2012 American Control Conference Fairmont Queen Elizabeth, Montral, Canada June 27-June 29, 2012

Zero-controllability and dead-beat control of discrete-time behaviors

Mauro Bisiacco and Maria Elena Valcher

Abstract In this paper the concepts of controllability and zero-controllability of a (to be controlled) variable w through a (control) variable c are introduced and characterized. By assuming this perspective, the dead-beat control (DBC) problem is posed as the problem of designing a controller such that, for the resulting controlled behavior, the to be controlled variable w goes to zero in a nite number of steps in every trajectory. Zero-controllability of w through c turns out to be a necessary and sufcient condition for the existence of admissible DBCs.

I. I NTRODUCTION As nicely described in some recent papers dealing with control in the behavioral setting [9], [15], the traditional perspective to control problems has deeply intertwined the idea of control itself with the concepts of input, output and feedback. This perspective, however, fails to provide the appropriate framework where to cast the controller design problem in a number of interesting cases, as, for instance, the design of passive car suspensions and of insulation equipments for noise abatement, or ships stabilization [14], [15]. The behavioral approach provides the natural framework where control problems can be addressed in the utmost generality and without any a priori assumption regarding input/output partition and feedback connection [13], [14]. Indeed, in this setting, the control target is that of restricting the behavior trajectories to a subset of good ones and this goal is achieved by interconnection, namely by constraining either all or a subset of the system variables to obey an additional family of laws, representing the controller laws. Starting from the contributions [13], [14], there has been quite a number of papers on the control within the behavioral framework (for instance, [1], [4], [6], [9]). In these papers two fundamental control set-ups have been explored: the full interconnection case and the partial interconnection case. In the former, it is assumed that all variables are the target of the control problem (for instance, the stabilization problem) and at the same time they are all available for interconnection. In the latter, the system variables are partitioned in two (or possibly more) groups, by distinguishing between to be controlled variables, typically denoted by w, which are the object of the control specications, and control variables, denoted by c, which are the means through which the control target is achieved. Indeed, the controller achieves the desired result by restricting the behavior of the variables in c which are the only ones available for interconnection.
Mauro Bisiacco and Maria Elena Valcher are with the Dipartimento di Ingegneria dellInformazione, Universit` a di Padova, via Gradenigo 6/B, 35131 Padova, Italy, {bisiacco,meme}@dei.unipd.it.

In investigating this problem, we take a perspective that differs both from the full interconnection and from the partial interconnection set-ups, in that we assume that only the to be controlled variable w must be driven to zero in a nite number of steps, but the control laws restrict the evolutions of both w and c. So, there is full interconnection for control purposes, but the control target is just the variable w. This viewpoint seems to us closer to the classical idea underlying the DBC design for state-space models, since the target is that of driving to zero (only) the state variable, but the control law involves both the state and the input. The investigation of the DBC problem under this perspective requires to introduce new concepts of controllability and of zero-controllability of the variable w through the variable c. Consistently, it will turn out that the possibility of designing DBCs with good properties (admissibility) that drive to zero in a nite number of steps all the trajectories w, by constraining both w and c, is just equivalent to the zero-controllability of w through c. The paper is organized as follows: at the end of this section and in section 2 background material about polynomial matrices and behaviors dened on Z+ is recalled. In section 3, controllability and zero-controllability notions referring to the whole system variable are recalled and introduced, respectively, and they are later characterized. Section 4 deals with the denitions and characterizations of controllability and zero-controllability of w through c. Dead-beat controllers are the focus of section 5, where some preliminary results are given. In section 6 admissible DBCs are presented and it is shown that such controllers are available if and only if w is zero-controllable through c. Notation. A polynomial matrix R(z ) R[z ]pq is right monomic [5] if rank R() = q for every C \ {0}. This means that R(z ) is of full column rank and the GCD of its maximal (i.e., q th) order minors is a monomial cz h , c R \ {0}, h Z+ . R(z ) is right monomic if and only if it admits a Laurent polynomial left inverse or, equivalently, the diophantine equation L(z )R(z ) = z N Iq , in the unknown polynomial matrix L(z ), is solvable for some integer N 0. R(z ) R[z ]pq is right prime if rank R() = q for every C. Right prime matrices are special cases of right monomic matrices. Actually, right primeness characterizations can be obtained by simply replacing in the previous equivalent conditions the word monomial with unit and the integer N by zero. Every full column rank matrix R(z ) (z )(z ), where R (z ) is can be expressed as R(z ) = R right prime and (z ) is nonsingular square. When so, (z ) is called a greatest right divisor of R(z ). Left monomic

