International Journal of Systems Science

ISSN: 0020-7721 (Print) 1464-5319 (Online) Journal homepage: https://www.tandfonline.com/loi/tsys20

Model order reduction of interval systems using

an arithmetic operation

Kranthi Kumar Deveerasetty & S. K. Nagar

To cite this article: Kranthi Kumar Deveerasetty & S. K. Nagar (2020): Model order reduction of
interval systems using an arithmetic operation, International Journal of Systems Science, DOI:
10.1080/00207721.2020.1746433

To link to this article: https://doi.org/10.1080/00207721.2020.1746433

Published online: 07 Apr 2020.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tsys20
INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE
https://doi.org/10.1080/00207721.2020.1746433

Model order reduction of interval systems using an arithmetic operation

Kranthi Kumar Deveerasettya,b and S. K. Nagarb
a Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China; b Indian Institute of Technology, Banaras Hindu
University, Varanasi, India

ABSTRACT ARTICLE HISTORY

The paper presents an extension of the differentiation method for model order reduction of large- Received 9 December 2018
scale interval systems. This is an alternative approach to the existing differentiation method of Accepted 18 March 2020
interval systems. The proposed method has been applied for both continuous and discrete-time KEYWORDS
interval systems. The reduction of discrete-time interval systems is achieved by using simple lin- Bilinear transformation;
ear transformation z = (w + 1) and bilinear transformation z = (1 + w)/(1 − w), where w = 1. The differentiation method;
proposed method always generates stable reduced-order models, and also it retains the zeroth- linear transformation;
order interval time moment. Four numerical examples exemplify the accuracy of the method and interval systems;
computational simplicity. Furthermore, the difficulties associated with the extension of Routh-based Kharitonov’s theorem
approximations to interval systems for obtaining stable reduced-order models are discussed. The
stability of interval systems is verified by using Kharitonov’s theorem.

1. Introduction such as system decomposition algorithm (S. K.

In the field of systems and control theory, model Singh & Nagar, 2004), balanced singular perturba-
order reduction (MOR) techniques have attracted tion approximation (Kumar et al., 2012a, 2012b),
increasing attention from researchers. These tech- and extended minimal degree optimal Hankel norm
niques analyse high-order system dynamical prop- approximation (Kumar & Nagar, 2014) have been used
erties in practical application aimed at reducing the to achieve MOR of unstable systems.
system complexity. The reduced-order system per- Research on interval systems has experienced
mits to study and to achieve a better understand- considerable attention since the initiating work of
ing of the higher order system. Two major advan- Kharitonov. Kharitonov’s theorem (Argoun, 1990;
tages of reduced-order models are easier design and Bose et al., 1988; Kharitonov, 1978) is useful for
faster simulations. The developments of fundamental studying the stability of interval systems. Kharitonov’s
concepts of the MOR area were started in the early method is applied to generate the four associated
1980s, and significant improvements were achieved Kharitonov polynomials. Each Kharitonov polynomial
in the 1990s (Fortuna et al., 1992; Glover, 1984; is tested using the Routh criterion to find the num-
Jamshidi, 1983; Jamshidhi & Malek-Zavarei, 1986; ber of its roots in the right half-plane. If none of the
Mahmoud & Singh, 1981; Obinata & Anderson, 2001; Kharitonov polynomials has roots in the right half-
Shamash & Feinmesser, 1978; Sinha & Kuszta, 1983). plane, then the linear system is stable. If the origi-
Later, new MOR techniques have been developed for nal interval system is n ≤ 6 (Anderson et al., 1987),
the reduction of nonlinear systems (Antoulas, 2005; then the number of Kharitonov’s polynomials required
Antoulas et al., 2006; Wilhelmus & Rommes, 2008). A to verify the stability is less than four. The signif-
MOR technique has been developed based on ratio- icant drawback of Kharitonov’s theorem is that it
nal interpolation (Popeea & Jora, 2004). L2 norm does not apply to discrete-time interval polynomials
order reduction technique (Popeea & Jora, 2007) (Barmish, 1994). To overcome this limitation, bilin-
was developed to obtain the error in the system ear transformation (Mastorakis, 1997) was introduced
as small as possible. Recently, different techniques in Kharitonov’s algorithm. Generalised Kharitonov’s

CONTACT Kranthi Kumar Deveerasetty [email protected]

© 2020 Informa UK Limited, trading as Taylor & Francis Group

2 K. KUMAR DEVEERASETTY AND S. K. NAGAR

theorem and robust control under the parametric subtraction rule has been changed to obtain stable
approach were clearly explained by Bhattacharyya reduced-order models.
et al. (1995). A hybrid method based on the generalised Routh
The entire range of operating conditions for many approximation and the Factor division methods were
practical systems exhibits parametric uncertainty. The proposed (Selvaganesan, 2007). The main limitation of
analysis and design of interval systems have received the hybrid method is that the factor division method
a great deal of attention. The Routh-Pade approxima- has not been used either in the entire article or
tion has been extended to interval systems for reduc- for the order reduction of a numerator polynomial.
ing higher order continuous interval systems (Bandy- Saini and Prasad (2010) considered the order reduc-
opadhyay et al., 1994). The reduced-order denomi- tion of the linear continuous interval system by min-
nator polynomial is obtained by direct truncation of imisation of the Integral Square Error (ISE) using
the Routh table, and the lower order numerator poly- genetic algorithm and particle swarm optimisation.
nomial is obtained by matching the coefficients of The main drawback of this method is to obtain sta-
the power series expansions of the interval systems. ble reduced-order denominator coefficients because
The concept of γ − δ Routh approximation has been they applied generalised Routh approximation, which
proposed for continuous interval systems (Bandy- fails to preserve the stability of the reduced-order
opadhyay et al., 1997). The two major limitations interval polynomial coefficients. An another hybrid
of the above-mentioned Routh-based approximations method was proposed based on the Mahailov crite-
claimed by Hwang and Yang (1999) are as follows: (1) ria for the reduction interval denominator polynomial
the interval Routh extension formula does not guar- and for the reduction of numerator polynomial Cauer
antee the successes in generating a full interval Routh second form (Kranthi Kumar et al., 2011a) and fac-
array and (2) some interval Routh approximations may tor division method (Kranthi Kumar et al., 2011b)
not be robustly stable, even if the original interval sys- applied for the model reduction of the continuous
tem is stable. To reduce the computational effort, γ interval systems. Recently, the above technique based
Table formulation (Sastry et al., 2000) has been intro- on the Mahailov criteria has been applied to discrete-
duced, instead of γ − δ Table formulation (Bandy- time interval systems (Choudhary & Nagar, 2018a).
opadhyay et al., 1997). However, the limitation of this To reduce the computation of the above algorithm, a
method is that they obtained a reduced interval model direct truncation method is applied to the continu-
that may be unstable for the stable original interval ous interval systems (Kranthi Kumar et al., 2013a). In
model. Dolgin and Zeheb have proven that a gener- this algorithm, all the highest degrees of polynomials
alised Routh algorithm to interval systems does not are removed directly. This technique was applied to
guarantee the stability of the reduced-order system. discrete interval systems (Choudhary & Nagar, 2013).
To overcome this problem, Dolgin and Zeheb (2003) The above algorithm has a limitation to guaran-
modified the generalised Routh array and claimed tee the stability of reduced interval systems. This
that this method could guarantee the stability of the problem was overcome by the extended differentia-
reduced-order system. Later, Yang (2005) proved that tion method (Kranthi Kumar et al., 2013b) to inter-
the Dolgin and Zeheb method does not guarantee the val systems, obtaining stable reduced-order models.
stability of the reduced-order interval system. Interval However, it required to use reciprocal transformation
Routh approximation fails to obtain stable reduced- before and after differentiation and multiplying by a
order interval models because the interval operations possible gain factor. To overcome this limitation, this
are implicitly overbounded. Furthermore, the depen- present paper presents an extended derivation of the
dency problem affects the interval Routh approxima- reduced interval models. Such derivation permits to
tion. Therefore, the non-interval Routh approximation readily build reduced interval models generated by
has a major advantage of stability preservation that a modified Routh-type array structure. Recently, an
cannot be claimed for interval Routh approximation. impulse energy approximation technique (Siva Kumar
To overcome this problem, Dolgin (2005) proposed a et al. 2015) has used to obtain a lower order interval
modified method of the Routh algorithm for obtaining model from the original higher order interval system.
stable reduced-order models. It is mentioned that the For reduction of discrete interval systems, dominant
limitation of this method is that the interval arithmetic pole-based methods were introduced (Choo, 2007;
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 3

