IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39, NO.

4, APRIL 1994 703

A Function Space Approach to Sampled Data

Control Systems and Trackmg Problems
Yutaka Yamamoto, Senior Member, IEEE

Abstract-This paper presents a new framework for hybrid Certainly the stability of a sampled-data control system for the
sampled-data control systems. Instead of considering the state plant ( 2 ) is determined by the behavior of (1) (and the discrete-
only at sampling instants, this paper introduces a function piece time controller), and this can be determined based only on the
during the sampling period as the state and gives an inhite-
dimensional model with such a state space. This gives the advan- information at the sampled instants. A drawback in discretizing
tage that sampled-data systems with built-in intersample behavior the continuous-time plant as above, however, is that it loses
can be regarded as linear, time-invariant, discrete-time systems. the intersample information (or makes it implicit, at least).
As a result, the approach makes it possible to introduce such time- Since the plant is continuous time, the overall performance
invariant concepts as transfer functions, poles, and zeros to the must be evaluated in continuous time, but this is difficult once
sampled-data systems even with the presence of the intersample
behavior. In particular, tracking problems can be studied in this we discretize the plant and make it discrete time. A typical
setting in a simple and unified way, and ripples are completely problem is that of ripples in servo control systems. In such
characterized as a mismatch between the intersample reference a situation, the system is subject to continuous excitation by
signal and transmission zero directions. This leads to the internal exogenous signals, and there can remain stationary ripples,
model principle for sampled-data systems.
even though the error tends to zero at sampled instants. One
can compute the intersample behavior (after the control system
I. INTRODUCTION design) via the modified 2-transform, but this is often very
T is well recognized that digital control provides various annoying due to the complexity of formulas [ 171.
I advantages over the usual time-invariant feedback controls. In view of this, recent investigations of sampled-data sys-
Deadbeat control, for example, makes a type of stabilization tems place more emphasis upon intersample behavior, and
that cannot be achieved with continuous-time linear time- there are now a number of investigations along this line: [5]
invariant feedback. We can also implement a much more and [3] for stability analysis; [3] and [ 141 for H2-optimization;
complex logic in control actions making use of the recent 191, [20], and [I] for HE-problem; and [7], [22], [191, 1151,
advances in computer technology. For example, it is recog- and [24] for tracking.
nized that multirate sampling and/or generalized hold functions The study of sampled-data systems with built-in intersample
provide much greater capability in control; see, for example, behavior induces another technical difficulty. Even though the
(continuous-time) plant and the (discrete-time) controller are
[81, [lo].
On the other hand, sampled-data systems give rise to a dif- both time-invariant, the underlying time sets are different, so
ficulty not encountered in the classical situation. In designing when combined together to form a hybrid sampled-data control
sampled-data systems, one usually designs the continuous-time system, the resulting system is not time-invariant and only
system and then designs a discrete-time controller over the periodically time-varying. This mixture of two types of time
discrete-time domain. For example, the fundamental work of sets makes the analysis of sampled-data systems technically
[12] shows that as far as the regulation of initial states is very awkward. In particular, it makes the classical state space
concemed, one needs only consider the discrete-time system formalism ineffective.
To remedy this, we introduce a new framework for sampled-
h data systems. Since the fundamental issue lies in incorporating
Wk+l = eAhWk + e A r B U k d7,
the intersample behavior, we take the basic idea of taking
the full histories of each sampling period of input/state/output
Yk = c w k (1)
functions as input/state/output vectors and then derive linear,
where h is the sampling period and the matrices ( A , B , C) time-invariant, discrete-time transition rules. The difference
are the system matrices of the linear time-invariant plant here is that since we take functions as states, inputs, etc.,
the system becomes injinite-dimensional, but on the other
+
i ( t )= Az(t) Bu(t), hand it has the advantage that continuous-time systems can
be viewed as linear, time-invariant, discrete-time systems.
y(t) = C z ( t ) . (2)
Thus it is particularly suitable for the study of sampled-data
Manuscript received September 13, 1991; revised April 23, 1993. Recom- systems in that both digital and analog components can be
mended by Past Associate Editor D. F. Delchamps. placed into the unified framework of linear, time-invariant,
The author is with the Division of Applied Systems Science, Faculty of
Engineering, Kyoto University, Kyoto 606-01 Japan. discrete-time framework. This technique is a generalization
IEEE Log Number 9215718. to continuous-time systems of lifting employed by 1131 for
0018-9286/94$04.00 0 1994 IEEE
704 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39. NO. 4, APRIL 1994

discrete-time periodic systems. It "lifts" the original system

to that considered in a discrete-time yet with larger in- I ' ' ' I

put/state/output spaces. Unlike the discrete-time case [ 131,

however, the present case involves infinite-dimensionalfunc-
tion spaces due to the continuous-time nature of the plant. 0
I I l

This new framework of lifting continuous-time systems

has been introduced to control theory by [20], [24] and by
[2], [11, independently. While we use infinite-dimensional
state space, [20], [2] and [l] use a finite-dimensional state
space. This difference actually induces some difference in
formulas in the lifted system, and it is to be investigated
in the future as to which is more advantageous. While [20]
Fig. 1. Mapping S .
and [l] studies H"-type problem for sampled-data systems,
we are here concerned with tracking problems, in particular,
characterization of ripples. we see that C [ O , co) and the sequence space X" (the count-
The paper is organized as follows: Section I1 introduces ably many direct product of X) are placed in one-to-one
a function-space valued sequence space and 2-transforms correspondence. This mapping S "lifts" the continuous-time
there. We then define the lifting of the continuous-time linear signal cp(t)to an element { c p k ) residing in the sequence space
system in Section 111. After giving some facts on stability, X". The space X" can also be identified with the space of
stabilizability, and transfer functions in Section IV, we discuss formal power series X[[z-']] with coefficients in X via the
ripple-free tracking conditions in Sections V and VI. Section 2-transform
V deals with the case of tracking signals generated by a simple 2 : X" + x[[z-l]]
pole. The key feature here is the notion of transmission zeros
and their associated zero directions. This directional vector
(naturally a function in the present setting) gives a proper (5)
intersample tracking signal, and roughly speaking, ripple-free k=l
tracking occurs when and only when this direction coincides With slight abuse of notation, we will also denote Z(cp) for
with the intersample behavior of the reference signal. Section Z(S(cp)) when cp E C [ O , m); it may also be denoted by @.
VI deals with the general case: Here the tracking problem for Note that it is a function of 0.
sampled-data systems becomes more delicate and interesting in Let us introduce the sampling and hold operations. Let S h
that combination of a digital compensator and a hold element be the delta function placed at h, i.e., (Sh, c p ) = cp(h), or
yields a continuous-time internal model, so that one can often Shcp = cp(h)for cp E X. The sampling operator S is identified
split the internal model of the exogenous signal to the digital with S h acting on X . It has a natural extens_ion_:
X" + R" (or
and analog parts. A necessity and also a sufficiency (under a S : C [ O , co) ---t W" with identification S : C[O,m) E X")
mild assumption) of the internal model principle is obtained. by
This paper is the full journal version of the conference paper
[24], in which the basic framework and preliminary tracking = {&(PIE1
S({cpli)): ={cpk(h))L. (6)
results were presented. The hold operator 3-t: W" + X" is defined by

