https://ntrs.nasa.gov/search.jsp?

R=19700032148 2018-01-02T11:49:27+00:00Z

L
EmR
ES ON CONTROLLABILITY AND OBSEBVABILITY'

Stanford University
Stanfa-ci, Calif. 94305, U.SA.
and
Centre dPAutomatique .
Emle national^ Suerieure des Mines de Paris
Paris, FRANCE

* Lectures derivered at CENTRO INTERNAZIONALE MATEMATIC0

ESTIVO (C.I.M.E.) Seminar on Controllability and Gbservability. Fon-
da.6.oneGuglieImo Marconi, Pontecchio Marcoai (E~logxta,ITALY),from
July 1 through July 9, 1%8.
** This rzsearch was supported in part by Contract AF 49 (63s)-1440 and
in part by Grant NGR-05420473.The writer is.indebted tq CI-M.E. .. for
a.small travel grant.
.- r.-
% .

(Y
0
9 (ACCESSION NUMBER)
I
- (CODE)

(CA'IEG~RYI
Y I
 CENTRO INTERNAZIONALE MATEN\TICO ESTIYO
(C. I. M. E. )

LECTURES ON CONTROLLABILITY AND OBSERXTABILITY

R.E. KALMAN

(Stanford- University)

Corso tenuto a S s s s o Marconi (Bologna) dal 1 a1 9 Luglio 1968

Introduction. 5

Classical and modern dynamical systems. 15

Standardizz;t.ion of def hitions and " c l a ~ s i c a l 'results.

~ 23

Definition of s t a t e s via Nerode equivclence classes. 35

Modules induced by l i n e a r inDut/output maps. 43

Cycliciky and related questioas- 59

Transfer functions. 78

Abstract construction of realizations.

1

Construction of realizztions .
Theory of p a r t i a l realizatiozs.

General theory of 03servability.

Historical notes.

References.
 INTRODUCTION

The theory of controllability and observability has been

developed, one might almost say reluctantly, in response to problems
generated br technological science, especially in areas related to
control, comunication, and camputers. It seems that the first
conscious steps to formalize these matters as a separate area of
(system-theoretic or mathematical) research were undertaken only as
late as 1959, by KALMAN bg60b-c] . There have been, however, m q l
scattered results before this time (see Sectiori 12 for some historical
comments and references), and one might confidently assert today thzt
s a w of the main results have bee3 discovered, more or less independ-
ently, in every country which has reached an advanced stage of
ndevelopmentlvand it is certain that these same results w i l l be
rediscovered again in still more places as other countries progress
on the road to development.
With the perspective afforded by ten years of happenings in
this field, we ought not hesitate to make some guesses of the signi-
ficance of what has been accomplished. I see kwo main trends:

(i) The use of the concepts of controlla5ility and observability

to study nonclassical questions in optimal control and optimal estima-
tion theory, sometimes as basic hypotheses securing existence, more
often as seemingly technical conl'Ltions which allow'a sharper statement
of results or shorter proofs.

(ii) Interaction between the concepts of controllability and

observability and the study of structure of dynamical systems, such
 R. E. Kalman

as: formulation and solution of the problem of realization,

canonicu forms, decomposition of systems.

The f i r s t of these topics i s older and has been studied

primarily from t h e point of view of analysis, although the basic

lemma (2.7 ) i s purely algebraic. The second group of topics

may be viewed as "blowing up" the ideas inherent in the basic

lemma (2.7 ), r e s u l t i n g i n a more and more s t r i c t l y algebraic poixt

of view.

There i s active research i n both areas.

I n the f i r s t , attention has shifted from the case of systems

governed by finite-dimensional l i n e a r d i f f e r e n t i a l equations with

constant coefficients (where success was quick and t o t a l ) t o systems

governed by infinite-dimensional l i n e a r d i f f e r e n t i a l equztions (delay

d i f f e r e n t i a l equations, c l a s s i c a l types of p a r t i a l d i f f e r e n t i e l

equations, etc.), t o finite-dimensioml l i n e a r d i f f e r e n t i a l equa-

tions with time-dependent coefficients, and f i n a l l y t o a l l s o r t s

and subsorts of nonlinear d i f f e r e n t i a l equaticns. The f i r s t two

topics a r e surveyed concurrently by WEISS i19691 while blAEWS [ 1955 1

looks at the nonlinear situation.

own current i n t e r e s t l i e s i n the second s t r e m , and these

lectures w i l l deal primarily with it, a f t e r a rather hurried over-

iriew of the general problem and of the llclassicalllresults.

Let us take a quick look at the most i m p r t a n t of these "classicalf1

results. For convenience I s h a l l describe them i n system-theoretic
(rather than conventional pure mathematical! language. The mathe-
matically trained reader should have no difficulty in converting
them into his preferred framework, by digging a little into the

references .
In area (i), the mort important results are probably those
which give more or less explicit and computable results for control-

lability and observability of certain specific classes of systems.

Beyond these, there seen to be two main theorems:

THEOREM.A. A real, continuous-time, n-dimensional, constant,

linear dynamical system C has the property "every set of n

eigenvalues may be produced by suitable state feedback" if and

only if & is completely controllable.

The central special case is treated in great detail by KALMCLN,

FAZIB, and ARBIB [ 1969, Chapter 2, Theorem 5-10]

; for a proof cf the

general case with background coments, refer to WONHAM [19671. As

a particular case, we have that eveq system satisf'ying the hypotheses

of the theoren can be 'lstabilized"(made to have eigenvalues with

negative real parts) via a suitable choice of feedback. This result

is the "existence theorem" for a1gori.h~used to construct control

systems for the past three decades, znd yet a conscious formulation
of the problem and its mathematical so:-:tion go back t:, about 1963!
(see Tlieorem D below.') The analogous problem for nonconstant linear

systems (g0verned.b~linear differential equations with variable

coefficients) is still not solved.
 R. E.Kalman

THEOREM B. duality Principle1') Every problem of control-

lability in a real, (continuous-time, or discrete-time), finite-
dimensional, constant, linear dynamical system is equivalent to
a controllability problem in a dual system.

This fact was first observed by KALMAN [ lg60aI in the solution

of the o~timalstochastic filtering problem for discrete-time
systems, and was soon applied to several problems in system theory by
KAIMAN [1960b-c1. See also many related comments by KMWN, FALB,
and &BIB [chapters 2 and 6, 19691. As a theorem, this principle
is not yet known to be valid outside the linear area, but as an
intuitive prescription it has been rather iwefKL In guiding system-
theoretic research. Ti-e problems involved here are those of fomula-
tion rather than proof. The basic difficulties seem to point toward
algebra and in particular category theory. System-theoretic
duality, like the categoric one, is concerned with "reversing
arrowsv. See Section 10 for a modern discussion of these poin%s
and a precise version of Theorem B.
Partly as a result of the questions raised by Theorem B and
partly because of the algebraic techniques needed to prove Theorem
A and related lemmas, attention in the early 19601s shifted toward
certain problems of a structural nature which were, somewhat sur-
prisingly at first, found to be related to controllability and
observability. The main theorems again seem to be two:

THEOFEM C . i.tj.on) Every real ( coctinuous-

(canonical ~ecomgos
time or discrete-time), finite-dimensional, consf?r.t, linear dynamical
 R. E. Kalman

system may be canonically decomposed into four parts, of which only

one part, t h a t which i s completely controllable and completely observ-

able, i s involved i n the input/output behavior of the system.

The proof given by m4AN [ 19621 applies t o nonconstafit systems

only under the severe r e s t r i c t i o n t h a t the dimensions of the sub-

space of all controllable and a l l unobservable s t a t e s i s constant

on t h e whole r e a l line. The r e s u l t represented by Theorem C i s f a r from

definitive, however, since f inite-dimensiorlal linear, r c n c o n s t a ~ tsystems

Mt at l e a s t four d i f f e r e x canonical dec~mpositi~ng:it i s
possible and f r u i t f u l t o dualize the notions of controllability

and observability, thereby arriving a t four properties, presently

called

reachability and controllability

as well as
const-ructibilitp and observability.

(see Section 2 definitions.) Any combimtion of a property from

the f i r s t l i s t with a property from the second l i s t gives a canoni-

c a l decomposition r e s u l t ans.logms t o Theorem C. The complexity of

the situation was f i r s t revealed by \TEISS ?ad U L N [1955]; t h i s

paper contributed t o a revival of i n t e r e s t (with hopes of success)

@ the special problems of nonconstant.linear systems. Recent

*.EISS [1969] uses "detenlinabilitytt instead of constructi-

b i l i t y . The new terminology used i n these lectures i s not yet
entirely standard.
 R. E. Kalman

progress is surveyed by WEISS [~9691. Intimately related to the

canonical structure theorem, and in fact necessary to f'ully clarify
the phrase "involved in the input/output behavior of the system': is
the last basic result:

THEOREM D, (uniqueness of I-finiml~ealization) Given the

impulse-response matrix W of a real, continuous-time, finite-
dimensional, linear dynamical system, there exists a res1,'continuous-
time, finite-dimensional, linear dyndcal system % which
. (a) realizes W: that is, the impulse-respogse matrix of
5 is equal to W;

(b) has minim1 dim~nsionin the class of linear systems

satisfying (a) ;

(c) is completely controllable and completely observable;

(d) is uniquely determined (modulo the choice of a basis

at each t for its state space) by requirement (8)

together with (b) or, independently, by (a) together with

(c) .
In short, for any W as described $hove, there is an llessentially
1 -
of the same "type" rrhich satisfies (a) through (c).

COROLLARY 1. If W comes from a constant ~ystem,there is a

constant 3 which satisfies (a) through (c), -
and is uniquelz

-
determined by (a) .t -
(b) or (a) + (c) (modulo a fixed choice of

basis for its state space).

 R. E.Kalman

COROLIA3Y 2. A11 claims of Corollary 1 continue tr, hcld if

"impulse-response niatrix of a constznt, finite-dimensional system"

is replaced by "transfer function matrix of a consta~t,finite-
dimensional system".

The first general discussion of the situation with an equiva-

lent statement of Theorem D is due to -UT [1963b, Theorems 7
and 81. (!Chis paper: does not include coaplete proofs, or even
an explicit statement of Corollaries 1 zzd 2, althougl; they are
implied by the general algorithn given in Section 7. An edited
version of the original unpublished proof of Tqeorem D is given
in KfUNlN, FALB, and ARBIB [ 1969, Chapter 10, Appendix C 1. )
These results are of great importance in .ecgineering system
theory since they relate methods based on t'ne &place transform
(using the transfer function of the systen) asd the time-damin
methods based on input/output data (the netrix bl) to the state-
varizblr (dynamical system) methods developed in 1955-1960. In
fact, by Corollary 1 it follows that the t~:o msthods ~ m s tyield
identical results; for instance, starting with a constznt impulse-
response matrix W, property (c) implies that; the existence
of a stable control lay is always assured by virtue of Theorem A.
~ h u sit is only after the development represented by Theorems A-D
that a rigorous justification is obtained. for tke intuitive design

methods used in control engineering.

As with he or em C, certain formulationel difficulties arise
in connection with a precise definitio~of a "r,onconstantlinear
 R . E. Kalman

dynamical system". Thus, it geems preferable at present to replace

in Theorm D vimpulse-responsematrix W" by "weighting pattern W"
(or ttabstractinput/output map W") and "complet? controllabilityw
by "complete reachability". The definitive form of the 1963 theorem
evolved through the works of WEISS and KAW4.N [19651, YCUA [~966],
and KADZAN; a precis2 :ormulation and modernized proof of Theorem D
in the weighting pattern case was given recently by K U M N , FALB,
and ARBIB [ 1969, Chapter 10, SecC,ion 13.1 A completely gerieral
discussion of what is meant by a 'Iminimal realization" of a non-
constant impulse-response matrix involver many technical coaplica-
tions due to the fact tLzt such a minimal reelization does not
exist in the class of linear differential equatiollswith Itnice"
coefficie~ltf'unctions. Fov the current status of thSs probleia,
consult especially DESOER and VARAIYA E19671, SILVFXMAY a-d MEADOWS
C1.9691, KA.IMAN, FFALB, and ARBIB C1969, Chapter 10, Section 131 anrl
WEISS [1969].
From the standpoint of the present idtures, by far -tihe most
interesting consequence of Theorem D is its influence, via efforts
to arrive at a definitive proof of Corollary 1, on the developmezlt
of the algebraic stream of system theory. The first proof of this
important result (in the special case of .distincteigenvalues) is
.that of GILBERT [19631. Immediately afterwards, a general proof

was given by W [1963b, Sectior,71. !!?cis proof, strictly

co~putationaland.linear algebraic in nature, yields no theoreti-
%

cal insight although it is u s e m as t-hebasis of a computer algorithm.

 R. E. Kalman

Using the classical theory of invariant factors, W.'U [ 196ja]

succeeded in shwiag +&t t'le solution of the mi- realization

problem can be effectively reduced t o the classical Invariant-

factor algorithm. This result i s of grezt theoretical interest

since it sGon&v suggests the m w stzndard M u l e theorctic

approach, b ~ it
t does not lead t o a slmple proof of CoroUary 1

and 5s not a practical ~iS'aodof cmputation.

The best known proof of Corollzsy 1 was obtained i n 1965 by

B. L. lk, with the aJd of z rermzzhble -orit&, which i s equduy 'Ioportant

from a theoreticzl a3ld caqcitationa3 vieh~ctjnt. The early famuB~-

tion of the algorithro uzs &scribed by 30 a& KAL?-5!-1Z [2966], with

b 3 e r refinements discussed i n HO an6 KAELfiX [~96g], KAIXMI, FALB,

and ARLE [1$g, Cmpter 12, Section U] znd E [ P Xt

.l~
96)cl.
A3acst s5xmiltaneously with the wrk of B. L. Ho, t5e basic results

also by Y O U i 4 m d YISSI [1g66l m d

were discovered Znd~pende~tQ-
by- SILI'E&YUT [ 19661 . The scbject goes k c k t o the 1 s h cen%my

and centers around .tine t-heoqr of Hvrkel mat rice^; 'aawevsr, my

of the results just referenced seem t o bs ~ ~ e n + , a new.

U y This

field i s currefitly i n a very w t i v e stage or" dcvelopent . Ve skzI-2

discuss the essentizl ideas invoived in Sectims 8-3. F ! other
topics, especially S i l v e r n ts geaerclization of tbe algorithm t o

noncocstm% system uricrtunately c m m t 2 colered due t o l . x k or"

th.
 R. E. Kalman

Acknowledgment

It is a pleasure to thank C. I. M. E. and its organizers,

es~eciallyProfessors E. Bampiani, E. Sarti, and E. i3elardinelli)

for arranging a special conference on these topics. The sunny

skies and hospitality of Italy, along with Polognese food played

a subsidiary but vital part in the success of this important

gathering of scientists.
 R . E. Kalman
i. CLASSICAL ANI MODERN DYW'CAL SYSTEW

I n mathematics the term dynamical system (synonyms: topological

dynamics, flows, abstract dynamics, etc.) usually connotes the action

of a one-parameter group T (t5e reals) on a s e t X, where X is

at l e a s t a topological space (more often, a differentiable &fold)

and the action i s at l e a s t continuous. This setup i s physically

motivated, but i n a very old-fashioned sense. A l'dynamical system"

as just defined i s an idealization, generalization, and abstraction

of Newton's world view of the Solar System as described v i a a f i n i t e s e t of

nonlinear o r d i ~ a r yd i f f e r e n t i a l equations. These equations represent

the positions and momenta o f t h e planets regarded as point masses and

a r e completely deternined by the laws of gravitation, i.e., they do

not contain any terms t o account f o r "external" forces t h a t may a c t

on the system.

Interesting as t h i s notation of a dynamical system may be (and

i s l ) i n pure mathematics, it i s much too limited f o r the study of

thc;e dynamical systems which are of contemporary interest. There

a r e a t l e z s t three different ways i n which the c l a s s i c a l concept

must be generalized:

(i) The time s e t of the system i s not r,ecesaarily r e s t r i c t e d

t o the reals;

- (ii) A state x E X of %h2 s y ~ t e mi s not merely acted upon by

the "passage of time" but also by inputs which a r e or could be mani-

pulated t o bring about a desired type of behavior;

 R . E.Kalman

(iii) The s t a t e s of t h e system cannot, i n general, be observed.

Rather, the physical behavior of t h e system i s manifested through

i t s outputs rhich a r e many-to-one f'unctions of the s t a t e .

The generalization of the time s e t i s of minor i n t e r e s t t o us

here. Tht sotions of input and output, however, a r e exceedingly

fundamental; i n fact, c o n t r o l l a b i l i t y i s r e l a t e d t o t h e input and

observability t o t h e output. With respect t o dynamical systems i n

the c l a s s i c a l sense, neither c o n t r o l l a b i l i t y nor observability a r e

meaningful concepts.

A much more detailed discussion of dynamical systems i n the modern

sense, together with rather detailed precise definitions, w i l l be

found i n KAINAN, FALB, and ARBIB [1969, C h p t e r 11.

From here on, we w i l l use the term "dynamical system" exclusively

i n the modern sense (we have already done so i n the ~ n t r o d u c t i o n ) .

The f o l l o w i r ~symbols w i l l have a fixed mea-xing throughout the

paper :

T = time set,

U = s e t of input. values,
X = s t a t e set,

(1-1) Y = s e t of output values,

R = input functj-ons,
9 = t r a n s i t i o n map,

q = readout map.

The following assumptions w i l l always jpply(otherwise the s e t s

above a r e a r b i t r a r y ) :
 R. E. Kalman

T = an ordered subset of the r e a l s R,

-
R = class of flrnctions T -, U such t h a t

(i) each function w i s undefined outside some

(1.2) f i n i t e interval JoCT dependent on w;

(ii) if Jw nJot = there i s a function

o E Q which agrees with w on Jo and

with wt on J
,.
For mosl-, _c-.->ses later, T w i l l be equal t o Z.
-= (ordered)

abelian group of integers; U, X, Y, Q w i l l be linear spaces; "unde-

fined" can be replaced by "equal t o 0"; and "functions undefined out-

side a f i n i t e interval" w i l l m e a n the same as " f i n i t e sequences".

The most general notion of a dynmical system f o r our present

needs i s given by the followi~lg

(1.3) DEFINITION. A dynamical s y s t s C is a comp&site object

consjsting of the maps cp, 7 defined on fhe s e t s T, U, R, X, Y

(as above):

9: T X T X X X R 4 X,

: (ti 7, X, a) ~3 ~ ( t 7,
; X, a)

undefined when~ver t >- 7 ;

1: T X X -, Y: (t, x) t, q(t, x).

The &ransition
-- mi! QI satisfye,: the following asSumptions:
 R . E. Kalman

(1.6) -
if u, = cuN on [T, tl, then f o r a l l sE [7, t]

v(S; 7, Xj u)) = v(S; 7, X, a').

The definition of a dynamical system on t h i s l e v e l of generality

should be regarded only a s a scaffolding f o r the terminology; interest-

ing mathemtics begins only a f t e r further hypotheses a r e made. For

instance, it i s usually necessary t o endow the s e t s T, U, a, X, and

Y with a topology and then require t h a t 9 and 7 be continuous.

(1.7) EXAE4PLE. The c l a s s i c a l setup i n topological dynamics may

be deduced from our Definition (1.3) i n the following way. Let

T = _R- = reals, regarded a s an abelian group under the usual addition

and having t h e usual topology; l e t consist only of the nowhere-

defined function; l e t X be topological space; disregard Y and q .entirely;

define -
g, f o r a l l t, 7 E T and write it as

g(t; T, x, w) = x-(t - T),

t h a t i s , a f'mction of x and t - T alone. Check (1.4-5); i n

the new notation Yley become

x-0 = x and x - ( s + t ) = (xhs)*t.

Finally, require t h a t the map (x, t) I+ x g t be continuous.

(1.8) INTEP~ZT+TIOPU'. The e s s e n t i a l idea of Definition (1.3) i s

t h a t it axiomatizes the notion of state. A dynamical system i s informally
 R. E. Kalman

a rule f o r s t a t e transitions (the function cp), together with suitable

means of expressing the effect of the input on the s t a t e and the effect

of the s t a t e on the output (the function 7). The map cp i s verbalized

as follows: Itan input a, applied t o the system C i n s t a t e x at

time -c produces the s t a t e cp(t; -r, x, co) a t time t." The peculiar

definition of an input flrnction (I, i s used here mainly f o r technical

convenience; by (1.6) only equivalence classes of inputs agreeing over

[ r , t] enter into the determination of cp(t; T, x, cu). "03 not definedtt
a t t means no input a c t s on C at time t -
The pair (7, x) E T X X w i l l be called an event of a dynamical

system C.

In the sequel, we s h a l l be concerned primarily with systems which

are finite-dkensicnal, linear, and continuous-time or discrete-time.

Often these systems w i l l be also real and constant (= stationary or

time-invariant). We 'leave the precise definition of these terms i n

the context of Definition (1.3) t o the reader (consult KALMAN, FAI;Bj

o r ARBIB [I-969, Chapter 11 a s needed) and proceed t o nabe some ad hoc

definitions without detailed explanation.

The following conventinns w i l l remain i n force thkoughout the

lectures whenever the l i n e a r case i s discussed:

(1.91 Continuous-time. - U=
T= E -FJm, X = FJ-n, Y = -g*,
. . . .
= a l l continuous f b c t i o n i R- -t -
Rrn which -ish out-

side a f i n i t e interval.-

(]..lo) Discrete-time. T= g
-, K = fixed f i e l d (arbitrary),
 R. E. Kalman

U = I?, X = X?, Y = ICP, Q = a l l functions

-
-
Z 4 which'are zero for a l l but a f i n i t e number of

t h e i r arguments.

Now we have, finally,

(1.l.l) DEFINITION. A real, continuous-time, n-dimensional, linear

dynamical system.-C i s a t r i p l e of continuous matrix f'unctions of

-
time (I?(-), G(*), ~ ( 0 ) ) where

F ) : -
- + {n X n matrices over g)
~(0): -+
R [n x m matrices over -
R),
H -1: - -t {p x n matrides over -FJ).

These maps determine the equations of motion of C i n the following

manner :

dx/dt +~(t)~(t),
= ~(t)x

yit! = ~(t)x(t),

-
where t E -
R, xE gn,
- -
urn($)E-gn, 'R, ~ ( t E) gP.
-
To check t h a t (1.12) indeed makes C i n t o a well-defined dyrmical

system i n t h e sense of Definition (1.3), it i s necessary t o r e c a l l the

basic f a c t s about f i n i t e systems of ordinary l i n e a r d i f f e r e n t i a l equations

with continuous coefficients. Define the map

$(t, r): . R- X R- -, (n X n matrices over g)

-
t o be the family of n X n matrix solutions of the l i n e a r d i f f e r e n t i a l
 R. E. Kalman

equation

dx/dt = F(t)x, xE R-

subject to the initial condition

%(T, T) = I = unit matrix, TE -

R.