Choudhary & Nagar, 2017, 2018a; Ismail et al., 1997; V. numbers, that is, o ∈ {+, −, ×, ÷}. The corresponding
P. Singh & Chandra, 2012). It is noted that the Routh operations for intervals [e] and [f ] are given by
approximation-based method results in the unstable     
reduced model when it is extended to interval systems [e] o f = eof /e ∈ [e] , f ∈ f . (1)
(Choudhary & Nagar, 2018b, 2018c). Recently, the
direct implementation of impulse-response Gramian The interval arithmetic rules (Alefeld & Mayer, 2000)
(Kranthi Kumar et al., 2019) applied for the reduction are defined as follows:
of large-scale interval systems and proved it fails to Addition:
obtain stable reduced-order interval models. To over-      
come the problem, the reduced-order model is devel- e − , e+ + f − , f + = e− + f − , e+ + f + (2)
oped based on Kharitonov’s theorem. The method is
evaluated using two techniques: one is Routh’s tech- Subtraction:
nique (α − β), and the other is Inner’s technique.  − +  − +  − 
e , e − f , f = e − f + , e+ − f − (3)
The main advantage of the impulse-response Gramian
technique is useful for system identification, and the
Multiplication:
method is computationally simple.
Essentially, the methods for MOR of interval sys-  − +  − + 
e , e × f , f = min(e− f − , e− f + , e+ f − ,
tems are derived from the existing method for systems 
with fixed coefficients. The differentiation method is e+ f + ), max(e− f − , e− f + , e+ f − , e+ f + ) (4)
considered, which is originally proposed by Gutman
Division:
et al. (1982) for an extension to interval systems.
 
The difficulty of reciprocal transformation before  − +
  − +
  − +
 1 1
and after the differentiation method is eliminated e ,e ÷ f ,f = e ,e × , (5)
f+ f−
by Lucas (1992) for computational simplicity. Direct
application of Lucas’s differentiation method to inter- From the specialised existent literature, it is clear that
val systems does not guarantee the stability of reduced- dependency problem (Hayes, 2003) is a major issue to
order systems. Hence authors consider using the mod- the application of interval arithmetic. Let us consider
ified Routh approximation (see step 3 in Section 3.2). an example to illustrate the square of an interval func-
From the literature, clearly, Routh-based approxima- tion. The following function [e− , e+ ]2 = [e− , e+ ] ×
tion fails to produce stable reduced-order models, but [e− , e+ ] seems to work in some cases, such as [2, 4]2 =
the proposed method guarantees the stability of the [4, 16]. However, from [−3, 3]2 = [−9, 9], we can see
reduced-order interval system if the original system is that the square of a real number cannot be nega-
stable. Four numerical examples are solved to examine tive. The correct answer is [−3, 3]2 = [0, 9]. From the
the validity of the proposed method. extension of classical MOR techniques to interval sys-
The paper is structured as follows: Section 2 pro- tems, dependency problem is one of the reasons for
vides the arithmetic rules, which are used in the obtaining an unstable reduced-order model. Let us
present investigation. Section 3 describes the proposed check the statement that [−1, 1] − [−1, 1] might be
method for continuous and discrete-time interval sys- zero or not. There is a dependency issue, well known
tems followed by the computation of ISE and Error in interval arithmetic (or reliable computing) circles,
index. In Section 4, the results are compared with in which the lack of ability to determine the source of
reduced-order models of other methods, and the con- an interval, i.e. whether two intervals are correlated,
clusions are reported in Section 5. leads to wider and therefore less useful results.
Consider the following example. If the expression
means the set {e − e|e ∈ [−1, 1]}, the answer is 0. If the
2. Preliminaries
two intervals are different {e − f |e ∈ [−1, 1]} and f ∈
Let us consider that e = [e− , e+ ] ≡ {e|e ∈ [e− , e+ ]}, [−1, 1], then the answer is the interval [−2, 2].
f = [f − , f + ] ≡ {f |f ∈ [f − , f + ]} are real intervals. o is It is noted that in the literature the arithmetic rules
one of the fundamental procedure addition, subtrac- were limited to only four rules (addition, subtraction,
tion, multiplication and division, severally, for actual multiplication and division).
4 K. KUMAR DEVEERASETTY AND S. K. NAGAR

Lemma 2.1 (Bhattacharyya et al., 1995): Let Theorem 2.3 (Kharitonov’s theorem): Every polyno-
mial in the family (s) is Hurwitz if and only if the four
Q1 (s) = Qeven (s) + Qodd
1 (s) polynomials are Hurwitz:
(6)
Q2 (s) = Qeven (s) + Qodd
2 (s) 1 (s) = x0− + x1− s + x2+ s2 + x3+ s3 + x4− s4
denote the same degree of the two stable polynomials + x5− s5 + x6+ s6 + · · ·
with equal even part Qeven (s) and distinct odd parts
2 (s) = x0− + x1+ s + x2+ s2 + x3− s3 + x4− s4
1 (s) and Q2 (s) satisfying
Qodd odd

+ x5+ s5 + x6+ s6 + · · ·
Qo1 (ω) ≤ Q02 (ω) (7)
3 (s) = x0+ + x1− s + x2− s2 + x3+ s3 + x4+ s4
for all ω ∈ [0, ∞].
Then, Q(s) = Qeven (s) + Qodd (s) is stable for every + x5− s5 + x6− s6 + · · ·
polynomial Q(s) with odd part Qodd (s) satisfying 4 (s) = x0+ + x1+ s + x2− s2 + x3− s3 + x4+ s4
Qo1 (ω) ≤ Qo (ω) ≤ Qo2 (ω) (8) + x5+ s5 + x6− s6 + · · · (12)

for all ω ∈ [0, ∞]. Proof: The proof is given in Bhattacharyya et al.
(1995), and it is omitted here. 
Proof: The proof is given in Bhattacharyya et al.
(1995), and it is omitted here. 
3. Proposed method
Lemma 2.2 (Bhattacharyya et al., 1995): Let Let Gn (s) the transfer function of the original stable
continuous interval systems of order n by
Q1 (s) = Qeven
1 (s) + Qodd (s)  − +  − +
(9) b0 , b0 + b1 , b1 s + · · ·
Q2 (s) = Qeven (s) + Qodd (s)  +  n−1
2
+ b− n−1 , bn−1 s
denote the same degree of the two stable polynomials Gn (s) =  − +   − + 
a0 , a0 + a1 , a1 s + · · ·
with equal odd part Qodd (s) and distinct even parts  
+ a− + n
n , an s
1 (s) and Q2 (s) satisfying
Qeven even
N (s)
Qe1 (ω) ≤ Qe2 (ω) (10) = , (13)
D (s)
for all ω ∈ [0, ∞]. where a− + −
j < aj < aj for j = 0, 1, 2, . . . ., n and bi <
Then, Q(s) = Qeven (s) + Qodd (s) is stable for every bi < b+ i are lower bounds and upper bounds for D(s)
polynomial Q(s) with even part Qeven (s) satisfying and N(s) of the original interval transfer function. Let
Rk (s) be the reduced-order model of the stable original
Qe1 (ω) ≤ Qe (ω) ≤ Qe2 (ω) (11)
system in the form
for all ω ∈ [0, ∞].  − +  − +
d0 , d 0 + d1 , d 1 s + · · ·
 − +  k−1
+ dk−1 , dk−1 s
Proof: The proof is given in Bhattacharyya et al. Rk (s) =  − +   − + 
(1995), and it is omitted here.  c 0 , c 0 + c1 , c 1 s + · · ·
 