11. SEQUENCE SPACE AND 2-TRANSFORMS 3-t({x/C)):= {vk(e)), cpk+l(~) = H(@k, 0 < Ih
(7)
Let h be a fixed sampling period. Our basic idea is to regard where H ( 0 ) is a fixed hold function. We remark that
a trajectory z ( t ) as the sequence of functions { z k ( B ) ) ~ . l Z(S(cp)) = p(kh)zPk, so it agrees with the usual
defined by 2-transform of the sampled data {cp(kh)). We also remark
5k(8)3 = x ( ( k - 1)h e)+ (3)
that

and give a discrete-time state transition rule of zk. To rigor-

ously ictroduce this, we need a few preliminaries.
Let C[O,co) denote the space of $1 piecewise continuous so that the sample-hold operation induces a unit-time delay
functions on [0, co). Similarly, let C(0, h] denote the space (forward shift) operator.
of all such functions p(t) on (0, h] with finite right limit at 0. We note that the space X admits an algebra structure with
The latter space will be denoted X in the sequel. X is clearly respect to convolution
a Banach space with supremum norm 0
(cp * !b)(O= cp(0 - T)!b1(7) d7. (9)
llcpll: = SUPo<tJ<hl'P(m
Defining the mapping s as This naturally induces an algebra structure to X[[z-']], which
we freely use in sequel. We remark that the multiplication
S:C[O,co)+X": cpH{cpk)g1: of z and 2 - l to an element in X[[z-l]] acts as the one-step
pk(e) = cp((lc - i ) h +e), o < e I h (4) backward and forward shift operators, respectively.
YAMAMO'IU SAMPLED DATA CONTROL SYSTEMS 705

111. FUNCTION SPACE MODEL

Using the lifting S : p(t) H { p k } , we can derive a discrete-
time, time-invariant system equation for a continuous-time
system. This has the obvious advantage that the underlying
time set can be made the same for digital and analog elements. I I
We start by giving the discrete-time, time-invariant model Fig. 2. Hybrid closed-loop system E,/.
for a continuous-time system. Let (A,, B,, C,)be a given
continuous-time system
+
k ( t ) = A , z ( t ) B,u(t),
y ( t ) = Cc2(t), U E R", 2 E R", y E RP. (10)
Let Z k ( 6 ) E X",Y k ( e ; ! E X P , U k ( e ) E X" be the functions
obtained by applying S to 2, y , U, respectively, i.e., Z k ( e ) =
+
z ( ( k - 1)h e), (cf. (4)), etc. Suppose that system (10) is
at time t = kh. Then the past h second history of the state t=kh t=(k+l)h
is Z k ( e ) E X". If input U k + 1 ( 8 ) E X" is applied on the Fig. 3. Signal timing.
+
interval (kh,(k l)/h], then the state trajectory % k + 1 ( 8 ) and
the output trajectory Y k ( 8 ) follows the transition rules

zk+1(8) =eAcexk(h) + 1
e
eA'(e-T)B&k+l(T) d7,
that the closed-loop system, denoted C,l in the sequel, obeys
the equations:

yk(8) =ccxk(e), 0 < 6 5 h. (11) wk+l =Adwk +Bdek(h),

Introducing the operators uk+l(e) H(e)Cdwk- (15)

F : X" + X" : .(e) H eAcez(h),

G : X" -+ X" : .(e) H 1 e

eAc(e-T)B,u(T) dT, (12) zk+l(e) = eAcesk(h)

'&(e) = C c X k ( e ) ,
+
0<8
Le eAc(e-')Bcuk+l(T)

5 h.
d7,

H : x" -, x~: .(e) H ccX(e)

(16)

(11) can be written simply as and

xk+l =Fxk +GUk+ir ek(e> =Tk(e) -yk(e).

Yk =HXk. (13)
The one-unit time shift in the output equation in (15) is a result
It is clear that (10) and (11) Or (13) give Precisely the Same of hold operation (see Fig. 3 for the precise correspondence
input-statehtate-output correspondence. There is, however, a of timing of each signal). We then obtain the following model
point that calls for more attention. On the right-hand side of for cC1
(13), the input term is ?&+I, not Uk. Therefore, in the strict
sense, the quantity l k + l ( e ) does not satisfy strict causality.
To remedy this, one can introduce a new state variable

and satisfies causality. In what follows, however, we will

use mainly (13) because the unit-time delay induced by the
hold operation just cancels this effect of U k + l . Also, when a Y k ( e ) = [O [2y8)] (19)
continuous-time system C, is viewed as a discrete-time system
under the lifting here, we will denote it by S(&) Or by E,. where B(0) is given by
Consider the hybrid control system depicted in Fig. 2.
Here E d and C, denote discrete-time and continuous-time
systems ( A d , Bd,_Cd) and (A,, B,, C,), respectively. Under
the identificationS introduced in the previous section, we note
B ( B )=:
I" eAC('-')B,H(.r) d ~ . (20)

that at t = kh, the sampled values of signals e k , 2 k , Y k , T k , We will denote the first operator on the right-hand side of (17)
etc., are e k ( h ) , Z k ( h ) , y k ( h ) , T k ( h ) , respectively. It fOllOWS by A.
706 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39, NO. 4, APRIL 1994

