Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Linear Dynamical System Kalman

Download as pdf or txt
Download as pdf or txt
You are on page 1of 41
At a glance
Powered by AI
The paper discusses the relationship between physical dynamical systems and their mathematical representations using differential equations and transfer functions.

The author aims to identify the linear dynamical system corresponding to a given impulse response matrix.

The author proposes axioms that provide a precise, abstract definition of a physical dynamical system, generalizing the Newtonian view of causality.

J.S.I.A.M. CONTROI Ser. A, Vol. 1, No.

Printed in U.,q.A., 1963

MATHEMATICAL DESCRIPTION OF LINEAR DYNAMICAL SYSTEMS*


R. E.

KALMAN

Abstract. There are two different ways of describing dynamical systems: (i) by means of state w.riables and (if) by input/output relations. The first method may be regarded as an axiomatization of Newtons laws of mechanics and is taken to be the basic definition of a system. It is then shown (in the linear case) that the input/output relations determine only one prt of a system, that which is completely observable and completely controllable. Using the theory of controllability and observability, methods are given for calculating irreducible realizations of a given impulse-response matrix. In particular, an explicit procedure is given to determine the minimal number of state varibles necessary to realize a given transfer-function matrix. Difficulties arising from the use of reducible realizations are discussed briefly.

1. Introduction and summary. Recent developments in optimM control system theory are bsed on vector differential equations as models of physical systems. In the older literature on control theory, however, the same systems are modeled by ransfer functions (i.e., by the Laplace transforms of the differential equations relating the inputs to the outputs). Two differet languages have arisen, both of which purport to talk about the same problem. In the new approach, we talk about state variables, transition equations, etc., and make constant use of abstract linear algebra. In the old approach, the key words are frequency response, pole-zero patterns, etc., and the main mathematical tool is complex function theory. Is there really a difference between the new and the old? Precisely what are the relations between (linear) vector differential equations and transferfunctions? In the literature, this question is surrounded by confusion [1]. This is bad. Communication between research workers and engineers is impeded. Important results of the "old theory" are not yet fully integrated into the new theory. In the writers view--which will be argued t length in this paperthe diiIiculty is due to insufficient appreciation of the concept of a dynamical system. Control theory is supposed to deal with physical systems, and not merely with mathematical objects such as a differential equation or a transfer function. We must therefore pay careful attention to the relationship between physical systems and their representation via differential equations, transfer functions, etc.
* Received by the editors July 7, 1962 and in revised form December 9, 1962. Presented at the Symposium on Multivariable System Theory, SIAM, November 1, 1962 at Cambridge, Massachusetts. This research was supported in part under U. S. Air Force Contracts AF 49 (638)-382 and AF 33(616)-6952 as well as NASA Contract NASr-103. Research Institute for Advanced Studies (RIAS), Baltimore 12, Maryland.

152

LINEAR DYNAMICAL SYSTEMS

153

To clear up these issues, we need first of all a precise, abstract definition of a (physical) dynamical system. (See sections 2-3.) The axioms which provide this definition are generalizations of the Newtonian world-view of causality. They have been used for many years in the mathematical literature of dynamical systems. Just as Newtonian mechanics evolved from differential equations, these axioms seek to abstract those properties of differential equations which agree with the "facts" oi classical physics. It is hardly surprising that under special assumptions (finite-dimensional state space, continuous time) the axioms turn out to be equivalent to a system of ordinary differential equations. To void mthemtical diificulties, we shll restrict our attention to linear differential equations. In section 4 we formulate the central problem of the pper: Given an (experimentally observed) impulse response matrix, how can we identify the linear dynamical system which generated it? We propose to call any such system a realization of the given impulse response. It is an irreducible realization if the dimension of its state space is minimal. Section 5 is a discussion of the "canonical structure theorem" [2, 14] which describes abstractly the coupling between the external variables (input and output) and the internal variables (state) of any linear dynamical system. As a immediate consequence of this theorem, we find that a linear dynamical system is an irreducible realization of an impulse-response matrix if and only if the system is completely controllable and completely observable. This important result provides a link between the present pper and earlier investigations in the theory of controllability and observability

[3-5].
Explicit criteria for complete controllability and complete observability
are reviewed in a convenient form in section 6. Section 7 provides a constructive computational technique for determining the canonical structure of a constant linear dynamical system.. In section 8 we present, probably for the first time, complete and rigorous theory of how to define the state variables of a multi-input/multi-output

constant linear dynamical system described by its transfer-function matrix. Since we are interested only in irreducible realizations, there is a certain unique, well-defined number n of state variables which must be used. We give a simple proof of a recent theorem of Gilbert [5] concerning the value of n. We give canonical forms for irreducible realizations in simple cases. We give constructive procedure (with examples) for finding an irreducible realization in the general case. Many errors have been committed in the literature of system theory by carelessly regarding transfer functions and systems as equivalent concepts. A list of these has been collected in section 9. The field of research outlined in this paper is still wide open, except

perhaps in the case of constant linear systems. Very little is known about irreducible realizations of nonconstant linear systems. It is not clear what dditional properties--besides complete controllability and complete observability--are required to identify the stability type of a system from its impulse response. Nothing is known about nonlinear problems in this context. Finally, the writer would like to acknowledge his indebtedness to Professot E. G. Gilbert, University of Michigan, whose work [5] predates this and whose results were instrumental in establishing the canonical structure theorem.
2. Axiomatic definition of a dynamical system. Macroscopic physical phenomena are commonly described in terms of cause-and-effect relationships. This is the "Principle of Causality". The idea involved here is at least as old as Newtonian mechanics. According to the latter, the motion of a system of particles is fully determined for all future time by the present

positions and momenta of the particles and by the present and future forces acting on the system. How the particles actually attained their present positions and momenta is immaterial. Future forces can have no effect on what happens at present. In modern terminology, we say that the numbers which specify the instantaneous position and momentum of each particle represent the state of the system. The state is to be regarded always as an abstract quantity. Intuitively speakig, the state is the minimM amount of information about the past history of the system which suffices to predict the effect of the past upon the future. Further, we say that the forces acting on the particles are the inputs of the system. Any variable in the system which ca be directly observed is an output. The preceding notions can be used to give a precise mathematical definition of a dynamical system [6]. For the present purposes it will be convenient to state this definition in somewhat more general fashion [14]. DFNTON 1. A dynamical system is a mathematical structure defined by the following axioms" There is given a state space and a set of values of time 0 at (D) is a topological which the behavior of the system is defined; space and 0 is an ordered topological space which is a subset of the real numbers. There is given a topological space t of function, s of time de(D) fined o (R), which are the admissible inputs to the system. For any initial time to in O, any initial state x0 in 2;, and any (D) input u in tl defined for __> to, the future states of the system are determined by the transition function X 0 X 0 X 2; Z, which is written as ,(t; to, x0) This functio is defined

."

LINEAR DYNAMICAL SYSTEMS

]55

only for => to. Moreover, any to _-< tl _-< t in O, any x0 in ;, and any fixed u in 2 defined over [to, tl] 91 (R), the following relations hold"

(D-i) (D-ii)
v

9u(t0;t0,x0) u(t2 ;0, X0)

x0,

u(t.- ;ti,u(tl ;to, X0)).

In addition, the system must be nonanticipatory, i.e., if u, andu von[t0,t] Owehave

(D-iii)

,,(t; to, x,,)

,,,(t; to, xo).

Every output of the system is a function :0 X 2 reals. The functions and are continuous, with respect to the defined for 2, 0, and t and the induced product topolotopologies gies. In this paper we will study only a very special subclass of dynamical systems: those which are real, finite-dimensional, continuous-time, and

(D4) (Ds)

linear.

n-dimensional real R" "Real, finite-dimensional" means that 2 set of real numbers. R linear space. "Continuous-time" means that 0 "Linear" means that is linear on 2 X 2; and f is linear on Z. By requiring and b to be sufficiently "smooth" functions, we can deduce from the axioms a set of equations which characterize every real, finitedimensional, continuous-time, and linear dynamical system. The proof of this fact is outside the scope of the present paper [14]. Here we shall simply assume that every such system is governed by the equations

(2.1) (2.2)

dx dt

F(t)x

+ G(t)u(t),

y(t)

where x, u, and y are n, m, defined on the whole real line < and p-vectors* respectively, and the matrices F(t), G(t), and H(t) are continuous functions of the time t. We call (2.1-2) the dynamical equations of the system. It is instructive to check whether the axioms are satisfied. (D) is obviR O R The state of the system is the vector x. ously true; we hve To satisfy (D.), we must specify the class of all inputs, that is, a subclass of all vector functions u(t) (u.(t),..., u,(t)). To define t, we shall assume that these functions are piecewise continuous; this is suiiciently

H(t)x(t),

Vectors will be denoted by small Roman letters, matrices by Roman cpitls. The components of vector x are xi components of a mtrix A re ai On the other are vectors, nd F .4, F B are matrices. A is the transpose of A. hand, x 1, x

