DSP Ch11 S1 5

Chapter 11
Design of Recursive (IIR) Filters

11.1 Introduction
11.2 Realizability Constraints
11.3 Invariant Impulse-Response Method
11.4 Modied Invariant Impulse-Response Method
11.5 Matched-z Transformation Method
Copyright c 2005- Andreas Antoniou
Victoria, BC, Canada
Email: aantoniou@ieee.org
December 4, 2008
Frame # 1 Slide # 1 A. Antoniou Digital Signal Processing Secs. 11.1 to 11.5
Introduction
t
Approximation methods for the design of recursive (IIR) lters
dier quite signicantly from those used for the design of
nonrecursive lters.
Introduction
t
nonrecursive lters.
t
The basic reason is that in recursive lters the transfer
function is a ratio of polynomials of z whereas in nonrecursive
lters it is a polynomial of negative powers of z.
Introduction
t
nonrecursive lters.
t
The basic reason is that in recursive lters the transfer
function is a ratio of polynomials of z whereas in nonrecursive
lters it is a polynomial of negative powers of z.
t
In recursive lters, the approximation problem is usually solved
through indirect or direct methods.
Introduction Contd
t
In indirect methods, a continuous-time transfer function that
satises certain specications is rst obtained by using one of
the standard analog-lter approximations.
Introduction Contd
t
t
The continuous-time transfer function obtained is then
converted into a discrete-time transfer function.
Introduction Contd
t
t
t
In direct methods, the design problem is formulated as an
optimization problem which is then solved using standard
optimization methods.
Introduction Contd
t
t
t
In direct methods, the design problem is formulated as an
optimization problem which is then solved using standard
optimization methods.
t
This presentation will deal with some indirect methods for the
design of recursive lters.
Introduction Contd
t
Several indirect approximation methods are available as
follows:
Invariant impulse-response method
Modied invariant impulse-response method
Matched-z transformation method
Bilinear transformation method
Introduction Contd
t
follows:
t
The most important among them are the invariant impulse-
response and bilinear transformation methods.
Introduction Contd
t
follows:
t
t
The rst three are closely interrelated and will form the
subject matter of this presentation.
Introduction Contd
t
follows:
t
t
The rst three are closely interrelated and will form the
subject matter of this presentation.
The bilinear transformation method will be discussed in the
next presentation.
Realizability Constraints
t
Before we discuss the available approximation methods for recursive
lters, it is important to mention the constraints that are imposed
on the transfer function, which are as follows:
t
It must be a rational function of z with real coecients.
This follows from the fact that digital-lters comprise unit
delays, adders, and multipliers with real multiplier constants.
t
Its poles must lie within the unit circle of the z plane to ensure
that the lter is stable.
t
The degree of the numerator polynomial must be equal to or
less than the degree of the denominator polynomial to ensure
that the lter is causal.
t
The degree of the numerator polynomial must be equal to or
less than the degree of the denominator polynomial to ensure
that the lter is causal.
t
These constraints will ensure that the transfer function is realizable
in the form of a stable digital-lter network and are, therefore, said
to be the realizability constraints.
Invariant Impulse-Response Method
t
Given an analog lter, a corresponding digital lter can be
obtained by constructing an impulse-modulated lter

F
A
as
shown in the gure where S is an ideal impulse modulator and
F
A
is an analog lter characterized by a continuous-time
transfer function H
A
(s).
F
A
S
F
A
^
Invariant Impulse-Response Method Contd
t
On the basis of the Poison summation formula (see Chap. 6),
the impulse-modulated lter can be represented by a
continuous-time transfer function

H
A
(s) or, equivalently, by a
discrete-time transfer function H
D
(z) as follows:
H
A
(j ) = H
D
(e
j T
) =
h
A
(0+)
2
+
1
T
k=
H
A
(j + jk
s
)
where
h
A
(t) = L
1
H
A
(s), h
A
(0+) = lim
s
[sH
A
(s)],
and
H
D
(z) = Zh
A
(nT)
t
Therefore, given an analog lter, a corresponding digital lter
can be obtained as follows:
t
Deduce the impulse response of the analog lter as
h
A
(t) = L
1
H
A
(s)
t
h
A
(t) = L
1
H
A
(s)
Replace t by nT in h
A
(t) to obtain h
A
(nT).
t
h
A
(t) = L
1
H
A
(s)
A
(t) to obtain h
A
(nT).
Obtain the z transform of h
A
(nT), i.e.,
H
D
(z) = Zh
A
(nT)
t
h
A
(t) = L
1
H
A
(s)
A
(t) to obtain h
A
(nT).
Obtain the z transform of h
A
(nT), i.e.,
H
D
(z) = Zh
A
(nT)
t
The method is referred to as the invariant-impulse response
method because the impulse response of the digital lter is
exactly equal to the impulse response of the analog lter at
t = nT for n = 0, 1, . . . , .
t
Obviously, we have a way of generating a discrete-time
transfer function from a given continuous-time transfer
function and, therefore, we can design a digital lter starting
with an analog lter.
t
t
However, the following important questions will immediately
arise:
t
t
arise:
1. How is the frequency response of the derived digital lter
related to that of the analog lter?
t
t
arise:
2. Would the digital lter obtained be stable if the analog lter is
stable?
t
t
arise:
2. Would the digital lter obtained be stable if the analog lter is
stable?
3. Would the discrete-time transfer function obtained have real
coecients?
t
Let us examine the issue of the frequency response.
t
As was mentioned earlier, the Poisson summation formula gives the
frequency response of the derived digital lter as
H
A
(j ) = H
D
(e
jT
) =
h(0+)
2
+
1
T
k=
H
A
(j + jk
s
)
t
As was mentioned earlier, the Poisson summation formula gives the
frequency response of the derived digital lter as
H
A
(j ) = H
D
(e
jT
) =
h(0+)
2
+
1
T
k=
H
A
(j + jk
s
)
t
If H
A
(j ) 0 for ||

