Design of Stable IIR Digital Filters with Specied

Group Delay Errors

Yasunori Sugita and Toshinori Yoshikawa
AbstractThe design problem of Innite Impulse Response (IIR)
digital lters is usually expressed as the minimization problem of
the complex magnitude error that includes both the magnitude and
phase information. However, the group delay of the lter obtained
by solving such design problem may be far from the desired group
delay. In this paper, we propose a design method of stable IIR digital
lters with prespecied maximum group delay errors. In the proposed
method, the approximation problems of the magnitude-phase and
group delay are separately dened, and these two approximation
problems are alternately solved using successive projections. As a
result, the proposed method can design the IIR lters that satisfy the
prespecied allowable errors for not only the complex magnitude but
also the group delay by alternately executing the coefcient update
for the magnitude-phase and the group delay approximation. The
usefulness of the proposed method is veried through some examples.
KeywordsFilter design, Group delay approximation, Stable IIR
lters, Successive projection method.
I. INTRODUCTION
Finite-impulse response (FIR) digital lters nd many ap-
plications in image processing, waveform transmission, etc. in
which phase distortion becomes a problem, because they can
easily realize an exactly linear phase characteristic and are
always stable [1]-[4]. However, since the resulting delay at the
output of the exactly linear phase FIR lters is half of the lter
order, its group delay may become unacceptably large when
high-order lters or narrow transition bands are required. On
the other hand, it is known that Innite Impulse Response (IIR)
digital lters can realize about the same magnitude response
by the lower order compared with the FIR lters. Hence, in
order to process a signal processing with high-speed and with
high-precision, it is very important to design the IIR lters.
The design problem of the IIR lters is usually expressed
as the approximation problem of the magnitude and the phase
responses [7]-[12]. However, the group delay response of the
lter obtained by those methods tends to become relatively
large, especially in the vicinity of the band-edge. This is
because that the complex magnitude error, which include both
the magnitude and phase responses, is minimized instead of
the group delay error. In general, an allpass lter is used
to equalize the phase response to make the group delay
approximately constant in the interested bandwidth. However,
the use of the allpass lter is not necessarily a good policy
because the lter coefcients are redundant. Therefore, it is
desirable to realize the lter that has an equalized group
Authors are with the Department of Electrical Engineering, Na-
gaoka University of Technology, Niigata, 940-2188 Japan (e-mail:
[email protected])
delay without using the allpass lter. This motivates the
investigation on designing the IIR lters with prespecied
maximum group delay errors by directly approximating the
group delay response.
By the way, it is desirable to have a unied approach that
is able to design various types of digital lters. Recently,
semi-denite programming (SDP) has been widely employed
designing various types of FIR and IIR lters [6],[12]-[14].
However, because the design method based on the SDP needs
to solve a large matrix, it often causes a problem that personal
computers cannot be used. On the other hand, the successive
projection method (SP) proposed by A. A. -Taleb et. al. [3],
which is one of convex projection algorithm, is a simple
iterative approximation method, and the amount of computing
memory is very small. The SP method also has been applied to
the design problem of various types of the lters such as one-
and two-dimentional IIR lters, minimum phase FIR lters,
the lters with time- and frequency-domain constraints, and so
on [3],[15]-[19].That is, the applicability of the SP method is
wider than that of the SDP algorithm. Moreover, it is known
that the lters designed by the SP method have better lter
characteristics than those by other design methods.
In [18], we presented a design method of stable IIR lters
using the SP method. However, since this method is also for-
mulated as the approximation problem of the magnitude-phase
responses, the group delay response of the lter obtained tends
to become a large. So, in this paper, we propose an improved
method to design the stable IIR lters with prespecied maxi-
mum group delay errors using the SP method. In our proposed
method, the approximation problems of the magnitude-phase
and the group delay responses are dened separately on the
different dimension, and these two approximation problems
are solved alternately using the SP method. As a result, the
proposed method allows the direct approximation of the group
delay response, and it can restrict the group delay response
within the preselected allowable errors. The usefulness of the
proposed method is veried through some examples.
This paper is organized as follows: in section 2, the itera-
tive algorithm for the approximation of the magnitude-phase
responses using the SP method [18] is described. The iterative
algorithm to directly approximate the group delay response are
described in Section 3. Several design examples are given in
Section 4. Conclusions of this work are drawn in Section 5.
II. MAGNITUDE-PHASE APPROXIMATION [18]
Let the transfer function of an innite impulse response
(IIR) digital lter with numerator degree m and denominator
International Journal of Information and Communication Engineering 6:1 2010
7
degree n be
H (z) =
N (z)
D(z)
=
m

