This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
(� + 211') + 4>(� - 211') and of 4>(�/2) for the Meyer multiresolution analysis; their product is 1 t,b (O I . (See also Figure 4.2.)
1 40
CHAPTER
and
� I Ck l 2
=
(21l") -1
5
127< d1" I � Ck e -ike l 2 ,
(5. 3 .1) is equivalent to 0 < (21l") - 1 A :::; Li I¢(� + 21l"£) 1 2 :::; (21l") -1 B < 00 a.e. We can therefore define ¢# E L 2 (1R) by so that
4># ( { )
�
(2. ) -' 1 '
12
(
[�
1 4> ( U 2 '£) l
f
' I'
(5. 3 . 2) (5. 3 . 3)
4> ( { ) .
¢ (. . ¢(
Clearly, L i I¢# � + 21l"£) = ( 21l")-1 a.e., which means that the # - k) are orthonormal. On the other hand, the space vt spanned by the # - k) is given by
vt
{I; I
}
� I!! ¢# ( - n) , ( f!! ) nEZ E £2 (Z) {f; i = v ¢# with v 21l"- periodic, v E L2 ( [0 , 21l"] )} { f; i = ¢ with 21l"- periodic, E L2 ([0, 21l"])} ( use (5. 3 . 2 ) and (5. 3 . 3 )) = � In ¢( . - n ) with ( fn) nEZ E £2 (Z) Vo ( since the ¢( . - n ) are a Riesz basis for Vo ) .
{I; I
=
'
VI
VI
VI
}
5.3.2. Using the scaling function as a starting point. As described in and a multiresolution analysis consists of a ladder of spaces a special function E Vo such that are satisfied ( with possibly relaxed as in One can also try to start the construction from an appropriate choice for the scaling function after all, Vo can be constructed from the - k) , and from there, all the other can be generated. This strategy is followed in many examples. More precisely, we choose such that
§5.1,
¢( .
¢ §5. 3 .1).
(5.1.1)- (5.1.6) l'J¢: ¢ ¢(x) Ln Cn ¢(2x - n) ,
(l'J )j E(5.1.6) Z (5.3.4)
(5. 3 . 5 ) I¢(� + 21l"£ W :::; f3 < fELZ We then define l'J to be the closed subspace spanned by the ¢j , k , k E Z, with ¢j, k (X) = 2-j /2 ¢(2 -j x - k) . The conditions (5. 3 . 4) and (5. 3 . 5 ) are neces sary and sufficient to ensure that {¢j , k ; k E Z} is a Riesz basis in each l'J, and o
< a :::;
00 .
141
MULTIRESOLUTION ANALYSIS
that the '0 satisfy the "ladder property" ( 5. 1 . 1 ) . It follows that the '0 satisfy ( 5.1.1 ) , ( 5.1.4 ) , (5.1.5), and (5. 1.6) ; in order to make sure that we have a mul tiresolution analysis we need to check whether 5.1. and (5. 1.3) hold. This is the purpose of the following two propositions. PROPOSITION 5 . 3 . 1 . Suppose E satisfies ( 5.3.5 ) , and define '0 = Span E ll} . Then = '0
{¢j, k ; k
nj E z
¢ L2 (JR) {o}.
( 2)
Proof.
Vo. ¢O, k Vo, 1 Vo, 1 1 1 2 ::; kELZ l (f, ¢o, k )1 2 1 1 1 2 (see Preliminaries) . Since '0 and the ¢j , k are the images of Vo and the ¢O , k -j /2 1 (2 -j x), it follows that, for all 2 under the unitary map (Dj f )(x) 1 E '0 , A 1 1 1 2 ::; L l ( f, ¢j , k )1 2 ::; B 1 1 1 2 , ( 5.3.7 ) kEZ
constitute a Riesz basis for In particular, they 1. By (5.3.5) , the constitute a frame for i.e., there exist A > 0, B < 00 so that, for all E A ( 5.3.6 ) ::; B =
with the same A, B as in ( 5.3.6 ) .
2. Now take 1 E nj Ez '0 . Pick E > 0 arbitrarily small. There exists a compactly supported and continuous i so that I l f -i l £ 2 ::; E. If we denote by Pj the orthogonal projection on Vj , then hence
I l f l ::; E + I l Pj il
M ¢> mo, nI l -+ oo
mo,¢> ¢>. 7
0,
¢>. 6
1(0) 0 L2 (JR), 12 L (JR),
3. If is bounded, and continuous in then the condition =I- is nec essary in Proposition 5.3.2. This can be seen as follows. Take E then If Uj E z Vj = =Iwith support C [-R, RJ , R < = lim J --+ oo But
11 0,
00.
p- J I · j
k
Since 1 is continuous, the first term tends to AThe- I second 27r 1 1(0)term J-+ 00 , byexactly the dominated convergence theorem. 1 2 1 1can1 2 forbe bounded in (5.3. 15), so that this term tends to zero for J -+00. It follows that as in (5.3. 13) .
as
I l f l =I- 0, this implies 1(0) =I- O. 4. The argument in points 3 and 4 of the proof can also be used to prove 1 1(0) 1 2 ::; B/27r. We have indeed B I II 1 2 � B I P_ J II 1 2 � kELZ 1 (1, ¢>- J, k )1 2 = 27r f �1 1 ( T J �) 1 2 I j (�) 1 2 + R, where I R I can be bounded by C 2 - J for nice I. The other term tends to 27r 1 1(0)2 l 2 1 1 1 2 (see 4) . Together with remark 3 above, this implies A/27r ::; 1 1(0)1 ::; B/27r. In particular, if the ¢>O, k are orthonormal, then A = B and 1 1 (0) 1 = (27r) - 1 / 2 . 5. The conditions 1 E L OO , 1(0) =I- 0 (with 1 continuous in 0) imply certain restrictions on the as well. Equation (5.3.4) can be rewritten as (5.3. 18 ) 1(�) = mo(�/2) 1(�/2) , Since
en
145
MULTIRESOLUTION ANALYSIS
n
n
with mo (�) = � L Cn ei ( In particular, ¢(O) implies mo (O) = 1 ( since ¢(O) =f. 0) or
mo(O) ¢(O), which
n
(5.3. 19)
Moreover, (5.3. 18) implies that mo is continuous, except possibly near the zeros of ¢. In particular, mo is continuous in � O. If, further more, I ¢(�) I :S C ( + I W - 1 / 2 - ., then the continuity of ¢ implies that L I I ¢(� + 27r£) 12 is continuous as well, so that ¢# ( as defined in §5.3.1) is also continuous; consequently, mt (�) = ¢# (2�)/¢# (�) satisfies mt ( O) = 1. Since Imt (0 12 + I mt (� + 7r) 12 = 1 , it follows that mt (7r) O. This implies mO (7r) = 0 (mt (�) = mO (�) [LI I ¢(� + 27r£) 12] 1 /2 . [LI I¢(2� + 27r£) 12t1/2) , or (5.3.20) Cn ( t = o.
l
=
=
Ln
n
-
l
n
n
Together with L Cn = 2, this implies L n C2 = 1 L C2n+ l . This is consistent with the admissibility condition for 1jJ. 8 Note also that Ln C2n = 1 = L C2nH is equivalent with Micchelli (1991)'s condition L l ¢(x-£) const. =f. 0 if 1¢(x) 1 :S C (1 + I x l ) - l - . and if ¢ is continuous.9 0
n
=
=
All this suggests the following strategy for the construction of new orthonor mal wavelet bases: •
Choose ¢ so that (1) ¢ and ¢ have reasonable decay, (2) (5.3.4) and (5.3.5) are satisfied, (3) J dx ¢ (x ) =f. 0
( by Propositions 5.3.1, 5.3.2 the V; then constitute a multiresolution anal ysis ) ;
•
•
If necessary, perform the "orthonormalization trick"
Finally, ;j; (�) = eie /2mt(� / 2 + 7r)¢# (� / 2), with mt (�) [LI I ¢(� + 27r£) I 2 P / 2 [LI I ¢(2� + 27r£) I 2 t 1 /2 , or equivalently 1jJ ( x )
=
Ln ( _ l) n h"!.nH ¢# (2x
-
n) ,
=
mo(�)
CHAPTER 5
146 5.4.
More examples: The Battl�Lemarie family.
The Battle-Lemarie wavelets are associated with multiresolution analysis ladders consisting of spline function spaces; in each case we take a B-spline with knots at the integers as the original scaling function. If we choose ¢ to be the piecewise constant spline, A. ( ) 0 ::; x ::; 1 , 0 otherwise, 'I' then we end up with the Haar basis. The next example is the piecewise linear spline,
X =
¢ (x)
=
{ I,
{ 01 - l x i ,
0 ::; Ixl ::; 1,
otherwise,
plotted in Figure 5.4a. This ¢ satisfies
¢ (x)
=
� ¢(2x
+ 1) + ¢(2x) + � ¢(2x - 1) ;
see Figure 5.4b. Its Fourier transform is
� 3
+
1 3
( sine/2e/2 )2 '
cos 0, this implies that there exists a, possibly smaller than ",(, so that ReG(�) :::: a/2 for 1 1m � I < a. Consequently, G -1/2 can be defined as an analytic function on 1 1m �I < a, which means that ¢# = G - 1 / 2 ¢ has an extension to a uniformly bounded analytic function on the strip 1 1m � I < a. 4. On the other hand, (5.4. 1 ) implies that
(6
(6
6
for 16 1 ::; a. It follows that on 1 1m � I bounded by Consequently,
1 + 1, and suppose that f E c , with f (l) i 1 for C :S Then (5. 5 .1) J dx xi j(x) 0 fod 0, 1 ,
such that
=
=
0:
m.
m
=
=
"
"
m .
Proof. The idea of the proof is very simple. Choose j, k, j', k' so that is rather spread out, and very much concentrated. (For this expository point only, we assume that has compact support.) On the tiny support of the slice of "seen" by can be replaced by its Taylor series, with as many terms as are well defined. Since, however, J dx = 0, this implies that the integral of the product of and a polynomial of order m is zero. We can then vary the locations of as given by k'. For each location the argument can be repeated, leading to a whole family of different polynomials of order m which all give zero integral when multiplied with 1. This leads to the desired moment condition. But let us be more precise as follows.
1.
fj, k
ij ' , k' i
h,k
ij ' , k'
h, k (X) i ', k'(x) i i ' , ' ,j jk
ij , , k'
prove (5.5.1) by induction on C. The following argument works for both 2 . We the initial step and the inductive step. Assume J dx x n i( x ) 0 for n E N, nis continuous C. (If C 0, then this amounts to no assumption at all.) Since f (i ) ( C :S ) and since the dyadic rationals 2 - j k , (j, k E Z) are dense in JR, there exist J, K so that f ( i ) (2 - J K) -=I- O. (Otherwise f ( i ) 0 would follow, implying f constant if C 0 or 1, which we know not to be the case, or, if C � 2, f polynomial of order C - 1 � 1, which would imply that f is not bounded and is therefore also excluded.) Moreover, for any t > 0 there exists 8 > 0 so that I f(X) - t,(n! ) -l f(n) (2-JK) (x - r JKt l t lx - 2-JKli if Ix - 2 - J KI :S 8. Take now j > J, j > O. Then o J dx f (x) j(2j x - 2j - J K) =
- I , n )
=
=
1.
Ln gn ¢> - I ,n ,
(_l) n h_ n+ l ' Consequently, 2 - j /2 ,¢(2 - j x - k) Tj /2 L gn 2 1 /2 ¢>(Tj +l x n n L gn- 2k ¢>j - l , n ( X ) . n
2k
-
n)
(5. 6 .1)
It follows that
n
i.e. , the (I, '¢ 1 , k ) are obtained by convolving the sequence ( (I, ¢>O , n ) )h E Z with g n ) n E Z , and then retaining only the even samples. Similarly, we have
(
-
(5. 6. 2)
n
which can be used to compute the (I, ,¢j , k ) by means of the same operation ( convolution with g, decimation by factor from the (I, ¢>j - l , k ) , if these are
2)
157
MULTIRESOLUTION ANALYSIS
known. But, by
(5.1.15) 0 (a necessary condition to have some regularity for 1/J). Not every such rno is asso ciated to an orthonormal wavelet basis, however, an issue addressed in §§6.2 and 6.3. The main results of these two sections are summarized in Theorem 6.3.6, at the end of §6.3. Section 6.4 contains examples of compactly supported wavelets generating orthonormal bases. The orthonormal wavelet bases thus obtained cannot, in general, be written in a closed analytic form. Their graph can be computed with arbitrarily high precision, via an algorithm that I call the "cas cade algorithm," which is in fact a "refinement scheme" as used in computer aided design. All this is discussed in §6.5. A lot of this material goes back to Daubechies (1988b); for many of the re sults, better, simpler, or more general proofs have been found since, and I have given preference to these new ways of looking at things. These different ap proaches are borrowed mainly from Mallat Cohen ( 1990), Lawton (1 9 , 1991), Meyer and Cohen, Daubechies, and Feauveau for the link with refinement equations the references are Cavaretta, Dahmen, and Micchelli and Dyn and Levin as well as earlier papers by these authors (see §6.5).
20
(1991)
(1990),
(1989),
(1992) ;
(1990),
90
6.1. Construction of mo .
In this chapter we are mainly interested in constructing compactly supported wavelets 1/J. The easiest way to ensure compact support for the wavelet 1/J is to choose the scaling function ¢ with compact support (in its orthogonalized version) . It then follows from the definition of the hn ' hn = v'2
J dx ¢(x) ¢(2x 167
n
)
,
168
CHAPTER
hn
6
that only finitely many are nonzero, so that 1/J reduces to a finite linear combination of compactly supported functions (see (5. 1.34)) , and therefore au tomatically has compact support itself. Choosing both and 1/J with compact support also has the advantage that the corresponding subband filtering scheme (see uses only FIR filters. For compactly supported the 211"-periodic function mo ,
¢
§5. 6)
¢
becomes a trigonometric polynomial. As shown in Chapter orthonormality of the implies
¢O , n
I mo ( e W
+ I mo (e + 1I"W = 1 ,
5 (see (5.1.20)), (6.1.1 )
where we have dropped the "almost everywhere" because mo is necessarily con tinuous, so that has to hold for all e if it holds a.e. We are also interested in making 1/J and ¢ reasonably regular. By Corol lary 5.5.4, this means that mo should be of the form
(6.1.1)
(6.1. 2 ) 1,
with N :::: and £. a trigonometric polynomial. Note that even without reg ularity constraint, we need with N at least P utting together, it follows that we are looking for
1. 1
(6.1. 2)
(6.1.1), (6.1. 2) (6.1. 3)
Mo ( e ) = Imo ( e W ,
a polynomial in cos e , satisfying Mo ( e )
and
(6.1.4)
+ Mo (e + 11") = 1
(6.1. 5)
where L( e ) = 1 .c(e W is also a polynomial in cos e. For our purpose it is conve nient to rewrite L( e ) as a polynomial in sin2 e /2 = ( 1 - cos e / ,
)2
(6.1. 6) (6.1. 4 )
In terms of P, the constraint becomes ( 1 - y) N P(y) yN P(l - y) = which should hold for all y E we use Bezout's theorem. 2
+
1,
(6.1. 7)
[0, 1]' hence for all y E JR. To solve (6.1.7) for P
169
COMPACTLY SUPPORTED WAVELETS
P I , P2
n l , n2 ,
THEOREM 6 . 1 . 1 . If are two polynomials, of degree respec tively, with no common zeros, then there exist unique polynomials q l , q2 , of degree - 1 , - 1, respectively, so that
n2
nl
(6.1.8)
Proof. 1 . We first prove existence; uniqueness follows later. We can assume that ( by renumbering, if necessary ) . Since degree � degree (pd,
nwel ::::cann2find polynomials a (x), b (x), with degree (a (P) 2 ) degree (P ) I 2 2 2 degree (P2 ), degree (b2 ) degree (P2 ), so that PI (X) a2 (x) P2 (X) + b2 (x) . 2. Similarly, we can find a3 (x), b3 (x), with degree (a3 ) -degree (b2 ) , degree (b3 ) degree (b2 ) , so that P2 (X) = a3 (x) b2 (x) + b3 (x) . We keep going with this procedure, with bn - l taking the role of P2 in this last equation, and bn the role of b2 , =
0 . This proves that (6. 3 . 2 ) is satisfied, and finishes the proof ( 1) (2). 4. We now prove the converse, (2) (1). Define f.L k (�) (271") - 1 /2 [n;=l mo(2 -j �)] . XK (2- k�), where XK is the indicator function of K, X K (�) 1 if � E K, 0 otherwise. Since K contains a neighborhood of 0, f.L k -t¢ pointwise for k-too. Since
2
2
'
:::}
:::}
=
=
185
COMPACTLY SUPPORTED WAVELETS
I mo(2- k �) 1 Cmo(�) k 1 andC'I� E K.henceOn Ithe other m o(�) mo(O) �, I � I ; 1 ::; 1 C'k I � I . ko. ko so that 2 - k C'I � 1 � if � E K e - 2x ::; ::; � , we find therefore, for ko I ¢(�)I (2rr) - 1 /2 k=II I mo( T k �) 1 k=kII+ I mo(2 - k �) 1 1 o1 > (2rr) - 1 / 2 C ko II exp [-2C' 2 - k l � 1 l k=ko+ 1 > (2rr) - 1 / 2 C ko exp [ _C' 2 - kO + 1 max I � I ] c" > 0 . t; EK
5. By assumption, � > 0 for hand, we also have, for any 1Since K is bounded we can find and � Using 1 - x � for 0 � E K,
�
�
O n . ( This is illustrated by the counterexample mo (O � ( 1 + e- 3ie) discussed above. ) The points � ± i play a special role for the following reason: + 2k 7l" ) 0 for all k E Z, contradicting (6.2.5 ). This mo (± i) 0 implies implication can be checked as follows. Take any k E N ( negative k can be treated similarly) . Then k has a binary representation k 2:7=0 Ej 2j , with Ej 0 or 1j for good measure we can add a couple of zeros at the front end of k En En - l . . . E I EO , so that we can assume En En - l = O. If k is even, k = 2£, then 271" ¢> 3 + 2k7f + 2£71" J> + 2£71" 0 mo
,
=
J>e;
=
=
=
=
=
=
A
(
=
)
(i
) (i
) (because mo (i) 0) =
=
We therefore need to check only what happens if k is odd, k Then ; + 2k7f 8; + 4£7I"j hence
2
(
=
271" ¢> 3 + 2k7f A
)
=
=
2£ + 1, or EO = 1.
