1680

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 9, SEPTEMBER 2007

Event-Triggered Real-Time Scheduling of Stabilizing Control Tasks

Paulo Tabuada, Senior Member, IEEE

AbstractIn this note, we revisit the problem of scheduling stabilizing control tasks on embedded processors. We start from the paradigm that a real-time scheduler could be regarded as a feedback controller that decides which task is executed at any given instant. This controller has for objective guaranteeing that (control unrelated) software tasks meet their deadlines and that stabilizing control tasks asymptotically stabilize the plant. We investigate a simple event-triggered scheduler based on this feedback paradigm and show how it leads to guaranteed performance thus relaxing the more traditional periodic execution requirements. Index TermsEvent-triggered, input-to-state stability, real-time scheduling.

I. INTRODUCTION Small embedded microprocessors are quickly becoming an essential part of the most diverse applications. A particularly interesting example are physically distributed sensor/actuator networks responsible for collecting and processing information, and to react to this information through actuation. The embedded microprocessors forming the computational core of these networks are required to execute a variety of tasks comprising the relay of information packets in multihop communication schemes, monitoring physical quantities, and computation of feedback control laws. Since we are dealing with resource-limited microprocessors, it becomes important to assess to what extent we can increase the functionality of these embedded devices through novel real-time scheduling algorithms based on event-triggered rather than time-triggered execution of control tasks. We investigate in this note a very simple event-triggered scheduling algorithm that preempts running tasks to execute the control task whenever a certain error becomes large when compared with the state norm. This idea is an adaptation to the scheduling context of several techniques used to study problems of control under communication constraints [6], [11], [18]. We take explicitly into account the response time of the control task and show that the proposed scheduling policy guarantees semi-global asymptotical stability. We also show existence of a minimal time between consecutive executions of the control task under the proposed event-triggered scheduling policy. This time, when regarded as a lower bound on release times of the control task, suggests that periodic implementations of control tasks can be relaxed to more exible aperiodic or event-triggered real-time implementations while guaranteeing desired levels of performance. The proposed approach is illustrated with simulation results. Real-time scheduling of control tasks has received renewed interest from the academic community in the past years [2], [5], [8][10], [14], [22]. Common to all these approaches, that followed from the initial work of Seto and co-authors [23], is the underlying principle that better control performance is achieved by providing more CPU time to control tasks. This can be accomplished in two different ways: letting control
Manuscript received May 30, 2006; revised May 17, 2007. Recommended by Associate Editor J. Hespanha. This research was partially supported by the National Science Foundation EHS Award 0712502. The author is with the Department of Electrical Engineering at the University of California at Los Angeles, Los Angeles, CA 90095 USA (e-mail: tabuada@ee.ucla.edu; URL: http://www.ee.ucla.edu/~ tabuada). Color versions of one or more of the gures in this note are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TAC.2007.904277

tasks run for longer amounts of time using anytime implementations or model predictive controllers; or by scheduling control tasks more frequently. All these approaches assume the existence of a performance criterion for the control task such as a cost function used to design an optimal linear quadratic regulator. Scheduling strategies are then obtained through optimization algorithms seeking to determine schedules maximizing the performance criterion. The work presented in this note does not resort to optimization and does not require a performance criterion. Instead, the decision to execute the control task is determined by a feedback mechanism based on the state of the plant. The above described work focused on optimizing schedules and sampling rates while leaving controllers unchanged. A further generalization of such ideas appears in [20], [21] where the optimization process also changes the feedback controllers in order to enforce stability of the controlled processes and schedulability. Close at the technical level, although addressing very different problems, is the recent work on stabilization under communication constraints [6], [7], [11], [15], [18], [19]. All these approaches are concerned with the stabilization of continuous systems under reduced communication and the employed techniques share with some of the techniques described in this note a common ancestor: the perturbation approach to stability analysis of control systems, described, for example, in [13]. Similar techniques have also been used in [16] and [27] to show how sample-and-hold implementations of stabilizing controllers guarantee stability under sufciently fast periodic executions. Unfortunately, these results are either nonconstructive or rely on assumptions much stronger than those used in this note. Finally, we would like to refer the reader to [1] where some advantages of event-driven control over time-driven control are presented in a stochastic setting. A preliminary version of the results presented in this note was reported in [25]. II. NOTATION AND PROBLEM STATEMENT We shall use the notation jxj to denote the Euclidean norm of an n element x 2 I R . Given matrices A and B , [AjB ] denotes the matrix formed by the columns of matrix A followed by the columns of matrix B . A function f : IRn ! IRm is said to be Lipschitz continuous on n R there exists a constant compacts if for every compact set S  I L > 0 such that jf (x) 0 f (y)j  Ljx 0 yj for every x; y 2 S . A + R0 , a > 0, is said to be of class K continuous function  : [0; a[! I if it is strictly increasing and (0) = 0. It is said to be of class K if a = 1 and (r) ! 1 as r ! 1. A. Notation

