Proceedings of the 2007 American Control Conference WeC09.

6
Marriott Marquis Hotel at Times Square
New York City, USA, July 11-13, 2007

Further results on stability of networked control systems: a Lyapunov

approach
Daniele Carnevale and Andrew R. Teel and Dragan Nesic

Abstract Simple Lyapunov proofs are given for an improved and input-output stability when the MATI [0, M AT I ] with
(relative to previous results that have appeared in the literature)
bound on the maximum allowable transfer interval to guarantee 1 1
M AT I ln 1 + (1)
global asymptotic or exponential stability in networked control L L +
systems and also for semiglobal practical asymptotic stability
with respect to the length of the maximum allowable transfer where [0, 1) characterized the contraction of the proto-
interval. We apply our results to emulation of nonlinear cols Lyapunov function at transmission times while L > 0
controllers in sampled-data systems. described its expansion between transmission times, and
> 0 captured the effect of the error signals on the behavior
I. I NTRODUCTION of the ideal system through an Lp gain.1
In this paper, we will give a simple Lyapunov proof of an
A networked control system (NCS) is composed of multi- improved (larger) MATI bound, expressed in terms , L and
ple feedback control loops that share a serial communication corresponding to the case of L2 gains. (A similar approach
channel. This architecture promotes ease of maintenance, can be taken for the general Lp case.) In particular, we
greater flexibility, and low cost, weight and volume. On the establish uniform asymptotic or exponential stability when
other hand, if the communication is substantially delayed or
infrequent, the architecture can degrade the overall system M AT I !

performance significantly. Results on the analysis of an
1 r(1 )

arctan >L
NCS include [12], [13], [14], [7], [8]. In an NCS, the
Lr


2 1+ L 1 +1+
delay and frequency of communication between sensors and 1 1
actuators in a given loop is determined by a combination = L (2)

L 1+ !
of the channels limitations and the transmission protocol


1 arctanh r(1 )
used. Various protocols have been proposed in the literature,

<L,
Lr 2 1+ L 1 +1+
including the round robin (RR) and try-once-discard
(TOD) protocols discussed in [12] and [13]. We note that and note that in the first and last expressions we use respec-
sampled-data systems are a special case of NCS since in tively the trigonometric and hyperbolic functions, where
this case all sensor and control signals are transmitted at s
2
each transmission instant. r := 1 . (3)
When the individual loops in an NCS are designed assum- L
ing perfect communication, the stability of the NCS is largely It is not obvious, except for the case = L, that (2)
determined by the transmission protocol used and by the so- provides a larger bound than (1). We will establish that it is
called maximum allowable transfer interval (MATI), i.e., an improvement by first noting that the bound in (1) is the
the maximum allowable time between any two transmissions value 1 satisfying
in the network. When specialized to sampled-data systems,
this controller design approach is called emulation (see [5], 1 = L1 , 1 (0) = 1 , 1 (1 ) = (4)
[2]). Following pioneering work of Walsh and co-authors whereas the bound in (2) is the value 2 satisfying
[13], [12], we consider the problem of characterizing the
length of the MATI for a given protocol to ensure uniform 2 = 2L2 (22 + 1) , 2 (0) = 1 ,
2 (2 ) =
global asymptotic or exponential stability. We also demon- (5)
strate that our results apply in a straightforward manner to and that necessarily 2 > 1 for all L > 0, > 0 and
emulation of nonlinear controllers for digital implementation. (0, 1). The above equations can be obtained from appropriate
This appears to be the first result in the literature that Lyapunov arguments that are presented later in our proofs.
provides an explicit formula for the computation of MATI The difference in the bounds for the batch reactor system
in the context of emulation of nonlinear controllers. considered in [7] is reported in Table I. (For more discussion,
In [7] the authors were able to improve on the initial MATI see Remark 1.) The improvement p is on the order of 10%.
bounds given in [13], [12] by efficiently summarizing the (The values L = 15.73, = 1/2, = 15.9222 for the
properties of protocols through Lyapunov functions and char- 1 For convenience, in a minor departure from the description in [7], we
acterizing the effect of transmission errors through Lp gains. use an inequality rather than a strict inequality in (1) but take to be any
They established uniform asymptotic or exponential stability number that is strictly greater than the Lp gain used in [7].

