Scribd Logo
Upload

Welcome to Scribd!

  • VetterliH92 PDF

    Uploaded by

    Snail on a Blade
    16 views26 pages

    Original Title

    VetterliH92.pdf

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    16 views26 pages

    VetterliH92 PDF

    Uploaded by

    Snail on a Blade

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    of 26
    Wavelets and Filter Bank: : Theory and Design Martin Vetterli, Senior Member, IEEE, and Cormac Herley applied mathematics, signal processing, and computer vision, respectively, have recently converged to form a single theory. Tn this paper, we rst compare the wavelet transform with the ‘more classical short-time Fourier transform approach to signal ‘analysis. Then we explore the relations between wavelets, filter banks, and multiresolution signal processing. We brielly re- view perfect reconstruction filter banks, whieh can be used both for computing the discrete wavelet transform, and for deriving ‘continuous wavelet bases, provided Strain known as regularity. Given a low-pass Alter, we derive necessary and sulllent conditions for the existence of a com plementary high-pass filter that will permit perfect reconstruc- tion. Posing the perfect reconstruction condition as a Bezout to design highly regular filter banks, which generate blorthog- ‘onal continuous wavelet bases with symmetries. 1. Istropuction All this time the guard was looking at her, Bist ohrough @ Telescope, then through « microscope, and then through an opera pas, Lewis Caroll, Through the Looting Glass HE analysis of nonstationary signals often involves a ‘compromise between how well transitions or discon: tinuities ean be located, and how finely long-term behav~ ior can be identified. A typical example is the choice of window length in the short-time Fourier transform. In wavelet analysis one looks at the signal st different "scales" or resolutions": a rough approximation ofthe signal might look stationary, while at a detailed level (when using a small window) discontinuities become ap- parent, This multiresolution, or multiscale view of signal analysis is the essence of the wavelet transform, which has recently become quite popular [1]. The goal is “to see the wood and the trees. ‘The wavelet analysis is performed using a single pro: totype function called a wavelet, which can be thought of Marci eosved ug 24, 1950; vied June 28. 191. The woo ‘of Vener wa supported inp the Nationa Since Foundation ther Orang. ECD SETITI1, MP-S6GR577. and MP0 14189. The Wonk of. Hey nas sapped in pay te National Seine Foun Aiton under Grape ECD 861111 "ears are ith he Depaneao Leta Eanes tel Tlesommuncions Rewarh, Columbia Univers. Sew You. Nona 93 Loss serxiases 4s a bandpass filter Fine temporal analysis is done with ‘contracted (high-frequency) versions ofthe wavelet, while fine frequency analysis uses dilated (low-frequency) ver- sions. The bandpass filters have thus constant relative bandwidth or “constant-O."" The importance of constant relative bandwidth when perceptual processes like the au 1, b#0 which corresponds to a = af! and b = nip, Note that the translation step depends on the dilation, since long. ‘wavelets are advanced by large steps, and short ones by small steps. On this discrete grid, the wavelet transform is thus Kal, m) = a5) ® Of panicular interest isthe discretization on a dyadic grid, which occurs for ap = 2, by = 1. It is possible to construct functions f(-) so that the set Fal) = a5! + Whag"t — nb), mneZ, a is orthonormal. That is, hol), Hal DD = By Bu ‘A classic example is the Haar basis (which is not conti wows, but is of interest because of its simplicity), where: 1 osr made up of ¢x) and its Integer translates, then, we can find coeficien's cy such that 60) = Bg bx ~ ny as, because Vp C Ys. Then E_CWe nr m 06) and its integer translates form an orthonormal basis for Wy, So (Yu) = 2/°VQ%e j,i, j €Z) constitutes an conthonormal basis for L?(R), following (11). The impli cation of (15) is given graphically in Fig, 7 where itis shown how a sealing function ean be obtained from a lin- fear combination of its sealed versions. The first example is for illustration purposes only, while the second is an actual scaling function from a wavelet bass C. The Discrete Time Case Assume now that we deal with discrete time sequences x(n), m/€ Z which are square summable; that is, the space F°@). Now, let us derive a coarse half-resolution approx- imation to the original sequence. This can be done by fl- tering with a halfband low-pass filter, followed by sub- sampling by 2 (discarding every odd-indexed sample). In matrix notation, and assuming forthe sake of simplicity thatthe filter is FIR, the convolution witha filer having impulse response (hy(0), fy(I), = , hy(L ~ 1) followed by subsampling by 2, corresponds to matrix multiplica- tion of the infinite signal vector xf «+ + x(—1), :€0),x(0), by hl ~ 1) afl. ~ 2) Ho ° 0 hele Assume further that the impulse response and its shifted versions by even shifts (i.e, the rows of the above ma- trix) form an orthonormal set, that is, Chol = 21), fn = 2) = by ke Le Z. In matrix notation the equivalent of (18) is Ho ‘The projection of the original sequence x(n) onto the sub- ais) We} Wy Wo mere a » Fig. 5. Meal division of spectra using sine ter.) Dison imo V, fact. Note that CV, 0) Dison io W, sacs, space spanned by the rows of Hy is given by HS Hy [Note that multiplication by H corresponds to upsampling by 2 followed by convolution with a filter having impulse response fin) = thy(L ~ 1), hall ~ 2). °°" s hal). ‘h(O)], (that is the time reversed impulse response of ‘hn)). Note also that in order for the set {hgin ~ 24), & ‘€ Z} to form an orthonormal basis has to be even, For if L were odd, then (y(n), f(t = L + 1)) = holOVho(L = 1) # Ounless either ho(0) oF hy(L ~ 1) is zero. Call Vy the space [°(2) and call Vy the subspace of V. spanned by the rows of Hp. Then call Wy the orthos- ‘onal complement of Vy in Vy va Vy ® My, Now the filter with impulse response fin) = (—1)" Ag — 1 ~ n) and its even shifted versions form an ot- thonormal basis for Wy. First note that orthogonality of ‘(n) and hy(n) with respect to even shifts is easly veri- ng 0 0 = 1) + eg) af) lO) a fied, because of the sign change in y(n) hin = 2), hla = 2) =0 kyle Z (19) In matrix notation calling HY, a matrix based on h(n) in the same way that Hin (17) was based on hy(n) we have HoH = 0 Then from the onthonormality of thus Vo 4 Fi, 6 isison of spectrum with areal wane a} Dyalie secs ofthe scaling function, () Fourier transform of the fiched scaling funtone neating te nesting of he spaces (e) Dye sche of be mal. (Fear estar of "Re trteed nels ning te aangarien of the Hopes. : i) =| Ale Fig. 7. Sealing anton) 8 iar combvnaton of ssl and sie versions (2x ~ (2) Hat anton example this "no amonhononnal xu) ()Exampe based oes mga ropa a fiter 12) ‘agin ~ 21) i follows Gyn = 2, hyn = hy = by ke LEZ since the sign change in h(n) is cancelled. In other words, based on an orthonormal basis for Vp, we were able to ‘construct one for Wo, just asin the continuous ease. Again AT 20 and the projection of the original sequence onto My is sven by WH, Note tha since we have projections onto orthonormal and ‘complete subspaces: HSH, + HEH, = 1 Iki also clear that by interleaving Hy and Hy, we obtain a block Toeplitz orthonormal mateix 7: gl ~ 1) halle a nl = 1) hy ~2) o 0 yt o 0 me Which satisfies T= TT = 1 that is, the two fiter impulse responses fy(n) and h(n), together with their even translates, form an orthonormal basis for 172). Because this is such a fundar-ental con- copt we will illustrate it with an example. Instead of the rather trivial case where hg(n) = 1/-V2 - [1,1] and Ay(n) = 1/2 [1, =1] (which corresponds to the Walsh-Had- amard