B. Problem Statement We consider a control system

n m x _ = f (x; u); x 2 I R ; u2I R

for which a feedback controller

(1)

u = k (x )
has been designed rendering the closed-loop system

(2)

x _ = f (x; k (x + e)):

(3)

Input-to-State Stable (ISS) with respect to measurement errors e 2 n I R . We shall not need the denition1 of ISS in this note but rather the following characterization.
1See,

for example, [24] for an introduction to ISS and related notions.

0018-9286/$25.00 2007 IEEE

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 9, SEPTEMBER 2007

n Denition II.1: A smooth function V ! + 0 is said to be an ISS Lyapunov function for the closed-loop system (3) if there exist class K functions ; ;  and satisfying

1681

: IR

IR

(jxj)  V (x)  (jxj) @V f (x; k(x + e))  0 (jxj) + (jej): @x

(4) (5)

Closed-loop system (3) is said to be ISS with respect to measurement errors e 2 n if there exists an ISS Lyapunov function for (3). The implementation of the feedback law (2) on an embedded processor is typically done by sampling the state at time instants

IR

t0 ; t1 ; t2 ; t3 ; t4 . . .
computing u instants

(t ) = k(x(t )) and updating the actuator values at time

i i

t0 + 1 ; t 1 + 1 ; t 2 + 1 ; t 3 + 1 ; t 4 + 1 ; . . .
where  represents the time required to read the state from the sensors, compute the control law and update the actuators. This means that to a sequence of measurements

1 0

since implies that if the control task is executed at time ti we x ti 0 x ti will have e ti and je ti j thus enforcing > , the control task needs to be executed before the (8). When equality jej  jxj is satised in order to account for the delay between measuring the state and updating the actuators. Although the simple execution rule (9) guarantees global asymptotical stability by construction, there are three important questions that need to be answered in order to assess the feasibility of this scheduling policy. 1) Since the execution times are only implicitly dened, can we guarantee that they will not become arbitrarily close resulting in an accumulation2 point? 2) In the absence of accumulation points, can we compute an estimate of the time elapsed between consecutive executions of the control task? 3) How can we use the execution rule (9) when there are more tasks competing for processor time and still guarantee that no deadlines are missed? Answering the above questions is the objective of Sections IIIVI.

1=0 ( )= ( ) ( )=0 1 0 ( )= ( )

( ( ))=0

III. EXISTENCE OF A LOWER BOUND FOR INTEREXECUTION TIMES We start immediately with one of the main contributions of this note. Theorem III.1: Let x f x; u be a control system and let u k x be a control law rendering the closed-loop system ISS with respect to measurement errors. If the following assumptions are satised: n 2 m ! n is Lipschitz continuous on compacts; 1) f n m 2) k ! is Lipschitz continuous on compacts; 3) there exists an ISS Lyapunov function V for the closed-loop system satisfying (5) with 01 and Lipschitz continuous on compacts, then, for any compact set S  n containing the origin, there exists an " > such that for all response times 2 ; " there exists a time  2 + such that for any initial condition in S the inter-execution times fti+1 0 ti gi2 implicitly dened by the execution rule (9) with  < are lower bounded by  , that is, ti+1 0 ti   for any i 2 . Proof: Let R be the compact set dened by all the points x 2 n satisfying V x   where  > is large enough so that S  R. Such  always exists since by (4), V is proper or radially unbounded. Set R is forward invariant for the closed-loop system since the execution rule (9) guarantees V  . We now dene another compact set E by all the points e 2 n satisfying jej  01  jxj for all x 2 R. Since 01 and are Lipschitz continuous on compacts, then so is 01 jrj = . Let P be the Lipschitz constant for the compact set E so that j01 jrj = 001 jsj = j  P jr 0sj. If r e and s we obtain 01 jej =  P jej. Note that by enforcing the more conservative inequality P jej  jxj we guarantee 01 jej =  jxj (which is (8)) so that it sufces to show that the inter-execution times are bounded for the execution rule P jej jxj. As a rst step towards showing boundedness we note that it follows from Lipschitz continuity on compacts of f x; u and k x that f x; k x e is also Lipschitz continuous on compacts, that is

x (t 0 ); x (t 1 ); x (t 2 ); x (t 3 ); x (t 4 ); . . .
there corresponds a sequence of actuation updates