978-1-4577-1094-0/12/$26.00 2012 AACC

6685

matrices, left prime matrices and greatest left divisors are similarly dened and characterized. Every polynomial matrix (z )(z )R r (z ), R(z ) factorizes over R[z ] as R(z ) = R where R (z ) is right prime, (z ) nonsingular square and r (z ) is left prime. R If R(z ) is a p q polynomial matrix of rank r, a polynomial matrix M (z ) is a left annihilator of R(z ) if (z ) of R(z ) is a M (z )R(z ) = 0. A left annihilator M minimal left annihilator (MLA, for short) if it is of full row rank and for any other left annihilator M (z ) of R(z ) we (z ) for some polynomial matrix P (z ). have M (z ) = P (z )M It can be easily proved that, unless R(z ) is of full row rank, an MLA always exists, it is a (p r) p left prime matrix and is uniquely determined modulo a unimodular left factor. In the following, w will denote the size of the vector w and c the size of c. II. BASIC RESULTS ABOUT BEHAVIORS WITH TRAJECTORIES IN (Rw )Z+ In this paper, all trajectories will be assumed dened on the time set Z+ of nonnegative integers. The left (backward) shift operator on (Rv )Z+ , the set of trajectories dened on Z+ and taking values in Rv , is dened as : (Rv )Z+ (Rv )Z+ : (v0 , v1 , v2 , ) (v1 , v2 , v3 , ).
i p q If R(z ) = is a polynomial matrix, i=0 Ri z R[z ] we associate with it the polynomial matrix operator R( ) = L i i=0 Ri . Results about polynomial matrix operators acting on (Rq )Z+ can be found in [12], where these results have been derived with (and compared to) those about the more common set-up of polynomial matrix operators acting on (Rq )Z . The interested reader can refer to [3] for more details. It can be proved that R( ) describes an injective map from (Rq )Z+ to (Rp )Z+ if and only if R(z ) is a right prime matrix, and a surjective map if and only if R(z ) is of full row rank. L

deduce that every autonomous behavior can be expressed as the kernel of a nonsingular square polynomial matrix. In general, an autonomous behavior B = ker R( ) includes nite support trajectories if and only if rank R(0) < w. Autonomous behaviors for which there exists some N s.t. all their trajectories have (nite) supports included in [0, 1] are called nilpotent (autonomous) and they are kernels of polynomial matrix operators R( ) corresponding to right monomic matrices [10]. In particular, if R(z ) is nonsingular square, ker R( ) is nilpotent if and only if det R(z ) = c z , for some c R \ {0} and some Z+ . If an autonomous behavior is not nilpotent, it includes at least one innite support trajectory. III. C ONTROLLABILITY AND ZERO - CONTROLLABILITY We here recall the denition of controllability of a behavior and introduce that of zero-controllability. The former has been investigated quite in detail. In particular, in [16] several denitions of controllability have been introduced and compared, thus showing that some of them are independent of the choice of the independent variable set (namely they hold both for Z+ and for Z, and they extend to all the possible multidimensional cases), while others are not. Denition 1: [16] A behavior B = ker R( ) (Rw )Z+ is said to be controllable if there exists some nonnegative integer L s.t. for every N N and every pair of trajectories B s.t. w |[0,N 1] = w1 , w2 B, there exists w |[N +L,+) = w2 |[0,+) . w1 |[0,N 1] and w Controllable behaviors are endowed with very strong properties [16]. Among them, in this paper, we will mainly refer to the fact that every controllable behavior can be described ( ), for some left prime polynomial matrix as B = ker R (z ) R[z ]pw . R We now dene zero-controllability of a behavior. Denition 2: A behavior B = ker R( ) (Rw )Z+ is said to be zero-controllable if there exists some nonnegative integer L s.t. for every N N and every trajectory w B s.t. w |[0,N 1] = w|[0,N 1] and B, there exists w |[N +L,+) = 0. w The following proposition provides a characterization of zero-controllability. Due to the page constraint we omit here the proof and refer the reader to [2]. Proposition 1: Given a behavior B = ker (R( )) (Rw )Z+ , with R(z ) R[z ]pw , the following facts are equivalent ones: i) B is zero-controllable; (z ), for some right monomic L(z ) and ii) R(z ) = L(z )R (z ). some left prime matrix R IV. C ONTROLLABILITY AND ZERO - CONTROLLABILITY
OF A VARIABLE THROUGH ANOTHER

