SOP TRANSACTIONS ON APPLIED MATHEMATICS

VOLUME 1, NUMBER 1, March 2014

SOP TRANSACTIONS ON APPLIED MATHEMATICS

MIMO System Using Eigen Algorithm and

Improved Pade Approximations
Jay Singh1 *, Kalyan Chattterjee2 , C. B. Vishwakarma3
1 Indian School of Mines Dhanbad, Jharkhand, India
2 Dept. of Electrical Engg, Indian School of Mines, Dhanbad, Jharkhand, India
3 Dept. of Electrical Engg, Galgotias College of Engg. & Tech. Greater. Noida, U. P. India

*Corresponding author: jaysinghism@gmail.com

Abstract:
A mixed method by combining a Eigen algorithm and Improved pade approximations is proposed
for reducing the order of the large-scale multi-input multi-output (MIMO) systems. The most
dominant Eigen value of both original and reduced order systems remain same in this method.
The proposed method guarantees stability of the reduced model if the original high-order system
is stable and is comparable in quality with the other well known existing order reduction methods.
The viability of the proposed method is shown through numerical examples taken from the
literature and compared with few well-known order reduction methods.
Keywords:
Eigen algorithm; Order reduction; improved pade approximations; Stability; Transfer function

1. INTRODUCTION

Every physical system can be translated into a mathematical model. The mathematical procedure
of system modeling often leads to comprehensive description of a process in the form of high order
differential equations which are difficult to use either for analysis or controller synthesis. Hence, it is
useful and sometimes necessary to find the possibility of finding some equation of the same type but of
lower order that may be considered to adequately reflect the dominant characteristics of the system under
consideration. Some of the reasons for using reduced order models of high order linear systems could be :
(i) To have a better understanding of the system.
(ii) To reduce computational complexity.
(iii) To reduce hardware complexity.
(iv) To make feasible controller designs.
(v) To reduce online processing real time.
Numerous methods are available in the literature for order reduction of linear continuous systems in
time domain as well as in frequency domain [1–4]. Further, several methods have also been suggested
by combining the features of two different methods [5–7]. Pal et. al. proposed [8] pole-clustering using
Inverse Distance Criterion and time moment matching. Vishwakarma [9] modified the pole clustering by
an iterative method, the complexity with these methods is in selecting poles for the clusters. A method
based on Eigen Spectrum Analysis suggested by Mukherjee [10], in which both the pole centroid and

60
SOP TRANSACTIONS ON APPLIED MATHEMATICS

system stiffness of the original and reduced order systems are kept exactly same to obtain the reduced
order system. Parmar et al [11] proposed a mixed method using Eigen Spectrum Analysis with Factor
division algorithm to determine the numerator of the reduced model with known denominator. In some
cases, the difficulty with these methods [10, 11] are tendency to become non-minimum phase due to
equalization of system stiffness. Every methods have their advantages and disadvantages when tried on a
particular system. Major number of methods available in the literature but no approach constantly gives
the best results for all systems. Approximately all methods having aim to get an accurate reduced models
for a low computational cost.
In the present work, authors have directly taken poles from the proposed Eigen algorithm while the
zeros are taken from the improved pade approximations to obtain the reduced order system. Proposed
method reduces the problem of non-minimum phase in the reduced models. The reduction procedure is
simple and computer oriented. It is shown that the method has several advantages e.g., good approximation
of reduced models retain the steady state value, transient value and stability of the original system.
In subsequent, the proposed algorithm is described in detail with the help of numerical examples.
The paper is organized as following manner. In Section 2, the proposed algorithm and improved pade
approximations [12] are described. In Section 3, two numerical examples have been presented with
computer simulation results. Finally performance indices, qualitative comparisons of methods and
conclusions are prepared in Sections 4 and 5 respectively.

2. DESCRIPTION OF THE METHOD

Let the transfer function of high order multivariable system of the order ’n’ be :
 
a21 (s) a22 (s) · · · a2p (s)
a31 (s) a32 (s) · · · a3p (s) 
 
1 
[Gn (s)] =  . .. ..  (1)
 ..
Dn (s)  
. . 
aq1 (s) aq2 (s) · · · aqp (s)

where, Dn (s) = a11 + a12 s + a13 s2 + . . . + a1,n+1 sn , p and q are number of input and output variables
respectively.
Let the transfer function of the reduced system with the same number of inputs and outputs of the order
’k’ be:
 
b21 (s) b22 (s) . . . b2p (s)
b31 (s) b32 (s) . . . b3p (s) 
 
1 
[Rk (s)] = .. (2)
 
.. ..
Dk (s) 
 
. . .

 
bq1 (s) bq2 (s) . . . bqp (s)

where, Dk (s) = b11 + b12 s + b13 s2 + . . . + b1,k+1 sk .