156

R.E. KALMAN

general for most applications. We have exactly p observations on the system (the components of the vector y) and by (2.2) they are functions of t, x. Hence (D) is satisfied. To check (D3), we recall that the general solution of (2.1.) is given by

(2.:)

(t; to, zo) =- z,

(t, to)zo

(t, )()u()

c,

where (t, r) is the transition matrix of the free differential equation defined by F (t) [4, 7] I. Since (2.3) is valid for any >= to (in fact, also for < to), is well defined. Property (D3-i) is obvious. (D.ii) follows from the composition property [4, 7] of the transition matrix:

(t, )(, ), which holds for every set of real numbers t, , o-. Indeed, (2.4) is simply the linear version of (D-ii). (D-iii) is obvious from formula (2.3). The continuity axiom (D) is satisfied by hypothesis. Evidently e given by (2.3) is linear on the cartesian product of 2; with
(t, )
the linear space of vector-valued piecewise continuous functions. We call a linear dynamical system (2.1-2) constant, periodic, or analytic whenever F, G, and H are constant, periodic, or analytic in t. It is often convenient to have a special name for the couple 0 X 2. Giving a fixed value of (t, x) is equivalent to specifying (t, x) at some time (t) the state (x) of the system. We shall call (t, x) a phase and 0 X 2 the phase space. (Recall the popular phrase" "phases" of the

(2.4)

Moon.) To justify our claim----implicit in the above discussion--that equations (2.1-2) are a good model of physical reality, we wish to point out that
these equations can be concretely simulated by a simple physical system" a general-purpose analog computer. Indeed, the numbers (or functions) constituting F, G, and H may be regarded as specifying the "wiring diagram" of the analog computer which simulates the system (2.1-2) (see, for instance, [8]).
3. lquivalent dynamical systems. The state vector x must always be regarded as an abstract quantity. By definition, it cannot be directly measured. On the other hand, the inputs and outputs of the system (2.1-2) have concrete physical meaning. Bearing this in mind, equations (2.1-2) admit two interpretations" () They express relations involving the abstract linear transformations F(t), G(t), nd H(t). (b) At any fixed time, we take an arbitrary but fixed coSrdinate system
I.e.,
I
q, is a solution of d/dt unit matrix for all 7.

F(t), subject to the initial condition (7, 7)

LINEAR DYNAMICAL SYSTEMS

157

in the (abstract) vector space 2. Then the symbol x (x, x) is interpreted as the numerical n-tuple consisting of the coSrdiuates of the abstract state vector which is also denoted by x. F, G, and H are interpreted as the matrix representations of the abstract linear transformations denoted by the same letters under (a). To describe the behavior of a dynamical system in concrete terms, the second point of iew must be used. Then we must also ask ourselves the question" To what extent does the description of a dynamical system depend on the urbitrary choice of the coordinate system in the state space? (No such arbitrariness occurs in the definition of the numerical vectors u, y since the input and output variables u nd y. are concrete physical quantities.) This question gives rise to the next definition. DEFINITION 2. Two linear dynamical systems (2.1-2), with state vectors x, are algebraically equivalent whenever their numerical phase vectors are related for all as

(3.1)

(t, )

(t, T(t)x),

where T(t) is a n X n matrix, nonsingular for all and continuously differentiable in t. In other words, there is a 1-1 differentiable correspondece 2; and 0 between the phase spaces 0 Remark: We could generalize this definition of equivalence to (, 2) (-(t), T(t)x) where is un increasing function of t. But this involves distortion of the time scale which is not permitted in Newtonian physics. Algebraic equivalence implies the following relations between the defining matrices of the two systems"

,.

(t, -)
(3.2)

T(t)(t, -)T-(-),

(t) (t) IZI(t)

.(t)T-(t)
T(t)G(t),

H(t)T-(t).

In general, lgebmic equivalence does not preserve the stability properties of a dynamical system [7, 9, 10]. For this it is necessry and sufficient to have topological equivalence" algebraic equivalence plus the condition

T(t)F(t)T-(t),

(3.3)

liT(t)

--< c

and

l]T-(t) =<

c,

is the euclidean norm*. where c and c are fixed constants, and A nonconstnt system may be algebraically and even topologically equivalent to constant system. The latter case is called by Markus [11]
Let O, 2, and E have the usual topologies induced by the euclidean norm. Then the product topologies induced on O X 2: and O X are equivalent if and only if (3.3) holds.

]58

R.E. KALMAN

"kinematic similarity". Moreover, two constant systems may be algebraically and topologically equivalent without T(t) being a constant. To bypass these complications, we propose DEFINITION 3. Two constant linear dynamical systems are strictly equialent whenever their numerical phase vectors are related for all as (t, 2) (t, Tx), where T is a nonsingular constant matrix. Evidently strict equivalence implies topological equivalence.

4. The impulse-response matrix and its realization by a linear dynamical system. Sections 2-3 were concerned with mathematics, that is, abstract matters. If we now take the point of view of physics, then a dynamical system must be "defined" in terms of quantities which can be directly observed. For linear dynamical systems, this is usually done in the following way. We consider a system which is at rest at time t0 i.e., one whose input and outputs have been identically zero for all =< to. We apply at each input in turn a very sharp and narrow pulse. Ideally, we would take (t 0), where is the Dirac delta function, t. is the Kronecker symbol, and 1 __<= i, j =< m. We then observe the effect of each vector input u((t) on the outputs, which are denoted by u(t; j). The matrix S(t, to) [y(t; j)] so obtained is called the impulse-response matrix of [si.(t, t0)] the system. Since the system was at rest prior to to, we must define S(t, to) =- 0 for < t0. We also assume, of course, that S is continuous in

tandt0fort >
With these conventions, the output of a linear system originally at rest is related to its input by the well-known convolution integral"

(4.1)

y(t)

S(t, v)u(-) dr.

In much of the literature of system theory [12] (and also at times in physics) formula (4.1) is the basic definition of a system. The Fourier transform of S is often called "the system function" [13, p. 92]. Unfortunately, this definition does not explain how to treat systems which are not "initially at rest". Herme we may ask, "To what extent, if any, are we justified in equating the physical definition (4.1) of a system with the mathematical one provided by (2.1-2)?" Suppose that the system in question is actually (2.1-2). Then (2.3) shows that

(4.2)

S(t, -)

H(t)(t, r)G(r),
=0,

t=>

r,

t<r.

The right-hand side of the first equation (4.2) is defined lso for ; then the left-hand side my be regarded s the "bckwrd impulse response", whose physical interpreta, tion is left to the reder.

LINEAR

I)YNAMICAL

SYSTEMS

159

Thus it is trivial to calculate the impulse-response matrix of a given linear dynamical system. The converse question, however, is non trivial and interesting. Whez and how does the impulse-response matrix determine the dynamical equations of the system? This problem is commonly called the identification of the system from its impulse-response matrix. Having been given an impulse-response matrix, suppose that we succeed in finding matrices F, G, and H such that (4.2) holds. We have then identified a physical system that may have been the one which actually generated the observed impulse-response matrix. We shall therefore call (2.12) a realization of S(t, r). This terminology is justified because the axioms given in section 2 are patterned after highly successful models of classical macroscopic physics; in fact, the system defined by (2.1-2) can be concretely realized, actually built, using standard analog-computer techniques in existence today. In short, proceeding from the impulse-response matrix to the dynamical equations we get closer to "physical reality". But we are also left with a problem" Which one of the (possibly very many) realizations of S(t, r) is the actual system that we are dealing with? It is conceivable that certain aspects of a dynamical system cannot ever be identified from knowledge of its impulse response, as our knowledge of the physical world gained from experimental observation must always be regarded as incomplete. Still, it seems sensible to ask how much of the physical world can be determined from a given amount of experimental

data. The first clear problem statement in this complex of ideas and the first results appear to be due to the writer [2, 14]. First of all we note TIEOEM 1. An impulse-response matrix S(t, r) is realizable by a finitedimensional dynamical system (2.1--2) if and only if there exist continuous matrices P(t) and Q(t) such that

(4.3)

S(t, r)

P(t)Q(r) for all t,

r.

side of (4.2) as H(t)q(t, 0)(0, v)G(r), with the aid of (2.4). Sufficiency is equally obvious. We set E(t) O, G(t) Q(t), and tt(t) P(t). Then (t, r) I

Proof. Necessity follows by writing the right-hand

and the desired result follows by (4.2). A realization (2.1-2) of S(t, -) is reducible if over some interval of time there is a proper (i.e., lower-dimensional) subsystem of (2.1-2) which also realizes S(t, r).As will be seen later, a realization of S (particularly the one given in the previous paragraph) is often reducible. An impulse-response matrix S is stationary whenever S(t, r) S(t q- o5 r q- r) for all real numbers t, r, and a. S is periodic whenever