+ ck− , ck+ sk
2.1. Kharitonov’s theorem Nk (s)
= , (14)
Dk (s)
Consider the family (s) of real polynomials of degree
n with the form (s) = x0 + x1 s + x2 s2 + · · · + xn sn , where cj− < cj < cj+ for j = 0, 1, 2, 3, . . . , k and di− <
where the coefficients lie within the given ranges x0 ∈ di < di+ for i = 0, 1, 2, 3, 4, . . . ., k − 1 are lower
[x0− , x0+ ], x1 ∈ [x1− , x1+ ], . . . , xn ∈ [xn− , xn+ ]. bounds and upper bounds for the reduced-order
Kharitonov’s theorem provides a surprisingly sim- Dk (s) and Nk (s) numerator polynomial interval trans-
ple necessary and sufficient condition for the Hurwitz fer function. The order of the reduced-order interval
stability of the entire family. system is k with k < n.
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 5

Table 1. Modified Routh table. obtained by using the modified Routh approxima-
row 1 [a− + − +
11 , a11 ] = [an , an ] [a− + − +
12 , a12 ] = [an−1 , an−1 ] ··· tion. The third order coefficients are obtained by the
− +
row 2 [a21 , a21 ] = (n)[a− +
n , an ] [a22 , a22 ] = (n − 1)[an−1 , a+
− + −
n−1 ] ··· following relation:
row 3 [a− ,
n−1 n−1a +
] = [a− +
a
31 31 ]
, [a− + − +
n−2 , an−2 ] = [a32 , a32 ] ···

a− + − +
ij , aij = ai−2,j+1 , ai−2,j+1
3.1. Modified differentiation method for continuous αi−2,1 −
− a , a+ , (19)
time interval systems αi−1,1 i−1,j+1 i−1,j+1
Theorem 3.1: If the reduced-order interval polynomial where i = 3 and j = 1, 2, 3, . . . , n − 1. To overcome
Rk (s) satisfies Kharitonov’s theorem then original inter- the limitations of a dependency property, midpoints
val polynomial D(s) is also stable. are used in Equations (18) and (19).
The reduced-order interval polynomial is stable if it
Proof: Let the n order polynomial satisfies Kharitonov’s theorem. Based on Kharitonov’s
   − +  n−1 theorem, the reduced-order interval polynomial Rk (s)
D(s) = a− + n
n , an s + an−1 , an−1 s + ··· can be written as four polynomials
 − +  − +
+ a 1 , a1 s + a 0 , a0 (15)
1 (s) = a− − + 2 + 3 − 4 − 5
0 + a1 s + a2 s + a3 s + a4 s + a5 s + · · ·
be a higher order polynomial. Its differentiated
polynomial is 2 (s) = a− + + 2 − 3 − 4 + 5
0 + a1 s + a2 s + a3 s + a4 s + a5 s + · · ·

   3 (s) = a+ − − 2 + 3 + 4 − 5
0 + a1 s + a2 s + a3 s + a4 s + a5 s + · · ·
+  n−2
Ḋ(s) = n a− + n−1
n , an s + (n − 1) a−
n−1 , an−1 s
 4 (s) = a+ + − 2 − 3 + 4 + 5
0 + a1 s + a2 s + a3 s + a4 s + a5 s + · · ·
+
+ · · · + a− 1 , a1 . (16) (20)

To obtain the reduced-order polynomial, the modified Let us bring in the hyper rectangle or box  of coeffi-
Routh approximation is applied by using the above two cients of the perturbed polynomials
equations (15) and (16):  
  = a a ∈ n , a− +
i ≤ ai ≤ ai , i = 1, 2, 3, . . . , n − 1 .
+  n−1  +  n−2
Rk (s) = a−n−1 , an−1 s + a−
n−2 , an−2 s (21)
 +
+ · · · + a−0 , a0 . (17) Four Kharitonov’s polynomials are constructed from
two different even parts even
max (s) and min (s) and two
odd

The construction of Modified Routh approximation different odd parts odd

max (s) and min (s), which are
odd

(Dolgin, 2005) is given in Table 1. defined as

To improve the calculation in Table 1, let us consider even + − 2 + 4 − 6
max (s) = a0 + a2 s + a4 s + a6 s + · · ·
[a− + − +
n−1 , an−1 ] = [a31 , a31 ], [a− + − +
n−2 , an−2 ] = [a32 , a32 ] (22)
− + 2 − 4 + 6
and so on. even
min (s) = a0 + a2 s + a4 s + a6 s + · · ·
Therefore,
and

a− +
11 +a11 + − 3 + 5 − 7
 − +  − +  + odd
max (s) = a1 s + a3 s + a5 s + a7 s + · · ·
a31 , a31 = a12 , a12 −  − +
2
a−
22 , a22 ; (23)
a21 +a21 − + 3 − 5 + 7
2 odd
min (s) = a1 s + a3 s + a5 s + a7 s + · · ·
 − +
a11 +a11
 − +  − +  +
The motivation for obtaining the subscripts max and
a32 , a32 = a13 , a13 −  − +
2
a−
23 , a22 ; · · · min is as follows. Let an arbitrary polynomial a(s) with
a21 +a21
2 its coefficients lying in the box be  and let aeven (s) be
(18)
its even part. Then
The midpoint of the interval [a− , a
11 11
+
] and [a− +
21 21 ] is
, a
− +
defined by α11 = mid([a11 ]) = (a11 + a11 )/2; α21 = emax (ω) = a+ − 2 + 4 − 6
0 − a 2 ω + a4 ω − a 6 ω + · · ·
mid([a21 ]) = (a− +
21 + a21 )/2.
Note that the first line is the original polynomial ae (ω) = a0 − a2 ω2 + a4 ω4 − a6 ω6 + · · ·
D(s) and the second row is the differentiation of row emin (ω) = a− + 2 − 4 + 6
0 − a2 ω + a4 ω − a6 ω + · · ·
1. The third row is the reduced-order polynomial (24)
6 K. KUMAR DEVEERASETTY AND S. K. NAGAR

so that Table 2. Denominator table for continuous time interval systems.

  [a− + − + [a− + − +
11 , a11 ] = [an , an ] 12 , a12 ] = [an−1 , an−1 ] ···
emax (ω) − ae (ω) = a+ − 2 row 1
0 − a0 + a2 − a2 ω row 2 − + − +
[a21 , a21 ] = (n)[an , an ] [a22 , a22 ] = (n − 1)[an−1 , a+
− + −
n−1 ] ···
 [a− +
[a− +
+ a+ 4
4 − a4 ω + · · · (25) row 3
row 4 [a− +
,
31 31a ]
− +
[a− +
32 , a32 ]
− +
···
41 , a41 ] = (n − 1)[a31 , a31 ] 42 , a42 ] = (n − 2)[a32 , a32 ] ···
row 5 [a− ,
51 51a +
] [a − +
,
52 52a ] ···
and
  +
ae (ω) − emin (ω) = a0 − a−
0 + a2 − a2 ω
2
 (i) The first row as shown in Table 2 is formed from
+ a4 − a− 4
4 ω + ··· (26) denominator coefficients of Gn (s) (higher order
Therefore, coefficients).
(ii) The second row is obtained by differentiation of
emin (ω) ≤ ae (ω) ≤ emax (ω) ; ω ∈ [0, ∞] . row 1.
(27) (iii) The third row is obtained by applying the modi-
Similarly, if aodd (s) denotes the odd parts of a(s), it can fied Routh approximation. This process will give
be verified that reduced-order denominator of order n−1.
omin (ω) ≤ ao (ω) ≤ omax (ω) ; ω ∈ [0, ∞] . (iv) The fourth row is obtained by differentiation of
(28) row 3.
To proceed, note that the Kharitonov polynomials in (v) The fifth row is obtained by the modified Routh
Equation (20) can be rewritten as approximation using rows 3 and 4. This will give
reduced denominator of order n−2.
1 (s) = even odd
min (s) + min (s) ,

