“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

Chapter 5
Design of IIR Filters
5.1 Introduction
IIR filter design primarily concentrates on the magnitude response of the filter and regards the phase response as
secondary. The most common design method for digital IIR filters is based on designing an analogue IIR filter and then
converting it to an equivalent digital filter.
There are many classes of analogue low-pass filter, such as the Butterworth, Chebyshev and Elliptic filters. The classes
differ in their nature of their magnitude and phase responses. The design of analogue filters other than low-pass is based
on frequency transformations, which produce an equivalent high-pass, band-pass, or band-stop filter from a prototype
low-pass filter of the same class. The analogue IIR filter is then converted into a similar digital filter using a relevant
transformation method. There are three main methods of transformation, the impulse invariant method, the backward
difference method, and the bilinear z-transform.

5.2 IIR Filter Basics

A recursive filter involves feedback. In other words, the output values are calculated using one or more of the previous
outputs, as well as inputs. In most cases a recursive filter has an impulse response which theoretically continues forever.
It is therefore referred to as an infinite impulse response (IIR) filter. Assuming the filter is causal, so that the impulse
response h[n] = 0 for n < 0, it follows that h[n] cannot be symmetrical in form. Therefore, an IIR filter cannot display
pure linear-phase characteristics like its adversary, the FIR filter.

The finite difference equation and transfer function of an IIR filter is described by Equation 3.3 and Equation 3.4
respectively. In general, the design of an IIR filter usually involves one or more strategically placed poles and zeros in
the z-plane, to approximate a desired frequency response. An analogue filter can always be described by a frequency-
domain transfer function of the general form, shown in Equation 5.1.

( s − z1 )( s − z 2 )( s − z3 ) ⋅ ⋅ ⋅
H ( s) = K (5.1)
( s − p1 )( s − p 2 )( s − p3 ) ⋅ ⋅ ⋅

Where s is the Laplace variable and K is a constant, or gain factor. The filter is characterised by its poles p1, p2, p3…,
and its zeros z1, z2, z3…, which can be plotted in the complex s-plane. The frequency response of the filter H(ω), can be
obtained by replacing s = j ω into Equation 5.1. The complete response of the filter is then generated by varying ω in
Equation 5.2 between 0 and ∞.

( jω − z1 )( jω − z 2 )( jω − z 3 ) ⋅ ⋅ ⋅
H (ω ) = K (5.2)
( jω − p1 )( jω − p 2 )( jω − p3 ) ⋅ ⋅ ⋅

5.3 Analogue Low-pass Filters

There are several classes of analogue low-pass filter, three of which are the Butterworth, Chebyshev and Elliptic. These
filters differ in the position of their and in the nature of their magnitude and phase responses. Their frequency responses
are illustrated in Figure 5.1 below. The Butterworth filter is said to be monotonic at all frequencies (i.e. no local
maxima or minima), the Chebyshev is monotonic in the stop-band and equiripple in the pass-band, and an Elliptic filter
is equiripple in all bands.

University of Newcastle upon Tyne Page 5.1

“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

Pass-band ripple = 0.5 dB

Filter order n =3
1 ωo = 0.5

Chebyshev filter
0.707

H(ω)
Ideal low-pass filter Butterworth filter

Elliptic filter

0
ωo

Frequency ω

Figure 5.1: Typical frequency responses of various analogue low-pass filters.

5.4 The Bilinear z-transform

One of the most effective and widely used techniques for converting an analogue filter into a digital equivalent is by
means of the bilinear z-transform. In this method, we replace s in equation (5.1) by the bilinear z-transform:

z −1
F ( z) = (5.3)
z +1
to give the following function of z:

 z − 1    z − 1    z − 1  
 z + 1  − z 1   z + 1  − z 2   z + 1  − z 3  ⋅ ⋅ ⋅
       
H (z ) = K   (5.4)
 z − 1    z − 1    z − 1  
 z + 1  − p1   z + 1  − p 2   z + 1  − p 3  ⋅ ⋅ ⋅
        

The frequency response of this z-transfer function is obtained by substituting z = e jΩ in Equation (5.4). The result of
doing this is most easily seen by making this substitution first in the function F(z) in equation (5.3):

e jΩ − 1 e jΩ 2 − e − jΩ 2 j 2 sin (Ω 2 ) Ω
F (Ω) = = jΩ 2 = = j tan   (5.7)
jΩ
e +1 e +e − jΩ 2
2 cos(Ω 2 ) 2
Substituting this into equation (5.4) we obtain:

