PROCEEDINGS OF THE IEEE, VOL.6164, NO.

1976
1, JANUARY

parameter systems,” Trans. ASME J. Basic Eng., pp. 67-79, [ 1451 T. K. Yu and J. H. Seinfeld, “Obsewability of a class of hyper-
Mar. 1964. bolicdistributedparametersystems,” IEEE TransAutomat.
[ 142 ] E. D. Ward, “Identification of parameters in nonlinear boundary Contr., (corresp.),vol. AC-16, pp. 495-490, Oct. 1971.
conditionsofdistributedsystemswith linear fields,”Ph.D. [ 1461 V. P. Zhivoglyadovand V. Kh. Kaipov,“Application ofthe
dissertation, Purdue Univ., Lafayette, IN,Aug. 1971. method of stochastic approximations in the problem of identifi-
[ 1431 E. D. Ward and R. E. Goodson, “Identification of nonlinear cation,” Automat. Telemek., vol. 27, pp. 54-58, Oct. 1966.
boundary conditions in distributed systems with linear fields,” [ 1471 -, “Identification of distributedparameterplants in the
Trans.ASME J. DynamicSyst. Meas. Contr., vol. 9 5 , pp. presence ofnoises,” in PreprintsIFAC Symp., June, 1967,
390-395, Dec. 1973. Paper 3-5.
[ 1 4 4 ] M. J . Wozny, W. T. Carpenter, and G. Stein, “Identification of [ 1481 V. P. Zhivoglyadov, V. Kh.Kaipov,and I. M. Tsikunova,
Green’s functionfordistributed parameter systems,” IEEE “Stochastic algorithms of identification and adaptive control of
Trans. Automat.Contr. (comesp.),vol.AC-15, pp. 155-157, distributed parameter systems,” in Preprints IFAC Symp. Con-
Feb. 1970. trol of Distributed Parameter Systems, June 1971, Paper 13-1.

Nonlinear Systems and Differential Geometry

ROGER W. BROCKETI’, FELLOW, IEEE

Abstract-Nonlinear systems which are governed by a fmite number In spite of the rather general form of nonlinear systems we
of ordinary differential equations with controls present constituk a work with here, most of the results given here could be stated
large andimportant c k d model for proclical purposegInthe in even greater generality. For example, the restriction to one
last few years, there h x been considerable progress m our understmd-
ing of this class of model. This is an expository paper devoted to sur- input is completelyunnecessary; many resultsapply to i =
wying and explaining m e of the main results currentlyavailable. f ( x , u ) , etc.It seemed counterproductiveto move in this
@ackgoundmaterial on m d d d theory is included in ader to make direction, however,ina paperintended as an expository
&e paper more nearly seifcontained. introduction.
We introduce a few standardnotations. R n standsfor
I.INTRODUCTION
n-dimensional Cartesian space with itsordinary topology,
HE PURPOSE of this paper is to describe some of the if x E R n then IlxII = ( x : + x : + . - * x ; ) ’ ’ * . We use letters
main theoretical results onthe class of input/output such as M and N for differentiablemanifolds (see appendix
(I/O) models which take the form for definition). We always assume that f(-), g(.), and h ( . ) in
(1) are infinitelydifferentiable andoften we assume even
more, that they depend analytically on their arguments. We
call systems of this latter type linear-analytic systems.
The most significant results relate to a) controllability theory, The special case of bilinear systems
b) 1/0 theory based on Volterra series, c) isomorphism t h e e
rem and bilineaxization, and d) stochastic theory. i ( t )= A x ( t ) + u ( t ) B x ( t ) ; y ( t ) = (c, x ( t ) > : x ( 0 ) = x o
Our intention is to give an introduction to this area. We will has received a great deal of attention in the literature in the
h e a number of the most important results but we do not past few years.Although this theory is,itself, quite a suc-
give proofs unless there are special circumstances which make cessful generalization of the linear theory, it now appears that
it desirable. The material to be discussed has a great deal of it should be viewed in the context of the more general class
intuitive content but it also requires some technical develop of systems described here. The reasons for this are that many
ments. In fact, the solution of some of the central problems of the principleresults on bilinear systems are truefor this
required the development of some new pure mathematics. In wider class of systemsand the class of bilinear systems
an attempt to present a blend of intuition and solid theory, has the unfortunate propertyof not being closed under compo-
we have included in an appendix the definition of many tech- sition and feedback.This is not to say that bilinearsystems
nical terms and also precise statements of a few background are of littleinterest. On thecontrary,they are the most
theorems. interesting specialization of this theory, and, in fact, they a p
pear from the point of view of the theory of linear-analytic
ManuscriptreceivedApril 23,1975; revised July18, 1975. This systems to be even more general, and hence of greater interest,
paper waswrittenwhilethe authorheldaGuggenheim Fellowship.
This work was supported in part by the U S . Office of Naval Research than was first expected.
undertheJointServicesElectronics Program Contract N00014-75-C- With regard to the literature, there is a recent conference
0648.
The authac is with the Division of Engineering and Applied Physics, proceedings [ 11 which contains papers on most of the topics
Harvard University, Cambridge,M A 02138. to be discussed here. In particular, the paper by Lobry is rele
horized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions ap
 62 PROCEEDINGS
JANUARY OF THE IEEE, 1976

vant for Section I1 on controllability. The paper by Sussmann

is appropriate for Section IV. The papers by Clark and Elliott
will provide the details behind the material discussed in Sec-
tionIV. Specific citations will begiven in theappropriate
places.
Fig. 1 . Motivating the definition of ubiquitous bracket of Lie.
11. CONTROLLABILITY
We begin with a discussion of aresult of Frobenius.This
theorem is fundamental to the geometric theory of control. is the secalled Lie bracket (also sometimes called the Jacobi
In fact, onecould say that it is t o control theory what the bracket, also sometimes defined with the opposite sign) of the
Cauchy-Lipschitz existence theorem for ordinary differential vector fields f1 and f 2 . Thiscalculation,which“everyone
equations is to thetheory of autonomous models found in should doonce in his life” is most significant. Everything
analytical mechanics, electrical circuit theory, etc. else depends on it. If [ f l , f2 ] is not a linear combination of
Suppose we have a single differential equation in R” -
fl ,f 2 , * , f m then [fl , f2 ] represents a “new” direction in
i ( t ) = f [ x ( t )xl (; =
0 )x o E R” (2) which the solution can moveand the original problem of
-
finding a manifold such that f 1 , f 2 *, * ,f , span the tangen
with f smooth enough to definea uniquesolution.Inthe space will not be solvable.
theory of differential equations, oneordinarilyindicates the We introduce a definition which is useful when we want to
value of the solution at lime r , by @ ( r , X O ) unless f is linear in rule out this behavior. One calls a set of vector fields { f l ,
which case the exponential function is displayed. I t is more f 2 , * * ,f m } involuative if there exists yip such that
common in differential geometry to write (exp t f ) x o instead
of @ ( t ,x o ) to indicate the solutionof ( 2 ) regardless of whether
f is linear or not. We will see that the disadvantage of the
k=l
slight ambiguity of this notation is offset by a savings in the
number of parentheses onemust use. One way to describe Some thought about the above construction should persuade
the usual existence theorem of Cauchy-Lipschitz is to say the reader that the property of being involuative is necessary
that there exists a one-dimensional subset of R“,M = { x : x = inorder to be able to “integrate” the set of vector fields
(exp t f ) x o ; It1 < e } such that f [ x ] is tangent t o M at each
point. (A glance at the appendix atthis point is recommended
( f i ,f 2 , - ,f m ) t o get a “solution” manifold. More than
60 years ago, Frobenius established his theorem on “complete
for those who want a definition of a manifold, vector fields integrability” which showed that, under some mild regularity
on manifolds, etc.) Stated in this way, the Cauchy-Lipschitz assumptions, this condition is sufficient as well. Frobenius
theorem answers a special case of the following more general [ 2 ] stated his result as a local result (and actuallya “dual”
question. Consider totheform we discuss here)butsubsequently it has beep
m establishedinaglobal form which asserts the existence of
i(t)=x
u i ( t ) f i [ x ( tx) (l O
, )=xo ER“. (3) maximal (with respect t o theoretic inclusion) solutions.
i- 1 Just as there are many versions of the existence theorem for
ordinary differential equations, there are many versions of
Under what circumstances can we find a smooth p-dimensional
the Frobenius theorem. We give two examples which playa
subset M of R“ such that the f i ( x ) span the tangent space of
role in the control literature.
M at each point? Why is this question of any interest? Well
Theorem 1: (Versions of theFrobeniustheorem.) Let
if the ui’s are controls which can be turned on and off, re-
{ f ( x l ) , f ( x 2 ) ,* * , f ( x , ) } be an involuative collection of
versed, etc., at will, and if such an M exists, then apparently x
vector fields which are
will be able to move anywhere inM but not outof M .
At the first level of complexity, the existence of M hinges a)analyticon an analytic manifold M . Then given any
on the following observation. Ifwe follow the integral curve point x . E M there exists a maximal submanifold N con-
of i = f l ( x ) for t units of time, then i = f 2 ( x )f o r t units, then taining x . such that { f i ( x ) } spans the tangent space of
i = -fl ( x ) for t units and for t units of time and i = -f2 [ x ] for N at each point of N .
t units of timethenthe resulting point is (exp - t f 2 ) (exp b) C” on a C” manifold M with the dimension of the spai
-tfl ) (exp t f 2 ) (exp t f l ) x o (see Fig. 1). Working out the value of { f i ( x ) } being constant on M . Then given any point
of this product based on repeated use of the second-order x . E M there exists a maximal submanifold N containing
expansion x . suchthat { f i ( x ) } spans thetangent space of N at
each point of N .
Example 1: Consider the vector field on R defined by the
function x . R is an analytic manifold and this vector field is
yields analytic. According to a) for each point in R ,there is a maxi-
mal submanifold of R such that x spans the tangent space of
this submanifold at each point. In this case, we see that there
are three distinct submanifolds of. R which arise in this way.
If x . is positive, then N is the positive half-line (0, m); if x0 is
negative, N is the negative half-line (-00,0);and if x0 = 0, it
is the zerodimensional manifold consisting of 0 alone.
af2 afl Example 2: We give asecondexample with a littlemore
[f2,fl 1 = Zfl - Z f 2
geometric content. Consider threeanalytic vector fields in
thorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions appl
 BROCKETT: NONLINEAR
DIFFERENTIAL
SYSTEMS AND GEOMETRY 63

