daubechiesfilters
TRANSCRIPT
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 1/52
D AUBECHIES F ILTERS
Patrick J. Van Fleet
Center for Applied MathematicsUniversity of St. Thomas
St. Paul, MN USA
Joint Mathematical Meetings, 8 January 2007
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 1 / 5
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 2/52
R EVIEW T HE H AA R WAVELET T RANSFORM
In Session 1, we constructed the orthogonal Haar Wavelet Transform :
W 8 =
h 1 h 0 0 0 0 0 0 00 0 h 1 h 0 0 0 0 00 0 0 0 h 1 h 0 0 00 0 0 0 0 0 h 1 h 0
g 1 g 0 0 0 0 0 0 00 0 g 1 g 0 0 0 0 00 0 0 0 g 1 g 0 0 00 0 0 0 0 0 g 1 g 0
where
h = ( h 0 , h 1 ) = ( √2/ 2, √2/ 2), g = ( g 0 , g 1) = ( √2/ 2, −√2/ 2)
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 2 / 5
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 3/52
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
h = ( h 0 , h 1) = ( √2/ 2, √2/ 2) is called a lowpass lter .Lowpass lters tend to preserve non-oscillatory data and either
dampen or annihilate oscillatory data.Lowpass lters are best analyzed using a Fourier series.
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 3 / 5
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 4/52
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
h = ( h 0 , h 1) = ( √2/ 2, √2/ 2) is called a lowpass lter .Lowpass lters tend to preserve non-oscillatory data and either
dampen or annihilate oscillatory data.Lowpass lters are best analyzed using a Fourier series.
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 3 / 5
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 5/52
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
h = ( h 0 , h 1) = ( √2/ 2, √2/ 2) is called a lowpass lter .Lowpass lters tend to preserve non-oscillatory data and either
dampen or annihilate oscillatory data.Lowpass lters are best analyzed using a Fourier series.
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 3 / 5
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 6/52
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
Pretending we’re engineers for the moment, let’s “form” theFourier series
H (ω) =1
k = 0
h k e ik ω = √22
+ √22
e i ω
We wish to plot |H (ω)| so simplifying gives
H (ω) = √22
+ √22
e i ω
=√22
e i ω/ 2 e −i ω/ 2 + e i ω/ 2
= √2e i ω/ 2
cos (ω/ 2)So
|H (ω)| = √2cos (ω/ 2)
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 3 / 5
L O ASS /H G ASS F S L O ASS F S
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 7/52
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
Pretending we’re engineers for the moment, let’s “form” theFourier series
H (ω) =1
k = 0
h k e ik ω = √22
+ √22
e i ω
We wish to plot |H (ω)| so simplifying gives
H (ω) = √22
+ √22
e i ω
=√22
e i ω/ 2 e −i ω/ 2 + e i ω/ 2
= √2e i ω/ 2
cos (ω/ 2)So
|H (ω)| = √2cos (ω/ 2)
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 3 / 5
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 8/52
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
Pretending we’re engineers for the moment, let’s “form” theFourier series
H (ω) =1
k = 0
h k e ik ω = √22
+ √22
e i ω
We wish to plot |H (ω)| so simplifying gives
H (ω) = √22
+ √22
e i ω
=√22
e i ω/ 2 e −i ω/ 2 + e i ω/ 2
= √2e i ω/ 2
cos (ω/ 2)So
|H (ω)| = √2cos (ω/ 2)
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 3 / 5
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 9/52
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
Plotting |H (ω)| = √2cos (ω/ 2)
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 3 / 5
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 10/52
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
Suppose h = ( h n , . . . , h m ) with Fourier series
H (ω) =m
k = n
h k e ik ω
We will say h is a lowpass lter if
H (0) =m
k = n
h k = √2
and
H (π) =m
k = n
h k e ik π =m
k = n
h k (−1)k = 0
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 3 / 5
L OWPASS /H IGHPASS F ILTERS L OWPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 11/52
L OW SS /H G SS F S L OW SS F S
Suppose h = ( h n , . . . , h m ) with Fourier series
H (ω) =m
k = n
h k e ik ω
We will say h is a lowpass lter if
H (0) =m
k = n
h k = √2
and
H (π) =m
