Professional Documents
Culture Documents
North-Holland
163
Abstract: This paper studies the problem of characterization and computation of the H~-norm of sampled-data systems using the
time-invariant function space model via lifting. With the advantage of time-invariance, the treatment gives an eigenvalue-type
characterization, first in the operator form in the frequency domain and then in the Hamiltonian-type finite-dimensional form. The form
obtained can be adopted for use with the bisection algorithm for actual computation.
Keywords: Sampled-data systems; H -norm; singular value equation; generalized eigenvalue problem.
1. Introduction
The H-type problem for sampled-data systems has attracted much interest recently, just for a few see
[1, 3, 5-10, 14, 17]. Francis [5] and Chen and Francis [3] gave a way of computing the L2-induced norms of
certain types of systems; for the H design problem, Toivonen 117], Hara and Kabamba [6], Kabamba and
Hara [8,9], Bamieh and Pearson [11 Hayakawa et al. [7] derived norm-equivalent finite-dimensional
discrete-time systems for a given standard H~-control problem. In contrast to these, Sun et al. [15],
Sivashankar and Khargonekar [14] derived directly a Riccati equation type solution through a limiting
procedure of finite-horizon Riccati differential equations.
A somewhat different but closely related problem is that of computing the L2-induced (or H -) norm of
sampled-data systems [3, 5]. While a time-domain solution is obtained by Hara and Kabamba [6, 8, 9], this
paper attempts a direct characterization of the H-norm of a sampled-data system in the frequency domain,
using a time-invariant function space model via lifting [18,17, 20,1,21]. The resulting advantages of
time-invariance are the following:
Transfer matrix becomes available so that H~-norm is directly characterized as the maximal singular
value in the frequency domain.
The singular value equation, along with the time-invariance of the adjoint system, naturally leads to
a Hamiitonian characteristic equation.
As we will see below, these advantages of time-invariance turn out to be a crucial asset for clarifying
various technical but delicate issues involved in the eigenvalue and singular value problems in the characterization of H~-norms.
Correspondence to: Y. Yamamoto, Division of Applied Systems Science, Faculty of Engineering, Kyoto University, Kyoto 606,
Japan
* Research supported in part by the Tateishi Science and Technology Foundation.
0167-6911/93/$06.00 1993 - Elsevier Science Publishers B.V. All rights reserved
164
f~ y*(O)x(O)dO,
(1)
where denotes conjugate transposition. Note that the order of arguments is reversed on the right; this
follows from the requirement that the inner product (x, y) must be, by definition, conjugate-linear in the
second variable.
0).
(2)
The kth element represents, in general, an intersample signal at the kth step. When considered over L 2 [0, oo),
this mapping gives rise to a norm-preserving isomorphism between L 2 [0, ~ ) and ~.'to,
2 hl, where the latter is
equipped with the norm
II{ek } II :=
IICPkII2L 2 1 0 . h ]
k=O
In what follows, the function space valued signal ~p~(0) will be denoted by ~p[k](0), etc., to clarify the
distinction between the digital and analog variables. In accordance with this, pure discrete-time signals may
be denoted as x[k], etc.
Let
y(t) =
Cexc(t) + D~u(t)
where 6(0 is the delta function. This is a time-invariant discrete-time system where the intersample
(continuous-time) variable 0 comes into play only as a parameter describing operators. Combining this with
a digital controller, we consider in this paper a generalized sampled-data system in the following form:
Xd['k + l'l
/A~s
As J [ x d [ k ] J +
rxotkJ1
'
(3)
(4)
165
where xc[k] and Xd[k] denote, respectively, the continuous and discrete state variables and belong,
respectively, to C "c and C "~. The matrices C1(0) and C1 (0) are continuous functions in O, and operators B,
D are of the following form:
B: L2[0, hi ~ C"c'u(" ) ~'~ f~ K(h - z)u(z)dr,
(5)
(6)
W ( O - z)u(z)d~.