1-4244-0989-6/07/$25.00 2007 IEEE. 1741

 WeC09.6

TABLE I
II. N OTATION AND DEFINITIONS
B OUNDS C OMPARISON FOR TOD/RR PROTOCOLS : BATCH R EACTOR IN
[7] We denote by R and Z the sets of real and integer numbers,
respectively. Also R0 = [0, +), and Z0 = {0, 1, 2, . . . }.
Definition TOD RR The Euclidean norm is denoted | |. A function : R0
theoretical bound on MATI 0.01 0.0082 R0 is said to be of class K if it is continuous, zero at
M AT I computed in [7] zero and strictly increasing. It is said to be of class K if
theoretical bound on MATI M AT I 0.0108 0.009 it is of class K and it is unbounded. A function : R0
via Theorem 1 R0 R0 is said to be of class KL if (, t) is of class
percentage of increment 8.4 % 9.76% K for each t 0 and (s, ) is nonincreasing and satisfies
using the new M AT I limt (s, t) = 0 for each s 0. A function : R0
R0 R0 R0 is said to be of class KLL if, for each
r 0, (, r, ) and (, , r) belong to class KL.
TOD protocol and = 21.5275 for the RR protocol are We recall definitions given in [3] that we will use to
reported in [7].) For some systems, the improvement could develop a hybrid model of a NCS. The reader should refer to
be over 50%. See the figures below, which address separately [3] for the motivation and more details on these definitions.
the case < L and L. Definition 1: A compact hybrid time domain is a set D
We emphasize that the contribution of this paper is not R0 Z0 given by :
only a (modest) improvement in the MATI bound relative J1
[
to [7] but also a simple Lyapunov proof. At the same D= ([tj , tj+1 ], j )
time, we give a direct Lyapunov proof of a result in [8] j=0
which states that if an NCS is asymptotically stable with
perfect communication then it is semiglobally practically where J Z0 and 0 = t0 t1 tJ . A hybrid
asymptotically stable with respect to M AT I . This proof time domain is a set D R0 Z0 such that, for each
also generalizes easily to the case, addressed in [8], where (T, J) D, D ([0, T ] {0, . . . , J}) is a compact hybrid
there are exogenous inputs to which the system with perfect time domain.
communication is input-to-state stable. Definition 2: A hybrid trajectory is a pair (dom , )
consisting of hybrid time domain dom and a function
defined on dom that is continuously differentiable in t on
(dom ) (R0 {j}) for each j Z0 .
60
Definition 3: For the hybrid system H given by the open
Percentage of Increment

state space O Rn and the data (F, G, C, D) where F :

50

40

30
O Rn is continuous, G : O O is locally bounded,
20 and C and D are subsets of O, a hybrid trajectory : dom
10 O is a solution to H if
0

1) For all j Z0 and for almost all t Ij := dom

0
100 0.5
80
60
40
20
~

0 1

(R0 {j}), we have (t, j) C and (t, j) =
F ((t, j)).
2) For all (t, j) dom such that (t, j + 1) dom ,
Fig. 1. Percentage improvement in the MATI bound using Theorem 1 we have (t, j) D and (t, j + 1) = G((t, j)).
compared to using [7, Theorem 4],
= /L 1.
Hence, the hybrid system models that we consider are of the
form:
j) =
(t, F ((t, j)) (t, j) C
45
(tj+1 , j + 1) = G((tj+1 , j)) (tj+1 , j) D .
40
Percentage of Increment

35

30 We sometimes omit the time arguments and write:

25

20

15
= F () C (6)
+
10
= G() D ,
5

0 0
1 0.8 0.6
~ 0.4 0.2 1
0.5
where we denoted (tj+1 , j + 1) as + in the last equation.
We also note that typically C D 6= and, in this case,
if (0, 0) C D we have that either a jump or flow
Fig. 2. Percentage improvement in the MATI bound using Theorem 1, is possible, the latter only if flowing keeps the state in C.
compared to using [7, Theorem 4],
= /L < 1.
Hence, the hybrid model we consider may have non-unique
solutions.

