Discrete Wavelet Transform_* 55 FIGURE 4.7 Nested spaces. Consider the relation %? c R? where R? is the space spanned by the basis vectors 7 and j. SR? is the space spanned by the basis vectors 7, j and k. We now ask, what is missing in R that makes %? a subset of %°. Yes, you guessed it right. The basis vector ¢. To answer the first question, let us explore another function called Haar wavelet function. 4.3. HAAR WAVELET FUNCTION Haar wavelet function y(#) (see Figure 4.8) is given by 1 Osrsi/2 Yo = {-1 W2sts1 (4.10) 0 elsewhere yo ft 2 7 ae ‘ FIGURE 4.8 Haar wavelet function.56 © Insight into Wavelets—From Theory to Practice - It is quite easy to see that the set of functions {y(r-k), k € N} form an orthonormal set of basis functions. Like Haar scaling function, there is no overlap betwecn the translates of y(t) (see Figure 4.9). ' Mo 12 + FIGURE 4.9 Haar wavelet function and its translates. Therefore, = Jvove-pvar=o0 In general, ~ Jve-m Y(t —n) dt = Syn (4.1) ‘That is, the integral is 1 if m =n and 0 if m # n. Figure 4.10 shows that 7 vo yt) dr =. vO = yoyo “ao “on v2] Jr m2] [1 ’ T =) L_ SI + ’ FIGURE 4.10 Finding norm of Haar wavelet function. Let Wo be the space spanned by the orthonormal set of bases {y(t -k), k € N}-i Discrete Wavelet Transform _* 57 Formally we define, Wy = Span {yir—H} (4.12) a How does a signal which is an element of Wo look like? Figure 4.11 shows a signal in Wo 12.5 FIGURE 4.11 A signal in space Wo What is the speciality of signal in Wo (spanned by Haar wavelet functions)? In every unit interval, if the function value in first half is +m units then in the next half it must necessarily be Fm units. Otherwise, we cannot express the signal using the basis of Wo, that is, y(t) and its translates. Thus, the space Wo is a highly restricted space. The signal in Figure 4.11 is expressed using the bases of Wo as: y(t) + 2y(t — 1) - 2.5ylr- 2) + yr - 3) 4.3.1 Scaled Haar Wavelet Functions Consider a scaled version of yA#). We will go, only for dyadic scaling. That is, scaling by the integer power of two. Consider at first (21). Given y(?) as defined in Eq. (4.10), y(2r) can easily be shown to be 1 Osrsl4 W2s1s V2 w(2t) = 0 elsewhere58_+ Insight into Wavelets—From Theory to Practice - We denote the space spanned by the set of orthogonal bases {y(2r-k), ke N} as W; Formally we define, W, = Span {y2r— 0} x Figure 4.12 shows some of the bases of W). FIGURE 4.12 Some bases in W,. Is Wo c W,? Or in other words, can we represent a signal in Wo by using the bases of Wy. The answer is ‘No’. See Figure 4.13. an) uo 0 12 0 1 FIGURE 4.13 (21) and its translates cannot express a Y(t).ransform_¢ 59 Discrete Wavelet Wo Since linear combination of 42) and its translates cannot make # YA". 1 teeta can be expressed using bases of W,. Therefore, Wo < W;. However we aa Pets iiae Wy 1 W1. This means that bases of Wo are orthogonal to bases of W,. Figuie 4.14 s fvo v(2r) dt =0 also [y() yOr-1) dr =0 ee “a °lo 1 ° 1 y y2n), war- 1) Tol ia 12] Jt yt) y2r) ya) yar- 0 FIGURE 4.14 Product function yt) y(21) and yr) W2r > Note that the area under the function, which is the product of i) aml ¥{2!), 8 260%. Similarly, the area under the function which is the product of yA) and yA2" 1) 18 ale 20 This implies that space Wo is orthogonal to the space W;. Again it can be proved that Wi 1 Wa. Thus, we have another most important relation in wavelet theory. 