Lecture 6: March 26, 2007

Topics:
1. IIR Filter design: Impulse-Invariant Method
2. IIR Filter design: The Matched Z-Transform
3. IIR Filter design: Bilinear Transformation
Method
4. Frequency Transformations

1
Lecture 6: March 26, 2007
Topic:
1. IIR Filter design: Impulse-Invariant Method

• basic principle: sampling of impulse response of an

analogue filter,
• mapping: HA(p) -> H(z),
• resulting filter implementation as a parallel bank of two-pole
filter,
• aliasing effect following from sampling process and its
impact,
• p -> z mapping and its impact, 2
Lecture 6: March 26, 2007
Topic:
2. IIR Filter design: The Matched Z-Transform

• basic principle: direct mapping the poles and zeros of

HA(p) (p-plane) into poles and zeros of H (z) (z-plane),
• mapping: HA(p) -> H(z),
• method limitations.

3
Lecture 6: March 26, 2007
Topic:
3. IIR Filter design: Bilinear Transformation
Method
• basic principle: application of the trapezoidal formula for
numerical integration of differential equation,
• mapping: HA(p) -> H(z),
• summary on digital filter design.

4
Lecture 6: March 26, 2007
Topic:
4. Frequency Transformations

• frequency transformations in analogue domain,

• frequency transformations in digital domain.

5
5.3. Impulse-Invariant Method
(Impulse Invariant Transformation)
Objective: to design an IIR filter having an impulse
response h(n) as the sampled version of the
impulse response of the analogue filter hA (t ):

hA (t )  hA (nT )  h(nT )  h( n) n  0,1,2,...

where T is the sampling interval.

In consequence of this result, the frequency response of

the digital filter is an aliased version of the frequency
response of the corresponding analogue filter.
6
Let the transfer function of the analogue filter be given:

L  y (t )  B ( p ) k
b p k

H A ( p)    k 0
L  x(t )  A( p ) N

 k
a
k 0
p k

7
Let us assumed that the order M of the numerator is less
that the order N of the denominator and that all poles
of H A ( p ) are simple. If the poles of H A ( p ) are not simple,
the discussion in this section can be appropriately
modified. Then, we rewrite the transfer function of the
analogue filter in its partial expansion, as follows
M

k
b p k
N
ck
H A ( p)  k 0

N
p  dk
 k
a
k 0
p k k 1

where ( d k ) is the location of the k-th pole and

ck  H A ( p )( p  d k ) p  d 8
k
The impulse response of the analogue filter hA (t ) :
hA (t )  L 1  H A ( p )  

 N
c  N
1  ck 
 L 
1 k
  L   
 k 1 p  d k  k 1  p  dk 
N
  ck e  dk t t  nT
k 1

The impulse response of the digital filter h(nT):

N
h(n)  h(nT )  hA (nT )   ck e  dk nT
k 1

n  0,1,2,3, ,  9
Transfer function of the digital filter:
H ( z )  Z h( n)  

  h( n) z  n 
n 0

 N
  z  n  ck e  dk nT 
n 0 k 1

N 
  ck  e  
 dkT n
1
z 
k 1 n 0

N
1
  ck
k 1 1  e  pkT z 1 10
Transfer function of Transfer function of
the analogue filter: the digital filter:
N
ck N
ck
H A ( p)   H ( z)    d k T 1
k 1 p  dk k 1 1  e z

Comparing HA(p) and H(z) it can be seen that H(z) can

be obtained from HA(p) by using the mapping relation:
ck ck

p  dk 1  e  dkT z 1
11
 Realization: parallel bank of two-pole filters
(second-order section filters):
N
ck
H ( z)    d k T 1
k 1 1  e z

Intention: to implement filters with real-valued

coefficients instead of complex-valued coefficients

12
With the previous given expressions for the transfer
function H(z), the IIR filter is easily realized as a
parallel bank of single-pole filters:
N
ck
H ( z)    dkT
k 1 1  e z 1

