Slides14

# Slides14 - Smoothness of wavelet bases Convergence of the...

This preview shows page 1. Sign up to view the full content.

This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: Smoothness of wavelet bases: Convergence of the cascade algorithm (Condition E); Splines. Bases vs. frames. Use eigenvalue analysis to study convergence of the cascade algorithm and smoothness of resulting scaling function. The cascade algorithm revisited: �(i + 1) (t) = 2� h0[k] �(i)(2t - k) k Consider the behavior of the inner products � a(i)[n] = � �(i)(t) �(i)(t + n) dt -� as i � � to understand convergence. 2 � a(i + 1)[n] = �{2�h0[k]�(i)(2t - k)}{2�h0[�]�(i)(2t + 2n - �)}dt -� k = 2� h0[�] � h0[k] a(i)[k + 2n - �] � � k ��� ��� = 2� h0[�] � h0[m œ 2n + �] a(i)[m] ��� m � m = 2� h0[2n œ r] � h0[- (r œ m)]a(i) [m] r -r m Filter with h0[n] Filter with h0[-n] and then downsample 3 In matrix form: a(i+1) = (�2) 2 H0H0T a(i) ; ��� H0 � Toeplitz matrix T Iteration converges if the eigenvalues of the transition matrix T satisfy ��� � 1 ��� with only a simple eigenvalue at � = 1. 4 Splines Splines are scaling functions whose filters only have zeros at � i.e. H0(�) = ( h0[n] = 1 + e-i� p 2) 1p () 2p n Consider p = 1 �(t) = ; n = 0, 1, …, p binomial coefficients 1 0 ^ �(�) = e-i�/2 1 sin �/2 �/2 5 What happens when p = 2 ? H0(�) = ( 1 + e-i� 1 + e-i� 2) 2 )( H1 (�) H2(�) 0 0 ��� ��� ��� ��� � ^(�) = � H (�/2j) � 0 j=1 � =� j=1 H1(�/2j) 0 � 2 . � H0(�/2j) j=1 = ^1(�) . ^2(�) � � = �(t) = -i�/2 sin �/2 )2 (e � 0 � 0 1 1 = 0 1 2 6 More generally �(�) = (e-i�/2 sin �/2 p ) �/2 �(t) = �box(t) � �box(t) � … � �box(t) 0 1 p=1 0 1 p=2 2 01 23 p=3 (p terms) 0 123 4 p=4 �(t) is piecewise polynomial of degree p œ 1. The derivatives, �(s)(t), exist for s � p œ 1 and they are continuous for s � p œ 2. e.g. Cubic spline (p = 4) is C2 continuous. 7 Alternatively, measure smoothness in L2 sense: ^ ||�(s)(t)||2 = = � ^ ��(i�)s �(�)�2 d� 2� -� 1 1� p 2p 2s 4 �sin �/2� � �2p 2� -� � �� (by Plancherel) d� when 2s œ 2p � -1 � Note: � 1 -� � d� is limiting case So, �(t) has s derivatives in the L2 sense for all s � smax, where smax = p - ² Valid for splines 8 Non-spline Scaling Functions In general, we have H0(�) = ( so that 1 + e-i� p ) 2 Q(�) �(t) = �p(t) � �q(t) ��� ��� ��� ��� pth order ugly spline Notice that the approximation power of �(t) comes entirely from �p(t): Suppose that we write � ck�p(t œ k) = t� for some � (0 � � � p). k 9 Then we have � ck �(t œ k) = �q(t) � t� k � = � �q(�)(t - �)� d� -� � = �( i=0 � � i )- � �q(�)(� �i �)�-i d� . ti ����� ����� = polynomial of degree �. 10 What about smoothness (in L2 sense)? Smoothness is given by smax = p - ² log2��max(TQ)� where TQ = (�2)2QQT Transition matrix for Q(�) Alternatively, look at the transition matrix for H0(�), T = (�2)2H0H0T T has 2p special eigenvalues due to the zeros at �: � = 1, ², ³, …, ( 1 )2p-1 2 Disregard these eigenvalues and look at the largest non-special eigenvalue, �max. Then the smoothness is given by 1 smax = - 2 log2��max(T)� �max(T) = 4-smax 11 ...
View Full Document

{[ snackBarMessage ]}

Ask a homework question - tutors are online