transform on successive blocks of 2 samples), we will choose fin) = [apa}, dob, bobs, ~byay] where a, c0s aj and b, = sin a, Because aj + b? = | itis clear that Iigtnly = 1. Also Ciy(n + 2), hy(a)> = O-and thus ‘h(n and its even translates form an orthonormel set. Now choosing in(n) = {Boay, byby, —doby, aya] makes y(n + 21) orthogonal to y(n) while keeping it of norm 1, and ‘onthogonal to its even translates, Thus the set [yi + 21), y(n + 2k); 1, k € Z} is an onhonormal basis for PQ). In Fig. 8, we recapitulate the above relations in the usual digital signal processing notation, using filters and sampling rate changes to denote the operators used so fa: First the inner products Hy » x and HY «x are calculated Then the projections onto Yq and Wo are found. Finally, the original signal is reconstructed by adding the projec= tions from the two orthogonal subspaces, ‘What worked once will work again; that is we can de- {Ze oof Fie. 8, Decomposition of Yn Vp and W, using mine Sten, and Feeombinaiontoachve perc meonmstion HGH, athe projection ‘the gnome Vand MPH = isthe projection ont My ‘compose Vp into V; and 1%, and so on. Hence we find Woven F=01, Yiu @W, F=O1 from which we have woKVewevs and W@W, BW, Bs Oa nO 0 0 ig2) gl) hgO) QA) yO) -1) y+ ‘That is, the direct sum of all the W's is the space of square summable functions 1°(Z). Note that in contrast to the ‘continuous case there is a “*maximum” resolution in the dliserete case given by V_, which is determined by the original sampling rate. The decomposition of V_, into Wo, W), Wa, ete, is es sentially @ wavelet transform on diserete sequences, since it splits the original space in two, and then splits one of the resulting half spaces in two, etc. Ths is shown in Fig, 9 using filters and subsamplers. Actually, if the filter ‘oln) isthe ideal halfband low-pass fiter given by (12), then fin) is the ideal halfband high-pass filter. Thus if Vis the space of functions band limited to (—2", 22) as considered in the discussion of multiresolution analysis in Section II-B, then Vand Ware the spaces of functions band limited to (—x, x) and (—2x, —) U (x, 2m), re- spectively. That is, by iteration, the diserete system in Fig. 9 computes exactly the discrete wavelet transform into octave bands (although with noncausal infinite im- pulse response filters). [Note that the two concepts of scale and resolution i troduced above, while related, are not the same. The no~ tion of resolution of a signal is related to its bandwidth This holds also in the sampled domain, but it is best thought of as the bandwidth of the equivalent continuous time signal, This definition indicates that an oversampled version will not have more resolution than a critically Fig. 9 Discrete wavelet ransom on seqdnses. Thalia lowpass ‘nd hppa tre and) respetivly and? Lea sabe ‘Seog y 2 Soppng samples with odes) we LO te $3 sampled version of the same signal. The notion of scale is related (0 the size of a signal. We will adhere to the convention used in the wavelet literature, that large scale denotes contraction ofthe signal, while small scale stands for a dilated signal. Fig. 10 shows various multirate si- nal processing operations, and their effect upon resolution and scale (for simplicity only changes by factors of two are considered) D. Orthogonal Pyramids and Critical Sampling In computer vision and image coding [8], [9]. 9 suc- cessive approximation or multiresolution technique called an image pyramid is sometimes used. It consists of deriv- ing a low resolution version ofthe original, then predit- ing the original based on the coarse version, and finally taking the difference between the original and the predic- tion Se Fig. 11(a). Ache reconstruction, the prediction is added back wo the diference,guarantecing perfect re- construction. A shortcoming ofthis scheme isthe over sampling, since we end up with a low resolution version and a fall resolution difference (at the initial rte). We bretly show below that, ifthe system is linear and the low-pass filter is orthogonal to its even translates, then one can actually subsample the difference signal ater fil- tering i. In that case, the pyramid reduces exactly to a critically subsampled orthogonal subband coding scheme like the one discussed in the previous subsection First, the prediction ofthe orginal based on the coarse version is simply the projection onto the spave spanned by {h(n — 21), 162}. That i, calling the predition Hi Hy x. ‘The ditfrence signal is thos aad HH) Bt, because its perfect reconsrition system 1 HSH = HH, that is, d is the projection onto the space spanned by th(-n ~ 2,162}. Therefore, we can flter and sub- sample d by 2, since, following (20) AHH, = Hy, In that case, the redundancy of d is removed (d is now critically sampled), and the pyramid is equivalent to an orthogonal subband coding system (see Fig. 11(b)). “0 ema —— Seas — a “le! oo eee eae —o{ Se} eee Fig. 10. Relation and sae changes in dete tne by facta of 2 te vale of gel dened sm geopapies! maps.) Hal itrng reduces the resoaion by? cle eachanged) iowpus Seng floes by sampling by? doubles the teu fad halve he esto a (a) c) Upsampting 6) 2 fale ‘5 hlud ow-pos Sitenng aves he scale estton sundae) b= aie (Meee ee zt ee =) a Fig. 1. Pyamil scheme ievoling a couse low-pass appoxination, and 1MiteraceSetwecs te cou pprosiaation ad the ginal (0) Over Sample (Cray sole ease, whch isequvalet oa soba Hae GY Geb = [ne HDD @) [om =e} —@)— ie Le }@ Fig. 12 Derivation of he euitlet teat le, (a) Subsumplng by 2 efor ter) canbe writen seen by Is) flowed by aban Ping.) Cascade ofits ech flowed by subsanpling by 2 (0 Eunale er, flowed by sbsampling by 2 The signal d can be reconstructed by upsampling by 2 and filtering with fy(n). Then we have Hy HTH) «x = HEH, and this, added to = Hf Hy - x is indeed equal tox. In the notation of the multiresolution scheme described in Section I-C the prediction ¥ is the projection onto the space Vy, and d isthe projection onto We. We have thus a shown that pyramidal schemes can be critically sampled as well. A similar analysis is given in [29], It interest- ing to note that although pyramid coding schemes have ‘been in use for a decade or So, it had not been areviously noted that they could be subsampled under the conditions above. The discussion of the more general, bingthogonal tase (see also Section II-C) i given in Appendix 4.1 and p24) E. FIR Filter Banks and Compactly Supported Wavelets ‘We now briefly show the connection betwee wavelets of finite length and filter banks, as originally irvestigated bby Daubechies [12]. Note that it is common to use the term compact support instead of finite length. First, as sume we have an orthonormal basis of such functions (x) ‘and ¥(x) which obey two-scale difference equations as in (15) and (16) (or +1, 6 +MY = by en (Yer +, Yer + DY = by a) or +). ver tH) =O. @) ‘We will show these relations lead to perfect reconstruc- tion FIR filterbank. The finite support of 6(2) means that itcan be written a finite linear combination of the terms (2x — m); that is, finitely many of the c, are different from zero. From (21) we get (ex = 1, 62x ~ BY $y ea Now, using (15) and (24), 21) ean be written as (with n= 21, and m’ =m ~ 2k): (oar + D, ote +) (Eeeer +2 =n), Dede + Uk — m) (Bevo ~ 1), Bow ns 00) 5 o from which it follows that lc nother words the discrete filter, with impulse response hifn) = c,/V2 is orthogonal to its even translates, and with hm)"= (= Dhl. ~ m ~ 1) we obtain an onthogonal perfect reconstruction FIR filterbank with orthogonal im= pulse responses. Thus, compactly supported wavelet bases lead to perfect reconstruction FIR filter banks. While the ‘converse does not always hold, and is not as immediate to analyze, we discuss it here because it i the basis for the construction of compactly supported wavelets [12] Considering the discrete time wavelet transform in Fig. 9, one notices thatthe lower branch is an infinite cascade of filters hn) followed by subsampling by 2. Note that subsampling the signal with z-transform X{c) by 2 results ina new signal with z-transform ¥(2): We) = 1/2 KE!) 4X27 CH) It's easily verified that subsampling by 2 followed by fi tering with H(@) is equivalent to filtering with (23) fol- lowed by subsampling by 2 (see Fig. 12(a). ‘Therefore the cascade of i blocks of filtering operations followed by subsampling by 2 is equivalent to a filter He) with z transform Ht ae") followed by subsampling by 2, We define HG) = 1 Assuming tha the filter H(2) has an impulse response of Tength L, the length of the filter H(z) is L!? = @! = DL ~ i) + Las can be checked from (27) Of course as i 0@ we get L! ~+ o>, Now instead of considering the discrete time filter, we are going to consider the function #0) which is piecewise constant on intervals of length 1/2, and equal to Pee) = 2 = HM HG) on n/t sxc v/2. 