We simply choose

fo n = 0,1
h(n) = < if 2k <n<2k + 2k_1 (29)
if 2k + 2k_1 < n < 2k+!

With these definitions, h(n) satisfies (28) on Z n [0, oo), and, hence, h(n) defines a

solution h of (28) on D n [0,go).

113
Theorem S.3.2.6
The function h defined by (29) has a unique continuous extension to (0, oo) as
a solution to (28) and moreover, the extension h is piecewise linear continuous on
(0, oo). In addition h is discontinuous at t = 0 but is bounded on [0, oo), and
h e L2 (0, oo), but hgL1 (0, oo).

Proof
For subintervals on R of the form (k, k +1], k>2,ke Z, one can verify that the
linear interpolation of the function values of h(n), n = 2, 3, ..., is a piecewise linear
function which is a solution to (28) on [2, <x>). For the interval [1,2], one must add the

. 5 ., ,,5 4
vertex point-, with h(-) , and then h / ti >2| is the Linear interpolation of the
4 4 8

three points h(l) = 0, h(-) , and h(2) = -. Define the function h / , by


4 4 2 l^J

h(x) = h(2x) + h(2x + 1). Since h (2x) and h (2x+ 1) are piecewise linear on [1,2] and
[2,3], respectively, we see that h / , will also be piecewise linear, and a continuous

extension of the solution to (28) for [1, oo) to [—,co). In a similar manner, one defines

h on[|,i], etc., and we obtain the desired piecewise linear continuous solution to

(28) on (0, oo). It is easy to verify the other assertions in the theorem, so the

verifications will be omitted.

We can now construct (p, a solution to (27), by defining q>R(x)2s\A. (pi(x) by


<|>R(x):=h(x-l), xeR (30)

<j>L(x) :=h(-x), xeL (31)

The functional equation will then determine the values of 4>T on 1 from its
values on L and R. Then, we obtain a solution to (27) which is in L2(R) but which

114
does not have compact support and cannot be the same as the classical compactly
supported Haar scaling function.

Such a (p is a generalization of the classical Haar function, and its graph (see
Figure 5.5) is reminiscent of the graph of sine nx [which is also in L2(R), but not in

Figure 5.5 A noncompactly supported solution of the Haar scaling equation

5.3.3 Special Wavelets


5.3.3.1 Sine Wavelets
The sine function is defined as usual by
sin x if x*0
sine x:=( x
if x = 0

Sine wavelet matrices have infinite genus and exist for every rank m, 1 < m < oo (cf.
Example 5.3.2.4). They arise naturally from the Whittaker-Shannon sampling
theorem which uses the sine function to interpolate a band-limited function
x-> f(x) in terms of its values on a lattice. When this formula is applied
to ffx) = sine 7i x itself, for any fixed m with 1 < m < co, one finds

115
Figure 5.6. The universal sine scaling function, genus = oo.

This equation shows that sine % x is universal of genus g = co, since the vector with

components

a° (m) := sine
(7dO

is a well-defined element of / (Z) and a well-defined scaling vector of infinite genus.


The corresponding wavelet matrices for such a scaling vector were discussed in more
detail in Example 5.3.2.4. They depend on the choice of a Haar matrix of rank m. For
rank 2, we have the following simple example of a sine wavelet vector (there are only
two choices differing in sign). Namely,

a1 =(....,(-l)ksinc(y),.....)

The graph of the sine scaling function is shown in Figure 5.6, and the graph of the
square of its modulus measured in decibels (dB) is shown in Figure 5.7. This form
will be familiar to every communication engineer.
The corresponding wavelet function is illustrated in Figure 5.8.
5.3.3.2 Haar Wavelets
For rank m = 2, the Haar wavelet basis corresponds to the wavelet matrix

fj1 n
-b

116
Figure 5.7. Logarithmic modulus of the sine scaling function.

Figure 5.8. A wavelet function of rank 2 and infinite genus corresponding to the
universal sine scaling function.

The Haar scaling function is a compactly supported solution of the corresponding


scaling equation
cp(x) = (p(2x) + (p(2x -1) (32)

The compactly supported solutions of (32) differ only at the endpoints of (0, 1). Up
to a normalizing factor (see Theorem 5.3.2.2 and Example 5.3.2.3), the general
solution is
cp(x) = cp(0) + cp(l) for0<x<l with cp (0), tp (1) arbitrary.

The Haar scaling function is defined by

if 0 < x < 1
cp(x) := (33)
otherwise

117
1.5,

1
0.5

-0.5

-1

~i,S -0.5 0 0.5 1 1.5 2

Figure 5.9. Standard Haar scaling function with compact support.

1.5

0.5

-0.5

-1

~1'5 -0.5 0 0.S 1 1.5 2

Figure 5.10. Standard Haar wavelet function with compact support.

Figure 5.9 illustrates the Haar scaling function on the interval from -1 to 2, and
Figure 5.10 shows the graph of the Haar wavelet function on the same interval. Haar
is a universal scaling function of genus g =1. Its scaling equation for rank m is
<poo- 2>(mx~k)
0<k<m

and it is easy to check that the Haar scaling function satisfies this equation for every

m, 1 < m < oo.

This corresponds to the scaling vector (1, 1 . . . , 1) of rank m for a Haar matrix of
rank m. We recall that the real Haar matrices of rank m are parameterized by
O(m-l), and thus for each choice of such a Haar matrix, we have wavelet functions

..........v|/m_1 associated with the universal Haar scaling function q>(x).

118
The Haar scaling function is integrable and square integrable, so we can
calculate its Fourier transform term by term in (33), obtaining

A ( i m-1 Aa
-1 ^ g27iik^/m
<P(0 = cp(^/m)
Vm k=0

Recall that this is the same as

(p(^) = Fo(^/m)(P(^/m)

where Fo(£, ) is the Fourier transform of the scaling vector (1, 1) that is,
i m-l
F„© = -Ze2^ (34)
mk-o
Taking the limit, we find

9(y = nFo(^n'n).
n=l

which is a special case of the infinite product formula used by Daubechies in the
proof of Theorem 5.3.2.2. But we can calculate the Fourier transform of (p explicitly

to find

<P(S) = ^ sine (tcS),

and thus we have an interesting infinite product formula for sine (x),

sine (7i£) = e_7t1^ ]”[F0 (£, / m11).


n=l

Recall that Fo is an explicit trigonometric polynomial given by (34), and thus we


have an explicit infinite product representation of sine (). For the case m- 2,

sine m =e-'*fWl
n=l Z
+ e2"'y2n)

e-^J-Je2n COS
n=l

119
e ^ne”4'2nricos«/2n)
n=I n=l
1 00
inncosW/2»)
n=l

J~[cos(7r£/2n).
n=l

Letting t, = ^, we obtain
— = J^[cos(7t/2n+1).
^ n=l
By repeated use of 2 cos2x = cos 2x+l we obtain a formula discovered by Vieta in

the sixteenth century:


J I--------- r==\
fv5Y\/2W2YV2 + V2W2
% 2 v2A
A

120

You might also like