()

_= ( ) IR

u(t0 + 1); u(t1 + 1); u(t2 + 1); u(t3 + 1); u(t4 + 1); . . . :
Between actuator updates the input value u is held constant according to

: IR : IR

IR IR

t 2 [ti + 1; ti+1 + 1[=) u(t) = u(ti + 1):

(6)

is typically peFurthermore, the sequence of times t0 ; t1 ; t2 ; t3 ; t4 ; riodic meaning that ti+1 0 ti T where T > is the period. We can thus regard the execution of the control task implementing the control law (2) as being time-triggered. In this note we consider instead event-triggered executions where the sequence t0 ; t1 ; t2 ; t3 ; t4 of execution times is no longer periodic neither specied in advance but rather implicitly dened by an execution rule based on the state of the plant. To introduce this execution rule we dene the measurement error e to be

...

...

0 IR 1

IR

1 [0 ]

IN

()

IN IR

t 2 [ti + 1; ti+1 + 1[=) e(t) = x(ti ) 0 x(t):

_ 0 IR

( ( ))

(7)

We can thus describe the evolution of (1) under the implementation (6) of control law (2) by

(( ) ) (( ) ) (( ) ) 0 (( ) ) ( ) ()

x _ = f (x; k(x + e))

with e 2 n as dened in (7). with the single Let us rst consider the hypothetical case purpose of explaining the execution rule. From (5) we see that if we restrict the error to satisfy

= = (( ) )

IR

1=0

( ( + ))

(jej)  (jxj);  > 0

the dynamics of V is bounded by

(8)

jf (r; k(r + s)) 0 f (r0 ; k(r0 + s0 ))j  Lj(r; s) 0 (r0 ; s0 )j

by taking r

= x, s = e and r0 = 0 = e0 we obtain

can be enforced by executing the control task when

@V f (x; k(x + e))  ( 0 1)(jxj) @x thus guaranteeing that V decreases provided that  < 1. Inequality (8) (jej) = (jxj)
(9)

state as well as the error dynamics. We can now bound the interevent
2In the context of hybrid systems this corresponds to another example of the infamous Zeno behavior [4].

jf (x; k(x + e))j  Lj(x; e)j  Ljxj + Ljej (10) when (x; e) 2 R 2 E . Note that R 2 E is forward invariant for the

1682

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 9, SEPTEMBER 2007

times by looking at the dynamics of jej=jxj

d je j dt jxj

je k x j jx k x j jx j _ j je j j e k _ j + jx k x  jekx x j jx k x j jx j _j ej j x = 1 + jjx satisfying j jx j (; 0) =  (14) ej Ljxj + Ljej by (10)  1 + jjx j jx j where (t;  ) is the solution of: j e j j e j _ = jA + BK j + (jA + BK j + jBK j) + jBK j (11) = L 1 + jx j 1 + jx j :  (15) for 1 > 0 and for any desired If we denote jej=jxj by y we have the estimate y _  L(1 + y) and satisfying (0;  ) =  . Furthermore, 0 we conclude that y (t)  (t;  ) where (t;  ) is the solution of  > 0, the execution rule jej =  jxj with _ = L(1 + ) satisfying (0;  ) =  .  1 [A + BK jBK ] ( + 1) Assume now that 1 = 0. Then, the inter-execution times are  0  (01; ) (16) bounded by the time it takes for  to evolve from 0 to 1=P , that 1 0 1 [ A + BK jBK ] ( + 1) R is, the inter-execution times are bounded by the solution  2 I of (; 0) = 1=P . Since (; 0) = L=(L 0 1) we obtain enforces for any i 2 IN and for any t 2 [ti +1; ti i +1[ the following
+1 0 2 2 0 0 2 0 0 0 0 +

d (eT e) = = dt (x T x ) = 0 = T T = (x T x )0 = x T x T _ (e T e ) = = (e e) e e_ (x x) x0 Tx T _ je j _ 0 xT x =0 e x
1 2 1 2 1 2 1 2 1 2