The objective of this paper is to realize the reduced order models in the form of (2) from the original
system (1), such that it retains the significant features of the original high order system.

2.1 Reduction procedure

The reduction procedure for getting reduced models consist of the following two steps:

61
 MIMO System Using Eigen Algorithm and Improved Pade Approximations

Figure 1. Proposed Eigen Algorithm

Step 1: Determination of the denominator polynomial for the kth - order reduced model using Eigen
Algorithm.
The following are the simple rules for making Eigen Algorithm.
1) Firstly retain the one most dominant real pole.
2) Make a clusters of real poles and complex poles separately.
3) Clusters of the poles in the left half s-plane should not contain any pole of the right half s-plane and
vice-versa.
The interactive algorithm for synthesizing the reduced denominator polynomial using Eigen Algorithm
is stated as following.

2.1.1 Polynomial consisting of only real poles

Let there be n real poles are p1 , p2 , p3, . . . , pn where |p1 | < |p2 | < |p3 | . . . < |pn | .
Estimate the real Eigen values for reduced denominator polynomial as follows :

pe1 = |p1 | (3)

62
SOP TRANSACTIONS ON APPLIED MATHEMATICS

" #
n
1
pe2 = ∑ Repi , i = 1, 2, ...n (4)
Repn i=1

Now using (3,4), we can write reduced denominator polynomial as

Dk (s) = (s + pe1 )(s + pe2 ) (5)

2.1.2 Polynomial consisting of only complex poles

Let ’m’ pair of complex conjugate poles are [(Reλ1 ± Imδ1 ) , (Reλ2 ± Imδ2 ) , . . . (Reλm ± Imδm )]
where,
|λ1 | < |λ2 | ......... < |λm |

Estimate the complex Eigen values for reduced denominator polynomial as follows :

Reλe1 = Re |λ1 | , i f Reλ1 is repeating (6)

" #
m
1
Reλe1 = ∑ Reλ j , j = 1, 2, ...m, i f Reλ1 is not repeating (7)
Reλm j=1
" #
m
1
Imδe1 = ∑ Imδ j , j = 1, 2, ...m (8)
2Imδm j=1

Now using (6,7,8), we can write reduced denominator polynomial as:

Dk (s) = (s + Reλe1 + Imδe1 )(s + Reλe1 − Imδe1 ) (9)

Hence the denominator polynomial Dk (s) is obtained as

Dk (s) = do + d1 s + d2 s2 + ...dk sk (10)

Step 2: Determination of the numerator of the reduced model using improved Pade approximations
[12].
The original high order system can be expanded in power series
∞
G(s) = ∑ Mi s−i−1 (about s = ∞) (11)
i=0
∞
G(s) = − ∑ Ti si (about s = 0) (12)
i=0

Where Mi and Ti are the ith - Markov parameter and Time moment of G(s) respectively. The ith - order
reduced model is taken as:

k−1
∑ ci si
Nk (s) i=0
Rk (s) = = (13)
Dk (s) k
∑ di si
i=0

The coefficients of the numerator Nk (s) may be obtained by using the following set of equations.
63
 MIMO System Using Eigen Algorithm and Improved Pade Approximations


co = do To 


c1 = d0 T1 + d1 T0





c2 = d0 T2 + d1 T1 + d2 T0 



...





...





cα−1 = d0 Tα−1 + d1 Tα−2 + ... + dα−2 T1 + dα−1 T0 
(14)
ck−β = dk Mβ −1 + dk−1 Mβ −2 + ... + dk−β +2 M1 + dk−β +1 M0 


ck−β +1 = dk Mβ −2 + dk−1 Mβ −3 + ... + dk−β +3 M1 + dk−β +2 M0





...





