4 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO.

1, JANUARY 1997

Dynamic Stabilization of Regular Linear Systems

George Weiss and Ruth F. Curtain, Fellow, IEEE

Abstract— We consider a general class of infinite-dimensional

linear systems, called regular linear systems, for which convenient
representations are known to exist both in time and in frequency
domain. For this class of systems, we investigate the concepts
of stabilizability and detectability, in particular, their invariance
under feedback and their relationship to exponential stability. Fig. 1. The closed-loop system 6p ;c .
We introduce two concepts of dynamic stabilization, the first
formulated as usual, with the plant and the controller connected
in feedback, and the second with two feedback loops. Even for (equivalently, if has a proper inverse). This happens
finite-dimensional systems, the second concept, stabilization with if and only if is invertible (equivalently,
an internal loop in the controller, is more general. We argue is invertible). An admissible feedback transfer function which
that the more general concept is the natural one, and we derive
sufficient conditions under which an observer-based stabilizing
is constant is called an admissible feedback operator.
controller with an internal loop can be constructed. In the sequel we assume that is an admissible feedback
transfer function for . Then, the feedback connection of
Index Terms— Detectability, internal loop, observer, regular
linear system, stabilizability, stabilizing controller.
Fig. 1 defines a new linear system , called the closed-
loop system, with input function , state , and output
function . The transfer function from to is
I. INTRODUCTION

T HIS PAPER deals with the dynamic stabilization of a

large class of infinite-dimensional systems called regular
linear systems (RLS’s). To make our paper easier to read,
and the transfer function of is .
in this section, we present the finite-dimensional version of
The system (or the matrix ) is called stable if
our results, some of which might be new even in this simple
as . is stable iff its eigenvalues are in the open
context.
left half-plane. The transfer function is in if it is
Let be a linear time-invariant system, called the plant,
analytic and bounded on the open right half-plane , which
described by
is the case iff the poles of are in the open left half-
plane. We call input–output stable if . is
(1) called stabilizable (respectively, detectable) if there exists a
matrix (respectively, ) of appropriate dimensions such
where is the input function, is the state of that (respectively, ) is stable. According to
at time , and is the output function. and a well-known proposition, is stable if and only if it is
are matrices of appropriate dimensions. Let be a similar stabilizable, detectable, and input–output stable.
system, called the controller, described by In all feedback system design methods, one of the objectives
is to make the closed-loop system stable. This is possible
(2) only if both the plant and the controller are stabilizable and
detectable. If both and are stabilizable, then the same is
We assume that the dimension of equals that of and true for , and a similar statement holds for detectability. It
similarly for and . We want to connect these two follows that if both and are stabilizable and detectable,
systems in feedback, as shown in Fig. 1. Let and denote then is stable if and only if
the transfer functions of and , respectively, so that
(3)

For this reason, if and are proper, rational, and (3) holds,
and similarly for . We say that is an admissible feedback then we say that stabilizes . In the “frequency domain
transfer function for if has a proper inverse theory” the systems are defined via their transfer functions,
realizations (1) and (2) are ignored (it is implicitly assumed
Manuscript received May 5, 1995; revised June 13, 1996. Recommended
by Associate Editor, B. Lehman. that they are minimal), and by the stability of the closed-loop
G. Weiss is with the Center for System and Control Engineering, School system is meant (3); see, for example, Vidyasagar [33].
of Engineering, University of Exeter, Exeter, EX4 4QF, U.K. If is stabilizable and detectable, then we can find
R. F. Curtain is with the Mathematics Institute, University of Groningen,
9700 AV Groningen, The Netherlands (e-mail: [email protected]). controllers such that is stable. Such a controller is
Publisher Item Identifier S 0018-9286(97)00487-X. called a stabilizing controller for . There are infinitely many
0018–9286/97$10.00  1997 IEEE
WEISS AND CURTAIN: DYNAMIC STABILIZATION OF REGULAR LINEAR SYSTEMS 5

stabilizing controllers for of various orders. Conceptually,

the simplest ones are obtained by taking matrices and
such that

are stable and then choosing

y u
Fig. 2. The observer-based controller with input and output . From the
additional output w v
to the additional input , we may connect an arbitrary
(4) stable system whose transfer function is the Youla parameter. For simplicity,
in the diagram we have assumed that D = 0.

Such controllers are called observer based, and they were

introduced by Luenberger [18]. These stabilizing controllers
are not of the lowest possible order, but they have a special
importance both in linear-quadratic-Gaussian theory (where
and are determined via Riccati equations) and because of
their connection to state-space formulas for doubly coprime
factorizations and to controller parameterization. For a beau-
tiful and insightful account of these ideas, we recommend
P
Fig. 3. A feedback connection with internal loop. If the conditions in
Maciejowski [19, ch. 6] (where the relevant references are
K
Proposition 1.2 hold, then the connection between the blocks (the plant)
listed) and also Francis [9, ch. 4]. and (the controller) should be realized first, and the lower “internal” loop
A block diagram of the controller defined by (4) is shown of the controller should be closed afterwards. Closing the internal loop first

K K
could result in an improper and/or unstable controller. In most applications,
in Fig. 2, with the input denoted by and the output denoted 11 = 0 and is stable.
by . One reason for this change of notation is that we think
of as being connected to as in Fig. 1, but with
and so that is the input of and is its output. For which, together with the four transfer matrices obtained by
the sake of simplicity of the diagram, we have assumed that regarding and as inputs, constitute a doubly coprime
. If , then a block has to be added (from the factorization of . More precisely, this is the canonical doubly
input of the block to the output of the block ), destroying coprime factorization from Nett et al. [21]. It is amazing
the symmetry. that all this can be achieved without actually changing the
The additional input appearing in Fig. 2 may be consid- relationship between the four signals and . It is only
ered to be zero, and the additional output may be ignored. our subjective way of regarding them either as inputs or as
However, these additional signals may also be very useful outputs which change. The extended plant, the observer-based
since they allow us to obtain all the transfer functions controller, the canonical left-coprime factorization, and the
which stabilize . This is achieved in the following simple canonical right-coprime factorization are different represen-
way, called Youla parameterization: let be rational tations of the same system if we define systems via behaviors,
and of the same dimensions as a feedback matrix for , let in the sense of Willems [41].
and assume that We propose another definition of a stabilizing controller
which is more general than the one we have defined earlier
is invertible (5) (even in the finite-dimensional context which we are now
using). This new concept enables a simpler Youla parame-
Then connecting from to in Fig. 2, so that (denoting
terization and has other advantages which will turn out to be
the Laplace transformation by a hat)
very important for infinite-dimensional systems. It makes the
theory of dynamic stabilization simpler and more natural, as
we shall see in Section V. Roughly speaking, we are no longer
the resulting transfer function from to is proper and looking for a controller which, when connected to the plant in
it stabilizes . Moreover, all the transfer functions which feedback gives rise to a stable system, but we are looking for
stabilize are obtained in this way. Naturally, we may a stable system which incorporates the plant as a subsystem.
construct a stabilizable and detectable realization of and The relevant picture is now Fig. 3 (instead of Fig. 1).
interconnect it with the observer-based controller of Fig. 2 in Definition 1.1: Let and be finite-dimensional vector
the proper way, obtaining a stabilizing controller for . spaces. Let and be proper rational transfer functions, from
There is even more hidden in the simple diagram of the to and from to , respectively. is partitioned
observer-based controller, if we look at it in different ways, naturally into and . We say that is an
regarding other signals as inputs (instead of and ). Indeed, admissible feedback transfer function with internal loop for
regarding and as inputs (and the remaining two signals if has a proper inverse, where
as outputs—this is possible, by changing the direction of
some arrows), we obtain an extension of the plant (more
precisely, we get the plant with an additional input and an
additional output ). The transfer function from to is . If
we regard and as inputs, we obtain four transfer matrices
6 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 1, JANUARY 1997