H (Ω ) = K
[ j tan (Ω 2) − z1 ][ j tan (Ω 2) − z 2 ][ j tan (Ω 2) − z 3 ]⋅ ⋅ ⋅ (5.8)
[ j tan (Ω 2) − p1 ][ j tan (Ω 2) − p 2 ][ j tan (Ω 2) − p3 ]⋅ ⋅ ⋅
The frequency response of a desirable analogue filter was given by Equation (5.2). The function H(Ω) in equation (5.8)
takes all values of the frequency response of the analogue filter, but compressed into the range 0 ≤ Ω ≤ π. Note that the
compression of the frequency scale is non-linear. The shape of the tan function, as depicted in Figure 5.2, means that
the “warping” effect is small near Ω = 0, but increases greatly towards Ω = π/2.

University of Newcastle upon Tyne Page 5.2

“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

Function ω=tan(Ω/2)
8

ω -2

-4

-6

-8
0 π/2 π 3π/2 2π
(Ω/2)

Figure 5.2: The "warping" effect of the tan function.

There are several advantages in using the bilinear z- transform. Firstly, the equiripple amplitude properties of the filters
are preserved when the frequency axis is compressed. Secondly, there is no aliasing of the original analogue frequency
response. As a result, the response of a low-pass filter falls to zero at Ω = π. This is an extremely important feature in
many practical applications. The principle of the bilinear z-transform, by making the substitution of Equation 5.6, is
illustrated in Figure 5.3 below. It shows that the imaginary axis in the s-plane (s = j ω) maps into the unit circle of the z-
plane.

Imag Imag

Real Real

s-plane z-plane

Figure 5.3: Illustration of s-plane to z-plane mapping using the bilinear z-transform.

The substitution maps the left-hand side of the s-plane to the inside of the unit circle in the z-plane. This ensures that the
Nyquist stability criterion is obeyed and therefore filter stability is preserved. To overcome the frequency “warping”
introduced by the bilinear z-transform, it is common practice to pre-warp the specification of the analogue filter, so that
after warping they will be located at the desired frequencies. For example, suppose we wish to design a digital low-pass
filter with a cut-off frequency Ωc. We first transform this frequency to the analogue-domain cut-off frequency ωac, using
the pre-warping relationship of Equation (5.9).

Ω  2
ω ac = k ⋅ tan c  k =1 or (5.9)
 2  T

We then proceed to design the analogue filter using the corresponding cut-off frequency, obtained from Equation (5.9).
After the analogue filter has been transformed using the bilinear z-transform, the resulting digital filter will have its cut-
off frequency in the correct place. Since pre-warping is performed in the beginning of the design procedure, and bilinear
transformation is performed at the end, the value of k is immaterial.

University of Newcastle upon Tyne Page 5.3

“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

5.5 Frequency Transformations

The design of analogue filters other than low-pass is usually achieved by designing a low-pass filter of the desired class
e.g. Butterworth, Chebyshev, or Elliptic, and then transforming the resulting filter to get the desired frequency response
e.g. high-pass, band-pass, or band-stop. This is accomplished by substituting the frequency-domain transfer function
H(s) with one of the relevant frequency transformations listed below. Where ω2 and ω1 are the band-edge frequencies of
the desired filter and are also positive parameters satisfying ω2 > ω1.

s
Low-pass to Low-pass transformation: s⇒ (5.10)
ω ac

ω ac
Low-pass to High-pass transformation s⇒ (5.11)
s

s 2 + ω 1ω 2
Low-pass to Band-pass transformation s⇒ (5.12)
(ω 2 − ω 1 ) s

(ω 2 − ω 1 ) s
Low-pass to Band-stop transformation s⇒ (5.13)
s 2 + ω 1ω 2

5.6 Summary of IIR Filter Design Using the Bilinear z-transform

• Use the digital filter specification to determine a suitable normalised frequency-domain transfer function H(s).

• Determine the cut-off frequency of the digital filter Ωc.

• Obtain the equivalent analogue filter cut-off frequency ωac using the pre-warping function of Equation 5.9.

• Denormalise the analogue filter by frequency scaling H(s), with one of the appropriate frequency transformations
e.g. s ⇒ s/ ωac etc.

• Apply the bilinear z-transform to obtain the digital filter transfer function H(z) by replacing s with (z - 1)/(z + 1).

