Stability of networked control systems with variable sampling and delay
Payam Naghshtabrizi and Joao P. Hespanha
Abstract We consider Networked Control Systems (NCSs)
consisting of a LTI plant; a linear static or dynamic feedback
controller; a collection of sensors that provide measurements
to the controller; and a collection of actuators that are used
to control the plant. The different elements of the control
system are spatially distributed, but interconnected through
a communication network. Due to the shared and unreliable
channel used to connect the subsystems, the sampling intervals
are uncertain and variable. Moreover, samples may be dropped
and experience uncertain and variable delays before arriving
at the destination. We show that the resulting NCSs can be
viewed as a MIMO sampled-data system with variable sampling
intervals and delay, which can be modeled by linear infinite-
dimensional impulsive systems. The infinite dimensionality of
the system arises from the existence of delays. We provide
conditions for the stability of the closed-loop expressed in terms
of LMIs. By solving these LMIs, one can determine positive
constants related to each entity sent through the network that
determines an upper bound between the sampling time and
the next update time at the destination of that entity, for which
stability of the closed-loop system is guaranteed.
I. INTRODUCTION
Network Control Systems (NCSs) are spatially distrib-
uted systems in which the communication between plants,
sensors, actuators, and controllers occurs through a shared
band-limited digital communication network. Using network
as a medium to connect spatially distributed elements of
the system results in flexible architectures and generally
reduces wiring and maintenance cost, since there is no need
for point to point wiring. Consequently, NCSs have been
finding application in a broad range of areas such as mobile
sensor networks, remote surgery, haptics collaboration over
the Internet and unmanned aerial vehicles [4]. However, the
use of a shared network, in contrast to using several dedicated
independent connections, introduces new challenges: the
sampling intervals are uncertain and variable, samples may
be dropped and experience uncertain and variable delays
before arriving at the destination.
We start by considering an abstract single-input single-
output (SISO) sampled-data system of the form
x(t) = Ax(t) + Bx(sk ), tk t < tk+1 , k N, (1)
which models the closed-loop system in Fig. 1, where sk
denotes the k-th sampling time instant and tk is the time at
which the k-th sample arrives to the destination. If k denotes
the total delay that the k-th sample experiences in the loop,
This material is based upon work supported by the Institute for Collab-
orative Biotechnologies through grant DAAD19-03-D-0004 from the U.S.
Army Research Office, and by the National Science Foundation under Grant
No. CCR-0311084.
P. Naghshtabrizi and J.P. Hespanha are with the Department of Electrical
Engineering at the University of California, Santa Barbara.
sk
z(t) x(t)
Hold x = Ax + Bu
Delay k
x(sk )
Fig. 1. Sampled-data system with variable sampling intervals and delays
where u(t) = z(t) = x(sk ) for sk + k t < sk+1 + k+1
then tk := sk + k . Equation (1) can be used to model a NCS
in which a linear plant x(t) = Ax(t) + Bu u is in feedback
with static state-feedback remote controller Kx where B =
Bu K in (1). We introduce a new discontinuous Lyapunov
functional to establish stability of the closed-loop based on
the theorems in [9] developed for a general nonlinear time-
varying delay impulsive systems . The Lyapunov functional
is discontinuous at the impulse times but its decrease is
guaranteed by construction. The stability test is presented
as LMIs that can be solved numerically. By solving these
LMIs, one can find a positive constant that determines the
upper bound between the sampling time sk and the next input
update time tk+1 = sk+1 + k+1 for which the stability of the
closed-loop system is guaranteed, assuming a given lower
and upper bound on the total delay in the loop. When there is
no delay this upper bound simply determines the maximum
sampling interval which is often called MAT I in the NCS
literature, e.g., in [14]. We use the terminology MAT I for the
case when there are delays in the systems too, and one can
state our result as follows: the system (1) is exponentially
stable if sk+1 + k+1 sk MAT I and min max for
k N where MAT I appears in our LMIs.
We also consider an abstract multi-input multi-output
(MIMO) sampled-data system of the form
x(t) = Ax(t) + Bu(t), y(t) = Cx(t), (2)
which models the closed-loop system in Fig. 4. The input and
the output are partitioned as y := [ y01 y0m ]0 and u := [ u01 u0m ]0
and the partition of u matches the partition of y. At time
sk , k N the ith output of the system, yi (t), 1 i m
is sampled and yi (sk ) is sent to update the input ui , to be
used as soon as it arrives until the next update arrives. This
framework is general enough to model both one-channel
NCSs and two-channel NCSs with dynamic output-feedback
controllers that may or may not be antecipative [7]. For LTI
processes and controllers, we present two stability tests in
terms of LMIs. The first one is less conservative but the
number of LMIs grows exponentially with m. The second II. S AMPLED - DATA SYSTEMS WITH VARIABLE SAMPLING
stability condition is based on the feasibility of a single LMI INTERVALS AND DELAYS
with dimention that grows linearly with m. For small m the Consider a linear delay sampled-data system of the form
first stability test is more desired, because it leads to less
conservative results, but the second stability test is more x(t) = Ax(t) + Bx(sk ), tk t < tk+1 , k N, (3)
adequate for large m. which models the closed-loop system in Fig. 1 where sk
denotes the k-th sampling time and tk is the instance that
By solving the LMIs that guarantee stability, one obtains the k-th sample arrives to the destination. If k denotes the
positive constants i max , 1 i m that determine the upper total delay that the k-th sample experiences in the loop, then
bound between the sampling time sk and the next input tk := sk + k . The closed-loop system can be modeled by the
update time tk+1 = sk+1 + k+1 for each i which the stability following impulsive system with delay
of the closed-loop system is guaranteed for a given lower
and upper bound on the total delay in each loop. In practice, (t) = F (t), tk t < tk+1 , (4a)
 