In this paper, by a behavior B (Rw )Z+ we mean the linear and left shift invariant set of solutions w = {w(t)}tZ+ of a system of difference equations R0 w(t)+ R1 w(t +1)+ + RL w(t + L) = 0, t Z+ , (1) with Ri R
p w

. This system is equivalently described as R( )w = 0,

L i=0 i pw

(2)

Ri z belongs to R[z ] , and this leads where R(z ) := to the short-hand notation B = ker R( ). It has been shown in [12] that ker R1 ( ) ker R2 ( ) if and only if R2 (z ) = P (z )R1 (z ) for some polynomial matrix P (z ), and that every behavior B can be described as the kernel of a full row rank polynomial matrix. A behavior B is said to be autonomous if there exists Z+ such that (s.t.) if w1 , w2 B and w1 (t) = w2 (t) for t [0, ], then w1 = w2 . B = ker R( ) (Rw )Z+ , with R(z ) R[z ]pw , is autonomous if and only if R(z ) is of full column rank w [10], [12]. From the previous comment about full row rank kernel representations, we

In the behavioral approach to system modeling [7], [13] there is a traditional lack of symmetry between the concept of controllability and that of observability. Indeed, controllability is dened in terms of the (trajectories of the) whole

6686

system variable w, while observability is dened by referring to a partition of the system variables. Indeed, we talk about observability (or reconstructibility) of the variable w2 from some other variable w1 [3], [11], where often w1 represents the latent variable of the system [13]. In this section we propose denitions of controllability and of zero-controllability that refer to a partition of the system variables in two subsets: a variable that is the target of the control action and a variable that represent the means of the control action. Clearly, this requires to introduce such a partition in the original kernel description of the behavior. So, we consider a behavior described for every t Z+ by the following difference equation: Rw ( )w(t) = Rc ( )c(t), t Z+ , (3)

full row rank. When so, it is well known (see [7]), that if Mc (z ) represents a minimal left annihilator of Rc (z ), then B = Pw Bf ull = ker(Mc ( )Rw ( )). Consequently, we have the following results. Proposition 2: Given a behavior Bf ull , described as in (3), the following facts are equivalent: i) the variable w is controllable through the variable c; ii) the behavior B is controllable; iii) either Rc is of full row rank or Mc Rw factorizes as where the left factor L is right prime, Mc Rw = LR, is left prime. while the right factor R Proof: The equivalence of i) and ii) follows trivially from the denition: the control variable c is just a means, but what we are trying to do is to stir any trajectory w of B into any other trajectory w B, at any point after a sufciently large number of steps. The equivalence of ii) and iii) follows from the characterization of controllability of a behavior derived in the previous section (by treating separately the case when B = (Rw )Z+ , namely Rc is of full row rank). A similar result holds for the zero-controllability of w through c. Proposition 3: Given a behavior Bf ull , described as in (3), the following facts are equivalent: i) the variable w is zero-controllable through the variable c; ii) the behavior B is zero-controllable; iii) either Rc is of full row rank or the matrix Mc Rw where the left factor L is factorizes as Mc Rw = LR, is left prime. right monomic, while the right factor R Proof: The equivalence of i) and ii) follows trivially from the denition. The equivalence of ii) and iii) follows from the characterization of zero-controllability of a behavior derived in the previous section (again, by treating separately the case when B = (Rw )Z+ ). As it is easily seen, (zero-)controllability through c is a weaker property with respect to (zero-)controllability tout court. This will allow us to solve control problems even when standard (both fully and partially interconnected) controllers do not work, as it will be claried in the sequel. E XAMPLE 1 Consider the behavior Bf ull described as in (3), with Rw (z ) = Rc (z ) = z 1. Clearly, Bf ull = ker [ Rw ( ) Rc ( ) ] is neither controllable nor zerocontrollable, as the full row rank matrix [ Rw (z ) Rc (z ) ] is not left monomic (and hence not even left prime), however it is easily seen that since Rc (z ) is of full row rank, w is both controllable and zero-controllable through c. V. D EAD - BEAT CONTROLLERS By a controller of a given behavior (3) we mean a system (a set of difference equations) that constrains the trajectories of both the control variable c and of the to-be-controlled