1742
 WeC09.6

III. P ROBLEM STATEMENT
In this section, we formally state the problem that we
consider and summarize the model of NCS from [7]. In the
next section, we will embed this model within the hybrid
framework of [3] by representing it in the form (6) that is
useful in our proofs.
We pursue the controller design technique proposed in
[12], [13] and further developed in [7], [8]. The plant model
Note that if NCS has links, then the error vector can be
partitioned as follows e = [eT1 eT2 . . . eT ]T . The functions
hu and hy are typically such that, if the ith link gets access
to the network at some transmission time tj we have that
the corresponding part of the error vector has a jump. For
several protocols, such as the round robin and Try-Once-
Discard protocols (see [7]), we typically assume that ei is
reset to zero at time t+ +
j , that is ei (tj ) = 0. However, we

is given by equations: emphasize that this assumption is not needed in general. This
allows us to write the models hu , hy for protocols commonly
x P = fP (xP , u) (7) found in the literature (see [7], [8] for more details).
y = gP (xP ) . We combine the controller and plant states into a vector
x := (xP , xC ) and using the error vector defined earlier
The first step in controller design is to ignore the network e = (ey , eu ), we can rewrite (9) as a system with jumps that
and design a stabilizing controller for the plant: is more amenable for analysis:
x C = fC (xC , y) (8) x = f (x, e) t [tj1 , tj ] (10)
u = gC (xC ) . e = g(x, e) t [tj1 , tj ] (11)
The second step in the design is to implement the above e(t+
j ) = h(j, e(tj )) , (12)
controller over the network and determine the value of a
network parameter (MATI) that guarantees that the same where tj+1 tj for all j Z0 , x Rnx ,
controller implemented over the network will yield stability. e Rne and f , g and h are obtained using straightforward
Note that this approach is very similar to the emulation calculations from (9), see [7]. In order to write (11), we
approach to controller design of sampled-data systems. assumed that functions gP and gC in (9) are continuously
Now we describe the model of NCS. Let the sequence differentiable (this assumption can be relaxed). We refer to
tj , j Z0 of monotonically increasing transmission times (12) as a protocol. The protocol determines the algorithm by
satisfy tj+1 tj for all j Z0 and some which access to the network is assigned to different nodes
fixed , > 0. Note that is arbitrary and it is used to in the system. For more details on protocol modelling in this
prevent Zeno solutions in the model given below. At each manner, see [7], [8]. Note that
tj , the protocol gives access to the network to one of the
x = f (x, 0) (13)
nodes i {1, 2, . . . , }. We refer to as the maximum
allowable transmission interval (MATI). Using the plant (7) represents the closed loop system (7), (8) without the net-
and controller (8), we introduce the nonlinear NCS of the work. We consider the following problem:
following form
Problem: Suppose that the controller (8) was de-
x P = fP (xP , u) t [tj1 , tj ] signed for the plant (7) so that the closed loop
y = gP (xP ) system (7), (8) without network (equivalently, the
x C = fC (xC , y) t [tj1 , tj ] system (13)) is globally asymptotically stable. De-
u = gC (xC ) termine the value of M AT I so that for any
(9)
y = fP (xP , xC , y, u
) t [tj1 , tj ] (0, M AT I ] and all [, M AT I ], we have that
u = fC (xP , xC , y, u
) t [tj1 , tj ] the NCS described by (10), (11), (12) is stable in
y(t+ j ) = y(tj ) + hy (i, e(tj )) an appropriate sense.
u(t+ j ) = u(tj ) + hu (i, e(tj )) Moreover, we show that the value of M AT I computed
in [7] and given by (1) is always smaller than the value
where xP and xC are respectively states of the plant and
of M AT I given by (2). Hence, our new result provides
the controller; y is the plant output and u is the controller
a less conservative analytical bound for M AT I that is
output; y and u are the vectors of most recently transmitted
very important in implementing the controller (8) via the
plant and controller output values via the network; e is the
network in the manner described by (9). Indeed, this bound
network induced error defined as
shows that stabilization is possible with lower bandwidth
y(t) y(t) ey of the communication channel (since M AT I is inversely
e(t) := = .
u(t) u(t) eu proportional to the channel bandwidth).
We often use the choice fP = 0 and fC = 0 which means IV. M AIN RESULTS
that the networked version of the output y and control u
are kept constant during the transmission intervals (i.e., the In order to streamline the proofs, we map the model (10),
network nodes operate in a similar manner to a zero order (11), (12) of an NCS that was introduced in the previous
hold). section into a hybrid system of the type (6) discussed in the

