ECE503: Digital Filter Design

ECE503: Digital Filter Design

Lecture 9

D. Richard Brown III

WPI

26-March-2012

WPI D. Richard Brown III 26-March-2012 1 / 33

 ECE503: Digital Filter Design

Lecture 9 Topics

Within the broad topic of digital filter design, we are going to focus on
the design of LTI IIR discrete-time filters from continuous-time filters.

h̄(t) h[n]

continuous-time discrete-time
H̄(s) H̄(Ω) H(ω) H(z)

1. Impulse invariance
2. Bilinear Transform
3. Specification of filter characteristics
4. Frequency transformation of lowpass IIR filters

WPI D. Richard Brown III 26-March-2012 2 / 33

 ECE503: Digital Filter Design

Motivation/Applications

Application 1: You are given an LTI continuous-time system (block

diagram, impulse response, transfer function, ...) and asked to design a
discrete-time system that emulates its behavior. How should you do this?
voltage
x(t) multiplier F̄ (s) controlled y(t)
oscillator

Application 2: You are given a set of discrete-time filter specifications, for

example “design a lowpass filter with less than one dB ripple in the
passband and at least 30 dB rejection in the stopband, where the passband
is |ω| < π/8 and the stopband is |ω| > π/4”, and asked to design a DT
filter to meet these specs. How should you do this? One approach:
1. Convert the DT filter specifications to CT filter specifications
2. Design a CT IIR filter to meet the specifications (using
long-established relatively simple methods for CT IIR filter design)
3. Transform the CT filter to a DT filter
WPI D. Richard Brown III 26-March-2012 3 / 33
 ECE503: Digital Filter Design

Why IIR Filters?

Chapter 10 of your textbook covers the design of DT FIR filters (we will
not be covering this). Our focus on DT IIR filter design is motivated by
the following facts:
1. All practical LTI CT systems are IIR.
2. To meet a given set of filter specifications, the order of an IIR filter
that meets the specifications is typically much lower than the order of
an FIR filter that meets the same specifications. Hence the
computational complexity of DT IIR filters is typically much less than
DT FIR filters that meet the same specifications.

WPI D. Richard Brown III 26-March-2012 4 / 33

 ECE503: Digital Filter Design

General Requirements for CT/DT Transformation

Requirement 1: Points on the imaginary axis in the s-domain must be

mapped to points on the unit circle in the z-domain.
Requirement 2: Stable causal CT transfer functions must be transformed
to stable causal DT transfer functions.
Recall that stable causal CT transfer functions have their poles in the
left-half plane.
s-domain z-domain

WPI D. Richard Brown III 26-March-2012 5 / 33

 ECE503: Digital Filter Design

Impulse Invariance

Given a causal LTI CT system impulse response h̄(t), we can write the DT
impulse response

h[n] = h̄(t)|t=nT = h̄(nT ) for n = 0, 1, . . .

Note: Your textbook defines impulse invariance slightly differently in

problem 9.6: (
1
h̄(0+ ) n = 0
h[n] = 2
h̄(nT ) n = 1, 2, . . .
The only difference is in h[0]. The boxed definition is more common.

Does this CT→DT transformation satisfy the general requirements?

WPI D. Richard Brown III 26-March-2012 6 / 33

 ECE503: Digital Filter Design

Impulse Invariance Example

1
Suppose you are given a causal LTI CT system with H̄(s) = s−a . The
inverse Laplace transform gives h̄(t) = e µ(t). Hence h[n] = e n µ[n] (or
at aT

h[n] = 21 δ[n] + eaT n µ[n − 1] using your textbook’s definition).