5.6.1 Example
Design a digital filter equivalent of a 2nd order Butterworth low-pass filter with a cut-off frequency fc = 100 Hz and a
sampling frequency fs = 1000 samples/sec. Derive the finite difference equation and draw the realisation structure of the
filter. Given that the analogue prototype of the frequency-domain transfer function H(s) for a Butterworth filter is:

1
H (s) = 2
s + 2 ⋅ s +1
The normalised cut-off frequency of the digital filter is given by the following equation:

2πf c 2π 100
Ωc = = = 0.628
fs 1000

Now determine the equivalent analogue filter cut-off frequency ωac, using the pre-warping function of Equation 5.9.
The value of K is immaterial so let K = 1.

Ω   0.628 
ω ac = K ⋅ tan c  = 1 ⋅ tan 
 2   2 

ω ac = 0.325 rads / sec

University of Newcastle upon Tyne Page 5.4

“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

Now denormalise the frequency-domain transfer function H(s) of the Butterworth filter, with the corresponding low-
pass to low-pass frequency transformation of Equation 5.10. Hence the transfer function of the Butterworth filter
becomes:

1
H ( s) = 2
 s   s 
 0.325  + 2 ⋅  0.325  + 1
   
Next, convert the analogue filter into an equivalent digital filter by applying the bilinear z-transform. This is achieved
by making a substitution for s in the transfer function.

z − 1 1 − z −1
s= ≡
z + 1 1 + z −1

1
H ( z) = 2
1 1 − z 
−1
2 1 − z −1 
⋅ −1 
+ ⋅ −1 
+1
0.325 2 1 + z  0.325 1 + z 

Y ( z ) 0.067 + 0.135 z −1 + 0.067 z −2

H ( z) = =
X ( z) 1 − 1.1429 z −1 + 0.4127 z −2
The finite difference equation of the filter is found by inverting the transfer function.

y (n ) = 1.1429 y (n − 1) − 0.4127 y (n − 2 ) + 0.067 x(n ) + 0.135 x(n − 1) + 0.067 x(n − 2 )

The transfer equation H(z) above, resembles the direct structure of Equation 3.13, from Chapter 3. So the realisation of
this filter follows the same format as Figure 3.9, where the corresponding coefficients a1, a2, b0, b1, and b2 are taken
from the Equation above.
0.067
x[n] + + y[n]

z-1

1.1429 0.135
+ +

z-1

-0.4127 0.067

Figure 5.4: Direct realisation for a 2nd order Butterworth equivalent filter.

5.7 Z-Plane Poles and Zeros

A very useful representation of a z-transform is obtained by plotting its poles and zeros in the complex plane. It is quite
easy to visualise the frequency response from such a diagram and it also gives a good indication of the degree of
stability of a system.
The frequency-selective properties of first and second-order systems can be controlled by the appropriate choice of the
pole-zero locations. Poles are particularly effective in this respect because when they are placed close to the unit circle
they produce sharp, well-defined peaks in the frequency response. Usually an equal number of zeros are then placed at
the z-plane origin (0, 0) to ensure that the impulse response begins at n = 0. The frequency response of a first-order
system is defined by Equation 5.14 below; it has one real pole at the location z = α and one real zero at the origin z = 0.
The frequency response of a second-order system is also defined by Equation 5.15; it has two poles (either both real or a

University of Newcastle upon Tyne Page 5.5

“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

complex conjugate pair) at the locations z = r·exp(jθ) and z = r·exp(-jθ). In addition, it also has two zeros at the origin z
= 0.
(a )
4
(a ) r = 0 .9 , θ = 0
3

h [n ] 2

0
0 50 100 150 200 250 300 350 400 450 500

n
100

80

60
 H (Ω ) 
40

20

0
0 π

(b )
4

(b ) r = 0 .9 9 , θ = 2 5
2

h [n ] 0

-2

-4
0 50 100 150 200 250 300 350 400 450 500

n
150

100

 H (Ω ) 
50

0
0 π

(c )
1
(c) r = 0 .8 , θ = 1 10
0 .5

h [n ] 0

-0 .5

-1
0 50 100 150 200 250 300 350 400 450 500

n
3

2
 H (Ω ) 
1

0
0 π

(d )
4
(d ) r = 0 .9 , θ = 18 0
2

h [n ] 0

-2

-4
0 50 100 150 200 250 300 350 400 450 500

n
100

80

60
 H (Ω ) 
40

20

0
0 π

University of Newcastle upon Tyne Page 5.6

“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

Figure 5.5: The impulse and frequency responses of several second-order systems.