The intuitive meaning of this concept is that the inter- Example 1.4: We assume that all signals are one-
connection shown in Fig. 3 makes sense. Denoting dimensional, that is, , and that
, it is clear that the condition in Definition 1.1 is
equivalent to the invertibility of .
Proposition 1.2: With the notation of Definition 1.1, sup-
pose that the following conditions are satisfied.
1) is an admissible feedback transfer function for .
2) is an admissible feedback operator for the transfer We take to be the simplest realization of , described
function by (this is unstable, of course ).
is a static system with trivial (zero dimensional) state space.
(6) This is an extreme example illustrating the impossibility of
closing the internal loop first: to compute the transfer function
Then is an admissible feedback transfer function with of the resulting controller by (7), we would have to divide
internal loop for . something by zero. However, the conditions of Proposition
For the proof of this proposition (in a somewhat more 1.2 are satisfied so that we can connect to first and
general context) we refer to Proposition 4.8. The expression close the internal loop afterwards. The state of the closed-loop
in (6) is called an upper linear fractional transformation of system is , and it satisfies (i.e., )
and ; see, for example, [19, p. 317]. The converse of so that is a stabilizing controller with internal loop for
Proposition 1.2 is not true, i.e., Conditions 1) and 2) are not .
necessary. Examples to illustrate this fact are easy to find. This controller is particularly well suited for tracking:
Intuitively, is the transfer function of the plant, and is suppose that an external signal is subtracted from so
the transfer function of the controller, but with a connection that the first input of is (instead of ). The
within the controller missing. This missing connection, called reader is invited to draw the relevant diagram. Then very
the internal loop, should go from the second output of simple computations show that (after closing the loops) we
to its second input. In all the applications that we have in have , that is, we get perfect tracking. Intuitively,
mind, is stable and the two conditions of Proposition 1.2 we may think of our controller (with the internal loop closed)
are satisfied. The first condition says that we are allowed to as an amplifier with infinite gain from to . However,
connect to as long as the internal loop is not connected. this does not mean that there are hidden infinite signals in
After connecting to , the transfer function seen from the our system. Indeed, if is bounded (respectively, in ),
internal loop is , and the second condition says that then the signals and are also bounded (respectively,
now we are allowed to connect the internal loop. If we connect in ); and moreover, is continuous. In the terminology of
the internal loop first, then usually we get a controller with [41], the behavior of the controller is defined by the equation
transfer function .
It is easy to give a physical interpretation to our system. The
(7) plant is a unit resistor connected in series with a unit capacitor,
is the current, is the voltage on the capacitor, and
However, the above expression might be unstable, or improper, is the total voltage. The reference voltage is generated
or (even worse) it might be undefined, as we illustrate in a by an independent (time-varying) voltage source. We connect
simple example below. this source to our plant, which obviously results in perfect
Let and be finite-dimensional vector spaces. Let tracking (we invite the reader to draw the circuit). One of the
and be linear time-invariant systems with input spaces wires connecting the source to the plant may be regarded as
and , respectively, and with output spaces and the controller. Indeed, it has a practically infinite gain from
, respectively. We denote by and the transfer the voltage between its extremities, , to the current
functions of these systems so that is partitioned naturally through it, . Of course, for this circuit, the above solution of
into and . Suppose that is an admissible the tracking problem is what any engineer would have done
feedback transfer function with internal loop for . Then regardless. We just wanted to draw the reader’s attention to
connecting with , as shown in Fig. 3, we obtain a new the fact that using control theory based on Fig. 1 could not
system with no inputs and no outputs. The state of this lead to this (realistic) solution.
system has two components, namely the state of and the Design Procedure: We now describe a procedure for ob-
state of , and satisfies a differential equation of the form taining stabilizing controllers with internal loop for any stabi-
. Of course, we could easily define inputs lizable and detectable plant. In fact, we return to the idea of
and outputs for this closed-loop system, but since we are only the Youla parameterization described earlier, but we no longer
interested in its stability, this is not necessary. Computing impose (5).
is not needed at this stage. In a more general context, this We assume that the (possibly unstable) plant is described
computation will be carried out in Section V. by the equations
Definition 1.3: With the notation of the preceding para- . Its transfer function is .
graph, we say that is a stabilizing controller with internal We assume that is stabilizable and detectable. We denote
loop for if is stable. by and its input, state, and output spaces.
WEISS AND CURTAIN: DYNAMIC STABILIZATION OF REGULAR LINEAR SYSTEMS 7

The new state equations are

Fig. 4. A stabilizing controller with internal loop. The plant should be

connected from u y
to . Afterwards, the connection from 0 to u u should The transfer function of the above system is , defined
be restored. The block with transfer function Q represents any stable linear in (6). By simple computation
system with appropriate input and output spaces. For simplicity, in the diagram
we have not shown the block . D (9)
We see that has no feedthrough term (i.e., its limit
Let be an arbitrary stable linear system with input space at infinity is zero), and hence Condition 2) of Proposition 1.2
, state space , and output space , described by is satisfied. We can therefore close the internal loop, obtaining
the equations

where is the input function, is the state, and is the (10)

output function. We denote by the transfer function of . The block matrix above is similar to from Definition
We define , the controller with internal loop, as follows: it is 1.3. From its special structure it is easy to see that its
obtained by connecting from the output to the input of eigenvalues are eigenvalues of one of the matrices ,
the observer-based controller shown in Fig. 2 (with the matrix , or . Therefore, is stable and so is
present, although not shown in the picture) and interrupting a stabilizing controller with internal loop for . Note that
the output of the summator giving , turning the interrupted we have not assumed (5) so that closing the internal loop
signal into a new output (see Fig. 4). The internal loop goes first might lead to an improper or undefined controller. For
from to (now is both an input and an output). Note that example, by taking
when (5) is satisfied and the internal loop is closed, then we and , we get and as described in Example 1.4.
get the stabilizing controller described earlier, before Fig. 2.
The first input of is , and its first output is . These There is a dual procedure for obtaining a stabilizing con-
have to be connected to the plant. Thus has as input space troller with internal loop, as follows: instead of interrupting
, as state space , and as output space . the signal at the output of the summator giving , interrupt
(The space introduced before Definition 1.3 is now .) the signal at the output of the summator giving , thus
If we denote the state of the observer-based controller by obtaining the new output and the new input . The
, then the state of is . The equations of are internal loop is now from to (and ). The
resulting (dual) controller satisfies , and
. The transfer functions of
the controllers obtained by the two dual procedures described
above are most easily expressed in terms of a doubly coprime
(8a) factorization of ; see Section V or Curtain et al. [7], [8].
The aim of this paper is to extend the above results to RLS’s.
These form a large class of infinite-dimensional, linear, time-
invariant systems. They include systems described by partial
differential equations with boundary control and/or boundary
(8b)
observation as well as systems described by delay-differential
equations. They can be described by state equations similar
Note that basically, is the cascade connection of two stable to (1), but the operators and are unbounded in
systems and, of course, is stable. It is clear from the general. For the precise definition of RLS’s and many of their
equations that so that Condition 1) of Proposition properties, we refer to the papers by Weiss [37], [38]. Some
1.2 is satisfied. Note also that . minimal background material will be given in Section II.
Let us see what happens if we interconnect with The difficulties we shall encounter stem mainly from the
(with the internal loop open ). As is well known, contains unboundedness of and , and herein lies the true motivation
an observer, more precisely, approximates with an for introducing the new stabilization concept of Definition 1.3.
error that converges to zero. It is advantageous to introduce the The definitions of stabilizability and detectability become more
estimation error and then make a change complicated since we allow and to be unbounded as
of state variables. The new state of the interconnected system well. It is not obvious how to make sense of expressions like
is , where the exponent stands for transpose. , and it is not necessarily true that the operator
8 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 1, JANUARY 1997