where w is the to be controlled variable, while c is the control variable. We can think of (3) as a latent variable representation, where w is the manifest variable, while c is the latent variable. According to the standard notation for control problems in the behavioral setting (see, for instance, [1], [4], [6]), we dene the full system behavior as Bf ull := {(w, c) (Rw )Z+ (Rc )Z+ satisfying (3)}, (4) and the manifest behavior as B := {w (Rw )Z+ : (w, c) (Rw )Z+ (Rc )Z+ satisfying (3)}. (5) The matrices Rw (z ) and Rc (z ) involved in the kernel description are polynomial matrices of sizes p w and p c, respectively. We now dene the controllability (zerocontrollability) of the variable w through the variable c. Denition 3: Given a behavior Bf ull described as in (3), we say that the variable w is controllable through the variable c if there exists some nonnegative integer L s.t. for every N N, every (w, c) Bf ull , and every w B, ,c ) Bf ull s.t. one can nd (w |[0,N 1] = w|[0,N 1] , and w |[N +L,+) = w |[0,+) . w Similarly, we introduce the zero-controllability of w through c as a special case of the previous property, namely by assuming as w the zero trajectory. Denition 4: Given a behavior Bf ull described as in (3), we say that the variable w is zero-controllable through the variable c if there exists some nonnegative integer L s.t. for every N N and every (w, c) Bf ull , one can nd ,c ) Bf ull s.t. (w |[0,N 1] = w|[0,N 1] , w and |[N +L,+) = 0. w

The characterization of both properties can be obtained in terms of B, the projection of the full behavior Bf ull on its manifest variables w (and hence denoted by Pw Bf ull ). To this end, it is worthwhile noticing that if Rc (z ) is of full row rank, then B = (Rw )Z+ and this ensures that w is both controllable and zero-controllable through c. So, we have only to investigate the non-trivial case when Rc (z ) is not of

6687

variable w, and hence is described by a difference equation of the following type P ( )w(t) = Q( )c(t), t Z+ , (6)

for suitable polynomial matrices P (z ) and Q(z ). We dene the controller behavior as C := {(w, c) (Rw )Z+ (Rc )Z+ satisfying (6)}. The overall behavior of the system obtained by interconnecting the system (3) with the controller (6) is described by Rw ( ) Rc ( ) P ( ) Q( ) w(t) = 0, c(t) t Z+ , (7)

interconnected controllers [9], the problem is unsolvable. However, dead-beat controllers in the sense of the previous denition exist. It is easy to verify that the choice P (z ) = 1 and Q(z ) = 0 leads to w(t) = 0, t 0. The prize to pay consists in the fact that c(t) does not vanish, in general, since c is constant but not necessarily zero. A characterization of the DBCs can be easily found, and it requires the following preliminary lemma. Lemma 1: Given a behavior described as in (3), a polynomial pair (P (z ), Q(z )) denes a DBC (6) for the system if and only if (z ) := M (z ) Rw (z ) P (z ) Rc (z ) . Q(z ) (8)

and it will be denoted by Kf ull . Clearly, Kf ull = Bf ull C . The target of the control problem, however, is not the whole behavior Kf ull , but only its projection on the tobe-controlled variable w, and accordingly we dene the controlled behavior as K := {w (Rw )Z+ : c (Rc )Z+ s.t. (w, c) Bf ull C} = Pw Kf ull . Note that this perspective is different both from the one adopted in the full interconnection case and from the one adopted in the partial interconnection case [14], in that the controller acts both on c and on w, as in the full interconnection case, but the target is only w, as in the partial interconnection case. It must be remarked, however, that the set-up we have chosen is consistent with the one normally adopted with state-space models, where the control law involves both the input and the state variables, but the target is only the state. Within the class of controllers we are interested in those that make the resulting controlled behavior autonomous and nilpotent: dead-beat controllers. Denition 5: Given a behavior Bf ull described as in (3), a controller (6) is said to be a dead-beat controller (DBC) for the system if in every trajectory (w, c) Kf ull , the component w(t), t Z+ , has nite support, which amounts to saying that the behavior K is nilpotent. The concept of DBC is rather intuitive and very much in line with the philosophy underlying the denitions introduced in the previous section: through a dead beat controller, that makes use of both the control variable c and of the to be controlled variable w, we aim at ensuring that all the trajectories of w go to zero in a nite number of steps. What happens of the control variable is not strictly relevant, and we may even accept that, in order to ensure that w goes to zero in a nite number of steps and remains zero, the control action may last forever. E XAMPLE 1 (continued) Consider the behavior Bf ull described in Example 1 and corresponding to Rw (z ) = Rc (z ) = z 1. We have already seen that Bf ull = ker [ Rw ( ) Rc ( ) ] is neither controllable nor zerocontrollable, while w is both controllable and zerocontrollable through the variable c. If we attempt to solve the dead-beat control problem by resorting either to (regular) full interconnected controllers or to (regular or not) partially