Remark 3.1: There are now several different frameworks these vectors are found, (22) can be trivially solved as
incorporating intersample behavior to sampled-data systems.
Francis and Georgiou [5] and Chen and Francis [3], [4] gave a
finite-dimensional,periodically time-varying model. Hara and
Kabamba [9], [l 11 used a mixed discrete-timekontinuous-time
model. A similar lifting as presented here is also employed by This completes the proof. 0
[20], [2] and [l] with a finite-dimensional state space. In this We now study the internal stability of C,I. We say that C,Z
model, xk:= z ( k h ) is taken as the state, and as a result one is (asymptotically} stable if
needs an infinite-dimensionaldirect transmission term from U
to y to describe intersample behavior. On the other hand, this is
built into 2 k (e) in our framework. While finite-dimensionality
of the state has obvious merit-typically the input operator is The norm 11 . 11 is the one induced from that of X". This
of finite rank-our framework also has the (dual) advantage condition holds if and only if A k z decay to zero uniformly
that the output equation is simply expressed as yk(e) = for all x in the unit ball of X".
Ccxk(e), whereas an infinite-dimensional operator is needed The following theorem is well known (e.g., [3]), stating that
in the former approach. The present framework also naturally the internal stability of C,I is determined by the behavior at
exhibits the nature of sampling: it is an evaluation operator the sampled instants, but is given in the time-invariant operator
S h (delta function). Of course, these two realizations yield theoretic setting here.
entirely the same input/output behavior and hence the transfer Theorem 4.2: The closed-loop system C,1 (17)-(19) is in-
function. Actually, in the definition of zeros (Definition 5.2), ternally stable if and only if the matrix A0 in (21) is a stable
they yield almost the same relations. Again the difference lies matrix.
in whether we allow infinite-dimensionaloperator either in the Proof: Since every nonzero spectrum is an eigenvalue,
state or in the output equation. the condition is clearly necessary. Conversely, suppose A0 is
a stable matrix. Theorem 4.1 implies that the spectral radius
IV. SPECTRUM,STABILITY, of rU(A)of A coincides with the magnitude of the largest
AND TRANSFERMATRICES
STABILIZABILITY eigenvalue of (21), which is less than one by hypothesis. Recall
the well-known formula for spectral radius ([18, page 2351)
Since our lifted model (17)-(19) is, in principle, infinite-
dimensional, we need to investigate some basic system prop-
erties. Let us start by characterizing the spectrum of system
Ccl.
Theorem 4.1: The spectrum of system C,l (17)-(19) is the Take any E > 0 such that rU(A)+ E < 1. Then
union of (0) and the spectrum of the matrix

for all sufficiently large k, and hence the conclusion follows.0

Furthermore, except 0, they are all eigenvalues. In other words, Theorem 4.2 shows that if the sampled digital sys-
the spectrum of the hybrid system is entirely determined by tem ( e A c h ,B(h),C,) is stabilized by a digital controller
the evaluation at 0 = h. (Ad, Bd, Cd) as a finite-dimensional discrete-time system,
Proof: Let us note that the operator A in (17) is a then the closed-loop hybrid system C,l (17)-(19) is also
bounded operator. It is clearly a finite-rank operator, so it internally stabilized. This agrees with the classical knowledge,
is also compact [18, Theorem 4.181. Then A is not contin- and it shows the consistency of the present framework.
uously invertible because X is infinite-dimensional so that We now discuss the preservation of stabilizability. In doing
zero belongs to the spectrum o(A), and any other point in so, we make the following standard assumption:
the spectrum is an eigenvalue ([18, Theorems 4.18, 4.251). To Assumption A (Spectrum Nondegeneracy Assumption}:
compute the nonzero eigenvalues, suppose that X # 0 and None of the eigenvalues of A, with nonnegative real parts
(w',~ ( 0 ) ' ) 'satisfy differ by 2k7rj/h7 k # 0.
Now consider the sample-holded state feedback

This yields the feedback system

To solve these equations, first set B = h
(XI - A d ) W f BdCcx(h) = O
-B(h)Cdw + (XI - e A c h ) x ( h=) 0.
and as in Theorem 4.2, system (24) is stable if and only if
These have a nonzero solution if and only if X is an eigenvalue eAch + B ( h ) K is a stable matrix. Therefore, we readily obtain
of (21) and (w'?~ ( h ) ' )is' a corresponding eigenvector. Once the following theorem.
Y A M A M m SAMPLED DATA CONTROL SYSTEMS 707

Theorem4.3: The system (11) is stabilizable by sampled operator of system C,t. Combining (27) with (28), we see
state feedback (23) if the original continuous-time system +
x ( h ) = ( l / X ) { B ( h ) w o eAchxo} = 20. Therefore, we have
(Ac, B,) is stabilizable and Assumption A is satisfied. the sampled-transfer matrix
Before closing this section, we give some facts on transfer
matrices. Let ( F , G , H ) be a given system over X : 2(h) = P(h) - [O C,]

This shows that, at sampled instants, the closed-loop transfer

Using the algebra structure of X [ [ z - ’ ] ] ,especially { z X } k = function is obtained by composing the discrete-time (finite-
X k + l , we can compute the 2-transform of both sides of (25) dimensional) transfer functions in the usual way, and this
to obtain guarantees consistency of the present method with the classical
approach.

where denotes 2-transform. Allowing formal division for v. ZEROS AND TRACKING
( z I - F)-’, H ( z I - F ) - l G agrees with the 2-transform of One of the problems in digital control is that it may induce
the sequence ripples. If we focus our attention only on the discrete-time
H G , H F G , . .. , H F k - ’ G , . model describing the behavior at the sampled instants, it is
difficult to analyze ripples, since during the intersample peri-
and multiplication of H ( z I - F ) - l G with U gives the ods the system works as an open-loop system. This makes it
output in the 2-transform domain. On the other hand, by the difficult to obtain the internal model principle for servo control
standard technique of considering characteristic solutions of problems as obtained by [6] (see [7], [22] for some extensions
type X k = X k - b 0 ( e ) ,one can consider substituting a complex to the case of ripple-free tracking). Taking advantage of the
number X to (26), and using the fact that Z X k = X k + l = X X k , present framework, however, we can give a much clearer view
it is obvious that the solution is obtained by solving on the ripples. In this section, we first study the case where
the reference signals are generated by a simple pole and give
a ripple-free tracking condition for this case. The general case
where the reference signals are generated by systems with
In this case ( X I - F)-’ appearing in the solution is the resol- multiple poles is studied in the next section.
vent operator. Therefore, by this formal agreement, the transfer As it turns out, making use of the fact that the present
function (26) may be computed by formally substituting z into approach allows transfer matrices, we can naturally talk about
H(XI - F ) - l G and vice versa. zeros of the transfer matrix, and then the ripple behavior can
We now compute the transfer function of C,t (17)-(18). be clearly understood as a result of the mismatch between a
This will be given by solving directional function intrinsically associated with a transmission
zero and the intersample function. To see the idea, let W ( z )be
a transfer matrix from the reference signal r to the error e of a
finite-dimensional,discrete-time closed-loop system. Suppose
that the tracking signal r is w/(z - A) where v is a vector
giving the direction into the input channels. It is well known
e(0) = r ( 6 ) - Ccx(8). that the error e tends to zero if and only if
Suppose X belongs to the resolvent set of matrix A. (21). W(X)v = 0.
Let (tu;, xb)’ be the solution of (XI - Ao)(wb, x;)’ =
That is, not only W ( z )has X as a transmission zero, but also
((Bdr(h))’,0)‘: should admit w as a directional vector associated to this zero A.