defined as in (4) is the generator of a strongly continuous A well-posed linear system is a linear time-invariant system
semigroup, as illustrated in Example 6.5. One advantage of our such that on any finite time interval, the operator from the
concept of a stabilizing controller is that we do not consider initial state and the input function to the final state and the
the controller alone with the internal loop connected, and so output function is bounded. The input, state, and output spaces
we do not need to be the generator of a semigroup. are Hilbert spaces, and the input and output functions are of
The first paper on stabilizing compensators for infinite- class . For the detailed definition we refer to Salamon [26],
dimensional systems was by Gressang and Lamont [11] [27] or to Staffans [31], [32], or to Weiss [37], [38] (in [37]
who gave a generalization of the Luenberger compensator and [38], these systems were called abstract linear systems).
from [18]. They considered an infinite-dimensional system The input-to-output operator of any well-posed linear sys-
described by (1), where is the generator of a strongly tem can be described by a transfer function which is an analytic
continuous semigroup and and are bounded operators. operator-valued function defined on some right half-plane in
In the early eighties, results about compensators for this . The transfer function of any well-posed linear system is
class of systems were published by Balas [1], Nett et al. well-posed, meaning that it is bounded on some right half-
[20], Schumacher [28], [29], and others; see the survey paper plane (not necessarily the same where it is defined). We do
by Curtain [4]. The problem of finding finite-dimensional not distinguish between two transfer functions defined on two
stabilizing compensators for systems with unbounded and different right half-planes, if one function is a restriction of
bounded has been addressed by Curtain and Salamon [5], the other (thus, by a transfer function we mean in fact an
and more recently Lasiecka [14], [15] has addressed this equivalence class of analytic functions). Well-posedness is the
problem with both and unbounded. natural generalization of properness to irrational functions.
This paper is organized as follows. In Section II we provide Let be a well-posed linear system, with input space ,
some background on RLS’s and on static output feedback. state-space , and output space . Let be the semigroup
In Section III we investigate the concepts of (exponential) of , i.e., the strongly continuous semigroup on which
stabilizability and detectability introduced in Rebarber [24]. describes the evolution of the state of if the input function
We show that they are feedback invariant and, for stabilizable is zero. Let denote the generator of . The Hilbert spaces
or detectable systems, we derive equivalent conditions for and are defined as follows: is with the
exponential stability, in terms of transfer functions. norm , where is fixed
In Section IV we discuss dynamic stabilization, without and and is the completion of with respect to the norm
with internal loop, for RLS’s. We show how the exponential . We have
stability of the closed-loop system is related to certain transfer
functions being in .
In Section V we generalize the design procedure to RLS’s. It and the semigroup can be extended to a semigroup on .
turns out that constructing a stabilizing controller with internal These two semigroups are isomorphic, and we shall denote
loop is possible under more general and natural assumptions them by the same symbol. The generator of on is an
than those needed to do it in the classical sense (without extension of to , also denoted .
internal loop). The precise conditions are listed in Proposition The state of at any moment can be expressed by
5.3, and the equations of the controller are given in Theorem the formula
5.4, which is the main result of this paper.
In Section VI we address the following question: when
is a stabilizing controller with internal loop reducible to a
stabilizing controller in the usual sense? The main point of
(see [26] and [34]). Here is the input
the section is an example of an unstable system with
function and is the control operator of .
feedthrough operator and of a stabilizing controller
We have , and depends continuously on , on
with internal loop , obtained by the design procedure with
, and on . The transfer function from the input to the
Youla parameter . This controller cannot be reduced
state is .
to a stabilizing controller (by closing its internal loop).
If and , then the output function of on
Indeed, if this were possible, then the semigroup generator of
is (see [26] and [36])
would be something like [according to
(4)]. However, the latter expression is not a generator.

II. SOME BACKGROUND ON FEEDBACK Here, is the observation operator of . The

FOR REGULAR LINEAR SYSTEMS -extension of is defined by
In this section we recall very briefly some facts about RLS’s
(11)
and especially about feedback for such systems, as needed
in later sections. We do not give proofs, but only indicate
( is real), for all in the domain
references. We assume that the reader has some familiarity
with the concept of an RLS, as presented in [38]. We do not the limit in (11) exists
assume the reader is familiar with the results on feedback
contained in the more recent paper by Weiss [39]. (In [38] this operator was denoted .)
WEISS AND CURTAIN: DYNAMIC STABILIZATION OF REGULAR LINEAR SYSTEMS 9

Let denote the transfer function of . is called regular

if the following limit exists for all v

(12)
Fig. 5. The closed-loop system 6k .
( is real). Then is called the feedthrough
operator of . If is regular, then is called an RLS. The Now we turn to feedback, following [39]. Let and be
regularity condition (12) can be formulated in many different Hilbert spaces, suppose is an -valued well-posed
ways, of which we mention the following: is regular if and transfer function, and let . is an admissible
only if the product makes sense for some feedback operator for if is invertible on some right
(hence for every) , i.e., half-plane and its inverse is a well-posed transfer function
(equivalently, if has the same property). Then the
function defined by

(see [38]). Another equivalent condition is that the average of (15)

any step response of over has a limit as (the
original definition in [37]). is called the closed-loop transfer function corresponding to
If is regular, then and . We have
(16)
(13)
By combining [39, Proposition 3.14] with (16), we obtain
for every with Re sufficiently large. Moreover, in the following proposition.
the time domain, is completely described by the following Proposition 2.2: Let and be admissible feedback op-
equations: erators for , and let and be the corresponding
closed-loop transfer functions. Then is an admissible
(14a) feedback operator for , the corresponding closed-loop
(14b) transfer function is , and we have that
which hold for almost every (in particular, (17)
for almost every ). is the unique strong
solution (in ) of (14a), for any given Now suppose that is regular, and let be its feedthrough
and . The output function belongs to operator. If is an admissible feedback operator for ,
. then [given by (15)] is regular iff is invertible
The operators and , appearing in (14), are called (equivalently, iff is invertible). If is regular, then
the generating operators (GO’s) of . We mention that in [38], its feedthrough operator is
(13) and (14) were stated (and proved) with in place of
(18)
, where is another extension of . is an extension of
, and so it is clear that (13) and (14) are valid as stated here. Under certain additional assumptions (e.g., if at least one of
We prefer to work with because its definition is somewhat the spaces and is finite-dimensional), the invertibility of
simpler than that of , and because it is not known if formula follows from the admissibility of .
(23), below, is true for as well. Let be a well-posed linear system with transfer function
Let and be Hilbert spaces and and linear . We say that is an admissible feedback operator for if
operators. We say that is a regular triple on is an admissible feedback operator for . In this case there
and if for some (hence for every) exists a unique closed-loop system corresponding to
are the GO’s of an RLS with input, state, and output spaces (the open-loop system) and , represented in Fig. 5. is a
and . Combining results from Curtain and Weiss [6] well-posed linear system with the same state space as , and
and from [38], we get the following proposition. its transfer function is from (15). For the precise definition
Proposition 2.1: is a regular triple if and only if of we refer to [39]. Proposition 2.2, without formula (17),
the following five conditions are satisfied. can be reformulated in terms of systems (instead of transfer
1) is the generator of a strongly continuous semigroup . functions). In particular, for we get that is an
2) is an admissible control operator for (as defined admissible feedback operator for , and the corresponding
in [34]). closed-loop system is . Any interconnection of finitely many
3) is an admissible observation operator for (as defined well-posed linear systems (starting with the simplest cascade
in [36]). connection) can be represented as a closed-loop system in the
4) makes sense for some (hence for every) above sense.
. Now suppose that is an RLS with GO’s and
5) is bounded on some right , and is as before. Let be invertible, then
half-plane. (as mentioned earlier) is regular. We can compute the
10 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 1, JANUARY 1997

GO of via formulas which are similar to those for finite- III. STABILIZABILITY AND DETECTABILITY
dimensional systems. Let denote the GO of We introduce some notation. denotes the open right half-
. As we already know, is given by (18). The formula plane in , where Re . For any Banach space ,
for is is the Banach space of -valued bounded analytic functions
(19) on with the sup norm. Sometimes we write instead of
, if is clear from the context. Let be a well-posed
defined for all in the domain linear system with input space , state space , and output
1 3 space . Let be the transfer function of , and let be
(20) its semigroup.
We call defined above the natural domain of The system is called input–output stable if
since it is its domain as a semigroup . As is well known, input–output stability is
generator on (formula (19) enables us to extend to equivalent to the following condition: if the state of at
larger domains, if needed). is zero and the input function is in ,
The operators and are given by then the output function is in . is called
exponentially stable if the growth bound of is negative.
(21) Exponential stability implies input–output stability (see [38]),
(22) but the converse is not true (even in finite dimensions). Let
be the generator of . Instead of “ is exponentially
The formulas for , , and given in [39] use
stable,” we also say that “ is exponentially stable” or “ is
instead of , but it is easy to see that this makes no
exponentially stable.” We recall the following definition from
difference. As already mentioned, in this paper we consistently
Rebarber [24], [25].
use only . To understand (22) better, it should be pointed
Definition 3.1: Let and be Hilbert spaces, let be the
out that , where is the analogue of
for , but in fact there exists a Banach space such that generator of a strongly continuous semigroup on , and let
. Then is stabilizable if there exists
such that:
and both and are in . depends only on i) is a regular triple;
and , is dense in , and has a continuous ii) is an admissible feedback operator for
extension to . ;
The extensions and [defined as in (11)] are related iii) is exponentially stable.
by In this case we say that stabilizes .
Some comments are in order. Note that by i) above and by
(23) Proposition 2.1, must be an admissible control operator for
For determining the domain of the closed-loop generator, it , and must be an admissible observation operator for .
is sometimes useful to have another version of (20) available. Condition ii) means that has a well-posed
For this we introduce the space inverse, but it follows from iii) that this inverse is in fact in
. Indeed, the inverse of is
(24) the transfer function of an RLS whose semigroup generator
where (it is easy to see that is independent of ). is , as we shall see in the proof of Proposition 3.2.
It follows from the regularity criterion mentioned before (13) In iii) it is understood that is defined on its natural
that . It is easy to see that , whence domain; see formula (20) with and .
It was shown in [24] that this definition of stabilizability is
more general than others which have appeared in the literature
(25) and that (unlike previous definitions) it does not imply that
It follows from (18), (22), and (23), by simple substitution if is finite-dimensional, then can be decomposed into
into (13), that a finite-dimensional unstable part and a stable part. Relevant
(26) examples of systems which are stabilizable in the above sense
were given in [25]. Recently, Staffans [32] has given a more
and
general definition of stabilizability, from which the regularity
(27) assumption has been eliminated.
Proposition 3.2: Let and be as in Definition
The resolvents of and are related by
3.1, and suppose that is stabilizable. Then is expo-
nentially stable if and only if the function is in
.
(28) Proof: First we show the necessity. Let denote the
growth bound of , the semigroup generated by . We
for any with Re sufficiently large. recall from Weiss [40] that the admissibility of implies the
WEISS AND CURTAIN: DYNAMIC STABILIZATION OF REGULAR LINEAR SYSTEMS 11