28) Clearly, f°%(e) is supported on the interval 0, L~ Me Note that the normalization by 2/7 ensures that if E (Hay) = L then {FUN de = 1 as well. Also, it can be checked that lz = 1 when H's = 1. A fundamental question isto find out whether and to what the function /'"(8) converges as i + &. Fig. 13 shows two examples of such iterations. In Fig. 13(a) the first six iterates of the filter with impulse response [1, 3, 3, 1] show that it converges rapidly to a continuous function while in Fig. 13() the iterates ofthe iter [~1, 3,3, ~ tend to a discontinuous function. In other words, different filers exhibit very different behavior. OF course when constructing wavelets of compact support one would like them to be continuous functions, perhaps possessing con- tinuous derivatives also. This can be achieved ifthe filter rmeets certain regularity constraints; so that the limit fune tion f°") is continuous. In {12} Daubechies gives such ‘condition, which we now review. First, assume thatthe filter H(z) has a zero atthe half sampling frequency, or H(e) = 0. This is not unreason- able since H() isto be a halfband low-pass filter in fact if f'"Gx), 1 = is 0 converge to a continuous function at least one such zero is necessary [34], [35]. This to gether with the fact thatthe filter impulse response is or- thogonal to its even translates is equivalent to EAC) H(1) = V2. Define mlz) = 1/-¥2 ~ H(z), that is mo(1) = 1, Now factor m2) into its roots atx (there i at Teast, one by assumption) and a remainder polynomial R(z), in the following way: mde) = (0 + £9 /2PR) [Note that R(1) = 1 from the definitions. Now call B the supremum of [R()| on the unit etce Fee ‘Then the following result from [12] holds Proposition 2.1 [12]: IB < 2*-', then the piecewise constant function f''2) defined in (28) converges point- wise 10 a continuous function f"). o o Fig. 13. Neation (28 fo two single hers a) [1.3.3.1] wh om ‘eres to comin ection (B}{ = 3,3. 1] whi converges to . By taking the limit in G1), #(2) itself satisies a two-scale difference equation: 2, hala) 92x ~ m. ot) = 2" (32) We an define also the bandas futon He) = 29 E hed 62x. G3) ‘Assume now that we have filters such that the impulse responses are orthogonal with respect to even shifts: chan = 20), in = Wy kez G4) We show now that (2) is orthogonal to integer translates of itself Vi. From the definition (28) f(s) is just the i dicator function on the interval [0, 1); so we immediately et orthogonality at the Oth level, that is, < fx — 1) Je ~ BY = by. Now we assume orthogonality at the ith level: y by CFG = Di LMC = bY = by 69) and prove that this implies orthogonality at level: + 1st ct — ny fee =) 2B Z hdd) (F2x ~ 2 ~ nd. FRx — k—m) = SS htndidn + 21 ~ 28) bu Hence, by indvction (35) holds forall. So i the limit (oer =), 6 = b> 66) ‘The other orthogonality relations between Ya) and d(x) now follow easly from (34) War ~ D, Moe — WY = by a =, Ye bY =O. on ‘To end the section we show that ¥(x) is oxthogonal across scale, First, across one scale’ He =D, We = BY = (Ei - 6@s ~ 21+ 0. 90x ~ 8) Enum (oes ~ 21+ m). ¥Ox— 0) =o Similar reasoning shows that across several scales Y2'x ~ 1) is orthogonal to YO'x = k) forj # iand k # f. We hhave shown the correspondence between the orthogonal ity relations of the filter impulse responses, and those of | their infinite iterates. Besides onhogonality. complete- ness is required in order that (27 °Y(2'x ~ 1, j, eZ) be an orthonormal basis for Z°(R), and this it proved in [12]. Thus any function f(x) € 17(R) can be writen as J) = BE YO — p,fey2 UA ~ 0. IIL, Gesekat, FIR PeRrecr RECONSTRUCTION FILTER Banks aNb BIORTHOGONAL WAVELETS Since we will be using perfect reconstruction filter banks (PREB's) to construct wavelets, we brisfly review the salient points here. For greater detail, we refer for example t [19]-[22]. Assume that we have a filter bank as in Fig. 8, with analysis fiters Ho(2) and Hy) but with general synthesis filters Gy(z) and Gy(2) instead of H(z) and Az). Note that upsampling by 2 (inserting a zero between every two samples) corresponds to simply re- placing z by 2? in the z transform, Using (26) itis easily seen that the output of the analysisisynthests system is sven by Xe) = 1G) GI _ [Fae Ho ae ] 8) me) HA-2) [Xa Call the above 2. 2 matrix H(z), where the subscript im indicates tht it contains modulated versions of the fil ters Ho(z) and H(z). Note that from the above equation in order to elimiaate the contribution to the reconstructed, signal from X(—2) (which represents an aliased version of the signal it is necessary and sufficient that the synthesis filters are related tothe analysis iters as follows’ 1G), Gi] = COMHY-2. ~ Ho-aI- 8% Note that det [H(@)] = Hola) H(—2) = Hol—2) 2) (40) = Pla) ~ Pl-2) ay, where P(z) = Ho) Hi(—2). We introduce the following. polyphase notation for the filters: Hye) = Hla?) + 2 HEC) that is, H(2) contains the even-indexed coefficients ofthe filter (2), while H,(2) contains the odd ones. Thus [Hate?) aoa [Hol2) Bee|: fe | lived) med) “Ela mally -allo « ‘ li ‘| 1 i] where H(z) i called the polyphase matrix. In particular, det (Hy(0)) = ~22 7! det (29) @ and et (Hy2)] = Hele) Hy) ~ Ho) #2) 1/2 2! 1pG') ~ Pe). 3) A. FIR Filter Banks For convenience we repeat the following well-known fact [37], [38], [21] Fact 31: For perfect reconstruction with FIR synthesis {filters after an FIR analysis, itis necessary and sufficient that det Hy] = c+ 27! where Ie Z. [Note that det [H,(2] is thus a pure delay as well. In order {o make the results that follow less arbitrary, we shall fix ¢ = 2. This involves no loss of generality, since filters that differ only by a multiplicative constant can be re= ‘garded as equivalent. From this, and (40) it follows that P(2) can have only 4 single nonzero odd-indexed coefficient: PO) ~ P-2) = Dyes a“ and we would normalize py.) = 1. A polynomial P(z) satisfying this constraint is termed a valid polynomial. Tt is then clear that any factorization of a valid P(2) Hy(z) H\(—2) gives a possible FIR PREB. Note that Hy(2) and H\(~2) ate interchangeable. Given an FIR filter Hi(2) we use the term complementary filter to denote any Ay(2) such that P(z) = H(z) Hy(~2) isa valid polynomial; that is Ho(2) and Hy(2) form a PR filter pai. ‘The above discussion indicates two possible design ‘methods for PRFB's [21]. First the factorization method consists of finding a valid (2) satisfying design criteria, and then factoring it into Ho(2) and A—2). Second, the complementary filter method stars with a filter Hy(z) and then solves a system of linear equations to find a comple~ mentary filter leading to a valid P(z). (Note that this P(2) ‘can then be refactored as desired.) Of course once we have designed Pz) and factored it in terms of the analysis fl- ters Ho(z) and H(~2) the synthesis filters follow directly from (39), with Clz) = © = B. Orthogonal or Paraunitary Filter Banks In Section ID-C we have seen how to construct unitary ‘operators based on filters which were orthogonal to their even translates. In filtering notation this means that the even terms of the autocorrelation function are all zero. withthe exception of the central one (which equals unity for normalized filters). The autocorrelation ofa filter is Hi(2)H,(2 "ys thus the eondit Hee Yy M(-a (2 (45) Further, the two filters H(z) and H(z) were orthogonal to each other at even translates (19), so the even terms of | the cross correlation are all zero HAH) + Hf-DH-e) = 0. 