Here we have also made, and will continue to make, the following assumptions:
(1) The input term to the discrete state is zero; this is necessary to assure the boundedness of the H~-norm
of the system; see, e.g., [6, 8].
(2) B: L 2 [0, hi ~ L z [0, hi arises from an integral operator with an L z kernel function K (.), and there is no
directly sampled input to the continuous state xc[k]. This is also usually required to assure the existence of
the H~-norm (the input term must be preceded by a low-pass filter before sampling) [3, 6, 8, 9]; note that
B has finite-dimensional range, so it is a compact operator.
(3) The system is exponentially stable, i.e., the sample-time transition matrix
A :-" LAds
Ad J
(7)
must satisfy A" ~ 0 as n --. ~ , i.e., the eigenvalues of A must be all inside the unit circle. This is also equivalent
to the poles of the transfer function (as appropriately defined) of (3) and (4) all being inside the unit circle
(assuming no pole-zero cancellation); see, e.g., [18].
Although seemingly a little restrictive, the assumptions above are general enough to handle the H~-norm
problem of all sampled-data systems of interest. Typically, A, Ci(O), K(O), W(O) are of the following form:
A,s - eao~ + ~2 eao'h-"Hod~) dT,
K(O) = eA~Bc,
Acd=f2ea='n-"H
)dT,
d(
(8)
(9)
:= Z ,p z
k=0
(10)
we can also define the z-transform of (8) and (9) as D + C(zl - A)-lB. This admits the Neumann series
expansion
oo
(11)
k=!
at least for sufficiently small z- x. When Assumptions 1-3 are also satisfied, for each fixed z with Iz- 11 < 1,
G(z) is a bounded linear operator from L 2 [0, hi to L 2 [0, hi. Since A is stable and since operators C and B are
166
continuous (as a consequence of Assumption 2 above), the operator G (z) is uniformly bounded for Iz- 11 < 1,
and according to [1 I] (see also [1]), its H~-norm
IIGII~:=
sup t sup
I~-'1<1 t~L2to.hJ
IIG(z)vlI2}
(12)
Ilvll2
is finite. It is also known that this norm is equal to the L2-induced norm in the time-domain. That is, if
u e L 2 [0, ~ ) is an L 2 input, and if we denote by Gu its corresponding output, 1 then (12) is known ([11, 1]) to
be equal to the L2-induced norm
IIG II = t
(13)
Furthermore, by the exponential stability, its domain of analyticity extends to Iz-ll _< 1, and by the
maximum modulus principle, (12) becomes
IIGIl~:=
sup
sup
O<~<2n t v~L2[O.h]
(14)
IlG(ei~)vll2}
Ilvll2
"
Also, in this case, for each fixed z with Iz- 11 < 1, Go(z) in (11) converges in norm because A k --, 0. Since B is
a compact operator as noted above, each CAkB is also compact so that, as a uniform limit of compact
operators, Go(z)is compact (but D is never compact unless Dc is zero) and G(z) is the sum of D and a compact
operator. This decomposition plays a crucial role in what follows.
Remark 2.1. We have employed here the finite-dimensional state space model as in [1] (as opposed to [18]
where one also takes state trajectory as a state), so that the state space is also infinite-dimensional. On the
other hand, in the finite-dimensional model here it is necessary to introduce an infinite-dimensional direct
transmission operator D to describe the continuous-time plant. However, from the external viewpoint, they
describe the same behavior in lifting, so both lead to the same notion of transfer functions.
(15)
where W is compact.
We now prove the following result.
1The notation is a little sloppy, but here we regard G as an operator in the time domain also.
2 The author is indebted to S. Hara and P.P. Khargonekar for valuable comments and discussions concerning the subsequent material in
the paper.
167
L e m m a 3.1. Suppose II vii = 3,2 > iiDll 2. Then 3,2 is the maximum eigenvalue of V. Hence, IlG(eJ~')lJ is also
attained as the maximum singular value. 3
Proof. Let
(16)
R~ := 3,21 - D*D.
dS..
