International Journal of Electronics, Communications and Computer Engineering 1:2 2009

Design of Digital IIR filters with the Advantages of

Model Order Reduction Technique
K.Ramesh, A.Nirmalkumar, and G.Gurusamy

Abstract—In this paper, a new model order reduction detail. This enables the modeling of coupled complex
phenomenon is introduced at the design stage of linear phase digital phenomena in a reasonable time. First of all, these models
IIR filter. The complexity of a system can be reduced by adopting the should be smaller than the original model, meaning that they
model order reduction method in their design. In this paper a mixed are computationally less demanding. Their behavior must in at
method of model order reduction is proposed for linear IIR filter. The least one way be comparable to the behavior of the original
proposed method employs the advantages of factor division technique
to derive the reduced order denominator polynomial and the reduced
model. Preferably, the smaller models are physically as well as
order numerator is obtained based on the resultant denominator mathematically motivated. This makes the models
polynomial. The order reduction technique is used to reduce the delay interpretable and efficient. The host of mathematical processes
units at the design stage of IIR filter. The validity of the proposed to derive smaller models forms the field of what we call
method is illustrated with design example in frequency domain and Model Order Reduction (MOR). The reason for the name is
stability is also examined with help of nyquist plot. that the model is reduced in size by some technique. The
model is assumed to be already present, being derived by
Keywords—Error index (J), Factor division method, IIR filter, physical laws and assumptions. Sometimes the methods are
Nyquist plot, Order reduction. abusively called Reduced Order Modeling., which should then
account for the task of modeling in such a way that the derived
I.INTRODUCTION models are smaller right away.

A N important step in the development of a digital filter is

the determination of a realizable transfer function G(z)
approximating the given frequency response specifications. If
In many cases it is very well possible to reduce the size of a
model. We already saw that mathematical rules of
discretization may lead to models which are too large and
an IIR filter is desired, it is also necessary to ensure that G(z) have too much detail for the required precision. Also the final
is stable. The process of deriving the transfer function G(z) is output of the model has some flexibility in the required
called digital filter design. After G(z) has been obtained, the accuracy. In applications a small error is often admissible.
next step is to realize it in the form of a suitable filter This flexibility gives room for smaller approximations and
structure. There are two major issues that need to be answered faster methods. Furthermore, we are often only interested in
before one can develop the digital transfer function G(z). The certain states of the model, for instance only in the output due
first and foremost issue is the development of a reasonable to a certain input. This, while the model contains information
filter frequency response specification from the requirements of inputs that are never imposed. Model Order Reduction tries
of the overall system in which the digital filter is to be to quickly capture the essential features of a structure. This
employed. The second issue is to determine whether an FIR or means that in an early stage of the process, the most basic
IIR filter is to be designed. The order reduction technique is properties of the original model must already be present in the
then introduced to obtain an equivalent reduced IIR linear smaller approximation. At a certain moment the process of
phase filter. reduction is stopped. At that point all necessary properties of
The mathematical models obtained from the theoretical the original model must be captured with sufficient precision.
criteria those relevant with the given system normally results The order reduction of a linear time invariant system is
as a higher order model. Such systems are difficult to be applied in almost all fields of electrical engineering. The use
analyzed and also the design of controller becomes too of order reduced models for test simulations of complex
difficult. Obviously, there is a need for smaller models. systems is a lot easier than utilizing full order models. This is
Smaller models are models that describe the behavior of due to the fact that the lower order transfer function can be
a system accurately, without the disadvantage of unnecessary analyzed more easily. Therefore, order reduction algorithms
are standard techniques in the integrated circuits community
for analysis, approximation and simulation of models arising
from interconnect and electromagnetic structure analysis.
K.Ramesh is Lecturer with the Department of Electrical and Electronics Conventional reduction methods are available in literature
Engineering, Velalar College of Engineering and Technology, Erode,
for reducing the order of large-scale linear MIMO systems in
Tamilnadu, India-638012. (e-mail: [email protected])
A.Nirmalkumar is Professor and Head with the Department of Electrical frequency domain [1–4]. Some of the existing methods are
and Electronics Engineering, Bannari Amman institute of Technology, based on the global properties of G(s) have been proposed in
Sathyamangalam, Tamilnadu, India-638401. the literature amongst which the frequency domain method
G.Gurusamy is Dean with the Department of Electrical and Electronics
using the continued-fraction expansion of G(s) and the
Engineering, Bannari Amman institute of Technology, Sathyamangalam,
Tamilnadu, India-638401. classical Pade approximations are well known for their
computational simplicity and matching of few initial time-