We say that the Lie algebra is finitedimensional if the real

t x3
vector space {fi} is finite dimensional. For example, the col-
lection of all polynomials form an infinite-dimensional Lie
algebra of vector fields on R' since [ xp, x4 ] = ( p - q ) x P + r l .
On R", the set of all vector fields of the form Ax with A a
constantmatrixforms an n'dimensional Lie algebra since
[Ax,Bx] = ( A B - BA)x-thus giving closureunder Lie
bracket-and since the space of n by n matrices is an n'-
Fig. 2. Maximal integral manifold for the example. dimension vector space.
One can see the reason for introducing the Lie algebra in
R 3 represented by controllability problems. The previous calculationstrongly
suggests that not only is (exp t f l )xo and (exp tf2)x0 in the
reachable set from x0 but also (exp t [ f l ,f 2 ] ) x 0 . In fact,
iterating on thistheme suggests that { exp rf}xo is in the
reachable set if f can be expressed as abracketedcombina-
tionofthe given fi. Let { f i } ~denote ~ the Lie algebra of
vector fields generated by {fi}-that is, take all linear combi-
This collection is involuative since [ f l ,f2 ] = f 3 , [ f 2 , f3 ] =
fl and [ f 3 , f l ] = f 2 , and a t each point inR 3 - { 0) it contains nations of elements of {fi}, take Lie brackets, take all linear
combinations, take Lie brackets, etc. t o arrive at the smallest
exactly two linearly independent vectors ( x l f l + x2f2 +
x 3 f 3 = 0). If we look at any nonzero point in R 3 , sayx = Lie algebra of vector fields which contains {fi} and call this
a,
5 (a,a), then we can integrate this distribution
through that point. For the given point, we get the set
{ fi } L A .
Supposethatthe vectorfieldsin { f i } ~are
~ complete,
Le., that(exprf)x is defined for all - m < t < 00. (If the
N = { x : llxll = 1) manifold is R", this means no finite escape times in forwards
or backwards time.) In this case, there is a group of mappings
as thesolution of thisintegralmanifoldproblem. Why this
of M into itself which is closely connected with { f i } ~and ~
set? Ifwe integrate along one of the vector fields, say f l ,
the integral curve is acircle;x1 is constant, x f + x $ is con-
which is obtained by"exponentiating" { f i } ~ We
~ . explain
this as follows. The set of all C" one to one and onto map-
stant (see Fig. 2). Now given any point on the circle, we can
pings of a C"-manifold onto itself, having theproperty
integrate along .- = fi
f 3 to get a band around this circle, etc.
that the inverse mappings are also C", is called the group of
In fact, this is how one sometimes proves the Frobenius theo-
diffeomorphisms of M and is written diff (M). The set of
rem: riding successive vector fields to fill out a neighborhood
such maps is clearly closed under inversion and composition,
of a given point. Notice that, in this case, the vectors f l , f z ,
3 , are, in fact, tangent to the spherical shell N at each point
justifying the label group. Given f, for each t (exp r f ) defines
a map of M into itself which is just the mapping produced by
as required.
The relevance of the Lie bracket andFrobenius' theorem the flow on M defined by the differential equation i =f(x).
forcontrollabilitystudies comesin via a theorem of Chow We denote the smallest subgroup of diff ( M )which contains
[ 3 ] anditsrefinement by others [ 4 ] - [ 8 ] . Given a wntrol exp tf for all f in {fi} by { exp {fi} } G . It is clear that any
point x in M of the form x = {exp (fi}}~x, can be reached
system in R"
from x0 along solution curves of (3) because in this case
m
i ( t )= ui(t)fi
[x(t)I, ~ ( 0 =x0
) x = (exp t1fa, 1 (exp t 2 f a , )* * . (exp tmfa, 1x0
i=l
and piecewise constantcontrols suffice to carry outthe
we can certainly reach any pointof the form transfer.
A question of major wncern can now beasked. What
x = (exp t l f a , ) (exp f2f%(exp t 3 f a , (exp tmfam)xo the
is differencebetween {exp { f i } } ~ x ~ and
{exp
simply by letting all but one of the inputs bezero with the r e { f i } }GXO? ~ ~ The former is reachable and the latter has the
maining one being one. In particular, given two vector fields appearance of something larger. The theorem of Chowsays
f l and fz we can reach that these setsare,
in fact,
equalunder
rather weak
assumptions.
x = (exp - t f 1 ) (exp - V i (exp tft 1 (exp tf2 1x0 Theorem 2 (Versions of Chow'sTheorem): Let { f l ( x ) ,
fz(x), - -
,f,(x)} be acollection of vector fields such that
This suggests that we can move in the direction [ f l , fi ] even
the collection {fl (x), f2 ( x ) , -
,f, (X)}LA is
thoughthis direction maynot be in the linear span of the a)analytic on an analytic manifold M . Then given any
{fi}. To push this idea to its logical conclusion, we need the point x. E M , there exists maximal
a submanifold
following definition. N C M containing x0 such that { exp { x i ) } ~ x=~
If { f i } is a collection of C" o r analytic vector fields on a { ~ X P{XiILA 1 ~ x =0 N .
manifold M , then we say that {fi} is a Lie algebra of vector b) C" on a C w manifold M with dim(span { f i ( x ) } ~ ~ )
fields provided that constant on M . Then given anypoint x. E M , there
existsamaximalsubmanifold N C M containingx0
i) { f i } is a real vector space with respect to ordinary vec- suchthat { exp {xi}}~xo = { exp }cx0 = N .
tor addition andscalar multiplication,
ii) if fi and fk belong to {fi}, then the Lie bracket [ fi, f i ] Clearly this result goes a very long way toward answering
belongs to {fi}. the controllability question for the system in (3). Most often
thorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions app
 64 PROCEEDINGS
JANUARY OF THE IEEE, 1976