1743
 WeC09.6

preliminaries section. In particular, we consider systems of for all (t, j) in the solutions domain.
the form In order to guarantee asymptotic or exponential stability,
we make the following assumption:
x = f (x, e)
Assumption 1: There exist a function W : Z0 Rne
e = g(x, e)
[0, M AT I ] R0 that is locally Lipschitz in its second argument, a locally
= 1

= 0 Lipschitz, positive definite, radially unbounded function V :
x+ = x (14)

Rnx R0 , a continuous function H : Rnx R0 , real
+
e = h(, e) numbers (0, 1), L 0, > 0, W , W K and a
[, )
+ = 0
continuous, positive definite function such that, Z0
+
= +1 and e Rne
where > 0 can be arbitrarily small, M AT I and x W (|e|) W (, e) W (|e|) (17)
Rnx , e Rne , R0 and Z0 .
In what follows we will consider the behavior of all W ( + 1, h(, e)) W (, e) (18)
possible solutions to the hybrid system (14) subject to
and for all Z0 , x Rnx and almost all e Rne ,
(0, 0) 0. Since the derivative of is positive (equal
to one) and when jumps it is reset to zero, it follows W (, e)
, g(x, e) LW (, e) + H(x) ; (19)
that will never take on negative values. According to the e
definition of solution for a hybrid system, the error vector e moreover, for all e Rne , all Z0 , and almost all
can jump, following the rules of the protocol, after seconds x Rnx ,
have elapsed from the previous jump. This is because at
the previous jump was reset to zero, when the system is hV (x), f (x, e)i (|x|) (W (, e)) H 2 (x)
not jumping we have = 1, and the D set, which enables + 2 W 2 (, e) . (20)
jumps, is the set {(x, e, , ) : [, )}. On the other
hand, if M AT I seconds have elapsed from the previous
jump then the error vector e must jump. This is because Remark 1: This assumption is essentially the same as the
the C set is {(x, e, , ) : [0, M AT I ]}, and thus flows main assumption of [7, Theorem 4] when considering L2
are not allowed after reaches M AT I . In this way, the gains. The condition on x = f (x, e) is expressed here in
time-invariant hybrid system (14) covers all of the possible terms of a Lyapunov function that establishes an L2 gain
behaviors described by (10), (11), (12). from W to H whereas in [7, Theorem 4] it is directly in
Standing Assumption 1: f and g are continuous and h is terms of the L2 gain . However, in practice the L2 gain is
locally bounded. often verified with a Lyapunov function V that satisfies (20).
We will give an upper bound on M AT I to guarantee For example, the results in the first row of Table I, which
asymptotic or exponential stability. come from [7], use values (, L, ) that admit functions W ,
Definition 4: For the hybrid system (14), the set H and a positive definite, quadratic function V that satisfy
{(x, e, , ) : x = 0 , e = 0} is uniformly globally asymp- (17)-(20) with (s) = s2 for some > 0 sufficiently small.
totically stable if there exists KLL such that, for each
initial condition (0, 0) R0 , (0, 0) Z0 , x(0, 0) Theorem 1: Under Assumption 1, if M AT I in (14) satis-
Rnx , e(0, 0) Rne , and each corresponding solution, fies the bound (2) and 0 < M AT I then, for the system
(14), the set {(x, e, , ) : x = 0, e = 0} is uniformly glob-
x(t, j)
x(0, 0) , t, j (15) ally asymptotically stable. If, in addition, there exist strictly
e(t, j) e(0, 0) positive real numbers W , W , a1 , a2 , and a3 such that
for all (t, j) in the solutions domain with > 0 (it avoids W |e| W (, e) W |e|, a1 |x|2 V (x) a2 |x|2 , and
Zeno solutions). The set is uniformly globally exponentially (s) a3 s2 then this set is uniformly globally exponentially
stable if can be taken to have the form (s, t, k) = stable.
In the proof of Theorem 1, Sec. VI-A it is shown that
M s exp((t + k)) for some M > 0 and > 0.
V (x) + W 2 (, e) is a strict Lyapunov function for the
Definition 5: For the hybrid system (14) the set
discrete-time system that is generated as the composition of
{(x, e, , ) : x = 0 , e = 0} is uniformly semiglobally
flows and jumps in the system (14).
practically asymptotically stable (USPAS) with respect to
Theorem 2: Consider the hybrid NCS (14). Suppose that
M AT I if there exists KLL and for any pair of positive
the following conditions hold.
real numbers (, ) there exists M AT I > 0 such that for
each 0 < M AT I , each initial condition (0, 0) R0 , 1) There exist a function W : Z0 Rne R0 that is
(0, 0) Z0 , |x(0, 0)| , |e(0, 0)| and each locally Lipschitz in its second argument, a continuous,
corresponding solution we have positive definite function and class-K functions
W , W , such that, Z0 and e Rne ,
x(t, j)
max x(0, 0) , t, j , ,
e(t, j) e(0, 0) W (|e|) W (, e) W (|e|) (21)
(16) W ( + 1, h(, e)) W (, e) (e) (22)