What is H(z)? Rewrite

n
h[n] = eaT n µ[n] = eaT µ[n] = αn µ[n]
Table lookup tells us
1
H(z) =
1 − eaT z −1
with ROC |z| > eaT .
Remarks:
◮ Suppose a is on on the imaginary axis. What is the magnitude of the
pole of H(z)?
◮ Suppose a is complex with negative real part. What can you say
about the magnitude of the pole of H(z)?
WPI D. Richard Brown III 26-March-2012 7 / 33
 ECE503: Digital Filter Design

Impulse Invariance Example

% CT system 1
% H(s) = 1/(s-a) CT
a = -0.5; h[n]
0.9 Matlab impulse response of H(z)
tfinal = 20;
ctnum = 1;
0.8
ctden = [1 -a];
ctsys = tf(ctnum,ctden);
[y,t] = impulse(ctsys,tfinal); 0.7

plot(t,y,’r’)
impulse response 0.6
% DT system
T = 0.5; 0.5
alpha = exp(a*T);
dtnum = [1 0]; 0.4
dtden = [1 -alpha];
dtsys = tf(dtnum,dtden,T); 0.3
[dty,dtt] = impulse(dtsys,tfinal);
hold on
0.2
n=0:length(dtt)-1;
plot(dtt,alpha.^n,’b+’);
0.1
stem(dtt,dty);
hold off
xlabel(’time’); 0
0 2 4 6 8 10 12 14 16 18 20
ylabel(’impulse response’); time
legend(’CT’,’h[n]’,’Matlab impulse response of H(z)’);

WPI D. Richard Brown III 26-March-2012 8 / 33

 ECE503: Digital Filter Design

Impulse invariance doesn’t imply invariance to all inputs

% step responses 5
% note these results are consistent with CT
% the final value theorems Matlab step response of H(z)
4.5
[ys,ts] = step(ctsys,tfinal);
[dtys,dtts] = step(dtsys,tfinal);
4
figure(2)
plot(ts,ys,’r’)
hold on 3.5

stem(dtts,dtys);
hold off 3

step response
xlabel(’time’);
ylabel(’step response’); 2.5
legend(’CT’,’Matlab step response of H(z)’);
2

1.5

0.5

0
0 2 4 6 8 10 12 14 16 18 20
time

In general, you can pick a particular input and match the response of the
DT system to the response of the CT system for that input. But the
response to other inputs will be different.
WPI D. Richard Brown III 26-March-2012 9 / 33
 ECE503: Digital Filter Design

Extending the Previous Impulse Invariance Example

Now suppose you have a causal LTI CT system with transfer function
N
X bk
H̄(s) =
s − ak
k=1

where each ak is distinct (no repeated roots). Using linearity, we can write
N
!
X
ak t
h̄(t) = bk e µ(t).
k=1
1
When we sample this at rate FT = T , we get the DT impulse response
N
! N
!
X X
h[n] = b k e ak T n µ[n] = ak T n
bk (e ) µ[n]
k=1 k=1
which has z-transform
N
X bk
H(z) =
1 − eak T z −1
k=1
with ROC extending outward from the largest magnitude pole.
WPI D. Richard Brown III 26-March-2012 10 / 33
 ECE503: Digital Filter Design

A Repeated Pole Example

Suppose you have a causal LTI CT system with transfer function
b1 b2
H̄(s) = + .
s − a1 (s − a1 )2
From a Laplace transform table, we have for Re(s) > −α
tk−1 −αt 1
e µ(t) ←→
(k − 1)! (s + α)k
Hence, using linearity, we can write
h̄(t) = b1 ea1 t µ(t) + b2 tea1 t µ(t).
1
When we sample this at rate FT = T , we get the DT impulse response
h[n] = b1 ea1 T n µ[n] + (b2 T )nea1 T n µ[n]
which has z-transform
b1 (b2 T )ea1 T z −1
H(z) = +
1 − ea1 T z −1 (1 − ea1 T z −1 )2
with ROC |z| > |ea1 T |.
WPI D. Richard Brown III 26-March-2012 11 / 33
 ECE503: Digital Filter Design

