Lecture Notes in

Control and
Information Sciences
Edited by M.Thoma

72

A. Isidori

Nonlinear Control Systems:

An Introduction

Springer-Verlag
Berlin Heidelberg New York Tokyo
Series Editor
M. Thoma

Advisory Board
A.V. Balakrishnan L. D. Davisson A. G. J. MacFarlane
H. Kwakernaak .1. L. Massey Ya Z. Tsypkin A. J. Viterbi

Author
Prof. Alberto Isidori
Dipartimento di Informatica e Sistemistica
Universit& di Roma ,,La Sapienza-
18 Via Eudossiana
00184 Rome (Italy)

ISBN 3-540-15595-3 Springer-Verlag Berlin Heidelberg New York Tokyo

ISBN 0-38?-15595-3 Springer-Verlag NewYork Heidelberg Berlin Tokyo

Library of Congress Cataloging in Publication Data

Isidor[, Alberto:
Nonlinear control systems: an introduction/A. Isidori.
Berlin; Heidelberg; New York; Tokyo: Springer 1985.
(Lecture notes in control and information sciences; Vol. 72)
NE: GT
This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically those of translation, reprinting, re-use of illustrations, broadcasting,
reproduction by photocopying machine or similar means, and storage in data banks, under
54 of the German Copyright Law where cop!es are made for other than private use, a fee is
payabte to "VerwertungsgeseHschaft Wort =, Munich.
Springer-Verlag Berlin, Heidelberg 1985
Pdnted in Germany
Offsetprinting: Mercedes-Druck, Berlin
Binding: L~deritz und Bauer, Berlin
2161/3020-543210
 PREFACE

This volume was p l a n n e d as a textbook for a graduate course on

nonlinear multivariable feedback systems. Most of it was p r e p a r e d
while the author was t e a c h i n g a s i m i l a r course at the D e p a r t m e n t of
Systems S c i e n c e s and M a t h e m a t i c s of the W a s h i n g t o n U n i v e r s i t y in
St. Louis, in the 1983 fall semester. The purpose of -the volume is to
p r e s e n t a s e l f - c o n t a i n e d d e s c r i p t i o n of the fundamentals of the theory
of n o n l i n e a r feedback control systems, with special e m p h a s i s on the
differential-geometric approach.
In the last decade, differential g e o m e t r y has p r o v e n to be as
s u c c e s s f u l to the study of n o n l i n e a r systems as Laplace t r a n s f o r m and
c o m p l e x functions theory were in the '50s to the study of s i n g l e - i n p u t
s i n g l e - o u t p u t linear systems and linear algebra in the '60s to the
study of m u l t i v a r i a b l e linear systems. Typical "synthesis" problems
like d i s t u r b a n c e isolation, noninteraction, shaping of the i n p u t - o u t p u t
response via feedback, can be d e a l t with relative ease, w i t h tools that
are w e l l w i t h i n the reach of a ( m a t h e m a t i c a l l y oriented) control
engineer. The purpose of this volume is to make the reader a q u a i n t e d
with m a j o r methods and results, and to make h i m able to e x p l o r e the
c o n s t a n t l y g r o w i n g literature.
The book is o r g a n i z e d as follows. C h a p t e r I introduces invariant
distributions, a fundamental tool in the analysis of the internal struc-
ture of n o n l i n e a r systems. W i t h the aid of this concept, it is shown
that a n o n l i n e a r s y s t e m locally e x h i b i t s K a l m a n - l i k e decompositions
into " r e a c h a b l e / u n r e a c h a b l e " parts and/or " o b s e r v a b l e / u n o b s e r v a b l e "
parts. C h a p t e r II e x p l a i n s to what extent there may exist global de-
compositions, c o r r e s p o n d i n g to a p a r t i t i o n of the whole state space
into "lower dimensional" r e a c h a b i l i t y and/or i n d i s t i n g u i s h a b i l i t y sub-
sets. C h a p t e r III d e s c r i b e s various "formats" in w h i c h the i n p u t - o u t p u t
map of a n o n l i n e a r system may be represented, and p r o v i d e s a short de-
s c r i p t i o n of the f u n d a m e n t a l s of r e a l i z a t i o n theory. C h a p t e r s IV and V
deal with the s y n t h e s i s of f e e d b a c k control laws. In the first of these,
d i s t u r b a n c e d e c o u p l i n g and n o n i n t e r a c t i n g control are d e a l t w i t h , a l o n g
the s o - c a l l e d "geometric approach", that p r o v e d to be quite successful
for the s o l u t i o n of s i m i l a r s y n t h e s i s p r o b l e m s in l i n e a r m u l t i v a r i a b l e
systems. In C h a p t e r V it is shown that n o n l i n e a r s t a t e - f e e d b a c k may be
used in o r d e r to make a given system to behave, i n t e r n a l l y and/or ex-
 IV

ternally, like a l i n e a r one. In particular, feedback may be used in

orde r to shape the input-output behavior in some p r e s c r i b e d way.
The r e a d e r is s u p p o s e d to be familiar with the b a s i c concepts
of l i n e a r systems theory. Moreover, some k n o w l e d g e of the fundamentals
of d i f f e r e n t i a l geometry is required. There are several excellent text-
books available to this end, and some of them are q u o t e d a m o n g the refer-
ences. However, in o r d e r to make the volume as much as p o s s i b l e self-
contained, and p a r t i c u l a r l y to u n i f y the notations, the m o s t important
notions and r e s u l t s of f r e q u e n t u s a g e are c o l l e c t e d - without proof -
in the A p p e n d i x .
The a u t h o r of this book is p a r t i c u l a r l y grateful to P r o f e s s o r
A. Ruberti, for his c o n s t a n t encouragement, to P r o f e s s o r s J. Zaborszky
and T.J. Tarn for their interest and g en e r o u s support, to P r o f e s s o r
A.J. Krener who, especially in the c o u r s e a joint research venture,
was a source of i n s p i r a t i o n for m a n y of the ideas developed in this
volume. The author w o u l d also like to thank Professor M. T h o m a for his
encouragement during the p r e p a r a t i o n of this w o r k and P r o f e s s o r s C.
Byrnes, M. Fliess, P. K o k o t o v i c and S. M o n a c o for m a n y stimulating
discussions.

Rome, March 1985

 TABLE OF CONTENTS

Chapter I - LOCAL DECOMPOSITIONS OF CONTROL SYSTEMS ........ I

I. Introduction .......................................... I
2. Distributions on a Manifold ........................... 5
3. Frobenius Theorem ..................................... 12
4. Invariant Distributions ............................... 24
5. Local Decompositions of Control Systems ............... 28
6. Local Reachability .................................... 33
7. Local Observability ................................ ... 47

Chapter II - GLOBAL DECOMPOSITIONS OF CONTROL SYSTEMS ...... 55

I. Sussmann Theorem and Global Decompositions ............ 55

2. The Control Lie Algebra ............................... 61
3. The Observation Space ................................. 67
4. Linear Systems, Bilinear Systems and Some Examples .... 72

Chapter III- INPUT-OUTPUT MAPS AND REALIZATION THEORY ..... 83

I. Fliess Functional Expansions .......................... 83

2. Volterra Series Expansions ............................ 92
3. Output Invariance ..................................... 96
4. Left-Invertibility .................................... 102
5. Realization Theory .................................... 106
6. Uniqueness of Minimal Realizations .................... 120

Chapter IV - DISTURBANCE DECOUPLING AND NONINTERACTING

CONTROL ....................................... 124

I. Nonlinear Feedback and Controlled Invariant

Distributions ......................................... 124
2. The Disturbance Decoupling Problem .................... 134
3. Some Useful Algorithms ................................ 141
4. Noninteracting Control ................................ 152
5. Controllability Distributions ......................... 159
6. More on Noninteracting Control ........................ 165

Chapter V - EXACT LINEARIZATION METHODS .................... 178"

I. Linearization of the Input-Output Response ............ 178

2. The Internal Structure of the Linearized System ....... 191
- 3. Some Algebraic Properties ............................. 200
4. Linear Model Matching ................................. 206
"5. More on Linear Model Matching, Output Reproducibility
and Noninteraction .................................... 215
6. State-Space Linearization ............................. 222
-7. Observer with Linear Error Dynamics ................... 244
 VI

Appendix - BACKGROUND MATERIAL FROM DIFFERENTIAL GEOMETRY ... 254

I. Some Facts from Advanced Calculus ...................... 254

2. Some Elementary Notions of Topology .................... 256
3. Smooth Manifolds ....................................... 258
4. Submanifolds ........................................... 264
5. Tangent Vectors ........................................ 268
6. Vector Fields .......................................... 277

BIBLIOGRAPHICAL NOTES ....................................... 289

REFERENCES .................................................. 292

 CHAPTER I
LOCAL DECOMPOSITIONS OF CONTROL SYSTEMS

I. I n t r o d u c t i o n

In t h i s section we review some basic results from the theory of

linear systems, with the purpose of d e s c r i b i n g some fundamental pro-
perties which find close analogues in t h e theory of nonlinear systems.
Usually, a linear control system is d e s c r i b e d by equations of the
form
= Ax + Bu

y = Cx

in w h i c h the state x belongs to X, an n - d i m e n s i o n a l vector space and

the i n p u t u and the o u t p u t y belong respectively to a n m - d i m e n s i o n a l
vector space U and -dimensional vector space Y. T h e m a p p i n g s
A : X ~ X, B : U ~ X, C : X ~ Y a r e linear mappings.
Suppose that there exists a d-dimensional subspace V of X with
the f o l l o w i n g property:

(i) V is invariant under the m a p p i n g A, i.e. is s u c h t h a t A x 6 V for

all x 6 V;

then, it is k n o w n from linear algebra that there exists a basis for X

(namely, any basis (v I .... ,v n) w i t h the p r o p e r t y that ( V l , . . . , v d) is
also a basis f o r V) in w h i c h A is r e p r e s e n t e d by means o9 a b l o c k -
triangular matrix

A11 A12

0 A22

whose elements o n the lower (n-d) rows and left d columns are vanishing.
Moreover, if t h i s subspace V is such that:

(ii) V c o n t a i n s the image of the mapping B, i.e. is s u c h that Bu E V

for all u E U;

then, choosing again the same basis as b e f o r e f o r X, regardless of the

choice of basis i n U, the mapping B is r e p r e s e n t e d by a matrix

0
 2

whose last n-d rows are vanishing.

Thus, if t h e r e exists a subspace V which satisfies (i) and (ii),
then there exists a choice of c o o r d i n a t e s for X in w h i c h the control
system is d e s c r i b e d by a set of d i f f e r e n t i a l equations o f the form

Xl = AIIXl + A12x2 + BIu

~2 = A 2 2 x 2

By x I a n d x 2 we denote the d - v e c t o r and, respectively, the n - d vector

formed by taking the first d and, respectively, the last n-d coordina-
tes of a p o i n t x of X in the selected basis.
The representation thus obtained is p a r t i c u l a r l y interesting when
studying the behavior of the system under the a c t i o n of the control
u. A t time T, the coordinates of x(T) are

f
T
x I (T) = e x p ( A 1 1 T ) x I (0) + ~ e x p ( A 1 1 (T-T))AI2exp(A22T)dTx2(0) +
0

t
T
+ i e x p (At I ( T - T ) ) B l U (T) dT
0

x 2(T) = e x p ( A 2 2 T ) x 2(0)

From this we see that the set of c o o r d i n a t e s denoted with x 2 does

not depend on the input u but only on the time T. The set of p o i n t s
that c a n be r e a c h e d at time T, starting f r o m x(0), under the a c t i o n of
the input lies inside the set of p o i n t s of X w h o s e x 2 coordinate is
equal to e x p ( A 2 2 T ) x 2 ( 0 ) . In o t h e r words, if we let x(T) denote the
point of X r e a c h e d at t i m e T when u(t) = 0 for all t E [0,T} , we ob-
serve that the state x(T) may be e x p r e s s e d as

x(T) = x(T) + v

where v is a v e c t o r in V. Therefore, the set of p o i n t s that c a n be

reached at t i m e T, starting from x(0), lies inside the set

S T = x(T) + V

Let us n o w m a k e the additional assumption that the subspace V,

which is the starting point of o u r considerations, is s u c h that:

(iii) V is the smallest subspace which satisfies (i) a n d (ii) (i.e.

is contained in a n y o t h e r subspace of X w h i c h satisfies both (i)
 and (ii)).

It is k n o w n from the linear theory that this happens if and o n l y

if
n-1
V = ~ Im(AiB)
i=0

and, moreover, that in this case the pair (AII,B I) is a r e a c h a b l e pair,

i.e. satisfies the condition

r a n k ( B I A 1 1 B I ... A~;IBI) = d

or, in o t h e r words, for e a c h x I E ~d there e x i s t s an i n p u t u, defined

on [0,T], such that
T
xI = [ exp(A11 (T-T))B~U(T)dT
0
Then, if V is s u c h that the condition (iii) is a l s o satisfied,
starting f r o m x(0) we can reach at time T any state of the form
x(T) + v with v E V or, in o t h e r words, any state belonging to the
set S T . T h i s set is t h e r e f o r e exactly the set of the states reachable
at time T starting f r o m x(0) .
This result suggests the following considerations. Given a linear
control system, let V be the smallest subspace of X s a t i s f y i n g (i) a n d
(ii). Associated with V there is a partition of X i n t o subsets of the
form
x+V

with the p r o p e r t y that each one of t h e s e subsets coincides with the

set of p o i n t s reachable at some time T starting from a suitable point
of X. M o r e o v e r , these subsets have the structure of a d-dimensional
flat submanifold of X.
An a n a l y s i s similar to the one developed so far c a n be c a r r i e d
out by e x a m i n i n g the interaction between state and output. In this
case we c o n s i d e r a d-dimensional subspace W of X s u c h that

(i) W is i n v a r i a n t under the m a p p i n g A

(ii) W is c o n t a i n e d into the kernel of the m a p p i n g C (i.e. is s u c h
that C x = 0 for all x 6 W)
(iii) W is the largest subspace which satisfies (i) a n d (ii) (i.e.
contains any other subspace of X w h i c h satisfies both (i) a n d
(ii)).

Then, there is a c h o i c e of c o o r d i n a t e s for X in w h i c h the control

system is d e s c r i b e d by e q u a t i o n s of the form
 4

Xl = A11Xl + AI2X2 + BIU

x2 = A 2 2 x 2 + B2u

y = C2x 2

From this we see that the set of c o o r d i n a t e s denoted with x I has

no influence on the o u t p u t y. Thus any two initial states whose last
n-d coordinates coincide produce two identical outputs under any input,
i.e. are indistinguishable. Actually, any two s t a t e s whose last
n-d coordinates coincide are such that their difference is an e l e m e n t
of W and, then, we m a y conclude that any two states belonging to a set
of the f o r m x + W are indistinguishable.
Moreover, we k n o w that the c o n d i t i o n (iii) is s a t i s f i e d if a n d
only if
n-1
W = ~ k e r ( C A i)
i=0

and, if this is the case, the p a i r (C2,A22) is o b s e r v a b l e , i.e. sa-

tisfies the c o n d i t i o n

rank(C~ A 2 C ~ ... (A~2)d-Ic~) = d

or, in o t h e r words,

C22exp(A22t)x 2 - 0 ~ x 2 = 0

Then, if two initial states are such that their difference does
not belong to W, t h e y m a y be d i s t i n g u i s h e d from e~ch other by the out-
put produced under zero input.
Again we m a y synthesize the above discussion with the following
considerations. Given a linear control system, let ~ be the largest
subspace of X s a t i s f y i n g (i) and (ii). Associated with W there is a
partition of X i n t o s u b s e t s of the f o r m
x+W

with the p r o p e r t y that each one of these subsets coincides with the
set of p o i n t s that are indistinguishable from a fixed point Qf X. Mo-
reover, these subsets have the structure ~f ~ d-dimensional flat sub-
manifold of X.
In the following sections of this chapter and in the following
chapter we shall deduce similar decompositions for nonlinear control
systems.