... 



ck−2 = dk M1 + dk−1 M0





ck−1 = dk M0


the coefficients c j ; j = 1, 2...(k − 1) can be found by solving the above ’k’ linear equations. Hence, the
numerator Nk (s) is obtained as:

Nk (s) = c0 + c1 s + c2 s2 + ... + ck−1 sk−1 (15)

3. NUMERICAL EXAMPLES

Two numerical examples have been taken from the literature to illustrate the algorithm of the proposed
method. The examples are solved in details to get second order reduced model. An integral square error
(ISE) is calculated between the transient parts of the original and reduced model using MATLAB to
measure the goodness of the reduced model i.e. lower the ISE closer the Rk (s) to G(s) which is given by :

Z∞
ISE = [gi j (t) − ri j (t)]2 dt (16)
0

Z∞  
IAE = gi j (t) − ri j (t) dt (17)
0

where, gi j (t) and ri j (t) are the unit step responses of original and reduced system respectively.
Example 1: Consider a 6th order multivariable system from Bistriz and Shaked [13] described by the
transfer function matrix as:

2(s+5) (s+4)
" # " #
(s+1)(s+10) (s+2)(s+5) 1 a11 (s) a12 (s)
[G(s)] = (s+10) (s+6) =
(s+1)(s+20) (s+2)(s+3)
D(s) a21 (s) a22 (s)

where,

D(s) = s6 + 41s5 + 571s4 + 3491s3 + 10060s2 + 13100s + 6000

a11 (s) = 2s5 + 70s4 + 762s3 + 3610s2 + 7700s + 6000

64
SOP TRANSACTIONS ON APPLIED MATHEMATICS

Table 1. Comparison of The Proposed Method With Others For Example 1

ISE for ri j (s)[i = 1, 2; j = 1, 2]
Reduced Model r11 r12 r21 r2
Proposed Methods 7.862 × 10−4 7.983 × 10−5 7.012 × 10−5 0.003876
C.B. Vishwakarma [14] 0.03605 3.12x10-4 0.01033 0.028940
Prasad et al.[15] 0.030689 0.000256 0.261963 0.021683
Safonov and Chinang [16] 0.590617 0.037129 0.007328 1.066123
Prasad and Pal [17] 0.136484 0.002446 0.040291 0.067902

a12 (s) = s5 + 38s4 + 459s3 + 2182s2 + 4160s + 2400

a21 (s) = s5 + 30s4 + 331s3 + 1650s2 + 3700s + 3000

a22 (s) = s5 + 42s4 + 601s3 + 3660s2 + 9100s + 6000

The poles of the system are located at: - 1, - 2, - 3, - 5, - 10, - 20.

The poles of required system i.e., second order reduced system is obtained using proposed algorithm
pe1 = −1 and pe2 = −2.05.
The first few Time Moments and Markov Parameters are:
" # " #
0.999 0.400 −0.899 −0.179
T0 = , T1 = , ...
0.499 0.999 −0.475 −0.666

" # " #
2.000 1.000 −11.99 −2.969
M0 = , M1 = , ...
1.000 1.000 −10.99 −0.999

Using step 1 and step 2, the Second order reduced model is obtained as
" #
1 1.203s + 2.0479 0.85s + 0.82
R21 (s) = (α = 2, β = 0)
D2 (s) 0.552s + 1.0230 1.68s + 2.047
" #
1 2s + 2.0479 1s + 0.820
R22 (s) = (α = 1, β = 1)
D2 (s) 1s + 1.02300 1s + 2.047

where, D2 (s) = s2 + 3.05s + 2.05, and α =Number of Time Moments, β =Number of Markov Param-
eters.
The step response of the 2th order reduced model i.e., R21 (s) and original system G(s) are plotted in
Figure 2 and error index ISE is calculated between the reduced model and original system as shown in
the Table 1 .
Example 2: Consider a transfer function matrix taken by Shamash [18].
" # " #
s+20
(s+1)(s+10) 1 a11 (s)
G(s) = s+10
=
(s+2)(s+7)
D(s) a21 (s)
65
 MIMO System Using Eigen Algorithm and Improved Pade Approximations

Figure 2. Comparison of the step responses for example 1

where,

D(s) = s4 + 18s3 + 97s2 + 180s + 100

a11 (s) = s3 + 27s2 + 150s + 200

a21 (s) = s3 + 21s2 + 120s + 100

The poles of the system are located at: (−1, −2, −5& − 10).
The poles of required system i.e., second order reduced system is obtained using proposed algorithm
pe1 = −1 and pe2 = −1.8.
The first few Time Moments and Markov Parameters are:
" # " #
2 −2.1
T0 = , T1 = , ...
1 −0.6
" # " #
1 9
M0 = , M1 = , ...
1 3

By using the proposed method, 2nd - order reduced model is obtained as :

66
SOP TRANSACTIONS ON APPLIED MATHEMATICS

Table 2. ISE Comparisons With Others For Example 2

ISE for ri j (s)[i = 1, 2; j = 1]