is right monomic, where M (z ) is an MLA of Proof: Clearly Kf ull = ker If Rw ( ) Rc ( ) . P ( ) Q( )

Rc (z ) were of full row rank then K = (Rw )Z+ and Q(z ) hence (P (z ), Q(z )) could not dene a DBC. So, an MLA M (z ) of this matrix exists and K can be expressed as K = ker M ( ) Rw ( ) P ( ) .

Thus K is nilpotent if and only if (z ) is right monomic. Theorem 1: Given a behavior described as in (3), a pair (P (z ), Q(z )) denes a DBC (6) for the system if and only if the equation X (z, z 1 ) Rw (z ) Rc (z ) = [ Iw P (z ) Q(z ) 0] (9)

has an Laurent polynomial solution X (z, z 1 ). Proof: If the pair denes a DBC then, by the previous Rc (z ) lemma, we know that, given an MLA M (z ) of , the Q(z ) matrix (z ) given in (8) is right monomic. So, (z ) admits an L-polynomial left inverse, say L (z, z 1 ). But then L (z, z 1 )M (z ) Rw (z ) Rc (z ) = [ Iw P (z ) Q(z ) 0],

and hence equation (9) has the L-polynomial solution X (z, z 1 ) = L (z, z 1 )M (z ). Conversely, if (9) holds, then, in particular, X (z, z 1 ) is a Rc ( z ) left annihilator of . Consequently, Q(z ) X (z, z 1 ) = F (z, z 1 )M (z ), for some L-polynomial matrix F (z, z 1 ). This implies F (z, z 1 )M (z ) Rw (z ) = Iw , P (z ) Rw (z ) P (z )

and therefore the polynomial matrix (z ) = M (z )

6688

has the left L-polynomial inverse F (z, z 1 ), which ensures that (z ) is right monomic. By the previous lemma, this implies that the pair (P (z ), Q(z )) denes a DBC for the system. VI. A DMISSIBLE DBC S In the previous section we have introduced the concept of DBC and showed how to characterize the pairs (P (z ), Q(z )) that correspond to the DBCs of a given behavior (3). It is worthwhile noticing that in doing so we did not introduce any assumption on the system, and indeed every system admits a dead-beat controller. In fact, the goal of forcing to zero all the trajectories of K in a nite number of steps, is always achievable, independently of the behavior properties: it is sufcient to choose, for instance, the controller P (z ) = Iw and Q(z ) = 0, to ensure that K is the zero behavior, and hence the nilpotency of K is achieved. The reason for this contradiction is that it is always possible to restrict the behavior trajectories to the zero set, but this is obtained by resorting to meaningless controllers that essentially rule out any system evolution but the trivial one. For this reason it is fundamental to understand what are the features that an admissible DBC should reasonably endow the resulting controlled system with. Dead beat controllers for standard state-space models provide good suggestions in this direction. Given a statespace model, if we decide to apply the DBC starting at some time N > 0, we expect to nd among the trajectories of the controlled behavior at least one nite support trajectory that coincides with the original trajectory w B in the initial window [0, N 1]. When moving to the general behavior setting, what we expect is that the DBC performs its task without constraining the initial portion of the trajectories in B: so if it starts working at time t = N , it does not affect the samples of the trajectory in some initial window [0, M 1] provided that N M is large enough. By assuming this perspective, we want to introduce the concept of admissible DBC. To this end, we need some mathematical preliminaries. Consider the difference equation (7), describing the behavior Kf ull of the overall system. As we enlightened within the proof of Lemma 1, if C denes a Rc (z ) DBC for system (3), the matrix cannot be of full Q(z ) row rank. If so, K = (Rw )Z+ , and hence it could not be a Rc ( z ) nilpotent behavior. An MLA M (z ) of can always Q(z ) 1 be described as follows : M (z ) = Mc (z ) 0 , M1 (z ) M2 (z )

