such that (Brauer and Nohel, 1989)
In particular for t = x (t)
T
qn+1 = Rn
zn = &qn
~n
(T, 0) =
(9) (10)
where 81 , 82,Cl, C2 , 13 11 , 1312 , 1321 , 1322 are bounded linear operators defined by &:X--+ H qn --+ &qn
(5)
(&lin) (t) =
Consider the zero-input state response, at time t + T, the state is given by
8j : ~
(11)
Ci (t) qn (t)
(12)
~~
(13)
--+
rn --+ Bjr n
x (t + T) =
=
(8j rn) (t)
and moreover for n ;::: 1
=
J
(t , () Bj (() rn (() d( (14)
o
x (nT + t) =
13ij : ~
--+ ~ H~ rn --+ 'Dijrn
The significance of equation (6) is that if a solution is known over one full period, then the solution is known for all time. This is in fact a restatement of Floquet theorem for periodic differential equations in matrix form. A symbolic way of writing equation (6) is to view x (t) as a trajectory composed of successive "segments" over one full line period. Precisely, let denote the lifting of x , i.e.,
i,j = 1,2 and 0 ::; t
Xn (t) = x (t + nT), 0 ~ t < T, nE N then from equation (6), the components of the lifted state verify the abstract first-order homogeneous difference equation
Xn+1
= rxn
X
-+
X
1-+
Xn+l
Remark 1. A different infinite-dimensional model specialized to sampled-data systems is proposed in (Yamamoto, 1994)where the state-transition operator is an operator on C (0, T] (the Banach space of continuous functions with the uniform norm). The state-transition operator as described
with the state-transition operator, which we call the discrete semigroup generator of the system
xn
(15)
and rn stands for wn or Un. The state variable qn in equation (8) is the lifting of a new variable (see appendix - equation (27) ) and in fact , it embeds the original state xn of the homogeneous system since for zero-inputs wn = Un = 0, qn reduces to xn . It is also interesting to note that the values of qn and xn match exactly at t = O. Since all the signals and the state in equations (8), (9) and (10) are distributed discrete-time signals, the distributed parameter being the continuoustime variable t (0 ::; t < T), it is clear that these equations establish a ''faithfull'' correspondence with the full behavior of the original continuoustime periodic system.
x
F:
+ 13 n wn + 1312 un
(4)
(8)
the output equations are given by
(3)
= 0, (3) yields
+ 81wn + ihfin
481
°: :;
x
for (t) i= 0 almost everywhere on t < T, if and only if A = Aj , j = 1, 2, ... ,p; so the set of eigenvalues of the matrix cP is the point spectrum of the discrete semigroup generator J . Let A i= Aj, for j = 1,2, ... , p , then (16) is solvable and
in (Yamamoto, 1994)contains a delta like distribution (in fact, an N-tuple functionals on C (0, TJ) and acts as a "point operator" in the point t = T; actually this operator is finite-dimensional and it would be unbounded if it is to be defined on Lp spaces. Thus compared to Yamamoto (1994), our model is less restrictive and it is well suited to modem control problems when considering energy cost on the state.
x= (AI -
J) -1 Y
{=}
x(t) = (>..IN - cp)-l y(t) (18)
J)
-1 defines for 0 :::; t < T.Clearly (>..I _ a bounded linear operator on all of X, i.e.,
4. STRUCTURE OF THE SPECTRUM
J)
(>..I -1 E B (X) .Thus p and the spectrum is
We study the distribution of the complex values A for which the operator-valued function >..I - J has an inverse. First, we give the following definitions (Naylor and Sell, 1982) , (Yoshino, 1993)
Observe that J is not a compact operator on X since cP being always a non-singular matrix and the point spectrum being a finite set, A = 0 is neither an eigenvalue nor a point of accumulation of a sequence of eigenvalues.The main result of this section is the characterization of the spectrum of the discrete semigroup generator in the following theorem.
