ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS...
Transcript of ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS...
![Page 1: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/1.jpg)
![Page 2: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/2.jpg)
![Page 3: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/3.jpg)
![Page 4: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/4.jpg)
![Page 5: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/5.jpg)
![Page 6: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/6.jpg)
![Page 7: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/7.jpg)
![Page 8: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/8.jpg)
![Page 9: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/9.jpg)
![Page 10: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/10.jpg)
![Page 11: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/11.jpg)
![Page 12: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/12.jpg)
![Page 13: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/13.jpg)
![Page 14: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/14.jpg)
![Page 15: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/15.jpg)
![Page 16: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/16.jpg)
![Page 17: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/17.jpg)
![Page 18: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/18.jpg)
![Page 19: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/19.jpg)
![Page 20: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/20.jpg)
![Page 21: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/21.jpg)
![Page 22: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/22.jpg)
![Page 23: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/23.jpg)
![Page 24: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/24.jpg)
![Page 25: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/25.jpg)
![Page 26: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/26.jpg)
![Page 27: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/27.jpg)
![Page 28: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/28.jpg)
![Page 29: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/29.jpg)
![Page 30: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/30.jpg)
![Page 31: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/31.jpg)
![Page 32: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/32.jpg)
DESIGN OF DIGITAL FILTERS
AND FILTER BANKS BY
OPTIMIZATION:
APPLICATIONS
Tapio Saramaki and Juha Yli-Kaakinen
Signal Processing Laboratory,Tampere University of Technology,
Finlande-mail: [email protected]
Tampere University of TechnologySignal Processing Laboratory
![Page 33: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/33.jpg)
ABSTRACT
This paper emphasizes the usefulness and theflexibility of optimization for finding optimized digitalsignal processing algorithms for various constrainedand unconstrained optimization problems. This isillustrated by optimizing algorithms in six differentpractical applications:
• Optimizing nearly perfect-reconstruction filter bankssubject to the given allowable errors,
• minimizing the phase distortion of recursive filterssubject to the given amplitude criteria,
• optimizing the amplitude response of pipelinedrecursive filters,
• optimizing the modified Farrow structure with anadjustable fractional delay,
• finding the optimum discrete values for coefficientrepresentations for various classes of lattice wavedigital filters, and
• finding the multiplierless coefficient representationsfor the linear-phase finite impulse response filters.
Tampere University of Technology
Signal Processing Laboratory 1
![Page 34: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/34.jpg)
WHY THERE IS A NEED TO USE
OPTIMIZATION?
Among others, there exist following three reasons:
1) Thanks to dramatic advances in VLSI circuittechnology and signal processors, more complicatedDSP algorithms can be implemented faster and faster.
• In order to generate effective DSP products, oldalgorithms have to be reoptimized or new onesshould be generated subject to the implementationconstraints.
2) All the subalgorithms in the overall DSP productshould be of the same quality:
• In the case of lossy coding, nearly perfect-reconstruction filter banks are more beneficial: loweroverall delay and shorter filters.
3) There are various problems where one response isdesired to optimized subject to the given criteria forother responses:
• A typical example is to design recursive filters suchthat the phase is made as linear as possible subjectto the given amplitude criteria.
Tampere University of Technology
Signal Processing Laboratory 2
![Page 35: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/35.jpg)
TWO-STEP PROCEDURE
It has turned out that the following procedure is veryefficient:
1) Find in a systematic simple manner a suboptimumsolution.
2) Improve this solution using a general-purposenonlinear optimization procedure:
• Dutta-Vidyasagar algorithm
• Sequential quadratic programming
Desired Form: Find the adjustable parametersincluded in the vector Φ to minimize
ρ(Φ) = max1≤i≤I
fi(Φ) (1)
subject to constraints
gl(Φ) ≤ 0 for l = 1, 2, . . . , L (2)
and
hm(Φ) = 0 for m = 1, 2, . . . ,M. (3)
Tampere University of Technology
Signal Processing Laboratory 3
![Page 36: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/36.jpg)
COMMENTS
• The proposed two-step procedure is very efficientwhen a good start-up solution being rather close tothe optimum solution can found.
• For each problem under consideration the way ofgenerating this initial solution is very different.
• A good understanding of the problem at hand isneeded.
• If a good enough start-up solution cannot be foundor there are several local optima, then simulatedannealing or genetic algorithms can be used.
Tampere University of Technology
Signal Processing Laboratory 4
![Page 37: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/37.jpg)
NEARLY
PERFECT-RECONSTRUCTION
COSINE-MODULATED FILTER BANKS
x(n) MH0(z) M F0(z)
MH1(z) M F1(z)
MHM--1(z) M FM--1(z)
y(n)
Linear-phase prototype filter:
Hp(z) =N∑
n=0
hp(n)z−n, (4)
Filters in the bank:
hk(n) = 2hp(n) cos
[(2k + 1)
π
2M
(n−
N
2
)+ (−1)kπ
4
]
(5)
fk(n) = 2hp(n) cos
[(2k + 1)
π
2M
(n−
N
2
)− (−1)kπ
4
].