( ) mo ( 271" + £71") ¢> ( 271" £71") .
471" mo 3
3
A
3
+
CHAPTER 6
188 (a)
o 1
-7t
( b)
21 �
21 � 27t + -
7t [ - 7t, 7t 1 \
o
- 7t
t
1
21 �
+
+L______________��--------------�+
t
7t
K
FIG. 6 . 2 . This figure assumes that mo has only one zero in 11' / 3 < I{I :S 11' / 2, namely in {t = �; . We choose It = ] �� , Ii: [ ; hence 2It = ] �; , It; [. According to (6.3.6) , the compact set K is then [_ 1;; , _ It; ] u [-11', �; ] u [ 1;; , 11'] .
f
mo( 5; ) mo( ¢e;
moe;
O. O. mo
= -�) = It follows If is odd, i.e., E l = 1, then + hI' ) = that we only need to investigate further what happens for E l = 0, or e even. We can continue this further, showing that only those k with binary representation ending in 010101· · ·01 do not automatically lead to + 2k7r) = But if we work back far enough, then we will hit En En - l = 0 0, so that we indeed have + 2k7r) = This whole argument uses that the zero set of contains G , - n = [ e ; , -;11' } + 71'] mod 271' ) and that e ; , -r } is an invariant cy cle under the operation � t-+ mod 2 71' ) , mapping [ - 71' , 71'] into itself. In his Ph.D. Thesis, Cohen (1990b) proves that such invariant cycles are the root of the problem. T HEOREM 6 . 3 . 3 . Assume that is a trigonometric polynomial satisfying (6. 1.1) and = 1, and define ¢ as in (6.2.2) . Then conditions (1) and (2) in Theorem 6.3. 1 are also equivalent to
¢e;
O.
2� ( ( mo
mo(O)
3.
or the oper {6, " 'm, �o(�n } in) [ 1 ] allinvariant I j 1 for j 1, , f, , ,
there is no non-trivial cycle ation 2 mod 2 ) such that
� � t-+
(
71'
- 71' , 71'
=
=
n.
R EMARK S .
1. Because of (6. 1.1),
I mo(�j ) 1
=
1 is of course equivalent to
I mo(�j + 71') 1
=
o.
{O} , which is always an invariant cycle. 2; 3. In our example above, 6 2; , 2. Non-trivial means different from =
6
=
-
.
0
For a proof of this theorem and related results, we refer to Cohen (1990b); one of the two implications is in fact proved in step 6 of the proof of Theorem 6.3.5 below.
189
COMPACTLY SUPPORTED WAVELETS
A very different approach to the derivation of conditions on mo that ensure (6. 2 . 5) was initiated by Lawton (1990). Let us assume that mo is of the form N " hn e - i ne , mo ( then the only trigonometric polynomials invariant under are the constants .
0 Iml
=
k, n
hn hn - k fU - k
=
(6. 3 . 8 )!
Po
(1991)), 6.3.5. mo(O) 1. mo (6.1.1) I mo - �) 1 0,
1, 1990 (1992)
0,
Po
COMPACTLY SUPPORTED WAVELETS
191
R EMARK . This is sufficient to prove equivalence. If we denote Lawton's orig inal condition by (L), Cohen ' s condition by (C), Lawton ' s condition rephrased in terms of by and the orthonormality of the The 2 y are obtained by shifting the decimal point to the left. Since is 211'-periodic, only the "tail," i.e., the part of the expansion of 2 y to the right of the decimal point, decides whether 211'2 - y vanishes or not. If = db then y 2 would have the same decimal part as hence 211'y 2 = 0 would follow. Since 0, we therefore have = Similarly, e L+ n we conclude = = etc. It follows that are also successively equal to for some k E = . . . = {I, 2, . . . , n}. Since the are not all equal to 0, whereas = 0, this is a contradiction. This finishes the proof. •
/
eL+n
With Theorem 6.3.5 we end our discussion of necessary and sufficient con ditions on The following theorem summarizes the main results of §§6.2 and 6.3. T HEOREM 6 . 3 . 6 . Suppose is a trigonometric polynomial such that by = 1 . Define + + 11' 1 = 1 and
mo. I mo(�) 1 2 I mo(� ) 2
mo mo(O)
C/>, 'IjJ
00
j=l ¢(�) _e - if,,/2 mo(�/2 + 11') ¢(�/2) . Then C/>, 'IjJ are compactly supported L 2 -functions, satisfying n n where hn is determined by mo via mo(�) 0 L n hn e - inf" . Moreover, the -j x - k), j, k E Z constitute a tight frame for L2 (JR) with 2 -j/2 1'IjJ(2 'ljJj, k (X)constant . This tight frame is an orthonormal basis if and only if mo =
=
frame satisfies one of the following equivalent conditions: There exists a compact set K, congruent to [-11', 11'] modulo 211', containing a neighborhood of 0, so that >0. inf inf •
•
•
k> O f"E K I mo(2 - k �) 1 There exists no nontrivial cycle { , ' " } in [0, 211'[, invariant under � 2� modulo 211', such that mO (�6j + )�n 0 for all j 1" " n. The eigenvalue 1 of the [2 (N2 - N1 ) - 1] [2 (N2 - NI ) - l] -dimensional matrix A defined by N2 ARk n=N1 L hn hk - 2/+n , -(N2 - N1 ) + 1 £, k (N2 - NI ) + 1 (where we assume hn 0 for n N1 , n > N2 ) is nondegenerate. I--->
11'
=
=
x
::;
=
=
0 .5
s 'l' o
o -1 -0 .5 ��_�_�_�.-J o 2 4 6 8
-4
-2
o
2
4
1 .0 0.5
o
o -0.5
�___--.J L...__ ...
o
-1
�__�__��
-5
10
5
5
o
1 .0
9 '1'
0.5
o
o -
0.5
L--"::"'�_�_�--.l
o
5
10
15
-1 -5
o
5
FIG. 6 . 3 . Plots of the scaling functions N fiJ and wavelets N 'ifJ for the compactly supported wavelets with maximum number of vanishing moments for their support width, and with the extremal phase choice, for = 2, 3, 5, 7, and 9.
N
198
CHAPTER 6 TABLE 6 . 3
N
The low-pass filter coefficients for the "least asymmetric" compactly supported wavelets wi maximum number of vanishing moments, for = 4 to 10. Listed here are the CN, n v'2 hN, n ; one has L n C N, n = 2 . N _ 4
N = 5
N = 6
n
0
- 0 . 1 07 1 4890 1 4 1 8
1
- 0 . 04 1 9 1 0965 1 2 5
5
- 0.038493 5 2 1 263
2
0 . 703739068656
6
- 0 .073462 50876 1 0 . 5 1 5 398670374
eN,"
n
N _ 8
3
1 . 1 36658243408
7
4
0 . 4 2 1 234534204
8
1 .099106630537
5
- 0 . 1 403 1 7624 1 79
9
0 .68074 534 7 1 90
- 0 . 0 \ 782470 1 4 4 2
10
- 0 . 0866536 1 5406
7
0.045570345896
11
- 0 . 202648655286
0
0.038654 7959 5 5
12
0 . 0 1 07586 1 1 7 5 1
13
0.044 8236230 4 2
1
0.04 1 7468644 2 2
14
- 0 .000766690896
2
- 0.055344 1 861 1 7
15
- 0 .004 7834585 1 2
3
0 . 2 8 1 990896854
4
1 .023052966894
5
0 .896581 648380
N = 9
0
0 . 00 1 5 1 2487309
1
- 0 . 000669 1 4 1 509 - 0 .0 1 4 5 1 5 5 78553
6
0 .023478923136
2
7
- 0 . 2 4 795 1 3626 1 3
3
0 . 0 \ 2 5 28896242
8
- 0 .02984 2499869
4
0.087791 2 5 1 554
9
0.027632 1 5 2958
5
- 0 . 0 2 5 78644 5930
6
- 0 . 2 70893783503 0.049882830959 0 . 8730484U7349
0 . 0 2 1 784700327
7
1
0.0049366 1 2372
8
2
- 0 . 1 668632 1 54 1 2
9
1 .0 1 5 2 59790832
3
- 0.068323 1 2 1 587
10
0 . 337658923602 .- 0.0771 7 2 1 6 1 097
0
4
0.694 4 5 7972958
11
5
1 . 1 1 3892783926
12
U.0008 2 5 1 40929
6
0 . 4 779043 7 1 333
13
0 . 0 4 2 7 4 4 4 33602
7
- 0 . 1 02724969862
14
- 0 . 0 1 630335 1 2 26
8
- 0.02978375 1 299
15
- 0 . 0 1 8 769396836
9
0 .063250562660
16
0 . 000876502539
10
0 .002499922093
17
0.00 1 98 1 1 93736
11
- 0 . 0 1 1 03 1 867509
0
0 . 003792658534
0
0 . 00 1 089 1 704 4 7
1
0.000 1 35245020
2
- 0.00 1 48 1 2 2 59 1 5
2
- om 2 2 2064 2630
- 0.01 78704 3 1 6 5 1
3
- 0.002072363923
3
0.043 1 5 54 52582
4
4
0 . 0960 1 4 767936
5
+ 0.0649509 2 4 5 79
5
- 0.0700782 9 1 2 2 2
6
- 0 . 2 2 5 5 58972234
1
N = 8
eN."
0.0694904659 1 1
6
N = 10 N = i
4
0 . 0 1 64 1 8869H6
6
0.024665659489
7
- 0 . 1 002402 1 5031
7
0 . 7581 6260 1 964
8
0 . 6670 7 1 338 1 5 4
8
1 .0857827098 1 4
9
1 .088 2 5 1 530500
9
0.4081 83939725
10
0 . 5 4 2 8 1 30 1 1 2 1 3
10
- 0 . 1 98056706807
11
- 0.050256540092
11
- 0 . 1 5 24638 7 1 896
12
- 0.0452407722 1 8
12
0 .00567 1 3 4 2686
13
0.070703567550
13
0 .0 1 4 5 2 1 394762
14
0.008 1 5 2 8 1 6799
15
- 0.028786231926
0
0 .0026727!J3393
16
- 0 . 00 1 1 375353 1 4
1
- 0 . 0004 28394300
17
0 . 0064 95728375
2
- 0 . 02 1 1 4 5686528
18
0 . 00008066 1 204
3
0.005386388754
19
- 0 .000649589896
199
COMPACTLY SUPPORTED WAVELETS
2
1.5
1
N=4
I
0
-0.5
I
0
-1
N=6
1 . 5 ,---- --
1.0
0.5
I
f I
6
4
2
-2
2 ,--
N=6
.
1 .5 1.0 0.5
0 I
I
f
5
N=8
-
\
\
0
1.5 0.5
----N =-----
0
5
10
10
I
f 0 15
0 15
-1
-
'I'
5
0
N=8
1.-
-1 2
\
-5
2
- --�-
/1 1 \\
I
�
0 -0 . 5
5
10
1.0
10
�
0 -0.5
�
-1
-� �
VV�
0
0 -0.5
�-�--
4
2
0
-2
-5
-�� 'I'
�( 0
5
�
!N �-� --�-
��H
I
•
-5
0
5
N
FIG. 6.4. Plots of the scaling function ¢ and the wavelet 1/J for the "least asymmetric " = 4, 6, 8, compactly supported wavelets with maximum number of vanishing moments, for and 10.
200
CHAPTER 6
--- - -- ------ -
o
�
o
n
1 6 mO ( � ) I
o n
o
- - - - - - ""----I
o
n
o
FIG. 6 . 5 .
Imo ({) 1
for
N
=
2, 6 and 1 0 , corresponding t o the filters in Table 6 . 1 o r 6.3.
1 .0 I------�
0.5
o
_
o FIG. 6 . 6 .
_
'-_--=--,,�_.J
_ _ _ _ _ _ _ _ _ _
�
Plot of Imo ({) 1 for the 8-tap filter corresponding to
N
n =
2 and
mo (77r /9)
=
o.
201
COMPACTLY SUPPORTED WAVELETS
0 (6.1.11).
orthonormal bases are indeed very flat at and 7r, but very "round" in the transition region, near 7r /2. The filters can be made "steeper" in this transition region by a judicious choice of R in Figure shows the plot of corresponding to = 2 and R of degree 3 chosen such that has a zero at � = 77r/9 ( = This is much closer to a "realistic" subband coding filter. The corresponding "least asymmetric" function ¢ is shown in Figure 6.7; it is less smooth than 4 ¢ (which has the same support width, but corresponds to = and R == but turns out to be smoother than (for which has a zero of the same multiplicity, i.e. , 2, at � = 7r ) . In Chapter 7 we will come back in greater detail to these regularity and flatness issues. The hn corresponding to Figure are listed in Table
N 140°). 0),
N 4
6. 6 mo( 2 I 01 2¢
6. 7
I mo l
mo
6.4.
1 .5
1 .0
0.5
0
-0.5
0
2
4
6
2
o
- 1 L-__L-______�__�__�______� 4 2 -2
o
FIG . 6 . 7 . The "least asymmetric " scaling function 4> and wavelet .p corresponding to as plotted in Figure 6.6.
Imo l
I¢I l ?,b l II mo 0mo(± 2; ) 1, mt(O mo 2; ). mt 1. (6.1.1), mo j i e mt(2 �), I1� 1 + f,/2
All these examples correspond to real hn ' ¢ and 'ljJ, i.e., to and symmet ric around � = O. It is also possible to construct (complex) examples with ¢ , l ?,b l of the concentrated much more on � > than on � < O. Take for instance the (� _ = = and define previous example, which satisfies since does, and ( O ) = We can there obviously satisfies This ?,b = = # (O fore construct ¢# (�) mt (�/2 7r) ¢# ( � /2) ;
mt
202
CHAPTER
6
TABLE 6 . 4
The coefficients for the low-pass filter corresponding t o the scaling function in Figure 6 . 7 . n
01 342 657
hn -0. 0 802861503271 243085969067 -0.0. 3062806341592 50576616156 0.0. 5229036357075 -0. 00644368523121 -0. 115565483406 0. 0381688330633
these are compactly supported L 2 -functions, and the 'lj;tk ' j, k E Z constitute a tight frame for L 2 (JR) , by Proposition Moreover, since the only zeros 7 of ma on are in � = ± ; , ±11", it follows that mt (O = only for or 7; . Consequently, Imt (� ) 1 � C > for I � I :::; i , and the 'lj;tk � = ±11", constitute an orthonormal wavelet basis, by Corollary Figure plots oo Imt (O I , I ¢# ( O I and 1 ¢ # (O I ; it is clear that fa d� 1 ¢ # (O I 2 is much larger than a L oo d� 1 'Ij; # (O I 2 . Note that the negative frequency part of 'Ij; # is much closer to the origin than the positive frequency part, as required by the necessary condition oo The existence of such fa � 1 � I - l l ¢ # (O I 2 = f� oo � 1 � I -l l ¢ # ( � ) 1 2 (see "asymmetric" ¢ was first pointed out in Cohen in fact, for any E > one can find an orthonormal wavelet basis such that La oo � 1 'Ij;(O I 2 < E.
6. 2 . 3 .
[-11",11"] - 5; -
0 6. 3 . 2 .
A
0
6. 8
A
§3.4). (1990);
A
0
6.5. The cascade algorithm: The link with subdivision or refinement schemes. It can already be suspected from the figures in that there is no closed form analytic formula for the compactly supported ¢(x), 'Ij;(x) constructed here (except for the Haar case). Nevertheless, we can, if ¢ is continuous, compute ¢(x) with arbitrarily high precision for any given X; we also have a fast algorithm to compute the plot of ¢ .8 Let us see how this works. First of all, since ¢ has compact support, and ¢ E L l (JR) with f dx ¢(x) = we have P ROPOSITION If f is a continuous function on JR, then, for all X E JR,
§6.4
6.5.1.
lim
3 - 00
1,
2j J dy f(x + y) ¢(2j y)
=
f(x)
.
(6. 5 .1)
If f is uniformly continuous, then this pointwise convergence is uniform as well. If f is Holder continuous with exponent a , If(x) - f(y) 1 :::; G l x - y l " ,
203
COMPACTLY SUPPORTED WAVELETS 1 .0
0.5
o
-27t
27t
o
o �---�
-67t
-47t
-27t
0
27t
47t
67t
0
27t
47t
67t
O f------' -67t
-47t
-27t
FIG . 6 . 8 . Plots of Imo l , 14>1, and 1,z,1 for an orthonormal wavelet basis where tmted more on positive than on negative frequencies.
then the convergence is exponentially fast in j :
,z,
is concen
(6. 5 . 2) I f (x) - 2j I dy f (x + y) 4>(2j y)I � CTjQ . Proof. All the assertions follow from the fact that 2j 4> (2j . ) is an "approximate 8-function" as j tends to More precisely, 00.
I f (x) - 2j I dy f(x + y) 4>(2j y) I 1 2j I dy [J (x) - I (x + y)] 4>(2j y) I I I dz [J (x) - f (x + 2-j z)] 4>(z) I
204
CHAPTER
:::; I ¢ I £ 1 '
6
I f (x) - f (x + u) 1 (where we suppose support ¢
f
sup
l u l 9 -i R
C
[- R, R]) .
If is continuous, then this can be made arbitrarily small by choosing ] suf ficiently large. If is uniformly continuous, then the choice of ] can be made independently of and the convergence is uniform. If is Holder continuous, then (6.5.2) follows immediately as well. •
x,f
f
¢
Assume now that itself is continuous, or even Holder continuous with exponent a. (We will see many techniques to compute the Holder exponent of in the next chapter.) Take to be any dyadic rational, = 2 - J Then Proposition 6.5.1 tells us that
x
¢ (x )
x
K.
