702 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO.

3, MARCH 2010

ACKNOWLEDGMENT A Lyapunov Approach to Cascade Interconnection

of Integral Input-to-State Stable Systems
The authors wish to thank the anonymous referees for the careful
reading and helpful suggestions, as well as Associate Editor for the Hiroshi Ito, Senior Member, IEEE
efficient processing of this technical note.

Abstract—This technical note deals with the stability of cascade inter-

connection of integral input-to-state (iISS) time-varying systems. A new
technique to construct smooth Lyapunov functions of cascaded systems is
REFERENCES proposed. From the construction, sufficient conditions for internal stability
and stability with respect to external signals are derived. One of the de-
[1] K. Back, “Incomplete and asymmetric information in asset pricing rived conditions is a trade-off between slower convergence of the driving
theory, stochastic methods in finance,” in Proc. Lecture Notes system and steeper input growth of the driven system. The trade-off is no
CIME-EMS, Bressanone/Brixen, Italy, Jul. 6–12, 2003, pp. 1–23. more necessary if the speed of convergence of the driven system is not ra-
[2] R. Frey and W. J. Runggaldier, “Pricing credit derivatives under in- dially vanishing. The results are related to trajectory-based approaches
complete information: a nonlinear-filtering approach,” Fin. Stoch., to and small-gain techniques for feedback interconnection. The difference be-
be published. tween the feedback case and the cascade case is viewed from the require-
[3] M. Fujisaki, G. Kallianpur, and H. Kunitac, “Stochastic differential ment on convergence speed of autonomous parts.
equations for the nonlinear filtering problem,” Osaka J. Math., vol. 9,
Index Terms—Cascaded systems, integral input-to-state stability, Lya-
pp. 19–40, 1972.
punov function, nonlinear systems.
[4] H. Jeffreys, Theory of Probability. London, U.K.: Oxford University,
1961.
[5] M. S. Johannes, N. G. Polson, and J. R. Stroud, “Optimal filtering of
jump diffusions: Extracting latent states from asset prices,” Rev. Fin. I. INTRODUCTION
Stud., vol. 22, pp. 2759–2799, 2009.
[6] T. Kailath and H. V. Poor, “Detection of stochastic process,” IEEE Stability and stabilizability of cascaded nonlinear systems are often
Trans. Inform. Theory, vol. 44, no. 6, pp. 2230–2259, Oct. 1998. related to growth rate conditions on functions describing the interaction
[7] G. Kallianpur, Stochastic Filtering Theory. Heidelberg, Germany: between systems (See [9], [14]–[19], [21], references and literature re-
Springer Verlag, 1980. view therein). The cascade of input-to-state stable (ISS) systems is ISS
[8] R. E. Kass and A. E. Raftery, “Bayes factor and model uncertainty,” J.
Amer. Stat. Assoc., vol. 90, pp. 773–795, 1995. since a growth rate condition can be always satisfied[23]. However,
[9] P. E. Kloden and E. Platen, Numerical Solution of Stochastic Differen- for broader classes of systems, stability of their cascade is not always
tial Equations. New York: Springer, 1992. guaranteed. Seibert and Suárez [19] derived global asymptotic stability
[10] M. A. Kouritzin and Y. Zeng, “Bayesian model selection via filtering (GAS) of a cascade of two time-invariant systems from individual GAS
for a class of micro-movement models of asset price,” Int. J. Theor.
Appl. Fin., vol. 8, pp. 97–121, 2005.
properties of the driving system and the disconnected driven system
[11] T. G. Kurtz and P. Protter, “Weak limit theorems for stochastic inte- assuming that all solutions are bounded. To guarantee boundedness of
grals and stochastic differential equations,” Annu. Probab., vol. 19, pp. the solutions, trade-off conditions between the decay rate of the driving
1035–1070, 1991. system and the growth of the interconnection term in the driven system
[12] H. J. Kushner and P. Dupuis, Numerical Methods for Stochastic Con- have been used in the literature. Although most of the trade-off condi-
trol Problems in Continuous Time, 2nd ed. New York: Springer, 2001.
[13] J. Nielsen and M. Vestergaard, “Estimation in continuous-time sto- tions impose exponential decay rate on the driving system, the integra-
chastic volatility models using nonlinear filters,” Int. J. Theor. Appl. bility condition on the driving signal can relax the exponential-decay
Fin., vol. 3, pp. 279–308, 2000. constraint[16]. Roughly, the result in [16], [17] shows that integrability
[14] N. G. Polson and G. O. Roberts, “Bayes factors for discrete observa- of the perturbing trajectory of the driving system is sufficient to en-
tions from diffusion processes,” Biometrika, vol. 81, pp. 11–26, 1994.
[15] G. O. Roberts and O. Stramer, “On inference for partially observed
sure GAS of the cascade. This observation has been re-interpreted and
nonlinear diffusion methods using the Methopolis-Hasting algorithm,” re-written in [2] in terms of integral input-to-state stable stability (iISS)
Biometrika, vol. 88, pp. 603–621, 2001. for a time-invariant cascade in which an iISS system is driven by a
[16] J. Xiong, An Introduction to Stochastic Filtering Theory. London, GAS system. The required trade-off condition is that the iISS gain of
U.K.: Oxford University Press, 2008. driven system needs to be steep satisfactorily in the direction toward
[17] M. Zakai, “On the optimal filtering of diffusion processes,” Z.
Wahrschein. Verw. Geb., vol. 11, pp. 230–243, 1969. the equilibrium if the convergence of the driving system is slow. Note
that the set of iISS systems is larger and contains the ISS systems as a
subset. The idea of the growth-order and decay-rate trade-off result has
been improved further in [3] which shows additional conditions and
states the trade-off in terms of Lyapunov-like inequalities (dissipation
inequalities) of the two individual subsystems. These results dealing
with time-invariant cascades of iISS and iISS/GAS systems are based
on estimation of trajectories and give no explicit interpretation in terms
of constructing Lyapunov functions of the cascades.