s
2
then
k=,k=0
H
A
(j + jk
s
) 0 for || <

s
2
i.e., the side-bands contribute a negligible amount of aliasing error.
t
If , in addition,h
A
(0+) = 0 then
H
A
(j ) = H
D
(e
jT
)
1
T
H
A
(j ) for || <

s
2
t
If , in addition,h
A
(0+) = 0 then
H
A
(j ) = H
D
(e
jT
)
1
T
H
A
(j ) for || <

s
2
t
In eect, given an analog lter that satises the stated bandlimiting
conditions, a digital lter can be derived which would have
approximately the same frequency response as the analog lter to
within a multiplier constant 1/T.
t
If , in addition,h
A
(0+) = 0 then
H
A
(j ) = H
D
(e
jT
)
1
T
H
A
(j ) for || <

s
2
t
t
Since
h
A
(0+) = lim
s
[sH
A
(s)],
the condition h
A
(0+) = 0 is satised if the denominator degree in
H
A
(s) exceeds the numerator degree by at least two.
t
If , in addition,h
A
(0+) = 0 then
H
A
(j ) = H
D
(e
jT
)
1
T
H
A
(j ) for || <

s
2
t
t
Since
h
A
(0+) = lim
s
[sH
A
(s)],
the condition h
A
(0+) = 0 is satised if the denominator degree in
H
A
(s) exceeds the numerator degree by at least two.
t
The multiplier constant 1/T can be eliminated by multiplying the
discrete-time transfer function obtained, H
D
(z), by T.
40 30 20 10 0 10 20 30 40
0
0.1
0.2
40 30 20 10 0 10 20 30 40
0
0.1
0.2
H
A
(j - j
s
)
H
A
(j+ j
s
)
H
D
(e
jT
)
H
A
(j)
H
A
(j)
Design Procedure
Given an analog lter characterized by a transfer function H
A
(s)
that satises the bandlimiting conditions stated, a digital lter can
be obtained by applying the following design procedure:
1. If the transfer function H
A
(s) is given in terms of its
coecients, i.e.,
H
A
(s) =
M
i =0
a
i
s
i
N
i =0
b
i
s
i
express it in terms of its zeros and poles as
H
A
(s) = H
0
M
i =1
(s z
i
)
N
i =1
(s p
i
)
Design Procedure Contd

H
A
(s) = H
0
M
i =1
(s z
i
)
N
i =1
(s p
i
)
2. Express the transfer function in terms of partial fractions as
H
A
(s) =
N
i =1
A
i
s p
i

H
A
(s) = H
0
M
i =1
(s z
i
)
N
i =1
(s p
i
)
2. Express the transfer function in terms of partial fractions as
H
A
(s) =
N
i =1
A
i
s p
i
3. Deduce the impulse response of the analog lter as follows:
h
A
(t) = L
1
H
A
(s) =
N
i =1
A
i
e
p
i
t

h
A
(t) =
N
i =1
A
i
e
p
i
t
4. Replace t by nT in h
A
(t) to obtain h
A
(nT) as
h
A
(nT) =
N
i =1
A
i
e
p
i
nT

h
A
(t) =
N
i =1
A
i
e
p
i
t
4. Replace t by nT in h
A
(t) to obtain h
A
(nT) as
h
A
(nT) =
N
i =1
A
i
e
p
i
nT
5. Obtain the z transform of h
A
(nT) as
H
D
(z) = Zh
A
(nT) =
N
i =1
A
i
z
z e
Tp
i

H
D
(z) =
N
i =1
A
i
z
z e
Tp
i
6. Multiply the discrete-time transfer function obtained in Step 5 by
T, i.e.,
H
D
(z) = TH
D
(z)