].60

R.E. KALMAN

the preceding relation holds for all t, r, and some An iinpulse-response matrix is analytic whenever S is analytic in and r; if (4.3) holds, then P and Q must be analytic in t. The main result, whose proof will be discussed later, is the following [14]: THEOREM 2. Hypothesis: The impulse-response matrix S satisfies (4.3) and is either periodic (and continuous) or analytic. Conclusions: (i) There exist irreducible realizations of S, all of which have the same constant dimension n and are algebraically equivalent. (if) If S is periodic [analytic] so are its irreducible realizations. Topological equivalence cannot be claimed in general. It may happen that S has one realization which is asymptotically stable and another which is asymptotically unstable [15]. Hence it may be impossible to identify the stability of a dynamical system from its impulse response! This surprising conclusion raises many interesting problems which are as yet unexplored [15]. If S is not periodic or analytic, it may happen that the dimension n(t) of an irreducible realization is constant only over finite time intervals. In the stationary case, Theorem 2 can be improved [14]. THEOREM 3. Every stationary impulse-response matrix S(t, r) W(t r) satisfying (4.3) has constant irreducible realizations. All such realizations are strictly equivalent. In view of this theorem, we may talk indifferently about a stationary impulse-response matrix or the dynamical system which generates it--as has long been the practice in system, theory on intuitive grounds. But note that we must require the realization to be irreducible. For nonconstant systems, such a conclusion is at present not justified. The requirement of irreducibility in Theorem 3 is essential; disregarding it can lead--and has led--to serious errors in modeling dynamical systems. (See section 9.) In many practical cases, it is not the weighting-function matrix W (t r) (see Theorem 3) which is given, but its Laplace transform, the transferfunction matrix Z (s) [W (t) ]. Then condition (4.3) has an interesting equivalent form, which is often used as a "working hypothesis" in engineering textsTHEOREM 4. A weighting-function matrix W(t r) satisfies (4.3) if and O, only if its elements are linear combinations of terms of the type te (i n n). Hence every element of the transfer-function 1, j 1, 1, matrix is a ratio of polynomials in s such that the degree of the denominator polynomial always exceeds the degree of the numerator polynomial. This result is provedd in [14]. I implies that the realization of an impulseresponse matrix is equivalent to expressing the elements of F, G, and H as functions of the coefficients of the numerator nd denominator polynomials of elements of Z (s). (See section 8.) In the remainder of the paper, we wish to investigate two main problems

LINEAR

l)YNAMICAL

SYSTEMS

161

arising in the theory sketched above" (i) Explicit criteria for reducibility. (ii) Construction of irreducible realizations. Remark. Elementary expositions of system theory often contain the statement that the operator d/dt (--s) is a "system." Is a it system in the same sense as that word is used here? The answer is no. To define such a system rigorously in accordance with the axioms introduced in section 2, one must proceed as follows. The output of the system, which by definitio is the derivative of the input, is given by

(3.4)

y(t)

du(t)
dt

(t, x(t)),

so that at any fixed t, u(t) must be a point function of (t, x(t) ). Therefore the state space 2; must include the space 2 of functions on which the opera2. Then 2 is usually infinite tor d/dt is defined. It is simplest to let 2 dimensional because 2 is. Thus we define the state x x(t) as the function t. The mapping (t; to, xt0) assigns to the function u(r), defined for all defined for -<_ to the function xt, which is equal to Xto on to and equal x0

=<

r =< t. to uont0 In this paper, the finite dimensionality of 2; is used in an essential way, which rules out consideration of the "system" d/dt in all but trivial cases.

=<

5. Canonical structure of linear dynamical systems. The concept of irreducibility can be understood most readily with the help of the writers "canonical structure theorem" for linear dynamical systems [2, 14]. Before presenting and illustrating this central result, it is necessary to recall some definitions and facts concerning the controllability and. observability of linear dynamical systems. DEFINITION 4. A linear dynamicnl system (2.1-2) is completely controllable at time to if it is not algebrtically equivalent, for all to, to a system of the type

(a) dx/dt F(t)x Fr(t)x + G(t)u(t) (5.1) (b) dx:/dt F (t) x y(t) Hi(t)x(t) He(t)x(t). (c) (In (5.1), x and are vectors of n and n n n components respec-

x:

>=

.-

tively.) In other words, it is not possible to find a coSrdinate system iu which the state variables x are separated into two groups, x (x, x) and x (Xnl+I,, Xn), such that the second group is not affected either by the first group or by the inputs to the system. If one could find such a

162

E. KALMAN

FIGURE 1.

coSrdinate system, we would have the state of affairs depicted schematically in Fig. 1. Clearly, controllability is a system property which is completely independent of the way in which the outputs of the system are formed. It is a

property of the couple F(t), G (t) }. The "dul" of controllability is observbility, which depends only on the outputs but not on the inputs. DEFINITION 5. A linear dynamical system (2.1-2) is completely observable at time to if it is not algebraically equivalent, for all <= to, to any system of the type

dxl/dg
(5.2) (b)
(c)

Fll(g)x1(g)

dx2/dt
y(t)

F2(t)x(t) + F2(t)x + G(t)u(t) Hl(t)xl(t).

(Again, x is an nl-vector and x is an (n nl)-vector.) In other words, it is not possible to find a coSrdinate system in which the state variables xi are separated into two groups, such that the second group does not affect either the first group or the outputs of the system. If such a coSrdinate system could be found, we would have the state of affairs depicted in Fig. 2. The above definitions show that controllability und observbility are preserved under algebraic equivalence. These properties are coSrdintefree, i.e., independent of the particular choice of basis in the state space. The equivalence of the present definitions with other more abstract

LINEAR DYNAMICAL SYSTEMS

163

FIGURE 2.

definitions of controllability may be found in [4]. As to observability,


we note that the duality relations

t-- to

to--

(5.3)

(b) (c)

F(t- to) v= F(toG(t- to) :* H(toH(t- to)

(d)

transform the system (5.2) into (5.1). Itence all theorems on controllability can be "dualized" to yield analogous results on observability. It can be shown that in applying definitions 4-5 to constant systems it is immaterial whether we require algebraic or strict equivalence [14]. IIence-as one would of course expect--for constaut systems the notions of complete controllability and complete observability do not depend on the choice of to. EXAMPLE 1. A simple, well-known, and interesting case of a physical system which is neither completely controllable nor completely observable is the so-called constant-resistance network shown in Fig. 3. Let Xl be the magnetic flux in the inductor and x2 the electric charge on the capacitor in Fig. 3, while ul(t) is a voltage source (zero short-circuit resistance) and yl(t) is the current into the network. The inductor and capacitor in the network may be time-varying, but we assume--this is the constant-resistance condition--that L (t) and C (t) are related by:

t), t), G(to- t),

L(t)/C(t)

(L(t), c(t) > o).

164

n.E. KALMAN

x/

L (l)

x2

C (t)

-I GURE 3.

The differential equations of the network are

dx/dt dx,2/dt
y(t)
If we let

-[1/L(t)]x + ul(t), -[1/C(t)]x,2 + u(t), [1/L(t)]xt [1/C(t)]x2

+ u.(t).

x
the dynamical equations become

+x

/2,

32-- (Xl-

(5.4)

d2l/dt d2/dt
y(t)

--[1/L(t)]2 + ul(t), [1/L (t)],, 2[1/L(t)]22 + u(t).*

Here the state variable 2 is controllable but not observable, while 2: is observable but not controllable. For obvious reasons, the subsystem (b) of (5.1) may be regarded as (completely) uncontrollable, while subsystem (b) of (5.2) is (completely) unobservable. In view of linearity, it is intuitively clear that it must be possible to arrange the components of the state vectorreferred to a
Note that this equation does not correspond to (2.2) but to y(t) H(t)x(t) minor point. In fact, Axiom (D ) may be generalized to "(D)" Every output is a function of t, x(t), and u(t)." This entails only minor modifications

+ J(t)u(t). This is

as far as the results and arguments of the present paper are concerned.

LINEAR DYNAMICAL SYSTEMS

165

suitable (possibly time-varying) co6rdinate system into four mutally exclusive parts, as follows: Part (A): Completely controllable but unobservable. Part (B): Completely controllable and completely observable. Part (C): Uncontrollable and unobservable. Part (D) Uncontrollable but completely observable. The precise statement of this idea is [2, 14]: THEOREM 5 (Canonical Structure Theorem,). Consider a fixed linear dynamical system (2.1-2). (i) At every fixed instant of time, there is a co6rdinate system in the state space relative to which the components of the state vector can be decomposed into four mutually exlusive parts
X

(X A,

B, X c, xD),

which correspond to the scheme outlined above. (ii) This decomposition can be achieved in many ways, but the number n z(t) in each part is the same for any such of state variables nt(t), decomposition. (iii) Relative to such a choice of co6rdinates, the system matrices have the canonical form

F(t)

F(t) 0

FA(t) F Ac (t) 0 F(t) 0 Fee(t)


0 0

F(t) FC)(t) F))(t)

G( t)
and

G t) | 0

H(t) [0 H(t) 0 HD(t)]. In view of this theorem, we shall talk, somewhat loosely, about "Parts (D) of the system." Thus the system (5.4) consists of Parts (A), (A) and (D). The canonical form of F, G, and H can be easily remembered by reference
to the causal diagram shown on Fig. 4. It is intuitively clear (and can be easily proved) that algebraically equivalent systems have the same canonical structure. Unfortunately, the coSrdinate system necessary to display the canonical form of F, G, and H will not be continuous in time unless n(t), n,(t) are constants. If these dimension numbers vary, we cannot call the various

166

R.E. KALMAN

FIGURE 4.

parts of the canonical structure "subsystems." For constant systems this diiIiculty does not arise. More generally, we have" THEOREM 6. For a periodic or analytic linear dynamical system (2.1-2) n, are constants, and the canonical decomthe dinension numbers nA position is continuous with respect to t. An illustration of the canonical structure theorem is provided by EXAMPLE 2. Consider the constant system defined by
-3 26 30 30
3 -2 0 0

-3 36 39 43
3 -1 0 1

0 -3 -2 -3

1 -25

-27
-32

H
where

[-5

-8

1
2

5].

