trol of multirate sampled-data systemsweb.abo.fi/fak/tkf/rt/reports/rep97_4.pdfh 1 con trol of...
TRANSCRIPT
H1 Control of Multirate Sampled-Data Systems:
A State-Space Approach
Mats F. S�agfors� and Hannu T. Toivonen
Department of Chemical Engineering,�Abo Akademi University,
FIN{20500 �Abo, FINLAND
[email protected], [email protected]
Bengt Lennartson
Control Engineering Laboratory
Chalmers University of Technology
S{412 96 Gothenburg, SWEDEN
Abstract
A state-space solutionof the H1 control problem for periodic multirate sampled-
data systems is presented. The solution is characterized in terms of a pair of discrete
algebraic Riccati equations with a set of associated matrix positive de�niteness condi-
tions and coupling criteria. The solution is derived using two di�erent approaches. In
the �rst approach, the solution is obtained by discretizing two coupled periodic Ric-
cati equations with jumps which characterize the H1-optimal controller. The second
approach is based on the lifting technique. The H1-optimal controller for a lifted rep-
resentation of the multirate system is characterized by a two-Riccati equation solution,
in which the so-called causality constraints are represented by a set of nonstandard
positive de�niteness conditions and coupling criteria.
�Supported by the Academy of Finland
1
1 Introduction
During the last few years there has been an increased interest in the direct design of digital
controllers using continuous-time performance measures. Such systems arise naturally when
a digital controller is applied to a continuous-time plant. In contrast to standard continuous-
time or discrete-time techniques, direct sampled-data methods address both the intersample
behavior of the process, as well as the e�ect the sampling frequency has on the performance.
Multirate systems arise when various measurement sampling functions and control signal
hold elements operate at di�erent rates. The development of a theory for such multirate
systems started in the 1950's with the pioneering work by Kranc [20], Kalman and Bertram
[18] and Friedland [13]. More recent work in the �eld is concerned with e.g. system analysis
and stability [3, 16, 23, 24, 25, 38], optimal control of multirate systems with a quadratic cost
function [1, 2, 6, 8, 11, 21, 22, 26, 39], and H1 control of multirate sampled-data systems
[10, 39].
The H1 problem for single-rate sampled-data systems has recently been studied in a
number of papers [5, 9, 15, 17, 30, 32, 33, 34, 35, 36]. One common approach is to apply
the lifting technique [5] to reduce the hybrid continuous/discrete-time problem to a norm-
equivalent discrete-time H1 problem, to which standard methods can be applied. Another
method is to state a solution directly for the hybrid problem [32, 33]. This leads to a solution
in terms of Riccati equations with jumps. The approaches have recently been shown to give
equivalent solutions [29, 37].
The lifting technique has also been applied to the multirate sampled-data problem [10, 39],
where the periodically time-varying sampled-data system is lifted to give an equivalent time-
invariant discrete system. The main di�culty encountered in this approach is caused by the
causality constraint, which is required for the lifted controller to correspond to a causal
periodic controller. The causality constraint has limited the applicability of standard state-
space solutions on the lifted system description. Previous solutions have therefore been
derived through constrained model matching [10, 39].
In this work, a state-space solution to the periodic multirate sampled-data H1 problem
is derived. The solution is obtained through two di�erent approaches. The �rst approach
is based on a direct Riccati-equation solution, and the second approach is based on lifting.
The direct method has, to our best knowledge, not previously been applied to the multirate
sampled-data H1 problem. The lifting method was used in [10, 39], but no time-domain
solution was derived. The use of both approaches gives insight into the relationship between
the two methods. Both approaches studied here generally give rise to the same synthesis
equations and the same controller.
In the �rst approach, the H1-optimal controller is characterized by a pair of coupled
2
Riccati equations with jumps. Using the results in [29, 37], the Riccati equations can
be discretized and the solution expressed by two periodic Riccati di�erence equations and
discrete-time coupling conditions. The solutions to these periodic equations can then be
expressed in terms of a pair of coupled standard algebraic Riccati equations. This direct
approach allows for a very general problem statement, including non-synchronized sampling
and hold elements.
In the second approach, we adopt the lifting technique to construct a time-invariant
discrete-time system associated with the periodic multirate problem. A solution of the
causally constrained H1-optimal control problem is derived, which consists of two standard
discrete algebraic Riccati equations, a set of associated coupling conditions and a set of
matrix positive de�niteness criteria. An interesting feature of the result is that the alge-
braic Riccati equations are exactly those associated with the unconstrained discrete-time
H1 problem for the lifted time-invariant system description. Thus, standard discrete-time
solution methods can be applied. The additional causality constraint associated with the pe-
riodic multirate system appear in the solution as non-standard positive de�niteness criteria
and coupling conditions, which are easy to check.
The paper is organized as follows. The next section contains the problem de�nition and
a description of the notation used in the paper. In Section 3 a solution to the multirate
sampled-data problem is derived through periodic Riccati equations. A state-space solution
through the lifting technique is presented in Section 4. A simple example is used to illustrate
the results in Section 5, and some concluding remarks are given in Section 6.
2 Problem statement
Consider a linear time-invariant �nite-dimensional continuous-time plant described by
_x(t) = Ax(t) +B1w(t) +B2u(t); x(0) = 0
z(t) = C1x(t) +D12u(t) (1)
where x(t) 2 Rn is the state vector, w(t) 2 Rm1 is the process disturbance, u(t) 2 Rm2 is
the control signal and z(t) 2 Rp1 is the controlled output.
The sampling of the measurements is assumed to be described by a periodic function S
with the period Ty. The outputs may be sampled with di�erent rates or possibly at non-
equidistant time-instants. The discrete signal yd := fyd(iTy + �j); j = 0; 1; : : : ; Ny � 1; i =
0; 1; : : :g available for feedback is given by
yd(iTy + �j) = C2(�j)x(iTy + �j) +D21(�j)vd(iTy + �j); (2)
where vd(�) is a discrete measurement disturbance and iTy + �j (�0 � 0, �j+1 > �j, �Ny=
�0 + Ty) are the times at which a measurement is available. It is assumed that the signal
3
yd(iTy + �j) contains only those measurements obtained at time instant iTy + �j. This
means that the discrete output vector has a periodic and time-varying dimension.
The control signal is assumed to be generated by a zero-order hold function H with the
overall period Tu. The number of control signal updates during one period Tu is Nu. The
individual control signals may be updated at faster rates (with Tu as a common multiple of
the hold times) or the hold times may be periodically time-varying. The multirate zero-order
hold is described by
u(t) = [I �H(�l)]u(kTu + �l) +Hu(�l)ud(kTu + �+l); (3)
t 2 (kTu + �l; kTu + �l+1] ; l = 0; 1; 2; : : :Nu � 1; k = 0; 1; 2; : : ::
where ud := fud(kTu + �+l); l = 0; 1; 2; : : :Nu � 1; k = 0; 1; 2; : : :g is a discrete-time control
signal. The matrixH(�l) := diagf�1(�l); �2(�l); : : : ; �m2(�l)g selects the control signals to be
updated at time instant kTu + �l (�0 � 0, �l+1 > �l, �Nu= �0 + Tu). The functions �m(�l)
take the value 1 if the m:th control signal is updated at time instant kTu + �l, and zero
otherwise. The discrete control signal ud(kTu + �+l) contains only those signals updated at
time kTu + �l. The matrix Hu(�l) is therefore a submatrix of H(�l), which consists of the
nonzero columns of H(�l).