(6)
Tampere University of Technology
Signal Processing Laboratory 5
![Page 38: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/38.jpg)
INPUT-OUTPUT RELATION
Y (z) = T0(z)X(z) +
M−1∑
l=1
Tl(z)X(ze−j2πl/M), (7a)
where
T0(z) =1
M
M−1∑
k=0
Fk(z)Hk(z) (7b)
and for l = 1, 2, . . . ,M − 1
Tl(z) =1
M
M−1∑
k=0
Fk(z)Hk(ze−j2πl/M). (7c)
Here, T0(z) is the reconstruction transfer function andthe remaining ones are aliased transfer functions. It isdesired that T0(z) = z−N and the remaining transferfunctions are zero.
Tampere University of Technology
Signal Processing Laboratory 6
![Page 39: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/39.jpg)
STATEMENT OF THE PROBLEMS
Problem I: Given ρ, M , and N , find the coefficientsof Hp(z) to minimize
E2 =
∫ π
ωs
|Hp(ejω)|2dω, (8a)
where
ωs = (1 + ρ)π/(2M) (8b)
subject to
1 − δ1 ≤ |T0(ejω)| ≤ 1 + δ1 for ω ∈ [0, π] (8c)
and for l = 1, 2, . . . ,M − 1
|Tl(ejω)| ≤ δ2 for ω ∈ [0, π]. (8d)
Problem II: Given ρ, M , and N , find the coefficientsof Hp(z) to minimize
E∞ = maxω∈[ωs, π]
|Hp(ejω)| (9)
subject to the conditions of Eqs.(8c) and (8d).
Tampere University of Technology
Signal Processing Laboratory 7
![Page 40: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/40.jpg)
COMPARISONS BETWEEN FILTER
BANKS WITH M = 32 and ρ = 1
Boldface numbers indicate that these parameters havebeen fixed in the optimization.
Criterion K N δ1 δ2 E∞ E2
Least 8 511 0 0 1.2 · 10−3 7.4 · 10−9
Squared −∞ dB −58 dB
Minimax 8 511 0 0 2.3 · 10−4 7.5 · 10−8
−∞ dB −73 dB
Least 8 511 10−4 2.3 · 10−6 1.0 · 10−5 5.6 · 10−13
Squared −113 dB −100 dB
Minimax 8 511 10−4 1.1 · 10−5 5.1 · 10−6 3.8 · 10−11
−99 dB −106 dB
Least 8 511 0 9.1 · 10−5 4.5 · 10−4 5.4 · 10−10
Squared −81 dB −67 dB
Least 8 511 10−2 5.3 · 10−7 2.4 · 10−6 4.5 · 10−14
Squared −126 dB −112 dB
Least 6 383 10−3
0.00001 1.7 · 10−4 8.8 · 10−10
Squared −100 dB −75 dB
Least 5 319 10−2
0.0001 8.4 · 10−4 2.7 · 10−9
Squared −80 dB −62 dB
Tampere University of Technology
Signal Processing Laboratory 8
![Page 41: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/41.jpg)
PERFECT-RECONSTRUCTION FILTER
BANK with N = 511
−150
−100
−50
0A
mpl
itude
in d
B
Angular frequency ω
Prototype Filter
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−150
−100
−50
0
Filter Bank
Am
plitu
de in
dB
Angular frequency ω0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−1
0
1x 10
−10
Am
plitu
de
Angular frequency ω
Amplitude Error T0(z)−z−N
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
0
1
2
3x 10
−15
Am
plitu
de
Angular frequency ω
Worst−Case Aliased Term Tl(z)
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
Tampere University of Technology
Signal Processing Laboratory 9
![Page 42: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/42.jpg)
NEARLY PR FILTER BANK with
N = 511 for δ1 = 0.0001
−150
−100
−50
0A
mpl
itude
in d
B
Angular frequency ω
Prototype Filter
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−150
−100
−50
0
Filter Bank
Am
plitu
de in
dB
Angular frequency ω0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−1
−0.5
0
0.5
1x 10
−4
Am
plitu
de
Angular frequency ω
Amplitude Error T0(z)−z−N
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
0
0.5
1
1.5
2
2.5x 10
−6
Am
plitu
de
Angular frequency ω
Worst−Case Aliased Term Tl(z)
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
Tampere University of Technology
Signal Processing Laboratory 10
![Page 43: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/43.jpg)
NEARLY PR FILTER BANK with
N = 511 for δ1 = 0
−150
−100
−50
0A
mpl
itude
in d
B
Angular frequency ω
Prototype Filter
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−150
−100
−50
0
Filter Bank
Am
plitu
de in
dB
Angular frequency ω0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−1
0
1x 10
−10
Am
plitu
de
Angular frequency ω
Amplitude Error T0(z)−z−N
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
0
0.2
0.4
0.6
0.8
1x 10
−4
Am
plitu
de
Angular frequency ω
Worst−Case Aliased Term Tl(z)
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
Tampere University of Technology
Signal Processing Laboratory 11