1744
 WeC09.6

and for all Z0 and almost all e Rne , Below, by abuse of notation, we consider the quantity
hU (), F ()i even though W is not differentiable with
W (, e)
(|e|) . (23) respect to . This is justified since the component in F ()
e
corresponding to is zero. It is easy to check that for all
2) The origin of x = f (x, 0) is globally asymptotically (, ) and that for almost all (x, e)
stable.
U ( + ) = V (x+ ) + ( + )W 2 (+ , e+ )
Then, for (14), the set {(x, e, , ) : x = 0, e = 0} is US-
PAS with respect to M AT I . V (x) + W 2 (, e) U (). (27)
hU (), F () i (|x|) (W (, e)) .
V. E MULATION IN SAMPLED - DATA SYSTEMS
In this section, we specialize Theorem 1 to the case of Since is positive definite, V is positive definite and radially
emulation of continuous-time controllers. We believe that the unbounded, and Claim 1 holds, it follows that there exists a
explicit formula which we provide is the first one reported continuous, positive definite function e such that
in the literature in this context. For space reasons, we do not hU (), F ()i e
(U ()) . (28)
re-state the result of Theorem 2.
First, we note that sampled-data systems are a special case Then, by standard results for continuous-time systems (see,
of NCS (see (14)) and they can be described by the following for example, [10]) and using (27), we have the existence of
model: KL satisfying
)
x = f (x, e) (s, t1 + t2 ) = ((s, t1 ), t2 ) (s, t1 , t2 ) R0R0R0 ,
e = g(x, e) [0, M AT I ] U ((t, j)) (U ((0, 0)), 0.5t + 0.5j) (t, j) dom .
= 1
x+ = x (24) Then, using that V is positive definite and proper, using (17),
+
e = 0 [, ) Claim 1, and the definition of U in (26), uniform global
+ = 0 asymptotic stability of the set {(x, e, , ) : x = 0, e = 0}
where the main difference with (14) is the simplified model follows.
of the protocol (e+ = 0) and the absence of the equations. Under the assumptions made in the theorem to guarantee
In other words, u and y are transmitted at each transmission uniform global exponential stability, it follows that e can be
instant, or equivalently, there is only one link. A straight- taken to be linear and can be taken to be of the form
forward consequence of this special structure is that for any (s, t) = M s exp(t). Then uniform exponential stability
function W that satisfies (17), (19) and (20), we have that it follows from the quadratic upper and lower bounds on V (x)
also satisfies (18) for any [0, 1) (in particular, we can let and W 2 (, e). The proof will be complete after we prove
= 0). Using this, a direct consequence of Theorem 1 is the Claim 1, which we will do in Section VII-B.
following result on emulation of controllers in sampled-data B. Proof of Theorem 2
systems: Using (21) and (22), one can combine the ideas in [6] and
Corollary 1: Suppose that (17), (19) and (20) in As- [9, p. 22-23] to get a continuously differentiable function
sumption 1 hold. If MATI in (24) satisfies the bound (2) K and > 0 such that with W f (, e) := (W (, e))
with = 0 and 0 < M AT I then, for the system we have
(24), the set {(x, e, ) : x = 0, e = 0} is uniformly globally
asymptotically stable. If, in addition, there exist strictly W f (, e) .
f ( + 1, h(k, e)) e W (29)
positive real numbers W , W , a1 , a2 , and a3 such that Using the last assumption of the Theorem, let the smooth
W |e| W (e) W |e|, a1 |x|2 V (x) a2 |x|2 , and function V be the one obtained from Kurzweils converse
(s) a3 s2 then this set is uniformly globally exponentially Lyapunov theorem [4], satisfying
stable.
The proof of Corollary 1 follows directly from the proof hV (x), f (x, 0)i V (|x|) (30)
of Theorem 1 by letting 0+ in the formula (2). for some V K . Using the definition of and F () from
VI. P ROOFS OF MAIN RESULTS the proof of Theorem 1, define
A. Proof of Theorem 1 f (, e) .
U () := V (x) + e /M AT I W (31)
Let : [0, M AT I ] R be the solution to Then, using (29), (14), and (31), we get
= 2L (2 + 1) (0) = 1 . (25) f (, e) U () .
U ( + ) V (x) + e W (32)
We will establish the following claim in the next section: Using the continuity of f , (30), (21) and (23), we also have
Claim 1: ( ) [, 1 ] for all [0, M AT I ]. the existence of a continuous function satisfying (x, 0) =
T
We will use the definitions := xT , eT , , and F () := 0 for all x and such that
T
f (x, e)T , g(x, e)T , 1, 0 . Define hU (), F ()i V (|x|) + (x, e)
1 /M AT I (33)
U () := V (x) + ( )W 2 (, e) . (26) M AT I e W (|e|) .