46) If the two filters satisfy (45) and (46), their impulse re- ‘sponses and even translates form an omthogonal basis for TZ). We are now ready to prove a fact on the form of orthogonal FIR solutions: Fact 3.2: Consider an FIR perfect reconstruction filter ‘bank such that H(z) satisfies (45). Then the length of hain} thas 10 be even. In order that hgin ~ 2k) and hy(n ~ 21) form an orthogonal basis (that is H,(z) should satisfy (4S) sand (46)), itis necessary and suffciens thar Hye) = 2") “ Proof: ‘That L must be even was already shown in Section II-C. That a complementary fiter given by (47) satisfies (45) and (46) and leads to perfect reconsteuction is easly verified by substitution [Necessity is shown as follows: from (46) it is known that H(z) Hy(2") is a polynomial with only odd coetti- cients, that is Hs aw. So the zetos appear in pairs at (a, —a). Proposition 4.3 below states that H(z) cannot have such a pair of zeros if PR isto be possible; thus for every zero at = a in Hy(2), there must be & corresponding zero at ¢ = ain F(z. ‘This means that H(z) has a zero at z = —1/a: so (2) has the form given in (47) a ‘This necessary form of H\(2) means that me Mya) + 2 Mele Choosing & = 0 leads t Hale) — Hoo) Because itis a perfect reconstruction system et [Hy (2)] = Hl) Hole!) + Hoy (2) Hoy 1 9) since det [H,(z)] has to be a delay, and is symmetric. This is the polyphase equivalent of (45), because the even cocf- ficients of the autocorrelation Ho(z) H(z ') are given by (49). On the unit circle z =e”, this means that [Hegle™))? + [Hpyfe)? = 1 that is they form a power complementary pair [38]. The necessary form (47) also means that, substituting into (40) 28 MHela) Hye") + Hola) Hl Or, on the unit circle (50) Hyp = 2! Hoe) + [Het = [Note thatthe matrix Hf (2) in (48) has the following prop- eny: Ue" He) ee en Hye!) Hl) [0 '] because of (49). In other words, it i unitary on the unit circle; such a matrix is called @ paraunitary matrix [38]: if iis also stable it is called lossless [39], (40). This is the extension of the allpass filter concept [41] to matrices with polynomial entries, ‘The above discussion indicates two possible design ap- proaches. The first is to find an autocorrelation function that has only a single even-indexed coefficient different from zero. This must be the central one, since an auto- correlation function is symmetric. Such a function can be factored into its “square roots” H(z) and Ho(z~!). In particular, zeros on the unit circle have to be double, since Hale) poe Hole) Hake) EE TRANSACTIONS OW SIGNAL PROCESSING, VOL, 40. ND. 4 SEFTEMRE 00> the fanction must not change sign. This method was frst, used by Smith and Barnwell {37] and Mintzer 18] 0 syn- thesize orthogonal filter banks. The second metiod uses lattice structures to synthesize paraunitary matrices, for Which complete factorizations have been given by Vai- dyanathan and Hoang (20). Note that designs in the filter bank literatare did not address the question of regularity (sce proposition 2-1). Daubechies [12] specifically designed filters satisfying the above orthogonality relations and having a maximum umber of zeros at. C. Biorthogonal or General Perfect Reconstruction Filter Banks From fact 3.1, we know that perfect reconstruction re quires det [27q(2)] t be an od delay, and the synthesis filters are given by (39) with C(e) = ¢ ‘Choose ! = 1 forthe purposes of this discussion. First note that Gy(2)s(2) and G(c) H(z) have only odd coet- ficients, that is, ( goln — 28), yin = 2) on (ain ~ 24), hain = 2D) (2) (note the time reversal in the inner product). In matrix notation HG, = 0 = HyGy (53) where H, and G; have been defined in a fashion similar to 17). Now G2) Hifa) = 2~'Hy(z) Hy(—2) has a single nonzero odd coefficient (because it is a PR system). Sim: ilanly Gy(2)Hy(2) has also only a single even-indexed coef ficient different from zero. That is, (gun = 21), hd = 6 co In operator notation HyGy = 1 = HYG, 5) Since we have a perfect reconstruction systert we get GyHy + GH, = 1 (56) Of course the last equation above indicates that no non- zero vector ean lie in the column nullspaces of both Hy and H,. Note that (55) implies that GyH ant Gy #1, are each projections (since G,H,G;H, = G,H,). They project ‘onto subspaces which are notin general orthogonal, Note that (S1) and (52) indicate, however, that tFere are ot thogonality relations between the filter impulse responses. Because of (51), (52), and (54) the analysis synthesis sys- tem is termed biorthogonal. Inthe special case where we have a paraunitary solution one finds: Gy = Hi and G, =H}. and (53) gives that we have projections onto sub- spaces which ate mutually orthogonal. Further consider- ations on biorthogonal systems can be found i [28], (29) ‘One reason to use biorhogonal rather than orthogonal ‘bases is that the additional freedom allows us have ar- bitrary length linear phase filters. 1 is well known [12]. [23), [22] that the ony orthogonal (or paraunitary) real IK niter bank duvig seat passe 4s wy ro nonzero taps, and is given by Hota) = 24 218 | and Hie) S12 tye, (Note that paraunitary filters cam be factorized so as to approximate linear pase {17] 142}, (291) To olaain longer real FIR filters, and still have exact linear phase, one has to give up omthogonalty. The classes of linear phase solutions are indicated below Proposition 3.3: Linear phase perfect reconstruction real FIR filter banks using filters He) and H, 2) have one ofthe following forms 4) Bosh filters are symmetric and of odd lengths, dif fering by an odd multiple of 2 1) One filter is symmetric and the other is antisym- meric; both lengths are even, and are equa or difer by din even multiple of 2. ‘) One filter is of odd length, the other is even: both have all zeros onthe unit circle. Either both fters are symmetric, or one is symmetric and the other is ant ‘symmetric Proof: See Appendix A.2. In class) we find that P(2) = He) Hy(~2) bas zeros only atthe 2n = Urools of £1; 80 the filters are of very Tile practical interest, We will not consider this trivial solution further. Tris also possible wo get lattice strictures that generate linear phase FIR filter banks [23}, 22]. Appendix A.3 indicates the lattice that will generate the low-pass Ho(2) 1+ 21) (hati the “ost regular” low pass, fl- lowing proposition 2.1) and its complementary ‘filter Hi). A general discussion of complementary filters in the biorthogonal case is given in Appendix A.3. The dis cussion of the biorthogonal pyramid is presented in Ap- pendix A.1 Note that while in the FIR case it is necessary to give up ombogonality to get nontrivial linear phase solutions, in the TI case iti possible to have both the filters are roneausal, but good designs are posible [25]. D. Biorthogonal Wavelets Based on Filter Banks otic together the results of this section we now show thatthe infinitely iterated biorthogonal perfect reconstruc- tion filter banks generate biorthogonal sets of functions. We denote by H(z) and