IMA Journal of Mathematical Control & Information (1998) 15, 53-62

A Lyapunov analysis of stability robustness for discrete linear descriptor systems

D. LJ DEBEUKOVIC 1 , V. B. BAJIC 2 , T. N. ERIC 1 , AND S. A. MILINKOVIC3

department of Control Engineering, Faculty of Mechanical Engineering, University of Belgrade, 27 Marta 80, 11000 Beograd, Yugoslavia 2 Centrefor Engineering Research, Technikon Natal, P.O. Box 953, Durban 4000, Republic of South Africa 3[ System Control Group, Faculty of Technology and Metallurgy, University of Belgrade, Karnegijeva 4, Beograd, Yugoslavia [Received 12 July 1995 and in revised form 25 March 1996] Discrete descriptor systems are those for which the dynamics are governed by a mixture of algebraic and difference equations. This paper examines the existence of solutions that are attracted by the origin of the phase space, for regular and irregular discrete linear descriptor systems. By a suitable transformation, the original system is transformed to a convenient form that enables development and easy application of Lyapunov's direct method for the existence analysis of a subclass of solutions characterized by convergence to the origin. A potential (weak) domain of attraction of the origin is underestimated on the basis of a symmetric positive definite solution of a reduced-order discrete Lyapunov matrix equations. Also, it has been shown that the same result can be efficiently used in determining quantitative measures of robustness for a class of perturbed discrete linear descriptor systems.

Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012

1. Introduction We consider a class of discrete linear descriptor systems (DLDS) for which the dynamics are governed by Ey{k+\)
nxn

= Ay{k)

(k = ko,ko + l,...),

y(k0) = y0

(1.1)

with E,A e R , where y e R" is the phase vector (i.e. the generalized state-space vector) and where the matrix E may be singular. Discrete systems with models of the form (1.1) are known as linear descriptor (as well as singular, semistate, generalized state-space, algebraic-difference) systems. The notion of discrete descriptor system was introduced by Luenberger (1977). The collections of results relating to different aspects continuous and discrete descriptor systems can be found in the books of Aplevich (1991), Bajic (1992), Campbell (1980, 1982), Dai (1989a), and in two special issues of the journal Circuits, Systems and Signal Processing (1986, 1989). The complex nature of descriptor systems causes many difficulties in numerical and analytic treatment that do not appear when systems in the normal form are concerned. Qualitative and stability analysis of discrete descriptor systems have been treated by several authors. Campbell & Rodrigues (1985) investigated bounds of response of discrete nonlinear descriptor systems. With the use of Lyapunov's
53
1' Oxford Um'nft)< Pra I99S

54

D. LJ. DEBELJKOVIC, V. B. BAJlt, T. N. ERIC, AND S. A. MILINKOVIC

direct method, some particular classes of nonlinear nonstationary discrete descriptor systems were studied by Milic & Bajic (1984) and Bajic et al. (1990). Large-scale discrete descriptor systems were also analysed by Milic & Bajic (1984). Owens & Debeljkovic (1985) derived some necessary and sufficient conditions for asymptotic stability of a DLDS on the basis of geometric considerations. Finite-time and practical stability results for a DLDS were given by Debeljkovic & Owens (1986), Owens & Debeljkovic (1986), Bajic (1995), and Debeljkovic et al. (1995). Owens & Debeljkovic (1985) investigated the geometric description of initial conditions that generate solution sequences [y(k) : k = 0,1,...]. The results were expressed directly in terms of matrices E and A, and avoid the need to introduce algebraic transformations as a prerequisite for stability analysis. In that sense, the geometric approach provided a possibility for a basis-free analysis of dynamic properties of this class of systems. However, at the moment, there are no feasible methods for testing the conditions they obtained. The problem stems from the fact that the properties of certain matrices are required to hold only on a linear submanifold of the system's phase space and not in the whole phase space. Physical systems are very often modelled by idealized and simplified models, so that information obtained on the basis of such models is not always sufficiently accurate. This motivates investigating the robustness of properties of the examined system with respect to model inaccuracies. Quantitative measures of robustness for multivariable systems in the presence of time-dependent nonlinear perturbations were first investigated in Patel & Toda (1980). The bounds were obtained for the perturbation vector such that the nominal system remains stable. The stability robustness of linear discretetime system in the time domain using Lyapunov's approach was treated by Kolla et al. (1989). Bounds on linear time-dependent perturbations that maintain the stability of an asymptotically stable nominal system are obtained for both structured and unstructured independent perturbations. A general overview of results concerning problems of stability robustness in the area of nonlinear time-dependent descriptor systems was published by Bajic (1992), while some other robustness results for linear descriptor systems are presented by Dai (1989a). This paper presents the new results on the attraction of the origin for a DLDS. The results are similar to those given for continuous linear singular systems by Bajic et al. (1992). In the second part of the paper (Section 3) the Lyapunov stability robustness of both regular and irregular DLDSs is considered. In Section 4, the bounds on the perturbation matrix are determined so that the attraction property of the origin of the nominal system is preserved for all perturbation matrices of a specific class. The results presented here rely to some extent on those given by Bajic (1995) and Debeljkovic et al. (1995). 2. Preliminaries The model (1.1) of a DLDS can be transformed into a more convenient form for the intended analysis by a suitable transformation. For that purpose, consider a DLDS (1.1). It is assumed that the matrix E is in the form E = diag (l B l ;O n j ), where \p and Op stand for the p x p identity matrix and the p x p null matrix, respectively. If the matrix E is not in this form, then the transformation i- TEQ, where Tand

Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012

LYAPUNOV STABILITY ROBUSTNESS FOR DISCRETE LINEAR DESCRIPTOR SYSTEMS

55

Q are suitable nonsingular matrices (Dai 1989a), can convert it to that form, and a broad class of systems (1.1) in this way can be brought to the form xl(k+l) = Aixi(k) + A2x2(k), 0 = AiXl(k) + A4x2(k), (2.1a,b)

where the column vector x = ( i | , x 2 ) R" is composed of subcolumns x{(k) R"1 and x2{k) = R"2, with n = nx + n2. Let k e K. denote the current discrete moment; here K = {k0, ko + 1,...} is a discrete time interval, where k0 is an integer. The matrices At (i = 1,...,4) are of appropriate dimensions. The form (2.1) for DLDS is also known as the second equivalent form (Dai 1989a). For the system (2.1), d e t = 0. Since the system considered is time-invariant, it is sufficient to consider its current solution value x(k) as depending only on the current discrete moment k and initial value x 0 at the initial moment k0. Hence let x(k, x0) denote the value of a solution x of (2.1), at the moment k e AC, which emanated from JC0 at k = k0. In an abbreviated form the value of solution x at the moment k will be denoted by x(k). When matrix pencil {cE - A : c C} is regular, i.e. when there exists c G C such that det(cE-A)?0, (2.2)

Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012

then solutions of (1.1) exist and are unique for the so-called consistent initial values y0. Moreover a closed form of the solutions exists (Campbell 1980). If A4 is nonsingular, then the regularity conditions (2.2) for the system (2.1) is considerably simplified and reduces to det(clfl, - Ax)det [-A4- A3(clHl - AX)~XA2] = (-iy"det/l4det[(cI n | -Al)+ A2A4
l

A3] ^ 0.

(2.3)

It was proved by Owens & Debeljkovic (1985) that, under the conditions of an appropriate lemma, y0 is a consistent initial condition for (1.1) if ,y0 belongs to a certain subspace W of consistent initial conditions. Moreover, y0 generates a discrete solution sequence [y(k) : k 0,1,...] (in this case k0 = 0) such that y(k) e W for all k = 0,1,.... The subspace is given by W = N(1 EED), where ED is the so-called Drazin inverse of E = (cE - A)~x, with N ( ^ ) denoting the kernal (null space) of the operator X; note that W is independent of the particular choice of c, which can be any complex number such that cE - A is nonsingular. Remark. The following discussion of the consistent initial values is taken from Bajic (1995). Let us denote the set of the consistent initial values of (2.1) by M\. Consider the manifold M C R" determined by (2.1 b) as M = {x R" : 0 = A3xt +AAx2}. For systems (2.1) in the general case, M \ Q M. Thus a consistent value x 0 = (x| 0 , x M ) has to satisfy 0 = /l 3 x 10 + A^-^, or equivalently x o e.M, C M = However, if rank [A3, A4] = rank A4, (2.5)