k = n
h k e ik π =m
k = n
h k (−1)k = 0
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 3 / 5
L OWPASS /H IGHPASS F ILTERS H IGHPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 12/52
If we write down the Fourier series for g = ( √22 , −
√22 ), we have
G (ω) =√22 −
√22 e −i ω
Simplifying gives
G (ω) =
√22 −
√22 e −
i ω
=√22
e −i ω/ 2 e i ω/ 2 −e −i ω/ 2)
= √2ie −i ω/ 2 sin (ω/ 2)
So
|G (ω)| = √2|sin (ω/ 2)|
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 4 / 5
L OWPASS /H IGHPASS F ILTERS H IGHPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 13/52
If we write down the Fourier series for g = ( √22 , −
√22 ), we have
G (ω) =√22 −
√22 e −i ω
Simplifying gives
G (ω) =
√22 −
√22 e −
i ω
=√22
e −i ω/ 2 e i ω/ 2 −e −i ω/ 2)
= √2ie −i ω/ 2 sin (ω/ 2)
So
|G (ω)| = √2|sin (ω/ 2)|
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 4 / 5
L OWPASS /H IGHPASS F ILTERS H IGHPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 14/52
If we write down the Fourier series for g = ( √22 , −
√22 ), we have
G (ω) =√22 −
√22 e −i ω
Simplifying gives
G (ω) =
√22 −
√22 e −
i ω
=√22
e −i ω/ 2 e i ω/ 2 −e −i ω/ 2)
= √2ie −i ω/ 2 sin (ω/ 2)
So
|G (ω)| = √2|sin (ω/ 2)|
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 4 / 5
L OWPASS /H IGHPASS F ILTERS H IGHPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 15/52
Plotting |G (ω)| = √2|sin (ω/ 2)|
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 4 / 5
L OWPASS /H IGHPASS F ILTERS H IGHPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 16/52
Suppose g = ( g n , . . . , g m ) with Fourier series
G (ω) =
m
k = n g k e
ik ω
We will say g is a highpass lter if
G (0) =m
k = n
h k = 0
and
G (π) =m
k = n h k e ik π =
m
k = n h k (−1)k = √2
Highpass lters tend to preserve oscillatory data and eitherdampen or annihilate non-oscillatory data.
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 4 / 5
L OWPASS /H IGHPASS F ILTERS H IGHPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 17/52
Suppose g = ( g n , . . . , g m ) with Fourier series
G (ω) =
m
k = n g k e ik ω
We will say g is a highpass lter if
G (0) =m
k = n h k = 0
and
G (π) =m
k = n h k e ik π =
m
k = n h k (−1)k = √2
Highpass lters tend to preserve oscillatory data and eitherdampen or annihilate non-oscillatory data.
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 4 / 5
L OWPASS /H IGHPASS F ILTERS H IGHPASS F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 18/52
Suppose g = ( g n , . . . , g m ) with Fourier series
G (ω) =
m
k = n g k e ik ω
We will say g is a highpass lter if
G (0) =m
k = n h k = 0
and
G (π) =
m
k = n h k e ik π =
m
k = n h k (−1)k = √2
Highpass lters tend to preserve oscillatory data and eitherdampen or annihilate non-oscillatory data.
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 4 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 19/52
We can build longer orthogonal lters h , g .Orthogonal lters are those that give rise to an orthogonal wavelettransformation W N (N even).It turns out length 3 doesn’t work, so we will try length 4.Let lowpass h = ( h 0 , h 1 , h 2 , h 3 ) and highpass g = ( g 0 , g 1 , g 2 , g 3).
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 20/52
We can build longer orthogonal lters h , g .Orthogonal lters are those that give rise to an orthogonal wavelettransformation W N (N even).It turns out length 3 doesn’t work, so we will try length 4.Let lowpass h = ( h 0 , h 1 , h 2 , h 3 ) and highpass g = ( g 0 , g 1 , g 2 , g 3).
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 21/52
We can build longer orthogonal lters h , g .Orthogonal lters are those that give rise to an orthogonal wavelettransformation W N (N even).It turns out length 3 doesn’t work, so we will try length 4.Let lowpass h = ( h 0 , h 1 , h 2 , h 3 ) and highpass g = ( g 0 , g 1 , g 2 , g 3).
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 22/52
We can build longer orthogonal lters h , g .Orthogonal lters are those that give rise to an orthogonal wavelettransformation W N (N even).It turns out length 3 doesn’t work, so we will try length 4.Let lowpass h = ( h 0 , h 1 , h 2 , h 3 ) and highpass g = ( g 0 , g 1 , g 2 , g 3).
