SIAM J. CONTROL OPTIM.

c 1998 Society for Industrial and Applied Mathematics

Vol. 36, No. 5, pp. 17031749, September 1998 010

ON QUADRATIC DIFFERENTIAL FORMS

J. C. WILLEMS AND H. L. TRENTELMAN

Abstract. This paper develops a theory around the notion of quadratic differential forms in
the context of linear differential systems. In many applications, we need to not only understand
the behavior of the system variables but also the behavior of certain functionals of these variables.
The obvious cases where such functionals are important are in Lyapunov theory and in LQ and H
optimal control. With some exceptions, these theories have almost invariably concentrated on first
order models and state representations. In this paper, we develop a theory for linear time-invariant
differential systems and quadratic functionals. We argue that in the context of systems described by
one-variable polynomial matrices, the appropriate tool to express quadratic functionals of the system
variables are two-variable polynomial matrices. The main achievement of this paper is a description
of the interaction of one- and two-variable polynomial matrices for the analysis of functionals and
for the application of higher order Lyapunov functionals.

Key words. quadratic differential forms, linear systems, polynomial matrices, two-variable poly-
nomial matrices, Lyapunov theory, positivity, spectral factorization, dissipativeness, storage functions

AMS subject classifications. 93A10, 93A30, 93D05, 93D20, 93D30, 93C05, 93C45

PII. S0363012996303062

1. Introduction. In the theory of models for dynamical systems, it has been

customary to consider both external input/output as well as state space models. Also,
there is a well developed theory for passing from one type of model to another. Thus,
there are efficient algorithms for passing from a convolution, to a transfer function,
to a state model, and back. Even for stochastic and nonlinear systems, there are
methods for associating a first order state representation to a high order model.
However, in addition to understanding the interaction between system variables,
we need in many applications to understand also the behavior of certain functionals of
these variables. The obvious cases where such functionals are crucial are in Lyapunov
theory, in the theory of dissipative systems, and in optimal control. In these contexts
it is remarkable to observe that the theory of dynamics has almost invariably con-
centrated on first order models and state representations. Thus, in studying system
stability using Lyapunov methods, we are constrained to consider state representa-
tions, and optimal control problems invariably assume that the cost is an integral
of a function of the state and the input. The question thus occurs of whether it is
possible to develop an external theoryfor example, Lyapunov theoryfor systems
and functionals so that analysis of stability and passivity, for instance, could proceed
on the basis of a first principles model instead of first having to find a state repre-
sentation. In this paper, we consider models that are not in state form (even though
some proofs use state representations). Our models are externally specified yet they
are not completely general first principles models in that we concentrate on models
in kernel or in image representation.
It is the purpose of this paper to develop such a theory. We do not, however,
set our aims too high and start with a very well-understood class of systems and
functionals: linear time-invariant differential systems and quadratic functionals in the
Received by the editors May 6, 1996; accepted for publication (in revised form) September 9,

1997; published electronically June 22, 1998.

http://www.siam.org/journals/sicon/36-5/30306.html
Research Institute for Mathematics and Computing Science, P.O. Box 800, 9700 AV Groningen,

The Netherlands ([email protected], [email protected]).

1703
1704 J. C. WILLEMS AND H. L. TRENTELMAN

system variables and their derivatives. We shall see that one-variable polynomials
are the appropriate tool in which to parametrize the model (see also, among others,
[16], [17]) and two-variable polynomials are the appropriate tool for parametrizing
the functionals. Thus, the paper presents an interesting interplay between one- and
two-variable polynomial matrices. Two-variable polynomials turn out to be a very
effective tool for analyzing linear systems with quadratic functionals.
This paper consists of a series of general concepts and questions, combined with
some specific results concerned with Lyapunov stability and with dissipativity, i.e.,
with positivity of (integrals of) quadratic differential forms. As such, the paper aims
at making a contribution to the development of the very useful and subtle notions of
dissipative and lossless (conservative) systems.
In companion papers, these ideas will be applied to LQ and H problems. The
main achievement of this paperthe interaction of one- and two-variable polynomial
matrices for the analysis of functionals and application in higher order Lyapunov
functionsappears to be new. However, seeds of this have appeared previously in the
literature. We mention especially Brocketts early work on path integrals [7], [8] in
addition to classical work on RouthHurwitz-type conditions (see, for example, [6]),
and early work by Kalman [13], [14].
2. Review. In order to make this paper reasonably self-contained, we first in-
troduce some notation and some basic facts from the behavioral approach to linear
dynamical systems. References in which more details can be found include [31], [32],
and [33].
We will deal exclusively with continuous-time real linear time-invariant differential
dynamical systems. Thus, the time axis is R, the signal space is Rq (the number of
variables q, of course, depends on the case at hand), and the behavior B is the solution
set of a system of linear constant coefficient differential equations
 
d
(2.1) R w=0
dt
in the real variables w1 , w2 , , wq , arranged as the column vector w; R is a real
polynomial matrix with, of course, q columns. The number of rows of R depends,
as do its coefficients, on the particular dynamical system described by (2.1). Hence
we denote this as R Rq [], where denotes the indeterminate. Thus, if R() =
R0 + R1 + + RN N , then (2.1) denotes the system of differential equations
dw dN w
(2.2) R 0 w + R1 + + RN N = 0.
dt dt
For the behavior, i.e., for the solution set of (2.1) or (2.2), it is usually advisable to
consider locally integrable ws as candidate solutions and to interpret the differential
equation in the sense of distributions. However, it is our explicit intention to avoid
mathematical technicalities as much as possible in this paper. In keeping with this,
we assume that the solution set consists of infinitely differentiable functions, even
though many of the results are valid without this assumption. Hence the behavior of
(2.1) is defined as
   
q d
(2.3) B = w C (R, R ) | R w=0 .
dt
We denote the family of dynamical systems obtained this way by Lq . Hence elements
of Lq are dynamical systems = (R, Rq , B) with time axis R, signal space Rq , and
 ON QUADRATIC DIFFERENTIAL FORMS 1705

behavior B described through some R Rq [] by (2.3). Note that instead of writing

Lq we may as well write B Lq , and we prefer to use this notation in this paper.
As explained in the previous paragraphs, each R Rq [] unambiguously defines
a system B Lq . However, there are always many Rs defining the same B Lq . For
example, if U is any unimodular polynomial matrix such that the product U R makes
sense, then R and U R induce the same element of Lq . Also, there are many other
d
ways of specifying a given B Lq . Note that (2.1) describes B as B = ker(R( dt ))
d q rowdim(R)
with R( dt ) viewed as a map from C (R, R ) into C (R, R ). For obvious
reasons we hence refer to (2.1) as a kernel representation of B Lq . We will meet
other representations, in particular image, latent variable, input/output, state, and
input/state/output representations. These are now briefly introduced.
A system B Lq is said to be controllable if for each w1 , w2 B there exists
a w B and a t0 0 such that w(t) = w1 (t) for t < 0 and w(t) = w2 (t t0 ) for
t t0 . It can be shown that B is controllable iff its kernel representation satisfies
rank(R()) = rank(R) for all C. Here, rank(R) is defined as the rank of R
considered as a matrix with elements in the field R() of real rational functions. On
the other hand, for a given C, R() is a matrix with elements in C. Accordingly,
rank(R()) denotes the rank of the complex matrix R(). It is easy to see that
rank(R) = maxC rank(R()).
Controllable systems are exactly those that admit image representations. More
concretely, B Lq is controllable iff there exists an M Rq [] such that B =
d d
im(M ( dt )), with M ( dt ) viewed as a mapping from C (R, Rcoldim(M ) ) into C (R, Rq ).
The resulting representation
 
d
(2.4) w=M `
dt
is called an image representation of B.
An image representation is a special case of what we call a latent variable repre-
sentation of B. The system of differential equations
   
d d
(2.5) R w=M `
dt dt
is said to be a latent variable representation of B Lq if
B = {w C (R, Rq ) | ` C (R, R ) such that (2.5) holds}.
d d
A latent variable representation is said to be observable if (R( dt )w = M ( dt )`1 and
d d
R( dt )w = M ( dt )`2 ) implies (`1 = `2 ). Observability is equivalent to the condition
that M () is of full column rank for all C. A controllable system, it turns out,
always allows an observable image representation.
Of special interest in section 4 will be the observability of the system
   
d d
(2.6) A ` = 0, w = C `.
dt dt
Of course, the definition of observability applies to (2.6). If this is the case, then
we call the pair of polynomial matrices (A, C) with the same number of columns an
observable pair. Hence (A, C) is an observable pair iff
 
A()
(2.7)
C()
1706 J. C. WILLEMS AND H. L. TRENTELMAN

is of full column rank for all C.

Systems in Lq admit many other useful representations. We already encoun-
tered kernel and image representations. Next, we introduce state and input/output
representations.
In [22] the notion of state models and their construction has been discussed in
detail. Here we limit ourselves to the bare essentials. Let B Lq . A latent variable
representation (with the latent variable denoted by x this time) of the form (2.5) is
said to be a state model if, whenever (w1 , x1 ) and (w2 , x2 ) are C -solutions of (2.5)
with x1 (0) = x2 (0), then the concatenation (w1 , x1 )(w2 , x2 ) also satisfies (2.5). Since
this concatenation need not be in C , it need only be a weak solution of (2.5), that
is, a solution in the sense of distributions. State models are governed by equations
of the form (2.5) with special structure. In fact, (2.5) is a state model iff there exist
matrices E, F , and G such that
dx
(2.8) Gw + F x + E =0
dt
is equivalent to (2.5) in the case of state models. Thus (2.8) is called a state repre-
sentation of the behavior B if

B = {w C (R, Rq ) | x C (R, Rn ) such that (2.8) is satisfied}.

Here n denotes the dimension of the vector x. The important feature of (2.8) is
that it is an (implicit) differential equation containing derivatives of order at most
one in x and zero in w. We call a state representation state minimal if among all
state representations of B, n is as small as possible. It is possible to prove that
(2.8) is state minimal iff it is state trim (meaning that for all a Rn there exists
(w, x) C (R, Rq+n ) such that x(0) = a) and observable. The dimension of the state
space of a state minimal representation of B Lq is called the McMillan degree of B.
The notion of McMillan degree usually refers to properties of polynomial matrices.
Actually, for the case at hand this correspondence holds in terms of full row rank
kernel representation matrices R or observable image representation matrices M , but
we do not need this correspondence in this paper.
Every system B Lq also admits an input/output representation. By reordering
the components of the vector w, if need be, we can decompose w into
 
u
(2.9) w=
y

with, in terms of R, rank(R) components for y and q rank(R) components for u,

such that B Lq admits the special kernel representation
   
d d
(2.10) P y=Q u,
dt dt

with P square, det P 6= 0, and P 1 Q a matrix of proper rational functions. Thus, in

(2.10) u has the usual properties of input and y those of output. Therefore (2.10) is
called an input/output representation.
Actually, for controllable systems, we can also recover the input/output structure
in terms of the image representation. Thus the image representation
   d

u U ( dt )
(2.11) = d `
y Y ( dt )
 ON QUADRATIC DIFFERENTIAL FORMS 1707

is an input/output representation if U is square, det U 6= 0, and Y U 1 is a matrix

of proper rational functions. The number of input components of a system in image
representation (2.4) equals rank(M ).
It is possible to combine the above, leading to the familiar input/state/output
representation
dx
(2.12) = Ax + Bu, y = Cx + Du.
dt
This representation is state minimal iff it is observable, i.e., iff (A, C) is an observable
pair of matrices (not to be confused with an observable pair of polynomial matrices).
Summarizing, given any w B, we may partition the components of w into
inputs and outputs. Also, there exists an X Rq [] such that
 
d
(2.13) x=X w
dt
is a (minimal) state map for B. For a system in image representation (2.4) this leads
to a state representation of the form
 
0 d
(2.14) x=X `.
dt
The resulting relation between u and y is as in (2.10); that between w and x is as in
(2.8); and that between u, y, and x is as in (2.12).
We need a few more details about the state construction for systems in image
representation (2.4). Assume that M is of full column rank. Then after permutation
of the components of w (i.e., of the rows of M ), if need be, M is of the form
 
U
M= ,
Y
with U square, det(U ) 6= 0, and Y U 1 a matrix of proper rational functions. The
resulting system
   d

u U ( dt )
= d `
y Y ( dt )
is then an input/output representation. Consider all polynomial row vectors F
R1 [] such that F U 1 is strictly proper. It can be shown (see [22]) that this set is
a vector space. Now
 
d
(2.15) x=X `
dt
is a state map for (2.4) iff the rows of X span this vector space. It is a minimal state
map iff the rows of X form a basis for this vector space.
Next, consider associated with (2.4) the variable v governed by
 
d
v=L `.
dt
Then it follows from the above that there exist matrices P and Q such that
v = P x + Qu
iff LU is proper. Moreover, Q is zero iff LU 1 is strictly proper, and Q is invertible
1

iff LU 1 is biproper.
1708 J. C. WILLEMS AND H. L. TRENTELMAN

3. Quadratic differential forms. Differential equations and one-variable poly-

nomial matrices play an essential role in describing the dynamics of systems, as we
have seen in section 2 and the references given therein. When studying functions of
the dynamical variables, as in Lyapunov theory, studying dissipation and passivity, or
specifying performance criteria in optimal control, we invariably encounter quadratic
expressions in the variables and their derivatives. As we shall see, two-variable poly-
nomial matrices are the proper mathematical tool to express these quadratic func-
tionals. We aim to illustrate throughout this paper that linear dynamical equations
expressed through one-variable polynomial matrices, and quadratic functionals ex-
pressed through two-variable polynomial matrices fit as a glove fits a hand.
Let Rq1 q2 [, ] denote the set of real polynomial matrices in the (commuting)
indeterminates and . Explicitly, an element Rq1 q2 [, ] is thus given by
X
(3.1) (, ) = k` k ` .
k,`

The sum in (3.1) ranges over the nonnegative integers and is assumed to be finite,
and k` Rq1 q2 . Such a induces a bilinear differential form (BLDF), that is, the
map
(3.2) L : C (R, Rq1 ) C (R, Rq2 ) C (R, R)
defined by
X  dk v T  `
dw

(3.3) (L (v, w))(t) := (t) k` (t) .
dtk dt`
k,`

If q1 = q2 (=: q), then induces a quadratic differential form (QDF)

(3.4) Q : C (R, Rq ) C (R, R)
defined by
(3.5) Q (w) := L (w, w).

Define the asterisk operator by

(3.6) : Rq1 q2 [, ] Rq2 q1 [, ]; (, ) := T (, ),
where T denotes transposition. Obviously L (v, w) = L (w, v). If Rqq [, ]
satisfies = , then is called symmetric. The symmetric elements of Rqq [, ]
are denoted by Rsqq [, ]. Clearly
(3.7) Q = Q = Q 12 (+ )
This shows that when considering quadratic differential forms, we can hence in prin-
ciple restrict our attention to s in Rsqq [, ]. However, both bilinear and quadratic
forms are of interest to us.
Associated with Rq1 q2 [, ], we can form the matrix

00 01
10 11

.. .. ..
(3.8)
= . . . .


k`
.. .. ..
. . .
 ON QUADRATIC DIFFERENTIAL FORMS 1709

Note that, although is an infinite matrix, all but a finite number of its elements are
zero. We can factor as = N T M , with N and M infinite matrices having a finite
number of rows and all but a finite number of elements equal to zero. This decompo-
sition leads, after premultiplication by [Iq1 Iq1 Iq1 2 ] and postmultiplication by
col[Iq2 Iq2 Iq2 2 ], to the following factorization of :

(3.9) (, ) = N T ()M ().