these constants produce deadlines for sampling that can be x(tk+1 )
(tk+1 ) = , k N, (4b)
used to design protocols for communication scheduling. x(sk+1 )
where
Significant work has been devoted to finding MAT I ([4]    
A B x(t)
and references therein). First we review the work which F := , (t) := ,
consider NCSs that can be presented as (1). In [3, 16, 8] 0 0 z(t)
the effect of delay is ignored and the problem of finding tk := sk + k , z(t) := x(sk ), tk t < tk+1 .
MAT I is formulated as LMIs. In the presence of variable We will seek for a class S of impulse-delay sequences
delays in the control loop, [2, 7, 15] show that if min , the {sk , k } such that the infinite-dimensional impulsive sys-
lower bound on the delay in the control loop, is given, the tem (4) is globally uniformly exponentially stable. We say
stability is guaranteed for a less conservative MAT I . Our that the system (4) is globally uniformly exponentially stable
result depends not only on min but also on an upper bound over a class S of impulse-delay sequences, if for every se-
on the delay in the control loop, which we denote by max . quence of impulse-delay in the S and every initial condition
Through examples we show that assuming an upper bound x(t0 ), x(t0 0 ) the solution to (4) is globally defined and
max on the delay reduces the conservativeness greatly. satisfies |x(t)| c max(|x(t0 )|, |x(t0 0 )|) e (tt0 ) , for some
c, > 0 and t t0 .
When the delay in the feedback loop is small (min , max
Consider the Lyapunov functional
0), our LMIs reduce to the ones presented in [8] which
8
are less conservative than [3, 16]. The following references
V := Vi , (5)
consider a more general framework where there are many i=1
nodes sending data to the network. To orchestrate network
where
access, Walsh et al. [13, 14] consider Round-Robin (sta-
tic) and Try-Once-Discard (TOD) (dynamic) protocols and V1 := x0 Px,
they find MAT I which determines the maximum deference
Z t
V2 := (max t + s)x0 (s)R1 x(s)ds,
between any consecutive sampling times. Nesic and Teel t
[10, 11] study the input-output stability properties of general Z t
nonlinear NCSs using an argument based on small gain V3 := (max t + s)x0 (s)R2 x(s)ds,
t
theorem to find MAT I for NCSs. Z t
V4 := (min t + s)x0 (s)R3 x(s)ds,
t
In section II we consider the sampled-data system in Fig. 1 Z tmin
min
and we model it as a linear delay impulsive system, then V5 := (max t + s)x0 (s)R4 x(s)ds,
we present stability test. To enlarge the class of NCSs we t
Z t
consider MIMO sampled-data system in Fig. 4 and then we V6 := (max min ) x0 (s)R4 x(s)ds,
present the stability test and classes of NCSs that can be tmin
Z t
presented by the MIMO sampled-data system in Fig. 4. We
V7 := x0 (s)Zx(s)ds,
finish by conclusions and future work in section IV. tmin
V8 := (max )(x w)0X(x w),
Notation: We denote the transpose of a matrix A by A0 and
we write P > 0 (or P < 0) when P is a symmetric positive with P, R1 , R2 , R3 , R4 , X, Z symmetric positive definite matri-
(or
 negative)
  definite
 matrix. We write a symmetric matrix ces and
A B A B
as . The notations 0i j , Ik are used to denote (t) := t sk , (t) := t tk , tk t < tk+1 ,
B0 C C
a i j matrix with zero entries and a k k identity matrix max := sup (t), max := sup (t),
t0 t0
and when there is no confusion about the size of the matrices
we drop the dimensions. w(t) := x(tk ) tk t < tk+1 .
 1.1 1.1

1.05 1.05
MATI

MATI
1 1

0.95 0.95

increases
max
0.9 0.9

0.85 0.85
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
min
min

Fig. 2. MAT I with respect to min based on Theorem 1 for max equal to Fig. 3. MAT I with respect to min from [7], [15] are shown by + and
0,0.1, ,0.7 and 1. respectively. The worse case (max = MAT I ) and the best case (max = min )
from Theorem 1 are shown by and o respectively.