2 (s) = even odd The algorithm is illustrated by Table 2 for obtaining

min (s) + max (s) ,
(29) a reduced-order denominator
3 (s) = even odd
max (s) + min (s) ,
a− + − +
ij , aij = ai−2,j+1 , ai−2,j+1
4 (s) = even odd
max (s) + max (s) .
αi−2,1 −
If all the polynomials with the coefficients in the box  − a , a+ , (30)
αi−1,1 i−1,j+1 i−1,j+1
are stable, it is clear that the Kharitonov polynomials
in Equation (21) are also stable since their coefficients where i = 3, 5, 7, . . . and j = 1, 2, 3, . . . , n − 1.
lie in . Assume that the Kharitonov polynomials are
a− +
i−2,1 + ai−2,1 a− +
i−1,1 + ai−1,1
stable, and let a(s) = aeven (s) + aodd (s) be an arbitrary αi−2,1 = ; αi−1,1 =
polynomial with coefficients in the box  with its even 2 2
part aeven (s) and its odd part aodd (s). is the midpoint of the coefficients. = [a− [a− +
31 , a31 ]
12 ,
Since 1 (s) and 2 (s) are stable and Equation (28) a+ ] − (α 11 /α 21 )[a − +
, a ].
12 22 22
holds, we conclude from Lemma 2.1 applied to 1 (s) The denominator polynomial Dk (s) for k = n−1
and 2 (s) in Equation (29) that even min (s) + a
odd (s) is
has coefficients (see row 3 from Table 2)
stable.  − +  − +   − +   +
Similarly, Lemma 2.1 is applied to 3 (s) and 4 (s) ck , ck = a31 , a31 ; ck−1 , ck−1 = a− 32 , a32 ; · · ·
in Equation (29), and we conclude that even max (s) +
(31)
aodd (s) is stable. The above algorithm is repeated for reducing the
Now, since Equation (27) holds, we can apply numerator polynomial.
Lemma 2.2 to the two stable polynomials even max (s) +
Kranthi Kumar et al. (2013b) proposed the differen-
a (s) and min (s) + a (s), and we conclude that
odd even odd tiation interval system reduction method, which pre-
aeven (s) + aodd (s) = a(s) is stable.  serves stability and is based upon the simple idea of dif-
ferentiating the higher order numerator and denomi-
nator polynomial to reduce their order of the interval
3.2. Algorithm
system. However, to obtain adequate reduced-order
The procedure to obtain reduced-order polynomials models, they determined as an essential step to ini-
Dk (s) and Nk (s) in Equation (7) can be summarised tially use a reciprocal transformation on the higher
as follows. order interval system, i.e. Gn (s) = Gn (s)/si (where i is
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 7

the pole-zero excess) which has the effect of revers- 3.4. Modified differentiation method for
ing the order of the numerator polynomial coefficients discrete-time interval systems
and denominator polynomial coefficients. Differentia-
Let the transfer function of a higher order discrete-
tion is implemented to the necessary interval reduced-
time interval systems be
order model and the inverse reciprocal transforma-
tion is employed to the reduced-order interval system.
Eventually, the reduced-order model is multiplied by a [b− + − + − +
0 , b0 ] + [b1 , b1 ]z + · · · + [bn−1 , bn−1 ]z
n−1
Gn (z) =
proper dc gain factor to obtain steady-state gain. [a− + − + − + n
0 , a0 ] + [a1 , a1 ]z + · · · + [an , an ]z

In this article, it is demonstrated that by redevel- N(z)

= . (34)
oping the differentiation technique by neglecting the D(z)
consideration of reciprocal transformation and multi-
plication by dc gain factor from the algorithm proce- The kth order reduced model of Gn (z) is expressed as
dure, computational simplicity is improved. Applying
a modified Routh array for computing the reduced      −  k−1
d0− , d0+ + d1− , d1+ z + · · · + dk−1 +
, dk−1 z
interval models also allows for a quick link with other Rk (z) =  − +  − +  − +
c0 , c0 + c1 , c1 z + · · · + ck , ck z k
stability preserving methods.
Nk (z)
= . (35)
Dk (z)
3.3. Direct extension of the Lucas (1992) method to
interval systems
Next, we present the algorithm to obtain reduced-
The algorithm is the same as the above-mentioned order discrete interval models.
algorithm except for step (iii) where the generalised
Routh approximation is used instead of the modified 3.4.1. Case 1: Modified differentiation method
Routh approximation (see Equation (30)). step1: The first row formed from the original
The generalised Routh approximation is denominator coefficients of Gn (z) (higher
 −  − order coefficients).
ai−1,1 , a+ +
i−1,1 ai−2,j+1 , ai−2,j+1 step2: The second row is obtained by differentiation
  −
− a−i−2,1 , a+i−2,1 ai−1,j+1 , a+
i−1,j+1
of row 1.
− +
aij , aij =  −  , step3: The third row is obtained by applying the
ai−1,1 , a+
i−1,1 modified Routh approximation. This process
(32) will give a reduced-order denominator of
where i = 3, 5, 7, . . . and j = 1, 2, 3, . . . , n − 1. order n−1.
Another alternative method is using the Dolgin and step4: The fourth row is obtained by differentiation
Zeheb method instead of the generalised Routh of row 3.
approximation step5: The fifth row is obtained by the modified
Routh approximation using row 3 and row 4.
a− + − +
ij , aij = ai−2,j+1 , ai−2,j+1 This will give reduced denominator of order
 − 
ai−2,1 , a+ n−2.
− − i−2,1
 a− +
i−1,j+1 , ai−1,j+1 ,
ai−1,1 , a+i−2,1
3.4.2. Case 2: Applying linear transformation
(33)
step1: Apply linear transformation (z = w + 1) to
where i = 3, 5, 7, . . . and j = 1, 2, 3, . . . , n − 1. the higher order system Gn (z), then the
Via numerical example 4.1 (see Equation (41) first row is obtained as Gn (w + 1) and then
and (42)), it has been proved that the direct extension denominator coefficients can be formed in
of the Lucas method to interval systems fails to obtain terms of w, then the procedure to obtain
stable lower order interval models. reduced interval models are the same as
In terms of non-interval coefficients, Equations (32) shown in Table 2.
and (33) are equal; however, in terms of interval coef- step2: The second row is obtained by differentiation
ficients Equations (32) and (33) are not equal. of row 1.
8 K. KUMAR DEVEERASETTY AND S. K. NAGAR

Figure 1. Block diagram of the modified differentiation method.

step3: The third row is obtained by applying the step3: The third row is obtained by applying the
modified Routh approximation. This process modified Routh approximation. This process
will give reduced-order denominator of order will give reduced-order denominator of order
n−1. n−1.
step4: The fourth row is obtained by differentiation step4: The fourth row is obtained by differentiation
of row 3. of row 3.
step5: The fifth row is obtained by the modified step5: The fifth row is obtained by the modified
Routh approximation using row 3 and row 4. Routh approximation using row 3 and row 4.
This will give reduced denominator of order This will give reduced denominator of order
n−2 and so on. n−2 and so on.
step6: Substitute (w = z − 1) in the reduced-order step6: Substitute (w = (z − 1)/(z + 1)) in the
model. reduced-order model.

Figure 1 illustrates the above algorithm.