117
 International Journal of Electronics, Communications and Computer Engineering 1:2 2009

moments. However, stability of such reduced-models is not

1- δp ≤ G(e jω ) ≤ 1+ δp , for |ω| ≤ ωp (1)
guaranteed even though the original system is stable. Further,
some mixed methods have been suggested by combining the
algorithm of two different reduction methods [5–7]. In spite of In the stop band, defined by ωs ≤ ω ≤ π, we require that
having several reduction methods, none always gives the the magnitude approximants zero with an error of δs , i.e.,
satisfactory results for all the systems.
In last decades, papers are came out in model order
G(e jω ) ≤ δs , for ωs ≤ ω ≤ π. (2)
reduction based on concepts of dominant pole retention or
dominant eigen value retention, Routh approximation,
stability-equation method. An alternative approach suggested The frequencies ωp and ωs are respectively called the pass
by a number of researchers is to use the methods belonging to band edge frequency and the stop band edge frequency. The
the second category mentioned above to obtain better reduced- limits of the tolerances in the pass band and stop band, δp and
models and extend the applicability of the methods. In this δs are usually called the peak ripple values. Note that the
connection the works of Shamash [8], Chen, et al [9], [10], frequency response G(e jω ) of digital filter is an even function
Pal, et al [11] and Gupta, et al [12], are worth mentioning. S K of ω. As a result, the digital filter specifications are given only
Bhagat [13] proposed some new mixed methods for the for the range 0 ≤ ω ≤ π.
simplification of single-input single-output systems of Digital filter specifications are often given in terms of the
Stability preserving methods, such as, (g-d) canonical loss function, A(ω)= -20log10 G (e jω ) , in dB. Here the peak
expansion.
In the proposed method, a mixed method is proposed for pass band ripple αp and the minimum stop band attenuation αs
the reduction of linear phase digital IIR filters. The obtained are given in dB, i.e., the loss specification of a digital filter are
filter transfer function as per given design specification is given by
converted in to transfer function in‘s’ domain before let it in to
model order reduction stage. The denominator polynomial of αp = -20log10 (1- δp ) dB
reduced order model is obtained by using the factor division αs = -20log10 (δs ) dB (3)
method and numerator polynomial is obtained by proposed
method. The reduced order model obtained by the proposed The pass band and stop band edge frequencies, in most
method will preserves the stability of the original system. The applications, are specified in Hz, along with the sampling rate
proposed method is explained in the following sections. of the digital filter. Since all filter design techniques are
developed in terms of normalized angular frequencies ωp and
II. DIGITAL IIR FILTER SPECIFICATIONS ωs , the specified critical frequencies need to be normalized
before a specific filter design algorithm can be applied. Let FT,
As in the case of an analog filter, either the magnitude FP and FS denote respectively the sampling frequency, the pass
and/or the phase (delay) response is specified for the design of band and stop band edge frequencies respectively.
a digital filter for most applications. In some situations, the
unit sample response or the step response may be specified. In Ωp 2πFP
most practical applications, the problem of interest is the ωP = = (4)
development of a realizable approximation to a given FT FT
magnitude response specification. The phase response of the
designed filter can be corrected by cascading it with an all Ω S 2πFS (5)
pass section. In this paper, we restrict our attention to the ωs = =
FT FT
magnitude approximation problem only.
In the case of IIR filter design, the most common practice is
For the design of an IIR digital low pass filter G(z) based on
to convert the digital filter specifications in to analog low pass
the conversion of an analog low pass filter Ha(s), the filter
prototype filter specifications, to determine the analog low
order of Ha(s) is first estimated from its specifications. The
pass filter transfer function Ha(s) meeting these specifications,
order (n) of butter worth low pass filter is obtained from the
and then to transform it into the desired digital filter transfer
equation
function G(z). This approach has been widely used for many
reasons such as the extensive tables are available, yields
closed-form solutions and they are highly advanced. log 10 (1 / k 1 ) (6)
n=
The magnitude response specifications of a digital filter in log 10 (1 / k )
the pass band and in the stop band are given with some Ωp ε
acceptable tolerances. In addition, a transition band is Where, k = and k1 =
specified between the pass band and the stop band to permit Ωs A2 − 1
the magnitude to drop off smoothly.
Consider the low pass filter with the magnitude of G(ejω) , in The resultant IIR filter transfer function in ‘Z’ domain is
used for reduction of IIR filter complexity.
the pass band defined by 0 ≤ ω ≤ ωp, we require that the
magnitude approximates unity with an error ±δp, i.e.,