H
D
(z) =
N
i =1
A
i
z
z e
Tp
i
6. Multiply the discrete-time transfer function obtained in Step 5 by
T, i.e.,
H
D
(z) = TH
D
(z)
7. Combine partial fractions with complex conjugate poles to obtain
the modied discrete-time transfer function
H
D
(z) =
N/2
i =1
a
1i
z + a
2i
z
2
b
0i
+ b
1i
z + b
2i
z
2
for even N
or
H
D
(z) =
a
11
z
b
01
+ z
+
(N1)/2
i =2
a
1i
z + a
2i
z
2
b
0i
+ b
1i
z + b
2i
z
2
for odd N
Notes
t
The bandlimiting condition implies that the numerator degree in
H
A
(s) is less than the denominator degree.
Consequently, there is no constant term in the partial fraction
expansion of Step 2.
Notes
t
H
A
t
In Step 5, we note that a pole p
i
=
i
+ j
i
in the analog lter will
yield a pole p
i
in the digital lter where
p
i
= e
Tp
i
= e
T(
i
+j
i
)
= e
i
T
e
j
i
T
Notes
t
H
A
t
i
=
i
+ j
i
yield a pole p
i
p
i
= e
Tp
i
= e
T(
i
+j
i
)
= e
i
T
e
j
i
T
Hence | p
i
| = e
i
T
and if
i
< 0 then | p
i
| < 1.
Notes
t
H
A
t
i
=
i
+ j
i
yield a pole p
i
p
i
= e
Tp
i
= e
T(
i
+j
i
)
= e
i
T
e
j
i
T
Hence | p
i
| = e
i
T
and if
i
< 0 then | p
i
| < 1.
Therefore, if the analog lter has poles in the left-half s plane, the
poles of the digital lter obtained will be located in the unit circle of
the z plane.
Notes
t
H
A
t
i
=
i
+ j
i
yield a pole p
i
p
i
= e
Tp
i
= e
T(
i
+j
i
)
= e
i
T
e
j
i
T
Hence | p
i
| = e
i
T
and if
i
< 0 then | p
i
| < 1.
Therefore, if the analog lter has poles in the left-half s plane, the
poles of the digital lter obtained will be located in the unit circle of
the z plane.
That is, a stable analog lter will yield a stable digital lter.
Notes Contd
t
In Step 6, we multiply the discrete-time transfer function obtained
in Step 5 by T in order to cancel the multiplier constant 1/T
introduced by the Poisson summation formula.
Notes Contd
t
Thus we obtain a digital lter that has a baseband frequency
response which is approximately the same as the frequency response
of the analog lter.
Notes Contd
t
of the analog lter.
t
In Step 7, the coecients of
H
D
(z) =
K
i =1
a
1i
z + a
2i
z
2
b
0i
+ b
1i
z + b
2i
z
2
turn out to be real as required.
Notes Contd
t
of the analog lter.
t
In Step 7, the coecients of
H
D
(z) =
K
i =1
a
1i
z + a
2i
z
2
b
0i
+ b
1i
z + b
2i
z
2
turn out to be real as required.
This is due to the fact that complex conjugate poles give complex
conjugate residues.
Notes Contd
t
If the transfer function is of odd order, the continuous-time
transfer function H
A
(s) has a real pole which will produce a
rst-order partial transfer function in Step 7.
Notes Contd
t
transfer function H
A
t
Each of the partial transfer functions in Step 7 represents a
causal lter since the numerator degree does not exceed the
denominator degree in each case.
Consequently, H
D
(z) also represents a causal lter.
Notes Contd
t
transfer function H
A
t
Each of the partial transfer functions in Step 7 represents a
causal lter since the numerator degree does not exceed the
denominator degree in each case.
Consequently, H
D
(z) also represents a causal lter.
t
A lter designed by using the invariant-impulse-response
method can be conveniently realized by using the parallel
realization since the overall transfer function is a sum of rst-
or second-order transfer functions.
Features
t
The method gives good results for lowpass and bandpass
Butterworth, Bessel, and Chebyshev lters.
Features
t
This is because, these lters have a bandlimited frequency
response.
Features
t
response.
t
The results are not so good for lowpass and bandpass elliptic
lters.
Features
t
response.
t
lters.
In these lters, the frequency response is not bandlimited
since the stopband gain oscillates between zero and a
specied maximum.
Features
t
response.
t
lters.
In these lters, the frequency response is not bandlimited
since the stopband gain oscillates between zero and a
specied maximum.
t
It does not work at all for highpass lters since these lters
are not bandlimited by denition.
Features Contd
t
If the bandlimiting condition is satised, then the frequency
response is preserved which includes both the amplitude
response and phase responses.
Features Contd
t
Therefore, the method tends to preserve the phase response
of the analog lter.
Features Contd
t
of the analog lter.
Thus if a linear-phase continuous-time transfer function is
used (e.g., Bessel-Thomson), a linear-phase digital lter is
obtained.
Features Contd
t
of the analog lter.
obtained.
t
To apply the method, all one needs to do is to nd the
residues of the continuous-time transfer function and calculate
the poles of the discrete-time transfer function using the poles
of the continuous-time transfer function.
Features Contd
t
of the analog lter.
obtained.
t
To apply the method, all one needs to do is to nd the
residues of the continuous-time transfer function and calculate
the poles of the discrete-time transfer function using the poles
of the continuous-time transfer function.
That is, the method is relatively simple to apply.
Example
Starting with a third-order normalized lowpass Chebyshev transfer
function, obtain a discrete-time transfer function using the
invariant-impulse response method.
Assume a maximum passband loss A
p
= 1.0 dB and a sampling
frequency
s
= 10.0 rad/s.
Example
Starting with a third-order normalized lowpass Chebyshev transfer
function, obtain a discrete-time transfer function using the
invariant-impulse response method.
Assume a maximum passband loss A
p
= 1.0 dB and a sampling
frequency
s
= 10.0 rad/s.
Solution The required Chebyshev transfer function can be readily
obtained as
H
A
(s) =
H
0
(s p
1
)(s p
2
)(s p
2
)
where
H
0
= 0.4913, p
1
= 0.4942, and p
2
, p
2
= 0.2471 j 0.9660
(See Chap. 10.)
Example Contd
On expanding H
A
(s) into partial fractions as in Step 2 of the
design procedure, we obtain
H
A
(s) =
A
1
s p
1
+
A
2
s p
2
+
A
3
s p
2
where
A
1
=
H
0
(s p
2
)(s p
2
)
s=p
1
=
H
0
(p
1
p
2
)(p
1
p
2
)
= 0.4942
A
2
=
H
0
(s p
1
)(s p
2
)
s=p
2
=
H
0
(p
2
p
1
)(p
2
p
2
)
= 0.2471 j 0.0632
A
3
= A
2
= 0.2471 + j 0.0632
Example Contd

