Professional Documents
Culture Documents
Handout 20
Basics of Wavelets
The first theoretical results in wavelets are connected with continuous wavelet decompositions of L2 functions and go back to the early 1980s. Papers of Morlet et al. (1982) and Grossmann and Morlet (1985) were
among the first on this subject.
Let a,b (x), a R\{0}, b R be a family of functions defined as translations and re-scales of a single
function (x) L2 (R),
1
a,b (x) = p
|a|
Normalization by 1
|a|
xb
a
(1)
ensures that ||a,b (x)|| is independent of a and b. The function (called the
wavelet function or the mother wavelet) is assumed to satisfy the admissibility condition,
Z
|()|2
C =
d < ,
(2)
||
R
R
where () = R (x)eix dx is the Fourier transformation of (x). The admissibility condition (2)
implies
Z
0 = (0) =
(x)dx.
R
R
Also, if (x)dx = 0 and (1 + |x| )|(x)|dx < for some > 0, then C < .
Wavelet functions are usually normalized to have unit energy, i.e., ||a,b (x)|| = 1.
For any L2 function f (x), the continuous wavelet transformation is defined as a function of two variables
Z
CWT f (a, b) = hf, a,b i =
f (x)a,b (x)dx.
Here the dilation and translation parameters, a and b, respectively, vary continuously over R\{0} R.
Resolution of Identity. When the admissibility condition is satisfied, i.e., C < , it is possible to find the
inverse continuous transformation via the relation known as resolution of identity or Calderons reproducing
identity,
1
f (x) =
C
Z
R2
da db
.
a2
|()|2
d < ,
(3)
1
da db.
a2
(4)
CWT f (a, b), then g(x) = 1s f xs has the continuous wavelet transformation CWT g (a, b) = CWT f as , sb .
Both the shifting property and the scaling property are simple consequences of changing variables under
the integral sign.
Energy Conservation. From (4),
Z
Z Z
1
1
2
|CWT f (a, b)|2 2 da db.
|f (x)| dx =
C 0
a
Localization. Let f (x) = (x x0 ) be the Dirac pulse at the point x0 . Then, CWT f (a, b) =
1 ( x0 b ).
a
a
Reproducing Kernel Property. Define K(u, v; a, b) = hu,v , a,b i. Then, if F (u, v) is a continuous
wavelet transformation of f (x),
F (u, v) =
1
C
1
da db,
a2
i.e., K is a reproducing kernel. The associated reproducing kernel Hilbert space (RKHS)
defined as a
R isR
1
2
CWT image of L2 (R) the space of all complex-valued functions F on R for which C 0 |F (a, b)|2 daa2db
is finite.
R
Characterization of Regularity. Let (1 + |x|) |(x)| dx < and let (0) = 0. If f C (Holder space
with exponent ), then
|CWT f (a, b)| C|a|+1/2 .
Conversely, if a continuous and bounded function f satisfies (5), then f C.
Example 1.1 Mexican hat or Marrs wavelet. The function
(x) =
d2
2
2
[ex /2 ] = (1 x2 )ex /2
2
dx
(5)
a6
rrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr
r r r r r r r r r r r r r r r r r r r r r r r r r
r
r
r
rr
r
r
r
rr
r
r
r
r
rr
r
r
-
The continuous wavelet transformation of a function of one variable is a function of two variables.
Clearly, the transformation is redundant. To minimize the transformation one can select discrete values of
a and b and still have a transformation that is invertible. However, sampling that preserves all information
about the decomposed function cannot be coarser than the critical sampling.
The critical sampling (Fig. 1) defined by
a = 2j , b = k2j , j, k Z,
(6)
will produce the minimal basis. Any coarser sampling will not give a unique inverse transformation; that
is, the original function will not be uniquely recoverable. Moreover under mild conditions on the wavelet
function , such sampling produces an orthogonal basis {jk (x) = 2j/2 (2j x k), j, k Z}.
There are other discretization choices. For example, selecting a = 2j , b = k will lead to nondecimated (or stationary) wavelets. For more general sampling, given by
j
a = aj
0 , b = k b0 a0 , j, k Z, a0 > 1, b0 > 0,
(7)
numerically stable reconstructions are possible if the system {jk , j, k Z} constitutes a frame. Here
jk (x) =
j/2
a0
x k b0 aj
0
aj
0
!
j/2
= a0 (aj0 x k b0 ),
1.1
Multiresolution Analysis
A multiresolution analysis (MRA) is a sequence of closed subspaces Vn , n Z in L2 (R) such that they lie
in a containment hierarchy
V2 V1 V0 V1 V2 .
(8)
The nested spaces have an intersection that contains the zero function only and a union that is dense in L(R),
n Vj = {0}, j Vj = L2 (R).
[With A we denoted the closure of a set A]. The hierarchy (8) is constructed such that (i) V -spaces are
self-similar,
f (2j x) Vj iff f (x) V0 .
