IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 46, NO.

7, JULY 2001

1093

set of controllable modes and the set of uncontrollable modes do not

intersect. Some examples illustrate the result.

Asymptotic Behavior of Nonlinear Networked Control

Systems

REFERENCES

Gregory C. Walsh, Octavian Beldiman, and Linda G. Bushnell

[1] B. D. O. Anderson, An algebraic solution to the spectral factorization

problem, IEEE Trans. Automat. Contr., vol. AC-12, pp. 410414, Aug.
1967.
, A system theory criterion for positive real matrices, SIAM J.
[2]
Control, vol. 5, pp. 171182, 1967.
[3] B. D. O. Anderson and S. Vongpanitlerd, Network Analysis and Synthesis. Englewood Cliffs, NJ: Prentice-Hall, 1973.
[4] N. E. Barabanov, A. Kh. Gelig, G. A. Leonov, A. L. Likhtarnikov, A. S.
Matveev, V. B. Smirnova, and A. L. Fradkov, The frequency theorem
(KalmanYakubovich lemma) in control theory, Automat. Rem. Control, vol. 57, pp. 13771407, 1996.
[5] P. Ioannou and G. Tao, Frequency domain conditions for strictly
positive real functions, IEEE Trans. Automat. Contr., vol. AC-32, pp.
5354, Jan. 1987.
[6] R. Johansson and A. Robertsson, Stability analysis of adaptive output
feedback systems, in 37th IEEE Conf. Decision Control, Tampa, FL,
1998, pp. 40084009.
[7] R. E. Kalman, Lyapunov functions for the problem of Lure in automatic control, in Proc. Nat. Acad. Sci., vol. 49, 1963, pp. 201205.
[8] H. K. Khalil, Nonlinear Systems, 2nd ed. Upper Saddle River, NJ:
Prentice-Hall, 1996.
[9] H. Kimura, Chain-Scattering Approach to H -Control. Boston, MA:
Birkhuser, 1997.
[10] R. Lozano and S. M. Joshi, Strictly positive real transfer functions revisited, IEEE Trans. Automat. Contr., vol. 35, pp. 12431245, Nov. 1990.
[11] K. R. Meyer, On the existence of Lyapunov functions for the problem
of Lure, SIAM J. Control, vol. 3, no. 3, 1966.
[12] K. S. Narendra and J. H. Taylor, Frequency Domain Criteria for Absolute
Stability. New York: Academic, 1973.
[13] V. M. Popov, Absolute stability of nonlinear systems of automatic control, Automat. Rem. Control, vol. 22, pp. 857875, 1962.
[14] A. Rantzer, On the KalmanYakubovichPopov lemma, Syst. Control
Lett., vol. 28, pp. 710, 1996.
[15] W. Rugh, Linear Systems, 2nd ed. Upper Saddle River, NJ: PrenticeHall, 1996.
[16] G. Tao and P. Ioannou, Strictly positive real matrices and the LefschetzKalmanYakubovich lemma, IEEE Trans. Automat. Contr.,
vol. 33, pp. 11831185, Dec. 1988.
[17] M. Vidyasagar, Nonlinear Systems Analysis, 2nd ed. Englewood
Cliffs, NJ: Prentice-Hall, 1993.
[18] V. A. Yakubovich, Solution of certain matrix inequalities in the stability
theory of nonlinear control systems (in English), Soviet. Math. Dokl.,
vol. 3, pp. 620623, 1962.
[19] D. C. Youla, On the factorization of rational matrices, IEEE Trans.
Inform. Theory, vol. IT-7, pp. 172189, July 1961.
[20] V. M. Popov, Absolute stability of nonlinear systems of automatic control, in Frequency-Response Methods in Control Systems, A. G. J. MacFarlane, Ed. New York: IEEE Press, 1979, pp. 163181.

AbstractThe defining characteristic of a networked control system

(NCS) is having a feedback loop that passes through a local area computer
network. Our two-step design approach includes using standard control
methodologies and choosing the network protocol and bandwidth in order
to ensure important closed-loop properties are preserved when a computer
network is inserted into the feedback loop. For sufficiently high data
rates, global exponential stability is preserved. Simulations are included
to demonstrate the theoretical result.
Index TermsAsynchronous packets, networked control systems.