118
 International Journal of Electronics, Communications and Computer Engineering 1:2 2009

III. STATEMENT OF PROBLEM further in second step based on the resultant denominator
th polynomial.
Consider an n order linear phase digital IIR filter denoted
in Z domain as,
Step 1
Obtaining the denominator polynomial from the given
X (z) A 0 + A1 z + A 2 z 2 + ........ + A m z m (7)
G(z) = = system transfer function
Y (z) B 0 + B 1 z + B 2 z 2 + ......... + B n −1 z n −1 + B n z n
The nth order original system given in equation (8) is
While applying the proposed model order reduction equated to the rth order reduced model represented by the
technique, the linear transformation technique (Z=S+1) is equation (9)
applied to obtain the transfer function in‘s’ domain.
a 0 + a1 s + a 2 s 2 + ........ + a n −1 s n −1 =
N ( s) a0 + a1 s + a 2 s 2 + ........ + a m s m (8) b0 + b1 s + b2 s 2 + ......... + bn −1 s n −1 + bn s n
G(s) = =
D( s) b0 + b1 s + b2 s 2 + ......... + bn −1 s n −1 + bn s n
d 0 + d1 s + d 2 s 2 + ........ + d r −1 s r −1 (10)
m e0 + e1 s + e2 s 2 + ......... + er −1 s r −1 + er s r
∑a
k =0
k sk
On cross multiplying and rearranging the equation (10)
= n

∑b s
k =0
k
k
a0e0 + (a0e1 + a1e0 )s + (a0e2 + a1e1 + a2e0 )s 2 + .........+ an −1ek s n −1+ r =
b0 d 0 + (b0 d 1 + b1 d 0 ) s + ............... + bn d r −1 s n −1+ r (11)
where ak = [a ]k
ij pxq i= 1, 2,….p
By equating the coefficients of the same power of‘s’ on
j= 1, 2,….q both sides in the equation and simplifying the relations, the
k= 0, 1, 2, … m. reduced order denominator polynomial is obtained as,
and bk = a 0, k (k= 0, 1, 2, ….. n-1).
Dr (s ) = e0 + e1s + e2 s 2 + .............+ er −1s r−1 + er s r (12)
The corresponding rth (r<n) order reduced order is of the form

N ( s) d 0 + d1 s + d 2 s 2 + ........+ d r −1 s r −1 (9) Step 2

Gr(s) = = r −1 Consider the given system transfer function given in (8)
D(s) e0 + e1 s + e2 s + .........+ er −1 s
2
+ er s r

G(s) = a 0 + a1 s + a 2 s 2 + ........ + a m s m
r −1 b0 + b1 s + b 2 s 2 + ......... + b n −1 s n −1 + b n s n
=
∑d
k =0
k s k

(1 + A1s + A2 s 2 + ........ + Am s m ) ;n≥m (13)

r
=K
∑
k =0
ek s k
(1 + B1s + B2 s 2 + ......... + Bn −1s n −1 + Bn s n )

where d k = a ij [ ] k
pxq
i= 1, 2,….p Where, K=
a0
b0
,
j= 1, 2,….q
k= 0, 1, 2, … r-1. ai
Ai = ; i = 0,1,2,3…m and
a0
and ek (k= 0, 1, 2, ….. r-1,r)
bj
Bj = ; j = 0,1,2,3…n
Inverse linear transformation(s=z-1) is then applied at b0
equation (9) to form the reduced order linear phase IIR filter.
In this paper, the obtained reduced model retains the important Let the transfer function of the approximating low-order
characteristics of the original digital IIR filter system and system be,
approximates its response as closely as possible for the same
type of inputs. (1 + C 1 s + C 2 s 2 + ........ + C q s q ) (14)
Gr(s) = K
(1 + D1 s + D 2 s 2 + ......... + D p −1 s p −1 + D p s p )
IV. PROPOSED METHOD where n≥p≥q
The reduced order obtained by the proposed method
consists of two steps. In first step, the reduced order The following relation should be satisfied as closely as
denominator polynomial is obtained by using the method as possible
described in [14]. The numerator polynomial is obtained