Then b is of class c1 in both arguments. It is called the

transition matrix o$ (the system C whose ffinfinitesimaJI'
transition
matrix is) ~(0). From this stnnilard result we get easily also the
fact that the transition map of C is explicitly given by

while the readout ma^ is given by

It is instructive to verie that cp indeed depends only on the equiva-

lence class of s which agree on [T, t]
UI~ .
In view of the classical terminology "linear differential equa-
tions with constant coefficients",we introduce the nonstandard

(1.15) DEFINITION. Areal, continuous-time, finite-dim3nsional

\

linear dynamical system C = (F(-), G ( * ) , H(-)) is called constant

if'f all three matrix f'unctions are constaht.

In strict analogy with (l.l5), we say:

(1.16) DEFINITION. . A discrete-timeZ finite-dimensional, l i ~ s

-
constant,dynamical system_ C over K is a triple (F, G, H) of
 R. E. Kalman

n X. n, n X m, p X n matrices over the f i e l d K. These m%ps deter-

mine the equations of motion of i i n the following manncr:

-
where -,
t Eg x E m(t) E K ~ , ~ ( t )E K ~ .

I n the sequel, we s h a l l use the notations (F, G- j - or

F - H) t o denote systems possessing certain properties which

a r e t r u e f o r any H or G.

Finally, we adopt the folloicing convention, which i s already

implicit i n the preceding discussion:

(1.18) DEFINITION. The dimension

-
n of a dynamical system

X i s equal t o the. d i h n s i o n of Xz as a vector space.

 R , E, Kalman

2. STANDAR9IZATION OF DEFINITIONS AND "CL4SSIC~'FGSULTS

In t h i s section, Tie s h a l l be mainly interested i n f i n i t e -

dimensional l i n e a r dynamical systems, although the f i r s t two

definitions w i l l be quite general.

Let C be an a r b i t r a r y dynamical system a s defined i n

Section 1. We assume t h e following s l i g h t l y special property:

There e x i s t s a s t a t e + and an input such t h a t

cp(t; 7, fi, u+) = f o r a l l t, T E T and t --

> T.

For simplicity, we write and @ as 0. (When X

and SZ have additive structure, 0 w i l l have the usual mean-

ing.) The next two definitions r e f e r t o dynamical systems

with t h i s extra, property.

(2.1) DEFINITION. An event ( ) i s controllable iff.S

there e x i s t s a t E T and an w E SZ -t -
(both
and w may depend

-
on (7, X) ) such t h a t

In words: an event i s controllable i f f it can be 'transferre

. t o 0 i n f i n i t e time by an a p p r ~ p ~ i a choice
te of t h e input f'unction

w. Think of t h e path fron? (7, x) to (t, 0) as t h e graph of a

function defined over [T, t].

'The technical word i f f means lf and only i f .

 .' .
R E Kal man

Consider now a reflection of t h i s graph about r. This .

suggests iz new definition which i s a kind of "adjoint" of the

definition of controllability:

(2-2) DEFINITION. An event (T, x) i s reachable i f f there

-
%san s E T and an- u, E Cl -s -
(both
and o, may depend on

(7, x)) -
such t h a t

We e.mp:iasize: controllability and reachability are entirely

different concepts. A strzking example of t h i s f a c t i s encountered

below i n Proposition (4 -26) .

We s h a l l now review b r i e f l y some well-known c r i t e r i a f o r and

re1ation:;'betveen reachability and controllability i n l b e a r systems.

(2.3) PROPOSITION. I n a reel, cont inuous-time, f inite-dimensiona~,

linear dynmical system C = (F( * ), G( ), 9 - ), an event (T, x)

(a) reachable i f and only i f x E range Q(s, r ) -

for

-
some s€ R,
- s < r, where

G(s, T) = .lT Q*(T, C)G(U)G'(C)O;(T, U)&

. s
(b) controllable if an only if x E range ~ ( r t, ) -
for

- -t, t > T, -
some t E I where

The original proof of ( 3 ) i s i n KALMAN [1960b]; both cases

are t r e a t e d i n & t a i l i n KAMANt FALB, and L W B [1969, Chzpter 2,

 R . E. Kalman

Section 21. Xote t h a t i f G(*) i s i d e n t i c a l l y zero on (- =, rj

we cannot have rezchability, and i f G(-) is idznticzlly

zero on (T, + a ~ ) we cannot have c o n t r o l k b i l i t y .
For a constvlt syskm, t h e h t e g r a l s above depend only on

t h e difference of Yne limits; ker-ce, i n p r t i c u l z r

(2.4) P R O ~ I T I O N . In a real, continuous-tlbe, finite-dinensional,

linear, constant ~ - m L c a 1
system a2 even'; (T, x) i s rezcbble

for a l l r i f and only i f it i s rezchable f o r one T; sn ev2nt

i s reachable if and only i f 1'; i s cc\~itrolla.'Dle.

From (2.3) one can o b G h i n a streigbtfor.~-a-dfashion z l s o

tile following m c h stronger result:

THEOEI4. a real, c o ~ t i a n ~ u s - t i n en-&ensj.o~2!,

,

i s reachabfe (or, equivzlently, controllcble) 9 r tl R--

i f and o a -- -
v if

i f t h i s condition i s sztisfied. we can choose s=T - 6: t = r i- 6,

-
with 6 > 0 arbitrzry. (me sgan of E. seqv-ence of m t r i c e s 5s t o

be i n t e q h e t e d as t h e vector space generated by the columns of

these r;s.tricrts.)
 A proof o f (2.5) m y be found iil KAlXQJ, HO, and IIILWNDRA

[19631 and i n gllLMAN, FAD, and ARBIB [I$@, Chapter 2, Section

31. A trivial but noteworthy consequence i s the f a c t t h a t t3e

definition of reachable s t a t e s of C i s "co~rdinate-free":

- -

(2.6) COROLiAQY. The s e t of reachzbie (or controllable)

s t a t e s of C i n Theorem (2.5) i s a subspce of the r e a l vector
space Xz, t h e s t a t e s p c e of Z.

Very of'ten the a t t e n t i o n t o individuzl s t a t e s i s unnecessary

II
end therefore mvly authors prefer t o use the terniaology C is

c m p l e t e l y reachable at T" for "every event (-c, x), .r = fixed,

It
xE i s reachable", or 2 conpleteljr reacbsble" f o r "evesy

event; i n Z i s reachable", etc. Thus (2.5), together with t h e

-ley-Hamilto?l theorem, i n p l i e s the

(2-7) BASiC L!Jil+lA. A real, continuous-t ime, n-db.e?lsional,

l i n e a r , constant Q W c a l systen C = (F, G, -) s co~pletely

reachable i f a n only i f

Condition (2.8) i s ~ e r y
well-&own; it o r equivalent f o m s of

it have bsen discovered, e x p l i c i t l y used, or h p l i c i t l y assumed by

many a-rlthors. A trivially equivalent form of (2.7) f s given by

(2.9) COROLMRY 1. A constant system C = (F, G, -) -

is
completely reachable i f and only if the s ~ s l l e s F-invariant
t

subspace of conteining ( a l l colu~mvectors of) G i s ]t

-
--
itself.
 R. E. Kalman

A useful variant of the l a s t f a c t i s given by

(2.10) COROLLARY 2. (w. ~ahn) Aconstant system C = (F, G, -)

i s completely reachable i f and only i f there i s no nonzero elgen-

vector of F which i s orthogonal t o (every column vector of) G.

FinaUy, l e t us note tht, far from b e i q a technical condi-

tinn, (2.5) has a d i r e c t system-theoretcc interpretation, as

f oilows :

(2.~) PROWSITION. The s t a t e sozce % of a reel, continuous-

time, n-dimensional, l h e a r , constant dy-nmical sys';en~ Z = (F, G, -)

may be written a s a direct sum

-
which induces a decomposition cf the equztions of notion as (obvious

notations)

The subsystem 5 = ( F ~ ,GI, -) i s completelv reachable. Hence

a state x = (2,
x2) E 3 i s reachable if and only if x2 = 0.

PROOF. We define X1 t o be the set of reachable s t a t e s

of 2; - br (&5) t h i s i s w F-invariant subspace of %. Hence, by

finite-dhensionalityj XI i s a direct s m d i n . construc-

tion, every s t a t e i n q i s reachzble, and (every column vector of)

 R. E. Kalman

G belongs t o Xl. Ihe F-invariance of X implies t h a t

1
FU = 0, which implies the asserted form of the equations of

motion. 0

(223) -. Note t h a t X2 i s not i n t r i n s i c a l w - d c f i n e d

( i t depends on an a r b i t r a r y choice i n completing the direct sum).

Hence t o say t h a t "(0, x2) i s an unreachable (or uncontrollable)

state i f x2 # 0" i s an abuse of language. More precisely: thp

s e t of a l l reachable ( o r controllable) s t a t e s has the structure of

-.
a vector suaeg b d t h e s e t of a11 unreachable (or uncon-i;rollzbl.e)

s t a t e s does not have such structure. This f a c t i s imgortaat t o

bear i n mind f o r the algebraic deve10,oment which follows after

t5is section znd a l s o in t h e definition of observability and

c o n s t r u c t i b i l i t y below. I n general, the d i r e c t sum cannot be

chosen in such -a way t h a t -

F,,, = 0.

While condition (2.8) has been frequently used as a technical.

requirement i n the soluticn of various optimal control problems i n

the l a t e 1950's, it was only i n 1959-60 t'at t h e r e l a t i o n between

(2.8) and system theoretic questions was c l a r i f i e d by lYALMAN [lgbOt~-c]

via Definition (2.2) and Propositions (2.5) and (2.l.l). (see Section

11 f o r further details.) I n other words, without the preceding

- discussion the a s e of (2.8) may appear t o be 2-rtificial, but i n f a c t

it i s not, a t l e a s t i n problems i n which control enters, because,
by (2.12) control problems stzted f o r % a r e nontrivial only with

respect t o the intrh>.sicsubspace

*1
.
 R. E..Kalman

The hypothesis "constant" i s by no means e s s e n t i a l f o r

Proposition (2.11), but we must forego further comments here.

For l a t e r purposes, we s t a t e some f a c t s here f o r discrete-

time, constant l i n e a r systems a ~ a l c g o ~

.to
s those already developed

f o r t h e b continuous-time counterparts. The proofs a r e straight-

forward and therefore omitted (or given . l a t e r , f o r i l l u s t r a t i v e

(2.14) PROPOSITION. A s t a t e x of a real, discrete-time,

n-mensional, linear, constant dyxanical system C = (F, G, -)

is reachable i f and only i f

(2.15) X E span(G, FG, ..., lT1G).

Thus such a system i s completely reachable i f a2d only i f (2.8)

-
holds

(2.16) PRORX1'TION. A state x of t h e system Z described

i n Proposition (2.14) i s controllable i f and only i f

(2.17) 3, E span (F-'G, ..., F%),

-
where

(2.18) PROR3SITION. I n a real, discrete-time, finite-dimensional,

linear, constant dynamical system C = (I?, G, -) a reachable s t a t e

i s always controllable and the converse i s always t r u e whenever

det F # 0.
 Note a l s o t h a t Propositions (2.11) and i t s proof contime
If
t o be correct, without any modification, when continuous-the"

i s replaced by "discrete-time".

Now we t u r n t o a discussion of observability.

The o r i g i m l defilnition of observability by W i ~ ~ [ l g 6 0 b ,

Definition (5 -23) 1 was concocted i n such a way as t o take advan-

tage of vector-space d w l i t y . The conceptual problems surround-

ing duality a r e easy t o handle i n t h e l i n e a r case but a r e s t i l l

by no means fully understood i n t h e nonlinear case (see Section

10). I n order t o g e t a t t h e main f a c t s quickly, we s h a l l consider

here only t h e l i n e a r case and even t3en we s h a l l use t k - u n d e r -

lying idea of vector-spzce duali%y in a r ~ t l n e rad-hoc fashion.

The reader wishing t o do so can e a s i l y t u r n our r m r k s i n t o z.

s t r i c t l y dual treatment of f a c t s (2.1)-(2.12) with the a i d of

5he setup i n t r ~ d u c e di n Section 10.

(2.19) DEFIPITION. An event (T, x) i n a real, continuo~s-

time, finite-dimensional, l i n e a r dynamical system C = (F(*), -,H(*))

i s unobservable i f f

H(s)$(s, T)X = 0 for a l l s E [r, a

).

(2.20) DEFINITION. With respect t o the same system, an event

(7, z) is unconstructible* i f f

*In the older l i t e r a t u r e , s t a r t i n g with KAL2AN [ ~ g a b ,

-
Definition (5.23)], it i s t h i s r . ept which i s called "observabilityl'.
By hindsight, t h e present choice ci;words seems t o be more natural
t o the m i t e r ,
 R . E.Kalman

H(C)%(U, r ) x = 0 for a l l u € (- w, TI.

The motivation f o r the f i r s t defir-ition i s obvious: the

11
occurrence" of an unobservable ?vent cannot be detected by look-

ing at the output of the system a f t e r time T. (The definition

subsumes LD = 0, but this i s no l o s s of generality because of

linearity.) The motivation f o r the second definition i s l e s s

obvious but i s i n fzct strongly s u z e s t e d by s t a t i s t i c a l f i l t e r i n g

theory (see Section 10). In any case, Definition (2.21) comple-

ments Definition (2.20) i n . exactly the same way as Definition (2.1)

complements Definition (2.2).

From these definitions, it is very easy t o deduce the follow-

ing c r i t e r i a :

(2.21) PROIQSITION. I n a real, continuous-time, finite-dimensional,

linear dynamical system Z = (F(*), -, ~ ( 0 ) ) an event (T, x) -

is

(a) unobservable i f and anfy i f x E kernel %(T, t )

for a l l t € -, t > 7,
3 where

(b) unconstructible i f and only i f x E kernel ~ ( s 7)

,

_for a l l sE -
R, s < T, -
where
 - 32 - R . E.Kalman

PROOF. B r t (a) follows immediately from the observation:

x E kernel ~ ( r ,t) @ H(s)@~(s, r)x = 0 for all s E [ r, t]. Part
(b) follows by an analogous argument.

(2.22) REMARK. Let us compare thfs result with Proposition (2.3),

and let us indulge (only temporarily) in abuss of language of the
following sort:*

('F, x) = unreachable @ x E kernel $(T, t)

for all t >r
and

(r, $ = observable x E range d(~, t)

for some t > T.

From these relatims we can easily deduce the so-called "duality

rules"; that is, problems involving observability (or constructibil-
ity) are converted into problems involviag reachability (or ~ontrol-
.,

lability) in a suitably defined dual system. See KAIMAN, FALB,

and ARBIB [1969, Chapter 2, Proposition (6.12)l and the broader
discussion in Section 10.

We will say, by slight .&use of language, that a system is

completely observable whenever 0 is the only unobservable state.
Thus the Basic Lemma (2.7) "dualizes" to the

(2.23) PXOPOSITION. A real, continuous-time or discrete-time,

n-dimensional, linear, constant dynaruical system Z = (F, - , H)
* A l l this would be strictly correct if we agreed to replace
11
It
direct sum" in Proposition (2.11) and its counterpart (2.25) by
orthogonal.direct suint' ; but this would be an arbitrary convention
which, while convenient, has no natural system-theoretic justifica-
tion. - ,Rerezd'~er&$k.'(2.. 13).
is completely observable if and only if

(2.24) r ( F ... ( F ) ) = n.

By duality, com2lete constructibility in a continuous-time

system is equivalent to observability; in a discrete-time system
this is not true in general but it is true when det F # 0.

It is easy to see also that (2.11) " .afizes" to:

(2.25) PROPOSITION. !The state space 3 of a real, continuous-

time or discrete-time , n-dimensional, linear, constant dynamical

system C = (F, -,H) m y be written as a.direct sum

and the equations of C are decomposed correspondingly as

PROOF. Proceed dually to the proof of Froposition (2.11)'

beginning with the definiticn of X1 -
as the set of all unobservable

states of C.

Combining Propositions (2.11) and (2.25) gives Theorem C as in

KALMA.N [ 1962I .
This completes our survey of the "classical" results related
 R . E. Kalman

t o reachabilitjr, controllability, observability, and

constructibility.

The r e m i n i n g lectures w i l l be concerned exclusively with

discrete-time systems. The main motivation f o r t h e succeeding

will be t h e algebraic c r i t e r i a (2.8) and (2.24)

develop~~ents

as well as a deeper e x a n h a t i o n of m-eorens C and D of t h e

Introduction.
 R..E. Kzlman

3. DEFINITION OF STATES VIA NERODE EQUIVALENCE CLASSES

A c l a s s i c a l dynamical system i s essentially the action of the

time s e t T (= reals) .on the s t a t e s X. I n other words, the

s t a t e s a r e acted on by -an abelian group, namely -
(f! + usual
definition of addition). This i s a t r i v i a l fact, but it has deep

consequences. A (modern) dy-namical system i s the action of the

inputs on X; i n exact analogy with the c l a s s i c a l case, t o

the abelian structure on T there corresponds an (associative

but noncomutative) semigroup structure on Q. The idea t h a t Q

always admits such a structure was apparent';; overlooked u n t i l

the l a t e 1950~swhen it becam* fashionable i n automata theory

.
. !Phis seens t o be the "right" way
(school of SCHUTZENBERGER)

of translating the i n t ~ i t i v enotion of dynamics i n t o mathematics,

and it w i l l be fundamental i n our succeeding investigations.

I'Gi s convenient t o assume fron now on, u n t i l the end of

these Lectures, t h a t

(3.1) T = time s e t = - -
Z = additive (ordered) group of
integers.

Since we shall Le only interested i n constant systems from

here on, we s h a l l adopt the following normalizaticn convention:*

*In the discrete-time nonconstant case, we would have t o deal

with Z copies of Q, ezch normalized with respect t o a different
p a r t i c a r value of T E 8. -
 R eE. Kalman
(3.2) No element of 0 i s defined for t > T = 0.

I n view of (3.2), we can define the lo1 of (I, by

Iwl = max [ - t € -
g: w i s fiat definad f o r -&q s < t).

Biifore defining the semigroup on 52, we introduce another

fundamentsl notion of dynamics: the ( l e f t ) shift operator ui,

defined for a l l Q >- 0 in Z_- by

Note t h a t the definition of on i s comgatible with the normaliza-

t i o n (3.2).
If TUnJu, = empty f o r cu, u' E Q, we define the join
of w and w' as the function

" on Jd
(3.4) w y 0' =
C"' on J
,.

When Q has an additive structure, then we replace u, co' by u, + cur.

(3.5) DEFINITION. There i s an associative operatia

o: X 0 3 J2, called concatecation, defined by

.
Note that, by (3.2) through (3.4), i s well -defined.

Note also t h a t t h e asserted existence of concatenation r e s t s

on the fact t h a t R' ir: nade up of flmctions defined over f i n i t e
intervals i n . T. We might express the content of (3.5) also as:
Q i s a semigroup with valuation, since e15dently fuov1 = + 1~1.
 In view of (3.5), it i s natural t o use an abbreviated n o t a t i o s
also for the transition fhmtian, ss follows:

ITOWwe come t o an important nonclsssical concept in dyxmmical

system, u b s e evolution was strongly influenced by problems i n

c o m m + - ~ ~ t i oand
n s autaata theory: a discrete-time constant
input/output map

We k t e r p r e t t h i s map as follows: y(1) is the output of same

system C (say, a &igi&l c-uter) uhen C i s subjected t o .

the (f5nite) ingut sequence o, assuming that C is some fixed

initial equilibrium state before the application of a- 'fhis
definition automatically incorporates the notions of Udiscrete-
tinen as well as "causaln or "dynamics" (the l a t t e r because
y(t) i s not defined for t < 1). However, (3.7) does noz;
clearly imply "constancy" (implicitly, however, t h i s i s clear from

the nomaUzation a s s q t i o n (3.2) on Q) . To nake the definition

more forceFul, we dend f t o the map

.(3.8) -f: a + I' = Yx Y ... (infinite cartesian prcc2uct)

: a t+ ( f ( 4 , f(c$~), -*- 1 = ( ~ ( 1 1 ,~!2), 0-• 1.

Interpretation:
-f gives the output sequence Y = (y(l), ~ ( 2 1 ,

of the system Z af'ter t = 0 resulting from tha applicatior cf an

Wbserve that xw i s the s t r i c t analog of the notation xt

custmary i n topalogical dynmics. The action of u, oh x satis-
f i e s xo(u,ot.: = ( x a ) o v in view of (1.5)
 R. E. Kalman
input w which stops at t = 0.
Tbis definition e-pressescausality more forceflrlly m d
incorporates consfaac:r, provided we define the ( l e f t ) shift
operator ur on I' so as to be comptible w i t h (3.3). So,

for any r >- 0: c € Z-, let

Note: the operator "appendsttan undefined term at 0, the

operator cr "discardsw tfle term y(1).
Now, drcppkg the bar over f, we adopt

-
(3 10) DEFSLJITION. A discrete-time, constant input/output map

(of some underlying ~vnanicalsystm Z) is any map f such that

the following diagram

is comutztive. We seg %hat f is linear iff it is a K-~ctor

7-

s ce h m ~ r n o r u ? ~ i
J z . - . t ' "sm
- *

It w i l l be convenient to regard (3.10) as the exker~~l-

.dsfinition of a dymmical system, in contrast to the internal
definition set up in Section 1.
Intuitively, we shculd think of f as a highly idealized
kind of experimzntal data; azmely, f incorporates all possible
infomation that could be gained by subjecting the underlying
 R. E. Kalman

system t o eqerinents i n which only input/output data i s avail-

able. This point of view i s related t o experimntd physics the

same kqv as the classical notion of a dymmlrld,l system i s related

t o Newtonian (axiomatic) pwsics.

The basic question which m t i v a t e s much of what -~ll

follow

can now be formulated as follows:

(3. EL) PROBISM OF REALIZATION. Given only the knowledg~of

f (but of course also of Z-, R, -

and )'I how can we discover,

i n a =thematically consistent, rigoross, and nztural way, the

propertles of the system wh5-ch i s sup~osedt o underlie the

piven input/output map f ?

This suggests immediately the follcsiing f b d m e n t a l concept:

(3.12) DEFINITIOII. A fixed dynamical systex C (internal

definition, a s i n Section 1) i s a realization of a fixed input]

output map f O fo = fZ , t h a t is,' f 0 i s identical with

0
the input/output map of k.
I n view of the notations of Section 1plus the special-con-

vention (3.6), the explicit. form o f the realization condition i s

simply that

(3.u) fob) = - % (%
0 0
(0; - 14, *, 4)

for all w Q. ~ n symbol

e * stands for an arbitrary equili- -

brim state in viich Eo remeins, by definition, u n t i l the

application of o. (Later we simply tahe * t o be 0.) ,

 R. E.Kalman

To solve the realization problem, the c r i t i c a l step i s t o

induce a definition of X (of some zo) from the given fo.

