0% found this document useful (0 votes)
112 views102 pagesChap-7 Filter Design Techniques
filter design
Uploaded by
ujwa princeCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
112 views102 pagesChap-7 Filter Design Techniques
filter design
Uploaded by
ujwa princeCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 102
7
FILTER DESIGN TECHNIQUES
7.0 INTRODUCTION
Filters are a particularly important class of linear time-invariant systems Strictly speak-
ing, the term frequency-selective filter suggests a system that passes certain frequency
components and totally rejects all others, but in a broader context any system that
modifies certain frequencies relative to others is also called a filter. While the primary
emphasis in this chapter is on the design of frequency-selective filters, some of the tech-
niques are more broadly applicable. Also, we concentrate on the design of causal filters,
although in many conteats filters need not be restricted to causal designs, Very often,
noncausal filters are designed and implemented by modifying causal designs.
The design of filters involves the following stages: (1) the specification of the
desired properties of the system, (2) the approximation of the specifications using a
causal discrete-time system, and (3) the realization of the system. Although these three
steps are certainly not independent, we focus our attention primarily on the second
step, the first being highly dependent on the application and the third dependent on the
technology to be used for the implementation. In a practical setting, the desired filter is
generally implemented with digital computation and used to filter a signal thatis derived
from a continuous-time signal by means of periodic sampling followed by analog-to-
digital conversion. For this reason, it has become common to refer to discrete-time
filters as digital filters, even though the underlying design techniques most often relate
only to the discrete-time nature of the signals and systems.
When a discrete-time filter is to be used for discrete-time processing of continuous-
time signals in the configuration of Figure 7.1, the specifications for both the discrete-
time filter and the effective continuous-time filter are typically (but not always) given
in the frequency domain, This is especially common for frequency-selective filters such
439 440 Filter Design Techniques Chap. 7
ol em ‘| pe Le
tn) ytnl Js.
Figure 7.1 Basic system for
discrete-time filtering of
a continuous-time signals.
as lowpass, bandpass, and highpass filters. As shown in Section 4.4, if a linear time-
invariant discrete-time system is used as in Figure 7.1, and if the input is bandlimited
and the sampling frequency is high enough to avoid aliasing, then the overall system
behaves as a linear time-invariant continuous-time system with frequency response
ja) = { Hei"), IQ) x/Ta, (76)
then
H(e!®) = He (2). lol <7; (7)
Ta
i.e., the discrete-time and continuous-time frequency responses are related by a linear
scaling of the frequency axis, namely, w = 27, for || < x. Unfortunately, any practical
continuous-time filter cannot be exactly bandlimited, and consequently, interference
between successive terms in Eq. (7.5) occurs, causing aliasing, as illustrated in Fig-
ure 7.3. However, if'the continuous-time filter approaches zero at high frequencies, the
aliasing may be negligibly small, and a useful discrete-time filter can result from the
sampling of the impulse response of a continuous-time filter.
In the impulse invariance design procedure, the discrete-time filter specifications
are first transformed to continuous-time filter specifications through the use of Eq. (7.7).
Assuming that the aliasing involved in the transformation from He(j2) to H(e!) will
be negligible, we obtain the specifications on H.(j&) by applying the relation
R=w0/Ty (7.8)
Figure 7.3 Illustration of aliasing in
the impulse invariance design technique. Sec. 7.1 Design of Discrete-Time IIR Filters from Continuous-Time Filters 445
to obtain the continuous-time filter specifications from the specifications on H(e/).
After obtaining a suitable continuous-time filter based on these specifications, the
continuous-time filter with system function H,(s) is transformed to the desired discrete-
time filter with system function H(z). We develop the algebraic details of the transfor-
mation from H,(s) to H(z) shortly. Note, however, that in the transformation back
to discrete-time frequency, H(e’*) will be related to H(j&) through Eq. (7.5), which
again applies the transformation of Eq. (7.8) to the frequency axis. As a consequence,
the “sampling” parameter Ty cannot be used to control aliasing. Since the basic specifi-
cations are in terms of discrete-time frequency, if the sampling rate is increased (i.e.,if Ta
is made smaller), then the cutoff frequency of the continuous-time filter must increase
in proportion. In practice, to compensate for aliasing that might occur in the transfor-
mation from H.(s) to H(2), the continuous-time filter may be somewhat overdesigned,
i.e., designed to exceed the specifications, particularly in the stopband.