2 Way LW LW, LW, 1... dent espect to So far we have learned two very important concepts in wavelet theory with respe function spaces. They are: * Space spanned by scaling function bases are nested, i.e., Vy EW CU Cw. ch cmselves. Thus, © Space spanned by wavelet function bases are orthogonal amony themselves. Th Way LW LW, LW. ob60 + Insight into Wavelets—From Theory to Practice We are now ready to answer the question we have posed earlier. What is that missing in Vo to make it a subset of Vj. To answer this question we consider a signal in V, as shown in Figure 4.15. o xan oY FIGURE 4.15 A signal in space V Since it is a signal in V,, we can express it using the basis of V;. Thus, S(O) = 4G(20) + 29(2t — 1) + 6Y(2t — 2) — 29(2e — 3) + 4G(2r — 4) + 6G(2t — 5) + 26(21 ~ 6) + 29(2r -7) It may be observed that the signal segment in the first unit interval can be represented using the basis of Vo and Wo, ie., - 4-2 4920 + 2921-1) = “2? gs yt) = 39) + WO Pictorially this can be visualized as in Figure 4.16. Similarly, S 69¢21 ~ 2) = 29(2r-3) = 252 ge - 2) + 7 VU) = 29 - D+ aye Again 4- So ye 3) = 6940 -2)— yer-2) 49021 —4) + 6921-5) = “28 9-3)Discrete Wavelet Transform_¢_ 61 3e0) 20-1) wo, 0 1 ch FIGURE 4.16 Decomposition of signal in V; into signal in Vo and Wo, Lastly, 29(2t — 6) + 2924-7 22 oq) + 2-2 yay = 29(t -3) + OWE -3) Thus, fi) = 4921) + 29(2t — 1) + 6G(2t — 2) — 29(2 = 3) + 4G2t — 4) + G25) + 2(2t - 6) + 29(2 - 7) (4.14) = 36(1) + 2p(r — 1) + 691-2) + 2600-3) + YD + AYE -—D— YeE=2) FOYE 3)_ 4.140) We could express a signal in V, space, in terms of bases of Vo space and Wo space. If we combine the bases of Vo and Wp space, we can express any signal in V, space. We express this truth mathematically as: Y=YWyow We call Wo as the complementary space of Vo. Now, we found out the “thing” that was missing in Vo, that made Vp a proper subset of Vj. We call Vo and Wo space as complementary space because Vo and Wo space are orthogonal and their bases together can represent any signal in the next ‘higher’ or finer space Vj. We call this as the decomposition of a finer signal into two coarser signals. Figure 4.17 shows the standard representation through diagram. Yo vy 3, 2, 6,2 4, 2, 6, -2, 4, 6 2,2 1, 4,-1.0 m FIGURE 4.17 Visualization of Eqs. (4.14a) and (4.14b).62 _+ Insight into Wavelets—From Theory to Practice 4.4 ORTHOGONALITY OF ¢(¢) AND y(t) How shall we prove that VoL Wo? We must show that the bases of Vp are orthogonal to bases of Wo. We need to show the orthogonality only in the case where both the bases overlap (exist in the same interval). If the bases do not overlap. naturally they are orthogonal because the Product of the two functions will identically be zero at every point. We know that 1) and y(1) are defined in the same interval (0, 1] and hence overlap. So we will try to prove that 441) is orthogonal to ys), os 1 (00. wen) = Foy dr= fi-tdr+ fiyar=05-05=0 * 0 és Pictorially this is depicted in Figure 4.18. Product of (2) and y(t) gives y(t) whose net area is zero. 1 «) oO I 1 way 0 T a j foovod = G 0 WO 1 ew 0 1 FIGURE 4.18 Orthogohality of (2) and y(n). Let us go back to our previous result, We have proved that the bases Vp and Wy together span V;. That is, V, = Vo ® Wo Using the same argument, we may write Vz = V, ® W, In general, V; = Vj. ® W,. (4.15) Let us now express V, using wavelet spaces alone. (Except the last term) Vj= Vin ® Whi But Vj. = Vi @ Who. heDiscrete Wavelet Transform _* 63 Therefore, Vi = Wp © Wir ® Vir Vj = Wy © Wy ® Wy @--® Wo ® Vo (4.16) Equation (4.16) tells that any signal in Vj can be expressed using the bases of W)-1. Wr. «s+ Wo and Vo, You will recognize later that Eq. (4.16) is the foundation of wavelet based decomposition of the signal. Schematically, this process is shown in Figure 4.19. FIGURE 4.19 Schematic representation of wavelet decomposition of signals. 