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 23/52
We form the matrix
W 8 =
h 3 h 2 h 1 h 0 0 0 0 00 0 h 3 h 2 h 1 h 0 0 00 0 0 0 h 3 h 2 h 1 h 0h 1 h 0 0 0 0 0 h 3 h 2
g 3 g 2 g 1 g 0 0 0 0 00 0 g 3 g 2 g 1 g 0 0 00 0 0 0 g 3 g 2 g 1 g 0g 1 g 0 0 0 0 0 g 3 g 2
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 24/52
The order of the lters is not important - it turns out reectionswork as well.The reverse order I’ve used reects the fact that the wavelet matrixwas rst motivated using convolution (see Section 5.1 of the text).Note the “wrapping of rows”. This makes it easy to set uporthogonality conditions for W 8 , but is not practical in applications.
In this case, the fourth and eighth elements of the transformeddata will be built using elements from the beginning and end of theinput vector.This makes sense if the data are periodic ...
But data are not typically periodic!We can periodize input data, but then we need more from ourlters ... that’s coming!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 25/52
The order of the lters is not important - it turns out reectionswork as well.The reverse order I’ve used reects the fact that the wavelet matrixwas rst motivated using convolution (see Section 5.1 of the text).Note the “wrapping of rows”. This makes it easy to set uporthogonality conditions for W 8 , but is not practical in applications.
In this case, the fourth and eighth elements of the transformeddata will be built using elements from the beginning and end of theinput vector.This makes sense if the data are periodic ...
But data are not typically periodic!We can periodize input data, but then we need more from ourlters ... that’s coming!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 26/52
The order of the lters is not important - it turns out reectionswork as well.
The reverse order I’ve used reects the fact that the wavelet matrixwas rst motivated using convolution (see Section 5.1 of the text).Note the “wrapping of rows”. This makes it easy to set uporthogonality conditions for W 8 , but is not practical in applications.
In this case, the fourth and eighth elements of the transformeddata will be built using elements from the beginning and end of theinput vector.This makes sense if the data are periodic ...
But data are not typically periodic!We can periodize input data, but then we need more from ourlters ... that’s coming!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 27/52
The order of the lters is not important - it turns out reectionswork as well.
The reverse order I’ve used reects the fact that the wavelet matrixwas rst motivated using convolution (see Section 5.1 of the text).Note the “wrapping of rows”. This makes it easy to set uporthogonality conditions for W 8 , but is not practical in applications.
In this case, the fourth and eighth elements of the transformeddata will be built using elements from the beginning and end of theinput vector.This makes sense if the data are periodic ...
But data are not typically periodic!We can periodize input data, but then we need more from ourlters ... that’s coming!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 28/52
The order of the lters is not important - it turns out reectionswork as well.
The reverse order I’ve used reects the fact that the wavelet matrixwas rst motivated using convolution (see Section 5.1 of the text).Note the “wrapping of rows”. This makes it easy to set uporthogonality conditions for W 8 , but is not practical in applications.
In this case, the fourth and eighth elements of the transformeddata will be built using elements from the beginning and end of theinput vector.This makes sense if the data are periodic ...
But data are not typically periodic!We can periodize input data, but then we need more from ourlters ... that’s coming!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 29/52
The order of the lters is not important - it turns out reectionswork as well.
The reverse order I’ve used reects the fact that the wavelet matrixwas rst motivated using convolution (see Section 5.1 of the text).Note the “wrapping of rows”. This makes it easy to set uporthogonality conditions for W 8 , but is not practical in applications.
In this case, the fourth and eighth elements of the transformeddata will be built using elements from the beginning and end of theinput vector.This makes sense if the data are periodic ...
But data are not typically periodic!We can periodize input data, but then we need more from ourlters ... that’s coming!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 30/52
The order of the lters is not important - it turns out reectionswork as well.
The reverse order I’ve used reects the fact that the wavelet matrixwas rst motivated using convolution (see Section 5.1 of the text).Note the “wrapping of rows”. This makes it easy to set uporthogonality conditions for W 8 , but is not practical in applications.
In this case, the fourth and eighth elements of the transformeddata will be built using elements from the beginning and end of theinput vector.This makes sense if the data are periodic ...