I. INTRODUCTION
Using a (local area) networked control architecture has many advantages over a traditional point-to-point design including low cost of
installation, ease of maintenance, lower cost, and greater flexibility [3],
[4]. For these reasons the networked control architecture is already used
in many applications, particularly where weight and volume are of consideration, for example in automobiles [2] and aircraft [5], [6]. The introduction of a computer network in the feedback loop unfortunately
invalidates the traditional analytic stability and performance guarantees that control design typically produces. In this note, we reconnect
the analysis of the control design to the networked control context, and
provide guarantees of stability and certain levels of asymptotic performance to the control systems employing networked feedback loops.
We focus on a multiple-inputmultiple-output (MIMO) nonlinear
plant with a nonlinear controller connected by a communication network. A block diagram of this system is presented in Fig. 1.
We assume that the controller is designed without regard to the network, meaning that if the input to the controller is connected directly
to the output of the plant the system would be globally (or locally) exponentially stable. We provide conditions under which these stability
properties are preserved when the communication network is inserted
into the loop between the outputs of the plant and the controller input.
Each output, or group of outputs, is assumed to be monitored by a smart
sensor with a network interface. Specifically, in the laboratory we us
a Controller Area Network (CAN-II) operating at 1 Mb/sec because
CAN-II is commonly used in automobiles and manufacturing plants.
Each smart sensor must compete with the others for access to the network. The resulting communication constraint is the primary focus of
this note, hence propagation delays, communication errors and observation noise will not be treated.
The general system consists of the time-varying plant, the
time-varying controller, and the network. We denote the plant
dynamics by x_ p (t) = fp (t; xp (t); up (t)), y (t) = gp (t; xp (t)),

Manuscript received April 5, 1999; revised December 23, 1999 and August
16, 2000. Recommended by Associate Editor L. Dai. This work was supported
in part by the United States Army Research Offic under Grant DAAG55-98-D0002, and in part by the National Science Foundation under Grant ECS-9702717.
G. C. Walsh is with the Department of Mechanical Engineering, University
of Maryland, College Park, MD 20742 USA (e-mail: gwalsh@eng.umd.edu).
O. Beldiman is with Mitusbishi Electric and Electronics, Inc., Durham, NC
27613 USA (e-mail: beldiman@msai.mea.com).
L. G. Bushnell is with the Department of Electrical Engineering, University of Washington, Seattle, WA 98195-2500 USA (e-mail: bushnell@ee.washington.edu).
Publisher Item Identifier S 0018-9286(01)06614-4.
00189286/01$10.00 2001 IEEE

1094

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 46, NO. 7, JULY 2001

Fig. 1. Configuration of networked control system.

and let y^(t) denote the input to the controller. This is the vector of
the most recently transmitted plant output values, and has the same
dimension as y (t). At each transmission time a component(s) of y^(t)
is set to the corresponding component(s) of y (t). The controller is
described by the following equations: x_ c (t) = fc (t; xc (t); y^(t)),
up (t) = gc (t; xc (t)). We define the network induced error as
e(t) = y(t) 0 y^(t). Between two successive transmission times we
have e_ (t) well defined as long as gp is continuously differentiable.
Note that the network may also be used to transmit the control signal
u(t) by augmenting the error vector.
We now combine the controller and plant states into the vector x(t),
and we relabel the vector fields to obtain the following dynamics for
the closed-loop system in between two successive transmission times

x_ (t) = f (t; x(t); e(t))

e_ (t) = g (t; x(t); e(t)) :

(1)

At each transmission time the vector e has a discontinuous jump,

namely the component corresponding to the transmitting output sensor
is set to zero. We assume that the functions f and g are globally
Lipschitz with associated Lipschitz constants kf and kg . We then
define

z (t) = (x(t); e(t))0

z_ (t) = h (t; z (t))
and let kh denote the global Lipschitz constant of the function h =
g)0 .
Our approach is to model the error e(t) as a perturbation on the
system. If e(t)  0, then y^(t) = y (t) and from the control design
we have a Lyapunov function V (t; x) and constants c1 , c2 , c3 and c4
greater than zero with