We introduce new co6rdinates by letting


2 1 --2 --6
0 1 3 1
3 1 --3 --9 3 -2 0 0

Tx,
-2 1 3 6

T
and

0 0 0 1

1 0 0 1

0 0 1 0

With respect to these new coSrdintes the system matrices assume the

LINEAR DYNAMICAL SYSTEMS

167

canonical form"

17

TFT-I=
0

2 0 0 0

4 -1 0 0

1 0 -3 0

--1 1 -2 1

=
and

TG=

tt

HT

On the other hand, if we define the new co6rdinates by

3 1

[0
4 1

1].
--3 --1 6 6

T=

0 0

-5
-6 0 1 3 1

--7.5
9 3 --3 --3 --1

O.5 1
--0.5 0 1 --0.5

T_i__

1 0 0 1

then the system matrices become

fi=

2 0 0 0 1 1 0 0

1 --1 0 0 2 1 0 0
]

1 0 --3 0

0 1 0 1

and

fl

[o

]].

The numerical values of these two canonical forms are different, yet Theorem 5 is verified in both cases. In the second case the connections from Part (D) to Parts (A) and (C) are missing. This is not a contradiction since Theorem 5 does not require that all the indicated casual connections in Fig. 4 be actually present. The transfer-function matrix of the system is easily found from the canonical representation. The co6rdinate transformations affect only the

168

n.E. KALMAN

internal (state) variables, but not the external (input and output) variables; consequently the impulse response matrix is invariant under such transformations. We get by inspection:

s+l

s-t-1

It would be rather laborious to determine these transfer functions directly from the signal-flow graph [16] corresponding to F, G, and H. EXAMPLE 3. A far less trivial illustration of the canonical decomposition theorem is provided by the following dynamical system, which occurs in the solution of a problem in the theory of statistical filtering [17]. Let A be an arbitrary positive function of and define
F

t/4A ta/2A t2/2A t4/4A]


|ta/2A|,

0 0

lo]
1 0

L?/eAJ
H= [0
We introduce new state variables
1

0].

2(t)
where

(t)x(t),
0 2 0
0 -t 1

T(t)

T-(t)
Then

rte/2

1/2

t/2]
t4/4A_ ta/4A_ t/4A

l(t)

T(t)F(t)T-(t)- 5/(t)T-l(t)

o t/2A
d(t)
T(t)G(t)

LINEAR DYNAMICAL SYSTEMS

169

H(t)

H(t)T-l(t)

It 1,01 11.

Hence the system consists of Parts (B D), with n, nc n, 1. It is interesting that the canonical decomposition is of constant dimension, even though the system may be neither periodic nor analytic.
The preceding examples illustrate special cases of a noteworthy general relationship which exists between the canonical structure of a dynamical system and irreducible realizations of an impulse-response matrix. The main facts here are the following: THEOREM 7. (i) The impulse-response matrix of a linear dynamical system (2.1-2) depends solely on Part (B) or the system and is given explicitly by

(5.5)

S(t, r)

H(t)"(t, r)G(r),

where is the transition matrix corresponding to (ii) Any two completely controllable and completely observable realizations of S are algebraically equivalent. (iii) A realization of S is irreducible if and only if at all times it consists of Part (B) alone; thus every irreducible realization of S is completely controllable and completely observable. Proof. The first statement can be read off by inspection from Fig. 4. The second statement is proved in [14]. The necessity of the third statement follows from Theorem 5, while the sufficiency is implied by (ii). It is clear that Theorem 2 is a consequence of Theorems 5-7. We can now answer the question posed in section 4 in a definite way: THEOREM 8 (Main Result). Knowledge of the impulse-response matrix S(t, r) identies the completely controllable and completely observable part, and this part alone, of the dynamical system which generated it. This part ("B" in Theorem 5) is itself a dynamical system and has the smallest dimension among all realizations of S. Moreover, this part is identified by S uniquely up to algebraic equivalence. Using different words, we may say that an impulse-response matrix is .faithful representation of a dynamical system (2.1-2) if and only if the latter is completely controllable and completely observable. Remarlc. It is very interesting to compare this result with Theorem 4 of E. F. Moore, in one of the early papers on finite automata [26]: "The class of all machines which are indistinguishable from a given strongly connected machine S by any single experiment has a unique (up to isomorphism) member with a minimal number of states. This unique machine, called the reduced form of S, is strongly connected and has the property that any two of its states are distinguishable." "Indistinguishable machines" in Moores terminology correspond in ours to alternate realizations of the same input/output relation. "Strongly con-

"

1.70

a.E. KALMAN

neeted" in his terminology means completely controllable in ours. "Indistinguishable states" in our terminology corresponds to states whose difference, not zero, is an unobservable state in the sense of [3]. Evidently the two theorems are concerned with the same abstract facts, each being stated in a different mathematical framework.
6. Explicit criteria for complete controllability and observability. The canonical structure theorem is so far merely an abstract result, since we have not yet given a constructive procedure for obtaining the co6rdinate transformation which exhibits the system matrices in canonieM form. We shall do this in section 7. The meth()d rests on the possibility of finding explicit criteria for complete controllability and complete observability. The following lemmas, proved in [4], play a central role: rank W(to, tl) jor I1 > lo ,sufficiently large, LEMMA ]. na(t0) -t- n,(to) where

(6.1)
or

W(t,), h.)

dW/dto F(to)W + WF(to) G(to)G(to), W(t) O. (6.2) rank M (to, t_.) for t_ < to suciently LEMMA 2. n (t0) nD (t0) small, where
(6.3)
or

M(to t-l)

*
fro (,

(6.4) -dM/dto

F(to)M

For constant systems, the preceding lemmas can be considerably improved [4]: LEMMA 3. For a constant system,

t0) H (r)H(r)(r, to) dr

MF(to)

H(to)H(to), M(t_.)

O.

(6.5) (6.6)

n -t- n

rank [G, FG,

F-G].

LMM 4. For
nc

a constant system,

-t-

n,

rank

[H, FH,..., (F)-IH].


3 8 18 6
o

EXAMPLE 4. For F and G defined in Example 2, the matrix (6.5) is

(6.7)

3 --2 0 0

3 --1
o

3 6 12 4

2 12 4

3 6 24: 8

3 14 36 12

22

20

LINEAR DYNAMICAL SYSTEMS

171

The rank of this matrix is 2, which checks with the fact that n 1 and 1 in Example 2. n. The determination of the rank of (6.7), while eleinentary, is laborious. For practical purposes it might be better to compute W; for instance, by solving the differential equation (6.2). In the constant case, there is another criterion of complete controllability which is particularly useful in theoretical investigations. The most general form )f this theorem (which may be found in [14]) is complicated; we state here a simplified version which is adequate for the present purposes: LEMMA 5. Hypothesis: The matrix F is similar to a diagonal matrix. In Tx with other words, there is a nonsingular coordinate transformation the property that in the new coordinate system F has the form

? TFTwhere

I
i--=1

}r Iq,

is a q

><

q. unit matrix,

and the matrix G has the

form

0 ()

q rows

5= TG=

L (r)J

rOWS.