(9)
and (ii) there exists a scaling function V0 whose integer-translates span the space V0 ,
(
V0 =
f L2 (R)| f (x) =
)
ck (x k) ,
hk
2(2x k),
(10)
kZ
for some coefficients hk , k Z. This equation is called the scaling equation (or two-scale equation) and it
is fundamental in constructing, exploring, and utilizing wavelets.
ADD BASES FOR Vj .
ANY L2 can be projected on Vj
In the wavelet literature, the reader may encounter an indexing of the multiresolution subspaces, which
is the reverse of that in (8),
V2 V1 V0 V1 V2 .
(11)
FORM OF jk (x).
1
It is possible to relax the orthogonality requirement. It is sufficient to assume that the system of functions {( k), k Z}
constitutes a Riesz basis for V0 .
or, equivalently,
where () is Fourier transformation of ,
(0) = 1,
R (x)e
ix dx.
The coefficients hn in (10) are important in connecting the MRA to the theory of signal processing. The
(possibly infinite) vector h = {hn , n Z} will be called a wavelet filter. It is a low-pass (averaging) filter
as will become clear later by considerations in the Fourier domain.
To further explore properties of multiresolution analysis subspaces and their bases, we will often work
in the Fourier domain. Define the function m0 as follows:
1 X
1
m0 () =
hk eik = H().
2 kZ
2
(12)
The function in (12) is sometimes called the transfer function and it describes the behavior of the associated
filter h in the Fourier domain. Notice that the function m0 is periodic
with the period 2 and that the filter
taps {hn , n Z} are the Fourier coefficients of the function H() = 2 m0 ().
In the Fourier domain, the relation (10) becomes
() = m0
(13)
(x)eix dx
Z
X
=
2 hk
(2x k)eix dx
() =
Z
X hk
ik/2
e
=
(2x k)ei(2xk)/2 d(2x k)
2
k
X hk
eik/2
=
2
2
k
= m0
.
2
2
By iterating (13), one gets
() =
m0
n=1
2n
(14)
which is convergent under very mild conditions on rates of decay of the scaling function . There are
several sufficient conditions for convergence of the product in (14). For instance, the uniform convergence
5
on compact sets is assured if (i) m0 () = 1 and (ii) |m0 () 1| < C|| , for some positive C and . See
also Theorem 1.2.
Next, we prove two important properties of wavelet filters associated with an orthogonal multiresolution
analysis, normalization and orthogonality.
Normalization.
X
hk =
2.
(15)
kZ
Proof:
Z
X Z
(x)dx =
2
hk (2x k)dx
k
Z
X
1
=
2
hk
(2x k)d(2x k)
2
k
Z
2X
hk (x)dx.
=
2
k
Since
hk hk2l = l .
(16)
Proof: Notice first that from the scaling equation (10) it follows that
(x)(x l) =
X
hk (2x k)(x l)
2
(17)
X
X
hk (2x k) 2
hm (2(x l) m).
2
m
l = 2
hk
XX
k
"
X
m
1
hm
2
hk hm k,2l+m
hk hk2l .
h2k = 1.
(18)
One consequence of the orthogonality condition (16) is the following: the convolution of filter h with
itself, f = h ? h, is an a` trous.2
The fact that the system {( k), k Z} constitutes an orthonormal basis for V0 can be expressed in
the Fourier domain in terms of either () or m0 ().
(a) In terms of ():
|( + 2l)|2 = 1.
(19)
l=
By the [ PAR ] property of the Fourier transformation and the 2-periodicity of eik one has
Z
k =
=
=
(x)(x k)dx
Z
1
()()eik d
2 R
Z 2 X
1
|( + 2l)|2 eik d.
2 0
R
(20)
l=
The last line in (20) is the Fourier coefficient ak in the Fourier series decomposition of
f () =
|( + 2l)|2 .
l=
Remark 1.1 Utilizing the identity (19), any set of independent functions spanning V0 , {(x k), k Z},
can be orthogonalized in the Fourier domain. The orthonormal basis is generated by integer-shifts of the
function
F 1 qP
()
l= |(
+ 2l)|2
(21)
This normalization in the Fourier domain is used in constructing of some wavelet bases.
2
The attribute a` trous (Fr.) ( with holes) comes from the property f2n = n , i.e., each tap on even position in f is 0, except
the tap f0 . Such filters are also called half-band filters.
(b) In terms of m0 :
|m0 ()|2 + |m0 ( + )|2 = 1.
Since
l= |(2
(22)
(23)
l=
Now split the sum in (23) into two sums one with odd and the other with even indices, i.e.,
1 =
k=
k=
To simplify the above expression, we use relation (19) and the 2-periodicity of m0 ().