Remark 2.1 The time-varying dimensions of the signals yd and ud guarantee the existence
of matrix inverses which appear in the solution. For an alternative way to guarantee non-
singularity, see [11].
The system de�ned by (1) { (3) is assumed to be periodic with the period T , such that T
is the least common multiple of the period of the sampler Ty and the period of the hold
function Tu. (The solution can, however, with minor modi�cations be extended to cover
the case when the periods of S and H are asynchronously related, cf. [28].) The multirate
sampled-data problem de�ned here is very general. Note that the only assumption on S and
H is their periodicity, otherwise the discrete events may occur at any (non-synchronized)
instants. The solution presented in Section 3 can also be generalized to the case with non-
periodic time-varying sampling. In connection with the lifting technique in Section 4 we
will make some additional technically simplifying assumptions on the synchronization of the
sampling and the hold elements.
The system (1) { (3) can be represented as a mixed continuous/discrete system with
jumps,
�G :
8>><>>:
_xe(t) = Aexe(t) + B1ew(t); t 6= kTu + �lxe(kTu + �+
l) = Ad(�l)xe(kTu + �l) +Bd(�l)ud(kTu + �+
l)
z(t) = C1exe(t)
yd(iTy + �j) = C2d(�j)xe(iTy + �j) +D21(�j)vd(iTy + �j)
(4)
4
where we have extended the state vector, xe(t) := [x0(t) u0(t)]0, and
Ae :=
�A B2
0 0
�; B1e :=
�B1
0
�; C1e := [C1 D12 ] (5)
Ad(�l) :=
�I 0
0 I �H(�l)
�; Bd(�l) :=
�0
Hu(�l)
�; C2d(�j) := [C2(�j) 0 ] : (6)
We need the following de�nitions, cf. [32],
De�nition 2.1 Let �Ad
Ae(�; �) denote the state-transition matrix of the system �G. Then �G
is said to be exponentially stable if there exist c1; c2 > 0 such that
k�(t2; t1)k � c1e�c2(t2�t1); all t2 � t1: (7)
De�nition 2.2 The system �G is said to be stabilizable (respectively, detectable) if there
exists a bounded function Kd(�l) (respectively, Ld(�j)) such that �Ad+BdKd
Ae(respectively,
�I+LdC2
A) satis�es (7).
The following assumptions are made in the paper:
A1 The system �G is stabilizable.
A2 The system �G is detectable.
A3 The matrix D012D12 is non-singular.
A4 The matrices D21(�j)D021(�j) are non-singular.
A5 The continuous-time system (A;B1; C1) is stabilizable and detectable.
The set of admissible controllers is assumed to consist of exponentially stabilizing discrete
causal control laws ud = Kyd such that ud(kTu + �+l) is a function of past measurements
fyd(iTy + �j) : iTy + �j � kTu + �lg only.
The following worst-case performance measure induced by the disturbances w 2 L2[0;1)
and vd 2 l2(0;1) is introduced,
J1(K) := sup(w;vd)6=0
24 kzkL2�kwk2
L2
+ kvdk2l2�1=2
35 : (8)
The multirate sampled-data H1 control problem is de�ned as follows:
Find an admissible controller K, if one exists, which achieves a speci�ed attenuation level
> 0, such that J1(K) < .
5
3 A Riccati equation solution
The following solution to the multirate sampled-dataH1 problem stated in terms of coupled
periodic Riccati equations with jumps stands as a basis for the rest of the treatment in this
section (Theorem 3.1). In Theorem 3.2, the Riccati equations are discretized, and the
solution is given in terms of discrete periodic Riccati equations. Theorem 3.3 shows the
fact that solutions to the periodic equations can be obtained in terms of a pair of coupled
algebraic Riccati equations.
Theorem 3.1 Consider the multirate sampled-data system �G. Suppose that the assump-
tions A1 { A5 hold. Then there exists a discrete admissible controller K which achieves the
performance bound J1(K) < if and only if the following conditions are satis�ed:
(i) There exists a bounded Tu-periodic symmetric positive semide�nite matrix function
Se(t) :=
�S S1S01 S2
�(t); t 2 [0;1) (9)
which satis�es the following periodic Riccati di�erential equation with jumps,
� _Se(t) = A0eSe(t) + Se(t)Ae + �2Se(t)B1eB
01eSe(t) +C01eC1e; t 6= kTu + �l
Se(�l) = A0d(�l)Se(�
+l)[Ad(�l) �Bd(�l)Kd(�l)] (10)
Se(kTu + �l) = Se(�l) l = 0; 1; : : :Nu � 1; k = 0; 1; : : :
where
Kd(�l) := [B0d(�l)Se(�
+l)Bd(�l)]
�1B0d(�l)Se(�
+l)Ad(�l) (11)
such that the Tu-periodic system
_xe(t) = [Ae + �2B1eB01eSe(t)]xe(t); t 6= kTu + �l
xe(kTu + �+l) = [Ad(�l)� Bd(�l)Kd(�l)]xe(kTu + �l) (12)
is exponentially stable.
(ii) There exists a bounded Ty-periodic symmetric positive semide�nite matrix function Q(�)
which satis�es the following periodic Riccati di�erential equation with jumps,
_Q(t) = AQ(t) + Q(t)A0 + �2Q(t)C01C1Q(t) +B1B01; t 6= iTy + �j
Q(�+j) = [I � Ld(�j)C2(�j)]Q(�j) (13)
Q(iTy + �j) = Q(�j) j = 0; 1; : : :Ny � 1; i = 0; 1; : : :
where
Ld(�j) := Q(�j)C02(�j)[D21(�j)D
021(�j) +C2(�j)Q(�j)C
02(�j)]
�1 (14)
6
such that the Ty-periodic system
_x(t) = (A+ �2Q(t)C01C1)x(t); t 6= iTy + �j
x(iTy + �+j) = [I � Ld(�j)C2(�j)]x(iTy + �j) (15)
is exponentially stable.
(iii)
�[S(t)Q(t)] < 2; all t 2 [0;1) (16)
where �(�) denotes the spectral radius.
Moreover, when these conditions hold, the performance bound J1(K) < is achieved by the
T -periodic controller
_x(t) = �A(t)x(t) + �B2(t)u(t) t 6= iTy + �j;
x(iTy + �j) = x(iTy + �j) + �Ld(�j)[yd(iTy + �j)�C2(�j)x(iTy + ��j)] (17)
ud(kTu + �+l) = �[H0
u(�l)S2(�l)Hu(�l)]�1H0
u(�l) [S01(�l)x(kTu + �l)
+S2(�l)[I �H(�l)]u(kTu + �l)]
where u(t) is given by (3),
�A(t) := A+ �2B1B01S(t)
�B2(t) := B2 + �2B1B01S1(t) (18)
�Ld(�j) = N (�j)C02(�j)[D21(�j)D
021(�j) + C2(�j)N (�j)C
02(�j)]
�1 (19)
and N (t) is given by
N (t) = Q(t)(I � �2S(t)Q(t))�1: (20)
Proof: The multirate sampled-data system de�ned by (4) can be regarded as a periodic
time-varying system. The multirateH1 problem can then be solved by adapting continuous-
time and discrete-time H1 theory of time-varying systems [19, 27] to this case. This leads
to a two-Riccati equation solution of the sampled-data problem. The theorem is a fairly
straightforward generalization of the corresponding single-rate result in [30, 33] and the
dual-rate result in [36].