A Repeated Pole Example

% CT system 1
% H(s) = b1/(s-a1)+b2/(s-a1)^2 CT
a1 = -0.5; h[n]
Matlab impulse response of H(z)
b1 = 1;
b2 = -2; 0.5
tfinal = 20;
ctnum = b1*[1 -a1]+[0 b2];
ctden = [1 -2*a1 a1^2];
ctsys = tf(ctnum,ctden);

impulse response
0
[y,t] = impulse(ctsys,tfinal);
plot(t,y,’r’)

% DT system
−0.5
T = 0.5;
alpha = exp(a*T);
dtnum = b1*[1 -alpha 0]+b2*T*alpha*[0 1 0];
dtden = [1 -2*alpha alpha^2];
−1
dtsys = tf(dtnum,dtden,T);
[dty,dtt] = impulse(dtsys,tfinal);
hold on
n=0:length(dtt)-1;
plot(dtt,b1*alpha.^n+b2*T*n.*alpha.^n,’b+’); −1.50 2 4 6 8 10 12 14 16 18 20
stem(dtt,dty); time
hold off
xlabel(’time’);
ylabel(’impulse response’);
legend(’CT’,’h[n]’,’Matlab impulse response of H(z)’);

WPI D. Richard Brown III 26-March-2012 12 / 33

 ECE503: Digital Filter Design

Impulse Invariance: Frequency Response

Recall that h[n] is related to h(t) through ideal sampling. What is the
relationship between H(ω) and H̄(Ω)?
We’ve covered the relationship between the CTFT and the DTFT a few times:
∞  
1 X ω − k2π
H(ω) = H̄
T T
k=−∞

Remarks:
◮ If H̄(Ω) is bandlimited such that H̄(Ω) ≈ 0 for all Ω ≥ Tπ , then there will be
negligible overlap in the sum of shifted spectra and
1 ω
H(ω) ≈ H̄ for |ω| ≤ π
T T
Hence, the DT system designed via impulse invariance accurately emulates
the CT system’s frequency response (and can be scaled if desired).
◮ If H̄(Ω) is not bandlimited, e.g. H̄(Ω) is a notch filter, there will be aliasing
and the resulting DT system’s frequency response is not likely to be an
accurate emulation of the CT system’s frequency response.
WPI D. Richard Brown III 26-March-2012 13 / 33
 ECE503: Digital Filter Design

Impulse-Invariant Lowpass Butterworth Filter Design Ex.

|H(ω)|
1
1 − δ1

δ2
ω
ωp ωs π

We start with the desired specifications of the DT filter. For this example,
we will use ωp = 0.2π, ωs = 0.3π, 1 − δ1 = 0.89125, and δ2 = 0.17783.
WPI D. Richard Brown III 26-March-2012 14 / 33
 ECE503: Digital Filter Design

Impulse-Invariant Lowpass Butterworth Filter Design Ex.

We will use the Butterworth filter approach in this example. Some facts
about Butterworth CT lowpass filters (from Appendix A.2):
1
|H̄(Ω)|2 =  2N
Ω
1+ Ωc

where Ωc is the cutoff frequency (radians/second) and N is the filter order.

Note:
◮ |H̄(Ω = 0)|2 = 1.
◮ |H̄(Ω)|2 is monotonically decreasing in Ω.

We will use the following steps to design our DT filter:

1. Convert DT filter specs to CT filter specs.
2. Determine the required filter order N and CT cutoff frequency Ωc .
3. Determine H̄(s).
4. Determine H(z) via impulse invariance.
WPI D. Richard Brown III 26-March-2012 15 / 33
 ECE503: Digital Filter Design

Impulse-Invariant Lowpass Butterworth Filter Design Ex.