It i s rather surprising thst this step turns out t o be trivial,

on the a b s t r k t level. (on the concrete level, however, there are
many unsolved problems i n actually computing what X is. In
Section 8, we s h a l l solve this problem, too, but only i n the
line= case.) The essential idea seems t o have been published
first by NERODE [19~81:

(3.14) DEFINITIQPJ. Make the concatenatron semigroq i2 into -

a monoid by adjoining a neutral elptueilt
.- @ (which i s the nowhere-
defined flrnction on Z).
- -
Then u, =
-f at : o i s Nerode

equivalent t o ot Gth respect t o f ) iff

f(oov) = f(co8o~) for a l l v E Q.

There are n;-w~r intuitive, physical, historical, and technical

reasons (which are scattered throughout the literafare and cone=-a
trated especially strongly i n ICABUI?, FALB, and ARBIB [19691) for
using t h i s as the

(3.15) MAIN DEElNl!CION. The s e t of equivalence classes under

-
=f' denoted a s Xf =. ( ( ~ 0 ) ~cu
: E Q), i s the state set of the
iput/output ma2 f.

Let us verify inimediately.that (3. S ) makes sathemstical

sense:
 R. E.Kalman

(3.16) ' PROPOSITION. For each linear, constant input/output ma2

f there exists a dynamical system Zf-. such that
(a) Ef realizes f;

PROOF. We show how t o induce C p given f. We

define the s t a t e s e t of Zf by (b). Ftnther, we define the

transition f'unction of C by
f

We must check that on the l e f t of a

= i s well defined (note

two different uses of ! ) that is, fudependent of the repre-

sentation of x as - (4 . T U s follows trivia- from (3.14) .
llov we define the readout map of Ti by

Again, this map i s well defined since we can . W e V = $8.. as a

special case i n (3-14) . Then
('L*V) = % = f(a.v),
f f

and the realization conditiw~(3.6) i s verified. Hence claim ( e )

. i s correct. 0

(3.19) COMMEXCS. In' autanata theory, Zf - i s known as the

; reduced form of.ary system which realizes f. Clearly, say two

reduced forms are isomorphic, inthe set-theoretic sens?, since
the set Xf is intrinsically defined by f. his observation
is a weak version of Theorem D of the Introduction; here "unique-
ness'' mcans flmoduloa pernutation of the labels of elements in
the set x~".) Kootics also that LC is con~letelyreachable
since, by Definition (3.15 I , every element x = of Xf
is reac-habie via any elmefit cut - in the Xerode equivzleact class
( u ) ~ . As to observcbility c? .Zf, see Section 10.
 .
K. E Kalman

4. MODULE3 INDUCED BY LINEAR INHJT/oIPTPUT MAPS

'

We are now yea* to enhark on the main topics of these lectures.

It is assumt, that the reader is conversant w 5 t modern
~ algebra (espe-
cially: abelian groups, commutative rings, fields, modules, the ring
of polynomials in one variable,and the theory of elementary divisors),
on the level of, sag, VAN DER WAERDEN, LAI?G [1965I, W [ 19651 or
ZARISKI and SAMUEL [1958, Vol. 11. The material covered fram here
on dates from 1965 or later.
Standing assumptions until Secticr-.10:

(4.1) All systems C = (F G H) are discrete-time, linear,

constant, defined over a fixed field K (but not necessarily
finite-dimensional) .
Our immediate objective is to provide the setup and proof for the

(4-2) THEOREM OF ILNEAR SYSTEM THEORY. The natural

state set Xf associated with a discrete-time, linear, constant input-
output map f over a fixe6 field K admits the structure of a finitely
-
senerated module over the ring K[z] of polynomials (35th rndeterminate
z and coefficients in K) .
(4.3) CWME31TS. Since tke 5ing ~[z;] will be seen to be related
to'the inputs to C, this result bss a superficial resemblance to the
fact that in an arbitrary m e a l system C the state set 3 admits

the action of a semigr~up,namenamely (see (3.5) and related footnote).

It turns out, however, that this action of R on X, which results
fram camblrAng the concatenation product in R with the d~f~iition
of
 R. E.Kalman

states via Nerode equivalence, is incompatible with the addit;ive

structure of S2 [KAI&WJ, 1967, Section 31. O u r theorem asserts the
existence of an entirely different kind of structure of X. This
structure, that of a K[ zl-module, is not just a consequence of
dynamics, but-dependscritically on the additive structure on Q
and on the linearity of f. The relevant multiplication is not
(noncommutative) concatenation but (commutative) convolution (because
convolution is the natural product in KLz;]); dynamics is thereby
restated in such a way that the tosls of commutative algebra become
apglicable. In a certain rather definite sense (see also Rzmark
(4.30)), Theorem (4.2) expresses the algebraic content of the method
of the Laplace transformation, especially as regards the practices
developed in electrical engineering in the U.S. during the 1950's.
The proof of Theorem (4.2) consists in a long sequence of canoni-
cal constructions and the verification that everything is well defined
and works as needed.
In view of (4.1) and the conventions made in Section 1, R may
be viewed as a K-vector space and m(t) = 0 for almost all t E Z_-
and all o E Q. convention (3.2 ), we have assumed also that '

m(t) = 0 .for all t > 0. As a result, we have that:

(a) Q Km[z] gs a ' zctor space. Let us exhibit the isonor-
phism explicitly as follows:

By (3.2 ), the sun in (4.4) is always finite. The isomorphism

 P.E,Kalman

obviously preserves the K-linear structure on Q. In the sequel, we

8hall not distinguish sharply between m as a function T -r K? and '

a as an in-vector polynomial.

(b) Sl is a Pree az]-module with mp -

Sl Icm[z] also i n the K[z]-module sense. In fact, we define the

action of a z ] on Q by s c a h r multi-plication a s

0: I[[=] X n + n: (a, 01 f-3 7r.U

(ma E ~ [ z ] ,j = 1, ..., m).

The product of a with the components of the vector m i s the

product i n a z ] . W e write the scalar product o i the left, to'avoid

any cornsion with notation (3.6 ) .' It i s easy t o see that the module
axioms are verified; SZ i s obviously free, w i t h generators---:

+- j-th position, j = 1, ..., m.

(c) Q the action of the shift operator on i s represented

by .multiplication by z. !RI&s, .of course,. i s the nuin -reason_for

(4.4) i n the f i r s t place.

-ine~ducucjngthe isomorphf.~~
 R. E.Kalman

(d) Each element of r is a formal power series in z-l. In fact,

(4.4) suggests viewing d as an abstract representation of - t E Z;
-
hence we define

By (3.8 ) and (4.1), r(t) E KP for each t > 1 and is zero.(or

not defined) for t < 1. In general the sum is W e n over infinitely many
nonzero terms; there is no question of convergence and the right-hand side
of (4.7) is to be interpreted stictly aLgebraically as a formal power
series. Since Y(Q) is always zero (see (3.8)), .wecan say also .:

that

(e) l? is isomorphic to the K-vector subspace of ICP[ 1'(-Z 1

(form1 power series in i1 w i t l coefficients in K ~ )~onsistine
of a11 power series with 0 first terr?.

The first nontrivial construction is the following:

(f) r has the structure of a K[z] rliodule, with scalar
multi~licationdefined as

This product may be interpreted an the ordinary product of a power

-1
series in z by a poQmmialin z, .followedby the deletion of .

a l l terms containing no. negative powers of z. The verification of

the module axioms is straightforward.
 R. E,Xalman

(g) .f is K[z] homomorphism. This i s an inrmediate conse-

quence of the fact that f = constant (see (3.10))'and that multipli-

cation by z corresponds t o the l e f t s h i f t operators on Q and I'-.

(h) ' The Nerode equivalence classes of f are isomorphic with
Qlkernel f. !Chis i s an easy but highly nontrivial lemma, connecting

Nerode equivalence with the module structure on 0. The proof i s an

immediate consequence of the formula

In fact, by K-linearity of f, .(4.9) implies

~ ( U Q V=) f(w1ov) for all v E R

i f am3 only i f

f(8.d = f(zk*ub) for a ~ h

. >- o in -P. '

The proof of Theorem (4.2) i s now complete, since the l a s t

lemma identifies Xf as defined by (3.15) w i t h the K[z] quotient

We write elements of the l a t t e r as [uIf = w + kernel f; then

.. ..
it i s clear that Xf as a Uzl-module i s generated by [ellP ...., [em]*,
since Q i t s e l f i s generated by e , ..., em (see (4.6) ) . Note also
thpt the scalar product in Q/kernel f is

The l a s t product abom (that in Q) has already been defined i n (4.5).

The reader sllould verify directly that (4.10) gives a well-defined
scalar product.
(4.11) KEMARK. There is a strict duality in the setup used to
define f. From the point of view of homological algebra [KGLANE
19633, this duality looks ss follows. Since every free module is
-
projective, the natural map

ex3ibits X as th.e imge of a projective module. On the other

f
hand, there is a bijection between the set X and the st:.
f

fif is clearly a K[ el-subnodule of I' (with zm f(m) = f(z-m)),

and so Xf and $. a r e isomorphic also as K[c]-modules. It is
known that e M U W E 1963, page 95,
is an injective F.. . ~ l [
Exercise 21 So the natural map :if -t : ["If H ?(a) efiibits

..
Xf as a submodule of an injective mocl...:,:.' This fact i~ basic in the
co?-stsuctionof the :'transfer flm~ticn~~ associated with f (section 7),
but its rull implications are not yet understood at present.

There is an easy counterpart of Theorem (4.2) which concerns a

dynamical system given in "internal" form:

(4.12) PROPOSITION. The state set of every discrete-time,

fi'nite-dimensional, linear, constant -dynamicalsystem C = (F, G, -)
-
admits the structure of' a K[z]-module.-
FWXlF. definition.(see (1.10)), X = K" is already a
K-vector space. We make it into a K[ z 1-mdule by defining
(4.13) : K[Z]X?+ K?: (a, x)- T(F)X.

(4.14) CoME.IENT. The constructi~nused 5 . the proof of (4.12) i s

the classical t r i k o f studying the proper$ies of a fired u n e a r map

F: K? + via the Uzl-module stracture that F induces on

by (4.13). In view cf tbe canonical construction of Z provided by
f
Bopositian (3.16), the stzte set X can be treat& as a Uzl-
module irrespective ss t o whethe1 T i s constructed f r a f (X = xf)
or given a priori ss part of the specification of Z (X = %) ).
Thus

the K[z]-mdule s t r u c t u e on X i s a nice of W t h g the "e-uternalw

and the l'ini;ernall' definitions of a dymIr&calsystem. Henceforth we

shall t a l k about a (33 screte-time, linear, constant dynanical) system

Z somewhat imprecisely via properties of its associated K[ z]-module 3.

We s U no* give s m examples of p i n g &e-theoretic language

t o eKJress standard facts encomtered before.

(4.15) PROPOSI'SIIQN. -
If X is the state-module of C, the map .

-
FC i s g i v e r ~ 5X->
~ X: xr-.. z-x.

PROOF. This i s nbviaus from ( 4 . ~ 1 if X = T . If X = Xf = sf,

then we find tihat, by (1.1'7),

s b c e x(@] results frarr, i- put 9, x(1) results from icput -z-6 + ~ ( 0 )

 R. E. Kalman
and we get

So the assertion is again verified. 0

Now we can replace Proposition (2.16) by the much more elegant

(4.16) PROPOSITIO~. A system C = (F, G, -) is completely reachable

if and only if tht colurmis of .G p e ~ e r a t e
s.
PROOF. The claim i s that complete reachability i s equZva-

lent t o the fgct that every element x E - i s expressible as

In view of (4-lS), t h i s i s t h e sane as requiring that x be eqressible

this last condition i s equivalent to com_nlete reachability by (2.14). 0

(4.17) COROIUEtY. The reachable states of C are precisely

-those of the si~hm"Ct1'Ie of 5 generated.by (the columns of) G.

(4.18) P . . The statement that "C i s not conqIetely reachable"

-
~imolymeans that X is not generated by those vectors wf3ch make up

the mt.r.ijr G in thz specificatf on of the iuput side tt +;he system C.

 R: E, Kalman

It does not folLow that X cannot be f i n i t e l y generated by some other

vectors. In fact, t o avoid unnecessary generality, we shall. Lenceforth

assume that

X i s always f i n i t e l y generated over K[ z] .

From the system-theoretic point of view, the case when we need

infinitely many generators, that is, infinitely lnany input channels,

seems rather bizzare a t present. .

(4.19) PROWSITION. The syshsrm Xf i s completely reachable.

PROOF. Obvious from the notation: a s t a t e x = [ Elf

i s reached by 5 E a. n

(4.2Q) PROPOSITION. The system Xi i s completely observable.

PROOF. Obvious feom Lema (h) above: q( ["If) = f (u) = 0

iff o i [0If, which ws that the only unobservable state of Xf

Let us geceralize Vas Ust result t o obtain a rm21.XLe-thecretic criterion

for complete observability. There are two technically different ws of

doing this. The first depends on the observation that the '+iualI1of a

subnodule (see CcroUary (4.17)) i s a quotient module. The second defines

observability via the "dual1system ( F Kt, ) associated with (F, -,H).

Consider a meal system -,H) and the c o r r e s ~ n d i n g
= (F,.
-
C

K[ zl-module % and K-hommrphisrn H: 3 + Y .'K We can extend H

 R. E.Kalman
t o a K[z I-horromorp~sm (look back a t ( 2 .8)) by setting

From Definition (2.19) we see that no nonzero element of the quotient

module %/kernel ?i i s unobserwkle. Hence, by abuse of lengmge, we

can s a y that %/kernel i s the sodule of observable states of E.

Thus we arrive at was* the comterparts of (4.16-17) i n the follow-

ing language:

(4.21) PROPOSITION. A system C = (F, -,

H) .is comletely observable
if adonly i f the quotient module %/kern21 is isomorphic with t.
(4.22) CCROLLARY. The observable states of C are t o be identified

with the elements of the quotient module %/kernel E.

(4 -23) ~ T O I C G Y . The preceding considerztions suggest v i e h i -

a system Z as essentially the sane ltthbg" as a module X. Strictly

speaking, however, !mowing Z = (I?, G j H) gives us not only &

Y
= %
(see ( 4 . ~ ) )bct also a quotiant module! (over kernel E) of a sub-

module {thet generated by 0 ) of + that i s

If $ 5 we say that 5 i s canonice1 {relative t o the given 4 H) .

To be more precise, l e t us observe the following stronger version

of (4.19-20) :
 R. E. Knlman

(4.24) CORRESPONDENCETHEORE2~1. Thereisabijectivecorrespondence

between K[z]-homomorphisms f: S! -, F and the equivalence class of

completely reachable and completely observable systems C modulo a

basis change i n 3.
Detailed discussion of t h i s result is postponed u n t i l

Section 7.

A s t r i c t e r observatioa of the Itduality principlet1 leads t o

(4.25) IEFINITION. The K-linezr dual of C = (I?, G, H) -

is
C* = (F', Ha, G I ) ( 8 = matrix t r a n ~ ~ o s i t ~ o nThe
) . s t a t e s of

C* a r e called costates of C.

The following fact i s an M e d i a t e conse.yence of this definition:

(4.26) PROPDSITIOI?. The s t a t e s e t X&* of E* may be given the

structure of K[c'] module, a s follows: ( i ) es a vector s p c e H2,

i s the dual of 5 regarded a s a K - ~ c t cspace, ( i i ) t h e scalar

product i n s* i s defined by

(b.26A) RWWC. %e ccanrd define %* as H o ~ [ ~~[ z ~

l ) equal
( ~ to

Uz]-linear dual of 5, because every^toraion module M over an i a t e g r s l

anmain D has a trivial D-dual. However, the reader can verify (using

the ideas t o be developed in Section 6) t h a t *.X, defined above i s iso-

) ~SOURBRRI
noqhic with ~ Z l ( ~ , ~ ( ~ See ~ ] ) [Alg&bre,
. Chapter 7

(
2
' dd. ), Section 4, &.81.
 . R. E. i\alman
Now we v e r i a easily the follo-wing dual statements of (4.16-17) :

(4.27) PROPOSITION. A system C = (F, -,8) i s completely observable

i f and only i f HI jenerztes %+.

(4.23) C O O : The observable COstates of C* are precisely

. ---
th5 reachable --;+.Z;eb of C*, that is, those of the submodule of -
* generated bx HI.

We have eliminated the abiise of languzge incurred by t a l k i l g

about "observzble sta%estrthough introduction of the new notion of

"observable COstatesl'. The full explication of why t h i s i s necessary

(as well as natural) is postp~nedu n t i l Section 10.

The preceding simple fzcts depend only on the notion of a module

and are immediate once we recognize the fact that F may be eliminated

from statenients such as (2.8) by passing t o the mdule induced by F

a ( 1 ) But module theory yields many other, l e s s obvious results

as well, which derive mainly fram the fact that . I(lz] is a principal-

ideal domain.

We recall: sn element m of an R-module W (R = arbitrary

commutative ring) has torsion i f f there i s a r E R such that

r - m = 0. I f t h i s i s nf>tthe case, m is free. Similarly, M i s
said t o be a torsion module i f f every element of M has torsioil.
M i s a free module i f no nonzero element has +,orsion. I f L C M

is any s ~ b s e tof K, the annihilator t of L i s the s e t

A . = (r: 1-01 = 0 for a , l l 1 E L};

it follows imediately that t i s an irieal i n 2. Note also that

the statement I'M i s a torsion mdule" does not imply i n general

that AL i s nontrivial, that is, A,-, # 0. (~ounterexample: take

an M which i s not f i n i t e l y generated.)

Coupling these notions with the special f a c t that, for us,

R = S z l , we get a number of interesting syste-ntttheoretic results:

(4.29) PROPOSITION. Z -
i s f :nite-dimensional if and only i f T
i s a torsion Uzl-module.

COROIURY. If $ i s free, C i s infinite dimensional.

PBOOF. We r e c a l l t h a t "C = firslte-dimensiodll' i s defined

t o be 5 = finite-dimensional as a K-vector spce". See (1.18).

Sufficiency. By assumption X i s - f i n i t e l y generated

by, sqr, q nonzero elements ... x

Q
of (which are not-

necessarily the columiis of G) . Hence

Since H z ] is riprincipal-i&al domzh, each of the A i s a princi-

xJ4
p a ideal, Says ~ f [ z l ~ 5 t hr E a z l . I f
3
t
is a torsion mdule,
then d e g r = n > O
3 S
f o r a l l j=l, .-., Q. Forotherwise rS
is either zero (and then x i s free, which i s a contradiction) or
3
a unit yvhich implies x = 0- contr y t o assumption. .Hence we can
j - ,

replace each expression

 R. E . Kalman

by the simpler. one

9
X = J 1 & [vJ (mod rj)]-x J'

which shows t h a t $, as a K-module, i s generated by the f i n i t e s e t

Necessity. Let 4fF be the minimal polynomial of the map

F: x H Z-X. If % i s finite-dimensions:- as a K-module, deg JrF > 0.
This means (by the usual definition of the minimrrl polynomial i n matrix

theory or more generally i n l i n e a r algebra) t h a t tF - annihilates every

xC % so that YE is 3 torsion IS[ 21-module.

Notice, from the second half of t h e proof, that the notion of a

m i n i m a l polynomial can be extended from K-linear algebra t o KEzl-modules.

I n fact, the same argument gives us a l s o the well-known

(4.30) PROPOSITION. Every f i n i t e l y generated torsion module M

over a principal-ideal domain R has a nontrivial mininiil D ynomia7,

qM given by %=

(4.31) COROLLARY. If KK[z]-module X i s fs'nitely generated with

q penerators a ~ minim~l
d polynomial qX, then

'dim X (ss K--rotor s w e ) <- q-deg IX.

(4.32) FEMARK. The f a c t $bat Zf + s completely reachable and i s

therefore generated.by m vectors allows us t o estimate the aimension

of 2, by (4.11) knowing only deg rU but without having com2uted

xf
Xf itself. ( ~ n o v i n Xf
~ explicitly means knowing F: x H z.x, etc. f

I n other words, the module-theoretic setup considerably enhances the

content of Proposition (3.16). Guided by these observations, we s h a l l

develop i n Section 8 explicit slgorithms for calculating dim Lf directly

fiom f without f i r s t ha-~ngt o compute F.

(4.33) PROPOSITION. i s a free K[z]-module, no s t a t e of

C can be simultaneously reachable ant5 controllable.

PROOF. We r e c s l l t h a t = free" means that % is

(isomorphic t o ) a f i n i t e sum of copies of K[z] . Suppose f o r

simplicity that % = K[ 21 Then x = reachable means that x = 5 -1

for some 5 E K[ z] . Similarly, x = controlLFLble means that

z"'I -r + 0.1 = 0 f o r some m E K[ z ] . Hence i f x has both properties,

This shows t h a t 1 i s annihilated by 5&, the input 5 followed

by cu, which contradicts the assumption t h a t i s free.

The most .important consequence of Theorem (4.2) i s due t o t.'. .

fact that through it we can apply t o iinear dynamical systems the well-known

(4.34) IilNUWXm STRUCTUW TIYEOREX FOR FINITELY ~~ MODULES

A PRINCIPh IDEllL DO^^ R ( ~ n v a r i a n tFactor Tllesrem for Modules).

Every such moaule M wPkh m generators i s isomorphic t o

 R. E. Kalman

where the R / V ~ =re

~ quotient rings of R viewed as modules over R,

-
the qi (called the invariant factors of M) are uniquely determined

g M up to units in R, qi 1 qiq1? i = 2, ..., q, and, as usual, R'

denotes the free R-module with s generators; finally, r+ s <- m..
Vasious proofs of this theorem are referenced in KAIWU, FALB,
and ARBIB [1969, page go], and one is given later in Section 6.
Note: The divisibility conditions imply that M is a torsion
module iff s = 0 and then $N--
One im~ortantconsequence of this theorem (others in Section 7)
is that it gives us the most general situation whet % is not a
torsion module C. For instance, combining (4.33) with (4.34), we

(4-36) PROWITION. d system cannot be simultaneously completely

--
reachable and completely controllable if its K[ 21-module X has any
=-dimensional components ( i. e ., -
s > 0 in (4.35 j ) .
(4.37 1 REMARK. Although our entire devel3pment in this section may

be regarded es a deep examination of Proposition (2.14), most of our

comments apply equally well to (2.7), since both statements rest on
the same algebraic condition (2.8). In fact, the only remaiiu
thing to be lta+lgebraizedwis the wtion of vcontinuous-timew. We
shall not do this here. Once *is last step is taken, the algebraization
of the Laplace transform (as related to ordinary linear differential
equations) w i l l be complete.
 5 , CYCLICITY AND RELATED QUESTIONS

We r e c a l l that an R-module M (R = a r b i t r a r y ring) i s cyclic

i f f there i s an element m E M such t h a t M = IZm. [ ~ would

t be
.
,

b e t t e r t o say t h z t such a module i s momgenic; generated by one

element m.]
If M i s cyclic, the map R -, M: r H r-m i s epimorphism
and has kernel Am, the annihilzticq ideal of m. This plus the

homonorphism theorem gives the well-.ho-m

(5 -1) PROPOSITION. Every cyclic R - e 1.1 with g e c e r ~ t ~LOr

i s isomorphic with the quotient rinq X / A ~ vie.vst 9 s on 2-mdlee.