The LMIs in the next theorem guarantee that the system (4)
with the Lyapunov functional (5) satisfies the conditions in the following polytopic condition:
[9] of global uniform exponential stability over the set S n
f j j , 0 f j 1, fj = 1
o
,
such that sk+1 + k+1 sk MAT I (or equally tk+1 tk + j=1 j=1
k MAT I ) for a given MAT I > 0 and that min k max .
The Lyapunov functional is a positive functional where its where the vertices of the polytope are described by
decrease at the jump points is guaranteed by the construction j := A j B j . Stability of the system can be checked by
and the following LMI conditions guarantee the decrease of solving the LMIs in Theorem 1 for each of the individual
the Lyapunov functional along the solution of the system (4). vertices with the same matrix variables.
The proof of this result can be found in [9]. A. NCSs modeled by a sampled-data system
Theorem 1: The system (4) is exponentially stable for any Equation (3) models NCSs in which a linear plant with
delay and sampling interval satisfying min k max and state-space
sk+1 + k+1 sk MAT I , k N, provided that there ex- x p (t) = A p x p (t) + B pu p
ist symmetric positive definite matrices P, R1 , R2 , R3 , R4 , X, Z
and (not necessarily symmetric) matrices N1 , N2 , N3 , N4 that where x p Rn , u p Rm are the state and the input of the
satisfy the following LMIs: 1 plant, is in feedback with a static feedback gain K. At time sk ,

M1 +max (M2 +M3 ) max N1 min N3
 k N the plants state, x(sk ), is sent to the controller and the
max R1 0 < 0, (6a) control command Kx(sk ) is sent back to the plant to be used
min R3 as soon as it arrives until the next control command update.
M1 +max M2 max N1 min N3 max (N1 +N2 ) max N4

max R1 0 0 0
The total delay in the control loop that the k-th sample
min R3 0 0

< 0, (6b) experiences is denoted by k where min k max , k N.
max (R1 +R2 )

0 Then the closed-loop system equations can be written as (3)
max R4
with
where max = MAT I min , F := A B 0 0 , and
 
P x := x p , A := A p , B := B p K, (8)
M1 := F [ P 0 0 0 ] + 00 F + min F 0 (R1 + R3 )F
0
and exponential stability of the system can be concluded
0 from Theorem 1.
 I   I 0  I   I 0  0   0 0
0 0 Remark 2: We only index the samples that get to the
I X I + 00 Z 00 00 Z 00
destination, which enables us to capture packet drops [16].
0 0
 I 0 0 I I
Consequently, even if the sampling intervals are constant,
I 
0
I
N1 [ I I 0 0 ] 0 N1 N2 [ I 0 I 0 ] I 0
N20 because of the packet dropouts in the network, the NCS
0 0
I  0 can be represented by a sampled-data system with variable
0 0
N3 [ I 0 0 I ] 0 N3 N4 [ 0 I 0 I ] I
0
0 N4 ,
sampling intervals.
I I Example 1: Consider the state space plant model [1]
M2 := F 0 (R1 + R2 + R4 )F,       
x1 0 1 x1 0
I  = + u,
M3 := I 0
X F + F 0 X [ I 0 I 0 ] . (7) x2 0 0.1 x2 0.1
0  
with state feedback gain K = 3.75 11.5 , for which we
Remark 1: Suppose that the system matrices := A B
 
have
are not exactly known and instead they are specified through 
0 1
  
0  
A= , B= 3.75 11.5 .
1 All zero matrices and identity matrices are n n 0 0.1 0.1
   sk
I 0 Fh
By checking the condition that eig( e ) < 0 on z1
I 0 x = Ax + Bu
y1 (t)