Conclusion" The system is completely controllable

and only

if

rank () rank ((r) q, q" (6.8) We leave it to the reader to dualize this result to complete observability. 1 of Lemma 5. EXAMPLE 5. Consider the special case q q is satisfied, every The eigenvalues of F are then distinct. If condition (6.8)
element of the one-column matrix is nonzero; by a trivial transformation, Thus we all of these elements can be made equal to 1, without affecting have the representation: can choose a coordinate system in which F, G

(6.9)

17
0

(X

hi

j), d

1.72

R.E. KALMAN

This is the canonical form of Lure [18]. It is closely related to the partialfraction expansion of transfer functions. To illustrate this, consider the 1 transfer-function matrix" 1

s+2
s+3
This transfer function is realized by the system"
0 --3 0

s+4"

(6.10)

0 0

0 --4

(6.11)

[11
H

01

[---] which is in the canonical form of Lure. By Lemma 5, (6.10-11) is completely controllable; by the dual of Lemma 5, (6.1.0-12) is completely observable. We can double-check these facts by means of Lemmas 3-4. For (6.9) the matrix (6.5) is

(6.12)

(6.13)

where the he are the diagonal elements (= eigenvalues) of F in (6.9). But the determinant of (6.13) is the well-known Vandermonde determinant. The latter is nonzero if and only if all the Xi are distinct, which is what we have assumed.
7. Computation of the canonical structure. We show now how to determine explicitly the change of coSrdinates which reduces F, G, H to the canonical form. We consider only the constant case of (2.1-2). The computations are elementary; it is not necessary to diagonalize the matrix F or even to determine its eigenvalues. The procedure is as follows: W(0, 1)* given by (a) We compute the controllability matrix W
* It can be shown [4, Theoren 10] that in the constant case one may choose any

> to in Lemna 1.

LINEAR DYNAMICAL SYSTEMS

173

a nonsingular matrix

(6.1);for instance, by solving the differential equation (6.2). Then we find T such that

mtrices of pproprite size. Clemqy n n.4 -t- n is the number of controllble state vribles. The matrix T defines the change of coSrdintes

(7.2)

T;

in terms of the new coSrdinates, the system matric are

(7.3)
(7.4)
2

T-FT,
,F

(
0

T-1G,

I
()

HT,
and

E.
tt

/?J

Ill I?].
n, i.e., when

This decomposition is trivial (and therefore omitted) if n the system is completely controllable. (b) Next we consider the two subsystems defined by

(7.5)

1, (.,
I?, 0,

and
and

;
B

We compute the observability matrices (0, 1) and (0, 1 given by (6.3) for both of these subsystems. Then we determine two nonsingular matrices

1, 02

such that

(7.6)

(1)]101 ---1---

InB

0]
I,,,

(7.7)

02),/0.2 2

/?2

0]
n or

These results define another change of coSrdinates


0

One or the other of these trmsformtions is superfluous if n


nd

hi.,

After the coSrdinate changes (7.2) and (7.8), we obtain the following

174
matrices

n.E. KALMAN

F Ac
B

F Ad

(7.9)
xa

-IF,?

FBB
0

Fda

/.

llV

[0

H"

Ha],

Clearly, n, is the number of state variables which are both controllable and observable. But, in general, na nD and nc > n(. (c) It remains to transform the element /7"" into 0, if this is not already the case. (If /7"c n, and (7.9) has the desired 0, then nc nc, na canonical structure.) We consider the subsystem

F BB

(7..0)

d
0

t*

The corresponding observability matrix given by (6.3) is

21*(0, 1)

*
*

(Q

nonnegative definite.)

The upper lefg elemeng of is I, in view of (7.9); all we know aboug he oher elements is their symmetry properties.) Lein.g

we find that

?*)YI*?*
where R

InB
0

AA

is a symmetric, nonnegative-definite matrix.

LINEAR DYNAMICAL SYSTEMS

175

Now let if** be a nonsingular matrix such that

(?**),**?**
where n rank R. Let lY**. Since relative to the partitioning in (7.10), so is upper triangular form

upper * and lY**will take /Y*triangular into the which


are

Fee
where nc n ne. But these transformations decompose /* into a completely observable and an unobservable part. Hence Fee F ee O.

Moreover,

*=[H"
F, G, and H into the
X

0]=

IN"

H1

THEOREM 9. The explicit transformation which talces the constant matrices canonical form required by Theorem (5-iii) is given by ---) ?--1_--1--1X. We partition

F.C= [FAC
and partition

Then we define

n)

+n

and find

i,,= [F

F"= [Fee
F (:"

[F c

F], F], F],


Fdd

Hv= [H
8. Construction of irreducible realizations. Now we give an expliei procedure for the construction of an irreducible realization of a weighging-funetion matrix W( r). In view of Theorem 7,

part (iii), we can do this in two stages: (i) We construct a realization of W, then (II-A) we prove, using Lemmas 1-5, that the resultant system is completely controllable and completely observable, hence irreducible; or (II-B) we carry out explicitly the canonical decomposition and remove all parts other than (B). Instead of the weighting-function matrix W, it is usually more convenient to deal with its Laplace transform Z. Let us consider the problem with Method A in order of increasing difficulty. 1. This is equivalent to the problem of simulating a Case 1. m p single transfer function on an analog computer. There are several wellknown solutions. They may be found in textbooks on classical servomechanism theory or analog computation. Without loss of generality (see Theorem 4) we may consider transfer functions of the form

(8,1)

ZlI(8)

a, s
s"

A-

A- al

b, s "-

+b

N(s) D(s)

where the am,", a, b,..., bl are real nmnbers. Of course, at least one of the a+ must be different from zero. We assume also that the numerator N(s) and denominator D(s) of zn(s) have no common roots. There are two basic realizations of (8.1). See Figs. 5-6, where the standard signal-flow-graph notation [1.6] is used. In either case, one verifies almost by inspection that the transfer functions relating y to u are indeed given by Zn. In Fig. 5, the system matrices are
0 0

1 0
0

0 1
0

0 0 0

0 0
1

(s.2)

F
0

bl

b2

b:
0 0

b-

(8.3)

(8.4)

[a

ax

a,_

a,].

LINEAR DYNAMICAL SYSTEMS

177

Ul

Xn

bl
FIGURE 5.

In Fig. 6, the systen matrices arc


0 1
0 0 0 0

0 0

(s.5)

0. 1. 0.
0

0
1

--bll 0. --b..
--b2
1
--b,

j,

(8.6)

I]
0

and

(8.7)

t0

].

It is very easy to check by means of (6.5) and (6.6) that the system (8.2, 3) is completely controllable and (8.5, 7) is completely observable. However, if we attempt to check the controllability of (8.5, 6) by means of (6.5) we get a matrix whose elements are complicated products of the coefficients of N(s) and D(s). To prove that the determinant of this matrix does not vanish, we have only one fact at our disposal" the assumption that N(s) and D(s) have no common roots. Guided by this observation, we find
that the following is true" LEMMA 7. Suppose F has the form (8.5) and G has the form (8.6). Then (i) we have the relation

(s.s)

K(F, G)

[G FG

I,"-GJ

N(F),

178

n.E. KALMAN

-b

and (ii) the polynomials N(s) and D(s) have no root in common if and only if O. det K(F, G) The main fact to be proved is (ii), for then the complete controllability of (8.5, 6) follows by Lemma 3. A straightforward way of establishing (ii) is to transform the standard Euler-Sylvester determinantal criterion [19, p. 84] for the nonexistence of common roots of N(s) and D(s) (the so-called resolvent of N(s) and D (s)) into the form (8.8). This can be easily done, but the details arc not very transparent. Therefore we prefer to give another n, be the set of n-vectors in which the j-th Proof. Let el, i 1, Since F is given by (8.5), we see that e+l Fe, component of ei is 1 =< n en] I. Hence K(F, e) Iv1, e2, 1, and K(F, el) K (F, F-el) F-IK (F, el) F when I =< i _-< n. Then (8.8) follows by linearity. n denote the eigenvalues (not necessarily distinct) Let [A], i 1, of a square matrix A. Then

a..

get K(F,

G)

II

i1

i=l

II

where the second equality follows from (8.8) by a well-known identity in 0 for some 0 if and only if N(Xi[F]) matrix theory. Thus det K(F, G) i; that is, when an eigenvalue of F is a root of N(X). Since the eigenvalues of F are roots of D (h), this proves (ii).* It is interesting that (8.8) provides a new representation for the resolvent, which is preferable in some respects to the Euler-Sylvester determinant. The latter is a 2n X 2n determinant, whereas det K (F, G) is n X n. The complete observability of (8.2, 4) is proved similarly. The systems given by (8.2-4) and (8.5-7) are duals of one another in
The present proof of Lemma 6 was suggested by Drs. John C. Stuelpnagel and W. M. Wonham of RIAS.