Definition 1. Given A E B (X) , a point A E C is called a regular point of A if >..I - A is invertible, i.e., (AI - A)-l exists as a bounded operator with domain X . The set p (A) of regular points is called the resolvent set of A. The spectrum i:T (A) of A is the complement of p (A) . The spectrum can be sorted into more precise classes as follows
Theorem 2. Each point in the spectrum an eigenvalue of infinite multiplicity.
Definition 2. i:Tc
(A)
t::. = {>.. E C; R(>..I -
A)
(A) ~ {>.. E C; R (>..I - A)
i= X
and (>..I - A) is injective} i:Tp
(A) ~ {A E C; (AI - A) is not injective }
are said to be the continuous spectrum, the residual spectrum and the point spectrum of A respectively. These sets are mutually disjoint and i:T
(A) =
i:Tc
i:T
(J) is
Proof. Indeed pick one eigenvalue of J , say AI , and let us assume that it has finite multiplicity. Thus the corresponding eigenspace EA! is a finite-dimensional subspace of X. Let tP1' tP2' ... , tPr be any orthonormal basis in it. By the Riesz's lemma (Naylor and Sell, 1982)and the Hilbert space structure of X , using the Gram-Schmidt process it is possible to construct an infinite orthonormal sequence {tPn}~ such that the first r vectors are tP1' tP2' ... , tPr . Since by hypothesis Al is of finite multiplicity, there exist a compact operator on X such that Al E P + Then
= X,
(>..I - A) -1 exists and is unbounded}
i:T r
(J) = C\ {AI, ... , Ap}
fC (J fC) . All -J -fC is invertible and in particular we have
(Yoshino, 1993) , for some 8 > 0
(A) U i:T r (A) U i:Tp (A)
811xl~ :::; 11 ( All - J -
and A E i:TP (A) if and only if there exists a nonzero vector x E X such that Ax = ,Xx.
fC) xii,
for all x E X (19)
Now consider the sequence {tPn}~ c X , we obtain with (19)
Let AI, A2, ... , Ap ( p :::; N) be an enumeration of the eigenvalues of the matrix cP (with possible multiple eigenvalues) . Consider the following equation on X
811tPn - tPmll :::; 11 (All - J) (tP n - tPm)11 + IIfC (tPn - tPm)1I (20)
(16)
and for n, m > r , the sequence vectors tPn ~ and therefore there exist "Y > such that
°
Now (( >..I -
J) x) (t) = (>..IN -
CP) x(t)
= 0 (17) 482
EAl
We can choose 'Y such that 'Y < 6, and the following inequality follows from (20) and (21)
6. REFERENCES Barnieh, B. and Pearson J.B. (1992), A general framework for linear periodic systems with applications to ?too sampled-data control, IEEE Trans. Automat. Contr. , vol. 37 , pp. 418-435 Brauer, F. and Nohel J.A. (1989) The Qualitative The
But this contradicts the compactness of operator
JC and thus the finite multiplicity of the eigenvalue ).1 ' Since ).1 has been chosen arbitrarily, the same reasoning applied to all others eigenvalues in the spectrum. In the course of the proof of the above theorem, we have thus shown the following
Theorem 3. For every ). E
(J'
(1),
there is a
sequence {
(i) lI
()'I - 1)
A sequence {
(1)
are verified is called the essential spectrum of 1 and it is denoted by (J' ess Thus a complete characterization of the spectrum can be stated as
7. APPENDIX-DERIVATION OF THE DISCRETE-TIME MODEL
(1).
To avoid cumbersome notations, define an input r = (w ,u)T and an input matrix B = (Bl B2) . The zero-state response of system (1) at time (t + nT) is given by (see formula (2))
Corollary 4. The spectrum of the discrete semigroup generator 1 consists entirely in an essential spectrum.
J
t+nT
x(t+nT)=
o
5. CONCLUDING REMARKS
(23)
In this note, we have presented a state-space representation for continuous-time periodic systems on an abstract space and the emphasis has been made on the internal behavior as opposed to the external approach. A complete characterization of the spectrum of the system has been obtained. Our motivation for introducing such a framework comes from a desire to understand system theoretic properties in the study of sampled-data control systems from a continuous-time viewpoint. Indeed, the assumption of finite-dimensionality of the state in the lifting framework hides important questions regarding systems theoretic concepts such as stabilizability and controllability and it is not at all clear what are the limitations of the performance of sampled-data controls inherent to their structure. These questions remain important subjects under current investigation and we believe that the state-space representation provided in this note is an appropriate one for the study of such questions.