![Page 44: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/44.jpg)
NEARLY PR FILTER BANK with
N = 511 for δ1 = 0.01
−150
−100
−50
0A
mpl
itude
in d
B
Angular frequency ω
Prototype Filter
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−150
−100
−50
0
Filter Bank
Am
plitu
de in
dB
Angular frequency ω0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−0.01
−0.005
0
0.005
0.01
Am
plitu
de
Angular frequency ω
Amplitude Error T0(z)−z−N
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
0
1
2
3
4
5
6x 10
−7
Am
plitu
de
Angular frequency ω
Worst−Case Aliased Term Tl(z)
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
Tampere University of Technology
Signal Processing Laboratory 12
![Page 45: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/45.jpg)
NEARLY PR BANK with N = 383 for
δ1 = 0.001 and δ2 = 0.00001
−150
−100
−50
0A
mpl
itude
in d
B
Angular frequency ω
Prototype Filter
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−150
−100
−50
0
Filter Bank
Am
plitu
de in
dB
Angular frequency ω0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−1
−0.5
0
0.5
1x 10
−3
Am
plitu
de
Angular frequency ω
Amplitude Error T0(z)−z−N
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
0
0.2
0.4
0.6
0.8
1x 10
−5
Am
plitu
de
Angular frequency ω
Worst−Case Aliased Term Tl(z)
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
Tampere University of Technology
Signal Processing Laboratory 13
![Page 46: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/46.jpg)
NEARLY PR BANK with N = 319 for
δ1 = 0.01 and δ2 = 0.0001
−150
−100
−50
0A
mpl
itude
in d
B
Angular frequency ω
Prototype Filter
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−150
−100
−50
0
Filter Bank
Am
plitu
de in
dB
Angular frequency ω0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−0.01
−0.005
0
0.005
0.01
Am
plitu
de
Angular frequency ω
Amplitude Error T0(z)−z−N
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
0
0.2
0.4
0.6
0.8
1x 10
−4
Am
plitu
de
Angular frequency ω
Worst−Case Aliased Term Tl(z)
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
Tampere University of Technology
Signal Processing Laboratory 14
![Page 47: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/47.jpg)
DESIGN OF APPROXIMATELY LINEAR
PHASE RECURSIVE DIGITAL FILTERS
• It is shown how the minimize the maximum deviationof the passband phase of a recursive digital filtersubject to the given amplitude criteria.
• The filters under consideration are conventionalcascade-form filters and lattice wave digital (LWD)filters (parallel connections of two all-pass filters).
• There exist very efficient schemes for designing initialfilters in the lowpass case.
• Before stating the problems, we denote the overalltransfer function by H(Φ, z), where Φ is theadjustable parameter vector.
• The unwrapped phase response of the filter isdenoted by argH(Φ, ejω).
Tampere University of Technology
Signal Processing Laboratory 15
![Page 48: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/48.jpg)
STATEMENT OF THE PROBLEMS
Approximation Problem I: Given ωp, ωs, δp, and δs,as well as the filter order N , find Φ and ψ, the slopeof a linear phase response, to minimize
∆ = max0≤ω≤ωp
| argH(Φ, ejω) − ψω| (10a)
subject to
1 − δp ≤ |H(Φ, ejω)| ≤ 1 for ω ∈ [0, ωp], (10b)
|H(Φ, ejω)| ≤ δs for ω ∈ [ωs, π], (10c)
and
|H(Φ, ejω)| ≤ 1 for ω ∈ (ωp, ωs). (10d)
Approximation Problem II: Given ωp, ωs, δp, and δs,as well as the filter order N , find Φ and ψ to minimize∆ as given by Eq. (10a) subject to the conditions ofEqs. (10b) and (10c) and
d|H(Φ, ejω)|
dω≤ 0 for ω ∈ (ωp, ωs). (10e)
Tampere University of Technology
Signal Processing Laboratory 16
![Page 49: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/49.jpg)
EXAMPLE: ωp = 0.05π, ωs = 0.1π,
δp=0.0228 (0.2-dB passband ripple), and
δs = 10−3 (60-dB attenuation)
Elliptic Filter: The minimum order is five.
Cascade-Form Filter for Problem I: For the filter oforder seven the maximum deviation from φave(ω) =−47.06ω is 0.29 degrees.
Cascade-Form Filter for Problem II: For the filter oforder seven the maximum deviation from φave(ω) =−47.56ω is 0.50 degrees.