[r:] XI-& Bdc,

= [-B(h)cd X I - e A c h ]
The second row of (27) yields
-1
Bdr(h) .
0 [ ]
Precisely the same result holds in the present situation, only
(28) that vector v must be replaced by a function .(e). We start
with the following lemma.
Theorem 5.1: Let Wer(z)be the closed-loop transfer ma-
1 trix operator of c , ~
from the reference r to the error e given
A
+
.(e) = - { B ( 0 ) w o eAc8xo}. by (29). Suppose it is stable. Then its response to the input
Hence {Xk-lw(e)}, 1x1 2 1, asymptotically approaches

In particular,
1) there is no stationary ripple if and only if We,(X)w( S)
0;
This correspondence P(z) H e(.) (or $ ( z ) with suitable 2) when X = 1, the stationary ripple is given by
changes) will be referred to as the closed-loop transfer matrix We,(X)v( 0);
708 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39, NO. 4, APRIL 1994

3) if 1x1 > 1 and if WeT(X)w(8) $ 0, then the response (33). The last row of (33) yields
will always diverge.
Proof: Since Wer(z) is stable, we can disregard the
BdC,Z(h) + BdW(h) = 0
initial state response. Expand Wer(z)(w(8)/(z - A)} as so that we must have
(XI - Ad)W = 0, (34)

According to Theorem 4.2 and formula (29), the second term

-B(8)Cdw + XZ(O) - eAcez(h)= 0. (35)
R ( z ) is analytic outside the unit disk { z ; IzI < 1). Therefore, Existence of a nonzero solution to (34) and (35) is a necessary
the response (32) asymptotically approaches X"'Wer(X)t.(8>. iind sufficient condition for X to be an transmission zero of
This readily yields the conclusions. 0 Wer(z) with w(8) = -Ccx(8) an associated zero direction
In summary, function. Now (34) and (35) can admit a solution in the
1) asymptotic tracking to X L - l w ( 8 ) is possible only when following way:
wer(x)w(e) 0; 1) there exists a solution (w', x(8)')' with nonzero w, or
2) when this condition does not hold, and if 1x1 > 1, ripples 2) a solution of type (0, z(8)')' with .(e) $ 0 exists.
always diverge. In the first case, w is an eigenvector of -and X is a pole
The second property is suggestive. In practice, what we can of E d . In the second case, X is a pole of C, and z(8) is
implement precisely are almost exclusive_ly digital devices. an eige_nfunction. This argument also shows that a pole of,
Then, unless the continuous-time plant C, possesses X as E d or C, always yields a transmission zero of Wer(z)under
a pole, the tracking condition Wer(X)w(0) 0 is hardly Assumption B. Therefore, we have the following proposition.
satisfied, so that the intersample ripples always diverge; see Proposition 5.3: Let Wer(z) be the closed-loop transfer
also [16]. (Tracking at sampled points is still possible by a matrix operator from T to e of E,(, and let E d and 3, be
digital compensator.) This suggests that from the engineering its digital and continuous parts, respectively. Assume closed-
point of view, tracking to e p t , Re p > 0 is impractical, and the loop stability. Then, under Assumption B, unstable poles of C d
only practical cases are tracking to steps, ramps, and sinusoids. or Cc induce a transmission zero of Wer(z)and vice versa.
In what follows we specialize Theorem 5.1 to the particular This proposition says nothing about the zero directions. To
form of (29). We start with the definition of zeros. We first elaborate more upon this, we now confine ourselves to the case
make the following assumption throughout. of zero-order hold H ( t ) 1 and r ( t ) = e%o, R e p 2 0.
Assumption B: Consider the closed-loop system C,I in Fig. The following theorems may be regarded as an internal model
2. Let C ( z ) be the transfer function of E d , and P ( s )the transfer principle for the special case of tracking to signals generated
function of E,, with lifted (discrete-time) transfer function by a single pole p.
P ( z ) of the lifted system Cc of E,. We assume that We first consider the tracking to steps: p = 0.
for any unstable pole X of C ( z ) , no w E X satisfies Theorem 5.4: Consider the closed-loop system C,I in Fig.
@ ( X ) w = 0, and 2. Assume that H ( t ) 1, and and C, are stabilizable and
for any unstable pole p of P ( s ) ,ep'" is not a transmission detectable. Consider the tracking to the step reference signal
zero of C(z). (represented in the continuous-time).
Since the poles of p(z) are precisely e p h with p being T(t) = 210.
poles of P(s), this assumption simply requires that-there be
no unstable pole-zero cancellation between E d and E,. Assume the closed-loop stability. Then tracking without rip-
Definition 5.2: A complex number X is an invariant zero ples is possible either by incorporating 0 into Cc or by
and a function w(8) E X" is an associated zero direction of incorporating one into E d as a pole.
the closed-loop system (17)-(18) if there exists (w', z(8)')' Proof: The 2-transform of r ( t ) is given by VO/(Z - 1).
Put w(8) E WO. Let us first consider the case where zero is