2
The periodic Riccati di�erential equations with jumps in (10) and (13) can be discretized
to periodic Riccati di�erence equations. Similarly, the coupling condition (16) can be ex-
pressed as a discrete-time criterion. Since the sampling and hold functions may have di�er-
ent periods the discretization points of the Riccati equations involved can occur at di�erent
7
time-instants. The solution can then be expressed in terms of two discrete periodic Riccati
equations with di�erent periods.
Introduce the 2(n +m2)� 2(n+m2) matrix
�(t) :=
��11 �12
�21 �22
�(t) := exp
��
�Ae �2B1eB
01e
�C01eC1e �A0e
�t
�(21)
where Ae, B1e and C1e are de�ned by (5), and de�ne
~Ae(hl; �l) := �11(hl)�1Ad(�l)
~B1e(hl) ~B01e(hl) := � 2�11(hl)
�1�12(hl);
~B2e(hl; �l) := �11(hl)�1Bd(�l) (22)�
~C01e(hl; �l)~D012(hl; �l)
�[ ~C1e(hl; �l) ~D12(hl; �l) ] :=
�A0d(�l)
B0d(�l)
��21(hl)�11(hl)
�1 [Ad(�l) Bd(�l) ]
where hl := �l+1 � �l. Note that the structure of the matrices in (22) associated with the
augmented system (4) allows for the partition
~Ae(�; � ) =:
�~A(�) ~B2(�)0 I
�Ad(� )
~B1e(�) =:
�~B1(�)0
�(23)
~C1e(�; � ) =: [ ~C1(�) � ]
where ~A is n� n, ~B1 is n�m1 and ~C1 is p1 � n. Denote further
~C2(�) := C2(�); ~D21(�) := D21(�): (24)
It is assumed that the matrices �11(�) in (22) are invertible, which is the case if there exist
bounded solutions to (10) and (13), see [37].
Theorem 3.2 Consider the multirate sampled-data system �G. Suppose the assumptions
A1 { A5 hold. Then the conditions in Theorem 3.1 (i){(iii) are equivalent to the following
conditions:
(i) The matrices �11(hl) are invertible and there is an Nu-periodic symmetric positive
semide�nite matrix function Se(�l) satisfying the following periodic Riccati di�erence equa-
tion
Se(�l) = ~A0e(hl; �l)Se(�l+1) ~Ae(hl; �l) + ~C01e(hl; �l)
~C1e(hl; �l)�
E0S(hl; �l)M�1S
(hl; �l)ES(hl; �l) (25)
Se(kTu + �l) = Se(�l); l = Nu � 1 : : : ; 1; 0; k = 0; 1; : : :
8
where
ES(hl; �l) :=
�~B01e(hl)
~B02e(hl; �l)
�Se(�l+1) ~Ae(hl; �l) +
�0
~D012(hl; �l)
�~C1e(hl; �l) (26)
MS (hl; �l) :=
�� 2I 0
0 ~D012(hl; �l)
~D12(hl; �l)
�+
�~B01e(hl)
~B02e(hl; �l)
�Se(�l+1)
�~B01e(hl)
~B02e(hl; �l)
�0
such that
2I � ~B01(hl)S(�l+1) ~B1(hl) > 0; (27)
where S(�l) and ~B1(hl) are given by the partitions in (9) and (23), respectively, and the
discrete Nu-periodic system
xe(kTu + �l+1) = Acl(hl; �l)xe(kTu + �l) (28)
where
Acl(hl; �l) := ~Ae(hl; �l)� [ ~B1e(hl) ~B2e(hl; �l) ]MS(hl; �l)�1ES(hl; �l) (29)
is exponentially stable.
(ii) The matrices �11(sj), where sj := �j+1� �j, are invertible and there is an Ny-periodic
symmetric positive semide�nite matrix function Q(�j) satisfying the following periodic Ric-
cati di�erence equation
Q(�j+1) = ~A(sj)Q(�j) ~A0(sj)+ ~B1(sj) ~B
01(sj)�EQ(sj ; �j)M
�1Q
(sj ; �j)E0Q(sj ; �j)
Q(iTy + �j) = Q(�j); j = 0; 1; : : :Ny � 1; i = 0; 1; : : : (30)
where
EQ(sj ; �j) := ~A(sj)Q(�j) [ ~C01(sj)
~C02(�j) ] (31)
MQ(sj ; �j) :=
�� 2I 0
0 ~D21(�j) ~D021(�j)
�+
�~C1(sj)~C2(�j)
�Q(�j)
�~C1(sj)~C2(�j)
�0
such that
2I � ~C1(sj)[I � Ld(�j) ~C2(�j)]Q(�j) ~C01(sj) > 0 (32)
where Ld(�) is given by (14) and the discrete Ny-periodic system
x(iTy + �j+1) = [ ~A(�j) �EQ(sj ; �j)M�1Q
(sj ; �j)
�~C1(sj)~C2(�j)
�]x(iTy + �j) (33)
is exponentially stable.
(iii)
�[S(t)Q(t)] < 2; for t = iTy + �j; such that t � T;
j = 0; 1; : : : ; Ny � 1; i = 0; 1; : : : (34)
9
where S(t) is given by (10) and T is the period of the multirate system �G. Moreover, when
these conditions hold, the performance bound J1(K) < is achieved by the controller in
(17).
Proof: For the single-rate case, a similar discretization of the equations corresponding to
(10) and (13) was done in [29, 37] and the proof is therefore not given in its full length here.
The most obvious di�erence is the periodic time-varying nature of the resulting discrete-
time equations (25) and (30). By the analysis in [29, 37], the existence of bounded solutions
to (10) and (13) between the discretization points is equivalent to the conditions (27) and
(32), respectively, and the stability criteria (12) and (15) are equivalent to the discrete-time
criteria in (28) and (33), respectively.
An important step is the result that the coupling condition (16) can be discretized too.
The discrete-time coupling condition is obtained from the observation that the quantity
�[S(t)Q(t)] � 2 can change sign at the discontinuous jumps only, i.e. at the instants f�lg
and f�jg [29, 37]. From the equations describing the discontinuities in (10) and (13) we know
that S(�+l) � S(�l) and Q(�j) � Q(�+
j). By the property of the spectral radius function,
�(UV ) � �(ZV ) i� U � Z, where U;Z; V are positive de�nite matrices of compatible
dimensions, it then follows that if �(S(�j)Q(�j)) � 2 then �(S(t)Q(t)) � 2 is true in the
whole interval t 2 (�j�1; �j]. Thus, also the coupling condition can be discretized. The
periodicity of the Riccati equations then give the set of coupling conditions (34) . 2
Remark 3.1 Observe that S(t) in (34) is obtained from the matrix exponential (21)
Se(t) = [�21(�l � t) + �22(�l � t)Se(�l)] [�11(�l � t) + �12(�l � t)Se(�l)]�1
;
�l�1 < t � �l (35)
and the partition in (9).