Lattice Wave Digital Filter for Problem I: Forthe filter of order nine the maximum deviation fromφave(ω) = −40.38ω is 0.094 degrees.
Lattice Wave Digital Filter for Problem II: Forthe filter of order nine the maximum deviation fromφave(ω) = −42.92ω is 0.27 degrees.
Linear-Phase FIR Filter: Minimum order is 107 anddelay is 53.5 samples.
Tampere University of Technology
Signal Processing Laboratory 17
![Page 50: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/50.jpg)
Cascade-Form Filter for Problem I
−100
−80
−60
−40
−20
0
Angular Frequency ω
Am
plitu
de in
dB
0 0.2π 0.4π 0.6π 0.8π π
−0.2
−0.1
0
0 0.05π
Angular Frequency ω
−2π
−1.5π
−1π
−0.5π
0
Pha
se in
Rad
ians
0 0.025π 0.05π
−1
0
1
−0.2
0
0.2
P
hase
Err
or in
Deg
rees
Angular Frequency ω0 0.01π 0.02π 0.03π 0.04π 0.05π
−1 −0.5 0 0.5 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Real Part
Imag
inar
y P
art
Tampere University of Technology
Signal Processing Laboratory 18
![Page 51: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/51.jpg)
Cascade-Form Filter for Problem II
−100
−80
−60
−40
−20
0
Angular Frequency ω
Am
plitu
de in
dB
0 0.2π 0.4π 0.6π 0.8π π
−0.2
−0.1
0
0 0.05π
Angular Frequency ω
−2π
−1.5π
−1π
−0.5π
0
Pha
se in
Rad
ians
0 0.025π 0.05π
−1
0
1
−0.4−0.2
00.20.4
Pha
se E
rror
in D
egre
es
Angular Frequency ω0 0.01π 0.02π 0.03π 0.04π 0.05π
−1 −0.5 0 0.5 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Real Part
Imag
inar
y P
art
Tampere University of Technology
Signal Processing Laboratory 19
![Page 52: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/52.jpg)
Lattice Wave Digital Filter for Problem I
−100
−80
−60
−40
−20
0
Angular Frequency ω
Am
plitu
de in
dB
0 0.2π 0.4π 0.6π 0.8π π
−0.2
−0.1
0
0 0.05π
Angular Frequency ω
−2π
−1.5π
−1π
−0.5π
0
Pha
se in
Rad
ians
0 0.025π 0.05π
−0.5
0
0.5
−0.05
0
0.05
P
hase
Err
or in
Deg
rees
Angular Frequency ω0 0.01π 0.02π 0.03π 0.04π 0.05π
−1 −0.5 0 0.5 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Real Part
Imag
inar
y P
art
Tampere University of Technology
Signal Processing Laboratory 20
![Page 53: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/53.jpg)
Lattice Wave Digital Filter for Problem II
−100
−80
−60
−40
−20
0
Angular Frequency ω
Am
plitu
de in
dB
0 0.2π 0.4π 0.6π 0.8π π
−0.2
−0.1
0
0 0.05π
Angular Frequency ω
−2π
−1.5π
−1π
−0.5π
0
Pha
se in
Rad
ians
0 0.025π 0.05π
−0.5
0
0.5
−0.2
0
0.2
P
hase
Err
or in
Deg
rees
Angular Frequency ω0 0.01π 0.02π 0.03π 0.04π 0.05π
−1 −0.5 0 0.5 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Real Part
Imag
inar
y P
art
Tampere University of Technology
Signal Processing Laboratory 21
![Page 54: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/54.jpg)
MODIFIED FARROW STRUCTURE
WITH ADJUSTABLE FRACTIONAL
DELAY
y(n)
--2
G0(z)G1(z)GL(z)
x(n)
G2(z)
1
µ
Fixed linear-phase filters for l = 0, 1, . . . , L:
Gl(z) =N−1∑
n=0
gl(n)z−n (11a)
where N is an even integer and
gl(n) =
gl(N − 1 − n) for l even
−gl(N − 1 − n) for l odd.(11b)
Delay: N/2−1+µ, where the fractional delay 0 ≤ µ <1 is directly the adjustable parameter of the structure.
Tampere University of Technology
Signal Processing Laboratory 22
![Page 55: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/55.jpg)
TRANSFER FUNCTION
The overall transfer function is given by
H(Φ, z, µ) =N−1∑
n=0
h(n,Φ, µ)z−n, (12a)
where
h(n,Φ, µ) =L∑
l=0
gl(n)(1 − 2µ)l (12b)
and Φ is the adjustable parameter vector
Φ =[g0(0), g0(1), . . . , g0(N/2 − 1), g1(0), g1(1), . . . ,
g1(N/2 − 1), . . . , gL(0), gL(1), . . . , gL(N/2 − 1)].