1 2

: IRn ! IR satisfying a jx j  V ( x )  a jx j (12) @V (Ax + BKx + BKe)  0 ajxj + bjekxj (13) @x with a; a; a; b 2 I R . Corollary IV.1: Let x _ = Ax + Bu be a linear control system, let u = Kx be a linear control law rendering the closed-loop system globally asymptotically stable and assume that 1 = 0. For any initial condition in I Rn the interevent times fti 0ti gi2IN implicitly dened by the execution rule jej =  jxj are lower bounded by the time 
thus have a Lyapunov function V
2 + 0 2 2 +

= 1=(L + LP ). inequality: For 1 > 0 we need a more detailed analysis. First, pick 0 satisfying 0 0 0  <  < 1 (take for example  =  + (1 0 )=2) and let P be je ( t) j   jx ( t) j R satisfy the Lipschitz constant for 0 ( (jej)= 0 ). Let now " 2 I 0 (" ; 1=P ) = 1=P . Such " always exists since  is continuous, with interexecution times bounded by 1 +  where time  satises _ > 0 and 1=P < 1=P 0 . Then, by executing the control task at time  ti , dened by P jej = jxj, we guarantee that for t 2 [ti ; ti + " [ we 1 [A + BK jBK ] ( + 1) have jej  jxj=P 0 and thus also (jej)   0 (jxj). Since  0 < 1  ; = 0 : (17) asymptotic stability is still guaranteed. The interexecution times are 1 0 1 [ A + BK j BK ] (  + 1) now bounded by 1+  where  is the time it takes for  to evolve from je(ti +1)j=jx(ti +1)j = jx(ti ) 0 x(ti +1)j=jx(ti +1)j to 1=P . We Proof: Estimate (15) follows directly from (11) by using jx _j  thus need to pick 1 small enough so that je(ti + 1)j=jx(ti + 1)j < _ > 0. It now follows from continuity3 of jx(ti ) 0 x(ti + jA + BK kxj + jBK kej. Inequality 0  (01; ) guarantees that 1=P since  1)j=jx(ti + 1)j with respect to 1 the existence of " > 0 such that if jej = 0 jxj holds at time ti , triggering an execution of the control for any 0  1  " we have jx(ti ) 0 x(ti +1)j=jx(ti +1)j < 1=P . task, then the inequality jej   jxj will hold for t 2 [ti ; ti + 1[. To The proof is now nished by taking " = minf" ; " g. conclude boundedness of interexecution times we need to ensure that Theorem 3.1 shows that the simple execution rule (9) results in a se- je(ti + 1)j=jx(ti + 1)j   0 since if this inequality fails to hold, an quence of interexecution times for the control task that is guaranteed to execution of the control task is requested before the termination of the be lower bounded provided that 1 is sufciently small. The techniques previous execution. Let e0 denote the error dened by e0 (t) = x(t ) 0 i used in the proof rely on Lipschitz continuity and are necessarily con- x(t) for t 2 [t ; t + 1] and note that je(t + 1)j=jx(t + 1)j is by i i i i servative for general nonlinear systems. However, for linear systems denition equal to je0 (ti +1)j=jx(ti +1)j. To estimate je0 (t)j=jx(t)j they provide reasonable estimates and we can even provide computable we compute d=dtje0 (t)j=jx(t)j following the same argument as in the bounds for 1 as discussed in Section IV. proof of Theorem 3.1 to obtain:

+ 1 1 + 1 1 1 2 2 1 2

IV. THE LINEAR CASE In this section, we discuss the linear case in some detail since the arguments in the proof of Theorem can be rened in order to provide a computational estimate for  . We thus assume the control system to be of the form with A and B matrices of appropriate dimensions. We also assume Kx rendering the closed-loop the existence of a linear feedback u system globally asymptotically stable and where K is a matrix of appropriate dimensions. Note that in the linear case any such K renders the closed-loop system ISS with respect to measurement errors. We

by

0j ej 1 + jjx (18)  L 1 + jje xj j with L = j[A + BK jBK ]j. The evolution of je0 j=jxj is thus bounded

d je 0 j dt jxj

x _ = Ax + Bu; x 2 IRn ; u 2 IRm

(t) which is the solution satisfying (ti ) = 0 of: _ = L(1 + ) _ = L(1 + )(1 + ) 
2

since, as shown in the proof of Theorem (3.1),  t . Since

()

jej=jxj is bounded by

( t) =

3Note that x t is never zero since the closed-loop system converges asymptotically to zero and thus never reaches zero in nite time.