H
A
(s) =
A
1
s p
1
+
A
2
s p
2
+
A
3
s p
2
From Step 3, the impulse response of the analog lter is obtained
as
h
A
(t) = L
1
H
A
(s) = A
1
e
p
1
t
+ A
2
e
p
2
t
+ A
2
e
p
2
t
Example Contd

H
A
(s) =
A
1
s p
1
+
A
2
s p
2
+
A
3
s p
2
From Step 3, the impulse response of the analog lter is obtained
as
h
A
(t) = L
1
H
A
(s) = A
1
e
p
1
t
+ A
2
e
p
2
t
+ A
2
e
p
2
t
Replacing t by nT in h
A
(t), as in Step 4, gives
h
A
(nT) = A
1
e
p
1
nT
+ A
2
e
p
2
nT
+ A
2
e
p
2
nT
where T = 2/
s
= 2/10.0 = 0.6283 s.
Example Contd

h
A
(nT) = A
1
e
p
1
nT
+ A
2
e
p
2
nT
+ A
2
e
p
2
nT
On applying the z transform to h
A
(nT), as in Step 5, we get the
discrete-time transfer function as
H
D
(z) = Zh
A
(nT) =
A
1
z
z e
Tp
1
+
A
2
z
z e
Tp
2
+
A
2
z
z e
Tp
2
Example Contd
Now on multiplying H
D
(z) by T (to adjust the gain of the lter)
and then combining partial fractions with complex conjugate poles,
the discrete-time transfer function can be expressed as
H
D
(z) =
A
1
z
z e
Tp
1
+
A
2
z
z e
Tp
2
+
A
2
z
z e
Tp
3
=
A
1
z
z e
p
1
T
+
(A
2
+ A
2
)z
2
(A
2
e
p
2
T
+ A
2
e
p
2
T
)z
z
2
(e
p
2
T
+ e
p
2
T
)z + e
p
2
T
e
p
2
T
=
A
1
z
z e
p
1
T
+
2Re (A
2
)z
2
2Re (A
2
e
p
2
T
)z
z
2
2Re (e
p
2
T
)z +|e
p
2
T
|
2
=
a
11
z
z + b
01
+
a
22
z
2
+ a
12
z
z
2
+ b
12
z + b
02
Example Contd