1 = |m0 ()|
|( + 2k)| + |m0 ( + )|
|(( + ) + 2k)|2
k=
k=
2
(24)
such that {jk (x), j-fixed, k Z} is an orthonormal basis of the difference space Wj = Vj+1 Vj . The
function (x) = 00 (x) is called a wavelet function or informally the mother wavelet.
Next, we detail the derivation of a wavelet function from the scaling function. Since (x) V1 (because
of the containment W0 V1 ), it can be represented as
(x) =
gk
2(2x k),
(25)
kZ
(26)
By mimicking what was done with m0 , we obtain the Fourier counterpart of (25),
() = m1 ( )( ).
2
2
(27)
(x)(x k)dx =
1
2
1
2
Z
()()eik d
Z
( + 2l)( + 2l)eik d.
l=
( + 2l)( + 2l) = 0.
l=
By taking into account the definitions of m0 and m1 , and by mimicking the derivation of (22), we find
m1 ()m0 () + m1 ( + )m0 ( + ) = 0.
(28)
(m1 (), m1 ( + ) ) = () m0 ( + ), m0 () .
(29)
(30)
is 2-periodic.
Any function () of the form ei S(2), where S is an L2 ([0, 2]), 2-periodic function, will satisfy
(28); however, only the functions for which |()| = 1 will define an orthogonal basis jk of L2 (R).
To summarize, we choose () such that
(i) () is 2-periodic,
(ii) () = ( + ), and
(iii) |()|2 = 1.
Standard choices for () are ei , ei , and ei ; however, any other function satisfying (i)-(iii) will
generate a valid m1 . We choose to define m1 () as
m1 () = ei m0 ( + ).
9
(31)
since it leads to a convenient and standard connection between the filters h and g.
The form of m1 and the equation (19) imply that {( k), k Z} is an orthonormal basis for W0 .
Since |m1 ()| = |m0 ( + )|, the orthogonality condition (22) can be rewritten as
|m0 ()|2 + |m1 ()|2 = 1.
(32)
(33)
In signal processing literature, the relation (33) is known as the quadrature mirror relation and the filters h
and g as quadrature mirror filters.
Remark 1.2 Choosing () = ei leads to the rarely used high-pass filter gn = (1)n1 h1n . It is
sometimes convenient to define gn as (1)n h1n+M , where M is a shift constant. Such re-indexing of g
affects only the shift-location of the wavelet function.
1.2
Haar Wavelets
In addition to their simplicity and formidable applicability, Haar wavelets have tremendous educational
value. Here we illustrate some of the relations discussed in the Section 1.1 using the Haar wavelet. We start
with (x) = 1(0 x 1) and pretend that everything else is unknown.
The scaling equation (10) is very simple for the Haar case. By inspection of simple graphs of two scaled
Haar wavelets (2x) and (2x + 1) stuck to each other, we conclude that the scaling equation is
(x) = (2x) + (2x 1)
1
1
= 2(2x) + 2(2x 1),
2
2
which yields the wavelet filter coefficients:
1
h0 = h1 = .
2
Now, the transfer functions become
10
(34)
1
m0 () =
2
1
ei0
2
1
+
2
1
ei1
2
1 + ei
.
2
and
m1 () = e
m0 ( + ) = e
1 1 i
e
2 2
=
ix
1 ei
.
2
ix
() =
1 ei/2 1 1 i/2
=
e
,
2
2
2
2
2
2
1.3
Daubechies was first to construct compactly supported orthogonal wavelets with a preassigned degree of
smoothness. Here we present the idea of Daubechies, omitting some technical details. Detailed treatment of
this topic can be found in the monograph Daubechies (1992),R Chapters 6 and 7.
Suppose that has N ( 2) vanishing moments, i.e., xn (x)dx = 0, n = 0, 1, . . . , N 1. The
m0 () has the form:
m0 () =
1 + ei
2
N
L(),
(35)
where L() is a trigonometric polynomial. Indeed, since has N vanishing moments, then (n) (0) =
0, n = 0, 1, . . . , N 1. By differentiating (27), we get
(n)
[m1 ()()]=0 = 0, n = 0, 1, . . . , N 1.
(n)
(n)
In terms of
N
M0 () = |m0 ()|2 = cos2
|L()|2 ,
2
the orthogonality condition (22) becomes
M0 () + M0 ( + ) = 1.
(36)
(y = sin2
).
2
(37)
By Bezouts result (outlined below), there exists a unique solution of the functional equation (37). It can
be found by the Euclidean algorithm since the polynomials (1 y)N and y N are relatively prime.
Lemma 1.1 (Bezout) If p1 and p2 are two polynomials of degree n1 and n2 , respectively, with no common
zeroes, then there exist unique polynomials q1 and q2 of degree n2 1 and n1 1, respectively, so that
p1 (x)q1 (x) + p2 (x)q2 (x) = 1.