Manuscript received April 05, 2009; revised July 15, 2009. First published
January 19, 2010; current version published March 10, 2010. This work was
supported in part by Grants-in-Aid for Scientific Research of The Japan Society
for the Promotion of Science under grant 19560446. Recommended by Asso-
ciate Editor D. Angeli.
The author is with the Department of Systems Design and Informatics,
Kyushu Institute of Technology, Iizuka, Fukuoka 820-8502, Japan (e-mail:
hiroshi@ces.kyutech.ac.jp).
Digital Object Identifier 10.1109/TAC.2009.2037457

0018-9286/$26.00 © 2010 IEEE

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:504 UTC from IE Xplore. Restricon aply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 3, MARCH 2010 703

As for feedback interconnection, stability conditions for iISS sys- fulfilling such assumptions. Indeed, for the simple system (1) and (2),
tems have been derived in [5], [7], and [8]. The development is based the choice @ V1 =@ x1 f1 1 0 2 2
0:5x1 =(x1 + 1) + x2
4 allows Theorem j j
on explicit construction of smooth Lyapunov functions of feedback sys- 1 in [3] to prove GAS of x = 0. Although the approaches in [2], [3] do
tems. The relation between the feedback results [5], [8] and the afore- not aim at providing a Lyapunov function of the overall system, there
mentioned cascade results has never been discussed in the literature yet. should be a Lyapunov function we can construct explicitly. In fact
Constructing Lyapunov functions of cascades to fill the gap is the main
aim of this technical note. Developing constructive Lyapunov coun- 1
V
e
2s 0 1
4
j 2j3
terparts of [2], [3] by specializing the idea of [5], [8] in the cascade
V (x) =
4 0 e
2s ds + x (4)

case, this technical note covers time-varying systems which have not
is a Lyapunov function establishing the GAS property of (1) and (2).
been straightforward in the previous iISS cascade approach. The au-
The idea of Corollary 1 (ii) in this technical note is to show that, for
thors of [2], [17] showed examples of cascades which are not GAS
general systems including time-varying ones, such a Lyapunov func-
when their trade-off condition is not fulfilled. It is, however, also known
tion V (x) can be directly constructed from the dissipation inequalities
that trade-off conditions are not always necessary. This technical note
in (3) whenever the dissipation inequality of the x1 -system contains
demonstrates that the construction of Lyapunov functions can elucidate
a radially non-vanishing convergent term. Constructing a Lyapunov
this fact in terms of non-vanishing convergence rate of the subsystems.
function V (x) is useful for robustness analysis with respect to external
In addition to addressing all these points, this technical note corrects
an error in [5].
0
inputs. Consider x_ 1 = x1 + x1 x2 + r13 and x_ 2 = x32 + r2 . The 0
time-derivative of V (x) in (4) along the trajectories of the cascade is
In this technical note, the set of positive definite functions from
1
+ := [0; ) to + , i.e., (0) = 0 and (s) > 0, s + 0 , 8 2 nf g 0
21 5
0 j 2j5 j 1j3 j 2j5 3
P
is denoted by . A function is said to belong to class if it is in K _ (x)
V
1 x
21 + 1
7
x +
1
r +
9
r
=
: (5)
P and increasing. A class function is said to be of class if it K K1 20 x 4 8 5

tends to infinity as its argument approaches infinity. For h , we 2P Therefore, the cascade system is iISS with respect to input (r1 ; r2 ) and
write h 2O (> L) with a non-negative real number L if there exists a state (x1 ; x2 ), and V (x) is an iISS Lyapunov function. Theorem 2 in
positive real number K > L such that lim sups 0+ h(s)=sK < .
We write h 2O (L) when K = L. Note that (S) holds
0O!  O 1 this technical note demonstrates this point.

_
(L) The next example of cascade is the following:
for L > S . The identity map on is denoted by Id. The symbols
^
and denote logical sum and logical product, respectively. A system x
_1= 0 x
x1 2
21 + 1 + x2 + r1 ; 1
x (0); x (0)2 2 + (6)
is said to be GAS if it has a globally asymptotically stable equilibrium
4
at the origin of the state space. UGAS stands for uniformly global
asymptotic stability in the case of time-varying systems.
x
_2= 0 4
2x2

x2 + 1
+
r2

r2 + 1
; 1 2
r (t); r (t) 2 +; 82 t + : (7)

The solutions x(t) = [x1 (t); x2 (t)]T evolve only in the positive or-
II. MOTIVATING EXAMPLES thant 2+ for all t 2+ . For r1 (t) = r2 (t) 0, the system has 
Consider the stability of the cascade a unique isolated equilibrium at x = 0 on the boundary of 2+ . Its
GAS property is defined by simply restricting the domain to 2+ . The
x
_ 1= 01
x 1 2
+ x x (1) x1 -system is not ISS, but only iISS with respect to input (x2 ; r1 ) and

_2 =
x 0 32
x (2) state x1 . The x2 -system is ISS with respect to input r2 and state x2 .
The convergence rate of x2 -system near the origin is much slower than
motivated by similar examples extensively studied in the literature such LES. Since the system (6) and (7) is non-negative, the simplest choice
as [16], [17]. The x1 -system is not ISS. Only iISS property holds. The of dissipation inequalities is
x2 -system does not have the LES property which had been used con-

stantly in the late 90 s including Corollaries 2 and 3 in [2]. It is worth

@V

@x
1f
1
1 0 x
x1 2
21 + 1 + x2 + r1 ;
stressing that the LES constraint can be circumvented if the perturbing 4
signal is integrable in the sense of [16], [17], which is not fulfilled
@ V2

@ x2
f2 0 4
2x2

x2 + 1
+
r2

r2 + 1
(8)
by x2 (t) of (2). If one considered x1 x2 g2 (x2 ) instead of x1 x2 in (1),
the integrability of x2 (t)g2 (x2 (t)) from t = 0 to j
satisfied by an j 1 for V1 (x1 ) = x1 and V2 (x2 ) = x2 . However, simple choices are not
appropriate g2 could guarantee the GAS of x = [x1 ; x2 ]T = 0 of
the modified system[16], [17]. Define V1 (x1 ) = 0:5 log(x12 + 1) and
often useful in proving the stability of the cascade. Searching for suit-
2 able V1 (x1 ) and V2 (x2 ) is needed in the application of [2] and [3] for
V2 (x2 ) = 0:5x2 . Then, the dissipation inequalities
proving GAS and iISS of the cascade, respectively. What is a Lyapunov
2
@V 1f
@ x1
1 0 2
x1