(12c)
Tampere University of Technology
Signal Processing Laboratory 23
![Page 56: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/56.jpg)
AMPLITUDE AND PHASE DELAY
RESPONSES
The frequency, amplitude, and phase delay responsesof the proposed Farrow structure are given by
H(Φ, ejω, µ) =N−1∑
n=0
h(n,Φ, µ)e−jωn, (13a)
|H(Φ, ejω, µ)| =∣∣∣N−1∑
n=0
h(n,Φ, µ)e−jωn∣∣∣, (13b)
and
τp(Φ, ω, µ) = −arg(H(Φ, ejω, µ))/ω, (13c)
respectively.
Tampere University of Technology
Signal Processing Laboratory 24
![Page 57: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/57.jpg)
STATEMENT OF THE PROBLEM
Optimization Problem: Given L, N , Ωp, and ǫ, findthe adjustable parameter vector Φ to minimize
δp = max0≤µ<1
[maxω∈Ωp
|τp(Φ, ω, µ) − (N/2 − 1 + µ)|]
(14a)
subject to
δa = max0≤µ<1
[maxω∈Ωp
||H(Φ, ejω, µ)| − 1|]≤ ǫ. (14b)
Tampere University of Technology
Signal Processing Laboratory 25
![Page 58: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/58.jpg)
EXAMPLES
Example 1: Ωp = [0, 0.75π], ǫ = 0.025, and δp ≤0.01.
δp = 0.00402 is achieved by N = 8 and L = 3.
Example 2: Ωp = [0, 0.9π], ǫ = 0.01, and δp ≤ 0.001.
The criteria are met by N = 26 and L = 4.
Tampere University of Technology
Signal Processing Laboratory 26
![Page 59: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/59.jpg)
RESPONSES FOR EXAMPLE 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
Angular Frequency ω
Am
plitu
de R
espo
nse
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
3
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
4
Angular Frequency ω
Pha
se D
elay
Res
pons
e
Tampere University of Technology
Signal Processing Laboratory 27
![Page 60: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/60.jpg)
RESPONSES FOR EXAMPLE 2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
Angular Frequency ω
Am
plitu
de R
espo
nse
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
12
12.1
12.2
12.3
12.4
12.5
12.6
12.7
12.8
12.9
13
Angular Frequency ω
Pha
se D
elay
Res
pons
e
Tampere University of Technology
Signal Processing Laboratory 28
![Page 61: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/61.jpg)
DESIGN OF LATTICE WAVE DIGITAL
FILTERS WITH SHORT COEFFICIENT
WORDLENGTH
• It is shown how the coefficients of the variousclasses of lattice wave digital (LWD) filters canbe conveniently quantized.
• The filters under consideration are
– the conventional LWD filters,– cascades of low-order LWD filters providing a very
low sensitivity and roundoff noise, and– LWD filters with an approximately linear phase in
the passband.
• There exist very efficient schemes for designing initialfilters in the lowpass case.
• Before stating the problems, we denote the overalltransfer function as H(Φ, z), where Φ is theadjustable parameter vector.
Tampere University of Technology
Signal Processing Laboratory 29
![Page 62: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/62.jpg)
Overall Transfer Function
The most general form of the transfer function is givenby
H(Φ, z) =
K∏
k=1
Hk(Φ, z), (15a)
where
Hk(Φ, z) = αkAk(z) + βkBk(z). (15b)
Here, Ak(z)’s and Bk(z)’s are stable all-pass filters oforders Mk and Nk, respectively.
For conventional and approximately linear-phase LWDfilters K = 1.
++ +
A1(z) A2(z) AK(z)
B1(z) B2(z) BK(z)
α1 α2 αK
β1 β2 βK
In Out
Tampere University of Technology
Signal Processing Laboratory 30
![Page 63: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/63.jpg)
Statement of the Problems
In VLSI applications it is desirable to express thecoefficient values in the form
R∑
r=1
ar2−Pr, (16)
where each of the ar’s is either 1 or −1 and the Pr’sare positive integers in the increasing order.
The target is to find all the coefficient values includedin Φ, in such a way that:
1. R, the number of powers of two, is made as smallas possible and
2. PR, the number of fractional bits, is made as smallas possible.
Tampere University of Technology
Signal Processing Laboratory 31
![Page 64: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/64.jpg)
Problem I: Find K, the number of subfilters, the Mk’sand Nk’s, as well as the adjustable parameter vector Φin such a way that:
1. H(Φ, z) meets the criteria given by
1 − δp ≤ |H(Φ, ejω)| ≤ 1 for ω ∈ [0, ωp] (17a)
|H(Φ, ejω)| ≤ δs for ω ∈ [ωs, π]. (17b)
2. The coefficients included in Φ are quantizedto achieve the above-mentioned target for theirrepresentations.
Problem II: Find Φ as well as τ , the slope of thelinear-phase response, in such a way that:
1. H(Φ, z) meets the criteria given by Eq. (17) and
| arg[H(Φ, ejωi)] − τω| ≤ ∆ for i = 1, 2, . . . , Lp,
where arg[H(Φ, ejω)] denotes the unwrapped phaseresponse of the filter and ∆ is the maximumallowable phase error from the linear-phase response.