2. D i s t r i b u t i o n s on a Manifold

The easiest way to i n t r o d u c e the n o t i o n of distribution A~ o n a

manifold N is to c o n s i d e r a mapping assigning to e a c h point p of N a
subspace A(p) o f the tangent space T N t o N at p. T h i s is n o t a r i g o -
P
rous d e f i n i t i o n , in t h e sense t h a t w e h a v e o n l y d e f i n e d the d o m a i n N
of A w i t h o u t giving a precise characterization of its codomain. Defer-
ring for a m o m e n t the n e e d for a m o r e rigorous definition, we proceed
by a d d i n g some conditions of regularfty. This is i m p o s e d by assuming
that for e a c h point p of N there exist a neighborhood U of p and a set
of s m o o t h vector fields defined o n U, d e n o t e d {~ : i C I}, with the
1
property that,

A(q) = span{Ti(q):i E I}

for all q E U. Such an o b j e c t will be c a l l e d a smooth distribution on

N. U n l e s s otherwise noted, in the following sections we will use the
term "distribution" to m e a n a smooth distribution.
Pointwise, a distribution is a l i n e a r object. Based on this pro-
perty, it is p o s s i b l e to e x t e n d a number of elementary concepts re-
lated to the notion of subspace. Thus, if {T.:i 6 I} is a set o f v e c t o r
l
fields defined on N, their span, written s p { T i : i E I}, is the d i s t r i b u -
(*)
tion defined b y the rule

sP{Ti:i 6I}: p ~ span{Ti(P):i 6I}

If A I a n d A 2 a r e two dLstributions, their sum A I + A 2 is d e f i n e d by

taking

A I + A 2 ": p ~ A I (p) + A 2(p)

and their intersection g I ~ A 2 by t a k i n g

A I M A 2 : p ~ AI (p) ~ A2(P)

(*) In order to avoid confusions, we ,use the symbol span{-} to denote any B-linear
combination of elements of some ~-vector space (in particular, tangent vectors
at a point). The symbol sp{'} is used to denote a distribution (or a codistribu-
tion, see later).
 6

A distribution A I i8 contained in t h e d i s t r i b u t i o n A 2 and is w r i t t e n

A 1C A 2 if At(p) C A2(p) f o r a l l p 6 N. A v e c t o r field T beZong8 to a
distribution A and is w r i t t e n T 6 A if T(p) E A(p) for all p 6 N.
The dimension of a distribution A at p E N is the d i m e n s i o n of
the subspace A(p) o f TpN.
Note that the s p a n of a g i v e n set of smooth vector fields is a
smooth distribution. Likewise, the s u m of two smooth distributions is
smooth. However, the intersection of t w o such distributions may fail
to be s m o o t h . This m a y be seen in the following example.

(2.1) Ezamp~e. Let M = 2 2 , and

= sp{_ 8 + --~ }
AI ~x I 8x 2

A 2 = sp{(1+Xl)~1 + ~x 2

Then we have

(A I N A2 ) (x) = {0} if xI ~ 0

(A I 63 A 2) (x) = A I (x) = A2(x) if xI = 0

This distribution is n o t smooth because it is n o t possible to find a

smooth vector field on 2 2 which is zero everywhere but on the line
x I = 0. []

Since sometimes it is u s e f u l to t a k e the intersection of smooth

distributions A I and A 2 , one may overcome the p r o b l e m t h a t A I N A2 is
possibly non-smooth with the a i d of the following concepts. Suppose A
is a m a p p i n g which assigns to e a c h point p ~ N a subspace A(p) of T p N
and let M(A) be the set of all smooth vector fields defined on N which
at p t a k e values in A(p), i.e.

M(A) = {T e V ( N ) : T ( p ) 6 A(p) for a l l p 6 N}

Then, it is n o t difficult to see that the s p a n o f M(A), in the

sense defined before, is a s m o o t h distribution contained in A.

(2.2) Remark. Recall that the s e t V(N) of a l l smooth vector fields de-
fined on N may be g i v e n the structure of a v e c t o r space over ~ and,
also, the structure of a m o d u l e over C~(N), the r i n g o f all smooth
real-valued functions defined o n N. The set M(A) defined before (which
is n o n - e m p t y because the zero element of V(N) belongs t o M(A) for any
4) is a s u b s p a c e of the v e c t o r space V(N) and a submodule of the m o d u l e
V(N). From this is it e a s i l y seen that the span of M(4) is c o n t a i n e d
in 4. []

Note that if 4' is a n y smooth distribution contained in 4, then

4' is c o n t a i n e d in the s p a n of M(4), so the span of M(4) is a c t u a l l y
the largest smooth distribution contained in 4. To i d e n t i f y this distri-
bution we shall henceforth use the n o t a t i o n

smt(A) =4 sp M(A)

i.e. we look at the span of M(A) as the "smoothing" of 4. N o t e also

that if ~ is smooth, then smt(~) = 4.
Thus, if A I N 42 is n o n - s m o o t h , we shall rather consider the dis-
tribution smt(41N 42).

(2.3) Remark. N o t e that M(A) may not be the unique subspace of V(N),
or s u b m o d u l e of V(N), whose span coincides with smt(4). But if M' is
any o t h e r subspace of V(N), or submodule of V(N), with the property
that sp M' = smt(&), then M' c M(~).

(2.4) Example. L e t N = ~ , and

4 = sp{x ~}

Then M(4) is the set of all vector fields of the form c(X)~x where
c(x) is a s m o o t h function defined on ~ which vanishes at x = 0 . C l e a r l y
is s m o o t h and coincides with smt(A). There are many submodules of
V(~) which span 4, for i n s t a n c e

M~ = {T 6 V ~ ) : T ( x ) = c(x)x~-~ and c 6 C~)}

M2 = {T 6 V ~ ) : T ( x ) = c ( x ) x 2 ~-~
~ and c 6 C~(~)}

Both are submodules of M(A), M 2 is a s u b m o d u l e of M I but M I is not a

submodule of M 2 b e c a u s e is n o t p o s s i b l e to e x p r e s s every function
c(x) x as 6 ( x ) x 2 w i t h 6 6 C ~) . []

(2.5) Remark. The previous considerations enable us to give a rigorous

definition of a s m o o t h distribution in the following way. A smooth
distribution is a s u b m o d u l e M of V(N) with the following property: if
0 is a s m o o t h vector field such that for all p E N

8(p) 6 span{T(p):T 6 M}
 8

then 8 belongs to M. []

Other important concepts associated with the notion of distribu-

tion are the ones related to t h e "behavior" of a given A as a "func-
tion" of p. We have already seen how it is possible to c h a r a c t e r i z e
the quality of being smooth, but there are other properties to b e con-
sidered.
A distribution A is n o n ~ i n g u l a r if there exists an integer d such
that

(2.6) dim A(p) = d

for all p E N. A singular distribution, i.e. a distribution for which

the above condition is n o t satisfied, is sometimes called a distribu-
tion of v a r i a b l e dimension. If a d i s t r i b u t i o n A is such that the con-
dition (2.6) is satisfied for all p belonging to an o p e n subset U of
N, then we say that A is n o n s i n g u l a r on U. A p o i n t p is a regular p o i n t
of a distribution A if there exists a neighborhood U of p with the pro-
perty that A is n o n s i n g u l a r on U.
There are some interesting properties related to these notions,
whose proof is left to the reader.

(2.7) Lemma. Let A be a smooth distribution and p a regular point of

g. Suppose dim g(p) = d. Then there exist an open neighborhood U of
p and a set {T I ..... T d} of smooth vector fields defined on U with the
property that every smooth vector field T belonging to A a d m i t s on U a
representation of the form

d
(2.8) = ~ ci~ i
i=I

where each c. is a real-valued smooth function defined on U. []

1
A set of d v e c t o r fields which makes (2.8) satisfied will be cal-
led a set of local generators for g a t p.

(2.9) Lemma. The set of all regular points of a distribution A is an

open and dense submanifold of N.

(2.10) Lemma. Let A 1 and A 2 be two smooth distributions with the pro-
perty that A 2 is nonsingular and At (p) C A2(p) at each point p of a
dense submanifold o f N. Then A I C A2"

(2.11) Lemma. Let A I and A 2 be two smooth distributions with the pro-
perty that A I is nonsingular, A I C A 2 and At(p) = A2(p) at each point
p of a dense submanifold of N. Then A 1 = A 2. []
 We h a v e seen before that the intersection of t w o smooth distribu-
tions may fail to be smooth. However, around a regular point this
cannot happen, as w e see from the following result.

(2.12) Lemma. Let p be a regular point of ~1 , A 2 and A 1 Q A 2. T h e n

there exists a neighborhood U of p w i t h the p r o p e r t y that A1 n A2
restricted to U is s m o o t h . []

A distribution is n v o l u t i ~ e if the L i e bracket [ T I , T 2] o f a n y

p a i r of v e c t o r fields T I and T2 belonging to A is a v e c t o r field which
belongs t o A, i.e. if

ql ,5, [ 2 C A ~ ['~1"t2] E A

(2.13) Remark. I t is e a s y to see that a nonsingular distribution of

dimension d is i n v o l u t i v e if a n d o n l y if, at e a c h point p, any set of
local generators TI,...,T d defined on a neighborhood U o f p is s u c h
that
d
[ 4 = ~ c~.~

where each c~. is a r e a l - v a l u e d smooth function defined on U. []

z3
If f is a v e c t o r field and A a distribution o n N we d e n o t e by
[ f,A] the distribution

(2.14) [f,A] = sp{[ f,T] V(N):T e A}

Note that [ f,A] is a s m o o t h distribution, even if A is not. Using this

notation, one can say that a distribution is i n v o l u t i v e if a n d o n l y if
[f,A] C A for all f &.
Sometimes, it is u s e f u l to w o r k with objects that are dual to the
ones defined above. In the same spirit of the d e f i n i t i o n given at t h e
beginning of this section, we say that a eodistribution ~ o n N is a
mapping assigning to e a c h point p of N a subspace ~(p) o f the cotangent
s p a c e T (N). A s m o o t h c o d i s t r i b u t i o n is a c o d i s t r i b u t i o n ~ o n N w i t h
P
the p r o p e r t y t h a t for e a c h p o i n t p o f N t h e r e e x i s t a n e i g h b o r h o o d U
of p a n d a s e t o f smooth covector fields (smooth one-forms) defined on
U, d e n o t e d {wi:i 6 I}, such that

~(q) = s p a n { e i (q) :i E I}

for a l l q C U.
In the same manner as w e did for distributions we may define the
 10

dimension of a codistribution a t p, and construct codistributions by

taking the span of a given s e t of c o v e c t o r fields, or else by adding
or intersecting two given codistributions, etc. always looking at a
pointwise characterization o f the objects we are dealing with.
Sometimes, one can construct codistributions starting from given
distributions and conversely. The natural way to do t h i s is the fol-
lowing: given a distribution A o n N, the annihiZator o f A, d e n o t e d A i,
is the c o d i s t r i b u t i o n on N defined by the r u l e

A i : p ~ { v* E T p*N : (v*,v)= 0 for all v 6 A(p)}

!
Conversely, the annihilator o f ~, d e n o t e d ~-, is the d i s t r i b u t i o n
defined by the rule

~i : p ~ {v E TpN: (v* ,v ) = 0 for all v *E ~(p)}

Distributions and codistributions thus related possess a number

of interesting properties. In p a r t i c u l a r , the s u m o f the d i m e n s i o n s of
A a n d o f A is e q u a l to the dimension o f N. T h e inclusion A I C A 2 is
satisfied if a n d only if the inclusion A I D A 2 is s a t i s f i e d . The an-
nihilator (A I A &2 ) o f a n i n t e r s e c t i o n of distributions is e q u a l to
I
the s u m A 1 + A 2.
Like in t h e case of the distributions, some care is r e q u i r e d when
dealing with the quality of being smooth for c o d i s t r i b u t i o n s construc-
ted in s o m e of the w a y s we described before. Thus it is e a s i l y seen
that the s p a n of a g i v e n set o f smooth covector fields, as w e l l as s u m
of two smooth codistributions is a g a i n smooth. But the intersection of
two such codistributions may not need to be s m o o t h .
Moreover, the annihilator of a smooth distribution may fail to be
smooth, as it is s h o w n in the following example.

(2.15) Example. Let N =

A = sp{x ~8'
}

Then
Ai(x) = {0} if x ~ 0

AA(x) = T N if x : 0
x

and we see t h a t A i is n o t smooth because it is n o t p o s s i b l e to f i n d a

smooth covector field o n I~ w h i c h is zero everywhere but on the p o i n t
 11

x = 0. []

Or, else, the annihilator of a s m o o t h codistribution may n o t be

smooth, as in the following example.

(2.16) EmampZe. Consider again the two distributions A I and A 2 describ-

ed in the E x a m p l e (2.1). One may easily check that

A I = s p { d x I - dx2}

A i2 = sp{dXl _ (I + x l ) d x 2}

The intersection At A A2 is not smooth but its annihilator AtL + A 2I is

smooth, because both A I and A 2 are smooth. []

One may easily extend Lemmas (2.7) to (2.12). In p a r t i c u l a r , if

p is a r e g u l a r point of a codistribution e and dim S(p) = d, then it
is p o s s i b l e to f i n d an o p e n neighborhood U of p and a set {m1'''''md }
of s m o o t h covector fields defined o n U, such that every smooth covector
field m belonging to ~ can be e x p r e s s e d on U as

d
e = [ ci~ i
i=I

where each c i is a r e a l - v a l u e d smooth function defined o n U. The set

{ ~ 1 ' ' ' ' ' ~ d } is c a l l e d a set of local generators for ~ at p.
We h a v e seen before that the annihilator of a s m o o t h distribution
A may fail to be smooth. However, around a regular point of A this
cannot happen, as we see from the following result.

(2.17) Lemma. L e t p be a r e g u l a r point of A. Then p is a r e g u l a r p o i n t

of A i and there exists a neighborhood U of p w i t h the property that A~
restricted to U is smooth. []

We conclude this section with some notations that are frequently

used. If f is a v e c t o r field and ~ a codistribution o n N we d e n o t e by
Lf~ the smooth codistribution

(2.18) Lf~ = sp{Lfw V~(N):~ ~}

If h is a r e a l - v a l u e d smooth function defined o n N, one may as-

sociate with h a distribution, written ker(h~), defined by

ker(h~) : p ~ {v 6 T p N : h ~ v = 0}

One m a y also associate with h a codistribution, taking the span of the

 12

covector f i e l d dh It is e a s y to v e r i f y that the two objects thus de-

fined are one the annihilator of t h e other, i.e. that

(sp(dh)) = ker(h#) .