x1 + 1
+ x2 ; j j @V

@x
2f
2
2  0j 2j4
x (3)
function characterizing the stability properties of the overall system?
Under what condition is the construction possible? The idea of The-
orem 1 in this technical note is to provide answers in terms of the orig-
are satisfied by (1) and (2). The input x2 appears in the x1 -dissipation inal V1 (x1 ) and V2 (x2 ). Theorem 1 with (s) = s2 yields an iISS
inequality in a first order fashion, which violates the coupled condition Lyapunov function
on convergence rate and gain growth required by Theorem 1 in [2]. This 1 2 1 5
violation suggests that there exists an x1 -system such that its intercon- V (x) =
5
log(x 1 + 1) + 5 x2 + x2 (9)
nection term is the same as (1) and that it generates unbounded x1 (t)
when (2) is connected. In fact, x_ 1 = sgn(x1 ) min 1; x1 + x1 x2 0 f j jg for (6) and (7) as might be expected. When the convergence rate in
whose convergence term is saturated is such an example[2], [17]. It the dissipation inequality of the x1 -system decreases toward zero in
would be natural to expect some cascades to have GAS equilibriums the radial direction of x1 as in (8), cascade connections are not always
without the help of the LES, integrability and growth rate assumptions iISS. In this technical note, a sufficient condition for the construction
if the x1 -system has non-saturated terms of the convergence. Alterna- of an iISS Lyapunov function V (x) will be derived and shown to be
tively, one would be able to search for a better bound of @ V1 =@ x1 f1 1 consistent with the result of [3].

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:504 UTC from IE Xplore. Restricon aply.
704 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 3, MARCH 2010

from the general result. To prepare for the explicit formula of a Lya-
punov function for the cascade 6 given in Definition 1, we define  ^1
as a class K function satisfying

^1 (s) = 1 (s); 8s 2 [0; 1) (17)

Fig. 1. Cascade system 6. ^1 (s)  1 (s); 8s 2 [1; 1) (18)
1 2 K1 _ 1 ; 2 2 K1 , ^1 2 K1 : (19)

III. CASCADE OF IISS SYSTEMS

Unless r2 (s)  0, let 2 be a class K 1 function satisfying
Consider the nonlinear interconnected system 6 shown in Fig. 1. 2  r2 (s) = (!2 + 1)r2 (s); 8s 2 [0; 1) (20)
The subsystems 61 is driven by 62 . The state vector of 6 is x = 2  r2 (s) > r2 (s); 8s 2 [1; 1) (21)
[xT1 ; xT2 ]T 2 n . The exogenous signals r1 and r2 are packed into
r = [rT ; rT ]T 2 k . This technical note considers the following sets
1 2 0 and
for some real number !2 >
of systems.
Definition 1: Given 1 ; 2 2 P , 1 2 K and r1 ; r2 2 K [ f0g, 2 2 K1 _ 1 62 K1 ) 2 (s) > s; 8s 2 (0; 1) (22)
we write 61 2 S1 (n1 ; 1 ; 1 ; r1 ) and 62 2 S2 (n2 ; 2 ; r2 ) if 61 2 62 K1 ^ 1 2 K1 ) lim 2 (s)  lim 2  r2 (s):
and 62 are described by !1
s0 s0!1 (23)

We will not need 2 and !2 if r2 (s)  0. The existence of such

x_ 1 = f1 (t; x1 ; x2 ; r1 ); xi 2 n ; ri 2 k (10) a function 2 will be ensured by (27) _ (28) _ (29). For arbitrary
x_ 2 = f2 (t; x2 ; r2 ) (11) c1 2 (0; 1=2), define
fi (t; 0; . . . ; 0) = 0; t 2 + ; i = 1; 2 (12)
q(s) = min c1 1 (w):
fi is locally Lipschitz in (x; ri ) uniformly in t w2[ (s); (s)]
and piecewise continuous in t (13)
The following is the main theorem which constructs a Lyapunov func-
tion of the cascade 6.
which admit the existence of functions Vi : + 2 C1 n 0! and i ,
Theorem 1: Consider the cascaded systems 6 consisting of 61 2
1
i 2 K , i = 1; 2, such that
S1 (n1 ; 1 ; 1 ; r1 ) and 62 2 S2 (n2 ; 2 ; r2 ) for given 1 2 P ,
i (jxi j)  Vi (t; xi )  i (jxi j); i = 1; 2 (14)
2 ; 1 2 K and r1 ; r2 2 K [ f0g and positive integers n1 and n2 .
Assume that there exists  2 K1 such that
@V1
+ @V 1
f  01 (jx1 j) + 1 (jx2 j) + r1 (jr1 j) (s)
= 1 _ (ss) 2 K1
(15)
@t @x1 1 (24)
@V2 s
@t
+ @V 2
@x2 2
f  0 2 (jx2 j) + r2 (jr2 j) (16) 1   q(s)
ds = 1
q(s)
(25)
1
hold for all x 2 2 k and t 2 + . n, r 01
The Lipschitzness imposed on fi guarantees the existence of a lim0+   1  012 (s) < 1
s0
! (26)
2   (s)
unique maximal solution of 6 for locally essentially bounded ri .
2

The inequalities (15) and (16) are often referred to as dissipation are satisfied. Then, the following hold true.
inequalities, and their right hand sides are called supply rates. The i) If one of
individual system 6i fulfilling the above definition is said to be
integral input-to-state stable (iISS)[22]. The function Vi is called a 1 C 2 2 K1 (27)
iISS Lyapunov function [1]. Under a stronger assumption i 2 K , 1 1 62 K1 (28)
the system 6i is input-to-state stable (ISS) [20], and the function Vi lim
C
is a 1 ISS Lyapunov function [24]. The trajectory-based definition
1 > s0 ! 1 2 (s)  ssup
2
r2 (s) (29)
of ISS (iISS) and the Lyapunov-based definition this technical note
adopts are equivalent in the sense of the existence of ISS (iISS, holds, the system 6 is iISS with respect to input r and state x.
respectively) Lyapunov functions [22], [24]. By definition, an ISS Furthermore, an iISS Lyapunov function V of 6 is
system is always iISS. The iISS property guarantees GAS in the V (t;x ) V (t;x )
absence of the exogenous signal. The right hand side of (15) and (16) V (t; x) = 1 (s)ds + 2 (s)ds (30)
is independent of t, which means that this technical note deals with 0 0
c   q(s)
iISS and ISS properties uniform in time [12]. ISS Lyapunov functions 1 (s) = 1
q(s)
; 0 < c1 < 12 (31)
considered here is strict in the sense of [13]. In this technical note, the
convergence speed of the system 6i is said to be radially vanishing if   ^1  0
2 (w )
1
2 (s) = max :
w2[0;s] (2 0 Id)  2  2  
0 02 1 (w)
(32)
1
lim inf s 0! i (s) = 0. 1