This decomposition is not unique, but if we take N and M surjective, then their
number of rows is equal to the rank of . The factorization (3.9) is then called a
canonical factorization of . Associated with (3.9), we obtain the following expression
for the BLDF L :
   T  
d d
(3.10) L (w1 , w2 ) = N w1 M w2 .
dt dt

Next we discuss the case that is symmetric. Clearly = iff is symmetric. In

that case, it can be factored as = M T M M with M an infinite matrix having a
finite number of rows and all but a finite number of elements equal to zero, and M
a signature matrix, i.e., a matrix of the form
 
Ir+ 0
.
0 Ir

This decomposition leads to the following decomposition of :

(3.11) (, ) = M T ()M M ().

Also, this decomposition is not unique but if we take M surjective, then M is unique.
We denote this M as and the resulting pair (r , r+ ) by ( , + ). This pair is
called the inertia of . The resulting factorization

(3.12) (, ) = M T () M ()

is called a symmetric canonical factorization of . Of course, a symmetric canonical

factorization is not unique. However, they can all be obtained from one by replacing
M () by U M () with U Rrank()rank() such that U T U = .
Associated with (3.11), we obtain the following decomposition of Q into a sum
of positive and negative squares:
   
d 2 d
(3.13) Q (w) = kP wk kN wk2 ,
dt dt

where N, P Rq [] are obtained by partitioning M conform M as:

 
P
(3.14) M = .
N

For a given symmetric (, ) we are also interested in the symmetric two-variable

polynomial matrix ||(, ), the absolute value of , which we define as follows. For
a given real symmetric matrix A Rnn define its absolute value, |A| Rnn , as the
unique symmetric nonnegative definite matrix X Rnn such that X 2 = A2 . This
1710 J. C. WILLEMS AND H. L. TRENTELMAN

matrix |A| can be computed as follows. Factor A = U T U , where is the diagonal

matrix with, on its diagonal, the nonzero eigenvalues of A in decreasing order, and
where U U T = I, and define |A| := U T ||U with || defined in the obvious way. Let
be the symmetric matrix associated with (, ). Let || be the absolute value of
. Next, define ||(, ) as the symmetric two-variable polynomial matrix associated
with ||:
T
I I
I I

||(, ) := 2 I || 2 I .

.. ..
. .

Note that a factorization = U T U immediately

p yields a symmetric canonical fac-
torization of (, ). Indeed, define Mc := ||U . We then have
p p
(3.15) = U T || ||U = McT Mc ,

with Mc surjective. The corresponding Mc () := Mc col [I I 2 I . . .] then

yields a canonical factorization (, ) = McT () Mc (). This particular canonical
factorization has the property that

(3.16) ||(, ) = McT ()Mc ().

In general, if M () is any canonical factor of , then we have U M () = Mc () with

U satisfying U T U = , and hence ||(, ) = M T ()U T U M ().
One of the conveniences of identifying BLDFs and QDFs with two-variable poly-
nomial matrices is that they allow a very convenient calculus. One instance of this
d
is differentiation. Obviously if L is a BLDF, so is dt L , and if Q is a QDF, so
d
is dt Q . The result of differentiation is easily expressed in terms of the two-variable
polynomial matrices and leads to the dot operator defined as

(3.17) : Rq1 q2 [, ] Rq1 q2 [, ]; (, ) := ( + )(, ).

It is easily calculated that

d d
(3.18) L = L and Q = Q
dt dt

In the following, an important role is played by certain one-variable polynomial ma-

trices obtained from two-variable polynomial matrices by means of the delta operator
, defined as

: Rq1 q2 [, ] Rq1 q2 []; () := (, ).

Note that, among other things, this allows one to associate a differential operator
d d
( dt , dt ) with a QDFthis is one of the key ingredients in LQand variational
problems.
Introduce the star operator ? acting on matrix polynomials by
?
: Rq1 q2 [] Rq2 q1 []; R? () := RT ().
d d
The importance of this operation stems from the fact that M ( dt ) and M ? ( dt ) are
qq
formal adjoints as differential operators. A polynomial matrix M R [] is called
 ON QUADRATIC DIFFERENTIAL FORMS 1711

para-Hermitian if M = M ? . Note that ()? = ( ). Hence if Rsqq [, ], then

is para-Hermitian.
In addition to studying BLDFs and QDFs as maps to C (R, R), we are interested
in their integrals. In order to make sure that those integrals exist, we assume in
this case that the arguments have compact support. As is common, we denote by
D(R, Rq ) := {w C (R, Rq ) | w has compact support}. Let Rq1 q2 [, ]. Then
obviously L : D(R, Rq1 ) D(R, Rq2 ) D(R, R). Consider the integral
Z
(3.19) L : D(R, Rq1 ) D(R, Rq2 ) R

defined as
Z Z +
(3.20) L (v, w) := L (v, w)dt.

R
The notation Q follows readily from this. Furthermore, consider the same integral
over a finite interval [t1 , t2 ]
Z t2
(3.21) L (v, w)dt
t1
R t2
denoted as t1 L . We call this integral independent of path if for any t1 and t2
the result of the integral (3.21) depends only on the values of v and w and (a finite
number of) their derivatives at t = t1 and t = t2 but not on the intermediate path
used to connect these endpoints, assuming, of course that v C (R, Rq1 ) and w
C (R, Rq2 ). R
The questions of when the map L is zero and when path independence holds
are studied next.
Theorem 3.1. Let Rq1 q2 [, ]. Then the following statements are equiva-
lent: R Rt
1. L = 0, equivalently t12 L is independent of path.

2. There exists a Rq1 q2 [, ] such that = , equivalently, such that
d
L = dt L . Obviously is given by

(, )
(3.22) (, ) = .
+
3. = 0, i.e., (, ) = 0.
The same equivalence holds for QDFs. Simply assume Rsqq [, ] and replace
the Ls by Qs in 1 and 2.
Proof. For the proof, see the appendix.
The importance of this theorem is that condition (3) gives a very convenient way
of checking (1) or (2). Path integrals and path independence featured prominently in
Brocketts work in the sixties (see [7], [8]), and indeed some of our results can be viewed
as streamlined versions of this work. Another potentially interesting connection of
the above theorem and our paper with the existing literature is [3] where, in our
notation, (, ) = R()M () is studied, with R and M associated with a kernel
representation (2.1) and an image representation (2.4) of a controllable system. This
defines an intriguing path independent BLDF that can be associated with any
controllable B.
1712 J. C. WILLEMS AND H. L. TRENTELMAN

In this paper we also study the behavior of QDFs evaluated along a differential
behavior B Lq . In order to do so, it is convenient to introduce an equivalence
relation on both the one- and two-variable polynomial matrices modulo a given B
Lq .
B d d
Let D1 , D2 Rq []. Define (D1 = D2 ) : (D1 ( dt ) D2 ( dt ))B = 0. Let
B
1 , 2 Rsqq [, ]. Define (1 = 2 ) : (Q1 (w) = Q2 (w) for all w B). These
equivalencies are easily expressed in terms of a kernel or an image representation of
B.
Proposition 3.2. Let R Rq [] define a kernel representation of B Lq .
B
Then D1 = D2 iff

(3.23) D1 D2 = F R
B
for some F R [] and 1 = 2 iff

(3.24) 2 (, ) = 1 (, ) + RT ()F (, ) + F (, )R()

for some F Rq [, ]. Let M Rq [, ] define an image representation of

B
B Lq . Then D1 = D2 iff

D1 M = D2 M
B
and 1 = 2 iff

M T ()1 (, )M () = M T ()2 (, )M ()

Proof. For the proof, see the appendix.

The first equivalence in the above proposition was already proven in [23], with an
account of the history of the result, which goes back to 1895. We will return to the
second equivalence at the end of section 4.
We now briefly discuss positivity of QDFs. This will be a major issue in the
following; here we restrict our attention to the basic definitions.
Definition 3.3. Let Rsqq [, ]. We call the QDF Q nonnegative, denoted
0, if Q (w) 0 for all w C (R, Rq ), and positive, denoted > 0, if 0
and if the only w C (R, Rq ) for which Q (w) = 0 is w = 0.
Using the matrix representation of , it is easy to see that 0 iff there exists
D Rq [] such that (, ) = DT ()D(). Simply factor as = DT D and take
D() = D col [Iq Iq Iq 2 ]. Moreover > 0 iff this D has the property that D()
d
is of rank q for all C; in other words, iff the image representation w = D( dt )`
qq
defined by D is observable. Note that, for Rs [, ], we always have || 0.
We are also interested in QDFs which are zero or positive along a behavior B Lq .
B
Definition 3.4. We call zero along B, denoted = 0, if Q (w) = 0 for all
B B
w B. The notions of nonnegative () and positive (>) along B follow readily.
d
Note that it immediately follows from Proposition 3.2 that, if R( dt )w = 0 is a
B
kernel representation of B, then = 0 iff it can be written as (, ) = F (, )R()+
RT ()F (, ). A similar result holds for positivity as follows.
Proposition 3.5. Let Rsqq [, ], B Lq , and R Rq [] induce a kernel
representation of B. Then
B B
(i) 0 iff there exists 0 Rsqq [, ] with = 0 and 0 0;
 ON QUADRATIC DIFFERENTIAL FORMS 1713

B B
(ii) > 0 iff there exists 0 Rsqq [, ] with = 0 and 0 (, ) = DT ()D(),
with (R, D) an observable pair.
Proof. For the proof, see the appendix.

4. Lyapunov theory. Lyapunov theory is a firmly established and very use-

ful technique for establishing stability. It pertains to systems described by explicit
first order differential equations. However, as argued in [33], models obtained from
first principles are seldomly in first order form, will contain latent variables, and
may contain high order derivatives. Writing them in explicit first order form without
introducing spurious solutions may not be an easy matter. Moreover, stability con-
siderations do not require systems to be in first order form. In fact, historically the
very first stability questions and results, as Maxwells statement of the stability prob-
lem and the RouthHurwitz conditions, pertain to high order differential equations.
Oddly enough, to our knowledge, no attempts seem to have been made to establish
Lyapunov theory for high order differential equations. This is the purpose of this
section. We limit our attention, however, to linear differential systems and to Lya-
punov functions that are quadratic differential forms, but we recognize the urgency of
generalizing this work to nonlinear systems. We should remark that in this section for
stability, we only consider systems in which the latent variables have been eliminated,
although latent variables do not cause essential difficulties in the context of stability.
First, we introduce the notion of stability. We say that a system B Lq is
asymptotically stable if (w B) (w(t) 0) and stable if (w B) (w is
t
bounded on the half-line [0, )). For a system B Lq to be (asymptotically) stable
it has to be autonomous. A system B Lq is said to be autonomous if (w1 , w2 B)
and (w1 (t) = w2 (t) for t < 0) imply (w1 = w2 ). It is easy to see that the system with
d
kernel representation R( dt )w = 0 is autonomous iff rank(R) = q; in particular, if R
is square and det(R) 6= 0.
Definition 4.1. Let R Rq []. The complex number C is said to be a
singularity of R if rank(R()) < rank(R); R is said to be Hurwitz if rank(R) = q and
if R has all its singularities in the open left half of the complex plane.
Thus a square R Rqq [] is Hurwitz iff det(R) is a Hurwitz polynomial, i.e., a
nonzero polynomial with its roots in the open left half-plane. We record the following
classical result for easy reference.
Proposition 4.2. The system with kernel representation (2.1) is asymptotically
stable iff R is Hurwitz.
Our most basic Lyapunov theorem regarding high order systems is the following.
Theorem 4.3. Let B Lq . Then B is asymptotically stable iff there exists
B B
Rsqq [, ] such that 0 and < 0.
Proof. For the proof, see the appendix.
2
Example 4.4. Consider the scalar system described by w + dw d w
dt + dt2 = 0. Consider
2 2
the QDF w2 + ( dt )2 . Its derivative QDF is 2(w + dt2 ) dt . Since w B iff w + ddtw
dw d w dw
2 =
dw dw 2
dt , we see that this QDF is B-equivalent to the QDF 2( dt ) . Finally observe that

(1 + + 2 , 2) is an observable pair. Hence 2( dw 2
dt ) is negative on B. Theorem
4.3 establishes asymptotic stability (a rather trivial matter for the case at hand).
Our aim was to show the use of Lyapunov theory without getting involved with state
representations (admittedly also a trivial matter).
1714 J. C. WILLEMS AND H. L. TRENTELMAN

Example 4.5. Consider the multivariable system

dw d2 w
(4.1) Kw + D + M 2 = 0,
dt dt
with K, D, M Rqq , K = K T 0, D + DT 0, and M = M T 0. Such second
order equations occur frequently as models of (visco-)elastic mechanical systems. Take

(, ) = K + M . Then (, ) = K( + ) + M ( 2 + 2 ) which is obviously
B-equivalent to (D + DT ). Thus, asymptotic stability follows if
 q 
(4.2) K + D + M 2 , (D + DT )

is an observable pair. This is the case, for example, if {0} = ker(K) ker(D + DT )
ker(M ). Indeed, under this condition
 
Kp+ D + M 2
(D + DT )
has full column rank for all C.
State representations of autonomous systems take a very special form. Indeed, it
is easy to see that B Lq is autonomous iff it admits a state representation of the
form dxdt = Ax, w = Cx. Such state representations are automatically state trim.
If (A, C) is observable, then they are state minimal. It also follows that for every
d B
D Rq [] there exists a matrix H Rn such that D( dt )w = Hx, i.e., every
q
linear differential operator acting on an autonomous B L is B-equivalent to an
instantaneous function of the state. An analogous statement holds, of course, for
QDFs.
Viewed from this perspective, one can regard Theorem 4.3 as being about state
systems and in this sense not very different from classical Lyapunov theorems. The
point of Theorem 4.3 is twofold:
1. It avoids the state construction which algorithmically (and conceptually) is
not always easy in the multivariable case; and
2. It has the usual Lyapunov theory as a special case by applying it to systems
in first order form and using memoryless QDFs. For the sake of completeness,
we record this as a corollary.
Corollary 4.6. Let B be the behavior of dw dt = Aw. Let (, ) = 0 with
B
0 Rqq , 0 = T0 0. Then (, ) = A0 + 0 AT =: 0 . Whence, if
0 = T0 0 and if (A, 0 ) is an observable pair of matrices, B is asymptotically
stable.
Proof. For the proof, see the appendix.
In section 3, we discussed B-positive QDFs. When B is autonomous, it is useful
to consider also a stronger concept. Let B Lq and Rsqq [, ]; we call strongly
B B
B-positive (denoted  0) if 0 and if (w B and Q (w)(0) = 0) imply (w = 0).
B B B
It is easy to see that  0 implies > 0 and that in order for  0, B must be
B
autonomous. In fact, ((Q (w)(0) = 0) (w = 0)) by itself already implies (  0
B
or  0). Using this notion we arrive at the following refinement of theorem 4.3.
B B
Proposition 4.7. If 0 and < 0, then B is asymptotically stable and
B
 0.
 ON QUADRATIC DIFFERENTIAL FORMS 1715

Proof. For the proof, see the appendix.

We now formulate a stronger version of the only if part of Theorem 4.3.
Theorem 4.8. Assume that B Lq is asymptotically stable. Then for any
B
Rsqq [, ] there exists a Rsqq [, ] such that = ; is unique up to
B B B B
B-equivalence in the sense that, if 1 = and 2 = , then 1 = 2 . If 0,
B B B
then 0, and if < 0, then  0.
In order to compute from , the following algorithm may be used. Let R
Rq [] induce a kernel representation of B. Consider the polynomial matrix equation

(4.3) X T ()R() + RT ()X() = (, )

in the unknown X Rq []. Then (4.3) has a solution. Let X0 be a solution. If R

is square, then all its solutions can be obtained from this one as

(4.4) X() = X0 () + F ()R(),

where F ranges over all polynomial matrices of appropriate size satisfying

(4.5) F T () = F ().