¢
lim 2j J dY ¢ ( T J K + y ) ¢ (2j y ) J-+OO lim 2j / 2 J dZ ¢(z) ¢ - J' 2i-J K (Z) J-+OO lim 2j / 2 ( ¢, ¢ - j , 2i-J K ) . J-+OO
Moreover, for ] larger than some ]0 ,
(6.5.3)
K. 2j - J K (¢, ¢_j,2i-J K ) ¢O , n rno ¢ (6.5.4) (I, ¢O,n ) OO,n , for ] > 0, k E Z . (6.5.5) (I, 1P-j,k) We can use this as input for the reconstruction algorithm of the subband filter ing associated with rno (see §5.6) . More specifically, we start with a low pass sequence c� OO , n and a highpass sequence � 0, and we "crank the machine" to obtain (6.5.6) C� 1 Lk hn - 2k c� . We then use d� 1 0, to obtain, after another cranking, (6.5.7) Cm- 2 '""n hm - 2n Cn- 1 , etc. At every stage, the c:;; j are equal to ( ¢, ¢ - j , n ) ' Together with (6.5.3) , this means that we have an algorithm with exponentially fast convergence to
where are dependent on J or If is integer, which is automatically true if ] � J, then the inner products are easy to compute. Under the assumption that the are orthonormal (which can be checked with any of the necessary and sufficient conditions on listed in Theorem 6.3.5), is the unique function f characterized by
C, ]o
°
=
=
=
=
= L...J
COMPACTLY SUPPORTED WAVELETS
¢
205
compute the values of at dyadic rationals. We can interpolate these val ues and thus obtain a sequence of functions approximating We can, for instance, define to be the function, piecewise constant on the intervals such that = . An other possible choice is piecewise linear on the so that = For both choices we have the following proposition. P ROPOSITION 6 . 5 . 2 . If is Holder continuous with exponent 0:, then there exists > 0 and jo N so that, for j ?: jo,
'f/j ¢.9 'f/J (2 -j k) -j 2j /-2j (¢, ¢-j, k ) [2 n, 2 (n + 1)], n E Z,
'f/jJ (x) j 1/2), 2 [2 (n(n+ 1/2)[' n E Z, 'f/} (x), 'f/; (2 -j k) 2j /2 (¢,¢_j, k ) . C E ¢
( 6.5.8 )
Take any x E For any j, choose n so that 2 - j n :S x -2 jProof. (n + 1). By the definition of 'f/j , 'f/j (x) is necessarily a convex linear com bination of 2j / 2 ( ¢, ¢ - j , n ) and 2j / 2 ( ¢, ¢ - j , n + l ) , whether 0 or 1. On the other hand, if j is larger than some jo, I ¢(x) - 2j /2 (¢, ¢-j,n )1 :S I ¢ (x) - ¢( T j n) 1 + 1 ¢ (2 -j n) - 2j / 2 ( ¢, ¢ - j , n )1 :S C Ix - 2 - j nl o + C T j o :S C T j o the same is true if we replace n by n + 1 . It follows that a similar estimate holds for any convex combination, or I ¢ (x) - 'f/j (x) I C 2 - j o . Here C can be chosen independently of x, so that ( 6.5.8 ) follows. This then is our fast algorithm to compute approximate values of ¢(x) with arbitrarily high precision: 1. Start with the sequence · · · 0 · · · 010 · · · 0 · · · , representing the 'f/o (n), n E Z. 2. AtCompute the 'f/j (2 -j n), n E Z, by "cranking the machine" as in ( 6.5.7) . every step of this cascade, twice as many values are computed: values at "even points" 2 - j (2k) are refined from the precious step, J+ l l!) , 'f/ ( 'f/j ( Tj 2k) V22::> T ( 6.5.9 ) l j _ k ) l 2 ( I and values at the "odd points" 2 - j (2k + 1) are computed for the first time, 'f/j ( Tj (2k + 1 )) V22:l h2( k -I) + 1 'f/j _ l ( Tj + l l!) . ( 6.5.10 ) K
3. The computation of for which :S Computation of these, in turn, involves only the or 3/4 :::; :S with Working back to j = J - 4, we see that to compute on m, ��] we only need the for 28 :S :S 34. We can therefore start the cascade from . . · 0 · . · 010 · . · 0 · . . , go five steps, select the seven values :S 34, use only these as the input for a new cascade, with four steps, and end up with a graph of on For larger blowups on even smaller intervals, we simply repeat the process; the blowup graphs in Chapter 7 have all been computed in this way. l1 The arguments leading to the cascade algorithm have implicitly used the orthonormality of the we have or equivalently (see §6.2, 6.3) , of the characterized as the unique function satisfying (6.5.4), (6.5.5) . The cascade algorithm can also be viewed differently, without emphasizing orthonormality at all, as a special case of a stationary subdivision or refinement scheme. Refinement schemes are used in computer graphics to design smooth curves or surfaces going through or passing near a discrete, often rather sparse, set of points. An excellent review is Cavaretta, Dahmen, and Micchelli (1991). We will restrict ourselves, in this short discussion, to one-dimensional subdivision schemes. 12 Suppose that we want a curve y = taking on the preassigned values One possibility is simply to construct the piecewise linear = graph through the points this graph has the peculiarity that, for all
¢
17] ,
¢, 'l/J
17]
175 (2 -5n), ¢ 17J(2 - Jn) [��, ��]
¢.
17J 2J- 4 :::; n :::; 2 J-4 h17Jn (2 - J+ 1nk) n (n - 3)/2 :::; k n/2.17J(2 - Jn) _l 17J_ 2 (2 - J+2t') (k - 3)/2 :::; t' :::; k/2, n/4 - 3/2 - t' n/4. 179 17�(2 -5 m) m 17�(2 -5 m), 28 m :::; 179 [��, ��]. 'l/Jj, k ,
¢
I (n) In .
¢O , n :
I
I (x)
(n, In ); n, 2n +-1 ) = "21 f (n) + "21 1 (n + l) , (6.5. 11 ) I (2 which gives a quick way to compute I at half-integer points. The values of I at quarter-integer points can be computed similarly, (6.5.12)
and so on for Z/4 + Z/8, etc. This provides a fast recursive algorithm for the computation of at all dyadic rationals. If we choose to have a smoother spline interpolation than by piecewise linear splines (quadratic, cubic or even higher
I
COMPACTLY SUPPORTED WAVELETS
(6. 5 . 9), (6. 5 .10),
207
order splines ) , then the formulas analogous to computing the 1(2- j n + 2- j -1 ) from the 1(2- j k) , would contain an infinite number of terms. It is possible to opt for smoother than linear spline approximation, with inter polation formulas of the type
I (Tjn + 2-j -1 ) =
1(2 -j (n - k)) , :�:�:> k k
(6.5.13)
ak are nonzero; the resulting curves are no longer splines. - 161 [/(TJ· (n - 1)) + 1(2 -J. (n + 2)] . (6. 5 .14) + 169 [J (T J n) + I(T J (n + 1))] . This example was studied in detail in Dubuc (1986), Dyn, Gregory, and Levin (1987), and generalized in, for example, Deslauriers and Dubuc (1989) and Dyn and Levin (1989); it leads to an almost C2 -function I . ( For details on methods to determine the regularity of I, see Chapter 7.) Formula (6. 5 .14) describes an with only finitely many An example is
.
interpolation refinement scheme, in which, at every stage of the computation, the values computed earlier remain untouched, and only values at intermediate points need to be computed. One can also consider schemes where at every stage the values computed at the previous stage are further "refined," corresponding to a more general refinement scheme of the type
Lk Wn - 2k /j (Tjk) .
(6.5.15)
Lk W2 (n - k ) /j (Tjk)
(6.5.16)
/j +l (Tj -1 n) =
(6. 5 .15)
Formula corresponds in fact to two convolution schemes ( with two masks, in the terminology of the refinement literature ) ,
/j +l (Tjn) =
( the refinement of already computed values ) , and
/j +l (Tjn + Tj - l ) =
Lk W2 (n - k )+l /j (Tjk)
(6. 5 .17)
( computation of values at new intermediate points ) . In a sensible refinement scheme, the /j converge, as j tends to 00, to a continuous ( or smoother; see defines the fJ only on the discrete Chapter function 100. Note that set 2- j Z. A precise statement of the "convergence" of /j to the continuous function 100 is that
7)
(6. 5 .15)
(6. 5 .18)
208
CHAPTER
A
6
It(n) A An . h(2W -j k);b o , 2 k,
where the superscript indicates the initial data, The refinement = scheme is said to converge if (6.5. 18) holds for all E £OO (Z) ; see Cavaretta, Dahmen, and Micchelli (1991). ( It is also possible to rephrase (6.5. 18) by first introducing continuous functions interpolating the see below. ) A general refinement scheme is an interpolation scheme if k = leading to
Ij ,
-j n) h(2-j n). hH(2 In both cases, general refinement scheme or more restrictive interpolation =
scheme, it is easy to see that the linearity of the procedure implies that the limit function 100 ( which we suppose continuous 13 ) is given by
loo (x) = Ln lo(n) F(x - n) ,
F F hex)
(6.5.19)
where = 00 is the "fundamental solution," obtained by the same refinement scheme from the initial data This fundamental solution obeys = a particular functional equation. To derive this equation, we first introduce functions interpolating the discrete
Fo(n) bn,o . h(2 -j k): hex) = Lk h( Tj k) w(2j x - k) , (6.5.20) where w is a "reasonable,, 1 4 function so that wen) = bn, o . Two obvious choices are w(x) = 1 for - � x � , 0 otherwise, or w(x) = 1 - I x l for I x l 1, 0 otherwise. ( These correspond to the two choices in the exposition of the cascade algorithm above. ) The convergence requirement (6.5. 18) can then be rewritten as l i lt - I� IIL'''' ->O for j->oo. For the fundamental solution Foo , we start from Fo(x) = w(x). The next two approximating functions Fl , F2 satisfy Fl (X) = Ln Fl (n/2) w(2x - n) (by (6.5.20) ( use (6.5. 15) and Fo(n) = bn, o ) = L Wn w(2x - n) n ( 6.5.21) Ln Wn Fo(2x - n) , F2 (X) Ln F2 (n/4) w(4x - n) ( use (6.5. 15)) L Wn - 2 k Fl (k /2) w(4x -- n) n, k ( becauseFl (k/2) = Wk ) L W k L Wi w( 4x - 2k - £) k i = L W k Fl (2x - k) . k This suggests that a similar formula should hold for all Fj , i.e., Fj (x) = Lk Wk Fj _ l (2x - k) (6.5.22) :S
:S
N2 , and hNl -# 0 -# hN2 . Then (6. 5 . 24) already implies that either N1 = 0 or N2 O. Suppose N1 = 0 (N2 0 is analogous) ; N2 is necessarily odd, N2 = 2L + 1. Take k 2L in (6. 5 . 24) . Then ==
n
O. In that case I sin n(1 S; In(lmin( I ,£) leads to a similar bound. 00
5. We use here the classical formula sm x _ 00 cos(2 -j x) . x j=1 An easy proof uses sin 2a = 2 cos a sin a to write J J sin(2 - j +1 x) . sin x -:cos(TJ x) = j x) = --:---:----:J sin(2 - J x) 2 sin(2 2 j= 1 j= 1 •
II
II
II
which tends to Si� X for J---t oo . In Kac (1959) this formula is credited to Vieta, and used as a starting point for a delightful treatise on statistical independence.
rno
6. This is true in general: if satisfies (6. 1.1) and ¢, as defined by (6.2.2), generates a non-orthonormal family of translates ¢O,n , then necessarily Ll l ¢ (� + 2 11"£) = 0 for some �. (See Cohen (1990b) .)
12
212
CHAPTER
6
J dx 'ljJ(x) 'ljJ(x -
7. The condition k) = 8k ,o may seem stronger than 11 'ljJ 11 = 1, but since the 'ljJj , k constitute a tight frame with frame constant 1, the two are equivalent, by Proposition 3.2. 1.
'ljJ ( x)
8. Since is a finite linear combination of translates of ¢(2x) , fast algo rithms to plot ¢ also lead to fast plots for Throughout this section, we restrict our attention to ¢ only.
'ljJ.
L2
9. If ¢ is not continuous, then the 'r/j still converge to ¢ in (see §6.3) . More over, they converge to ¢ pointwise in every point where ¢ is continuous. 10. The choice f. = 1 was used in the proof of Proposition 3.3 in Daubechies (1988b) , because the are absolutely integrable, whereas the are not. In Daubechies (1988b) the convergence of the 'r/j to ¢ was actually proved first (using some extra technical conditions) , and orthonormality of the ¢O,n was then deduced from this convergence.
ryJ
ryJ
1 1 . Note that there exist many other procedures for plotting graphs of wavelets. Instead of a refinement cascade one can also start from appropriate ¢( n) and then compute the ¢(2 -j k) directly from ¢( ) = n) . (In fact, when ¢ is not continuous, the cascade algorithm may diverge, while this direct use of the 2-scale equation with appropriate ¢( n ) still converges. I would like to thank Wim Sweldens for pointing this out to me.) This more direct computation can be done in a tree-like procedure; a different way of looking at this, avoiding the tree construction and leading to faster plots, uses a dynamical systems framework, as developed by Berger and Wang (see Berger (1992) for a review) . The "zoom in" feature is lost, however.
x v'2Ln hn ¢(2x -
12. Many experts on refinement or subdivision schemes find the multidimen sional case much more interesting! 13. This is not a presentation with fullest generality! We merely suppose that the are such that there exists a continuous limit. This already implies = = 1.
W L W2nk L W2n+ l
14. For example, any compactly supported be "reasonable" here.
w
with bounded variation would
15. The following stretched Haar function shows how the F(x-n) can fail to be independent. Take = = 1 , all other = O. The solution to (6.5.23) is then (up to normalization) = 1 for 0 � < 2, 0 otherwise. In this =0 case the lOO -sequence defined by = leads to a.e.
Wo W2 F(x) Wn x A An (_ I ) n
Ln An F(x-n)
16. This is no coincidence. If we fix the length of the symmetric filter Mo = Imo l 2 , then the choice R == 0 means that Mo is divisible by (l + cos � ) with the highest possible multiplicity compatible with its length and the constraint Mo( � ) + Mo (� +IT) = 1. On the other hand, Lagrange refinement
213
COMPACTLY SUPPORTED WAVELETS
schemes of order 2N - 1 are the interpolation schemes with the shortest length that reproduce all polynomials of order 2N - (or less) exactly from their integer samples. In terms of the filter W(�) = � L n Wn ein� , this means
1
W(�) + W(� + 11") = 1
(interpolation filter:
W2n
= 8n , o )
and 1 + 0 ((1 - cos �) N ) (see Cavaretta, Dahmen, and Micchelli (1991), or Chapter
8).
The two requirements together mean that W (� + 11" ) has a zero of order 2N in � = 0, i.e., that W(� + 11" ) is divisible by (1 - cos �) N ; hence W(O by + cos �) N . It follows that W = Mo .
(1
CHAPTER
7
M ore A b o ut the Reg u l a rity of Com pactly S u p ported Wave l ets
The regularity of the Meyer or the Battle-Lemarie wavelets is easy to assess: the Meyer wavelet has compact Fourier transform, so that it is Coo , and the Battle Lemarie wavelets are spline functions, more precisely, piecewise polynomial of degree k, with (k - 1) continuous derivatives at the knots. The regularity of compactly supported orthonormal wavelets is harder to determine. Typically, they have a non-integer Holder exponent; moreover, they are more regular in some points than in others, as is already illustrated by Figure 6.3. This chapter presents a collection of tools that have been developed over the past few years to study the regularity of these wavelets. All of these techniques rely on the fact that ¢(x) = en ¢(2x - n ) , (7.0. 1) where only finitely many en are nonzero; the wavelet ,¢, as a finite linear combina tion of translates of ¢(2x) , then inherits the same regularity properties. It follows that the techniques exposed in this chapter are not restricted to wavelets alone; they apply as well to the basic functions in subdivision schemes (see §6.5) . Some of the tools discussed here were in fact first developed for subdivision schemes, and not for wavelets. The different techniques fall into two groups: those that prove decay for the Fourier transform ¢, and those that work directly with ¢ itself. We will illustrate each method by applying it to the family of examples N ¢ constructed in §6.4. It turns out that Fourier-based methods are better suited for asymptotic estimates (rate of regularity increase as N is increased in the examples, for instance) ; the second method gives more accurate local estimates, but is often harder to use. References for the results in this chapter are Daubechies (1988b) and Cohen (1990b) for §7. 1 . 1 ; Cohen (1990b) and Cohen and Conze (1992) for §7. 1.2; Cohen and Daubechies (1991) for §7. 1.3; Daubechies and Lagarias (1991, 1992), Mic chelli and Prautzsch ( 1989), Dyn and Levin (1990) , and Rioul (1992) for §7.2; Daubechies (1990b) for §7.3.
L
7. 1 .
Fourier-based methods. The Fourier transform of equation (7.0. 1) is ¢(� ) = mo (� /2) ¢(� /2) , 215
(7. 1 . 1 )
2 16
CHAPTER
Ln cn e - in�
7
is a trigonometric polynomial. As we have seen where mo(�) � many times before, (7. 1 . 1) leads to 00
j=1 1 and J dx ¢(x) ( 1 +;-t.� ) N £(�) ,
where we have assumed mo (O) mo can be factorized as mo (�)
=
=
(7. 1.2) 1, as usual. Moreover, (7. 1.3)
£ is a trigonometric polynomial as well;Nthis leads to . (7.1.4) ¢(�) (271') - 1 /2 ( l -i�e - t� ) JII= £(Tj �) . 1 A first method is based on a straightforward estimate of the growth of the infinite product of the £(2 - j �) as I � I ---> oo . Brute force methods. For n + {3, n E N, 0 ::; {3 < 1, we define Co. to be the set of functions j which are n times continuously differentiable and such that the nth derivative j ( n ) is Holder continuous with exponent {3, i.e., I j (n) (x) - j (n) (x + t) 1 ::; Cltl ,6 for all x, t .
where
00
=
7. 1 . 1 .
Q =
It is well known and easy to check that if
J d� I j (�) 1 (1 + I W o. < 00 , then j E Co. . In particular, if I j (�) 1 ::; C ( 1 + 1 � 1 ) - 1 - 0. - oo of II;: 1 £(2 - j � ) in (7. 1.4) can N be kept in check, then the factor ( (1 - e - i� )/i�) ensures smoothness for ¢. N sup� I £ (�)I < 2 - 0. - 1 , then ¢ E Co. . 7. 1 . 1 . If L EMMA
Proof. 1. Since mo (O)
q
=
1,
=
£(0)
=
00
Now take any �, with I � I 2. Hence II 1 £ ( Tj �) 1 j=1
1 ; hence
2':
I £(�) I ::; 1 + ClH Consequently, 00
2J- 1 ::; I � I < 2J. J II 1 £ ( Tj �) 1 II 1 £ ( Tj T J �)I j=1 j=1 eC ::; c'N2 J(N1- 0. - 1 - OO
0
follows that JC = limj -> oo JCj .