3. F r o b e n i u s Theorem

In t h i s section we shall establish a correspondence between the

notion of d i s t r i b u t i o n on a manifold N and the e x i s t e n c e of partitions
of N into lower dimensional submanifolds. As we h a v e seen at the be-
ginning of t h i s chapter, partitions of the state space into lower di-
mensional submanifolds are often encountered when dealing with reach-
ability and/or observability of c o n t r o l systems.
We begin our analysis with the following definition. A nonsing~lar
d-dimensional distribution A on N is completely integrable if a t e a c h
p 6 N there exists a cubic coordinate chart (V,~) with coordinate func-
tions ~1,...,~n , such that

(3.1) A(q) = s p a n { ( ~ I ) q ..... ( ~ d ) q }

for all q C V.
There are two important consequences related t o the n o t i o n of
completely integrable distribution. First of all, observe that if t h e r e
exists a cubic coordinate chart (V,~), with coordinate functions
~I .... '~n ' such that (3.1) is s a t i s f i e d , then any slice o f V p a s s i n g
through any point p of V and defined by

(3.2) Sp = {q E V : ~ i (q) = ~i (p) ; i = d+1 ..... n}

(which is a d - d i m e n s i o n a l imbedded submanifold o f N), has a tangent

space which, at any point q, coincides with the subspace &(q) of T N.
q
Since the set of a l l such slices is a partition of V, w e m a y see
that a completely integrable distribution A induces, locally around
each point p E N, a partition into lower dimensional submanifolds, and
each of these submanifolds is s u c h that its t a n g e n t space, at each
point, agrees with the distribution A at t h a t p o i n t .
The second consequence is t h a t a completely integrable distribu-
tion is inuoiutive. In o r d e r to see t h i s we u s e the d e f i n i t i o n of in-
volutivity and compute the Lie bracket of any pair of vector fields
belonging t o A. For, recall that in the ~ coordinates, any vector
field defined o n N is r e p r e s e n t e d by a vector of t h e form
 13

~(~) = (~1(~) . . . T n ( ~ ) ) '

The c o m p o n e n t s of this vector are related to the v a l u e of the vector

field at a p o i n t q by the expression

~(q) = T 1 (~(q))(~1)q +-.-+Tn(~(q)) (~n)q

If T is a v e c t o r field of A a n d (3.1) is s a t i s f i e d , the last n-d

components Td+1 ( ~ ) , . . . , T n ( ~ ) must vanish. Moreover, if @ is any o t h e r
vector field of 4, also the last n-d components of its local repre-
sentation

e(() = (@1(~) . . . @ n ( ( ) ) '

m u s t vanish. From this one deduces immediately that also the last n-d
components of the v e c t o r

9e 3~

are vanishing. Since this vector represents locally the vector field
IT,e] one may conclude that [T,e] belongs to ~, i.e. that A is invo-
lutive.
We have seen that involutivity is a n e c e G s a ~ y condition for the
complete integrability of a d i s t r i b u t i o n . However, it c a n be p r o v e d
that this condition is a l s o sufficient, as it is s t a t e d below

(3.3) 2~eorem (Frobenius). A nonsingular distribution is c o m p l e t e l y

integrable if a n d o n l y if it is i n v o l u t i v e

Proof. Let d denote the dimension of A. Since A is n o n s i n g u l a r , given

any p o i n t p @ N it is p o s s i b l e to find d vector fields ~1,...,Td E
with the p r o p e r t y that TI(q) ..... Td(q) are linearly independent for
all q in a s u i t a b l e neighborhood U of p. In o t h e r words, these vector
fields are such t h a t

~(q) = s p a n { T 1(q) ..... Td(q) }

for all q 6 U.
Moreover, let rd+1,...,Tn be a n y o t h e r set of v e c t o r fields with
the p r o p e r t y that
s p a n { T i ( p ) : i = I .... n} = T N. W i t h e a c h v e c t o r
' p Ti
field T i , i = 1,...,n, we a s s o c i a t e its f l o w #t and we c o n s i d e r the
mapping
 14

F : C (0) >N
E

T2 ~n
(~1 ,~n) ~ T 1
: . . . . . . %.1 ~ 2 " ~n(p)
where Ce(O) = {~ ~ ~n:l~il < e, 1 i 3 n}.
If 8 is s u f f i c i e n t l y small, this mapping:

(i) is d e f i n e d for all ~ 6 Ce(O ) and is a d i f f e o m o r p h i s m onto its

image
(ii) is such that for all ~ 6 Co(O)

F,(~i) ~ e A(F(~)) i = I ..... d (*)

We show now that (i) and (ii) are true and, later, that both imply
the thesis.

Proof of (i). We know that for each p 6 N and s u f f i c i e n t l y small Itl

the flow ~ ( p ) of a vector field T is defined and this makes the func-
tion F d e f i n e d for all (~I ..... ~n ) with s u f f i c i e n t l y small l~iI.More-
over, since a flow is smooth, so is F. lle prove that F is a local dif-
feomorphism by showing that the rank of F at 0 is equal to n.
To this purpose, we first compute the image under F,) of the
tangent vector (~) at a point ~ E C (0) Suppose F is e x p r e s s e d in
~i ~ ~ "
local coordinates. Then, it is known that the coordinates of F ~ ( ~ ) p
in the basls
{(~-~)
~ ~
,..., (~--~-) } of the tangent space to N at the
~i ~
I q ~n q
point q = F(~) coincide with the e l e m e n t s of the i-th column of the
jacobian m a t r i x

~F

By taking the partial derivative of F w i t h respect to ~i we ob-

tain

T1 (#~i-1 ~ (.Ti ~Tn(p)) =

8F~i = (#~1)* "'" ~i-I )~ ~ i . w ~ i o .. . o ~n

T1 (Ti-1 ~i CTn
= (~I)~ "'" ~i_1 )~ Yi o ~i ..... ~n(p) =

= (~i), ... (~Ti-1

~i_1), Ti o =-~i-1
~i-1 . "" o ~-~I(F(~))

(*} Note that (~i)~

~ is a tangent vector at the point ~ of CE(O)
 15

In particular, at ~ = 0, since F(0) = p,

F,(%)0 = Ti (p)

The tangent vectors Tl(p),...,Tn(p) are by assumption linearly inde-

pendent and this proves that F, has rank n at p.

Proof o~ (ii). From the previous computations, we deduce also that, at

any ~ 6 Ce(0),

F,(~i) = TI (~Ti-1 .Ti-1 T1

(~{1)~ ... {i_I )~ T i o9_[i_ I ..... #_~1 (q)

where q = F(~).
If we are able to prove that for all q in a neighborhood of p, for Itl
small, and for any two vector fields T and @ belonging to A,

( ~ } , T o~St(q)_ E ~(q)

i.e. that ( ~ ) , ~ o#St is a (locally defined) vector field of 4, then

we easily see that (ii) is true.
To prove the above, one proceeds as follows. Let 8 be a vector
field of A and set

V i(t) = (~8t)_ ~ T i o8t(q)

for i = 1,...,d.
Then, from a well known property of the Lie bracket we have

dVi =
dt (#@-t)~[ 8'Ti] ~St(q)

Since both T i and 8 belong to A and A is involutive, there exist func-

tions I. defined locally around p such that
l]
d
I 0,~ i] = ~ lij~ j
j=1

and, therefore,

d d
dVi = (~St)~[
dt j~1 lij (~t8 (q))] Tj oSt(q) = j~lliJ (~8t(q))vj(t)
 16

The functions Vi(t) are seen as solutions of a linear differential

equation and, therefore,it is possible to set

[V1(t)...Vd(t)] = [VI(0)...Vd(0)]X(t)

where X(t) is a dxd fundamental matrix of solutions. By multiplying on

the left both sides of this equality by (~)~ we get

[ T1o~(q)...TdO~(q)] =[ ( ~ ) ~ T 1 ( q ) . . . (~)~d(q)]X(t)

and also, by replacing q with %~t(q)

Since X(t) is nonsingular for all t we have that, for i = 1,...,d,

e
(~t),Tio#8_t(q) C span{Y I (q) ..... Tp(q) }

i.e.

~
(t),~io~@t(q) C A(q)

This result, bearing in mind the possibility of expressing any

vector T of A in the form

d
T = ~ ciT i
i=1

completes the proof of (ii).

From (i) and (ii) the thesis follows easily. Actu~lly, (i) makes
it possible to consider on the neighborhood v = F(CE{0)) of p the
coordinate chart (V,F-I). By definition, the tangent vector (~)q at

a point q E V coincides with the ima~e~under F, of ~the tangent vector

(~)~ at the point ~ ='F-1(q) E C(0). From (ii) we see that the

tangent vectors

are elements of A(q). Since these vectors are linearly independent,

they span A(q) and (3.1) is satisfied. []
 17

There are several interesting system-theoretic consequences of

Frobenius' Theorem. The most important one is found in the c o r r e s p o n d -
ence, established by this Theorem, between involutive distributions
and l o c a l partitions of a m a n i f o l d into lower dimensional submanifolds.
As we have seen, given a nonsingular and completely integrable, i.e.
involutive, d-dimensional distribution & on a m a n i f o l d N, around each
p 6 N it is p o s s i b l e to f i n d a coordinate neighborhood V on w h i c h A
induces a partition into submanifolds of d i m e n s i o n d, w h i c h are slices
(and,therefore, imbedded submanifolds) of V. C o n v e r s e l y , given any
coordinate neighborhood V, a partition of V i n t o d-dimensional slices
defines on V a n o n s i n g u l a r completely integral distribution of d i m e n -
sion d.
We e x a m i n e some examples in o r d e r to further clarify these con-
cepts

(3.4) Example. L e t N = A n and let x = (Xl,...,Xn) be a point on ~ n .

Suppose V is a s u b s p a c e of ~ n , of dimension d, spanned by the v e c t o r s

vi = (Vil ..... Vin) I ~ i ~ d

We m a y a s s o c i a t e with V a distribution, denoted AV , in the following

way. At each x 6 A n, ~ v ( X ) i s the subspace of T ~ n s p a n n e d by the
tangent vector~

n,
j Ivij ( ~ x I _< i < d

It fs e a s i l y seen that this distribution is n o n s i n g u l a r and in-

volutive, thus completely integrable.
Now, suppose we p e r f o r m a (linear) change of c o o r d i n a t e s in
~n, ~ = ~(x) such that

~i(vj ) = ~ij

In the ~ c o o r d i n a t e s , the. s u b s p a c e V will be spanned by vectors of the

form (1,0 ....... 0), (0,1 .... ,0), etc., while the subspace Av(X) by the
tangent vectors (~)x ..... ( ~ ) x " Thus, we see that the condition

(3.1) i~s s a t i s f i e d globally on ~ n in the ~ coordinates.

The slices~

S = {x e A n : ~i(x) = ci , i = d + I ..... n}

characterize a global partition of ~ n anrd e a c h of these is s u c h that

 18

its t a n g e n t space, at e a c h p o i n t x, is e x a c t l y ~v(X). It is w o r t h

noting that e a c h o f t h e s e slices corresponds to a set of the f o r m

x+V

Thus, the p a r t i t i o n s of the s t a t e space X discussed in s e c t i o n I may

be t h o u g h t of as g l o b a l partitions induced by a d i s t r i b u t i o n associated
with a given subspace of X.

(3.5) Ezample. L e t N = ~ 2 a n d let x = (Xl,X 2) he a p o i n t on ~ 2 . C o n -

sider the o n e - d i m e n s i o n a l nonsingular distribution

)__~ +__!_~}
& = sp{ (exp x 2 ~x I ~x 2

If we w a n t to f i n d a c h a n g e of c o o r d i n a t e s that m a k e s (3.1) satisfied,

we m a y p r o c e e d as follows. Recall that, given a coordinate chart with
coordinate functions ~i,~2 , a tangent vector v at x m a y be r e p r e s e n t e d
as

v Vl ( )x x

where the c o e f f i c i e n t s v I a n d v 2 are s u c h t h a t v I = Lv~ I a n d v 2 = Lv~ 2.

Since the t a n g e n t vector

T(x) = (exp x 2) ( ~ 1 ) x + (~2)x

spans A(x) at e a c h x E ~ 2 , if we w a n t that (3.1) is satisfied we have

to h a v e

T(x) = (LT~ I) (-~1)x + (LT 2) ( )x = ( )x

for all x @ U, or

~I 2I
1 = (LT~ I) = (exp x2)~--~i + ~x 2

~2 ~2
0 = (LT~2) = (exp x 2 ) ~ I + ~x2

A solution o f this set of p a r t i a l differential e q u a t i o n s is g i v e n

by
~I = ~I (x) = x2

~2 = ~2 (x) = Xl - exp(x2)
 19

The mapping ~ = ~(x) is a d i f f e o m o r p h i s m ~ : ~2 ~2 and solves

the problem of finding the change of coordinates that makes (3.1) sa-
tisfied. Note that ~ 2 is globally p a r t i t i o n e d into o n e - d i m e n s i o n a l
slices, each one being the locus where the function {2(x) is constant,
i.e. the locus of points (xl,x2) such that

x I = exp(x2) + constant O

The procedure described in the Example (3.5) may easily be ex-

tended. For, let A be a n o n s i n g u l a r involutive distribution of dimen-
sion d. Let (U,~) be a coordinate chart with coordinate functions
~1,...,~n. Given any point p E U it is possible to find d vector fields
TI ..... Td 6 A with the p r o p e r t y that TI(q) .... ,Td(q) are linearly in-
dependent for all q in a suitable n e i g h b o r h o o d U' C U of p. In other
words, these vectors are such that

A(q) = span{T I (q) ..... Td(q) }

for all q 6 U'.

In the c o o r d i n a t e s ~1,...,~n , each of these v e c t o r fields is
locally e x p r e s s e d in the form

If (V,~) is another coordinate chart around p with coordinate

functions ~l,...,~n the corresponding expressions for Ti has the form

n
X i = j=1
[ (LTi~J) ( % )

For (3.1) to be satisfied, i.e. for

sP{T I T d} = s P { ~ I
. . . . . . . . . . 3~d }
on V, we must have

LT. ~j = 0
l

on V, for i = 1,...,d and j = d + 1 , . . . , n and, m o r e o v e r

LT I LT 1
rank = d

L~d~ 1 ... LTd~ d

 20

on V. These conditions characterize a set of partial differential equa-

tions on V, which has to be satisfied by the new coordinate functions

~I ..... ~n"
Setting, as usual

-I
~ij(x) = (LTi~j)o ~ (x)

where x = (Xl,...,x n) E ~n, it is possible to express the functions

LT.~ j involved in the previous conditions as follows
l
n ~(~jo-1)
LTi~j (x) = k:1[Tik(X) SXk

-I
Therefore, using just ~i(x) to denote the composite function ~jo9 (x) ,
one has

n ~j
LTi~ j (x) = k=~iTik(X)~x k

Setting

T11 (x)...Tdl (x)

T(x) =

Tin (x) . . . Tdn (x)

the previous equations for L ~ j become

1

K(x)
(3.6) $---~
~ T(x) =

0(n-d)d

in which K(x) is some dd matrix of real valued functions, nonsingular

for all x @ 9(V).
Thus, we may conclude that finding a coordinate transformation
= ~(x) that makes (3.1) satisfied corresponds to solving a partial
differential equation of the form (3.6).
Note that the matrix T(x) is a matrix of rank d at x = ~(p) be-
cause the tangent vectors TI(p),...,Td(p) are linearly independent.
Therefore the matrix ~-~ can be n o n s i n g u l a r at x = ~(p) and this, ac-
cording to the rank Theorem, guarantees that ~ = ~(x) is a local dif-
feomorphism.

(3.7) Remark. There are alternative ways to describe the equation (3.6).
 21

For instance, one m a y e a s i l y check that solving these e q u a t i o n s cor-

responds to f i n d n - d f u n c t i o n s 1 1 , . . . , l n _ d d e f i n e d on a n e i g h b o r h o o d
V of p w i t h v a l u e s in ~ with the following p r o p e r t i e s

(i) the t a n g e n t eovectors dll(p) , .... d l n _ d ( p ) are l i n e a r l y independ-

ent
(ii) ( d l i ( q ) , T j ( q ) ) = 0 for all q 6 V, i = I .... ,n-d and j = 1 , . . . , d .

In fact, if (V,~) is a c o o r d i n a t e chart that m a k e s (3.6) satis-

fied, then the f u n c t i o n s

li = ~i+d

will s a t i s f y (i) a n d (ii). C o n v e r s e l y , if 1 1 , . . . , I n _ d is a set of func-

tions that s a t i s f i e s (i) a n d (ii), then it is a l w a y s possible to find
d functions ~1,...,~d defined on V a n d w i t h v a l u e s in ~ w h i c h , to-
gether w i t h the functions ~d+1 = h1'''''~n = in-d ' define a coordinate
chart (V,~) w i t h ~ solving the e q u a t i o n s (3.6).
From (ii) we d e d u c e also that there is a set of covector fields
{ d l l , . . . , d l n _ d} w i t h the p r o p e r t y t h a t at e a c h q 6 V, < d l i ( q ) , v ) = 0
for all v E A(q). Thus

dl i(q) 6 A (q) i = 1 ..... n - d

Moreover, the t a n g e n t covectors dll ( q ) , . . . , d l n d(q) are l i n e a r l y in-

-
dependent for all q in a n e i g h b o r h o o d of p and A (q) has e x a c t l y di-
m e n s i o n n-d. Therefore, we m a y c o n c l u d e that the set of c o v e c t o r fields
{ d l l , . . . , d l n _ d} s p a n s A l o c a l l y a r o u n d p.
In short, we m a y state this r e s u l t by s a y i n g that a nonsingular
distribution of d i m e n s i o n d is i n t e g r a b l e if a n d o n l y if its a n n i h i -
lator is l o c a l l y s p a n n e d by n - d exact o n e - f o r m s .

(3.8) Remark. We n o t e that the i n v o l u t i v i t y of A c o r r e s p o n d s to the

property t h a t a n y two c o l u m n s Ti(x) a n d Tj(x) of the m a t r i x T(x) are
such that

i 3
(-~-~ ~ j ( x ) - - ~ x ~i(x))e Im(T(x))

for all x E ~(V).