H
D
(z) =
a
11
z
z + b
01
+
a
22
z
2
+ a
12
z
z
2
+ b
12
z + b
02
where
a
11
= A
1
= 0.3105
b
01
= e
Tp
1
= 0.7331
a
22
= 2Re (A
2
) = 0.4942
a
12
= 2Re (A
2
e
p
2
T
) = 0.4093
b
12
= 2Re (e
p
2
T
) = 1.4065
b
02
= |e
p
2
T
|
2
= 0.7331
Example
Design a digital lter by applying the invariant impulse-response
method to the Bessel-Thomson transfer function
H
A
(s) =
105
105 + 105s + 45s
2
+ 10s
3
+ s
4
Employ a sampling frequency
s
= 8 rad/s; repeat with
s
= 16
rad/s.
Example
Design a digital lter by applying the invariant impulse-response
method to the Bessel-Thomson transfer function
H
A
(s) =
105
105 + 105s + 45s
2
+ 10s
3
+ s
4
Employ a sampling frequency
s
= 8 rad/s; repeat with
s
= 16
rad/s.
Solution The poles and residues of H
A
(s) are given by
p
1
, p
1
= 2.896211 j 0.8672341
p
2
, p
2
= 2.103789 j 2.657418
R
1
, R
1
= 1.663392 j 8.396299
R
2
, R
2
= 1.663392 j 2.244076
Example Contd
Steps 1 to 7 of the design procedure yield
TH
D
(z) =
2
j=1
a
1j
z + a
2j
z
2
b
0j
+ b
1j
z + z
2
s
j a
1j
a
2j
b
0j
b
1j
1 6.452333E-1 2.612851 1.057399E-2 -1.597700E-1
8
2 -8.345233E-1 -2.612851 3.671301E-2 1.891907E-1
1 3.114550E-1 1.306425 1.028299E-1 -6.045080E-1
16
2 -3.790011E-1 -1.306425 1.916064E-1 -4.404794E-1
Example Contd
t
Loss characteristic (i.e., 20 log[1/M()] versus ):
0 0.5 1 1.5 2 2.5 3 3.5 4
2
0
2
4
6
8
10
12
14
, rad/s
L
o
s
s
,

d
B
Digital filter
s
=8 rad/s
Digital filter
s
=16 rad/s
Analog filter
(a)
Example Contd
t
Delay characteristic (i.e., group delay versus ):
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
0
0.2
0.4
0.6
0.8
1.0
D
e
l
a
y
,
s
, rad/s
Digital filter
s
=8 rad/s
Digital filter
s
=16 rad/s
Analog filter
(b)
Modied Invariant Impulse-Response Method
t
Aliasing errors tend to restrict the application of the invariant
impulse-response method to the design of allpole lters, i.e.,
lters that have no zeros in the nite s plane.
Modied Invariant Impulse-Response Method
t
Aliasing errors tend to restrict the application of the invariant
impulse-response method to the design of allpole lters, i.e.,
lters that have no zeros in the nite s plane.
t
However, a modied version of the method is available, which
can be applied to lters that also have zeros in the nite s
plane.
Modied Invariant Impulse-Response Method Contd
t
Consider the transfer function
H
A
(s) =
H
0
N(s)
D(s)
=
H
0
M
i =1
(s z
i
)
N
i =1
(s p
i
)
where N M.
t
Consider the transfer function
H
A
(s) =
H
0
N(s)
D(s)
=
H
0
M
i =1
(s z
i
)
N
i =1
(s p
i
)
where N M.
t
We can write
H
A
(s) = H
0
H
A1
(s)
H
A2
(s)
where
H
A1
(s) =
1
D(s)
and H
A2
(s) =
1
N(s)

H
A1
(s) =
1
D(s)
and H
A2
(s) =
1
N(s)
t
With M and N 2, we have
h
A
(0+) = lim
s
[sH
A
(s)] = 0

H
A1
(s) =
1
D(s)
and H
A2
(s) =
1
N(s)
t
With M and N 2, we have
h
A
(0+) = lim
s
[sH
A
(s)] = 0
t
Also
lim
s
[H
A1
(s)] 0 and lim
s
[H
A2
(s)] 0
and, consequently,
H
A1
(j ) 0 and H
A2
(j ) 0 for ||

s
2
for some suciently high value of
s
.
t
In eect, by using a suciently high sampling frequency, functions
H
A1
(s) and H
A2
(s) can be considered to be bandlimited analog-lter
transfer functions, and for each a discrete-time transfer function can
be obtained, as follows, by using the invariant impulse-response
method:
H
D1
(z) =
N
1
(z)
D
1
(z)
=
N
i =1
A
i
z
z e
Tp
i
1
T
H
A1
(s) =
1
T
1
D(s)
H
D2
(z) =
N
2
(z)
D
2
(z)
=
M
i =1
B
i
z
z e
Tz
i
1
T
H
A2
(s) =
1
T
1
N(s)
t
In eect, by using a suciently high sampling frequency, functions
H
A1
(s) and H
A2
(s) can be considered to be bandlimited analog-lter
transfer functions, and for each a discrete-time transfer function can
be obtained, as follows, by using the invariant impulse-response
method:
H
D1
(z) =
N
1
(z)
D
1
(z)
=
N
i =1
A
i
z
z e
Tp
i
1
T
H
A1
(s) =
1
T
1
D(s)
H
D2
(z) =
N
2
(z)
D
2
(z)
=
M
i =1
B
i
z
z e
Tz
i
1
T
H
A2
(s) =
1
T
1
N(s)
t
Therefore, a discrete-time transfer function can be formed as
H
D
(z) = H
0
H
D1
(z)
H
D2
(z)
= H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
H
0
N(s)
D(s)
= H
A
(s)