Moreover, (26) and (32) can be replaced by

IV. MAIN RESULTS 1
  1  0 2 (s) ds < 1
1
0
2   2 (s)
1 (33)
A. A Unified Characterization 0
^  (s)
This subsection considers the cascade 6 in the most general setting   (s)
; s 2 [0; 1)
in this technical note, and a constructive theorem for the stability of 6 2 (s) = (34)
is presented. The subsequent subsections derive specialized corollaries max ^  (w) ; s 2 [1; 1)
w2[1;s]   (w)

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:504 UTC from IE Xplore. Restricon aply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 3, MARCH 2010 705

respectively, if r2 (s)  0 or equivalently r2 (t)  0. involves the K1 bounds on Vi ’s, while the K1 bounds are not involved
ii) If 1 ; 2 2 K holds, the system 6 is ISS with respect to input 1 in a result of [3]. It is a natural consequence of constructing a Lyapunov
r and state x. Furthermore, an ISS Lyapunov function of 6 is function of the whole system. It is worth noting that this technical note
(30) given with (31) and (32). does not show iISS Lyapunov functions in the case of 2 (1) = 0
Theorem 1 includes the ISS cascade addressed in [23] as the special although a Lyapunov functions for UGAS is derived in Corollary 1
case (ii). The construction of a Lyapunov function in [23] is less explicit presented later on. It is remarkable that the iISS property is proved in
and not completely specified. Theorem 1 can be considered as a result [3] without constructing a Lyapunov function of the cascade in the case
showing how to exploit the remaining flexibility for the ISS cascade to of 2 (1) = 0. The question of how to construct an iISS Lyapunov
encompass the iISS cascade. function for 2 (1) = 0 remains open. According to Remark 3, we
Remark 1: The assumption (24) ^ (25) ^ (26) ^f(27)_ (28) _ (29)g need to search for a Lyapunov function in which V1 and V2 are coupled.
is fulfilled if
B. Radially Non-Vanishing Case
9c2 > 0; k  1 The subsection deals with 61 and 62 whose convergence speed is
s:t: c2 1 02 1(s)  [2  02 1 (s)]k ; 8s 2 + (35) not radially vanishing. The following demonstrates that the trade-off
condition (24) ^ (25) ^ (26) is not necessary when 1 ; 2 2 K. In
holds. In fact, the property 1  02 2 K implies that (35) ensures
1
other words, there always exists  2 K1 such that (24) ^ (25) ^ (26)
! 1 1  02 1(s)=2  02 1 (s) < 1. Thus, the conditions (24), holds.
lims 0
(25) and (26) hold with (s) = s. Furthermore, the condition (35) Theorem 2: Consider the cascaded systems 6 consisting of 61 2
always implies f(27) _ (28) _ (29)g. Hence, if (35) holds, the cascade S1 (n1 ; 1 ; 1 ; r1 ) and 62 2 S2 (n2 ; 2 ; r2 ) for given 1 ; 2 ; 1 2
system 6 is iISS with respect to input r and state x. The condition (35) K and r1 ; r2 2 K [ f0g and positive integers n1 and n2 . If one
is used by Corollary 3 i) in [5]. Theorem 1 relaxes (35). Note that the of (27), (28) and (29) is satisfied, the system 6 is iISS with respect to
statements of ii)-v) of Corollary 3 in [5] are incorrect since their proofs input r and state x. Furthermore, an iISS Lyapunov function of 6 is
are based on the non-vanishing assumption 1 ; 2 2 K. Theorem 1 in (30) given with (31), (32), (s) = [2   01
 2  2  
01
1 (s)]s and any
this technical note not only corrects the error, but also provides us with a 1 2 K1 satisfying
more flexible Lyapunov function and a less restrictive proof specialized
in the cascade system. 1 (s) = ^1 (s); 8s 2 [0; 1) (38)
Remark 2: In the presence of the disturbance r2 , the assumption 1 (s)  ^1 (s); 8s 2 [1; 1): (39)
(27) _ (28) _ (29) cannot be removed for the choice of Lyapunov
functions in the form of (30) whatever 1 and 2 are. To see this, sup-
pose that (28) does not hold. Then, the property C. Null Exogenous Signal Case
If the exogenous signal r1 is not involved, we can remove the con-
1 (V1 (t; x1 ))[01 (jx1 j) + 1 (jx2 j)] straint 2 2 K from Theorem 1. When neither r1 nor r2 is involved,
02 (V2(t; x2 ))2 (jx2 j)  0; 8x1 2 n
; x2 2 n
(36) the trade-off condition is removed using Theorem 2 if the convergence
speed of the driven system is not radially vanishing.
! 1 2 (s) = 1 _ lim sups 0! 1 2 (s) = 1.
requires lim sups 0 Corollary 1: Consider the cascaded systems 6 consisting of 61 2
Next, suppose that lim sups 0
! 1 2 (s) < sups2 r2 (s) holds. S1 (n1 ; 1 ; 1 ; r1 ) and 62 2 S2 (n2; 2 ; r2 ) for given 1 ; 2 2 P ,
Then, the existence of U 2 P0 and c 2 + satisfying 1 2 K and r1 ; r2 2 K[f0g and positive integers n1 and n2 . Then,
1 (V1 (t; x1 ))[01 (jx1 j) + 1 (jx2 j)]
the following hold true.
i) If there exists " > 0 and  ^ 2 2 K such that
+2 (V2 (t; x2 ))[02 (jx2 j) + r2 (jr2 j)]
c + U (jr2 j); 8x1 2 n ; x2 2 n ; r2 2 k ^2 (s)  2 (s); s 2 [0; ") (40)
1  0
(37) 1 1
2 s) ds < 1
(