K can be equivalently described as K = Pw Kf ull = ker Mc ( ) 0 M1 ( ) M2 ( ) Mc ( )Rw ( ) = ker M1 ( )Rw ( ) + M2 ( )P ( ) ker [ Mc ( )Rw ( ) ] = B. Rw ( ) P ( )

(10) Starting from C , we introduce the delayed controllers Ci , i Z+ , described by the difference equation i P ( )w(t) = i Q( )c(t), t Z+ .

If we denote by Ki the controlled behavior obtained corresponding to Ci , we can describe it as Ki = = where Mi ( z ) = is an MLA of M (z ) = M0 (z ). The controller Ci acts on the trajectories (w, c) of Bf ull as the original controller C does, but instead of performing the control action from t = 0 onward, it starts at t = i. Note, however, that this does not mean that the controlled trajectories are unconstrained at the time instants preceding t = i. It is possible to show that if C is a DBC for the given system, then every Ci is. Due to space constraints we omit the proof and refer the interested reader to [2]. Lemma 2: Given a behavior described as in (3), if the controller C described by the difference equation (6) is a DBC for the system, then such is any controller Ci , i Z+ , described by i P ( )w(t) = i Q( )c(t), t Z+ . Mc ( z ) 0 ( i) (i) M1 (z ) M2 (z ) ker Mc ( ) 0 Rw ( ) (i) ( i) i P ( ) M1 ( ) M2 ( ) Mc ( )Rw ( ) ker , (11) (i) (i) M1 ( )Rw ( ) + i M2 ( )P ( )

Rc ( z ) . Clearly, C = C0 , K = K0 and z i Q(z )

The result of the previous lemma allows us to introduce the following denition of admissible DBC. Denition 6: Given a behavior Bf ull described as in (3) a dead-beat controller C described as in (6) is said to be admissible if there exists L Z+ s.t. for every w B = KL+N , the Pw Bf ull and every N N, there exists w nilpotent behavior obtained corresponding to the controller (t)|[0,N 1] = w(t)|[0,N 1] . CL+N , s.t. w We are in a position to relate zero-controllability of w through c to the existence of an admissible DBC. Theorem 2: A behavior Bf ull described by Rw ( )w(t) = Rc ( )c(t), t Z+ ,

where Mc (z ) is an MLA of Rc (z ). Accordingly, the behavior

1 If

Rc (z ) is of full row rank [ Mc (z )

0 ] is an empty matrix.

admits an admissible DBC if and only if w is zerocontrollable through c. If this is the case, then every DBC is admissible.

6689

Proof: Assume, rst, that the system admits an admissible DBC C , described by the pair (P (z ), Q(z )), and, hence K = ker Mc ( )Rw ( ) M1 ( )Rw ( ) + M2 ( )Q( )

hence admissible DBCs do not exist for behaviors for which w is not zero-controllable through c. E XAMPLE 2 Consider the behavior described by the following difference equation: 1 0 0 1 w(t) = c(t), t Z+ . 1 0

is a nilpotent behavior. This implies that there exists M Z+ s.t. all trajectories in K are zero for t M . On the other hand, we have shown that for every i Z+ Ki ker Mc ( )Rw ( ) i (M1 ( )Rw ( ) + M2 ( )Q( )) Mc ( )Rw ( ) M1 ( )Rw ( ) + M2 ( )Q( ) ,

ker i

and the behavior on the right hand-side is nilpotent, with trajectories which are identically zero at least for t i + M . So, also the trajectories of Ki have nite support included in [0, i + M 1], and this is true for every i Z+ . Since C is an admissible DBC, there exists L Z+ s.t. for KL+N every N N and every w B a trajectory w B can be found, coinciding with w in [0, N 1]. Such a is surely zero for t L + N + M . So, we have trajectory w proved that there exists L N, specically L := M + L, KL+N s.t. for every w B there exists a trajectory w B coinciding with w in [0, N 1] and zero for t N + L . This proves that w is zero-controllable through c. We have already pointed out that a DBC always exists, independently of the behavior properties. We want to show that when w is zero-controllable through c, every DBC is admissible (this, obviously, implies that there exists an admissible one). Let (P (z ), Q(z )) be the pair of polynomial matrices that describes a DBC. We have only to verify that it is admissible. By the zero-controllability property, there exists a nonnegative integer L s.t. for every N N and ,c ) Bf ull s.t. every (w, c) Bf ull , one can nd (w |[N +L,+) = 0. w (12) We want to show that this same nonnegative integer L makes the denition of admissible DBC satised. To this end we have to show that for every N N and every w B, KL+N coinciding with w in [0, N 1]. By there exists w resorting to the same reasonings we previously used, we can claim that and
KL+N = ker = ker Mc ( ) (L+N ) M1 ( )
(L+N ) M1 ( )