Since m is positive, this square root satisfies
IIR~/2vll 2 > mllvll z.
(17)
(18)
Here R~ ~/2 W R ~ 1/2 is a compact operator since the composition of a continuous and compact operator is
again compact. Since 3,2 is the norm of V, the left-hand side of (18) is nonnegative-definite, and so is
[I - R 71/2 W R ~ 1/21. Therefore, ][R~- 1/2 W R ~ 1/2 II -< 1.
We claim that IIU ; 1/2 W R ~ 1/2 II - 1. Assume the contrary, i.e., I]R~" 1/2 W R ( 1/2 II = 6 < 1. We have
( x , U ~ 1/2 WR~I/2x) <- 6llx[I 2
(19)
R~/2p)
yields
(20)
D * D - W i v e = O.
[]
a If the assumption here is not satisfied, then the conclusion here need not follow. The author is indebted to P.P. Khargonekar for this
and other relevant discussions here.
168
Lemma 3.2. There exists a Zo on the unit circle such that l[G 1[2 = l[ V(zo)11.
Proof. By (14), there exists a sequence {Zk}, all of modulus 1, such that l[ V(Zk)l[ approaches IIGII2 from
below. Since the unit circle is compact, there exists a convergent subsequence; without loss of generality,
assume z~ itself is convergent, to Zo. Clearly, IIV(zo)ll -< IIGII~.
Now let Vk, Vo be normalized eigenvectors that give rise to the eigenvalues 2k (= II V(zD II), ;to(= II V(zo)II).
Since II V(e j~') II is continuous in tO by the stability of G, for any ~ > 0, there exists N such that, for all k > N,
20 = IIV(zo)(Vo) II>- II V(zo)(Vk)II > IIV(zD(vD II - ~ > IIG II~ - 2~.
Since ~ is arbitrary, this yields
[]
Theorem 3.3. Let G(ei~')= D + Go(e j~') and V(e i~') be as above, and suppose ~,> max{llDH, /z:=
info_<,o<2~llG(eJ~')l[}. Then [IGII~ < T if and only if there exists no real tOo such that
2ma,(G * G(e ZOO))= ~)2.
(21)
(22)
To reduce (22) to a finite-dimensional equation, we first need the state space form for the adjoint G*. So we
start with the following lemma, which is a rather straightforward consequence of the standard duality for
discrete-time systems.
Lemma 4.1. The dual system of equations (3) and (4) is given by
r,,o,,-. [..,,.
~r,. UPd[k]l =
A*d A* JLPd[k + 1]
+ L(C,,v[k](0))* '
(23)
~69
Proof. We readily see that the dual system of (8) and (9) with input v[k] and output z[k] is
,,,-,<-,
] r,,<'-,:, ] r,,<,-,<+ ,-,]
pd[k] = LA*a A~ LPd[k + 1] + C*v[k](O),
(24)
(25)
To exhibit this dual system in a concrete form, we need only compute the dual operators B*, C* and D*. By
Lemma A.I, we have
c*o(o)= ILtc=,
-(c'' v(o))*]
vie))* e c " + " "
LPdJ
(26)
LPdJ
and
rh
(28)
[]
(y2I - G*G(eJ'))u = 0.
(29)
To express this in terms of the state space equations, let us write down y = G(eJ~)u and v = G*(eJ~')y, u, y,
v ~ L 2 [0, hi, and set v = y2u. With state space equations (3), (4) and (23), (29) admits a nontrivial solution if
and only if there exist u, v, y, not all zero, such that
G(eJ~')u: eJ~'x = Ax + Bu,
y - Cx + Du,
(30)
(31)
where x:= [x~, xT] "r, p:= [p[, pTo]T (recall also D = D).