While the impulse invariance transformation from continuous time to discrete
time is defined in terms of time-domain sampling, itis easy to carry out as a transforma-
tion on the system function. To develop this transformation, let us consider the system
function of the continuous-time filter expressed in terms of a partial fraction expansion,
so that?
N
HAs) =o Ae, (79)
kal
The corresponding impulse response is
N
Ae", 120,
he(t) = = ieee (7.10)
0, 1<0.
‘The impulse response of the discrete-time filter obtained by sampling Tyh-(t) is
N
hn] = Tah(nTa) = D> Ta Ane" un]
i
mi (7.11)
=o TAde™"
The system function of the discrete-time filter is therefore given by
N
Ta Ax
H(z) = Lira eatiga (7.12)
In comparing Eqs. (7.9) and (7.12), we observe that a pole at s = s; in the s-plane
transforms to a pole at z = e**”: in the z-plane and the coefficients in the partial fraction
expansions of H,(s) and H(z) are equal, except for the scaling multiplier Ty. If the
continuous-time filter is stable, corresponding to the real part of s, being less than zero,
then the magnitude of e*” will be less than unity, so that the corresponding pole in the
discrete-time filter is inside the unit circle. Therefore, the causal discrete-time filter is
1 For simplicity, we assume in the discussion that all poles of H(s) are single order. In Problem 7.24,
we consider the modifications required for multiple-order poles. 448 Filter Design Techniques Chap. 7
also stable. While the poles in the s-plane map to poles in the z-plane according to the
relationship zx = e*", it is important to recognize that the impulse invariance design
procedure does not correspond toa simple mapping of the s-plane to the z-plane by that
relationship. In particular, the zeros in the discrete-time system function are a function
of the poles and the coefficients 7 Ax in the partial fraction expansion, and they will not
in general be mapped in the same way the poles are mapped. We illustrate the impulse
invariance design procedure with the following example.
Example 7.2 Impulse Invariance with a Butterworth Filter
Let us consider the design of a lowpass discrete-time filter by applying impulse invari-
ance to an appropriate Butterworth continuous-time filter.” The specifications for the
liscrete-time filter are
0.89125 <|H(e*) <1, O< |u| < 0.27, (7.13a)
1H(e!*)| < 0.17783, 03m < lol <7, (7.136)
Since the parameter Ty cancels in the impulse invariance procedure, we can choose
Ta = 1,80 that » = &. In Problem 7.2, this same example is considered, but with the
parameter Ty explicitly included to illustrate how and where it cancels.
In designing the filter using impulse invariance on a continuous-time Butter-
worth filter, we must first transform the discrete-time specifications to specifications
on the continuous-time filter. Recall that impulse invariance corresponds to a linear
mapping between & and w in the absence of aliasing. For this example, we will assume
that the effect of aliasing is negligible. After the design is complete, we can evaluate
the resulting frequency response against the specifications in Eqs. (7.13a) and (7.13b)..