! 1 2 (s) < 1. Therefore, when neither (28) nor ^2  0

(41)
implies lim sups 0 0
1
2 (s)
(29) holds. the function V (t; x) of the form (30) cannot be an iISS
Lyapunov function unless (27) is satisfied. hold, the system 6 with r2 (t)  0 is iISS with respect to input
Remark 3: Note that Theorem 1 does not allow r1 and state x. Furthermore, an iISS Lyapunov function of 6 is
! 1 2 (s) = 0. As far as lim inf s 0! 1 2 (s) > 0
lim inf s 0 V (t;x ) (V (t;x ))
is concerned, we can assume 2 2 K without loss of generality V (t; x) = 1 (s)ds + 2 (s)ds (42)
since there always exists a class K function bounding from below. It
0 0
   (s) ;
 ^ 
s 2 [0; 1)
is worth mentioning that, in the presence of the disturbance r2 , the   (s)
^ 
 
assumption lim inf s 0! 1 2 (s) > 0 is necessary for the choice of 2 (s) = ^   (w)
(43)
Lyapunov functions in the form of (30) whatever 1 and 2 are. To
maxw2[1;s]
^ 
   (w)
; s 2 [1; 1)
see this, suppose that lim inf s 0
! 1 2 (s) = 0. Since 1 is of class s 0
^2  2 (w)
1
K, the property (36) requires lim sups 0! 1 2(s) = 1. In addition,  (s) = max
 01 ; 1 dw
2   (w )
(44)
due to 0 < sups2 r2 (s), the existence of U 2 P0 and c 2 +
0 2

satisfying (37) implies lim sups 0 ! 1 2 (s) < 1 which contradicts with (31) and (s) = s.
the above consequence of (36). Thus, no function V in the form of (30) ii) If 1 2 K, the system 6 is UGAS for ri (t)  0, i = 1; 2.
is an iISS Lyapunov function when lim inf s 0! 1 2 (s) = 0 holds. Furthermore, a UGAS Lyapunov function of 6 is (42) with (31),
Remark 4: Theorem 1 generalizes a similar result in [3] by covering (43)–(44), (s) = [^2  
01
 2  2  
01
1 (s)]s and any 1 2 K1 ,
time-varying systems and explicitly providing Lyapunov function of ^2 2 K satisfying (38), (39) and (40).
the whole system. It is stressed that the conditions (24) and (25) are The condition (41) constrains the growth of interconnection term in
fulfilled by (s) = s. The remaining condition (26) for (s) = s is the the driven system to be slow enough to cope with a low speed conver-
growth order restriction used in [3]. The growth order restriction (26) gence of the driving system near the equilibrium. This type of trade-off

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:504 UTC from IE Xplore. Restricon aply.
706 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 3, MARCH 2010

VI. CONCLUSION

This technical note has investigated stability of cascade interconnec-

tion of subsystems which are not necessary ISS. It has been shown how
to construct iISS/UGAS Lyapunov functions of the cascades, thereby
developing constructive Lyapunov counterparts of [2], [3]. This tech-
Fig. 2. Feedback system 6 . nical note has demonstrated that a smooth Lyapunov function can be
constructed explicitly if the condition of the trade-off between con-
vergence rate and input growth rate holds, which is consistent with
condition (41) conforms to the main result of [2] which imposes re- [2], [3]. The role of radially non-vanishing convergence rate of the
striction on growth of the interconnection term to establish GAS of driven system removing the trade-off has also been elucidated. Further-
time-invariant 6. The condition (41) is essentially the same as the cor- more, this technical note has addressed the difference between feed-
responding condition in [3]. More precisely, the condition (41) is less back interconnection and cascade interconnection. In distinction from
restrictive than the latter when 2  
2 although (41) become conser- the small-gain techniques, the proposed Lyapunov function specialized
vative if the gap between 2 and  2 around zero is large. in cascade interconnection can allow the convergence speed of indi-
vidual subsystems to be radially vanishing. Finally, it is mentioned that
this technical note does not show iISS Lyapunov functions in the case
of 2 (1) = 0 in contrast to the GAS Lyapunov function in Corol-
V. DIFFERENCE BETWEEN CASCADE AND FEEDBACK lary 1. Recently, the iISS property in the case of 2 (1) = 0 has
been proved by [3] without driving a Lyapunov function of the cas-
Consider the feedback system shown in Fig. 2, where 61 2 cade. Construction of an iISS Lyapunov function explicitly in the case
S1 (n1 ; 1 ; 1 ; r1 ) and 62 2 S2 (n2 ; 2 ; 2 ; r2 ) defined with an of 2 (1) = 0 is a subject of future study. This is also an important
additional term +2 (jx1 j) in (16) for 2 2 K. A special case of the step toward constructing Lyapunov functions iteratively for cascades
feedback is cascade connection. Indeed, if we assume 2 (s)  0 in constituted of more than two iISS systems as general as the non-con-
Fig. 2, the system is identical with Fig. 1. This fact surely implies that structive method [3]. In addition, for ISS systems, the relation between
stability of cascade systems can be verified by means of the small-gain the approach in this technical note and the construction of locally Lip-
technique. If 61 and 62 are individually ISS, the small-gain condition schitz continuous Lyapunov functions, e.g., [4], [10] is worth investi-
in [11], [25] is met by zero loop-gain since the loop is broken. Thus, gating.
the cascade of ISS subsystems is always proved to be ISS. However,
in the case of iISS subsystems, the application of the zero loop-gain to
the small-gain result in [5], [8] does not yield a tight stability condition
for the cascade. The following indicates this fact. APPENDIX A
Theorem 3: Consider the feedback system 6F shown in Fig. 2. Let PROOF OF THEOREM 1
n1 , n2 be positive integers. Assume that 1 ; 2 2 P , 1 ; 2 2 K and
r1 ; r2 2 K [ f0g are 1 and satisfy C i) The properties (24), (25), (26), (17) and (20) guarantee that there
 2 K such that
exist ;  1
i 2 O(> 1); i ; ri 2 O(> 0); i = 1; 2: (45)
(jxj)  V (t; x)  (jxj); 8x 2 n ; t 2 + (47)
Then, the system 6F with ri (t)  0, i = 1; 2, is GAS for all 6i 2
Si (ni ; i ; i ; ri ), i = 1; 2, only if holds for V defined by (30), (31) and (32). In the case of 2 62
K ^ 1 62 K , the properties lims 0! 2 (s) > 0, (19),
1 1 1
lim (20) and (32) imply sups2 2 (s) < 1. Since the properties
s0!inf
1 i (s) > 0; i = 1; 2: (46)
Id
2 2 K1 and (22) yield s=(2 0 )  201 (s)  1 for all
The above theorem demonstrates that, as long as we derive GAS, s 2 + , there exists hr2 > 0 satisfying
iISS and ISS from supply rates of the subsystems, the convergence rate
of each subsystem needs to be radially non-vanishing (1 ; 2 2 K
without loss of generality) if the interconnection forms a closed loop. 2 (V2 )f02 (jx2 j) + r2 (jr2 j)g
The smoothness of functions and (45) are only for proving the neces-  0  ^1  02 1(V2) + hr2 r2 (jr2 j) (48)
sity in Theorem 3 among subsystems having unique maximal solutions.
If systems are defined on the positive(or negative) orthant n + , the as- in the case of 2 62 K1 ^ 1 62 K1 . Next, suppose 2 2 K1 .
sumption i 2 O(> 1) can be relaxed into  2 O(1).
Define r2 (s) = 2  
2  02 1  2 r2 (s). From (32) and its
The necessity of 1 ; 2 2 K does not hold any more in the cas-
non-decreasing property, we obtain
cade case. There are pairs of supply rates from which we can derive
stability of their cascade even if i are only positive definite. A simple
example is the cascade connection of 61 2 S1 (n1 ; 1 ; 1 ; r1 ) and
62 2 S2 (n2 ; 2 ; r2 ) satisfying lim inf s 0! 1 (s) = 0 and 1 = 1 2 (V2 )f02 (jx2 j) + r2 (jr2 j)g
c2 for a constant c > 0 since V (t; x) = V1 (t; x1 ) + 2cV2 (t; x2 ) is 02 (V2 )f2 (jx2j) 0 201  2 (jx2 j)g
an iISS Lyapunov function. The results in Section IV cover this case.  if 2 (jx2 j)  2 r2 (jr2 j)
The proposed approach specialized in cascade connection is far beyond 02 (V2 )2 (jx2 j) + r2(jr2 j)r2 (jr2 j) otherwise
the application of small-gain technique in [5], [8] to cascaded systems
since Theorem 1 and Corollary 1 allow the driven system to have radi-
 0[(2 0 Id)  201  2  02 1 (V2)]2(V2 )
ally vanishing convergence rate. In the case of r2 (t)  0, the driving + r2 (jr2 j)r2 (jr2 j)
system is also allowed to have radially vanishing convergence rate.  0  ^1  02 1(V2 ) + r2 (jr2 j)r2(jr2 j): (49)