Consider any Y Rq [, ] such that

(4.6) Y (, ) = X()

and compute

(, ) Y (, )R() RT ()Y (, )
(4.7) (, ) = .
+
B B B B
Then = . Since any two 1 , 2 such that 1 = and 2 = satisfy 1 = 2 ,
any other solutions of (4.3) and/or (4.6) yield s in (4.7) that are B-equivalent.
Proof. For the proof, see the appendix.
Theorem 4.8 is more than a mouthful and so we illustrate it for ordinary state
space systems dw T
dt = Aw. Let = R
nn
. Then (4.3) becomes

(4.8) X T ()(A I) + (AT + I)X() = .

This equation has a constant solution which must be symmetric, X0 = X0T , the
solution of the ordinary Lyapunov equation

(4.9) X0 A + AT0 X = .

Choose Y = X0 and verify that (4.7) reduces to = X0 , whence the Lyapunov

function is Q (w) = wT X0 w and for w B its derivative is Q (w) = wT w. Because
of this analogy, we refer to (4.3) as the polynomial matrix Lyapunov equation.
The above shows that it seems to suffice to consider Lyapunov functions and
their derivatives that are of lower degree than that of R. That is, in fact, a general
feature of the equations in Theorem 4.8. However, in order to formalize this, we
return first to the notion of B-equivalence of differential operators in the case that
B Lq is autonomous.
1716 J. C. WILLEMS AND H. L. TRENTELMAN

Let B Lq be autonomous. Then there always exists a square kernel repre-

d
sentation for it. Let R Rqq [] be such that B = ker(R( dt )). We assume in the
remainder of this section that R is square.
Let D Rq []. We call D R-canonical if DR1 is a matrix of strictly proper
rational functions. Let Rsqq [, ]. We call R-canonical if (RT ())1 (, )
(R())1 is a matrix of strictly proper two-variable rational functions. (Note that
there is no ambiguity about what strictly proper means for these two-variable ra-
tional functions.) Since for autonomous systems all differential operators can be seen
as instantaneous functions of the state, it is clear that for any D there exists a canon-
ical D0 that is R-equivalent to D0 . The aim of the next result is to derive this also for
QDFs.
Proposition 4.9. Let D Rq []. Among all differential operators B-equiva-
lent to D, there is exactly one, D0 , which is R-canonical. This D0 can be computed
as follows. Compute DR1 Rq () and write it as DR1 = P + S, with P
the polynomial part and S the strictly proper rational part of DR1 . Then D0 =
D P R. Let Rsqq [, ]. Among all QDFs B-equivalent to there is exactly
one, 0 , which is R-canonical. This 0 can be computed as follows. Write as
(, ) = M T ()N (). Compute the R-canonical representatives M 0 of M and N 0 of
N . Then 0 (, ) = M 0T ()N 0 ().
Proof. For the proof, see the appendix.
The following proposition shows that B-positivity reduces to positivity of the
B-canonical representative.
Proposition 4.10. If is R-canonical, then we have
B
(i) ( = 0) ( = 0),
B
(ii) ( 0) ( 0) ((, ) = DT ()D() with D R-canonical),
B
(iii) ( > 0) ( > 0 and (, ) = DT ()D() with (R, D) observable)
((, ) = DT ()D() with (R, D) observable and D R-canonical).
Proof. For the proof, see the appendix.
We immediately obtain the following consequence of Theorem 4.3.
Corollary 4.11. B Lq is asymptotically stable iff there exists a Rsqq [, ],
0, such that the R-canonical representative of ( + )(, ), computed as in
Proposition 4.9, is 0 and factors as DT ()D() with (R, D) observable.
Our next result is perhaps the most useful of all. It shows how to walk through
the algorithm of Theorem 4.8 and preserve canonicity.
Theorem 4.12. Assume that B Lq is asymptotically stable and has kernel rep-
resentation (2.1) with R square. Assume that is R-canonical. Then the polynomial
matrix Lyapunov equation (4.3) has a unique R-canonical solution. Denote it by X 0 .
Then
(, ) X 0 T ()R() RT ()X 0 ()
(4.10) (, ) =
+
B
is the unique R-canonical such that = . Hence if 0, then 0, and if, in
B
addition, (, ) = DT ()D() with (R, D) observable, then  0.
Proof. For the proof, see the appendix.
We make a short comment relating these results to state representations. The
state maps (2.13) associating a minimal state to B are uniquely defined up to B-
equivalence. There is, consequently, a minimal state map (unique up to premultipli-
d
cation by a nonsingular matrix that is R-canonical, say, x = X( dt )w). An R-canonical
 ON QUADRATIC DIFFERENTIAL FORMS 1717

Rsqq [, ] is of the form Q (w) = xT x, i.e., (, ) = X T ()X(), with =

B
T an (n n) matrix. Of course for this there holds ( 0) ( 0) ( 0);
B
furthermore, ( > 0) ( 0 and observability of the pair of matrices (A, ) (with
B
A associated with the state x)), and finally (  0) ( > 0).
The above results allow generalizations to unstable systems. Let us briefly men-
B
tion a few. We have seen that (B asymptotically stable) ((, ) such that ( 0
B B B
and < 0)). There also holds (B stable) ((, ) such that ( > 0 and 0))
B B
and (an autonomous B is not stable) ((, ) such that ( 6 0 and < 0)).
Furthermore, the result that a Lyapunov function can be constructed so that it
has a given derivative (Theorem 4.8) can be generalized to autonomous systems,
as long as they have the property that if is a singularity of R then will not
be a singularity of R. As such this theorem extends in this sense to a large class of
unstable systems.
We close this section with two extensive examples.
Example 4.13. In this first example we use Theorem 4.8 in order to give a Lya-
punov proof of the RouthHurwitz test for stability of scalar systems. Let R R[]
d
be a Hurwitz polynomial. Hence R( dt )w = 0 defines an asymptotically stable scalar
system. Take, for the derivative of the Lyapunov function,
1 1
(4.11) (, ) = R()R() = R? ()R? ().
2 2
Then obviously, since R has no imaginary axis roots, (R, R? ) is an observable (i.e., a
coprime) pair. The polynomial matrix Lyapunov equation (4.3) yields X() = 14 R()
as a solution. Take Y (, ) = X(). Then (4.7) yields

1 R()R() R()R()
(4.12) B(, ) =
2 +
as a Lyapunov function. Note that this Lyapunov function can be written directly
from the system parameters, without having to solve linear equations! This fact is
actually well known, even though it is not presented in the vein of providing a higher
order Lyapunov function ([6], [13], [14]).
The two-variable polynomial B defined by (4.12) is called the Bezoutian of R.
B
Note that B (, ) = 12 R()R(); B is R-canonical, but B is not. However, B
is B-equivalent to 12 R()R() + 12 R()R(), which is. If we take this for the
in Theorem 4.8, then the Lyapunov equation yields X = 0. Taking Y = 0 then
also yields the Bezoutian (4.12) as the corresponding (hence R-canonical) Lyapunov
function B.
A close examination of the arguments involved yields the equivalence of the follow-
ing three conditions on a polynomial R of degree n and the corresponding B R[, ]
given by (4.12):
1. R is Hurwitz,
2. B 0 and (R, R? ) is coprime,
3. B (the constant matrix associated with B) has rank n and is 0.
The Lyapunov function (4.12), the Bezoutian, is a very useful one for deriving
various stability tests. It is a classical concept in stability (see [11] for a recent
reference). Let us illustrate its usefulness by deriving the Routh stability test from it.
1718 J. C. WILLEMS AND H. L. TRENTELMAN

Let R R[] be a polynomial of degree n. Decompose R in its even and odd parts as

R() = E0 ( 2 ) + E1 ( 2 ).

Form the Routh table by computing the polynomials E2 , E3 , . . . , En as

Ek () = 1 (Ek1 (0)Ek2 () Ek2 (0)Ek1 ()).

Assume for simplicity that R(0) = E0 (0) 0. Rouths stability criterion states that
R is Hurwitz iff all elements of the Routh array E0 (0), E1 (0), . . . , En (0) are positive.
Define Rk () = Ek1 ( 2 ) + Ek ( 2 ) for k = 1, . . . , n, and let Bk be the Bezoutian
associated with Rk . Examining expression (4.12) yields, after a simple calculation

(4.13) Ek (0)Bk (, ) = Bk+1 (, ) + Ek1 (0)Ek ( 2 )Ek ( 2 )

for k = 1, . . . , n (define Bn+1 = 0). Assume that E0 (0), E1 (0), . . . , En (0) are all
positive. Then we obtain (note that B = B1 )
n
X
B(, ) = k k1 k1 Ek ( 2 )Ek ( 2 ),
k=1

where k = Ek1 (0)/E1 (0)E2 (0) . . . Ek (0). Obviously B 0 and has rank n. There-
fore R is Hurwitz. To show the converse, assume that R is Hurwitz. Then E0 (0) > 0.
Also, B 0 and has rank n. Therefore, by (4.13), B2 0 and has rank n 1. Hence
R2 is Hurwitz, and E1 (0) > 0. Now proceed by induction.
The key point thus is that
n
X   2  k1 2
d d
k Ek w
dt2 dtk1
k=1

is a QDF which is well defined and nonnegative definite when the Routh conditions
are satisfied. It has derivative
   2    2 !
1 d d
R w R w ,
2 dt dt

d
which is obviously nonnegative definite along solutions of R( dt )w = 0.
Example 4.14. Let E1 , E2 , . . . , EN and O1 , O2 , . . . , ON 0 be two sets of real polyno-
mials, and assume that Rk,` () := Ek ( 2 ) + O` ( 2 ) is Hurwitz for all k = 1, 2, . . . , N
and ` = 1, 2, . . . , N 0 . Then any combination
0
N
X N
X
2
R() = k Ek ( ) + ` O` ( 2 )
k=1 `=1

is also Hurwitz whenever all the k s and ` s are positive. In order to see this, simply
observe that, in the obvious notation, (4.12) yields
0
N X
X N
B(, ) = k ` Bk,` (, )
k=1 `=1

and the conclusion follows.

 ON QUADRATIC DIFFERENTIAL FORMS 1719

This may be applied to interval polynomials. Assume that R R[] is given by

R() = R0 + R1 + + Rn n , with Rk [ak , Ak ]. The question arises under what
conditions all these polynomials are Hurwitz. The weak Kharitonov test states that
this is the case iff the 2n extreme polynomials, that is, those obtained by replacing
each Rk by ak or Ak , are all Hurwitz. This result is an immediate consequence of the
above. With a little bit of extra work, we can also obtain the strong Kharitonov test
[15] which states that the interval polynomials are Hurwitz iff the four Kharitonov
polynomials obtained by taking the initial sequences a0 , a1 , A2 , . . ., or a0 , A1 , A2 , . . .,
or A0 , A1 , a2 , . . ., or A0 , a1 , a2 , . . ., and continuing by alternating between two con-
secutive maxima and minima, are all Hurwitz. Indeed (see [20]), observe that for all
, R(i) lies in the rectangle in the complex plane spanned by the four points ob-
tained by taking for R the Kharitonov polynomials. This rectangle does not contain
the origin since, by the above, the convex hull of the Kharitonov polynomials contains
only Hurwitz polynomials if the Kharitonov polynomials are themselves Hurwitz.
5. Average positivity. Up to now, we have considered positivity of QDFs and
its use in establishing stability through Lyapunov functions. However, in many appli-
cations, especially in control theory, we are
R interested in an average type of positivity.
In section 3, we already discussed when Q is zero. We now study when it is posi-
tive. With an eye towards applications in LQ and H control we have to distinguish
several (unfortunately not less than three) types of average positivity. All of them
have quite logical definitions.
Definition 5.1. Let Rsqq [, ]. The QDF Q (or simply ) is said to be
R R +
1. average nonnegative, denoted Q 0, if Q (w)dt 0 for all w
D(R, Rq ),
R R R +
2. average positive, denoted by Q > 0, if Q 0 and if Q (w)dt = 0
implies w = 0,
R per
3. strongly average positive, denoted Q > 0, if for all nonzero periodic w
RT
C (R, Rq ) there holds T1 0 Q (w)dt > 0, where T denotes the period of w.
Note that (3) looks somewhat different from the other definitions in this paper
since, for the first time, periodic functions are involved. Actually, wherever in the
paper definitions refer to compact support functions, they could have been written
just as well in terms of periodic functions. However, strong average positivity is the
only instance where the converse is not true.
Proposition
R 5.2. Let Rsqq [, ]. Then
(i) (R Q 0) ((i) 0 R).
(ii) ( Q > 0) ((i) 0 R and det() 6= 0).
R per
(iii) ( Q > 0) ((i) > 0 R).
Proof. For the proof, see the appendix.
Concerning the equivalence (ii), note that (i) 0 R and det() 6= 0
is equivalent to: (i) > 0 for all but finitely many R.
Intuitively, we think of Q (w) as the power going into a physical system. In
many applications, the power is indeed a quadratic differential
P form of some system
variables. For example, in mechanical systems, it is k Fk dq k
dt with Fk the external
force acting P on the system, and qk the position of the kth pointmass; in electrical
circuits it is k Vk Ik , with Vk the potential and Ik the current going into the circuit
at the kth terminal. Note that in these examples the variables are themselves also
related. When this relation is expressed as an image representation, then we obtain
a general QDF in terms of latent variables for the power delivered to a system.
1720 J. C. WILLEMS AND H. L. TRENTELMAN

Average nonnegativity states that the net flow of energy going into the system
is nonnegative: the system dissipates energy. Of course, sometimes energy flows into
the system, while at other times it flows out of it. This outflow is due to the fact
that energy is stored. However, because of dissipation, the rate of increase of storage
cannot exceed the supply. This interaction between supply, storage, and dissipation
is now formalized.
Definition 5.3. Let Rsqq [, ] induce the QDF Q . The QDF Q induced
by Rsqq [, ] is said to be a storage function for if

d
(5.1) Q Q .
dt
A QDF Q induced by Rsqq [, ] is said to be a dissipation function for if
Z Z
(5.2) 0 and Q = Q .

The next proposition shows that one can always interpret average positivity by an
instantaneous positivity condition involving the difference between the rate of change
of storage function and the supply rate.
Proposition
R 5.4. The following conditions are equivalent:
1. Q 0,
2. admits a storage function,
3. admits a dissipation function.
Moreover, there is a one-one relation between storage and dissipation functions,
and , respectively, defined by
d
Q (w) = Q (w) Q (w)
dt

equivalently, = , i.e.,

(, ) (, )
(5.3) (, ) = .
+

Proof. For the proof, see the appendix.

Of course, we should expect that a storage function is related to memory, to state.
The question, however, is: the state of which system? After all, we are considering a
QDF, not a dynamical system. However, the factorization of as

(5.4) (, ) = M T ()M M ()

discussed earlier in section 3 allows us to introduce a state for the QDF Q . Indeed,
(5.4) induces the dynamical system in image representation
 
d
(5.5) v=M w.
dt

Note that in (5.5) we are considering w as the latent variable and v as the manifest
one. This is in keeping with the idea that v T M v, the supply rate, is the variable
of interest and that w is a latent variable that explains it. We are hence considering
the behavior of the possible trajectories v. Assume that M has r rows, i.e., that
 ON QUADRATIC DIFFERENTIAL FORMS 1721

d
M Rrq []. Thus, (5.5) defines a system B Lr with B = im(M ( dt )) and
d q r
M ( dt ) viewed as a map from C (R, R ) to C (R, R ). Hence this system has a state
representation. Assume that
 
d
(5.6) x=X w
dt

induces such a state representation. Thus X Rq [] is a polynomial matrix defining

d
a state map for B = im(M ( dt )). Let Rsqq [, ]. Then the QDF Q is said to be
a state function (relative to the state of ) if there exists a real (symmetric) matrix
P such that
 
d
(5.7) Q (w) = kX wk2P .
dt

It is said to be a state/supply function if there exists a real (symmetric) matrix E

such that
 d

M ( dt )
(5.8) Q (w) = k d wk2E
X( dt )

where, as always, kak2A denotes aT Aa. Note that the factorization (5.4) is not unique.
However, any such factorization is related in a simple way to a canonical one, say, to

(5.9) (, ) = M T () M ()

by the existence of a matrix F R such that M () = F M (). This relation has as

a consequence that any (possibly nonminimal) state map X for the system in image
representation (5.5) is related in a static way to a minimal state map X associated
with the system in image representation
 
d
(5.10) v = M w
dt

based on a canonical factorization. Indeed, there exists a matrix L R such that

(5.11) X() = LX().