Because of the preceding considerations, we want to design a continuous-time
Butterworth filter with magnitude function |H-(j2)] for which
0.89125 <|H(JQ) <1, 0< 12} <02n, (7.14a)
|He(jQ)| < 0.17783, 0.3m <|Q\ 0, then |z| > 1 for all &. That is, if a pole of H-(s) is in the left-half s-plane, its
image in the z-plane will be inside the unit circle. Therefore, causal stable continuous-
time filters map into causal stable discrete-time filters. Sec. 7.1 Design of Discrete-Time IIR Filters from Continuous-Time Filters 451
Next, to show that the j9-axis of the s-plane maps onto the unit circle, we substi-
tute s = /2 into Eq, (7.22), obtaining
+ j2T4/2
= Tane (7.24)
z
From Eq. (7.24), it is clear that |z| = 1 for all values of s on the jS-axis, That is, the
axis maps onto the unit circle, so Eq. (7.24) takes the form
_ 1+ jQT/2
© T= jat2
To derive a relationship between the variables w and Q, it is useful to return to
Eq. (7.20) and substitute z =e. We obtain
jo
(7.25)
2 _ 7.26)
6-7, (een): ey
or, equivalently,
8 2 fre (jsinw/2)) _ 2f
saa4 j= 7 ae] 7, tanw/2). (727)
Equating real and imaginary parts on both sides of Eq. (7.27) leads to the relations
o =Oand
2= 2 tan(w/2), (7.28)
Ts
or
w = 2arctan(QT,/2). (7.29)
‘These properties of the bilinear transformation as a mapping from the s-plane to
the z-plane are summarized in Figures 7.6 and 7.7. From Eq. (7.29) and Figure 7.7, we
see that the range of frequencies 0 < 2 < co maps to 0 < w < 2, while the range
—co < Q < Omaps to —x < w < 0. The bilinear transformation avoids the problem
of aliasing encountered with the use of impulse invariance, because it maps the entire
imaginary axis of the s-plane onto the unit circle in the z-plane. The price paid for this,
however, is the nonlinear compression of the frequency axis depicted in Figure 7.7. Con-
sequently, the design of discrete-time filters using the bilinear transformation is useful
only when this compression can be tolerated or compensated for, as in the case of filters
that approximate ideal piecewise-constant magnitude-response characteristics. This is
illustrated in Figure 7.8, where we show how a continuous-time frequency response
and tolerance scheme maps to a corresponding discrete-time frequency response and
tolerance scheme through the frequency warping of Eqs. (7.28) and (7.29). If the critical
frequencies (such as the passband and stopband edge frequencies) of the continuous-
time filter are prewarped according to Eq. (7.28) then, when the continuous-time filter
is transformed to the discrete-time filter using Eq. (7.21), the discrete-time filter will
meet the desired specifications. Image of
left half-plane
me Plane
Image of
5 =02 (unit circle)
Figure 7.8 Mapping of the s-plane
‘onto the z-plane using the bilinear
transformation.
Figure 7.7. Mapping of the
continuous-time frequency axis onto the
discrete-time frequency axis by bilinear
transformation.
Figure 7.8 Frequency warping
inherent in the bilinear transformation of
a continuous-time lowpass filter into a
discrete-time lowpass fiter. To achieve
the desired discrete-time cutofl
frequencies, the continuous-time cutoff
frequencies must be prewarped as
indicated. Sec.7.1 Design of Discrete-Time IIR Filters from Continuous-Time Filters 453
Typical frequency-selective continuous-time approximations are Butterworth,
Chebyshev, and elliptic filters. The closed-form design formulas of these continuous-
time approximation methods make the design procedure rather straightforward. As
discussed in Appendix B a Butterworth continuous-time filter is monotonic in the pass-
band and in the stopband. A type I Chebyshev filter has an equiripple characteristic in
the passband and varies monotonically in the stopband. A type II Chebyshev filter is
monotonicin the passband and equiripple in the stopband. An elliptic filter is equiripple
in both the passband and the stopband. Clearly, these properties will be preserved when
the filter is mapped to a digital filter with the bilinear transformation. This is illustrated
by the dashed approximation shown in Figure 7.8.
Although the bilinear transformation can be used effectively in mapping a
piecewise-constant magnitude-response characteristic from the s-plane to the z-plane,
the distortion in the frequency axis also manifests itself as a warping of the phase re-
sponse of the filter. For example, Figure 7.9 shows the result of applying the bilinear
transformation to an ideal linear phase factor e~**. If we substitute Eq. (7.20) for s
and evaluate the result on the unit circle, the phase angle is —(2a/7,) tan(w/2). In Fig-
ure 7.9, the solid curve shows the function —(2a/ Ty) tan(w/2), and the dotted curve
is the periodic linear phase function —(wa/ Tz), which is obtained by using the small
angle approximation /2 ~ tan(w/2). From this, it should be evident that if we were
interested in a discrete-time lowpass filter with a linear phase characteristic, we could
not obtain such a filter by applying the bilinear transformation to a continuous-time
lowpass filter with a linear phase characteristic.