in applications, however, we are interested in systems which

have a drift term,Le., systems of the form
i ( t ) = f [ x ( t )+lu ( t ) g [ x ( t ) 4
l ,0=)x o . (4)
As a result about controllability of this class of systems, the
mainlimitation on Chow’s theorem is that it does not dis- N
tinguishbetween positive and negative time. That is, the Fig. 3. Controllability approach to parametric amplification.
submanifold whose existence is guaranteed by Theorem 2 may
include points which can only bereachedby passing back- We consider a particular model for a child pumping a swing.
wards along the vector field f ( x ) . This means that while the Thequestion being that of determining if controllability
reachable set from x . will always be contained in the mani- theory can predict the existence of pumpingmodes which
fold defined by Chow’s theorem, in general, it will be a proper will maintain or increase a given amplitude of oscillation (see
subset of this manifold. The mostaffirmative statementis Fig. 3). Themodel consists of a mass which moves up and
the following result proved at various levels of generality by down along a weightless rod (the child changes its center of
Krener [ 7 ] ,Lobry [ 6 ] ,and Sussmann and Jurdjevic [8]. mass). We ignore, to cut down the length of the formulae, the
Theorem 3: Suppose f and g are vector fields on a manifold mass ?f swing itself. The state space fo the problem is M
M and suppose that {f, g } meets either of the conditions for { ( 8 , 8 , 2 ) : 0 < 8 8 2 2 n ,O E R , O < I < I o } . Settingthe time
Chow’s theorem. Then the reachable set for (4) contains an rate of change of the angular momentum about the support
open subset of the manifold N = { exp {f,g}La } ~ x ~ . equal to the torque dueto gravity gives
A slick proof of this theorem is given by Krener (see also
Lobry’s survey in [ l ] ) . This theorem underscores the rela- d
tively meager state of our knowledge regarding the reachable - [ e ’ ( t ) P ( t ) m ]+ l ( t ) m sin e ( t )= 0.
dt
set of points for systems witha drift term.
Example: Consider the problem in R 2: Differentiating gives the system
e(t) + [ 2 i ( t ) / z ( t ) 1 e ’ ( t+) [ 1 / z ( t ) 1sine(t) = 0.
Now letting e’ = $ and regarding i = u as the control, we get
a set of three equations
if u ( t ) and u ( t ) are arbitrary, then an application of Chow’s
theorem shows that any point in R Z - {0} can be steered to
any point in RZ - (0). On the other hand, if u ( t ) is con-
strained t o be one, giving us a drift term, then the reachable
set is difficultto describe precisely. We can say that if x1 (0)>
0 and x z ( 0 ) > 0,then x l ( f ) , x z ( will
t ) lie in the positive quad-
rant for all f > 0 so that the Teachable set is not all of R 2 - Notice that with u e 0, the solution is periodic so we may a p
Io}. ply Theorem 3 in computing the reachable set (provided we
In order to be able t o presentsome interesting examples, stay away from a neighborhood of 8 = n where the period of
we state a rather special circumstance under which the drift the free motion is not bounded). Let the right side be f + ug.
term causes no difficulty. Then the Lie brackets include
Theorem 4: Supposethat f and g arevectorfields on a
manifold M. Suppose that { f , g } meet either of the con-
ditions of Chow’s theorem and suppose that for each initial
condition x . the solutionof
i(t)=f[x(t)l

is periodic with a least period T ( x o )< M . Then the reachable

set from x . for (4) is { exp {f,g}LA }GxO.
Proof (Sketch): The idea is that if we are at x. and we
want to pass backwards along the vector field f ( x ) ,we simply
let u be zero and allow the free periodic motion to bring x0
nearly back to x . along the integral curve of i = f ( x ) . If the
least period of the periodic solution through x0 is T, then,
by following = f ( x ) for T - E units of time, we have the
same effect as following i = - f ( x ) for E units of time. Thus
we can, given enough time, reach any point whichis reachable where a = ( I 2 - Z3 )/Il, 0 = ( 1 3 - ZIYZ2, and 7 = (11- I2 1/13
for with Zi being the moment of inertia about the ith principal
axis. Let these equations of motion be denoted by
x ( t ) = u ( t ) f [ x ( t ) l+ u ( t ) g [ x ( t ) l .
We now consider some applications to illustrate the ideas
and to contrast the theory with the well-known linear con- where the wi are the angular velocities and Zini are the a p
trollability theory. Our first example can be viewed as a Lie plied torques-both referred to principalaxes. Now s u p
theoreticexplanation of parametric amplificationand para- pose there is only one torque available, say, from a pair of
metric instability. gas jets, but that the torque vector is not necessarily aligned
uthorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions appl
 BROCKETT: NONLINEAR
DIFFERENTIAL
SYSTEMS AND GEOMETRY 65

with a principal axis. That is, we have f and g but not necessarily {f, g}LA. A particularly simple
special case of this kind of result is expressed by the following.
n = u(t)g Theorem 5: Let L = {f,g}LA and let LO be the smallest
for some constant vector g. Again, the motion with u = 0 is subalgebra of L which contains g and is closed under Lie
periodic (elliptic function theory!) thus mitigating the effect bracketing with f. Suppose thatfor all h in LO wehave
of the drift term. [ h ,g] = ahg for some constant a. Then
If we linearize this system about w1 = wz = w3 = 0, then R ( t )= { exp L O}G exp tfxo .
linear theory will predicta one dimensionalreachable set,
provided g is nonzero. We will see that the Lie theory, on the Proof: See [9, p. 945, Corollary 11 and make the neces-
other hand, answers the problem precisely. We begin by com- sary modifications to cover the present setup.
puting some of the vector fields belonging to the Lie algebra Example [ 10, p. 2771 : Consider the linear time invariant
of vector fields generated by f and g. Let a , b , and c denote system
the components of g. We have, for example, i ( t ) = A x ( t ) + b u (xt()0; ) E R " .

[f,gl =
[
a(bw3 + cw2 1
Kaw3 + c u l l
y(aw2 + bwl I]
[[f,gI,gl= 2

and, selecting a third constant vector field from {f,g } u ,

[I:] &c
A quick calculation shows that the vector fields of Theorem 5
are

and that
L = {b,Ab,-,A"-'b,Ax}
Lo = { b , A b ; . . , A " - ' b }

[ [ f , g l , [ [ f , g l , g l=
l 2 [Lo,bI= (0).
Thus thereachable set at time t is

To check controllability, we want to determine if we have, R(t)=exp { b , ~ b ; . . , ~ " - ' b } e ~ ' x ~ .

as yet, vector fields whichspan R 3 . In particular, we ask if Since ( b , A b , * ,A " - ' b ) represent constant vector fields,
g, [f,g1 and [ [f,gl ,gl span. A calculation shows that their exponentials correspond to translation. Thus this expres-
sion for R ( t ) is equivalent to
~ ( t ) {=x : C t x o + q , rlEspan(b,Ab,...,A"-'b)}
which is a well-known result.