Remark 3.2 An equivalent expression for the coupling condition (34) can be stated in terms
of the periodicity of Se(�):
�[S(t)Q(t)] < 2; for t = kTu + �+l; such that t � T;
k = 0; 1; : : : ; Nu � 1; l = 0; 1; : : : (36)
where S(t) and Q(t) are given by (10) and (13), respectively.
Remark 3.3 The controller (17) can be discretized by observing that the state transition
matrix � �A(�; �) associated with the matrix �A(�) and the matrix �(t2; t1) :=R t2t1�(t2; �) �B2(�)d�
can be obtained through the matrix exponential in equation (21) [36].
10
Remark 3.4 Observe that no separate causality constraint problem arises in this solution,
as opposed to the lifting solution [10, 39]. The controller (17) is periodically time-varying. A
lifted time-invariant controller description can be obtained by lifting the periodic controller
(17).
The periodic Riccati equations in (10) and (13) are discretized at those points only where
a discrete event actually occurs. An alternative, and perhaps more common formulation, is
to de�ne a small time interval, say h, which is a fraction of all sampling and hold intervals.
This way of discretizing would result in two discrete Riccati equations with the same period.
Further, the coupling condition would be expressed in terms of the discrete periodic solutions
instead of the expressions (34) or (36). We would also need one matrix exponentiation only
in order to obtain the matrices involved in the discrete Riccati equations. This way of
discretizing may, however, lead to a very long period of the periodic equations, since they
may be discretized at times where no action actually occurs. In particular, a common
sampling-time fraction may lead to a very long period for systems with non-synchronized
sampling, i.e. when there are discrete time actions both at times ft+kTg and ft+kT+�g for
some � < T . The present approach is also applicable to the case when � is irrational. This
case is not possible to solve using a constant discretization step. Our problem formulation
is stated in order to achieve the shortest possible period of the periodic Riccati di�erence
equations in (25) and (30), with the possible expense of more matrix exponentiations. The
alternative formulation with one single sampling-time fraction can be regarded as a special
case of this description.
The following theorem shows that the periodic Riccati di�erence equations in (25) and
(30) can be solved via standard discrete algebraic Riccati equations. See also [21] for an
alternative generation of an algebraic Riccati equation.
Theorem 3.3 Assume that there exists a bounded stabilizing positive semide�nite matrix
function P (�) satisfying the following discrete N -periodic Riccati equation
P (tk+1) = A(tk)P (tk)A0(tk) +R1(tk)� [A(tk)P (tk)C
0(tk) + R12(tk)]
� [C(tk)P (tk)C0(tk) +R2(tk)]
�1[C(tk)P (tk)A
0(tk) +R012(tk)] (37)
A(tk) = A(tk+N ); C(tk) = C(tk+N ); Ri(tk) = Ri(tk+N )
k = 0; 1; 2; : : :
Then there exists a solution P := P (tN ) to the following algebraic Riccati equation
P = �AP �A0 + �R1 ���AP �C0 + �R12
� ��CP �C0 + �R2
��1 � �CP �A0 + �R012�
(38)
where�A := AN ; �C := CN
�R1 := R1;N ; �R12 := R12;N ; �R2 := R2;N ;(39)
11
where the matrices are de�ned by the following recursive equations
Ak := A(tk)Ak�1; A0 := A(t0);
Ck :=
�Ck�1
C(tk)Ak�1
�; C0 := C(t0);
R1;k := A(tk)R1;k�1A0(tk) +R1(tk); R1;0 := R1(t0); (40)
R12;k := [A(tk)R12;k�1; A(tk)R1;k�1C0(tk) + R12(tk)] ; R12;0 := R12(t0)
R2;k :=
�R2;k�1; R12;k�1C
0(tk)
C(tk)R12;k�1 C(tk)R1;k�1C0(tk) +R2(tk)
�; R2;0 := R2(t0):
Proof: The result is obtained by induction, cf. [7, 12]. 2
By Theorem 3.3, we can obtain the solutions to the periodic Riccati equations (25) and
(30) at one point during their periods by solving algebraic Riccati-equations, to which there
exist algorithms with known convergence properties. The other solution points to (25) and
(30) at the remaining time instants can then be determined using the original equations, or
from the recursion
P (tk+1) = AkPA0k +R1;k � [AkPC
0k + R12;k][CkPC
0k +R2;k]
�1[CkPA0k + R012;k]: (41)
The solution to the multirate sampled-data H1 problem can thus be stated in terms of a
pair of standard discrete algebraic Riccati equations together with the set of non-standard
positive de�niteness criteria (27), (32) and the set of coupling conditions (34).
4 A lifting based solution
Periodic multirate control problems can be solved by the lifting technique, in which the
system is represented by a time-invariant, lifted system. The lifted system has as its inputs
and outputs the inputs and outputs during one period of the periodic system. The controller
can then be designed for the time-invariant lifted system. However, the requirement that
the controller should represent a causal periodic controller imposes conditions on the direct
coupling term of the lifted controller. Due to these causality constraints, standard H2 and
H1 synthesis methods are not directly applicable to the problem of designing controllers
for periodic systems via lifting. Various special methods have therefore been studied for the
problem of designing optimal controllers which satisfy a causality constraint [10, 26, 39, 40].
In this section a two-Riccati equation solution is presented to the H1 problem for lifted
systems with controller causality constraints. The result is closely related to the Riccati-
equation based solution given in Section 3, and thus it also gives insight into the connection
between the two solution procedures. A particular feature of the approach presented here
is that lifting is applied separately to the periodic state feedback and estimation problems,
which need not have the same periods.
12
In this section, it is assumed that all hold elements are synchronized, i.e. there exist times
kTu such that the function H(kTu) = I in (3). This enables us to formulate a particularly
simple solution of the associated state-feedback problem.
4.1 A multirate state-feedback problem
Introduce the �nite-dimensionalNu-periodic discrete system associated with the multirate
system (4), cf. [10, 39],
~xe(kNu + l + 1) = ~Ae(hl; �l)~xe(kNu + l) + ~B1e(hl) ~wu(kNu + l)
+ ~B2e(hl; �l)~ud(kNu + l)
~zu(kNu + l) = ~C1e(hl; �l)~xe(kNu + l) + ~D12(hl; �l)~ud(kNu + l); (42)
l = 0; 1; : : :Nu � 1; k = 0; 1; : : :
where hl = �l+1 � �l, ~ud(kNu + l) := ud(kTu + �+l), ~xe(kNu + l) := [~x0u(kNu + l) u0(kTu +
�l)]0 := xe(kTu + �l) and ~wu(�) denotes a discrete-time disturbance in l2. The system
matrices are given by (22). The system (42) is the equivalent discrete representation of (1)
and (3) obtained by applying lifting to describe the sampled-data system in the intervals
[kTu + �l; kTu + �l+1] [5, 15, 36]. Thus, a sampled state-feedback controller K stabilizes the
sampled-data system (4) and achieves the H1-type performance bound J1(K) < if and
and only if it stabilizes (42) and the l2-induced norm from ~wu to ~zu is less than .