13
With the previous given expressions for the transfer
function H(z), the IIR filter is easily realized as a
parallel bank of single-pole filters.
If some of poles are complex-valued, they may be
paired together and combined to form two-pole filter
sections with real-valued coefficients:
ck ck
 1  d k T
p  dk 1  z e
ck ck

p  d k 1  z 1e  dkT
ck ck ck ck
  1  d k T

p  dk p  dk 1  z e 1  z 1e  dkT 14
When: d k   k  jk and ck  g k  jhk d k , ck  C

ck ck 2 g k p  2  k g k  k hk 
 
p  dk p  dk p 2  2 k p   k2  k2 

ck ck
1  d k T
 
1 z e 
1  z 1e kd T


2 gk  z e 
1  k T
2 g k cos(k T )  2hk sin(k T ) 
1  2 z 1e  kT cos(kT )  z 2e2 kT
15
With the previous given expressions for the transfer
function H(z), the IIR filter is easily realized as a
parallel bank of single-pole filters.
If some of poles are complex-valued, they may be
paired together and combined to form two-pole filter
sections with real-valued coefficients.
In addition, two factors containing real-valued poles
may be combined to form two-pole filters with real-
valued coefficients.

16
With the previous given expressions for the transfer
function H(z), the IIR filter is easily realized as a
parallel bank of single-pole filters.
If some of poles are complex-valued, they may be
paired together and combined to form two-pole filter
sections with real-valued coefficients.
In addition, two factors containing real-valued poles
may be combined to form two-pole filters with real-
valued coefficients.
Consequently, the resulting filter may be realized as
a parallel bank of two-pole filters with real-valued
coefficients.
17
 Aliasing Effect:
When a continuous time signal hA (t ) with spectrum H A ()
is sampled with sampling frequency  S  2 FS , the
spectrum of the sampled signal is given by the following
expressions:
H ( j)  FT  hA  nT   FT  hA  n / FS 

1 
H ( j)   H A  j  jk  S 
T k 
1 
H ( f )   H A  f  kFS  where f   / 2
T k 

Aliasing occurs if the sampling frequency FS is less then

twice the highest frequency contained in hA (t ) . 18
The next figure depicting the frequency responses of the
low-pass analogue filter and the corresponding digital filter
illustrates the aliasing effect following from hA (t ) sampling
FS
with the sampling frequency .

19
 Magnitude response property illustration

H A  f  2 FS  Hf 

H A  f  FS  H A  f  FS 

ALIASING ! HA  f  H A  f  2 FS 

1 
H ( f )   H A  f  kFS 
T k 

20
2 FS  FS FS 2 FS
 Aliasing effect impact:

A. The digital filter will possess (approximately) the

frequency response characteristics of the corresponding
analogue filter if the sampling interval T is selected
sufficiently small to avoid completely or minimize the
effects of aliasing.

B. The impulse invariance method is inappropriate for

designing high-pass filters or stop-band filters due to
spectrum aliasing that results from the sampling process.

21
 p - > z mapping:
To investigate the mapping between the p-plane and the
z-plane implied by the sampling process, we rely on a
generalization of the expression relating Z-transform of
h(n) to the Laplace transform of hA(nT). This
relationship is given by
1   2 k 
H ( z ) z e pT  H ( p )   H A  p  j 
T k   T 

H ( z)   h (
n 
n ) z n


H ( z ) z e pT   h (
n 
n ) e  pTn
z  e pT
the p -> z mapping 22
Note that when,p  j
1   2 k 
H ( z ) z e pT   HA  p  j 
T k   T 
reduces to

1
H (e )   H A  j  jk  S 
j

T k 

23
The general characteristic of the p->z mapping defined
as
z  e pT
can be obtained by substitutions:

p    j z  re j

With these substitutions we can get

z  re j  e(  j )T  e T e jT

r  e T   T    /T
24
 Consequences:

a)   0  0  r  1   0r  1   0r  1
Then, the left-half of p-plane is mapped inside the
unite circle in z-plane and right-half of p-plane is
mapped into points that fall outside the unit circle
in z-plane. This is one of the desirable properties
of a good p -> z mapping.

b) j -axis is mapped into the unit circle in z-plane

as indicates above.