2. The coefficients are quantized to achieve the above-mentioned target for their representations.
Tampere University of Technology
Signal Processing Laboratory 32
![Page 65: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/65.jpg)
Quantization Algorithm
The coefficient optimization is performed in two stages:
1. A nonlinear optimization algorithm is used fordetermining a parameter space of the infinite-precision coefficients including the feasible spacewhere the filter meets the criteria.
R(max)R(min)
Θ(max)
Θ(min)
1
24
3
R(min)0
R(max)0
5 6
(a)
(b)
Γ(min)2l−1
Γ(max)2l−1Γ
(max)2l
Γ(min)2l
Tampere University of Technology
Signal Processing Laboratory 33
![Page 66: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/66.jpg)
2. The filter parameters in this space are searchedin such a manner that the resulting filter meetsthe given criteria with the simplest coefficientrepresentation forms.
The algorithm guarantees that the optimum finite-wordlength solution can be found for both the fixed-point and the multiplierless coefficient representations.
Tampere University of Technology
Signal Processing Laboratory 34
![Page 67: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/67.jpg)
Numerical Examples
Filter specifications: δp = 0.0559 (0.5-dB passbandvariation), δs = 10−5 (100-dB stopband attenuation),ωp = 0.1π, and ωs = 0.2π.
Ninth-order direct LWD filter is required to meet thecriteria (M1 = 5 and N1 = 4).
−140
−120
−100
−80
−60
−40
−20
0
Angular Frequency ω
Am
plitu
de in
dB
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−0.5
−0.25
0
Am
plitu
de in
dB
0 0.01π 0.02π 0.03π 0.04π 0.05π 0.06π 0.07π 0.08π 0.09π 0.1π
Tampere University of Technology
Signal Processing Laboratory 35
![Page 68: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/68.jpg)
Alternatively, cascade of four 3rd-order LWD filters isneeded to satisfy the amplitude specifications.
−140
−120
−100
−80
−60
−40
−20
0
Angular Frequency ω
Am
plitu
de in
dB
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−0.5
−0.25
0
Am
plitu
de in
dB
0 0.01π 0.02π 0.03π 0.04π 0.05π 0.06π 0.07π 0.08π 0.09π 0.1π
For the cascade of four LWD filters, only 5 fractionalbits are needed for coefficient implementation comparedto 9 bits required by the direct LWD filter.
The number of adders required to implement all thecoefficients are 12 and 6, for the direct and cascadeimplementations, respectively.
The price paid for this is a slight increase in the overallfilter order (from nine to twelve).
Tampere University of Technology
Signal Processing Laboratory 36
![Page 69: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/69.jpg)
Optimized Finite-Precision Adaptor
Coefficient Values for the Cascade of
Four LWD Filters
A(z) B(z)
γ(1,2)0 = 2−1 + 2−3
bγ(1,2)1 = −1 + 2−2 − 2−5
bγ(1,2)2 = 1 − 2−3 + 2−5
γ(3)0 = 2−1 + 2−3 + 2−5
bγ(3)1 = −1 + 2−2
bγ(3)2 = 1 − 2−3 + 2−5
γ(4)0 = 1 − 2−2 + 2−5
bγ(4)1 = −1 + 2−2 − 2−4
bγ(4)2 = 1 − 2−4
Tampere University of Technology
Signal Processing Laboratory 37
![Page 70: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/70.jpg)
Approximately Linear-Phase LWD Filter
Filter specifications: δp = 0.0228 (0.2-dB passbandvariation), δs = 10−3 (60-dB stopband attenuation),ωp = 0.05π, and ωs = 0.1π.
The minimum order of an elliptic filter to meet theamplitude specifications in five. An excellent phaseperformance is obtained by increasing the filter orderto nine.
For the optimal infinite-precision filter the phase erroris 0.09399 degrees.
To allow some tolerance for the quantization, themaximum allowable phase error is increased to 0.5degrees.
Tampere University of Technology
Signal Processing Laboratory 38
![Page 71: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/71.jpg)
Amplitude and Phase Responses for the
Quantized Filter
−100
−80
−60
−40
−20
0
Angular Frequency ω
Am
plitu
de in
dB
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−0.5
0
0.5P
hase
Err
or in
Deg
rees
Angular Frequency ω0 0.025π 0.05π
Pha
se in
Rad
ians
Angular Frequency ω0 0.025π 0.05π−2.5π
−2π
−1.5π
−1π
−0.5π
0
−0.2
−0.1
0
0 0.01π 0.02π 0.03π 0.04π 0.05π 0.06π
For the optimized filter, only 10 adders with 11fractional bits are required to implement all the adaptorcoefficients.