In this section we apply the lifting technique to solve the problem of �nding a causal
stabilizing Nu-periodic state-feedback controller for the periodic system (42) such that the
closed-loop system has the l2-induced norm from ~wu to ~zu less than . Motivated by the
multirate sampled-data problem, we consider the periodic discrete state-feedback problem
only. However, the multirate full-information problem can be solved in a similar way.
The Nu-periodic system (42) has the state ~xe(�). However, as the lower part of the state
~xe(�) consists of the control inputs, which are known to the controller, any Nu-periodic
state-feedback law ~ud = Ke~xe can be written as
~ud = K~xu; (43)
where K is a dynamic Nu-periodic controller, which includes the lower part of ~xe(�) as part
of its state.
Introduce the lifted signals
�wu(k) := [ ~w0u(kNu) ~w0
u(kNu + 1) � � � ~w0
u(kNu + Nu � 1) ]
0
�ud(k) := [ ~u0d(kNu) ~u0
d(kNu + 1) � � � ~u0
d(kNu + Nu � 1) ]
0(44)
�zu(k) := [ ~z0u(kNu) ~z0u(kNu + 1) � � � ~z0u(kNu +Nu � 1) ]0:
13
Assuming that H(kTu) = I, the associated time-invariant lifted system is given by
�xu(k + 1) = Au�xu(k) + B1;u �wu(k) + B2;u�ud(k)
�zu(k) = C1;u�xu(k) +D11;u �wu(k) +D12;u�ud(k) (45)
where �xu(k) := ~x(kNu) and
Au := ~A(Tu)
B1;u := [ ~A(�Nu� �1) ~B1(h0) ~A(�Nu
� �2) ~B1(h1) � � � ~B1(hNu�1) ]
B2;u := [ ~A(�Nu� �1) ~B2(h0)Hu(�0) ~A(�Nu
� �2) ~B2(h1)Hu(�1) � � � (46)
� � � ~B2(hNu�1)Hu(�Nu�1) ]
C1;u :=
2664
~C1(h0)~C1(h1) ~A(�1 � �0)
...~C1(hNu�1)
~A(�Nu�1 � �0)
3775
D11;u := [D11;u(p; q)]; p; q = 0; 1; : : : ; Nu � 1
D12;u := [D12;u(p; q)]; p; q = 0; 1; : : : ; Nu � 1
where
D11;u(p; q) :=
�0; if p � q~C1(hp) ~A(�p � �q+1) ~B1(hq); if p > q
D12;u(p; q) :=
8<:0; if p < q~D12(hp; �p); if p = q~C1(hp) ~A(�p � �q+1) ~B2(hq)Hu(�q); if p > q
(47)
and ~A(�), ~B1(�), ~B2(�), ~C1(�) and ~D12(�; �) are de�ned by (22) and (23). Note that for p > q,
~A(�p � �q) = ~A(hp�1) ~A(hp�2) � � � ~A(hq): (48)
Notice that the lifted system (45) has an n-dimensional state vector. This is due to the
assumption H(kTu) = I, i.e., that all inputs are updated at time instants kTu. If the state
of the lifted systems were de�ned at time instants where only part of the inputs are subject
to change, then the state of the lifted system would include those elements of u which are
not updated at these time instants, cf. the augmented state in (4).
From the construction of the lifted system (45) it follows that an Nu-periodic state-
feedback controller K for (42) can be characterized as a time-invariant full-information con-
troller KL := [Kx Kw] : l2 � l2 ! l2 for (45), such that
�ud(k) = ([Kx Kw ]
��xu�wu
�)(k) :=
264
(K~xu)(kNu)...
(K~xu)(kNu + Nu � 1)
375 (49)
14
By construction, the controller (43) stabilizes the periodic system (42) if and only if the
controller (49) stabilizes (45), and the l2-induced norm of the closed-loop system (42), (43)
is equal to the H1 norm of the closed loop (45), (49).
The problem of constructing a stabilizing periodic state-feedback controller for (42) which
achieves a speci�ed l2-induced norm bound can thus be reduced to an H1-optimal full-
information control problem for the time-invariant lifted system (45). However, the equiva-
lent H1 problem de�ned for (45) is not a standard H1 problem, because the requirement
that (43) should be causal imposes corresponding causality constraints on the structure of
the lifted controller in (49). In particular, partitioning Kw in compliance with �ud and �wd,
the direct coupling term of Kw should have zeros on and above the main block diagonal.
For future reference, it is convenient to introduce a connection between the Riccati-
equation solution in Section 3 and the lifted time-invariant system (45).
Theorem 4.1 Consider the multirate system (4) and the associated discrete Nu-periodic
Riccati equation (25). Let Hu(kTu) = I. Then
Se(kTu) =
�S 0
0 0
�(50)
where S is the symmetric positive semide�nite solution of the discrete algebraic Riccati
equation
S = A0uSAu + C01;uC1;u � E0
SM�1SES (51)
where
ES :=
�B01;uB02;u
�SAu +
�D0
11;u
D012;u
�C1;u
MS :=
�B01;uB02;u
�S
�B01;uB02;u
�0+
�� 2I +D0
11;uD11;u D011;uD12;u
D012;uD11;u D0
12;uD12;u
�(52)
Proof: The result follows from (25), Theorem 3.3 and the structures of the matrices in (22).
2
The following theorem gives a state-space solution to the causally constrained lifted full-
informationH1 problem associated with the multirateH1-optimal state-feedback problem.