25
 Comments on j -axis mapping:   T

a) The mapping of j -axis into the unit circle is not

one-to-one.

b)  / T     / T       

c) Mapping of the adjacent strips - frequency interval:

 / T    3 / T       

d) General case:
(2k  1) / T    (2k  1) / T       
26
 Conclusions:

A. The mapping from the analogue frequency domain to

the digital frequency domain is many-to-one. It reflects
simply the effects of aliasing due to sampling.

B. It follows for the frequency responses of analogue

filter and equivalent digital filter obtained by impulse
invariant transformation that the analogue filter must be
bandlimited to the range  / T     / T This
generally requires that the analogue filter has to be
suitably bandlimited prior to transformation.

27
 Mapping: z  e pT
j  j Im[ p ]
3 /T The mapping of strips of the width 2
Im[ z ]
p-plane
 /T

0 1
  Re[ p] 0 Re[ z ]

 /T

z-plane unit circle

3 /T 28
 Summary on digital filter design:
1. Digital filter specification:  P , S ,  1 and  2

2. Transformation of requirements to the digital filter

to the analogue filter:
   /T P   p S   S  1 and  2
N
ck
3. The analogue filter design: H A ( p)  
k 1 p  dk

4. The analogue filter conversion to the digital filter:

N
ck
H ( z)  T   d k T 1
k 1 1  e z 29
 Comment on scaling factor application (T):

The frequency response of the filter obtained by impulse

invariant transformation is given by

1
H (e j )   H A    k  S 
T k 
Under condition that

H A    k  S  ~ 0 for k  0

we can obtain:
1
H (e ) ~ H A   
j

T 30
If it is desired to get a digital filter with the same gain
as the analogue filter possesses, it is necessary to
transform the expression for H(z) originally given by

N
ck
H ( z)    d k T 1
k 1 1  e z
in the form
N
ck
H ( z)  T   d k T 1
k 1 1  e z
Then:

H (e j ) ~ H A   
31
 5.4. The Matched Z-Transform

Basic principle:
Mapping the poles and zeros of H A ( p ) (from the p-plane)
directly into poles and zeros of H ( z ) (in the z-plane).

32
 Transfer function of Transfer function of
the analogue filter: the digital filter:
M

 
M M

b p k
k
 p  z  k
1  e zkT z 1
H A ( p)  k 0
 k 1 H ( z)  k 1
N

 1  e 
N N

a  p  p 
pk T 1
k p k
k
z
k 0 k 1 k 1

T is sampling interval.
Comparing HA(p) and H(z) it can be seen that H(z) is
obtained from HA(p) by using the mapping relation:

p  a  1  e aT z 1

The matched Z-transformation 33

 Comments:

To preserve the frequency response characteristics of an

analogue filter, the sampling interval in the matched Z-
transformation must be selected properly to yield the pole
and zero locations at the equivalent positions in the z-
plane. Thus aliasing must be avoided by selecting T
sufficiently small.

34
 Comments:

To preserve the frequency response characteristics of an

analogue filter, the sampling interval in the matched Z-
transformation must be selected properly to yield the pole
and zero locations at the equivalent positions in the z-
plane. Thus aliasing must be avoided by selecting T
sufficiently small.

35
Although the matched Z-transformation is easy to apply,
there are many cases when it is not a suitable mapping:
A. If the analogue system has zeros with center
frequencies greater that half sampling frequency, their z-
plane positions will be greatly aliased.
B. Another case where the matched Z-transformation is
unsuitable is where the continuous transfer function is an
all-pole system. Then the digital transfer function is an
all-pole system that, in any cases, does not adequately
represent the desired continuous system.
In general, use of impulse-invariant transformation is
to be preferred over the matched Z-transformation.
36
 5.5. Bilinear Transformation
Method
The IIR filter design techniques described in the previous
sections have severe limitations in that they are
appropriate only for low-pass filter design and limited
class of band-pass filter design.
In this section we describe a mapping from the p-plane to
the z-plane, called the bilinear transformation, that
overcomes the limitation of the other three design
methods described previously.