119
 International Journal of Electronics, Communications and Computer Engineering 1:2 2009

| G ( jω ) | 2 =1 for 0≤ω≤∞ V. DESIGN EXAMPLE

(15)
| G r ( jω ) | 2 Consider the digital IIR filter with the transfer function in
direct form as H(z) in equation (25) and the general structure
2 m 2 p−1
G(s) (1+ A1s + A2s +........+ Ams )(1+ D1s + D2s +.........+ Dp−1s + Dps )
p of lattice-ladder IIR filter is shown in fig.1.
=
Gr (s) (1+ B1s + B2s2 +.........+ Bn−1sn−1 + Bnsn)(1+ C1s + C2s2 +........+ Cqsq)
1 + 2 z −1 + 2 z −2 + z −3 (25)
= (1 + m 1 s + m 2 s 2 + ........ + m u s u ) (16) H ( z) =
v −1 1 + 0.54167 z −1 + 0.625 z −2 + 0.333−3
(1 + l 1 s + l 2 s + ......... + l v −1 s
2
+ lv s )
v

z 3 + 2z 2 + 2z + 1 (26)
Where, u=m+p and v=n+q =
z 3 + 0.54167 z 2 + 0.625z + 0.333
| G ( jω ) | 2 G ( s )G (− s )
= (17)
| G r ( jω ) | 2
G r ( s )G r ( − s ) s = jω 1.45833z 2 + 1.375z + 0.667
=1+ (27)
z + 0.54167z 2 + 0.625z + 0.333
3

The equation produces the even powers of‘s’ and can be

written as,
X(n) fN-1(n) f2(n) f1(n) f0(n)
| G ( jω ) | 2 1 + L 2 s 2 + L 4 s 4 + ........ + L 2 u s 2 u (18) + + +
= fN(n)
| G r ( j ω ) | 2 1 + M 2 s 2 + M 4 s 4 + ........ + M 2 v s 2 v s = jω
kN
k2 k1

- kN -k2 -k1
| G( jω) |2 (L − M2 )s2 + (L4 − M4 )s4 + ........+ (L2u − M2v )s2u (19) g1(n)
= 1+ 2 g2(n)
| Gr ( jω) |2
1 + M2s2 + M4s4 + ........+ M2v s2v s = jω
gN(n) + z-1 + Z-1 + Z-1

To satisfy the above equation, CN CN-1 C2 C1 C0

L2 = M2
L4 = M4
. + + + +

.
. Fig.1 Structure of Lattice-ladder IIR filter oh Nth order
L2u = M2v ; if u=v (20)
Let,
If u<v, then the error generated by the lower order model is, 1.45833 z 2 + 1.375 z + 0.667 (28)
G( z) =
z 3 + 0.54167 z 2 + 0.625 z + 0.333
| G ( jω ) | 2 (21)
ε = −1
| G r ( jω ) | 2
By applying the linear transformation (Z=S+1), the
On comparing the equations (10) and (11), we get transfer function G(z) becomes G(s) and is given by,

x −1 1.45833s 2 + 4.29166s + 3.50033

L2 x = ∑ (−1) i 2mi m2 x −i + (−1) x m x G ( s) = (29)
2
(22)
i =0
s 3 + 3.54167 s 2 + 4.70834s + 2.49967

for x= 1,2, 3……..u and m0=1

and By applying the step-1 as stated in the proposed method, the
y −1
reduced order denominator polynomial is obtained as,
M 2 y = ∑ (−1) i 2li l 2 y −i + (−1) x l y
2
(23)
i =0 Dr ( s ) = 0.20499 s 2 + 0.43244 s + 0.333 (30)

for y= 1,2, 3……..v and l0=1 With help of denominator polynomial Dr (s ) , the numerator
From the equation (13), the reduced order model polynomial is obtained from the proposed method as,
numerator coefficients can be obtained. Finally the reduced
order model is in the form of N r ( s) = 0.2995s + 0.467 (31)