The linear transformation (z = w + 1) has the con- 3.5. Integral square error
sequence of shifting all the poles and zeros of Gn (s) by In order to verify the accuracy of the proposed
a distance −1 unit in the complex plane. Thus for a sta- method, the integral square error index between the
ble or minimum phase system Gn (z), the poles and/or transient parts of the original and reduced models is
zeros of the corresponding Gn (w) need not lie within calculated by using Matlab/Simulink.
the unit circle. The integral square error (ISE) is defined as

3.4.3. Case 3: Applying bilinear transformation  ∞ 2

step1: Apply bilinear transformation (z = (1 + w)/ ISE = g(t) − r(t) dt, (36)
(1 − w)), where w = 1 to the higher order 0
system Gn (z), then, the first row is obtained
as Gn ((1 + w)/(1 − w)), where w = 1 and where g(t) and r(t) are step responses of the original
then denominator coefficients can be formed and the reduced-order models, respectively.
in terms of w, then the procedure to obtain
reduced interval models are same as shown in
Table 2: 3.6. Error index

1+w Error index ‘J’ is specified by the following
Gn algorithm:
1−w
 
[b−0 , b +
0 ] + [b − +
1 , b 1 ] 1+w
1−w
 n−1 
M
   2
+ · · · + [b− , b +
n−1 n−1 1−w ] 1+w J= y tp − yr tp , (37)
=   p=0
[a0 , a0 ] + [a−
− + +
1 , a1 ] 1−w
1+w
 n
+ · · · + [a− ,
n n a + ] 1+w
1−w where y and yr are the outputs of the original interval
system Gn (z) and the reduced-order interval system
step2: The second row is obtained by differentiation Rk (z), respectively, at sampling instants tp , and M is
of row 1. the number of sampling periods.
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 9

Table 3. Modified differentiation method for Example 4.1.

row 1 Higher order [2.1,2.6] [76.1,76.7] [119.1,119.6] [111.0,111.6] [71.8,72.3] [31.0,31.7] [9.0,9.9]
row 2 Differentiation [12.6,15.6] [380.5,383.5] [476.4,478.4] [333,334.8] [143.6,144.6] [31.0,31.7]
row 3 5th order [12.17,13.27] [39.35,40.18] [55.18,56.08] [47.69,48.19] [25.71,26.52] [9,9.9]
row 4 Differentiation [60.8,66.35] [157.40,160.73] [165.56,168.26] [95.39,96.39] [25.71,26.52]
row 5 4th order [7.20,8.7] [21.53,22.97] [28.41,29.11] [20.41,21.37] [9,9.9]
row 6 Differentiation [28.81,34.81] [64.60,68.92] [56.83,58.23] [20.41,21.37]
row 7 3rd order [4.30,6.82] [13.85,14.9] [15.06,16.28] [9,9.9]
row 8 Differentiation [12.91,20.47] [27.71,29.81] [15.06,16.28]
row 9 2nd order [3.92,5.67] [9.63,11.26] [9,9.9]

4. Results and discussions Without applying a modified Routh approximation

(see step 3 in Section 3.2), i.e. extending directly differ-
Example 4.1: Consider a sixth-order interval polyno-
entiation method proposed by Lucas (1992) to interval
mial reported in Yang (2005)
systems, the reduced-order system cannot give any
P(s) = [2.1, 2.6] s6 + [76.1, 76.7] s5 guaranteed stability (see Equations (41) and (42)). By
+ [119.1, 119.6] s4 + [111.0, 111.6] s3 the Dolgin and Zeheb method (2003)

+ [71.8, 72.3] s2 + [31.0, 31.7] s + [9.0, 9.9] .

(38) P4 (s) = [115.6673, 116.1857] s4 + [63.98, 64.9237] s3

Following the algorithm in Section 3.1, Table 3 is con- + [70.8249, 71.3465] s2 + [24.4755, 25.7687] s
structed. The fifth-order reduced-order polynomial is + [9, 9.9] . (43)
formed from the third row of Table 3
P5 (s) = [12.1706, 13.2707] s5 + [39.3507, 40.1841] s4 It is noted that the above interval polynomial (P4 (s))
3
+ [55.1888, 56.0889] s +[47.6952, 48.1952] s 2 of Equation (43) is not stable since one of the
four Kharitonov vertex polynomials of P4 (s) of
+ [25.7156, 26.5223] s + [9, 9.9] . (39)
Equation (43) is not stable. Therefore, it is clear that the
The second-order reduced-order polynomial is Dolgin–Zeheb method does not assure the stability of
formed from the ninth row of Table 3 the reduced-order polynomial.
For the above problem, some of the existing
p2 (s) = [3.9205, 5.6705] s2 + [9.6356, 11.2632] s
methods (Bandyopadhyay et al., 1994, 1997; Dolgin
+ [9, 9.9] . (40) & Zeheb, 2003; Saini & Prasad, 2010; Sastry et al., 2000;
The reduced-order polynomial stability can be eas- Selvaganesan, 2007) fail to give a stable reduced-order
ily verified by the Kharitonov theorem (1978). Direct polynomial. It was reported by Yang (2005) that the
extension of the Lucas method (1992) to interval sys- Dolgin and Zeheb method will result in an unstable
tems is done by using Equation (32): reduced-order model.

P5 (s) = [−3.0349, 31.5452] s5 + [16.7961, 68.3428] s4

+ [33.8538, 82.6714] s3 +[33.8923, 65.5809] s2 Example 4.2: Consider a numerical example reported
by Hwang and Yang (1999), where the γ − δ method
+ [19.7551, 34.0809] s + [9, 9.9] . (41)
of the interval system proposed by Bandyopadhyay
By direct extension of the Lucas method (1992) to et al. (1997) fails to produce a fifth-order reduced
interval systems by using Equation (33) model. The proposed method successfully provides a
stable reduced-order system. The sixth-order polyno-
P5 (s) = [−3.016, 25.4847] s5 + [20.4061, 55.4766] s4
mial is as follows:
+ [41.9308, 66.7782] s3 + [41.96, 52.9714] s2
+ [24.4603, 27.5274] s + [9, 9.9] . (42) P6 () = [9, 9.5] s6 + [31, 31.5] s5 + [71, 71.5] s4
It is clear that from Equations (41) and (42) the + [111.0, 111.5] s3 + [119, 119.5] s2
Kharitonov polynomial vertex has roots in the right
half plan. + [76, 76.5] s + [2, 2.5] . (44)
10 K. KUMAR DEVEERASETTY AND S. K. NAGAR

The fifth-order reduced-order polynomial is obtained by the proposed method R2 (s), while this is
not the case for other reduced order interval mod-
P5 (s) = [4.7448, 5.6615] s5 + [23.3238, 24.1572] s4 els obtained by using impulse energy approximation
+ [55.2389, 55.9889] s3 of original interval systems using Kharitonov’s poly-
nomials, γ − δ and γ methods (see Equation (51)
+ [79.1587, 79.8254] s2
–(52)). If the zeroth-order time moment of the origi-
+ [63.2475, 63.8008] s + [2, 2.5] . (45) nal model and reduced-order model are matched, then
their steady-state values will surely match.
The second-order reduced-order polynomial is
A comparison with other existing methods avail-
p2 (s) = [6.7646, 9.1468] s2 + [24.8576, 25.9643] s
able in the literature is noted below to show the better
+ [2, 2.5] . (46) performance and efficiency of the proposed method.
By using the impulse energy approximation of orig-
The stability of the reduced-order interval polyno- inal interval systems using Kharitonov’s polynomials
mial can be easily verified by the Kharitonov theorem (Siva Kumar et al., 2015), the second-order model is
(1978). obtained as
The fifth-order reduced interval polynomial based
on the γ − δ method (Bandyopadhyay et al., 1997) and [1.172, 1.3682] s + [1.0269, 1.1097]
R2 (s) = . (50)
γ method (Sastry et al., 2000) is [1, 1] s2 + [2.344, 2.6232] s
+ [1.4033, 1.5164]
P5 (s) = [1, 1] s5 + [11.166, 67.3359] s4
By using the γ − δ-based Routh approximation
+ [29.9183, 222.3257] s3
method (Bandyopadhyay et al., 1997), the second-
+ [40.5241, 329.7038] s2 order model is obtained as
+ [25.9046, 213.8408] s + [0.6761, 7.0354] . [1.0091, 1.2554] s + [0.8409, 1.1168]
(47) R2 (s) = . (51)
[1, 1] s2 + [2.0181, 2.4430] s
It is noted that the above interval polynomial (P5 (s)) + [1.1492, 1.5007]
of Equation (47) is not stable since one of the four
By using the γ -based Routh approximation method
Kharitonov vertex polynomials P5 (s) of Equation (47)
(Sastry et al., 2000), the second-order model is
is not stable. Therefore, it is clear that the γ − δ
obtained as
method does not assure the stability of the reduced
order polynomial. [0.94, 1.35] s + [0.8409, 1.168]
R2 (s) = . (52)
[1, 1] s2 + [2.0181, 2.443] s
Example 4.3: Consider a third-order interval system + [1.492, 1.5007]
(Bandyopadhyay et al., 1997; Sastry et al., 2000)
The step response of the high-order system and
[2, 3] s2 + [17.5, 18.5] s + [15, 16] reduced-order models by the proposed method is
G(s) = . (48)
[2, 3] s3 + [17, 18] s2 + [35, 36] s shown in Figures 2 and 3. A comparison has been
+ [20.5, 21.5] made with the methods proposed by Bandyopad-
hyay et al. (1994, 1997), Sastry et al. (2000), Dol-
By using the proposed method, the second-order gin and Zeheb (2003), Chuan-qing and Yang (2010),
model is obtained as Pratheep et al. (2013) and Siva Kumar et al. (2015). The
[8.25, 9.75] s + [15, 16] comparison of the time-domain specifications of the
R2 (s) = . (49)
[5.12, 6.78] s2 + [23.12, 24.45] s proposed method with existing methods for a reduced
+ [20.5, 21.5] model is given in Table 4.
Figures 2 and 3 represent the original lower limit
The zeroth-order interval time moment of the original transfer function and upper limit transfer function,
higher order system [t0− , t0+ ] = ([b− + − +
0 , b0 ])/([a0 , a0 ]) respectively. The responses are compared with the pro-
of Gn (s) is equal to the reduced-order interval model posed and other existing techniques.
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 11