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:504 UTC from IE Xplore. Restricon aply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 3, MARCH 2010 707

In the case of 2 62 K ^ 1 2 K , the property (23) ensures 1 1 holds. Using [6, Lemma 1], choose a pair f1 (x1 ; u1 ; r1 ),
that r2 (s) is well-defined for all s 2 + , and we arrive at (49) f2 (x2 ; u2 ; r2 ): n 2 m 2 k 0 ! for which there exist 1 C
again. Consider the case of (s)=s = 1. We obtain functions V1 , V2 : n
0! and 1 , 1 , 2 , 2 2 K1 such that
6i 2 Si (ni ; i ; i ; ri ), i (jxi j) = Vi (xi ) = 
 i (jxi j) and

1 (V1 )f01 (jx1 j) + 1 (jx2 j) + r1 (jr1 j)g

(1 + i )i ( jxi j) < i (jx30i j) ) @Vi
f >   (jx j)
 0c1 1 (jx1 j) + c1   ^1  02 1 (V2 ) + c1 r1 (jr1 j) (50) @x i i i i
i

from (31) and  ^1 2 K . In the case of (s)=s 2 K1 , define Y 2 hold with ri (t)  0 for i = 1; 2. These systems 6i , i = 1; 2, defined
K1 by Y 01 (s) = c1 (s)=s. Then, using Y 01  q(s) = 1 (s), with x_ i = fi satisfy
(31) and  ^1 2 K , we can verify Y (1 (s))1 (s) = c1   q (s)
and Y 01 (^ 1 (jx2 j))^1 (jx2 j)= c1   ^1 (jx2 j)  c1   ^1  jxi j = li ; jx30i j  l30i ) @Vi
f >   (jx j):
@x i i i i
(52)
02 1 (V2 ). Recall that ab  Y (a)a + Y 01 (b)b holds for a; b 2
i

+ and Y 2 K1 . The definition (31) yields The pair of (52), i = 1; 2, implies that trajectories starting from
(x1 (0); x2 (0)) 2 f(x1 ; x2 ) 2 n 2 n : Vi (xi )  Vi (li ); i =
1 (V1 )f01 (jx1 j) + 1 (jx2 j) + r1 (jr1 j)g g
1; 2 stay there for all t 2 + . This implies that 6F is not GAS.
 0 (1 0 2c1 )  q(V1 (x1 ))
+c1   
01
^1  2 (V2 (x2 )) + c1   r1 (jr1 j): (51)
REFERENCES
Using (48), (49), (50), (51) and 0 < c1 < 1=2, we arrive at [1] D. Angeli, E. D. Sontag, and Y. Wang, “A characterization of integral
input-to-state stability,” IEEE Trans. Autom. Control, vol. 45, no. 6, pp.
1 (V1 )f01 (jx1 j) + 1 (jx2 j) + r1 (jr1 j)g 1082–1097, Jun. 2000.
+2 (V2 )f02 (jx2 j) + r2 (jr2 j)g
[2] M. Arcak, D. Angeli, and E. Sontag, “A unifying integral ISS frame-
work for stability of nonlinear cascades,” SIAM J. Control Optim., vol.
 0 o;1 (V1 ) 0 2   ^1  02 1 (V2 ) 40, pp. 1888–1904, 2002.
+o;1 (jr1 j) + o;2 (jr2 j)
[3] A. Chaillet and D. Angelli, “Integral input to state stable systems in
cascade,” Syst. Control Lett., vol. 57, pp. 519–527, 2008.
[4] S. N. Dashkovskiy, B. S. Rüffer, and F. R. Wirth, Small Gain Theorems
for some o;1 2 P , o;1 ; o;2 2 K[f0g and 2 > 0. Hence, V for Large Scale Systems and Construction of ISS Lyapunov Functions
in (30) is an iISS Lyapunov function of 6 with respect to input 2009 [Online]. Available: http://arxiv.org/pdf/0901.1842
r and state x. If r2 (s)  0 or r2 (t)  0, 2 and r2 vanish. [5] H. Ito, “State-dependent scaling problems and stability of intercon-
nected iISS and ISS systems,” IEEE Trans. Autom. Control, vol. 51,
Since it implies that the above arguments do not require 2 to be no. 10, pp. 1626–1643, Oct. 2006.
non-decreasing, the iISS can be established by (34). Note that [6] H. Ito and Z.-P. Jiang, “On necessary conditions for stability of in-
(33) and (34) ensure limV 0 ! 0 2 (V2 )2 (02 1 (V2 )) = 0 and terconnected iISS systems,” in Proc. Amer. Control Conf., 2006, pp.
(47). 1499–1504.
ii) By virtue of 1 ; 2 2 K1 , (19) and  2 K1 , we have   q;  
[7] H. Ito and Z.-P. Jiang, “Nonlinear small-gain condition covering iISS