(1 + C 1 s + C 2 s 2 + ........ + C q s q ) (24) On applying inverse linear transfer function, the reduced

Gr(s) = K p −1 order transfer function in ‘z’ domain is obtained as,
(1 + D1 s + D 2 s + ......... + D p −1 s
2
+ Dps )p

120
 International Journal of Electronics, Communications and Computer Engineering 1:2 2009

0.205 z 2 + 0.322 z + 0.2731 (32)

H r ( z) =
0.205 z 2 + 0.02246 z + 0.1056

The results were compared in both continuous and as well

as in discrete domain and are shown in fig.2 and fig.3
respectively. The error index (J) between the original and
reduced order filter is obtained as 0.093.

Fig.4 Frequency responses of original and reduced order IIR filters

Fig.2 Comparison of Step responses

Fig.5 Nyquist plot of original and reduced order IIR filters

The reduced order transfer function H r (z ) is also can be

Fig.3 Comparison of original and reduced order IIR filters in discrete stated as,
instants
0 .205 + 0 .322 z −1 + 0 .2731 z − 2 (33)
H r ( z) =
The closeness of the higher order filter with reduced order 0 .205 + 0 .02246 z −1 + 0 .1056 z − 2
IIR filter in frequency domain is shown in fig.4. The reduced
order IIR filter preserves closeness with original IIR filter and The direct forms of reduced order IIR filter can be
also preserves the closed loop stability. The reduced order IIR represented in reduced order Lattice-ladder IIR filter form
filter has the poles at -0.0548 + 0.7156i and -0.0548 – 0.7156i. some conversions. Consider the direct form of Lattice-ladder
There is no pole at right half of the s-plane and hence the IIR filter in the form of
reduced order system is stable. The stability criterion is hence
proved with the help of Nyquist plot and is shown in fig.5. bM ( z) 0 .205 + 0 .322 z −1 + 0 .2731 z −2 (34)
H r ( z) = =
a n ( z ) 0 .205 + 0 .02246 z −1 + 0 .1056 z − 2

121
 International Journal of Electronics, Communications and Computer Engineering 1:2 2009

[6] R. Prasad, A. K. Mittal, and S. P. Sharma, “A mixed method for the

reduction of multi-variable systems,” Journal of the Institution of
From the equation (34), the lattice-ladder filter co-efficient Engineers, vol. 85, pp. 177–181, 2005.
are obtained as a2(0) = 0.205, a2(1) = 0.02246, [7] L. Shieh and Y.Wei, “A mixed method for multivariable system
a2(2)=k2=0.1056, b2=0.205, b1=0.322 and b0=0.2731.The co- reduction,” IEEE Transactions on Automatic Control, vol. 20, no. 3, pp.
efficient ‘c’ is obtained with the equation of, 429–432, 1975.
[8] Y Shamash,” Model Reduction using the Routh Stability Criterion and
the Pade Approximation Technique”, International Journal of Control,
M

∑ C a (i − m )
vol 21, no 3, pp 475-484, 1975.
C m = bm − i i
; m= M, M-1,…. 0. (35) [9] T C Chen, C Y Chang, and K W Han, “Model Reduction using Stability
i = m +1 equation Method and the Pade Approximation Method”, Journal of
Frankline Institute, vol 309, 1980, pp 473-490.
and [10] T C Chen, C Y Chang, and K W Han, “Model Reduction using Stability
equation Method and Continued Fraction Method”, International Journal
of Control, vol 32, no 1, 1980, pp 81-94.
a m (k ) − a m (m )a m (m − k ) (36) [11] J Pal, A K Sinha, and N K Sinha, “Reduced-order Modeling using Pole
a m −1 ( k ) = Clustering and Time-moment Matching”, Journal of The Institution of
1 − a m2 ( m )
Engineers (India), Pt El, vol 76, 1995, pp 1-6.
[12] D K Gupta, S K Bhagat and J P Tewari, “A Mixed Method for the
The value of a parameter k1 is obtained from the above Simplification of Linear Dynamic Systems,” in Proceedings of
equation as k1 = a1(1) = 0.02. The reduced order structure of International Conference on Computer Applications in Electrical
Engineering Recent Advances, IIT, Roorkee, February 21-23, 2002, pp
lattice-ladder IIR filter is shown in fig.6. 455-459.
[13] S K Bhagat, J P Tewari and T Srinivasan, “Some Mixed Methods for the
Simplification of High-order Single-input Single-output Systems,” IE (I)
f1(n) f0(n)
+ + Journal.EL, Vol 85, pp.120-123, Sep 2004.
[14] K.Ramesh, A.Nirmalkumar and G.Gurusamy, “Order reduction by error
x(n)
k2= 0.1056 k1=0.02 minimization technique,” Proceedings of International Conference on
Computing, Communication and Networking 2008, Dec 2008, pp 1-6.
-k1
-k2
g1(n)
g2(n)
+ Z-1 + Z-1