] [$;I
such that
XI-& BdCc6h Bd6h a pole of E,. From (33), w(8) can be a transmission zero
direction if and only if
-B(8)Cd X I - eAce6h 0 =o. (33)
[ O C C I Cc.(8) +.(e) = 0 (36)
The operator on the left-hand side is a clear analog of the ( I - Ad)W = 0 (37)
system matrix in the finite-dimensional case. The difference -B(8)Cdw +.(e) - eAcez(h)= o (38)
here is that the matrix depends on the intersample parameter
8. We call this operator the closed-loop system operator. for some w, ~ ( 8 )Let
. zo be an eigenvector of A, correspond-
Under Assumption B and closed-loop stability, unstable X ing to 0. Then (36)-(38) can be easily solved as
(i.e., 1x1 2 1)is an invariant zero with associated zero direction .(e) = -cczo, = 0, .(e) = 20.
w(8) if and only if
In this case W O = -C,Q gives an allowable trackable direc-
wel.(x)w(e) =o tion. (Note that C,ZO# 0 by the detectability of ( A , , B,).)
i.e., it is also a transmission zero. Therefore, we can identify We next consider the case where one is a pole of but not
these two notions for tracking problems. Let us start by solving of 2,. As shown above, w(8) is a transmission zero direction
YAMAMOM: SAMPLED DATA CONTROL SYSTEMS 709

if af?d only if (36)-(38) are satisfied. Since one is not a pole so that .(e) above cannot contain the mode of e @ . Therefore,
of E,, w must be an eigenvector corresponding to one, and w(0) = Ccx(e)cannot be a zero direction associated to A, so
this in turn yields a unique solution stationary ripples exist.
It remains to show that trackicg without ripples is possible
.(e) = {eAce(I- e A C h ) - l B ( h+) B(e)}Cdw by implementing X as a pole of E,. In this case we can solve
and .(e) = -Ccx(e) is a corresponding trackable direction. (39)-(41) as
This x ( 0 ) is indeed a constant because 1
.(e) = -ccx(e),w = 0, .(e) = -eAcezo
X
%(e
Ace
(I - eAch) - l B ( h ) + B(0)) where 50 is an eigenvector of eAch.Now in view of Assump-
- eAce(I- eACh)-lA,B(h)+ eAceBc tion A, 20 is also an eigenvector of A, with respect to p. This
= eAce(l- eACh)-’(eAch- I ) B , + eAceBc implies eAcexO = e@xo, and hence
1
= 0. .(e) = --e@C,zo
X
Finally, if zero is not a pole of C, and one is not a pole of and such .(e) gives a trackable direction. (In particular, W O =
Ed. then one cannot be a transmission zero of the sampling -(l/X)Cozo.) Note also that C,xo # 0 by the detectability of
time closed-loop transfer matrix (30). Hence by the discrete- (A,, Bc). 0
time internal model principle, the closed-loop system does not Remark 5.6: Related results have been studied by [7], [22],
track the step reference as above. 0 etc. We note, however, that the present characterization in
We next consider the tracking to r ( t ) = e%O, R e p # 0. terms of poles and zeros has been made possible via the notion
Since the hold element is a zero-order hold, ripple-free tracking of transfer matrices resulting from lifting.
is possible only by incorporating p into C, as a pole and not so
by incorporating a digital internal model l / ( z - e P h ) into Ed.
VI. TRACKING AND l% INTERNAL MODELPRINCIPLE
Theorem 5.5: Consider the tracking to the reference signal
(represented in the continuous time) The results in the previous section show that tracking to
signals generated by a single pole can be well described by
r ( t ) = ePtwo, R e p 2 0 poles and zeros. In particular, when the tracking signal is
v(O)/(z - 1) and if the internal model is in the digital part,
under the same hypotheses on E,l and H ( t ) as in Theorem then the only trackable intersample signal .(e) is necessarily a
5.4. Assume p # 0. Then tracking without stationary ripples constant function for the case of zero-order hold; conversely, it
is possible only by incorporating p into Cc as a pole. is possible to track step functions by incorporating the pole one
Proofi The signal r ( t ) is represented as Sr = into the digital compensator Cd. This situation occurs because
{Ak-le”ewo}gl in the notation in Section 11, and its the cascade connection of l/(z - 1) and the hold element
2-transform is given by somehow works as a continuous-time internal model l/s. This
e@ is not, however, well described in the previous section.
-210. When the tracking signal is generated by a repeated pole, the
2-X
situation is more complicated and cannot be easily described
Put ~ ( 0 )=: e@wo. To prove that ripple-free tracking is not by simple pole-zero arguments. For example, take the ramp
possible by incorporating A:_= eph into C d , suppose that X signal r ( t ) = t. The 2-transform (defined in Section 11) of
is a pole of Cd but not of E,. As shown above, .(e) is a this signal is
transmission zero direction if and only if h e
+-
( 2 - 1)’
cc+) +.(e) =o (39) ( 2 - 1)2

( X I -A ~ ) w
=0 (40) Since this is an output of a continuous-time plant 1/s2, how-
+
-B(e)cdw h ( e ) - eAcex(h)= o (41) ever, neither 1 / ( ~ - 1 )nor
~ O/(z-l) can appear independently
as the output of this plant. Therefore, we cannot separate
for some w, .(e). Since x is not a pole of 5,, this implies the treatment of simple and double poles in such a case.
that w must be a (nonzero) eigenvector of Ad corresponding This situation is quite different from the standard situation
to A, and we have a unique solution in the internal model principle for the usual finite-dimensional
1 systems and requires a more elaborate treatment.
= -{eAce(XI
X +
- eAch)-’B(h) B(e)}Cdw. (42) Obtaining a tracking condition for the general case in the
sampled-data systems has attracted recent research interest:
Since X is not an eigenvalue of eAch, p is not an eigenvalue Using a mixed continuous-timddiscrete-timemodel, Hara
of A,, so that eAce does not contain the mode of e@. Since and Sung [19] derived a state-space necessary and sufficient
H ( t ) 1, we have condition for tracking. A geometric approach was taken by

B(8) =
e
1
eAcTBcdT
Kawano et al. [ 151 to get a necessary and sufficient condition.
Somewhat earlier than these, Franklin and Emami-Naeini [7]
gave an internal model principle in a more classical setting,
7 10 IEEE TRANSACTIONS ON AUTOMATIC CONTROL,VOL. 39, NO. 4, APRIL 1994