‘As mentioned previously, because of the frequency warping, the use of the
bilinear transformation is restricted to the design of approximations to filters with
tHe)
Figure 7.9 _ Illustration of the effect ofthe bilinear transf
characteristic. (Dashed line is linear phase and solid line is phase resulting from
bilinear transformation.) 454 Filter Design Techniques Chap. 7
piecewise-constant frequency magnitude characteristics, such as highpass, lowpass and
bandpass filters. As demonstrated in Example 7.2, impulse invariance can also be used
to design lowpass filters. However, impulse invariance cannot be used to map highpass
continuous-time designs to highpass discrete-time designs, since highpass continuous-
time filters are not bandlimited.
In Example 4.5, we discussed a class of filters often referred to as discrete-time
differentiators. A significant feature of the frequency response of this class of filters
is that it is linear with frequency. The nonlinear warping of the frequency axis intro-
duced by the bilinear transformation will not preserve that property. Consequently, the
bilinear transformation applied to a continuous-time differentiator will not result in a
discrete-time differentiator. However, impulse invariance applied to an appropriately
bandlimited continuous-time differentiator will result in a discrete-time differentiator.
7.1.3 Examples of Bilinear Transformation Design
In the following discussion, we present a number of examples to illustrate IIR filter de-
sign using the bilinear transformation. Example 7.3 serves to illustrate the design proce-
dure based on the bilinear transformation, in comparison with the use of impulse invari-
ance. Examples 7.4, 7.5, and 7.6 illustrate a Butterworth, Chebyshev, and elliptic filter,
respectively, each designed to the same specifications using the bilinear transformation.
Example 7.3 Bilinear Transformation
of a Butterworth Filter
Consider the discrete-time filter specifications of Example 7.2, in which we illustrated
the impulse invariance technique for the design of a discrete-time filter. The specifica-
tions on the discrete-time filter are
0.89125 <|H(e/")J <1, O 0.89125 (7.32a) Seo. 7.1
Design of Discrete-Time IIR Filters from Continuous-Time Filters 455
and
|H-(j2 tan(0.15x))| < 0.17783. (7.32b)
‘The form of the magnitude-squared function for the Butterworth filter is
1
UDP = ere 733)
Solving for N and Q- with the equality sign in Eqs. (7.32a) and (7.32b), we obtain
wo ye
denn) = (as) (7.34a)
2tan(0.15e)\"" _ (_1_\?
1+ (Zee) =(a) > (7.34b)
and
and solving for N in Eqs. (7.34a) and (7.34b) gives
2logltan(0.15)/tan(0.1x)] (735)
= 5.305.
Since N must be an integer, we choose N = 6. Substituting N = 6 into
Eq. (7.34b), we obtain 2. = 0.766. For this value of 2., the passband specifications
are exceeded and the stopband specifications are met exactly. This is reasonable for
the bilinear transformation, since we do not have to be concerned with aliasing. That
is, with proper prewarping, we can be certain that the resulting discrete-time filter will
meet the specifications exactly at the desired stopband edge.
In the s-plane, the 12 poles of the magnitude-squared function are uniformly
distributed in angle on a circle of radius 0.766, as shown in Figure 7.10. The sys-
tem function of the continuous-time filter obtained by selecting the left half-plane
Figure 7.10. s-plane locations for poles of He(s)H.(-s) for sixth-order
Butterworth fiter in Example 7.3. Fitter Design Techniques Chap. 7
My 02m 04m 06m O89 7
Radian frequency (w)
@)
n
0 O25 Oa 6% 08% 7
Radian frequency (w)
()
2
Samples
0 0.20 O40 0.60 0.80 7
Radian frequency (1)
©
Figure 7.11 Frequency response of sixth-order Butterworth filter transformed by
bilinear transform. (a) Log magnitude in dB. (b) Magnitude. (c) Group delay. Sec. 7.1 Design of Discrete-Time IIR Filters from Continuous-Time 457
poles is
Hels) = 0.20238
= GF 030965 + OSBTING + 1.08365 + OSBTINS? + 1.48028 + 05871)
(7.36)
‘The system function for the discrete-time filters then obtained by applying the bilinear
transformation to H-(s) with Ty = 1. The result is
0.0007378(1 + -")*
1 = 1.2686z~! + 0.7051z-)(1 — 1.0106z
MO=— + 0.358327)
(737)
1
* (0908421 0215522)"
The magnitude, log magnitude, and group delay of the frequency response of the
discrete-time filter are shown in Figure 7.11. At @ = 0.2% the log magnitude is
—0.56 dB, and at « = 0.3% the log magnitude is exactly ~15 dB.