^1  02 1 2 K1 .
systems: Necessity and sufficiency from a Lyapunov perspective,” in
Proc. IEEE Conf. Decision Control, 2006, pp. 355–360.
[8] H. Ito and Z.-P. Jiang, “Necessary and sufficient small gain conditions
for integral input-to-state stable systems: A Lyapunov perspective,”
APPENDIX B IEEE Trans. Autom. Control, vol. 54, no. 10, pp. 2389–2404, Oct. 2009.
PROOF OF THEOREM 2 [9] M. Janković, R. Sepulchre, and P. V. Kokotović, “Constructive
From 2 2 K and 1 2 K1 it follows that (s) 2 K1 Lyapunov stabilization of nonlinear cascaded systems,” IEEE Trans.

2 K1 . Due to 101  1 (s)  s, we have

Autom. Control, vol. 41, no. 12, pp. 1723–1736, Dec. 1996.
and (s)=s [10] Z.-P. Jiang, I. Mareels, and Y. Wang, “A Lyapunov formulation of the
  1  02 1 (s)=2  02 1 (s)  1  02 1 (s). Since 1 2 K nonlinear small-gain theorem for interconnected ISS systems,” Auto-
implies q = c1 1   1
01 2 K, we have (  q)=q 2 K. Thus, matica, vol. 32, pp. 1211–1215, 1996.
(24) ^ (25) ^ (26) holds. Hence, Theorem 1 completes the proof.
[11] Z.-P. Jiang, A. Teel, and L. Praly, “Small-gain theorem for ISS sys-
tems and applications,” Mathe. Contr. Signals Syst., vol. 7, pp. 95–120,
1994.
APPENDIX C [12] Y. Lin, Y. Wang, and D. Cheng, “On non-uniform and semi-uniform
input-to-state-stability for time varying systems,” in Proc. 16th IFAC
PROOF OF COROLLARY 1 World Congress, Prague, Czech Republic, Jul. 2005, [CD ROM].
Define V^2 =  (V2 ). Then,   2 (jx2 j)  V^2 (t; x2 )      2 (jx2 j) [13] M. Malisoff and F. Mazenc, “Further remarks on strict input-to-state
and  2 K1 hold. From (44) it follows that dV^2 =dt  0 ^ 2 (jx2 j).
stable Lyapunov functions for time-varying systems,” Automatica, vol.
41, pp. 1973–1978, 2005.
Substituting ^ 2 ,   2 and   
 2 for 2 , 2 and   2 , respectively. [14] F. Mazenc and L. Praly, “Adding integrations, saturated controls, and
we obtain Claim (i) from Theorem 1 (i) with (s) = s. Here, (41) stabilization for feedforward systems,” IEEE Trans. Autom. Control,
implies (33). Note that (27) _ (29) is met for r2 (t)  0, equivalently vol. 41, no. 11, pp. 1559–1578, Nov. 1996.
r2 (s)  0. Theorem 2 proves Claim (ii). [15] F. Mazenc, R. Sepulchre, and M. Jankovic, “Lyapunov functions for
stable cascades and applications to global stabilization,” IEEE Trans.
Autom. Control, vol. 44, no. 9, pp. 1795–1800, Sep. 1999.
APPENDIX D [16] E. Panteley and A. Loría, “On global uniform asymptotic stability of
nonlinear time-varying systems in cascade,” Syst. Control Lett., vol. 33,
PROOF OF THEOREM 3
pp. 131–138, 1998.
Consider 6F with ri (t)  0, i = 1; 2. Suppose that [17] E. Panteley and A. Loría, “Growth rate conditions for uniform asymp-
lim inf s 0! 1 i (s) = 0 holds for at least one of i = 1; 2. Due totic stability of cascaded time-varying systems,” Automatica, vol. 37,
pp. 453–460, 2001.
to 1 ; 2 2 K, there exist li > 0 and i > 0 for i = 1; 2 such that [18] A. Saberi, P. V. Kokotovicánd, and H. J. Sussmann, “Global stabiliza-
tion of partially linear systems,” SIAM J. Control Optim., vol. 28, pp.
jxi j = li ; jx30i j  l30i ) (1 + i )i (jxi j) < i (jx30i j) 1491–1503, 1990.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:504 UTC from IE Xplore. Restricon aply.
708 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 3, MARCH 2010