(f;

c1 kxk2  V (t; x)  c2 kxk2

@V
@t

@V
f (t; x; 0)  0c3 kxk2
@x
@V
 c4 kxk:
@x

(2)
(3)

There exist many models of network traffic. We assume packets can

arrive at any time but the amount of time between transmissions obeys
a deadline,  . Even a simple Poisson distribution allows for the possibility of arbitrary long delays between transmissions, making the stability analysis impossible. In the following derivations we assume that
there is a bound for the delays between transmission. We will call this
deadline the maximum allowable transfer interval (MATI).
II. PERTURBATION BOUNDS
The following stability results are based on ideas from perturbation
theory [1]. We consider the error e(t) as a vanishing perturbation and
we compute worst case bounds for it as a function of a bound  on
its derivative, that is, ke_ (t)k <  for t in some fixed interval of time.
At any given transmission opportunity, only one of the p information

sources may transmit its data, and so how the transmission opportunities are scheduled is critical.
Two scheduling algorithms are considered. The simplest is a static
scheduler. The transmission order is decided in advance and this order
is repeated indefinitely. This common method would typically be implemented by polling or by a token ring. The other method uses bit-wise
arbitration to dynamically schedule the use of the network at run time.
This is the Try-Once-Discard (TOD) protocol introduced in [7]. Some
portion of a nodes identifier (the CAN-II identifier is 29 bit long) is
used to transmit a local evaluation of the importance of the locally held
data. The information source or smart sensor that consider its data the
most important wins use of the network and all losing data packets are
discarded.
Lemma 1 (Scheduler Error Bound) [7]: Consider a static or a dynamic network scheduler starting at time t0 , visiting each node at least
once, with integer periodicity p, maximum allowable transfer interval
 , maximum growth in error e(t) in  seconds strictly bounded by
 2 (0; 1). Then, for any time t  t0 + p , the error is bounded
as ke(t)k < p(p 0 1)=2.
Proof: For the static scheduler, because the transmission is
p-cyclic, every node will transmit at least one time in p seconds, and
the bound follows.
For the dynamic scheduler, the error of the channels not transmitted
is bounded by the error of the channel that did transmit. Consequently,
if any channel repeats, it establishes a bound for the remaining channels. If no channel repeats, then the error bound is similar to the static
scheduler because in p seconds all channels will have transmitted at
least one time. Let t1 ; . . . ; tp be the last p transmission times (t1 the
most recent) and i1 ; . . . ; ip be the channels that were transmitted, respectively. Let im be the first channel that is repeated. If m = p then
the lemma is proved, since the error on channel ik at time t1 is bounded
by k .
Assume m < p. Let 1  l < m be the index of the time when
channel im was last transmitted. Then ei (tl )  (m 0 l) . Also, since
at time tl channel im was transmitted it must have had the highest error,
so all the other channels have error bounded by (m 0 1) . The channels
ik with 1  k  l have the usual bound k . Summing up and taking
the worst case will give the bound from the lemma statement.
III. STABILITY OF NONLINEAR NETWORKED CONTROL SYSTEMS
In this section, we present the main result of this note: that the stability property of a networked system is preserved if the network is fast
enough. There are no assumptions about the initial conditions of the
error e(t), and so only BellmanGronwall can bound the growth of
z = (x; e)0 during the first p seconds. At t = p , however, Lemma
1 provides a bound on ke(t)k. We then prove that for t  p , the error
remains bounded and kx(t)k decreases exponentially to a ball proportional to the bound on ke(t)k. The previous steps are repeated to show
overall exponential decay.
Theorem 1: Consider a NCS whose continuous dynamics are described by equation (1) with the control law designed so that equations
(2)(4) are satisfied, with p nodes of sensors operating under TOD or
static scheduling, a maximum allowable transfer interval (MATI) that
satisfies  < minf(ln(2)=kh p), (S=8), (Sc3 =32kf c4 )((c1 =c2 ))3=2 g
where S = (( c2 =c1 +1)kh pi=1 j )01 . Then, the networked system
is exponentially stable.
Proof: We first expand the bounds imposed on  for easy reference. Note that some of the bounds are used for several inequalities.