This r e s u l t i s much m x = i - ~ t e r e s t i n g'vihen, e s i n our case, R

i s a o t only conmutative and a principal-i8eal domin, but specifically

t h e polynomial r i n g K[ z1.
So l e t X be a cyclic K[z]-module x t t h ger~erator g and l e t

A = $ g ~Z]
[ where is the min-1 or snnihilatinp, polynomial of
g . g
g. By c o m t a t i v i t y and cyclicity, A =
g
5.
Hence Y i s a minimkl
g
polynomial a l s o f o r X. Write $ = riiX = q. I n view of (5.l)#
g
v x[z]/$K[ el. Let us r e c a l l sone features of tihe ring K[ z]/.t'i([ z]:

(i) I t s elements a r e the residue cins:es of polyno~dals u (mod o ) ,

.
n E ~ [ z ] Write these a s [rr] or [TI*. M u l t i p l i c e t i ~ nis d e f i e d as

[a]-[u] = fml.
(ii) Each [ a ] i s either a or e divisor of zero. I n Sect,
r
I
[a] i s a unit i f f (n; 9 ) = greatest 'cw.%n divisor of 71; r/r i s a
 R. E. Kalman

unft i n ~ [ z ]( t h a t is, (n, 9 ) € K) . Then

so t h a t [oj i s the inverse of [TI. On the other hand, i f

(n, JI) = C f unit i n K[z], then both :?TI and [q/Q] a r e zero

divisors since [n].[JI/@] = [(T/Q)JI] = 0.

(fii) If JI i s a prime i n Kl z] (that' is, an irreducible poly-

nomial with respect t o cof:fficients over the ground f i e l d K), then

by ( l i ) K[z]/JIK[z] i s a field. This i s a very standard construction

i n algebraic number theory.

Since it i s awkward t o cornpute with equivalence classes [TI, we

shall often prefer t o wcrk k i t h the siandard representative of [n],

namely a pobnomial % of l e a s t Z e g r ~ ei n [n]. 5 i s uniquely deter-
&.zd by [TI &d the condition deg ? < deg q. Heaceforth * dll

always be used i n this sense.

The next two assertions a r e immzdiate:

( 5 -2) PROPOSITTQN. -
K[ z ]/VICE z ] as a !;-vector space i s isomorvhic
t o the I(-vector spece = (x € K[z]: deg Z C n =. Ecg W J .
.
K[z]/JIK[z] i s also isomorphic t o @(n) - .
a s a K[z]-module, .-provided
we define the scalar product i n @ln) (riel) .
it n j
hC

.( 5 4 PROPOSITION. -
If 3 j,s cyclic with minimal po-c)mial $,

-
then dim C = deg v.
 I&khg back at Theorem (4.36), we see t h a t the most general

a d - w e is a direct sum of cyclic Uzl-modules. By combining

(5.3) and (4.34) and using the fact that dimension i s addifive under
direct summing, we can replace (k31) by the follow-. exact result:
..- .

(5-4) PROPOSITION. If 5 i s a torsion rnodsle w i t h invariant

factors $lJ ..., f9 - then
d i m C = deg $l + ...+ d.eg *q.
A siruple but 'LZgkly u s e m consequence of cyclicity i s the

so-called control c a n ~ n i c a lform I

-, FAI;B, slid ARBIBj 1969,

page 441 f ~ arcompletely reachable pair (F, g) where g i s an

tr X 1 mtrix. We shall now prpcee&'to deduce t h i s result. ..

Clbserve f i r s t that "(F, g) coiiletely reachable1' is equiva-

lent t o Ifg generates XpJ the module induced by F via (4.13) . It Let

then i s the characteristic ( a d also the) minimal polynomial for

$. [This i s a well -lm.!om~.fact of module theory. See for e x q l e

mJFALB, and L I B [196gJ Chapter 10, Section 71 for detailed

discussion. 1 As. in KALMAN 119621, consider the vectors

*
 e = g = l o g = $)(z)

i:
-g,
n
+q-g = (2) t.z)-g,
=n-1 = 2-g

(5.5)

e
1
= zn-1.g+4Zn-2.g+ ... = (n)(z1-e

in 5. [Fez cansiskacy, (-1) b) = s(3.1mese vectors are

easily seen t o be linearly independent over K. They generate
since 5 =dn) a v e s e o i t i (2)) - Hence

el, ---9 en are a basis f x 5 as a K-vector space. With

respect t o this basis, the K--msm

i s represented b y e s matrix

[This i s proved by direct computation. In paxticulsr, it i s

necessary t o use the fact that

 R. E. Kalman

Note that the last row of F in (5-6) consists of the coefficients

f . By definition, g = e
n'
Xeience g as a column vector in
has the representatioq

Conversely, suppose @, have the matrix representation (5.6-7)

with respect to some basis in I?.Then (by direct computation)
the rank condition (2.8) is satisfied and therefore (F, g) is
campletely reachable in both the continuouc-time snd discrete-
time cases (Proposikions (2.7) and (2.16)).
We have nos proved:

(5 08) PROPOSITION. The pair (F, g) is completely reachable

if and orly if there is a ksis relative to which F is given by

(5.6)s g (5.7).

(5.9) Cm0WY. Given an ar3itraq n-th degree polyriamial

n
h(z) = z + B1z
n-1
.
+

.
L -
+ B, in lS[ [a], -
K = arbitrary field. There
exists an n-vector I such %at h= s-ggif- and only if the
r (F, g) is completely reacha'ole.
 R. E.Kalman

PBOOF. Suppose that (F, g) i s completely reachable.

With respect t o the same basis (5.5) which exhibits the canonical

forms (5.6-7), defiae

men verify by direct camputation that h = s-gJ,.

Canirersely, suppose that (F, g) is not completely

reachable. Then, recalling Proposition (2.12) (which is an

algebrdc eollseqpence of (2.8) and hence equally vslid for both

continuous-time and discrete-time), dim X2 >0 and so is also

deg 5
22
. Since X1 fs an F - i n m i a n t subspace of X = I!?,
the polynsmial i s independent o f t h e c h o i c ~of basis i n
1
-I
II
I?and the same is true then also for
s22 = Y?Fu-(=
pa~ticular, does not depend on the a ~ b i t r a r ychoice of

X;,
".,
i n s a t i s m the condition X = ILL @ 3.) I n view of (2.12),

we have for a l l . n-vectors 1,

This contradicts the claim that h = $-gl, is true for any A

with suitable choice of 1. 0

In view of the importance of this l a s t result, we shall

rephrsse it ir. purely module fieoretic terms:
 R. E.Kalman

(5.~) !DEDFGH. -
Let K be an arbitrary f i e l d and X a cyclic

K(z]-module with generator g an6 minimal polynomial X of degree

n. There i s a bijection between n-th degree polynomials

A(z) =
n
2 + p1zn-1 +
... + B, -
K[z] and K-3omom0rphisms
1: IP -, e: ~ ( j ) - ~1.-g
u (j = 1, ..., n 9 x(j) defiaed
3
-
as i n (5.5)) such that A i s the minim1 polynomial ?or the

new module structure induced on X by the map z,: x I+ z-x - f ? ( x ).

Note that in (5. El) l(x) corresponds t o gf?I x in (5.10).
The nap I in (5 .Kt.) defines a control law for the system
C = (I?, g, -) corresponding t o the module X. The passage from
z t o z, i s the module-theoretic form of the well-known open-loop
to closea-loop transformation used i n classical linear control theory.

PROOF. Since the vectozs x . x form a

basis for e, 5 i s clearly a well-defined K-hamomorphiam. We
t r e a t 1 formaUy as an element of ~ i z ] (that is, an operator
on X i s a K-vector space), by writing 1.x = 3(Ewg), where
% represents the equivalence class [ 5 1 = ( 5: 5 g = x) . Unless
identically zero, f! i s never a a21-hnmnmnrphism and therefore
4 does not commute with nonunits i n ~ [ z ] .
Define f?J = B J - a
J'
j = 1 . . We prove f i r s t

that t h i s choice of . i implies h(3) ( a - k?) = x(')(z) for

j= 1 .., n + 1. Use induction on 3. By definition,
-A (a -) = x .
(z) ;In the general case,
 R . E. Kalman

= [(z - l ) x(3) (2) -I-Bjleg (inductive hypothesis) ,

;[ a ( J ) i z ) + B~ - l .J1 - g (def. of a),
; [ z ~ ( j ) ( z )+ a.1.g
J
(def. of a .),
J
-- x(~+l)(z)og (aef. of ~ ( 3 ~ ) )

It follows (case j = n + 1) that A annihilates X

regarded as a K[ z*] -module. On the other hand, the
(1)( 2 ) ., A(n)(2,) *g is a basis for X as a K-vector

space since x(')(z) -g -..-, (z! -g was such a b s i s . SO x

fs cyclf c with generator g slso as a K[ z, 1-moduie. Hence
C;i Fxgositions (5-1-2)the amihilatiq ideal of g' with respec:t
t o the K[z,]-msdule structure cannot be generated by a polynwial
of degree less than n, that is, h i s indeeB the minimal poly-
nomial with respect t o 2., B e correspondence A t, J i s obviously
bijective, 0

The proof ixnediateLv implies the following

(5 .El)
-
COROLLARY. Let x = g -g be any element of X viewed

- --
as s K[z] -module.
7

Then x has the repr5sentation S,*g

7
- with
_e

respect t o the K[q]-module structure on X, -

where 5 and 5,
are related as
 R. E. Kalman

So the open-loop/closed-loop transformation is essentially a

change in the canonical basis, provided X is cyclic.
It is interestir;g that the x(') have long been knom in
Algebra (they are related to the Tschirnhausen transformation
discussed extensively by WEBER [1898, $46, 54, 74, 85, 961), but their
present (very natural) use in module theory seems to be new.
*Theorem' ( 5 . ~ )hay be viewed as the central spe-
d a l case
--
of Theorem A of the Introduction. Let us restate the latter in
precise form as follows:

(5.l.3) THEOFWZ. Given an arbitrary n-th degree polynomial

h(z) = zn + plzn-l + ... + f3, & K[ z 1, K = arbitrary fiela.
There exists an n X m matrix L over K such that s-GL,
h =

-
if and only if (F, G) is cor~letelyreachable.

For same the, this result had the stat;- of a well-known folk theorem,
considered to be a straightforward consequence of (5.9). The latter
has been discovered indepelldently by many people. (-Ifirst heerd
of it in 1958, proposed as a conjecture by J. E. Bertrar; and proved
soon afterwards by the so-called root-locus method.) InGstG, the
passage from (5.11) to (5.1,7) is primarily a tzchniczl problem. A
proof of (5 -13) was given by L A I i I O P [1$4] md su5seqaerrtly
simplified by K O m - 1 [1967]. Tne first proof was
very long, but the second proof is also unsatisfactory; since
it-dependson arguments using a splitting field of K

---------------
*The material between these marks was aeded after the Surpmer
School.
 R. E.Kalman

and fail when K is a finite field. We shall use this situation

as an excuse to illustrate the power of the module-theoretic
approach and to give a proof of (5-13) valid for arbitrary fields.
The procedure of LANGENKOP asdlIQNHAM zests on %he following
fact, of which we give a module-theoretic .proof:

(5.14) LEMMA. && K be an arbitrary but infinite field. &t

&

F be cyclic* and (I?, G) completely reachable. Then there is

-
an -
m-vector a E I? such that (F, ~ a ) is also coz~letely
reachable,

We begin with a simple remark, which i s also useful in

reducing the proof of (5.13) to Lemma (5.18).

(5.15) SUBUMMA. Every submodule of a cyclic module over a

principal-ideal domain is cyclic.

PROOF OF (5.14). We use induction on m. The case

n = 1 is trivial. The general case amounts to the following.
Consider the submodule Y of X = 5 generated by the columns

8p *J G,l of G. In view of (5.15), Y is cyclic. By the

inductive hypothesis, we are given the existence of a cyclic
generator of Y of the form 5 7 4gl+ -* + am-l*q-l>
ai E'K.
We must prove: for suitable (2, @ E K the vector a - g y + pee,
is a cyclic generator for X.

*Of course, I3h.s means that the K[z]-module $ (see (4.13))

is cyclic.
 By hypothesis, X has an (abstract) cyclic generator
By cyclicity ue have the representations
%.

Hence our problem is reduced to proving the following: for suitable

a, p E K the polynomisl q + & is a unit in K[Z~/$K[Z]. This,
in turn, is equivalent to proving

where QlY ..., % in K[Z] are the unique prime factors of

5. Let " mean the reprasentative of least degree of equivalence
classes mod Qi. Then no pair ( ) i=1 ... r can be
zero. For if one is, then oil (s, q, p), that is, y Q i annihilates
the submodule XI = K[z]gy + K[Z]Q whence XI is a proper sub-
module of X, contradicting the fact that (F, G) is completely
'4 #
.
II

reachable. If all the Pi are zero, then every 01 SO 9

is a unit in K[ z]/]l~[ z], and gy is already a cyclic generator.
So let
..
a = 1. Then the condition \ + bi = G eliminates at most
r values of 6 from consideration. Since K is Infinite by
>

hypothesis, there are always some g which satisfy (5.16). o

Anessential part of the lemma is the stipulation that a E K?.

The hypothesis "F = cyclic + (F, G) = completely reachable" means that
 K. E. Kalman

that is, the lens is tfivi8Uy true for some a E $[z] since

g~ = Ga. But since we want a E K, -there must be interaction

between vector-space structure and module structure, and for this

reason the lemma is nontrivial. 'As a'matter of-fict, th= lenms is f d s e

when K = finite field. The simplest counterexanple is provided

when (5.12) rules out a single nonzero value of g, thereby ruling

-
out all p.

(5.1) C-E-LE. Let K = - -

g22, that is, the ring of
integers modulo the prime ideal 22.'
- Consider

Notice &at Xp = 5 (3 X2 @ X3 (as 8 K[ z]-module), where the

minimal polynomials of the direct summnds are

A l l these factors are relatively prime, ( 5 XP 5) = 1, hence

X is cyclic. ,Votice slso that gl generates XI @ X while g2
3
generates X @ X A cyclic generator f ~ rX is
2 3-
 R. E.Kalman

A simple calculatiori gives

Conditions (5 -16) are here

a-i+ p0o # o (ma%),

a-0+ 6-1 # 3 (mod x ~ ) ,

These conditions have no solution in -gg.

-
At this point, the followi~gis the situation concerning
Theorem (5. U) :

(1) Its cowiterpart, Theorem A of the Introduction, was

claimed to be true i3 the continuous-time case under the hypc'.lesis

of complete controllability.

(2) In the discrete-time case (5.13) with the preceding

hypothesis Theorem A is false, because-of the counterexample: the pair

(3 = n5lpotent, G = 0:) is completely controllable, but evidently

is independent of L. Kowever, in view of (5.11) ,.Theorem

S-GL~
(5.13) might be true also in the discrete-time case if "cmglete

controllability1 is replaced by Itcanpletereacha,bilLtytt,this modi-

fidtion being tmterial in the continucius-time case.

(3) Fkcause of (5.17), we might expect that a theorem like (5 -13)

is false for an arbitrary field K.
 ( k ) I f our generdl claim that reachability properties are
reflected i n module-tkcoretic propertizs i s true, then (5.13)

-
skould hol6 vi-khuut assmptions concerninq Kj kcause the principal
module-theoretrc fact, tkat IC[z] = pl incipal ideal domain, is

i n d e w e l l t of the specific choice of K.

We nov proceed t o establjsh !%ec.rem (5.13). That is, special

~ v p ~ t h e s eons K w i l l turn out t o be irrelevant.

EmF C2 [5.l3). Necessity i s proved e w c t l y as in (5.8).

Sufficiencx wil; +'32lo'r~&y induction on m, once we hzve m v e d it

in the special c2se rn = 2:

-
IBZJA. -Let Ii be m. arbitrary f i e l d -
l e t X be a

(of the t p defined i? (5.111 such tbzt if z =z

, -I induces a

a;lz,l-n~dule structure on X then X - i s cyclic wj..tfr respect t o t h i s

stficture er,d i s g e ~ s r a t e dby either % + g2 or 5.

mr'. let Y = lt[z]g,
- and Z = a z l ~ .

--
CaseI. Y ~ z = O , Wtis,
. .
X=Y@Z. In(5.u)

take an I such Wt J(x) = 0 -for all x f 2. Replacing z by

z =z
, -I w i l l change the Hz]-module structure on Y bat pre-

serve that on Z. Further, choose I so that the new minimal poly-

nomial h on Y is prime to the urchmged m i n i m a l p o w a l 5 = X
z.
on Z. Thus there exist polynmWss V, a such that vh + oX = 1.
BJ hypotI:esis, every x E X bas "be representation
 Nov veriiy that

E s c e gT .L t. is. in-1eed c cyclik - ge~:eratozf94 X a s a

.L -2
. z* ]-md*de.
=

there i s e 5 i I(I z: such thet w = Em% end thereiors, by

cyA.2city of Y, there is also a q E I([z] such that E-g2 = w =

Take SZZE w # 0. Then if q = writ (mad X;) we are done because

-
-
r, 'E-+ gemistes Y, and so Z = X. 51the nontr5viol case,

q # unit (nod )k). TO show: there i s a suik2ble neu module structure

on X sich t'nat & = M t (mod X,), X* being the ~&inai pa%--

n m i a l of X as a -IC[ z, ]-module.
The =in facts we need are the fcllcwi.iug: .

(5.19; . m -
Let X be a fixed clenent of a(jz] :%tt:h
.-

deg X = g -- ~ s t r i xof
Fx the c o ~ p n i m X given bE (5.6),
.
"., :
tkce cyclic rm6uIe induced by Fp +g a y c l f c gemrator 02'
. .

% .- Then rl € ~ [ z ]i s *mit reod11Lo- c" X

7'-

if ail&c-dy if . ..- g -
fs
also a cyclic gmerator of .-
'X

PROOF .. Obvious.
 R. E.Kalman

(5.20) SUBTB.P.iA. .; Same natations as i n (5.19). -

Write

-
Then 2s a mit KG~JU~S X i f and only i f -
(5.21) det (y, F$', ..., ql:-s)
0,

-
where g is the colmm vector

. S b c e X(1) .> , x("' is the basis for ttr

K-vector space of 8U p'jSq7-t'daJsof degree < T;, the n-tuple

(ql, .- ., In)
7 is Ujiqyely deteAWedby g. By defkition FX

is the mtris represe~tirgthe miMu2.e operator z: x e z - x relakive

to the special basis el, , en in ...given by ( 5 -5). Simi%Lrly,
using one of the n : d a e ~Aons,vs. verifE- th&t

ir other words: the :nmerical uecto-8- is. 22) represents the abstract
vector 6 ' ~in rel&tive .to t2ie sane basis el,
. - .. -
.
..., en' Xect~11
\
that ?wg generates iffX, (F,,,
? j ( ~ ~ ) gi)s conplete reachable.
'X
By (2.7) the l a t t e r condition is eqd.valefit t o ((5.21). The r e s t

f'ol-hws fron (5.15). [3

(5 -23) S-CJQ. Sme r.ot&tions as ir. (5 .s) (5.20). Given

any nonzero nmerical n-vector (5.22), there exists a polynomial X

such t h ~ (t5 .a)

is sat5sf ied.

PROOF. Let qr be the first nember of the sequence of

nmbers
-
ql, q2, .- . which i s nonzero. Write

an& determine %e first r coefficien+,s of X 'by the rule

(Giace all nmbers belong t o a fiel6, tke reqdred values of

ar, -*-, an eidst .) Now check, by congutation, t'nrt th2se con5itiom

redux the mtrix i n (5.21) t o the direct sum of two trianplzr

matrices, each with mazero elements on i t s diagonal.

t can
In view of (5.12), it follows from these facts t h ~ xe
alws choose a new -Xy = Xt such that % = unit mad X+.
 R. E.Kalman

The proof of Case 2 i s n ~ yet

t cozpiete, however, beczuse

we must s t i l l extend th2 az,l-m5ule structure from Y t o X. ThZs

i s easy. Write first Z = W 6, 2' znd then X = Y @ z', where the

direct sum i s noii with respect t o =e K-mciule structure of X. Extea

j frca .ir t o x par s%ttir,s&i
Z t = O. -- we ht7e
~;ow ni-1
.Z 3 . ; ~

p0lynaolra.l defined over X Since z = zi on Y,

, 5=\ By

(5.12), 6 is replaced by s- 6 such tkat

that is, our previuas re2reseatation of w 0 i n GI induces a

sinilw representation with respect t o the nev K[z,]-&de structuze

on XI Since rl, i s a unit owiiulo X,i we can write

% = 1+ T X ~ , with O, r E K[z*]-

By (5.24), we b v e , with respect t o the K[z,]-structrrre,

This' proves thzt % generates both Y and 2; tiat is, g

, is
a c;yclic generator f o r X endoxed k i t h the -1-structure. The

proof of (5.18)- I s now complete. .

 It should be clear that Theorem (5.13) i s not a purely modde-

theoretic result, but depends on the interplay between module theory,

vzctor-spaces, and elimination theory (via (5.21)) . For instance,

the fact that l? c m be extended from Y t o X, which was needed

i n the proof of Case 2, i s a typical vector-space argument.*

There a r e many ope8 (or forgctten) results concerning cyclic

modules which are of interest i n system theory. For instance, it

i s easy t o show that an n X n r e a l matrix i s cyclic i f f a certain

polynomial Y E g[zl,
- ..., zn2] i s nonzero at F; the polynomial

Y is roughly aniLlogous t o the p o l y n d a l det i n the same .ring,

but, u n l i k e i n the b t t e r case, the general- form of Y does not seem

t o be known.
We must not terminate this discission without pointing out

another consequeace of cyclicity which trmscends the module frame-

work. Since X = cyclic with generator g i s isomorphic with

~[z]/XgK[z], it i s clear that 7; alro has the structure of t h i s

conmutative ring, that is, the product i s defined as

If X= irreducible, then X i s even a field. Hence, i n particular,

g
X has a galois group. 1Jo one has ever given a dpamical interpreta-

tion of t h i s galois grou?. In other words, there are obvious algebraic

facts i n the theory of dyrmical systems which have never been examined

from the dynamical point of view. For same related comments i n the

setting of topological semigroups, see DAY an6 W U C E [1967].

 ..
R. E Kalman

6. TRANSFER FUNCTIONS

(6.0) PREZil4ELZ. There has been a vigorous t r a d i t i o n i n engineer-

ing (especially i n e l e c t r i c a l engineering i n the United States during

1940-1960) t h a t seeks t o phrase a l l r e s u l t s of %he theory of l i n e a r

constant meal systems i n the langpuage of the Laplace transform.

Textbooks i n t h i s area orten t r y t o motivate t h e i r biased point of

view by claimix?g that "t'ae Laplace transform reduces the analytical

problem of solving a d i f f e r e n t i a l equation t o an algebraic problem".