following: for any there is a such that Proof: Let be the RLS with GO’s . Its
transfer function is
(29)

for all with Re . In particular, if is

where is the transfer function of . Clearly is an
exponentially stable ( ), then the function
admissible feedback operator for , and the corresponding
is in .
closed-loop system is exponentially stable (since by (19)
To prove the sufficiency of the condition, let
and (20), its semigroup is generated by , restricted
be an operator that stabilizes , and put
to its natural domain). Another admissible feedback operator
(with its natural domain). Let denote the RLS with
for is , so let denote the corresponding
GO’s and let denote the closed-loop system
closed-loop system. Using the formulas (18)–(22), it is easily
corresponding to with the feedback . Then it follows from
the results stated in Section II that is an RLS and its GO’s verified that the GO’s of are , where
are (restricted to ) and zero. Using (28) (with denotes the restriction of to . By Proposition
) we obtain 2.2 and its reformulation in terms of systems (as explained
in Section II), is an admissible feedback operator for
, and the corresponding closed-loop system is . Since
(30) this latter system is exponentially stable, by Proposition 3.3,
By elementary semigroup theory, the exponential stability of we have that
implies that is in . By the dual
statement of (29) and by the exponential stability of , we
have that is in . If we assume
that is in , then (30) shows that stabilizes . By simple computation, we have for any
must be in . By a theorem of Prüss
[23] (see also Huang [13] or Weiss [35]), it follows that is
exponentially stable.
The above proposition is due to Rebarber [24] (not stated
explicitly, but only as part of a proof; see also the comment
after Theorem 3.8).
Proposition 3.3: Let be an RLS with GO’s
and let be an admissible feedback operator for such that After carrying out the matrix multiplications, it turns out that
is invertible. If the corresponding closed-loop system , the operator defined in the statement of the proposition.
is exponentially stable, then the operator
stabilizes . Now we can state a generalization of the well-known
Proof: Since is a regular triple, it follows from finite-dimensional result that stabilizability is invariant under
Proposition 2.1 that is also a regular triple, i.e., i) of feedback.
Definition 3.1 is satisfied. We notice that is an extension Corollary 3.5: With the notation of Proposition 3.4,
of , whence is stabilizable if and only if is stabilizable.
Indeed, as mentioned in Section II, the roles of and
can be reversed so that the stabilizability of implies
that of .
Definition 3.6: Let and be Hilbert spaces, let be
the generator of a strongly continuous semigroup on ,
where is the transfer function of . By the admissibility and let . Then is detectable if there
of , the last expression has a well-posed inverse so that exists such that:
ii) is also satisfied. Finally, iii) is satisfied, because by (19) i) is a regular triple;
is exactly the generator of the closed-loop system ii) is an admissible feedback operator for
corresponding to and . ;
Proposition 3.4: Let and be as in Propo- iii) is exponentially stable.
sition 3.3, and let be the corresponding closed-loop system In this case we say that detects .
(not assumed to be stable). Let be the GO’s This definition is taken from [24] and [25], and again there
of . If there exists an operator that stabilizes , and exists a more general concept defined in [32]. There is a partial
if we define the operator on by duality between stabilizability and detectability, as expressed
by the following proposition.
Proposition 3.7: Let and be as in the last
definition, and assume that is finite-dimensional. Then
then stabilizes . is detectable if and only if is stabilizable.
12 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 1, JANUARY 1997

The proof of this proposition, while not difficult, requires a

result about duality which is not yet available in the literature,
namely that the adjoint of is (this
should be in [38, Part II]). We shall not use this proposition
in this paper. The reason for requiring that should be finite- Fig. 6. The feedback connection of two well-posed transfer functions.
dimensional is that otherwise the regularity of does
not imply the regularity of . However, it might
feedback transfer function for is an -valued well-
be possible that is stabilized by another operator (not
posed transfer function such that is invertible on
by ), and so we do not know if Proposition 3.7 is true for
some right half-plane and its inverse is a well-posed -
infinite-dimensional , in general.
valued transfer function.
In spite of the fact that the duality is not perfect, it is
From we see that by
not difficult to formulate and prove the dual versions of
interchanging with , we get an equivalent definition. For
Propositions 3.2–3.4 and of Corollary 3.5. For example, the
rational, proper and the condition in the above definition
dual version of Proposition 3.4 states that if detects ,
is equivalent to the invertibility of , where and
then detects .
denote their respective limits at infinity.
In the sequel, if is an RLS with GO’s then we
The -valued transfer function
say that is stabilizable if is stabilizable, and similarly
is detectable if is detectable. The following theorem (32)
is due to Rebarber [24]. For the sake of completeness, and
also because our proof is very short, we include the proof. corresponds to and connected in parallel, as independent
Theorem 3.8: An RLS is exponentially stable if and only if channels. We see that is an admissible feedback transfer
it is stabilizable, detectable, and input–output stable. function for if and only if is an admissible feedback
Proof: The necessity follows from [38, Sec. 4]. To prove operator for . Indeed, denoting we have
the sufficiency, suppose is a stabilizable and detectable RLS,
with GO’s denoted . Suppose detects
and denote (with its natural domain). The
feedback operator is admissible for the RLS with the GO’s
. For this feedback, formula (28) reads (33)

The closed-loop transfer function [see

(31) (15)] is the transfer function from to in the feedback
By (22) the control operator of the closed-loop system is also connection shown in Fig. 6 (which is the transfer function
so that and are admissible control operators for analogue of Fig. 1). is the transfer function from
the semigroup generated by . By (29) and the exponential to in the same figure and, clearly, .
stability of , we have that and When two systems and are connected as in Fig. 1, it
are in . Now suppose that is input–output stable, i.e., is customary to say that is a realization of a dynamic output
is in . If we multiply all the terms in (31) feedback for (as opposed to static output feedback, shown
with from the right, we get an identity which shows that in Fig. 5). We define and analyze the closed-loop system
is in . By Proposition 3.2, is exponentially corresponding to a dynamic output feedback connection with
stable. the help of the corresponding results for static output feedback.
Proposition 3.2 can be thought of as a particular case of the Let be a well-posed linear system with input, state, and
above theorem, corresponding to and . output spaces and , and let be another well-posed
linear system, with input, state, and output spaces and
IV. DYNAMIC STABILIZATION . Let and denote the transfer functions of and ,
In this section we introduce the concept of a stabilizing respectively. The parallel connection of and is a well-
controller, with and without internal loop, along the same posed linear system with input and output space ,
lines as in finite dimensions (see Section I). We explore the state space , and transfer function given in (32).
connection with stabilizability and detectability. We leave it to the reader to supply the missing details of this
Not every pair of well-posed linear systems and can definition, based on the detailed definition of a well-posed
be connected as in Fig. 1. Even if the input and output spaces linear system, as formulated in [38]. If and are regular,
match, it might happen that the interconnection is not a well- then is more easily defined via its GO as follows:
posed linear system because its transfer function is not well
posed. Actually, we have seen that this can happen even in (34a)
finite dimensions. Thus we are led to the following concept
which is a generalization of the concept of an admissible (34b)
feedback operator.
Definition 4.1: Let and be Hilbert spaces, and let be Here and are the GO’s of and
an -valued well-posed transfer function. An admissible , respectively.
WEISS AND CURTAIN: DYNAMIC STABILIZATION OF REGULAR LINEAR SYSTEMS 13