Basic principle: application of the trapezoidal formula

for numerical integration of differential equation.
37
Let us consider an analog linear filter with transfer
function
b
H A ( p) 
pa

This system is also characterized by the differential

equation dy (t )
 ay (t )  bx(t )
dt

or by the differential equations:

dy (t )
  ay (t )  bx(t ) y '(t )   ay (t )  bx (t )
dt
38
Instead of substituting a finite difference for the
derivative, suppose that we integrate the derivative and
approximate the integral by the trapezoidal formula:
x2
1
 f ( x)dx   x2  x1  f ( x2 )  f ( x1 ) 
x1
2
Thus:
t nT

y (t )   y '( )d  y (t0 )  

nT T
y '( )d  y (nT  T )
t0

The approximation of the previous integral by trapezoidal

formula is
T
y (nT )   y '(nT )  y '(nT  T )   y (nT  T )
2 39
 Now the differential equation evaluated in t=nT:

y '(t )   ay (t )  bx(t ) y '(nT )  ay (nT )  bx(nT )

We use this expression to substitute for derivative and thus

obtain a difference equation for the equivalent discrete-
time system. Then:
T
y (nT )   y '(nT )  y '(nT  T )   y (nT  T )
2
y (nT ) 
T
  ay (nT )  bx(nT )  ay (nT  T )  bx(nT  T )   y (nT  T )
2
40
 The previous expression can be expressed in his form:

 aT   aT  bT
1   y (nT )  1   y (nT  T )   x(nT )  x(nT  T ) 
 2   2  2

The Z-transform of the previous equation is

 aT   aT  1 bT

1   Y ( z )  1   z Y ( z )   X ( z )  z 1
X ( z ) 
 2   2  2

Transfer function of the equivalent digital filter is

Y ( z)
H ( z) 
X ( z) 41
 bT
Y ( z) 2
1  z 1

H ( z)   
X ( z) aT  aT  1
1  1  z
2  2 
bT bT
2
1 z 
1

2 b
  
1  z   2 1  z  1  z 1   aT a  2 1  z 1 
aT 1 1
1 1

1  z  2 T 1  z 

b
H ( z) 
2 
1
1  z
a
T 1  z 1
 42
Transfer function of Transfer function of
the analogue filter: the digital filter:
b b
H A ( p)  H ( z) 
pa 2 1  z 
1

a
T 1  z 1

Comparing HA(p) and H(z) it can be seen that H(z) is
obtained from HA(p) by using the mapping relation:
2
p
2 1  z 1
z -> p: p p -> z: z  T
T 1  z 1 2
p
T 43
Investigation of the properties of the bilinear transformation:
2
p
z  re j
p    j zT
2
p
If p  j T
2
2 2
 j     2
j arctan
T
T
j T T
  e 2 j 2 arctan
z  re   T  e 2
2 2  j arctan
 j 2
     
2 e 2
T
T 

From this equation it can be found

T 2 
z  r  1   2arctan   tan
2 T 2 44
 T
  2arctan  z  1   0 z 1
2

Between these limits the angle of z varies from 0 to  .

Conclusion 1: j -axis unit circle

Conclusion 2: The entire range in  is mapped only

once into range      , the aliasing errors are
eliminated. However, the mapping is highly nonlinear.
We observe a frequency compression or frequency
warping, as it is usually called, due to these nonlinearity.