m
and Urikura and Nagata [22] gave a geometric condition for
the case of deadbeat tracking. While they are more of state-
k=l
space nature, we here look for a condition in the frequency
domain.
To see why the standard machinery does not work, let us
write the loop transfer function of Fig. 2 as P ( z , O)/q(z).
Here the denominator q ( z ) is independent of 0 by (29). This = e-hs . .&[2[4][Z]]1z=ehs.
means that if we consider intersample tracking, its continuous-
The interchange of integral and summation in the second
time behavior is reflected only upon the numerator P ( z , e ) line is justified by the absolute convergence of the Laplace
and not the denominator q(z). On the other hand, the usual transform. 0
treatment of the internal model principle for finite-dimensional Let us now see the finite Laplace transforms of the transfer
systems expresses it in terms of divisibility between the two functions of digital and analog components in Fig. 2. Let
denominators of the loop transfer function and the exogenous C ( Z ) = C~(ZI - Ad)-'Bd and P ( s ) = CC(SI - AC)-'B,
signal generator. This is clearly not suitable for the present be the discrete-time and continuous-time transfer matrices
case where the intersample information is represented by the of Ed and E,, respectively. By Lemma 6.2, the transfer
numerator.
function of the continuous-time plant in the sense above
To remedy this situation, it is desirable to recover the is ehsP(s).The finite Laplace transform L h [ ~ ] ( sof) the
continuous-time transfer function to prove the internal model hold function is an entire function of s by the Paley-Wiener
principle. To this end, let us first introduce the notion of finite
theorem [21]. The digital part C ( z ) becomes C(ehs) by the
Laplace transforms. finite Laplace transform. Therefore, the loop transfer function
Dejhition 6.1: Let cp be any function or distribution on the
(in continuous-time) becomes the product of these three:
interval [0, h].Thefinite Laplace transform, denoted &[(PI is G(s):= eh"C(eh")Lh[H](s)P(s). This falls into the category
defined by of pseudorational transfer functions studied in [23]. Roughly

~ h [ c p l ( s=
): 1 h
e-*ecp(e)do.
speaking, a transfer function W ( s )is said to be pseudorational
(43) if W ( s )admits a fractional representation W ( s )= p ( s ) / q ( s )
such that p ( s ) and q(s) are Laplace transforms of distributions
The integral must be understood in the sense of distributions with compact support in (-00, 01. (There is one more technical
if cp is a distribution. condition which does not concern us here [23].) In view of the
Note that, in view of the well-known Paley-Wiener theorem facts eh* = L[b-h](s), s = .C[S'], we see immediately that our
[21], Lh[cp](s)is always an entire function of exponential type. loop transfer matrix G is pseudorational.
The 2-transform of a function 4 on [0, m) and its Laplace Since C ( z ) and P ( s ) are both rational, they admit coprime
transform can be related by the following lemma. factorizations over polynomials (in z and s, respectively).
Lemma 6.2: Suppose that 4 satisfies the estimate Also, C h [ H ] ( sis) itself the Laplace transform of a distribution
with compact support in [0, h], and it admits a coprime
Il#'ll[kh, ( k + l ) h ] 5 CePk (44) factorization eh*Lh[H](s)/ehs (this looks odd, but eh* does
not appear as a common factor after multiplication with
for some C , p > 0, where I(f$l([kh,( k + l ) h ] is the L2-norm on & [ H ] ( s ) ) .Furthermore, since P ( s ) has only finitely many
+
[kh,(k l)h]. Then the Laplace transform L[4] exists for poles, the product of these three functions admits a coprime
Res > p, and factorization G(s) = Q - l ( s ) P ( s ) [23]. Now define
= eh*L[41(s).
L C ~ [ ~ [ ~ I I Z I I I ~ = ~ ~ S (45) XQ:= .{ CO) : Q
E LfOC[O, * ZI(0,).. = 0)

A remark on (45) is in order. Acgording to the definition of where Q is the inverse Laplace transform of Q(s), and *
(5), Z[C$][Z] = Cf,fJkZ-le, where S ( 4 ) = { + k ) g 1 . Here the denotes convolution. Then the following fact is known.
multiplication operator by z acts as the one-step left shift: Fact 6.3 [23]: If W = Q-'(s)P(s) is a coprime factor-
{ 4 k ) H { + k + l } . In the time domain, this clearly corresponds ization, then the closure of the set of all input-free outputs
to d ( t ) H q5(t+h),so that its Laplace transform is ehs. Identity generated by W ( s )is precisely XQ.Furtherm!re, for another
(45) claims that taking the finite Laplace transform of each such D ( s ) , X D c XQ holds if and only if Q = II * for
piece (bk and expanding the sum via the substitution z H eh' some matrix II over the ring of distributions with compact
actually yield the Laplace transform. The multiplication by eh' support in (--00, 01. When this holds, we write D(s)lQ(s).
on the right-hand side becomes necessary to account for our Let us now consider the tracking problem. Suppose that the
definition of Z-transform starting with rather than zo. reference signal generator is represented by a left coprime fac-
Pmofi That 4 is Laplace transformable and converges torization D - ' ( s ) N ( s ) over polynomials. By the fact above,
absolutely for R e s > is obvious from (44). To show (45), the input-free outputs of this system are precisely X D which
observe is the same as those generated by D - l ( s ) . Therefore, we may
as well take D - l ( s ) as the reference signal generator. The
following theorem gives the necessity of the internal model
for the hybrid sampled-data system Fig. 2.
YAMAMOm SAMPLED DATA CONTROL SYSTEMS 711

the practical cases; for example,

U
0-'(a)