Definition 4.2: Under the above assumptions, if is an coprime factorizations and graphs of input–output operators
admissible feedback transfer function for , then the closed- play a central role in modern linear control theory. We mention
loop system corresponding to and is the closed- only the recent papers by Georgiou and Smith [10], Logemann
loop system corresponding to with the feedback operator [16], and Ober and Sefton [22] which contain important results
(which is admissible for , as explained earlier). and further references on this subject. In this paper, we are not
Intuitively, is the system shown in Fig. 1, the state- dealing with algebraic or topological aspects of stabilization,
space version of Fig. 6. Now suppose that and are but we are interested in the interplay with the state-space
regular and, using the notation of (34), is invertible theory for RLS’s.
in . Then is regular as well. Indeed, the regularity of Definition 4.5: Let and be well-posed linear systems
is clear from (32) and, according to a result mentioned after with transfer functions and , and assume that is an
Proposition 2.2, is regular iff is invertible. More- admissible feedback transfer function for . Let be the
over, it is straightforward to compute the GO of using closed-loop system corresponding to and . We call
formulas (18)–(22). For example, the semigroup generator of a stabilizing controller for if is exponentially stable.
is shown at the bottom of the page with its natural domain Proposition 4.6: Under the assumptions and the notation
(see Section II). As usual, the subscript means -extension. of Proposition 4.3, assume that both systems and are
We leave it to the reader to compute the other GO. stabilizable and detectable. Then is a stabilizing controller
Proposition 4.3: Let and be RLS’s with feedthrough for if and only if stabilizes .
operators and and with transfer functions and , This follows immediately from the combination of Propo-
respectively. We assume that is an admissible feedback sition 4.3 with Theorem 3.8 (applied to ). Proposition 4.6
transfer function for and is invertible. Let is a generalization of a similar result for Pritchard–Salamon
be the closed-loop system corresponding to and . systems which appeared in [16].
To achieve our goal of generalizing the concept of a
i) If and are stabilizable, then is also
stabilizing controller with internal loop, introduced in Section I
stabilizable.
in the finite-dimensional context, to RLS, we need some
ii) If and are detectable, then is also de-
preparation. The following definition is almost identical to
tectable.
Definition 1.1 (see also Fig. 3).
Proof: As before, let denote the parallel connection Definition 4.7: Let and be Hilbert spaces. Let
of and . Suppose that and are stabilizable. Then and be well-posed transfer functions with values in
it is clear that is stabilizable as well. Since is obtained and in , respectively. is partitioned naturally
from by static output feedback, by Corollary 3.5 must into and . We say that is an admissible
be stabilizable. This proves statement i). The proof of ii) is feedback transfer function with internal loop for if is an
similar. admissible feedback operator for the transfer function , where
Conjecture: The converse of Proposition 4.3 is also true.
An equivalent statement to the converse of i) of Proposition
4.3 is the following: if the parallel connection of two RLS’s is
(35)
stabilizable, then both RLS’s are stabilizable. The converse of
ii) has an analogous reformulation. It seems to us that to prove
this conjecture, we would need linear quadratic optimal control
theory, but a complete theory for unstable RLS’s is lacking. The intuition behind this definition is the same as in finite
More precisely, we would need that with the standard quadratic dimensions.
cost on the input and the state, the finite cost condition implies Proposition 4.8: With the notation of Definition 4.7, sup-
that the minimal cost is achieved by a regular state feedback. pose that:
Definition 4.4: Let and be Hilbert spaces, and let 1) is an admissible feedback transfer function for ;
and be well-posed transfer functions with values in 2) is an admissible feedback operator for de-
and , respectively. We say that stabilizes fined in (6).
if is invertible on some right half-plane and Then is an admissible feedback transfer function with
[defined in (33)] is in . internal loop for .
We mention that since we do not distinguish between two Proof: We partition the operator into four
transfer functions if one is a restriction of the other, the fact blocks as follows:
that stabilizes does not imply that and are defined
on the right half-plane .
Various versions of the above concept of stabilization
(sometimes called internal stabilization) and its relationship to
14 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 1, JANUARY 1997

By Condition 1) of Proposition 4.8, has a well-posed

inverse. This enables us to factor

(36)
Fig. 7. The cascade connection of 6h and 6q through K . The inputs are
where u y v
and and the outputs are and z (in this order).

V. THE DESIGN PROCEDURE

A little computation reveals that, in fact The idea of the design is the same as in finite dimensions
(see Section I), but we have to proceed more carefully. It
is not obvious that the interconnection shown in Fig. 4 (the
controller) makes sense. We shall realize this interconnection
where is the expression defined in (6). By Condition using a feedback and a cascade connection. Then we have to
2) of Proposition 4.8, has a well-posed inverse. This, show that the controller can be connected to the plant, and (the
together with the factorization (36) implies that has a well- most difficult part) we have to show that the whole feedback
posed inverse. system is exponentially stable. For this, we have to compute
The next step is the definition of the closed-loop system. its semigroup generator.
Using the notation of Definition 4.7 (but without assuming the As mentioned in Section II, any interconnection of finitely
invertibility of ), let and be well-posed linear many well-posed linear systems can be represented as a closed-
systems whose transfer functions are and , respectively. loop system, where the open-loop system contains all the
We denote by and the state spaces of these systems. The components connected in parallel. We shall apply this principle
parallel connection of and is a well-posed linear system in the proof of the following lemma which concerns the
which corresponds to the two systems taken together, but cascade connection of two RLS, as shown in Fig. 7.
operating independently. Its state space is , and its Lemma 5.1: Let and be RLS’s with input spaces
inputs and outputs are arranged in such an order that its input , state spaces , and output spaces , respec-
and output spaces are equal, namely, . Its transfer tively. We denote by the GO’s of and
function is from (35). We leave it to the reader to supply similarly for . Let . Consider the operators
the missing details in the definition of , using the precise
definition of a well-posed linear system. If and are
regular, then it is easy to define via its GO, similarly to
(34).
Definition 4.9: Let and be well-posed linear sys-
tems whose transfer functions are and , respectively, and
assume that is an admissible feedback transfer function where and are defined on the natural domain of ,
with internal loop for . Let be the parallel connection which is
of and , with transfer function defined in (35). The
closed-loop system corresponding to and is the
closed-loop system corresponding to with the feedback
operator .
The relevant diagram for the above definition is Fig. 3, but
with three inputs added in the obvious way, similarly to Fig. 6. Then are the GO’s of an RLS with input
One external input should be added to each of , and as space , state space , and output space .
they enter one of the blocks marked or . The outputs are There holds
, and . According to (15), the transfer function of
is . (37)
Definition 4.10: Under the assumptions of Definition 4.9,
is a stabilizing controller with internal loop for if
where and are the transfer functions of and
is exponentially stable.
, respectively. If both and are exponentially stable,
It is easy to extend Proposition 4.6 to this situation.
then so is .
Proposition 4.11: Under the assumptions of Definition 4.9,
Proof: We form the parallel connection of and
if both and are regular, stabilizable, and detectable,
which is an RLS with input space , state space
and if is invertible, where denotes the feedthrough
, output space , and transfer function
operator of , then is a stabilizing controller with internal
loop for if and only if .
As mentioned in Section II, if the space on which
operates, , is finite-dimensional, then the invertibility
of need not be checked since it is a consequence of We leave it to the reader to write down the GO of , as in
the admissibility of the feedback operator . (34). It is easy to see that the operator
WEISS AND CURTAIN: DYNAMIC STABILIZATION OF REGULAR LINEAR SYSTEMS 15