When directed t o a mthematicizn, such claims a r e highly misleading

because the lcathematical ideas of t h e Laplace transform a r e never i n

f a c t used. The ideas which -

a r e actuclly used belong t o c l a s s i c a l

complex fmctlon theory: properties of rational fbnctions, the

partial-fraction expansion, residue calculus, e t c . More importantly,

the word "algebraic" i s used i n engineering i n an archaic sense and

the actual (modern) algebraic content of engineering education and

practice as relsted t o l i n e a r systems i S very nieager. For exGple,

the crucial concept of the transfer function i s usually introduced

via heuristic arguerits based on l i n e a r i t y or "defined" purely formally

as "the r a t i o of Laplace tr=sforms of the output over the inputtt. To

do the job right, and t o recognize the transfer function a s a natural

and purely.algebraic gadget, requires a Grastically new point of view,

which i s now a t hand a s the machinery s e t up in Sections 3-5. The

essential idea of our present treatment was f i r s t 9ublished i n

mmi Ilg65bl.
 R . E.Kalman

The first g q o s e of this sectlon is to give an intrinsically

algebraic definition of the transfer fbnction associated with a
discrete-time, constant, linear input/output map (see Definition (3.10)).
Since the applications of transfer functions are standard, we shall not
develop them in detail, but we do want to emphasize their role in relat-
ing the classical invariant factor theorem for polynomial matrices to
the corresponding module theorem (4.34) .
Consider an arbitrary K[ z]-homomorphism f: R + I' (see lemma
(g) following Theorem (4.2)) . object" f is
Then as a lfmathematical
equivalent to the set E f(e 5 ), i = 1, ... m e
5
defined by (4.6)),
since

h he scalar product on the right is that in the K[ 21-module , as

defined in Section 4.) By definition of I', each f(e ) is a ' f o r d
5
power series in z
-
' with vanishing fir-& term. We shall try to
represent these formal power series by ratios of polynomials (which
we shall call transfer functionsx.) and t3en we can replace formula (6.1)
-
by a certain specially defined product of a ratio of polynomials by a

polynomial. Some algebraic sophistication will be needed to find the

will consititute a
correct rules of calculations. These llruleslf
rigoroos (and simple) version of Heavisiders so-called wcalculuslf.
There are no conceptual complications of any sort. o ow ever, we are
dodging some difficulties by working solely in discrete-time ) .
*This entrenched terminology is rather unenlightening in the present
algebraic context.
 .
R'. E Kalman

Let Xf = Wkernel f be the s t a t e s e t of f regarded a s

a K[z]-mo~l.il.e. We assume t h a t Xf i s a torsion module with nontrivial

minimal yolynomial Jr. Then, f o r each j = 1, ..., m we have

By definition of the module structure on r, (6.2) means t h a t the

ordinary prodilct of the power s e r i e s f ( e .) by the polynomial Q is

J
a (vector) polynomial. ~ e n c e(6.2) i s equivalent t o (n~tation:

no dot = ordinary

---IntuLtively,
--.- we can solve this equztion br writing f (e .) = 8 ./Q.
J J
There a r e two w a p of making t h i s idea rigorous.

Method 1. Define

as t h e formal division of Q
3
by $ i n t ~ascending powers of z -1 .
Check t h a t the coefficient of z0 i s always 0. Verif'y by computation

t h a t the power s e r i e s so obtained s a t i s f i e s (6.21).

Method 2, Multiply both sides of (6.21) by z ' ~ . Write
( z ) = ( z ) and ;jj(z-') = ;%(z). Then 3E ~ [ z - l ]C K[[ z'l] I
and (6.21) becmes

Moreover, the O-th caefficient of $ i s 1 (because of the convention

 .
R. E Kalman

t h a t the leading coefficient of v i s 1)) hence $ 1s unit i n

K[ [z'll] and therefore

Note t h a t (6.3) and (6.3 ' ) actu.ally give s l i g h t l y different defini-

tions of f(e.), depending on whether we use a transfer function with

J
-1
respect t o the variable z or z , ( ~ 0 t hnotations have been used

in the engineering l i t e r a t u r e . : For us the formalism of Method 1 is-

preferable, (The calculations of Method 1 c m be r e d x e d by Method 2

t o t h e better-known calculations of the inverse in the ring K[ 1.)

Sumrriarizing, we have the easy but fundamental result:

(6.4) EXISTENCE OF TRPJXSFER FUNCTIONS. There i s a bijective

correspondence betveen K[ z ] -homomorphisr;_s f: R -, I? with minimal

polynomial lr and transfer function matrices of the type

where 8 . E xp[z],
J
2% 8. < iieieg*,
J . -
and t i s the l e a s t common
denominator of Z.

I n many contexts, it i s preferable t o deal with the Z corres-

f
ponding t o f rather than with f itself. Because the corresrondence

i s bijective, it i s clear t h a t a l l objects induced by f a r e well-

defined also for Z a.nd conversely. Thus, f o r ?.ns+.ance,
f
A A
dimz, = dimf = d i m Xp;

$z = i e a s t cammon denominator of Z,

= minimal polyllomi.al of fZ.

(6.5) HEMARK. I n view of Propositions (4.20-21), the natural
A
r . ~ - l i z a t i o nof 2, namely XZ =
* z, i s completely reachable as

well a s completely ~bservable. ~ o having

t this f a c t available before .',961,

has caused 3 great conf'usiozae Questions such as thostresolved by Theorem (5.13)

tended t o be attacked algorithmically, using specie1 t r i c k s amounting

t o elementary aJgebraic manipulations of elements of Z. Very few

theoretical r e s u l t s could be conclusively established by this route

u n t i l the conceptual foundations of the theory of reachability and

observability were developed.

The preceding r e s u l t s may be restated as "rulesv whereby the

values of f be computed using Z. We Mve i n fact, f(w) = Z-o, where

(6.6) z*m (Z)/q,

= multiply the polynomial matrix $Z consisting of
. t h e numerators of Z with cu, reduce t o rrinimal-
degree polynomials modulo 9 and then divide
formally by $ as i n Method 1 above.

We can a,lso compute t h e e n t i r e output of the system XZ ( t h a t is,

all output values folloving the application of the f i r s t nonzero input
value? by the r u l e. .

= sane as above, but do not reduce modulo 9.

I n this second case, the output sequence w i l l begiu with a positive

power of z. (l7.e coefficients of the positive powers of z . a r e
thrown away in the definition of f (see (3.7)) i n the definition
 .
R. E Kalman

of the scalar product i n I', i n order t o secure e simple formula

for Xf = ~ / h e r n e lf . )

Many other applications of transfer functio~lsmay be found i n

KAUGW, FALB, and ARBIB 11969, Chapter 10, Section 101.
It i s easy t o show t h a t the transfer f'unction associated x i t h

t h e system Zf = (F, G, H) i s given (3Y Z

f
-
= H(ZI F ) - ~ ~ .hi^ i s

just the f o r m 1 LapIace transform computed from the constsfit version

of (1.12) by setting z - d/dt or from (1.17) by setting

x(t + 1 ) = zx(t) .) Yr.obab1.y the simplest way oi compdLng Z is

v i a the formula

where i s the minimal polynomial of the matrix F and the super-

script denotes the special polynomials defined f n (5.5) . The m t r i x

i d e n t i t y (6.8) follows a t once from t h e c l a s s i c a l scalar identity

upon setting w = F, a = J~F, and iilvoking the Cayley-Hamilton theorem.

Much of c l a s s i c a l linear system theory was concerned with computing

In the modern context, t h i s problzm llfactorsv i n t o f i r s t solving

zf-
t h e realization problem f + Zf ard then applying formula (6.83. see

Sections 8 and 9.
One of the mysterious features of Rule (6.A) (as contrasted with

the conventional rule. (6.7)) is the necessity of reducing modulo @.

l'he simplest way of understanding the importance of t h i s
aspect of the p r o b l a i s t o sbow how tO rek%et t i g r?cdae i_n_m.rat

factors occurb! h the structure thtorez ( 4 . 3 ) t o the clcssiczl

facts comerring the ir,v"criar.t fictors or" a polyllozi21 si,rir.

(5.9) E L ' I3 F K C -
t -
P be z I! X n
mtrix with eleU&nts in ar- zrbitrary ~rlnctsal-lder,ldo=in -
R > Then

-
a r e A -
a ~ dB are
-- pXp zX EL zatr%ces(not necesssrily

\mique) -
~ 5 t htlezezts In R azd dei, A, 6et B wdts i n 3, p-hile

i s unique (E? to units in B) h i t 3 \I - hi+lJ 2 = lJ .*., q - lJ s

9 = ra,& P. Tke \
- are czued the i n - a ~ i z , r tf+ctors af P.

&, in particular, -bet.ioen the respective ic-rdi&?tfzctors +:,- ...: -.-'r

.

and \, ..., hq. let us zketch t?re s-Aerd g r c o T of ZSia fact follw-

ing mS and REDER 11962, 513.33 zho C s o give

I P, prmf r ~ f(5.9).

PROOF OF (1;. 34). Consider the R-Lammxyhiszi fr3a fin

onto M give2 by p: e. L, where -i;%e. e eze *e ska&rd
g.,
1 1 i
basis elementn of (recall (4.6)) pad the gi gener2.t.s I.

ClearlyJ H X where N = kernel p. It can he prove2 that

II %R' i s a free subcdule o f Rm, w i t h a besis cf at mst -n
lj
elments. Write each basis e l a e n t f of & es p5 ?ij E R.
j -
Apply is.?) t o S-?i: C - ~ t r i x P. ,"T-Ln@ "r" =r c..
LI sf., i: = B
-1
,
j - . rj r
a = C %-5.-.Q
A
(6. J.Q.-LI), fk = ?i-ci. A
&sce
.j .- . ..

%zt is, (4.34) h01k k i t h $:

a.
= hi and r = rank P = E.

I$ the size f y p of cz3-c-alatiozs, we ccn prove also

(6.12) EEGI!=Z. , ., A
9
be the ir~variaatfactars of

I given 3~ (6.9), _as 15%

A
$1 = G ~ , i = 1, ..., q. men the

in=riaet.. T G I S ~ Cef
~S % arc

-r
where 5s %IE ssillest -2nteger sucF- &t +jh, for.
.
i

Clearly, a E [0Iz = kernel.g iff Z*m = 9 (see (6.6)). . ~~uikently,

( =0 ( $1. Vsing the representa%ion whose existence is claimed

bx (6.9), write $Z = ~m (c,4 D = natrices over Hz!.) ikfke

V = D'-\ where

Tken = 0, ( ~ z ) w
= 0, and W has clearly mxird. ~ 0 %K [~]-

matrices with t h i s groperty. So the colmins of the ' b r i x W consti-

tute a besis f ~ rkercel p. The r e s t follows easi?y, as in the proof

of (4.34). 0

(6.13) FE249RIi. The preceding proof remains correct, kcithout bny

modification, i f the representatior, $Z = CADJ det C, det D = units

5s taken i n the ring K[z]/$K[z], rather t h a a i n ~ i z ] . The former

representation follows t r i v i a l l y from the l a t t e r but q be easier t o

compute.

(6.14) FEbLUK. Theoran (6.12) shows how t o com~utethe invarfaat

factors of 212 from those of $Z. W e must define the ia.variz;lt

factors of Z t9 -
be .;he same as those of XZ (because of ihe

bijective corrzsponi!ence Z e, %). Consistency with (6.12) denan*

t%t we m i t e -

where / i s defined as i n ( 6 . 3 ) . I n other words, qi

the denominators of tlie scalar transfer function AI afker
-.- cancellztion

of ali -omon factors.

Theorems (4.31:-) and (6.12) do not ~Zzllyreveal the significance

of invariant factors i n dynamical systems. l o r i s it conveniezf, t o

deduce sll. properties of ~ratrix-in-arizntfac-bors f ran-'the represenhtion

 R. E. Kalman

theorem (6.9). It i s interesting tkt the s b ~ p e n e dresults we present

bolow w e nu& i n t'se s p i r i t of the original work of ~ ' r r n ~ H.~ J.

S ,S,

S?.LITH, EEROIECKER, F R O B i I U S , a n 2 i-EZSE.L, a s summized in the well-known

monograph of ~~~H [ 18991.

(6.16) I)EFII?ITIOZ. A, B recta.=-mar matrices over a fact-

orizzt;ion d o ~ r , i n R. AlB (read: A divides 33) i f f there are mztrices

V, W (ov$r R, of a p p r o ~ r i a t esizes) such t h a t B = VAW.

This i s of course just the us& definition of "dividen in a ring,

specizlizedto the noncomi).tztive r i r g of =trices.

The follo-sing res-fit [bD!?II 1899, Tieorem IIIa-b, p. 521 shows

t h a t i n cese of principzl-iCizal the corresp1111elice between

zr-d their invarizfit factors preserves the divide relation

m%r%ces

( i s w l ~ u c t c r i a l 'with
' respect t o "di~ide"):

(6.17) w . -
Let B be 8 p r b e i p l - i d e a l domzin. Then A! B

if azd n31y i f A.1(A)/ A; (3) fcr 211 .i.

6

EECaF. BMficiezcy. Write %he re2resectation (6.10) as

A = V A1,
W B = v2% W 2.

By hypotnesls, there i s a % (diagonal) such t h r t = 4.Eence

B = v2fyyr2,
=

= ( ~ , v i l ) kA( l i ~ ~ ~ ~ I ? ~ ) .
 R. E. Kalman

Necessity. 'k;s i s jmt the followi%

(6.18) m . For an arbitrary x ~ ~ o u e - f z c t o r i z ~ t i o n

domain R, A!B implies hi@)! hi(3) .

(1899, meorem n, p. 16-17]. o

l ~ cm
s.letes the proof of Theoren (6.i~) u

(6.19) RBABK. Since (6.9) does not z=,g~y(w>y?) t o unique facto?i-

zatiorz donzins, for purposes of uskg ( 6 .u) we need k=ELmTRd-%I s

definition of invariznt factors: if A.(A) = greatest c-n factor of
3
all j X j minors of a natrix A, ~5th -I~(A) = 1, then

h1(A) = A ~ ( A ) / A ~ - ~ ( A ) . O f course, t h i s definition c a be shorn t o be

e q u i u e n t (over g~inci-g&-idea dmzill-~)to tht h q l i e d ty (6.9).

In analogy k i t h Definition ( 6 . ~ 6 )l~e t us %Tee (nete hnversioot ) 02

.-

(6.20) DEET.IllT!IOl?. -
Let Z1, Z2 be trensfer-function nztrlces
-L

ZliZ, (-read: Z1 divides Z2) iffthere a r e =trices -

V, Y over ~ [ z ]
such t h a t Z1 = E 2 W . ( ~ o t ethat ZII Z2 implies a t ome: Ib .)
z2

(6-2L) . 41Z2 if and only i f $i(~l) I qi(z2) for all i.

PROOF. This is the naturzl counterpart of Theorem (6.16)~

=id follows from it by a simple calculation using tbe definition of

$,(z) given by (6.15).

 R. E. Kalman

(6.22) DEFII<ITIO:I. - Zl can be simulated by z ~ )

51Z2 (read:
i f f -51 % .
- that is, i f f 5 i s isonorphic to L submodule of
1 2 1

t [or iso~3rphict o z quotient ~ 0 6 u l eof

It2I -
This definition i s zlso functoridly related t o the definition

of "divide" ovsr a prbci2al idezl d&n R because of the foUoviri

(6.23) TEXOE3.f. -
Let R be a >X, Y

R-mdules. -
men Y i s (isonorpvnic) t o a submodule or quotlent ~ o d u l e

-
of X i f and only i f

Sufficiency. Ta3te both X and Y in canonical

fom (k.3&), ~5Cn , x generating the cyclzc pieces of X,

Yl, ---,
Yr(x) (ayi = 0 if i > r(Y)) those of Y. The

assZgxeat y.1 u ( W ~ ( X ) / $ ~ ( Y ) )defines

X~ a mm@sm Y + X, that
is, e a b i t s Y as (isowr-hic to) a submodule of X. Similarly, the

~ c s * ~ fxi *ii yi epimcrphism X + Y exhibiting Y as

ciefhes a.~

(isamorpkic to) a quotlent nodule of X.

Necessity (foJloxbg WJIiBKrCI [.91ke3re, Chapter 7 (2e =do),

Section 4, Exercise 81). L e t P be a submdule of X. By (k.34),

X L/X where L, N N free R-mdules. By a classical isomorphism
meorem, Y i s isomorphic t o a quotient module 1 , where L 3 E4 3 N
and 1.5 i s free (since submodules of a free module are free).
 .. .
R E Kalman

From the l a s t re3-%tion, r (Y) <- r ( ~ ) . Now observe, again using (4.34)

that, f c r a ~ R-mdde
y X and' any ?rE R,

therefore

- R(rk(x) = ideal generated by (n: r (TX) < k] .

Since fl is a submodule of 71X for all n E R, it follows t h a t

R$ k(x) ~ K . + ~ ( Y ) , a d the proof is c q l e t e for the ease when Y is

a s ~ . m d z i eof X. !be proof of the other case i s similar.

PIIOOF. Inmediate fram the f a c t that

of C (see Section 7).

% is a submcdule

NOW we cul summrize main results of t h i s section as the

(6.25) - J??XDED E ~ ~ I T I O THEOREX

I ? FOR LINEXR DYNlY31CAL SYSTEI-is.
The f037owing conditio. s are equivalent:

i Z1 F v i d e s Z2.

(35) ~ ~ ( divldes
2 ~ ) qi(Z2) for all i.

(iii) can be simulaked by 3.

1 2

PROOF. This follows by cornbinirg Theorem (6.21) with Theorem

(6.23), since qi(Z) = qi(%) by definition.
. (6.26) - ~LWGWREWFION.
The definition of z11z2 means, i n s y s t e

theoretic terms, t h a t the inputs and outputs of the machine chose t r a s f e r

m c t i o n i s Z2 are t o be llrecodedll: the origioal input u2 i s replaced by

an input w2 = B(z)rul and the output r2 i s replaced by an output

rl = A ( z ) ~ ~with
; these "codiiig" operations, L2 w i l l act l i k e
a nmhine w i t h transfer Pmction Z1. In view of the definition of a
trausfer function, the equation Z1 = AZ2B i s a l q s satisfied vhenever

A, B are replaced by z, -reduced

modulo $ ) !L'his new that the
z2
coding operations c m be carried out physically given a delay of
d = deg $ units o i time (or more). NO feedback i s involved i n codine,
z2
it i s merely necessary t o store the d last elements of the input and
output sequences. Hence, in view of meorem (6.25) Corollary (6.24),

ue can sw that it i s possible t o a l t e r the dynamical behavior of a

system C a r b i t r a r i l y by external coding involving delay but not
2
feedback i f and only i f the invariant factors of the d2sired external

behavior (2 )
1 .
are divisors of invariant factors of the external -
behavior (2 of the given system. The invariant factors mby be
=2
c a l l e d t h e PRIMES of linear systems: they represent the atoms of system

behador which cannot be simulated from smaller units us- arbitrary

but feedback-free &rig. I n fact, there i s a close (bot not isomorphic)

relationship between the Hrohn-Rhodes primes of autonab theory (see

m,FAZIB, ard ARBIB [1969, Chapters 7-91) and ours. A w treat-

ment of this part of linear system theory w i l l be published elsewhere.

 R. E. Kalman
7. ABSTRACT THEORY OF REALIZATIOITS

The purpose of this short section is to review and expand those

portions of the previous discussion whicli are relevant to the detailed
theory of realizations to be pressnted in Sections 8 and 9. The same
issues are examined (from a different point of view) also in KAMAN,
.
FALB, and ARBIB [ 19691
Let f: a + 'I be a fixed input/output map. Let us recall the
construction of X' as a set and as carrying a K[ z]-module structure
f
'
(sections 3 and 4). It is cleu that (i) f = L ~ w where ~ ,

are K[ 21-homomorphisms, m d (ii) pf = epimarphism while L = nonor lorphism.

f
We have also seen that

rpf = epinwrphism W Xf is completely reachable;

1rf = monomorphism W Xf is completely observable.

These facts set up a "functorm tetween system-theoretic notions and

algebra whLch characterize X uniquely. Consequently, it is desirable
f
to replace also our system-theoretic definition of a r2alization (3.12)
by a purely algebraic one:

(7.2) DEFINITION. A realization of a K[ z]-homomorphism f: Q -, I'

is any factorization f that is, any commutative diagram
 R. E. Kalman

of Hz]-homo~orphisms.
5
as]-module X -
i n called the s t a t e
rno$ile of the realization. A r e a l i z a g n i s can_onical i f f it i s

completely reachable and completep~observable, t'r;at.is, IJ. &

surjective and t i s injectivz.

A realization always exists because we can take X = 0, p = 12,

r = f (or X = 1
: p = f, l =.I$.

(7-3) RENARK. It i s clear t h a t a realization i n the sense of (3.12)

can always be obtained from a r e m z a t i o n given by (7.2). I n fact,
define C = (F, G, H) by

G = p r z s t r i c t e d t o the submodule [a: Icul = 11.

H = L followed by the projection r t+ ~(1).

It i s easily verified t h a t these rules w i l l define a system with

f, x = i. Given any such Z, it i s a l s o clear that the rules

define a factorization of f. b2tween (3.12)

Hence the_correspctr~de~.ce

-
and (7.2) i s bijective.

!the quickest way t o exploit the algebraic consequences oP our

definition (7.2) i s via the following arrow-theoretic fact :

 R. E.Kalman

(7.4) ZEIGER FILL-IN LEMMA. Let A, B, C, D be sets and a!, p, y,

-
and 6 set maps for which the following diagram commutes:

-
If a is surjective and 6 is injective, there exists a unique set
cp corresponding to the dashed arrow which preserves comutativity.

This follows by straightforward "df agram-chasing", which proves

at the same time tke

(7.5) COROLLARY. The claim of the lemma remains valid if "sets"

are ~eplacedby "R-modules1tand "set mapsn by "R-homomorphisms1'.

Appl~ingthe module version of the 1- twice, we get

(706) PROPOSI'i!ION. Consider any two canonical realizations o f 2

-
fixed f: _the corresponding state-sets are isomorphic as K[ z!-modyles.
/
Since every K[z ]-module is automtically also a K-vector space, (7.6)
shows that she two state sets are K-isomorphic, that is, have the same
dimension as vector spaces. The fact that they are also K[z]-isomorphic
imp'lies, via Theorem (L.341, that they have t h e same invariant factors.
We have already employed the convention that (in view of the bijection
between f and ef), the invariant factors of f and 4 are to be
 R. E. Kalman

identified. I n view of (7.6), this i s now a genzral fact, not A-ependent

'

on the.Bpecia1 construction used t o get X Wz can therefore r e s t a t e

f
(7.6) as the

(7-7) ISOl4ORPHISM TI-EOREM FOR m01\31CAL REALIZATIOIIS. Any two

canonical realizations of a fixed. f have isoraorphic s t a t e modules.

The s t a t e module of a canonical realization i s uniqueu characterized

iuP t o isomorphism) by i t s Lnmriant factors, which may be also viewed

a s those of f,

A simple exercise proves also

(7.8) PROPOSITION. -
If X i s the s t a t e module of a cafioifical
realization f, then dim X (as a vector space) i s minimum i n the

class of a l l realizations of f.

