Multi-rate Signal Processing 7. M -channel Maximally Decmiated Filter - PowerPoint PPT Presentation

7 M -channel Maximally Decimated Filter Bank Appendix: Detailed Derivations Multi-rate Signal Processing 7. M -channel Maximally Decmiated Filter Banks Electrical & Computer Engineering University of Maryland, College Park Acknowledgment: ENEE630 slides were based on class notes developed by Profs. K.J. Ray Liu and Min Wu. The LaTeX slides were made by Prof. Min Wu and Mr. Wei-Hong Chuang. Contact: minwu@umd.edu . Updated: September 27, 2012. ENEE630 Lecture Part-1 1 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) M -channel Maximally Decimated Filter Bank M -ch. filter bank: To study more general conditions of alias-free & P.R. As each filter has a passband of about 2 π/ M wide, the subband signal output can be decimated up to M without substantial aliasing. The filter bank is said to be “maximally decimated” if this maximal decimation factor is used. [Readings: Vaidynathan Book 5.4-5.5; Tutorial Sec.VIII] ENEE630 Lecture Part-1 2 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) The Reconstructed Signal and Errors Created Relations between ˆ X ( z ) and X ( z ): (details) X ( z ) = � M − 1 ˆ l =0 A ℓ ( z ) X ( W ℓ z ) � M − 1 A ℓ ( z ) � 1 k =0 H k ( W ℓ z ) F k ( z ), 0 ≤ ℓ ≤ M − 1. M X ( W ℓ z ) | z = e j ω = X ( ω − 2 πℓ M ), i.e., shifted version from X ( ω ). X ( W ℓ z ): ℓ -th aliasing term, A ℓ ( z ): gain for this aliasing term. ENEE630 Lecture Part-1 3 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) Conditions for LPTV, LTI, and PR • In general, the M -channel filter bank is a LPTV system with period M . • The aliasing term can be eliminated for every possible input x [ n ] iff A ℓ ( z ) = 0 for 1 ≤ ℓ ≤ M − 1. When aliasing is eliminated, the filter bank becomes an LTI system: ˆ X ( z ) = T ( z ) X ( z ), � M − 1 where T ( z ) � A 0 ( z ) = 1 ℓ =0 H k ( z ) F k ( z ) is the overall transfer M function, or distortion function. • If T ( z ) = cz − n 0 , it is a perfect reconstruction system (i.e., free from aliasing, amplitude distortion, and phase distortion). ENEE630 Lecture Part-1 4 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) The Alias Component (AC) Matrix From the definition of A ℓ ( z ), we have in matrix-vector form: H ( z ): M × M matrix called the “Alias Component matrix” The condition for alias cancellation is   MA 0 ( z ) 0   H ( z ) ❢ ( z ) = t ( z ) , where t ( z ) =   :   0 ENEE630 Lecture Part-1 5 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) The Alias Component (AC) Matrix Now express the reconstructed signal as ˆ X ( z ) = A T ( z ) X ( z ) = 1 M ❢ T ( z ) H T ( z ) X ( z ) ,   X ( z ) X ( zW )   where X ( z ) =  .   :  X ( zW M − 1 ) Given a set of analysis filters { H k ( z ) } , if det H ( z ) � = 0, we can choose synthesis filters as ❢ ( z ) = H − 1 ( z ) t ( z ) to cancel aliasing and obtain P.R. by requiring   cz − n 0 0   t ( z ) =   :   0 ENEE630 Lecture Part-1 6 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) Difficulty with the Matrix Inversion Approach H − 1 ( z ) = Adj [ H ( z )] det[ H ( z )] Synthesis filters { F k ( z ) } can be IIR even if { H k ( z ) } are all FIR. Difficult to ensure { F k ( z ) } stability (i.e. all poles inside the unit circle) { F k ( z ) } may have high order even if the order of { H k ( z ) } is moderate ...... ⇒ Take a different approach for P.R. design via polyphase representation. ENEE630 Lecture Part-1 7 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) Type-1 PD for H k ( z ) Using Type-1 PD for H k ( z ): H k ( z ) = � M − 1 ℓ =0 z − ℓ E k ℓ ( z M ) We have E ( z M ): M × M Type-1 polyphase component matrix for analysis bank ENEE630 Lecture Part-1 8 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) Type-2 PD for F k ( z ) Similarly, using Type-2 PD for F k ( z ): F k ( z ) = � M − 1 ℓ =0 z − ( M − 1 − ℓ ) R ℓ k ( z M ) We have in matrix form: ❡ T B ( z ): reversely ordered version of ❡ ( z ) R ( z M ): Type-2 polyphase component matrix for synthesis bank ENEE630 Lecture Part-1 9 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) Overall Polyphase Presentation Combine polyphase matrices into one matrix: P ( z ) = R ( z ) E ( z ) � �� � note the order! ENEE630 Lecture Part-1 10 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) Simple FIR P.R. Systems ˆ X ( z ) = z − 1 X ( z ), i.e., transfer function T ( z ) = z − 1 Extend to M channels: H k ( z ) = z − k F k ( z ) = z − M + k +1 , 0 ≤ k ≤ M − 1 ⇒ ˆ X ( z ) = z − ( M − 1) X ( z ) i.e. demultiplex then multiplex again ENEE630 Lecture Part-1 11 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) General P.R. Systems Recall the polyphase implementation of M -channel filter bank: Combine polyphase matrices into one matrix: P ( z ) = R ( z ) E ( z ) If P ( z ) = R ( z ) E ( z ) = I , then the system is equivalent to the simple system ⇒ H k ( z ) = z − k , F k ( z ) = z − M + k +1 In practice, we can allow P ( z ) to have some constant delay, i.e., P ( z ) = cz − m 0 I , thus T ( z ) = cz − ( Mm 0 + M − 1) . ENEE630 Lecture Part-1 12 / 21

7.1 The Reconstructed Signal and Errors Created 7.2 The Alias Component (AC) Matrix 7 M -channel Maximally Decimated Filter Bank 7.3 The Polyphase Representation Appendix: Detailed Derivations 7.4 Perfect Reconstruction Filter Bank 7.5 Relation between Polyphase Matrix E ( z ) and AC Matrix H ( z ) Dealing with Matrix Inversion To satisfy P ( z ) = R ( z ) E ( z ) = I , it seems we have to do matrix inversion for getting the synthesis filters R ( z ) = ( E ( z )) − 1 . Question: Does this get back to the same inversion problem we have with the viewpoint of the AC matrix ❢ ( z ) = H − 1 ( z ) t ( z )? Solution: E ( z ) is a physical matrix that each entry can be controlled. In contrast, for H ( z ), only 1st row can be controlled (thus hard to ensure desired H k ( z ) responses and ❢ ( z ) stability) We can choose FIR E ( z ) s.t. det E ( z ) = α z − k thus R ( z ) can be FIR (and has determinant of similar form). Summary: With polyphase representation, we can choose E ( z ) to produce desired H k ( z ) and lead to simple R ( z ) s.t. P ( z ) = cz − k I . ENEE630 Lecture Part-1 13 / 21