defined by The following proposition contains the three assumptions

on the plant under which we know how to construct a sta-
bilizing controller with internal loop for . The proposition
introduces (among other things) four transfer functions
is an admissible feedback operator for . It follows that there and , which the reader might recognize as being one
is a closed-loop system corresponding to and . If half of a doubly coprime factorization of ; see the discussion
denotes the feedthrough operator of , then and the references at the end of this section.
is invertible (its inverse is ), which implies that Proposition 5.3: Suppose that there exist operators
is regular (see Section II). If we compute the GO of and such that:
using formulas (18)–(22), we find that they are exactly I) stabilizes (see Definition 3.1);
as defined in this lemma. The transfer function II) detects (see Definition 3.6);
of can be computed from (15) and we obtain exactly III) is a regular triple.
expression (37) in the lemma. Then there exists an exponentially stable RLS with input
Suppose that and are exponentially stable. Then space , state space , and output space , whose
the parallel connection is exponentially stable as well. GO’s are the following:
In particular, it is stabilizable and detectable. By Corollary
3.5 and its dual statement for detectability, is stabilizable (38a)
and detectable. From the expression for we see that (38b)
is input–output stable. By Theorem 3.8, is exponentially
stable. where and are defined on the natural domain of .
We call the system defined in the last lemma the We introduce the transfer functions and by
cascade connection of and through . We could have partitioning the transfer function of
defined a simpler cascade connection by taking the second
input of to be zero, ignoring the second output of , (39)
and taking . Then the transfer function of the cascade
connection would have been . However, the way we did
Then has a well-posed inverse and
it in Lemma 5.1 gives us added flexibility to construct more
complex interconnections. A result closely related to Lemma (40)
5.1 can be found in Logemann and Townley [17].
The following lemma complements Lemma 5.1, telling us Proof: It follows from the three assumptions in the
that the -extension of may be computed “component- proposition that the operators
wise.”
Lemma 5.2: With the notation of Lemma 5.1, the -
extension of is
are the GO’s of an RLS which we call (the extended plant).
Its transfer function is

with domain .
Proof: We use the notation from the proof of Lemma We partition into and . Then by Assump-
5.1. According to (23), we have tion II), has a well-posed inverse. Therefore

with , where is the ob-

servation operator of the parallel connection . It is easy is an admissible feedback operator for , giving rise to a
to see that , and by simple closed-loop system . Since
computation we get the formula for .
For the remainder of this section we use the following
notation: and are Hilbert spaces, and is an RLS
with input space , state space , and output space . is regular and its GO can be computed from (18)–(22). If
The GO’s of are and its transfer function is we do this simple computation, we obtain that
. is the plant that we want are given by (38). By Assumption II), is exponentially
to stabilize. stable. From
is an exponentially stable RLS with input space , state (41)
space , and output space . Its GO’s are
and its transfer function is . we see [by looking at the right lower corner and using (39)]
will play the role of the Youla parameter in our controller, that so that has a well-posed inverse
in a sense that we shall explain later. indeed. By looking at the left lower corner of (41), we obtain
16 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 1, JANUARY 1997

, which proves the first identity in (40). From the

right upper corner of (41) we obtain . Using this
fact, the left upper corner of (41) becomes ,
which is the same as the second identity in (40).
The system can be seen as a subsystem of another
system , shown in Fig. 8. Our main result is the following
[compare (42) below with (8)].
Theorem 5.4: We impose the assumptions and the notation 6 6
Fig. 8. The system g , the cascade connection of h and q through 6
of Proposition 5.3, together with the notation K = [0 0 I ] . As usual, the block D is not shown in the diagram. Note that
and introduced earlier. 6 g is very close to the controller shown in Fig. 4. To obtain this controller,
There exists an exponentially stable RLS with input
we only have to take 0 y =y u = r+v
and 0 .

space , state space and output space ,

whose GO’s are the following: introduce the operators and
given by
(42a)

(42b)

The transfer function of is ( and stand for left and right). is obtained from as
follows: the input of is multiplied by , becoming the
(43) input of (thus, ). The output of is multiplied by
and is added to it, then it becomes the output of
The system is a stabilizing controller with internal loop (thus, ). Expressed in formulas
for , and the closed-loop system (see Definition 4.9)
is regular.
Proof: The first part of the proof is to show that there
exists an exponentially stable RLS described by (42) and (43).
We introduce the cascade connection of and through
, as defined after Lemma 5.1. (Thus, in terms of If we carry out the simple computations, we obtain (42). The
the notation of Lemma 5.1, we have , , transfer function of is
.) As in the lemma, we denote this cascade
connection by and its transfer function by . According to
Lemma 5.1 and a very short computation, the GO’s of are
Again, if we perform the computations, we obtain (43). Since
is exponentially stable, so is . This concludes the first
part of this proof. The block diagram of the controller is
shown in Fig. 4, only this time and should be replaced
by and .
In the second part of the proof we determine how the -
extension of acts on a certain subspace of its domain. In
By Lemma 5.1, is exponentially stable. From (37) and fact, we find that it behaves just as expected, in view of the
(39) we have formula (42b) for . More precisely, we show that for any
and for any we have

The system is represented in Fig. 8. In this diagram, the

input of is , its output is , the input of is (44)
, its output is and, assuming zero initial
states, we have To prove (44), we notice that and
for any

where a hat denotes the Laplace transformation, as usual. According to Lemma 5.2, we have that
The system is almost the stabilizing controller we are
looking for, only some small adjustments are needed. We
WEISS AND CURTAIN: DYNAMIC STABILIZATION OF REGULAR LINEAR SYSTEMS 17

and for any and any In the fourth part of the proof we compute . According
to (19) and (25) we have
(47)
with domain
By definition, , which implies that
and, for any , .
(48)
Therefore, if and are as above, then (by a trivial
Here, is the state space of (and of )
computation) (44) holds.
The third part of the proof is to check that is an
admissible feedback transfer function with internal loop for
, and the closed-loop system from Definition 4.9 is and, according to (24)
regular. Condition 1) from Proposition 4.8 is trivially satisfied, with
since . To verify Condition 2), we compute
using (6) and (43) To obtain a more explicit formula out of (47) and (48),
we have to examine how acts on . Denoting
and
Using both identities in (40), this simplifies to with
(45) we see from the block diagonal structure of and that
[this should be compared with (9)]. By Assumption I), (49)
has a well-posed inverse, so Condition 2) holds as
well. According to Proposition 4.8, is an admissible feedback Denoting
operator for , the parallel connection of and , whose with
transfer function is from (35). Thus, the closed-loop system
from Definition 4.9 exists. For later use, we write down we see from the definition of that
the GO of . Using (42) we find that they are . Since , we obtain that
. On the other hand, since is regular, we have
that [see Section II, the text preceding (25)]
and, as already mentioned in the second part of the proof,
. Summarizing, we
have

Using (49), we can write

Taking into account that and

, we obtain
(50)
It is a little exercise in linear algebra to show that is
invertible (regardless if is invertible or not). More We denote by the space appearing in the middle of (50).
precisely Since we may replace by in (48) [see (20)], it is
clear from (50) that we may replace also by in (48).
This has the advantages that we know exactly what is (in
(46) terms of the original data on and ) and [with the help
of (44)] we know exactly how acts on
Therefore, the closed-loop system is regular.
According to Definition 4.10, it remains to prove that
is exponentially stable. We could show this using Proposition
4.11, but for this we would need a doubly coprime factoriza-
tion of , a concept which has not been introduced in this
paper. We shall present this transfer function approach in [8].
Here, we prefer to determine the semigroup generator
of , as in Section I, and to show that it is exponentially Thus, the determination of from (47) and (48) has been
stable. reduced to matrix algebra. Doing the computations, we obtain
18 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 1, JANUARY 1997

(51), shown at the bottom of the page, with its natural domain where
in , that is, all those for which .
The fifth and final part of this proof is to show that
is exponentially stable. To achieve this, we show that is
similar to the semigroup generator of the cascade connection
of two exponentially stable systems. More precisely, let
be the RLS with GO (so that has This should be compared with (43) in which, according to (39)
no output), and let be the cascade connection of the
controller [whose GO’s are given in (42)] and of
through . According to Lemma 5.1, the semigroup
generator of is