Figure 2. Comparison of step response (lower limit).

Figure 3. Comparison of step response (upper limit).

Table 4. Comparison of time domain specifications for lower limit.

Peak Rise Settling steady-state
Methods amplitude Overshoot time (s) time (s) final value
Original system 0.735 0.413 1.15 1.86 0.732
Proposed method 0.731 0 1.36 2.7 0.732
Siva Kumar et al. (2015) 0.74 1.08 1.2 1.84 0.732
Pratheep et al. (2013) 0.731 0 1.3 2.91 0.732
Chuan-qing and Yang (2010) 0.697 0 2.21 1.15 0.698
Dolgin and Zeheb (2003) 0.715 2.46 1.54 4.08 0.698
Sastry et al. (2000) 0.62 9.93 0.847 4.1 0.564
Bandyopadhyay et al. (1997) 0.747 2.13 1.29 3.47 0.732
Bandyopadhyay et al. (1994) 0.732 3.58 1.54 4.51 0.698

The Bode response of the high-order system and made with the methods proposed by Bandyopad-
reduced-order models by the proposed method is hyay et al. (1994, 1997), Sastry et al. (2000), Dol-
shown in Figures 4 and 5. A comparison has been gin and Zeheb (2003), Chuan-qing and Yang (2010),
12 K. KUMAR DEVEERASETTY AND S. K. NAGAR

Figure 4. Comparison of Bode response (lower limit).

Figure 5. Comparison of Bode response (upper limit).

Table 5. Comparison of time domain specifications for upper limit.

Peak Rise time Settling steady-state
Methods amplitude Overshoot (s) time (s) final value
Original system 0.745 0.416 1.05 1.67 0.7444
Proposed method 0.744 0 1.89 3.6 0.744
Siva Kumar et al. (2015) 0.734 0.358 1.11 1.83 0.732
Pratheep et al. (2013) 0.742 0 1.22 2.48 0.744
Chuan-qing and Yang (2010) 0.778 0 0.582 1.67 0.78
Dolgin and Zeheb (2003) 0.817 9.82 0.908 4.58 0.7444
Sastry et al. (2000) 0.79 1.49 1.09 1.64 0.778
Bandyopadhyay et al. (1997) 0.752 1.07 1.14 1.76 0.744
Bandyopadhyay et al. (1994) 0.81 8.86 0.883 4.56 0.744

Pratheep et al. (2013) and Siva Kumar et al. system and lower order models obtained by the
(2015). proposed method and with available methods from
Table 4 represents the comparison of the time- the literature. Table 5 represents the comparison of
domain specifications for the lower limit of the original the time domain specifications for the upper limit of
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 13

Table 6. Comparison of ISE.

S.No Methods Lower limit Upper limit
1 Proposed method 3.43581 × 10−6 6.63119 × 10−8
2 Siva Kumar et al. (2015) 2.93246 × 10−6 0.001205443
3 Pratheep et al. (2013) 1.19446 × 10−5 9.93222 × 10−7
4 Chuan-qing and Yang (2010) 0.010817047 0.010362791
5 Dolgin and Zeheb (2003) 0.008876522 8.01481 × 10−5
6 Sastry et al. (2000) 0.225673362 0.00949689
7 Bandyopadhyay et al. (1997) 2.36319 × 10−5 4.39626 × 10−6
8 Bandyopadhyay et al. (1994) 0.00878538 8.87354 × 10−5

the original system and lower order models obtained

by the proposed method and with available methods
from the literature. It has been noted from Tables 4 Figure 6. Comparison of impulse response (lower limit).
and 5 that the peak amplitude and steady-state final
value of the original system and reduced-order sys-
tem obtained by the proposed method are equal. This
shows that the proposed method closely matches with
the original system. The Integral Square Error values
for different lower order models are also tabulated in
Table 6.

Example 4.4: Consider a third-order discrete-time

interval system given by

[1, 2] z2 + [3, 4] z + [8, 10]

G(z) = . Figure 7. Comparison of impulse response (upper limit).
[6, 6] z3 + [9, 9.5] z2 + [4.9, 5] z + [0.8, 0.85]
(53)
By using the proposed method, the second-order
model is obtained as Case 1: Modified differentiation impulse response of the lower and upper bounds of the
method high-order system and reduced-order system obtained
by using the proposed method are shown in Figures 6
[1, 2.5] z + [8, 10] and 7. The step responses of the lower and upper
R2 (z) = .
[2.6678, 3.5006] z2 + [3.2335, 3.3668] z bounds of the high-order system and reduced-order
+ [0.8, 0.85] system obtained by using the proposed method are
(54) shown in Figures 8 and 9. In Table 7, the compari-
Case 2: Modified differentiation method by using lin- son of the error index has been verified and compared
ear transformation with the existing techniques. The important limitation
[1, 5.5] z + [6.5, 15] of Kharitonov’s theorem is that it cannot be applied
R2 (z) = . (55) directly to discrete-time interval polynomials. To over-
[8.685, 9.518] z2 + [7.878, 11.01] z
come this limitation, bilinear transformation is used in
+ [1.005, 3.954]
Kharitonov’s theorem for studying the stability of the
Case 3: Modified differentiation method by using discrete-time interval systems (Mastorakis, 1997).
bilinear transformation
[−11.9991, 16.0001] z2 5. Conclusions
+ [−4.0024, 30.6632] + [26.6695, 54.6687] The paper has presented the extension of the dif-
R2 (z) = .
[34.8844, 37.9344] z2 ferentiation method to interval systems. The differ-
+ [35.9658, 40.0658] z + [8.4498, 11.4998] entiation method has been proposed based on the
(56) modified Routh array. The comparison of the reduced-
By using bilinear transformation, the obtained reduced- order interval model step responses provides an excel-
order model is a non-minimum phase system. The lent accurate approximation to the original interval
14 K. KUMAR DEVEERASETTY AND S. K. NAGAR

Figure 8. Comparison of step response (lower limit).

Figure 9. Comparison of step response (upper limit).