H
D
(z)H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
H
A
(s)
t
Evidently, given an arbitrary analog lter with frequency
response H
A
(j ), a corresponding digital lter can be derived
with a frequency response
H
D
(e
T
) H
A
(j ) for || <

s
2

H
D
(z)H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
H
A
(s)
t
response H
A
H
D
(e
T
) H
A
(j ) for || <

s
2
t
In other words, the modied invariant impulse-response
method eliminates the aliasing problem associated with the
standard invariant impulse-response method.

H
D
(z)H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
H
A
(s)
t
response H
A
H
D
(e
T
) H
A
(j ) for || <

s
2
t
In other words, the modied invariant impulse-response
method eliminates the aliasing problem associated with the
standard invariant impulse-response method.
t
Unfortunately, it also introduces two other problems.

H
D
(z) = H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
H
A
(s)
t
Polynomials N
1
(z) and D
1
(z) are of degree N which is the
denominator degree in H
A
(s) and polynomials N
2
(z) and
D
2
(z) are of degree M which is the numerator degree in
H
A
(s).

H
D
(z) = H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
H
A
(s)
t
Polynomials N
1
(z) and D
1
A
2
(z) and
D
2
H
A
(s).
t
Therefore, the order of the derived digital lter is M + N
instead of N in the case of the standard invariant
impulse-response method.

H
D
(z) = H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
H
A
(s)
t
Polynomials N
1
(z) and D
1
A
2
(z) and
D
2
H
A
(s).
t
t
The zeros of D
1
(z) are located in the unit circle |z| = 1.