Ge) the filters which are equivalent to the cascade of i blocks of filtering/subsam- pling inthe analysis and synthesis sections, respectively ‘We assume that both of the filters involved are regula. From these define functions which are piecewise constant fn intervals of length 1/2 i 2AM) nf sx cine /E on n/P sx (2 auoy - Fx ~ 2 =m, Dhimsan~ 21m) =D gun) = hn + 20 = 28) = by In the limit this gives (Bee ~ D, 6 — by = By © Once this is established (eb, Ye = bY = by ws) follows immediately from (34); and the relations (lr D, Yar By = 0 ) (JO — D, oe — BY = 0 (66) ‘come from (51) and (52), respectively. We have shown that the conditions for perfect recon- struction on the filter coefficients lead to functions that have the biorthogonalty properties as shown above. Or- thogonality across scale is also verified, following the analysis of Section ILF: QYOx = 9,2 Pyar - By 1 Bu Thus the set (YQ!x — 1), Px — K). iJ. ke LEZ} is bionhogonal. That itis complete can be verified as inthe orthogonal case [28]. Hence any function from LR) can be written Fo) = EE YO's ~ Fo) 77YOx ~ D. Note that Ys) and Ys play interchangable rls. ‘So, regulrbiorhogonal FIR iter banks Ted to bior thogonal bases of functions of finite length; its easly shown atthe convene is also ue. Assume hat we have functions Ya) o(2), 82), He) satisfying (39)-(62) and (63)-(66). Then ty easy to verify that they can be veed to derive binhogonl ster banks. For example, sing 6) (6 — 1, oe — = (Eau - bax - 21~m, -m) = EE gum + him (42x ~ 20 =m, Z Ayim) g(x ~ 2k 2x — my = E goin)» hn + 20 ~ 2h) = dy ‘The other filter biorthogonality relations follow from (64)- (66). E. Filter Design Propositions 2.1 gives a sufficient condition to ensure pointwise convergence to a continuous funetion, which hinges on the numberof zeros atx. Its clear that after with a maximum numberof zeros atx will not necessarily bbe maximally regular. Ideally, to get a maximally regular filter (which is written H(z) = (1 + 2~")/2]* FC) as in Section II-E) we would maximize N while simultaneously minimizing sup,eio 2x) |F(@")|- There is a tradcof in- volved. However only NY, the number of zeros at x, can be easly controlled ‘The question of regularity is more involved for the biomogonal ease than it is inthe orthogonal case, since row we have to check the regularity of both the analysis (Hof3)) and synthesis (F(z) low-pass filters (see Sec- tion V-A). Fig. 14 illustrates the difficulty of making both Hi(@) and H\(—2) regular for the length 4 linear phase case. Recall that the impulse responses of Hotz) and H\(~2) are given by [1, e, a I] and [1, ~a, ~a, 1, respectively [22]. The figure shows the sealing function generated by [. @, @, 1] for ae [-3, 3]. Ata = 3 we have that H(2) is very regular, while H(z) is very it- regular (as shown also in Fig. 13). The difference be- tween the regularity ofthe two filters clearly decreases as a gets smaller, without ever leading to two regular linear phase perfect reconstruction filters of length 4, Actually at a = 0 there is neither pointwise nor L* convergence [12] (the figure which shows the sixth iteration gives therefore an erroneous impression atc Fig. 14, Sclng ston gerd by aning eatin (28) for aH] nda €[-3. 3. (st iterations shown) IV, Avcepraic SraUCTURE OF FIR SoL.stI0Ns. A. Bezout’s Identity From Section IIL it can be seen using (42), (40), and (43) that the condition to achieve perfect reconstruction ‘with FIR synthesis filters after an FIR analysis section can ‘be expressed in two equivalent forms: , Hae) Hy (2) — Hoc) Mio Hole) H\(—2) ~ Hy 2) 2 68) Each of these equations implies the other, and we use (67) or (68) depending on which is more convenient for our purpose, These requirements of course considerably con- strain the possible solutions; both (67) and (68) are spe cial forms of what is known as a Bezout identity (43). {48}, This identity arises in the Euclidean algorithm which caloulates the greatest common divisor of to polyno- ‘mals a(x) and a(x): ala) = ox pod (af), (80) where a(x) isthe last divisor of the algorithm and ois ‘constant. It is well known that we can always write (x) = pola) + pyCcbayix) for some pots), py 9) faglx) and ay(x) are coprime then (x) has 22r0s at O oF only: and the identification between (69), the Bezout identity, and (67) or (68) becomes clear. By examining the implications of this, new results emerge: also, from this angle certain results that previously had more in- volved proofs (e.g, [45}) now become simpler. A review ‘of the necessary properties ofthe Euclidean algorithm can be found, for example, in [46]. The importance of the algorithm in the context of biorthogonal syste ns has also been noted in [28] and [47]. The following easily proved fact will play an important role in the est of this section 146} Fact 4.1: Given atx) and bts) ats)pCx) + Boats) ots) thas a solution | p(x), qts)} © ged (a(x), blr) divides e(x) ‘We now make use ofthese observations to examine the constraints on the filter banks, Proposition 4.2: Assume thatthe filters Hy) and Hy(2) ‘are FIR and causal. Then given one of the pairs (Ho), Hole), Hyolc). Hit. Woult),. Hyd] or Lor) H,y(2] in order to caleulate the other pair necessary 10 achieve perfect reconstruction itis necessary and suff- cient that the given pair be coprime (except for possible reros atz = ©) Proof: From fact 4.1 the ged of each of these pairs ‘must divide the right-hand side of (67). Hence the only actors that they can have in common are zeros at =, a The above proposition is also proved using a different argument in [45] Proposition 4.3: A filter Hy) has a complementary fil ter if and only if it has no zeros in pairs at Proof: Ho(c) has a zero pair at (a, ~a) if and only if Ho(z) has a factor A(=*). This happens if and only if both Mip(2) and H(z) have a common factor A(z), that is, they are not coprime. Thus the absence of zero pairs of the form (a, ~4) and coprimeness are equivalent and the proof is completed by using proposition 4.2. Proposition 4.4: There is always a complementary il ter to the binominal filter: He) 0 Proof: If Hog(z) and Hy(2) had a common factor it ‘would appear asa pair of zeros of H(z) at (a, ~a); since Hola) has zeros only at 2 = —1 it cannot have such a factor. o It should also be clear that for Ho(z) and H(z) to form ‘a perfect reconstruction pair itis necessary that they be coprime. From fact 4.1 ged [Ho(2), Hy(2)] must divide the right-hand side of (68). This has a very clear signal pro- cessing interpretation: a zero common to H(z) and Hy(2) ‘would imply a transmission null in both channels of the filter bank at some frequency, making reconstruction im- possible B. An Analogy with Diophantine Equations ‘The conditions under which a complementary filter to some chosen Ho(2) exist were detailed in Section IV-A. It has already been pointed out that such a filter could be found using Euclid’s algorithm: another method of find- ing a complementary filter based on solving a set of linear equations is given in [21] and Section V-A. The comple- ‘ment isnot unique. If given filter Ho(2), we can calculate fone ofits complements H(z), itis natural to wonder how ‘we may find others. The results of this section will show that given any complementary filter, a simple mechanism cists for finding all others ‘The basic idea again stems from our interpretation of the condition for FIR perfect reconstruction as a Bezout identity. In number theory equations with integer coef cients for which integer solutions are sought are known ‘a diophantine equations [43]. For example, art bye a where all quantities are integers and we seek the solution (Gr, 3) is the most basic diophantine equation. Clearly, if solutions of the equation ax! + by! = m are available then we can add them to (x, y) and