For the proof of the lemma, we direct the reader to Daubechies ([?], 169-170). The unique solution of (37)
with degree deg(P (y)) N 1 is
N
1
X
k=0
N +k1 k
y , y = sin2 ,
k
2
(38)
and since it is positive for y [0, 1], it does not contradict the positivity of |L()|2 .
Remark: If the degree of a solution is not required to be minimal then any other polynomial Q(y) =
P (y) + y N R( 21 y) where R is an odd polynomial preserving the positivity of Q, will lead to a different
solution for m0 (). By choosing R 6= 0, one can generalize the standard Daubechies family, to construct
symmlets, complex Daubechies wavelets, coiflets, etc.
The function |m0 ()|2 is now completely determined. To finish the construction we have to find its
square root. A result of Riesz, known as the spectral factorization lemma, makes this possible.
Lemma 1.2 (Riesz) Let A be a positive trigonometric polynomial with the property A(x) = A(x). Then,
A is necessarily of the form
M
X
A(x) =
um cos mx.
m=1
P
imx such that |B(x)|2 =
In addition, there exists a polynomial B of the same order B(x) = M
m=1 vm e
A(x). If the coefficients um are real, then B can be chosen so that the coefficients vm are also real.
12
PA (z) =
N 1
1 X
a|k| z N 1+k .
2
(39)
k=1N
Since PA (z) = z 2N 2 PA ( z1 ), the zeroes of PA (z) appear in reciprocal pairs if real, and quadruples (zi , zi , zi1 , zi1 )
if complex. Without loss of generality we assume that zj , zj and rj lie outside the unit circle in the complex
plane. Of course, then zj1 , zj1 and rj1 lie inside the unit circle. The factorized polynomial PA can be
written as
PA (z) =
" I
#
Y
1
1
aN 1
(z ri )(z )
2
ri
i=1
J
Y
(z zj )(z zj )(z z 1 )(z z1 ) .
j
j
(40)
j=1
i=1
j=1
i=1
j=1
Y
Y
Y
Y
1
|aN 1 |
|ri1 |
|zj |2 | (z ri )
(z zj )(z zj )|2 .
2
Now, L() becomes
I
i=1
j=1
Y
Y
1
1
( |aN 1 |
|ri1 |
|zj |2 ) 2
2
|
J
I
Y
Y
(z ri )
(z zj )(z zj )|, z = ei ,
i=1
(41)
j=1
2
where the sign is chosen so that m0 (0) = L(0) = 1. Note that deg[P
A (z)] = deg[|L(z)| ] = N 1.
Finally, the coefficients h0 , h1 , . . . , h2N 1 in the polynomial 2 m0 () are the desired wavelet filter
coefficients.
13
Example 1.3 We
willfind m0 for N = 2.
P21
1cos
2
= 12 4 1 cos gives a0 = 4 and a1 = 1.
|L()| = k=0 2+k1
sin2 k
k
2 =1+2
2
The auxiliary polynomial PA is
1
1 X
a|k| z 1+k
2
PA (z) =
k=1
1
=
(1 + 4z z 2 )
2
1
= z (2 + 3) z (2 3) .
2
One square root from the above polynomial is
s
1
1
z (2 + 3)
(| 1|)
=
2
2+ 3
=
q
1
2 3 z (2 + 3)
2
1
( 3 1)z (1 + 3) .
2
The change in sign in the expression above is necessary, since the expression should have the value of 1 at
z = 1 or equivalently at = 0. Finally,
2
1
1 + ei
(1 3)ei + (1 + 3)
2
2
1
1 + 3 3 + 3 i 3 3 2i 1 3 3i
+ e
+ e
+ e
.
2
4 2
4 2
4 2
4 2
m0 () =
=
1.4
Regularity of Wavelets
There is at least continuum many different wavelet bases. An appealing property of wavelets is diversity
in their properties. One can construct wavelets with different smoothness, symmetry, oscillatory, support,
etc. properties. Sometimes the requirements can be conflicting since some of the properties are exclusive.
For example, there is no symmetric real-valued wavelet with a compact support. Similarly, there is no
C -wavelet function with an exponential decay, etc.
Scaling functions and wavelets can be constructed with desired degree of smoothness. The regularity
(smoothness) of wavelets is connected with the rate of decay of scaling functions and ultimately with the
number of vanishing moments of scaling and wavelet functions. For instance, the Haar wavelet has only the
zeroth vanishing moment (as a consequence of the admissibility condition) resulting in a discontinuous
wavelet function.
Theorem 1.2 is important in connecting the regularity of wavelets, the number of vanishing moments,
and the form of the transfer function m0 (). The proof is based on the Taylor series argument and the
scaling properties of wavelet functions. For details, see Daubechies (1992), pp 153155. Let
Z
Mk =
Z
k
x (x)dx and Nk =
14
xk (x)dx,
Table 1:
k
0
1
2
3
4
5
6
7
k
0
1
2
3
4
5
6
7
8
9
10
11
12
13
k
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
15
C1
, > N,
(1 + |x|)
and CN 1 (R), where the derivatives (k) (x) are bounded for k N 1.