i=0
b
i
z
i
n

i=0
a
i
z
i
, (1)
where a
i
and b
i
are the real-valued lter coefcients and a
0
=
1. Then, the frequency response H() of H(z) can be given
by
H () =
N ()
D()
=
m

i=0
b
i
e
ji
n

i=0
a
i
e
ji
= |H ()| e
j()
, (2)
where is the normalized angular frequency.
Now, we dene a desired frequency response H
d
() as
H
d
() = A() e
j
d
()
, (3)
where A() is the desired magnitude response and
d
()
is the desired group delay response. Then, the lter design
problem considered here is to nd the lter coefcients a
i
and b
i
satisfying the following complex magnitude error:
|H
d
() H ()| () (4)
for
H
, where () is the positive maximum allowable
deviation from the desired frequency response H
d
() and
H
is the interest region, which usually consists of the passband
and stopband.
The SP method is an iterative approximation technique.
Using the technique presented in [18], the error function
E
[k+1]
H
() in the (k+1)th iteration step can be written as
E
[k+1]
H
() =
Q
[k+1]
()
R
[k]
()
, (5)
where
Q
[k+1]
()=A()
n

i=0
a
i
cos
_

d
()+i2t
[k+1]
n
()
_

i=0
b
i
cos
_
i2t
[k+1]
n
()
_
(6)
t
[k+1]
n
() =
1
2
tan
1
y
[k+1]
n
()
x
[k+1]
n
()
(7)
x
[k+1]
n
() = A()
n

i=0
a
[k+1]
i
cos (
d
() +i)

i=0
b
[k+1]
i
cos (i)
(8)
y
[k+1]
n
() = A()
n

i=0
a
[k+1]
i
sin (
d
() +i)
+
m

i=0
b
[k+1]
i
sin (i)
(9)
R
[k]
() =
n

i=0
a
[k]
i
cos
_
i 2t
[k]
d
()
_
(10)
t
[k]
d
() =
1
2
tan
1
y
[k]
d
()
x
[k]
d
()
(11)
x
[k]
d
() =
n

i=0
a
[k]
i
cos (i) (12)
y
[k]
d
() =
n

i=0
a
[k]
i
sin (i) . (13)
The lter coefcient update by the SP method is equivalent to
solving the following optimization problem:
Minimize
_
_
a
[k+1]
a
[k]
_
_
2
+
_
_
_b
[k+1]
b
[k]
_
_
_
2
Subject to E
[k+1]
H
(
M
) (
M
) 0
where
_
_
_a
[k+1]
a
[k]
_
_
_
2
=
N

i=0
_
a
[k+1]
i
a
[k]
i
_
2
(14)
_
_
_b
[k+1]
b
[k]
_
_
_
2
=
N

i=0
_
b
[k+1]
i
b
[k]
i
_
2
, (15)
and
M
is the frequency point at which the specications are
violated the most and satises
E
[k]
H
(
M
) (
M
) = max
all H
E
[k]
H
() (). (16)
Consequently, for the magnitude-phase response approxima-
tion, we can get the following iterative algorithm:
a
[k+1]
i
= a
[k]
i

R
[k]
(
M
)
_
E
[k]
H
(
M
) (
M
)
_

[k]
1M
n

i=0
(
[k]
1M
)
2
+
m

i=0
(
[k]
2M
)
2
(17)
b
[k+1]
i
= b
[k]
i
+
R
[k]
(
M
)
_
E
[k]
H
(
M
) (
M
)
_

[k]
2M
n

i=0
(
[k]
1M
)
2
+
m

i=0
(
[k]
2M
)
2
,
(18)
where

[k]
1M
= A(
M
) cos
_

d
(
M
)
M
+i
M
2t
[k]
n
(
M
)
_
(19)

[k]
2M
= cos (i
M
2t
n
(
M
)) . (20)
In order to obtain the stable IIR lters, in [18], they
expanded eq.(17) using Rouch es theorem. Then, the iterative
algorithm is
a
[k+1]
i
= a
[k]
i