-
Fig. 4. Steady-state mode.
Remark 6.5: A similar hypothesis is made in [15]; see also
Example 6.8.
Let us now state and prove the sufficiency of our internal
Theorem 6.4: Consider the unity feedback system given model principle.
in Fig. 2. Suppose that the exogenous signal is generated Theorem 6.6: Let C ( z ) and P ( s ) be the transfer matrices
by D - l ( s ) with poles only in the closed right-half plane. of the digital and analog parts in Fig. 2, and C ( z ) =
Suppose also that the closed-loop system is internally stable, P c ( z ) Q c ( z ) - l and P ( s ) = Q p ( s ) - l P P ( s ) be their
and let G(s)= Q - l ( s ) P ( s ) be a coprime factorization of the coprime factorizations over polynomials. Let $(s) =
loop transfer function. If the closed-loop system asymptotically r(z)R2-l(z)~(s)-'IZ,,hsbe the transfer function of the hold
tracks any signal generated by D - l ( s ) , then D ( s ) divides element satisfying Assumption C. Suppose that the exogenous
&(SI* signal is generated by D - l ( s ) with polynomial matrix D ( s ) .
Proof: Let ~ ( be
t ) any signal generated by D-'(s), and Assume the following five conditions:
y ( t ) the corresponding output tracking ~ ( t Let
) . us also write 1) The closed-loop system C,l is stable.
T k and Y k in accordance with the notation in Section 11.
2) Q c ( z ) = r(z)II(z) for some polynomial matrix II(z).
We first consider the system in the discrete-time mode. If 3) Q ( s ) := Q p ( s ) Q ( s )= N ( s ) D ( s )for some polynomial
the asymptotic tracking occurs, then it follows that the error matrix N ( s ) , i.e., Q ( s ) contains D ( s ) as an internal
e ( t ) must also converge to zero at the sampling instants, i.e., model.
e ( k h ) = ek(h) --f 0, k + 00. By the linearity of the system 4) The minimal realizations of D - l ( s ) and Q - l ( s ) both
we can decompose the response as satisfy the spectrum nondegeneracy Assumption A.
5) For any pole p of D-'(s), e P h is not a pole of
rI-1 (z)R-1( 2 ) .
where yi(8) is the initial-state response (corresponding to the Then the closed-loop system asymptotically tracks any
state at time kh), and YE(@) is the response corresponding to signal generated by D-' (s).
the input e ( k h ) .Since the input e k ( h ) + 0, yE(e) approaches Proof: By property 2, the numerator F(z) of $ is can-
0 as k -+ 00. This means that celled by the denominator Q c ( z ) . This means that by suitably
y;(e) - Tk(e)-, o as k -, 00. (46) setting an initial state in C ( z ) , one can get a continuous-
time output generated by +(s)-l by combination of the
(See Fig. 4 for the steady-state interpretation.) Since T is digital element C ( z ) with the hold element. Therefore, we can
an output generated by D-'(s) and it is shift invariant and regard the loop transfer function as the cascade connection of
arbitrary, (46) and Fact 6.3 imply that there exists a sequence Pc (2 ) r I - l (z)R-' (Z ) and Q-l (s)Pp(s).
of states zk in XQ whose corresponding outputs approach Since D(s)IQ(s),if we look at only the sampled instants,
~ ( t )But
. since Q-lP is coprime, the closure of the set of the closed-loop system contains the discrete-time internal
outputs generated by this system is precisely XQ by Fact 6.3. model. Therefore, by the internal stability, at least tracking at
Since XQ is closed, it follows that sampled instants occurs. Therefore, the error input to the loop
x D c XQ. transfer function tends to zero, and the sampled output y ( k h )
approaches the sampled reference signal r ( k h ) . In the same
Again by Fact 6.3, this implies D(s)lQ(s). 0 way as in the proof of Theorem 6.4, we see that there must
The necessity theorem above is still unsatisfactory because exist a sequence z j of initial states in the forward-loop system
the delay element C(eh") contains too many poles and zeros. such that their corresponding responses approach the sampled
For example, l/(ehs - 1) has 2nnj/h, n = 0, f l , . . .as reference signal r ( k h ) .Therefore, there exists an initial state z
poles, but the corresponding digital compensator is just l/(z- in the forward-loop system such that its corresponding output
l ) , and it need not have the capability of producing continuous- agrees with ~ ( tat) sampled instants kh, IC = 1, 2 , . . .. Now
time signals that can be produced with l / ( e h s - 1). Only in by property 5, the digital part Pc(z)II-l(z)R-'(z> cannot
combination with the hold device can the digital part produce contribute to this output. Therefore, this initial state must lie
continuous-time signals. Therefore, we make the following in the continuous-time part Q-'(s)Pp(s). Thus we return to
assumption. the steady-state diagram Fig. 4. In this steady-state mode, the
Assumption C: The finite Laplace transform $(s) of the output ~ ( tmust ) be cancelled by y ( t ) at sampled instants.
hold function H ( t ) is expressible as the ratio $ ( s ) = Since both the exogenous signal generator D-'(s) and the
r(z)s2-1(z)Q(s)-11Z=,h, where r ( z ) and R(z) are poly- continuous-time part Q-l ( s ) satisfy Assumption A, however,
nomial matrices in z and Q(s) is a polynomial matrix in and since D ( s ) divides Q ( s ) , if y(t) agrees with ~ ( t at)
S. sampled instants, it must agree with ~ ( ton ) the intersample
Since $(s) must be entire, all zeros in +(s) must be intervals. (Otherwise, there would be ep' E X D that agrees
cancelled by r(eh").This assumption is satisfied for most of with some output in XQ at sampled instants but not on the
712 IEEE TRANSACTIONS ON AUTOMATIC CONTROL,VOL. 39, NO. 4. APRIL 1994

Fig. 5. Closed-loop system.

intersample intervals. But this is clearly impossible under

Assumption A, by the linear independence of exponentials.)O
Remark 6.7: The spectrum nondegeneracy condition on
both the exogenous signal generator and the augmented plant
Q-'(s) cannot be removed in Theorem 6.6. For example, if
we have s = 0, 27rj/h both as poles in the continuous-time
part, then it may well happen that the closed-loop system may
track to the sine wave even if the exogenous signal is a step,
because the closed-loop system does not have the capability
of knowing the intersample error. The same can be said of the
exogenous signal generator.
To summarize: Theorems 6.4 and 6.6 give variants of
the continuous-time internal model principle for sampled- 0 2 4 6 a 10
data systems. The z and s domains are linked by the finite Fig. 6. Tracking to ramp.
Laplace transform and the substitution z = eh*. With this, the
combination of some part of the digital compensator with the
hold element can work as a continuous-time system and hence function is
an internal model (cf. Fig. 5). 'Qpical examples now follow. +
h(4s 1)
Example 6.8: The conditions of the theorem are satisfied + +
ehss2(s2 4s 6 )
for the following system with D ( s ) = s2 w 2 : +
h so that it surely has the internal model 1/s2. Here one pole s =
C(z) = - 0 is in the plant P ( s )and the other is made by the combination
2-1
w = 27r/h of the hold (1 - e-hs)/s and the digital compensator h / ( z- 1)
H ( t ) = sin ( w t ) ( p ,y,
through the cancellation (1 - e - h s ) / ( z - l)lehs=+ = l/z.
Asymptotic tracking is achieved as seen in Fig. 6.