The phase error for the optimized filter is 0.458 55degrees.
Tampere University of Technology
Signal Processing Laboratory 39
![Page 72: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/72.jpg)
A SYSTEMATIC ALGORITHM FOR
THE DESIGN OF MULTIPLIERLESS
FIR FILTERS
In this application, we show how the coefficients of thelinear-phase FIR filters can be conveniently quantizedusing optimization techniques.
The zero-phase frequency response of a linear-phaseN th-order FIR filter can be expressed as
H(ω) =M∑
n=0
h(n) Trig(ω, n), (18)
where the h(n)’s are the filter coefficients andTrig(ω, n) is an appropriate trigonometric functiondepending on whether N is odd or even and whetherthe impulse response is symmetrical or antisymmetrical.Here, M = N/2 if N is even, and M = (N + 1)/2 ifN is odd.
Tampere University of Technology
Signal Processing Laboratory 40
![Page 73: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/73.jpg)
Desired Coefficient Representation Form
The general form for expressing the FIR filter coefficientvalues as a sums of signed-powers-of-two (SPT) termsis given by
h(n) =
Wn+1∑
k=1
ak,n2−Pk,n for n = 1, 2, . . . ,M, (19)
where ak,n ∈ −1, 1 and Pk,n ∈ 1, 2, . . . , L fork = 1, 2, . . . ,Wn + 1. In this representation form,each coefficient h(n) has Wn adders and the maximumallowable wordlength is L bits.
Tampere University of Technology
Signal Processing Laboratory 41
![Page 74: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/74.jpg)
Statement of the Problem
When finding the optimized simple discrete-valuerepresentation forms for the coefficients of FIR filters,it is a common practise to accomplish the optimizationin such a manner that the scaled response meets thegiven amplitude criteria.
In this case, the criteria for the filter can be expressedas
1 − δp ≤ H(ω)/β ≤ 1 + δp for ω ∈ [0, ωp] (20a)
−δs ≤ H(ω)/β ≤ δs for ω ∈ [ωs, π], (20b)
where
β =1
2
[maxH(ω) + minH(ω)
]for ω ∈ [0, ωp]
(20c)
is the average passband gain.
These criteria are preferred to be used when the filtercoefficients are desired to be quantized on a highlynonuniform discrete grid as in the case of the power-of-two coefficients.
Tampere University of Technology
Signal Processing Laboratory 42
![Page 75: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/75.jpg)
Optimization Problem
Given ωp, ωs, δp, and δs, as well as L, the numberof fractional bits, and the maximum allowed numberof SPT terms per coefficient, find the filter coefficientsh(n) for n = 1, 2, . . . ,M as well as β to minimizeimplementation cost, in such a manner that:
1. The magnitude criteria, as given by Eq. (20), aremet and
2. the normalized peak ripple (NPR), as given by
δNPR = max∆p/W, ∆s, (21)
is minimized.
Here, ∆p and ∆s are the passband and stopbandripples of the finite-precision filter scaled by 1/β andW = δp/δs.
Tampere University of Technology
Signal Processing Laboratory 43
![Page 76: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/76.jpg)
Coefficient Optimization
The coefficient optimization is performed in two stages:
1. For each of the filter coefficient h(n) for n =0, 1, . . . ,M − 1 the largest and smallest values ofthe coefficient are determined in such a mannerthat the given amplitude criteria are met subject toh(M) = 1.
This restriction, simplifying the overall procedure,can be stated without loss of generality since thescaling constant β, as defined by Eq. (20), can beused for achieving the desired passband amplitudelevel.
These problems can be solved conveniently by usinglinear programming.
2. It has been experimentally observed that theparameter space defined above forms the feasiblespace where the filter specifications are satisfied.After finding this larger space, all what is neededis to check whether in this space there exists acombination of the discrete coefficient values withwhich the overall criteria are met.
Tampere University of Technology
Signal Processing Laboratory 44
![Page 77: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/77.jpg)
Optimization of Infinite-Precision
Coefficients
The goal is achieved by solving 2M problems of thefollowing form. Find the filter coefficients h(n) forn = 0, 1, . . . ,M − 1 as well as β to minimize ψ subjectto the conditions
M−1∑
n=0
h(n) Trig(ωi, n) − β(δp + 1) ≤ −Trig(ωi,M),
−M−1∑
n=0
h(n) Trig(ωi, n) − β(δp − 1) ≤ −Trig(ωi,M),
for ωi ∈ [0, ωp] and
M−1∑
n=0
h(n) Trig(ωi, n) − βδs ≤ −Trig(ωi,M),
−M−1∑
n=0
h(n) Trig(ωi, n) − βδs ≤ −Trig(ωi,M),
for ωi ∈ [ωs, π].