-
then JCt < N 1 3. repeat If JC < N 1 for some £ E N. We can then the argument in the proof of Lemma 7.1.1, applying it to -
-
-
a,
-
00
00
j= 1
j =O
a
with Ct (�) = I1�:� C(2 - j �) , and with 2t playing the role of in This leads to I ¢(�) I :::; G(l + I W - N +Ke :::; G(l + I W -o- 1 -., Lemma hence ¢ E Go . •
2
7.1.1.
The following lemma shows that in most cases, we will not be able tlJ obtain much better by the brute force method. LEMMA There exists a sequence (�t ) tEN so that
7.1.3.
(1
+ I �t l ) - K
00
II C(Tj �t )
j=1
�G>0.
CHAPTER 7
218
Proof.
1. By Theorem 6.3.1, the orthonormality of the ¢> ( . - n ) implies the existence of a compact set K congruent to [-11", 11"] modulo 211", such that 1 ¢ (e ) 1 � C > 0 for e E K. Since K is congruent to [-11", 11"] and CI. is periodic with period 21.+ 1 11", we have
ql.
=
sup
ICI. (e ) 1
l e 1 91 .".
sup ICI. (e) 1 ,
=
e E21 K
i.e., there exists (I. E 21.K so that 1 CI. ((I.) 1 = ql. . Since K is compact, the 2-1. (I. E K are uniformly bounded. We therefore have
( 7. 1 . 7)
for 0 < Cf•
1 1+;;( 1 = 1 cos e/2 1 ::; 1, we have for all e E 2i K, II C( 2 -j e) � II mo ( T j e) = 1¢( 2 -l.e) 1 � C > 0 . l.+l l.+l
2. Moreover, since 00
00
j=
j=
Putting it all together we find for el. = 2(1. 00
00
I Ci ((£ ) 1 II l C (Tj (i )
j =l.+ > C ql. = C 2£ lCl
•
By (7.1.7),
(1 + l el. l ) - lC
00
j el.) ( T C j=1 II
� C 2£lC l CIf
TilC .
Since /C = inf£ /C£ , this is bounded below by a strictly positive constant.
•
Let us now turn to the particular family of N ¢> constructed in §6.4, and see how these estimates perform. We have
with
We start by establishing a few elementary properties of PN .
7.1.4.
LEMMA The polynomial the following properties:
Proof. If 0
1. :::; x :::; y, then x- ( N-1 ) PN (x) = >
219
PN (X) = E::01 ( N -� + n ) xn satisfies (7.1. 8 ) (7.1. 9)
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
}; ( N - � + n ) x - (N-1-n) }; ( N -� + n ) y-(N- 1-n) Y-(N- 1) PN (Y) ·
2. Recall (see §6.1) that PN is the solution to xN PN (I - x) + (1 - x) N PN (X) = 1 . On substituting x = �, it follows that PN (I/2) = 2 N -1 . For x :::; �, we have PN (X) :::; PN (�) = 2 N -1 because PN is increasing. For x ::::: �, applying (7.1. 8 ) leads to PN (X) :::; x N - 1 2 N- 1 PN (�) = 2 N - 1 ( 2x) N-1 . This proves (7.1.9). It is now easy to apply Lemmas 7.1.1 and 7.1. 2 . We have [}; ( N � + n ) r �p I C N ( D I •
,
�
1 immediately leads =
to sharper results. We have, for instance, sup q2
j
I .cN (�) .cN (201 sup [PN (y) PN (4y(1 - y)W / 2 O : 0 is independent of k. C
=
=
m =
=
=
Cl > 0 so that, for all k E N, (7.1.10) Indeed, 2kM �o �o (mod 211" ) , so that (7.1.10) follows if �o =I- 0 or ±11". We already know that �o =I- 0 ; if �o ±11", then 6 0 (mod 211" ) and hence �o 2M- l 6 = 0 (mod 211" ) , which is impossible.
Proof. 1. First note that there exists =
2. Now =
=
=
221
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
1,
Since C is a trigonometric polynomial and C(O) = there exists G2 so that IC( O I - G2 I� I ;::: e -2 C2 I� I for I�I small enough. Hence, for r large enough,
;::: 1
00
00
j=r M
j=r M
Hence
1 ¢(2 k M +l �0) 1
(r+ k ) M - I ( 2k M I�ol ) -N G3 II C(2 k M - £ �o ) £= 0 G4 IC(�o ) C(6) · · · C(�M _dl r+ k +l (1 + 1 2 k M �o l) -N G5 2 K M k (1 + 12 k M �ol ) -N G (1 + 1 2 k M +l �ol ) -N + K .
> Gf > > >
•
7.1. 5
We can apply this to the example at the end of the last subsection: Lemma implies 1¢( 2n 2nl ;::: G(l + 12n 2311" 1 )-N +K, with JC = log lC en C(- 2nl / 2 log 2 . If C has only real coefficients (as is the case in most applications of practical interest), then IC(- �.n l = I C en l , and JC = log I C en l/ log 2. The next short . . t cyc . 1es are { 2 11" 4 11" 411" } 211" ' 7411" ' - 7611" } , etc.,· each f them gIves mvanan 5' 5' -5' -5 ' {7 an upper bound for the decay exponent of ¢ . In some cases one of these upper bounds on a can be proved to be a lower bound as well. We first prove the following lemma. LEMMA Suppose that [-n, n] = D I UD2 · · ·UDM , and that there exists q > 0 so that I C(OI '5: q 21r
0
7.1.6.
IC(O C(20 1 '5: q2 IC(� ) C(20 · · · C(2 M - I O I '5: q M Then I¢(OI '5: G(l + I W -N+K, with K = log q/ log 2. Proof.
1. Let us jestimate 1 I1{:� C(2- k �) I , for some large but arbitrary j. (=
2- +l � E Dm for some
m
E
{ l , 2, · · · , M } , we have
Since
222
7 We can now apply the same trick to 2 m (, and keep doing so until we cannot go on. At that point we have CHAPTER
1
with at most M - different £-factors remaining ( Le., r
with q 1 defined as in
I nk=O £(Tko l ::; qj-M+l q�- l ,
::; M - 1). Hence
(7.1. 5 ). Consequently, with the definition (7.1.6), 1 Kj ::; -:-- [C + J. log q] , J 1og 2
and K = limj -+ oo Kj ::; 10g q / log 2. The bound on (p now follows from • Lemma
7.1.2.
In particular, one has the following lemma. LEMMA Suppose that
7.1.7.
I � I ::; 2; , (7.1.11) 2 2 1 £ (0 £(2�) 1 ::; l £ e3"' ) 1 for ; ::; I � I ::; 7r . Then 1 (P (�) 1 ::; C ( l + I W - N + K , with K = 10g l £ en l / log 2, and this decay is optimal. The proof is a straightforward consequence of Lemmas 7.1. 5 and 7.1. 6Proof. . Of course, Lemma 7.1.7 is only applicable to very special £; in most cases (7.1.11) will not be satisfied: there even exist £ for which £ ( �.n = O. Similar optimal bounds can be derived by using other invariant cycles as breakpoints for a partition of [ -7r, 7rJ , and applying Lemma 7.1. 6 . Let us return to our "standard" example N- _ To prove (7.1.18) it is therefore sufficient to prove N- l 9 N-l 2� (2y 1) 4 (7.1.20) (2y � 1) 2 Since both (2y - 1) - 2 and y(2y - 1) - 2 are decreasing on [�, 1] , it suffices to verify that (7.1.20) holds for y = � , i.e., that N 4 . � -l < _ 6 9 v'N This is true for N 2: 13.
[
]
IN (
()
)
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
225
5. It remains to prove (7.1.13) for � � Y � 1 and 1 � N � 12. We do this in two steps: Y � Yo = ¥, and Y Yo. For Y � ¥, 4y(1 - y) �, hence, again by Lemma 7.1. 4 , [16y(1 - y)] N - l . 2':
2':
< ( 20 ) N - l PN (�) 9
y2 (1 - y) (7.1.21)
[�, ¥].
,
(7.1.21 )
because is decreasing on One checks by numerical computation that for � N � is indeed smaller than
[PN (�)j 2 1
12.
6. For ¥ = YoN -�1 Y � 1 we usederive the bounds PN (4y(1-y)) � (2y-1) - 2 N and PN(y) � ( ;; ) PN (yO) to N- l < N + 2 o PN(4y(1 - y))PN(y) Y ! PN(yo)(2y - 1) [ (2Y � 1) 2 ] < 2 N PN(Yo) , (7.1. 2 2) where the last inequality uses that (2y - 1) - 2 and y (2y _ 1) - 2 are both decreasing in [Yo, 1]. One checks by numerical computation that (7.1. 2 2) is smaller than [PN (�W for 5 � N � 12. 7. small It remains to prove (7.1.13) for 1 � N � 4 and ¥ � y � 1. For these values of N the polynomial PN(y) PN (4y(1-y))-PN (�) 2 has degree at most 9, and its roots can be computed easily ( numerically ) . One checks that there are none in H , 1], which finishes the proof, because (7.1.13) is satisfied in y = 1. •
follows from Lemmas
N¢(�):
7.1.8 and 7.1.7 that we know the exact asymptotic decay (7.1.23)
CHAPTER 7
226
For the first few values of N this translates into N
. 1 / log 2 . 1, then F E eo:-< for all f > O. a
If 1 >' 1
' 1 . Since, for any b > 0, there exists e > 0 so that IIAn l l ::; C(p( A) + b) n for all n E N, it follows that (7. 1.30) 3. On the other hand, f (� ) � 1 for � ::; I�I ::; 11". Together with the bound edness of n;': 1 Mo(2 -j �) for I�I 11" (derived as usual from IMo ( � ) 1 < 1 + C l W , this implies
::;
j=1
::; e
J
2n - l "': � . For every binary sequence d = (dn ) n E N\ { O } we also define its right shift rd by n = 1 , 2" " . It is then clear that rd(x ) = d( 2x ) if 0 ::; x < � , rd(x ) = d( 2x - 1) iq < x ::; 1 . (For x = � , we have two possibilities: rd+ (�) = d(O ) , rd- (�) = d(1) . ) Although r is really defined on binary sequences, we will make a slight abuse of notation and write rx = y rather than rd(x ) = d(y ) . With this new notation, we can rewrite (7.2.5) , (7.2.7) as the single equation If the Vj have a limit v, then this vector-valued function fixed point of the linear operator T defined by (Tw) (x)
= Td1 (x )
v
will therefore be a
w (rx) ;
acts on all the vector-valued functions w : [0, 1 ] requirements T
( 7.2.9 )
--+
]RK that satisfy the
[w ( O ) h = 0, [W (1)] K = 0, [W ( O )] k = [W (1)] k - b k = 2" " , K .
( 7.2.10 )
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
235
( As a result of these conditions Tw is defined unambiguously at the dyadic rationals: the two expansions lead to the same result. ) What has all this recasting the equations into different forms done for us? Well, it follows from (7.2.9) that
Vj (x)
= Td 1 (x) Td2 (x)
.
.
•
Tdj (x)
vo( rj x) ,
which implies (7.2.1 1 )
In other words, information on the spectral properties of products of the Td matrices will help us to control the difference Vj - Vj +l , so that we can prove Vj ----+ v, and derive smoothness for v. But let us turn to an example. For the function 2 (7.2. 1) reads
2(X) =
3L Ck 2(2x - k) ,
(7.2.12)
k=O
with
Co =
1 + v'3 4
3 + v'3 4
1 - v'3 4
3 - v'3 4
Note that
(7.2. 13)
C Cl + 3C3 ,
and
22 =
(7.2. 14)
both of which are consequences of the divisibility of mo (�) = ! E!=o Ck e-ike by (1 + e-ie) 2 . The values 2 (1), 2(2) are determined by the system
( 22(2)(1) )
=
M
( 22(2)(1 ) ) '
with
M
=
( CC31 Coc2 )
Because of (7.2. 13), the columns of M all sum to 1 , ensuring that (1, 1) is a left eigenvector of M with eigenvalue 1. This eigenvalue is nondegenerate j the right eigenvector for the same eigenvalue is therefore not orthonormal to (1, 1), which means it can be normalized so that the sum of its entries is 1 . This choice of ( 2(1), 2(2)) leads to 1 + v'3 1 2(2) = - v'3 . 2 2 The matrices To, Tl are 3 x 3 matrices given by
2(1) =
CHAPTER 7
236
Because of ( 7 .2. 1 3) , To and T1 have a common left eigenvector e 1 with eigenvalue 1 . Moreover, for all x E [0, 1) ,
e1 . vo{x)
=
e1 · [(I - x) vo{O)
+
=
(I, 1, 1)
x vo{I))
(I - x) [ 24> (1) + 24> (2)) + x [ 2 4> (1) + 2 4> (2) ) (use (7.2.2))
=
1.
It follows that, for all x E [0, 1)' all j E N,
e 1 · vi (x)
=
=
e 1 · Td 1 ( x ) . . . Tdj (x ) vo{ ri x)
1.
Consequently, vo{y) - Vt{Y) E El = {Wj e 1 · w = W1 + W2 + W3 = O}, the space orthogonal to e1 . In view of (7.2.11), we therefore only need to study products of Td-matrices restricted to E1 in order to control the convergence of the vi . But more is true! Define e2 ( I, 2, 3). Then (7.2.14) implies =
(7.2.15) Co + 2C2 - � = 5 -/'3 , a1 = C1 + 2C3 - � eg = e2 - 2ao e l , then (7.2. 15) becomes
with ao
=
3 -4Y3 . If we define
On the other hand,
(I - x) eg . vo{ O ) + x eg . vo{ l )
eg . vo{x)
-
X
j
consequently,
eg . vi {x)
= =
eg . Td1 ( x ) vi - 1 {rx) 1 1 - 2 d 1 { x ) + 2 e02 . vi - 1 { rx)
i
=
-
I: 2- m dm {x)
m= l i
- I: 2- m dm {x) m= l
+ 2-i eg . vo{ ri x)
-x .
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
eg . [vo {x) - Vi {X)] Vj - Vj+l '
237
It follows that = O. This means that we only need to study products of Td-matrices restricted to E2 , the space orthogonal to and in order to control But, because this is a simple example, E2 is one dimensional, and Td l E2 is simply multiplication by some constant, namely the 1 - Y3 for T1 • Consequently, third eigenvalue of Td , which is ¥ for 4
el
To,
I l vj{x) - Vj+l{X)I I ::;
1 - L:� 1 1 J3 1 =
[ + J3]j 1 + 1 --
d (x) n
eg,
J3 Vi are uniformly bounded.6 Since I ��a I 1, I vj{x) - Vj+l{X) I ::; C Taj , with .550. It follows that the Vj have a limit 1 l0g « 1 + J3)/4 ) I / log 2 function V, which is continuous since all the Vj are and since the convergence is uniform. Moreover V automatically satisfies (7.2.1O), since all the Vj do, so that it can be "unfolded" into a continuous function F on [0 , 3] . This function solves ( 7.2.1 ) , so that 2 F, and it is uniformly approachable by piecewise linear spline functions Fj with nodes at the k 2 -j ,
4
c,
where we have used that the (7.2.16) implies a
=
(7.2.16)
. = 1/2, then the Uh derivative F(l) of F is almost Lipschitz: it satisfies
!p (l)
( x + t)
- F ( l)
(x) 1 �
G l t i l ln I t l l .
REMARK. The restriction >. 2:: � means only that we pick the largest possible integer £ � L for which (7.2.21) holds with >. < 1. If £ = L, then necessarily >. 2:: � (see Daubechies and Lagarias (1992)) ; if £ < L and >. < � , then we could replace £ by £ + 1 and >' by 2>', and (7.2.21) would hold for a larger integer £. 0 A similar general theorem can be formulated for the local regularity fluctua tions exhibited by the example of 2, 'l/J have sufficient decay; both conditions are trivially satisfied for the compactly supported wavelet bases as constructed in Chapter 6.) This was our motivation to construct the N1>, which lead to N'l/J with N vanishing moments. The asymptotic results in §7.1.2 show however that the N1>, N'l/J E C/L N with I-t .2. This means that 80% of the zero moments are "wasted," i.e., the same regularity could be achieved with only N/5 vanishing moments. Something similar happens for small values of N. For instance, 2 1> is contin uous but not C 1 , 31> is C 1 but not C2 , even though 2 'l/J, 3 'l/J have, respectively, two and three vanishing moments. We can therefore "sacrifice" in each of these two cases one of the vanishing moments and use the additional degree of freedom to obtain 1> with a better Holder exponent than 2 1> or 3 1> have, with the same support width. This amounts to replacing Imo (�) 1 2 = (cos2 � ) N PN (sin2 � ) by N I mo (�) 1 2 = (cos2 � ) - 1 [PN_ 1 (sin2 � ) + a(sin 2 � ) N cos �] (see (6.1.11)), and to choose a so that the regularity of 1> is improved. Examples for N = 2, 3 are shown in Figures 7.4 and 7.5; the corresponding hn are as follows: 3 ho N = 2 5 v2 �
h1
h2 h3 N
3
6
5 v2 2 5 v2 -1 5 v2
ho .3743 2841633/V2 h1 .109093396059/ V2 h2 = .786941229301/V2 h3 - .146269859213/V2 h4 -.161269645631/V2 h5 .0553358986263/V2
These examples correspond to a choice of a such that max[p(To I Et ) ' P(T1 I Et )] is minimized; the eigenvalues of To, T1 are then degenerate.s One can prove that the Holder exponents of these two functions are at least .5864, 1.40198 respectively, and at most .60017, 1.4176; these last values are probably the true Holder' exponents. For more details, see Daubechies (1993). 7.4.
Regularity or vanishing moments?
The examples in the previous section show that for fixed support width of 1>, 'l/J, or equivalently, for fixed length of the filters in the associated subband coding scheme, the choice of the hn that leads to maximum regularity is different from the choice with maximum number N of vanishing moments for 'l/J. The question
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
most
regular � for N
=
243
2
o o
3
2
FIG. 7.4. The scaling function 4> for the most regular wavelet construction with support width 3.