Thus, considering arbitrary (i.e., not necessarily canonical) factorizations and arbi-
trary (i.e., not necessarily minimal) state representations yields a (rather than the)
state of Q . Thus, the situation with the state is similar to the situation with the
state of a system B Lq .
We have the following R important result.
Theorem 5.5. Let Q 0, and let Rsqq [, ] be a storage function for ,

i.e., . Then is a state function. Let Rsqq [, ] be a dissipation function
for . Then is a state/supply function. In fact, if X Rq [] is a state map for
, then there exist real symmetric matrices P and E such that

(5.12) (, ) = X T ()P X(),

 T  
M () M ()
(5.13) (, ) = E .
X() X()
1722 J. C. WILLEMS AND H. L. TRENTELMAN

Equivalently
   T  
d d
L (w1 , w2 ) = X w1 PX w2 ,
dt dt
 d
T  d

M ( dt )w1 M ( dt )w2
L (w1 , w2 ) = d E d
X( dt )w1 X( dt )w2

for all w1 , w2 C (R, Rq ).

Proof. For the proof, see the appendix.
Let Rqq [] be para-Hermitian: ? = . An F Rqq [] is said to induce
a symmetric factorization of if () = F T ()F (). It is said to be a symmetric
Hurwitz factorization if F is square and Hurwitz and a symmetric anti-Hurwitz factor-
ization if F ? is square and Hurwitz. It is easy to see that for a symmetric factorization
to exist we need to have (i) 0 R and for an (anti-)Hurwitz one to exist we
must have (i) > 0 R. The converses are also true but not at all trivial in the
matrix case. This result is well known (see, e.g., [9], [10], [18], [21]), and we state it
for easy reference.
Proposition 5.6. Let Rqq [] be para-Hermitian. Then
(i) allows a symmetric factorization iff (i) 0 for all R.
(ii) allows a symmetric Hurwitz factorization iff (i) > 0 for all R. Such
a factorization () = F T ()F () is unique up to premultiplication of F ()
by an orthogonal matrix.
(iii) allows a symmetric anti-Hurwitz factorization iff (i) > 0 for all R.
Such a factorization () = F T ()F () is unique up to premultiplication of
F () by an orthogonal matrix.
An important issue of concern is the uniqueness of the storage function, and
therefore ofRthe dissipation function, because of the one-to-one relation between the
two. When Q = 0, then the associated storage function is unique ((, ) = (,) + )
and the dissipation function
R is zero. However, in general there are many possibilities.
Theorem 5.7. Let Q 0. Then there exist storage functions and + for
such that any other storage function for satisfies

(5.14) + .
R per
If Q > 0 then and + may be constructed as follows. Let () = H T ()H()
and () = AT ()A() be, respectively, Hurwitz and anti-Hurwitz factorizations of
. Then
(, ) AT ()A()
(5.15) + (, ) =
+
and
(, ) H T ()H()
(5.16) (, ) = .
+

Proof. For the proof, see the appendix.

We close this section with a few remarks.
Remark 5.8. In this section we have studied average positivity with, in Q (w),
w C (R, Rq ) or D(R, Rq ), but otherwise free. It is of interest to generalize these
 ON QUADRATIC DIFFERENTIAL FORMS 1723

concepts to the case that w B with B a given element of Lq . Of course, in section

4 we have considered precisely such situations for Bs that are autonomous. Actually,
it turns out that the theory of section 5 is immediately applicable to systems B Lq
that are controllable. Indeed, let B Lq be controllable and assume that we want to
study when
Z + Z +
(5.17) Q (w)dt 0 or Q (w)dt = 0

holds for all w B D(R, Rq ). Simply construct an image representation for B, say,
 
d
(5.18) w=M `.
dt

Upon substituting (5.18) in (5.17), we see that the issue then becomes one of studying
when
Z +     Z +    
d d
Q M ` dt 0 or Q M ` dt = 0
dt dt
d
for all ` D(R, R ). Since obviously Q (M ( dt )`) = Q0 (`) with

0 (, ) := M T ()(, )M (),

the problem reduces to studying 0 . For example, the existence of a storage function is
established as follows. Without loss of generality, take (5.18) to be an observable image
representation. Then M has a polynomial left inverse M . By Proposition 5.4, there
d
exists 0 such that dt Q0 (`) Q0 (`) for all ` D(R, R ). Now define (, ) :=
T 0 d d
M () (, )M (). Then for w = M ( dt )` we have Q (w) = Q0 (M ( dt )w) =
d
Q0 (`) and Q (w) = Q0 (`), so we obtain dt Q (w) Q (w).
The case that B Lq is neither controllable nor autonomous will be studied
in a later publication. The next comment is relevant to the question of what the
appropriate definition of dissipativity is in that case.
Remark 5.9. Finding an appropriate definition of a dissipative system is an issue
that has attracted considerable attention (see [29], [12], [24], [27]). Of course, this is
at the root of the issues discussed in the present article. Let B Lq . There are many
examples where the instantaneous rate of supply (say, of energy) into the system is
given, not by a static function of the external variables, but by a QDF, Q (w). The
study of supply rates that are themselves dynamic is one of the novel aspects of the
present paper. When would one want to call B dissipative with respect to Q (w)?
Lossless? Conservative? When would one want to say that B absorbs some R of the
supply? The definitions of average nonnegativity for dissipativeness, and Q = 0
for losslessness (= conservativeness), are fully adequate provided that B is controllable
(see Remark 5.8). However, Proposition 5.4 points to another definition which does
not need controllability and which, in the controllable case, reduces to it. Thus, we
arrive at the following definition as the most general: B Lq is said to be dissipative
d
with respect to the supply rate Q if there exists a Q such that dt Q (w) Q (w) for
all w B, and lossless, or conservative, if this holds with equality. The unfortunate
aspect of this definition is its existential nature it shares this notorious feature with
the first and second law of thermodynamics. It does not seem an easy matter in the
noncontrollable case to reduce this to a statement involving only Q , and without
1724 J. C. WILLEMS AND H. L. TRENTELMAN

invoking a to-be-constructed Q . In Theorem 5.5 we have unraveled this existence

question a bit by proving that this Q will be a state function.
Note that the proposed definition of dissipativity and losslessness is an interesting
generalization of the notion of a Lyapunov function since, for autonomous systems, it is
natural to take the external supply = 0. Also note that this definition holds for any
B and and does not require the introduction of the notion of state. In other words,
dynamical systems with free variables that allow interaction with the environment
relate to flows on manifolds, just as dissipative systems relate to Lyapunov functions.
Remark 5.10. Let B be controllable and assume that it is dissipative with respect
to the supply rate Q (w) (see Remark 5.9). Also in this general case every storage
function is a state function, and every dissipation function is a state/supply function.
However, this time, not simply the state of is involved, but the state of a system
obtained by combining the dynamics of and B. This is elaborated in [26].
Remark 5.11. Let (, ) = M T ()M M (). Consider the system in image
representation (5.5). Then it can be shown that for to be average nonnegative, there
must be an input/output partition for this system so that all the input components
correspond to +1s in . In other words, the supply rate v T M v is always of the
form kuk2 + ky1 k2 ky2 k2 , with u an input, and y1 , y2 outputs.
Remark 5.12. It follows from Theorem 5.5 that a factorization of the polynomial
matrix (, ) = M T ()M M () into F T ()F () always leads to a situation
in which the McMillan degree of M is equal to that of col(M, F ). This means that
the factorization is a regular factorization (as this property is called). In the H -
problem factorization, questions are encountered in which the existence of a regular
factorization poses a serious problem.
Remark 5.13. It is easy to see that the set of storage functions corresponding to a
given supply rate is convex. Moreover, in the case of average positivity, 6= + and
hence, in this case, there are an infinite number of possible storage functions. Actually,
in this respect it is worth mentioning the following refinement of Theorem 5.7, which
follows immediately from our proof of this theorem. If (, ) satisfies (i, i)
0 (but not (i, i) > 0) for all R, then a symmetric Hurwitz factorization does
not exist. In this case, there are two possibilities: either det() 6= 0 or det() = 0.
In the former case, (, ) allows a factorization (, ) = H T ()H() with H
almost Hurwitz (i.e., H has all its singularities in <e() 0). In the latter case,
there exists a unimodular matrix U such that

(, ) = U T ()0 (, )U (),

with 0 of the form

 
1 2
0 = ,
2 3

with det(1 ) 6= 0 and 2 = 0, 3 = 0. Factor 1 (, ) as before as H1T ()H1 ().

Then
 
H1 0
H= U1
0 0

yields an almost Hurwitz-like factorization of . Similarly, we can define an almost

anti-Hurwitz-like factorization = A? A of any satisfying (i, i) 0 for all
R. The computation of + and given in Theorem 5.7 holds unaltered with
 ON QUADRATIC DIFFERENTIAL FORMS 1725

this A and H. Note that in the lossless case ( = 0) this yields + = , whence
the uniqueness of .
Remark 5.14. It is easy to deduce from the proof of Theorem 5.7 that and +
d
have the following interpretations. Let x = X( dt )w be the state (see 5.6). Let a Rn .
q d
Consider all w D(R, R ) such that (X( dt )w)(0) = a. Denote this set by Ba . By
d
Theorem 5.5 we know that Q is a state function, say, Q (w) = kX( dt )wkK ,
nn
for some symmetric matrix K R . Hence for w Ba we have Q (w)(0) =
aT K a. Similarly Q+ (w)(0) = aT K+ a for some symmetric matrix K+ . Then it can
be shown that
 Z + 
(5.19) aT K a = sup Q (w)dt
wBa 0

and
Z 0 
T
(5.20) a K+ a = inf Q (w)dt .
wBa

For this reason, Q (w)(0) is called the available storage and Q+ (w)(0) the required
supply at t = 0 due to w. In this inf and sup, one keeps the past, respectively, the
future of w fixed.
R +
6. Half-line positivity. In section 5, we studied QDFs for which Q (w)dt
0. The intuitive idea was that this expresses that the net supply (of energy)
is directed into the system: energy is being absorbed and dissipated in the system.
There are, however,
Rt situations where at any moment in time the system has absorbed
energy, i.e., Q (w)( )d 0 for all t R. For example, electrical circuits and
mechanical devices at rest are in a state of minimum energy, and therefore the energy
delivered up to any time is nonnegative. This type of positivity is studied in this
section. It plays a crucial role in H problems.
Definition 6.1. Let Rsqq [, ]. The QDF Q (or simply ) is said to
Rt R0
be half-line nonnegative, denoted by Q 0, if Q (w)dt 0 for all w
Rt R0
D(R, Rq ), and half-line positive, denoted Q > 0, if in addition Q (w)dt = 0
implies w(t) = 0 for t 0.
Note that half-line nonnegativity implies average nonnegativity, and that half-line
positivity implies average positivity.
Write (, ) = M T ()M M () and partition M conform M as
 
P
(6.1) M=
N
d d
so that (, ) = P T ()P ()N T ()N () and hence Q (w) = kP ( dt )wk2 kN ( dt )wk2 .
In the following, for C, let denote its complex conjugate.
Proposition 6.2. Let Rsqq [, ]. Then
Rt
(i) ( Q 0) ((, ) 0 C, <e() 0)
Rt
(ii) ( Q > 0) ((, ) 0 C, <e() 0 and det() 6= 0).
Proof. For the proof, see the appendix.
As noted before, it immediately follows from the definitions that half-line nonneg-
ativity implies average nonnegativity, etc. Thus, Proposition 5.4 implies the existence
of a storage function. It is the nonnegativity of the storage function that allows us to
conclude the half-line positivity.
Theorem 6.3. Let Rsqq [, ]. Then the following statements are equivalent.
1726 J. C. WILLEMS AND H. L. TRENTELMAN

Rt
1. Q 0,
2. there exists a storage function 0 for ,
3. admits a storage function, and the storage function + defined in Theorem
5.7 satisfies + 0.
Proof. For the proof, see the appendix.
In order to check half-line nonnegativity, one could thus in principle proceed as
follows. Verify that (i, i) 0 for all R, compute + , and check whether
+ 0. In some situations, it is actually possible to verify this condition in a more
immediate fashion; for example, when (, ) = 0 , a constant matrix, with 0 0
(trivial, but that is the case that occurs in standard LQ theory!), or when in (6.1)
P is square and det(P ) 6= 0. Then, under the assumption that a storage function
exists (equivalently: N T (i)N (i) P T (i)P (i) for all R), all storage
functions are actually nonnegative if one of them is nonnegative. In fact, in this case
the following theorem holds.
Theorem 6.4. Let Rsqq [, ]. Assume it is factored as (, ) = P T ()P ()
N ()N () with P square and det(P ) 6= 0. Let X Rq [] be a minimal state
T

map for the B given in image representation by (6.1). The following statements are
equivalent:Rt
1. Q 0,
2. (, ) 0 for all C, <e() 0,
3. N P 1 has no poles in <e() 0 and (i, i) 0 for all R,
4. there exists a storage function 0 for ,
5. there exists a storage function for and every storage function for
satisfies 0,
d
6. there exists a real symmetric matrix K > 0 such that QK (w) := kX( dt )wk2K
is a storage function for ,
7. there exists a storage function for and every real symmetric matrix K such
d
that QK (w) := kX( dt )wk2K is a storage function for satisfies K > 0.
P
Furthermore, if [ N ] is observable, then any of the above statements is equivalent with
30 . P is Hurwitz and (i, i) 0 for all R.
Proof. For the proof, see the appendix.
7. Observability. One of the noticeable features of QDFs is that a number of
interesting systems theory concepts generalize very nicely to QDFs. We have already
seen that the state of a symmetric canonical factorization of functions as the state
of the QDF Q . In this section we introduce observability of a QDF. In a later section
we will discuss duality of QDFs.
For Rq1 q2 [, ] and w1 C (R, Rq1 ) fixed, the linear map w2 7 L (w1 , w2 )
is denoted by L (w1 , ). For w2 C (R, Rq2 ) fixed, the linear map w1 7 L (w1 , w2 )
is denoted by L (, w2 ). The BLDF is called observable if L (w1 , ) and L (, w2 )
determine w1 and w2 uniquely. Equivalently we have the following.
Definition 7.1. Let Rq1 q2 [, ]. We call observable if, for all w1
C (R, Rq1 ) and for all w2 C (R, Rq2 ), we have

L (w1 , ) = 0 w1 = 0
and
L (, w2 ) = 0 w2 = 0.
The following theorem gives necessary and sufficient conditions for observability
purely in terms of the two-variable polynomial matrix and in terms of the (one-
 ON QUADRATIC DIFFERENTIAL FORMS 1727

variable) polynomial matrices N and M occurring in any canonical factorization of