This r e s u l t has'keen t s e d i n some of the l i t e r a t u r e t o justify

the terminology ''minimal realizationt1 a s equivalent t o " c = o ~ c a l

.real%zation1'. We s h a l l see i n Section 9 t h z t t h e tva notions a r e

not almys equivalent; we prefer t o vies (7.2) as the basic defini-

t i o n a-ld (7.8) as a derived fat*.

(7 9) IWlARIC. Theorem (7.7) constitutes a proof cf the pxeviously

claimed (4.24) . To be more explicit: i f Z = (I?, G, H) and
h 'A h A
C = (F, G, H) . are two t r i p l e s of matrices defining canonical realiza-
.. . .
tions of Cne sane f, then (7.7) 5nplj.e~the existence of a vector-
n
spacc isomorphism A:.X 4 X such that
 E, E. Kaa!man

A
If we 2-ti% X m d X men A i s simply a -5s a e and it
'olhus + a t -the clsss of a77 mtrix t~iplesw3ickt *E canorLca7-

reaXzz;tio+s of a fix& f is i s m q ~ w%%h

c t n e qeneral Unear

proup over 3.

a factorizztfer; of f suck mt dim X < CO. ithis is -the~

q r e s s e d by ssghg tkat f hiss f i d t e rank.) Gfven my such r e -

&ion, a ohm a --onica.l

it 5s pss5Ue " one bx a process of

e Rare precisely, v;t Wve

(7.~) THEXEEX. s e ~ rge a l l z ~ t i s nof f w5CI &ate m25ule X

cor.bins a szbo~otler-Lt(2 uqu~tferntaf z submodule, cr equivalently,

a c=ozCeal rezEzat-ion of fS

P300F. %e re2cMbIe states Xr =image i; are a submodule

of X +nci so are the unobserye'ole 'states Xo = 5 r w L r. Hence
X, = Xr /X
'r
firo is a subqpatient af I. It fdlous imediately that

X; i s a C B ~ sw ~ d u~l e for
Oe - m f. [The proof m y be visualized

v i t~h e foU(?iP2cazmtztive d i a g r q where the j 1 s srzd prs are

cam= d injections m.6 projections.] CI
 3.E. Kalman

(7.12) lG34A.I- S i m e any s u b q i h ~ %of X is isanor9hic to a

that X can e a ;.ealizatiaiz

sizte-atzte ~ d u l of 3x13- if d, (?jl$.(x)
1 1

f ~ all
r i <re& .&o Corallaq {6.2;i-j ) . This condition, houevsr, is
no& e2o@ s-c the *i are 2 n k z i z n t s of &Gi? i s o m r p k i s s g~t.

imx@iss;of tAe comm';ative diagram (7.2).

!5e precedizg E s c ~ ~ t s sdf ~c o~- d dbe B q t in ad to &ah over-

 R . E. Kalman

8. CONSTRUCTION OF REALIZATIOI~S

Now we SaU develcp and generalize tEu2 h s i c algorithm, originally

due t o B. L. Eo (see EIO and KAI&!AN Ls661), for comguting a c-nical

realization C = (F, Gj 3) of 8 given input/output nap f. Host of

the discuss5on will be i n the language of n t r i x algebra.

Notatioas. Here a;ld i n Section 9 boldface capita3 letters* w i l l

denote block mai;rices or sequences of matrices; f b l t e Slock =trices

w i l l be denoted by SIIIELIL
Greek subscrists on boldface c z . ~ i k l s ;the

elexients of such =trices will be 6enoted by o r d i w ~

ccf,itals. TMs

i s intended t o make the practical asgects of the conguL&t.ions self-

evi&nt; no f l r r t h e r ez-ilanations will be mzde.

LeC f: 9 -, I' be a given, fixed z j-h~mamorplrisa. Using only

the K-linearity of f we have that

&re the % (k > 0) are p X m matrices over the fixed f i e l d K.

Xe demte t%e t o t a l i t y of these matrices by

!Ben 1% i s clezr that the specif icat.1on of a a z]-ho~omorphism f

i s e ~ i v z l z n t o the specificat,ion of i t s mtrix sequence &
-[f). &re-

over, if Z, r e a i z e s z :8.3 r,z2 be v r i t t e a e x p l i c i t u as

*?Tote t o t2rf~ter: IndLcztcd by double wdtrline.

 R. 5. Kalman

Camparing (8.1j and (8.2) we can translate (3.12) into an equivalent

matrix-language

(8-3) DEFIMTION. A dynamical systen Z = (F; G, H) realizes a

(matrix) i n f i n i t e seauence 4
- iff the relation

k t u now t r y t o obtain also a mztrix criterion for an i n f i n i t e

.. sequence A- ' t o have a finit-e-dimensionzl realizatioa. The s i q l e s t
-to do that is t o first m i t e down a &rix representation f o r the
map f: Q + l?, So let

and verify thzt g(A(f))

- - represents f when u, E R i s kiewed as an

Classically, I$-
(?) - i s Imo-a as the (infinite) Hzd~elm t r i x associated

kith -4,
- We d&n.oi;eby H
*,v the ir X V block submatrix of
- appear-

i .in'the Qpper 1eft:habd corner of It.

-
(8-4) PROPCSITIGH: -
Let C be any realization of -
6. -
Then

* > v =(A)
~rrnkFE <
= dimZ
.
for a l l p, v. >
=I
1.
(8-5) C O R O ~ .A n infi:.,te seausnce 4- has a finite-dinensional
realization o n l y i f rank H+, ,(4)- i s constant for all 14 V sufficiently

large .
PROW. If din C = =, the claim of the p o p s i t i o n i s

vacuocs (althollgh f o r m l i y correct!). Assume therefoze that dim C < 0

and define from Z the f i n i t e block matrices

and

% = [ ~ ..., H ~ ( F ~ ) ~ - ' I -
Hn ~ ~ F I ~

!hen

by the definition (8.3) of a realizatiolz. It is clear that rsnk R

=v

and rank $ are gt most n = dim C. !!&us our claim is reduced t o

the stanjlard matrix fact

rank (AB) <- min ( re.& Aj rank 91 . 0

Our next objective is tl;s proof of the converse of the c o r o n r y . This c m by

done in several ways. !5e original praof i s due t o HO and WIlMQI [ 19661;

similar results were obtained independently and concurrextly by YOUIA

and TISSI [19661 as we= as by SU;~W&Y [1g66]. a - o different proofs

are adalyzed and compsred i n w !FALB, and BRBIB [1969, Chapter 10,

Section I l l . All proofbdepend on certain finitzness arguments. We

s h a l l give here a variant of the prwf develoged in HO and m.LW [19691.
 R. E. Kalman

(8.6) TI^. The i n f i n i t e Ifankel =tr% .g

- associated Kith

- has f M t e 1eW.h A = (A1,

t h s sequence A - XI) i f f one of the follow-

ing two equivalent conditions holds:

At = min (ll: rank &t~,;v~=rankIll,*

9 .
for a l l K,V = 1, 2, ... ) < =
-
or

={lQiIl , --.raIlk$,jI,+~ for= K , p = 1,2, ...)

<a.
. .

he i s the row length of H- A'' is the calm length of 3.

-
The equi-ence of the two conditions i s immediate from the

e-ty af the row rank and cobmu rank of a f i u i t e matrix. Tne proof
of the folloubg result ( m t z-.Aed in the sequel) is l e f t f o r the reader
as an exercise in f-arizing himself with the s p e c i a pattern of the
el-ts of a Hankel matrix:

(8.7) PROPOSITION. For any l

-i, the f o l l ~ r r i minequalities are
either both true [g
- bas f i n i t e lewh1' D r bot5 false [otherwise]:

A l <- r a n k H .C mhn,
A" =
dtl,
. A" 5 rank lii,
- - ,ph) c
= PA',

The most direct consequence of the finiteness condition given by

(8.6) i s the existence of a finite-dimensional repesentation -s aa~

- of the shift operator a*. acting on a sequence 4.

- The llopkrand't
% i l
be thelHankel' xistrix associated with a given &
-.. A s we shall see

soon, + M s representation cf the shif'k operator W i e s a rule for

computing tke matrix --
F of a realization of 4. This i s e m t k v what

we would expect: module theory t e l l s us thzt, loosely speaking,

gg - u-" z "F.
N

(8.7) DE,FDX'IIO=L'. The shift opezator oA 02 m i r f i n i t e sequence

5 i s givm by

the corresponding s h i f i oyerator on Ra_nIcel =trices i s then

9: - - EI(A) t+ HH(~A).
= A=

(of course, d s o on s u b t r i c e s of a Hankel matrix.)

H is uell-defined
u

(8.8) WUH IJWW. - associated with an Lnfinite

A HcxGsel m t r i x Zi
-
sequsce A- has finite length i f and o d y If thc shift ooerator
uH
has f5nita-dimensio- l e f t and right m t r i x rextreseatat5.ons. Precisely:

H
= -k s M t e Zemh A= ((A': hw) if arrd only i f there e x i s t g8 X I'

-
cmd La X Im b1& matrrces S = Z
= -and sxch t h z t

and .murthem3re t'nc aini.mun size of these &trices satisf'Jl2n.g (8.9) i s -

At X At -
and hw X An.

PROOF., Sufficiency. Take any It' X 1" block matrk Z_

-
r i sstf sfies (8.9). Compute the last calm of $, r,,Z:
 R. E. Kalman

for all j = 0, 1, ... (where by i s the (p, v ) ~ 'element

~ block

of I-) . &tion (8.10) proves that

lank !&+l,2w = rank for- a l l K = 0, 1, .-.;

the general cane follows by regetition or" the same argument. Hence the

- implies that the co1m.u length An

existence of the c W d Z_ of g-
-- If actually A"
cannot exceed the size of -2. - is m r than @ size

of the -st -Z vhich works in (8.9), we get a contradiction fran

the necessity part of the proof. The concerning S_
- are proved
by a s t r i c t l y dual argument,

Hiessity. By the definition of A", each column of the

th
(A= 4- 1) block column of ,&I i s linezrly d e p d e n t on the
columns of the preceding block colurvls of H+, moreaver, this

property i s true for a l l integers v, no m&tter how large. So there

exist n X m matrices Z1, ..., ZA,, such that the relabron

hclds identically for a l l j = 0 1 . . How define 2_- t o be an

A" X h" block companion mtrix of m X m block made up from the
Zi
just -defined:
 R. E. Kalman

The verification of (8.9) i s iruneciiate, using (8. U) . The existence

of At x 2,' block matrix s-- verifying (8.9) follows by a s t r i c t l y

dual argument. 0

Row we have enou~blcaterial on hand t o prove the strong version

t
of Corollary (8 -5) :

(8.10) ~ 0 n f i n i t e. sequence 4
An i ~ - has a finite-dimensional
rea.lization of dimension n i f and only if t h e associated Iranlrel

-matrix
= H has f i n i t e Length h = ( A * , A").

PROOF. - Sufficiency. Let Ehl,, be a A" X 1 block

column matrix whose first block e l a e n t i s an m X m unit matrix md

the other blocks are m x m zero matrices. Us@ (8.9) with el' =A1',
 R. E.Kalman
Then, f o r e31 k >- 0, comput 5

t3e second step uses (8.9). By definition of uA and E, the l a s t

k
m t r i x i s just the (1, l)th element of -
-lI(a--(A)), ""lY %+ka
Hence the given C ' i s a realization of -
4.

Necessity. This i s immediate from Cor: ;ary (8.5). 0

ITOWwe wznt t o attack the problem of finding a canonical realiza-

t i o n of &-, since the realization given by (8.13) i s uuaally very far

f'rom canoi3ical. Olrc succeeding considerations here anr? i n ~ e d t i o n9 .

&re made more transparent i f we digress f o r a moment t o establish

another consequence of (8.8).

By ou5rageous abuse of m e , we sh&Ll say that 4- Us f i n i t e

length i f f --
H(A) has f i n i t e length. We note

(8.14) DEFINITION. An l n f i n i t e sequence B

- i s an extension of
-
order N of (the i n i t i a l part of) an i n f i n i t e sequencs A- iff
% = % for k = 1, 0 . 0 , N.

(8.15) THE:O-. No i n f i n t t e seauence of f i n i t e length (hl, A")

has d i s t i ~ c lenqth-pres:-rving
t extensions of any order N 1-. hl + A".

PROOF. Sugpose - is
.B a-'-length-preservinl: ext;ension of order

N of A,- the lengbh of both sequences being (At, A"), with N >
- A1 + A".
By (0.8), both sequences s a t i s f y r u e i o n (8.9), with s u i t a b ~ e & and 5.
 R . E. Kalman

The sequence 4- i s Uiquely determi~edby acting on llA,,A,,(h) -

Prom the l e f t &nd the sequence -a i s uniquely determined by ZB
acting on the matrix zhl,(E)
At, I *om t h e right. The twc m a i ~ f c e s

a r e equal by hypothesis on N. Moreover,

aze a l s o equal, since the matrices on the right-hand side depend anly

on the 2nd' .--,N-th member of each sequence. Using only this fact

and the a s s o c i a t i v i t j of t h e matrix ~roduct

Now we can hope f o r a realization algorithm which uses only the

first. At + A" terms of a seqzlense or' f i n i t e length. I n fact, we have

.!8.16) B. L. FP8s REALIZATIOIt W R I T H M . Consider any i n f i n i t e

-A of f i n i t e length with associated Hankel matrix H.=
sequence .- -
The

follotring steps w i l l lead t o a canonical realization of A: -

 (ii) Compute n = rank HA, i n doing so, determine

nonsinrndar ph' X pht mAt' X'mh" matrfces P, Q such t h a t

(iii) Compute

m are idempotent "editingf1mahrices

-
where Rp, C corresponding t o Cie

operations "retain only the f i r s t g ---

roks' ....L I "retain
~ only the first

m columns".

We claim the

(8 19) REALIZATION THEOROYI FOR IIWIKCTE SEQUENCES. -For an,, i n f i n i t e

sequence A- whose associated Hznkel matrix H

- has f i n i t e length
(A!, A"), . B. TU.
Hots forrnu12.s (8.1.7-18) yielij- a canolical realizaticn.

PROOF. If Z definsd by (8.17-18) i s a realization of A,

-
then it i s certainly cznonical: by (8.4) C ha.s minimch dimensicn' in

the class of all realizations of d

and so it 1s canor;ical by (7.8).
The required ~ e r i ~ c a t i oi sn interzsting. F i r s t , drop a l l

subscripts. Observe that - =

H# QGRP i s a peudo-inverse of -
H, that
 R, E,Kalman

is
---Fx_--- = 1-1.
.- Tlen, by d 2 Z W t i o n sl' -
Y
F G, 3, snd If*,

first n eel-- sf t$~, -fie&is-j u s t %+kt as re@reS. D

(8.~3, flr rzspets.

~ V G
. -
(5) It is m laager-a c e s k y to cqte
-
-
5: we s*r
- .

e t ( -A*%
A=A 5s *t of the of the $:c3lerc.

i Yo- s -{8,38) g 2 ~ Yse desired 2 esliza*f on is .w&l

W &pe& (just :.'~=e_ b e latter) Yre&-.- 09 (8;a)- . .

..
- < ~. .
 --
Br; 8nnarently seriocs 'imik t icn ef t h e a l g a r i m ( 3 . ~ 6 )is fie

necessity to vex ify ~ 3 s t r a c : t Utbzt "4- U s fizCte 1eq5tfsV. Of

cocrse, this am ke doue oslj- cr; the lksfs of c e r t z k s-=cia Qptheses

,- given ir d ~ r a r x e . (&angles: ( 2 ) % = O for f l ir > q;
(if)-% = ccexEi.cieEts of the r;y-lnr t z q a s f c ~ 1oz 2. ratio& fri?ctzcn.)

Pcxc&e&-, GfSculty is only spj?zrezt, fcr ~ c c e d i z g5.13-elo~-

e1ts a be ~ ~ n flmt'ber:
e d

q -+;r!j%e - m2 the
secuecce A
- H
-.
corrcs>z~t25qYm'Lel ~ t r h

-
with At = lB, An = I" @yes s c ~ ~ , o xedizaf.i3a
I ~ c ~ ~ of A.
-
6

Zxaetly as the ucessZtjr pzt cf me wocf ar'

(8.8), c-eon (9.22) -+ss ~2 existezc~of - Z- suci ~t
scse extel~fonof &
&reover,

by t r i a a,&

&
- for wMch

a h i s r-ze

kmiri,

(8.22)
l
iz)

If
By repeated .z.a?Ec~tioriof (8.231, i t follows thz+, we have

HOU it 1s c l e z , fmm (8.8),

(8.13) 5s st=

l3xarcm (I?.=) sq-s,

-5,

(8.22) is never satrsfied.

If gja,em

s o x exbensio~of

kowever, %w

~ x l b order cf W r extexicn of

The rer&-ning i-nterest*

that A:

even

i s s-e

is not satisfied for a f i n i t.e'oc

&-,
A

- i s 2;vzys possible as s c o ~as (8.22)

4.
-
5 el
-

t3110u&
md

(IwJ 2")

((822j & be used z s a ~rzcklcalcrZteria for c011stnicCuZag

error a camnicel r u z - t i o s of any

f.Inite 1eagt.h (but kit-mu&b e i q ~ i v e n As? AV')

(8.25) I .-i )
-
-re is m
.

azd has i t d l
i~ the sca3ar case), Uien (8.22) 2s a u b m . t i a J l y sa%irfiied-

(iii)
tiso conzerfing c&ticn

-
tbe al.goritb (8.16) is 2 3 , 0 ~ 2
with~u%
~~
nfi 5 I-.

a y informa-
5-

bf'ini* setpace

--
R. E. Kaiman

!BE uoique-

is set r-.cessarily

fo eTZ'ect, Chat, e m t 1 5 c a . I reUzztion of

i s satisfZed.

to b v e

(9 = n = 13

(for instzllcc,

( 8 . ~ ) t~h e s y s t a X &?2ned by (8.28j -sill

~t lease of order 1- It 2s not
to get a simple r'omfla w%-ch vcdd deternine i 2 ~

q~estioni s tbai:

of &ta
k?at czlr be sail! if

5,..., % and
any , ' satisfying + it'= Y. =s problem is the topic of

the next ~--2t.ion.

(8.25) 3- aC4EiE. An e s s e t i z l fezture of 5. L. .Hols da:orithm

is thax i s preserves the block structure of the d a t a cf the problem. Of
course, one can ob-n parellel results by treat- l&l,lm as811

ordinary mtrix, disregardiq i t s bloc%-make1structure. Scch a

-- of
procedure requires lw!shg at a minor of H rank, ard was

described explZcit3.y 5y SILmC%N [19663 SILWICSi and NEiWFiiS [l*].

There does not seem t o be any obvious c ~ u t s t i o a z a5iraatage

l associated

vith the second mi&&.

 9 m O R Y OF PARTIAL ~ ~ T I O N S

In oae obvious respect the theory of rezlizations developed

i n the previous section i s rather unsztisfactory: it i s concerned
viL& infinite sequences. F r a here on we c z l l a system satisfybg

(8.3) a c o q l e t e realization, t o distinguish it from the przcticalQv

more interesting case given by

(9-1) DEFINtTIOli. - -- = (AlJ

let A A2, ... ) be an i n f i n i t e

sequence of I> ;c m =trices over a fixed f i e l d K. A meal

systera C = - of order r of
(F', G, H) i s a p a r t i a l reail-%tion -

We sbSL us? the tel.zliwiogy if, i@tezd of ul ZzGinite

s >- r. P-e rezscjn for U - s convention w5.U Ye clear j3-m the dis-

cussion to Sollow. W e shall call tfie first r .i;erms of -

A aprtiaP

sequence (of wder r).

The cclmepts of canc112csl pzrkl2.'- rezlization and minim1

m ~ r t i a reali=&2oc
l vill .t understood Z n exzctly the s m e sense as for

a complete rezlization. iie u a r n t h e reader, bowever, that now these

txo notions w i l l t u r n out t o be inequ5ualeak, in that

but rot conversely.

Our =in h k e r e s t w i l l be t o deterdne 21l equivalence classc;j

of ~ Z ~ ~ T=tZzl
LL rez3zztiolls; i=i gaerzl, e given sequecce -111.
have i n f S t e 1 . y ESIY izequivalellt niW p = t i r l rezlizztions i f

r is sufficiently n.nall.
According t o the & ? i13eoren (9.21) of +&e theory of re-za-

tions, the r e d i z a t i o n probLem has e d q u e salutiori

pP.5~2

whenever the r h i cosdition (8.22) i s sal-isfied. I f the l@h r of the

p r 3 i a . l s e q ~ i w ei s prescribed 2 priori, it c q u2ll hapsen .tki (8.22)

does not hold. What do do? C l e z r l y , i f xe k v e "a- T ptia

realizatior (r", of order
G, £I) r we ca3 exte- &z

sequence of A
=r on ~ 3 1 5 ~
-,c 3 re23izztion is ' w e d t o a minite

Coaseq'il=ntly, w e have the pre-y'

re-slizztion fcr A i s equivalezt fo the dete-z-rZnatiza of zll

=r
extensfons of a 9 t Z a l sequence 1 such t B a t t%e eeen?ed
=r
sequezce is

- Z X C , Fore st?ong7j-;
~-dizenric?~l
(i) finit-
(ii)55s dimension i s mLnic21 i n the clzss of &I1 e x t e a s Z o ~ s .

It is trivial t o prow Wt f i a i t e - & e n s l o exte~sionsexist

(of f i n i t e le~gth,. Hema tbe pro'blen i s il.-i=rliaLely

for aoy prtlal seq-~=iln_e

reduced t o dete- extezsions ~ E c have

h mil.ilr-.l dinension. The

solution or" this l a t t e r problm coxsists of t x o steps. First, ve shov

by a trivial argumenc a t the m-i dirznsioa c a be bounded frm

below by an exmination of tke Hvkel array defined by the pzrtiaL

sequence, Second, and t h i s i s rather surprising, we shcw that the

lo',er bound cul be actually attained. For fhrther details, especially

the characterization of eqcivalence classes of the minima'l partial

realizations, see mL4N 1196% and 19703!.

(9-3) DEFINI!lZW. By the -el arzq g(!($) of a partisl

sequence A
=I
we mean that r X r block Emkel matrix whose ( 5 , j)th
%lochi s Aicg - i + j - 1- r
- -if and undefined otherwise.
In csbher words, the HaAel array of a parti& sequence A
=r
consists 05 block rows and columns made up sf subsequences

A
P
'
..., r A - p <I-)
(1 < - of A
2
andblaak spaces.