H
D
(z) = H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
H
A
(s)
t
Polynomials N
1
(z) and D
1
A
2
(z) and
D
2
H
A
(s).
t
t
The zeros of D
1
(z) are located in the unit circle |z| = 1.
However, the zeros of N
2
(z) may be located outside the unit
circle, which would render the derived digital lter unstable.
Stability Problem
t
The stability problem mentioned in the previous slide can be easily
circumvented without changing the amplitude response of the lter.
Stability Problem
t
t
Consider a discrete-time transfer function
H
D
(z) = H
0
N(z)
D(z)
and assume that it has poles p
1
, p
2
, . . . , p
K
that are located
outside the unit circle |z| = 1.
Stability Problem
t
t
Consider a discrete-time transfer function
H
D
(z) = H
0
N(z)
D(z)
and assume that it has poles p
1
, p
2
, . . . , p
K
that are located
outside the unit circle |z| = 1.
H
D
(z) can be stabilized by simply replacing poles p
1
, p
2
, . . . , p
K
by
their reciprocals 1/p
1
, 1/p
2
, . . . , 1/p
K
and then replacing the
multiplier constant H
0
by H
0
/
K
i =1
p
i
.
(See Chap. 11 for proof.)
Features
t
Unlike the standard invariant impulse-response method, the
modied method yields excellent results for elliptic lters.
Features
t
t
The increased lter order leads to more complicated
realizations which would be more expensive to implement.
Features
t
t
t
Because of the higher order, a software implementation of the
discrete-time transfer function obtained would require more
computation.
Features
t
t
t
computation.
t
The stabilization technique preserves the amplitude response,
as stated.
Features
t
t
t
computation.
t
as stated.
However, the phase response is changed which could be a
problem in certain applications.
Features
t
t
t
computation.
t
as stated.
However, the phase response is changed which could be a
problem in certain applications.
t
The method provides a theoretical basis for the matched-z
transformation method to be described later.
Example
The transfer function
H
A
(s) = H
0
3
j=1
a
0j
+ s
2
b
0j
+ b
1j
s + s
2
where H
0
, a
0j
, and b
1j
are given in the table, represents an analog
lowpass elliptic lter.
j a
0j
b
0j
b
1j
1 1.199341E + 1 3.581929E 1 9.508335E 1
2 2.000130 6.860742E 1 4.423164E 1
3 1.302358 8.633304E 1 1.088749E 1
H
0
= 6.713267E 3
Example Contd
The specications of the lter are as follows:
Passband ripple: 0.1 dB
Minimum stopband loss: 43.46 dB
Passband edge:
0.8 rad/s
Stopband edge: 1/
0.8 rad/s
Design a corresponding digital lter by employing the modied
invariant impulse-response method.
Assume a sampling frequency
s
= 7.5 rad/s.
Example Contd
The design can be obtained through the following steps:
1. Let
H
A1
(s) =
3
j=1
1
b
0j
+ b
1j
s + s
2
and H
A2
(s) =
3
j=1
1
a
0j
+ s
2
Example Contd
1. Let
H
A1
(s) =
3
j=1
1
b
0j
+ b
1j
s + s
2
and H
A2
(s) =
3
j=1
1
a
0j
+ s
2
2. Find the poles and residues of H
A1
(s) and H
A2
(s).
Example Contd
1. Let
H
A1
(s) =
3
j=1
1
b
0j
+ b
1j
s + s
2
and H
A2
(s) =
3
j=1
1
a
0j
+ s
2
A1
(s) and H
A2
(s).
3. Form
H
D1
(z) =
N
i =1
A
i
z
z e
Tp
i
and H
D2
(z) =
M
i =1
B
i
z
z e
Tz
i
Example Contd
1. Let
H
A1
(s) =
3
j=1
1
b
0j
+ b
1j
s + s
2
and H
A2
(s) =
3
j=1
1
a
0j
+ s
2
A1
(s) and H
A2
(s).
3. Form
H
D1
(z) =
N
i =1
A
i
z
z e
Tp
i
and H
D2
(z) =
M
i =1
B
i
z
z e
Tz
i
4. Form
H
D
(z) = H
0
H
D1
(z)
H
D2
(z)
= H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
Example Contd
1. Let
H
A1
(s) =
3
j=1
1
b
0j
+ b
1j
s + s
2
and H
A2
(s) =
3
j=1
1
a
0j
+ s
2
A1
(s) and H
A2
(s).
3. Form
H
D1
(z) =
N
i =1
A
i
z
z e
Tp
i
and H
D2
(z) =
M
i =1
B
i
z
z e
Tz
i
4. Form
H
D
(z) = H
0
H
D1
(z)
H
D2
(z)
= H
0
N
1
(z)D
2
(z)
N
2
(z)D
1
(z)
5. Replace poles p
1
, p
2
, . . . , p
K
of H
D
(z) (zeros of N
2
(z)) located
outside the unit circle by their reciprocals and multiplier constant
H
0
by H
0
/
K
i =1
p
i
in order to stabilize the transfer function.
Example Contd
The design procedure gives a transfer function of the form
H
D
(z) = H
0
5
j=1
a
0j
+ a
1j
z + z
2
b
0j
+ b
1j
z + z
2
where H
0
, a
ij
, and b
ij
are given in the table shown.
j a
0j
a
1j
b
0j
b
1j
1 1.0 1.942528 4.508735E-1 1.281134
2 1.0 7.530225E-1 6.903732E-1 1.303838
3 1.0 1.153491 9.128252E-1 1.362371
4 3.248990E+1 1.955491E+1 5.611278E-2 7.751650E-1
5 1.331746E-2 3.971465E-1 5.611278E-2 7.751650E-1
H
0
= 3.847141E-4
Example Contd
t
Loss characteristic with respect to the baseband:
Analog lter; modied impulse-
invariant response method.
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
0
10
20
30
40
50
60
L
o
s
s
,

d
B
, rad/s
(a)
Example Contd
t
Loss characteristic with respect to the passband:
Analog lter; modied impulse-invariant response
method.
0 0.2 0.4 0.6 0.8 1.0
0.10
0.05
0
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
L
o
s
s
,

d
B
, rad/s
(b)
Matched-z Transformation Method
t
It was noted early in the history of digital-lter design that the
invariant impulse-response method yields a discrete-time
transfer function whose poles, p
i
, bear a one-to-one relation
to the poles of the continuous-time transfer function, p
i
, of
the form
p
i
= e
p
i
T
where T is the sampling period.
Matched-z Transformation Method
t
It was noted early in the history of digital-lter design that the
invariant impulse-response method yields a discrete-time
transfer function whose poles, p
i
, bear a one-to-one relation
to the poles of the continuous-time transfer function, p
i
, of
the form
p
i
= e
p
i
T
where T is the sampling period.
t
It did not take too long for someone to explore calculating the
zeros of the discrete-time transfer function, z
i
, from the zeros
of the continuous-time transfer function, z
i
, using the same
relation i.e.,
z
i
= e
z
i
T
Matched-z Transformation Method Contd
t
The technique seemed to work well for some types of analog
lters but not in others, but it was soon discovered that
improved results could be obtained by adding a number of
zeros at the Nyquist point.
t
The technique seemed to work well for some types of analog
lters but not in others, but it was soon discovered that
improved results could be obtained by adding a number of
zeros at the Nyquist point.
t
Further heuristic eort identied the number of Nyquist-point
zeros needed for the various types of analog lters and the
matched-z transformation method was formulated as detailed
next.
t
Given a continuous-time transfer function of the form
H
A
(s) = H
0
M
i =1
(s z
i
)
N
i =1
(s p
i
)
a discrete-time transfer function can be obtained as
H
D
(z) = H
0
(z + 1)
L
M
i =1
(z e
z
i
T
)
N
i =1
(z e
p
i
T
)
where L is an integer.
t
The value of L depends on the type of lter and it is given by
the table shown.
Type of Filter LP HP BP BS
Butterworth N 0 N/2 0
Chebyshev N 0 N/2 0
Inverse-Chebyshev, N odd 1 0 n/a n/a
Neven 0 0 1 for odd N/2 0
0 for even N/2
Elliptic, N odd 1 0 n/a n/a
N even 0 0 1 for odd N/2 0
0 for even N/2
t
If we now compare the discrete-time transfer function given by
the modied invariant impulse-response method, i.e.,
H
D
(z) = H
0
N
1
(z)
N
2
(z)