ln(2)
kh p
S
<

<

) ek

p

) < S

<2

(4)
(5)

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 46, NO. 7, JULY 2001

 < Sc3 c1
32kf c4
c2

3 2

) 2S < 14

(6)

) kf c 2S < c
4

(7)

4kf c4 

c2
c1

c1

 2:
1

(8)

1095

where we used (9) and (2) to get the bound on kxk and (10) for the
bound on kek.
The growth of error in any  interval included in [t0 + p; t^] is
bounded by kh (2=S + 2 c2 =c1 )kz (t0 )k, which is less than  , according to (5). Applying Lemma 1 on [t^ 0 p; t^], we get

e(t^) < 2 kz (t0 )k

S

so we obtain the contradiction. Therefore

ke(t)k < 2S kz(t )k ;

V (t; x(t)) < 4c kz (t )k ;

Let z (t0 ) be the initial condition at the initial time t0 . For any t in
[t0 ; t0 + p ], using BellmanGronwall and (4) we have:

kz(t)k  ek

Then, on the interval [t0 ;

p kz (t

k  2 kz(t )k :

0)

t0 + p ], we have
ke_ (t)k  kz_ (t)k  kh (t; z(t))k  kh kz(t)k  2kh kz(t0 )k
and the maximum growth of error in a  interval is bounded
c2 =c1 )kz (t0 )k. We denote
by 2kh  kz (t0 )k < 2kh (1 +
 = 2kh (1 + c2 =c1 )kz (t0 )k and we use it to apply Lemma 1.
While the smaller bound 2kh kz (t0 )k is acceptable for the interval
[t0 ; t0 + p ], the larger bound 2kh (1 + c2 =c1 )kz (t0 )k is needed to
extend the result beyond this short transient period, as is shown later.
We can now use Lemma 1 to get that

ke(t

k

+ p )

j kz (t0 )k = 2 kz (t0 )k :
S
j =1

Next, we prove by contradiction that

V (t; x(t)) < 4c2 kz (t0 )k2

ke(t)k < 2S kz(t0 )k

8 t > t + p
8 t > t + p:
0

(9)
(10)

Suppose this is not true. Then there is a first time t^ when one of these
conditions (or both of them) fails.
Consider the case when (9) does not hold

t;^

x(t^) = 4c2

kz(t )k :
2

(11)

x(t^)
V_ (x) = @V
@t
@V
=
@t

 2 kz(t )k > kz(t )k

0

(12)

V_ (x)  0c3 kxk2 + c4 kxkkf kek

3

+ kf c4

2

2

(13)

kz(t )k :
0

kz(t )k :
0

ke_ (t)k  kh kz(t)k  kh (kx(t)k + ke(t)k)

2

c2
c1

2

kz(t )k
0

+ p ,

8t  t :
0

(15)

using (13), (14), and (15) we have

V_ (t; x(t))  0 c3 V (t; x(t)) + kf c4 4

c2
S
Denote  = c3 =c2 and   = 4kf c4 (4=S )

c2 kz (t )k2 :
0
c1
(c2 =c1 ). Using the
Comparison Lemma [1, p. 100], and the fact that V (t0 + p; x(t0 +
p )) < 4c2 kz (t0 )k2 , we obtain
V (t0 + t; x(t0 + t))  kz (t0 )k2 e0(t0p) (4c2 0  ) + 
for all t > p .
Fix t1 > p so that
e0(t 0p) < =(4c2 0  )
and then

V (t0 + t1 ; x(t0 + t1 ))  2  kz (t0 )k2 :

According to the Lyapunov bound (2) and to the bound on  (8)

kx(t

k  2c  kz(t )k  1=2 kz(t )k :

+ t1 )

(16)