1 .5
r--,----.---,--,
most regular � for N
=
3
1 .0 0.5
o -0.5
--'-
L--'-____
o
2
--'-
____
4
----'
__
FIG. 7.5. The scaling function 4> for the most regular wavelet construction with support width 5 .
then arises: what is more important, vanishing moments or regularity? The answer depends on the application, and is not always clear. Beylkin, Coifman, and Rokhlin (1991) use compactly supported orthonormal wavelets to compress large matrices, i.e., to reduce them to a sparse form. For the details of this ap plication, the reader should consult the original paper, or the chapter by Beylkin in Ruskai et al. (1991) ; one of the things that make their method work is the number of vanishing moments. Suppose you want to decompose a function F( x ) into wavelets (strictly speaking, matrices should be modelled by a function of two variables, but the point is illustrated just as well, and in a simpler way, with one variable). You compute all the wavelet coefficients (F, 1/Jj,k ) , and to compress all that information, you throw away all the coefficients smaller than some threshold € . Let us see what this means at some fine scale; j -J, J E N and J "large." If F is CL-1 and 1/J has L vanishing moments, then, for x near =
244
CHAPTER
7
2-J k , we have F(T J k) + F' (T J k)(x - T J k) F(x) + . . . + ( L _1 ) ! F ( L- l ) (T J k) (x - T J k) L- l + (x - T J k) L R (x) , I
where R is bounded. If we multiply this by 'l/J(2J x - k) and integrate, then the first L terms will not contribute because J dx xi'l/J(x) = 0, = 0, , , , , L - 1. Consequently,
f
I ( F, 'l/J - J, k) I
I f dx (x - T J k) L R(x) 2J/2 'l/J(2J - k) 1 T J(L- l /2) f dy lyl L I 'l/J( Y ) I . X
1 (2x - n) .
=
257
Then the same calculations as in Chapter 5 show that the functions 1/1�j , k (X) = 2 -j 1/1 1 ( 2 - 2j x - k), 1/1�j + 1 , k (X) = 2 -j - 1 / 2 1/1 2 ( 2 - 2j 1 x - k) (j, k E Z) constitute an orthonormal basis for L2 (1R) . Since the recursions above correspond to -
( p (�)
mo (�/2) mo (�/4) mo (�/8) mo (�/16) · · · CXl
]
[
II mo (T 2j - 1 � ) mo (2 - 2j - 2 � ) , j= l the phase of J1 can be expected to be closer to linear - -phase than that - of '" 00 mo (2 - J' �). Note also that ¢>2 (0 = ¢> 1 (� ) ' 1/12 (0 = 1/11 (�) ; ¢>(�) = TI)=l hence ¢>2 �X) = ¢> 1 ( -X) , 1/12 (X) = 1/1 1 ( -X). Figure 8.2 shows ¢> 1 , 1/11 computed from the for N = 2, i.e. , ho = ltg, 1 = 3tg, h2 = 3:;g, = l:;g. (Unlike the previous construction, this "switching" makes a difference even for N = 2.) For the "least asymmetric" given in Table 6.3, this switching technique lead� to slightly "better" ¢>, but seems to have little effect on 1/1 . 0 ,,
-
-.
•
h
hn
h3
.
hn
2 ,----�----.--�
o .0.5 '---_ -___L-___---' __---'-_ o
2
3
3 ,----�----,--�
2
-
1
-2 '-------'--L--� 2 o -1
FIG. 8 . 2 . Scaling function cPl and wavelet 'l/J l obtained by applying the "switching trick"
the 4-tap wavelet filters of §6.4.
258
CHAPTER
8
8.2.
Coiftets. In §7.4 we saw one advantage of having a high number of vanishing moments for 'ljJ: it led to high compressibility because the fine scale wavelet coefficients of a function would be essentially zero where the function was smooth. Since J dx ¢(x) = 1, the same thing can never happen for the ( I, ¢j , k ) . Still, if J dx x" ¢(x) = 0 for f = 1" " , L, then we can apply the same Taylor expansion argument and conclude that for J large, ( I, ¢ - J, k ) 2 J/2 ! (2 - J k), with an error that is negligibly small where ! is smooth. This means that we have a particularly simple quadrature rule to go from the samples of ! to its fine scale coefficients ( I, ¢ - J,k) . For this reason, R. Coifman suggested in the spring of 1989 that it might be worthwhile to construct orthonormal wavelet bases with vanishing moments not only for 'ljJ, but also for ¢. 4 In this section I give a brief account of how this can be done; more details are given in Daubechies (1993). Because they were first requested by Coifman (with a view to applying them for the algorithms in Beylkin, Coifman, and Rokhlin) , I have named the resulting wavelets "coiflets." The goal is to find 'ljJ, ¢ so that �
and
J dx x"'ljJ(x)
=
f = 0"
0,
J dx ¢(x) = 1 , J dx x"¢(x) = 0,
·
·
,L -
l
(8.2.1)
f = 1, , , , , L - 1;
(8.2.2)
L is then called the order of the coiflet. We already know how to express (8.2. 1) in terms of mo; it is equivalent with
mo (�)
=
( 1 +;- "e ) L C(�) . .
(8.2.3)
f!P 1
What does (8.2.2) correspond to? It is equivalent to the condition 4> =o = 0, e f = 1" " , L - 1. Let us check what 4>' (0) = 0 means for mo. Because 4> (�) = mo (�/2) 4> (�/2) , we have
� m� (�/2) 4>(�/2) + � mo (�/2) 4>' (�/2) ; 4>'(0) = � m� (O) (271") -1/ 2 + � 4>' (0) ,
4>'(�) = hence or
m� (O) = (271" ) 1 / 2 4>'(0) " Consequently, J dx x¢(x ) = 0 is equivalent with m� (O) = O. Similarly, one sees that (8.2.2) is equivalent with 4> = 0, = 1, . . . , L - 1 , or with e =o
( f!P l )
f
(8.2.4)
259
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES
where C is a trigonometric polynomial. In addition to (8.2.3) and (8. 2 .4) , mo will of course also have to satisfy Imo (�W + Imo (� 1. Let us specialize to L even (the easiest case, although odd L are not much harder), L 2K. Then (8.2.3) , (8.2.4) imply that we have to find two trigonometric polynomials P1 , P2 so that (8. 2 .5)
+ 7r)1 2 =
(
Because
=
(1 + e-i� ) 2K= e -'.�K (cos2 ) K, ( 1 e -'�. ) 2K= e -'.K� (2i sin ) 2K.) "2 "2 2 �
-
�
But we already know what the general form of such P1 , P2 are: (8.2.5) is nothing other than the Bezout equation which we already solved in §6. 1 . In particular, P1 has the form
where I is an arbitrary trigonometric polynomial. It then remains to taylor I in mo (�) ((1 + 1 is P1 ( � ) so that I mo (�W + I mo (� + 7r W satisfied. With the ansatz I( � ) = L ��;;- l it is shown in Daubechies (1990) how to reduce this "tayloring" to the solution of a system of K quadratic equations for K unknowns. A heuristic, perturbative argument suggests that this system will have a solution for large K, and explicit numerical solutions are computed for K = 1" " , 5. Figure 8.3 shows the plots of the resulting , 'l/J; the corresponding coefficients are listed in Table 8. 1. It is clear from the figure that , 'l/J are much more symmetric than the N, N'l/J of §6.4, or even than the , 'l/J in §8. 1, but there is of course a price to pay: a coiflet with 2K vanishing moments typically has support width 6K - 1, as compared to 4K - 1 for 2 K .
=
e - i� )/2) 2K
In e - in�,
=
In e - in�
REMARK. The ansatz I( � ) = L ��;;- l is not the only possible one, but it makes the computations easier. For small values of K (K = 1, 2, 3) , different ansatzes are also tried out in Daubechies (1993) . It turns out that the smoothest coiflets (at least at these small values for K) are not the most symmetric ones; for K 1, for instance, there exists a (very asymmetric) coiflet with Holder exponent 1 . 191814, whereas the coiflet of order 2 in Figure 8.3 is not C 1 ; both have support width 5. Similar gains of regularity can be found for K 2, 3. For graphs, coefficients and more details, see Daubechies (1990b) . 0
=
=
8.3. Symmetric biorthogonal wavelet bases. As mentioned above, it is well known in the subband filtering community that symmetry and exact reconstruction are incompatible, if the same FIR filters are used for reconstruction and decomposition. As soon as this last requirement is given up, symmetry is possible. This means that we replace the block diagram of Figure 5. 1 1 by Figure 8.4. Several questions naturally arise: what do€.s Figure 8.4
260
CHAPTER
8
3 ,--�--�----, L = 2 2
1
o
o
I--��
-1 -2
-1
3
2
0
,-------�--,
1 .5
-2
�-�--' ��-�
-1
-2
3
2
0
2 ,--------;
1 .0 0.5
o
�--.J
L-_�__
-0 . 5
_
o
1 .0
5
5
L = 6
0.5
o
'---�
-0.5
'--'-_��_��--.J
1 .5
,----.-----�
-4
-2
0
2
-2
-4
4
0
4
2
2 ,.-----,
1 .0 0.5
o I----�"""' -4
1 .5 1 .0 0.5
o
-0 . 5
-2
0
2
-4
4
-2
0
2
4
0
2
4
,----.-----�
L = 10
L = 10
'-----
o
��
'--'--��-��-
-4
-2
0
2
4
1-----.-'
-4
-2
FIG. 8.3. Coiftets "" and their corresponding scaling functions 4> for L The support width of 4> and "" is 3L - 1 in all cases.
=
2, 4, 6, 8, and 10
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES TABLE
8. 1
261
The coefficients lor coiflets 01 order L = 2K, K = 1 to 5 . ( The coefficients listed are normal ized so that their sum is 1; they are equal to the 2- 1/2hn .) n
hn/v'2
-2 -1 0 1 2 3
- .05 1429728471 .238929728471 .602859456942 .272140543058 - .05 1429972847 - .011070271529
-4 -3 -2 -1 0 1 2 3 4 5 6 7
-
. 0 1 1587596739 .029320137980 .047639590310 .27302 1046535 .574682393857 .294867193696 .054085607092 .042026480461 .016744410163 .0039678836 13 .00 1289203356 .000509505399
K=3
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11
- .002682418671 .005503126709 .016583560479 - .046507764479 - .043220763560 .286503335274 .561285256870 .302983571773 - .050770140755 - .058196250762 .024434094321 .01 1229240962 - .006369601011 - .001820458916 .000790205101 .000329665174 - .000050192775 - .000024465734
K=4
-8 -7 -6 -5 -4 -3 -2 -1
.000630961046 - .00 1 1 52224852 - .005 194524026 .011362459244 .018867235378 - .057464234429 - .0396526485 1 7 .293667390895
K = l
K=2
-
-
n K=4
K=5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 - 10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
h n /v'2
-
-
-
.553126452562 .3071 57326 198 .0471 12738865 .068038 12705 1 .027813640153 .01 7735837438 .0107563 185 17 .004001012886 .002652665946 .000895594529 .000416500571 .000183829769 .000044080354 .000022082857 .000002304942 .000001262175
- .0001499638 .0002535612 .001 5402457 - .00294 1 1 108 - .0071637819 .0165520664 .01991 78043 - .0649972628 - .0368000736 .2980923235 .5475054294 .3097068490 - .0438660508 - .0746522389 .0291958795 .023 1 107770 - .0139736879 - .0064800900 .0047830014 .001 7206547 - .001 1 758222 - .000451 2270 .0002137298 .0000993776 - .0000292321 - .0000150720 .0000026408 .0000014593 - .0000001 1 84 - .0000000673
262
CHAPTER 8
dj
mean in terms of multiresolution analysis? What do d and now stand for? ( They were coefficients of orthogonal projections in Chapter 5.) Is there an associated wavelet basis? How does it differ from the bases constructed earlier? The answer is that, provided the filters satisfy certain technical conditions, such a scheme corresponds to two dual wavelet bases, associated with two different multiresolution ladders. In this section we will see how to prove all this, and give several families of ( symmetric! ) examples. Except for an improved argument due to Cohen and Daubechies (1992) , all these results are from Cohen, Daubechies, and Feauveau (1992). Many of the same examples are also derived independently in Vetterli and Herley (1990) , who present a treatment from the "filter design" point of view; a useful factorization scheme for this type of filter bank is in Nguyen and Vaidyanathan (1989) .
cI - 1
FIG. 8.4. Sub band filtering scheme with exact reconstruction but reconstruction filters different from the decomposition filters.
8.3. 1 . Exact reconstruction. Since we have now four filters instead of two, we have to rewrite (5.6.5), (5.6.6) as
d�
L k
=
L k
Cl n [hl- -2n Cn 1U-2n dn]
and
_
-
°
1
L...J "
+
1
In the z-notation introduced in §5.6, this can be rewritten as
C
O (z) =
� [ii(z) h(z) g(z) g(z)] cO (z) � [ii(z) h( -z) g(z) g( -z)] +
+
+
co ( -z) .
Consequently, we require
ii(z) h (z)
+
g(z) g(z) = 2 ,
ii(z) h( -z) + g(z) g( -z) = 0 ,
(8.3.1)
(8.3.2) where we assume ii, g, h, 9 to be polynomials since the filters are all FIR. (For simplicity, we use the term "polynomial" in a slightly wider sense than usual:
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES
an z n h(z) -z)p(z) -z)] o:zk
263
we also allow negative powers. In other words, L��- Nl is a polynomial in h this terminology.) From (8.3. 1) it follows that and 9 have no common zeros; consequently, (8.3.2) implies that
g (z) z z p(z) -z)g(z) p(z) g(z) o:zk h(-z) , = h ( - )p ( ) ,
= -g(
(8.3.3)
for some polynomial p. Substitution into (8.3. 1) leads to
[h(
- h(z)g(
= 2.
The only polynomials that divide constants are monomials; hence for some 0: E
C,
=
k E Z, and (8.3.3) becomes
(8.3.4)
=
Any choice for 0: and k will do; we choose 0: = 1, k = 1, which makes the equations (8.3.4) for g and 9 symmetric. Substitution into (8.3.1) gives
h(z) h(z) h(-z) h(-z) n hn hn+2k Dk,O ,
= 2.
+
In terms of the filter coefficients,
all
(8.3.5)
this becomes
L
(8.3.6)
=
(8.3.7) where we have implicitly assumed that all the coefficients are real. These equa tions are obvious generalizations of (5. 1.39) , (5. 1.34) .
8.3.2. Scaling functions and wavelets. Because we have two pairs of fil ters, we also have two pairs of scaling function + wavelet: {lPj,k ; Wj {-¢;j,k; Wj Wj C
C
with Vii
VI I
C
C
Vo
C
V- I
C
C
C
C
C
·k E ll}, Span k E ll}. The spaces k E Il} are again complements Span k E ll}, Span of Vj , respectively, "Cj in Vj - I , respectively, "Cj - I , but they are not orthogonal complements: typically the angle7 between Vj , or "Cj , will be smaller than 90° . This is the reason why we have to prove (8.3. 12) in this case, whereas it was automatic in the orthonormal case. Another way of seeing this is the following. Because of the non-orthogonality we have
Wj
=
Span
Vo
=
=
[k 1 (1, ¢>j,kW 1 (1, lPj,k) 12] Lk 1 (1, ¢>j- l,kW [� 1 (I, ¢>j,kW �I (I' lPj'k )12l ' lPj,k Wj Wj Ej,k I (I, lPj,k ) 1 2 .
QL
:::;
+
:::; f3
+
with Q < 1, f3 > 1 ( in the orthogonal case, equality holds, with Q f3 1) . Unlike the orthonormal case, we cannot telescope these inequalities t o prove that the constitute a Riesz basis: telescoping would lead to a blowup of the constants. We therefore have to follow a different strategy. Note that (8.3.13) implies that .1 "Cj , .1 Vj . The two multiresolution hierarchies and their sequences of complement spaces fit together like a giant zipper, and this is what allows us to control expressions like But let us return to the conditions (8.3. 12) and (8.3. 14) . We already saw how to tackle condition (8.3.14) in §6.3, in the simpler orthogonal case. Our strategy here is essentially the same. We again define an operator Po acting on 27r-periodic functions, =
=
a second operator Po is defined analogously. In terms of the Fourier coefficients of f, the action of Po is given by
we will be mostly interested in invariant trigonometric polynomials for Po. This means that we can restrict our attention to the 2( + I-dimensional subspace of f for which It 0 if f > NI ( we assume hn 0 if n < I or n > on which Po is represented by a matrix. Theorems 6.3.1 and 6.3.4 have the following analog. T HEOREM 8 . 3 . 1 . The following three statements are equivalent:
N2) ,
=
N2 -
N2 - NI )
=
N
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES
1. , E L2 (JR.) and (O , k , O , l ) = Ok , l .
267
2. There exist strictly positive trigonometric polynomials fo, io invariant for Po, Po; there also exists a compact set K congruent to [-11", 11"] modulo 211" so that
3. There exist strictly positive trigonometric polynomials fo, io invariant for Po, Po, and these are the only invariant polynomials for Po, Po (up to normalization) . The proof is very similar to the proofs in Chapter 6, but a bit more compli cated. In §6.3, the functions fo, io were simply con!tant; in this case, they are essentially fo(�) = Ll I ¢(� 211"£W, io(�) = L l �(� 211"£) 1 2 . For details on how to adapt the proofs of §6.3 to the present case, see Cohen, Daubechies, and Feauveau (1992) . Condition (8.3.14) therefore simply amounts to checking that two matrices have a nondegenerate eigenvalue 1 and that the entries of the corresponding eigenvectors define a strictly positive trigonometric polynomial. (Note that if the trigonometric polynomial takes negative values, then ¢ L2 ( JR. ) . This happens for some exact reconstruction filter quadruplets.) Condition (8.3. 12) is something we had not encountered in the orthogonal case. It turns out that this condition is satisfied if any of the three conditions in Theorem 8.3. 1 holds. The proof of this surprising fact is in the following steps8:
+
•
•
•
+
First, one shows that the existence of an eigenvalue >. of Po with 1 >' 1 2: 1, >. =I- 1 would contradict the square integrability of . It follows there fore from Theorem 8.3.1 that all the other eigenvalues of Po have abso lute value strictly smaller than 1 if the eigenvalue 1 is nondegenerate and the associated eigenvector corresponds to a strictly positive trigonometric polynomial. The proof of this step uses Lemma 7. 1. 10. Since mO(1I") = 0 = mo(1I") , we have obviously MO (1I") = I mo(1I"W = 0 = I mo(1I") 1 2 = MO(1I") . We saw in Chapter 7 that this means that the columns of the matrix representing Po all sum to 1 , so that the row vector (of the appropriate dimension) with all entries 1 is a left eigenvector for Po with eigenvalue 1. It follows from the first point that p, the spectral radius of PO I E1 ' with El = {I; L n fn = O}, is strictly smaller than 1 . One then uses that f (�) = 1 - cos � is in El to prove (the estimates are analogous to those in the proof of Theorem 7.1. 12) that J2 n- l .".S;lel:5 2 n.". d� I ¢(�W � C (l¥) n .