1
P ( s )= - VII. CONCLUDING REMARKS
s- +, 1- '
We have given a function space approach to sampled-
+ +
Here Q ( s ) = (s 1)(s2 w 2 ) and contains the internal
data control systems. The introduction of a time-invariant,
+
model s2 w 2 . Hence the closed-loop system Fig. 5 can
though infinite-dimensional,model made it possible to discuss
track a sinusoidal wave with w = 27r/h. The internal model is
tracking and ripples in terms of the familiar notions of poles
made up with combination of 1/(z - 1) and the hold element
and zeros and associated zero directions. A new element here
+
(I - e - h s ) / ( s 2 w 2 ) . is that this theory enables us to regard intersample ripples as
One of the consequences of the above theorem is that we
the problem of transmission zero directions just as in the case
can incorporate an internal model for sampled-data systems in
of the finite-dimensional multivariable systems where zero
such a way that we can split poles to analog and digital parts.
directional vector also governs tracking properties.
A typical example is the case of tracking to ramp ~ ( t=) t,
The general case allowing multiple poles is more interesting
where we can incorporate one pole A = 1 into the digital
in that it exhibits a nonclassical internal model splitting
compensator and the other to the analog plant P ( s ) :
into digital and analog parts. Although we have made some
Example 6.9:
restrictive assumptions on the type of a hold device for the suf-
Consider the unity feedback sampled-data system Fig. 5
ficiency part of the internal model principle, this assumption is
with
satisfied for most practically encountered cases. Generalization
h to the case of more general hold devices is a topic for future
C(Z) = -
2-1 investigation.
+(s) = ~
l-edhs 2-1
- -. 1
I
z s ACKNOWLFDGMENT
P(8) =
+1 4s
The author wishes to thank Profs. M. Ikeda and S . Hara
s(s2 + 4s + 6)' for their valuable comments and discussions. He also wishes
It is easy to see that they satisfy the conditions of Theorem to thank an anonymous referee for calling his attention to the
6.6. The Laplace transform o f t is 1/s2, and the loop transfer references [7] and [16].
YAMAMOM: SAMPLED DATA CONTROL SYSTEMS 713

REFERENCES [ 151 A. Kawano, T. Hagiwara, and M. Araki, “Robust servo condition for
sampled-data systems,” (in Japanese), in Proc. SICE Contr. Theory
B. Bamieh and J. B. Pearson, “A general framework for linear periodic Symp., 1990, pp. 3540.
systems with applications to H, sampled-data control,” IEEE Trans. [16] J. B. Lewis and Y.-S. Jan, “On linear discrete-time deadbeat control,”
Automat. Contr., vol. 37, pp. 418-435, 1992. in Proc. 20th CDC, 1980, pp. 494499.
B. Bamieh, J. B. Pearson, B. A. Francis, and A. Tannenbaum, “A lifting [17] J. R. Ragazzini and G. F. Franklin, Sampled-Data Contml Systems.
technique for linear periodic systems with applications to sampled-data New York McGraw-Hill, 1958.
control systems,” Syst. Contr. Lett., vol. 17, pp. 79-88, 1991. [18] W. Rudin, Functional Analysis. New York McGraw-Hill, 1973.
T. Chen and B. A. Francis, “Input-output stability of sampled-data [19] S. Hara and H.-K. Sung, “Ripple-free conditions in sampled-datacontrol
systems,” IEEE Trans. Automat. Contr., vol. 36, pp. 50-58, 1991. systems,” to be presented in 30th CDC Brighton, 1991.
-, ‘“HZ-optimal sampled-data control,” IEEE Trans. Automat. Contr., [20] H. T. Toivonen, ‘Sampled-data control of continuous-time systems with
vol. 36, pp. 387-397, 1991. an 3-1, optimality criterion,” Automatica, vol. 28, pp. 45-54, 1992.
[21] F. Treves, Topological Vector Spaces, Distributions and Kemels. New
B. A. Francis and T. T. Georgiou, “Stability theory for linear time-
invariant plants with periodic digital controllers,” IEEE Trans. Automat. York: Academic, 1967.
[22] S. Urikura and A. Nagata, “Ripple-free deadbeat control for sampled-
Contr., vol. 33, pp. 820-832, 1988. data systems,” IEEE Trans. Automat. Contr., vol. 32, pp. 474482,
B. A. Francis and W. M. Wonham, “The internal model principle
1987.
for linear multivariable regulators,” J. Appl. Math. Optim., vol. 2, pp. [23] Y. Yamamoto, “Pseudo-rationalinput/output maps and their realizations:
17G194, 1975. A fractional representation approach to infinite-dimensional systems,”
G. F. Franklin and A. Emami-Naeini, “Design of ripple-free multivari- SIAM J. control Optim., vol. 26, pp. 1415-1430, 1988.
able robust servomechanism,” IEEE Trans. Automat. Contr., vol. 31, [24] -, “New approach to sampled-data systems: A function space
pp. 661-664, 1986. method,” in Proc. 29th CDC, 1990, pp. 1882-1887.
T. Hagiwara and M. Araki, “Design of a stable feedback controller based
on the multirate sampling of the plant output,” IEEE Trans. Automat.
Contr., vol. 33, pp. 812-819, 1988.
S. Hara and P. T. Kabamba, “Worst case analysis and design of sampled
data control systems,” in Proc. 29th CDC,1990, pp. 202-203. Yutaka Yamamoto received his Ph.D. degree in
P. T. Kabamba, “Control of linear systems using generalized sampled- mathematics from the University of Florida in 1978
data hold functions,” IEEE Trans. Automat. Contr., vol. 32, pp. 772-783, under the guidance of R. E. Kalman.
1987. From 1978 to 1987 he was with Department of
P. T. Kabamba and S. Hara, “Worst case analysis and design of Applied Mathematics and Physics, Kyoto Univer-
sampled data control systems,”IEEE Trans. Automat. Contr., vol. 38, sity. In 1987 he joined the Department of Applied
pp. 1337-1357, 1993. Systems Science as an Associate Professor.
R. E. Kalman and J. E. Bertram, “General synthesis procedure for Dr. Yamamoto is currently an Associate Editor
computer control of single and multiloop linear systems,” AIEE Trans., of Automatica and Mathematics of Control, Signals
vol. 77, pp. 602-609, 1959. and Systems. He is a member of the Society of
P. P. Khargonekar, K. Poolla, and A. Tannenhaum, “Robust control of Instrument and Control Engineers (SICE) and the
linear time-invariant plants using periodic compensation,” IEEE Trans. Institute of Systems, Control and Information Engineers. He received Sawaragi
Automat. Contr., vol. 30, pp. 1088-1096, 1985. memorial paper award in 1985, outstanding paper award of SICE in 1987 (joint
P. P. Khargonekar and N. Sivashankar, ‘“HZ optimal control for with S. Hara), and best author award of SICE in 1990. His current research
sampled-data systems,” preprint, Department of Electrical Engineering, interests are in realization and robust control of distributed parameter systems,
Univ. Michigan, 1991, to appear in SIAM J. Control and Optimiz. learning control, and sampled-data systems.