(3.9) Remark. We k n o w that, given a set of f u n c t i o n s {li:i E I}, de-

fined on N and w i t h v a l u e s in ~ , we can d e f i n e a codistribution
= sp{dli:i ~ I}. It is e a s i l y seen that if ~ is n o n s i n g u l a r then ~
is c o m p l e t e l y integrable. For, let d d e n o t e the d i m e n s i o n of ~, take
a p o i n t p 6 N a n d a set of f u n c t i o n s 11,...,I d with the p r o p e r t y that
 22

~(p) = s p a n { d l I (p) ..... dld(p))

If U is a n e i g h b o r h o o d of p w i t h the p r o p e r t y that dl1(q),...,dld(q)

are linearly independent at all q C U, it is seen that ~ is s p a n n e d on
U by the e x a c t forms d l 1 , . . . , d h d. As a c o n s e q u e n c e of our earlier
discussions, ~i is c o m p l e t e l y integrable.

The n o t i o n of c o m p l e t e integrability can be e x t e n d e d to a g i v e n

collection of d i s t r i b u t i o n s . There are two cases of special importance
in the a p p l i c a t i o n s .
Let A I , ~ 2 , . . . , A r be a c o l l e c t i o n of neeted d i s t r i b u t i o n s , i.e. a
set of d i s t r i b u t i o n s with the p r o p e r t y

41C A 2 C ... C ~r

A collection of n e s t e d nonsingular distributions on N is c o m p l e t e l y

integrable if at e a c h point p E N there exists a coordinate chart (V,~)
with coordinate functions ~i,...,~ n such that

A i(q) = span{ ( ~ 1 ) q ..... ( ~ d ) q }

1
for all q E V, w h e r e d i denotes the dimension of 4 i.
The following results extends Frobenius Theorem

3.10) Theorem. A c o l l e c t i o n 41C 42 C ... C 4 r of n e s t e d nonsingular

distributions is c o m p l e t e l y integrable if a n d o n l y if e a c h distribu-
tion of the c o l l e c t i o n is i n v o l u t i v e .

Preof. The s a m e c o n s t r u c t i o n described in the p r o o f of T h e o r e m (3.3)

can be used. []

A collection A I , . . . , 4 r of d i s t r i b u t i o n s on N is s a i d to be in-
dependent if

(i) A i is n o n s i n g u l a r , for a l l i = I, .... r

(ii) Ai N ( ~ A~) = 0, for all i = I ..... r

j~i ~
A collection of d i s t r i b u t i o n s ~1,...,~r is s a i d to span ~he tan-
gent space if for all q 6 N

41 (q) + A 2(q) + ... + ~ r ( q ) = TqN

An independent collection of d i s t r i b u t i o n s 41,...,& r which spans

the tangent space is said to be simultaneously integrable if at e a c h
point p E N there exists a coordinate chart (V,~), with coordinate
 23

functions ~ i , . . . , ~ n such that

(3.11) Ai(q) = span{ ( ~ ) . (~-~)q}

~ s . +z 1 q' "'' si+1

for all q 6 V, w h e r e sI = 0 and

s i = d i m ( g I + ... + A i _ I)

for i = 2 , . . . , r + I .
The following result is an a d d i t i o n a l extension of Frobenius
Theorem

(3.12) Theorem. An i n d e p e n d e n t collection of d i s t r i b u t i o n s AI,...,A r

which spans the tangent space is s i m u l t a n e o u s l y integrable if a n d o n l y
if, for all I < i < r, the d i s t r i b u t i o n

r
--
(3.13) Di j=~1 A .]

jT~i
is i n v o l u t i v e .

Proof. S u f f i c i e n c y . Let n i = dim(Ai). Using Theorem(3.3), at e a c h p o i n t

p one m a y find a n e i g h b o r h o o d V of p and, for e a c h I < i < r, a set of
coordinate functions ~% , I ~ j ~ n, defined on V w i t h the p r o p e r t y
that

D i = sp{ ~ : I < j < n-n.}

An e a s y computation shows that the covector fields

d~In_n I+1 . ." . . d~ In r

. ." . . d ~ n _ n r +I
,...
'
d r
~n

are l i n e a r l y independent at p. Thus, the set of functions

{~:n-ni+1 < j < n; I < i < r} defines on V a set of c o o r d i n a t e func-
tions.
Since D i is t a n g e n t to the slice of v w h e r e all the coordinate
functions ~in _ n . + 1 , . . . , ~ ni are h e l d constant, one deduces that A l is
t a n g e n t to t h e ~ s l i c e of V w h e r e all the coordinate functions
~k k
n _ n k + 1 , . . . , ~ n , for all k ~ i, are held constant. This yields (3.11).

The n e c e s s i t y is a s t r a i g h t f o r w a r d consequence of the definition.

 24

4. I n v a r i a n t Distributions

The n o t i o n of d i s t r i b u t i o n invariant under a vector f i e l d plays,

in the t h e o r y of n o n l i n e a r control systems, a role similar to the one
played in the t h e o r y of l i n e a r systems by the n o t i o n of s u b s p a c e in-
variant under a linear mapping.
A distribution A on N is invariant u n d e r a vector field f if the
Lie b r a c k e t [ f,T] of f w i t h e v e r y v e c t o r field T 6 A is a v e c t o r field
which belongs to A, i.e. if

(4.1) [f,~] a

(4.2) Remark. T h e r e is a n a t u r a l w a y to see that the p r e v i o u s defini-

tion generalizes the n o t i o n of i n v a r i a n t subspace. Let N = ~n, A a
linear mapping A : ~n ~An and V a s u b s p a c e of ~ n invariant u n d e r A,
i.e. s u c h t h a t A V C V. Suppose V is s p a n n e d by the v e c t o r s

Vi = (Vil ..... Vin) I ~ i ~ d

and consider, as in the E x a m p l e (3.4), the flat d i s t r i b u t i o n AV s p a n n e d

by the v e c t o r fields

n ~ I < i < d
~i = [ vij 8x -- --
j=1 j

With the m a p p i n g A we a s s o c i a t e a vector field fA r e p r e s e n t e d , in the

canonical basis ( ~ 1 ) x ..... ( ~ n ) X of T ~ n by the v e c t o r

fA (x) = A x

(note that the r i g h t - h a n d - s i d e of this e x p r e s s i o n represent

of c o o r d i n a t e s of an e l e m e n t of the t a n g e n t space at x to ~ n a n d n o t
a vector of coordinates of a p o i n t in ~ n ) .
It is e a s i l y seen that the d i s t r i b u t i o n ~V is i n v a r i a n t under the
vector field fA in the s e n s e of our p r e v i o u s definition. For, observe
t h a t any v e c t o r field ~ in A V can be r e p r e s e n t e d in the f o r m (2.8) w h e r e
c l , . . . , c d is a n y s e t of r e a l - v a l u e d functions defined locally a r o u n d x.
Computing the Lie b r a c k e t of fA a n d T we h a v e

d d d
[ fA'T] = i=I~[ fA'CiTi] = i=I[Ci[ fA'Ti] + i=I~ (LfACi)Ti

Moreover, ~Ti ~fA (x)

[ fA,~i] (x) = - - ~ f A ( x ) ~ri(x) =-ATi(x)
 25

Note that Ti(x), regarded as a point of IRn, is an element of V, so also

Aii(x) E V. Then, for each x, [fA, li] (x) 6 Av(X) and

[ fA, TI (x) e Av(X )

that proves the assertion. []

The notion of invariance under a vector field is p a r t i c u l a r l y

useful when referred to c o m p l e t e l y integrable distributions, because
it provides a way of s i m p l i f y i n g the local r e p r e s e n t a t i o n of the given
vector field.

(4.3) Lemma. Let & be a n o n s i n g u l a r involutive distribution of dimen-

sion d and assume that A is invariant under the vector field f. Then,
at each point p 6 N there exists a coordinate chart (U,~) with coord-
inate functions (I .... '(n ' in which the vector field f is r e p r e s e n t e d
by a vector of the form

fI(~I ..... ~d,(d+1 ..... ~n )

fd(~1 ..... ~d,~d+1 ..... ~n )

(4.4) f(() =
fd+1(~d+1 , .... ~n )

fn(~d+1,...,~n )

Proof. The d i s t r i b u t i o n A, being n o n s i n g u l a r and involutive, is in-

tegrable and, therefore, at each point p E N there exists a coordinate
chart (U,~) that makes (3.1) satisfied for all q E U. Now, let
fl (~),...,fn(~) denote the c o o r d i n a t e s of f(q) in the canonical basis
of T N a s s o c i a t e d with (U,~), and recall that
q

n
f(q) = [ fi(~(q)) ( )
i=I

The invariance condition (4.1) implies, in p a r t i c u l a r , t h a t

[ f, ~ j ] (q)E span{ ( ~ 1 ) q ..... ( ~ d ) q }

for all q E U and j = 1,...,d. Therefore we m u s t have that

[ f, %] = I fi ~i'~j] =- ( ) e sp{~ ~ . ~ }
i=I i=I ~ j ~i 51 . . . . ~ d
 26

~f.
From this we see that the c o e f f i c i e n t s ~ are such that

~f.
- 0

for al i = d+l,...,n and j = 1,...,d and all ~ E ~(U). The components

fd+1,...,fn are thus independent of the c o o r d i n a t e s ~1,...,~d , and
the (4.4) are proved. []

The following properties of i n v a r i a n t distributions will be a l s o

used later on.

(4.5) Lemma. Let A be a d i s t r i b u t i o n invariant under the v e c t o r fields

f~ and f2" Then ~ is also invariant under the v e c t o r field [fl,f2] .

Proof. Suppose T is a v e c t o r field in A. Then, from the J a c o b i identity

we get

[[ fl,f2] ,T] = [ f1,[ f2,T] ]-[ f2,[ f1,T]]

By assumption [ f2,T] 6 A and so is [ f1,[f2,T]] . For the v e r y same

reasons [f2,[f1,T]] 6 ~ and thus from the above equality we conclude
that [I fl,f21 ,T] e 4. []

(4.6) Remark. Note that the n o t i o n of invariance under a given vector

field f is s t i l l meaningful in the case of a d i s t r i b u t i o n A which is
not smooth. In this case, it is simply required that the L i e b r a c k e t
[f,T] of f with every smooth vector field in ~ be a v e c t o r field in 4.
Since [ f,T] is a s m o o t h vector field, it f o l l o w s that if ~ is a
(possibly) non-smooth distribution invariant under the v e c t o r field f,
then also smt(A) is i n v a r i a n t under f. []

When dealing with codistributions, one can as w e l l introduce the

notion of i n v a r i a n c e under a vector field in the following way.
A codistribution ~ on M is i n v ~ r i a n t under a vector field f if
the L i e derivative along f of any c o v e c t o r field ~ @ ~ is a c o v e c t o r
field which belongs to ~, i.e. if

4.7) Lf~ C

It is e a s i l y seen that this is the d u a l version of the n o t i o n of

invariance of a d i s t r i b u t i o n .

(4.8) Lemma. If a s m o o t h distribution A is i n v a r i a n t under the v e c t o r

field f, then the codistribution ~ = 4 is i n v a r i a n t under f. If a
 27

smooth codistribution ~ is i n v a r i a n t under the v e c t o r field f, then

the d i s t r i b u t i o n A = ~I is i n v a r i a n t under f.

Proof. We s h a l l m a k e use of the i d e n t i t y

(Lf~,T)= Lf(~,T )- (~,[ f,T] )

Suppose A is i n v a r i a n t under f and let T be any vector field of A. T h e n

[ f,T] E ~. L e t ~ be a n y c o v e c t o r field in ~. Then, by d e f i n i t i o n

(~,T > (p) = 0

for all p 6 N, and also

(~,[ f,T]) (p) = 0

This y i e l d s

( Lfw,T > (p) = 0

Since A is a s m o o t h distribution, given any vector v in A(p) we

may find a v e c t o r field T in A w i t h the p r o p e r t y that T(p) = v and,
then, the p r e v i o u s result shows that

(Lfw(p),v > = 0

for all v 6 A(p), i.e. that Lf~(p) E ~(p). From this it is c o n c l u d e d

that Lf~ is a c o v e c t o r field in ~.
The second part of the statement is p r o v e d in the same way. []

(4.9) Remark. N o t e t h a t in the p r e v i o u s Lemma, first part, we don't

need to a s s u m e that the annihilator A i of A is smooth, nor, in the
second part, that the annihilator ~ of ~ is smooth. However, if b o t h
A and A ~ are smooth, we c o n c l u d e from the L e m m a that the invariance of
A under f implies and is i m p l i e d by the invariance of A l u n d e r the same
vector field. In v i e w of L e m m a (2.17) this is true, in p a r t i c u l a r , when-
ever A is n o n s i n g u l a r . []

By m a k i n g use of t h e s e notions one m a y give a dual formulation of

Lemma (4.3). Instead of a nonsingular and involutive distribution A,
we have to c o n s i d e r (see R e m a r k (3.7)) a nonsingular codistribution
of d i m e n s i o n n-d with the p r o p e r t y that for e a c h p E N there exist a
neighborhood U of p and n - d functions ~d+1,...,{n defined on U w i t h
values in ~ such that
 28

~(q) = s p a n { d ~ d + I (q) ..... d~n(q) }

for all q E U.
If ~ s a t i s f i e s these assumptions and if ~ is a l s o invariant under
f, then it is p o s s i b l e to find d more real-valued functions ~I .... '~d
defined on U w i t h the p r o p e r t y that, choosing as l o c a l coordinates on
U the functions ~i ' I ~ i ~ n, for e a c h q 6 U the v e c t o r field f is
represented by a vector f(~) of the form (4.4).

5. L o c a l Decompositions of C o n t r o l Systems

Throughout these notes we deal with nonlinear control systems

described by e q u a t i o n s of the form

m
(5.1a) x = f(x) + ~ gi(x)u i
i=I

(5.1b) Yi = hi(x) (i = I ..... Z)

The state x of this system belongs to an o p e n subset N of I~n,

while the m c o m p o n e n t s u 1 , . . . , u m of the input and, respectively, the
i components yl,...,y i of the o u t p u t are real-valued functions of time.
We shall make later on some further assumptions on the class of a d m i s -
sible input functions to be c o n s i d e r e d . The vector fields f'g1''" "'gm
are smooth vector fields defined on N a n d assumed to be c o m p l e t e . The
output maps h l , . . . , h i are real-valued smooth functions defined on N.

(5.2) Remark. One may define systems with the same structure as (5.1),
with the state evolving on some abstract manifold N (not n e c e s s a r i l y
diffeomorphic to an o p e n subset of l~n). In t h i s case, instead of (5.1),
which is an o r d i n a r y differential equation defined on an o p e n subset
of l~n,one should consider a description based upon an o r d i n a r y dif-
ferential equation defined on the abstract manifold N. The vector
fields f'q1'''''gm will be d e f i n e d on N and so the output functions
h l , . . . , h i. If we let p d e n o t e a point in N then, instead of (5.1), we
may use a description of the form

m
(5.3a) ~ = f(p) + [ gi(P)U i
i=I

(5.3b) Yi = hi(P) (i = I ..... )

with the u n d e r s t a n d i n g that p stands for the tangent vector at the

 29

point p to the smooth curve w h i c h c h a r a c t e r i z e s the solution of (5.3a)

for some fixed initial condition.
If this is the case, then (5.1) may be r e g a r d e d as a local re-
presentation of (5.3) in some c o o r d i n a t e c h a r t (U,~) w i t h the under-
standing that x = ~(p). []

The theory d e v e l o p e d so far enables us to obtain for this class

of systems d e c o m p o s i t i o n s similar to those d e s c r i b e d at the b e g i n n i n g
of the Chapter. The r e l e v a n t results may be f o r m a l i z e d in the fol-
lowing way.