Combining them together leads to
(32)
so that
(33)
Suppose that ~ > [ID II from now on. Then R~ is invertible, and we can solve (33) as
(34)
"
(35)
Substituting this for u in (30) and (31), and eliminating y will complete the picture. This yields
(36)
(37)
Now write ~ := [x T pT]lr = [x~ xa~ p~ paT]T. Then by (3), (4) and. (23), we get the generalized eigenequation
e j % ~ = ..~,
(38)
170
where
I13 ]
0 I 0
$:= 0 0 ~f33 A~'~ '
0 0 $',*3 A*J
~13 = - B R ~ I K * ( h -
(39)
'~31
'-~32
'
~4~ ~42 0
").
'-~12 --
Acd + BR(
t D'C2('),
We now see that this eigenvalue problem admits a nontrivial solution if and only if the same is true of (22).
Theorem 4.2. Assume ? > l}D I}. There exists a nontrivial solution u to the equation
(4o)
(721 - G*G(eJ~))u = 0
if and only if
(41)
Proof. If there is a nonzero solution u to the first equation, then from the discussion above, it is clear that
there must exist a nonzero ~ (otherwise, by (34), u(O) derived from ~ is easily seen to be 0) that satisfies (38), so
that (41) follows.
Conversely, suppose there exists a nonzero ~ that satisfies (38). We must assure that u(O) induced by ~ is
not identically zero. Suppose the contrary; then all the integrals involving R~-i are zero and hence we must
have
ej~[;
,IE,l
(42)
/~:-(43)
171
This theorem gives a sampled-data version of [2, Theorem 2]; since the characteristic equation here is
finite-dimensional, we can readily adopt the bisection algorithm combined with an effective check of the
existence of an eigenvalue of modulus one, e.g., by converting this to 's-domain' via bilinear transformation,
and employ the well-known Sturm method ([2]; see also [12]). However, to guarantee the condition
? > [ID [[, we need the value of [[D 1[ in advance. This is related to an H~-problem for a delay system, and
a computational procedure is also obtained (see [4, 19, 1] for details). Observe also how the equation above
has the structure of Hamiltonian system, i.e., if2 is an eigenvalue to the above, then so is 1/2 (this is clear from
the form of (30) and (3 I)).
A similar but different Hamiltonian eigenvalue problem has been obtained in [8, 9]. The difference is
mainly due to the fact that here operator R~- 1 must be present as a result of one-step integration while in their
setting, the computation of R~- ~ is embedded into the framework of their hybrid state space model. The
mutual relationship is, however, not yet entirely clear.
The entries appearing in the characteristic equation above contain terms involving R~ ~ and the integrals
containing it. However, as shown in [1], this can be explicitly computed, and when the hold function is the
zero-order hold, all the integrals can be obtained via exponentiations (see, e.g. El, 7I). Details, along with
generalizations to a more general context, will be reported in a subsequent work.
Appendix
Lemma A.I. Let K(O) be a fixed (continuous)matrix function on [0, hi. Define the operators T~, i = 1,2,3 as
follows:
(1)
(2)
(3)
Then
(I)
(2)
(x,r*O)=(rlx, O)f(K(')x,O('))=
O*(O)K(O)xdO=
x,
O*(O)K(O)dO
=(x,(K, 0)*).
K(h -
(K(h-
T)x)d .
~(0)*
K(O - z)u(T)d~ dO =
- (u, ( fn
dz
K (O- dO)* )
[]
Acknowledgements
The author wishes to thank Professors S. Hara and P.P. Khargonekar for their helpful discussions on the
material in Section 3, especially for pointing out a flaw in the first draft. He also wishes to thank Professor
172
M. Ikeda for helpful comments on the manuscript, and the anonymous referees for valuable comments on the
final manuscript.
References
[1]
B. Bamieh and J.B. Pearson, A general framework for linear periodic systems with applications to H sampled-data control, IEEE
[2] S. Boyd, V. Balakrishnan and P.P. Kabamba, A bisection method for computing the ~