Since the bilinear transformation maps the entire j9-axis of the s-plane onto
the unit circle in the z-plane, the magnitude response of the discrete-time filter falls
off much more rapidly than that of the original continuous-time filter. In particular,
the behavior of H(e!*) at « = x corresponds to the behavior of H.(j2) at
‘Therefore, since the continuous-time Butterworth filter hasa sixth-orderzeroat
the resulting discrete-time filter has a sixth-order zero at z = —1
It is interesting to note that, since the general form of the Nth-order Butter-
worth continuous-time filter is as given by Eq. (7.33), and since w and @ are related
by Eq. (7.28), it follows that the general Nth-order Butterworth discrete-time filter has
magnitude-squared function
1
tan(w/2) \"""
+ (Goer)
|H(e!*)P? = (7.38)
where tan(w./2) = 74/2.
‘The frequency-response function of Eq. (7.38) has the same properties as the
continuous-time Butterworth response; i.e., it is maximally flat? and |H(e!*)|? = 0.5.
However, the function in Eq. (7.38) is periodic with period 2 and falls off more sharply
than the continuous-time Butterworth response.
We do not design discrete-time Butterworth filters directly by starting with
Eq, (7.38), because it is not straightforward to determine the z-plane locations of the
poles (all the zeros are at z = —1) associated with the magnitude-squared function of
Eq. (7.38). It is necessary to determine the poles so as to factor the magnitude-squared
8 The first (2N — 1) derivatives of |H(e/#)/? are zero at w = 0. 458 Filter Design Techniques Chap. 7
function into H(z)H(z"') and thereby determine H(z). It is much easier to find the
s-plane pole locations (alll the zeros are at infinity), factor the continuous-time system
function, and then transform the left half-plane poles by the bilinear transformation as
we did in Example 7.3.
Equations of the form of Eq. (7.38) may also be obtained for discrete-time
Chebyshev filters, but the same difficulties arise in their use. Thus, the two-step approach
just described has become the established method of designing IIR frequency-selective
filters.
‘The major approximation methods for frequency-selective IIR analog filters are
the Butterworth, Chebyshev, and elliptic function approximation methods. The details
of these methods can be found in Guillemin (1957), Daniels (1974), Weinberg (1975),
and Lam (1979). The methods are generally explained and developed in terms of lowpass
filter approximations. This is the approach followed in Appendix B, where we summarize
the essential features of some of the methods. In the next three examples, we illustrate
the realization of a set of filter specifications for each of these classes of filters. The details
of the design computations are not presented, since they are tedious and lengthy and are
best carried out by computer programs that incorporate the appropriate closed-form
design equations.
‘The lowpass discrete-time filter specifications for these examples are those used
in Example 7.1, ic.,
0.99 < |H(e!*)| < 1.01, \o| < 0.42, (7.39a)
and
|H(e/*)| < 0.001, 0.62 <|o| halnjeio", (7.40) 466 Filter Design Techniques Chap. 7
where hg[n] is the corresponding impulse response sequence, which can be expressed
in terms of Hg(e/*) as
Ha(e!)e!*"deo. (741)
Many idealized systems are defined by piecewise-constant or piecewise-functional fre-
quency responses with discontinuities at the boundaries between bands. As a result,
these systems have impulse responses that are noncausal and infinitely long. The most
straightforward approach to obtaining a causal FIR approximation to such systems is to
truncate the ideal response. Equation (7.40) can be thought of as a Fourier series repre-
sentation of the periodic frequency response H,(e!), with the sequence hg[7!] playing
the role of the Fourier coefficients. Thus, the approximation of an ideal filter by trunca-
tion of the ideal impulse response is identical to the issue of the convergence of Fourier
series, a subject that has received a great deal of study. A particularly important concept
from this theory is the Gibbs phenomenon, which was discussed in Example 2.22. In
the following discussion, we will see how this nonuniform convergence phenomenon
manifests itself in the design of FIR filters.
The simplest way to obtain a causal FIR filter from /a[n] is to define a new system
with impulse response h[71] given by®
ta {a O