(9-Q) PROKISITICH. Z ) % the number of rows of the

no{*Lr
IEankel array of M c h are linearly independent of the roxs
5
above then. !Then the dimension 6f a realization of A
=r
i s a t least

n0(A=r1-

PROW. The rsr;lk of ary IEankel ntrix of an infinite

sequence &
- is a lower bound on the dhxmsfon of realization

of A, by Propsitian (8.4 ) . ~y ~roposi-i;ion(9,2), it suffices

- of A=r: This -lies

to corrsfder a suitable extension A_, *'filliag

in" the blank spaces 5a the Iknkei. armv of &. Regardless of how

H(A
= ) is fiU& i;l, the rssk of the res-tlting r . i( r b h c k EEs.dsel
matrix i s b i d from below by %(ir), 0

By $he ,block syuzetry of the Flankel matrix, ire ~ ~ o u expect

ld

tc be able tz dete,-mine - o0(.A=1" ) by €a andlogous krniilation of' t h e

columns o f the -el =rag of -$, thereby obtaining the -
same

lower bound. This i s indeed true. We prefer not t o give a direct

proof, since the result wi-11 follow as a corollary of the Ma5n

'Phe c r i t i c a l f a c t i s givw by t h e

(9.5) L . For a partial sequence A

=r
define:

t A ) = smzllest integer such that f o r k'> ht every

"L

=r)
row of I-I ( A i s linearly de:?nsent on the

r o w above it.

hn(llr) = snnliest integer such tbat f o r kl'> A" every

c o l m i n the k-th - =r)

block coluwr of H-(A

i s linearly dependent on the c o l ~ n - st3 the

l e f t of it.

Every partizl seaugnce A may be extended t o an i r f i d t e

=II

sequence - i n zt l e a s t one way such that the cond5tion

(9.61 r& Ks ,($

-9
= 'LA(~LL) for a l l p > (A ), Y
=r > Vy(.Ar)

i s satisfied.

PRCOF. Tht: existence of fhe nm3ers At. - h" Is trivial.

It suffices t o show, f o r arbitrary r, hov t o select AHl in

such 9 way Wt the rr~mbers At, A", . -and no remaia c o ~ s t a n t .

Consider t!-2 f a s t raw of Aril -a& exzmine in turn a l l the

first rms of the first, secon&, Zlir6, ..., ALth b h d - r o w s in

- (4
2 -r) . lf the first row rE the first Slock row i s linezsly aepen-

Cent the r a i a aboxx it; (t&t is, 0), ve f i l l in the f i r s 6 row

 R. E. Kalman

of AHl using t U s linear dependence ( t h a t is, we make the f i r s t

row of AHl a l l zeros). This choice of the f i r s t row of AHl

will preserve linear depeadencies for the f i r s t row of every block

row below the second block row, by the definition of the Hankel

pattern. I f the firs+, row in the first block row is l i ~ e a r l y

independent of those zbove (that is, contributes 1 t o n (A )),

0 =r
we psss t o the second block row ami repeat the procedure. Eirentually

the first row of some block row will become linearly dependent an

€hose above it, except when hk = r; in that case, choose the first

row of AHl t o be lirearly dependent of the f i r s t rows of

5, ..., Ar. Rspeztiag t E s process for the second, third, ... rows

of each block rob*, evkntmlly AHl i s determined without increas-

To caaplete the >roof, we must show that the a b v e definition

Of AHl
also preserves the value of h': That is, we mst shcw

thzt nc new indepeadent co1un;;s are produced in the Ha.nkr:l axray of

A bfhen Ael i s f i l l e d in. !Chis i s verified immediately by noting

=r
that the definition of Am implies the conditions

rank EE
=r, 1 = r* E
s+
l.,!l

--------------- .. ..

"3f course, zo.. linear depel2e~cei n the f i r s t step does JoS

imply. t?zt tke =or:-~;,aondin@;
row of Artl will b' e C I i zeros.
 With th-.
aid or' this simple but subtle observation, .$he prablem
is reduced to that. covered by the Main Theorea fq.21) of Section 8. We have:

(9.7) W N THEOEiEE2 FOR b m 2A.lWALREALEATIONS.* 3 =r

A

be a m i a l seQueKce. Tf-en:
(i) Ever$ minimi. r<alization of A has dimension n
-F o r
). (A
(ii) Ail m i n h l realizations ~y be determined with the aid
- ( ~) and A': = htr(!- =r)
of B. L. Hols formulas (6.1'7-1E) with h' = h 1 =r

- A~(A.
(iii) If r 2 =r) + A"(A=r) then t'ne ninhsl realization
is unique. OtheMse there are ES ~ z ; z )S
- -1 s~zlizetf~ns
as

there m e extensions of -4 satiswng (9.61.

=r

PROOF. By the &in Le1m3 (9.5), every pzstial seqmaee A=r

has at least one W M t e extensioa vMch preserves hl, htr ~ ~ n d
n0, So we can apply the (8.21) of the prece- section.
It follot~sthat the m i M partial realization is uniqce if
r2
- A~(A=r) + =r)
htl(A (the A ~ (=IA). + hn(A=r) +1 IIa&sl matrix ca2 be
filled in completely wiYn the available dztz); in ?fie coz$rary case, the
m-i extensions will depend on the m e r in which the =trices

AM -9 have been d e t d a e d (su3jcct to the requirenent

(9.6) 1 ,

In via*- af the theorez, we are ;ustified- in c m i r g the integer

..----
*A sirnilar result vas obtzined similtaneously m d independently
by ?.! . Tether (stanford dissertztion, 1969).
 - 118 - R. E,Kalman
(9.8) lEMAEE. The essential point is bhat the qua~ities n0'
A', and A" &re Wquely detsrmined already from partial &ta,
irrespective o f the possible nonuniqueness of the minirea.l extensions
of the partial sequence. We warn, however, that this result does
not generalize to a l l invariants of the minirnaf realization. For

instance, one cannot determine from A* how q cyclic pieces z

=&

realization cr" A ..
=r will h2ve: same xrmumaJ realizations
be cyclic and others not [UT 1905!

JiSaKIy, let us note also a second consequence of the Main

Theorem:

(9.9) COROLIARY. u p s q(&)is the number of independent

columns of the Hankel array of A (defined analogously witk.
=r

Jf %(A=r) > no(A=r) then, csing the Main Theorein,

PFOOF.

we get a contradiction 30 the fact that ths rank of my Hankel matrix

of an infinite sequence is lower bourd fa- the dimension of any reali-
zation (Proposition (8.4) J . If =r)
q(A < no(A=r) theu exkending A=r
to we contradict the fact that r e - is at leas-;;

eq- to no(%, . n

In other words, the ~~arecteristic

progertr of rmk, that
cowting rank by row or cclumn dependence y.ief%sidentical results,

is preserved even f ~ r
incq1ete Hankel arrays.
It is u s e m to check a simpl~case which KUustrates some of
the techicalities of the proof 0-2thz Fain Lemma.

(9.1~) -PU3. The dimension of (O, 0, ..., 0, ~ l is) precisely

r X p, where ' p = rank Ar &mi ht = Aft = r.
 10. GENERAL THEORY OF OBSERVmILITY

In t h i s concluding section, we wish t o discuss the probleu of

obse~vabilftgCn a rather general setting: we will not assume

linearity, a t least i n the beginning. This i s an ambitious program

and leads t o many more problems than results. S t i l l , I think it i s

interesting t o give some indication of the difficulties which are

conceptual as well as mathamtical. This discussion can also --

serve as an introducsion t o very recent research E
m 1969a.
190aI on the observability problem I n certain classes of nonlinear

systems.

!l%e motivation for t h i s section, as indeed for the whole theo-ry

of observability, stems from the writerr s discovery [IQUXAN 1969a]

that the problem of (line=) s t a t i s t i c a l prediction and f i l t e r i n g

can be formulated and resolved very effectively by consistent use

of dymmical concepts and methods, and that t h i s whole theory i s a

s t r i c t dual of the theory of opt- control of linear systems with

q m a t i c Lagrangian. For those who are f d l i a r with the stan6ard

classical t h e o ~ yof statistick1 Filtering (see, for instance, YAGLGM

[1962]), we can summarize the situstion very simply by s:,! . ng t b C ,

Wiener-Iblmogorov f i l t e r

+ theory of finite-dimensional linear dyllamical systems

= Kalman f i l t e r .

For the h % t e r , the original pa2ers a r e ElCADfAN 1960~,i963al snd

C anci BUCY. 19613 .

The reader interested i n ?krther details and a m3dern exposition i s

referred especially t o t5e monograph of KAIN9N [ig69b].

We shall exanine here only one aspect of this theory ( ~ c h

does not involve ZII~ stochastic eleme~ts): the s t r i c t formulatFon

of the " M t y principle" betveen reachability and oSser-ability.

!Chis principle uzs formally stated fgr the f i r s t t i m e by gAIHW [ 196Cc 1, but

the pertbent discussion i n this paper i s limited t o the linear case and

i s samewhat ad-hcrc. Aided by research progress since I*, it i s

m w possible t o develop a completely general apprach t o th% "duality
principle". We shall do tkis md, as a by-product, we shsll obtain
a new md s t r i c t l y deductive proof of the principle in the now
classical linear case.

W e all introduce a generzl notion of the "dual1' system, am!

use it t o replzce tt=problem of observability by an eqlivalent

problem of 2eachabiEty. I n keeping with the poiat of view of the

earlier lectures, we shall view a system i n t z r m of i t s input/output

mag f and dualize ? (zather than z). The constructibility

problem w i l l not be of direct interest, since its theary i s similzr

t o t'mt of the observability problem.

Let R, r be the ozne sets as defined i n Section 4 and used

froffi then on. We ass- that botk B and 1' are K-vector s-es

(k = arbitrzry field) and recall the definition of the s3ift

operators o9 and c on R and r (see (3.10)). We denote

both shift operators by z but ignore, -anti1 later, the K[ 21-

module structure on R and r.
 .
R. E Kalnlan

By a constant (not necessariiy linear) input/output map

f: R +i' we shall mean any map f which commutes with the shift
&erators, that is,

Let tls now formulate the general problem of this section:

its canonical realfzation C, and an input sequence V € i2 applied

after t = 0. Determine the state x (jf C at t = 0 -
from
the knowledge ofthe output sequence of 2 after t = 0.-
This problem cannot be solved in general! To see this, recall
that the state set Xf of f nqy k *%wed as a set of fluictiors

since a' is Nerode-equivalent to u, iff

Giving V € S l and the corresponding output secuence amounts to

giving --ious values of f(a0 (1) (namely those correspondulg

0)

to the sequences $, Vr, zVp + Vr-l, ...,

2
V, zV, z V, ...), and

it may happen that these substitutions do not yield enough values of

the f'uuctZon, f(oo-)(I) to determine the h t i o n itself. This
situation has been recognized for a long time in automata theory,
where, i n an &host self-explanatory terminology, one sqrs t h a t

"E i s i n i t i a l - s t a t e determinable by an i a f i n i t e nailtipie experiment

(possibly infinitely many d i f l l r e n t v t s ) but not necessarily by a
single experiment (single V chosen at w i l l ) . See MOORE [ 19561.

The problen i s k t h e r complicc.ted by t h s f a c t t h a t it nay make a

ciifference &€%her or not we &ve a free choice of V. KAIWN,

F U , u d P-SBLB [1969, Section 6 . 3 ) ] give some related commznts.

A fhrtller a f f i c u l t y in3erent i n the preceding discussion i s

that the problem i s posed on a purely set-theoretic l e v e l and does

not lent2 i t s e l f t o the introduction of more.refined structural

assumptions. We shall therefore reformulate the problem in such

a way as t o focus attention on determining those properties of t h e

i n i t i a l s t a t e ufiic'n can be co~putzdfrom t h e combined knowledge of

thc input and ou%put sequence occurring a f t e r t = 0.

For simplicity, we shall fix the value of V at-..O (no loss of

generality, since f i s not linezr) . Then the output sqAence

resulting from x a f t e r t. = 0 i s given sinply as f(w), where

We &'s use the circumflex t o denote certain classes of

f h c t r a n s from a s e t i n t o the f i e l d K, For the moment, t l ~ i s

class w i l l be the class of a l l fhnctions. Thus

? = ( a l l functions I' +K}.

An element ;of - ?!. i s simply a " r u l e w ( i n practice, a computing

a l g o r i t w which a s s i g s t o each possible output seqL1t~-

--fie y in I'
 R. E. Kalman

a number in the field K. If f resulted frm the state x = ["If,

then

A
gives the value of a certain function in R and, by definition of
the state, also the M u e of a certain function in f. This suggests
the

(10.2) ~ i t W C I ~An
. element € 8 is hn c5sonable costate
iff there is a ?$ E f? such that we have identically for a l l
coE n

In other words, no matter what the initial state x = [uIf is,

the value of ;at x can always be determined by applying the
rule ?; to the output sequence f(m) resulting from x. Note,
care-, that this definition subsumes (i) a fixed choice of the..
class of functions denoted by the circumflex, and (ii) a fixed inpct
sequence.after t = 0 (here V = 0). For certain purposes, it
may be necessary to generzlize the definitfon in various ways
[W~YLN
1g0 a], but here we wish to avoid all unesseatisl complica-
.. .
tions.
According to Definition (10.2), we shaU see that a system is
completely observable iff every costate is observable. !Chis agrees

vith the point of view adopted earlier (see Section 4) in a,n ad-hoc
. .
fUhio3. Also, the vague requirement to "determine xI1 used in
(10.1) i s now replaced by a precise notion which c a be
~ manipulated

(via the actual definition of the circumflex) t o express limitations

on the algorithms that we m a y apply t o the output sequence of the

system.
The requirement "every costate i s observablett can,be often
replzced by a much simpler one. For instance, i f X i s a vector

space, it i s enough t o know that Itevery linear costate i s observablett

or even just that "every element of some dual basis i s an observable
costate"; i f X i s an algebraic variety, it i s natural t o interpret

"complete observability" as "every element of the coordinate ring of

X is an observable costatett [ K A L 2 .l%Oal.

We can now carry out a straightforward ttdualization" of the

setup involved in the definitioc of the input/out-out map f: S2 +r.

First, we adopt (again with respect t o a fixed interpretation of tine
circumflex) :

(10.3) DEFINITION. -
The dual of - an input/output mp f: R +I'
i s the map

Dote that ? i s well-defined, since the circumflex means the class

As t o the next step, we wish t o prove that constancy i s inherited

under dualization. To do this, ws h v e t o induce a definition of the

.a
shirt operator on 'I and 6. The only possible definitions are the

obvious ones :
 R. E. Kalman

. -Both of.these new shift operators will be denoted by z-la

The reason for this notation will become clear hter.

Now it is easy to verify:

(10.4) -
PROPOSITION. If f is constant, so is f.
A

PROOF. We apply the definitions in suitable sequence:

( z )(a) = (z-'-=?) (f(m) (def. of 3,

= ;(z.f(o)) (def. of q),
= ;(f(z-u)) (f is constant),

= ?(?)(z*u) (def. of ?j,

-1 A A
= (2 -f(Y))(u) (def. of u~),

A
and so we see that f commutes with z whenevzr f does. 0

A
At this stage, we cannot as yet view f as the input/output map
A
of a anwical system because concatenation is not yet defined on ,'I
m d therefore hr is not yet a properly defined "input setw.
In other words, it is necessary to check -;hat the notion of time is
also inherited under dualization. In general; this does not appear
A
to be possible without some strong limitation on the class r. Here
we shall look only.at the simplest
 R. E.Kalman

(10.5) HYRIZIIESIS. Every function j P s a t i s f i e s the

finiteness cond.ition: There i s an integer 1 (dependent on $)
such that for a l l r, 6 E r the condition

y
, = a, k = 1, - * * , I?/
implies

( 1 = ?(6)-

I n other words, we assume that the value of each y^ a t y

i s uniquely determined by same fin2te portion of the output sequence

r.
Assuming (10.5)~ !t i s immediate t h a t P admits a cor-catenation
multiplication which corret20nds ( a t l e a s t intuitively) t o the usual
one defined on R:

We can now prove the expected theorem, which may be regarded

as the precise f o m of the "duality" principle:

(10.7) THEOREX. Let f

.
be an arbitrary constant input/output

map and ? i t s dual. a y s e f'urther t h a t (10.5) holds. - Then

each observable costate of f (relative t o P satisfying (10.5))
may be viewed a s s reachable s t a t e of ?, an4 conversely.

- PROOF. F i r s t we determine the Herode equivalence classes on

F induced by 3. .By definition
 R. E. Kalman

h
for a l l 2E I?. Now 3 i s l i n e a r (!) ; i n fact, direct use of

the definition of Clf and (10.6) gives

So ?of and 8.f are equal as elements or i: =hey define the

sane observable ~ o s t a t e . Tr, fancier lenguage, the assignment

i s well defined and constitutes a bijection between t h e reacha3le

A
s t a t e s of f and those costates of f which a r e observable

relative t o the G c t i o n class ".

Thus (10.5) i s a sufficient. condition for the 5 a J i t y principle
n
t o hold. However, &e f a c t tha% t'ne canonical realization af f is
completeQ- reachable i s not quite the same a s saying thet the canonical

realization of f i s completely obserwble because the . l a t t e r depends

on the choice of
..r a d therefore is not an i n t r i n s i c property of f.

Moreover, Theorem (10.7) does not give aay i n d i c a t i m how l1bigv X? is

and it may certainly happen that the o b s e m b i l i t y problem f a r f .is

a.eh more d i f f i c u l t 'than the reachability problem. These matters w i l l

be i l l u s t r a t e d l a t e r by some exaniples..

Now we deduce the original form of the duality principle from

Theorem (10.7). The essential point i s t h a t (10.5) holds automati-

c a l l y as a r e s u l t of linearity.
New definition of the +tion class: l e t the circumflex denote

the class of a l l K-linear f'unctions. (All the u n d e r l y i q s e t s wttb the

K-vector smces, so the definition makes sense.)

 R. E.Kalman
The following facts axe well known:

(10.9) PROPOSITION. * denote duality i n the sense of

K-vector spaces. T3h:

Now we can s t a t e the

(10.10) MAIS THEORFSI. Suppose f K-linear, constant, f i n i t e -

-
--dhsnsional. Suppose further t h a t " means K-linear dvality. !!ken:

A
(i) f K-linear and constant, t h a t i s , a K[Z?']-homomorphism

(and therefore written a s - fl) and finite-dimensional.

(ii) The reachable s t a t e s of r" a r e isomorphic with the
K-linear dual of 'Xf; hence every costate of X i s observable.
f

PROOF. The f a c t t h a t l? i s K-linear implies, by (10.3))

that i s K-linear; the constancy of f always implies t h a t of

-
A h

f, by Proposition (10.4). (caution: f i s not t h e K[z]-linear

dual of the .K[z]-homomorphism f, and the construction given here

cannot be simplified. See Remrk (4.26.) ) .
To prove the second part, we note t h a t by Proposition (10.9)

Hypothesis (10.5) holds and thus 2 = f* is G well-defined input/output

map-of a dynamicall. system. We must prove t h a t t h e reachable s t a t e s

of P are isomorphic with q, the K-linear dual of Xf. This

amounts t o proving t h a t the K-vector space of functions

i s isomorphic with the k v e c t o r space XXf. It suffices t o prove
thst t h e K-vector space generated by the K-linear functions

i
(10.~) (A: x u [hfje .x)li, i = 0,1,... and3 = l,...,~]

i s isomorphic with .:x Suppose that, f o r fixed x, every A(x) = 0.

men x = 0, by definition of the Nerode eqqivalence relation induced

by f ( r e c a l l here $he discussion from Section 3 ) . Since Xf i s

finite-dimensional by m o t h e s i s , it follows from this property of
the functions ( h ) t h a t they generate .x: Obviously, d i r ~Xf46 = d i m Xp
so that everything i s proved. U

I n other terms, the f a c t t h a t f = K[z]-homomorphism together

with ?he appropriate definition of A implies that

i s a ~ [ ]-homomorphism.
f' Since (10.5) holds, we can interpret
h
f i n a system-theoretic way, a s follcxs: the output of the dual
system at t = -k due t o input i s given by t h e assignment

which i s a linear function defined on the k-th term of the input

sequence. I n fact, we have

 .E. Kalman

(10.12) REMARK. It i s essentially a consequence of Proposition (10.9)

that 2 turns out t o be the same kind of algebraic object a s f . Note,

however, %hat

=der duality the input and output terminals Ere

interchenged and. t i s replaced by -

-t (hence z

In terms of the p i c t o r i a l definition of a system, t h i s

statement simply amounts t o t t r e v e ~ . s ithe
r ~ directions of the arrowsv,

which i s the "right" way t o define duality i n the most general

mathematical context, namely i n category theory. We would exrgect

t h a t the duality principles of system theory- w i l l eventually become

a part of t h i s very general duality theory. This has not happened

yet beczuse the correct categories t o be considered i n the study of

dynamical systems have not yet been determined. It i s l i k e l y t h a t

eventually may different categories w i l l have t o be looked a t i n

studying dynamical problems.

We s h a l l now present an exmple which should help t o i n t e r i r e t

the previous results; We emphasize, however, t h a t the theory sketched

here i s s t i l l i n a very rudinentary form.

(10.13) EXAMPLE. Consider the system C defined by

 R. E. Kalman

with X = U = Y = I
-!mod 1, i.e., the interval 0 1 (1 i s to

be thought of s;s fdentified with 0.) Xe let u(t) = 0, We view

x thzougf: its bi-sa~-yrepreseatstion

It is clear frm tke dzf2?litian or' the s - s t a t h a t the output

sequmce due ta a y x i s preciseu

i Ccmeq~ent3S;i;e x's are fs,-~@&c w2r;h the Nerode eq~iva-

are obszr'i-&le ( m r e prec5s2u9 t.h25e zre Fmtfons which d e z e ~only

on s fLxs6 fiiLite sibset of the Sk(x) 1 s) . Tass:
.
eii2.5r
.
A
f does
na-5 6efi!l.5 *-mic~l system or n ~ t1111 coatai;zs-areabserwbl2,

E03.let us rg~*&c-i;t:e s+t

.. . . . 10, 1) 5y fts fn';ercect:cn

with the , r ~ ~ t i ~ MICis

s . - Fie=
.
tbti there 2s- ~
.
G Wa fznite %oi-lt.im
..

..

prbb~enis to rxpresa r (il(x1, - .

.. .
5pi$0 a-.r&io .

of' plynca'&Ls . ~ E . ' < z , [ ~ ] 3s a

- LaSwys-~br.. l e sSac~,each - x
.-
=d . .

). Itcrcpver,
5 . re;ti~nal,
~ ' .x -.is
r.kk Itel"i'. ,ti*..jr- . c c q ~ t & l . e ~in
~ -the-
s t r i c t sense since there i s no uay of knoving vhen the algorithm

has stopped. I=I other words, given an arbitrary costate -there exists

-
. ~

m fixed rule ?; such that the application of ?; t o rx gives

A
x(xj for a 3 1 x. OIL the other hand, substituting into A; the
results of t'ne partir.1-realization algorithm w i l l give an appro-xi-

mation t o thz value of :(X) which always converges i n a Sirtite

(but a priori unknown) number of steps a s more values of the output

sequewe are observed. In short, the costate-eetermination algorithm

ZELS certair, pseudo-random e l w n t s in it ssd t'nerefore cannot be
described through the machinery of dete-ministic dynamicaL systems.