Via Holder 's inequality this implies J � 1 ¢(�) 1 2 ( 1- 6 ) < 00 for sufficiently small O . This can then be used to prove a "discretized" version, i.e., Lm EZ I ¢(� 1I"m) 1 2 ( 1- 6') � C < 00 for all � E JR., again for sufficiently
+
CHAPTER 8
268 small
(/ .
Because m 1 is bounded, -0 satisfies a similar bound, L 1 -0(e + 211'm) 1 2 ( 1- 6' ) :::; G < 00 . mE:!:
•
(8.3.15)
On the other hand, one can also prove that (8.3. 16) Since -0 is entire and -0(0) 0, 1 -0(e) 1 :::; C 1el for sufficiently small l e i , so that �J= - oo 1 -0(2j e W6' is uniformly bounded for lei :::; 211', and we only need to concentrate on j � 0 in (8.3. 16) . But =
� !£ 1 -0(e W �
[
� 1 -0(O I
2j 1I" :5 l e l :5 2H 1 11"
:::; G
.
/
.
I� 1 ] . [/
d( I ¢( ( W
23 - 1 11" :5 1 and 3 , 3 'I/J are examples of the fact that the cascade algo
FIG. 8 . 7 . Spline examples with if = 3, N = 3, 5, 7, and 9. For N
[ - Nt 1 , Nt3 ] .
�-'
________�______
is not square integmble. Here support
=
=
=
=
The functions rithm may diverge while the direct algorithm still converges (see Note 1 1 at the end of Chapter 6).
276
CHAPTER
8
3 CP
O.S 0 - O .S -1
O.S
2
0
-
-
3,7 1/;
3,9 1/;
O.S
0
0
- O. S
- O. S
-1
-4
-2
0
2
4
2
-1 -S 2
3,7 CP 0
-1
-1
-S
3,7 I/;
0 -1
S
Ii
2
1v 0v
-2
0
2
-S
3,9 I/;
0 -1
\J
-2 -4
0
-2 4
-S
FIG. 8.7. ( continued) .
S f\
3,9 CP
0
2
0
i lt 0
10
S
1 Ar 0
S
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES
277
TABLE 8 . 2 List of Nino , N N mo for the first few values of N, N , with z = e - i{ . The corresponding filter coefficients N h� , N N h k are obtained by multiplying v'2 with the coefficient of z k in N N inO , , , respectively. Note that the coefficients of N N mO are always symmetric; for very long N N mO , , we only list about half the coefficients ( the others can be deduced by symmetry) .
N
fJ mO
N
N
N mo
1 - ( 1 + .) 2
1 - ( 1 + .) 2
%- 1 1 % .z 2 .z 3 % -2 - -- + -- + - + - + - - 16 16 2 2 16 16
3
_
256
z-4 _
� %-3
_
256
":':' % - 2 + ":':' % - 1 128
128
+
..: 2
+ .: + ":':' % 2 2 128
11 3 3 4 3 5 - -z - -% + -% 256 256 128 1 - (.- 1 + 2 + .) 4
--z 8 4
_
128
z -4 _
3 1 1 1 - 2 + - z - 1 + - + - z - - z2 4 4 8 4
� %-3 64
_
': % - 2 8
+
�%-l 64
+
� 64
+
�% 64
_ ': % 2 _ � % 3 + .2.... % 4 128 8 64 6
_ __ z - 6 + � % - 5 + � % - 4 _ � % - 3 _ � % - 2 1024 512 512 512 1024 1 75 + �%- l + + � % _ � %2 1024 256 256 256
8
2 - 1 5 ( 3 5 % - 8 - 70z - 7 - 300 % - 6 + 670z - 5 + 1228% - 4 - 3126%-
3
2 1 - 3796% - + 10718% - + 22050 + 10718% - 3796% 2 . . . )
3
9 3 7 45 45 7 - z - 3 - - z - 2 - - % - 1 + - + - % - - %2 64 64 64 64 64 64 _ � %3 + � %4 64 64 _ _ z - 5 + � .z - 4 + � % - 3 _ � % - 2 _ � % - 1 512 512 512 512 256 +
7
175 256
+
�.
256
_
� .2 256
2 - 14 (35% - 7 - 105.= - 6 - 195% - 5 + 865% - 4 + 336% - 3 - 3489.= -
2
1 - 307% - + 1 1025 + 1 1 025.= · · · )
2 - 1 7 ( _ 63.= - 9 + 189.z - 8 + 469 .= - 7 - 1 9 1 1 .= - 6 - 1308 .= - 5 + 9 1 88.= - 4 + 1 1 40.= - 3 - 29676.= - 2 + 190.= - 1 +873 18 + 87318 •
. . .
)
2 78
CHAPTER
8
The functions 1, 3 'l/J and 1, 3 "p were first constructed in Tchamitchian (1987) as an example of two dual wavelet bases with very different regularity prop erties. Here they constitute the first non-orthonormal example of the family eN = 1 = N gives the Haar basis ) . As in the orthonormal case, arbitrarily high regularity can be attained with these examples, for both 'l/J and "p . As a spline function, fir ' N "p is piecewise polynomial of degree if - 1 and is e fir - 2 at the knots; the regularity of fir , N 'l/J can be assessed with any of the techniques in Chapter 7 . Asymptotically, for large if, one finds that fir , N 'l/J E em if N > 4. 165 if + 5. 165 (m + 1). These spline examples have several remarkable features. For one thing, all the filter coefficients are dyadic rationals ; since divi sion by 2 can be done very fast on a computer, this makes them very suitable for fast computations. Another attractive property is that the functions fir N 'l/J(x) are known exactly and explicitly for all x, unlike the orthonormal co�pactly supported wavelets we saw before ,u One disadvantage they have is that mo and rho are very unequal in length, as is apparent from Table 8.2. This is reflected in very different support widths for ¢ and ¢; because they are determined by both mo and rho , 'l/J and "p always have the same support width, given by the average of the filter lengths of mo, rho, minus 1 . The large difference in filter lengths for mo, rho can be a nuisance in some applications, such as image analysis.
Examples with less disparate filter lengths. Even if we still take R 0, it is possible to find mo and rho with closer filter lengths by choosing an appropriate factorization of P ( sin2 � ) into qo ( cos e) and qo ( cos e) . For fixed f + i there is a limited number of factorizations. One way to find them is to use spectral factorization again: we determine all the zeros (real and pairs of conjugated complex zeros) of P, so that we can write this polynomial as a product of real first and second order polynomials, ==
P (x)
h
=
12
A II (x - Xj ) II(x 2 - 2Rezi x + lzi I 2 ) . i= l j=l
Regrouping of these factors leads to all the possibilities for qo and qo. Table 8.3 gives the coefficients for mo , rho for three examples of this kind, for f + i = 4 and 5. (Note that f + i = 4 is the smallest value for which a non-trivial factorization of this type is possible, with qo , qo both real.) For f + i = 4, the factorization is unique, for f + i = 5 there are two possibilities. In both cases we have chosen f, i so as to make the length difference of mo, rho as small as possible. The corresponding wavelets and scaling functions are given in Figures 8.8 and 8.9. In all cases the conditions of §8.4.2 are satisfied.
8.3.5. Biorthogonal bases close to an orthonormal basis. This first example of this family was suggested by M. Barlaud, whose research group in vision analysis tried out the filters in §6A, 6B for image coding (see Antonini et al. (1992)) . Because of the popularity of the Laplacian pyramid scheme ( Burt and Adelson (1 983 )) , Barlaud wondered whether dual systems of wavelets could
279
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES
TABLE 8 . 3
The coefficients of mo, mo for three cases of "variations on the spline case" with filters of similar length, coTTesponding to t + 1 4 and 5 (see text) . For each filter we have also given the number of (cos t; / 2) factors (denoted N, iii ) . As in Table 8.2, multiplying the entries below with v'2 gives the filter coefficients hn , hn . =
N, N N=4 N=4 N=5 N=5
N=5 N=5
1 .5
n
0 -1 -2 -3 -4 0 1, -1 2, -2 3, -3 4, -4 5, -5 0 1, - 1 2, -2 3, -3 4, -4 5, -5 1, 2, 3, 4,
coefficient of e-ine in rno .557543526229 .295635881557 - .0287717631 14 - .045635881557 0 .636046869922 .337150822538 - .066117805605 - .096666153049 - .001905629356 .00951533051 1 .382638624101 .242786343133 .043244142922 .000197904543 .015436545027 .007015752324 2
cp
coefficient of e-ine in rno .602949018236 .266864118443 - .078223266529 - .016864118443 .026748757411 .520897409718 .244379838485 - .038511714155 .005620161515 .028063009296 0 .938348578330 .333745161515 - .257235611210 - .083745161515 .038061322045 0
if;
0.5 0
0 - 0.5
1 .5
-2
-1
2
0
2
cp
0.5
-2
0
2
4
0
2
4
-
if;
0
0 - 0.5
-4
FIG.
8.8.
-2
0
2
4
-1
-2
The functions 4>, �, 1/J, ;P coTTesponding to the case N
=
4
=
iii in Table 8.3.
CHAPTER 8
280
(a)
2
2
cp
if;
0 0
-1 -4
-4
4
-2
cp
� V I[ 0
2
4
-2
0
2
4
-2
0
2
4
-2
0
2
4
-2
if;
0.5 0 0 -4
(b)
0.8 0.6
-2
0
2
4
-1 -4
cp
if;
0.4
0.5
0.2
0
0 - 0.2
- 0. 5 -4
-2
0
2
-4
4
6 4
cp
5
if;
2 0
0 -2 -4 -4
FIG . 8.9.
-2
0
2
4
-5
-4
The functions r/>, 4>, 'I/J, 1iJ corresponding to the two cases
N
=
5
=
N in Table 8.3
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES
281
be constructed, using the Laplacian pyramid filter as either rno or mo . These filters are given explicitly by
(8.3.27) For a = - 1/16, this reduces to the spline filter 4 mO as described under the "spline examples" above. For applications in vision, the choice a = .05 is es pecially popular: even though the corresponding � has less regularity than 4 � ' it seems to lead to results that are better from the point of view of visual per ception. Following Barlaud ' s suggestion, we chose therefore a = .05 in (8.3.27) , or
.6
.5 cos �
(cos �r ( 1 +
. 1 cos 2� +
� sin2 � )
(8.3.28)
Candidates for mo dual to this rno have to satisfy
rno (�) rno (�)
+
rno (� + ) rno (� + ) 11"
11"
=
1.
As shown in §8.4.4, such rno can be chosen to be symmetric (since rno is sym metric) ; we also opt for rno divisible by (cos �/2) 2 (so that the corresponding 'IjJ, ,(f both have two zero moments) . In other words,
where
(1 - x) 2
( 1 � x) P(x) +
+
x2
( � - � x) P(l - x)
=
1.
By Theorem 6. 1.1, together with the symmetry of this equation for substitution of x by 1 - x, this equation has a unique solution P of degree 2, which is easily found to be 24 6 P(x) = 1 + 5 x - x2 . 35 This leads to 24 . 4 � � 2 1 + 6 s. 2 � - s cos (8.3.29) - m 2 5 2 35 m 2 73 73 3 17 3 - e - 3ie - 56 e -2ie + 280 e - ie + 28 + 280 e ie 280 3 3 (8.3.30) - 56 e 2ie - 280 e 3ie . One can check that both (8.3.28) and (8.3.29) satisfy all the conditions in §8.4.2. It follows that these rno and rno do indeed correspond to a pair of biorthogonal
(
)
(
)
282
CHAPTER
8
wavelet bases. Figure 8. 10 shows graphs of the corresponding 4>, if>, 'I/J and ;j;. All four functions are continuous but not differentiable. It is very striking how similar if> and 4> are, or 'I/J and ;j;. This can be traced back to a similarity of rno and mo, which is not immediately obvious from (8.3.27) and (8.3.30), but becomes apparent by comparison of the explicit numerical values of the filter coefficients, as in Table 8.4. In fact, both filters are very close to the (necessarily nonsymmetric) filter corresponding to one of the orthonormal coiflets (see §8.3) , which we list again, for comparison, in the third column in Table 8.4. This proximity of rno to an orthonormal wavelet filter explains why the mo dual to rno is so close to rno itself. A first application to image analysis of these biorthogonal bases associated to the Laplacian pyramid is given in Antonini et al. ( 1992). 2 r-
--r
�----
I
�--,
----
2
I
j
-2
2
0
-I
I
t/;
0 -1
-2
-1
0
2
3
-1
o
2
3
2
-1
-2
2
o
FIG. 8 . 1 0 . Grophs of Adelson low-pass filter.
•
.(f
-2
for the biorthogonal pair constructed from the Burt
M. Barlaud's suggestion led to the accidental discovery that the Burt filter is very close to an orthonormal wavelet filter. (One wonders whether this closeness makes the filter so effective in applications?) This example suggested that maybe other biorthogonal bases, with symmetric filters and rational filter coefficients, can be constructed by approximating and "symmetrizing" existing orthonormal wavelet filters, and computing the corresponding dual filter. The coiflet coeffi cients listed in §8.3 were obtained via a construction method that naturally led to close to symmetric filters; it is natural, therefore, to expect that symmetric biorthogonal filters close to an orthonormal basis will in fact be close to these
283
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES
TABLE 8.4
Filter coefficients for (mo )Burt , for the dual filter (mo )Burt computed in this section, and for a very close filter (mo )coiflet corresponding to an orthonormal basis of coiftets (see the entries for K = 1 in Table 8 . 1 ) . n
-3 -2 -1 0 1 2 3
(O.mo)Burt -.05 .25 .6 .25 -.05
O.
(mO ) Burt
-.010714285714 -.053571428571 .260714285714 .607142857143 .260714285714 -.053571428571 -.010714285714
(mo )coiflet
O. -.051429728471 .238929728471 .602859456942 .272140543058 -.051429972847 -.011070271529
[� (K k k) [ ( K - k1 + k )
coiflet bases. The analysis in §8.3 suggests, therefore, K -1 -1+ 2K (sin �/2) 2k + O ((sin �/2) 2K ) �) = (cos �/2)
mo(
In the examples below we have chosen in particular K -1 (sin �/2) 2 k � ) = (cos �/2) 2K
mo(
E
+
and we have then followed the following procedure:
1 [:
a(sin �/2) 2K
1
- I mo(�
1
.
1
1. Find a such that dE" [1 - l mo(�) 1 2 + 11") 1 2 ] is minimal (zero in the examples below) . This optimization criterium can of course be replaced by other criteria (e.g., least sum of squares of all the Fourier coefficients of 1 + 11") 1 2 instead of only the coefficient of e ile with £ = 0) . For the cases = 1, 2, 3, the smallest root for a is .861001748086, 3.328450120793, 13. 113494845221, respectively. 2. Replace this (irrational) "optimal" value for a by a close value expressible as a simple fraction. 12 For our examples a = .8 = 4/5 was chosen for = 1, a = 3.2 = 16/5 for = 2 and a = 13 for = 3. For = 1, this reduces then to the example above. 3. Since is now fixed, we can compute If we require that be also divisible by (cos �/2) 2K , then �) = (cos �/2) 2K PK ((sin �/2) 2 ) , (8.3. 31) where PK is a polynomial of degree - 1. The same analysis as in Daubechies (1990) shows that
I mo(�) 1 2 - I mo(� K
K
K
mo
K
K mo
mo.
mo( (
PK x)
=
3K � ( K - � + k ) xk
+
O xK ) ,
(
284
CHAPTER
8
thereby determining already K of the 3K coefficients of PK . The others can be computed easily. For K = 2 and 3 we find
14 1 + 2x + 5 x
8024 4 3776 5 x + x (8.3.32) 455 455 1721516 1 + 3x + 6x 2 + 7x 3 + 30x4 + 42x5 x6 6075 1921766 7 648908 8 + (8.3.33) x x 6075 6075 2
+
8x 3 -
In Table 8.5 we list the explicit numerical values of the filter coefficients for mo , mo and the closest coiflet, for K = 2 and 3. We have graphed
o.
290
CHAPTER
and Proof. Choose L
a ::;
9
I Qk l - 1 1rQ k dx f (x) ::; 2a,
for all
kEN.
2£ so that 2 -£ IJR dx f (x) ::; a. It follows that -L 1 I��+ l )L dx f (x) ::; a for all k E Z. This defines a first partition of R a fixed interval Q [k L, (k+ l ) L [ in this first partition. Split it into 2. Take two halves, [k L, (k + � )L[ and [(k + � ) L (k + l ) L [ . Take either of the halves, call it Q', and compute lQ1 I Q' I - 1 IQ I dx f(x) . If lQ1 > then put Q' in the bag of intervals that will make up B. We have indeed 2 1 Q I - 1 h dx f(x) 2a . a < lQ1 ::; I Q ' I - 1 h dx f(x) 1.
=
=
,
=
a,
::;
=
If lQ I < a, keep going (split into halves, etc.) , if necessary, ad infinitum. Do the same for the other half of and also for all the other intervals l)L[. At the end we have a countable bag of "bad" intervals which all satisfy the equation at the top of this page; call their union B and the complement set C.
Q,
[kL, (k +
3. By the construction of B, we find that for any x (j B , there exists an infinite sequence of smaller and smaller intervals Q , Q3 , · · · so that for every n, and IQn dy f(y) ::; a. In fact, 2 = xE for every j , and C Because the Qn "shrink to" x,
-1 I Q nl Qj Qj - 1 . I Qn l - 1 1Qn dy f(y)
Qn
Q1 , Q Q I j l � I j-1 1
-+
f(x) almost surely .
Since the left side is ::; a by construction, it follows that f (x) ::; in C . •
2£
a
a.e.
Note that the choice L = implies that all the intervals occurring in this proof are automatically dyadic intervals, i.e. , of the form for some , j E Z. Next we define Calder6n-Zygmund operators and prove a classical property. DEFINITION. A Calder6n-Zygmund operator T on lR is an integral operator
k
[k2 -j , (k + 1)2 -j [
1
(Tf) (x)
=
J dy K(x, y) f(y)
(9. 1.1)
for which the integral kernel satisfies C I K(x, y) 1 ::; x - yI ' K( y) K( , y) ::; ' x X, y X I x YI 2 and which defines a bounded operator on L 2 (lR) .