Here ψ is −h(n) or h(n) where h(n) is one among thefilter coefficients h(n) for n = 0, 1, . . . ,M − 1.
Tampere University of Technology
Signal Processing Laboratory 45
![Page 78: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/78.jpg)
Optimization of Finite-Precision
Coefficients
In the above procedure, h(M) was fixed to be unity.The search for a proper combinations of discrete valuescan be conveniently accomplished by using a scalingconstant α ≡ h(M).
For this constant, all the existing values between 1/3and 2/3 are selected from the look-up table containingall the possible power-of-two numbers for a givenwordlength and a given maximum number of SPTterms per coefficient.
Then for each value of α = h(M), the largest andsmallest values of the infinite-precision coefficients arescaled in the look-up table as
h(n)(min) = αh(n)(min) for n = 0, 1, . . . ,M (23a)
h(n)(max) = αh(n)(max) for n = 0, 1, . . . ,M (23b)
and the magnitude response is evaluated for eachcombination of the power-of-two numbers in the ranges[h(n)(min), h(n)(max)] for n = 0, 1, . . . ,M to checkwhether the filter meets the amplitude criteria.
Tampere University of Technology
Signal Processing Laboratory 46
![Page 79: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/79.jpg)
Numerical Examples
Example 1: N = 37, δp = δs = 10−3, ωp = 0.3π, andωs = 0.5π.
Method δNPR (dB) No. Powers of Two No. Adders
Lim and Parker −62.08 43 –Chen and Willson −60.87 40 –
Proposed −60.48 34 48
Example 2: N = 24, δp = δs = 0.005, ωp = 0.3π,and ωs = 0.5π.
Method δNPR (dB) No. Powers of Two No. Adders
Samueli −42.17 24 35Li et al. −43.33 24 –
Chen and Willson −43.97 24 33Proposed −44.09 21 30
If redundancies within the coefficients are utilized only26 adders are required to meet the specifications.
Tampere University of Technology
Signal Processing Laboratory 47
![Page 80: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/80.jpg)
Responses for Examples 1 and 2
−80
−70
−60
−50
−40
−30
−20
−10
0
Angular Frequency ω
Am
plitu
de in
dB
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−0.01
−0.005
0
0.005
0.01
0 0.1π 0.2π 0.3π
−70
−60
−50
−40
−30
−20
−10
0
Angular Frequency ω
Am
plitu
de in
dB
0 0.1π 0.2π 0.3π 0.4π 0.5π 0.6π 0.7π 0.8π 0.9π π
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0 0.1π 0.2π 0.3π
Tampere University of Technology
Signal Processing Laboratory 48
![Page 81: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/81.jpg)
Optimized Finite-Precision Coefficient
Values for the FIR Filter in Example 1
h(0) = h(37) = −2−11
h(1) = h(36) = 0
h(2) = h(35) = +2−9 − 2−12
h(3) = h(34) = +2−9
h(4) = h(33) = −2−9 − 2−11
h(5) = h(32) = −2−7 + 2−9 − 2−11
h(6) = h(31) = 0
h(7) = h(30) = +2−6 − 2−8
h(8) = h(29) = +2−7 + 2−9
h(9) = h(28) = −2−6 + 2−8 − 2−10
h(10) = h(27) = −2−5 + 2−8 + 2−12
h(11) = h(26) = 0
h(12) = h(25) = +2−4 − 2−6 − 2−9
h(13) = h(24) = +2−5 + 2−8 + 2−10
h(14) = h(23) = −2−4 + 2−6 − 2−10
h(15) = h(22) = −2−3 + 2−6 + 2−8
h(16) = h(21) = 0
h(17) = h(20) = +2−2 + 2−6
h(18) = h(19) = +2−1
Tampere University of Technology
Signal Processing Laboratory 49
![Page 82: ts/sldsp_part6_talk.pdf · DESIGN OF DIGITAL FILTERS AND FILTER BANKS BY OPTIMIZATION: APPLICATIONS Tapio Saram¨aki and Juha Yli-Kaakinen Signal …](https://reader031.fdocuments.net/reader031/viewer/2022030511/5abba3057f8b9ad1768cee7c/html5/thumbnails/82.jpg)
Effective Implementation Exploiting
Coefficient Symmetry for the
Multiplierless FIR Filter in Example 1
z−1 z−1
z−1 z−1
z−1
z−1 z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−2 z−2
z−2z−2
z−2
z−2
z−2
z−2
2−1
2−1
2−12−2
2−2
2−2
−2−2
2−3
2−3
2−4
2−5
2−62−6
2−8
2−9
−2−10
−2−11
−2−12
x(n)
y(n)
Tampere University of Technology
Signal Processing Laboratory 50