[k]
i,a
, (21)
where

[k]
i,a
=
R
[k]
(
M
)
_
E
[k]
H
(
M
) (
M
)
_

[k]
1M
n

i=0
(
[k]
1M
)
2
+
m

i=0
(
[k]
2M
)
2
. (22)
In eqs.(21) and (22), (0 < 1) is chosen as much as
possible large value satisfying

i=1

[k]
i,a
z
i

<

D
[k]
(z)

for z = e
j
, (23)
International Journal of Information and Communication Engineering 6:1 2010
8
where (0 < < 1) is the maximum allowable radius of the
poles of the lter obtained.
The detail of this algorithm has been presented in [18].
By the way, the group delay response of the lter obtained
by the above iterative algorithm may be distant from the
desired group delay because this is the iterative algorithm for
the magnitude-phase approximation. In the next section, we
consider about the iterative algorithm to approximate directly
the group delay response.
III. GROUP DELAY APPROXIMATION
The group delay response of the frequency response H()
in eq.(2) is written by
() = Re

i=0
i b
i
e
ji
m

i=0
b
i
e
ji
+
n

i=0
i a
i
e
ji
n

i=0
a
i
e
ji

. (24)
Then, the problem considered here is to nd the lter coef-
cients a
i
and b
i
satisfying

d
() Re

i=0
i b
i
e
ji
m

i=0
b
i
e
ji
+
n

i=0
i a
i
e
ji
n

i=0
a
i
e
ji

()
(25)
for

. Where a nonnegative function () is the

maximum allowable error from the desired group delay re-
sponse
d
(),

is the frequency band where the desired

group delay is prescribed, which consists of the passband
in this paper. However, it is very difcult to solve directly
this design problem using the SP method because eq. (25)
is the rational function of the lter coefcients. The SP
method has a property that the lter coefcients become
a
[n+1]
i
a
[n]
i
and b
[n+1]
i
b
[n]
i
, i.e. D
[n+1]
() D
[n]
()
and N
[n+1]
() N
[n]
(), if the algorithm converges. And
so, we consider the following new design formula instead of
eq.(25):

d
()Re

i=0
i b
i
e
ji

N()
+
n

i=0
i a
i
e
ji

D()

()
(26)
where

N () =
m

i=0
b

i
e
ji
= |

N () |e
j

N()
, (27)

D() =
n

i=0
a

i
e
ji
= |

D()|e
j

D()
, (28)
and a

i
and b

i
are the previous lter coefcients in each
iteration step.
The coefcient update by the SP method is equivalent to
solving the following optimization problem:
Minimize
_
_
a
[k+1]
a
[k]
_
_
2
+
_
_
_b
[k+1]
b
[k]
_
_
_
2
Subject to

E
[k+1]

(
M
)

(
M
) 0
where
_
_
_a
[k+1]
a
[k]
_
_
_
2
=
N

i=0
_
a
[k+1]
i
a
[k]
i
_
2
(29)
_
_
_b
[k+1]
b
[k]
_
_
_
2
=
N

i=0
_
b
[k+1]
i
b
[k]
i
_
2
(30)
E
[k+1]

(
M
) =
d
(
M
)

i=0
ib
[k+1]
i
cos
_
i
M
+

N
(
M
)
_
|

N(
M
)|

i=0
ia
[k+1]
i
cos
_
i
M
+

D
(
M
)
_

D(
M
)

,
(31)
and
M
is the frequency point at which the specications are
violated the most and satises

E
[k]

(
M
)

(
M
) = max
all

E
[k]

()

(). (32)
Below, in order to simplify the designation,

D(
M
),

N(
M
),
E
[k]

(
M
), and (
M
) are written as

D
M
,

N
M
, E
[k]
,M
, and

M
.
Solving the above optimization problem, we can get the
following iterative algorithm.
a
[k+1]
i
= a
[k]
i

|

D
M
||

N
M
|
_

E
[k]
,M

M
_

[k]
2M
sign(E
[k]
,M
)
m

i=0
(
[k]
1M
)
2
+
m

i=0
(
[k]
2M
)
2
(33)
b
[k+1]
i
= b
[k]
i
+
|

D
M
||

N
M
|
_

E
[k]
,M

M
_

[k]
1M
sign(E
[k]
,M
)
m

i=0
(
[k]
1M
)
2
+
m

i=0
(
[k]
2M
)
2
,
(34)
where

[k]
1M
= |

D
M
|i cos
_
i
M
+

N
(
M
)
_
(35)