LINEAR DYNAMICAL SYSTEMS

179

the sense defined by (5.3). Fig. 6 is a reflection of Fig. 5 about the vertical axis, with all arrows reversed. A third type of realization in common use is obtained from the partialfraction expansion of z11(s) (see Example 5). Note, however, that this requires factorization of the denominator of zl.(s), whereas the preceding realizations can be written down by inspection, using only the coefficients of z.(s). These considerations may be summarized as the following result, which is a highly useful fact in control theory" THEOREM 10. Consider a linear constant dynamical system with m p 1, which is completely controllable and completely observable. Then one may always choose a basis in the state space so that F, G, H have the form (8.2-4) or (with respect to a different basis) (8.5-7). Proof. Let (8.1) be the transfer-function matrix of the given dynamical system. By Theorem 8, the given system is an irreducible realization of (8.1). So are the systems specified by (8.2-4) and (8.5-7). By Theorem (7-ii), all three systems are algebraically equivalent and by constancy (Theorem 3) they are even strictly equivalent. Extensions of this theorem may be found in [14]. For an interesting application to the construction of Lyapunov functions, see [25]. The procedure described here may be generalized to the non-constant case. Assuming the factorization (4.3) of S(t, r) is known (with m p 1), Batkov [20] shows how to determine the coefficients of the differential equation

(8.9)

dny/dtn

Laning and Battin [21, p. 191-2] show how one converts (8.9) into a system of first-order differential equations (2.1) with variable coefficients. We shall leave to the reader the proof of the irreducibility of the realization so obtained. Case 2-a. m 1, p > 1. We have a single-input/multi-output system. We can realize Z(s), without factoring the denominators of its transfer functions, by the following generalization of the procedure given by Fig. 5 and (8.2-4). First, we find the smallest common denominator of the elements of Z(s). (This can be done, of course, without factorization.) Z (s) assumes the form

- - bn
-1

y/a

b(t) y

an(t)dn-lul/dt -1

+ al(t)u.

z(s)
Then the following dynamical system provides an irreducible realization

]_80

R.E. KALMAN

of Z (s): F and G are as in (8.2-3), while H given by (8.4) is generalized to


-all

aln

Lapl

Complete controllability is trivial; complete observability is established by a straightforward generalization of Lemma 6. In this case we form p linear functions of the state, rather than merely one s in Fig. 5. 1. We can realize this multi-input/single-output Case 2-b. m > 1, p system nalogously to Cse 2- by generMizing the procedure given by Fig. 6 nd (8.5-8.7). Let us write the elements of Z(s) in terms of their smallest common denominator"

as + + aim b" + i,, s ,+ Then the desired irreducible realization consists of F ad H as defined by (8.5-6), while

z(s)

+ + b,s- +
anl S
n--I

a..

G__
[..anl

This case is the dual of Case 2-a. Even in Case 2, it is impractical to give a general formula which expresses the coefficients of F, G, and H in terms of the coeificients of the transfer functions in Z(s) if the denominators are not all the same. When we pass to the general case, determination of F, G, and H often requires extensive numerical computation. Case 3. m, p arbitrary. Here Method (A) is very complicated if any transfer function in Z(s) has multiple poles [1.4]. In most practical applications, however, such complications are of no interest,. Ruling them out, E. G. Gilbert gave an elegant and relatively simple solution [5]. Let sl, Sq be distinct complex numbers corresponding to the poles of all the elements of Z(s). Assume that all poles are simple. Then

R(])

lira (s

s)Z(s),

.1,

..., q

is the k-th residue matrix of Z(s). If se =Sk, then R(se) (s), where the bar denotes ghe eomplex eonjugage. In erms of ghe residue magriees, ghe weighting-funegion magrix W() corresponding go Z(s) has ghe explieig form

w(t)

-[z()]

().

We have then" THEOREM ]1. (Gilbert). Hypotheses: No element of the transfer-function

LINEAR DYNAMICAL SYSTEMS

18|.

matrix Z(s) has multiple poles. Z (s) has a total of q distinct poles 81. with corresponding residue matrices R (1), R q Conclusions: The dimension of irreducible realizations of Z s is
q

(8.11)
(ii) Write (8.12)

n
k=l

r where r

rank R(]).

lc H(k)G(lc), 1,..., q, where H(lc) is a p X r matrix and G(]c) is an r X m matrix, both of ranlc rl Then Z s has the irreducible realization
81

R(k)

Irk

(8.13)

F
0

(It
sq Irq

X r unit matrix),

(s.,)
and

IG(I) 1( L(;:q)
H
IN(l)

H(q)]. Proof. This is one of the main results in [5]. With the aid of machinery developed here, we can give a shorter (though more abstract) demonstration. The factorization (8.1.2) is well known in linear algebra. We give in the

(8.15)

Appendix various explicit formulae (which are easily machine-computable) for G(lc) and H(/c). Applying Lemma 5 shows that the dynamical system defined by (8.13--15) is completely cotrollable and completely observable. Hence it is irreducible, which implies formula (8.11). By elementary changes of variables, (8.13-15) can be transformed into matrices which have only real elements. A serious disadvantage of Method (A), as expressed by Theorem 11, is that the denominators of the transfer functions in Z(s) must be factored in order to determine the poles. This is not easily done numerically. Moreover, the residue matrices R(/c) corresponding to complex poles are complex, which makes the factorization (8.11) more complicated (see Appendix). Now we turn to Method (B). This method does not require computation of eigenvalues, and it is not bothered by multiple poles. This is a decided advantage itx numerical calculations. On the other had, the method is not convenient for simple illustrative examples. Nor is it possible to display the elements of F, G, and H as simple functions of the coefficients in z(s).

182

n.

]03.

KALMAN

An easy way of realizing Z(s) (without guaranteeing irreducibility) is the following. Let ci be the number of distinct poles (counting each pole with its maximum multiplicity) in the i-th row of Z(s), and let i be the number of poles in the i-th column. Then the maximum number no of state variables required to realize Z(s) by repeatedly using the scheme given under Case 2-a or 2-b is
no= min

o,

As before, we can determine the a and Bi without factoring the transfer functions of Z(s). There is in general no simple way in this method to determine the dimension n N n0 of irreducible realizations without performing the computations outlined in Section 7. The two methods are best compared via an example. This example must be of fairly high order, since we wish to provide accurate numerical checks. EXAMPLE 6. Consider the transfer-function matrix 3(s+3)(s+5) 6(s+l) 2s+5 2s+7
(s+l)(sW2)(s+4)

(sW2)(s+4)
1

(sW3)(s+4)

(sW2)(s+3)

Z(s)=

(s+3)(s+5)
2(s+ 7s + 18) (s+l)(s+3)(s+5)

(s+3)
2s

8(s+2) 2(s--5) (sW1)(s+2)(s+3) (sW1)(sW3)(s+5)


1

(s+l)(s+3)

s+3)

2 (Ss+ 27s + 34) (s+l)(s+3)(s+5)

Applying Method (A) first, we find that the residue matrices are:

R(1)

0 1

4 0
0 0

1 3
--6 0

3.

R(2)

R(a)

R(4)

-1 1
-0.g

0 0

0 0

2.

2 1

2 1

ra

2.

0 0

9 0 0

1 0 0

0 0 0

1.

R()
Thus
9.

-1 2

0 0

0 0

r= 1.

LINEAR DYNAMICAL SYSTEMS

1.83

Employing the procedure given in the Appendix, we find the following factors for matrices R(I) (the products are accurate up to four places beyond the decimal point)"

H()

8.0000 |o.oooo
L3.0000

.
a

0.0000

0.0000]
o.oooo|, 3.0774_]

0.7276
1

[ 1.0000
|0.0000

ko.oooo
,.ooo
H(2)
0.0000 0.0000

0.0000 0.0000 0.3249

0.0000 0.9701 -0.2294

0.0 0 J
0.2425 0.9175

o.oooo
.0000

O.O000_J

G(2)
H(3)
1.3416 3.1305 0.0000

1-0"8182
0.0000

-0.5455 0.0000

0.0000 -1.0000

0.18181.
0.0000

0.4472 --0.4472 4.4721

]
-0.6708
0.2236 -0.6708

G(3)
H(4)

0.6708
0.2236

0.6708] 0.22363;

F9.06927 |o.oooo l,
LO.OOOOj
0.0000

a(4)

[--0.0,551. 0.9924 0.1103 0.0000];

g(a)

-a.a
6.3246

a(,)

[o.a.

o.oooo o.oooo

o.7].

Using these numerical results, we find that the dynamical equations of the irreducible realization are given by
1 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0

F=

0 0 0 2 0 0 0 0 0

0 0 0 0 2 0 0 0 0

0 0 0 0 0 3 0 0 0

0 0 0 0 0 0 3 0 0

0 0 0 0 0 0 0

0-]
0 0 0 0 0 0/

4 0

O|
5

184 1.0000 0.0000 0.0000 --0.8182 0.0000 0.2236 --0.6708 --0.0551. 0.3162

R.E. KALMAN

4.4721_ 0.0000 6.3246 _J 3.0000 Now we apply Method (It). Virst of all we note that al a 4, aa 3, while # 5, a 4 (see p. 181). Hence it is best to choose for the preliminary realization three structures of the type discussed under Case 2-b. This will require no 11 dimensions. p(a + a + a) the least common denominator of the rows of Z(s). See Next, we find

0.0000 0.0000 0.3249 --0.5455 0.0000 0.2236 0.6708 0.9924 0.0000

0.0000 0.9701 --0.2294 0.0000 1.0000 0.6708 0.1103 0.0000

0.0000 0.2425 0.9175 0.1818 0.0000 0.6708 0.2236 0.0000 0.9487

=/o.oooo

[-8.oooo

0.0000 0.0000 5.5000 0.0000 1.3416 4.1231 0.0000 0.0000 6.0000 3.1305 0.7276 3.0774 0.0000 0.0000 0.0000

0.4472 -o.4472

o.oooo
o.oooo

-I

.,

Fig. 7.