56

D. U. DEBEUKOVIC, V. B. BAJIC, T. N. ERIC, AND S. A. MILINKOVIC

then Mi = M = N([Ait AA]), and the determination of Mi requires no additional computation, except to convert (1.1) into the form (2.1). In this case, under the assumption that rank AA = r <n2, when x0 N([y43,/l4]) (i.e. when x 0 satisfies (2.1b)), ! + n2 r components of the vector x0 can be chosen arbitrarily in order to achieve consistent initial conditions of the system governed by (2.1). More precisely, all components of x 10 and some (the case rank AA = r < n2) or none (the case rank A4 = n2) of the components of x2o are free for choice. In both these cases, (2.1) is reducible to a lower-order normal form system with Xi as the state variable and for the case rank A4 = r < n2 with some of the components of x2 as free. In the later case, the reduced-order model will be of the form x,(& + 1) = F,x,(A:) 4- F2x*2{k), where Fx and F2 are matrices of appropriate dimensions, and where x*2 represents a vector composed of those components of x2 that can be chosen as free. Thus the existence of solutions is guaranteed for such selected x 0 , and M\ - M = N([^3,y44]). Note that the uniqueness of solutions is not guaranteed when rank At, = r < n2. Since the transformation from (1.1) to (2.1) is nonsingular, the convergence of solutions y(k) of (1.1) toward the origin of the phase space of (1.1) and that of x{k) toward the origin of (2.1) are equivalent problems. Let | | denote the euclidean vector norm or induced matrix norm. The potential (weak) domain of attraction of the null solution x(k, 0) = 0 {k A ) of (2.1) is defined by C A = {x0 M\. there exists \x(k): k = 0,l,...]
*->oo

Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012

satisfying (2.1)

such that x(0) = x 0 and Urn \x(k)\ = 0}. The term 'potential' or 'weak' is used because solutions of (2.1) need not be unique; thus, for every x0 A, there also may exist solutions which do not converge toward the origin. Our task is to estimate the set A. Lyapunov's direct method will be used to obtain an underestimate Au of the set A (i.e. Au QA). Our development will not require the regularity conditions (2.2) of the matrix pencil {cE - A : c 6 C}. Some other aspects of irregular singular systems were considered by Dziurla & Newcomb (1987) and Dai (1989b).

3. Attraction of the origin and its potential domain This section introduces a stability result which will be employed for the robustness analysis of the attraction of the origin. We assume that the rank condition (2.5) holds, which implies M\ = N([/43,/<4]) for the systems (2.1). Consequently, there exists a matrix L which satisfies the matrix equation O = Ai + AAL, (3.1)

where O is the null matrix of the same dimensions as 4 3 . From (2.5) and (3.1), it follows that, if solutions of (2.1) exist, then there will be solutions x whose components satisfy Lxi(k) (keK). (3.2)

LYAPUNOV STABILITY ROBUSTNESS FOR DISCRETE LINEAR DESCRIPTOR SYSTEMS

57

Under the rank condition (2.5), it follows (Bajic 1995) that N([L, - L J ) C N([A3,A4}). To show this, consider an arbitrary x* N([L, - I n J ) , i.e. x2 Lx\, where L is any matrix that satisfies (3.1). Then, multiplying (3.1) from the right by xx and using (3.2), one gets 0 = A3x] + A4Lx\ = A^x\ + A4x2, which shows that x* 6 N([y43,/44]). Hence N([L, - I n J ) C N([i43,/44]). Consequently those solutions of (2.1) that satisfy (3.2) also have to satisfy the constraints (2.1b). For all solutions of (2.1) for which (3.2) holds, the following conclusions are important. The solutions of (2.1) belong to the set N([L, -L,J). If, under the rank conditions (2.5), the existence of a solution x of (2.1) which satisfies (3.2) and converges to the origin is proved, then the potential domain of attraction of the origin for (2.1) can be underestimated by Au=N([L,-ln2])CA. (3.3) Let pd and nd stand for positive definite and negative definite, respectively. For the system (2.1), the Lyapunov function can be selected as v[x(k)]=x](k)Hxl(k), (3.4)

Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012

where H is pd symmetric. The total time difference of v is defined by the expression Av[x(k)} = y[x(k+l)}V[x(k)}, (3.5)

and its value, calculated along the solutions of (2.1), is then Av[x(k)} = x](k+\)Hxl(k+\)-x](k)Hx](k) = a]+a1+a2+a3-xl(k)Hxl(k), where al=x](k)A]HAlxl(k), a2=xT2(k)AJ2HAixi(k), ai=x[{k)A12H A2x2(k). (3.7) Employing (3.2) and (3.6), one obtains Av[x(k)) = x](k)[(At +A2L)TH(Ai where Z=-(A{+ A2V?H{AX + A2L) + H; note that Z is a real symmetric matrix. Since H is pd matrix, v is a pd function with respect to x\. Thus, if Z is pd, then K[*(&)] would tend to zero as k oo, provided that the solutions x exist when k > oo. This would imply |xi(/c)| 0 when k * oo. Finding the pair of matrices H and Z to comply with these requirements can be achieved by means of a discrete Lyapunov matrix equation A\HAL-H = -Z, with AL = Ax + A2L, where Z can be an arbitrary real symmetric pd matrix, and the corresponding H, (3.8) + A2L) - //]*,(*) = -x}{k)Zx,{k), (3.6)