45


  2arctan 2T

T

46
If p    j we obtain for z:

2
2 2 2 
p    j       2
j arctan

2 / T 
 T  e
zT T  
2 2 2 
2 j arctan
p    j 2 e 2 / T 
T T     
T 

2
2 
      2

T 
z 
2
2 
      2

T 
47
Conclusion 3:
If   0 (left-half p-plane), we find z  1 (inside
unit circle)

Conclusion 4:
If   0 (right-half p-plane), we find z  1 (outside
unit circle)

48
 2 p
Mapping:z  T
2 p
T
j  j Im[ p ]
Im[ z ]

0 1
  Re[ p] 0 Re[ z ]

p-plane z-plane unit circle

49
 Summary on digital filter design:
1. Digital filter specification:  P , S ,  1 and  2

2. Transformation of requirements to the digital filter

to the analogue filter:
2 
  tan  P   p S   S  1 and  2
T 2
3. The analogue filter design: H A ( p )

4. The analogue filter conversion to the digital filter:

H ( z )  H A  p  p  2 1 z 1
T 1 z 1
50
6. Frequency Transformations

51
The treatment in the preceding sections was focused
primarily on the design of low-pass IIR filters. If we
wish to design a high-pass or a band-pass or a band-stop
filter, it is a simple matter to take a low-pass prototype
filter and perform a frequency transformation.
1. One possibility is to perform the frequency
transformation in the analogue domain and then to
convert the analogue filter into a corresponding digital
filter by a mapping of the p-plane into z-plane.
2. An alternative approach is first to convert the analogue
low-pass filter into a low-pass digital filter and then to
transform the low-pass digital filter into desired digital
filter by a digital transformation.
52
6.1. Frequency Transformations
in Analogue Domain

53
6.1.1. Low-Pass to Low-Pass Transformation
Suppose that we have a low-pass filter with cut off
frequency C and we wish to convert it to another
 '
low-pass filter with cut off frequency C . The
transformation that accomplish this is
C
p ' p
C
Thus we obtain a low-pass filter with system function
 C 
H L ( p)  H P  ' p 
 C 
where H P ( p ) is the system function of the prototype
low-pass filter with cut off frequency C . 54
6.1.2. Low-Pass to High-Pass Transformation
If we wish to convert a low-pass filter into a high-pass
filter with cut off frequency C , the desired
'

transformation is
C C' 1
p 
p p
C C'
The system function of the high-pass filter is
 C C' 
H H ( p)  H P  
 p 
where H P ( p ) is the system function of the prototype
low-pass filter with cut off frequency C . 55
6.1.3. Low-Pass to Band-Pass Transformation
The transformation for converting a low-pass analogue
filter with cut off frequency C into a band-pass filter,
having a lower cut off frequency  L and an upper cut
off frequency U , may be accomplished by means of
transformation
p 2   L U
p C
p  U   L 
Thus we obtain
 p 2   L U 
H B ( p )  H P  C 
 p  U   L  

where H P ( p ) is the system function of the prototype

low-pass filter with cut off frequency C . 56
6.1.4. Low-Pass to Stop-Band Transformation
The transformation for converting a low-pass analogue
filter with cut off frequency C into a stop-band filter,
having a lower cut off frequency  L and an upper cut
off frequency U , may be accomplished by means of
transformation
p  U   L 
p C 2
p   L U
It leads to
 p  U   L  
H SB ( p )  H P  C 2 
 p   
L U 

where H P ( p ) is the system function of the prototype

low-pass filter with cut off frequency C . 57
6.1.5. Comments on Frequency Transformations
in the Analogue Domain

Generally, the frequency transformations are non-linear

mappings and may appear to distort the frequency
response characteristics of low-pass filter. However, the
effects of the non-linearity on the frequency response are
minor, affecting primarily the frequency scale but
preserving the magnitude response of the filter. Thus, an
equiripple low-pass filter transformed into an equiripple
band-pass band-stop or high-pass filter.