generate new solutions to (71), If we replace integers by polyno: rials then (71) is analogous to the equation for perfect reconstruction that we must solve. We exploit this fact in the rest ofthis section by noting that polynomial solutions to an equation analogous to (72) are indeed easy to find this allows us to generate all possible complementary fil- ters if we are first given one. Infact, if we work with the ‘modulation version of the PR condition (68), and iden- tify: a = —Hy(e) and b = Ho(—2) itis easy to see that fx, y] = [Et2) Hy(—2), El) Hola] sets the right-hand side to zeto if Et) = B~: For the following proposition we make use of propo- sition A.3, which gives thatthe length NV ~ linear phase 2 otaet ats ay day 6a, day a a) 4a; 603 fa Hole) Wy) Giving for Hi(2) He) complementary iter oan od length near pate er is unique if it exists, Proposition 4.5: Given a linear phase Heft) of odd length N, and its lengih N ~ 2 linear phase complement H,(2) all higher degree odd length linear phase filters complementary 10 Ho(z) are ofthe form Hi = 2 HC) + BOM) where Za rene erm, Fe) Proof: Note that E(2) = E(~2), and that 2) Hy2) is symmetric about the point (N+ 4m ~ 3)/2 just as 2 Hi(2) is, Hence, 2-2"7%(2) + E(2) Hy) is easily seen to be a linear phase solution of length N + 4m ~ 2 by direct substitution ‘We now show that all length NV + 4m ~ 2 solutions are of this form. If Hi) is of length N+ 4m ~ 2 then Hye) Hi(—2) and Hy) Ete) Hy(~2) ate both valid and of Tength 20 + 4m — 3. So also is P"(c) = Hu 1Hi@) — Fa Hi—2)). We can choose the coefficients of Elz) 10 set some of the end terms of P'(2) to zero. That i, for Some choice of ay the coeficiens of z° and 2-¥"4—9 ‘become zero, s0 that P™(2) is reduced in length by 4 (the coefficients of ¢~! and 2-°""""~9 are already zero) Similarly, for each ofthe a; we can reduce the length of Pr) by 4, When cy, ">" oy have been appropriately fixed P*() has length 2N — 3, has powers of 27" in the range 2-2" to 2-2"! * and is still valid, Since i con tains H(2) as a factor it must have the form Peg) = 2 HY) (-2) since the length NV ~ 2 solution is unique by proposition A3. a From proposition 3.3 it follows that in the two nontsiv= ial cases of linear phase solutions, the length of PC) = Hije)H(~2) isdn ~ 1 In Appendix A.4 itis shoven that the solutions indicated in proposition 3.3 b) are special cases of those in ); that is, they can always be refactored Into the form a). T follows that all higher degree comple- mentary filters to a fixed H,(2) are given by proposition 4.5, unless they ate trivial, in the sense of belonging © class c) of proposition 3.3, Example: Consider Hie) = (1.4, 6,4, 1] and its unigue length 3 complementary filter Hic) ~ [1, 4, 1/16. Let m= 2. So we get fon oy 4a Ge day a ay day 6a; 4a eda, 6a + 05, dy + Ay Gh + ay + Pa, Bay +A + ay + Tone Ay + Ag Bay + an Aso) which is Tinear phase, and complementary to Ho(2). ‘A further result allows us to use the diophantine equa- tions to reach more general solutions. For this we use the result (from proposition A.2) that a length 1 filter has N= 2 length N ~ 2.complementary filters Proposition 4.6: All filters of length N + 2m — 2 which ‘are complementary 10 length N filter H(z) have the form HQ) = HQ) + Bette) where E(e) = E(~2) is a polynomial of degree 2m ~ 1), ke (0,1, ++, m}and Hy(e) isa length N ~ 2 comple- ‘mentary filter. Proof: That this is a solution is easily verified by substitution. If Be = E 62% then P"(2) = Hy(2)[H(-2) ~ Elz) Ho(~2)] is reduced in length by 2 for some choice of Bp oF fn» Again the argument is repeated; the length of P(2) ean be reduced, bby 2 at each stage by fixing one of the 8. In tais way we reduce the degree of P"(z) by 2m when all of the 8's are appropriately chosen. Since Ho(2) is sill a factor, the re ‘maining Factor must be one of the length NV ~ 2 solutions ‘of which there are only N ~ 2 a So far we have addressed only the problem of gener- ating higher degree solutions from lower ones; the next proposition shows that we can also go in the opposite di rection. We examine the linear phase case only: the ex- tension to the general case is obvious. For the linear phase case this says that given any complementary filter we can generate all others using the results of proposition 4.5. Proposition 4.7: If H? (2) and H(z) are vengih N + ding ~ 2, and length N + 4m, ~ 2 linear phese comple- mentary filters t0 the linear phase odd length N filter Hof), with ms > my then we can generate the lower de: {gree solution from the higher as follows: MQ) = 2TH RG) + Ez) Ho'2)] where Ey(a) = 2-™-"B)&) ~ Bylo Proof: Direct substitution, o ‘The subclass of all valid P(2)'s which can be factored as Pte) = Ho(2) Hye!) corresponds to parauritary FB’s, and generates orthonormal bases of wavelets. Since all such P(2)'s are symmetric and positive they form a sub- class of those generated by the construction of proposition 4.5 above. The case for which Ho(2) is chosen to be an even power of the binomial is treated in detail by Dau- bechies. In Appendix A.5 we establish the close relation between her results and proposition 4.5 above. The fact that higher degree solutions contain lower degree ones is also given in [47] €. Continued Fraction Expansions In Section IV-B it was shown that any solu:ion to (67) ‘or (68) could be written as the sum of lower degree so- lutions and trivial higher degree solutions (trivial in the sense that the right hand side of (44) becomes zero). In Section IV-B we dealt exclusively with the modulation domain; but the results there, like those of this section could be expressed in either modulation or polyphase no- tation, For the sake of definteness consider the polyphase version of the perfect reconstruction condition (67), and assume that Hpg(2) and Hoy(z) are given, It can be seen from proposition 4.6 that this lowest degree solution isin some sense fixed, and embedded atthe core of any higher degree solution. The strong connections between Euclid’s algorithm, the Bezout identity and continued fraction ex: pansions (CFE’s) is well known [43], [48]. In fact, we now show that the canonic CFE of Hjo/Mhy is the same a that of Hoo/Ho, except for the last term; and further that higher degree solutions are formed by adding terms. to the CFE, the frst terms remaining unchanged. We define D- = Hog, Dp = Hoy, A-1 = Hi, and Ay = Hy. For the Sake of simplicity we remove the phase factor in (67), In this notation (67) becomes. D_s(2)Ag2) ~ A-s@)D 2) = 1 m) [Now use Euclid’s algorithm starting withthe pair D_(2). Dz). The first step gives Dis(2) = bsepDy) + DG) ‘Also do one division of the pair A-1( the remainder 4,(2) ALA) = fe) Ade) + Ae) ‘Together these equations give T= Diy2)Ade) ~ ADA (le) — aga) Ale) Dalz) + Dy(zdAgl2) ~ Ay) However, since deg AyDy > deg D,Ay and deg AyDy > deg Dy must have ay = bo, and hence Daley Aye) ~ Ad2)D\C2 Since this is of the same form, but of lower degree, than the equation that we started with (73), we can compare the second step of Euclid’s algorithm with a division of ‘Ag(t),Ay(2) and this gives a2) = by(2). The result is that ‘we get a succession of Bezout identities: deg Dy > deg Dy )- Auf), denoting deg Ay > deg Ay -1 Dy-0a = Di2A,-f@) which are of decreasing degree. We find in tum that ay(2) ‘ole)s (2) = bys - >" 5 fe) = BL) >>» yl) ry(2)- Note that these outputs of Eucli’s algorithm are the "partial denominators of the canonical CFE of (D_@)/Da(2) (48), {43}: ley; ula), by), ° ++ bu) where we have used the standant notation [by: by, bos bby] to denote a terminating CFE [49]. It follows that the (CFE's of D(z) /Do(2) and A(z) / Ay) are identical for the first N +1 terms. The terminal equation for j = N gives * (2) Dy Remark that the D,(2)'s are known, and Dy is a scalar (ince itis the