[k]
2M
= |

N
M
|i cos
_
i
M
+

D
(
M
)
_
. (36)
And then, in order to design the stable IIR lters, eq.(33)
is modied using Rouch es theorem as follows:
a
[k+1]
i
= a
[k]
i

[k]
i,a
(37)
International Journal of Information and Communication Engineering 6:1 2010
9
where

[k]
i,a
=
|

D
M
||

N
M
|
_

E
[k]
,M

M
_

[k]
2M
sign(E
[k]
,M
)
m

i=0
(
[k]
1M
)
2
+
m

i=0
(
[k]
2M
)
2
(38)
and (0 < 1) is chosen as much as possible large value
satisfying

i=1

[k]
i,a
z
i

<

D
[k]
(z)

for z = e
j
. (39)
In the proposed method, the magnitude-phase approxima-
tion algorithm described in section 2 and the group delay
approximation algorithm described in this section are carried
out alternately. Note, however, that the direct approximation
for the group delay response should be executed after the
magnitude-phase approximation converged enough.
The design procedure of the proposed method is summa-
rized as follows.
The Design Procedure.
Step 0 Set the lter order m and n, desired frequency
response H
d
(), magnitude allowable error (),
group delay allowable error (), and an initial lter
coefcients a
[k]
|
k=0
and b
[k]
|
k=0
.
Step 1 Calculate the complex magnitude error E
[k]
H
() and
nd the frequency points
M
satisfying eq. (16).
Step 2 If E
[k]
H
(
M
) >(
M
), then compute the new lter
coefcients h
[n+1]
by eqs. (17)-(23). Otherwise, set
h
[n+1]
=h
[n]
Step 3 If (E
[k+1]
H
()())/() ( 0.1) for all

H
, then go to next step. Otherwise, go back to Step
1.
Step 4 Calculate the group delay error |E
[n+1]

()| and nd
the frequency points
M
satisfying eq. (32).
Step 5 If E
[k+1]

(
M
)>(
M
), then compute the new lter
coefcients h
[n+2]
by eqs. (33)-(39); otherwise, set
h
[n+2]
=h
[n+1]
.
Step 6 If h
[n+1]
= h
[n]
in Step 2 and h
[n+2]
= h
[n+1]
in
Step 4, then terminates. Otherwise, go back to Step
1.
IV. DESIGN EXAMPLES
In this section, the examples of the IIR lters with reduced
group delay errors are presented to illustrate the effectiveness
of the proposed method. In all the following examples, the
initial value of the lter coefcients was set to a
i
= 0 for
i = 1, 2, , n and b
i
= 1 for i = 0, 1, 2, , m.
A. Example 1
The specications are as follows:
n = m = 4
H
d
()=
_
e
j5.0
if 0 || 0.2
0 if 0.4 ||
()=
_

p
if 0 || 0.2

s
if 0.4 ||
()=

if 0 || 0.2
0 0.2 0.4 0.6 0.8 1
80
60
40
20
0
Normalized Frequency ( rad/sample)
M
a
g
n
i
t
u
d
e

(
d
B
)
(a)
0 0.1 0.2
0.4
0.2
0
0.2
Normalized Frequency ( rad/sample)
M
a
g
n
i
t
u
d
e