Then, has N vanishing moments,
Nk = 0, 0 k N 1.
If, in addition,
|(x)|
C2
, >N
(1 + |x|)
m0 () =
1 + ei
2
N
L(),
(42)
C
,
(1 + |x|)
for k = 0, 1, . . . , r.
The requirement that possesses N vanishing moments can be expressed in terms of , m0 , or equivalently, in terms of the filter h.
Assume that a wavelet function (x) has N vanishing moments, i.e.,
Nk = 0, k = 0, 1, . . . , N 1.
(43)
dk ()
= 0, k = 0, 1, . . . , N 1,
d k =0
which implies
(k)
(k)
m1 () |=0 = m1 (0) = 0, k = 0, 1, . . . , N 1.
16
(44)
(k)
m0 () |= = m0 () = 0, k = 0, 1, . . . , N 1.
(45)
The argument is inductive. The case k = 0 follows from (0) = m1 (0)(0) [(27) evaluated at = 0]
and the fact that (0) = 1. Since 0 (0) = 12 m01 (0)(0) + 12 m1 (0)0 (0) it follows that m01 (0) = 0, as well.
(N 1)
Then, m1
nk g n =
nZ
(1)n nk hn = 0, k = 0, 1, . . . , N 1.
(46)
nZ
How smooth are the wavelets from the Daubechies family? There is an apparent trade-off between the
length of support and the regularity index of scaling functions. Daubechies (1988) and Daubechies and
Lagarias (1991, 1992), obtained regularity exponents for wavelets in the Daubechies family.
Let be the DAUBN scaling function. There are two popular measures of regularity of : Sobolev and
be the supremum of such that
Holder regularity exponents. Let N
Z
(1 + ||) |()|d < ,
and let N be the exponent of the Holder space CN to which the scaling function belongs.
and H
Table 2: Sobolev N
older N regularity exponents of Daubechies scaling functions.
N
N
1
0.5
2
1
0.550
3
1.415
0.915
4
1.775
1.275
5
2.096
1.596
6
2.388
1.888
7
2.658
2.158
8
2.914
2.415
9
3.161
2.661
10
3.402
2.902
log 3
ln N
=N 1
+ O(
).
2 log 2
N
From Table 2, we see that DAUB4 is the first differentiable wavelet, since > 1. More precise bounds
on N yield that from the DAUB3 family is, in fact, the first differentiable scaling function (3 = 1.0878),
even though it seams to have a peak at 1. See also Daubechies (1992), page 239, for the discussion.
,
Remark 1.3 the Sobolev and Holder regularities are related, thus, Theorem 1.3 holds for the exponent N
as well.
17
1.5
djk jk (x),
j,k
L
and this unique representation corresponds to a multiresolution
decomposition L2 (R) =
j= Wj . Also,
L
for any fixed j0 the decomposition L2 (R) = Vj0 j=j0 Wj corresponds to the representation
f (x) =
cj0 ,k j0 ,k (x) +
XX
(47)
jj0 k
|cj0 ,k | C,
1
|dj,k | C 2j(s+ 2 ) , j j0 , k Z .
(48)
j,k
Even the general (non-homogeneous) Besov spaces, can be characterized by moduli of the wavelet
coefficients of its elements. For a given r-regular MRA with r > max{, 1}, the following result (see
Meyer 1992, page 200) holds
Theorem 1.4 Let Ij be a set of indices so that {i , i Ij } constitutes an o.n. basis of the detail space Wj .
0
There exist two constants
P C C > 0 such that, for every exponent p [1, ], for each j Z and for
every element f (x) = iIj di i (x) in Wj ,
1/p
|di |p
iIj
18
C 0 ||f ||p .
The following characterization of Besov Bp,q spaces can be obtained directly from this result. If the MRA
has regularity r > s, then
bases for all 1 p, q , 0 < < r.
P wavelet bases are
PRieszP
The function f = k cj0 k j0 k (x) + jj0 k djk jk (x) belongs to Bp,q space if its wavelet coefficients satisfy
!1/p
p
|cj0 ,k |
< ,
and
1/p
j(+1/21/p)
p
2
|di |
, j j0
iIj
hP
j(+1/21/p) P
i1/q
p )1/p q
is an `q sequence, i.e,
< .
2
(
|d
|
jj0
k j,k
The results listed are concerned with global regularity. The local regularity of functions can also be
studied by inspecting the magnitudes of their wavelet coefficients. For more details, we direct the reader to
the work of Jaffard (1991) and Jaffard and Laurencot (1992).
j
j=1 dj 2 ). By dyad(x, n), we denote the subset of the first n digits from dyad(x), i.e., dyad(x, n) =
{d1 , d2 , . . . dn }.