Using the bounds (15) and (14) in inequality (16), we get that
1

k  kx(t

+ t0 )

and by the bound on  (6)

In this case, t^ cannot be a transmission time because during a transmission the error can only decrease. For any time in the interval [t0 ; t^] that
is not a transmission time, we have

 kh

kz(t

^ x(t^)) < 0. This means that V is

By the choice of  [see (7)], V_ (t;
decreasing locally around t^ so it cannot cross the bound. The only possible case is, consequently, when (10) does not hold meaning that

e(t^)

t

(14)

k ke(t

+ t1 ) +

+ t1 )

k  3=4 kz(t )k :

+ t0 )

By repeating this procedure, we can prove that

and due to (12) and (10)

 x(t^) 0c

At a time t

Using the Lyapunov bounds (2) and (3) and the Lipschitz condition for
the function f , we get

V_ t;^ x(t^)

kz(t

@V f (t; x; e)
@x
@V
+
(f (t; x; 0) + f (t; x; e) 0 f (t; x; 0)) :
@x
+

kx(t)k < 2 cc kz(t )k ;

kz(t

Note that (11) and Lyapunov condition (2) imply that

8 t  t + p
8t  t

k  (3=4)k kz(t )k :

+ kt1 )

Thus, the system is exponentially stable.

Remark. [Local Exponential Stability]: We can prove that local exponential stability is preserved with a networked architecture, but the
estimate of the region of attraction is correspondingly reduced. In particular, if the Lyapunov function, its associated constraints and the Lipschitz constants kf , kg and kh are defined for kxk < c5 with c5 > 0,
one estimate of the region of attraction [1, p. 100] is fx: kxk <
c5 c1 =c2 g if no network is present.
With the network present, during the first p seconds, the proof provides only a BellmanGronwall bound on the growth and hence kz k
may as much as double. The proof guarantees this is the maximum that
kzk will attain; however this limits the region of attraction to half the
limit attained without the network to fz : kz k < (c5 =2) c1 =c2 g.
IV. SIMULATIONS
In this section, we use a single-link flexible-joint robot to illustrate
the theoretical result from the previous section. This is a nonlinear
system that can be globally linearized by feedback. We use it to show
how the stability is preserved for small enough maximum allowable
transfer interval (MATI) and lost for large MATI.

1096

Fig. 2.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 46, NO. 7, JULY 2001

Schematic of the single-link flexible-joint robot.

Fig. 4. Single link: stable ( = 0:01) (solid line) and unstable ( = 0:05)
(dashed line) trajectories.

50 simulations for each value of . The system was considered to be

stable if the last 40 samples of its output were less than 0.1.
In Fig. 3, we show the percent of stable cases versus the average  of
the Poisson distribution used to generate the transmission times. This
graph shows that the system loses stability when  is around 0.015.
The theoretical MATI for this system is about 1003 , meaning that the
bounds that we used in deriving this limit are too conservative.
In Fig. 4, we present a typical stable trajectory for  = 0:01 and an
unstable trajectory for  = 0:05.
V. CONCLUSION

a(x) = (MgL=I ) sin q1 [q_1 2 +(MgL=I ) cos q1 +(k=I )]

+(k=I )(q1 0 q2 )[(k=I ) + (k=J ) +(MgL=I ) cos q1 ]. The resulting
equations are in Brunowski cannonical form, z_1 = z2 , z_2 = z3 ,
z_3 = z4 , z_4 = v . We then design a state feedback controller for this

In this note, we analyze a fundamental problem: the stability of a

nonlinear networked control system, i.e., a nonlinear control system
having a feedback loop closed through a communication network. A
networked control architecture has many advantage compared to the
traditional point-to-point design, but there are also some new problems
due to the presence of a communication network. The main complication is the presence of the random time-delays, but there are also
undesired effects due to the discrete nature of the messages that can be
exchanged over the network.
Our paradigm in this note is that the controller for the plant is designed regardless of the network, and it exponentially stabilizes the
plant (locally or globally) if the network is not present. We study a
newly introduced network protocol (TOD) that resolves collisions by
transmitting the channel with highest error between its last transmission and its current data.
We prove that if certain assumptions hold, the plant will be stabilized
even in the presence of the network, although the region of attraction
may be smaller (and we give an estimate for this region). The main
assumptions that we make are: 1) that the intertransmission interval
is bounded, and the bound is sufficiently small,and 2)that the collision
resolution is perfect, meaning that the highest error channel will always
be transmitted.
Future research includes taking into account the presence of the communication network when the controller is designed.