=/?ya4(0cZ - y b 2 ) + ~ ~ y b ~ (- 7
a ca2~)+4?c4(ab2 - pa2). 111. I N P U T ~ U T P U TDESCRIFTIONS
One of the central ideas of system theory is the relationship
We now analyze the case of the symmetric rigid body. In
between empirical "external" descriptions of systems and the
this case, we can assume 7 = 0 and since physics demands that
various detailed "internal" models one can propose to account
a + p + y = 0, we see that the determinant is simply ka3c4 * for the observed phenomena.Inthe caseof linearsystems
(b2 + a Z ) . If this vanishes, then one of three things happens-
this circle of ideas is expressed in terms of the relationship
either a is zero, c is zero, or a and b are both zero. The first between theintegral equation description
possibility implies that p is zero, thus f " 0 and the reachable
set is one dimensional. The second implies, since y = 0, that
w 3 is a constant regardless of u and the reachable set is not y ( t ) = L ' w(t,u ) u ( u ) d u + t ( t ) (6)
three dimensional. The third implies that w: + wz is acon-
stant regardless of u , thus the reachable set is not three dimen-
and the linear differential equationdescription
sional. All cases are incompatible with controllability. If
none of these occur,thenthe system is controllable and i ( t ) = A ( t ) x ( t )+ B ( t ) u ( t ) , y(t) = c(t)x(t).
ius this is a complete analysis of the symmetriccase.
If y is nonzero, then theanalysis is apparently more tedious, In the present context, the internal descriptions will be by
and it is not clear that the particular Lie brackets displayed means of nonlineardifferential equations. What substitute
here tell the whole story. Of course, if they do not, then we can be found for (6) as an external description? We can re-
must look at additional elements of {f,g}LA. gard (6) as the fust two terms in a power series expansion for
y . The general series is the Volterra series:
We concludethis section with a discussion of one further
type of result on controllability which has been investigated
by Hirschorn [ 9 ] . The idea is, that under some circumstances,
it should be possible to "factor out" the effect of the drift v ( t )= w o ( t ) +
OD 1'lul ...~''wj(t,ul,uz,...,uj)
i=l
term for
.u(u')u(u~)...u(u~)du~du~
"'dUj.
i ( t )= f [ x ( t ) l + u ( t ) g [ x ( t ) l , x ( t ) = x0 (5)
This series, as it turns out, is just the right alternative for the
so as to be able to
express the reachable set at time t as integral equation (6) whendealing with linear analytic systems.
This is due inpart to thefollowing theorem from [ 1 11.
R ( t )= {exp L}G (exp t f h
Theorem 6: Suppose that f(*,*):R'X R" +R" and g(*,.):
where L is some Lie algebra of vector fields constituted from R' X R" +R" are continuous with respect to their first argu-
horized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions ap
 66 PROCEEDINGS OF THE
IEEE, JANUARY 1916

mentandanalyticwith respect to theirsecond. Given any are all of the form ip= A p x p + uBpxp + ubp, and the Vol-
interval [0, T] such that the solutionof terra series for these bilinear systems can be computed using
the Peano-Baker series.
i ( t )=f[t,x ( t ) ] , x0 = 0
As an example of Volterra series computation, we compute
exists on [0, T I , there exists an E > 0 and a Volterra series the frequency modulation Volterra series. The model is
for
X(f) + (a2 + u ( t ) ) x ( t ) = 0 , x ( 0 ) = 0, i ( 0 ) = 1,
i ( t ) = f [ t , x ( t ) l+ u ( t ) g [ t , x ( t ) l , x(O)=xo,
Y 0 ) = x (t).
Y ( t ) = h[ x ( t ) l (7) Writing this as a first-order system gives
with the Volterra series converginguniformly on [O,TI to
the solution of ( 7 ) provided lu(t)l < E . Moreover, the Vol-
terra series is unique.
We give someindicationofhow the proofgoes. For bi-
linear systems
i ( t ) = A ( t ) x ( f )+ u ( t ) B ( t ) x ( t ) , ~ ( 0= )
xO, Y ( t )= CX(~)
and
we can introduce z ( t ) via x ( t ) = @ ~ ( O
t ,) z ( t ) . Then z ( t )
satisfies
def
; ( f )= U ( t ) @ i ’ ( f , O ) B ( t ) @ A ( t O, ) Z ( f ) =
Y ( t ) = a-’ sin at
l’ a+ sin a(t - u ) sin aou(o) do

Jo” [’
U(f)G(f)Z(f).

Using the Peano-Baker formula, we have + a-3 sin a(t - u) sin n(a - p )

z ( t ) = [I+[‘g(ol)dol +IT‘’
0 0
g(ol)B”(oz)u(ol)u(02) sin a p u ( p ) u ( o ) dpdo + *

* dozdol + .
* * ] xo.
with the nth
kernel being
w , = a-” sin a(t - U l ) sin a(o1 - uz)
Then c x ( t ) is characterized by theVolterra kernels
-- sin a(onml- a,) sin ao,.
w k ( t , o1,o2 * * ’ (Tk)= &A ( t , 0 1 )B(ol ) @ ( D l , uZ)B(oZ) ’*
As stated, the existence theorem for Volterraseries is only a
@A (ok-1, ok)xO. locally valid result. Global representation of a nonlinear sys-
(seed’Alessandro et aZ. [ 121).Fortime invariantbilinear tem by means of an everywhere convergent Volterra series is
systems, we have not expected except for special cases. These special cases art
the analogsof the entirefunctions ofcomplexfunction
w,(tl o1 * 0,) = c&fe-AulB&u~ e-A%B& “3 - -
* theory-Le., functions of a complex variable which have the
whole complex plane as a domain of holomorphy. d’Alessan-
e-A ‘nB&‘nxo.
dro et aZ. [12] have pointed out that, forbilinearsystems,
An algorithmforconstructing the Volterra series inthe the Volterra series doesconverge for all locally bounded u.
general case is given in [ l l I . Itamounts to expanding all This is, perhaps, the firstlarge class of nonlinearsystems
functions in their Taylor series, forming a sequence bilinear for which global convergence of the Volterra series has been
approximations ofincreasingaccuracy,andcomputing the established.
Volterra series forthe bilinearapproximations.Theidea Anobviousnecessaryconditionforglobalconvergence of
behind the derivationcanbededuced fromthe following the Volterra series isthatthere should benofinite escape
example. times for the given initial condition of the system, regardless
Consider of the choice of u . Equivalently, one should ask that there be
nofiniteescapetimefromanypointreachablefrom XQ
i ( t ) = - x ( t ) + x Z ( t ) + u ( t ) [ I+ x Z ( t ) l , y(t)=x(t). As waspointed out before,avector field ona manifold
said t o be complete if the solution through any initial point
Now if we look at x 2 ( t ) , it satisfies a certain differential equa- canbe extendedforwardsandbackwardsintime to give a
tion which is linear in u ( t ) . More generally, x p satisfies a dif- solution on (-m, CO). Thus onemay reasonably restrict the dis-
ferential equation which is also linear in u. We have cussion to complete vector fields. Even s o , this is a problem
0
of some subtlety in that it can happen that f and g are corn
- 12 + u
plete, and yet for

=[:-: 4+2u \
-3
i ( t )= f [ x ( t ) l + u ( t ) g [ x ( t ) l ,

there is a bounded function of time u ( * )such that there exists

finite escape time because [f,g ] is not complete. Thus in at-
x ( 0 ) = x0

\ \ \ tempting t o “globalize” Theorem 1, it is natural to ask that

not only f and g be eomplete but that every vector field in
if J lu(o)l do is small. Then we can expand this about {f,g}LA becomplete.(Inthisconnection it is appropriate
-
Jlu(o) do = 0 and get an expansion of [x, * * , x p ] which is t o mention that if a finite number of complete vector fields
of order xP+’. Thus we can expand x t o any order by this generateafiiitedimensional Lie algebra, then everyvector
uthorized licensed“Carleman linearization”
use limited to: ] . The
[ 1 3Library
J.R.D. Tata Memorial approximate
Indian equations
Institute of Science
field in thealgebra is complete.)
Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions apply
 BROCKETT: NONLINEAR
DIFFERENTIAL
SYSTEMS AND GEOMETRY 67