But data are not typically periodic!We can periodize input data, but then we need more from ourlters ... that’s coming!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 31/52
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 32/52
Orthogonality Conditions
h 20 + h
21 + h
22 + h
23 = 1
h 0h 2 + h 1h 3 = 0
g 20 + g 21 + g 22 + g 23 = 1
g 0g 2 + g 1g 3 = 0
h 0g 0 + h 1g 1 + h 2g 2 + h 3g 3 = 0h 0g 2 + h 1g 3 = 0
h 2g 0 + h 3g 1 = 0
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 33/52
and Lowpass/Highpass Conditions
h 0 + h 1 + h 2 + h 3 = √2h 0 −h 1 + h 2 −h 3 = 0
g 0 + g 1 + g 2 + g 3 = 0g 0 −g 1 + g 2 −g 3 = √2
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 34/52
Look at the 5th equation:
h 20 + h
21 + h
22 + h
23 = 1
h 0h 2 + h 1h 3 = 0
g 20 + g 21 + g 22 + g 23 = 1
g 0g 2 + g 1g 3 = 0
h 0g 0 + h 1g 1 + h 2g 2 + h 3g 3 = 0h 0g 2 + h 1g 3 = 0
h 2g 0 + h 3g 1 = 0
Given h 0 , h 1 , h 2 , h 3 , can you nd g 0 , g 1 , g 2 , g 3 to satisfy it?
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 35/52
How about
g 0 = h 3 , g 1 =
−h 2 , g 2 = h 1 , g 3 =
−h 0
Then
h 0g 0 + h 1g 1 + h 2g 2 + h 3g 3 = h 0 (h 3)+ h 1(−h 2)+ h 2(h 1)+ h 3(−h 0) = 0
Moreover, if h 0 + h 1 + h 2 + h 3 = √2 and h 0 −h 1 + h 2 −h 3 = 0, then
g 0 + g 1 + g 2 + g 3 = h 3 −h 2 + h 1 −h 0 = 0
and
g 0 −g 1 + g 2 −g 3 = h 3 −(−h 2) + h 1 −(−h 0) = √2
g is highpass!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 36/52
How about
g 0 = h 3 , g 1 =
−h 2 , g 2 = h 1 , g 3 =
−h 0
Then
h 0g 0 + h 1g 1 + h 2g 2 + h 3g 3 = h 0 (h 3)+ h 1(−h 2)+ h 2(h 1)+ h 3(−h 0) = 0
Moreover, if h 0 + h 1 + h 2 + h 3 = √2 and h 0 −h 1 + h 2 −h 3 = 0, then
g 0 + g 1 + g 2 + g 3 = h 3 −h 2 + h 1 −h 0 = 0
and
g 0 −g 1 + g 2 −g 3 = h 3 −(−h 2) + h 1 −(−h 0) = √2
g is highpass!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 37/52
How about
g 0 = h 3 , g 1 =
−h 2 , g 2 = h 1 , g 3 =
−h 0
Then
h 0g 0 + h 1g 1 + h 2g 2 + h 3g 3 = h 0 (h 3)+ h 1(−h 2)+ h 2(h 1)+ h 3(−h 0) = 0
Moreover, if h 0 + h 1 + h 2 + h 3 = √2 and h 0 −h 1 + h 2 −h 3 = 0, then
g 0 + g 1 + g 2 + g 3 = h 3 −h 2 + h 1 −h 0 = 0
and
g 0 −g 1 + g 2 −g 3 = h 3 −(−h 2) + h 1 −(−h 0) = √2
g is highpass!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 38/52
How about
g 0 = h 3 , g 1 =
−h 2 , g 2 = h 1 , g 3 =
−h 0
Then
h 0g 0 + h 1g 1 + h 2g 2 + h 3g 3 = h 0 (h 3)+ h 1(−h 2)+ h 2(h 1)+ h 3(−h 0) = 0
Moreover, if h 0 + h 1 + h 2 + h 3 = √2 and h 0 −h 1 + h 2 −h 3 = 0, then
g 0 + g 1 + g 2 + g 3 = h 3 −h 2 + h 1 −h 0 = 0
and
g 0 −g 1 + g 2 −g 3 = h 3 −(−h 2) + h 1 −(−h 0) = √2
g is highpass!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 39/52
There’s more good news ...