C2 =0.205 C1 = 0.317 C0 = 0.311 Ramesh.K received B.E. degree in Electrical and

Electronics Engineering from K.S.R.College of
Technology, Tamilnadu, India in 2002 and M.E. in
y(n)
+ + + Applied Electronics from Kongu Engg. College, Anna
University, Chennai in 2005 and also received the
MBA in System from PRIDE, Salem (Tamilnadu),
Fig.6 Structure of Reduced order Lattice-ladder IIR filter India in 2005. Since then, he is working as a Lecturer
in Velalar College of Engineering and Technology
VI. CONCLUSION (Tamilnadu), India. Presently he is a Part time
(external) Research Scholar in the Department of Information and
This paper proposes the novel model order reduction for the Communication at Anna University, Chennai (India). His fields of interests
design of IIR filters. The reduced order model results in include Model order reduction, Controller design and Optimization
reduction in number of delay units and other components in Techniques.
the filter design stage. The reduced order model has closeness Dr.Nirmalkumar.A, received the B.Sc.(Engg.) degree
with respect to the higher order system in their performance from NSS College of Engineering, Palakkad in 1972,
characteristics. The proposed method is mathematically simple M.Sc.(Engg.) degree from Kerala University in 1975 and
completed his Ph.D. degree from PSG Tech in 1992.
and also applied to the large scale systems with substantial Currently, he is working as a Professor and Head of the
merits over other conventional methods. The simulation Department of Electrical and Electronics Engineering in
results show the good approximation and stability criterion of Bannari Amman Institute of Technology,
the reduced order model. Sathyamangalam, Tamilnadu, India. His fields of Interest
are Power quality, Power drives and control and System
optimization.
REFERENCES
Dr.G.Gurusamy obtained his Pre-University education
[1] R. Prasad, J. Pal, and A. K. Pant, “Multivariable system approximation at St Johns College, Palayamkottai, and Thirunelveli
using polynomial derivatives,” Journal of the Institution of district. He then joined PSG College of Technology,
Engineers, vol. 76, pp. 186–188, 1995. Coimbatore, in the year 1962 to pursue his Engineering
[2] Y. Bistritz and U. Shaked, “Minimal Pade model reduction for course. He was graduated in Electrical Engineering in
multivariable systems,” Journal of Dynamic Systems, Measurement and 1967. He latter obtained M.E (Applied Electronics) in
Control, vol. 106, no. 4, pp. 293–299, 1984. 1972 and Ph.D in Control Systems in 1983. He is
[3] C. F. Chen, “Model reduction of multivariable control systems by means currently working as a Dean of Department of Electrical
of matrix continued fractions,” International Journal of Control, vol. 20, and Electronics Engineering in Bannari Amman Institute
no. 2, pp. 225–238, 1974. of Technology, Sathyamangalam. His fields of Interest
[4] M. R. Calfe and M. Healey, “Continued fraction model reduction are Advanced control, Digital control, Optimization and Biomedical
technique for multivariable systems,” Proceedings of the Institution of electronics.
Electrical Engineers, vol. 121, no. 5, pp. 393– 395, 1974.
[5] R. Prasad, “Pade type model order reduction for multivariable systems
using routh approximation,” Computers & Electrical Engineering, vol.
26, no. 6, pp. 445–459, 2000.

122