Table 7. Comparison of error index. reduced-order model are equal, whereas the same is
S.No Methods Lower limit Upper limit not the case with other methods. ISE values are mea-
1 Proposed method (MDM) 107.2958 134.1428 sured and compared with the proposed method and
2 Proposed method 1.3428 7.4154 with existing techniques to show the high quality and
(MDM by using linear
transformation) efficiency of the proposed method. It is determined
3 Proposed method (MDM 12.0829 26.5229
by using bilinear
that the proposed reduced interval model demon-
transformation) strates good performance, and the majority of the
4 Ismail et al. (1997) 9.9116 2.4058
5 V. P. Singh 3.4249 0.7721 necessary characteristics of the original interval sys-
and Chandra (2012) tem are preserved. To prove the powerfulness of the
proposed method, four numerical examples of con-
tinuous interval systems and discrete interval systems
model. The proposed method proves that the zeroth-
are considered. The proposed method facilitates the
order interval time moments of the original model and
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 15

easy computation of the reduced interval models. The a professional member of ACDOS and IEEE member. His
illustrate examples conclude that the proposed method research interests include model order reduction, the stability
of extension of the differentiation method to interval of interval systems, modelling and control design for UAVs and
Sliding mode control.
systems will guarantee the stability of the reduced
model when compared with other available methods. S. K. Nagar was born in Varanasi, India,
In case of reduction of discrete-time interval sys- on 10 May 1955. He studied at Institute
of Technology (IT) BHU and received the
tems, three different cases are considered to study the
B.Tech and M.Tech degrees in electrical
stability and accuracy of the reduced-order models, engineering in 1976 and 1978, respec-
and the following three cases are (1) direct modified tively. He completed a PhD degree in elec-
differentiation method; (2) applying linear approxima- trical engineering from the University of
tion by substituting (z = w + 1); (3) applying bilinear Roorkee, Roorkee, India. He joined IT,
transformation by applying (z = (1 + w)/(1 − w)). A BHU in 1980 as a lecturer with the Department of Electrical
Engineering and as a professor in 2001 to till date. He was
illustrate example is used to show the stability and
National Advisory Committee member for the National Sys-
accuracy of the proposed method on discrete-time tem Conference in 2008 (NSC2008) and Technical Program
interval systems. Committee member for the Student’s Conference on Engineer-
The proposed method has the following advantages: ing and Systems (SCES-2014). He received Best Paper award at
(1) it is computationally simple; (2) it can guarantee the National System Conference (NSC-2011). He is the life member
stability of the reduced-order interval systems; (3) the of System Society of India and Fellow of Institute of Engineers.
His field of interest is the model reduction, digital control and
proposed technique is applicable for both continuous
discrete event systems. He has supervised few PhD theses in
and discrete-time interval systems; (4) the proposed the area of model reduction and controller design. He has pub-
technique can achieve zeroth-time moment match- lished numerous papers in regular peer-reviewed journals and
ing and (5) generates accurate reduced-order response National and International Conferences.
with respect to the original systems.

Acknowledgements References
This work was done at IIT-BHU, Varanasi, India. Authors Alefeld, G., & Mayer, G. (2000). Interval analysis: Theory and
are greateful to Dr. Shyam Kamal, IIT-BHU and Prof. Abner applications. Journal of Computational and Applied Mathe-
Ramirez Vazquez, Cinvestav- Guadalajara, Mexico for their matics, 121( 1–2), 421–464. https://doi.org/10.1016/S0377-
suggestions to improve the quality of the paper. 0427(00)00342-3
Anderson, B. D. O., Jury, E. I., & Mansour, M. (1987). On robust
Disclosure statement Hurwitz polynomials. IEEE Transactions on Automatic Con-
trol, 32(10), 909–913. https://doi.org/10.1109/TAC.1987.
No potential conflict of interest was reported by the author(s). 1104459
Antoulas, A. C. (2005). Approximation of large scale dynamical
systems. SIAM Publications.
Notes on contributors
Antoulas, A. C., Sorensen, D. C., & Gugercin, S. (2006). A
Kranthi Kumar Deveerasetty received the survey of model reduction methods for large scale systems.
bachelor’s degree in electrical and elec- Contemporary Mathematics, 280, 193–220. https://doi.org/
tronics engineering from Sir. C. R. Reddy 10.1090/conm/280
College of Engineering, Eluru, A.P, India Argoun, M. B. (1990). On the stability of low-order perturbed
in 2008. He completed M.Tech and the polynomials. IEEE Transactions on Automatic Control, 35(2),
PhD degree in electrical engineering from 180–182. https://doi.org/10.1109/9.45174
IIT (BHU), Varanasi, India in 2010 and Bandyopadhyay, B., Avinash, U., & Ismail, O. (1997). γ − δ
2016 respectively. He worked as a Post- Routh approximations for interval systems. IEEE Trans-
doctoral Fellow at Shenzhen Institutes of Advanced Tech- actions on Automatic Control, 42(8), 1127–1130. https://
nology, Chinese Academy of Sciences, Shenzhen China from doi.org/10.1109/9.618241
August 2017- November 2019. He is currently working as a Bandyopadhyay, B., Ismail, O., & Gorez, R. (1994). Routh-
Post-Doctoral Fellow at Kochi University of Technology, Japan. Pade approximation for interval systems. IEEE Transactions
He has published 29 journal articles and conference papers. on Automatic Control, 39(12),2454–2456. https://doi.org/
Achieved best application paper award at the 11th international 10.1109/9.362850
conference on modelling, identification and control (ICMIC Barmish, B. R. (1994). New tools for robustness of linear systems.
2019). He is a life member in the System Society of India (SSI), Macmillan.
16 K. KUMAR DEVEERASETTY AND S. K. NAGAR