The orthogonality conditions reduce toh 20 + h 21 + h 22 + h 23 = 1
h 0h 2 + h 1h 3 = 0
Add to that the lowpass conditions:h 0 + h 1 + h 2 + h 3 = √2h 0 −h 1 + h 2 −h 3 = 0
and we have a (quadratic) system to solve.The moral of the story is we can build g from h .
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 40/52
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 41/52
There’s more good news ...
The orthogonality conditions reduce toh 20 + h 21 + h 22 + h 23 = 1
h 0h 2 + h 1h 3 = 0
Add to that the lowpass conditions:h 0 + h 1 + h 2 + h 3 = √2h 0 −h 1 + h 2 −h 3 = 0
and we have a (quadratic) system to solve.The moral of the story is we can build g from h .
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 42/52
There’s more good news ...
The orthogonality conditions reduce toh 20 + h 21 + h 22 + h 23 = 1
h 0h 2 + h 1h 3 = 0
Add to that the lowpass conditions:h 0 + h 1 + h 2 + h 3 = √2h 0 −h 1 + h 2 −h 3 = 0
and we have a (quadratic) system to solve.The moral of the story is we can build g from h .
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 43/52
The bad news (for mathematicians) is that the system has aninnite number of solutions.The good news (for engineers) is that the system has an innitenumber of solutions.
We can add a condition!The condition Ingrid Daubechies added:“Flatten” H (ω) at ω = π .Set H (π) = 0!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 44/52
The bad news (for mathematicians) is that the system has aninnite number of solutions.The good news (for engineers) is that the system has an innitenumber of solutions.
We can add a condition!The condition Ingrid Daubechies added:“Flatten” H (ω) at ω = π .Set H (π) = 0!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
ORTHOGONAL
FILTERS
DAUBECHIES
4-TAP
FILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 45/52
The bad news (for mathematicians) is that the system has aninnite number of solutions.The good news (for engineers) is that the system has an innitenumber of solutions.
We can add a condition!The condition Ingrid Daubechies added:“Flatten” H (ω) at ω = π .Set H (π) = 0!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
ORTHOGONAL
FILTERS
DAUBECHIES
4-TAP
FILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 46/52
The bad news (for mathematicians) is that the system has aninnite number of solutions.The good news (for engineers) is that the system has an innitenumber of solutions.
We can add a condition!The condition Ingrid Daubechies added:“Flatten” H (ω) at ω = π .Set H (π) = 0!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 47/52
The bad news (for mathematicians) is that the system has aninnite number of solutions.The good news (for engineers) is that the system has an innitenumber of solutions.
We can add a condition!The condition Ingrid Daubechies added:“Flatten” H (ω) at ω = π .Set H (π) = 0!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 48/52
The bad news (for mathematicians) is that the system has aninnite number of solutions.The good news (for engineers) is that the system has an innitenumber of solutions.
We can add a condition!The condition Ingrid Daubechies added:“Flatten” H (ω) at ω = π .Set H (π) = 0!
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 49/52
Differentiating
H (ω) = h 0 + h 1 e i ω + h 2 e 2i ω + h 3 e 3i ω
givesH (ω) = ih 1 e i ω + 2ih ,e 2i ω + 3ih 3 e 3i ω
Plugging in ω = π and simplifying gives the condition
h 1 −2h 2 + 3h 3 = 0
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 50/52
Differentiating
H (ω) = h 0 + h 1 e i ω + h 2 e 2i ω + h 3 e 3i ω
givesH (ω) = ih 1 e i ω + 2ih ,e 2i ω + 3ih 3 e 3i ω
Plugging in ω = π and simplifying gives the condition
h 1 −2h 2 + 3h 3 = 0
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 51/52
So we want to solveh 20 + h 21 + h 22 + h 23 = 1
h 0h 2 + h 1h 3 = 0h 0 + h 1 + h 2 + h 3 = √2h 0 −h 1 + h 2 −h 3 = 0
h 1 −2h 2 + 3h 3 = 0
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5
O RTHOGONAL F ILTERS D AUBECHIES 4-T AP F ILTERS
8/6/2019 DaubechiesFilters
http://slidepdf.com/reader/full/daubechiesfilters 52/52
Let’s look at the notebook
DaubechiesFilters.nb
8 JANUARY 2007 (S ESSION 2) DAUBECHIES F ILTERS JMM M INICOURSE #5 5 / 5