linear system in order to stabilize the origin, and then map it back to
the original system.
The goal of this example is to show what happens to the stability of
the system when the average amount of bandwidth allocated to the control loop is varied. We are going to model this variation by changing the
average  of the Poisson distribution used for the transmission times.
Due to the random nature of the transmission times distribution, we run

[1] H. K. Khalil, Nonlinear Systems, second ed. Upper Saddle River, NJ:
Prentice-Hall, 1996.
[2] U. zgner, H. Goktag, and H. Chan, Automotive suspension control
through a computer communication network, in Proc. IEEE Conf. Control Applications, 1992.

Fig. 3. Single link: the percent of stable cases as function of  . The system
loses stability when  is around 0.015

Consider a link driven by a motor through a torsional spring in the

vertical plane as shown in Fig. 2. The equations of motion are as follows:

I q1 + Mg sin q1 + k(q1 0 q2 ) = 0

J q2 0 k(q1 0 q2 ) = u:

The system is linearized exactly with the following transformation:

z=

q1
q_1

0(MgL=I ) sin q1 0 (k=I )(q1 0 q2)

0(MgL=I )q_1 cos q1 0 (k=I )(q_1 0 q_2 )
IJ
u=
(v 0 a(x))
k

where

REFERENCES

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 46, NO. 7, JULY 2001

[3] D. Radford, Spread-spectrum data leap through ac power wiring, IEEE

Spectrum, pp. 4853, Nov. 1996.
[4] R. S. Raji, Smart networks for control, IEEE Spectrum, pp. 4955,
June 1994.
[5] A. Ray, Performance evaluation of medium access control protocols
for distributed digital avionics, ASME J. Dyn. Syst., Meas., Control,
vol. 109, pp. 370377, Dec. 1987.
[6] J. A. Sparks, Low cost technologies for aerospace applications, Microprocessors and Microsystems, vol. 20, pp. 449454, 1997.
[7] G. C. Walsh, H. Ye, and L. G. Bushnell, Stability analysis of networked
control systems, in Proc. Amer. Control Conf., June 1999.

Optimal Control of a Particular Class of Singularly

Perturbed Nonlinear Discrete-Time Systems
Rachid Bouyekhf, Abdelkhalek El Hami, and Abdellah El Moudni

AbstractThis note studies a class of discrete-time nonlinear systems

which depend on a small parameter. Using the singular perturbation theory
in a systematic way, we give a trajectory approximation result based on the
decomposition of the model into reduced and boundary layer models. This
decomposition is used to analyze optimal control via maximum principle
of such systems. As a result, significant order reduction of optimal control
problems is achieved.
Index TermsDiscrete-time nonlinear systems, optimal control, singular
perturbation.

I. INTRODUCTION
One of the most popular methods for model reduction of large-scale
systems is the singular perturbation method. This approach has attracted much attention because of its simplicity and good performance
in most experimental situations. In the case of discrete-time linear systems, the theory of the singular perturbation has been probably initiated
by [1], [2]. In these papers, the method is treated by using the similar
ideas used in continuous-time case. After these results, the singular perturbation discrete-time linear system has received a great deal of attention in the literature, and significant progress has been made toward the
optimal feedback control of large-scale linear systems. Among several
contributions to the subject, we would-like to point out those related to
[3]. In that paper, the authors have derived sufficient conditions for the
existence of a suboptimal solution, and a near-optimum design method
has also been proposed. Another notable work in this area is the book
of Naidu [4] in which the author gives a particularly excellent account
of the application of the singular perturbation in discrete-time linear
control theory.
This paper is entirely devoted to the singularly perturbed nonlinear
discrete-time systems and develops a discrete-time counterpart of certain results of singularly perturbed continuous-time nonlinear control
problems. More precisely, a method will be developed in Section II
which parallels the corresponding method in [6]. In particular, a mode
decoupling approach will be given for the solution to be approximated
Manuscript received October 1, 1999; revised December 11, 2000. Recommended by Associate Editor M. Krstic
R. Bouyekhf and A. El Moudni are with the University of Technology BelfortMontbeliard (UTBM), Laboratory SET, 90010 Belfort Cedex, France.
A. El Hami is with INSA of Rouen-LMR-UPRESA 6104, 76801 Saint-Etienne du Rouvroy Cedex, France.
Publisher Item Identifier S 0018-9286(01)06613-2.