Let h = (h0 , h1 , . . . , h2N 1 ) be the wavelet filter coefficients. We build two (2N 1) (2N 1)
matrices as:
19
(49)
Theorem 1.5
lim
(x)
(x + 1)
..
.
(50)
(x)
(x + 1)
...
...
(x + 2N 2) (x + 2N 2) . . .
(x)
(x + 1)
(x + 2N 2)
The convergence of ||Td1 Td2 Tdn Td1 Td2 Tdn+m || to zero, for fixed m, is exponential
and constructive, i.e., effective decreasing bounds on the error can be established.
Example 1.4 Consider the DAUB2 scaling function (N = 2). The corresponding filter is h =
According to (49) the matrices T0 and T1 are given as
T0 =
1+ 3
4
3 3
4
3+ 3
4
1 3
4
1+ 3
4
3 3
4
and
T1 =
3+ 3
4
1 3
4
1+ 3
4
3 3
4
3+ 3
4
1 3
4
1+ 3 3+ 3 3 3 1 3
,
,
,
.
4 2
4 2
4 2
4 2
Let us evaluate the scaling function at an arbitrary point, say x = 0.45. Twenty decimals in the dyadic
representation of 0.45 are dyad(0.45, 20) = { 0, 1, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1 }. In
addition to the value at 0.45, we get (for free) the values at 1.45 and 2.45 (the values 0.45, 1.45, and 2.45
are in the domain of , the interval [0,3]). The values (0.45), (1.45), and (2.45) may be approximated
as averages of the first, second, and third row, respectively in the matrix
The Daubechies-Lagarias algorithm gives only the values of the scaling function. In applications, most
of the evaluation needed involves the wavelet function. It turns out that another algorithm is unnecessary,
due to the following result.
Theorem 1.6 Let x be an arbitrary real number, let the wavelet be given by its filter coefficients, and let u
with 2N 1 be a vector defined as
u(x) = {(1)1b2xc hi+1b2xc , i = 0, . . . , 2N 2}.
If for some i the index i + 1 b2xc is negative or larger than 2N 1, then the corresponding component
of u is equal to 0.
Let the vector v be
20
v(x, n) =
1
10
2N 1
Ti ,
idyad({2x},n)
xk (x)dx = 0, for k = 0, 1, . . . , N 1.
(51)
As it was discussed in Section 1.1, some relevant properties of a multiresolution analysis can be expressed as relations involving coefficients of the filter h.
For example, the normalization property gave
2N
1
X
hi =
2,
i=0
(1)i ik hi = 0, k = 0, 1, . . . , N 1,
i=0
hi hi+2k = k , k = 0, 1, . . . , N 1.
i=0
That defines 2N + 1 equations with 2N unknowns; however the system is solvable since the equations
are not linearly independent. For example, the equation
h0 h1 + h2 h2N 1 = 0,
can be expressed as a linear combination of the others.
21
h0 + h1 + h2 + h3 = 2
2
h0 + h21 + h22 + h23 = 1
h1 + 2h2 3h3 = 0
,
h0 h2 + h1 h3 = 0
which has the familiar solution h0 =
For N = 4, the system is
1+ 3
, h1
4 2
3+ 3
, h2
4 2
3 3
,
4 2
and h3 =
1 3
.
4 2
h0 + h1 + h2 + h3 + h4 + h5 + h6 + h7 = 2
h20 + h21 + h22 + h23 + h24 + h25 + h26 + h27 = 1
h0 h1 + h2 h3 + h4 h5 + h6 h7 = 0
h0 h2 + h1 h3 + h2 h4 + h3 h5 + h4 h6 + h5 h7 = 0
h0 h4 + h1 h5 + h2 h6 + h3 h7 = 0
h0 h6 + h1 h7 = 0
0h0 1h1 + 2h2 3h3 + 4h4 5h5 + 6h6 7h7 = 0
0h0 1h1 + 4h2 9h3 + 16h4 25h5 + 36h6 49h7 = 0
0h0 1h1 + 8h2 27h3 + 64h4 125h5 + 216h6 343h7 = 0.
The above systems can easily be solved by a symbolic software package such as Maple or Mathematica.
1.6
Discrete wavelet transformations (DWT) are applied to the discrete data sets to produce discrete outputs.
Transforming signals and data vectors by DWT is a process that resembles the fast Fourier transformation
(FFT), the Fourier method applied to a set of discrete measurements.
Table 3: The analogy between Fourier and wavelet methods
Fourier
Fourier
Fourier Discrete
Methods Integrals
Series
Fourier Transformations
Wavelet
Continuous
Wavelet Discrete
Methods Wavelet Transformations Series
Wavelet Transformations
Discrete wavelet transformations map data from the time domain (the original or input data, signal
vector) to the wavelet domain. The result is a vector of the same size. Wavelet transformations are linear
and they can be defined by matrices of dimension n n if they are applied to inputs of size n. Depending on
boundary conditions, such matrices can be either orthogonal or close to orthogonal. When the matrix is
orthogonal, the corresponding transformation is a rotation in Rn space in which the signal vectors represent
coordinates of a single point. The coordinates of the point in the new, rotated space comprise the discrete
wavelet transformation of the original coordinates.