1745
 WeC09.6

Now the continuous-time arguments given in [1] or [11, sgn(0) = 0. The first formula in (2), R b when >
Lemma 2.1] can be used to assure that, for each pair of L, comes from using the fact that 1 a (Lr/) ds
=
h a i
2 +s2
strictly positive real numbers e <
e there exists M AT I > 0 1 b
Lr arctan Lr arctan Lr and that for all c2
such that, for almost all in the set
c1 0 we have arctan(c2 ) arctan(c1 ) = arctan((c2
x c1 )/(1 + c1 c2 )). The second formula
e
(x, e, , ) :
e
e
, [0, M AT I ] , Z0 Rb L = ,
in (2), when
follows from the fact that 1 a ds s 2 = 1 1
b 1
. The third
R a ds
1 b
we have formula in (2), when < L follows from a (Lr/)2 s2 =
h a i
1 b
hU (), F ()i 0.5V (|x|) 0.5W (|e|) . (34) Lr arctanh Lr arctanh Lr . Then, the last formula
The result follows using standard continuous-time arguments
in (2) is obtained by using the identity arctanh(c2 )
c2 c1
like in the proof of Theorem 1. arctanh(c1 ) = arctanh 1c 2 c1
.

VII. P ROOF OF C LAIM 1 AND THAT THE BOUND IS VIII. C ONCLUSIONS

BETTER We have provided a simple Lyapunov proof for certain
A. A race between differential equations results that have appeared previously in the literature on the
In this section we establish the following fact: stability of networked control systems. Along the way, we
Lemma 1: For each (0, 1), the value 1 in (4) is less have provided some modest improvements to the previous
than the value 2 in (5). results. We hope that the Lyapunov approach to proving
This lemma shows that Claim 1 in the proof of Theorem 1 stability for networked control systems will lead to better
holds when M AT I satisfies the bound given by the right- insight into the design of protocols for these systems and
hand side of (1). Thus, the proof of Theorem 1 is complete will also inspire even sharper analysis tools.
in this case. In the next subsection, we establish that the R EFERENCES
bound (2) is equal to the value 2 in (5). This will establish [1] A. Bacciotti. Linear feedback: the local and potentially global
Claim 1 and finish the proof of Theorem 1 as it is stated. It stabilization of cascade systems. In Proc. IFAC Nonlinear Control
will also confirm that the bound on M AT I reported here is Systems Design Symposium, pages 2125, Bordeaux, 1992.
[2] T. Chen and B.A. Francis. Input-output stability of sampled-data
larger than the bound reported in [7]. systems. IEEE Transactions on Automatic Control, 36:5058, 1991.
Proof of Lemma 1. Note that 2 = 2+ + 2 where 2+ [3] R. Goebel and A. R. Teel. Solutions to hybrid inclusions via set and
and 2 satisfy graphical convergence with stability theory applications. Automatica,
pages 573 587, April 2006.