1
H 1 ( z) = (5.14)
1 − αz −1

1
H 2 (z) = (5.15)
[1 − r ⋅ exp( jθ ) z ][1 − r ⋅ exp(− jθ ) z −1 ]
−1

By using the trigonometric representation of the exponential function, Equation 5.15 can be re-written as Equation 5.16,
after multiplying out the denominator.

1
H 2 ( z) = (5.16)
[1 − 2r ⋅ cosθ ⋅ z −1 + r 2 ⋅ z −2 ]

By changing the parameters r and θ, the impulse response h[n] and frequency response magnitude H(Ω) vary. Some
typical results are illustrated in Figure 5.3 for various second-order systems. In illustration (a) of Figure 5.3, the values
of r = 0.9, θ = 0 show that this configuration is a low-pass system with a second-order pole on the real axis in the z-
plane. The choice of r = 0.9 gives a moderately selective frequency response. In illustration (b), r = 0.99 and θ = 25 deg.
The poles are now much closer to the unit circle, giving a very selective frequency-domain characteristic H(Ω). In
the time-domain, the impulse response is prolonged, with the frequency of oscillation corresponding to θ = 25 deg,
which relates to 14 Hz. Diagram (c) illustrates the results for r = 0.8, θ = 110 deg. This system is much less selective in
the frequency domain, so its impulse response is short. Finally, (d) has r = 0.9 and θ = 180 deg, producing a high-pass
counterpart of the low-pass system shown in (a), but with the frequency response centered at Ω = π.

5.8 Finite Word Length Effects in IIR Filters

In general IIR filters are much more difficult to analyse than FIR filters because of the feedback structure. However,
both types of filter suffer from the same problems and have the same sources of noise due to finite word length effects.
The extent of filter degradation depends on the length of the word and the type of arithmetic (fixed or floating point)
used to perform the filtering operation. A summary of the main four sources of noise and their corresponding effects on
IIR filter performance are summarised in Table 5.1.

Source of noise Affect on performance Reduction techniques

• Increase number of bits.
A/D conversion. Quantisation noise = q2/12.
• Use multirate techniques.
• Use double word length for
Causes low level limit cycles i.e. intermediate results.
oscillations at the filter output, or • Optimise filter structure to
Arithmetic round off.
output stuck at a nonzero value, include error spectral shaping.
even when there is no input. • Add a dither signal before
rounding.
• Use sufficient Nos. of bits in
Modifies position of the poles and fixed-point representation.
Coefficient quantisation. zeros, may cause instability and a • Optimise selection of filter
change in the frequency response. coefficients.
• Use floating-point arithmetic.
• Scale filter coefficients (at cost
of reduced SNR).
Arithmetic overflow. Incorrect output signal. • Detect and use “maximum”
rather than “overflowed” value.
• Use floating-point arithmetic.
Table 5.1: Finite word length effects in IIR filters.

5.8.1 Arithmetic Round-Off

Arithmetic round off, can cause low-level limit cycles to occur in IIR filters. These can cause oscillations of the filter
output, or the output to remain stuck at a non-zero value, even when there is no input. For example, consider the

University of Newcastle upon Tyne Page 5.7

“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

following output of a 1st order IIR filter, as shown in Figure 5.6, with a 4-bit data and register length. Notice how the
output oscillates between [-2, 2].

6
y(n) = x(n) –0.75y(n-1)

0
2 4 6 8

-2

-4

-6

Figure 5.6: Low-level limit cycles caused by arithmetic round off.

Low-level limit cycles, as illustrated by Figure 5.6, can be reduced by using longer registers or by adding a dither signal
before rounding. In addition, arithmetic round off can also be reduced by utilising feedback and feedforward paths in
the 2nd order section, often known as error spectral shaping (ESS).

5.8.2 Coefficient Quantisation

If an error is introduced during coefficient quantisation, it can cause the poles and zeros to deviate from their expected
positions and changes the desired frequency response. If a pole position is moved outside of the unit circle then this will
cause instability in the filter. Now let us examine the effects of finite word lengths on the position of the pole-zero
placement in the z-domain of the unit circle.

5.8.2.1 First-Order System

Consider a first-order system with a single pole at position z = b and a zero at the origin, as depicted in Figure 5.7
below.
Imag

Real
b δ1

Figure 5.7: A single pole at z = b in the z domain, and a zero at the origin.