a tight grid of h and finding the largest h, it is possible Hold

y = Cx
to show that the closed-loop system remains stable if and zm ym (t)
only if the constant sampling interval is smaller than 1.7s. Sampler
On the other hand when the sampling interval approaches
Delay k
zero, the system is described by a DDE and we can find the
maximum constant delay for which stability is guaranteed.
The system is stable if and only if all the roots of character-
istic function det(sI A Be0s ) are negative. We use the Fig. 4. MIMO sampled-data system with variable sampling intervals and
delays where ui (t) = zi (t) = yi (sk ) for tk t < t j where k and j are the
0 /2
Pade approximation e0 s = 1s
1+s0 /2 to derive the determinant consecutive sampling indexes in Ki , 1 i m.
polynomial. By the Routh-Hurwitz test,one can show that
the system is stable for any constant delay smaller than 1.36.
Comparing these numbers to the maximum variable sampling of u. In NCSs each yi represents the output of nodes that
interval 1.1137 and maximum variable delay 1.0744 reveals can send their information to the network in a single packet.
the conservativeness of our method. For example sensors spatially close to each other, send their
No-delay and variable sampling: When there is no de- measurements in a single packet or controller outputs sent
lay but the sampling intervals are variable, MAT I determines to actuators spatially close to each other, send the control
the upper bound on variable sampling intervals sk+1 sk . commands in a single packet.
The upper bound given by [3, 16, 7] (when min = 0) is At time sk , k N the ith output of the system, yi (t), 1
0.8696 which is improved to 0.8871 in [15]. Theorem 1 and i m is sampled and yi (ski ) is sent through the network to
[8] gives the upper bound equal to 1.1137. update the input ui , to be used as soon as it arrives until the
Variable-delay and sampling: Fig. 2 shows MAT I as a next update arrives. We define sets Ki , 1 i m as the set
function of min for different values of max obtained form of indexes of sampling times that are related to the output
Theorem 1. Note that the curve that shows the largest MAT I yi ;
in Fig. 2 is part of the curve in Fig. 3 shown by o. Fig. 3
Ki := {k | yi (t) is sampled at sk }.
shows MAT I with respect to min where the results from [7],
[15] are shown by +, respectively. The value of MAT I The total delay in the control loop that the k-th sample
given by [2] lays between + and in Fig. 3 and we did experiences is denoted by k where i min k i max , k
not show them. In Theorem 1, MAT I is a function of min Ki , 1 i m. We define tk := sk + k which is the time instant
and max . To be able to compare our result to others in the that the value of ui is updated. The overall system can be
literature we consider two values for max and we obtain written as an impulsive system of the form
MAT I as a function of min based on Theorem 1. Fist we
consider max = min , which is the case that the delay is (t) = F (t), tk t < tk+1 , (10a)
constant and equal to the value of min . The largest MAT I for

x(tk+1 )
.........
a given min for this case provided by Theorem 1 is shown
z1 (tk+1 )
by o in Fig. 3. The second case is when max = MAT I , which

.
..
is the case that there can be very large delays in the loop in (tk+1 ) = , k + 1 Ki , 1 i m, (10b)
zi (tk+1 )
compare to the sampling intervals. The largest MAT I for a
..

given min for this case provided by Theorem 1 is shown by .

zm (tk+1 )
in Fig. 3. One can observe that when the delays in the control
loop are small, our method shows a good improvement in
..

 
compare to other results in the literature. A . B x(t)
where F := ... ... ... , (t) := ...... , and
.. z(t)
III. M ULTI -I NPUT M ULTI -O UTPUT (MIMO) 0 . 0
SAMPLED - DATA SYSTEMS "z
1 (t)
#
To enlarge the class of NCSs modeled by impulsive z(t) := .. , zi (t) := yi (sk ), tk t < t j ,
.
systems we consider the MIMO sampled-data system in zm (t)
Fig. 4 with state space of the form where k and j are the consecutive sampling indexes in
x(t) = Ax(t) + Bu(t), y(t) = Cx(t), (9) Ki , 1 i m.
Since the minimum of delay in the network is typically
where x Rn is the overall state of the system, u Rq is the small and for simplicity of derivations, we assume that
overall input of the system, and y Rq is the overall output i min = 0, 1 i m. We now present two theorems for the
of the system. The input is partitioned as u := [ u01 u0m ]0 stability of the system (10). The first one is less conservative;
where ui Rqi , 1 i m and m i=1 qi = q and the output however, the number of LMIs grows exponentially with m.
is partitioned as y := [ y01 y0m ]0 where yi Rqi , 1 i m and The second stability condition is based on the feasibility of
mi=1 qi = q, so the partition of y is matched with the partion a single LMI, but its size grows linearly with m. For small
  
m the first stability test is more adequate because it leads where F := A B 0nq and
to less conservative results, but the second stability test is  
P
more desired for large m. We present our results for m = 2, M1 := F 0 [ P 0nq 0nq ] + 0qn F T10 X1 T1 T20 X2 T2 N11 T3
0qn
but deriving the stability conditions for other values of m is
straightforward by following the same steps. T30 N11
0
N21 T1 T10 N21
0
N12 T4 T40 N12
0
N22 T2 T20 N22
0
,
Inspired by the Lyapunov functional (5), We employ the M2i := F 0Ci0 (R1i + R2i )Ci F, M3i := Ti0 XiCi F + F 0Ci0 XTi ,
Lyapunov functional G1i := i max (N1i + N2i ), G2i := i max (R1i + R2i )
(13)
V := V1 + V2 + V3 + V4, (11)
with
where
C1 := [ Iq1 0q1 q2 ]C, C2 := [ 0q2 q1 Iq2 ]C,