Suppose that {f,g}LA is a collection of complete analytic series of algebras J? d e f i e d by = L ], =[e, J?] is [ek,
vector fields on a manifold M and suppose that every point on the trivial algebra for some k . Checking for solvability involves
M can be reached from every other point on M (this assump- computing derivatives andlinear algebra. Solutions of dif-
tion will be weakened later). If ferential equationsare not required.
We say that aVolterra series is finitely generated if there
exists a finite number of separable kernels w 1 , w2,* . , w y and
an analytic function q5 such that y ( t ) = @ ( y l ( t ) , y 2 ( t ) ,* * , -
Y r ( t ) )with
u(ok) dol do2 . ' ' dok
* u((71) u ( ( J 2 )*

is the Volterra series for thelinear-analytic system

* *

y i ( t )= I
,'* * luni w i ( t ,0 1 , . . , ani)
i ( t ) = f [ ~ ( t )+l u ( t > g [ x ( t ) l , y ( t ) = h [ x ( t ) l , ~ ( 0 =) x 0
* -
u(o1) . * u(o,,) do,, * * dol.
and if for some u ( * ) defined on [ 0, TI the Volterra series
diverges, then we can still determine y ( t ) from a knowledge Naturally any Volterra series which is itself finite is finitely
-4 u ( * )and the kernels wk as follows. generated, but the converse is clearly false as exp u ( o ) do Ji
From Theorem 6, given anybounded u ( * ) , there exists e, demonstrates.
depending onlyontheboundon u , suchthatthe Volterra Theorem 8: Given (7),if the Lie algebra is solvable, then the
series converges on [0,e l . Moreover, given any u ( ' ) on [0, TI Volterra series is finitelygenerated. Conversely, given any
there exists a trajectory on M generated by u. At each point finitely generated Volterra series, there exists a realization of
on this trajectory, there is a Volterra series corresponding to it with {f,g}LA solvable.
the given differential equationandoutputmap,butwith Proof (Sketch): The second statement comes from [ 11,
altered initial state. At each point along the trajectory, there Theorem 61 with minor changes. The fmt part is an immediate
is an E such that the Volterra series based on that starting state consequence of an analysis of Chen [ 14, Section 41 ; we sketch
converges on [0,e ] for u ( t ) < M . It is not difficult to show how his argument goes. Let { f , ( x ) } i = l be a basis for the Lie
(using the proof of [ 1 1, Theorem 1] ) that the infimum of all algebra generatedby f and g. (This Lie algebra is finite
these E'S is positive, say eo. Now we sum the Volterra series dimensional since it is generatedbya f i i t e number of
by splitting up the interval [0,TI into subintervals of length elements and is solvable.) We look for a solution of
eo or less. Because of causality,i.e., because y ( t ) does not
i ( t )=f [ x ( t ) l + g [ x ( t ) l, x(0) = x 0
depend on u ( 7 ) for 7 > t , we can rearrange the Volterra series
at time neo as which is of the form
x ( t ) = (exp ml ( t )f~) (exp m2 0 ) f 2 ) * * (exp m A t ) f r ) x o .
Using the BakerCampbell-Hausdorff formula,i.e., the vector
field analog of the matrix identity
* w ; ( t , ol, ,ok) u(ol + neo) u(o2 + neo)
* * *
e F ~ e - F = ~ ++ $[ [~F , B~I I~ + * * a

* k neoj dol do2 -

* ~ ( a+ dok
to expand exp ( m i ( t )f ) exp ( m i ( t ) g )exp ( - m , ( t ) f ) , we see
where the altered Volterra kernels are obtained by integrating that for small time, solutions of this form exist. Moreover, we
out the effect of u on the interval [0, neo 1. At the same time, can find the mi as integrals of u( - ) and exponential functions
they are the kernels described by Theorem 6 with starting state of integrals of u . (This part uses the solvability assumption!
x(ne0). Thus the Vnlterra kemsls at one point determine the One can even order the basis so that this representation is valid
Volterra kemeh at all points which are reachable from that for all time if the vector fields are complete.)
point. It is not too difficult to see that if x is either reachable
from x. or a point from which x 0 can be reached, then the IV. THE STATE-SPACE ISOMORPHISM THEOREM
k l t e r r a series at x 0 determinesthe Volterra series at x . There are two very important facts about modeling finite-
gxtendingthisargument, we see thattheVolterra series dimensional linear 1 / 0 systems. The first asserts that any two
at x. determines the Volterra series at each point in models of the same 1/0 map are related by a change of basis,
{exp { f , g)LA }GXO * provided we assume thatthe systemsare controllable and
We say that aVolterraseries is e-summable if, for every observable-thus making sure that the system has no internal
bounded u( .), there exists an e > 0 such that by splitting the parts which are irrelevant totheexternaldescription.The
time axis up into intervals of length E or less and integrating second asserts that any external description which can be real-
outthe kernels successively on [ 0, E ] , [e, 2 ~ 1 * , we can ized bya fiiite dimensionallinearsystem can be realized
make the series convergent. by one which is controllable and observable. In other
Theorem 7: Given a linear analytic system (7)with {f,g}LA settings, one of these results can be true
without the
a collection of complete vector fields, the Volterra series for other. For example, in the group theoretical setup of [ 1 5 1 ,
starting state x. determines the Volterra series at any state in the appropriate analog of the first is true but not the second.
{exp {f,g}LA }cxo, and for any bounded u , the Volterra In the Hilbert space setup of [ 161, the second is true but not
series is esummable toy ( *). the first. In the present context, Sussmann [ 171 hasshown
An important aspect of this work is that the properties of that if arealization exists, aminimalrealizationsexistsand
the Lie algebra { f , g}LA mirror themselves in the propertiesof thatanytwo minimalrealizations of a given 1 / 0 map are
the Volterra series. We give an example of this which hinges related by a smooth change of coordinates. In this section,
uthorized the following
CL licensed definition.
use limited to: A Lie Library
J.R.D. Tata Memorial algebra is solvable
Indian Institute of if the Bengaluru.
Science we consider a special
Downloaded case of
on September theorems
his at
18,2024 and from
07:51:00 UTC someIEEE related
Xplore. Restrictions appl
 68 PROCEEDINGS
JANUARY OF THE IEEE, 1976

results. For a well motivated discussion of these elegant

results, see Sussman's contribution to [ 1] .
Suppose that
X(r) =a[x(t)l + u ( t ) b [ x ( t ) l , A t ) = c [ x ( t ) l , x ( 0 ) =xo
and
ip
are two linear-analytic realizations of the same 1/0 map. We
would like to show that, under reasonable assumptions, there
exist an analytic map @ such that x = @ ( z )and z = q5-l ( x ) . If
such a @ exists, then it follows that X = ( a @ ) / ( a zi) and thus
that

Fig. 4. Illustrating the proof o f isomorphism theorem.

We say that a linearanalytic realization is bilaterallycon- it is one to one by observability. (All points are reachable by
trollable on a manifold M if, given any two pointsX I , x2 on M , 6
going forward and backward in time, so is defined on all of
there exists r E { exp {f,g}LA }G such that m1 = x ? . This is M and is onto N , no two starting states give identical 1/0maps.)
the same as asking that there be a sequence of points x' ,x z , Clearly then, $is a one to one and onto map. We have only to
* -- ,x" with x 1 = x 1and x 2 = x" and x' reachable from x 1 establish its smoothness properties. We want to show that is, 6
and x 3 , x 4 reachable from x 3 and x 5 , . That is, bilateral -- in fact, analytic. Notice, however, that it must happen that for
controllability means one can get from any point onM to any each { t l , t 2 ; . - , t n } , w e h a v e
other point on M moving forward and backward in time. We
say that alinear analytic realization on amanifold M'is ij[(exp tlf) (exp t2g) (exp t 3 f ) * *) x01
observable if, given any two distinct starting statesx l , x2 in M , =(exptla)(expr2b)(expta)(~")x~l.
there exists an input u ( * )whose choice may depend on x1 and
x2 , such that the output fromx 1 differs from the output from If y e pick coordinates in M and coordinates in N , we can lo01
x 2 . Realizationswhich have these properties will be called at @ as a map ,of R" + R". The setup is diagramed in Fig. 4 .
minimal. Notice that the idea of dimension plays no role in The analyticity follows from the analyticity of the exp map
this definitionalthoughone would hopethat all minimal and its inverse.
realizations of the same 1/0 map would be realized on Theother(moredifficult)part of this circle of ideas is
manifolds of the same dimension, and this turns out tobe the related to theexistence of minimal realizationsand is expressed
case. by the following theorem.
Theorem 9 (Specialcase of Sussmann /I 71): Suppose we Theorem IO: Given any 1 / 0 map which is realizable by an
are given two linear analytic systems initialized linear-analytic
system on manifold
a M with
{ f , g } L A being complete,there exists arealization of the
i ( t )= f [x(t)l + u ( t ) g [ x ( t ) l , Y ( t )= hIx(t)l, x0 =x(0) same 1/0 map on amanifold N suchthatthe system is
and bilaterally controllable on N , and any two distinct states can
be distinguished from the output forsome input.
i ( t )'= a [ z ( t ) l + u ( t ) b [ z ( t ) l , y ( t )= c [ z ( r ) ~ , z o = Z(O) The proof of thistheorem requiredasharpening of the
conditions of Theorem 2. Inorder to show that{exo
which are minimal realizations on manifolds M and N , respec- {f,g}LA }G is always a differentiable manifold [ 181 , [ 191 am
tively, of the same 1/0map. Suppose that all the vector fields also a generalization of the closed subgroup theorem of dif-
in { f , g } L A and { a , b)LA arecomplete.Thenthere existsa ferential geometry [ 201. Both thesegeneralizations require
diffeomorphism @ :M + N such that new mathematical technique and yield new basic mathematics
in return.
It is of interest to contrast Theorems 9 and 10 withthe
corresponding results for bilinear systems. Bilinear systems in
R"
b [@-I ( x ) ] = g ( x )
ax i ( t ) = Ax(t)+ u(t) Bx(t), y(r) = (c, x ( t ) ) , x ( 0 ) = x o
c[@-'(x)l = h(x).
have a reachable set which is a proper subset of R". One says
Proof (Sketch): Since the realizations are minimal, there that they are minimal [ 121 , [ 21 ] if there is no subspace of
exists a map $:M+ N which sends x0 into zo and which sends R" which contains the reachableset for all time and if any
any point x in M wbich,is reachable from x0 to the point z in twodistinctstartingstates, x 1 and x2, yield different 1/0
N reached by the application of the same controlsequence t o maps. The analog of Theorem 9 is that any bilinear system
the z system. Controllability insures that this map is onto N ; can bereduced to one which is minimal, and the analog of
thorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions app
 BROCKETT: NONLINEAR SYSTEMS AND DIFFERENTIAL GEOMETRY 69