Split the integration in (23) from 0 to t+(n - 1) T and t+ (n -1) T to t+nT and use relation (3) to obtain
x (t + nT) =
+ (n -
1) T)
t+nT
J
+
t+(n-l)T (24)
for all t 2: 0, and n E N. Now assume that 0 $ t < T , and split the integration in (24) as the sum of integrals over the intervals [(n -1) T + t , nT] and [nT, nT + tj to yield nT+t
x(t+nT) -
J
nT
=
483
nT
+
J
T
q,(t+nT,T)B(T)r(T)dT
J
1Pq, (t + (n - 1) T, ( + (n - 1) T) .
(25)
o
(n-l)T+t
B (( + (n - 1) T) r (( + (n - 1) T) d(
Note that the left hand side of equation (25) is related to information up to time nT + t whereas the right hand side is related to information up to time nT. The integral in the right hand side of (25) can be written as (the integrand has been omitted for simplicity) nT
nT
Thanks to the T-periodicity of matrices B (t) and A (t), equation (28) reduces to q (t
+ nT) = IP . q (t + (n - 1) T) T
+
(n-l)T+t
J (-) J (-) - J (.) (n - l)T
(29)
(n-l)T
Equation (29) describes the same dynamics as (23) with a new state variable defined by (27) . For the output equations of system (1) , introduce the output vector s= (z, and matrices C =
and using relation (3) in the second integral of the right hand side of the above identity, equation (25) can be written as
( g~ ) and D
nT+t
J
{x(t+nT)-
1Pq,(t, ()B(()r((+(n-1)T)d(
o
=
(n-l)TH
J
=
yf (g~~ ~~~) , then at time
t + nT, one obtains the output equation in terms
q,(t+nT, T)B(T)r(T)dT}
of the new state q
nT
= IP . {x (t
+ (n -
1) T) -
s (t
(n-l)TH
J J
+ nT) =
+ nT)
nT+t
q,(t+(n-1)T,T)B(T)r(T ) dT}
+
(n-l)T
J
C(t)q,(t+nT,T)B(T)r(T)dT
nT
nT
+
C (t) q (t
+D (t) r (t + nT) q,(t+nT, T)B(T)r(T)dT
(26) which is written, after the change of variable T = nT + ( and assuming that 0 ::; t < T,as
(n-l)T
A key observation here is that equation (26) can be viewed as an evolution equation for the quantities in the parentheses. Define a new variable q as
s (t + nT) = C (t) q (t + nT) + t
J
{C (t) q, (t , () B (()
+ D (t) c5 (t -
()) r (( + nT) d(
o q (t
+ nT) =
+ nT)
x (t
- j+tq, (t
(30) where c5 (t) is the Dirac distribution. Let denote the lifting of q, rand s , i.e.,
+ nT, T) B (T) r (T) dT
nT
, qn (t) = q (t + nT) , Tn (t) = r (t + nT) , sn (t) = S (t + nT)
(27) and write simply equation (26) as q (t
+ nT) =
IPq (t
+ (n -
J (n-l)T
q,r, S
with 0 ::; t < T, n E N. Now, in terms of the lifted variables, equations (29) and (30) are written as
1) T)
nT
+
qn+l = Fqn + Brn sn = Cqn + 15rn
q,(t+nT,T)B(T)r(T)dT
(28)
(31)
where F, 8, Cand 15 are bounded linear operators defined on appropriate Hilbert spaces (see equations (11)-(15) for the details) .
Now, it is straightforward to show that the kernel of the integral operator in the right hand side of equation (28) is independant of n for all n E N. Indeed, use relation (3) and make the change of variable T = (n - 1) T + ( , dT = d( to express the integral in equation (28) as
484