[⋆ ⋆ ⋆ details on board ⋆ ⋆ ⋆]

Main result from step 3:

0.12093
H̄(s) =
(s2 + 0.3640s + 0.4945)(s2 + 0.9945s + 0.4945)(s2 + 1.3585s + 0.4945)

Main result from step 4:

0.2871 − 0.4466z −1
H(z) =
1 − 1.2971z −1 + 0.6949z −2
−2.1428 + 1.1455z −1
+
1 − 1.0691z −1 + 0.3699z −2
1.8577 − 0.6303z −1
+
1 − 0.9972z −1 + 0.2570z −2

WPI D. Richard Brown III 26-March-2012 16 / 33

 ECE503: Digital Filter Design

Impulse-Invariant Lowpass Butterworth Filter Design Ex.

10
CT Butterworth LPF
DT Butterworth LPF via impulse invariance
0

−10

−20
magnitude response

−30

−40

−50

−60

−70

−80
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized frequency (times π)

WPI D. Richard Brown III 26-March-2012 17 / 33

 ECE503: Digital Filter Design

Impulse Invariance: Final Remarks

1. Main idea is to preserve characteristics of the CT impulse response in

the DT impulse response, e.g. fast settling time, etc.
2. Frequency response of DT system is only a good replica of CT system
if the CT system is bandlimited (negligible aliasing).
3. Idea can be extended to other types of waveform invariance, e.g. step
invariance.
4. The impulse invariance technique maps poles from s = ak on the
s-plane to z = eak T on the z-plane.
5. It is tempting to say that z = esT , which is a one-to-one mapping,
hence s = ln(z)/T and
H(z) = H̄(s)|s=ln(z)/T .
This is incorrect in most cases, however.
6. The impulse response of a DT IIR filter is not directly useful for
implementation. We need a transfer function or a difference equation.
7. See Matlab function impinvar.
WPI D. Richard Brown III 26-March-2012 18 / 33
 ECE503: Digital Filter Design

Bilinear Transform
Idea: Given a causal stable LTI CT filter H̄(s), we simply substitute
 
2 1 − z −1
s=
T 1 + z −1
to get H(z).
Remark: This substitution is based on converting H̄(s) to a differential equation,
performing trapezoidal numerical integration with step size T to get a difference
equation, and then converting the difference equation to a transfer function H(z).
One appeal of the bilinear transform is that we have a direct method to go from
H̄(s) to H(z) that always works without going through the time-domain.
Your textbook shows:
◮ This is a one-to-one mapping between points in the s-plane and z-plane.
◮ Points in the left-half (right-half) s-plane are mapped to points inside
(outside) the unit circle on the z-plane.
◮ There is no aliasing even if H̄(s) is not bandlimited.
WPI D. Richard Brown III 26-March-2012 19 / 33
 ECE503: Digital Filter Design

Bilinear Transform: Simple Example

1
Suppose you are given a causal LTI CT system with H̄(s) = s−a . We can
compute H(z) straightforwardly with a little algebra:

H(z) = H̄(s)|s= 2  1−z−1 

T 1+z −1

1
=  
2 1−z −1
T 1+z −1
−a
T (1 + z −1 )
=
2(1 − z −1 ) − aT (1 + z −1 )
T (1 + z −1 )
=
(2 − aT ) − (2 + aT )z −1
β(1 + z −1 )
=
1 − αz −1
This is quite different than the H(z) we computed via impulse invariance.
WPI D. Richard Brown III 26-March-2012 20 / 33
 ECE503: Digital Filter Design

Comparing Impulse Invariant and Bilinear Transform

1
H̄(s) = s−a with a = − 21 and sampling frequency FT = 2.
1 5
CT CT
impulse invariant impulse invariant
0.8 bilinear xform 4 bilinear xform

magnitude response
impulse response

0.6 3

0.4 2

0.2 1

0 0
0 5 10 15 20 0 0.5 1 1.5 2
time frequency (Hz)

Note the impulse invariant magnitude response is not scaled to match the
CTFT. Even if it were, it should be clear there is some aliasing here.
WPI D. Richard Brown III 26-March-2012 21 / 33
 ECE503: Digital Filter Design