(IS thzre same relation here t o the concept-& d i f f i c u l t i e s of

QatupMerhanics?)
 R. E.Kalman
11. HISTORICAL CWIEhTS

It i s not an exaggeration t;o say that the a t i r e theory of linear,

constant (am?here, discrete-time) dynamical systems can be v5ewed as
a systematic development of the equivalent algebraic conditions (2.8)
and (2.15).

Of course, the use of modules (over KE zlj t o study a constant

sqmre &rix (see (4.13;) has been "standardtt since the 1920's under

-
the Ibderrs AQebrz of V-4N Dm RAEEUBT.
nust be also qyite cld. For instance, -
the irfluenze of E. NOETHER and especially a f t e r the publication of
Condition (2..15), by i t s e l f ,

11959, Vol. 1, p. 2031

a t w b u t e s t o KRYILlV [ 19311 the idea of computing the characteristic
polynamial of a square matrix A by choosing a random vector b and

computing successively b, Ab, A%, ... u n t i l linear dependence i s

obtained, which yields the coefficients of det (21 - A), he method

w i l l succeed iff XA i s cyclic w i t h generator g.) However, the

merger of ( 4 . ~ ) with (2.15), which i s the essential. idea in the alge-
braic tbeory of linear systems, was done explicitly first i n WMN [1965b].
A

We shall direct our remarks here nainly t o the history of conditions

-
(2.8; and (2.15) as related t o controllabilitx. See also earlier
comments i n KlllMAN [1960c, pp. 481, 4-83, 4841 and in W A N , HO, and
IURENDRA [1963, pp. 210-2123. We will have t o .bear in mind that the

development of modern-control theory camot be separated 'from the develop-

ment of the concept of controllability; mreover, f&e technological
problemsof the 1950ts and even earlier had a ajar influence on the
genesis of mathematical idzas (just as the l a t t e r have led t o many
new technological applications of control i n the 19601s).
 R. E. Kalrnan

The writer developed the mathematical definition of controllability

with applications to control thecq-, during the first part of 1959.
course notes at Johns Hopkhs UnLversitg; 1958/59. ) These
(~n~ublished
first definitic s were in the form of (2.17) and (2.3). F o r d presenta-

tions of the results were made in Mexico City (~e~tesber,

1959, see
KALMAN [lg6Ob]), 1969, see
University of California at Berkeley (~~ril,
[lg@d]), and Moskva (June, 196% see [19ac]), in
and
scientific lectures on many other concurrent occasions in the U.S. As

'far as the writer is aware, a conscious and explicit definition of

controllability which combines a control-theoretic wording with a
prscise mathematical criterion was first given in the above references.
There are of course many instances of similar ideas arisirg in related
corrtexts. Perhaps the comments below can be used as the starting point
of a more detailed examination of the situation in a seminar in the
tiistory of ideas.
The following is the chain of the writerls own ideas culminating
in the publications mentioned above:
(1) In KALMAN [ 19541 it is pointed out (using transform methods)
tbit continuous-time linear systems can be controlled by a linear
discrete-time (sampled-data) controller in finite time.*
---------------
*It is sometimes claimd in the nathematical Etcrature of optimal
-
control theory that this cannot be done with a linear system. This is false;
the correct statement is "cannot be done with a linear controller producing
control functions which are continuous (and not merely piecewise continuousl)
in time." Such a restriction ia~completely'irrelevantfrom the technological
point of view. As a matter of fact, computer-controlled systens have been
proposed a d built for many years on the basis of linear, time-optimal control.
 - 135 -
R. E,Kalman

(2) Transposing the result.of W I A N [ 1954I f'rop transfer functions

to state variables, an algorithm was sketched'for the solution of the
discrete-time time-optimal control of systems with bounded control and
linear continuous-time dpmics. [ KALMAN, 19571

(3) As a popularization of the results of the preceding work, the

same technique was applied to give a general method for the design of
linear sampled-data systems 3y KALMAN and BERTRAM [19581.
Some background comments concerning these papers are appropriate:

(1) The ideas am3 method presented in KALMAN [1954] descend

directly fram ezrlier (and very well known) engineering research on
tke-optimal control. (The main references in KiUHAN [1954]are:
V~~ [ 19501, HOYMN [ 19511, BOGNER and KAZDA [1954], as well as a
research report included in ~ ' 4 i V i[I9551 .) Although the results of .
KALMAN [1954] on linear time-optimal control were considered to be new

when published, it became clear later that simiiar ideas were at least
implicit iri 0I;DENBOURG and SARTORIUS 11951, $90, p. 2.91and in TSYPKINts
work in the early 19501s. The engineering idea, of nonlinear time-optimal
control goes back, at least, to DOLL [ 19431 and to OLDENBURGER in lgbl,
although the lattertsvork was unfortunately not widely known before 1957.
During the same time, there was much interest in the sameproblems in
other countries; see, for ins%ance, FEIl)BP;UM [19531 and UTTLEY and HAMMOND

[1953]. Mathematical work in these problems probably began with 'BUSHAWIS

dissertation [1952] in which, to quote from KAIyA&Y [1955, before equation
(40)1, It ...:.[ it was]rigorously proved that the intuition which led to
the formulation of the [engineering] theory [quoted above] was indeed
correct.I1 TSLENt s survey [19541 contains a lengthy account of this state
of affairs mi was ready by many. -- W.e emphasize*: none of this
extensive literature contains even a hint of the algebraic considerations
related to controllability.

(2-3) !Che critical insight gained and recorded in I C A U U 119571 is

the following: the solution of the discrete-time time-optimal control
problem is equivalent to expressing the state as a linear combination
of a certain vector se¶uence (related to control and dynamics) w i t h
coefficients bounded by 1 in absolute value, the coefficients being
the values of the optimaf control sequence. !l!helinear independence
ofthe first n vectors of the sequence guarantees that every point
in i neighborhood of zero cul be moved to the origin in at no& n
steps (hence the terminology of "camplete controllability"); and the
condition for this is identica w i t h (2.17) (stated in KkLMiUl [1957]
and KAWAN and BERTRAM 119581 only for the case det F #0 and m = 1).
A thorough discussion of these matters is found in KALMAN 11960~;see
especially Theorem I, p. 485 1. A scrims ccmegtual error in ISAlNd
[19571 occurred, however, in that complete controllability was not
assumed, as a hypothesis for the existence of time-opt- control law,
but an attempt was made to show that the controllability is almost

-
always colliplete [ ~ e m a a11. In fact, this lemma is true, with a small
technicel modification in the condition. Only much later did it become
c l e q (see the discussion of Theorem D in the ~ntroduction), however,
that a .dynamical'system is always completely:controllable (in the nonconstsnt

case,. completely, rezch&Z,le) if it is derived from an external descrip+'

don. It was
this difficulty, very mysterious in 1957, which led to the development
 R. E. Kalman

of a formalma@ine?y for the definition of controllability during the

next two years. The changing point of view is already apparent in
KADW and BEqTRAM [19581;the unpublished paper promised there was
delayed precisely because the algebraic machinery to prove Theorem D
wss out of reach in 1957-8. Consult also the findings of the biblio-
grapher RUDOLF [ 19691 .
IN SWIARY: under the stimulation of the engineering problems
of minimal-time optimal control, the researches begun by 'XAfiMAN [1954,
19571 and KAIJl4N and BERTRAM [ 19581 eventually evolved intoiuhat has
came to.be czlled the mathematical theory of controllabilitx (of linear
systems).

Beginning about 1955, ~ u sthulated

d by the same engineerir-8
problems, PONTRYAGIN.andhis school in the USSR developed their
mathematical theory of optimal control around the celebrated "
m upi

Principle". hey were well aware of the survey of TSIEN [1954]

mentioned . above, and referenced it both in Ehglish and in the Russian
translation of 1956.) We now h o w that a _ ~ ytheory of control, regard-
less of its particular mathenatical style, must contain ingredients
related to controllability. So it is interesting to examine how
explicitly the controllability condition appears in the work of WNTRYAGIN
and related research.'
GAMKRELIDZE 11957, $2; 1958 61, $21 calls the time optimal control
problem associzted with the system
"nonrlegenerate" i f f b i s not contained i n a proper' A-invariant

subspace of R
n
. He notes inrmediatelythat t h i s i s equivalent t o

(11.2) det (b, Ab, ..., A%) # 0

(i.e., the special case of (2.8) for m = 1). IIe then proves: &
the "degenerate" case the problem either reduces t o a sjmpler one or

the motion c a n n ~ tbe influenced by t h e control f'unction u(*). A l l

t h i s i s very close t o an explicit definition of controllability.

However, i n discussing t h e general case m > 1, m Z E [ 19.58,

53, Section 11 defines "nondegeneracy" of t h e system

as t h e condition

(ll.4) det (bi, Abi, ..., A ~ - % . )

1
# 0 f o r every column b. E B,
1

but he does not show t h a t t h i s generaiized condition of "nondegeneracy" for (11.3)

inherits the interesting characterization sroved f o r "nondegeneracy"

i n the case of (11.1). I n fact, condition (ll.4) is much too strong

t o prove t h i s ; the correct condition i s (2.8), t h a t is, complete

controllability. I n other words, in G P m I D Z E s work (ll.4) plays

the r o l e of a technical condition fez1 elimina+;9ng "degeneracy" ( a t u a l l y ,

lack of uniqueness) from a pazticular optimal control problem and i s

not,explicitly related t o the more basic notion of complete controllability.

Neither GAMKRELIDZE nor PONTRYAGIN [1958] give an interpretation o l

(11.4) a s a property of tile Qilsslical system (11.3), but employ (11.4)

only i n relation t o the particular problem of time-optimal control. See
 R. E. Kalman

dso [lg60c, p. 484 I. A siular point of view is taken by

USALLE (19601;he calls a dynamical system (31.3) satisoing (2.8)
''proper. but then goes on to requira (~1.4)(to rssure the uniqueness
of the time-optimal controls) and calls such systems vnom%ln.
The assungtion of some kind of "nondegeneracyl'conditioa was
appazently unavoidable in the early phases of research on the time-
optimal control qroblem. For example, ROSE [1953, pg. 39-58] examines
)by defidng "nondegeneracy" [ p . 411 by a
this problem for (u.1;
condition equiw.lelit ot (ll.2), he obtains most of MMKRELDZE~sresults
in the special case when A has real eigenvalues [meorem 121. ROSE
uses determinants closely related to the now familiar lemmas in cantrol-
lability theory but he, too, fails to formulate controllability as a
conce2t independent of the time-optimal control pr031em~
A similar situation exists in the calculus of kiations. The
so-called Caratheodory classes (af'ter CARATHEODORY [19331) correspond
to a kind of classification of controllability properties of nonconstant
systems. In fact, the standard notion of a normi family of extremals
of the calculus of varta5ions is closely related to condition (ll.4),
suitably generalized via (2.5) to nonconstant systerns.* Normality is
.
used in the calculus of variations mainly as alhondegeneracJftcondition.
It is iaports.n",o note that the lrnondegeneracyll
conditions
employed in optim.1.coritro~and the calculus 01 varlazlons play mainly the
role of eliminating annoyim- technicalities and simplifjring proofs.
---------------
by IaSALLE [1960j o r (-11.4)is only
*The use of the word llnorma,lll
accidentally coincident with the earlier use of the "normalw in the
calculus of variations.
With suitable formulation, however, the basic ?esults of time-optimal
control theory continue to hold without the assumption of complete
controllability. The same is not tme, however, of the four kinds of
theorems mentioned in the Intorduction, and therefore these results
are more relevant to the story of controllability than the time-opt-1
control discussed above.
There is a considerable body of literature relevant to controllability
theory which is quite independent of control theory. For instance, the
treatment of a reachability condition in partial differential equations
goes back at least to CHOW [ 19401 but perhaps it is fairer to atkribute
it to Caratheodoryts weil-known approach to entropy via the nonintegra-
bility condition. The current status of these ideas as related to
controllability is reviewed by WEBS [1969, Section 91. An independent
and very explic5t study of reachability is due to ROXLN [1960]; unfor-
tunately, his examples were purely geometric and therefore the paper
did.~ o help
t in clarifying the celebrated condition (2.8). The
Wronskian determinant of the classical theory of ordinary differential
equations with variable coefficients also has intersections with control-
lability theory, as pointed out recently with considerable success by
SILVERMAN [19661. Vany problems in control theory were misunderstood
or even incorrectly solved before the adve1l.L of controllability theory.
Some of these are menttoned in KALMAN [1963b, Section 91. For relations
with automata theory, see m I B [19651.
Let us conclude by stating the writer's own current-,positionas
to the significance of controllability as s subject in mathematics:
 R. E. Kalman

(1) Controllability is basically an algebraic concept. ( ~ s

c ~ applies
h of course also to the nonlinear controllability results
obtained via the PfafYim method.)

(2) The historical development of controllability was heavily

influenced by the interest prevailing in the 1950,s in optimal control
theory. Ultimately, however, controllability is seen as a relative*
minor component of that theory.

(3) Controllability as a t~nceptualtool is indispensable in

the discussion of the relationship between transfer functions and
differential equations and in questiohs relating to tae .four theorems
of the Introduction,

(4) The chief current problem in controllability theory is the

ekension to more elaborate algebraic structures.
For a survey of the historical background of observability,
which would Lake us too far afieldhere, the reader should consult
 R.E.Kalman

Sectisn A: General Teferences

[1965] A common framework for automata theory and control theory,

SIAM J. Contr., 3206-222.
C. W. CUHTIS and I. REINER
119621 Representation Theory of Finite Groups and Associative
Algebras, Interscience-Wiley.
E. M. DAY and A. D. WWACE

[ 19671 Multiplication induced in the state space of an act,

Math. System Theory, &:305-3lL.

C. 1. DESOEK end P. V1.RAIY.A

E 19671 The minimal realization of a nonanticipative impulse
-
response matrix, SIAM J. Appl. Math., 15:754-7&.
P

E. G. GILBERT
[ 19631 Controllability and observability in multi-xiable
control system, S m -
T J. Control, 1:128-151.
B. L. HO and R. E. KAIJIAN

.[ 19661 Effective construction of linear state-variable models

-
from input/output f'unctions, Regelungstechnik, 14:545-548.
[19691 The realization of linear, constant input/output maps,
. . I. Complete realizations, SIAM J. Contr., to appear.

[ 19651 Elements of Modern Algebra, Holden-Day.

[1960a] A new approach to linear filtering and prediction

.
-
problems, J. Basic Engr. (~rarrs.ASME), 82~335-45.

[1960b] Contributions to the theory of a2tjmal conti.01, Bol.

. -
Soc k t . Mexicana, 5:102-119.
 [lgh] On tlle general theory of ~ o n t r o systems,
i Roc. 1st
WAC Congress, Moscow; Butteruorths, Inndon.

j19621 Canorricsl s t m t u r e of liriear e c a 3 systems, Proc.

- Bat. had. of Sci. (USA), 63:5966CQ.

[1%3a] Hew methods in Wiener f i l t e r i n g theory, Proc. 1st Synp.

on Erigineeri~gAppZications of Ransom Function Theory
and Probabf E t y , Puwlue Uciversity, November 1960, pp 270-388,
Yilejy. (Abridge6 from EUS Technical Report 61-1.)

11963bI Bhthematical description of lineax dynamical systems, SIirN

J. Contr., &l!j2-1%.
[1965a] irreducible realizations and th.e degree of a rational
mtrix, SIAM J. Cantr*, 13:5X)-544*
[1%5b] Alge3rec structure of linesr Qmmical systems. I. The
-
M u l e of Z, Proc. Eat. Acad. Sci. (w), 54:1503-~08.

[i*?] Algebraic aspects of the theory of dynemical systems, 5n

1.i'ferent2al q q l s t i o ~ sand DynamicaZ Systems, 3, K.
and J. P. IaSalle (eds.), pp. U3-1k6, Academic Press.

[l-1 On multilinear mechines, J. Comp. and System Sci., to

appear-
f :$%! Dynamic Prediction and Filtering Theory, Spriwer, t o
appear.

[19&] On p a r t i a l realizations of a linear irlput/output map,

Guillemin f + + v e r s uVolume,
y IIolt, Winston and Rimhart.

flY?&j Observability in multilinear system, t o appear.

[1g0b3 The rezI3zation of linear, constant, input/output maps.

II. Partial recilizatiom, SIAM 3. Control, t o cppear.

B. E. W-Ui and 2. S. BUCY

[1961] Hew resv1ts i n linear prediction and f i l t e r i n g theory,
-
A S E 2 Ser. D), 8 3 ~ : ~ - 1 0 0 .
J. Basic h g r . (~r211~.

R. E. KAUWI, PI L. F .and H. A. m B

[19691 T o ~ isc i n Fathematice1 System Theary, HcGray-Hi=.

R. E. lm,rm$ Y. C. iIO W d K . EAREND-*

El9631 Controllability of linear dynamical systems, Contr. t o

D i f f. Equations, 1:1 8 9 - a .
7

[1964] an the stabilization of systems, Proc. Am. Ekth.

SW., 3:735-7$2. .
 R, E. Kzlman

[ 2965 1 Algebra, Addison-Wesley.

[ 13631 Hoaolom, Springer.

1
1s1 CmtroUability of nonlinear processes, SIAM 3.
-
Control, 3378-90.

[ 19563 Gedanken-experiments on sequential =chines, in Automzta

Studies, C. E. Shannon and J. UcCarthy (eds.), pp. 129-153,
Princeton University Press..

fa991 Theorie und Anwendung der Elenentartheiler, Teubzer, Leipzig.

A. NERODE

!
[ ~958 Linear mtozatori trazlsfon;lations, hot. Arner. Nth. Soc .,
-9:541-5&.

115661 Representation and realization of time-uzriable line=

system, Doctoral dissertation, Columbia University.
L. M. SILT4EEXAIu'and H. E. b!EADOl;iS

119691 Eqivalent realizations of linear sysi;ems, SIN-f

J. Zmtrol, to appar.
Y. IdEBEP

118981 Lehrbn& &r Algeixa, Vol. 1, 2nd Edition, reprinted by

Chelsea, h'ew York.

f 196.31 Lectures on Controllability and O>servzibilit,jr, C.LtM.E.

seminar.
L. WEISS and R. E. mu3
i 196 1 ~ontribut~ons to linear systen theory, Intern. 3. Engr.
Sci., 2:lkl-l7l.

[ 1$7 1 On pole assignment in multi-i~p~t

controllable linear
systems, IW3 Trans. A~to.Contr., AC-12:600-665.
 [1962] An Introduction to the Theory of Stationary Random
Functions, Prentice-fiall.

El9661 The synthesis of linear dynamical systems from prescribed

weighting patterns, S U M J. Appl. Mat'n., &:527-549.

C. C. YOULA and P. TISSI

[19661 n-port synthests via reactmce extraction, Part I,
Intern. Convention Record.

[ 15581 CommutatZve Algebra, Vol. 1, Van Nostrand.

Section B: References for Section 11

[1965] A common framework for automata theory and control

-
theory, IAM.J. ~ontr., 3 :206-222.
I. WNR and L. F. ICAZDA
E

119541 An investigation of the switching criteria for higher

order contactor s e r v a m e n ~ s m s ,Trans. AIEE, I1:ll8-127.

119521 Differential equations with a discontinuous forc-

term, doctoral dissertation, Princetan University.

# #

[ 19331 Uber die Einteilung der Variationsprobleme von Lagrange

nach Klasse~, Corn. Mat. &lv., -
5:l-19.

[ 19401 h e r Systeme von linearen partiellen Differentialgleichungen

erster Ordnung, Math. Annalen, :98-105.
If. G. DOLL

119431 Automatic -cont?ol system for vehicles, US Patent

2,463,362 -

[ 19571 On the theory of optimal processes in linear systems

-
(in Russian), DoR1. Akad. Nauk SSSR, =6:9-~.
119581 The theory uf optimal processes in linear systems
-
(in Russian), Izvestia Akad. Nauk SSSR, 2:449-474.

[1959] The T h e o u f V~trices, 2 vols., Chelsea.

 C 19511 A phase-plane approach t o the campensation of saturating
servomechanisms, Trans. AIEE, '/0:631-639.

119541 Discussion of a paper by Bergen and Ragazz-X, Trans.

e,
73 11: 24-5-246.
119551 Analysis and design principles of second and higher-
-
order saturating servomechanisms, Trans. AIEE, 74 11:29-310.

119571 ~ p t i m a lnonlinear control of saturating systems by

intermittent control, LRE IIESCON Convention Record,
-
1IV:l33-=>.

[1960b] Contributions t o the theory of o p t i m l control, Bol.

. .
Soc Y a t Mexizana, 5:102-llg .

[19&] On the general theory of control systems, Proc. 1st

UAC Congress, Moscow; ktterworths, London.

[1960d] Lecture notes .on control system theory (by M. Athans

and G. ~endaris),.Univ. of Calif. a t Berkeley.
, .

[ 1963bl bIathematf cdl descristion of linear dynamical systems,

-
SUM J. 'Contr., 1:152-192.

11965bI Algebraic structure of linear dynamical systems. I. The

Yadule of .C, Proc. Nat. Acad. Sci. (USA), &:1503-1508.

e 19@bI Dynaqic Prediction and Filtering !PeoryL Springer, t o appear.

R. E. YNJ'&J aL: %. E BERTRAM

[ 19581 General synthesis procedure for computer control of
single and rrulti-losp linear systems, Trans, AEE,
77 T?I:€t-'2-609.
R. E. =IAN, Y. C. HO and K. X C I R A

~19631 Controllability of linear dynamical systems, Contr. t o

Diff. Equations, 1:189-213.
 [19311 On the numerical solution of the equation by which
the frequency of smll oscillations i s determined in
technical problems ( i n Russian), Isv. Akad. Nauk SSSR
Ser. Sf x.--Mat., -
4:491-539.

[1 9 a1 The time-optiml control problem, Contr Nonlinear .

Oscillations, Vol. 5, Princeton Univ. Press.

11%ol Nonlinear techniques for improving servo performance,

Proc. mt. Electronics Conf. (USA), 6:400-421. -
R. C. OiJ)ENBOTJRG and K. SARMRNS

[1'311 Dynamik selbsGttiger Regelungen, 2nd edition,

Oldenbourg, Mumhen.

R. om-
[19571 @timum nonlinear control, ~rans. ~ O E , ~:527-546.

[19661 Optimal and Self-OptimizLng Control, MIT Press.

[l9581 optimal control processes ( i n .Russian), ~ s p e k n i at.

I&, &:3-20.

I. J. ROSE
[I953 1 Theoretical aspects of l i m i t control, Report 459,
Stevens I n s t i t u t e of Tech., Hoboken, N. J.

[lga Reachable zones i n autonorrous differential sys<e.m,

. - -
Bol. Soc. Mat. Mexicana, 5:125-=5.

119691 On same unpublished works of R. E. Kafman, not t o be

unpublished.
H. S. TSIEN
119541 Engineering Cybernetics, McGraw-Hill.
A. M. UTTLEY and P. H. HAMMONI)
[ 19531 The stabilization of on-off controlled servomechanisms,
in Autcmtic and Vanuzl Control, Achdenic Press.

19691 Lectures on Controllability and Ob~e~yability,C :I.M.E.

Seminar