V1 := x0 Px, T1 := [ C1 0q1 q I1 ]0 , T2 := [ C2 0q2 q I2 ]0 ,

2 Z t T3 := [ C1 I1 0q1 q ] , T4 := [ C2 I2 0q2 q ] ,
V2 := (i max t + s)y0i (s)R1i yi (s)ds,
i=1 ti I1 := [ Iq1 0 q1 q2 ] I2 := [ 0q2 q1 Iq2 ] (14)
2 Z t
Proof of Theorem 2. After taking the derivative of the Lya-
V3 := (i max t + s)y0i (s)R2i yi (s)ds,
i=1 ti
punov functional (11) and following the same steps as in the
2 proof of theorem 2 of [9] we conclude that the derivative is
V4 := (i max i )(yi wi )0 Xi (yi wi ), negative as long as
i=1 2
with P, R1i , R2i , Xi symmetric positive definite matrices and M1 + i max (M2i + M3i ) + i (M4i M3i ) < 0. (15)
i=1

i (t) := t sk , i (t) := t tk tk t < t j , Then we can prove that (15) is equivalent to (12). Details of
i max := sup i (t), i max := sup i (t), the proof can be found in [6].
t0 t0
It is possible to generalize Theorem 2 for an arbitrary m.
wi (t) := yi (tk ), tk t < t j However, the number of LMIs is 2m and the size of LMIs
and the number of scalar variables increases linearly. For
where k and j are the consecutive sampling indexes in
complex systems with large number of sending nodes, to
Ki , 1 i m. The next theorem guarantees that the Lya-
have a numerically tractable test, it is crucial that the number
punov functional (11) decreases along the solution to the
of LMIs grows linearly too. The next theorem presents
system (10).
another stability test which is more conservative; however,
Theorem 2: The system (10) is exponentially stable for
the stability test is based on the feasibility of a single LMI.
any delay and sampling interval satisfying 0 k i max and
Theorem 3: The system (10) is exponentially stable for
s j + j sk i max , where k and j are consecutive sampling
any delay and sampling interval satisfying 0 k i max
indexes in Ki , i {1, 2} provided that there exist symmetric
and s j + j sk i max , where k and j are the consec-
positive definite matrices P, R1i , R2i , Xi and (not necessarily
utive sampling indexes in Ki , i = 1, 2 provided that there
symmetric) matrices N1i , N2i that satisfy the following LMIs:
exist symmetric positive definite matrices P, R1i , R2i and
(not necessarily symmetric) matrices N1i , N2i that satisfy the
1 +1 max (M21 +M31 )+2 max (M22 +M32 ) 1 max N11 2 max N12 following LMIs:
M 
1 max R11 0 q1 q2 < 0,
2 max R12
1 +1 max M21 +2 max M22 1 max N11 2 max N12 G11
M G12

(12a) 1 max R11 0 q1 q2 0 q1 q1 0 q1 q2
2 max R12 0q2 q1 0 q2 q2 < 0,

M1 +1 max M21 +2 max (M22 +M32 ) 1 max N11 2 max N12

G11 G21 0 q1 q2
1 max R11 0q1 q2 0 q1 q1
< 0, G22
2 max R12

0 q2 q1
(16)
G21
(12b) where F := A B
 
0nq and
M1 +1 max (M21 +M31 )+2 max M22 1 max N11 2 max N12

G12  
1 max R11 0 q1 q2 0 q1 q2 P

2 max R12 0 q2 q2
< 0, M1 :=F 0 [ P 0nq 0nq ] + 0qn F N11 T3 T 0 N 0 N21 T1
3 11
0qn
G22
(12c)
0
T10 N21 N12 T4 T40 N12
0
N22T2 T20 N22
0
,
0 0
1 +1 max M21 +2 max M22 1 max N11 2 max N12 M2i :=F Ci (R1i + R2i )Ci F,
M G11 G12

1 max R11 0 q1 q2 0 q1 q1 0 q1 q2

2 max R12 0 q2 q1 0q2 q2 < 0,
G1i :=i max (N1i + N2i ), G2i := i max (R1i + R2i ).
G21 0 q1 q2
G22
(12d) with the matrix variables defined in (14).
 sk sk

z1 y p1 (t) y p1 (t)

Hold Controller Plant Hold Plant

zm p y pm p (t) y pm p (t)