Bilinear Transform: Frequency Response

Given a causal stable LTI CT filter H̄(s), we can compute H(z) with via
the bilinear transform. What is the relationship between H̄(Ω) and H(ω)?
Recall that H̄(Ω) = H̄(s)|s=jΩ and H(ω) = H(z)|z=ejω . Substituting
these into the bilinear transform formula, we get
 
2 1 − e−jω
jΩ =
T 1 + e−jω
!
2 ejω/2 − e−jω/2
=
T ejω/2 + e−jω/2
1 jω/2 − e−jω/2 )
!
2j 2j (e
= 1 jω/2
T 2 (e + e−jω/2 )
2j
= tan(ω/2)
T
2
Hence Ω = T tan(ω/2) or ω = 2 tan−1 (ΩT /2).
WPI D. Richard Brown III 26-March-2012 22 / 33
 ECE503: Digital Filter Design

Bilinear Transform: Frequency Warping

0.8

0.6

0.4

0.2
ω (times π)

−0.2

−0.4

−0.6

−0.8

−1
−4 −3 −2 −1 0 1 2 3 4
ΩT (times π)

WPI D. Richard Brown III 26-March-2012 23 / 33

 ECE503: Digital Filter Design

Bilinear Transform: Consequences of Frequency Warping

The good news is that we don’t have to worry about aliasing. The bad
news is that we have to account for frequency warping when we start from
a discrete-time filter specification.
Procedure:
1. Convert the DT filter specifications to CT filter specifications
including prewarping the band edge frequencies
2. Design a CT IIR filter to meet the specifications
3. Transform the CT filter to a DT filter via the bilinear transform

WPI D. Richard Brown III 26-March-2012 24 / 33

 ECE503: Digital Filter Design

Bilinear Transform Lowpass Butterworth Filter Design Ex.

As before, we start with the desired specifications of the DT filter. We will

use the same parameters as last time: ωp = 0.2π, ωs = 0.3π,
1 − δ1 = 0.89125, and δ2 = 0.17783.

Steps:
1. Convert DT filter specs to CT filter specs (including pre-warping
frequencies).
2. Determine the required filter order N and CT cutoff frequency Ωc .
3. Determine H̄(s).
4. Determine H(z) via impulse invariance.

WPI D. Richard Brown III 26-March-2012 25 / 33

 ECE503: Digital Filter Design

Impulse-Invariant Lowpass Butterworth Filter Design Ex.

[⋆ ⋆ ⋆ details on board ⋆ ⋆ ⋆]

Main result from step 3:

0.20238
H̄(s) =
(s2 + 0.3966s + 0.5871)(s2 + 1.0836s + 0.5871)(s2 + 1.4802s + 0.5871)

Main result from step 4:

0.0007378(1 + z −1 )6
H(z) =
(1−1.2686z −1 +0.7051z −2 )(1−1.0106z −1 +0.3583z −2 )(1−0.9044z −1 +0.2155z −2 )

WPI D. Richard Brown III 26-March-2012 26 / 33

 ECE503: Digital Filter Design

Bilinear Transform Lowpass Butterworth Filter Design Ex.

10
CT Butterworth LPF
DT Butterworth LPF via bilinear transform
0

−10

−20
magnitude response

−30

−40

−50

−60

−70

−80
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized frequency (times π)

WPI D. Richard Brown III 26-March-2012 27 / 33

 ECE503: Digital Filter Design

Bilinear Transform Lowpass Butterworth Filter Design Ex.