(5.4) Proposition. Let A be a n o n s i n g u l a r i n v o l u t i v e d i s t r i b u t i o n of

dimension d and assume that A is i n v a r i a n t u n d e r the v e c t o r fields
f,gl,...,g m. Moreover, suppose that the d i s t r i b u t i o n s p ( g l , . . . , g m} is
contained in 4. Then, for each point x 6 N it is p o s s i b l e to find an
open subset U of x and a local c o o r d i n a t e s transformation ~ = ~(x)
defined on U, such that, in the new coordinates, the control system
(5.1a) is r e p r e s e n t e d by e q u a t i o n s of the form

m
(5.5a) ~I = fl (~I'~2) + [ gil (61'~2)ui
i=I

(5.5b) ~2 = f2(~2 )

where (~I,~2) is a p a r t i t i o n of ~ and dim(~ I) = d.

Proof. F r o m L e m m a (4.3) it is k n o w n that there exists, a r o u n d each

e N, a c o o r d i n a t e chart (U,~) with c o o r d i n a t e functions ~i,...,{ n
with the p r o p e r t y that the v e c t o r fields f , g l , . . . , g m are r e p r e s e n t e d
in form (4.4). Moreover, since by a s s u m p t i o n gi 6 A for all i=1,...,m,
then the v e c t o r fields gl,...,g m in the same c o o r d i n a t e chart are re-
presented by vectors w h o s e last (n-d)-components are vanishing. This
coordinate chart (U,~) may o b v i o u s l y be c o n s i d e r e d as a local change
of coordinates around x and therefore the Proposition is proved. []

(5.6) Proposition. Let A be a n o n s i n g u l a r involutive d i s t r i b u t i o n of

dimension d and assume that A is i n v a r i a n t under the v e c t o r fields
f,gl,...,g m. M o r e o v e r , assume that the c o d i s t r i b u t i o n s p { d h l , . . . , d h Z}
is contained in the c o d i s t r i b u t i o n 4 . Then, for each x @ N it is pos-
sible to find an open subset U of x and a local c o o r d i n a t e s trans-
formation ~ = ~(x) d e f i n e d on U, such that, in the new coordinates,
the control s y s t e m (5.1) is r e p r e s e n t e d by e q u a t i o n s of the form

m
(5.7a) %1 = fI(~I'~2) + ~ gi1(~1'~2)ui
i=I
 30

(5.7b) 6 2 = f2(~2 ) + [ gi2(~2)ui

i=I

(5.7c) Yi = hi (~2)

where (~1,~2) is a p a r t i t i o n of ~ a n d d i m ( ~ I) = d.

Proo]'. As b e f o r e , w e k n o w t h a t there e x i s t s , a r o u n d e a c h x E N, a
coordinate chart (U,~), w i t h coordinate functions ~I' "'''~n ' with
the p r o p e r t y that the v e c t o r fields f'g1' "'''gm are represented in the
form (4.4). M o r e o v e r , we h a v e assumed that

A C [ s p { d h I ..... dh}] !

For e a c h p o i n t x of the s e l e c t e d coordinate c h a r t we h a v e in p a r t -

icular, for j = 1,...,d,

(~j)x E A(X) C [i=I~ s p a n { d h i ( x ) } ] ~ = i = I n [ s p a n { d h i ( x ) } ] ~

As a c o n s e q u e n c e , for j = 1 , . . . , d a n d i = I .... , and for all

x EU

( dh i (x) , (~-~j) x > = 0

or, in o t h e r w o r d s , we see t h a t the l o c a l representation of h i in the

selected coordinate chart is such t h a t

~h
1
~j = 0

for all j = 1,...,d and i = I,..., a n d for all ~ E ~(U). We c o n c l u d e

that h i d e p e n d s only on the l o c a l c o o r d i n a t e s ~d+1'''''~n on U a n d
this c o m p l e t e s the proof. []

The two l o c a l decompositions thus o b t a i n e d are v e r y u s e f u l in

understanding the i n p u t - s t a t e and s t a t e - o u t p u t behavior of the c o n t r o l
system (5.1).
Suppose that the inputs u i are p i e c e w i s e constant functions of
time, i.e. that there exist real n u m b e r s T o = 0 < T I < T 2 < ... such
that

u i ( t ) = ~ki for T k _< t < T k + I

Then, on the time i n t e r v a l [Tk,Tk+1), the s t a t e of the s y s t e m e v o l v e s

 31

along the integral curve of the v e c t o r field

f + glu +... +gmUm

passing through the p o i n t X(Tk). In particular, if the initial state

o
x at time t ~ 0 is c o n t a i n e d in some n e i g h b o r h o o d U of N, then for
small t the state x(t) evolves in U.
Suppose now that the a s s u m p t i o n s of the P r o p o s i t i o n (5.4) are
satisfied and that x O b e l o n g s to the domain U of the c o o r d i n a t e trans-
formation ~(x). If the input u is such that the x(t) evolves in U, we
may use the e q u a t i o n s (5.5) to d e s c r i b e the b e h a v i o r of the system.
From these we see that the local c o o r d i n a t e s (~1(t),~2(t)) of x(t) are
such that ~2(t) is not a f f e c t e d by the input. In particular, let x(T)
denote the p o i n t of U r e a c h e d at time T w h e n u(t) = 0 for all t6[0,T],
i.e. the p o i n t

o
x (T) = ~ x )

%Tf being the flow of the v e c t o r field f, and let (~(T),~(T)) denote
the local c o o r d i n a t e s of x(T). We see that the set of points that
can be r e a c h e d at time T, s t a r t i n g from x , lies inside the set of
points w h o s e local c o o r d i n a t e s ~2 are equal (T) " This set is
to ~2
actually a 8lice of U p a s s i n g t h r o u g h the p o i n t x(T).
Thus, we see that locally the s y s t e m d i s p l a y s a b e h a v i o r s t r i c t l y
analogous to the one d e s c r i b e d in section I. Locally, the state space
may be p a r t i t i o n e d into s u b m a n i f o l d s (the slices of U), all of dimen-
sion d, and the p o i n t s reachable at time T, along t r a j e c t o r i e s that
stay in U for all t 6 [0,T], lie inside the slice p a s s i n g through the
point x(T) r e a c h e d under zero input.
The P r o p o s i t i o n (5.6) is useful in s t u d y i n g s t a t e - o u t p u t interac-
tions. S u p p o s e we take two initial states x a and x b b e l o n g i n g to U
with local c o o r d i n a t e s (~,~) and (~,~) such that

a b
~2 = ~2

i.e. two initial states b e l o n g i n g to the same slice of U. Let x~(t)

and x~(t) denote the values of the states r e a c h e d at time t, s t a r t i n g
from x a and x b under the action of the same input u. From the e q u a -
tion (5.7b) we see i m m e d i a t e l y that, if the input u is such that xa(t)
U
and x~(t) both evolve in U, the ~2 c o o r d i n a t e s of x~(t) and of x~(t)
are the same, no m a t t e r w h i c h input u we consider. Actually these
 32

a
coordinates ~2(t) and ~ (t) are solutions of the same differential
equation (the e q u a t i o n (5.7b)) with the same initial condition. If we
take into account also the (5.7c) we h a v e the e q u a l i t y

hi ltl) o h l (tl)
which holds for e v e r y input u. We m a y conclude that x a a n d x b are in-
distinguishable.
Again, we find that locally the state space may be p a r t i t i o n e d
into submanifolds (the slices of U), all of d i m e n s i o n d, a n d p a i r of
points of e a c h slice both produce the same output (i.e. are indistin-
guishable) under any input u which keeps the state trajectory evolving
on U.
In the n e x t sections we shall reach stronger conclusions, showing
that if we a d d to the h y p o t h e s e s contained in the P r o p o s i t i o n s (5.4)
and (5.6) the further assumption that the distribution ~ is "minimal"
(in the case of P r o p o s i t i o n (5.4)) or " m a x i m a l " (in the case of P r o -
position (5.6)), then from the decompositions (5.5) and (5.7) one m a y
obtain more informations about the set of states reachable f r o m x and,
respectively, indistinguishable f r o m x .
We conclude this section with a remark about a dual version of
Proposition (5.6).

(5.8) Remark. Suppose that ~ is a n o n s i n g u l a r codistribution o f di-

mension n-d with the p r o p e r t y that for e a c h x E M there exist a neigh-
borhood U of x and n - d real-valued functions ~d+1,...,~n defined on U
such that

(x) = s p a n { d ~ d + l (x) .... ,d~ n ( x ) }

for all x 6 U. L e t ~1,...,~d be other functions defining, together

with ~d+1' .... ~n ' a coordinate transformation on U. In t h e s e coord-
inates, the o n e - f o r m dh i w i l l be r e p r e s e n t e d by a r o w v e c t o r

dhi(~) = (ii ( ~ ) ' ' ' X i n (~))

whose components are related to the v a l u e of dh i at x b y the e x p r e s s i o n

dhi(x) = Xil (~(x)) ( d ~ 1 ) x + ... + X i n ( ~ ( x ) ) (d~n) x

If we assume that the c o v e c t o r fields dhl,...,dh Z belong to ~, then,

since ~ is s p a n n e d by d ~ d + 1 , . . . , d ~ n on U, we m u s t have
 33

7ij(~) = o

for all I < i < , I < j < d and all ~ in ~(U). But since

~h.
ij (6) =

one c o n c l u d e s that hl,...,h Z are independent of ~1,...,~d o n U, like

in (5.7c). []

6. L o c a l Reachabilit~

In the p r e v i o u s section we have seen that if there is a n o n -

singular distribution A of dimension d with the p r o p e r t i e s that:

(i) A is i n v o l u t i v e
(ii) A contains the d i s t r i b u t i o n s p { g I .... ,gm }
(iii) A is i n v a r i a n t under the v e c t o r fields f,gl,...,g m

then at e a c h point x E N it is p o s s i b l e to f i n d a coordinate trans-

formation defined on a neighborhood U of x and a partition of U into
slices o f d i m e n s i o n d, such that the p o i n t s reachable at s o m e time T,
starting from some initial state x O E U, a l o n g trajectories that stay
in U for all t E [0,T] , lie inside a slice o f U. N o w w e w a n t to i n -
vestigate the actual "thickness" of the s u b s e t of p o i n t s of a slice
r e a c h e d at time T.
The o b v i o u s suggestion that comes from the d e c o m p o s i t i o n (5.5)
is to l o o k at the "minimal" distribution, if any, that satisfies (ii),
(iii) and, then, to e x a m i n e what can be said about the p r o p e r t i e s of
points w h i c h belong to the same slice in the corresponding local de-
composition o f N. It turns out that this program c a n be c a r r i e d out in
a rather satisfactory way.
We need first some additional results on invariant distributions.
If ~ is a f a m i l y of distributions on N, w e d e f i n e the smallest or
minimal element as the m e m b e r of D (when i t e x i s t s ) which is c o n t a i n e d
in e v e r y o t h e r element o f 0.

(6.1) Lemma. Let A be a g i v e n smooth distribution and TI,...,T q a

given set of v e c t o r fields. The family o f all distributions which are
invariant under TI,...,~ q and contain A has a minimal element, which
is a s m o o t h distribution.

Proof. The family in q u e s t i o n is n o n e m p t y because the d i s t r i b u t i o n

sp{V(N)} clearly belongs to it. Let A I and A 2 be two elements of this
family, then i t is e a s i l y seen that their intersection A I N A 2 con-
 34

tains A and, b e i n g i n v a r i a n t under TI,...,T q , is an e l e m e n t of the

same family. This a r g u m e n t shows that the i n t e r s e c t i o n ~ of all elements
in the family c o n t a i n s A, is i n v a r i a n t under TI,...,Y q and is con-
tained in any o t h e r e l e m e n t of the family. Thus is its minimal element.
m u s t be smooth because o t h e r w i s e smt(~) w o u l d be a smooth d i s t r i b u -
tion c o n t a i n i n g A (because A is smooth by assumption), invariant
u n d e r T 1 .... ,Tq (see Remark (4.6)) and p o s s i b l y c o n t a i n e d in 2. []

In w h a t follows, the s m a l l e s t d i s t r i b u t i o n w h i c h c o n t a i n s A and

is i n v a r i a n t under the v e c t o r fields TI,...,T q will be d e n o t e d by the
symbol

(~I' .... ~q I~ )

While the e x i s t e n c e of a m i n i m a l e l e m e n t in the family of distri-

butions w h i c h s a t i s f y (ii) and (iii) is always guaranteed, the non-
s i n g u l a r i t y and the i n v o l u t i v i t y require some a d d i t i o n a l assumptions.
We deal w i t h the p r o b l e m in the f o l l o w i n g way. G i v e n a d i s t r i b u t i o n A
and a set TI,...,T q of vector fields we define the n o n d e c r e a s i n g se-
quence of d i s t r i b u t i o n s

(6.2a) A0 =
q
(6.2b) A k = Ak_ I + [ [Ti,Ak_ I]
i=I

There is a simple c o n s e q u e n c e of this d e f i n i t i o n

(6.3) Lemma. The d i s t r i b u t i o n s A0,AI,... g e n e r a t e d with the a l g o r i t h m

(6.2) are such that

A k C (T I ..... TqlA )

for a l l k. If there exists an integer k such that /Xk, = A k , + l , then

Ak~ = (T I ..... TqIA )

Proof. If A' is any d i s t r i b u t i o n w h i c h c o n t a i n s A and is i n v a r i a n t

under T i , then it is easy to see that A' D A k implies A' D A k + I. For,
we have
q q
Ak+ I = A k + ~ [Ti,A k] = A k + ~ sp{[ ~i,T] :~ 6 Ak}
i=I i=I
q
C Ak + [ s p { [ y i , T ] : y e A'} C A'
i=I
 35

Since A' D A 0 , by induction we see that A' D A k for all k.

If Ak+ = Ak++1 for some k we easily see that Ak# D A (by defini-
tion) and Ak, is i n v a r i a n t under TI,...,~ q (because [Ti,Ak+] C Ak~+I =
= Ak~ for all I ~ i ~ q). Thus Ak+ must coincide with ( T I , . . . , T q l A > . O

The p r o p e r t y Ak~ = Ak~+1 expresses a sort of finiteness quality

of the sequence ~0,AI,..., and such a property is clearly useful from
a computational point of view. The simplest practical situation in
which the chain of d i s t r i b u t i o n s (6.2) satisfies the assumption of
Lemma (6.3) arises when all the d i s t r i b u t i o n s of the chain are non-
singular. In this case, in fact, since by c o n s t r u c t i o n

dim A k ~ dim Ak+ I ~ n

it is easily seen that there exists an integer k < n such that

Ak~ = Ak~+1.
If the d i s t r i b u t i o n s A0,AI,... are singular, one has the fol-
lowing weaker result.
+
(6.4) Lemma. There exist an open and dense subset N of N with the
property that at each point p E N

(~I ..... Tq IA > (p) = An-I(P)

Proof. Suppose U is an open set with the property that, for some k ,
Ak+(p) = Ak++1(p) for all p 6 U. Then, it is possible to show that
(~I,...,TqlA) (p) = Ak+(p) for all p 6 U. For, we already know from
Lemma (6.3) that ( T I , . . . , T q l A ) D Ak+ . Suppose the inclusion is
proper at some p 6 U and define a new d i s t r i b u t i o n A by setting

~(p) = Ak~(p) if p e U

~(p) = (T I ..... TqlA ) (p) if p ~ U

This distribution contains A and is invariant under T 1,...,Tq.