Theorem 10 is thatanytwo minimal bilinear systemsare over the set of a l l u’s which steer the deterministicsystem
isomorphic in the following very strong, sense. If, A , B , c, and
x. and F , G ,h , and zo are the data which specify the systems,
i ( t ) = A ( t )x ( t ) + b(t) u(t)
then there exists an invertible matrix P such that PAP-’ = F , from x = 0 at t = 0 t o x = x 1 at t = tl is
PBP-’ = G , cP’c = h , and Pxo = z o . Thus in the case of bilinear
systems, the structure oneassumes to begin with is ratherrigid, ~ * ( ~ ~ , t ~ ) = x ~ ~ - ~ ( t ~ ) x ~ .
but the conclusion is also very precise (linearisomorphism). Comparing this with the formula for the density, we see that
With linear-analytic systems the structure is less rigid and the if we start at x = 0 at t = 0, then the value of the probability
conclusion is less precise. densityatanypoint is inversely proportional to the ex-
We note in passing that the results we have given, together ponential of the “energy” required to getthere. In this
with a result of [ 121, allow us to give an answer to a question section,we.describetheextent to whichtheseresultscarry
which has received some attention in the literature. That is, over to linear analytic systems.
given a linear analytic systems, when does there exist a change Because of theproperties of theIto calculus, stochastic
of coordinates which puts it in bilinear form? The answer is equations on manifolds take a somewhat peculiar form. The
the following. Compute the Volterra series for the stationary basic well-posedness conditionforthe class of systems of
linear analytic system; if there exists a square matrix T ( t ) with interest here is expressed by the following theorem.
a rational Laplace transform, and vectors c and x o , such that Theorem 11: Let f , g : R ” + R” be analytic. Suppose that
the kthkernel is M = { x : x E R ” , & ( x ) = 0, i = 1, 2, * . , p ) is an analytic
variety in R ” . Given the Ito equation
w k ( t , 0 1 ,0 2 , * * * 9 0k)
-
= c‘T(t - ul) T ( U ~- u2). * T ( U ~ -- ~uk)x. dx = f ( x ) d t + g ( x ) d w , x(0)E M (8)
then the Volterra series is bilinearly realizable (cf. [ 121). If we can assert that x ( t ) belongs to M with probability one, if
no such T,c, x0 exists, then the system is not bilinearizable. for all x
The work of Fleiss [ 221 is devoted to the derivation of some C(x)&(x)=O, i=1,2;*.,p
of the results on bilinearsystems mentioned here using the
Kleene-Schutzenberger representation theorem and the theory
( ~ ( x -) 3 c Z ( x ) )~ ~ (=x0, ) i = 1;2, - * ,p
of formal power series. where F and G are given by
The paper of Porter [23] contains further material develop-
ing Volterra series ideas in adifferent way.

V. STOCHASTIC
EQUATIONS
One of themost convincingapplications of thecontroll-
ability concept for linear systems occurs in the study of linear
stochastic equationsof the Ito type Proof: (see [25] and [26] .)
As application of thisresult we look at an example from
d x ( t ) = A ( t ) x ( t ) dt + b(t) dw(t). [ 281. Consider the Ito equation
The density which results from a particular initial density can
d x = A x d t + B x d w , x ( 0 ) = xI oI x, o l l = 1.
be expressed in terms of the density which results from x(0) =
0. Thisis Under what circumstances can we ,assert that Ilx(t)ll = 1 with
probability one? In this case 4 = x x - 1 and G @ ( x )= 0 if and
only if B is skew symmetric. Likewise, ( F - (1/2) G 2 ) $ ( x )=
0 if A - (1/2) B2 is skew symmetric. Thus for example,
where W is a controllability Grammian,

w(r)=~f~~(f.(I)b(D)bf(0)~~(t,O)d.
defines aprocess which stays on the circle {x: + x f = constant}
almost surely in themean-square sense. The apparent tendency
and W is invertible, as required for the expression for p to be of the drift term to pull the process inside the circle is offset
meaningful, when the system has the controllability property. by the Ito correction term.
Thus controllability is exactly the right concept for deciding Given a stochastic differential equation, the question arises
if smooth densities exist. Moreover, if A and b are constant as to whether or not smooth transition densities exist as they
and if the eigenvalues of A have negative real parts, then, under do in the linear controllable case. If they exist, then we denote
thecontrollabilityhypothesis,thereis a unique invariant by p(t, x , y ) the probabilitydensity at time t and point x given
measure which is C” (Gaussian actually) such that a l l measures that at t = 0 the state was y .
approach it as t goes to infinity. The Fokker-Planck equationforthetransition densities
Infact,there is one more connectionbetweencontrol- associated with the Ito equation(8) is easily expressed in terms
lability and the probability density which is even more striking. of the above notation. It is
It is well known and easily seen [ 241 that the minimum value
of ($ 3
+ F ( x ) - G z ( x ) p ( t , x , y ) = 0. 1
This equation has the formal appearance of a heat equation
exceptthatthe second-order part, represented bythe G 2
uthorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions apply
 70 PROCEEDINGS OF THE
IEEE, JANUARY 1976