Theorem 4.2 Consider the lifted system (44), (45) associated with the Nu-periodic system
(42). There exists a stabilizing full-information controller KL described by (49) such that
the associated Nu-periodic controller K is causal and such that the closed-loop system (45),
(49) has H1-norm less if and only if there exists a stabilizing solution S to the discrete
algebraic Riccati equation (51) such that
2I � ~B01(hl)S(�l+1) ~B1(hl) > 0; l = Nu � 1; : : : ; 0 (53)
15
where the matrices fS(�l)g are de�ned recursively according to (25) with the partition in (9),
and
Se(�Nu) :=
�S 0
0 0
�: (54)
In this case, a stabilizing causal Nu-periodic state-feedback controller K which achieves the
norm bound is given by
~ud(kNu + l) = �V �1S
(hl; �l)[ ~B02e(hl; �l)S
�e(�l) ~Ae(hl; �l)
+ ~D012(hl; �l)
~C1e(hl; �l)]~xe(kNu + l) (55)
l = 0; 1; : : :Nu � 1; k = 0; 1; : : :
where ~xe(kNu + l) = [ ~x0u(kNu + l) u0(kNu + l) ],
S�e (�l) := Se(�l)[I � �2 ~B1e(hl) ~B01e(hl)Se(�l)]
�1; (56)
and
VS(hl; �l) := ~B02e(hl; �l)S�e (�l)
~B2e(hl; �l) + ~D012(hl; �l)
~D12(hl; �l): (57)
Proof: By standard discrete H1 control theory, a necessary condition for the existence of
a full-information controller KL which achieves the H1 performance bound for the system
(45) is that the discrete algebraic Riccati equation (51) has a stabilizing symmetric positive
semide�nite solution. In this case, with �xu(0) = 0, we have the expansion [14]
k�zuk2l2� 2k �wuk
2l2=
1Xk=0
Lk( �wu(k); �ud(k)) (58)
where
Lk( �wu(k); �ud(k)) :=
��wu � �w0
u
�ud � �u0d
�0(k)MS
��wu � �w0
u
�ud � �u0d
�(k) (59)
and ��w0u
�u0d
�(k) := �M�1
SES�xu(k) (60)
Moreover, the contribution to the quadratic cost (58) from the discrete time instant k can
be expressed as
�z0u(k)�zu(k) � 2 �w0u(k) �wu(k) = Lk( �wu(k); �ud(k)) + �x0u(k)S�xu(k)
��x0u(k + 1)S�xu(k + 1) (61)
From the de�nitions of the signals �zu(k) and �wu(k), and the system equation (42), which
describes the system behavior between the discrete time instants kNu and kNu + Nu, we
have similarly the expansion
16
�z0u(k)�zu(k) � 2 �w0u(k) �wu(k) =
Nu�1Xl=0
[~z0u(kNu + l)~zu(kNu + l)� 2 ~w0u(kNu + l) ~wu(kNu + l)]
=
Nu�1Xl=0
Il( ~wu(kNu + l); ~ud(kNu + l)) + ~x0u(kNu)S~xu(kNu)
�~x0u(kNu + Nu)S~xu(kNu + Nu) (62)
where
Il( ~wu; ~ud) :=
�~wu � ~w0
u
~ud � ~u0d
�0MS (hl; �l)
�~wu � ~w0
u
~ud � ~u0d
�(63)
and �~w0u
~u0d
�(kNu + l) := �MS (hl; �l)
�1ES(hl; �l)~xe(kNu + l) (64)
The result of the theorem then follows from (58),(61) and (62) by standard discrete H1
control theory [4, 14, 31]. 2
Remark 4.1 Notice that the solution of the causally constrained H1 state-feedback problem
is obtained in terms of the same algebraic Riccati equation (51) which is associated with the
lifted system (45) and H1 problem without causality constraints. The constraints appear
only in the nonstandard form of the positive de�niteness conditions (53) and the controller
structure (55).
Remark 4.2 Theorem 4.2 gives a causal Nu-periodic controller which achieves the spec-
i�ed l2-induced norm bound for the system (42). The corresponding lifted time-invariant
controller KL in (49) can be recovered by applying lifting to the periodic controller (55). It
is easy to see that the lifted version of (55) is a static full-information controller of the form
KL = [Kx Kw], where Kw has zeros on and above the main block diagonal.
4.2 A multirate estimation problem
An analogous dual result is obtained for the multirate estimation problem. De�ne the �nite-
dimensional Ny-periodic discrete-time system associated with the multirate system (4),
~xy(iNy + j + 1) = ~A(sj)~xy(iNy + j) + ~B1(sj) ~wy(iNy + j)
~zy(iNy + j) = ~C1(sj)~xy(iNy + j) (65)
~yd(iNy + j) = ~C2(�j)~xy(iNy + j) + ~D21(�j)~vd(iNy + j)
j = 0; 1; : : :Ny � 1 i = 0; 1; : : :
17
where sj := �j+1��j, ~xy(iNy+j) := x(iTy+�j), ~yd(iNy+j) := yd(iTy+�j), ~vd(iNy+j) :=
vd(iTy+�j) and ~wy(�) denotes a discrete disturbance in l2. The system matrices are de�ned
by (23) and (24). The multirate H1-optimal estimation problem consists of �nding an
Ny-periodic stable causal estimator
zy = F~yd (66)
which achieves the performance bound
sup( ~wy;~vd)6=0
k~zy � zykl2[k ~wyk2l2 + k~vdk
2l2]1=2
< (67)
The time-invariant system obtained by applying lifting to (65) over the period Ny of the
sampling function is given by
�xy(i + 1) = Ay�xy(i) +B1;y �w(i)
�zy(i) = C1;y�xy(i) +D11;y �w(i) (68)
�yd(i) = C2;y�xy(i) +D21;y �w(i)
where we have de�ned �xy(i) := ~xy(iNy), ~w0 := [ ~w0y ~v0
d]0; and the lifted signals
�w(i) := [ ~w0(iNy) ~w0(iNy + 1) � � � ~w0(iNy + Ny � 1) ]0
�zy(i) := [ ~z0y(iNy) ~z0
y(iNy + 1) � � � ~z0
y(iNy +Ny � 1) ]
0(69)
�yd(i) := [ ~y0d(iNy) ~y0
d(iNy + 1) � � � ~y0
d(iNy +Ny � 1) ]
0:
The system-matrices of the lifted system are given by
Ay := ~A(Ty)
B1;y := [ ~A(�Ny� �1) �B1(s0) ~A(�Ny
� �2) �B1(s1) � � � �B1(sNy�1) ]
C1;y :=
2664
~C1(s0)~C1(s1) ~A(�1 � �0)
...~C1(sNy�1)
~A(�Ny�1 � �0)
3775
C2;y :=
2664
~C2(s0)~C2(s1) ~A(�1 � �0)
...~C2(sNy�1)
~A(�Ny�1 � �0)
3775 (70)
D11;y := [D11;y(p; q)]; p; q = 0; 1; : : :; Ny � 1
D21;y := [D21;y(p; q)]; p; q = 0; 1; : : :; Ny � 1
where
D11;y(p; q) :=
�0; if p � q~C1(sp) ~A(�p � �q+1) �B1(sq); if p > q
D21;y(p; q) :=
8<:0; if p < q�D21(sp); if p = q~C2(sp) ~A(�p � �q+1) �B1(sq); if p > q
(71)
18
where we have introduced ��B1(sj)�D21(�j)
�:=
�~B1(sj) 0
0 ~D21(�j)
�: (72)
The Ny-periodic estimator (66) is equivalent to a time-invariant lifted estimator FL : l2 !
l2 described by
�zy(i) = (FL�yd)(i) :=
264
(F ~yd)(iNy)...
(F ~yd)(iNy +Ny � 1)
375 : (73)
Thus, FL achieves the performance bound
sup�w6=0
k�zy � �zykl2k �wkl2
< (74)
for the system (68) if and only if F achieves the bound (67) for the system (65). The
multirateH1-optimal estimation problem can thus be described in terms of a time-invariant
H1-optimal estimation problem for the lifted system (68) subject to a causality constraint
imposed by the causality of (66) and the structure of (73). More speci�cally, the direct
coupling term of FL is required to have a lower block-triangular structure.
The following auxiliary result gives the solution to the periodic Riccati equation (30) in
terms of an algebraic Riccati equation associated with the lifted system (68).
Theorem 4.3 Consider the multirate system (4) and the associated discrete Ny-periodic
Riccati equation (30). Then Q := Q(iTy+�0) is the symmetric positive semide�nite solution
of the discrete algebraic Riccati equation
Q = AyQA0y+ B1;yB
01;y �EQM
�1QE0Q
(75)
where
EQ := AyQ
�C1;y
C2;y
�0+B1;y
�D11;y
D12;y
�0
MQ :=
�C1;y
C2;y
�Q [C01;y C02;y ] +
�� 2I +D11;yD
011;y D11;yD
021;y
D21;yD011;y D21;yD
021;y
�: (76)
The following theorem gives a state-space solution to the causally constrained liftedH1 es-
timation problem associated with the multirate H1-optimal estimation problem.