Sampler Sampler

Network
Network
Sampler
yc1 (t) z1
Fig. 5. One channel NCSs with the plant (17) and the controller (18).
Controller Hold

ycmc (t) zm p
Proof of Theorem 3. After taking the time derivative of the sk

Lypunov functional and following the steps in [6], we con-

clude that the derivative is negative if (15) holds. Equation Fig. 6. Two channel NCSs with the plant (17) and the anticipative or
non-anticipative controller (18).
(15) with i = i max is a sufficient condition for (15). So we
obtain
x 
2 x := xpc . The closed-loop system can be written as the
M1 + i max (M2i + M4i ) < 0, impulsive system (10) where
i=1    
A B pCc B D
and by applying Schur Lemma we get (16). A := p , B := p c , C := C p 0 . (19)
 
0 Ac Bc
A. NCSs modeled by MIMO sampled-data systems
2) Two-channel NCS with non-anticipative controller:
The impulsive system (10) can be used to represent the Fig. 6 shows a two-channel NCS consisting of a plant with
distributed sensors/actuators configurations shown in Figs. 5, the state-space (17) in feedback with a non-anticipative con-
and 6. The LTI plant has the state space of the form troller with state-space (18) where Dc = 0. Non-anticipative
controllers are simply output-feedback controllers which a
x p (t) = A p x p (t) + B pu p (t), y p (t) = C p x p (t), (17)
single value control command is calculated. Now the con-
where x p Rn p , u p := [ u0p1 u0pmc ]0 Rmc , and y p := troller is located away from the actuators and the control
[ y0p1 y0pm p ] Rm p are the state, input and output of the commands should be sent through the network. The con-
plant. At time tk , k Ki , 1 i m p , sensor i sends y pi (tk ) trol signal for the actuator i, yci (t), is sampled at sk , k
to the controller, which arrives at the destination at time Ki , m p + 1 i m p + mc , and samples get to the actuator
tk := sk + k . When a new measurement of the sensor i i at tk := sk + k . Note that a non-anticipative control unit
arrives at the controller side, the corresponding value at the sends a single-value control command to be applied to the
hold block, zi , is updated and held constant until another actuator i of the plant and held until the next control update
measurement of the sensor i arrives (all other measurement of the actuator i arrives (all other actuator values remain
values remain unchanged when the value of the sensor i is unchanged while the value of actuator i is updated). Hence
updated). Hence uci (t) = zi (t) = y pi (tk ), tk t < t j where k u pi (t) = zi (t) = yci (tk ), tk t < t j , where k and j are the
and j are consecutive sampling indexes in Ki , 1 i m . consecutive sampling indexes in Ki , m p + 1 i m p + mc .
An output feedback controller (or a state feedback controller) So in this case
uses the measurements to construct the control signal. The
(
y pi , 1 i mp
controller has the state space of the form yi := ,
yci , m p + 1 i m p + mc
xc (t) = Ac xc (t) + Bc uc (t), yc (t) = Cc xc (t) + Dc uc (t), x 
(18) m := m p + mc and x := xpc . The closed-loop system with
state can be written as the impulsive system (10) where
where xc Rnc 1 , uc := [ u0c1 u0cm p ]0 Rm p 1 , yc :=      
A 0 0 Bp C 0
[ y0c1 y0cmc ] Rm p 1 Rmc 1 are the state, input and output A := p , B := ,C := p . (20)
0 Ac Bc 0 0 Cc
of the plant and matrices Ac , Bc ,Cc , Dc have the proper
dimensions. The main difference between the different 3) Two-channel NCS with anticipative controller:
NCS configurations in this section is the control signal Fig.6 can represent a two-channel NCS with a plant with the
construction. state-space (17) in feedback with a anticipative controller
1) One-channel NCS with dynamic feedback controller: with state-space (18). Anticipative controller attempts to
Fig. 5 shows a one-channel NCS consisting of a plant with compensate the sampling and delay introduced by the actu-
the state-space (17) in feedback with a dynamic output ation channels. For simplicity, we assume that the actuation
controller with state-space (18). Note that in one-channel channels are sampled with constant sampling intervals h =
NCS the controller is directly connected to the plant and tk t j , for any consecutive sampling indexes k, j Ki , m p +
only the measurements of the plant are sent through the 1 i m p + mc , and that the delay in the actuation channels
network. Hence in this case yi (t) := y pi (t), m := m p , and is constant and equal to = k , k Ki , m p + 1 i m p +
 h i
x
mc . At each sampling time sk , k Ki , m p + 1 i m p + mc m p and the closed-loop system with state x := xpc can be
the controller sends a time-varying control signal yci () to written as the impulsive system (10) where
the actuator i. This control signal should be used from the    
time sk + at which it arrives until the time sk + h + at A B pCc B D
A := p , B := p c , C := C p 0 . (25)
 