term, is not positive definite, so the classical existence theory property that for any twodistinct points x1 and x2 in X,there
does not apply. Fortunately, Hormander [27] has worked out are elements of r , N ( x l ) and N ( x 2 )such that N(x ) fl N(x2 ) is
a theory of such equations which exactly fits the present case. empty with x 1 E N ( x l ) and x2 E N ( x 2 ) . We ask that (X, r ) be
Hormander calls an operator separable, that is, thatthere exista countable subset of X
which is dense in (X,7). As a consequence of this, there is a
L=(F+G~) countable subset of r whichgenerates all of r under union;
hypoelliptic if, forany given Schwartz distributions 4 and that is (X,r ) is second countable. So much for the topology.
J, suchthat L $ ( x ) = $ ( x ) , it follows that is C" off the
$J
Now we consider the collection of maps @ = {@ilt$i: Mi + N i } .
support of the singular part of J, after a suitable modification We assume that there exists an integer n , a collection of open
of @ on a zero measure. A necessary and sufficient condition sets Ni, and continuous one to one maps $i :Ni + M i C R"
for hypoellipticity is that { f , g}LA should span .the tangent where Mi are open subsets of R" such that @T1 :Mi + Ni are
space of the manifold at each point. The a/& term in the also continuous. Thiscollection CP is assumed to have the
Fokker-Planck equation must be combined with L to give an following additional properties.
operator which is hypoelliptic on M X [ 0, 00) where [ 0,w) is (i) The domains of the @i form an open covering of X.
the time axis. Since a/at spans the time direction, we have (ii)Each 4i maps its domain bicontinuously onto an open
hypoellipticity if the smallest Lie algebra, which contains g subset of R",
and is closed under bracketing with f , spans the tangent space (iii)If Ni nNi is not empty and if 4i : Ni n Ni + Pii and
of M at each point of the manifold. These ideas were put 4j :Ni nNi + Qii, then @j 0 4zT1 is an infinitely differ-
together by Elliott [ 261, [ 271 to get the following results. entiable map of Pii into Qij.
Theorem 11: Suppose that { f , g}LA consists of complete (iv) This collection @ is maximal with respect to properties
vector fields on a manifold M . Suppose that the smallest Lie (ii)and (iii).
algebra, which contains g and which is closed under bracketing
with f , spans the tangent space of M at each point. Then the The locally defined maps & are called charts and @ is called,
corresponding Ito equation ratherappropriately,an atlas. The last condition (iv) is not
too important in that if we have some charts that we want to
use, then we just declare Q, to be the collection of all charts
which are compatible with the given ones. If in the previous
paragraph we replace "infinitely differentiable" by "analytic",
d e f i e s C" transition densities on M . then we obtain an analytic manifold. (Incidentally, because we
Notice that,in statingthis theorem, we have takenthe assume that the manifolds are second countable, they are auto-
opposite point of view from Theorem 10 in that here f and g matically paracompact, a condition which is sometimes assumed
definevectorfields onthe manifold andthe "correction in place of second countable.)
term" (b/2) ( a g / a x )g is what is needed to insure that the give0 In addition to R" itself, a standard source of manifolds are
Ito equation evolves on the manifold. In Theorem 9,g and f - the so-called differential(analytic)varieties, in R" described
(1 /2) ( a g / a x )g were the vector fields on the manifold. by the following theorem.
Specific instances of processes meeting these conditions have Theorem A: Let @1@2,* * * ,, ,@, be C" (analytic) functions
been discussed in the literature. For example, in [ 281 the case on R" suppose that
of processes on spheres is studied in some detail.
Theorem 11 then asserts that some essential features of the Rank[(V41,V4z,"',V4nI=m
linear stochastic differential equation can betransferred to at each point on the setM = { x : & ( x ) = 0). Then this Set with
manifolds. Only the last fact citedat the beginning of the the topology r consisting of all sets of the form 0 nM 0 open
section regarding theproportionality of the density to the in R n can be given thestructure of a c" (analytic)differ-
exponential of the control energy has failed to be generalized.
entiable manifold.
In special cases (essentially Abelian Lie groups), some progress
Proof: Cee Singer and Thorpe [ 3 1, p. 1201 .)
has been made on finding a suitable extension of this result,
A submanifold of a C"-manifold M is a pair ( N , 4) where N
but a general theory seems to require new developments.
is a C"-manifold and I#J is a one to onemapping q5 : N + M sucb
There is agrowing literature on the formulation and solution
that ( d $ ) / ( d x )has a trivial kernel at each point of N . In some
of filtering and stochastic control problems on manifolds (see, cases, one defines an n-dimensionaldifferentiablemanifold
for example, [ 301 ).
M as a connected subset of the standard pdimensional Car-
APPENDIX tesian space subjecttotherequirementthat,at each point
x of R" which belongs to M , it is possible to establish a (gen-
The purpose of this appendix is to give the basic definition erally curvilinear) coordinate system for RPwith basis elements
from the subject of differentiable manifolds in a form which y , , y 2 ,* * ,y p (RP-valued C"-functions on R p ) such that in a
is most compatable with the main body of this paper. We do neighborhood N of x , the intersection of M and N is exactly
so both to provide a convenient reference for the reader and the set of points which satisfy
also because there are many slight but potentially annoying
differences in the general literature. Yn+1 = o
By a C" differentiablemanifold, we understand a triple
(X,r , a) whereby X is a set, r is topology on X,and is a set Yn+2 = 0
of continuous maps from open subsets of X into R" with its
usual topology. Both r and CP are subject to certain restrictions
which we now describe. We ask thatthe topologicalspace
(X,r ) be a Hausdorffspace, i.e., theopensets have the y p = 0.
uthorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions appl
 BROCKETT: NONLINEAR SYSTEMS AND DIFFERENTIAL GEOMETRY 71

functions JI which are constant on M , then in any local co-

ordinate system of the type described in Theorem A we have

fi-
a =0, i = 1,2;..,p.
axi
Thus we can restrict F to T(M)and get a vector field on M.
The symbols a / ( a x i ) can be thought of as unit vectors to be
Fig. 5 . illustrating submanifold of R 3. comparedwiththe symbols i, j , k, etc. in classical vector
analysis. This notation associates witheach vector field a
partial differential operator whose domain is the differentiable
functionson M. It is particularlyuseful, for example, in
discussing the Fokker-Planck equation.
By the Whitney imbedding theorem,any manifold is dif-
feomorphic to a submanifold of R ” . In R” , there are admissi-
ble coordinate systemswhich are global-e.g., thestandard
coordinate system for R n , It also serves to identify points in
Fig. 6. Tangent space of M at m . any submanifold M C R ” and thus we may describe vector
fields on M in terms of this single global coordinate system.
When we speak of a differentiable dynamical system being
The remaining coordinate functions y 1 , y 2 , . * ,Y n then serve on a manifold, what we mean is that at each point x0 of M ,
to define points in M unambiguously. Thus adifferentiable we can pick local coordinates and express the evolution in a
manifold is a pointsetand a set of coordinatecharts. This patch about x . by
definition is fully consistent with the more abstract definition
given above and, in fact, defines the differentiable structure i ( t ) = a [ x ( t ) ]+ u ( t ) b [ x ( t ) ] .
alluded to in Theorem A (see Fig. 5). However, it may happen that M is a subset of R m , and we look
Example: We may realize the n-dimensional sphere shell, as only at the redundant coordinates
asubmanifold of Rn+’ asfollows.Let 4 1 of Theorem A be
x ~ + x ~ * * * x ~ + T l -h e1 n. a t e a c h p o i n t o f M = { x : 4 ( x ) = =f [ z ( t ) l + u(t) g[z(t)l
0) the rank of V@ is one. Thus M can be given amanifold in R” space. It is in this second way that one is most likely to
structure. Since V@ has a trivial kernel, thismakes (M, 4) a encounter a practical problem.
submanifold of R “ + l .
According to the Whitney imbeddingtheorem[32] , any REFERENCES
@anifold is diffeomorphic to one which is imbeddedin a D. Q. Mayne and R. W. Brockett, Eds., Geometric Methods in
Euclidean space. System Theory, e r d r e c h t , Holland: Reidel, 1973.
G. Frobenius,“Uber die unzerlegbaren diskreten, Bewegungs-
We now set up the definition of a vector field on a manifold. gruppen,” Sitzungskrichte d . Konig. Peruss. Akd. W h . vol. 29,
Let M be amanifold and let Fo(M) be theset of all C” pp. 654-665, 1911.
(analytic) functions on M ,F’(M) = { f l F :M -P R , f~ 6” or W. L. Chow, “Uber Systeme Von Linearen Partiellen Differential-
gleichungen ester Ordinung,” Math. Ann., vol. 117, pp. 98-105,
analytic}. A tangent vector at m E M is an operator 1939.
u : F o ( M )+ R such that forall f,g , E Fo(M) R. H e m a n n , “On the accessibdity problem in control theory,”
in NonlinearDifferentialEquationsandNonlinearMechanics,
J. P. LaSalle and S. Lefschetz, E&. New York: Academic Press,
(i) o ( f + g ) = ~ ( f+)vk) 1963.
(ii) 4 f g ) = ~ ( f* d) m ) + v(g) f h ) . H. Hermes and G. W.Haynes, “On the nonlinear control problem
with control appearing linearly,” J. SIAM Contr., vol. 1, pp. 85-
This definition does not look verygeometrical, but observe 108,1963.
first of all that the space of all tangent vectors at m is a real C. Lobry, “Quelquesaspectsqualitatifs de la theories de la
commande,” L’Universite Scientifique et Medicale de Grenoble,
vectorspace, andsecond, if x 1 , x ? , * ,x, definesalocal pour obtenir le titre de Docteur es Sciences Mathematiques, May
coordinate system in the manifold, thenfor each i, u ( f ) = 19, 1972.
a f ) / ( a x i ) J mdefines a tangent vector. In fact, one can see on A. Krener, “A generalization of Chow’s theorem and the bang-
bang theorem t o nonlinear control problems,” SIAM J. Contr.,
the basis of a Taylor series argument that these vectors form a vol. 12, no. 1, pp. 43-52,1974.
basis for the set of all tangent vectors at m , i.e., the tangent H. Sussmann and V. Jurdjevic, “Controllabilityofnonlinear
space at m . Thisvector space is denoted by Tm(M)(see Fig. 6 ) . systems,” J.DifferentialEquations, vol. 12, pp. 95-116, July
1972.
By a vector field on manifold, we understand a map which R. Hirschorn, “Topological semigroups,setsofgenerators, and
assigns to each point of m an element of Tm(M) controllability,” Duke Math. J.,vol. 40, pp. 937-947, Dec. 1973.
R. W. Brockett, “System theory ongroupmanifolds and coset
spaces,”SZAMJ. Contr.,vol. 10, no. 2, pp. 265-284, 1972.
-, “Volterra series and geometric control theory,” 1975 Pre-
printsZFAC, Philadelphia, Pa.: ISA.
P. d’Alessandro, A. Isidori,and A. Ruberti,“Realization and
structure theory of bilinear systems,” SZAM 1. Contr., vol. 12,
If the ai are C” (analytic), then this is called a C” (analytic) no. 3, pp. 517-535, 1974.
vector field. A. J. Krener, “Linearization and bilinearization of control
In R ” , we may then associate with every function f : R” -+ systems,” in h o c . 1974 AllertonConf. Circuit and System
Theory, 1974.
R” a vector field on R” K. T. Chen,“Decompositionofdifferential equations,” Math.
Ann., vol. 146, pp. 263-278, 1962.
R. W. Brockett and A. S. Willsky, “Finite group homomorphic
sequentialsystems,” ZEEE Tmns. Automat.Contr., vol. 17,
pp. 483-490, Aug. 1972.
YE (M,4) is a submanifold of R n such that FJI vanishes for all J . Baras and R. W. Brockett, “H’-functions and infinite dimen-
uthorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions appl
 12 PROCEEDINGS OF THE IEEE, VOL. 64,NO. 1, JANUARY 1976