58

D. LJ. DEBEUKOVIC, V. B. BAJIC, T. N. ERIC, AND S. A. MILINKOVIC

which is a symmetric pd matrix, can be found as a unique solution of (3.8), if and only if AL is a discrete stable matrix, i.e. a matrix whose eigenvalues lie in the open unit circle of the complex plane. Note that the order of H is nx x nit so that (3.8) can be considered as reduced-order discrete Lyapunov matrix equation when compared to the number of the components of x for the system (2.1). We are now in a position to state the following result. 1 Let (2.5) hold. Then the underestimate A u of the potential domain A of attraction of the null solution DLDS (2.1) is determined by (3.3), provided that L is any solution of (3.1) and AL= A] + A2L is a discrete stable matrix. Moreover, Ay contains more than one element.
THEOREM

Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012

Proof. From (2.5), it follows that M , = H([AitA4]). Let L be any solution of (3.1). Note that such L always exists when (2.5) holds. Now select x0 N[L, - I n J ) . This is a consistent initial value at k = k^, since N([L - L,J) C N([i43,i44]) = M\. Then solutions x(k,x0) of (2.1) that emanate from the point x0 exist. Let us show that some of them have the form (3.2). Thus assume that (3.2) is valid for a solution x(k,x0) = (xl(k,x0),x2{k,x0)). Note that (3.2) implies that the system (2.1) can be reduced to a linear state-space with xi as the state variable and some (the case rank A4 = r < n2) or none (the case rank A2 = n2) of the components of x2 chosen as free. In both cases, the solutions of the reduced-order model exist. For the case rank AA = r < n2, the remaining components of x2 could, for each k, be calculated from (2.1b) on the basis of the values obtained for xt(k) and the values of the free components of x2. Note also that the free components of x2 could always be selected to equal zero. Thus there are solutions of (2.1) that have the form (3.2). To examine the behaviour of these solutions x(k,x0), we use the function v denned in (3.4) and utilize the expression (3.7). Select Z to be any symmetric real pd matrix. Since AL A\ + A2L is a discrete stable matrix, there exists a unique symmetric pd matrix H satisfying (3.8). Hence K(X) denned by (3.4) is a pd function with respect to xu and its total difference along the solutions of (2.1) that satisfy constraints (3.2) is nd. So \xi(k,xo)\ > 0 as k* oo, as long as x0 NQL.-LJ). But (3.2) implies also |x 2 (*,x 0 )| = |Lx,(Mo)l < |L||x,(fc,xo)| - 0 as k > oo. Since N([L, - I n J ) is not a singleton in the space of x, then there are solutions of (2.1) with the initial value x o ^ O e R " that converge toward the origin of the phase space as k oo. Thus Av has more than one element.

4. Robustness of attraction This section gives results on the robustness of attraction of the origin under unstructured and structured perturbation in the model of DLDS. We consider a perturbed version of (1.1) which has the form Ey{k + 1) = Ay{k) + APy(k), j f o ) = y0, (4.1) where the matrix AP represents the perturbations in the model. To analyse the robustness

LYAPUNOV STABILITY ROBUSTNESS FOR DISCRETE LINEAR DESCRIPTOR SYSTEMS

59

of attraction of the origin of (4.1), we consider (4.1) transformed to the form x,(*+ 1) = M, +B,)x,(*) + (A2 + B2)x2(k), 0 = Ayxx (k) + A4x2(k) + B^xik), (4.2a) (4.2b)

where x {xx,x2) need not represent the original variables y of the system (4.1). To simplify the formulation of results on stability robustness, we introduce the following assumption.
ASSUMPTION