For, if T is a vector field in ~, then [ Ti,~] e (~I ..... Tq IA )
(because ~ C (~I ..... Tq IA ) ) and, moreover, [ Ti,T ] (p) e Ak~(p ) for
all p 6 U (because, in a n e i g h b o r h o o d of p, T 6 Ak~ and [Ti,A k] C ~ ) .
Since ~ is p r o p e r l y contained in ( T I , . . . , T q l A ) , this would con-
tradict the m i n i m a l i t y of (T I,...,TqlA ) .
Now, let N k be the set of regular points of A k. This set is an
open and dense submanifold of N (see Lemma (2.9)) and so is the set
N = N 0 D N I P ... A N n _ I. In a n e i g h b o r h o o d of every point p E N the
 38

distributions A 0 , . . . , A n _ I are n o n s i n g u l a r . T h i s , t o g e t h e r w i t h the pre-

vious d i s c u s s i o n and a d i m e n s i o n a l i t y argument, shows that
An_ I = ( T I , . . . , T q l A } on N and c o m p l e t e s the proof. []
(6.5) R e m a ~ . If the d i s t r i b u t i o n A is s p a n n e d by some of the v e c t o r
fields of the set {TI,...,Tq}, then, it is p o s s i b l e to show that there
exists an o p e n and dense s u b m a n i f o l d N of N with the f o l l o w i n g pro-
perty. For each p E N there e x i s t a n e i g h b o r h o o d U of p and d vector
fields (with d = dim (TI,..., TqlA } (p)) 81,...,8 d of the form

@i = [ Vr' [ Vr-1 ..... [v1'v0] ] ]

where r ~ n-1 is an integer w h i c h may depend on i and v 0 , . . . , v r are

v e c t o r fields in the set {TI,...,Tq}, such that

<T I ..... TqlA } (q) = span{81(q) ..... 8d(q)}

for all q 6 U.
This fact may be p r o v e d by i n d u c t i o n using as N the subset of
N d e f i n e d in the p r o o f of L e m m a (6.4). Let d O denote the d i m e n s i o n of
A0 (which may depend on p but is c o n s t a n t locally around p). Since,
by assumption, A 0 is the span of some v e c t o r fields in the set
{TI,...,Tq}, there e x i s t e x a c t l y d O vector fields in this set that
span A 0 locally around p. Let d k denote the d i m e n s i o n of A k (constant
a r o u n d p) and suppose A k is spanned l o c a l l y a r o u n d p by d k vector
fields 8 1 , . . . , 8 d k of the form

8 i = [Vr,[Vr_ I ..... [vl,Vo]]]

where v 0 ..... v r (with r <_ k and p o s s i b l y d e p e n d i n g on i) are v e c t o r

fields in the set {T1,...,~q}. Then, a similar result holds for
Ak+ I. For, let T be any v e c t o r field in A k. From L e m m a (2.7) it is
known that there exists r e a l - v a l u e d smooth functions c l , . . . , C d k de-

fined locally a r o u n d p such that T may be expressed, locally a r o u n d

p, as T = c181 + ... + c a 8_ . If rj is any v e c t o r in the set {r I ..... Tq}
-k dk
we have

[Tj,c181 +.--+cdkedk] = C l [ ~ j , e ] + . . . + C d k [ ~ J , e d k ] + ( L ~ .~c 1 ) 8 1 +J' ' ' + ( L T 3 c ~ ) 8

As a c o n s e q u e n c e
 37

Ak+1 = Ak + [ TI'Ak] + "'" +[Tq'Ak] =

sP{@i,[Tl,@ i] ..... [Tq,0i] : i = I ..... d k}

since Ak+ I is nonsingular around p, then it is possible to find ex-

actly dk+ I vector fields of the form

e i = [Vr,[Vr_ I ..... [Vl,Vo]]]

where Vo,...,v r (with r ~ k+1 and possibly depending on i) are vector

fields in the set {TI,...,Tq}, which span ~k+1 locally around p. []

The previous remark is useful in getting involutivity for the

distribution (TI,...,TqlA) .

(6.6) Lemma. Suppose ~ is spanned by some of the vector fields

TI,...,T q and that (T I .... ,TqlA ) is nonsingular. Then ( ~ I , . . . , T q l A )
is involutive.

Proof. We use first the conclusion of Remark (6.5) to prove that if

~I and T 2 are two vector fields in An_ I , then their Lie bracket
[~i,T2] is such that [TI,T21 (p) 6 An_1(p) for all p E N . Using again
Lemma (2.7) and the previous result we deduce, in fact, that in a
neighborhood U of p

d I ! c1.%j] 6sP{Si,Sj,[Si,0j] :i,j =I d}

[TI'T2] = [ i = 1[ C 8 i , j 1 J .....

where ei,e j are vector fields of the form described before.

In order to prove the claim, we have only to show that [ 8i,8 j] (p)
is a tangent vector in An_ I (p). For this purpose, we recall that on
N the distribution An_ I is invariant under the vector fields
TI,...,T q (see Lemma (6.4)) and that any distribution invariant under
vector fields ~I and T 2 is also invariant under their Lie bracket
[TI,T 2] (see Lemma (4.5)). Since each 8 i is a repeated Lie bracket of
the vector fields TI,...,Tq, [Si,An_ I] (p) C An_ I (p) for all 1<i<d and,
thus, in particular [ 8i,8 j] (p) is a tangent vector which belongs to
An_ 1 (P).
Thus the Lie bracket of two vector fields T I ,T 2 in An_ 1 is such that
[TI,~2] (p) 6 An_ I (p). Moreover, it has already been observed that
(TI,...,~q[A) = An_ I in a neighborhood of p and, therefore, we con-
clude that at any point p of N the Lie bracket of any two vector
fields TI,T 2 in (T I ..... TqlA > is such that [TI,~ 2] (p) e ( T I ..... TqlA>(p).
Consider now the distribution
 38

= (T I ..... TqIA ) +sp{[8i,Sj] :8i,8 j 6 (rl, .... TqI~ )}

which, by construction, is such that

O (~I .... 'Tq IA )

From the previous result it is seen that A(p) = ( T I , . . . , T q I 6 ) (p)

at each point p of N , which is a dense set in N. By assumption,
(T1, .... TqlA ) is nonsingular. So, by Lemma (2.11) we deduce that
= ( T 1 , . . . , T q I 6 ) , and, therefore, that [8i,8 j] e ( ~ I , . . . , T q l A )
for all pair 8i,8 j 6 {T1,...,TqlA ) . This concludes the proof. []

(6.7) Remark. From Lemmas (6.4),(6.6) and (2.11) it may also be

deduced that if A is spanned by some of the vector fields TI,...,T q
and An_ I is nonsingular, then

(~1'''''Tq IA ) = An-1

and ( ~ 1 , . . . , ~ q l A ) is involutive.

We now come back to the original problem of the study the

smallest distribution which contains sp{gl,...,gm} and is invariant
under the vector fields f, gl,...,g m. From the previous Lemma it is
seen that if (f,gl,...,gml sp{gl,...,gm}) is nonsingular, then it is
also involutive and, therefore, the decomposition (5.5) may be per-
formed. We will see later that the minimality of (f'gl ..... gmlSp{gl ..... gm })
makes it possible to deduce an interesting topological property of
the set of points reached at some fixed time T starting from a given
point x . However, before doing this, it is convenient to analyze
some other characteristics of the decomposition (5.5).
Consider the distribution (f,gl,...,gmlsP{f,gl,...,gm}), i.e.
the smallest distribution invariant under f,gl,...,g m and which con-
tains sp{f,gl,...,g m} (note that now not only the vector fields
gl,...,g m but also the vector field f is assumed to belong to this
distribution).
If this distribution is nonsingular, and therefore involutive by
Lemma (6.6), it may indeed be used in defining a local decomposition
of the control system (5.1) similar to the decomposition (5.5). We
are going to see in which way this new decomposition is related to
the decomposition (5.5) and why it may be of interest.
In order to simplify the notation, we set

(6.8a) P = (f'gl ..... gmlsp{gl ..... gm })

 39

(6.8b) R = (f,gl,...,gmlsP{f,gl,...,gm})

The relation between p a n d R is d e s c r i b e d in the following state-

ment

(6.9) Lemma. The d i s t r i b u t i o n s P a n d R are such that

(a) P + sp{f} C R

(b) if x is a r e g u l a r point of P + sp{f}, then

(P + sp{f}) (x) = R(x)

Proof. B y d e f i n i t i o n , P C R and f E R, so (a) is true.

It is k n o w n from the p r o o f of L e m m a (6.6) that, around each point
x of an o p e n dense submanifold N of N, R is s p a n n e d by vector fields
of the f o r m

8 i = (v r ..... [Vl,V0]]

where r ~ n-1 is an integer which may depend on i, and V r , . . . , V l , V 0

are v e c t o r fields in the set { f , g l , . . . , g m }.
It is e a s y to see that all such vector fields belong to P + s p { f } .
For, if 8 i is j u s t one o f the v e c t o r fields in the set { f , g l , . . . , g m } it
either belongs to P (which contains g l , . . . , g m ) or to sp{f}. If 8 i has
the g e n e r a l form shown above we may, without loss of g e n e r a l i t y , assume
that v 0 is in the set { g l , . . . , g m }. For, if v 0 = f a n d v I = f, then
8 i = 0. O t h e r w i s e , if v 0 = f a n d v I = gj , then -8i = [ V r ' ' ' ' ' [ f ' g j ] ]
has the d e s i r e d form. Any vector of the form

8 i = Iv r ..... [vl,gj]]

with v r ..... v I in the set {f'gl .... ,gm } is in P b e c a u s e P contains gj

and is i n v a r i a n t under f,gl,...,g m and so the c l a i m is p r o v e d .
From this f a c t we deduce t h a t on an o p e n and dense submanifold
N of N,

R C p + sp{f}

and t h e r e f o r e , since R D p + sp{f} on N, that on N

R = P + sp{f}

Suppose that P + span f has constant dimension on some neighbor-

 40

hood u. Then, from Lemma (2.11) we conclude that the two d i s t r i b u t i o n s

R and P + sp{f} coincide on U. []

(6.10) Coro~Zary. If P and P + sp{f} are nonsingular, then

dim(R) - dim(P) < I. []

If P and P + sp{f} are both nonsingular, so is R and, by Lemma

(6.6), both P and R are involutive. Suppose that P is properly con-
tained in R. Then, using Theorem (3.10),one can find, locally around
each x 6 N, a n e i g h b o r h o o d U of x and a coordinate transformation
= ~(x) defined on U such that

(6.11a P(x = x ..... )x }

r-1

(6.11b) R(x) = s P { ( ~ ) x ' ' 1 "'' (9)~%r-I x ' (7~r)X ~

for all x 6 U, where r = dim(R).

In the ~ coordinates the control system (5.1a) is r e p r e s e n t e d by
equations of the form

%1 = fI(~I ..... ~n ) + ~ gi1(~l ..... ~n)Ui

i=1

~r-1 = fr-1 (61 ..... ~n )+I~I`=gi,r-1 (~I .... '~n)Ui

(6.12) ~r = fr(~r ..... ~n )

~r+1 = 0

n-0
The last components of the vector field f are v a n i s h i n g because,
by c o n s t r u c t i o n , f 6 R. In the p a r t i c u l a r case where R = P also
the r-th component of f vanishes and the c o r r e s p o n d i n g equation for ~r
is
~r = 0
From the equation (6.12) we see that any trajectory x(t) e v o l v i n g
on the n e i g h b o r h o o d U actually belongs to an r-dimensional slice of U
 41

passing through the initial point. This slice is in turn partitioned

into (r-1)-dimensional slices, each one including the set of p o i n t s
reached at a p r e s c r i b e d time T.

(6.13) Remark. A f u r t h e r change Of local coordinates makes it p o s -

sible to b e t t e r understand the role of the time in the b e h a v i o r of
the c o n t r o l system (6.12). We m a y assume, without loss of g e n e r a l i t y ,
that the i n i t i a l point x is such that ~(x ) = 0. T h e r e f o r e we h a v e
~i(t) = 0 for all i = r+1 .... ,n a n d

%r = f r ( ~ r '0 .... ,0)

Moreover, if we m a k e the assumption that f ~ P, then the function fr

is n o n z e r o everywhere on the n e i g h b o r h o o d U. ~ow, let ~r(t) denote
the s o l u t i o n of this differential equation which passes through 0 at
t = 0. C l e a r l y , the m a p p i n g

P : t ~--+ ~r(t)

is a d i f f e o m o r p h i s m f r o m an o p e n interval (-e,c) of the time axis

onto the o p e n i n t e r v a l of the ~r a x i s (~r(-e),~r(e)). If its inverse
-I
p is u s e d as a l o c a l coordinate transformation on the ~r axis one
easily sees that the n e w coordinate

~r = p-1 (~r) = t

satisfies the differential equation

~r = I

In t h e s e new coordinates, points on the r-dimensional slice of U

passing through the i n i t i a l state are parametrized by (~1,...,{r_1,t).
In p a r t i c u l a r , the p o i n t s reached at time T belong to t h e ( r - 1 ) - d i m e n -
sional slice

S = {x 6 U: ~r(X) = T , ~r+1 (x) = 0 ..... ~n(X) = 0}. []

(6.14) Remark. If f is a v e c t o r field of P then the local representa-

tion (6.12) is such that fr v a n i s h e s on U. T h e r e f o r e , starting from a
point x s u c h that ~(x ) = 0 we shall have ~i(t) = 0 for all i=r,...,n
and the s t a t e x(t) shall evolve on a (r-1)-dimensional slice of U p a s -
sing t h r o u g h x . []
 42

By d e f i n i t i o n the d i s t r i b u t i o n R is the s m a l l e s t d i s t r i b u t i o n
w h i c h contains f , g l , . . . , g m and is i n v a r i a n t u n d e r f , g l , . . . , g m. Thus,
we may say that in the a s s o c i a t e d decomposition (6.12) the d i m e n s i o n
r is "minimal", in the sense that it is not p o s s i b l e to find another
set of local c o o r d i n a t e s ~i,...,~
'''''~n ' w i t h ~ strictly less than
r
r, w i t h the p r o p e r t y that the last n-r c o o r d i n a t e s remain constant
with the time. We shall now show that, from the p o i n t of v i e w of the
i n t e r a c t i o n b e t w e e n input and state, the d e c o m p o s i t i o n (6.12) has even
s t r o n g e r properties. Actually, we are going to prove that the states
reachable from the initial state x fill up at least an open subset of
the r - d i m e n s i o n a l slice of in w h i c h they are contained.

(6.15) Theorem. Suppose the d i s t r i b u t i o n R (i.e. the s m a l l e s t distribu-

tion i n v a r i a n t u n d e r f , g l , . . . , g m w h i c h c o n t a i n s f'gl .... ,gm ) is non-
singular. Let r denote the d i m e n s i o n of R. Then, for e a c h xO 6 N it is
p o s s i b l e to find a n e i g h b o r h o o d U of x O and a c o o r d i n a t e t r a n s f o r m a -
tion 6 = {(x) d e f i n e d on U w i t h the f o l l o w i n g p r o p e r t i e s

(a) the set R(x ) of states r e a c h a b l e s t a r t i n g from x along trajec-

tories e n t i r e l y c o n t a i n e d in U and under the action of p i e c e w i s e
c o n s t a n t input f u n c t i o n s is a subset of the slice

S o = {x e U : ~ r + I (x) = ~ r + 1 (xO) ..... ~n (x) = ~n (xO) }

x

(b) the set R(x ) c o n t a i n s an open subset of S

o
x
Proof. The p r o o f of the s t a t e m e n t (a) follows from the p r e v i o u s discus-
sion. We p r o c e e d d i r e c t l y to the p r o o f of (b), a s s u m i n g t h r o u g h o u t the
proof to operate on the n e i g h b o r h o o d U on w h i c h the c o o r d i n a t e trans-
formation ~(x) is defined. For convenience, we b r e a k up the proof in
several steps.