last divisor and D_(2), Dylz) are assumed to be coprime); hence we can choose Ay(2) to be any polynomial and we are assured that 4(2)s are polynomial N—1,-77,0, “1. Therefore we get a valid ‘complementary filter (A(z), A_(2)] for any appropriate ‘Ax(2). This can be expressed by writing ay (2) CD" Ay(2)Dq. In summary, 5) Avi) Hod) 62) = tole); bb rl 6 Fite toga bu), buh “= By) 2 totes By), Bd, “~~ «By =D "Alera ™ It can be shown that choosing Ay(z) = 0 gives the same solution as that produced by Euclid’s algorithm. That is, Hida) (byf2); uC), «= + by (2), y(2, O} Hye = Utes bes + by Ie we divide (4) by Dy(2)D, 2) we get AQ An (0 Diz) Diz) Dy) Dy ee ‘Which isthe expression elating successive convergents to 4 CFE [30]. This is precisely the continued fraction in (76), (77). Note that since the successive convergemts sat- isfy the decreasing degree Bezout identities (74), truncat ing (16) and (77) gives a lower degree FIR PREB. There is more structure yet. Dividing (74) by D,-2)4)-(0 gives A@ Be 0 A@ Dyk) Like (78) this relates successive convergents of a contin ued fraction; but the CFE ere is that ofa different group- ing of the polyphase pairs, 1 turns out that we get FAO] Hy, . FH [—(—1)FAyledDys by, by -il2) * Fg, 7 DEAD leds bye» Bla Hy) FH) [—(—1)NAyl2)Dyi Bu 2h, by =O). "BWC ae "Ay 2)Dy; Byl2). by (2 => BY Of course, these continued fraction relations give an- ‘other method for generating higher degree solutions to the ‘perfect reconstruction condition. By varying Ay(2) in (77), for example we generate different complementary filters. Clearly, this structure is complete; all possible comple- mentary filters can be generated by appropriate choice of Av) it warrants reiteration thatthe above analysis using the polyphase notation, can be used to give equivalent results in the modulation notation, Y. Desion Resuts A. Linear Phase Filters with a Maximum Number of Zeros at x We now design a linear phase P(z) = H(z) #y(~2) sat- isfying (68) and having the maximum possible number of| zeros at 2 = —1, Note that if P(2) has even a single zero at z = —1 it cannot have any at z = 1 and vice versa This should be clear since we wish to have P(2) ~ P(—2) equal to 22%" (see Section III-A). We consider only the ease where P(e) is of odd length (see proposition 3.3): hence it has an even number of zeros; since P(2) is linear ‘phase its zeros must occur in quads, and pairs on the unit circle of real axis. It follows that P(z) must have an even number of zeros at z = —1 or none a all Note that by proposition 4.4 we are assured that there exists a complementary filter to the binomial of any de agree. We wish now to calculate Ro,(—2) which is the ‘complementary filter to (1 + 2 |) In other words, we wish to find the polynomial such that Pi) = (1+ 2)" Rax(2) is valid, as defined by (44), That Roa(2) is Tinear phase is an immediate consequence of the fact that Piz) and (1 + 27") are, Because of this symmetry, on ‘equating the appropriate terms (0 zero it turns out that only k of the equations are independent; so we get a set ‘of K by k equations to be solved for the coefficients (rp, rath) A ‘Example: If we choose F(z) = (1 + 271)? we solve the 3 by 3 system found by imposing the constraints on the coefficients of the odd powers of 2! of, + me + se Pay = (yt net + ne? tne (+ 6 + Se? + 202 + 6% 42" So we solve 6 1 /n\ (0 2» 6 6}(n}=fo 12 0 20) \ny \ giving r= 3/2, -9, 19)/128. In general therefore we solve the system Fyors a) where Fa is the k x k matt Fa y)and exis the length & vector (0, 0. ‘Having found the coefficients of Ry (2) we factor it imo Li Fig. 15, Bionhogoal wavelets gers y he length Ste given in Tables Ind Te ies ne x maxima ube BioKTHOCONAL M3) THE INFINITE Tremston or Tues rens Ove te Waveuers Show 13 a a oourzs0 ‘nat aan fh osoueo Na) 2) Soniusis wha) iO) ‘hotiess1 ws i Sonsosr has) i) —olnsaonot fui Ki) To baat fentny Puy O.6927386 ‘eu m8) fo rataze ai linear phase components; and then regroup these factors ff Ray(2), and the 2k zeros at z = —1t0 form «wo filters: Hy2) and Hy(~2), both of which are to be repular TH turns out that for small k, ensuring that both Ho(2) and H;(~2) meet the bound of proposition 2.1 ean force ‘one © choose filters of quite unequal length. For larger k, however, this problem eases, and it becomes possible 10 {get filters ofthe same, or approximately the same, length that generate regular symmetric wavelets, For example, ret Time foctn and Spc of een want () Tie futon and spect of eytbess mal. Inrise Reatnst CoerecIsT oF Boro Ho Ise TemarnoeTwise Pras Gr ooons680 SaoeaoTs -tonsie Bollelaes Bautrasis Det 8 noeans 8.500805 Fig, 15 shows the analysis and synthesis wavelets and their spectra for one particular factorization of the k = 9 case, and Tables I and Il list the coefficients ofthe filters H(z) and Hy(—2). Each of these filters has a factor (1 + z), We get Ho) = [C+ 2 ')/20R(2) where SuPscjoae [Fufe?}| = 115.06 < 28; so Hotz) satisfies the bound of proposition 2,1. Hence the sealing function (x) and wavelet Y(s) generated by the infinite iterations, fol lowing (59) and (61), converge to continuous functions. ® % Fig. 16, Omboncmal wavelet ese by length 1 Her om [12 The Herb maximum member of e052 = "2 Time fncton (0 Spectrum @ » Fg. 17. Blonogoal wales gevete by tte of length 20 and 2¢ given in Tales and LV. The ers were designed tang the dopamine sppoach, and have Reer oped performance tha owe op. 181) Tie eton eh Spectra ‘ot aysis wale () Time fostion an speci of sess wavelet (9, can_be: similarly factored, and supseyory 8: More General Solutions: Diophantne Approach |F\(e***)] = 211.3 < 2 hence oa) and Ye) converge to continuous Funetions also While the design procedure of the previous section is ‘The method outlined in [35] yields the following esti-_ very simple, the spectra of the scaling function (u) and mates forthe regularity index B: for Yix) we find 2.46 < the wavelet (a) are not as one might wish from low-pass 2, and for lx) we get 3.55 < B and bandpass filters. For comparison purposes we show in Fig. 16 the cor- Proposition 4.5 showed how to generate any valid lin responding plots for the orthonormal wavelet generated ear phase P2) conning a given factor. For example, to by after ofthe same length, as presented in [12 design a P(e) with 2k zeros at z, we can calculate Ry(~2). TABLE luriseRestonse Covers lure Boxtoconat Mere ee tga ‘or Tass ens Grete Waves Shown 17 ma) rn Dontsases re in ~0.0n201228 ma nm “oimsis79 mi By ‘howeaess a mi homes? eit ms) morass mits) me “onions: mit oy Goso0273 mits) mis -ons1630 m5 ni) Sonatas mrt io oss2zie2 my nant auaisiorie my as in the previous section. This solution, which has de- gree 2k — 2, can then be used to generate all possible solutions of degree 2 — 2 ++ 2m. {A second approach isto note that we need n>t place all of the zeros at 2 = —1 to star with. We could for example calculate the complementary filter to a factor (I+ 2°) fuy(z)us(2) *** a (2)) where wi(2) represents a 20 pair on the unit circle. We are then assured cf having a factor (+ 2°!)fuy@)unt2) +++ w/2y] to pace in the stopband of each of the filters Tn Fig. 17 we show the «wo symmetric wevelets ¥0) ‘and 90) and their spectra. These were generated by linear phase filters of lengths 24 and 20, the coefficients of which ae listed in Tables III and IV. The filters were designed using a combination of the two approaches describes above. Each of the filters has only a single zero at z “1, and neither of them mocts the bound given in prop- osition 2.1, However again using the estimation methods of [35] we find 0.79 < Band 0.96 < B. Note that the stopband performance is much better than in Figs. 1S and 16. Fig. 18 shows the wavelet and its spectrum for a better crthogonal set of fiers. The associated filter bank is