1 The matrix B^ in (4.2b) is null.

Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012

To perform analysis of robustness for the system (4.2), we employ the Lyapunov function v defined by (3.4). Let the rank condition (2.5) hold. Then, taking into account (3.2) and (3.7), the expression Av given by (3.5) along the solutions of (4.2) is obtained as AK[x(*)] = xT(*)[(C, + C 2 L) T //(C, + C2L)]xx{k) - x]{k)Hxx{k) = x}(*)ZPx,(*), Z P = (C, + C2L)TH(Cl + C2L) - H, Ci = Al + Bi (i=l,2); (4.3a) (4.3b,c)

note that Z P is a real symmetric matrix. Define the singular values of a real matrix X to be the square roots of the eigenvalues of XTX. Let oM(X) denote the maximal singular value of a matrix X, while XM(S) and ^ ( S ) stand for maximal eigenvalue and minimal eigenvalue of a symmetric matrix 5 respectively. Note thatCTM(S) is also the spectral norm of S. Now we are in position to state the following result which concerns the unstructured perturbations in (4.2).
THEOREM 2

Let the rank condition (2.5), Assumption 1 and all conditions of Theorem 1 hold. Let Z and H be two real, symmetric, and pd matrices satisfying discrete Lyapunov matrix equation (3.8), and let AL be the matrix from (3.8). Then the underestimate An of the potential domain of attraction of system (4.2) is determined by (3.3) if aM(BL) < -oM(AL) + ( ^ ( A J + Y ^ ) , (4.4)

where BL = B\ + B2L. Moreover, An contains more than one element. Proof. The proof is based on the proof of Theorem 1. The only difference is that, along the solutions of (4.2), we have Ac[j(fe)] given by (4.3). Then it follows from (4.3b) and (4.3c) that Z P = (C, + C2L)JH(C] + C2L) - H = (A{+ A2L?H{A\ + A2L) - H + (5, + B2L)TH(Bl + B2L) + (Ax + ^ 2 L) T //(5, + B2L) + (Bt + B2L)JH{A\ + A2L) = -Z + BTHBL + A\HBL + B\HAU

60

D. LJ. DEBEUKOVIC. V. B. BAJIC, T. N. ERIC, AND S. A. MILINKOVIC

so that we get xjZpx, = -x]Zxi + X\B\HBLXX However,

x]BJLHBLx\ + 2x]BTLHALxl

+ 2x]BTLHALXi.

(4.5)

(4.6)
Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012

and also so that, combining (4.5)-(4.7), one gets that Z P is nd, provided that
^M(H)[M(BL) + 2GM(BL)OM(AL)] < ^(Z).

The last inequality is satisfied if (4.4) holds (the reasoning for this is the same as in the proof of Theorem 1 in Kolla et al. 1989). Thus (4.4) implies that Z P is nd and consequently Af/[x(/c)] is nd. The rest of the proof is the same as in the proof of Theorem 1. In order to cater for the structured perturbations in the model (4.2), we introduce the following assumption.
ASSUMPTION

2 Let BL = Bx + B2L = [>,, : i,j = l,...,/ii], where L is any solution of (3.1). Let the constraints \bti\ < ntj hold. Then we have our next result.

3 Let the rank condition (2.5), Assumptions 1 and 2, and all conditions of Theorem 1 hold. Let Z and H be two real symmetric pd matrices satisfying the discrete Lyapunov matrix equation (3.8), and let AL be the matrix from (3.8). Let K = maxi^^^rt, 7T/y. Then the underestimate A u of the potential domain of attraction of system (4.2) is determined by (3.3) if
THEOREM

*<-\-aM{AL)L

(4.8)

Moreover A u contains more than one element.

Proof. The proof uses the relations (see proof of Theorem 2 in Kolla et al. 1989) CT M(^i) = "i 7r f r t n e matrix />,, whose elements are all equal to n, and ^ M ( ^ ) < ^M (^i) because 7r/7 < n. The rest of the proof follows from Theorem 2.
5. Conclusion

The use of Lyapunov's direct method allows simple sufficient algebraic conditions to be derived for testing the existence of solutions of a discrete linear descriptor

LYAPUNOV STABILITY ROBUSTNESS FOR DISCRETE LINEAR DESCRIPTOR SYSTEMS

61

system (DLDS) which converge toward the origin of the system's phase space. The determination of the potential domain of attraction of the origin is also analysed for a class of time-invariant regular and irregular DLDSs. The results could be a basis for further development of similar existence analysis for completely general nonlinear and time-dependent discrete descriptor systems. Results presented in this paper give an indication for a possible convenient approach in that sense. The results are adapted to cater for the robustness of the attraction property of the phase-space origin for two different classes of perturbed DLDS.
REFERENCES

Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012

APLEVICH, J. D., 1991. Implicit linear systems. Springer-Verlag, Berlin. BAJIC V. B., 1992. Lyapunov's direct method in the analysis of singular systems and networks. Shades Technical Publicaitons, Hillcrest, RSA. BAJIC, V. B., 1995. Existence of practically stable solutions of singular linear systems. Technical Report TR95-O2, Control Laboratory, Technikon Natal, RSA. BAJIC V. B., MILIC, M. M., & ZrvKovic, J. S., 1990. Bounds on solutions of discrete singular systems. Proceedings of IEEE-ISC AS 90, May 1-3, New Orleans, Louisiana, USA. Vol 2, pp. 1482-4.
BAJIC, V. B., DEBEUKOVIC, D. LJ., GAJIC, S., & PETROVIC, B., 1992. Weak domain of attraction

and existence of solutions convergent to the phase space of singular linear systems. Univ. Belgrade, Publ. ETF (series on automatic control) No. 1, 53-62. CAMPBELL, S. L., 1980. Singular systems of differential equations. Pitman, Marshfield, Mass., USA. CAMPBELL, S. L., 1982. Singular systems of differential equations II. Pitman, Marshfield, Mass., USA. CAMPBELL, S. L., & RODRIGUES, J., 1985. Bounded solutions of discrete singular systems on infinite time intervals. IEEE Trans AC 30, 165-8. Circuits, Systems and Signal Processing, 1986. Special issue on semistate systems 5(1). Circuits, Systems and Signal Processing, 1989. Special issue on recent advances in singular systems 8(3). DAI, L., 1989a. Singular control systems. Springer-Verlag, Berlin. DAI, L., 1989b. The difference between regularity and irregularity in singular systems. Circuits, Systems and Signal Processing 8, 43544. DEBEUKOVIC, D. LJ., & OWENS, D. H., 1985. On practical stability of singular systems. Proceedings of MELECON 85, October 9-10, Madrid, Spain. Pp. 103-105. DEBEUKOVIC, D. LJ., & OWENS, D. H., 1986. On non-Lyapunov stability of discrete descriptor systems. Proceedings ofEUROCON86, April 21-23, Paris, France. Pp. 406-9.
DEBEUKOVIC D. L J . , BAJIC, V. B., GRGIC, A. U., & MILINKOVIC, S. A., 1995. Non-Lyapunov

stability and instability robustness consideration for linear singular system. Proceedings of the 3rd European Control ConferenceECC95, 3-5 September, Rome, Italy. Vol. 4, part 2, pp. 3702-7. DZIURLA, B., & NEWCOMB, R. W., 1987. Non-regular semi-state systems: examples and inputoutput pairing. IEEE Proceedings of the 26th Conference on Decision and Control, Los Angeles, CA, USA. Pp. 1125-6. GANTMACHER, F. R., 1977. Theory of matrices. Chelsea, New York.
KOLLA, S. R., YEDAVALLI, R. K., & FARISON, J. B., 1989. Robust stability bounds on time-varying

perturbations for state-space models of linear discrete-time systems. Int. J. Control 50,151 - 9 . KUENBERGER, D. G., 1977. Dynamic equations in descriptor forms. IEEE Trans AC 22, 312-21. MILIC, M. M., & BAJIC, V. B., 1984. Solution behavior of the semi-state model for large-scale time-discrete systems. Proceedings of the 5 th Intern. Symposium on Network Theory, ETAN and IEEE CAS Society, Sarajevo, Yugoslavia. Pp. 106-11.

62

D. LJ. DEBELJKOVIC, V. B. BAJIC, T. N. ERIC, AND S. A. MILINKOVIC

OWENS, D. H., & DEBEUKOVIC, D. LJ., 1985. Consistency and Lyapunov stability of linear

descriptor systems: a goemtric analysis. IMA J. Math. Control & Information 2, 139-51.
OWENS, D. H., & DEBELJKOVI6, D. LJ., 1986. On non-Lyapunov stability of discrete descriptor

systems. Proceedings of the 25th CDC, Athens, Greece, 2138-9. PATEL, R. V., & TODA, M., 1980. Quantitative measures of robustness for multivariable systems. Proceedings ofJACC, San Francisco, CA, USA, Paper TP8-A.

Downloaded from http://imamci.oxfordjournals.org/ at Universiti Technoligi Malaysia on January 1, 2012