.
Theorem 7.2. Let (, ) = N T ()M () be a canonical factorization. The
following statements are equivalent:
1. is observable,
2. for every C, the rows of (, ) Rq1 q2 [], and the columns of (, )
Rq1 q2 [] are linearly independent over C,
3. N () and M () have full column rank for all C; equivalently, the image
d d
representations v1 = N ( dt )w1 and v2 = M ( dt )w2 are observable,
Proof. For the proof, see the appendix.
If is symmetric, then we have T (, ) = (, ), so condition (ii) above can
be replaced by a single statement on the independence of the rows of (, ). Also,
in this case the maps L (w, ) and L (, w) coincide. Furthermore, for the particular
symmetric canonical factorization (, ) = McT () Mc () obtained from (3.15),
we have ||(, ) = McT ()Mc (). Hence observability of is also equivalent with
|| > 0 and with the condition ||(, ) > 0 for all C. This immediately yields
the following.
Corollary 7.3. Let Rsqq [, ] and let (, ) = M T () M () be a
symmetric canonical factorization. Then the following statements are equivalent:
1. is observable,
2. L (w, ) = 0 w = 0,
3. for every C, the rows of (, ) Rq1 q2 [] are linearly independent over
C,
4. M () has full column rank for all C, equivalently, the image representa-
d
tion v = M ( dt )w is observable,
5. || > 0,
6. ||(, ) > 0 for all C.
8. Strict positivity. Throughout this section we assume that Rsqq [, ] is
observable. We now introduce and develop the notion of strict positivity. The concept
of strict half-line positivity given here is very analogous to that used by Meinsma [19].
Definition 8.1. Let Rsqq [, ] be observable. We call the QDF Q strictly
positive, denoted  0, if there existsR  > 0 such that || 0. We call it
strictly average positive, denoted by Q  0, if there exists  > 0 such that
Z + Z +
(8.1) Q (w)dt  Q|| (w)dt

Rt
for all w D(R, Rq ). We call it strictly half-line positive, denoted Q  0, if
there exists an  > 0 such that
Z 0 Z 0
(8.2) Q (w)dt  Q|| (w)dt

q
for all w D(R, R ). Note that (because of observability) strict positivity implies
positivity, and similarly for the other cases.
These notions of strict positivity involve || which may be difficult to evaluate.
However, it is possible to relate it to any canonical factorization of . This is stated in
the next proposition. For simplicity we state only the case of strict average positivity.
However, completely analogous statements hold for simple strict positivity or for strict
half-line positivity.
1728 J. C. WILLEMS AND H. L. TRENTELMAN

Proposition 8.2. Let Rsqq [, ] be observable and let (, ) = P T ()P ()

T P
N ()N () be a symmetric canonical factorization of (see (3.12)). Denote M = [ N ]
The following are equivalent:
1. is strictly average positive.
2. There exists an  > 0 such that
Z +   Z +  
d 2 d
(8.3) kM wk dt  kM wk2 dt
dt dt

for all w D(R, Rq ). Here kak2 denotes aT a.

3. There exists an < 1 such that
Z +   Z +  
d 2 d
(8.4) kN wk dt kP wkdt
dt dt

for all w D(R, Rq ).

Moreover, also for a noncanonical factorization (3.11), (2) and (3) (with replaced
by M ) are equivalent and imply (1).
Proof. For the proof, see the appendix.
9. A Pick matrix condition for half-line positivity. It is surprisingly dif-
ficult to establish some type of analogue of Proposition 5.2 for half-line positivity,
and earlier attempts [28], [30], [1]) turned out to be flawed. In Proposition 6.4
such an analogue of Proposition 5.2 was given but only in the special case where
(, ) = P T ()P () N T ()N () with det(P ) 6= 0. In this section we give a neces-
sary and sufficient condition for strict half-line positivity in terms of .
As is well known, the Pick matrix plays an important role in system and circuit
theory, in particular in connection with passivity properties of linear dynamical sys-
tems; see [34], [4], [5]. We derive a Pick-matrix-type test for nonnegativity of + .
This test is perhaps the most original specific result of this paper. For simplicity we
consider only the case of strict half-line positivity. First, however, we need to define
the Pick-type matrix which may be computed effectively from a Rsqq [, ]. Let
F Rqq [], and assume that det(F ) 6= 0. We call F semisimple if for all C the
dimension of the kernel of F () is equal to the multiplicity of as a root of det(F ).
Note that F is certainly semisimple if det(F ) has distinct roots. We now define the
matrix T . Since the expression is much simpler in the semisimple case, we explain
that case first.
Definition 9.1 (semisimple case). Let Rsqq [, ] be observable, and assume
that det() has no roots on the imaginary axis. Let 1 , 2 , . . . , n C be the roots of
det() with positive real part and let a1 , a2 , . . . , an Cq be such that (i )ai = 0,
and such that the ak s associated with the same i form a basis of ker((i )). Then
the Pick matrix of is defined as
 T 
ai (i , j )aj
(9.1) T := .
i + j i,j=1,...,n

In order to define the matrix T in the general case, we need to take into account
the algebraic multiplicities of the roots i .
Definition 9.2 (general case). Let Rsqq [, ] be observable, det() 6= 0,
and assume that det() has no roots on the imaginary axis. Let 1 , 2 , . . . , k C be
the distinct roots of det() with positive real part, and denote by ni the multiplicity
 ON QUADRATIC DIFFERENTIAL FORMS 1729

of i as a root of det(). For i = 1, 2, . . . , k, there are ni linearly independent vectors

ai,0 , ai,1 , . . . , ai,ni 1 determined by the (ni ) linear equations

i 1 
nX 
j
()(j`) (i )ai,j = 0, (` = 0, 1, . . . , ni 1).
`
j=`

Here, ()(k) () denotes the kth derivative of the polynomial matrix .

For i = 1, 2, . . . , k, define

ai,0 0 0
..
ai,0 ai,1 . . . .

Ai := . Cni qni
. .
.. .. .. 0
ai,0 ai,1 . . . ai,ni 1

Also, define i,j Cni qnj q by defining its (r, s)th block to be the q q matrix

(i,j )r,s := (r,s) (i , j ), r = 1, 2, . . . , ni ; s = 1, 2, . . . , nj ,

where (k,`) means taking the kth partial derivative with respect to and the `th with
respect to .
Then we define the Pick matrix of as the matrix T whose (i, j)th block is given
by Ti,j Cni nj , with

1
Ti,j := AT i,j Aj .
i + j i
Pk
Note that the sum i=1 ni of the multiplicities is equal to n := 12 deg det(),
and that T is a complex Hermitian matrix of size n n.
The next theorem is the most refined result of this paper. It shows, on the
one hand, the relation between strict half-line positivity and positivity of a storage
function, and, on the other hand, the relation with the positivity of the Pick matrix
T .
We have seen in Theorem 5.5 that a storage function is a quadratic state function,
d
i.e., Q (w) is of the form xT Kx, K = K T , with x = X( dt )w a minimal state map
for . We call this state function positive definite if K > 0.
Theorem 9.3. Let Rsqq [, ] be observable. The following are equivalent:
Rt
1. QR  0,
2. (a) Q  0,
(b) there exists a storage function that is a positive definite state function,
3. (a)  > 0 such that (i, i) ||(i, i) for all R,
(b) T > 0.
Proof. For the proof, see the appendix.
Remark 9.4. It follows from the proof of Theorem 9.3 that half-line nonnegativity
implies that the Pick-type matrix T (see (9.1)) is 0 whenever any set of i s in
the right half of the complex plane and any set of ai s are chosen. It is possible to
prove that if T is nonnegative definite for any such choice for the i s and ai s, then
we have half-line nonnegativity. The remarkable thing about Theorem 9.3 is that it
suffices to evaluate T at the set of special i s and ai s obtained from the singularities
of .
1730 J. C. WILLEMS AND H. L. TRENTELMAN

Remark 9.5. It is well known that solvability of a certain NevanlinnaPick in-

terpolation problem is equivalent to positive definiteness of a given Pick matrix. In
fact, in [34] the necessity of the positive definite Pick matrix is shown using a half-line
positivity argument. Theorem 9.3 states that positive definiteness of a given Pick
matrix is also sufficient for half-line positivity. R
Remark 9.6. It can be shown that if is observable, then Q  0 implies that
+ is a positive definite state function, with + and as defined in Theorem
5.7.
Remark 9.7. It is possible to generalize the T -test of Theorem 9.3 to half-line
positive (instead of strictly half-line positive QDFs) by including infinite zeros of
det(). However, the notation gets very involved, and therefore we will not do this
here.
10. Duality. In the present section we discuss some remarkable relations be-
tween positivity of QDFs and their duals. These relations are of interest in their own
right and will be of crucial importance in our treatment of the H -problem [25].
Let B1 and B2 Lq . We call B1 and B2 complementary if B1 B2 = C (R, Rq ).
It is easy to see that this implies that both B1 and B2 must be controllable: uncon-
trollable Bs in Lq have no complement in Lq . We call them dual if they are com-
plementary and if hw1 , w2 i = 0 for all w1 B1 D(R, Rq ) and w2 B2 D(R, Rq ),
R +
where hw1 , w2 i denotes the usual inner product w1T w2 dt. If this is the case,
then we denote B2 as B
1 , since B1 defines B1 uniquely. Obviously we also have

(B1 ) = B1 .
d
It is easy to see that R( dt )w = 0 is a minimal kernel representation of the con-
d
trollable B iff v = R ( dt )` is an observable image representation of B (a kernel
T
d
representation R( dt )w = 0 of B is called minimal if R has full row rank). Conse-
d d
quently, w = M ( dt )` is an observable image representation of B iff M T ( dt )v = 0

is a minimal kernel representation of B . This duality can also be extended to state
representations, in the following sense. If

dx
(10.1) E + F x + Gw = 0,
dt

is an n-dimensional minimal state representation of B, then B admits an also n-

dimensional minimal state representation (thus the dimensions of the minimal state
representations are the same), say,

dz
(10.2) E0 + F 0 z + G0 v = 0,
dt
having the property that for all (w, x) C (R, Rq+n ) satisfying (10.1), and for all
(v, z) C (R, Rq+n ) satisfying (10.2) there holds the following kind of duality in-
volving the state

d T
(10.3) z x = v T w.
dt
In fact, this is an immediate consequence of the following.
d d
Proposition 10.1. Let R( dt )w = 0 and w = M ( dt )` be a minimal kernel
representation and an observable image representation, respectively, of the controllable
system B Lq . Assume that X Rn [] defines a minimal state map for B, i.e.,
 ON QUADRATIC DIFFERENTIAL FORMS 1731

d
x = X( dt )` defines a minimal state of B. Then there exists a Z Rn [] defining
d
a minimal state map Z( dt ) for B , such that for all `, `0 C (R, R ), we have
   T      T  
d d d d d
(10.4) Z `0 X ` = RT `0 M `.
dt dt dt dt dt
If we define (, ) := Z T ()X() and (, ) := R()M (), then (10.4) is equiva-

lent to = .
Proof. For the proof, see the appendix.
If a pair of minimal state maps (X, Z) of B and B satisfies (10.4), then we call
it a matched pair of state maps.
We now associate with a QDF a dual one and relate their average nonnegativity
and average positivity. Let Rsqq [, ] and let (, ) = M T () M () be a
symmetric canonical factorization, with
 
Ir+ 0
= .
0 Ir
Let us assume that M Rrq []. Partition M conformably to as
 
P
M=
N
so that is written as
(10.5) (, ) = P T ()P () N T ()N ().
Consider the dynamical system B Lr with image representation
 
d
(10.6) w=M `.
dt
There are a number of integers associated with M that are of interest to us:
(10.7) r+ = rowdim(P ),
(10.8) r = rowdim(N ),
(10.9) m = rank(M ) .
The number r+ corresponds to the number of positive squares in Q , r to the num-
ber of negative squares, while m equals the number of inputs in any input/output
or input/state/output representation of B. Since it is defined by an image repre-
sentation, B is a controllable system and, as such, it admits a dual, B Lr . Let
d
R( dt )w = 0 be a minimal kernel representation of B. Then
 
d
(10.10) v = RT `0
dt
is an observable image representation for B . Let 0 (, ) := R() RT (). Note
d d
that the QDFs Q (`) = (M ( dt )`)T M ( dt )` and
   T  
d d
Q0 (`0 ) = RT `0 R T `0
dt dt
are in a sense also dual. Their positivity properties are very much related, as shown
in the following theorem.
Theorem 10.2. Assume that r+ = m. Then
1732 J. C. WILLEMS AND H. L. TRENTELMAN

R R
(i) R Q 0 R Q0 0,
(ii) Q > 0 Q0 < 0, R R
(iii) (assume is observable) Q  0 Q0  0,

R Z) be a matched pair of minimal state maps for B and B . Assume
(iv) Let (X,
that Q 0, and let define a storage function for . By Theorem 5.5,
is a state function, i.e., there exists a real symmetric matrix K such that
   T  
d d
(10.11) Q (`) = X ` KX `.
dt dt

Assume that K is nonsingular. Then

   T  
0 d 0 1 d
(10.12) Q0 (` ) = Z ` K Z `0
dt dt

R 0 .
is a storage function for Q R R
(v) (assume is observable) t Q  0 t Q0  0. Here t Q0  0 is
R
defined as the property that there exists  > 0 such that 0 Q0 (w)dt
R
 0 Q0 | (w)dt for all w D(R, Rq ) (i.e., half-line positivity over the posi-
tive half-line).
Proof. For the proof, see the appendix.
We close this section by pointing out that it is of interest to generalize the notion
of duality by using, instead of the usual inner product, an inner product that is itself
induced by a QDF. These ramifications are a matter of future research.
11. Conclusions. In this paper we studied two-variable polynomial matrices
and their role in a number of problems in linear system theory. The basic premise set
forward is the following. Dynamic models lead naturally to the study of one-variable
polynomial matrices. By substituting the time derivative for the indeterminate, and
by letting the resulting differential operator act on a variable, one arrives at a dynam-
ical system, which may then be in kernel or in image representation. The study of
quadratic functionals in a variable and its derivative, on the other hand, leads to two-
variable polynomial matrices. Important instances where dynamical systems occur in
conjunction with functionals are, for example, Lyapunov theory, the theory of dissi-
pative systems, and LQ and H control. We developed the former two applications
in the present paper. The latter two will be discussed elsewhere.
Appendix.
Proof of Theorem 3.1. We prove the equivalence of the two statements in (1) at
the end of the proof and proceedR with the first statement by
R running the circle (1)
(3) (2) (1). Assume that L = 0. Then obviously L (v, w)dt = 0 for all
v D(R, Cq1 ) and w D(R, Cq2 ), with L (v, w) in this case (for complex functions)
P k `
defined by k,` ( ddtkv )T k,` ( ddtw` ). Then
Z
v T (i)(i, i)w(i)d = 0

for all v L2 (C, Cq1 ), w L2 (C, Cq2 ) that are Fourier transforms of v D(R, Cq1 ),
and w D(R, Cq2 ). This implies that = 0. Assume to the contrary that there
exist 0 R, a Cq1 , b Cq2 such that aT (i0 , i0 )b 6= 0. Define vN D(R, Cq1 )
 ON QUADRATIC DIFFERENTIAL FORMS 1733

for N = 1, 2, . . . , by
2N
ei0 t a |t| 0 ,




2N
(A.1) vN (t) = v(t + 0 ) t < 2N
0 ,




2N 2N
v(t 0 ) t> 0 .