pparaunitary; so H(z) = Hg(z'). The coeficients are listed in Table V. The regularity estimate here is 0.97 < 8. C. Root Loci of Higher Degree Solutions While we have shown that higher degree solutions can sive better results, itis still clear from Sections IV-B and IV-C that these solutions are nonetheless very con- strained. To give a concrete example we briefly examine the case where H(z) = (1 + 2_')* and the unique linear phase degree 4 complementary filter has impulse response coefficients hy(n) = [3, 18, 38, 18, 3]/256. We examine the m = 1 solution (from proposition 4.5) Hye) = 27H) + al + 7) — (BO) TaRLe Ww Awrose Resrons Cosmin Inmrownd Boor ooosat i). em Ione Ite n08 wow Far? me ho) oaosesan7 mb Sooro2071 ia) 001389087 no) Comms i Sosessz48 A) S.o02i5e3 Ai) oes4i5t a 13347083 i © 37810 ne) 1898118 Fig. 18. Ontonormal wave ened y length 22 fier given ia Tale ‘VEie ferme eign ung the daphamine approach hat beter Siopbnd pevommance tha tht in Fg 1. () Tne acon (6 Spe and plot the trajectory in the & plane of the roots as a is varied, Fig. 19 shows the root locus for the region: a € [~2, 2}, While the increased degree solution has more free ‘dom, itis clear tha the roots move along very constrained paths, In other words, it would be necessary to look at Solutions of considerably higher degree to get substantial design freedom, Incise Resnovee or eove {fuels austen Ges nossr39 1 ease bates Basomees raeKe 8.390803, Cotas ‘e2si0 48389 os Conaes S.nsoes Sansess ~dm2893 Sonieat On1si99 poms Zones ‘.0H0 Doves S001 Fi. 19. Locus of he overeat of he tf 80) in the: pane fora € 1-2. VI. Coxctusion ‘The relationships between wavelets, multiresolution signal analysis, and filter banks have been developed, em- phasizing the equivalence between the fundamental ideas of each of these fields, which is that functions or signals be broken down into component signals at different reso- lutions. In addition, strong similarities between the de- tails of these techniques have been pointed out, and the interplay of ideas has enabled us to indicate new results in each of the areas, In particular, we derived biorthogonal compactly sup- ported wavelets bases with symmetries using regular FIR PREB's. If compact support isnot desired, similar tech- nigues, using IR filters generate onhogonal wavelet bases symmetries [25], [51] ‘Appenpix A 1. Biorthogonal Pyramid This Appendix extends the result of Section II-D to the biorthogonal case (see also Section III-C). In this case, the approximation ¥ is equal to F= Gyly =x ‘The difference signal becomes d= = Gybhyx = GH, Because we have perfect reconstruction: Gly + GH, = 1 ‘That is, we can filter the difference signal with h(n) and subsample by 2, because HGH, = Hy from which d can be reconstructed by upsampling and i terpolation with G,, following the fact that GM, GH = GH, and the reconstructed signal, obtained by adding ¥ to d, {is equal to the original. That is, the biorthonormal pyr ‘mid is equivalent to critically sampled biorthogonal fil ter bank, Note that linear processing has been assumed throughout, which would not be the case in coding appli 2. Proof of Proposition 3.3 Proof: Since H(z) and Hy(~2) have linear phase so does P(z), If Plz) has odd Iength, then both filters have ‘odd length or both have even length. Note that Poy — P-2) = Det = (I apy aye Since P(2) is of odd length, both P(z) and P(—z) are syn metric or antisymmetric about the point {= N. Hence, (2) ~ P(=2) is symmetric about this point; so the single nonzero coefficient is py, the central one, Hence P(z) is symmetric and not antisymmetric; which implies that Hy{z) and H(~2) are both symmetic or antisymmetsic. 'a) Ly and Ly both odd. Now Ly + Ly ~ 1 is odd. The center of symmetry, which is (lp + L;)/2 — 1 samples away from the end points, has to be odd. Thus (Ly + L)/2 = Ly + (Ly ~ L,)/2 has to be even. Thus (Lp — L,)/2 is odd, and the length difference Ly ~ Ly isan odd multiple of 2. In particular, there are no same Length so- lions. Suppose that Ai(2) and Hy(~2) can be both sym metric or both antisymmetric. The latter possibility is ruled out, however, because the two polyphase compo- nents are also antisymmetric when the length is odd, and are therefore not coprime. Perfect reconstruction is then impossible following proposition 4.2. b) Loand L, both even. Again (Ly + L,)/ 'be odd so thatthe center term is not in the san end terms. Then Ly + (Ly ~ L;)/2 has to be even, and since Lo is even, this means thatthe length difference Lo ~ L, is an even multiple of 2; and, for example, same length solutions do exist. Also it was assumed that Ho(2) and Hy(~2) were both symmetric, but since #%(—2) has ‘opposite symmetry to H(z) when the length is even, the even length solution leads to a symmetrie/antisymmetric pair, ©) IF P(z) has even length, then one filter has even length, and one odd. Now we get Bnet Pls) = P(-2) ‘P(2) has only a single nonzero odd-indexed coefficient; bbut note that 7) = pay, = 0, *** .N ~ 1. So odd indexed coefficients are paired with even-indexed-ones. It follows that P(2) also has only one nonzero even-indexed coefficient, and is of the form Poa) = pz il & 2-1"), (s) Since P(2) has all its zeros on the unit circle (at the 20 i) ~ 1 roots of +1), Ho(z) and Hy(—2) also have all zeros on unit cirle. If Hy(2) and H(z) are both to be an- tisymmetric, both must have zeros at z = 1. This will force Hy(—2) to have a zeroatz = Lorz = ~1 depending (on whether it is of odd or even length, respectively. This implies that P(z) has either a double zero at z = I, or a pair at z = | and z = —1. Since P{@) contains only the 2(V ~ j) ~ 1 roots of +1 both possibilities are ruled out Hence the filters must have opposite symmetry. 4. Lattice Structures Generating the Binomial Filter ‘and its Complementary Filter Proposition A.1: Let the polyphase components of Ho(2) ‘and Hc) be generated by he following lattice structure: wo-[) TED SIE} oo Li =. Lo 2“Yla, 1. = (QN + 1/QN — 1), QN ~ 1)/QN ~ 3), . late Holz) = (1 + 21" ‘An elegant proof of this proposition was pointed out to us bby Gopinath [52]. It is clear thatthe complementary filter has only rational coefficients. Now P(z) can be written as PO) = Hye) PR Since Ros(re") is linear phase itcan be converted into 1 polynomial of degree k — 1 in cos (@ — j In 7). Thus the size of the factorization to be performed is halved, and the accuracy of the solution is improved. If the ex tional coefficients of Riy(2) are used and this is factored in cos 2, the solution can be considerably more accurate than if standard linear equation solver is used and direet factorization is performed. We will consider k = 6 0 give an example ofthe importance ofthese considerations. De- note P,(2) which is exact, Pj(2) which is found by multi- plying out after using the lattice ovtputs given by propo- sition A.1 and the efficient factorization; PZ (2) is found by multiplying out after using an ordinary Tinear equation solver, and direct factorization. We find | psn) — itr) = 0.00892 while I pain) — p8 (nlx = 17.83739. 4. Restrictions on Complementary Filter Forms Proposition A.2: A length N filter has at most N ~ 2 complementary filters of length N-~ 2; and ar most one of length N — i. Proof: If Hy(z) has length N 2 then Pt) = H(z) Hy(—2) has length 2N ~ 3; hence ithas N ~ 2 odd- indexed coefficients. Thus we solve a square system of linear equations of size NV ~ 2. The nonzero odd-indexed coefficient can be placed in any of N ~ 2 positions. If H,G) has length N ~ 2 then P(2) = Ho(z) Hi(—2) has length 2N ~ 2; one ofthe end terms is odd indexed while the other is even, There ate N — 1 od-indexed terms Thus we solve a size N — 1 square system, but there is only one possible position for the single nonzero odd-in-

    You might also like