The z-transfer function of this first-order filter is given by the equation below:

1
H ( z) =
(1 − bz −1 )

The change δ1 in b that would cause the pole to lie at z = 1, is defined below:

1 − (b + δ 1 ) z −1 = 0

1 − (b + δ 1 ) = 0

δ1 = 1− b

University of Newcastle upon Tyne Page 5.8

“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

As an example, let us assume that the position of the pole was located at b = 0.95. Then, from the equation above, δ1 =
0.05. Now let us assume that the specification of the filter coefficient is not permitted to exceed 1% of the value of δ1.
Therefore the precision of the filter coefficient has to be accurate to within 0.0005. The minimum number of bits that is
required to meet this specification, after rounding, is given below:

 1  log10 (2000)
x = log 2   = log 2 (2000 ) = = 10.966 ≈ 11
 0.0005  log10 (2)

Furthermore, an additional bit has to be added to x for the mantissa (or sign) of the filter coefficient, so the total number
of bits required to meet this specification is 12.

5.8.2.2 Double Pole at z=b

Now let us consider a second-order system with a double pole at position z = b and two zeros at the origin. The z-
transfer function of this second-order filter is given by this equation:

1 1 1
H ( z) = −1 2
= −1 2 −2
=
(1 − bz ) 1 − 2bz + b z 1 + a1 z + a 2 z − 2
−1

The δ2 change in coefficient a1 that would cause one of the two poles to lie at z = 1, is defined by equation

1 + (a1 + δ 2 ) z −1 + a2 z −2 = 0 , evaluated at z = 1 :

1 + (a1 + δ 2 ) + a2 = 0

δ 2 = −(1 + a1 + a )2 = −(1 − 2b + b 2 ) = −(1 − b) 2

Using the same value for b = 0.95, then δ2 can be evaluated:

δ 2 = −(0.05) 2 = −0.0025
with a corresponding coefficient wordlength requirement of:

 1  log10 (40000)
x = log 2   = log 2 (4000 ) = = 15.287 ≈ 16 (+1 for the sign bit)
 0.000025  log10 ( 2)

We can therefore conclude that fewer bits are required by implementing the filter as a cascade of two first-order
sections rather than a single second-order section. It is fairly easy to generalise this result to higher orders and state that
implementing a digital filter as a cascade of first or second-order sections always results in shorter coefficient
wordlength requirements than if it were implemented as a single high order section.

5.8.2.3 Second-Order System

Now let us examine the case of a second-order complex conjugate pole pair with a double zero at the origin, as
illustrated in Figure 5.8. The z-transfer function of this second-order filter is given by the equation below:

1 1
H ( z) = −1 −2
= where r = a 20.5 and θ = cos −1 (− a1 2r )
1 + a1 z + a 2 z [1 − 2r ⋅ cosθ ⋅ z −1 + r 2 ⋅ z − 2 ]

For stability the poles must lie within the unit circle, satisfying the conditions:

0 ≤ a 2 < 1 and a1 ≤ 1 + a2 (derivation not given here, and is not required for this course)

As an example, let us consider the second-order Butterworth filter designed in section 5.6.1. The coefficient values
turned out to be a1= -1.1429 and a2 = 0.4127. Calculate δ1 and δ2 (corresponding to the two conditions above) that
would put the poles on the unit circle.

0 ≤ 0.4127 + δ1 < 1 resulting in the condition δ 1 ≤ 0.4127

University of Newcastle upon Tyne Page 5.9

“EEE305”, “EEE801 Part A”: Digital Signal Processing Chapter 5: Design of IIR Filters

1.1429 + δ 2 ≤ 1 + 0.4127 resulting in the condition δ 2 ≤ 0.2698

Imag

2
π Real

Figure 5.8: A complex conjugate pole pair and a double zero at the origin.

Now let us assume that the specification of the filter coefficient is not permitted to exceed 1% of the lowest value of δ
(δ2 = 0.2698). Therefore the precision of the filter coefficient has to be accurate to within 0.002698. The minimum
number of bits that is required to meet this specification, after rounding up, is given below:

1.1429 2
⇒ = 741.2898 = 2 x
0.002698 0.002698
log10 (741.2898)
x= = 9.5339 = 10 (rounded up)
log10 (2)

Furthermore, an additional bit has to be added to x for the mantissa (or sign) of the filter coefficient, so the total number
of bits required to meet this specification is 11.

University of Newcastle upon Tyne Page 5.10