2 = 2L2 (22 + 1), (35) [4] J. Kurzweil. On the inversion of Lyapunovs second theorem on
stability of motion. American Mathematical Society Translations,
2 (0) = 1, 2 (2+ ) = , 2 (2 ) = 1 . Series 2, 24:1977, 1956.
[5] D.S. Laila, D. Nesic, and A. R. Teel. Open and closed loop dissipation
Define 1+ := 1 . Let 1 (), respectively 2 (), denote inequalities under sampling and controller emulation. European
the solution of (4), respectively (35). Since 1 (1+ ) = Journal of Control, 18:109125, 2002.
[6] D. Nesic and A. R. Teel. Changing supply functions in input to
2 (2+ ) = and 2 (2 ) = 1 , we have 1 = did (i+ )
= state stable systems: the discrete-time case. IEEE Transactions on
d(i+ ) di+ 2 d ( ) d( ) d
di+ d and = 2
d
2
= d2
2 2
d . These Automatic Control, 46(6):960 962, 2001.
[7] D. Nesic and A. R. Teel. Input output stability properties of
equations yield networked control systems. IEEE Transactions on Automatic Control,
d1+ 1 d2+ 1 49(10):16501667, 2004.
= , = , [8] D. Nesic and A. R. Teel. A note on input-to-state stability of networked
d L + d 2L + (2 + 1) control systems. Automatica, 40(12):21212128, 2004.
d2 1 1 [9] L. Praly and Y. Wang. Stabilization in spite of matched unmodeled
= 2 = . dynamics and an equivalent definition of input-to-state stability. Math-
d (2L1 + (2 + 1)) 2L + (2 + 1) ematics of Control, Signals, and Systems, 9:133, 1996.
[10] E. D. Sontag. Smooth stabilization implies coprime factorization.
Using 2 + 1 < 2, 1+ = 1 , and 2 = 2+ + 2 gives IEEE Transactions on Automatic Control, 34(4):435443, 1989.
[11] A.R. Teel and L. Praly. Tools for semiglobal stabilization by partial
d2 d1
< . (36) state and output feedback. SIAM J. Control Optim., 33(5):14431488,
d d 1995.
[12] G. C. Walsh, O. Beldiman, and L. G. Bushnell. Asymptotic behavior of
Since 1 = 2 = 0 when = 1, the condition (36) nonlinear networked control systems. IEEE Transactions on Automatic
establishes the lemma. Control, 46(7):10931097, 2001.
[13] G. C. Walsh, H. Ye, and L. G. Bushnell. Stability analysis of
B. Proof of Claim 1 networked control systems. IEEE Transactions on Control Systems
Technology, 10(3):438446, 2002.
Claim 1 follows immediately from the following lemma. [14] W. Zhang and M. S. Branicky. Stability of networked control
Lemma 2: The right-hand side of (2) is equal to the value systems with time-varying transmission period. In Allerton Conf.
2 in (5) (cf. (25)). Communication, Control, and Computing, Urbana, IL, October 2001.
R d
Proof. By definition we can write 2 = 1 2 +2L+ =
R
1 +
L
ds L
1 + L 2 Lr 2
, where s := + , r
s sgn(L)( )
is defined in (3) and sgn() is the sign function with

1746