2(s

2)

9s

25s

15

F(n3IE 7.

The desired realization 7, using (8.5) and (8.6)" 0 0 0 0 0 0 0 0 0 1.

of Z(s) can be read off by inspection from Fig.


-24
-35 -10 0
0

0 0 0 1 0 0 0 1 0 0 0

-30 -61 -41. -11


0 0 1 0 0 1

--15 --23 --9

LINEAR DYNAMICAL SYSTEMS

185

135

117
33 3

18 42 3.0 6 10

14 25 13 2

20 33 15 2
32 32 8 0

4
6 2 0

17
8 1 0 --10 --2

50 20 2 0

36 14 2 0 0

5 68 6 54
1

10

0 0 0 0 0 0 0

H=

0 0 0 0 0 0

0 0 0

0 0 0 0 0 0 0 0 0

By virtue of its construction, this system is completely observable but


we cannot tell by inspection whether or not it is completely controllable. (From the results obtained above with Method (A), we know that the system is not completely controllable since 11 no > n 9.) Therefore the

canonical decomposition may contain Parts (B) and (D). To see what the dimensions of these parts are, we compute numerically the decomposition of the system into completely controllable and uncontrollable parts according to the method described in Section 8. These culculations involve only the matrices F and G, but the resulting transformations must be applied also to the matrix H.
-0. 331,6

0.1182

0.01[[)

-0.0299

0.0097

-0.0001

-0.0663 -0.0113
O. 07
O. 0120

0.0000
O.O001

0.0000

0.83]9

0.2455 -0.2029 -0.0119

0.0268 -0. 0101


1.0p59

0.0000

o. 0oo0

o. 7189

-0.8333
-0.2943

-0. 890

-O.102J
O. 0361

-0.2290

-0.0009

-0.9998
O. 1773

0.027 -0.0120

O. 2032

O. 0022

-0.0610
-0. 4287

-0.0044 -0.0279

-0.0969
0.0024
O. 97

O. 0194 O. 0089

2.9469 i. 6479
i. 1882

0.8726
2.9962
i0

-0.8896
0.2477

0.8321
i. <)97

0.188 2.0439
o.ooo9
o.1114

O. 4999
-0. 5777

1.1199

-0.9689 -0.4969 -0.4046


o.059

0.0199
-o.oo+/-o

-462.2221

-7 3.0068

-o, z98

o.o49

-o.o32+/-

0.0016 -o. 292 -o.o92

o. oooo
-7].0.9771
0.","2

o. 4698
-2.19i

i.694 -o.o29o

1.649 -o.4239 -0.0196


2.46o4 -1.98o -o.28Ol
0.0000 0.0000
O. 0000

0.4o97 -o.o48o -o.oo4o -29.877

-o.2787 -0.6894
0.0000

2.46o4.
0.0000

o.8992
0.0000

o.8863 -0.2649
0.0000
O. 0000

-32.0706
0.0003
O. 299

0.0000

0.0000

0.0000
O. 0000

O. 0000

O. 0000

O. 0000

O. 0000

O. 0000

O. 0000

O. 0401

]il G U RE 8.

186
i. 2011
i. 6822

R.E. KALMAN

-1.7962 -0.1683 -0.0665


G

1.0412

-0.1117

3 -2.8589 -2.876 -1.7637


0.0000 0.0000 0.0000 0.0000

-0.7886 -0.3459 -0. 320 O. 6039 O. 394 O. 474 0.8640 2.387 -1.7927 -0.2420 2.2981 O. 40-) 0.9596 1.4137 2.7344 ]-.2734 m.800 -2.9 -0. 46 -0.0977 i.
.9
-0.

x i0

0.9281 -2.5137
0.0000 0.0000 0.0000

0.0000

-0.2928
0599

0.4076
0.2173

0.0160 -0.053

0.0141

0.0000
0.0022

-0.0981 -0.0167
-O.08i)

-0.0001

0.0000
1.0000 0.0000

i. 0000

0.0023 -0.0116

-0.0169

0.0373

-0.0033 -0.0007

0.0000

0.0161

0.0394 -0.0603 -0.0039

0.6736

0.0178

0.0024

0.000_

-I

FGUE 9.

The final results may be seen in Figs. 8-9, which give the matrices F, G, and/. Elements in the lower left-hand corner of # should be exactly zero. In fact, they are zero to at least the number of digits indicated in Fig. 8. To check the accuracy of these two irreducible realizations of the transfer function matrix on p. 181, we have computed the corresponding weightingW (2) (t) function matrices W (1) (t) and W (2) (t). The equality W (1) (t)
was found to be correct to at least four significant digits.

9. Other applications to system theory. The literature of system theory contains many instances of errors, incomplete or misleading solutions of problems, etc., which can be traced to a lack of understanding of the issues discussed in this paper. This section presents some cases of this known to the writer; other examples may be found in the pper of Gilbert [5]. Analog computers. According to Theorem 8, a linear dynamical system (2.1-2) is a "faithful" realization of an impulse-response matrix if and only if it is irreducible. Suppose the dynamical equations (2.1-2) are programmed on an analog computer. (See [8].) Then it is clear from Theorem 8 that the computer will simulate the impulse-response matrix correctly if and only if a minimal number of integrators are used. Otherwise the system programmed on the analog computer will have, besides Part (B), t least one of the Parts (A), (C), or (D). Since the impulse-response matrix determines Part (B), and that alone, the nature of the redundant parts will depend not on the impulse-response matrix but on the particular method used to ob-

LINEAR DYNAMICAL SYSTEMS

1_87

rain the dynamical equations. It should be borne in mind that the canonical decomposition is an abstract thing; usually it is not possible to identify the redundant integrators without a change of variables. The writer is not aware of any book or paper on analog computation where this is explicitly pointed out. But the facts of life seem to be well known (intuitively) to practitioners of the analog art. That redundancy in the number of integrators used can cause positive harm is quite clear from the canonical structure theorem. EXAMPLE 7. Let the simulated system consist of Parts (A) and (B) and suppose that Part (A) is unstable. Because of noise in the computer, Part (A) will be subject to perturbations; they will be magnified more and more, because of the instability. As long as assumptions of linearity hold exactly, the unstable (A) component of the state vector will not be noticed, but soon the computer will cease to function because its linear range will be exceeded. Lre canonical form. In his book on the Lure problem, Letov implies [18; equation (2.4) and (2.23)] that every vector system

(9.1)
can be reduced

dx/dt

Fx

+ g.

scalar)

o the canonical form

i n 1, dxi/dt Xx Jr- r, whenever the eigenvalues X of F are distinct. Since (9.2) is completely controllable, this assertion, if true, would imply that (9.1) is also completely controllable, which is false. In fact, the system defined by

(9.2)

is obviously not equivalent to

g. whenever X In examining the derivation originally given by Lure for his canonical form [27; Chapter 1, 2-3], it is clear that the last step before equation (3.5) 0 (in the notation of Lure [27].) It is is valid if and only if det [H(X)] easy to show that this condition is equivalent to complete controllability, whenever the eigenvalues of F are distinct.. Unfortunately, the condition det [H(X)] 0 was not emphasized explicitly by Lure [28] in the original publications. We may thus conclude that when F has distinct eigenvalues and there is a

188

R.E. KALMAN

single control variable, the Lure-Letov canonical form exists if and only if the pair IF, g is completely controllable. It is interesting to note that (9.3) can be transformed into (9.4) when }, ; in other words, when the eigenvalues are not distinct the Lure canonical form may exist even if the system is not completely controllable. Cancellations in the transfer-function. When a mathematical model is derived from physical principles, the equations of the system are in or near the form (2.1-2). Regrettably, it has become widespread practice in system engineering to dispense with differential equations and to replace them by transfer functions Z(s). Later, Z(s) must be converted back into the form (2.1-2) for purposes of analog computation. In the process of algebraic manipulations, some transfer functions may have (exactly or very nearly) common factors in the numerator and denominator, which are then canceled. This is an indication that a part of the dynamics of the system is not represented by the transfer function. Such cancellations are the basic idea of some elementary design methods in control theory. These methods do not bring the system under better control but merely "decouple" some of the undesirable dynamics. But then the closed-loop transfer function is no longer a faithful representation of the (closed-loop) dynamics. Stability difficulties may arise. Similar criticisms may be leveled against the large, but superficial, literature on "noninteracting" control system design. EXAMPLE 8. Consider the system defined by the matrices

(.%)

F=

,
--2

0 0

a=
2
ul

H=[-2
0.5

0].

The transfer function relating y to

is the sum of two terms"

y(s) (.)