Bhattacharyya, S. R., Chapellat, H., & Keel, L. H. (1995). Robust Ismail, O., Bandyopadhyay, B., & Gorez, R. (1997). Discrete
control the parametric approach. Prentice Hall PTR. interval system reduction using Pade approximation to allow
Bose, N. K., Jury, E. I., & Zeheb, E. (1988). On robust Hur- retention of dominant poles. IEEE Transactions on Circuits
witz polynomials. IEEE Transactions on Automatic Control, and Systems, 44(11), 1075–1078. https://doi.org/10.1109/
33(12), 1166–1168. https://doi.org/10.1109/9.14442 81.641771
Choo, Y. (2007). A note on discrete interval system reduc- Jamshidi, M. (1983). Large scale systems, modelling and control.
tion via retention of dominant poles. International Journal Elsevier.
of Control, Automation, and System, 5(2), 208–211. Jamshidi, M., & Malek-Zavarei, M. (1986). Linear control sys-
Choudhary, A. K., & Nagar, S. K. (2017). Novel arrange- tems, A computer-aided approach. International Series on
ment of Routh array for z-domain uncertain system. Systems and Control (pp. 161–200). Pergamon Press.
System Science & Control Engineering, 5(1), 232–242. Kharitonov, V. L. (1978). Asymptotic stability of an equilibrium
https://doi.org/10.1080/21642583.2017.1311239 position of a family of system of linear differential equation.
Choudhary, A. K., & Nagar, S. K. (2018a). Model order reduc- Differential Equations, 14(11), 1483–1485.
tion of discrete-time interval system based on Mikhailov sta- Kranthi Kumar, D., Nagar, S. K., & Tiwari, J. P. (2011a). Model
bility criterion. International Journal of Dynamics and Con- order reduction of interval systems using Mihailov criterion
trol, 6(4), 1558–1566. https://doi.org/10.1007/s40435-018- and Cauer second form. International Journal of Computer
0404-4 Applications, 32(6), 17–21. https://doi.org/10.5120/ijca
Choudhary, A. K., & Nagar, S. K. (2018b). Order reduction in Kranthi Kumar, D., Nagar, S. K., & Tiwari, J. P. (2011b). Model
z-domain for Interval system using an arithmetic operator. order reduction of interval systems using Mihailov criterion
Circuits, Systems, and Signal Processing, 38(3), 1023–1038. and factor division method. International Journal of Com-
10.1007/s00034-018-0912-7 puter Applications, 28(11), 4–8. https://doi.org/10.5120/ijca
Choudhary, A. K., & Nagar, S. K. (2018c). Order reduction tech- Kranthi Kumar, D., Nagar, S. K., & Tiwari, J. P. (2013a, Febru-
niques via Routh approximation. IETE Journal of Research, ary 21–22). Direct truncation and factor division method for
65(3), 1–8. https://doi.org/10.1080/03772063.2018. order reduction of interval systems. National Conference on
1530074 Recent Developments in Control, Automation and Power
Choudhary, A. K., & Nagar, S. K. (2013, December 13–15). Engineering, Amity University, Noida, India.
Direct truncation method for order reduction of discrete inter- Kranthi Kumar, D., Nagar, S. K., & Tiwari, J. P. (2013b). A
val system. IEEE Annual INDICON Conference, Mumbai, new algorithm for model order reduction of interval systems.
India. Bonfring International Journal of Data Mining, 3(1), 6–11.
Chuan-qing, G. U., & Yang, J. (2010). Stable Routh-Pade https://doi.org/10.9756/BIJDM
type approximation in model reduction of interval sys- Kranthi Kumar, D., Zhou, Y., Kamal, S., & Nagar, S. K. (2019).
tems. Journal Shanghai University, 14(5), 369–373. Springer. Computation of impulse-response Gramian for interval
https://doi.org/10.1007/s11741-010-0661-1 systems. IETE Journal of Research. https://doi.org/10.1080/
Dolgin, Y. (2005). Author’s reply. IEEE Transactions on 03772063.2019.1690592
Automatic Control, 50(2), 274–275. https://doi.org/10.1109/ Kumar, D., & Nagar, S. K. (2014). Model reduction by extended
TAC.2005.843849 minimal degree optimal Hankel norm approximation.
Dolgin, Y., & Zeheb, E. (2003). On Routh-Pade model Applied Mathematical Modelling, 38(11–12), 2922–2933.
reduction of interval systems. IEEE Transactions on Auto- https://doi.org/10.1016/j.apm.2013.11.012
matic Control, 48(9), 1610–1612. https://doi.org/10.1109/ Kumar, D., Tiwari, J. P., & Nagar, S. K. (2012a). Controller
TAC.2003.816999 reduction by balanced approach with guaranteed closed
Fortuna, L., Nunnari, G., & Gallo, A. (1992). Model order reduc- loop performance. International Journal of Automation and
tion techniques with applications in electrical engineering. Control, 6(2), 105–119. https://doi.org/10.1504/IJAAC.2012.
Springer- Verlag. 048644
Glover, K. (1984). All optimal Hankel-norm approxima- Kumar, D., Tiwari, J. P., & Nagar, S. K. (2012b). Reducing order
tions of linear multivariable systems and their L∞ error of large scale system by extended balanced singular pertur-
bounds. International Journal of Control, 39(6), 1115–1193. bation approximation. International Journal of Automation
https://doi.org/10.1080/00207178408933239 and Control, 6(1), 21–33. https://doi.org/10.1504/IJAAC.
Gutman, P., Mannerfeltand, C. F., & Molander, P. (1982). Con- 2012.045438
tributions to the model reduction problem. IEEE Transac- Lucas, T. N. (1992). Some further observations on the dif-
tions on Automatic Control, 27(2), 454–455. https://doi.org/ ferentiation method of model reduction. IEEE Transactions
10.1109/TAC.1982.1102930 on Automatic Control, 37(9), 1389–1391. https://doi.org/
Hayes, B. (2003). A lucid interval. American Scientist, Sigma Xi, 10.1109/9.159578
91(6), 484–488. https://doi.org/10.1511/2003.6.484 Mahmoud, M. S., & Singh, M. G. (1981). Large scale systems
Hwang, C., & Yang, S. F. (1999). Comments on the computa- modelling. (pp. 156–166). Pergamon Press.
tion of interval Routh approximants. IEEE Transactions on Mastorakis, N. E. (1997). Robust stability of polynomials: New
Automatic Control, 44(9), 1782–1787. https://doi.org/10.1109/ approach. Journal of Optimization Theory and Applications,
9.788553 93(3), 635–638. https://doi.org/10.1023/A:1022603415620
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 17

Obinata, G., & Anderson, B. D. O. (2001). Model reduction for Sinha, N., & Kuszta, B. (1983). Modelling and identification
control system design. Springer- Verlag. (p. 133). Van Nostrand Reinhold.
Popeea, C., & Jora, B. (2004). Model order reduction with dis- Siva Kumar, M., Vijay Anand, N., & Srinivasa Rao, R.
sipativity preservation. Journal of Control Engineering and (2015). Impulse energy approximation of higher order
Applied Informatics, 6(1), 27–33. http://www.ceai.srait.ro/ interval system using Kharitonov’s polynomials. Transac-
index.php?journal = ceai&page = article&op = view&path. tions of the Institute of Measurement and Control, 1–11.
Popeea, C., & Jora, B. (2007). L2 -normorder reduction. Journal https://doi.org/10.1177/0142331215583326
of Control Engineering and Applied Informatics, 9(2), 5–10. Wilhelmus, H. A., & Rommes, J. (2008). Model order reduction.
http://www.ceai.srait.ro/index.php?journal = ceai&page = Theory, research aspects and applications. Springer Verlag.
article&op = view&path. Yang, S. F. (2005). Comments on Routh-Pade model reduction
Pratheep, V. G., Ramesh, K., & Venkarachalam, K. (2013). of interval systems. IEEE Transactions on Automatic Control,
Reduced order modeling of uncertain systems by pole clus- 50(2), 273–274. https://doi.org/10.1109/TAC.2004.841885
tering technique using genetic algorithm. IEEE-Fourth Inter-
national Conference on Computing, Communications and
Networking Technologies, Tiruchengode, India.
Appendix
   
Saini, D. K., & Prasad, R. (2010). Mixed evolutionary tech- Interval e = e− , e+  = e ∈: e− ≤ e ≤ e+ 
niques to reduce order of linear interval system using gen- Interval f = f − , f + = f ∈: f − ≤ f ≤ f +
eralized Routh array. International Journal of Engineering, Midpoint m (ai.1 ) = 12 a− +
i,1 , ai,1
n 
Science and Technology, 2(10), 5197–5205.  + i
Sastry, G. V. K. R., Raja Rao, G., & Mallikarjuna Rao, Interval Polynomial  (s) = a−
i , a+ s
i=0
P. (2000). Large scale interval system modelling using Lower limit e− , f −
Routh approximants. Electronics Letters, 36(8), 768–769. Upper limit e+ , f +
https://doi.org/10.1049/el:20000571 Order of the higher order systems n
Selvaganesan, N. (2007). Mixed method of model reduction for High-order interval transfer function Gn (s)
uncertain systems. Serbian Journal of Electrical Engineering, High-order numerator interval polynomial N (s)
4(1), 1–12. https://doi.org/10.2298/SJEE0701001S High-order denominator interval polynomial D (s)
Shamash, Y., & Feinmesser, D. (1978). Reduction of dis- Reduced-order interval transfer function Rk (s)
crete time systems using a modified Routh array. Interna- Reduced-order numerator interval polynomial Nk (s)
tional Journal of Systems Science, 9(1), 53–64. https://doi.org/ Reduced-order denominator interval polynomial Dk (s)
10.1080/00207727808941678 Order of the reduced interval system k
Singh, V. P., & Chandra, D. (2012). Reduction of discrete inter- Interval midpoint αi,1     
val system using clustering poles with Pade approximation: Arithmetic operations [e] o f = eof /e ∈ [e] , f ∈ f ;
A computer aided approach. International Journal of Engi- o ∈ {+, −, ×, ÷}
neering, Science and Technology, 4(1), 97–105. Endpoint formulas for arithmetic operations
Singh, S. K., & Nagar, S. K. (2004). An algorithmic approach e + f = [e− + f − , e+ + f + ]
for system decomposition and balanced realized model e − f = [e− − f + , e+ − f − ]
reduction. Journal of Franklin Institute, 341(7), 615–630. e · f = [min Q, max Q], Q = {e− f − , e− f + , e+ f − , e+ f + }
https://doi.org/10.1016/j.jfranklin.2004.07.005 e/f = e · (1/f ); 1/f = [1/f + , 1/f − ], 0 ∈
/f