Reducted Model r11 r21
Proposed Method 9.5 × 10−4 1.7 × 10−3
Shamash [18] 0.00071 0.00080
C.B. Vishwakarma [14] 0.00930 0.01230

Table 3. Qualitative Comparison with the Original System for Example 2

Rise Time tr (sec.) Peak Overshoot M p (%) Settling Time ts (Sec.)
Systems
r11 r21 r11 r21 r11 r21
4th order Original System 2.21 1.2 00 00 3.97 2.1
2nd order Reduced Model with (t=2, m=0) 2.28 1.32 00 00 4.01 2.44

" #
1 1.82s + 3.6
R21 (s) = f or (α = 2, β = 0)
D2 (s) 1.72s + 1.8
" #
1 9s + 3.6
R22 (s) = f or (α = 1, β = 1)
D2 (s) 3s + 1.8

where, D2 (s) = s2 + 2.8s + 1.8.

The time and frequency response of the 2th order reduced model i.e., R21 (s) and original system G(s)
are plotted in Figure 3 and Figure 4 respectively. The ISE and qualitative comparisons are calculated
between the reduced model and original system and shown in the Table 2 and 3 respectively.

4. COMPARISON OF METHODS

The projected technique consists of the pole combination of the reduced dynamic systems by means of
Eigen algorithm and the improved pade approximations has been in second-hand to find the coefficient of
the numerator of reduced order system function. In the proposed method, the most dominant real pole has
to be retain to avoid the tendency of non minimum phase which may reduce the designing complexity of
the system. Though, the method is computationally multifaceted, since the improved pade approximations
has been used, but it only requires linear algebraic equations to be solved to turn up at reduced order
models. Furthermore, the proposed technique shows the way to a stable low order system, if the original
high-order system is stable. For a stable system, the system structure must have all the poles in the left
half of the complex s-plane. consequently, if the original system is stable with all its real poles lying in
the left of the complex s-plane, the poles obtained by Eigen algorithm always give a stable reduced order
model as shown in Figure 2 and Figure 3 . In additional, the most important skin tone of the proposed
method are as follows:
(i) Forming ability of transient, stability and steady state values of the original system are sealed. As a
result, the proposed method overcomes the drawback associated with the continued fraction expansion
based methods [17, 19], in which the stability equation method has been used to preserve the stability of
the original high order system in connection with the continued fraction expansion methods.
67
 MIMO System Using Eigen Algorithm and Improved Pade Approximations

Figure 3. Comparison of Step response R21 (s) with G(s) for example 2

(ii) Steady state value is not required to be matched in the proposed method as in the case of authors [20],
where one of the method has been used to derive the reduced order models after reciprocal transformation.
(iii) The method can be easily applied to the systems having real or complex poles only and no error
minimization technique [2, 4] is required in the proposed method in order to match the time responses of
original and reduced order systems.
(iv) Few methods do not require the formation of the Routh stability array as in [6, 7, 21]. Also, the
alpha and beta tables [22] are not required to be computed to derive stable reduced order models.
(v) Proposed method overcomes the negative aspect coupled with [8, 9], in which the complexity with
these methods is in selecting poles for the clusters.
(vi) The performance Indices i.e. ISE and qualitative comparison are compared with the proposed and
the other well known existing methods.

5. CONCLUSIONS

The authors presented a mixed reduction method for reducing the order of the large scale MIMO system .
In this method, the denominator polynomial is determined by using Eigen algorithm while the coefficients
of the numerator are obtained by improved pade approximations. This method has been tested on two
numerical examples chosen from the literature. Time and frequency responses of the original and reduced
systems are compared graphically and shown in the Figure 2 , Figure 3 and Figure 4 respectively. ISE
and qualitative comparisons are shown in the Table 1 , Table 2 and Table 3 respectively. From these
68
SOP TRANSACTIONS ON APPLIED MATHEMATICS

Figure 4. Comparison of bode response R21 (s) with G(s) for example 2

comparisons, it has been concluded that proposed method is mathematically simple, straight forward,
computer oriented, as good as in quality and capable to generate more dominant Eigen values needed
to synthesize the efficient reduced order models also proposed method is not applicable for the systems
having real and complex poles only. The introduced method reduces the problem of non-minimum phase
in the reduced models. This method guarantees stability of the reduced models, if the original high-order
system is stable, and also capable of making transient as well as steady state region of the original system.