( + 1)

e(t )j tL jjx +1 (t )j e(t )j 1 0 tL jjx +1 (t )j 0 tL( + 1)  tL( + 1) = 10 tL(0 + 1) 1 0 tL( + 1)

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 9, SEPTEMBER 2007

1683

where the inequality follows from @

Assumption (16) thus guarantees that je ti The interexecution times are now bounded by time necessary for  to evolve from L  = 0 , that is, the solution  of  ; L 
0

(14). Although the formulas in Corollary 4.1 only provide estimates, they are sufciently accurate to be useful in practical situations as described in Section V. V. AN ACADEMIC EXAMPLE

( + 1)j=jx(ti + 1)j <  . 1 +  where  is the 1 ( + 1)=(1 0 1L( + 1)) to ( 1 ( +1) (1 0 1L( +1))) =  . Note that for 1 = 0 we can chose  =  in (16) and (17) becomes
0 0

=@ > 0, we have 1L( + 1) : (1) < 1 0 1L( + 1)

0

+ 0 (19) 1 u stabilized by the linear feedback u = x1 0 4x2 . Using V = xT P x as a Lyapunov function, we obtain @V =@x(Ax + BKx) = 0xT Qx
x1 x2
with P and Q dened by:

We now illustrate the previous results on the linear control system

x _1 x _2

1 = 00 2 3

1 4 = 1 1 4 1

; Q=

1 2 1 4

1 4 3 2

(20)

Taking the measurement error into account we have

@V Ax + BKx + BKe @x
where

 0ajxj2 + bjekxj

a = m (Q) > 0:44 b = jK T B T P + P BK j = 8

and m Q is the smallest eigenvalue of Q. Since b needs to be smaller that 0.44, we select  : and obtain the execution rule jej : jx0j. For : s and according to Corollary 4.1 we can select any  satisfying :  0  : 0 . The theoretical value for the interexecution time corresponding to  : is

( ) = 0 05

= 0 04 1 +  = 0:005 s + 0:0091 s = 0:0141 s

1 = 0 005 0 0235

= 0 05

0 0405

0:04.

Fig. 1. Evolution of jej, 

xj and  jxj for 1 = 0:005, 

= 0:05 and  =

while simulation results for the initial conditions

In Fig. 1 we can see how the error norm never reaches  jxj even though it goes beyond  0 jxj which is used as execution rule. Although for small values of time there is a large gap between the maximum value reached by jej and  jxj, this gap decreases as the state approaches the origin. For different values of  0 we obtained the interexecution time estimates reported in Table I. We can see that the estimated values, although conservative, do not overestimate the values obtained through simulation by more than a factor of 3. VI. COSCHEDULABILITY OF STABILIZING CONTROL TASKS The lower bound for interexecution times of the control task, presented in Theorem 3.1, can be used to analyze schedulability of a set fTi gi2I . We shall assume a preemptive scheduler in of tasks T which the control task has the highest priority and thus cannot be pre-

 2 (x1 (0); x2 (0)) = 10cos 2 30  ; 10sin 30   = 1; . . . ; 30 provided a lower bound of 0:0237s.

can be captured in the proposed framework by suitably enlarging . We regard execution of the control task as a timing overhead imposed on the tasks Ti by the scheduler in the sense that Ti may need to be interrupted to execute the control task. When a set of software tasks T can be scheduled despite the overhead associated with the control task we say that T is coschedulable with the control task. Coschedulability is now ensured by the following sufcient condition where dr e denotes the smallest integer greater than r 2 . f x; u be a control system, let u Proposition VI.1: Let x k x be a control law rendering the closed-loop system ISS with respect fTi gi2I be a set of tasks with to measurement errors and let T execution times f i gi2I . If the set of tasks T is schedulable with new execution times 0 i given by

empted by any other task and is executed without delays when (jej) = (jxj). Note that timing overheads associated with context switching

()

1 1

IR _= ( ) =

i 1 i = 1i + 1 1 
0

then T is coschedulable with the control task under the execution rule (9).

1684

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 9, SEPTEMBER 2007

TABLE I SIMULATED AND ESTIMATED INTER-EXECUTION TIMES FOR SEVERAL VALUES

OF

be scheduled is not processor time but rather the transmission medium. Similar ideas have been more or less explicitly explored in [17], [19], [26]. ACKNOWLEDGMENT The author would like to thank his former colleague Mike Lemmon at the University of Notre Dame for insightful discussions about some of the issues addressed in this note.