Theorem 4.4 Consider the lifted system described by equations (68) and (69) associated
with the Ny-periodic system (65). There exists a stable estimator FL described by (73) such
that the associated Ny-periodic estimator F is causal and such that the H1 performance
19
bound (74) is achieved if and only if the there exists a stabilizing solution Q to the discrete
algebraic Riccati equation (75) such that
2I � ~C1(sj)[I � Ld(�j) ~C2(�j)]Q(�j) ~C01(sj) > 0 j = 0; 1; : : : ; Ny � 1 (77)
where Ld is given by (14) and the matrices fQ(�j)g are de�ned recursively according to (30)
with
Q(�0) = Q: (78)
In this case, a stable causal Ny-periodic estimator F which achieves the norm bound (67) is
given by
xy(iNy + j + 1�) = ~A(sj)xy(iNy + j) (79)
xy(iNy + j) = xy(iNy + j�) + Ld(�j)[yd(iTy + sj) � ~C2(�j)xy(iNy + j�)]
zy(iNy + j) = ~C1(sj )xy(iNy + j)
Proof: The proof is analogous to the proof of Theorem 4.2.
In analogy with the state-feedback problem, the time-invariant lifted estimator FL asso-
ciated with (79) can be determined by applying lifting to the periodic estimator (79).
4.3 The multirate output-feedback problem
The causally constrained lifted state-feedback controller and estimator results of Theorems
4.2 and 4.4 can be combined to provide a lifting-based solution of the multirate output
feedback problem. Although the result can be stated for general periodic sampling and hold
functions, it will for simplicity be assumed that the sampling and hold functions operate
synchronously.
Theorem 4.5 Consider the multirate sampled-data system described by (1){(3). Assume
that the sampling and hold functions operate synchronously, i.e. Nu = Ny =: N and �l = �l,
l = 0; 1; : : : ; N . There exists a causal N -periodic controller K which stabilizes the sampled-
data system and which achieves the performance bound J (K) < if and only if the following
conditions are satis�ed:
(a) The conditions of Theorem 4.2 are satis�ed,
(b) the conditions of Theorem 4.4 are satis�ed, and
(c) �[S(�l)Q(�l)] < , l = 0; 1; : : : ; N � 1.
20
Moreover, when conditions (a){(c) hold, a stabilizing multirate controller which achieves
the H1 performance bound is given by
xe(kN + l + 1�) = Acl(hl; �l)xe(kN + l)
xe(kN + l) = xe(kN + l�)
+
�I
0
��Ld(�l)[~yd(kN + l)� C2d(�l)xe(kN + l�)]
~ud(kN + l) = �V �1S
(hl; �l)[ ~B02e(hl; �l)S
�e(�l+1) ~Ae(hl; �l)
+ ~D012(hl; �l)
~C1e(hl; �l)]xe(kNu + l) (80)
where Acl(�; �) is de�ned in (29), VS (�; �) is given by (57), and �Ld(�) is given by (19).
Proof: The �rst part of the proof follows directly from Theorems 4.1 { 4.4 and standard
H1 control theory. The equivalence of the controllers (17) and (80) can be shown in analogy
with the single-rate case [28, 37]. 2
Remark 4.3 Notice that Theorem 4.5 gives the multirateH1-optimal controller in terms of
the discrete algebraic Riccati equations of Theorems 4.1 and 4.3, which are the Riccati equa-
tions of the standard H1 problem associated with the lifted system. The causality relations
of the multirate system appear in the solution in the form of the nonstandard positive de�-
niteness conditions (53) and (77) and the coupling conditions in Theorem 4.5 (c). Theorem
4.5 gives a periodic controller which achieves the H1 performance bound for the multirate
system. The corresponding time-invariant lifted controller can be determined by lifting the
periodic controller (80) over the period N .
5 Example
The following simple numerical example is used to illustrate some of the results in this paper.
Consider the system
_x(t) = �0:8x(t) + w(t) + 2u(t)
z(t) =
�1
0
�x(t) +
�0
1
�u(t): (81)
Assume that there are two measurement devices that give sampled data with two di�erent
rates. Both devices measure the state, but the more frequently available measurement is
corrupted by a larger disturbance
yd1(is) = x(is) + 0:7vd1(is); i = 0; 1; 2; : : :
y2d(js) = x(js) + 0:1v2d(js); j = 0; 2; 4; : : : (82)
21
where s = 1:5, such that yd(is) = [yd1(is) y2d(is)]0 for i = 0; 2; 4; : : : and yd(is) = yd1(is) for
i = 1; 3; 5; : : :. The control signal
u(t) = ud(kh+); t 2 (kh; kh+ h] ; k = 0; 1; 2; : : : (83)
is updated with the period h = 2. The sampling function has the period Ty = 3 and the
single-rate hold function has the period Tu = 2. The period of the hybrid multirate system
de�ned by (81){(83) is T = 6.
By separate discretization of the feedback and the estimation problems we obtain a
time-invariant discrete Riccati-equation (25) associated with the feedback problem and a
2-periodic discrete Riccati-equation (30) associated with state estimation, cf. Theorem 4.3.
The latter one can further be lifted to a time-invariant equation. Direct application of the
lifting technique in order to construct a time-invariant description of the hybrid system
(81){(83) with a constant discretization interval would require the interval to be 0:5, and
the associated periodic discrete Riccati-equations would have the period 12.
The solutions S(t) and Q(t) to (10) and (13), respectively, together with the spectral radius
condition (16) are illustrated in Figure 1. Note that although the spectral radius condition
must be satis�ed for all times, it is su�cient to check the coupling condition at three points
during the period T , namely at fkh+; k = 0; 1; 2g, cf. Remark 3.2. Alternatively, we
may apply (34) which gives four coupling conditions at fis; i = 0; 1; 2; 3g. The optimal
cost is inf = 0:971 and the solutions to the discrete Riccati equations (25) and (30) are
S(kTu) = 0:507, Q(iTy + s) = 0:810 and Q(iTy) = 0:971.
6 Conclusion
A state-space solution to the multirate sampled-data H1 problem has been derived. The
solution is expressed in terms of a pair of discrete algebraic Riccati-equations together with a
set of matrix positive-de�niteness conditions and a set of coupling constraints. The result has
been derived both through a direct Riccati-equation approach and via the lifting technique.
Both methods generally lead to the same equations and the same controller. Although this
paper has focused on the H1 problem, a similar relationship between lifting and a two-
Riccati solution can be shown to hold in the multirate Linear Quadratic/H2 problem as
well.
The results of this paper provide a closed state-space solution to the lifted time-invariant
H1 problem with causality constraints. A particular feature of the solution is that the
algebraic Riccati equations are the same as those associated with the unconstrained lifted
H1 problem. The causality constraint of the multirate problem appear here as a set of
additional, nonstandard positive-de�niteness constraints and a set of coupling conditions.
22
References
[1] H. Al-Rahmani and G. F. Franklin. A new optimal multirate control of linear periodic
and time-invariant systems. IEEE Trans. Autom. Contr., 35(4):406{415, 1990.