(i) Let 81, .,8 k be a set of v e c t o r fields, w i t h k < r, and let

'%t denote the c o r r e s p o n d i n g flows. C o n s i d e r the m a p p i n g

F : (_,)k ~ N

,#k I (x o)
(tl,...,tk) ' tk ..... tI

where x is a p o i n t of N and suppose that its d i f f e r e n t i a l has

rank k at some Sl,...,s k , w i t h 0 ~ s i < e for I ~ i ~ k. For
e s u f f i c i e n t l y small the m a p p i n g
 43

(6.16) : (S1,S)x...(Sk,e) -~ N

(tl,...,tk) i , F(tl,...,tk)

is an e m b e d d i n g .
Let M d e n o t e the image of the m a p p i n g (6.16) (which depends on
the p o i n t x). Consider the slice of U

o = {x e U : ~i(x) = ~i(x), r + I ~ i ~ n}
S
x

If the v e c t o r fields 81,...,8 k have the form

m
J
8j = f + [ giui
i=1

with u z3. E ~ for I .< i .< m . and. I < j < k, then for e small M is an
embedded submanifold of S o" T h i s implies, in p a r t i c u l a r , t h a t for e a c h
x
x6M

(6.17) T x M C R(x)

where R, as b e f o r e , is the smallest distribution invariant under

f,gl,...,g m which contains f,gl,...,g m (recall that R(x) is the tangent
space to S o at x).
x
(ii) S u p p o s e that the v e c t o r fields f , g l , . . . , g m are such that

(6.18a) f(x) 6 TxM

(6.18b) gi(x) e Tx M I ~ i ~ m

for all x 6 M. W e shall show that this contradicts the assumption k < r.
For, consider the d i s t r i b u t i o n ~ defined by setting

&(x) = TxM for all x 6 M

&(x) = R(x) for all x 6 (N\M)

This d i s t r i b u t i o n is c o n t a i n e d in R (because of (6.17)) and contains

the v e c t o r fields f,gl,...,g m (because these vector fields are in R
and, moreover, it is a s s u m e d that (6.18) are true).
Let T be a n y v e c t o r field of ~. T h e n T 6 R and since R is i n v a -
riant under f,g1,-'-,gm , then for all x 6 (N\M)
 44

(6.19a) If,T] (x) 6 ~(x)

(6.19b) [gi,T] (x) 6 ~(x) I ! i ! m

Moreover since T , f , g 1 , . . . , g m are v e c t o r fields w h i c h are t a n g e n t

to M at e a c h x 6 M, we have also that (6.19) hold for all x 6 M, and
therefore for all x E N.
Having shown & is i n v a r i a n t under f , g l , . . . , g m and c o n t a i n s
f , g l , . . . , g m , we deduce that A m u s t coincide with R But this is a
contradiction since for all x ~ M

dim &(x) = k

dim R(x) = r > k

(iii) If (6.18) are not true, then it is p o s s i b l e to find m real

n u m b e r s u~ +1 ,...,u km + l and a p o i n t x E M such that the v e c t o r f i e l d

m k+l
ek+ I = f + ~ giui
i=I

satisfies the c o n d i t i o n
0 k + 1 ( x ) ~ T_ M.
x .k+1
Let x = F(s~ ..... s~) be this p o i n t and ~t denote the flow of
8k+ I. Then the m a p p i n g

k+1
F' : (-s,e) ~ N

(t I ,t k tk+1) I , ~k+1 o F ( t l , . . . , t k )
,... , tk+ I

at the p o i n t (s{ ..... s~,0) has rank k+1.

For, note that

( F ' ) * ( ~ t i) (s~,. .., s'k,0) = (F) * ( ~ t i) (s~ ..... s k)

for i = 1,...,k a n d that

(F'),(~--~I) (s~ ..... s~,0) = 9k+I (~)

The first k tangent vectors at x are l i n e a r l y independent, be-

cause F has rank k at all p o i n t s of (s1,s)x...(Sk,e). The (k+1)-th
one is i n d e p e n d e n t from the f i r s t k by c o n s t r u c t i o n and therefore F'
has rank k+1 at (Sl,...,Sk,0).
 45

Since s i' > s i , we m a y c o n c l u d e t h a t the m a p p i n g F' has r a n k k+1

at a p o i n t (s~ .... ,Sk+
' 1 ) , with 0 ~ s i' < e for I _< i _< k+1.
Note t h a t g i v e n any r e a l n u m b e r T > 0 it is a l w a y s possible to
choose the p o i n t x in such a w a y that

(s~-Sl)+...+(s~-s k) < T

For, o t h e r w i s e , we h a d that any v e c t o r f i e l d of the f o r m

m
8 = f +
i=Igiui

would be t a n g e n t to the image u n d e r F of the o p e n set

{(t I ..... tk) E ( S l , e ) . . . ( S k , e ) : ( t l - S l ) + . . . + ( t k - S k ) < T}

and this, as in (ii),would be a c o n t r a d i c t i o n .

(iv) We can n o w c o n s t r u c t a sequence of m a p p i n g s of the form (6.16).

m I
Let @ 1 = f + ~ l g i u i be a v e c t o r field which is n o t z e r o at x
i
(such a v e c t o r field can always be f o u n d b e c a u s e , otherwise, we would
have R(x O) = {0}) a n d let M 1 d e n o t e the image of the m a p p i n g

F~ : (0,E) ~ N

tl . , ~1 (x O)
t1

L e t x = F1(s~) be a p o i n t of M I in w h i c h a vector f i e l d of the

form 82 = f + m[ giui2 is s u c h t h a t 82(x) ~ T _ M I. T h e n we m a y d e f i n e

the m a p p i n g i=I x

F2 : (s11'e)x(0'c) ~ N

(tl,t 2 ) I , ~22oQI~tl (x )

Iterating this p r o c e d u r e , at s t a g e k we start with a mapping

Fk : (s k - l , c ) x . . . x ( s kIk-l,)x(0,c)
_ ~ N

(tl, ,tk_1,tk ) I , #k I (xO)

.. t k ..... ~tl
 46

a n d we f i n d a p o i n t x = F k ( S ~ ..... s~) o f its i m a g e ~ and a vector

m k+1
field 8k+ I = f + [ giui s u c h t h a t 8k+1(x) ~ T_~. This makes it
i=I x
possible to d e f i n ethe n e x t m a p p i n g Fk+1" Note t h a t s2 > sk-1
i for
i = I,...,k-I and sk > 0
k
The p r o c e d u r e clearly stops at the s t a g e r, w h e n a mapping Fr is
defined

Fr : (s~-1'e)x'''(srr--I '6)x(0'6) ~ N

r
(t1'''''tr-1'tr) ' ' ~t ..... ~ (x)
r 1

(v) Observe that a point x = Fr (tl .... 'tr) in the i m a g e M r of the

embedding Fr can be r e a c h e d , starting f r o m the s t a t e x a t time t=0,
under the a c t i o n of the p i e c e w i s e constant control defined by

ui(t) = u k for t 6 | t 1 + . . . + t k _ 1 , t 1 + t 2 + . . . + t k )

Thus, we k n o w f r o m our p r e v i o u s discussions t h a t M r m u s t be c o n t a i n e d

in the s l i c e of U

S o = {x 6 U: ~i(x) = ~i(x), r+1 ~ i ~ n}

x

The i m a g e s under Fr of the o p e n sets of

Ur (s[ -1 c)x r-l,e)x(0,e)

= , . . (Sr_ I

are open in the t o p o l o g y of M r as a s u b s e t of U ( b e c a u s e F r is an e m b e d -

ding) and therefore t h e y are a l s o o p e n in the t o p o l o g y of M r as a sub-
set of S (because S is an e m b e d d e d s u b m a n i f o l d of U). T h e r e f o r e we
o o
x x
h a v e that M r is an e m e b e d d e d s u b m a n i f o l d of S o a n d a d i m e n s i o n a l i t y
x
a r g u m e n t t e l l us t h a t M r is a c t u a l l y an o p e n s u b m a n i f o l d of S o " []
x
(6.20) Theorem. Suppose the d i s t r i b u t i o n s P (i.e. the smallest distri-
bution invariant under f,gl,...,g m which contains g1' .... g m ) a n d P + s p { f }
are n o n s i n g u l a r . Let p denote the d i m e n s i o n of P. Then, for e a c h x 6 M
it is p o s s i b l e to f i n d a n e i g h b o r h o o d U of x and a c o o r d i n a t e trans-
formation ~ = ~(x) defined on U w i t h the f o l l o w i n g properties:

(a) the set R(x,T) of s t a t e s reachable at time t = T starting from x

at t = 0, a l o n g trajectories entirely contained in U a n d u n d e r the
 47

action of p i e c e w i s e c o n s t a n t input functions, is a subset of the

slice

S ={x E U: (p+1=~p+1 ({~ (x)) ; (p+2 (x)=(p+2 (x) ,... , ~n (x) =~n (x) }
x,T

(b) the set R(x,T) contains an open subset of S

o
x ,T
Proof. We k n o w from Lemma (6.9) that R is nonsingular. T h e r e f o r e one
can repeat the c o n s t r u c t i o n u s e d to prove the part (b) of T h e o r e m
(6.15). Moreover, from C o r o l l a r y (6.10) it follows that r, the di-
mension of R, is equal e i t h e r to p+1 or to p.
Suppose the first s i t u a t i o n happens. Given any real n u m b e r
T ~ (0,), c o n s i d e r the set

UTr = {(tl'''''tr) e Ur : tl+ ...+t r = T}

where U r is as d e f i n e d at the step (v) in the proof of T h e o r e m (6.15).

From the last remark at the step (iii) we k n o w that there exists al-
r-1 r-1
ways a suitable choice of s I ,...,Sr_ I after w h i c h this set is not
empty.
Clearly the image Fr(U T)~ consists of points r e a c h a b l e at time T
and therefore is c o n t a i n e d in R(x,T). Moreover, using the same
arguments as in (v) , we deduce that the set Fr (UT) is an open subset
of S
,T
If p = r, i.e. if P = R, the proof can be c a r r i e d out by simply
adding an e x t r a state v a r i a b l e s a t i s f y i n g the e q u a t i o n

~n+l = 1

and showing that this reduces the p r o b l e m to the p r e v i o u s one. The

details are left to the reader. []

7. Local O b s e r v a b i l i t y

We have seen in section 5 that if there is a n o n s i n g u l a r distri-

bution A of d i m e n s i o n d with the p r o p e r t i e s that

(i) A is i n v o l u t i v e
(ii) A is c o n t a i n e d in the d i s t r i b u t i o n sp{dhl,...,dhz}l
(iii) A is i n v a r i a n t under the v e c t o r fields f,gl,...,g m

then, at each point x 6 N it is p o s s i b l e to find a c o o r d i n a t e trans-

 48

formation d e f i n e d in a n e i g h b o r h o o d U of x and a p a r t i t i o n of U into

slices of d i m e n s i o n d, such that p o i n t s on each slice produce the
same o u t p u t under any input u w h i c h keeps the state t r a j e c t o r y evolv-
ing on U. We want now to find c o n d i t i o n s under w h i c h points belong-
ing to d i f f e r e n t slices of U produce d i f f e r e n t outputs, i.e. are
distinguishable.
In this case we see from the d e c o m p o s i t i o n (5.7) that the right
o b j e c t to look for is now the "largest" d i s t r i b u t i o n w h i c h satisfies
(ii), (iii). Since the e x i s t e n c e of a n o n s i n g u l a r d i s t r i b u t i o n &
which satisfies (i), (ii), (iii) implies and is implied by the ex-
istence of a c o d i s t r i b u t i o n ~ (namely 4 ) w i t h the p r o p e r t i e s that

(i') ~ is spanned, locally a r o u n d each point p e N, by n-d exact

c o v e c t o r fields
(ii') ~ c o n t a i n s the c o d i s t r i b u t i o n sp{dh I .... ,dh Z}
(iii') ~ is i n v a r i a n t under the v e c t o r fields f , g l , . . . , g m

we may as well look for the "smallest" codistribution which satisfies

(ii'), (iii').
Like in the p r e v i o u s section, we n e e d some b a c k g r o u n d material.
However, m o s t of the results stated b e l o w require proofs w h i c h is are
similar to those of the c o r r e s p o n d i n g results stated before and, for
this reason, will be omitted.

(7.1) Lemma. Let ~ be a given smooth c o d i s t r i b u t i o n and TI,...,T q a

given set of vector fields. The family of all c o d i s t r i b u t i o n s w h i c h
are i n v a r i a n t u n d e r T I , . . . , T q and c o n t a i n ~ has a m i n i m a l element,
w h i c h is a smooth c o d i s t r i b u t i o n . []

We shall use the symbol < ~ 1 , . . . , ~ q l ~ ) to denote the s m a l l e s t

codistribution which contains ~ and is i n v a r i a n t under TI,...,Tq.
Given a c o d i s t r i b u t i o n ~ and a set of v e c t o r fields TI,...,T q
one can c o n s i d e r the f o l l o w i n g dual v e r s i o n of the a l g o r i t h m (6.2)

(7.2a) ~0 = ~
q
(7.2b) ~k = ~k-1 + [ L~i~k-1
i=I

and have the f o l l o w i n g result.

(7.3) Lemma. The c o d i s t r i b u t i o n s ~0,~i,... g e n e r a t e d w i t h the algo-

rithm (7.2) are such that

~k C (T I ..... Tq{fl )
 49

for all k. If there exists an integer k such that ~k~ = ~ k + + 1 , then

~k = (TI' .... Tq I~ ) []

The dual version of Lemma (6.4) is the following one

(7.4) Lemma. There exists

an open and dense subset N of N with the
%
property that at each point p 6 N

(T I .... ,~ql~ ) = ~n_1(p)

(7.5) Remark. If the c o d i s t r i b u t i o n ~ is spanned by a set dll,...,dl s

of exact covector fields, then there exists an open and dense sub-
manifold N of N with the following property. For each p 6 N there
exists a n e i g h b o r h o o d U of p and d exact covector fields (with
d = dim(T1,...,Tql~ ) (p)) ~ 1 , . . . , ~ d which have the form

~i = d(Lvr...LvllJ)

where r ~ n-1 is an integer w h i c h may depend on i, v l , . . . , v r are

vector fields in the set {TI,...,T q} and lj is a function in the set
{ll,...,Is }, such that

(T I ..... Tql~ ) (q) = sP{~1(q) ..... ~d(q)}

for all q E U.
This may easily be proved by induction as for the c o r r e s p o n d i n g
statement in Remark (6.5). []

(7.6) Lemma. Suppose ~ is spanned by a set dll,...,dk s of e x a c t co-

vector fields and that ( T I , . . . , T q I ~ ) is nonsingular.
Then ( T I , . . . , T q I ~ ) is involutive.

Proof. From the previous Remark, it is seen that in a n e i g h b o r h o o d of

each point p in an open and dense submanifold N , the c o d i s t r i b u t i o n
(T I, .... rqlA ) is spanned by exact covector fields.
Therefore, the Lie b r a c k e t of any two vector fields TI,T 2 in
(T 1 ..... Tql~ )I is such that [TI,T2] (p) 6 ( T I , . . . , T q I ~ ) ( p ) (see
Remark (3.9)).
From this result, using again Lemma (2.11) as in the proof of
Lemma (6.6), one deduces that { T I , . . . , T q I ~ )I is involutive. D

(7.7) Remark. From Lemmas (7.4),(7.6) and (2.11) one may also deduce
that if ~ is spanned by a set d l l , . . . , d l s of exact covector fields
 50

and ~n-1 is nonsingular, then

{ T I ' ' ' ' ' T q I > = ~n-1

and ( T 1 , . . . , T q l ~ ) I is involutive. []

In the study of the s t a t e - o u t p u t i n t e r a c t i o n s in a control system

of the form (5.1), we c o n s i d e r the d i s t r i b u t i o n

Q = (f'gl ..... gmlsp{dhl ..... dhz})

From Lemma (4.8) we deduce that this d i s t r i b u t i o n is i n v a r i a n t u n d e r

f , g l , . . . , g m and we also see that, by definition, it is c o n t a i n e d in
sp{dhl,...,dh}L. If nonsingular, then, a c c o r d i n g to Lemma (7.6) is
also involutive.
Invoking Proposition (5.6), this d i s t r i b u t i o n may be u s e d in order
to find locally a r o u n d e a c h x 6 N an open n e i g h b o r h o o d U of x and a
c o o r d i n a t e t r a n s f o r m a t i o n y i e l d i n g a d e c o m p o s i t i o n of the form (5.7).
Let s denote the d i m e n s i o n of Q. Since Q is the s m a l l e s t c o d i s t r i b u -
tion i n v a r i a n t under f , g l , . . . , g m w h i c h c o n t a i n s dhl, .... dh , then in
this case the d e c o m p o s i t i o n we find is maximal, in the sense that it
is not p o s s i b l e to find a n o t h e r set of local c o o r d i n a t e s
I ' ' ' ' ' ~ ' ~ + I ..... ~n w i t h ~ s t r i c t l y l a r g e r than s, with the pro-
%
p e r t y that only the last n-s coordinates influence the output. We show
now that this c o r r e s p o n d s to the fact that points b e l o n g i n g to dif-
ferent slices of the n e i g h b o r h o o d U are d i s t i n g u i s h a b l e .