Example 1.6 Let the vector be {1, 2} and let M (1, 2) be the point in R2 with coordinates given by the data
vector. The rotation of the coordinate axes by an angle of 4 can be interpreted as a DWT in the Haar wavelet
basis. The rotation matrix is
!
1
1
cos 4
sin 4
2
2
W =
,
=
1
1
cos 4 sin 4
2
2
22
and the discrete wavelet transformation of (1, 2)0 is W (1, 2)0 = ( 32 , 12 )0 . Notice that the energy
(squared distance of the point from the origin) is preserved, 12 + 22 = ( 21 )2 + (
3 2
2 ) ,
since W is a rotation.
Example 1.7 Let y = (1, 0, 3, 2, 1, 0, 1, 2). If Haar wavelet is used, the values f (n) = yn , n =
0, 1, . . . , 7 are interpolated by the father wavelet, the vector represent the sampled piecewise constant function. It is obvious that such defined f belongs to Haars multiresolution space V0 .
The following matrix equation gives the connection between y and the wavelet coefficients (data in the
wavelet domain).
1
1
1
1
0
0
0
0
2
2
2
2
2
2
1
c00
1
1
1
1
0
0
0
0 2
2
0
2 2
2 2
d00
1
1
1
1
0
0
0
0
3 2 2
2
2 2
2
10
1
1
1
1
2
2 0
0
2
0
0
d11
2 2
.
= 212
1
1
1
1
d
0
0
0
0
20
2 2
2
2 2
2
0
1
1
1
1
21
0
0
0
2
2 2
2 2
2
1
d22
1
1
1
1
0
0
0
2 2
2
2 2
2
d23
2
1
1
1
0
0
0
2
2 2
2 2
2
The solution is
c00
d00
d10
d11
d20
d21
d22
d23
2
2
1
1
= 1
2
5
2
1
2
12
Thus,
f
(52)
1
1
1
1
1
2 2 + 1 + = 1/2 + 1/2 = 1 (= y0 ).
2
2 2
2 2
2
2
Performing wavelet transformations by multiplying the input vector with an appropriate orthogonal matrix is conceptually straightforward, but of limited practical value. Storing and manipulating transformation
matrices when inputs are long (> 2000) may not even be feasible.
In the context of image processing, Burt and Adelson (1982a,b) developed orthogonal and biorthogonal
pyramid algorithms. Pyramid or cascade procedures process an image at different scales, ranging from fine
to coarse, in a tree-like algorithm. The images can be denoised, enhanced or compressed by appropriate
scale-wise treatments.
23
Mallat (1989a,b) was the first to link wavelets, multiresolution analyses and cascade algorithms in a
formal way. Mallats cascade algorithm gives a constructive and efficient recipe for performing the discrete
wavelet transformation. It relates the wavelet coefficients from different levels in the transformation by
filtering with h and g. Mallats algorithm can be viewed as a wavelet counterpart of Danielson-Lanczos
algorithm in fast Fourier transformations.
It is convenient to link the original signal with the space coefficients from the space VJ , for some J. Such
link is exact for interpolating wavelets (Haar, Shannon, some biorthogonal and halfband-filter wavelets) and
close to exact for other wavelets, notably coiflets. Then, coarser smooth and complementing detail spaces
are (VJ1 , WJ1 ), (VJ2 , WJ2 ), etc. Decreasing the index in V -spaces is equivalent to coarsening the
approximation to the data.
By a straightforward substitution of indices in the scaling equations (10) and (25), one obtains
X
X
j1,l (x) =
hk2l jk (x) and j1,l (x) =
gk2l jk (x).
(53)
kZ
kZ
(54)
hk2l cj,k .
P
Similarly dj1,l = k gk2l cj,k .
The cascade algorithm works in the reverse direction as well. Coefficients in the next finer scale corresponding to Vj can be obtained from the coefficients corresponding to Vj1 and Wj1 . The relation
cj,k = hvj , j,k i
X
X
=
cj1,l hj1,l , j,k i +
dj1,l hj1,l , j,k i
l
X
l
cj1,l hk2l +
X
l
dj1,l gk2l ,
(55)
3
X
hk 2(2n k).
k=0
3 , 3 3 , 3+ 3 , 1 3 }.
= { 1+
4 2 4 2 4 2 4 2
Since (0) = 2h0 (0) and 2h0 6= 1, it follows that (0) = 0. Also, (3) = 0. For (1) and (2)
we obtain the system
(1)
h1 h0
(1)
= 2
.
h3 h2
(2)
(2)
P
From k (x k) = 1 it follows that (1) + (2) = 1. Solving for (1) and (2) we obtain
1+ 3
1 3
(1) =
and (2) =
.