[2] N. Amit and J. D. Powell. Optimal digital control of multirate systems. AIAA Guidance
and Contr. Conf., pages 423{429, 1981.
[3] M. Araki and K. Yamamoto. Multivariable multirate sampled-data systems: State-
space description, transfer characteristics, and nyquist criterion. IEEE Trans. Autom.
Contr., AC-31(2):145{154, 1986.
[4] T. Ba�sar and P. Bernhard. H1-optimal control and related minimax design problems.
A dynamic game approach. Birkh�auser, Boston, 2 edition, 1995.
[5] B. A. Bamieh and J. B. Pearson. A general framework for linear periodic systems with
applications to H1 sampled-data control. IEEE Trans. Autom. Contr., 37(4):418{435,
1992.
[6] M. C. Berg, N. Amit, and J. D. Powell. Multirate digital control system design. IEEE
Trans. Autom. Contr., 33(12):1139{1150, 1988.
[7] S. Bittanti, P. Colaneri, and G. D. Nicolao. An algebraic Riccati equation for the
discrete-time periodic prediction problem. Systems & Control Letters, 14:71{78, 1990.
[8] J. R. Broussard and D. P. Glasson. Optimal multirate ight control design. Proc. Joint
Autom. Cont. Conf., pages WP1{E, 1980.
[9] T. Chen and B. Francis. Optimal Sampled-Data Control Systems. Springer-Verlag,
1995.
[10] T. Chen and L. Qiu. H1 design of general multirate sampled-data control systems.
Automatica, 30(7):1139{1152, 1994.
[11] P. Colaneri and G. De Nicolao. Multirate LQG control of continuous-time stochastic
systems. Automatica, 31:591{596, 1995.
[12] A. Feuer and G. C. Goodwin. Sampling in Digital Signal Processing and Control.
Birkh�auser, Boston, USA, 1996.
[13] B. Friedland. Sampled-data control systems containing periodically varying members.
In IFAC World Congress, pages 361{367, Moscow, 1961.
[14] M. Green and D. J. N. Limebeer. Linear Robust Control. Prentice Hall, Englewood
Cli�s, N. J., 1995.
23
[15] Y. Hayakawa, S. Hara, and Y. Yamamoto. H1 type problem for sampled-data con-
trol systems { A solution via minimum energy characterization. IEEE Trans. Autom.
Contr., 39(11):2278{2284, 1994.
[16] H. Ito, H. Ohmori, and A. Sano. Stability analysis of multirate sampled-data control
systems. IMA J. of Math. Contr. and Information, 11:341{354, 1994.
[17] P. Kabamba and S. Hara. Worst-case analysis and design of sampled-data control
systems. IEEE Trans. Autom. Contr., 38:1337{1357, 1993.
[18] R. E. Kalman and J. E. Bertram. A uni�ed approach to the theory of sampling systems.
J. Franklin Inst., 267:405{436, 1959.
[19] H. Katayama and A. Ichikawa. H1-control with output feedback for time-varying
discrete systems. Int. J. Control, 63(6):1167{1178, 1996.
[20] G. M. Kranc. Input{output analysis of multirate feedback systems. IRE Trans. Au-
tomat. Control, 3:21{28, 1957.
[21] B. Lennartson. Periodic solutions of Riccati equations applied to multirate sampling.
Int. J. Control, 48(3):1025{1042, 1988.
[22] B. Lennartson. Combining infrequent and indirect measurements by estimation and
control. Ind. Eng. Chem. Res., 28:1653{1658, 1989.
[23] O. Lindg�arde and B. Lennartson. Performance and robust frequency response for mul-
tirate sampled-data systems. In Proc. Amer. Control Conference, Albuquerque, New
Mexico, June 1997 (To appear).
[24] S. Longhi. Structural properties of multirate sampled-data systems. IEEE Trans.
Autom. Contr., 39(3):692{696, 1994.
[25] D. G. Meyer. A parametrization of stabilizing controllers for multirate sampled-data
systems. IEEE Trans. Autom. Contr., 35(2):233{236, 1990.
[26] L. Qiu and T. Chen. H2-optimal design of multirate sampled-data systems. IEEE
Trans. Autom. Contr., 39(12):2506{2511, 1994.
[27] R. Ravi, K. M. Nagpal, and P. Khargonekar. H1 control of linear time-varying systems:
A state-space approach. SIAM J. Control and Optimization, 29(6):1394{1413, 1991.
[28] M. F. S�agfors and H. T. Toivonen. Optimal H1 and LQG control of asynchronous
sampled data systems. IFAC World Congress,, pages 337{342 Accepted to Automatica,
1996.
[29] M. F. S�agfors and H. T. Toivonen. The sampled-data H1 problem: The equivalence
of discretization-based methods and a Riccati equation solution. In Conf. Decision and
Control, pages 428{433, 1996.
24
[30] N. Sivashankar and P. P. Khargonekar. Characterization of the L2-induced norm for
linear systems with jumps with applications to sampled-data systems. SIAM J. Control
and Optimization, 32(4):1128{1150, 1994.
[31] A. A. Stoorvogel. The H1 Control Problem. A State Space Approach. Prentice Hall
International, 1992.
[32] W. Sun, K. M. Nagpal, and P. P. Khargonekar. H1 control and �ltering for sampled-
data systems. IEEE Trans. Autom. Contr., 38(8):1162{1175, 1993.
[33] W. Sun, K. M. Nagpal, P. P. Khargonekar, and K. R. Poolla. Digital control systems:
H1 controller design with a zero-order hold function. In Proc. 31st Conf. Decision and
Control, pages 475{480, Tuscon, Arizona, December 1992.
[34] G. Tadmor. H1 optimal sampled-data control in continuous time systems. Int. J.
Control, 56:99{141, 1992.
[35] H. T. Toivonen. Sampled-data control of continuous-time systems with an H1 opti-
mality criterion. Automatica, 28(1):45{54, 1992.
[36] H. T. Toivonen. Digital control with H1 optimality criteria. In C. Leondes, editor,
Control & Dynamic Systems, pages 215{262. Academic Press, 1995.
[37] H. T. Toivonen andM. F. S�agfors. The sampled-dataH1 problem: A uni�ed framework
for discretization based methods and Riccati-equation solution. Int. J. Control, 66:289{
309, 1997.
[38] A. W. Truman. An analysis of sub-rate control systems. Int. J. Control, 63(2):365{392,
1996.
[39] P. G. Voulgaris and B. Bamieh. Optimal H1 and H2 control of hybrid multirate
systems. Systems & Control Letters, 20:249{261, 1993.
[40] P. G. Voulgaris, M. A. Dahleh, and L. S. Valavani. H1 and H2 optimal controllers for
periodic and multirate systems. Automatica, 30(2):251{263, 1994.
25
0 1 2 3 4 5 60
1
2
t
S(t)
0 1 2 3 4 5 60
0.5
1
t
Q(t)
0 1 2 3 4 5 60
0.5
1
t
S(t)Q(t)
γinf2
Figure 1: The solutions S(t), Q(t), and the coupling condition S(t)Q(t) for = inf over
one period [0; T ] of the multirate system (81) { (83).
26