REFERENCES
Proof: Since the time between executions of the control task  , each task Ti can be interrupted at is lower bounded by most d i = e times. Each interruption will delay the execution of Ti by units of time resulting in a combined execution time of d i = e . i Coschedulability is ensured by the existence of a schedule where the duration i of each task Ti has been inated to 0 i in order to accommodate the timing overhead imposed by the control task. This result is a very simple illustration of how the interexecution time  can be used for schedulability analysis of the proposed event-triggered policy. Much less conservative results can be obtained when more structure is assumed on the set of tasks T (periodicity for example) and on the scheduler.

1 1 1+1 1 1

1+

VII. DISCUSSION A. Real-Time Scheduling Policies for Nonpreemptible Tasks The simple preemptive scheduling strategy presented in Section VI relied on the possibility to preempt all but the control task. If there are other non-premptible tasks and its execution times are small compared with , then the same scheduling strategy can still be used by suitably enlarging to account for execution delays. In many, if not most, situations this assumption may not hold and more elaborate scheduling strategies are necessary. The results presented in this note are also relevant in this more general context since the lower bound on the interexecution times can be used to construct a timed-automaton model for the control task. This model can then be composed with models of the remaining tasks and a scheduler, regarded as a supervisor, can be synthesized by resorting to control and/or game theoretic techniques for timed automata [3], [12].

B. ISS With Respect to Actuation Errors and Networked Control Systems The results presented in this note are based on the assumption that the controller to be implemented renders the closed-loop system ISS with respect to measurement errors. From a theoretical point of view it is not difcult to see that Theorem 3.1 still holds, mutatis mutantis, if ISS with respect to actuation errors is assumed in place of ISS with respect to measurement errors. This is an assumption that can be made without loss of generality for control afne systems. In this case, the k x ti 0 k x t execution rule would be based on the error e t which makes this approach uninteresting from the practical point of view as k x t needs to be computed to determine if the control task computing k x t should be executed! Under the ISS with respect to measurement errors assumption used in this note we only need to compute jx ti 0 x t j and jx t j to determine if the control task should be executed and this can be done in hardware without requiring processor time. However, the ISS with respect to actuation errors, does make sense in a networked control setting where the shared resource to

( ) = ( ( ))

( ( ))

( ( )) ( ( )) ( ) ()

()