2 CT Butterworth LPF
DT Butterworth LPF via bilinear transform
0

−2

−4
magnitude response

−6

−8

−10

−12

−14

−16

−18
0.14 0.16 0.18 0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34
normalized frequency (times π)

WPI D. Richard Brown III 26-March-2012 28 / 33

 ECE503: Digital Filter Design

General Filter Design Procedure

The bilinear transform lowpass filter design procedure is straightforward:

discrete-time
filter specifications

prewarp
DT frequency design CT
lowpass filter bilinear transform H(z)
specifications
to CT

If you want a different type of filter, e.g. bandpass, there are two options:
convert CT
discrete-time prototype
filter specifications lowpass filter to bilinear transform H(z)
desired filter type
convert CT (Appendix B)
prewarp frequency specs
DT frequency design prototype
to a prototype CT lowpass filter
specifications CT lowpass convert DT prototype
to CT filter (Appendix B) bilinear transform lowpass filter to desired
to get prototype filter type via H(z)
DT lowpass filter spectral transformation
(table 9.1)

WPI D. Richard Brown III 26-March-2012 29 / 33

 ECE503: Digital Filter Design

Bilinear Transform Bandpass Butterworth Filter Design Ex.

Specifications: ωp1 = 0.45π, ωp1 = 0.65π, ωp1 = 0.3π, ωp1 = 0.75π,
passband ripple less than 1 dB, minimum stopband attenuation 40 dB.
% This code follows example 9.4
% set sampling period
T = 1;

% prewarp frequencies
omega = [0.3 0.45 0.65 0.75]*pi;
Omega_prewarped = (2/T)*tan(omega/2);

% convert CT frequency specs to a prototype CT lowpass filter

% we want the two stopband frequencies to be GEOMETRICALLY symmetric
% around the GEOMETRIC center frequency of the passband
Omega_0 = sqrt(Omega_prewarped(2)*Omega_prewarped(3));
BW = Omega_prewarped(3)-Omega_prewarped(2);

% adjust lower stopband edge so that product of stopband edge frequencies

% matches Omega_0^2
Omega_prewarped(1) = Omega_0^2/Omega_prewarped(4);

% create prototype CT LPF filter specs (see Appendix B)

Omega_p = 1;
Omega_s = (Omega_0^2-Omega_prewarped(1)^2)/(Omega_prewarped(1)*BW);

% design prototype CT LPF (can also do this as shown in on board)

[N,Wn] = buttord(Omega_p,Omega_s,1,40,’s’);
[B,A] = butter(N,Wn,’s’);
[BT,AT] = lp2bp(B,A,Omega_0,BW);
[num,den] = bilinear(BT,AT,T);

WPI D. Richard Brown III 26-March-2012 30 / 33

 ECE503: Digital Filter Design

Bilinear Transform Lowpass Butterworth Filter Design Ex.

10

−10
magnitude response

−20

−30

−40

−50
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized frequency (times π)

WPI D. Richard Brown III 26-March-2012 31 / 33

 ECE503: Digital Filter Design

Bilinear Transform: Final Remarks

1. The bilinear transform only attempts to emulate the magnitude

response of the CT filter. The phase response of a DT filter derived
via the bilinear transform may be totally different than the CT filter.
2. See section 9.2.2 for several low-order filter examples.
3. See Matlab function bilinear.
4. The spectral transformations in Table 9.1 are also handled
conveniently with Matlab functions iirlp2lp, iirlp2hp, iirlp2bp,
and iirlp2bs.
5. Other CT filter design methods can also be used, e.g. Chebychev,
elliptic, ...

WPI D. Richard Brown III 26-March-2012 32 / 33

 ECE503: Digital Filter Design

Conclusions

1. This concludes our coverage of Chapter 9. Our focus was on

Sections 9.1-9.6, plus some additional material on impulse invariance.
We also referred to material in Appendices A-B of your textbook.
2. We are skipping Chapter 10, but it may be of interest to you if you
want to know how to design good FIR digital filters. The techniques
in Chapter 10 are not based on analog filter design.
3. Please read FFT basics (Section 11.3) and Number Representation
(Section 11.8) before the next lecture.
4. The next lecture is on Monday 02-April-2012 at 6pm.
5. The bonus midterm exam (60 points, 60 minutes, based on
fundamental material from Chapters 1-7) will begin at 7:30pm on
Monday 02-April-2012.

WPI D. Richard Brown III 26-March-2012 33 / 33