[19] P. Seibert and R. Suárez, “Global stabilization of nonlinear cascaded tainties and stochastic uncertainties in [8], to uncertain descriptor sys-
systems,” Syst. Control Lett., vol. 14, pp. 347–352, 1990. tems in [4], [12], and so on. Recently, it is found in [11] that a robust
[20] E. Sontag, “Smooth stabilization implies coprime factorization,” IEEE estimator derived from prediction error sensitivity penalization has al-
Trans. Autom. Control, vol. AC-34, no. 4, pp. 435–443, Apr. 1989.
[21] E. D. Sontag, “Remarks on stabilization and input-to-state stability,” in most the same structure as that of [6]. These studies and extensions
Proc. IEEE Conf. Decision Control, 1989, pp. 1376–1378. have clarified many theoretical and practical issues in robust estimator
[22] E. Sontag, “Comments on integral variants of ISS,” Syst. Control Lett., design. On the other hand, some important properties of the estimation
vol. 34, pp. 93–100, 1998. procedure of [6] are still not very clear. For example, general conclu-
[23] E. D. Sontag and A. Teel, “Changing supply functions in input/state
stable systems,” IEEE Trans. Autom. Control, vol. 40, no. 8, pp.
sions are still not available for its convergence and boundedness.
1476–1478, Aug. 1995. Using routine matrix operations, some new formulas are obtained in
[24] E. Sontag and Y. Wang, “On characterizations of input-to-state stability this technical note for matrix updates in the above recursive state es-
property,” Syst. Control Lett., vol. 24, pp. 351–359, 1995. timation procedure. From these relations, certain general conclusions
[25] A. Teel, “A nonlinear small gain theorem for the analysis of control are reached for the convergence of the estimation and the stability of
systems with saturation,” IEEE Trans. Autom. Control, vol. 41, no. 9,
pp. 1256–1270, Sep. 1996. the estimator. It is proved that if certain stabilizability and detectability
conditions are satisfied and nominal system matrices are time invariant,
then, the robust estimator converges to a time invariant stable system
with the increment of time, provided that the filter design parameter
keeps constant and external inputs and disturbances are stationary.
These conditions are consistent with those of [6], and turn out to be
On the Convergence and Stability both necessary and sufficient for guaranteeing these properties.
of a Robust State Estimator The rest of this technical note is organized as follows. In the next sec-
tion, the time and measurement update form of the robust estimator of
Tong Zhou [6] is briefly summarized, as well as the existing convergence results.
Main results of this technical note are reported in Section III, in which
some general properties are established for the convergence of that es-
Abstract—Convergence and stability of the robust state estimator ob- timator. Some concluding remarks are given in Section IV. Finally, 3
tained in [6] is reinvestigated in this technical note. Some new relations have appendices are included to give proofs of some technical results.
been established for matrix updates in the recursive state estimation. It is
The following notation is adopted in this technical note. In repre-
sents the n 2 n dimensional identity matrix. The subscript n is often
proved that under certain stabilizability and detectability conditions, this
robust estimator converges to a stable time invariant system, provided that
plant nominal parameters are time invariant and the filter design param- omitted when the dimension is clear from context or not important.
eter is fixed. These results are consistent with existent ones, but different When the expression for X is long, XW X T is often abbreviated as
XW (?)T . diag fX1 ; X2 ; 1 1 1 ; XL g denotes a block diagonal matrix
from them at the point that there are no orthogonality constraints on un-
 (1) stands for the maximum
certainty related system matrices, and therefore widen theoretical guaran-
with its k -th diagonal block being Xk . 
singular value of a matrix, E(1) the mathematical expectation of a
tees for the effectiveness of the estimation procedure.
Index Terms—Recursive estimation, regularized least-squares, robust-
random variable/vector, while ij the Kronecker delta function which
equals to 1 when i = j and to zero whenever i 6= j . The other notation
ness, state estimation, unstructured parametric uncertainty.

and symbols are relatively standard.

I. INTRODUCTION Throughout this technical note, the subscript i is used as a time
index. When a variable/matrix, for example, Ai , becomes independent
Stimulated by the fact that in actual engineering problems, modelling
of time, its subscript i will be deleted.
errors are generally unavoidable and important physical variables can
not always be directly measured, robust state estimation has attracted II. BRIEF SUMMARY OF THE ROBUST ESTIMATOR
extensive attentions and become one of the basic problems in system
Based on a deterministic interpretation of Kalman filtering and re-
theory and signal processing [2], [3], [7], [10]. Among the various
sults on regularized least squares with norm bounded unstructured un-
frameworks proposed for robust estimator design, the methodology
certainties under a worst case framework, a robust state estimation pro-
suggested in [6] has many particularly attractive properties, in which,
cedure has been derived in [6] for regular state space models with norm
rather than deregulation, regulation is adopted to guarantee the robust-
bounded parametric modelling errors. To be more specific, concerning
ness of the designed estimator. For example, this robust estimator has
the following uncertain plant dynamics description:
a computational complexity comparable with the well known Kalman
filter, it can be recursively realized without verifying any existence con- xi+1 = (Fi +Fi )xi +(Gi +Gi )ui ; yi = Hi xi +vi ;  0 (1)
i
ditions, etc. [5], [9].
Attracted by its rigorous theoretical developments and appreciative in which [Fi Gi ] = Mi 1i [Efi Egi ], 1i 2 Rr 2s ,   (1i )  1,
implementation simplicity, the results of [6] have been extended to var- i = 0; 1; 1 1 1, assume that the plant initial state vector x0 , the plant
ious other uncertainty descriptions. For example, to polytopic uncer- external input vector ui and the measurement disturbance vector vi
are uncorrelated random vectors with E(ui ) = 0, E(vi ) = 0 and
E(col(x0 0 E(x0 ); ui ; vi )(?)T ) = diagf50 ; Qi ij ; Ri ij g. Here,
Manuscript received May 07, 2009; revised June 17, 2009 and October 26, col(1) represents a column vector formed by stacking all the column
2009. First published January 19, 2010; current version published March 10, vectors in the parentheses, 50 , Qi and Ri are some known real positive
2010. This work was supported in part by the National Natural Science Foun- definite matrices with 50 2 Rn2n , Qi 2 Rp2p , and Ri 2 Rq2q .
dation of China under Grants 60625305 and 60721003, and by the 973 Program Then, the time and measurement update form of the robust estimator
under Grant 2009CB320602. Recommended by Associate Editor X. Chen.
The author is with the Department of Automation and TNList, Tsinghua Uni-
of [6] consists of the following 4 steps.
• Initialization. Designate P0j0 and x
01 and x^ = P0j0 H T R01 y .
versity, Beijing 100084, China (e-mail: tzhou@mail.tsinghua.edu.cn). ^ respectively as P0j0 =
Digital Object Identifier 10.1109/TAC.2009.2037459 01 T 01
(50 + H0 R0 H0 ) 0j0 0j0 0 0 0

0018-9286/$26.00 © 2010 IEEE

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:504 UTC from IE Xplore. Restricon aply.