2
2
Now, one can refine ,
1
2+ 3
hk 2(1 k) = h0 2(1) =
=
,
2
4
k
X
3
hk 2(3 k) = h1 2(2) + h2 2(1)
=
2
k
3+ 3 1 3 3 3 1+ 3
=
= 0,
4
2
4
2
5
2 3
hk 2(5 k) = h3 2(2) =
=
,
2
4
k
or ,
(1) = (2) = 0,
X
1
1
=
gk 2(1 k) = h1 2(1) = , [gn = (1)n h1n ]
2
4
k
X
3
= h2 2(1) =
,
4
etc.
Let the length of the input signal be 2J , and let h = {hs , s Z} be the wavelet filter and let N be an
appropriately chosen constant.
Denote by Hk is a matrix of size (2Jk 2Jk+1 ), k = 1, . . . with entries
hs , s = (N 1) + (i 1) 2(j 1) modulo 2Jk+1 ,
(56)
Hk
The matrix
is a basis-change matrix in 2Jk+1 dimensional space; consequently, it is unitary.
Gk
Therefore,
Hk
0
0
I2Jk = [Hk Gk ]
= Hk0 Hk + G0k Gk .
Gk
and
I=
Hk
Gk
[Hk0
G0k ]
Hk Hk0 Hk G0k
Gk Hk0 Gk G0k
That implies,
Hk Hk0 = I, Gk G0k = I, Gk Hk0 = Hk G0k = 0, and Hk0 Hk + G0k Gk = I.
Now, for a sequence y the J-step wavelet transformation is d = WJ y, where
H2
H1
H1
,
W1 =
, W2 = G2
G1
G1
H3
H
2
G3
H1
,...
W3 =
G2
G1
h1 h2 h3 0 0 0 0 h0
0 h0 h1 h2 h3 0 0 0
H1 =
0 0 0 h0 h1 h2 h3 0
h3 0 0 0 0 h0 h1 h2
h2 h1 h0 0
0
0
0
h3
0
h3 h2 h1 h0 0
0
0
.
G1 =
0
0
0
h3 h2 h1 h0 0
h0 0
0
0
0
h3 h2 h1
26
1+ 3 3+ 3 3 3 1 3
,
,
,
.
4 2
4 2
4 2
4 2
Since,
H1 y = {2.19067, 2.19067, 1.67303, 1.15539}
G1 y = {0.96593, 1.86250, 0.96593, 0.96593}.
W1 y = {2.19067, 2.19067, 1.67303, 1.15539 | 0.96593, 1.86250, 0.96593, 0.96593}.
H2 =
h1 h2 h3 h0
h3 h0 h1 h2
G2 =
h2 h1 h0 h3
h0 h3 h2 h1
In this example, due to lengths of the filter and data, we can perform discrete wavelet transformation for two
steps only, W1 and W2 .
The two-step DAUB2 discrete wavelet transformation of y is
W2 y = {1.68301, 0.31699 | 3.28109, 0.18301 | 0.96593, 1.86250, 0.96593, 0.96593}, because
H2 H1 y = H2 {2.19067, 2.19067, 1.67303, 1.15539}
= {1.68301, 0.31699}
G2 H1 y = G1 {2.19067, 2.19067, 1.67303, 1.15539}
= {3.28109, 0.18301}.
For quadrature mirror wavelet filters h and g, we define recursively up-sampled filters h[r] and g[r]
h[0] = h, g[0] = g
h[r] = [ 2] h[r1] , g[r] = [ 2] g[r1] .
In practice, the dilated filter h[r] is obtained by inserting zeroes between the taps in h[r1] . Let H[r] and G[r]
be convolution operators with filters h[r] and h[r] , respectively. A non-decimated wavelet transformation,
NDWT, is defined as a sequential application of operators (convolutions) H[j] and G[j] on a given time
series.
Definition 1.2 Let a(J) = c(J) and
a(j1) = H[Jj] a(j) ,
b(j1) = G[Jj] a(j) .
The non-decimated wavelet transformation of c(J) is b(J1) , b(J2) , . . . , b(Jj) , a(Jj) , for some j
{1, 2, . . . , J} the depth of the transformation.
If the length of an input vector c(J) is 2J , then for any 0 m < J, a(m) and b(m) are of the same
length. Let j (x) = j,0 (x) and j (x) = j,0 (s). If the measurement sequence c(J) is associated with the
P (J)
function f (x) = k ck J (x 2J k) then the kth coordinate of b(j) is equal to
Z
bjk = j (x 2J k)f (x)dx.
Thus, the coefficient bjk provides information at scale 2Jj and location k. One canthink of a nondecimated
j
wavelet transformation as sampled continuous wavelet transformation hf (x), 1a xb
a i for a = 2 , and
b = k.
27