(
d
B
)
(b)
0 0.1 0.2
4.5
5
5.5
6
Normalized Frequency ( rad/sample)
G
r
o
u
p

D
e
l
a
y

(
S
a
m
p
l
e
s
)
(c)
Fig. 1. Frequency responses of the proposed IIR lters with = 0.500
(solid lines) and = 0.250 (dashed lines) and of the previous IIR lter
(dotted lines) in Example 1. (a) Overall magnitude response (b) Magnitude
response in the passband (c) Group delay response in the passband
The maximum allowable pole radius to guarantee the sta-
bility was set to 0.940.
First of all, for comparison, the IIR lter with the above de-
sign specications was designed using the previous SP method
[18]. Note that this method [18] is not possible to specify
the maximum group delay allowable error (). The resulting
maximum group delay error of the lter obtained using this
method was 0.630 in passband, and the maximum passband
ripple and the maximum stopband ripple were 0.02130 and
0.02130, respectively. And, the maximum pole radius was
0.895.
On the other hand, the main advantage of the proposed
method is that it can directly specify the maximum group delay
allowable error (). The resulting lters for many different
() are summarized in Table 1. Moreover, the frequency
International Journal of Information and Communication Engineering 6:1 2010
10
TABLE I
PERFORMANCE COMPARISON WITH [18] (EX. 1)
Previous SP [18] Proposed Method
p 0.0213 0.0215 0.0247
s 0.0213 0.0215 0.0247
0.6302 0.5000 0.2500
Maximum pole radius 0.8951 0.9028 0.9018
responses of the obtained lter are depicted in Figures. 1(a)-
(c). In Figure 1, the dotted line is the frequency response of
the lter designed by [18], and the solid line and the dashed
line are the frequency response in the case of () = 0.500
and () = 0.250 of the proposed method, respectively. It is
seen from Figure 1 and Table 1 that the performance of the
lter obtained using the proposed method is much better in
group delay response, while the magnitude response is poorer
ripples. Also, we can see that the magnitude and the group
delay errors have the relation of trade-off.
B. Example 2
We attempt the design of the IIR lters that satisfy the
following specications:
H
d
()=
_
e
j10.0
if 0 || 0.5
0 if 0.6 ||
()=
_
0.0132 if 0 || 0.5
0.0132 if 0.6 ||
()=0.250 if 0 || 0.5
= 0.960
The performance of the resulting lters are summarized in
Table 2, and the frequency responses of the IIR lter with
(n, m) =(8, 12) obtained by the proposed method are shown
in Figure 2. From Table 2, it can see that, in the previous SP
method, the passband ripple becomes small with an increase
in order n but the group delay response in the passband cannot
satisfy the given specication. Whereas, the lter with order
(n, m)=(8, 12) obtained by the proposed method satises all
the given specications. That is, the proposed method has a
possibility that the lter that can not be designed using the
previous SP method can be designed.
V. CONCLUSION
In this paper, we have proposed the improved method to
design the stable IIR lters with prespecied group delay
errors using the successive projection method. In our proposed
method, the approximation problems of the magnitude-phase
and the group delay responses are dened separately on the
different dimension, and these two approximation problems
are solved alternately using the successive projection method.
With the proposed method, it is possible to directly approxi-
mate of the group delay response, and it can restrict the group
delay response within the preselected allowable error. As a
result, the proposed method can design the lters that can
not be obtained using the previous method. Finally, the design
examples are demonstrated to illustrate the effectiveness of the
proposed method.
0 0.2 0.4 0.6 0.8 1
80
60
40
20
0
Normalized Frequency ( rad/sample)
M
a
g
n
i
t
u
d
e

(
d
B
)
(a)
0 0.1 0.2 0.3 0.4 0.5
0.15
0.1
0.05
0
0.05
0.1
0.15
Normalized Frequency ( rad/sample)
M
a
g
n
i
t
u
d
e

(
d
B
)
(b)
0 0.1 0.2 0.3 0.4 0.5
9.7
10
10.3
Normalized Frequency ( rad/sample)
G
r
o
u
p