Define wN D(R, Cq2 ) analogously by replacing a by b. Note that v and w can be

chosen independent of N , and obtain smoothness for all N : indeed, if v1 is smooth,
then by the periodic nature of vN for |t| 2N
0 , vN will also be smooth.
R
Next evaluate L (vN , wN )dt and observe that this integral equals
4N T
a (i0 , i0 )b + E
0
R
with E independent of N . It follows that L (vN , wN )dt 6= 0 for N sufficiently
large. In order to obtain this for real-valued functions, consider the real and imaginary
parts of vN , wN and the integrals. This establishes the contradiction. Hence (1)
implies (3).
To prove (3) (2), view (, ) as a one-variable polynomial in and carry out
the division by + . This yields (, ) = ( + )d(, ) + r(, ). Hence = 0
implies r = 0. This yields (2). R R d
To show that (2) (1), observe that L (v, w)dt = dt L (v, w)dt. The
last term obviously vanishes since v and w have compact support.
To show theR equivalence of the two statements in (1), observe that it follows
trivially that L = 0 implies path independence. Conversely, if = 0 then,
d
according to (3), there exists Rq1 q2 [, ] such that L = dt L . Thus, for any
q1 q2
pair of functions v C (R, R ) and w C (R, R ), and for any t1 and t2 , we have
Z t2 Z t2
d
L (v, w)dt = L (v, w)dt = L (v, w)(t2 ) L (v, w)(t1 ).
t1 t1 dt
Hence the integral depends only on the values taken on by v and w and their deriva-
tives at the endpoints t1 and t2 .
Proof of Proposition 3.2. This proposition is proven following the standard proofs
used in behavioral theory: reduce the problem to the scalar case using the Smith
d
form. Let R = U V with U, V unimodular and diagonal. Define B0 = V ( dt )B.
0 d
Then B has ( dt )w = 0 as kernel representation. To prove the proposition, note
that the if parts are immediate.
d
To show the first only if part, we show that D( dt )B = 0 implies that there
d
exists F such that D = F R or equivalently, with D = D0 V , that D0 ( dt )B0 = 0
implies that there exists F such that D = F . Let = diag(d, ), let d0 be the
0 0 0 0
d
first column of D0 , and let w1 be a solution of d( dt )w1 = 0. Since col[w1 , 0, . . . , 0] B0
0 d 0 0 d d
and D ( dt )B = 0, it follows that d ( dt )w1 = 0. It is easily seen that d( dt )w1 = 0
d
implies d0 ( dt )w1 = 0 iff each element of the polynomial vector d0 is a factor of d.
Proceeding this way column by column yields D0 = F 0 .
To show the second only if part, we prove first the analogous result for BLDFs.
This states that with Rq1 q2 [, ], the BLDF L (w1 , w2 ) = 0 for all w1 B1 and
w2 B2 iff there exists F1 , F2 such that
(A.2) (, ) = R1T ()F2 (, ) + F1 (, )R2 (),
1734 J. C. WILLEMS AND H. L. TRENTELMAN

where R1 and R2 induce kernel representations of B1 and B2 . The if part is once

again obvious. To prove the only if part, consider first the following lemma, which
proves the scalar case q1 = q2 = 1.
Lemma A.1. Let r1 , r2 R[] and R[, ]. Let Bm L1 , m = 1, 2 be given
d
in kernel representation by rm ( dt )wm = 0. Then Q (w1 , w2 ) = 0 for all wm Bm
iff there exists fm R[, ] such that

(A.3) (, ) = r1 ()f2 (, ) + f1 (, )r2 ().

Proof. The if part is obvious. To show the only if part, let r1 have degree n1
and r2 have degree n2 , and assume that they are monic. Consider the term k,` k ` of
k `
(, ). In the quadratic form Q (w1 , w2 ) this term contributes k,` d w 1 d w2
. If w1
dttk dt`
d
satisfies r1 ( dt )w1 = 0, and if k n1 , then the contribution in Q (w1 , w2 ) of k,` k `
is equivalent to that of k,` ( k kn1 r1 ()) ` . Proceeding
P analogously with the `s
and the other terms shows that there exists 0 (, ) = k,` 0k,` k ` with 0k,` = 0
for k n1 or ` n2 such that

(A.4) (, ) = r1 ()f2 (, ) + f1 (, )r2 () + 0 (, ).

Obviously Q0 (w1 , w2 ) = Q (w1 , w2 ) for wm Bm . Therefore Q0 (w1 , w2 ) = 0 for

wm Bm . Consider Q0 (w1 , w2 )(0) and observe that this is a quadratic form in
dn1 1 w1 dn2 1 w2
w1 (0), dw dw2
dt (0), . . . , dtn1 1 (0) and w2 (0), dt (0), . . . , dtn2 1 (0). These initial con-
1

ditions can be chosen arbitrarily in the sense that for any values of wm (0), dw dt (0),
m

dnm 1 wm
. . . , dtnm 1 (0) there exist wm Bm having these initial values. It follows that
0 = 0.
Now return to the proof of the case for general q1 , q2 . Bring R1 and R2 in Smith
form, showing that it suffices to prove (A.2) for R = 1 and R2 = 2 with 1 and 2
in Smith form. Let d1 be the (k1 , k1 )th element of 1 and d2 the (k2 , k2 )th element
of 2 . Examine (A.2) and observe that we need to show that the (k1 , k2 )th element
of , k1 k2 , can be written as

(A.5) k1 ,k2 (, ) = d1 ()f2 (, ) + f1 (, )d2 ()

d d
whenever it holds that d1 ( dt )v1 = 0 and d2 ( dt )v2 = 0 implies that Lk1 k2 (v1 , v2 ) = 0.
Now use the previous lemma.
In order to prove Proposition 3.2 for Rns 1 n2 [, ], use the -operator on
(A.2), and add.
The image representation part of Proposition 3.2 is proven analogously.
Proof of Proposition 3.5. The proof follows exactly along the same lines as the
proof of Proposition 3.2, and we can therefore be very brief. The Smith form once
again implies that it suffices to prove the case q1 = q2 = 1. Denote a kernel repre-
d
sentation of B by r( dt )w = 0. Using (A.4) with r1 = r2 = r shows that Q 0 on
B iff Q0 0 on B. However, again by the arbitrariness of the initial conditions,
(Q 0 on B) iff the matrix 0 associated with 0 is nonnegative definite. Part (i)
of the proposition follows.
To show part (ii), factor 0 (using 0 ) as 0 (, ) = DT ()D() with D R1 []
having elements whose degree is less than that of r. It thus suffices to find conditions
d d
for r( dt )w = 0 and D( dt )w = 0 to imply w = 0. That, however, is exactly equivalent
to the observability of the pair (r, D).
 ON QUADRATIC DIFFERENTIAL FORMS 1735

Proof of Theorem 4.3. The if part is shown as follows. By Proposition 3.5 we

B B
know that < 0 implies that (, ) = DT ()D() with D Rq [] such that
(R, D) is observable, with R Rq [] a kernel representation of B. It also holds that
d
(A.6) Q (w) = Q (w).
dt

Integrate this from 0 to T along a w B and obtain

Z T Z T  
d
(A.7) Q (w)(T ) Q (w)(0) = Q (w)dt = kD (w)k2 dt.
0 0 dt
B
Using 0, this yields
Z T  
d
(A.8) kD (w)k2 dt Q (w)(0).
0 dt
Therefore
Z  
d
(A.9) kD (w)k2 dt < .
0 dt
This implies the asymptotic stability of B. Assume that aet B, a 6= 0. Then
R()a = 0 and by (A.8) there must hold that either D()a = 0 or <e() < 0. (Note
that we silently use the obvious fact that (A.8) also holds for the complexification of
B.) However, by observability of (R, D), R()a = 0 and D()a = 0 imply a = 0.
d
Hence all exponential solutions aet of R( dt )w = 0 must have <e() < 0. It is well
known from the theory of differential equations that this implies that all solutions
approach zero as t . The only if follows from the stronger Theorem 4.8 and
will be proven then.

Proof of Corollary 4.6. (, ) = ( + )0 . In the case at hand, R() = A I.
B
Using Proposition
3.2, ( + ) = A0 + 0 AT . Finally, observe that observability
of (A I, 0 ) (as a pair of polynomial matrices) is equivalent to that of (A, 0 )
(as a pair of matrices) which is equivalent to that of (A, 0 ).
Proof of Proposition
R 4.7. dExamine formula (A.8) in the proof of Theorem 4.3. It
d
implies Q (w)(0) 0 kD( dt )wk2 dt. Therefore Q (w)(0) = 0 implies D( dt )w = 0.
d
However, by observability of D, D( dt )w = 0 in turn implies w = 0.
Proof of Theorem 4.8. The proof is organized as follows. First, we prove that
(4.3) is solvable; second, that if R is square (4.4), (4.5) gives all Sits solutions; third,
B B B B
that (4.7) yields = ; fourth, that 1 = 2 implies 1 = 2 ; fifth, that 0
B B B
yields 0; and sixth, that < 0 yields  0.

(i) First put R in Smith form: let

 
D
R=U V,
0
with D diagonal and U, V unimodular. Observe that it suffices to prove (4.3) with
R = D. The (k, `)th component of the matrix equation (4.3) in the obvious notation
takes the form
(A.10) x`k ()d` () + dk ()xk` () = k` (, ).
1736 J. C. WILLEMS AND H. L. TRENTELMAN

Since dk and d` are Hurwitz, d` () and dk () are coprime and hence, by Bezout,
(A.10) has a solution. This then yields a solution of the matrix version.
(ii) Again use the Smith form. Obtain that the difference of two solutions must
satisfy

(A.11) x`k ()d` () + dk ()xk` () = 0.

Hence again using coprimeness of d` () and dk (), there exists a polynomial fk` such
that xk` () = fk` ()d` (). This yields (4.4). To show (4.5), obtain

(A.12) RT ()(F () + F T ())R() = 0.

If R is square and det(R) 6= 0, (4.5) follows by pre- and postmultiplying by (RT ())1
and (R())1 .
(iii) This proof is obvious.
B B
(iv) Let w B and assume that 1 = 2 , i.e., = 0, where = 1 2 . Then
Z t Z t
d
(A.13) Q (w)dt = Q (w)dt = Q (w)(t) Q (w)(0).
0 0 dt

Since Q (w) = 0, asymptotic stability of B implies Q (w)(t) 0 as t and

B
hence that Q (w)(0) = 0. Therefore = 0.
(v) This follows immediately from (A.7), and Proposition 4.7 yields (vi).
Proof of Proposition 4.9. Existence of both D0 and 0 follows from the algorithm
given in the statement of the proposition. To show uniqueness of D0 observe that
B
D0 = D00 , i.e., D00 D0 = F R, and D0 R1 , D00 R1 strictly proper, implies F = 0, i.e.,
B
D0 = D00 . In the two-variable case assume 0 = 00 , i.e.,

0 (, ) = 00 (, ) + F T (, )R() + RT ()F (, ).

Thus,

(RT ())1 (0 00 )(, )(R())1 = (RT ())1 F T (, ) + F (, )(R())1

Strict properness again implies F = 0.

B
Proof of Proposition 4.10. Let = 0. Then (, ) = F T (, )R()+RT ()F (, ).
Pre- and postmultiply by (RT ())1 and (R())1 , respectively, and conclude that
B
F = 0. The result follows. If 0, use the same reasoning and Proposition 3.5 on
B B
(, ) = DT ()D() with D R-canonical. The case > 0 is similar.
Proof of Theorem 4.12. We first show that (4.4) has an R-canonical solution. Let
X be any solution. Factor XR1 as XR1 = P + S with P polynomial and S strictly
proper. First observe that it follows from (4.3) that P () + P T () = 0. Next, show
that X P R is a canonical solution. Uniqueness of this R-canonical solution follows
from (4.4).
Next, we show that (4.10) yields an R-canonical . Simply pre- and postmultiply
B
by (RT ())1 and R1 () and observe properness. Uniqueness follows from (1 =
B B
and 2 = ) = (1 = 2 ). Now apply Proposition 4.9.
The remaining statements follow from Proposition 4.10.
 ON QUADRATIC DIFFERENTIAL FORMS 1737

Proof of Proposition 5.2. The proof of all three statements is analogous. Therefore
we only give the proof of (i). To prove (), let w D(R, Rq ) and let w be its Fourier
transform. Observe, using Parsevals Theorem, that
Z + Z +
1
(A.14) Q (w)dt = w(i)T (i, i)w(i)d,
2
whence (). To show the converse, as in the proof of Theorem 3.1, we silently switch
from Rq as signal space to Cq . Assume that there exists a Cq and 0 R such that
aT (i0 , i0 )a < 0. Consider the function wN D(R, Cq ) for N = 1, 2, . . ., defined
R +
exactly as vN was in the proof of Theorem 3.1. Next evaluate Q (wN )dt and
observe (using the idea in the proof of Theorem 3.1) that this integral can be made
negative by taking N sufficiently large.
Proof of Proposition 5.4. We will run the circle (3) (2) (1) (3). To see
that (3) (2), assume that is a dissipation function. Then (, ) = (, ),
by Theorem 3.1. Define
(, ) (, )
(A.15) (, ) = .
+

Hence = . Use 0 to conclude that is a storage function. To see

that (2) (1), use and Theorem 3.1 to conclude (1). To see that (1) (3),
use Propositions 5.2 and 5.6 to construct a D such that (, ) = DT ()D().
Observe that (, ) := DT ()D() defines a dissipation function. The one-one
relation between and is given by (A.15).
Proof of Theorem 5.5. By (5.11) it suffices to consider minimal state representa-
tions obtained from a canonical factorization of . Let
 
d
v=M = w
dt
d
be obtained from such a factorization, and let x = X( dt )w be a minimal state. There
exists a permutation matrix P such that
 
U
PM = ,
Y

with det(U ) 6= 0 and such that Y U 1 is a matrix of proper rational functions. Denote
d d
u = U ( dt )w. Consider f = F ( dt )w, where F is an arbitrary polynomial matrix. Then
(see section 2) f is a state function, (i.e., there exists a matrix K such that f = Kx)
iff F U 1 is strictly proper and a state/input function (i.e., there exists matrices L, J
such that f = Kx + Ju) iff F U 1 is proper.
We first prove the second part of the theorem, i.e., that every dissipation function
is a state/supply function. Let (, ) = DT ()D() be a dissipation function. Then

(A.16) M T () M () = DT ()D().

Pre- and postmultiply by U 1 , to obtain

(A.17) (M ()U 1 ())T M ()U 1 () = (D()U 1 ())T D()U 1 ().

Since the left-hand side is proper, so is the right-hand side. This obviously implies
d
that D()U 1 () is proper. Hence D( dt )w is a state/input function and equivalently,
1738 J. C. WILLEMS AND H. L. TRENTELMAN

D() = KX() + JU () for suitable constant matrices K, J. From this it is readily

seen that there exists a matrix E such that (5.13) holds.
We now prove the first part of the theorem, i.e., that every storage function is a
state function. Let (, ) define a storage function for . Let D() be such that

( + )(, ) = M T () M () DT ()D().

By redefining
 T   
M () 0 M ()
M T () M () = ,
D() 0 I D()

and observing that a minimal state for the system with image representation v =
d
M ( dt )w is also a minimal state for the system with image representation
 d

M ( dt )
v= d w,
D( dt )

it suffices to prove the claim in the lossless case, i.e., when = .
Assume that (, ) = M T () M () is a symmetric canonical factorization of
and that (, ) = N T () N () is a symmetric canonical factorization of .

Postmultiply the identity = by U 1 () to obtain

(A.18) ( + )N T () N ()U 1 () = M T () M ()U 1 ().