The functions constitute a doubly

coprime factorization of over ; see [9], [19], or (closer
According to the last part of Lemma 5.1, is exponentially to our context) [7]. For a study of stabilizing controllers with
stable. internal loop, carried out entirely in the frequency domain, we
The components of the state of are (the state of refer again to [8].
), (the state of ), and (the state of ), in this
order. As in Section I, we prefer to replace by ,
VI. ELIMINATING THE INTERNAL LOOP
and we also permute the components
In this section we consider the following problem: suppose
that is a plant for which the assumptions in Proposition 5.3
are satisfied. Let be a stabilizing controller with internal
loop for which has been obtained by the design procedure,
We denote by the matrix appearing above. In the new i.e., is described by (42). Under what conditions can
variables, the semigroup generator is . we close the internal loop of (not connected to ),
If we do the computations [starting from (51)] we see that obtaining a well-posed system ? If this is possible, then
, which shows that is exponentially stable. is a stabilizing controller for , in the classical sense of
Definition 4.5?
From the last theorem we see that the Conditions I)–III) Our notation and assumptions are the same as in Section V:
from Proposition 5.3 are sufficient for the existence of a the GO’s of are and its transfer function is .
stabilizing controller with internal loop. They are probably The operators and are such that Conditions I)–III) from
not necessary; the regularity assumptions could be eliminated, Proposition 5.3 hold. is an exponentially stable RLS with
in the spirit of Section V of Staffans [32]. transfer function . The stabilizing controller with internal
If Conditions I)–III) are satisfied, then Theorem 5.4 gives loop is defined via its GO in (42), and its transfer function
us a family of regular stabilizing controllers with internal is from (43).
loop, parameterized by the stable (but otherwise arbitrary) Proposition 6.1: With the above notation, the following
regular system . Obviously, this family does not contain conditions are equivalent.
all the regular stabilizing controllers with internal loop for the 1) is an admissible feedback operator for .
given plant. However, if we impose the additional restrictions 2) has a well-posed inverse.
and , and if we look only at transfer
The proof is very simple and hence omitted. If either
functions (ignoring the internal structure), then in a certain
of the conditions in this proposition is satisfied, then (as
sense this family contains them all. For the precise statement
explained in Section II) there exists a closed-loop system
we refer to our paper [8].
corresponding to and . Both the input and the output of
We could also impose the restrictions and
this closed-loop system have two components, corresponding
, and this would lead us to a dual family of stabilizing
to the two components of the input and of the output of
controllers, as explained at the end of Section I, for which
. We define the reduced controller by eliminating the
. The transfer function of
second component of both the input and of the output of
these controllers is of the form
(i.e., taking the second component of the input to be zero
and ignoring the second component of the output). Intuitively,
in Fig. 4, we close the internal loop from to and then

(51)
WEISS AND CURTAIN: DYNAMIC STABILIZATION OF REGULAR LINEAR SYSTEMS 19

consider this system with the input and the output . In Remark 6.4: The two propositions in this section can be
general, is not stable. According to (7) and (43), its transfer reformulated for controllers obtained by the dual design pro-
function is cedure which was outlined at the end of Section V. The
interesting fact is that Condition 2) from Proposition 6.1 is
(52) equivalent to its dual counterpart: has a well-posed
inverse. This is proved in [8]. Moreover, the reduced controller
Proposition 6.2: If either of the conditions in Proposition obtained via the dual design procedure is the same as the
6.1 holds, then the reduced controller is a stabilizing reduced controller obtained via the initial design procedure.
controller for . This should not be too surprising. In both cases, the reduced
Proof: Let be the parallel connection of and , controller is what we see in Fig. 4 after closing the internal
with transfer function from (35) (we have encountered this loop (from to ). The dual formula for (52) is
system in Sections IV and V). According to Theorem 5.4
and Definition 4.10, is an admissible feedback operator
for , and the corresponding closed-loop system is
Example 6.5: We construct a simple stabilizing controller
exponentially stable. On the other hand, it follows from
with internal loop, for which the conditions in Proposition
Condition 1) in Proposition 6.1 that has a well-posed
6.1 are not satisfied, so that it cannot be reduced to a classical
inverse, which implies that
controller. The input, state, and output spaces of our system
are equal, namely . The generating operators and are
represented by block-diagonal infinite matrices (and )

is also an admissible feedback operator for . Let denote

the closed-loop system corresponding to and . By simple
computations, its transfer function is of the form
where and are defined by

(53)

where is the transfer function of ; see (52), and

We check that is a regular triple, following the
denotes an irrelevant entry. According to Proposition 2.2 and
criteria in Proposition 2.1. Clearly, generates a (diagonal)
its reformulation in terms of systems (see the text after the
semigroup on . We see that zero is an eigenvalue of
definition of the closed-loop system in Section II), is an
(with infinite multiplicity) so that is unstable. The operators
admissible feedback operator for and the corresponding
and are admissible for , according to the operator
closed-loop system is . (Intuitively, the feedback
Carleson measure criterion from Hansen and Weiss [12]. To
followed by the feedback leads to the same closed-loop
check Conditions 4) and 5) from Proposition 2.1, we compute
system as the feedback alone.)
the product
Let denote the system obtained by eliminating the third
component of both the input and the output of . As it can
be guessed from (53), is simply the parallel connection of
and , as described before (34), and its transfer function
is from (32). Similarly, let be the system obtained from
by eliminating the third component of both its input and It is easy to see that in the half-plane Re , each diagonal
its output. The fact regarding proved in the previous entry of the above product is bounded by .
paragraph implies that is an admissible feedback operator This implies that
for , and it leads to the closed-loop system . Since
for (54)
is exponentially stable, the conclusion follows according
to Definition 4.5. so that is a regular triple.
Remark 6.3: If either of the conditions in Proposition 6.1 We claim that stabilizes . We have to verify
holds and, moreover, is invertible, then is regular. the three conditions in Definition 3.1. Condition i) is obvious,
This follows from the theory surveyed in Section II; see the and ii) follows from (54). By a simple computation
text after (17). In this case, the GO’s of can be computed
from (19)–(22). The formulas turn out to be rather messy and
we do not reproduce them here. By comparison, the formulas
of given in Theorem 5.4 are relatively simple. In the If we denote , then from
particular case when does not exist (i.e., and we conclude that for all , whence
), the formulas for become simple. They are
as in (4), but with the subscript for and .
20 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 1, JANUARY 1997

This implies that generates an exponentially stable Our main result was that if an RLS with generating operators
semigroup which proves condition iii). By the same argument, is such that stabilizes , detects
detects (indeed, ). and is a regular triple, then there exist stabilizing
Thus, Conditions I) and II) from Proposition 5.3 are satisfied. controllers with internal loop such that the closed-loop system
Condition III) holds as well: is a regular triple since is exponentially stable and regular. This was done by means of
it is the same as . a constructive procedure for designing stabilizing controllers
According to Theorem 5.4, we can construct stabilizing in terms of and an arbitrary exponentially
controllers with internal loop for , based on the above stable RLS whose transfer function plays the role of the
and and any exponentially stable system with transfer Youla parameter. Finally, we gave sufficient conditions for
function . We choose to be the trivial system, i.e., its eliminating the internal loop to obtain a controller of the
state space is and . Then the controller is rather classical type, as well as a finite-dimensional and an infinite-
simple. According to (42), its GO’s are dimensional example for which the internal loop cannot be
eliminated. Further aspects of controllers with internal loop
are investigated in [8].