1097

in a form of a reduced and boundary-layer solutions. On the other hand,

sufficient conditions will be given for proving the validity of the approximation. In Section III, we turn our attention to optimal control
problems of discrete-time singularly perturbed nonlinear systems. Following the spirit and philosophy of Section II, we provide significant
order reduction of optimal control problems of such systems.
II. SYSTEM DESCRIPTION AND DECOMPOSITION
Consider a discrete-time nonlinear systems of the form

1 f (x(k); y(k))
x(k + 1) = f (x(k); y(k); ) =
(1)
1
y(k + 1) = g(x(k); y(k); ) = g(x(k); y(k))
(2)
n
where f (1) and g (1) are smooth functions for x 2 Dx 
and y 2
Dy  m satisfying f (0; 0) = g(0; 0) = 0. Here, Dx and Dy are
open convex neighborhoods of the origins of the respective Euclidian
spaces. k 2 and  is a smallpositive perturbation parameter. This
system will be studied under the following assumption.
Assumption 1: For all (x; y )T 2 Dx 2 Dy we assume that
i) supk0 kx(k)k  a1 < 1;
ii) supk0 ky (k)k  a2 < 1;
iii)  ((@f=@x)(x; y )) <  where  (1) denotes the spectrum of
a matrix and 0 <  < 1.
System (1) and (2) is characterized by an explicit separation of variables x(k) and y (k), owing to the presence of the small parameter
 that multiplies the state y(k). This fact allows decomposing such
system into separate reduced-order subsystems. Indeed, the reduced
subsystem is obtained by setting  = 0, then the dimension of the
system (1) and (2) reduces from n + m to n because the system (1)
and (2) degenerates into the n-dimensional system

xs (k + 1) = f (xs (k); 0)
ys (k + 1) = g(xs (k); 0)

(3)
(4)

where xs and ys refer to the reduced (or slow) solutions.

Assumption 2: Assume that, the Jacobian matrix (@f (x; 0))=@x is
nonsingular in Dx .
Thus, in view of the implicit function theorem implies that there exists a unique smooth function h(1) mapping Dx onto itself such that

xs (k) = h(xs (k + 1)):

(5)

To obtain the reduced model we substitute (5) into (4)

1 f (x (k))
xs (k + 1) = f (xs (k); 0) =
(6)
s s
1
ys (k) = g(h(xs (k)); 0) = gs (xs (k)):
(7)
It is straightforward to see that xs starts from the same initial condition
as x. By contrast ys is not free to start from y0 , hence the approximation
of x by xs may be uniform for all k 2 that is
x(k) = xs (k) + O():
(8)
However, the approximation

y(k) = ys (k) + O()

(9)

hold only on an interval excluding 0, that is, for k 2 fk1 ; k1 + 1; . . .g,

where k1 > 0. Thus, following the nomenclature of the singular perturbation literature [5], (1) and (2) is said to be the singularly perturbed
form and the boundary layer occurs at k = 0. Layer correction terms
must thus be added to the reduced solution ys (k) for the approximation
(9) to be valid over the entire interval. To investigate this initial value
problem we introduce the following transformation as suggested by [4]
for nonlinear systems
x(k) = k+1 xf (k); y(k) = k yf (k)
(10)

00189286/01$10.00 2001 IEEE