4.5 NORMALIZATION OF HAAR BASES AT DIFFERENT SCALES Our basis vectors in ° are orthonormal. In addition to orthogonality, they satisfy the relation 7-7=]-j=k-&=1. Hence orthonormality. In function space, corresponding to dot product we have multiplication and integration. Consider bases 9(r—k)s of Vo. We know they are orthogonal. In addition, we have, Joe ae = or Jou-b Ott —K)dt = making the bases of Vp orthonormal. We say, bases of Haar at scale level *0" are orthonormal. The actual value of scale is given by 2/ where j is the scale level. What about orthonormality of bases of V;, where scale level is 1? The bases of space V, are translates of @(21)s. Let us check the value of integral Joen o(2r)dr. Its value is 1/2. Figure 4.20 depicts the process.64__* Insight into Wavelets—From Theory to Practice _ 4 4 2) 1 20) 28) =f (21) (28) dt = 1/2 > > 0 12 0 v2 FIGURE 4.20. Finding the normalizing constant for $20. So to normalize 4421) and its translates, we multiply each by L/V/2. Refer Figure 4.21 for visualization at V29(21) V2921) a 20020) =] V29(28) J29(21) d= 2-31 — > o 1 On=2 FIGURE 4.21 Normalized base for Vj. Span | J29(2r-k 4.17 pan {Yap} 4) Consider bases for V>. Here the scale level is 2. Hence we call bases for V2 as level 2 bases. Since J 641) 641) dt = 1/4 , we should multiply each base by Va =2. See Figure 4.22 Now, we redefine V; as: M for visualization. aa feo wan = J (41) (41) dt = 1/4 > ow ° ow FIGURE 4.22 Finding normalization constant for space V2.i Discrete Wavelet Transform _» 65 We redefine V> as: Vy = Span {2047 =} = Span {200% - »} K K Let us now generalize. Consider un-normalized bases for Vj The un-normalized bases for Vj are translates of (2/1). Since l Jeeineainar= 5; ‘The normalizing constant is V2) = 27. Therefore, Vj is redefined as: V, = Span (2? 0a%-»} (4.18) i 4.6 STANDARDIZING THE NOTATIONS When we express a signal in terms of normalized bases, we have to repeat the normalizing constant in each term. To avoid this repetition, an abbreviated notation is introduced. We write 27 g(2/t—k) as oj 4(t). The first subscript denotes the scale level or simply level and the second subscript denotes translation in that scale level. Thus, the basis of Vo are: {.--4.-100) 6.009 $10 ---+Fo4 Os» a} The basis of V; are given by {6 Pj. $019 Fas Pies } Similarly, bases in W; are normalized. The basis of W; are expressed as: {Wi 1s ¥p0(O) Vian -oo Wyn -onsf> where Yjx(®) are given by we yr -k) 4.7 REFINEMENT RELATION WITH RESPECT TO NORMALIZED BASES With respect to un-normalized bases, we have the relation: (1) = 921) + 921-1) Similarly it is quite easy to see that wit) = (24) = $2t— 1)66 _+ Insight into Wavelets—From Theory to Practice Both these relation follows from the fact that Vo and Wp are Subset of Vj. As a result of that, any base in Vo and Wo can be expressed using that bases of V,. With respect to normalized bases, this relation can be re-expressed as: 9) = L390 +L. agar 1 ¥ Hold +e Oa) (4.19) I I 1 vee Y20Q2n - V2924 = 1 Fe MO - FP (4.20) These two relations. which seems quite obvious with respect to Haar wavelet system, are the corner stone of wavelet theory. When we go on other wavelet systems, this relation may not be that obvious since most of other wavelet systems are fractal in nature. We cannot draw directly the functions for visualization. Generally, we write the above relation as: 90 = F100 6.400 = Fh) V262r- a2n t * WO = Y) 8) 9.400 =D, ath): V202r-#) (4.22) t . Specifically for Haar 1 1 HO) = YAH O4(0) = Yh) V2G2r — Bb) m m 1 1 VO =) £6) 9.400 = D) ath V29(2r—& & a 1 1 1 1 eh (0) = =, g(l) = ie (1) ye 8 Rita a Later we will interpret {h(k), k € N}, {g(k), ke N} as low pass scaling filter and high ass wavelet filter coefficients. In conventional signal processing what really matters is the property of filters (frequency response, phase response, etc). We will study such properties of these filter coefficients in detail later. Before proceeding further, we would like to introduce some other types of wavelet systems. By wavelet system we mean both the scaling function and wavelet function with their refinement relation defined through a known set of coefficients. To introduce other type of wavelet, at first we need to familiarize with another concept called support of the wavelet system. where h(0) = 4.8 SUPPORT OF A WAVELET SYSTEM Support basically means the range of the interval over which the scaling function and wavelet function is defined. Beyond this interval, the functions should be identically zero