[1] K. J. Astrom and B. Bernhardsson, Comparison of riemann and lebesgue sampling for rst order stochastic systems, in Proc. 41st IEEE Conf. Decision and Control, Dec. 2002. [2] P. Albertos, A. Crespo, I. Ripoll, M. Valles, and P. Balbastre, RT control scheduling to reduce control performance degrading, in Proc. 39th IEEE Conf. Decision and Control, Sydney, Australia, 2000. [3] K. Altisen, G. Gssler, and J. Sifakis, Scheduler modeling based on the controller synthesis paradigm, Real-Time Systems: The International Journal of Time-Critical Computing Systems, vol. 23, no. 1, pp. 5584, 2002. [4] A. D. Ames, P. Tabuada, and S. Sastry, J. P. Hespanha and A. Tiwari, Eds., On the stability of zeno equilibria, in Proc. Hybrid Syst.: Comput. Control 2006, Santa Barbara, CA, USA, 2006, vol. 3927. [5] G. Buttazzo and L. Abeni, Adaptive workload management through elastic scheduling, Real-Time Syst., vol. 23, no. 1, pp. 724, 2002. [6] R. Brockett and D. Liberzon, Quantized feedback stabilization of linear systems, IEEE Trans. Automat. Control, vol. 45, pp. 12791289, 2000. [7] M. Branicky, S. Phillips, and W. Zhang, Scheduling and feedback co-design for networked control systems, in Proc. 41st IEEE Conf. Decision and Control, Las Vegas, NV, Dec. 2002. [8] A. Cervin and J. Eker, Feedback scheduling of control tasks, in Proc. 39th IEEE Conf. Decision and Control, Sydney, Australia, 2000. [9] A. Cervin, J. Eker, B. Bernhardsson, and K.-E. Arzen, Feedback-feedforward scheduling of control tasks, Real-Time Syst., vol. 23, no. 1, pp. 2553, 2002. [10] R. Chandra, X. Liu, and L. Sha, On the scheduling of exible and reliable real-time control systems, Real-Time Syst., vol. 24, no. 2, pp. 153169, 2003. [11] N. Elia and S. Mitter, Stabilization of linear systems with limited information, IEEE Trans. Autom. Control, vol. 46, pp. 13841400, 2001. [12] E. Fersman, L. Mokrushin, P. Pettersson, and W. Yi, Schedulability analysis of xed priority systems using timed automata, Theor. Comput. Sci., vol. 354, no. 2, pp. 301317, 2006. [13] H. K. Khalil, Nonlinear Systems, 2nd ed. Upper Saddle River, NJ: Prentice Hall, 1996. [14] D. Liu, S. Hu, M. Lemmon, and L. Qiang, Firm real-time system scheduling based on a novel qos constraint, IEEE Trans. Computers, vol. 55, pp. 320333, 2006. [15] D. Liberzon, Hybrid feedback stabilization of systems with quantized signals, Automatica, vol. 39, no. 9, pp. 15431554, 2003. [16] D. Laila, D. Nesic, and A. Teel, Open and closed loop dissipation inequalities under sampling and controller emulation, Eur. J. Contr., vol. 18, pp. 109125, 2002. [17] L. Montestruque and P. Antsaklis, Stability of model-based networked control systems with time-varying transmission times, IEEE Trans. Autom. Control, vol. 49, pp. 15621572, 2004. [18] G. N. Nair and R. J. Evans, Stabilization with data-rate-limited feedback: Tightest attainable bounds, Syst. Contr. Lett., vol. 41, no. 1, pp. 4956, 2000. [19] D. Nesic and A. Teel, Input to state stability of networked control systems, Automatica, vol. 40, pp. 21212128, 2004. [20] L. Palopoli, C. Pinello, A. Bicchi, and A. Sangiovanni-Vincentelli, Maximizing the stability radius of a set of systems under real-time scheduling constraints, IEEE Trans. Autom. Control, vol. 50, pp. 17901795, 2005. [21] L. Palopoli, C. Pinello, A. S. Vincentelli, L. Elghaoui, and A. Bicchi, C. Tomlin and M. R. Greenstreet, Eds., Synthesis of robust control systems under resource constraints, in Proc. Hybrid Systems: Computation and Control, 2002, pp. 337350. [22] L. Sha, X. Liu, M. Caccamo, and G. Buttazzo, Online control optimization using load driven scheduling, in Proc. 39th IEEE Conf. Decision and Control, Sydney, Australia, 2000.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 9, SEPTEMBER 2007

1685

[23] D. Seto, J. P. Lehoczky, L. Sha, and K. G. Shin, On task schedulability in real-time control systems, in Proc. IEEE Real-Time Technol. Appl. Symp. (RTAS), 1996, pp. 1321. [24] E. Sontag, Input to state stability: Basic concepts and results, in Proc. CIME Summer Course on Nonlinear and Optimal Control Theory, 2004, pp. 462488. [25] P. Tabuada and X. Wang, Preliminary results on state-triggered scheduling of stabilizing control tasks, in Proc. 45th IEEE Conf. Decision and Control, San Diego, CA, 2006. [26] J. Yook, D. Tilbury, and N. Soparkar, Trading computation for bandwidth: Reducing communication in distributed control systems using state estimators, IEEE Trans. Contr. Syst. Technol., vol. 10, pp. 503518, Jul. 2002. [27] Y. Zheng, D. H. Owens, and S. A. Billings, Fast sampling and stability of nonlinear sampled-data systems: Part 2. sampling rate estimations, IMA J. Math. Contr. Inf., vol. 7, no. 1, pp. 1333, 1990.

was discussed recently [11], [12]. But the controllability of positive systems is different from that of ordinary systems with positive controls because both of inputs and states are restricted to be nonnegative in positive systems. This note is constructed as follows. In Section II, some denitions are given for preliminaries. In Section III, a necessary and sufcient condition is given for a continuous-time linear system to be positive controllable based on the Jordan canonical form and a method is proposed to test the positive controllability. In Section IV, an example is presented to illustrate the proposed method. Finally in Section V, concluding remarks are stated. II. PRELIMINARIES Consider a multiple input continuous-time linear system described by
S:x _ (t) = Ax (t) + Bu (t)

(1)