Assume that Lk k is the term of degree k in the polynomial part of the matrix of
rational functions N ()U 1 (). Using that the right-hand side of (A.18) is proper in
, by equating powers of yields N T () Lk = 0. Express N () as

I
 
I

N () = N0 N1 . . . NL .
..
I L
 T
and use (A.18) to obtain N0 N1 . . . NL Lk = 0. Since the factorization
(, ) = N T () N () is canonical, N0 N1 . . . NL is surjective. Hence
(A.18) yields Lk = 0. This shows that N ()U 1 () is strictly proper and hence that
d
N ( dt )w is a state function as desired. Thus there exists a constant matrix K such
that N () = KX(). This shows that there exists a matrix P such that (5.12) holds.
This completes the proof of the theorem.
Proof of Theorem 5.7. We first prove the second part, the part regarding strong
average positivity. In this case it follows from Proposition 5.3 that (i, i) > 0
for all . Hence by Proposition 5.6, (, ) has a Hurwitz and an anti-Hurwitz
factorization. The associated storage functions, + and , satisfy
   
d d 2 d
(A.19) (Q+ (w) Q (w)) = kH wk kA wk2 .
dt dt dt
d
Let x = X( dt )w be a minimal state associated with a canonical factorization of .
By Theorem 5.5, there exist real symmetric matrices, say K+ and K , such that for
 ON QUADRATIC DIFFERENTIAL FORMS 1739

all w C (R, Rq ), we have

 T  
d d
Q+ (w) = X w K+ X w,
dt dt
 T  
d d
Q (w) = X w K X w.
dt dt
d d
Then if for all a there exists a solution w of A( dt )w = 0 such that (X( dt )w)(0) = a,
we obtain, by integrating along this solution,

Z 0  
d
(A.20) aT K+ a aT K a = Q+ (w)(0) Q (w)(0) = kH wk2 dt,
dt

whence K+ K , so + .
d
The problem is that there may not be a solution of A( dt )w = 0 for all a such that
d
(X( dt )w)(0) = a. In order to circumvent this difficulty we first prove the statements
of the second part under the additional assumption that || for some  > 0, in
addition to the assumption that (i, i) > 0 for all . Next, we modify to 
such that these conditions hold for  > 0, and, finally, take the limit for  0.
Assume that (i, i) > 0 for all R and || for some  > 0. The
system (5.5) allows an I/O representation, in the sense that there exists a permutation
matrix P such that
 d

U ( dt )
(A.21) Pv = d w,
Y ( dt )
d d
with det(U ) 6= 0 and G := Y U 1 proper. Let u = U ( dt )w, y = Y ( dt )w. There
exist constant matrices A, B, C, and D such that u, x, and y are related by dx dt =
1 d
Ax + Bu, y = Cx + Du. Since AU is biproper, A( dt )w is of the form F x + Lu with
L nonsingular. Using u = L1 F x and x(0) = a in these equations then results in a
d
solution of A( dt )w = 0. To show that AU 1 is indeed biproper, use Proposition 5.2
to obtain
 T  
U (i) U (i)
P P T =
Y (i) Y (i)
 T  
T U (i) U (i)
A (i)A(i)  .
Y (i) Y (i)

After pre- and postmultiplying by (U 1 (i))T and U 1 (i), respectively, we obtain

that
 T  
I I
(A.22) P P T = ((AU 1 )(i))T (AU 1 )(i) I.
G(i) G(i)

Since G is proper, AU 1 is proper, by the equality on the left. The inequality on the
right gives biproperness.
Consider a general and define  by  = + || + I. Then  satisfies the
above conditions and hence there exists (in the obvious notation) +
 and  such
+
that    . Observe that for 0 < 1 2 there holds 1 2 and deduce
1740 J. C. WILLEMS AND H. L. TRENTELMAN

from + + +
1 1 2 that 2 1 . Similarly, 2 1 . Consequently 2
+ +
1 1 2 . Prove (using for example the associated matrix representations)
+
that this monotonicity implies the existence of lim0 +
 =: 0 and lim0  =: 0 .

We now prove that + +
0 and 0 satisfy 0 and 0 , and subsequently
+
that any storage function of satisfies 0 0 . To prove the first part,

observe that +
  and   for  > 0 and take the limit for  0. To prove

the second part, assume that . Then  . Therefore   .
+

Now take the limit for  0.

We still have to prove the formulas (5.15) and (5.16) for the computation of
and + for the case that we only have (i, i) > 0 for all R and not
necessarily || for some  > 0. Let H be a symmetric Hurwitz factor of
 (, ):  (, ) = HT ()H
p (), as discussed in Proposition 5.6. In order to
make it unique, normalize H to  (0). It holds that

HT ()H () =  (, ) ( + )
 (, ).

Since  as  0 and  0 as  0, we also have that H converges. Clearly
the limit H0 satisfies (, ) = H0T ()H0 () and must be Hurwitz. The formula
for
+ is treated analogously.
0 (= ) follows. The situation for R
0
Proof of Proposition 6.2: (i) Compute Q (w) for w(t) = et a with <e() > 0
T
and a Cm . This integral equals a ( ,)a
+
. This w is not of compact support, but
an approximation argument can be used to complete the proof of (i). For <e() = 0
the result follows from Proposition 5.2. (ii) is proven similarly.
Proof of Theorem 6.3 : We prove that (3) (2) (1) (3). That (3) (2)
d
is trivial. In order to see that (2) (1), integrate dt Q (w) Q (w) from
to 0. We now prove that (1) (3). Assume first that satisfies the assumptions
(i, i) > 0 for all and || for some  > 0. By Theorem 5.7 we then have

(, ) AT ()A()
(A.23) + (, ) = .
+
d d d
This yields dt Q+ (w) = Q (w) kA( dt )wk2 for all w. Let x = X( dt )w be a minimal
d 2
state map of . By Theorem 5.5, Q+ (w) = kX( dt )wkK+ for some real symmetric
matrix K+ . Using this expression in (A.23) and integrating from to 0 yields that,
d d
R0
for all a such that X( dt )w(0) = a and A( dt )w = 0, we have aT K+ a = Q (w)dt.
This integral is 0, so we must have aT K+ a 0 (actually, such w does not have
compact support but, by an approximation argument, the integral cannot be < 0).
As in the proof of Theorem 5.7, it can be shown that for any initial condition a such
w exists. This proves that + 0. Take a general . As in the proof of Theorem 5.7,
first replace by  . By applying the previous to  , we can conclude that (in the
obvious notation) + 0. Then take the limit for  0.
Proof of Theorem 6.4. We will first run the circle (1) (2) (3) (7) (4)
(1).
(1) (2). This was proven in Proposition 6.2.
(2) (3). We have P T ()P () N T ()N () for C, <e() 0. Assume
that N P 1 has a pole such that <e() 0. Then there exists a vector v 6= 0 such
that P ()v = 0 while N ()v 6= 0. This, however, contradicts the above inequality.
 ON QUADRATIC DIFFERENTIAL FORMS 1741

d
(3) (7). Let QK (w) = kX( dt )wk2K be a storage function. We want to show
that K > 0. First we show that N P 1 is a proper rational matrix. We have
P T (i)P (i) N T (i)N (i) 0. Define G := N P 1 . Then for all such that
P (i) is nonsingular, we have GT (i)G(i) I, which shows that G is proper.
There exist matrices A, B, C, and D such that all smooth x, w1 and w2 satisfying
d d
x = Ax + Bw2 , w1 = Cx + Dw2 can be written as x = X( dt )w, w1 = N ( dt )w, and
d 1
w2 = P ( dt )w for some w C (R, R ). Since N P has no poles in <e() 0, the
matrix A can be chosen such that its eigenvalues are in the open left half of the complex
plane. Moreover, we may assume that the pair (C, A) is observable. Let a Rn .
d
Choose w2 = 0, let x satisfy dt x = Ax, x(0) = a, and let w1 = Cx. This shows that
d d
there exists w C (R, R ) such that (X( dt )w)(0) = a and w2 = P ( dt )w = 0. Also,
d d d
X( dt )w L2 [0, ) since A is a Hurwitz matrix. Since w1 = N ( dt )w = CX( dt )w, we
d
also have that N ( dt )w L2 [0, ). Thus we can integrate the dissipation inequality
from 0 to to obtain
    2 Z
d d
(A.24) X w (0)
kN ( )wk2 dt.
dt K 0 dt

This shows that aT Ka 0. Assume that aT Ka = 0. Then we must have w1 =

d
N ( dt )w = 0. By observability of the pair (C, A) this implies that a = 0.
(7) (4). Let Q be any storage function. Since X is a state map, by Theorem
d
5.5 there exists a real symmetric matrix K such that Q (w) = kX( dt )wk2K for all w.
By assumption, K is positive definite.
(4) (1). This was proven in Theorem 6.3.
The implications (7) (5), (5) (4), (7) (6), and (6) (4) are obvious.
Finally, if we assume observability, then the poles of N P 1 coincide with the
singularities of the polynomial matrix P . This shows that, under this assumption, (3)
and (30 ) are equivalent. This completes the proof.
Proof of Theorem 7.2. Consider the representation (3.10) of L . From the
fact that the factorization is canonical, it is easily seen that the mappings w1 7
d d
(N ( dt )w1 )(0) and w2 7 (M ( dt )w2 )(0) are surjective. Thus we have
   T    
d d d
L (w1 , ) = 0 N w1 M w2 = 0 for all w2 N w1 = 0.
dt dt dt
d
Similarly, L (, w2 ) = 0 iff M ( dt )w2 = 0. From this, the equivalence of (1) and (3)
is immediate.
To prove (1) (2), assume that, for some C, aT (, ) = 0, where a is a
complex vector. Define w1 (t) := et a. For any w2 and for all t, we then have
   
d
L (w1 , w2 )(t) = et aT , w2 (t) = 0.
dt

This implies w1 = 0, so a = 0, which proves that the rows of (, ) are linearly

independent over C. Similarly, we can prove that the columns of (, ) are linearly
independent.
Finally, we prove that (2) implies (3). Let C and put M ()a = 0 for some
complex vector a. We want to prove that a = 0. We clearly get N T ()M ()a = 0
so (, )a = 0. Since the columns of (, ) are linearly independent over C, this
yields a = 0. Likewise we can prove that N () has full column rank for all .
1742 J. C. WILLEMS AND H. L. TRENTELMAN

Proof of Proposition 8.2. Write (8.3) as

Z +   Z +  
d d
(A.25) (1 + ) kN wk2 dt (1 ) kP wk2 dt.
dt dt

Put = 1
1+ and conclude that (2) and (3) are equivalent (also in the noncanonical
case). To show that (1) (2), observe that ||(, ) = McT ()Mc () for the special
symmetric canonical factorization of (, ) corresponding to (3.15). Hence statement
(1) of the theorem is actually statement (2) for this special canonical factorization of
. It thus suffices to prove that if (2) holds for one canonical factorization, then it
holds for any. From matrix theory, it follows that two canonical factorizations

(A.26) M1T () M1 () = M2T () M2 ()

are related by M1 () = SM2 (), with S a nonsingular matrix. Hence (2) for M2
implies
Z +   Z +  
d d
kM1 wk2 dt = kM2 wk2 dt
dt dt
Z +  
d
2 kM2 wk2 dt
dt
Z +  
2 d
2
kM 1 wk2 dt
kSk dt

and (2) for M1 follows. Obviously this proof can be reversed with M1 playing the role
of M2 . If M2 comes from a noncanonical factorization, then S may not be nonsingular
and the proof goes through (but cannot be reversed).
Proof of Theorem 9.3. The proof is structured as follows. We first prove that (1)
(2). Subsequently, we show that (1) (3) and finally that (3) (1).
Rt
(1) (2). That Q  0 implies (2a) is obvious. In order to prove (2b) we
need the following lemma. Recall that a quadratic state function Q (or simply ),
d
Q (w) = kX( dt )wk2K , is called positive definite if K > 0.
Lemma A.2. Let M Rq [] be observable. Then there exists a positive definite
state function such that
 
d d
(A.27) Q (w) kM wk2 .
dt dt

Proof. We show that + , the supremal storage function associated with M T ()

d
M (), fits the bill. By Theorem 5.5, + is a state function, say, Q+ = kX( dt )wk2K+ .
Here we take X to be any minimal state map of M . Obviously + 0, since = 0
satisfies (A.27) and + = 0. In order to show that K+ > 0, let a 6= 0 be
arbitrary. We show that aT K+ a > 0. Factor M T ()M () = AT ()A(), with
A() anti-Hurwitz. Then
   
d d d
(A.28) Q+ (w) = kM wk2 kA wk2
dt dt dt

for all w C (R, Rq ). As in the proof of Theorem 5.7, it is easily seen that there
d d
exists w 6= 0 such that A( dt )w = 0 and X( dt )w(0) = a (show that AU 1 is biproper,
 ON QUADRATIC DIFFERENTIAL FORMS 1743

with M = col(U, Y ) an I/O partitioning). For this w in (A.28) we obtain aT K+ a =

R0 d

kM ( dt )wk2 dt > 0, where the strict inequality follows from the observability of
M.
We now return to the proof of (1) (2b) of Theorem 9.3. Let (, ) = M T ()
M () be the symmetric canonical factorization such that ||(, ) = M T ()M () (i.e.,
the one obtained by factoring = U T U , with p
the diagonal matrix consisting of
the nonzero eigenvalues of , and putting M := ||U ). By the above lemma there
d d
exists a positive definite state function such that dt Q (w) kM ( dt )wk2 . Now,
Rt Rt
Q  0 implies that there exists  > 0 such that Q 0, where  := ||.

Let + be the supremal storage function associated with  . Clearly +  ,
so + is also a storage function for . This immediately yields + + . Consider
the two-variable polynomial matrix + + . Clearly this defines a storage function
for as well, so + + +  + . According to Theorem 6.3, + 0. Since
is a positive definite state function, this implies that + is a positive definite state
function.
We show that (2) (1). Let X Rnq [] define a minimal state map for the
d
system v = M ( dt )w. By (2b) + is a positive definite state function. Hence there
T d
exists K+ = K+ > 0 such that Q+ (w) = kX( dt )wk2K+ . Factor M T ()M M () =
T
A ()A() with A anti-Hurwitz. Then we have
Z 0     Z 0  
d d d
(A.29) kM k2 dt = kX w(0)k2K+ + kA wk2 dt
dt dt dt
for all w D(R, Rq ). There exists a permutation matrix P such that P M = col(U, Y ),
d d
with det(U ) 6= 0 and Y U 1 proper. Write u = U ( dt )w, y = Y ( dt )w, and x =
d q
X( dt )w, with w ranging over C (R, R ). Write the associated input/state/output
representation. Hence there are constant matrices A1 , B1 , C1 , and D1 such that
these u, y, and x are exactly those that are related by the equations
(A.30) x = A1 x + B1 u, y = C1 x + D1 u.
As in the proof of Theorem 5.7, by strict positivity we have that AU 1 is biproper.
d
Thus there exist constant matrices F and L, det(L) 6= 0 such that A( dt )w = F x+Lu.
Solving this equation for u and substituting the result in (A.30) yields that the relation
d d
between a := A( dt )w, v = P col(u, y), and x = X( dt )w is given by linear equations of
the form
(A.31) x = A2 x + B2 a v = C2 x + D2 a
with (C2 , A2 ) observable and where the eigenvalues of A2 coincide with the singular-
ities of the spectral factor A(). This shows that for given a L2 ((, 0], R ) and
final condition x(0) = x0 , the corresponding v is in L2 ((, 0], R ). In other words
(A.31) defines a bounded operator from L2 ((, 0], R ) Rn to L2 ((, 0], R ),
mapping (a, x0 ) to v. Hence there exists a constants C1 and C2 such that
Z 0 Z 0
kvk2 dt C1 kak2 dt + C2 kx0 k2 .

Since K+ > 0, there exists  > 0 such that 1 K+ > C2 I and 1 > C1 . For this  we
have
Z 0 Z 0 
2 1 2 T
kvk dt kak dt + x0 K+ x0

1744 J. C. WILLEMS AND H. L. TRENTELMAN

which, by (A.29), is equivalent to

Z 0   Z 0  
d 2 1 d
kM k dt kM k2 dt.
dt  dt

This shows that is strictly half-line positive. Whence, (2) (1).