58
6.1.5. Comments on Frequency Transformations
in the Analogue Domain

Generally, the frequency transformations are non-linear

mappings and may appear to distort the frequency
response characteristics of low-pass filter. However, the
effects of the non-linearity on the frequency response are
minor, affecting primarily the frequency scale but
preserving the magnitude response of the filter. Thus, an
equiripple low-pass filter transformed into an equiripple
band-pass band-stop or high-pass filter.

59
6.1.5. Comments on Frequency Transformations
in the Analogue Domain

Generally, the frequency transformations are non-linear

mappings and may appear to distort the frequency
response characteristics of low-pass filter. However, the
effects of the non-linearity on the frequency response are
minor, affecting primarily the frequency scale but
preserving the magnitude response of the filter. Thus, an
equiripple low-pass filter transformed into an equiripple
band-pass band-stop or high-pass filter.

60
6.2. Frequency Transformations
in Digital Domain

61
As in the analogue domain, frequency
transformations can be performed on a digital low-
pass filter to convert it to either a band-pass, band-
stop, or stop-band filter. The transformation involves
replacing the variable z by a rational functiong  z ,
1 1

which must satisfy the following properties:

1. The mapping z 1
 g   must map points inside
z 1

the unit circle in the z-plane into itself.

2. The unit circle must also be mapped into itself.

62
The second condition implies that for r=1,

 g e   g ( ) e
 j  j j arg g ( ) 
e

It is clear, that we must have

g ( )  1 for all 

That is, the mapping must be all-pass. Hence it is of the

form n
z 1
 k
g z   
1

k 1 1   k z 1

where  k  1 ensures that a stable filter is transformed

into another stable filter (i.e. satisfies condition 1).
63
6.2.1. Low-Pass to Low-Pass Transformation

1
1 z a
z 
1  az 1

sin  C   C'  / 2 

a
sin  C   C'  / 2 

 C' : cut off frequency of new filter

 z a 
1
H z   H 
1
1 
 1  az 
64
6.2.2. Low-Pass to High-Pass Transformation
1
1 z a
z 
1  az 1

cos  C   C'  / 2 

a
cos  C   C'  / 2 

 C' : cut off frequency of new filter

 z a 
1
H z   H  
1
1 
 1  az  65
 6.2.3. Low-Pass to Band-Pass Transformation
2 1
1 z  a z  a2
z  1
a2 z 2  a1 z 1  1

 L : lower cut off frequency U : upper cut off frequency

a1  2 K /  K  1 a2   K  1 /  K  1

cos U   L  / 2  U  L C
 K  cot tan
cos U   L  / 2  2 2

 z 2
 a z 1
 a2 
H z   H  
1
2
1
1 
 a2 z  a1 z  1  66
 6.2.4. Low-Pass to Stop-Band Transformation
2 1
1 z  a z  a2
z  1
a2 z 2  a1 z 1  1

 L : lower cut off frequency U : upper cut off frequency

a1  2 /  K  1 a2  1  K  / 1  K 

cos U   L  / 2  U  L C
 K  tan tan
cos U   L  / 2  2 2

 z 2
 a z 1
 a2 
H z   H 
1
2
1
1 
 a2 z  a1 z  1  67
 Comments on frequency transformations
in digital domain

Since, the frequency transformations may be performed

either in the analogue domain or in the digital domain,
a filter designer has a choice which approach will take.

68
We know that the impulse invariance method and
the mapping of derivatives are inappropriate to use
in designing high-pass and many band-pass filters
due to the aliasing problem.
Consequently, one would not employ analogue
frequency transformations followed by conversion of
the result into digital domain by use of these two
mappings.
Instead, it is much better to perform the mapping from
an analogue low-pass filter into a digital low-pass filter
by either of these mappings and then to perform the
frequency transformations in the digital domain. Thus
the problem of aliasing is avoided. 69
In the case of bilinear transformation, where aliasing is
not a problem, it does not matter the frequency
transformation is performed in the analogue domain or
in the digital domain. In fact, in this case only, the two
approaches result in identical digital filters.

70
THANK YOU
VERY MUCH
FOR YOUR
ATTENTION
71