M
i =1
(z e
z
i
T
)
N
i =1
(z e
p
i
T
)
with that obtained by using the matched-z transformation
method, i.e.,
H
D
(z) = H
0
(z + 1)
L
M
i =1
(z e
z
i
T
)
N
i =1
(z e
p
i
T
)
we note that the only dierences is that the ratio of
polynomial N
1
(z)/N
2
(z) is replaced by the polynomial
(z + 1)
L
.
t
For the classical types of lters (elliptic and inverse-Chebyshev
lters), it turns out that N1(z) and N
2
(z) are mirror image
polynomials with zeros on the negative real axis of the z plane
clustered near the Nyquist point and, consequently,
N
1
(z)
N
2
(z)
(z + 1)
L
t
For the classical types of lters (elliptic and inverse-Chebyshev
lters), it turns out that N1(z) and N
2
(z) are mirror image
polynomials with zeros on the negative real axis of the z plane
clustered near the Nyquist point and, consequently,
N
1
(z)
N
2
(z)
(z + 1)
L
t
In eect, at least, for classical lters, the discrete-time
transfer function obtained with the matched-z method is an
approximation of that obtained with the modied invariant
Advantages
t
Simple to apply the design can be done with a calculator.
Advantages
t
t
Works moderately well not only for lowpass and bandpass
lters but also for highpass and bandstop lters including
elliptic lters, i.e., no aliasing problems.
Advantages
t
t
Works moderately well not only for lowpass and bandpass
lters but also for highpass and bandstop lters including
elliptic lters, i.e., no aliasing problems.
t
The absence of N
2
(z) eliminates the stability problem
associated with the modied invariant impulse-response
method.
Disadvantages
t
The passband loss characteristic of the digital lter is
seriously distorted relative to that of the analog lter.
Disadvantages
t
t
A high sampling frequency is usually necessary to achieve
good results which can introduce certain problems.
A high sampling frequency corresponds to a reduced sampling
period and, therefore, the amount of processing that can be
done between samples is reduced.
Disadvantages
t
t
A high sampling frequency is usually necessary to achieve
good results which can introduce certain problems.
A high sampling frequency corresponds to a reduced sampling
period and, therefore, the amount of processing that can be
done between samples is reduced.
t
The multiplier constant needs to be adjusted at the end of
the design (see Chap. 11) for details).
Example
t
The matched-z transformation method was used to redesign the
elliptic lter considered earlier and the designed obtained is as
follows:
H
D
(z) = H
0
5
j=1
a
0j
+ a
1j
z + z
2
b
0j
+ b
1j
z + z
2
where H
0
, a
ij
, and b
ij
are given in the table shown.
j a
0j
a
1j
b
0j
b
1j
1 1.0 1.153491 9.128252E-1 1.362371
2 3.248990E+1 1.955491E+1 5.611278E-2 7.751650E-1
3 1.331746E-2 3.971465E-1 5.611278E-2 7.751650E-1
H
0
= 3.847141E-4
Example Contd
t
Loss characteristics with respect to the baseband:
method; - - - - - matched-z transformation method.
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
0
10
20
30
40
50
60
L
o
s
s
,

d
B

, rad/s
(a)
Example Contd
t
Loss characteristics with respect to the passband:
method; - - - - - matched-z transformation method.
0 0.2 0.4 0.6 0.8 1.0
0.10
0.05
0
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
L
o
s
s
,

d
B

, rad/s
(b)
This slide concludes the presentation.
Thank you for your attention.

DSP Ch11 S1 5

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

DSP Ch11 S1 5

Caricato da

Copyright:

Formati disponibili

Chapter 11

Design of Recursive (IIR) Filters

Potrebbero piacerti anche