sional realization theory,” =AM J. Contr., vol. 13,pp. 221-241, J. Wiley, 1970.
Jan. 1965. [25] J. M. C. Clark, “An introduction to stochastic differential equa-
H. J. Sussmann, “Existence and uniqueness of minimal realiza- tions on manifolds,” in Geometric Methods in System Theory,
tions of nonlinear systems-Part I: Initialized systems,” J. Math. D. Q. Mayne and R. W. Brockett, Eds. Dordrecht, Holland:
Syst. Theory, to be published. Reidel, 1973.
-, “Orbits of families of vector fields and integrability of [26] D. Elliott, “Diffusions onmanifolds arising from controllable
systemswith singularities,” Bull. Amer. Math. SOC., vol. 79, systems,” in Geometric Methods in System Theory, D. Q. Mayne
pp. 197-199, 1973. and R. W. Brockett, Eds. Dordrecht, Holland: Reidel, 1973.
1191 P. Stefan, “Two proofs of Chow’s theorem,” in Geometrk I271 - , “Controllable systems driven by bite noise,” Ph.D. dis-
Methods in System Theory, D. Q. Mayne and R. W. Brockett, sertation, Univ. of Calif., Los Angeles, 1969.
Eds. Dordrecht, Holland: Reidel, 1973. I281 R. W. Brockett, “Lie theoryandcontrol systemsdefined on
H. J. Sussmann, “On quotients of manifolds: A generalization spheres,” SZAMJ. Appl. Math.,vol.25, pp. 213-225,Sept. 1973.
of the closed subgroup theorem,” J . Differential Geo., vol. 10, [ 29 1 L. Hormander,“Hypoellipticsecond-orderdifferential equa-
Mar. 1975. tions,”Acta. Math., vol. 119, pp. 147-171, 1967.
R. W. Brockett, “On the algebraic structure of bilinear systems,” I301 J. T. L o and A. Willsky, “Estimation for rotational processes with
in TheoryandApplications of VariableStructureSystems, R. one degree of freedom-Parts I-II1,”ZEEE Trans. Automat. Contr.,
Mohler and A. Ruberti, Eds. New York: Academic Press, 197:; vol. 20, pp. 10-33,Feb. 1975.
M. Fleiss, “Sur la rialization des systimes dynamiques biliniares, 131 1 I.M. Singer and J . A. Thorpe, LectureNotesonElementary
C. R . Acad. Sc. Paris, vol. A-277,pp. 923-926, 1973. Topology and Geometry. Glenview, Illinois: Scott, Foresman
W. A. Porter, “An overview of polynomic system theory,” this and Co., 1967.
issue. [32] L. Auslander and R. E. MacKenzie,Zntroduction to Differentiable.
R. W. Brockett, Finite Dimensional Linear Systems. New York: Manifolds. New York: McGraw-Hill, 1963.

linear Passive Networks: Functional Theory

Abstmct-Lineaa passivetime-varirblenetworksareinvestigated P i - [6,Sec. 3.51 and sometimes in the time domain [ 71.Many of
the use Of distniutionai kernels as applied to the the results based upondistributiontheory to date are sum-
scattering matrix treated in the time domain. Necessary and sufficient
conditions for are obtained,md the matrix is shown marized in the books of Doleial [8]
and, more recently,
to be ameasure s~tisfyingan energy formconstraint. Lossless con- ikmanian [91.
stnints pertinent to synthesis are deveioped while networks consisting For time-invariant networks, there are also available more or
of a finite number of circuit eknents are considezed in some less classical-type works [lo]-[ 151 based upon frequency do-
Examples illustrating interesting behavior are presented.
mainconcepts.These concepts have been taken over to the
I. INTRODUCTION timedomaincharacterizations, primarily throughthestate
[ 161, such that extensions to time-variable synthesis based
HE MATHEMATICAL FIELD of functional analysis is upon the state [ 171, [ 181 become straightforward, albeit with
now recognized asa rich one witha varied, though strangeresults (as instabilities of passive structures [ 19 I).
modem,history [ 11. Within functional analysis, the Likewise, there are recent applicable developments in operato!.
theory of distributions [2]-[4] plays a particularly interesting theory [ 201, especially with regard to resolution space con-
role, especially for physicalsystems. Indeed,motivationfor cepts [21]-[ 241 as well as some distributional treatments of
the theory came,in part, from the circuit theory aspects of time-variable networks [25]-[28]. However, whenoneturns
Heaviside’s work [ 51 while more recent applications have led toward synthesis of time-variable networks,thefunctional
to the development of the properties of passive networks in a analysis results are scarce [29]-[33].Thusit seems that a
distributional framework, sometimes in the frequency domain functional analysis treatment of linear passive time-variable
networks with an emphasis upon results important for synthe-
Manuscript received January 3, 1975; revisedMarch 21, 1975.This sis is in order.
work was supported in part by the U.S. Air Force Office of Scientific In this paper, some of themostimportantproperties of
Research underGrants AF-AFOSR-337-63and AFOSR 70-1910,in
part by the Australian Government Services Canteens Trust Fund, in linear passive networks aredevelopedin terms of the time-
part by the Australian Research Grants Committee, and in part by the varying scattering matrix [34]~ ( 7t ),; this distributional kernel
U.S. Educational Foundation in Australia under a Fulbright Traveling s appears to be one of the most generally useful descriptions
Grant.
B. D. 0. Anderson is with the Department of Electrical Engineering, available, especially for synthesis. The paper is structured such
University of Newcastle, Newcastle, N.S.W., Australia. that Sections IV and V contain the general results. Section I1
R. W. Newcomb is with the Department of Electrical Engineering, essentially serves as a review section where the underlyingcon-
College of Engineering, University of Maryland, College Park, MD
20742. cepts of interest are d e f i e d ; among these are a network an4
uthorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 18,2024 at 07:51:00 UTC from IEEE Xplore. Restrictions appl