-: x(s) + x:(,)
2
s-2s- 5s-- 6

(,.()

+s

(s- 2) (s -t- 1)(s 2)(s -t- 3)

Thus, by cancellation, the transfer function is reduced from the third to


the second order. The system has an unstable "natural mode" (corresponding to sa 2) about which the transfer functions gives no information. -Using (6.5) we see that the system (9.5) is completely controllable. By Theorem 5, the system cannot be completely observable: n 2 from (9.6) and Case 1, section 8. The canonical structure consists of Parts (A)

(s

2s

5s-- 6

)(s

+ 3)"

LINEAR DYNAMICAL SYSTEMS

189

and (B). In canonical co6rdinates the system matrices can be taken as


-1 0 0

0 0 -3 0 0 2

[0.5

-0.5 0].

We can easily calculate the change of coSrdinates 4---- Tx by the method of partial fractions discussed in [8]. First we find T-1, then T.
The results are
--9 6

Lo

4 -1

--1 4

2 2

Loss oj" controllability and observability due to sampling. Consider a singleinput/single-output constant linear system. Suppose the output is observed kT (It only at the instants integer, T > 0), and that the input is constant over the intervals kT =< (It q- 1 )T. This situation is commonly called "sampling"; it arises when a digital computer is used in control or data processing. T is the sampling period. We can regard such a setup as a discrete-time dynamical system. We define here 0 (Axiom (D1)) as the set of integers and replace (2.1) by a difference equation. All theorems carry over to this situation with small modifications. The analysis of discrete-time systems by conventional techniques requires the computation of the so-called z-transform of Z(s) [22]. The analysis using z-transforms then proceeds in close analogy with analysis based on Laplace transforms. A constant linear system which is completely controllable and completely observable will retain these properties even after the introduction of sampling if and only if [4!

(9.7)

Res

Rest.

implies Im (s

s)

qr/T

n and q where i, j positive integer. 1, If this condition is violated (the sampling process "resonates" with the system dynamics) then cancellations will take place in the z-transform. The z-transform will then no longer afford a faithful representation of the system, so that if (9.7) is violated, results based on formal manipulation of z-transforms may be invalid. This point is not at all clear in the literature. True, Barker [23] has drawn attention to a related phenomenon and called it "hidden oscillation." The textbooks, however, dismiss the problem without providing real insight

[22, 5-3; 24, 2.13].

190

..

KALMN

A practical difficulty arises from the fact that near the "resonance" point given by (9.7) it is hard to identify the dynamical equations accurately from the z-transform. Small numerical errors in the computation. of the z-transform may have a large effect on the parameters of the dynamical equations.
REFERENCES
[1]
[2]

[3] [4]
[5]

R. E. KALMAN, Discussion of paper by I. Fligge-Lotz, Proc. 1st International Conference on Automatic Control, Moscow, 1960; Butterworths, London, 1961, Vol. 1, pp. 396-7. I. E. KALMAN, Canonical structure of linear dynamical systems, Proc. Nat. Acad. Sci. USA, 48 (1962), pp. 596-600. 1. E. KALMAN, On the general theory of control systems, Proc. 1st International Congress on Automatic Control, Moscow, 1960; Butterworths, London, 1961, Vol. 1, pp. 481.-492. R. E. KALMAN, Y. C. Ho, AND K. S. NARENDRA, Controllability of linear dynamical systems, (to appear in Contributions to Differential Equations, Vol. 1, John Wiley, New York.) E. G. GILBERT, Controllability and observability in multivariable control systems, J. Soc. Indust. Appl. Math. Ser. A: On Control, Vol. 1, No. 2 (1963),
pp. 128-151.

[6] V. V. NEMITSKII AND V. V. STEPANOV, Qualitative Theory Of Differential Equations, Princeton Univ. Press, Princeton, 1960. [7] R. E. KhLMAN AND J. E. BERTRhM, Control system analysis and design via the second method of Lyapunov, J. Basic Engr. (Trans. A.S.M.E.), 82 D (1960), pp. 371-393. [8] R. E. KnLMnN, Analysis and design principles of second and higher-order saturating servomechanisms, Trans. Amer. Inst. Elect. Engrs., 74, II (1955), pp.
294-310.

[9] W. HAHN, Theorie und Anwendung der direkten Methode yon Ljapunov, Springer, Berlin, 1959. [10] J. P. LASALLE AND S. LEFSCHETZ, Stability By Lyapunovs Direct Method, Academic Press, New York, 1961. [11] L. MARKUS, Continuous matrices and the stability of differential systems, Math. Z., 62 (1955), pp. 310-319. [12] L. A. ZADEH, A general theory of linear signal transmission systems, J. Franklin Inst., 253 (1952), pp. 293-312. [13] D. MIDDLETON, An Introduction To Statistical Communication Theory, McGrawHill, New York, 1960. [14] R. E. KALMAN, On controllability, observability, and identifiability of linear dynamical systems, (to appear). [15] ]7. E. KALMAN, On the stability of time-varying linear systems, Trans. I.R.E. Prof. Gr. Circuit Theory, (CT-9 (1962), pp. 420-422.). [16] S. J. MASON, Feedback theory: some properties of signal flow graphs, Proc. I.R.E., 41 (1953), pp. 1144-56; Further properties of signal flow graphs, ibid., 44 (1956), pp. 920-926. [17] R. E. KALMAN, New results in filtering and prediction theory, RIAS Report 61-1, Research Institute for Advanced Studies (RIAS), Baltimore, 1961.. [18] A. M. LETOV, Stability In Nonlinear Control Systems, Princeton Univ. Press, Princeton, 1961.

LINEAR DYNAMICAL SYSTEMS

191_

[19] B. L.

VAN DER

WAERDEN, Modern Algebra, Vol. 1, 2nd Ed., Ungar, New York,

1949.

[20] A. M. BATKOV, On the problem of synthesis of linear dynamic systems with two parameters, Avtomat. Telemeh., 19 (1958), pp. 49-54. [21] J. H. LANING, JR. AND R. H. BATTIN, Random Processes In Automatic Control, McGraw-Hill, New York, 1956. [22] J. R. RAGAZZN AND G. F. FRANKN, Sampled-Data Control Systems, McGraw.Hill, New York, 1958. [23] R. H. BARKER, The pulse transfer function and its application to sampling servo systems, Proc. Inst. Elec. Engrs. 99 IV (1952), pp. 302-317. [24] E. I. JURY, Sampled-Data Control Systems, John Wiley, New York, 1957. [25] R. E. KALMAN, Lyapunov functions for the problem of Lure in automatic control, Proc. Nat. Acad. Sci. USA, 49, (1963), pp. 201-205. [26] E. F. MOORE, Gedanlcen-experiments on sequential machines, Automata Studies. Princeton Univ. Press, Princeton, 1956. [27] A. I. LuRE, Certain Nonlinear Problems in the Theory of Automatic Control. (in Russian), Gostekhizdat, Moscow, 1951; German translation AkademieVerl,g, Berlin, 1957.

[28] Private communication, Academician A. I. Lure.

APPENDIX
Factorization of rectangular matrices. Given an arbitrary, rel, p X m rain (m, p). We wish to find a p X q mtrix H nd matrix R of rank q HG. The existence of a q X m mtrix G, both of rank q, such that R H and G follows lmost immediately from the definition of rank. We describe below constructive procedure for determining H and G numerically from numerical values of R. Let p m. Form the p X p matrix S RR. As is well known, there exists nonsingulr matrix T such that

<=

=<

(A-l)

TRRT

TST

E,

where precisely q diagonal elements of E re 1, 11 other elements re 0. T cn be clculated by steps similar to the gussin elimination procedure. Compute the generalized iaverse R (in the sense of Penrose [4]) of R. R is n m X p mtrix. Using the properties of R ([4]) we obtain

T-]T-IR (T-E) (T-1E)R . SR (A-2) R RRR RRR Now T-1E is a matrix which contains precisely p q zero columns. DeH. Similarly, leting these columns, we obtain p X q matrix (T-I) (RT-1E) we obtain a m X q deleting p q zero rows from (T-1E)R matrix G (RT-E) Evidently R HG. Since the ranks of H nd G
are obviously less than or equal to q, both ranks must be exactly q for otherwise rank R q, contrary to hypothesis. Alternately, let T, U be nonsingular matrices such that

TRU

E;

1.92

.
R

E. KALMAN

fhen

(A-3)

is the desired decomposition. However, the computation of (A-3) may require more steps than that of (A-2). R/ RR* A iB is Suppose now that R is complex. Then S complex hermitian; it corresponds to the 2n )< 2n nonnegative matrix

(A-4)
whereA

z*RR*z

A and B -B.Infaet, ifz z + iy, the hermitian form (which is real-valued) is equal to the quadratic form

=I-BA AB1

TST*

As is well known, there exists a nonsingular complez matrix T such that (7 -5- iV, i-t, follows further tha E. If 7 0 U A B U

Hence the determination of the complex n X n matrix has been reduced to the determination of a real 2p X 2p matrix. Similar remarks apply to the calculation of R Thus the problem of factoring complex p m matriees can be embedded in the problem of factoring real 2p X 2m matrices.

You might also like