Next we show that (1) (3). We will only consider the semisimple case. That
R0
(1) (3a) follows from Proposition 5.2. To prove (3b), calculate Q (a)dt for

n
X
a(t) = k ek t ak
k=1

and obtain the result

T
1 1
Z 0 2 2

(A.32) Q (a)dt = .. T .. .
. .
n n
R0 R0
We know that, for some  > 0,
Q (w)dt  Q|| (w)dt for all w of compact

R0
support. An approximation argument yields that this implies Q (a)dt > 0 for
a 6= 0, equivalently for col (1 , 2 , . . . , n ) 6= 0. Hence, T > 0.
Finally, we turn to (3) (2). The implication (3) (2a) follows from Proposition
5.2. To show that (3) (2b), we show that T > 0 implies that the supremal storage
function + defines a positive definite state function. Let () = AT ()A() be an
anti-Hurwitz factorization. We claim that 1 , 2 , . . . , n are exactly the singularities
of A(), with associated vectors a1 , a2 , . . . , an in the kernel of A(k ), k = 1, 2, . . . , n.
Indeed, if has <e() > 0, then AT () is nonsingular. Hence (k , k )ak = 0
and <e(k ) > 0 implies A(k )ak = 0.
According to Theorem 5.7, it holds that
 
d d
(A.33) Q+ (w) = Q (w) kA wk2 .
dt dt
Pr d
For any solution w = k=1 k ek t ak of A( dt )w = 0, we thus have
T
1 1
Z 0 2 2

(A.34) Q+ (w)(0) = Q (w)dt = .. T .. .
. .
n n

Also, there exists a real symmetric matrix K+ such that

 
d
Q+ (w) = kX wk2K+ ,
dt
d
where X( dt ) is a minimal state map. Since T > 0, (A.34) implies that K+ > 0.
d
Indeed, let a 6= 0 be arbitrary. Let w C (R, Rq ) be such that A( dt )w = 0, say,
 ON QUADRATIC DIFFERENTIAL FORMS 1745
Pr d
w = k=1 ak ek t k , and X( dt )w(0) = a. Such w exists by strict positivity (see the
proof of Theorem 5.7). Thus we have
T
1 1
  2
d 2 2
aT K+ a =
X w(0)
= .. T .. > 0.
dt K+ . .
n n

This completes the proof of Theorem 9.3.

d d
Proof of Proposition 10.1. Let R( dt )w = 0 and w = M ( dt )` be, respectively,
a kernel and an observable image representation of B. Then we have R()M () =
d
0. Furthermore, v = RT ( dt )`0 is an image representation of B . Consider the
T d 0 T d
BLDF (R ( dt )` ) M ( dt )`. Note that this is the BLDF associated with the two-
variable polynomial matrix R()M (). By Theorem 3.1, there exists (, ) such
d d d
that dt Q (`0 , `) = (RT ( dt )`0 )T M ( dt )`, and by Theorem 5.5 Q (`0 , `) is a state
d d
function; in other words, if X( dt ) and Z( dt ) are minimal state maps of B and B ,
d d
then Q (`, `0 ) = (Z( dt )`0 )T KX( dt )` for some matrix K. The proposition follows by
T
taking Z := K Z if we can show that K is nonsingular. To show this, assume to the
contrary that Ka = 0. Let w B D(R, Rq ) be a trajectory emanating at t = 0 from
x(0) = a. It follows that
Z
v T wdt = 0
0

for all v B D(R, Rq ). Consider the function w : R Rq such that w(0) = 0 for
t 0 and w(t) = w(t) for t 0. Obviously it holds that
Z +
v T wdt = 0

for all v B D(R, Rq ). Therefore w belongs to the L2 (R, Rq ) closure of B (this

is the one point in this paper where C solutions are inadequate). Since w(t) = 0
for t 0, it must hold that x(0) = 0. Hence Ka = 0 implies a = 0, yielding the
result.
In order to prove Theorem 10.2 we use the following lemma. Recall that the inertia
of an nn complex Hermitian matrix H is the triple ( , 0 , + ), with the number
of negative eigenvalues, + the number of positive eigenvalues, and 0 (= n + )
the multiplicity of the zero eigenvalue.
Lemma A.3. Let L be a linear subspace of Rn . Consider the quadratic form
x Qx on Rn with Q = QT nonsingular. Let the inertia of Q be ( , 0, + ) and
T

assume that + = dim(L). Then aT Qa > 0 for all 0 6= a L iff aT Q1 a < 0 for
0 6= a L , and aT Qa 0 for all a L iff aT Q1 a 0 for a L .
Proof. Let L = ker(R) = im(M ) with R surjective and M injective. Then
L = im(RT ) = ker(M T ). Furthermore, aT Qa > 0 for 0 6= a L means M T QM > 0.
Consider the relations
   
MT  1 T
 M T QM 0
Q M Q R = ,
RQ1 0 RQ1 RT
   
MT   M T QM 0
Q M Q1 RT = .
R RQM RRT
1746 J. C. WILLEMS AND H. L. TRENTELMAN

The second relation shows that [M Q1 RT ] is nonsingular. The first shows that
in(M T QM ) + in(RQ1 RT ) = in(Q).
Hence M T QM > 0 implies RQ1 RT < 0. To get the case, replace Q by Q + I
and let  0.
Proof of Theorem 10.2. To prove (i) and (ii), combine Proposition 5.2 and Lemma
A.3 in the following way. For R fixed, define L := im(M (i)) = ker(R(i)).
Define Q := . Note that Q1 = as well. Using that M (i) and RT (i) are
injective, we get the equivalence
M T (i) M (i) > 0 R(i) RT (i) < 0
which yields statement (ii). Statement (i) follows from the second assertion of Lemma
A.3, which yields the same equivalence with nonstrict inequalities.
We now prove (iii). Again this can be proven using Lemma A.3, this time with
Q = I. We then get
M T (i)( I)M (i) 0 R(i)( I)1 RT (i) 0.
Using the formula + I = (1 2 )( I)1 , the latter is equivalent with
R(i) RT (i) R(i)RT (i).
This shows (iii).
In order to prove (iv), we need the following lemma.
Lemma A.4. Let (X, Z) be a matched pair of minimal state maps for B and B .
Define subspaces L Rr+2n , M Rr+2n by
d

w w M ( dt )`
(A.35) L := x Rr+2n | ` C (R, R ) x = X( dt d
)` (0) ,
d d
a a dt X( dt )`

T d

v v R ( dt )`0
(A.36) M := b Rr+2n | `0 C (R, R ) b = dt
d d
Z( dt )`0 (0) .
d
z z Z( dt )`0

Then dim(L) = n + m and L = M.

Proof. There exists a permutation matrix P such that
 
U
PM =
Y
d
with U Rmm [] and Y U 1 a proper rational matrix. If we define u = U ( dt )` and
d
y = Y ( dt )`, then u has the usual properties of input and y has the usual properties of
output of B. There exist matrices A Rnn , B Rnm , C Rpn , and D Rpm
d d
(with p = r m, the number of outputs) such that x = X( dt )`, u = U ( dt )`, and
d dx
y = Y ( dt )` are exactly related by dt = Ax + Bu, y = Cx + Du. Thus for all
` C (R, R ) we have:

0 Im
P 0 0 d
M ( dt )`  
d
0 I 0 X( d )` (0) = C D . X( dt )`
(0).
d
dt
d
In 0 d
U ( dt )`
0 0 I X( )`
dt dt A B
 ON QUADRATIC DIFFERENTIAL FORMS 1747

This implies

0 Im
P 0 0 C D
0 I 0 L im
In
.
0
0 0 I
A B

Here, in fact, equality holds. Indeed, given col (x0 , u0 ), take any x C (R, Rn ) and
u C (R, Rm ) such that x(0) = x0 and u(0) = u0 , and such that dx dt = Ax + Bu,
d d
y = Cx + Du. There exists ` C (R, R ) such that x = X( dt )`, u = U ( dt )`. This
shows that equality holds and that dim(L) = n + m.
We now prove that L = M. For all ` C (R, R ) and `0 C (R, R ) we
have that (10.4) holds. By evaluating this for t = 0, we immediately obtain that
L M. Thus it suffices to show that dim L = n + (r m). This is, however, an
immediate consequence of the fact that the number of inputs of B , m(B ) is equal to
r m. R
We now return to the proof of (iv) of Theorem 10.2. Assume that Q 0 and

let (, ) = X T ()KX(), with K = K T a storage function for , i.e., . In
d d
terms of w = M ( dt )`, x = X( dt )`, dx d d
dt = dt X( dt )` this inequality yields, in particular,
T
w(0) 0 0 w(0)
(A.37) x(0) 0 0 K x(0) 0.
dx dx
dt (0)
0 K 0 dt (0)

Denote the symmetric matrix in (3.16) by Q. Note that (A.37) says that aT Qa 0
for all a L, with L defined by (A.35). Since dim(L) = n + m = n + r+ , which
is exactly the number of positive eigenvalues of Q, it follows from Lemma A.3 that
aT Q1 a 0 for all a L = M. More explicitly,

0 0
(A.38) aT 0 0 K 1 a 0
1
0 K 0

for a L . A typical element of M has the form

T d

R ( dt )`0
a = dt
d d
Z( dt )`0 (t),
d
Z( dt )`0

where `0 C (R, R ). By letting t R be arbitrary, the inequality (A.38) yields

exactly the dissipation inequality
   
d d 0 2 T d
kZ ` kK 1 kR `0 k2
dt dt dt

which is the content of (4). To show (v), use (iv) and Theorem 9.3.

REFERENCES

[1] B.D.O. Anderson, Algebraic properties of minimal degree spectral factors, Automatica, 9
(1973), pp. 491500.
1748 J. C. WILLEMS AND H. L. TRENTELMAN

[2] B.D.O. Anderson and S. Vongpanitlerd, Network Analysis and Synthesis, PrenticeHall,
Englewood Cliffs, NJ, 1973.
[3] B.D.O. Anderson and E.I. Jury, Generalized Bezoutian and Sylvester matrices in multivari-
able linear control, IEEE Trans. Automat. Control, 21 (1976), pp. 551556.
[4] D.Z. Arov, Passive linear stationary dynamical systems, Siberian Math. J., 20 (1979), pp. 149
162.
[5] J.A. Ball, J.W. Helton, and J. William, Shift invariant subspaces, passivity, reproducing
kernels and H -optimization, in Contributions to Operator Theory and its Applications,
Oper. Theory Adv. Appl. 35, I. Gohberg, J.W. Helton, and L. Rodman, eds., Birkhauser,
Basel, 1988, pp. 265310.
[6] S. Barnett, Polynomials and Linear Control Systems, Marcel Dekker, New York, 1983.
[7] R.W. Brockett and J.L. Willems, Frequency domain stability criteria, parts 1 and 2, IEEE
Trans. Automat. Control, 10 (1965), pp. 255261; 401413.
[8] R.W. Brockett, Path integrals, Lyapunov functions, and quadratic minimization, in Proc.
4th Allerton Conference on Circuit and System Theory, University of Illinois, Monticello,
IL, 1966, pp. 685698.
[9] F.M. Callier, On polynomial spectral factorization by symmetric extraction, IEEE Trans.
Automat. Control, 30 (1985), pp. 453464.
[10] W.A. Coppel, Linear Systems, Notes in Pure Mathematics 6, Australian National University,
Canberra, 1972.
[11] P.A. Fuhrmann, Algebraic methods in system theory, in Mathematical System Theory, The
Influence of R.E. Kalman, A.C. Antoulas, ed., Springer-Verlag, New York, 1991, pp. 233
265.
[12] D.J. Hill and P.J. Moylan, Stability of nonlinear dissipative systems, IEEE Trans. Automat.
Control, 21 (1976), pp. 708711.
[13] R.E. Kalman, On the Hermite-Fujiwara theorem in stability theory, Quart. Appl. Math., 23
(1965), pp. 279282.
[14] R.E. Kalman, Algebraic characterization of polynomials whose zeros lie in certain algebraic
domains, in Proc. Nat. Acad. of Sci., 64 (1969), pp. 818823.
[15] V.L. Kharitonov, Asymptotic stability of an equilibrium position of a family of systems of
linear differential equations, Differentsialnye Uravneniya, 14 (1978), pp. 20862088.
[16] V. Kucera, Discrete Linear Control, The Polynomial Approach, John Wiley, Chichester, 1979.
[17] H. Kwakernaak, The polynomial approach to H optimal regulation, in H -Control Theory,
Lecture Notes in Math. 1496, E. Mosca and L. Pandolfi, eds., Springer-Verlag, Berlin, 1991,
pp. 141221.
[18] H. Kwakernaak and M. Sebek, Polynomial J-spectral factorization, IEEE Trans. Automat.
Control, 39 (1994), pp. 315328.
[19] G. Meinsma, Frequency Domain Methods in H Control, Ph.D. thesis, University of Twente,
The Netherlands, 1993.
[20] R.J. Minnichelli, J.J. Anagnost, and C.A. Desoer, An elementary proof of Kharitonovs
stability theorem with extensions, IEEE Trans. Automat. Control, 34 (1989), pp. 995998.
[21] A.C.M. Ran and L. Rodman, Factorization of matrix polynomials with symmetries, SIAM
J. Matrix Anal. Appl., 15 (1994), pp. 845864.
[22] P. Rapisarda and J.C. Willems, State maps for linear systems, SIAM J. Control Optim., 35
(1997), pp. 10531091.
[23] J.W. Schumacher, Transformations of linear systems under external equivalence, Linear Al-
gebra Appl., 102 (1988), pp. 133.
[24] H.L. Trentelman and J.C. Willems, The dissipation inequality and the algebraic Riccati
equation, in The Riccati Equation, S. Bittanti, A.J. Laub, and J.C. Willems, eds., Springer-
Verlag, New York, 1991, pp. 197242.
[25] H.L. Trentelman and J.C. Willems, H control in a behavioral context: The full informa-
tion case, IEEE Trans. Automat. Control, to appear.
[26] H.L. Trentelman and J.C. Willems, Every storage function is a state function, Systems
Control Lett., 32 (1997), pp. 249259.
[27] S. Weiland and J.C. Willems, Dissipative systems in a behavioral context, Math. Models
Methods Appl. Sci., 1 (1991), pp. 125.
[28] J.C. Willems, Least squares stationary optimal control and the algebraic Riccati equation,
IEEE Trans. Automat. Control, 16 (1971), pp. 621634.
[29] J.C. Willems, Dissipative dynamical systemsPart I: General theory, Part II: Linear systems
with quadratic supply rates, Arch. Rational Mech. Anal., 45 (1972), pp. 321351; 352393.
[30] J.C. Willems, On the existence of a nonpositive solution to the Riccati equation, IEEE Trans.
Automat. Control, 19 (1974), pp. 592593.
 ON QUADRATIC DIFFERENTIAL FORMS 1749

[31] J.C. Willems, From time series to linear system. Part I: Finite dimensional linear time in-
variant systems; Part II: Exact modelling; Part III: Approximate modelling, Automatica,
22 (1986), pp. 561580; 675694; 23 (1987), pp. 87115.
[32] J.C. Willems, Models for dynamics, Dynamics Report, 2 (1989), pp. 171269.
[33] J.C. Willems, Paradigms and puzzles in the theory of dynamical systems, IEEE Trans. Au-
tomat. Control, 36 (1991), pp. 259294.
[34] D.C. Youla and M. Saito, Interpolation with positive-real functions, J. Franklin Inst., 284
(1967), pp. 77108.