which the next control update of the actuator i will arrive. 0 Ac Bc
This leads to Remark 5: Equation (19) that represents one-channel
NCS with dynamic output feedback is similar to equation
u pi (t) = yci (t), t [sk + , sk + h + ), ki Ki ,
(25) that represents two-cannel NCS with anticipative con-
where m p + 1 i m p + mc . However, the prediction of troller. Consequently for analysis perpose one can model a
control signal yci (t) needed in the interval [sk + , sk + h + ) two-cannel NCS with anticipative controller as a one-channel
must be available at the transmission time sk , which requires NCS with fictitious delays equal to the sum of the delay in
the control unit to calculate the control signal up to h + the sensor to actuator channels, the delay in the actuator to
time units into the future. sensor channels, and the sampling of the actuator channel.
Remark 3: Anticipative controllers send actuation signals
to be used during time intervals of duration h, therefore the B. Stability of NCSs modeled by MIMO sampled-data sys-
sample and hold blocks in Fig. 6 should be understood in a tems
broad sense. In practice, the sample block would send over
Exponential stability of one-channel NCS, two-channel
the network some parametric form of the control signal uci ()
NCS with non-anticipative controller and two-channel NCS
(e.g., the coefficients of a polynomial approximation to this
with anticipative controller can be concluded from Theo-
signal).
rem 2 or 3 (for m = 2 or generalization of these theorems
Remark 4: Anticipative controllers are similar to predic- for arbitrary m) with the choice of (19),(20) and (25) respec-
tive controllers in the sense that both calculate the future tively. By solving the LMIs one can find positive constants
control actions. However in predictive controllers only the i max , 1 i m which determines the upper bound between
most recent control prediction is applied until the new control the sampling time sk and the next update time s j + j of
commands arrive. Anticipative controllers are predictive con- the output i for 1 i m, where k, j are two consecutive
trollers that send a control prediction for a certain duration. sampling instances of output i, for which the stability of the
In the expense of sending more information to the actuators closed-loop system is guaranteed for a given lower and upper
in each packet, one expect fewer packets to be transmitted bound on the total delay in the loop i.
to stabilize the system. However most of the work in the literature has been
devoted to finding MAT I ([13, 14, 10, 11, 4] and references
An estimate xc (t) of xc (t + h + ) is constructed as follows:
therein) which determines the upper bound between the
xc (t) = Ac xc (t) + Bcuc (t), (21) sampling time sk and the next input update time tk+1 =
sk+1 + k+1 for any consecutive sampling instances which the
where
stability of the closed-loop system is guaranteed for a given
uci (t) = y pi (tk ), t [tk + h + ,t j + h + ), (22) lower and upper bound on the total delay in the loop. One
can expect that having m constants i max , 1 i m instead
where k, j are two consecutive sample indexes in Ki , 1 i of one single constant MAT I , reveals more information about
m p . To compensate for the time varying delay and sampling the system as we explore in the following example.
intervals in the actuation channels, xc would have to be Example 2: This example appeared in [14, 10, 5, 12] and
calculated further into the future. Hence the assumptions of considers a linearized model of the form (17) for a two-input,
constant delay and sampling interval for actuation channel two-output batch reactor where
can be relaxed by predicting xc further into the future. With " #
such a controller state prediction available, the signal yci (t) 1.38 0.2077 6.715 5.676  0 0

0.5814 4.29 0 0.675
A p := 1.067 4.273 6.654 5.893 , B p := 5.679 0
1.136 3.146 ,
sent at times sk , k
Ki , m p + 1 i m p + mc , to be used in 0.048 4.273 1.343 2.104 1.136 0
sk + , sk + h + , is then given by  1 0 1 1 
C p := 010 0 .
yci (t) = Cci xc (t h ) + Dcuci , t [sk + , sk + h + ),
(23) This system is controlled by a PI controller of the form (18)
where
which
 only requires the knowledge of xc (.) in the interval    
Ac := 00 00 , Bc := 01 10 ,


t (sk h, sk , and therefore is available at the transmission
Cc := 2 DC := 50 2
 0  
times sk . The closed-loop system can be written as 0 8 , 0 .

Following the assumptions of [14, 10, 5, 12], we assume

x p (t)
      
A B pCc x p (t) B D
= p + p c uc (t), (24) that only the outputs of the plant are transmitted over the
xc (t) 0 Ac xc (t) Bc
network, there is no dropouts and the outputs are sent in a
where x p (t) := x p (t + h + ) and the elements of uc (t) are round-robin fashion and consecutively. We compare MAT I of
defined by (22). Hence in this case yi (t) := y pi (t), m := this example given by the stability results in [14, 10, 5, 12],