Positive Controllability Test for Continuous-Time Linear Systems

Hiroshi Yoshida, Member, IEEE, and Tetsuro Tanaka, Member, IEEE
AbstractThis note presents a method to test the positive controllability of a continuous-time linear system. Based on the Jordan canonical form, the controllability of a given system can be reduced to those of its subsystems with real eigenvalues. Because the dimension of the subsystem is smaller than that of the given system, the controllability test can be simplied. It is pointed out that the conditions for a continuous-time system to be positive controllable have almost the same expressions as those for a discrete-time system. Index TermsContinuous-time linear system, controllability test, Jordan canonical form, positive control.

where x(t) 2 n , u (t) 2 m , A 2 n2n , B 2 n2m . The control input that satises u(k) 2 m i=1 [ 0; 1) is called a positive control. Denition 1: Let x 0 and x f be any initial and nal states, respectively. Then system S is called positive controllable if there exist a nite time  and a positive control u(t)  0; 0  t   , which will bring the system from x(0) = x0 to x( ) = xf . Denition 2: If xi > 0 or xi  0 for all i = 1; 2; . . . ; n, then T [x1 ; x2 ; . . . ; xn ] with T x > 0 or x  0, respectively, where x n 2 n denoting the transposition. Further we use I n 2 as the identity matrix and the following notations:
C (A; B ) Em e m1 B ; AB ; . . . ; A
[1; 1; . . . ; 1]

T [1; 0; 0; . . . ; 0] m; . . . :

n01 B

m; 0

n T n [0; 0; . . . ; 0] m; e T m2 [0; 1; 0; . . . ; 0]

2 n2nm

I. INTRODUCTION The concept of controllability of systems was introduced by Kalman [1], which is a signicant and fundamental subject in control system theory. As for the controllability under positive input, necessary and sufcient conditions for continuous-time linear systems were obtained which arise frequently in the practical problems, such as antivibration control of pendulums system [2], optimal control of economic system [3], and others [4], [5]. It is very important to test the controllability in designing systems. But unfortunately the subject of testing the positive controllability of continuous-time linear systems has not been investigated completely as yet. On the other hand, the controllability of discrete-time linear systems with input constraints was also discussed in [6][10]. A necessary and sufcient condition for single input discrete-time linear systems with positive input was obtained by Evans et al. [7]. The authors presented a necessary and sufcient condition for multiple input discrete-time linear systems to be positive controllable, where a method to test the positive controllability was proposed [10]. The purpose of this note is to present a method to test the positive controllability of continuous-time linear systems based on the results of discrete-time linear systems [10]. By the way, the controllability of positive discrete-time linear systems
Manuscript received October 25, 2005; revised July 19, 2006 and May 7, 2007. Recommended by Associate Editor A. Hansson. The authors are with the Department of Electrical and Electronics Engineering, Kagoshima University, Kagoshima 890, Japan (e-mail: tetsu@eee.kagoshima-u.ac.jp). Digital Object Identier 10.1109/TAC.2007.904278

Denition 3: Let fk1 ; k2 ; . . . ; km g be one of the rearrangements m2m is [e mk ; e mk ; . . . ; emk ] 2 of f1; 2; . . . ; mg. Then W called a permutation matrix. Denition 4: By a nonsingular real transformation of the state vari01x(t), system S described by (1) is transformed into ~ ( t) = V able, x
~:x ~x ~ u ( t) ; x _ ( t) = A S ~ ~ ( t) + B ~ ( t)

(2)

where
~ A

01AV ; B ~
~ 11 A ~ 21 A

01B ; V
0

2 n2n ; det(V ) 6= 0:

(3)

If the following holds:

~ A ~ B ~ 11 A

V V

01AV 01B
n ~ n ~

~1 B ~2 B

~ 22 A

;V

01x(t)
~ 22 ;A

x ~ ( t)

x ~ 1 ( t)

2 ~2 2 B

2n ~ ~ 2n ~ ~ 21 2 n ;A 2m ; n = n ~1 + n ~2

n ~

2n ~

x ~ 2 ( t)
~1 ;B

2
n ~

n ~

2m
(4)

~1 described by then system S ~ 11 x ~ 1 u ( t) ; x ~1 : x _ 1 ( t) = A ~ ~ 1 ( t) + B ~ 1 ( t) S

(5)

is called a subsystem of S . ~ in (4), V can be calcuNote that in obtaining the triangular form A lated based on an A -invariant subspace [13].

0018-9286/$25.00 2007 IEEE