\:
-I
\ + \:
\
�
(9.1.2)
( 9. 1. 3 )
291
CHARACTERIZATION OF FUNCTIONAL SPACES
THEOREM 9. 1 .2 . A Calder6n-Zygmund operator is also a bounded operator from L 1 (JR) to L�eak(JR) . The space L�eak(JR) in this theorem is defined follows. DEFINITION. I E L�eak(JR) il there exists C > 0 so that, lor all a > 0, as
(9. 1.4 )
The infinum of all C for which ( 9. 1.4 ) holds (for all a > 0) is sometimes called IIIII L'weak . 2 EXAMPLES .
1. If I E L 1 ( JR) , then (9. 1.4) is automatically satisfied. So. = {x; I / (x) 1 � a} , then
a · I Sa l :::; { dx I/ (x) l :::; ( dx I/ (x) 1
lJi
lsc>
hence
1IIII Ll
weak
Indeed, if
IIIII L' ;
:::; IIIII L' .
I x l -1 is in L�eak ' since I {x; I x l -1 � a} 1 = � However, I x l - !3 is not in L�eak if f3 > 1. The name L�eak is justified by these examples: L�eak extends L 1 , and contains the functions I for which J III "just" misses to be finite because of logarithmic singularities in the primitive of II I . We are now ready for the proof of the theorem. 2. I (x) I (x)
=
=
Prool ol Theorem 9. 1.2. 1. We want to estimate I {x; I TI (x) 1 � a} l . We start by making a Calderon Zygmund decomposition of JR for the function III , with threshold a. Define now
{ {:
if x E G ,
I (x)
g(x)
b(x)
I Q k l -1
=
1
Qk
dy I (y)
(X) - I Q . I -1
1" dy f(y)
if x E interior of Q k , if X E G , if x E interior of Q k
•
Then I (x) = g(x) + b(x) a.e.; hence TI = Tg + Tb. It follows that I TI (x) 1 � a is only possible if either ITg(x) 1 � a / 2 or I Tb(x) 1 � a / 2 (or
CHAPTER 9
292 both) ; consequently,
I {x; ITI (x) 1 � a} 1 � x; I Tg(x) 1 �
I{
�}I + I { x; I Tb(x) 1 �}I ·
(9.1.5)
�
The theorem will therefore be proved if each of the terms in the right-hand side of (9. 1 .5) is bounded by � 1 1 / 11 Ll .
2. We have
( � f I { x; ITg(x) 1 �}I {X; J dx �
�
I Tg(xW
I Tg (x) I � V
h. dx I Tg(xW = II Tg ll 1.2 � C ll g ll 1.2 , (9. 1.6) because T is a bounded operator on L2 . Moreover, I Ig ll 1. 2 = fa dx I g(xW + 1 dx I g(xW 2 a fa dx I/ (x) 1 + � I Q k l � I k l � k dy l(y) 1 �
I
{::::::>
[� [� 3 ,k
3 ,k
I ( I, tPj ,k W I tPj ,k(X W
]
00 ,
1 /2
E V(JR)
I ( I, tPj ,k W Tj X!2ik , 2i ( k + l )] (X)
]
1 /2
E V (JR)
.
For a proof that these are indeed equivalent characterizations of V' (JR) , see Meyer (1990). Similarly, wavelets provide unconditional bases and characterizations for many other functional spaces. We list a few here, without proofs. The Sobolev spaces W8 (lR) .
The Sobolev spaces are defined by
CHARACTERIZATION OF FUNCTIONAL SPACES
299
Their characterization by means of wavelet coefficients is
f E WS (lR) {::} L I ( I, 1/1j , k ) 1 2 (1 + T 2j S ) j,k
The Holder spaces CS (lR) . For 0 < CS (JR)
=
For s = n + Sf , 0 < S f
{
I S; C for ] � 0, but the sum over ] between ]o < 0 and 0 then still leads to a term in Ix Xo I I In Ix Xo That is why one has to be more circumspect for integer a , and why the Zygmund class enters.
f,
-
'IjJ
- I I. 'IjJ
3. Theorems 9.2.1 and 9.2.2 are also true if has infinite support , and and have good decay at 00 (see Jafl'ard ( 1989b) ) . Compact support for 'IjJ makes the estimates easier. 0
'IjJ'
Local regularity can therefore be studied by means of wavelet coefficients. For practical purposes, one should beware, however: it may be that very large values of ] are needed to determine a in (9.2.5) reliably. This is illustrated by the following example. Take
f (x - a)
2 e -1 x -al
e -1x-al e - (x-a) [(x - a - 1 ) 2 + 1]
x x
if S; a - I , if a - I S; S; a + 1 , if � a+1 ;
x
this function is graphed in Figure 9 . 1 (with a = 0) . This function has Holder exponents 0, 1, 2 at x = a 1, a, a + 1 , respectively, and is Coo elsewhere. One can then, for each of the three points o = a - 1, a, or a + 1, compute
-
X
302
CHAPTER
9 f(
1 .0
x)
0.5
O L-�----�____-L____-L____�� -2 -1 o 2
F I G . 9 . 1 . This function is f" are discontinuous.
Coo
except at x
=
-1, 0
and
1,
where, respectively, f, f' , and
Aj = max { ! (I, 'If- j ,k) ! ; xo E support ('If -j , k ) } , and plot log Aj / log 2. If a = 0, then these plots line up on straight lines, with slope 1/2, 3/2 and 5/2, with pretty good accuracy, leading to good estimates for Q. A decomposition in orthonormal wavelets is not translation invariant, however, and dyadic rationals, particularly 0, play a very special role with respect to the dyadic grid {2- j k; j, k E fl.} of localization centers for our wavelet basis. Choosing different values for a illustrates this: for a = 1 / 1 28, we have very different (I, 'lfj ,k) , but still a reasonable line-up in the plots of log Aj / log 2, with good estimates for Q; for irrational a, the line-up is much less impressive, and determining Q becomes correspondingly less precise. All this is illustrated in Figure 9.2, showing the plots of log Aj / log 2 as a function of j, for Xo = a - I , a, a + 1 and for the three choices a = 0, 1 / 1 28 and - 1 1 /8 (we subtract 1 1 /8 to obtain a close to zero, for programming convenience) . To make the figure, ! (I, 'If j ,k) ! was computed for the relevant values of k and for j ranging from 3 to 10. (Note that this means that f itself had to be sampled with a resolution 2 - 17 , in order to have a reasonable accuracy for the j = 10 integrals. ) For a = 0, the eight points line up beautifully and the estimate for Q + is accurate to less than 1.5% at all three locations. For a = 1 / 1 28, the points at the coarser resolution scales do not align as well, but if Q + is estimated from only the finest four resolution points, then the estimates are still within 2%. For the irrational choice a = - 1 1/8 no alignment can be seen at the discontinuity at a - I (one probably needs even smaller scales) , and the estimate for Q + at a, where f is Lipschitz, is off by about 13% (interestingly enough, the estimate would be much better if the scale10 point were deleted) ; at a + 1 , where f' is Lipschitz, the estimate is within 2.5%. This illustrates that to determine the local regularity of a function, it is more useful to use very redundant wavelet families, where this translational non-invariance is much less pronounced (discrete case) or absent (continuous case) . (See Holschneider and Tchamitchian ( 1990) , Mallat and Hwang (1992) .) Another reason for using very redundant wavelet families for the characterization of local regularity is that then only the number of vanishing moments of 'If limits
v'2
�
�
�
v'2
303
CHARACTERIZATION OF FUNCTIONAL SPACES
SLOPE = -0.50522
3
a=O
2 1 0 -1 C\I Ol 0 ,
« Ol
.Q
SLOPE = -1 .49477
0
a=O
-5 -10 0 -5 -1 0 -1 5 -20 -25
~ 4
6
10
8
INDEX j OF THE SCALE SLOPE =
3
4
-0. 50596
a = 1 /1 28
2
0
C\I
Ol 0 , « Ol
.Q
a = -J2 - 1 .375
3
2
0
SLOPE = - 1 . 50367
2 0 -2 -4 -6 -8
a = 1 /1 28
SLOPE = - 1 . 70 1 84 a = -J2 - 1 .375
0 -5 -10
SLOPE = -2.45071
-5
a = 1 /1 28
-10
SLOPE = -2.44 1 46 a = -J2 - 1 .375
-5 -10 -15
-15
-20
-20 4
6
8
10
��--�----�--�
4
INDEX j OF THE SCALE
6
8
10
FIG. 9.2. Estimates of the Holder exponents of f(x - a) (see Figure 9 . 1 ) at a - I (top) , a (middle) , a + 1 ( bottom) , computed from log Aj / log 2, for different values of a. ( This figure was contributed by M. Nitzsche, whom I would like to thank for her help.)
304
CHAPTER
9
the maximum regularity that can be characterized; the regularity of 1/J plays no role ( see §2.9 ) . If orthonormal bases are used, then we are necessarily limited by the regularity of 1/J itself, as is illustrated by choosing f 1/J. For this choice we have indeed ( J, 1/J -j , k ) 0 for all j > 0, all k; it follows that with orthonormal wavelets we can hope to characterize only regularity up to Cr - f if 1/J E cr . =
=
9.3. Wavelets for L l ( [O, 1]). Since L l -spaces do not have unconditional bases, wavelets cannot provide one. Nevertheless, they still outperform Fourier analysis in some sense. We will il lustrate this by a comparison of expansions in wavelets versus Fourier series of L 1 ( [0, I ]) -functions. But first we must introduce "periodized wavelets." Given a multiresolution analysis with scaling function ¢ and wavelet 1/J, both with reasonable decay ( say, 1 ¢(x) l , 1 1/J(x) I ::; C (I + Ixl)-2- f ), we define
¢r,"{ (x)
=
L ¢j, k ( X + i) ,
1/Jj,"{
iEZ
=
L 1/Jj , k (X + £) ;
iEZ
and
Vjper
=
Span {¢r.�r ; k E Z},
Since LiEZ ¢(x + £) 1,6 we have, for j � 0, ¢r,"{ (x) 2 -j /2 p j j j 2 /2, so that the Vj er, for j � 0, are all identiL i ¢(2- x - k + 2- £) cal one-dimensional spaces, containing only the constant functions. Similarly, because Li 1/J(x + £/2) 0,7 Wrr {O} for j � 1. We therefore restrict our attention to the Vjper, Wrr with j ::; O. Obviously Vjper, Wrr c VJ�� , a property inherited from the non-periodized spaces. Moreover, Wrr is still orthogonal to VJer, because =
=
10 1 dx 1/Jj,"{ (x) ¢r,"{, (x) L Tj
i,i'EZ
1
10 dx 1/J(Tj x + 2-j £ - k) ¢(2 -j x + 2 -j £, - k')
i' + 1 L 21j l 1i'r dy 1/J(2 1j 1 y + 21i 1 (£ - £') - k) ¢(2 1i1 y - k' ) i,i' EZ
L ( 1/Jj , k + 2 Ii l r , ¢j , k ' )
rEZ
( because j ::; 0 ) 0.
It follows that, as in the non-periodized case, VjP�� = Vjper EI1 Wrr• The . · dImen · · ce 'f'-I,p,ekr m Ii l = 'f'-I,p,erk clor m E ILJ'71 , spaces Vjper , Wjper are a11 fimteSlOnal : s m j j + 2 and the same is true for 1/Jper , both VJer and WJer are spanned by the 21 jl functions obtained from k = 0, 1, · · · , 21i1 - 1. These 21 j l functions are moreover
30 5
CHARACTERIZATION OF FUNCTIONAL SPACES
orthonormal; in e.g. , Wrr we have, for 0 ::; k, k ' (1/J),"; , 1/J),";, ) =
L
::;
21 j l - 1,
(1/Jj,k+21i 1 r , 1/Jj ,k' ) = 8k,k'
rEZ
We have therefore a ladder of multiresolution spaces, with successive orthogonal complements wg er (of vter in V��r) , Wrr , . . . , and 21 j l orthonormal bases { 0/A.)' , k " k = 0 2 U I l } in VJper ' {-"0/)' " k ' k = 0 " e L 2 ([0, 1]) (this follows again from I } in Wrr . Since Uj E _NVjp r the corresponding non-periodized version) , the functions in {¢�� } U { 1/Jr,,,; ; -j E N, k = 0" " , 21 j l - I } constitute an orthonormal basis in L 2 ( [0, 1] ) . We will relabel this basis as follows: ·
gO (x ) g l (x) g2 (X) g3 (X )
·
.
·
·
-
·
·
-
1 = ¢��� (x )
g4 (X)
1/J�,"c{ (x) 1/J�,e{,o (x ) 1/J�e{, l (X) 1/J��, o (x)
g2i (x)
1/J���o (x)
g2i +k ( X)
1/J�j,k ( x)
1/J -p er1 , 0 ( X
_ .1 )
2
g2i (x k T j ) -
g2 (X -
�)
for 0 ::; k ::; 2j - l
Then this basis has the following remarkable property. THEOREM 9 . 3 . 1 . II I is a continuous periodic function with period 1, then there exist an E C so that
( 9.3. 1) Proof.
1. Since the gn are orthonormal, we necessarily have an SN by N- 1 L (f, gn ) gn . SNI n =O
(f, gn ) . Define
306
CHAPTER
9
In a first step we prove that the SN are uniformly bounded, i.e. , (9.3.2)
with C independent of l or N. 2. If N
=
2j , then S2j
Projvp�r ; hence
=
-3
21j l _ 1 ( S2j f) (X)
( I,
L
=
k =O
. correspond exactly to the wavelet coefficients ( F, 'lIr' n ) , with F In an image, horizontal edges will show up in d1 , h', vertical edges � n1 , in d v , diagonal edges in d1 , d, as illustrated in the image example below. (This justifies the h, v, d superscripts.) Note that if the original image consists of an N x N array, then (apart from border effects; see also §1O.6) , every array d1 , >. consists of I¥- x I¥- elements, and can therefore be represented by an image (magnitudes of the coefficients corresponding to grey levels) of one quarter the size of the original. The whole scheme can therefore be represented as in Fig ure 10.2. Of course, one can decompose even further if more multiresolution layers are wanted. Figure 10.3 shows this decomposition scheme on a real image, with three multiresolution layers. =
c�. more apparent. I would like to thank M. Barlaud for providing this figure.
317
GENERALIZATIONS AND TRICKS
eralizations of (5. 1 . 1 )-(5. 1.6)) in which Vo is not a tensor product of two one dimensional Vo-spaces. 1 Some (but not all!) of the constructions done in one dimension can be repeated for this case. More precisely, the multiresolution structure of the Vj implies that the corresponding scaling function cI> satisfies (10. 1 . 1 ) for some sequence (hn )n E 2? Orthonormality of the cI>o; n forces the trigonomet ric polynomial (10. 1.2) to satisfy To construct an orthonormal basis of wavelets corresponding to this multireso lution analysis, one has to find three wavelets w 1 , w 2 , w 3 in V - 1 . orthogonal to V0 and such that the three spaces spanned by their respective integer translates are orthogonal; moreover the W A ( . - n ) should also be orthonormal for each fixed A. This implies that
where the m 1 , m2 , m3 are such that the matrix
m 1 (e , ( )
mo (e , ( ) mo (e +
11" ,
()
m 1 (e +
mo (e, ( + ) 11" ,
()
m2 (e +
( + ) m 1 (e 11"
+ (+ 11" ,
11" ,
m3 (e, ( ) ()
m3 (e +
11"
) m2 (e
11"
+ (+ 11" ,
11" ,
()
m3 ( e , ( + )
m2 ( e , ( + )
m 1 ( e , ( + 1I")
11"
mo ( e +
11" ,
m2 (e , ( )
11"
) m3 ( e
+
11"
11" ,
(+ ) (10. 1.4) 11"
is unitary. The analysis leading to this condition is entirely similar to the one dimensional analysis in §5. 1; see, e.g., Meyer (1990, §III.4) . 2 Note that the number of wavelets to be constructed can be determined by an easy trick. In two dimensions, for example, Vo is generated by the translates of one function cI>(x, y) , over 1;2 ; the space V - 1 is generated by the translates of cI>(2x, 2y) over � Z2 , or equivalently, by the Z2 -translates of the four functions cI>(2x, 2y), cI>(2x - 1 , 2y) , cI>(2x, 2y - 1) , cI>(2x - 1, 2y - 1). V - 1 is therefore "four times as big" as Vo. On the other hand, each of the Wb -spaces is generated by the Z2 -translates of a single function wj (x, y) , and is therefore "of the same size" as Vo. It follows that one needs three ( = four minus one) spaces Wb (hence three wavelets wj ) to make up the complement of Vo in V - 1 . This rule may sound like "hand-waving," but we can also rephrase (and prove) it in more
318
CHAPTER
10
mathematical terms: the number of wavelets is equal to the number of different cosets (different from 712 itself) of the subgroup 712 in the group � 712 . In the general n-dimensional case, the same rule shows that there are 2n - 1 different functions mj to determine; they have to be such that the 2n x 2n _ dimensional matrix
(10.1.5) is unitary, with r = 1" " , 2n , and s = (S l , " " s n ) E {0, 1} n . 3 In fact the unitarity requirement of (10. 1 .4) or (10. 1.5) calls for a tricky bal ance: m 1 , m2 , m3 have to be found so that the first row of (10. 1.4) has unit norm, which seems harmless enough, but we also simultaneously need orthogl> nality with and among the other rows, which are all shifted versions (in e or ( ) of the first row. These correlations between the rows may be hard to juggle in practice. It is useful to untangle them first, which can be done via the sl>-called polyphase decomposition. We write, e.g., 2mo (e , () = mo , o (2e , 2 () + e - i� mo ,l (2e , 2( ) + e - i (mo , 2 (2e, 2 ( ) + e - i (H ( ) mo ,3 (2e, 2( ) ;
ml ,j , j = 0" " , 3, are defined similarly from mi, f = 1" " , 3. One easily checks that (10. 1.3 ) is equivalent to I mo , o (2e , 2 ( ) 1 2 + I mO, 1 (2e , 2( W + I mO, 2 (2e, 2 ( W + I mO,3 (2e , 2 ( ) 1 2 = 1 .