Maximum deterministic time interval between samples from [12]
Maximum stochastic arbitrary inter-sampling time distribution from [5]
Maximum stochastic uniform inter-sampling time distribution from [5]
Maximum deterministic time interval between samples from Theorem 2
C OMPARISON OF MAT I FOR
where in all the references the effect of the delay is ignored.
From Theorem 2 we compute 1 max = 0.081, 2 max = 0.113
when there is no delay. We can show that if the upper bound
between any sonsecutive sampling, MAT I , is smaller that
2 min(1 max , 2 max ), then the upper bound between the sam-
1
ples of y p1 or y p2 are smaller than min(1 max , 2 max ) and the
system is stable, so we get MAT I = 21 min(1 max , 2 max ) =
0.0405 Table I shows the less conservative results in the
literature and our MAT I for comparison. Only MAT I for a
(stochastic) uniform inter-sampling time distribution give by
[5] is less conservative than MAT I given by (2). However, for
fair comparison our result should be compared to stochastic
arbitrary inter-sampling time distribution give by [5]. When
the maximum of delay is 0.05 then 1 max = 0.097, 2 max =
0.126.
IV. CONCLUSIONS AND FUTURE WORK
We considered a large class of Networked Control Sys-
tems (NCSs) and we show that the resulting NCSs can
be viewed as abstract MIMO sampled-data systems with
variable sampling intervals and delay. We modeled the linear
delay sampled-data system as a linear infinite-dimensional
impulsive systems and we provided conditions for the stabil-
ity of the closed-loop expressed in terms of LMIs. By solving
these LMIs, one can determine positive constants related to
each variable sent through the network that determines an
upper bound between the sampling time and the next update
time at the destination of that variable.
Although in this paper we focused on the stability prob-
lem, it is possible to derive LMI conditions that lead to H
designs. Another direction for future research is the medium
access scheduling problem and the co-design of feedback
controllers and medium-access protocols.
R EFERENCES
[1] M. Branicky, S. M. Philips, and W. Zhang. Stability of
networked control systems: explicit analysis of delay.
In Proc. of the 2000 Amer. Contr. Conf., pages 2352
2357, June 2000.
[2] E. Fridman. A new Lyapunov technique for robust
control of systems with uncertain non-small delays.
J. of Math. Contr. and Info., 22(3), Nov 2005.
[3] E. Fridman, A. Seuret, and J. P. Richard. Robust
sampled-data stabilization of linear systems: an input
delay approach. Automatica, 40(8):14411446, Aug.
2004.
[4] J. P. Hespanha, P. Naghshtabrizi, and Y. Xu. Networked
control systems: analysis and design. To appear in
no drop
0.0123
0.0279
0.0517
0.0405
TABLE I
E XAMPLE 2 WHEN THERE IS NO DELAY.
Special Issue of the Proc. of the IEEE on The Emerging
Tech. of Networked Contr. Syst.
[5] J. P. Hespanha and A. R. Teel. Stochastic impulsive
systems driven by renewal processes. Proc. of the
Int. Symp. on the Mathematical Theory of Networks
and Syst., pages 606618, July 2006.
[6] P. Naghshtabrizi. Network control systems: analysis
and design. PhD thesis, University of California Santa
Barbara, 2006.
[7] P. Naghshtabrizi and J. Hespanha. Designing observer-
based controller for network control system. In Proc. of
the 44th Conf. on Decision and Contr., volume 4, pages
28762880, June 2005.
[8] P. Naghshtabrizi, J. Hespanha, and A. R. Teel. On
the robust stability and stabilization of sampled-data
systems: A hybrid system approach. To appear in
Proc. of the 45th Conf. on Decision and Contr., Dec.
2006.
[9] P. Naghshtabrizi, J. P. Hespanha, and A. R. Teel.
Stability of infinite-dimensional impulsive systems with
application to network control systems. Submitted to
Proc. of the 2007 Amer. Contr. Conf., June 2007.
[10] D. Nesic and A. R. Teel. Input-output stability prop-
erties of networked control systems. IEEE Trans. on
Automat. Contr., 49(10):16501667, Oct. 2004.
[11] D. Nesic and A. R. Teel. Input-to-state stability of
networked control systems. Automatica, 40(12):2121
2128, Dec 2004.
[12] M. Tabbara, D. Nesic, and A. R. Teel. Input-output
stability of wireless networked control systems. Proc. of
the 44th Conf. on Decision and Contr., pages 209 214,
Dec 2005.
[13] G. C. Walsh, H. Ye, and L. Bushnell. Stability analysis
of networked control systems. In Proc. of the 1999
Amer. Contr. Conf., volume 4, pages 28762880, June
1999.
[14] G. C. Walsh, H. Ye, and L. Bushnell. Stability analy-
sis of networked control systems. IEEE Trans. on
Contr. Syst. Tech., 10(3):438446, May 2002.
[15] D. Yue, Q. L. Han, and J. Lam. Network-based robust
H control of systems with uncertainty. Automatica, 41
(6):640644, 2005.
[16] D. Yue, Q.-L. Han, and C. Peng. State feedback
controller design for networked control systems. IEEE
Trans. on Automat. Contr., 51(11):640644, Nov. 2004.