References

[1] M. S. Mahmoud and M. G. Singh, Large scale systems modelling, vol. 1. Pergamon press Oxford,
1981.
[2] A. Mittal, R. Prasad, and S. Sharma, “Reduction of linear dynamics systems using an error mini-
mization technique,” JOURNAL-INSTITUTION OF ENGINEERS INDIA PART EL ELECTRICAL
ENGINEERING DIVISION, vol. 84, pp. 201–206, 2004.
[3] S. Mukherjee, R. Mittal, et al., “Model order reduction using response-matching technique,” Journal
of the Franklin Institute, vol. 342, no. 5, pp. 503–519, 2005.
[4] S. Mukherjee and R. Mishra, “Order reduction of linear systems using an error minimization
technique,” Journal of the Franklin Institute, vol. 323, no. 1, pp. 23–32, 1987.
[5] C. Vishwakarma, R. Prasad, et al., “Clustering method for reducing order of linear system using
pade approximation,” IETE journal of research, vol. 54, no. 5, p. 326, 2008.
69
 MIMO System Using Eigen Algorithm and Improved Pade Approximations

[6] Y. Dolgin and E. Zeheb, “On routh-pade model reduction of interval systems,” IEEE transactions on
automatic control, vol. 48, no. 9, pp. 1610–1612, 2003.
[7] V. Singh, D. Chandra, and H. Kar, “Improved routh-padé approximants: a computer-aided approach,”
IEEE transactions on automatic control, vol. 49, no. 2, pp. 292–296, 2004.
[8] J. Pal, A. Sinha, and N. Sinha, “Reduced order modelling using pole-clustering and time-moment
matching,” Journal of the Institution of Engineers(India): Electrical Engineering Division, vol. 76,
pp. 1–6, 1995.
[9] C. Vishwakarma, “Order reduction using modified pole clustering and pade approximation,” World
Academy of Science Engineering and Technology, vol. 56, 2011.
[10] S. Mukherjee, “Order reduction of linear systems using eigenspectrum analysis,” JOURNAL-
INSTITUTION OF ENGINEERS INDIA PART EL ELECTRICAL ENGINEERING DIVISION, pp. 76–
79, 1996.
[11] G. Parmar, S. Mukherjee, and R. Prasad, “System reduction using factor division algorithm and
eigen spectrum analysis,” Applied mathematical modelling, vol. 31, no. 11, pp. 2542–2552, 2007.
[12] J. Pal, “Improved pade approximants using stability equation method,” Electronics Letters, vol. 19,
no. 11, pp. 426–427, 1983.
[13] Y. Bistritz and U. Shaked, “Minimal pade model reduction for multivariable systems,” Journal of
dynamic systems, measurement, and control, vol. 106, no. 4, pp. 293–299, 1984.
[14] C. Vishwakarma, “Model order reduction of linear dynamic systems for control system design,”
Indian Institute of Technology Roorkee, 2009.
[15] R. Prasad, J. Pal, and A. Pant, “Multivariable system approximation using polynomial deriva-
tives,” JOURNAL-INSTITUTION OF ENGINEERS INDIA PART EL ELECTRICAL ENGINEERING
DIVISION, vol. 76, pp. 186–186, 1995.
[16] M. Safonov and R. Chiang, “Model reduction for robust control: A schur relative error method,”
International Journal of Adaptive Control and Signal Processing, vol. 2, no. 4, pp. 259–272, 1988.
[17] R. Prasad and J. Pal, “Use of continued fraction expansion for stable reduction of linear multi-
variable systems,” JOURNAL-INSTITUTION OF ENGINEERS INDIA PART EL ELECTRICAL
ENGINEERING DIVISION, vol. 72, pp. 43–43, 1991.
[18] Y. Shamash, “Model reduction using minimal realisation algorithms,” Electronics Letters, vol. 11,
no. 16, pp. 385–387, 1975.
[19] A. V. Agrawal and A. Mittal, “Reduction of large scale linear mimo systems using eigen spectrum
analysis and cfe form,” in XXXII National Systems Conference (December 2008), 2008.
[20] P.-O. Gutman, C. Mannerfelt, and P. Molander, “Contributions to the model reduction problem,”
Automatic Control, IEEE Transactions on, vol. 27, no. 2, pp. 454–455, 1982.
[21] V. Krishnamurthy and V. Seshadri, “Model reduction using the routh stability criterion,” Automatic
Control, IEEE Transactions on, vol. 23, no. 4, pp. 729–731, 1978.
[22] M. Hutton and B. Friedland, “Routh approximations for reducing order of linear, time-invariant
systems,” Automatic Control, IEEE Transactions on, vol. 20, no. 3, pp. 329–337, 1975.

70