Similarly, all the other conditions ensuring unitarity of ( 10. 1.4) can be recast in terms of the ml ,j ; one finds that (10.1.4) is unitary if and only if the polyphase matrix mO, o (e , ( ) m 1 , O ( e , ( ) m2 , O ( e , ( ) m3,O (e, ( )
mo ,l (e, ( ) m 1 , 1 (e , ( ) m2 , 1 (e , ( ) m3, 1 (e , ( ) mO , 2 (e , ( ) m 1 , 2 (e , ( ) m2 , 2 (e , ( ) m3, 2 (e , ( ) mo ,3 ( e, ( ) m 1 ,3 (e , ( ) m2 ,3 (e , ( ) m3,3 (e, ( )
(10. 1.6)
is unitary. In n dimensions, one similarly defines 2 n /2 mr (6 , · · · , en ) = L e - i ( S 1 6 + ··· + s n�n) mr, s (26 , · · · , 2en ) , SE{O,l}n and the unitarity of U is equivalent to the unitarity of the polyphase matrix {; defined by (10. 1.7) The construction therefore boils down to the following question: given mo (from (10. 1 . 1) , (10. 1.2)), can m l , " ' , m2n - 1 be found such that (10. 1.6) is uni tary? In the tWl>-dimensional case, and if mo (e , ( ) happens to be a real trigono metric polynomial, then one can even dispense with the polyphase matrix: it is
319
GENERALIZATIONS AND TruCKS
easy to check that the choice m i (e, ( ) = e-ie mo (e + 11" , ( ) , m2 (e , ( ) = e-i (HC; ) mo(e, ( + 11" ) , m3 (e, ( ) = e - iC; mo ( e + 11" , ( + 11" ) makes (10.1.4) unitary. If mo is not real, then things are more complicated. At first sight one might even think the task is impossible in general in the n-dimensional situation, where (10. 1.7) is a 2n x 2n -matrix: after all, we need to find unit vectors, depending continu ously on the ei (namely the second to last columns of (10. 1.7)), orthogonal to a unit vector (the first column of (10. 1.7) ) , i.e., tangent to the unit sphere. But it is well known that "it is impossible to comb a sphere," i.e., there exist no nowhere-vanishing continuous vector fields tangent to the unit sphere, except in real dimensions 2, 4, or 8. The first column in (10. 1 .7) does not describe the full sphere, however; in fact, because it is a continuous function of n variables (the 6 , " " en ) in a 2n -dimensional space, and 2 n > n, it only describes a compact set of measure zero. This fact saves the day and makes it possible to construct m l , ' . . , m2n- 1 , as shown by Gr6chenig (1987) ; see also §III.6 in Meyer ( 1990) . Gr6chenig ' s proof is not constructive; a different, constructive proof is given in Vial (1992) . Unfortunately, these constructions can not force compact support for the 'l1i : even if mo is a trigonometric polynomial (only finitely many hn =I- 0) , the mi are not necessarily.
10.2. One-dimensional orthonormal wavelet bases with integer dilation factor larger than 2. For illustration purposes, let us choose dilation factor 3. A multiresolution anal ysis for dilation 3 is defined in exactly the same way as for dilation 2, i.e. , by (5.1.1)-(5. 1.6) , except that (5.1 .4) is replaced by
We can use the same trick again as above: Vo is generated by the integer translates of one function, i.e., by the ¢(x - n) , while V-I is generated by the ¢(3x - n) , or, equivalently, by the integer translates of three functions, ¢(3x), ¢( 3x - 1), and ¢(3x - 2) . V-I is "three times as big" as Vo , and two spaces of the "same size" as Vo are needed to complement Vo and constitute V-I : we will need two spaces WJ , W6, or two wavelets, 'lj; 1 and 'lj; 2 . We can again introduce mo, m! , m2 by
� (e) = mo(e / 3) � (e / 3),
£ = 1, 2 .
Orthonormality of the whole family {¢O,n , 'lj;6 ,n ' 'lj;5,n ; n E il}, where ¢i ,n is now defined by ('lj;J. n are defined analogously), again forces several orthonormality conditions on the ml , which can be summarized by the requirement that the matrix
320
CHAPTER
10
( 2; ) ml (� + 2; ) m2 (� + 2; ) mo (� + � ) m l (� + � ) m2 (� + � ) mo � +
(10.2. 1)
is unitary. Again, one can restate this in terms of a polyphase matrix, removing the correlations between the rows. Explicit choices of mo, m l , m2 for which (10.2.1) is indeed unitary have been constructed in the ASSP literature (see, e.g., Vaidyanathan (1987)). The question is then again, as in Chapter 6, whether these filters correspond to bona fide L 2 -functions ¢, 'lj; l , and 'lj;2 , whether the 'lj;J,k constitute an orthonormal basis, and what the regularity is of all these functions. We know, from Chapter 3, that 'lj; l and 'lj; 2 must necessarily have 0 integral zero, corresponding to m l ( O) m2 (0). Since the first row of (10.2. 1) must have norm 1 for all �, it follows that mo(O) 1 (which is necessary anyway for the convergence of the infinite product n;: l mo( 3 -j �) which defines ¢(�)). The first column of ( 10.2. 1) must also have norm 1 for all �, so that mo(O) 1 implies moe311" ) 0 mo( 4;), i.e. , mo (�) is divisible by l± e - i�±e - 2iE . If, moreover, any smoothness for 'lj; l , 'lj; 2 is desired, then we need additional vanishing moments of 'lj; l , 'lj; 2 , which by exactly the same argument as before, lead to divisibility of mo(�) by ((1 + e-ie + e-2ie)j3)L if 'lj;\ 'lj; 2 E CL - 1 . One is thus led to looking for mo of the type mo(�) ((1 + e-ie + e-2ie)j3) N £ (0 such that Imo(�)12 + Imo(� + 2; ) 12 + Imo(� + 4; ) 12 1. If mo is a trigonometric polynomial, this means that L 1£12 is again the solution to a Bezout problem. The minimal degree solution leads to functions ¢ with arbitrarily high regularity; however, the regularity index only grows logarithmically with N (L. Villemoes, private communication) . 4 Once mo is fixed, m l and m2 have to be determined. The design scheme explained in Vaidyanathan et al. (1989) gives a way to do this. In this scheme, the matrix (10.2. 1) (or rather, its z-notation equivalent) is written as a product of similar matrices the entries of which are much lower degree polynomials, with only a few parameters determining each factor matrix. 5 If one imposes that the first column of a product of such matrices is given by the mo we have fixed, then the values of these parameters are fixed likewise, and mb m2 can be read off from the product matrix.6 If the compact support constraint is lifted, then other constructions are pos sible. In Auscher (1989) one can find examples where ¢ and 'lj;" are Coo functions with fast decay (and infinite support). One final remark about dilation factor 3. We have seen that mo must neces sarily be divisible by ( 1 + e-ie + e-2ie)j3. This factor does not vanish for � = 7r (unlike the factor (1 + e-ie)j 2 for the dilation factor 2 case). However, if we want to interpret mo as a low-pass filter, then mO( 7r) 0 would be a good idea. To ensure this, we need £(7r) 0, which means going beyond the lowest degree solution to the Bezout equation for 1£12. =
=
=
=
=
=
=
=
=
=
=
GENERALIZATIONS AND TRICKS
321
Similar constructions can be made for larger integer dilation factors. For non-prime dilation factors a, one can generate acceptable ml from constructions for the factors of a, although not all possible solutions for dilation a can be obtained in this way. For a = 4, e.g., one can start from a scheme with dilation 2 and filters mo and m I , and one can define the filters mo , ml , m2 , m3 (still orthonormalj the - distinguishes them from the dilation factor 2 filters) by
mo (O
=
mo (�)mo (�/2) ,
ml (�) = mo (�)ml (�/2) ,
m2 (�) = ml (�)ml (�/2) , m3 (�)
=
ml (�)mo (�/2) .
(It is left to the reader as an exercise to prove that this leads indeed to an orthonormal basis. One easily checks that the 4 x 4 analogue of (10.2. 1) is unitary.) Note that the function ¢ is the same for the factor 4 and the factor 2 constructions! We will come back to this in §1O.5.
10.3. Multidimensional wavelet bases with matrix dilations. This is a generalization of both §1O. 1 and §1O.2: the multiresolution spaces are subspaces of L 2 (JRn ), and the basic dilation is a matrix D with integer entries (so that DZn C zn ) such that all its eigenvalues have absolute value strictly larger than 1 (so that we are indeed dilating in all directions) . The number of wavelets is again determined by the number of cosets of DZn j one introduces again mo, m I , · · · , and the orthonormality conditions can again be formulated as a unitarity requirement for a matrix constructed from the mo, m I , . . . . The analysis for these matrix dilation cases is quite a bit harder than for the one dimensional case with dilation 2, and, depending on the matrix chosen, there are a few surprises. One surprise is that generalizing the Haar basis (i.e., choosing mo so that all its nonvanishing coefficients are equal) leads in many cases to a function ¢ which is the indicator function of a selfsimilar set with fractal boundary, tiling the plane. For two dimensions, with D = ( I -I ), e.g., one finds that ¢ can be the indicator function of the twin dragon set, as shown in Gr6chenig and Madych (1992) and Lawton and Resnikoff (1991) . Note that such fractal tiles may occur even for D = 2 Id if mo is chosen "non-canonically" (e.g., mo (�, ( ) = i (1 + e- i( + e- i (H() + e- i (H 2() ) in two dimensions-see Grochenig and Madych (1992)). For more complicated mo (not all coefficients are equal) , the problem is to control regularity. Zero moments for the 'ljJj do not lead to factorization of mo in these multidimensional cases (because it is not sufficient to know zeros of a multi-variable polynomial to factorize it), and one has to resort to other tricks to control the decay of J. A particularly interesting case is given by the "quincunx lattice," i.e., the two-dimensional case where DZ2 = {(m, n) j m + n E 2Z}. In this case there is only one other coset, and therefore only one wavelet to construct, so that the choice for ml is as straightforward as it was for dilation 2 in one dimension. The
322
CHAPTER
10
conditions on mo , m l reduce to the requirement that the 2
( mo (�, () mo (�
+ (+ 1r ,
1r
x
2 matrix
)
be unitary. It is convenient to choose
Note that any orthonormal basis for dilation factor 2 in one dimension auto matically gives rise to a pair of candidates for mo, m l for the quincunx scheme: it suffices to take mo (�, () = mt (�) (where mt is the one-dimensional filter). 7 Different choices for D can be made, however. Two possibilities studied in de tail in Cohen and Daubechies (1993b) and KovaCevic and Vetterli (1992) are D l = n n and D2 = n n The same choice for mo leads to very differ ent wavelet bases for these two matrices; in particular, if one derives, via the mechanism explained above, the filter mo from the "standard" one-dimensional wavelet filters N mO in §6.4, then the resulting ¢ are increasingly regular if D2 is chosen (with regularity index proportional to N) , whereas choosing D l leads to ¢ which are at most continuous, regardless of N. Other choices for D may lead to yet other families, with different regularity properties again. One can of course also choose to construct two biorthogonal bases rather than one orthonormal ba sis, as in §8.3; for the choices D1 , D2 several possibilities are explored in Cohen and Daubechies (1993b) and KovaCevic and Vetterli (1992) . In this biorthogo nal case, one can again derive filters from one-dimensional constructions. If one starts from a symmetric biorthogonal filter pair in one dimension, where all the filters are polynomials in cos �, then it suffices to replace cos � by (cos � + cos () in every filter to obtain symmetric biorthogonal filter pairs for the quincunx case. 8 Because of the symmetry of these examples, the matrices D l and D2 lead to the same functions ¢, ¢ in this case. One finds again that symmetric biorthogonal bases with arbitrarily high regularity are possible (see Cohen and Daubechies (1993b) ) . The quincunx case is of interest in image processing because it treats the different directions more homogeneously than the separable (tensor-product) two-dimensional scheme: instead of having two favorite directions (horizontals and verticals) , the quincunx schemes treat horizontals, verticals, and diagonals on the same footing, without introducing redundancy to achieve this. The first quincunx subband filtering schemes, with aliasing cancellation but without ex act reconstruction (which had not been discovered even for one dimension at the time) are given in Vetterli ( 1984) ; Feauveau ( 1990) contains orthonormal and biorthogonal schemes, and links them to wavelet bases; Vetterli, KovaCevic, and LeGall (1990) discusses the use of perfect reconstruction quincunx filter ing schemes for HDTV applications. In Antonini, Barlaud, and Mathieu (1991) biorthogonal quincunx decompositions combined with vector quantization give spectacular results for image compaction. -
_
·
�
323
GENERALIZATIONS AND TRICKS
10.4.
One-dimensional orthonormal wavelet bases with non-integer dilation factors.
In one dimension, we have so far only discussed integer dilation factors ;::: 2.9 Non-integer dilation factors are also possible, however. Within the framework of a multiresolution analysis, the dilation factor must be rationallO (for a proof, see Auscher (1989» . It had already been pointed out by G. David in 1 98 5 that the construction of the Meyer wavelet could be generalized to dilation a = � , for k E N, k ;::: 1 ; Auscher (1989) contains constructions for arbitrary rational a ( see also Auscher's paper in Ruskai et al. (1992») . Let us illustrate for a = � how the factor 2 scheme has to be adapted. We start again from a multiresolution analysis, defined as in (5. 1. 1)-(5.1.6) , with � instead of 2 for the dilation factor. We have again (
)
corresponding to n = 3£, n = 3£ + 1 , and n = 3£ + 2. The space Vo is generated by the 2Z translates of two functions, x - U) and x - 2£ - 1) , £ E Z. It follows that the complement space Wo is generated by the 2Z translates of a single func tion, Wo = Span {'IjI ( . - 2n) ; n E Z}. ( "Wo is half the size of Vo." ) We expect therefore an orthonormal basis of the type 'IjIj , k (X) = ( � ) -j/ 2 'IjI ( ( � )j x - 2k) , j, k E Z. This function 'IjI can also be written as a linear combination of the 4>(� x - n) , and orthonormality of the 'IjI (x - 2n ) , plus orthogonality with respect to the ( x - 2n) , ( x - 2n - 1) implies
4>
4>
L L n
n
9n 9n - 3£
=
8£0 ,
9n h�_3£
=
0,
� .jj,
(10.4.6)
L n
9n h;'_3£ = o .
(10.4.7)
With the definition m l ( ) = L n 9n e-inf. , the conditions (10.4.2) , (10.4.4) (10.4.7) are equivalent to the unitarity of the matrix
m8 (�) moo
(� + 327r )
I
mo
(� + 327r )
(10.4.8)
This matrix looks identical to (10.2. 1), but this similarity is deceptive: in (10.4.8) the first two columns are both given by low-pass filters, because they are both related to the scaling function (m8 (0) = 1 = mA CO)) , whereas the second column in (10.2. 1) corresponds to a high-pass filter. Such mi, m l can indeed be constructed ( see Auscher (1989) for details and graphs ) . Note that mij and m8 are closely related. The Fourier transforms of (10.4. 1), (10.4.3) are
4>
4>(�) = m0o ( 32 �) 4>. ( 32 �) , •
implying
¢(Oe - if. = m6
(� �) ¢ (� �) ,
(10.4.9)
325
GENERALIZATIONS AND TRICKS
which should hold for almost all (. If ¢ is continuous, then the following argument shows that ¢ vanishes on some intervals. Since ¢(O) = (27r) - 1/2 , there exists Q so that for 1 ( 1 ::; Q , I ¢( ( ) I � (27r) - 1/2 /2. Consequently, for 1 ( 1 ::; Q , m8 (() = e3i(, /2 mA ( ( ) , or
m8 (( + 27r) = _ e 3i (, /2 mA ( ( + 27r) ; since mg , mA are also 27r-periodic, this implies mg (( + 27r) = 0 = mA (( + 27r) for 1(1 ::; Q . It follows that I ¢( � ( + 37r) 1 = 0 for 1 ( 1 ::; Q. In particular, this means that ¢ cannot be compactly supported (compact support for ¢ means that ¢ is entire, and non-trivial entire functions can only have isolated zeros). Nevertheless, subband filtering schemes with rational noninteger dilation factors, in particular with dilation � , have been proposed and constructed by KovaCevic and Vetterli (1993) , with FIR filters. The basic idea is simple: start ing from co , one can first decompose into three subbands, by means of a scheme as in §1O.2, and then regroup the two lowest frequency bands by means of a synthesis filter corresponding to dilation 2; the result of this operation is c l , while the third, highest frequency band after the first decomposition is d i . The corresponding block diagram is Figure 10.4. If all the filters are FIR, then the whole scheme is FIR as well. But didn ' t we just prove that there does not exist a multiresolution analysis for dilation factor � with FIR filters? The solution to this paradox is that the block diagram above does not correspond to the construction described earlier. A detailed analysis of Figure 10.4 shows that this scheme uses two different functions ¢ I and ¢2 , with Vo generated by the ¢ I ( X - 2n) , ¢2 (X - 2n ), n E Z. The argument used to prove that ¢ cannot have compact support then no longer applies, and ¢ I , ¢2 can indeed have compact support. The analog of (10.4.9) is now an equation relating the two-dimensional vectors ( ¢ I (�), ¢2(�)) and ( ¢ I (� �) , ¢2 ( � �)), however, and it is hard to see how to formulate conditions on the filters that result in regularity of ¢ I , ¢2 .
�----- d l
FIG. 10.4. Block diagmm corresponding to a subband filtering with dilation factor constructed in Kovacevic and Vetterli (1993) .
�,
as
One may well wonder what the rationale is for these fractional dilation factors. The answer is that they may provide a sharper frequency localization. If the
3 26
CHAPTER
10
dilation factor is 2, then ,(j; is essentially localized between 7r and 27r, as illustrated by the Fourier transform of a "typical" 'IjJ in Figure 10.5. For some applications, it may be useful to have wavelet bases that have a bandwidth narrower than one octave, and fractional dilation wavelet bases are one possible answer. A different answer is given in Cohen and Daubechies (1993a) , summarized in the next section. 1 .0
0.5
o It
o
FIG. 1 0 . 5 .
21t
41t
31t
Modulus of 1 1O�(�) 1 , with N'I/J
as
defined in §6.4.
10.5.
Better frequency resolution: The splitting trick Suppose that hn, 9n are the filter coefficients associated to wavelet basis with dilation factor 2, i.e., 1 mo (