According to Proposition 6.1, we could transform this

controller into a reduced controller if were to have a REFERENCES
well-posed inverse. A little computation yields that [1] M. Balas, “Modal control of certain flexible dynamic systems,” SIAM
, where J. Control Optim., vol. 16, pp. 450–462, 1978.
[2] R. F. Curtain, “A synthesis of time and frequency domain methods
for the control of infinite-dimensional systems: A system theoretic
approach,” in Control and Estimation in Distributed Parameter Systems,
H. T. Banks, Ed. Philadelphia, PA: SIAM, 1992, pp. 171–224.
[3] , “Robust stabilizability of normalized coprime factors: The
infinite-dimensional case,” Int. J. Contr., vol. 51, pp. 1173–1190, 1990.
Since the sum of the zeros of tends to , does not [4] , “A comparison of finite dimensional controller designs for
have a well-posed inverse. An indirect way of obtaining the distributed parameter systems,” Contr. Theory Advanced Technol., vol.
same result is by looking at the operator which would be the 9, pp. 609–628, 1993.
[5] R. F. Curtain and D. Salamon, “Finite dimensional compensators for
semigroup generator of the reduced controller, if the latter infinite dimensional systems with unbounded input operators,” SIAM J.
were to exist. According to Remark 6.3, this generator should Contr. Optim., vol. 24, pp. 797–816, 1986.
be . We have [6] R. F. Curtain and G. Weiss, “Well posedness of triples of operators in the
sense of linear systems theory,” in Control and Estimation of Distributed
Parameter Systems, F. Kappel, K. Kunisch, and W. Schappacher, Eds.,
vol. 91. Basel: Birkhäuser Verlag, 1989, pp. 41–59.
[7] R. F. Curtain, G. Weiss, and M. Weiss, “Coprime factorization of regular
linear systems,” Automatica, to appear.
[8] R. F. Curtain, M. Weiss, and G. Weiss, “Stabilization of irrational
transfer functions with internal loop,” submitted.
We see that the trace of these blocks tends to so that [9] B. A. Francis, A Course in H1 -Control Theory, vol. 88. Berlin:
Springer-Verlag, 1987.
their eigenvalues are not bounded from above. Therefore, the
1
[10] T. T. Georgiou and M. C. Smith, “Graphs, causality and stabilizability:
infinite matrix built from these diagonal blocks (which would Linear, shift-invariant systems on L2 [0; ),” Math. Contr., Signals
correspond to ) cannot represent a semigroup Syst., vol. 6, pp. 195–223, 1993.
[11] R. V. Gressang and G. B. Lamont, “Observers for systems characterized
generator. by semigroups,” IEEE Trans. Automat. Contr., vol. AC-20, pp. 523–528,
1975.
[12] S. Hansen and G. Weiss, “The operator Carleson measure criterion for
admissibility of control operators for diagonal semigroups on l2 ,” Syst.
VII. CONCLUSION Contr. Lett., vol. 16, pp. 219–227, 1991.
[13] H. Falun, “Characteristic conditions for exponential stability of linear
In this paper, we investigated the concepts of stabilizability, dynamical systems in Hilbert spaces,” Ann. Differential Equations, vol.
detectability, and dynamic stabilization for a general class 1, pp. 43–56, 1985.
of infinite-dimensional systems called RLS’s. This class of [14] I. Lasiecka, “Galerkin approximations of infinite-dimensional compen-
sators for flexible structures with unbounded control action,” Acta
systems allows for unboundedness in the input and output op- Applicandae Mathematicae, vol. 28, pp. 101–133, 1992.
erators, and the concept of stability considered is exponential [15] , “Finite element approximations of compensator design for
stability. Our choice for the definitions of stabilizability and analytic generators with fully unbounded controls/observations,” SIAM
J. Contr. Optim., vol. 33, pp. 67–88, 1995.
detectability was motivated by the desire to obtain as close [16] H. Logemann, “Stabilization and regulation of infinite-dimensional sys-
a generalization as possible of the known finite-dimensional tems using coprime factorizations,” in Analysis and Optimization of Sys-
theory to the class of RLS’s. In particular, we showed that sta- tems: State and Frequency Domain Approaches for Infinite-Dimensional
Systems, R. F. Curtain, Ed., vol. 185. Berlin: Springer-Verlag, 1993,
bilizability is invariant under state feedback, and detectability pp. 102–139.
is invariant under output injection. [17] H. Logemann and S. Townley, “Low-gain control of uncertain regular
linear systems,” SIAM J. Contr. Optim., to appear.
To obtain a natural generalization of dynamic stabilization, [18] D. G. Luenberger, “Observers for multivariable systems,” IEEE Trans.
we introduced a new concept of stabilization by a controller Automat. Contr., vol. AC-11, pp. 190–197, 1966.
with internal loop. Although this concept is new even for finite- [19] J. M. Maciejowski, Multivariable Feedback Design. Wokingham:
Addison-Wesley, 1989.
dimensional systems, in infinite dimensions, it is essential to [20] C. N. Nett, C. A. Jacobson, and M. J. Balas, “Fractional representation
guarantee that the controller be well-posed. theory: Robustness with applications to finite-dimensional control of a
WEISS AND CURTAIN: DYNAMIC STABILIZATION OF REGULAR LINEAR SYSTEMS 21

class of linear distributed systems,” in Proc. 22nd IEEE Conf. Decision [40] , “Two conjectures on the admissibility of control operators,” in
Contr., San Antonio, TX, 1983, pp. 268–280. Estimation and Control of Distributed Parameter Systems, F. Kappel,
[21] , “A connection between state space and doubly coprime frac- K. Kunisch, and W. Schappacher, Eds., vol. 100. Basel: Birkhäuser
tional representations,” IEEE Trans. Automat. Contr., vol. AC-29, pp. Verlag, 1991, pp. 367–378, 1991.
831–833, 1984. [41] J. C. Willems, “Paradigms and puzzles in the theory of dynamical
[22] R. J. Ober and J. A. Sefton, “Stability of control systems and graphs of systems,” IEEE Trans. Automat. Contr., vol. 36, pp. 259–294, 1991.
linear systems,” Syst. Contr. Lett., vol. 17, pp. 265–280, 1991.
[23] J. Prüss, “On the spectrum of C0 -semigroups,” Trans. Amer. Math. Soc.,
vol. 284, pp. 847–857, 1984.
[24] R. Rebarber, “Conditions for the equivalence of internal and external
stability for distributed parameter systems,” IEEE Trans. Automat.
Contr., vol. 38, pp. 994–998, 1993. George Weiss received the Control Engineering
[25] , “Exponential stability of coupled beams with dissipative joints: degree from the Polytechnic Institute of Bucharest,
A frequency domain approach,” SIAM J. Contr. Optim., vol. 33, pp. Romania, in 1981 and the Ph.D. degree in applied
1–28, 1995. mathematics from the Weizmann Institute, Rehovot,
[26] D. Salamon, “Realization theory in Hilbert space,” Math. Syst. Theory, Israel, in 1989.
vol. 21, pp. 147–164, 1989. Since 1991 he has been working at the Ben-
[27] , “Infinite dimensional systems with unbounded control and Gurion University in Beer Sheva, Israel, and has
observation: A functional analytic approach,” Trans. Amer. Math. Soc., recently moved to the School of Engineering at
vol. 300, pp. 383–431, 1987. the University of Exeter, U.K. His research inter-
[28] J. M. Schumacher, “Dynamic feedback in finite- and infinite- ests include distributed parameter systems, operator
dimensional linear systems,” Ph.D. dissertation, Vrije Univ., Ams- semigroups, optimal and robust control, sampling
terdam, 1981; also MC Tract 143, Mathematisch Centrum, Amsterdam, theory, and power electronics.
1981.
[29] , “A direct approach to compensator design for distributed pa-
rameter systems,” SIAM J. Contr. Optim., vol. 21, pp. 823–836, 1983.
[30] M. C. Smith, “On stabilization and the existence of coprime factoriza-
tions,” IEEE Trans. Automat. Contr., vol. 34, pp. 1005–1007, 1989.
[31] O. J. Staffans, “Quadratic optimal control of stable well-posed linear Ruth F. Curtain (M’78–SM’89–F’91) was born
systems,” Trans. Amer. Math. Soc., to appear. in Melbourne, Australia, in 1941. She received the
[32] , “Coprime factorizations and optimal control of well-posed linear B.Sc. degree (hons.) in 1962, the Dip.Ed. degree in
systems,” SIAM J. Contr. Optim., to appear. 1963, and the M.A. degree in mathematics in 1965,
[33] M. Vidyasagar, Control System Synthesis: A Factorization Approach. all from the University of Melbourne, Australia. She
Cambridge, MA: MIT Press, 1985. received the Ph.D. degree in applied mathematics
[34] G. Weiss, “Admissibility of unbounded control operators,” SIAM J. from Brown University, Providence, RI, in 1969.
Contr. Optim., vol. 27, pp. 527–545, 1989. She was an Assistant Professor at Purdue Uni-
[35] , “Weak Lp -stability of a linear semigroup on a Hilbert space versity, Lafayette, IN, from 1970–1971. During
implies exponential stability,” J. Differential Equ., vol. 76, pp. 269–285, the period from 1971 to 1977 she worked at the
1988. University of Warwick, U.K., first as a Research
[36] , “Admissible observation operators for linear semigroups,” Israel Fellow at the Control Centre and subsequently as a Lecturer. Since 1977
J. Math., vol. 65, pp. 17–43, 1989. she has been with the Mathematics Institute at the University of Groningen
[37] , “The representation of regular linear systems on Hilbert spaces,” in the Netherlands, where she is a Professor in Mathematical Systems
in Control and Estimation of Distributed Parameter Systems, F. Kappel, Theory. Her research interests include the area of infinite-dimensional linear
K. Kunisch, and W. Schappacher, Eds., vol. 91. Basel: Birkhäuser systems theory, particularly state and frequency domain representations,

[38]
Verlag, 1989, pp. 401–416.
, “Transfer functions of regular linear systems—Part I: Character-
1
approximation, robust control, and H -control. She is a coauthor of three
books.
izations of regularity,” Trans. Amer. Math. Soc., vol. 342, pp. 827–854, Dr. Curtain has served as an Associate Editor for the international journals
1994. Systems and Control Letters and Automatica. She is currently an Associate
[39] , “Regular linear systems with feedback,” Math. Contr., Signals Editor for Mathematics of Control, Signals and Systems and for the Journal of
Syst., vol. 7, pp. 23–57, 1994. Mathematical Systems, Estimation and Control and an Editor for Automatica.