D
e
l
a
y

(
S
a
m
p
l
e
s
)
(c)
Fig. 2. Frequency response of the proposed IIR lters with the order
(n, m) = (8, 12) in Example 2. (a) Overall magnitude response (b)
Magnitude response in the passband (c) Group delay response in the passband
REFERENCES
[1] J. H. McClellan, T. W. Parks and L. R. Rabiner, A Computer program
for designing optimum FIR linear phase digital lters, IEEE Trans. Audio
Electroacoust., vol. AU-21, no.6, pp.506-526, Dec. 1973.
[2] T. W. Parks and J. H. McClellan, Chebyshev approximation for nonre-
cursive lters with linear phase, IEEE Trans. Circuit Theory, vol. CT-19,
no.2, pp.189-194, Feb. 1972.
[3] A. A. -Taleb, and Moustafa M. Fahmy, Design of FIR Tow-Dimensional
International Journal of Information and Communication Engineering 6:1 2010
11
TABLE II
PERFORMANCE COMPARISON WITH [18] (EX. 2)
Previous SP [18] Proposed Method
(n, m) (6,12) (7,12) (8,12) (9,12) (8,12)
p 0.0132 0.0100 0.0079 0.0062 0.0131
s 0.0132 0.0132 0.0132 0.0132 0.0131
0.8644 0.7579 0.4804 0.4223 0.2500
Maximum pole radius 0.9463 0.9433 0.9596 0.9599 0.9583
Digital Filters by Successive Projections, IEEE Trans. Circuits and
Systems, vol. CAS-31, no.9, pp.801-805, Sept. 1984.
[4] R. Ito, T. Fujie, K. Suyama and R. Hirabayashi, Design method of
FIR lters with signed power of two coefcients using a new linear
programming relaxation with triangle inequalities, International Journal
of Innovative Computing, Information and Control, vol. 2, no. 2, pp. 441-
448, April 2006.
[5] M. Akazawa, and M. Ikehara, Simultaneous Approximation of Magnitude
and Group Delay in FIR Digital Filters, IEICE Trans., vol.J87-A, no. 11,
pp. 1395-1402, Nov. 2004.
[6] Z. Lin and Y. Liu, Design of Complex FIR Filters With Reduced Group
Delay Error Using Semidenite Programming, IEEE Signal Processing
Letters, vol.13, no.9, Sept. 2006.
[7] A. G. Deczky, Synthesis of recursive digital lters using the minimum
p-error criterion, IEEE Trans. Audio Electroacoust., vol. AU-20, pp. 257-
263, Oct. 1972.
[8] W.-S. Lu, S.-C. Pei, and C.-C. Tseng, A weighted least-squares method
for the design of stable 1-D and 2-D IIR digital lters, IEEE Trans. Signal
Processing, vol. 46, pp. 1-10, Jan. 1998.
[9] X. Chen and T.W. Parks, Design of IIR lters in the complex domain,
IEEE Trans. Acoust., Speech and Signal Process., vol. 38, pp. 910-920,
June. 1990.
[10] C.-C. Tseng, Stable IIR diferentiator design using iterative quadratic
programming approach, Signal Process., vol. 80, no. 5, pp. 857-866, 2000.
[11] W.-S. Lu, Design of recursive digital lters with prescribed stability
margin: a parameterization approach, IEEE Trans. Circuits and Syst. II,
vol. 45, pp. 1289-1298, Sept. 1998.
[12] W.-S. Lu, Design of stable minimax IIR digital lters using semidenite
programming, in Proc. IEEE Int. Symp. Circuits and Syst., vol. 1, Geneva,
Switzerland, pp. 355-358, May 2000.
[13] W.-S. Lu, A unied approach for the design of 2-D digital lters via
semidenite programming, IEEE Trans. Circuits and Syst. I, Fundam.
Theory Appl., vol. 49, no. 6, pp. 814-826, June 2002.
[14] S.C. Chan and K. M. Tsui, On the design of real and complex FIR lters
with atness and peak error constraints using semidenite programmingin
Proc. IEEE int. Symp. Circuits and Syst., vol. III, pp. 125-128, May 2004.
[15] Y. Sugita and N. Aikawa, Design Method of FIR Digital Filters with
Specied Group Delay Errors using Successive Projection, International
Journal of Innovative Computing, Information and Control, vol. 4, no. 2,
pp. 445-455, Feb. 2008.
[16] S. Fukae, N. Aikawa and M. Sato, Complex Chebysev approximation
using successive projections method, IEICE Trans., vol.J80-A, no.7,
pp.1192-1196, July 1997.
[17] S-. C. Pei and I-. I. Yang, Design of a class of time-constrained FIR
digital lters by successive projections, IEEE Trans. Circuits and Systems,
vol. 36, no. 1, pp 164-167, Jan. 1989.
[18] Y. Sugita and N. Aikawa, Design Method of Stable IIR Digital Filters,
The Transactions of the Institute of Electronics, Information and Com-
munication Engineers A, vol.J85-A, no. 6, pp. 735-740, June 2002.
[19] Y. Sugita and N. Aikawa, A direct design method of minimum phase FIR
Filters, IEICE Trans., Fundamentals, vol.E87-A, no. 10, pp. 1284-1292,
Oct. 2004.
International Journal of Information and Communication Engineering 6:1 2010
12