For this behavior, w is not zero-controllable through c. Let C be any DBC for this system. Despite each delayed controller Ci (i Z+ ) seems to start its control action only at t = i, it actually imposes the constraint w2 (t) = 0 for any t 0. This is due to the fact that the plant equations constrain w2 (t) to assume a constant value, and the DBC action ensures that w2 (t) = 0 after a nite number of steps. Therefore no DBC can be admissible, since it constrains the second component of the trajectory w(t) to be zero-valued independently of the time when the DBC starts to act. In other words, there is no possibility of driving to zero a trajectory w(t) unless w2 (0) = 0. R EFERENCES
[1] M.N. Belur and H.L. Trentelman. On stabilization, pole placement and regular implementability. IEEE Trans. Aut. Contr., 47:735744, 2002. [2] M. Bisiacco and M.E. Valcher. Dead-beat control in the behavioral approach. to appear in IEEE Trans. Aut. Contr., 2012. [3] M. Bisiacco, M.E. Valcher, and J.C. Willems. A behavioral approach to estimation and dead-beat observer design with applications to state-space models. IEEE Trans. Aut. Contr., 51, no.11:17871797, 2006. [4] S. Fiaz and H.L. Trentelman. Regular implementability and stabilization using controllers with pre-specied input/output partition. IEEE Trans. Aut. Contr., 54, no.7:15611568, 2009. [5] P.A. Fuhrmann and J. Trumpf. On observability subspaces. Int. J. Control, 79:11571195, 2006. [6] A.A. Julius, J.C. Willems, M.N. Belur, and H.L. Trentelman. The canonical controller and regular interconnection. Syst. Control Lett., 54,no.8:787797, 2005. [7] J.W. Polderman and J.C. Willems. Introduction to Mathematical Systems Theory: A behavioral approach. Springer-Verlag, 1998. [8] P. Rocha. Structure and Representation of 2-D Systems. PhD thesis, University of Groningen, The Netherlands, 1990. [9] H.L. Trentelman. Behavioral methods in control. In William S. Levine, editor, The Control Handbook, pages 558581. CRC Press, Taylor & Francis, Boca Raton, FL, 2011. [10] M.E. Valcher and J.C. Willems. Dead beat observer synthesis. Systems & Control Letters, 37:285292, 1999. [11] M.E. Valcher and J.C. Willems. Observer synthesis in the behavioral approach. IEEE Trans. Aut. Contr., 44, no.12:22972307, 1999. [12] J.C. Willems. From time series to linear system, part I: Finite dimensional linear time invariant systems. Automatica, 22:561580, 1986. [13] J.C. Willems. Paradigms and puzzles in the theory of dynamical systems. IEEE Trans. Aut. Contr., AC-36:259294, 1991. [14] J.C. Willems. On interconnections, control, and feedback. IEEE Trans. Aut. Contr., AC-42:326339, 1997. [15] J.C. Willems. The behavioral approach to open and interconnected systems. IEEE Control Systems Magazine, 27, no.6:46 99, 2007. [16] J. Wood and E. Zerz. Notes on the denition of behavioural controllability. Systems & Control Letters, 37:3137, 1999.

|[0,N 1] = w|[0,N 1] , w

0
(L+N ) M2 ( ) (L+N ) M2 ( )

Rw ( ) N +L

P ( )

M c ( )R w ( ) N +L P ( )

B that So, it is easy to see that the same trajectory w satises (12), and whose existence is ensured by the zerocontrollability property, is necessarily a trajectory of both B = ker(Mc ( )Rw ( )) and ker( L+N P ( )). Therefore KL+N and this makes the denition of admissible DBC w satised. The following example illustrates why a delayed DBC unavoidably constraints the whole controlled trajectory, and
6690