(7.8) Theorem. Suppose the d i s t r i b u t i o n Q (i.e. the a n n i h i l a t o r of

the s m a l l e s t c o d i s t r i b u t i o n i n v a r i a n t u n d e r f , g l , . . . , g m and w h i c h con-
tains dhl,...,dh) is nonsingular. Let s denote the d i m e n s i o n of Q.
Then, for each x E N it is p o s s i b l e to find a n e i g h b o r h o o d U of x and
a coordinate transformation ~ = ~(x) d e f i n e d on U w i t h the f o l l o w i n g
properties

(a) Any two initial states x a and x b of U such that

~i(x a) = ~i(x b) , i = s + 1 .... ,n

produce identical output f u n c t i o n s under any input w h i c h k e e p s the

state t r a j e c t o r i e s e v o l v i n g on U

(b) A n y initial state x of U w h i c h c a n n o t be d i s t i n g u i s h e d from

under p i e c e w i s e c o n s t a n t input functions b e l o n g s to the slice
 51

S_ = {x 6 U:~i(x) = ~i(x), s+1 < i < n}.

X

Proof. We need only to prove (b). For simplicity, we break up the

proof in various steps.

(i) consider a piecewise-constant input function

ui(t) = uk for t 6 [t1+...+tk_1,t1+...+tk)

Define the vector field

m k
@k = f + [ giui
i=I
k
and let ~t denote the c o r r e s p o n d i n g flow. Then, the state reached at
time t k starting from x at time t = 0 under this input may be ex-
pressed as

x(tk ) = ~t
kk o...o~ItI (xo)

and the c o r r e s p o n d i n g output y as

Yi(tk) = hi(X(tk))

Note that this output may be regarded as the value of a m a p p i n g

o )k
FX : (-E,E ~
1

(t I ,..., tk ) ~__+h i o ~kt k O . . .o #ItI (xo)

If two initial states x a and b are such that they produce two
identical outputs for any possible p i e c e w i s e constant input, we must
have

a
Fx xb
i (t1'''''tk) = Fi (t1'''''tk)

for all possible (tl,...,tk), with 0 ~ t i < ~ for I < i < k. From this
we deduce that
a
~F x ~FX b
1
(~t 1...~t k ) t 1 = . . . = t k = O (~t1.~.~tk)t1=...=tk=O

An easy c a l c u l a t i o n shows that

 52

O
~F x

(~tl...~tk)t1=...=tk=Ol
= (Let "''L0khi(X))x

and, therefore, we must have

(L 6 I . "'L0khi (x)) x a = (L81 "''L6khi (X))xb

(ii) Now, remember that @j , j = I ..... k, depends on (u4 ..... u~) and
that the above equality must hold for all possible choices of
(u~ ..... u~) E ~ m By appropriately selecting these (u~ ..... u~) one
easily arrives at an equality of the form

(7.9)
(Lvl...Lvkhi)xa = (Lvl...Lvkhi)xb

where vl,...,v k are vector fields belonging to the set {f,gl,...,gm }.

For, set Y2 = L@2. ..Lskh. From the equality (L@IY2)xa = (L81Y2)xb
we obtain

m 1 m
(Lf~2) a + ~ (Lgi~2) aUi = (Lf~2)xb + ~ (LgiY2)xbU~
x i=I x i=I
I I
This, due to the arbitrariness of the ul,...,u m , implies that

(LvY2)Xa = (LvY2)Xb

where v is any vector in the set {f'gl' .... gm }. This procedure can be
iterated, by setting Y3 = L63"''L0kh" From the above equality one gets

m 2
(LvLfY3) a + m~ (LvLgi73) a u2i = (LvLfY3) b + ~ (LvLgz73) u
x i=I x x i=I . xb I

and, therefore,

(Lv ILv2Y 3 )x a = (LvILv2Y3) xb

for all Vl,V 2 belonging to the set {f,gl,...,gm }. Finally, one arrives
at (7.9).

(iii) Let U be a neighborhood of x on which a coordinate transformation

~(x) is defined which makes the condition

(7.10) Q(x) = span~( )x ..... ( )x

 53

satisfied for all x E U. F r o m Remark (7.5), we k n o w that there exists

an open subset U of U, d e n s e in U , with the p r o p e r t y that, around
each x' 6 U it is p o s s i b l e to f i n d a set of n - s real-valued functions
ll,...,In_ s w h i c h have the form

(7.11) Ii = Lv ...Lvlh j
r

with v I ..... v r v e c t o r fields in { f , g l , . . . , g m } and 1 ~ j ~ ~, s u c h that

Q (x') = s p a n { d l 1 ( x ' ) ,... ,dln_s (x')}

Since Qi(x') has dimension n-s, it f o l l o w s that the tangent covectors

dl1(x') ..... d l n _ s ( X ' ) are linearly independent.
In the l o c a l coordinates which satisfy (7.10), l l , . . . , l n _ s are
functions only of ~s+l,...,$n (see (5.7)). Therefore, we m a y deduce
that the m a p p i n g

A : (~s+1 ..... ~n )I ~ (11($s+I .... '~n ) ..... l n - s ( s + 1 ..... ~n ))

has a j a c o b i a n matrix which is s q u a r e and nonsingular at

(~s+1 (x') ..... ~n(X'))"

The m a p p i n g A is thus locally injective. We m a y use this pro-
perty to d e d u c e that, for some suitable neighborhood U' of x', any
other p o i n t x" of U' such that

li(x') = li(x" )

for I < i < n-s, must be such that

~s+i(X") = ~s+i (X !
)

for I < i < n-s, i.e. must belong to the slice of U p a s s i n g through
x'. This, in v i e w of the results proved in (ii) completes the p r o o f
in the c a s e where x 6 U .
#
(iv) S u p p o s e x ~ U . Let x(x,T,u) denote the state reached at time t=T
under the action of the p i e c e w i s e constant input function u. If T is
sufficiently small, x(x,T,u) is still in U. Suppose x(x,T,u) 6 U .Then,
using the c o n c l u s i o n s of (iii), we deduce that in some neighborhood U'
of x' = x ( x , T , u ) , the states indistinguishable f r o m x' lie on the
slice of U p a s s i n g through x'.
Now, recall that the m a p p i n g
 54

: xO ~ x(x,T,u)

is a local diffeomorphism. Thus, there exists a neighborhood U of

whose (diffeomorphic) image under # is a n e i g h b o r h o o d U" C U' of x'.
Let ~ denote a point of U i n d i s t i n g u i s h a b l e from x under piece-
wise constant inputs. Then, clearly, also x" = x ( ~ , T , u ) is i n d i s t i n -
guishable from x(x,T,u) = x'. From the p r e v i o u s discussion we k n o w that
x" a n d x' belong to the same slice of U. B u t this implies also that
and ~ b e l o n g to the same slice of U. T h u s the p r o o f is c o m p l e t e d , pro-
vided that

(7.12) x(x,T,u) U

(v) A l l w e have to s h o w n o w is t h a t (7.12) can be s a t i s f i e d . For,

suppose R(x), the set of s t a t e s reachable from x under piecewise con-
stant control along trajectories entirely contained in U, is such that

(7.13) R(x) N U* =

If this is true, we k n o w from Theorem (6.15) that it is p o s s i b l e to

find an r - d i m e n s i o n a l embedded submanifold V of U e n t i r e l y contained
in R(x) and therefore such that V ~ U ~ = @. For any choice of func-
tions l l , . . . , I n _ s of the form (7.11), at a n y p o i n t x 6 V the c o v e c t o r s
dl1(x),...,dln_s(X) are linearly dependent. Thus, without loss of ge-
nerality, we m a y assume that there exist d < n-s functions y1,...,yd
still of the form (7.11) such that, for some open subset V' of V,

- s p a n { d h 1 ( x ) ..... d h ( x ) } C s p a n { d ~ 1 ( x ) .... ,dYd(X)} for all x V'

- dY1(x) .... ,dYd(X) are linearly independent covectors at all x E V',
- dL v yj(x) span{dYl(X),...,dYd(X)} for all x V' a n d v 6 { f , g I .... ,gm }

Now, we define a codistribution on N as follows

~(x) = Qk(x) for x ~ V'

(x) = s p a n { d Y 1 (x) ..... dYd(X) } for x 6 V'

Using the fact that f'g1'''''gm are tangent to V', it is n o t d i f f i c u l t

to v e r i f y that this codistribution is i n v a r i a n t under f,gl,...,g m ,
contains sp{dhl, .... dh~} and is smaller than (f'gl ..... gmlsp{dhl .... 'dhz})'
This is a c o n t r a d i c t i o n and therefore (7.13) must be false. []
 CHAPTER II
GLOBAL DECOMPOSITIONS OF C O N T R O L SYSTEMS

I. Sussmann's Theorem and Global Decompositions

In the p r e v i o u s chapter, we have shown that a n o n s i n g u l a r and in-

volutive distribution induces a local partition of the m a n i f o l d N into
lower d i m e n s i o n a l submanifolds and we have used this result to o b t a i n
local d e c o m p o s t i o n s of control systems. The d e c o m p o s i t i o n s thus ob-
tained are very useful to u n d e r s t a n d the b e h a v i o r of c o n t r o l systems
from the p o i n t of v i e w of i n p u t - s t a t e and, respectively, state-output
interaction. However, it m u s t be s t r e s s e d that the e x i s t e n c e of de-
compositions of this type is s t r i c t l y related to the a s s u m p t i o n that
the dimension of the d i s t r i b u t i o n is c o n s t a n t at least over a neigh-
borhood of the p o i n t a r o u n d w h i c h we w a n t to i n v e s t i g a t e the beha-
viour of our c o n t r o l system.
In this section we shall see that the a s s u m p t i o n that A is non-
singular can be r e m o v e d and that g l o b a l partitions of N can be ob-
tained. To b e g i n with, we n e e d the following definitions. A submani-
fold S of N is said to be an integral submanifold of the d i s t r i b u t i o n
A if, for e v e r y p 6 S, the t a n g e n t space TpS to S at p c o i n c i d e s with
the subspace A(p) of T N. A maximal i n t e g r a l s u b m a n i f o l d of A is a
P
connected i n t e g r a l s u b m a n i f o l d S of A with the p r o p e r t y that e v e r y
other c o n n e c t e d integral submanifold of A w h i c h contains S coincides
with S.
We see i m m e d i a t e l y from this d e f i n i t i o n that any two m a x i m a l in-
tegral s u b m a n i f o l d s of A p a s s i n g through a point p 6 N m u s t coincide.
This m o t i v a t e s the following notion. A distribution A on N has the
maximal integral manifolds property if t h r o u g h e v e r y point p 6 N
passes a m a x i m a l integral submanifold of A or, in o t h e r words, if
there e x i s t s a partition of N into m a x i m a l i n t e g r a l s u b m a n i f o l d s of
A.
It is e a s i l y seen that this is a global version of the n o t i o n of
complete integrability for a d i s t r i b u t i o n . As a m a t t e r of fact, a
nonsingular and c o m p l e t e l y integrable distribution is such that for
each p E N there e x i s t s a neighborhood U of p w i t h the p r o p e r t y that
A restricted to U has the m a x i m a l integral manifolds property.
A simple consequence of the p r e v i o u s definitions is the fol-
lowing one.
 56

(1.1) Lemma. A d i s t r i b u t i o n ~ which has the m a x i m a l integral manifolds

property is involutive.

Proof. If T is a v e c t o r field which belongs to a d i s t r i b u t i o n A with

the m a x i m a l integral manifolds property, then T must be t a n g e n t to
every maximal integral submanifold S of 4. As a consequence, the Lie
b r a c k e t [ T i , T 2] of two v e c t o r fields r I and T2 both belonging to ~ m u s t
be tangent to e v e r y maximal integral submanifold S of 4. T h u s [ Ti,Y 2]
belongs to A. []

Thus, involutivity is a n e c e s s a r y condition for A to h a v e the

maximal integral manifolds property but, unlike the n o t i o n of c o m p l e t e
integrability, this condition is no l o n g e r sufficient.

(1.2) Example. L e t N = ~ 2 and let A be a distribution defined by

A(x) = span{ ( k ) x , l ( x I) (~x2)x }

where l(x I) is a C ~ f u n c t i o n such that l(x I) = 0 for x I < 0 and

l(x I) > 0 for x1> 0. This distribution is i n v o l u t i v e and

d i m ~(x) = I if x is such that xI ~ 0

dim A(x) = 2 if x is s u c h that x I > 0

Clearly, the o p e n subset of N

{ (x I,x2) 6 I~2 : x I > 0}

is an i n t e g r a l submanifold of A (actually a maximal integral sub-

manifold) and so is any subset of the form (a,b){c} with a < b < 0.
However, it is n o t possible to f i n d integral submanifolds of A p a s -
sing through a point (0,c). D

Another important point to be stressed, which emphasizes the

difference between the general problem here considered and its
local version described in s e c t i o n 1.3, is that the e l e m e n t s of a
global partition of N induced by a d i s t r i b u t i o n which has the in-
tegral manifolds property are immersed s u b m a n i f o l d s . On the con-
trary, local partitions induced by a nonsingular and c o m p l e t e l y
integrable distribution are always made of s l i c e s of a c o o r d i n a t e
neighborhood, i.e. of imbedded s u b m a n i f o l d s .

(1.3) Example. C o n s i d e r a torus T 2 = SIXS 2. We define a vector

field on the torus in the following way.
 57

Let T a v e c t o r f i e l d on ~ 2 defined by setting

T(x1'x2) = -x2( )x + x I ( ~ ) 2 x

At each p o i n t (Xl,X2) E $I t h i s m a p p i n g defines a tangent vector in

T(xl,x2)S I , a n d t h e r e f o r e a vector f i e l d on S I w h o s e f l o w is g i v e n

by
)T, 0 O, 0 0 0 0
ttXl,X2) = (x I c o s t - x 2 sint,x I sint-x2 cost)

In order to s i m p l i f y the n o t a t i o n we m a y r e p r e s e n t a point (Xl,X 2)

of S I w i t h the c o m p l e x number z = x I + jx 2 ,Izl = I, and have ~(z) =
= ejtz. Similarly, by s e t t i n g

8 (x 1,x 2) = -x2~(~-~i) x + X l e ( ~ 2 ) x

we define a n o t h e r v e c t o r f i e l d on S I , w h o s e f l o w is n o w g i v e n by
e
~t(z) = e3atz.
F r o m T a n d @ we m a y d e f i n e a vector field f on T 2 by s e t t i n g

f ( z l , z 2) = (T(Zl),@(z2))

and we r e a d i l y see t h a t the f l o w of f is g i v e n by

#f(z 1,z 2) = (eJtzl,eJ~tz2)

If e is a r a t i o n a l number, then there exists a T s u c h that

#tf = %t+kTf for all t 6 ~ a n d all k 6 Z. O t h e r w i s e , if ~ is i r r a t i o n a l ,
f
for each f i x e d p = (zl,z2) E T 2 the m a p p i n g Fp : t ~+ ~ t ( Z l , Z 2 ) is an
injective immersion of ~ i n t o T 2 , and F p ~ ) is an i m m e r s e d submani-
fold of T 2 .
F r o m the v e c t o r field f we can d e f i n e the o n e - d i m e n s i o n a l distri-
bution ~ = sp{f} and see that, if ~ is i r r a t i o n a l , the m a x i m a l in-
tegral s u b m a n i f o l d of A p a s s i n g through a p o i n t p 6 T 2 is e x a c t l y
Fp~) and A has the m a x i m a l integral manifold property.
Fp~R) is an i m m e r s e d b u t n o t an i m b e d d e d submanifold of T 2. For,
it is e a s i l y seen that given any p o i n t p C T 2 and any o p e n (in the to-
pology of T 2) n e i g h b o r h o o d U of p, the i n t e r s e c t i o n Fp~) D U is dense
in U and this e x c l u d e s the p o s s i b i l i t y of f i n d i n g a coordinate cube
(U,~) a r o u n d p w i t h the p r o p e r t y that F p ~ ) N U is a slice of U. []

The f o l l o w i n g theorem establishes the d e s i r e d n e c e s s a r y a n d suf-

 58

ficient condition.

(1.4) Theorem (Sussmann). A d i s t r i b u t i o n A has the maximal integral

m a n i f o l d s p r o p e r t y if and only if, for e v e r y vector f i e l d T E A and
T
for e v e r y pair (t,p) E ~ N such that the flow @t (p) of T is de-

fined, the d i f f e r e n t i a l (@~)~ at p maps the subspace A(p) into the

subspace A(~ (p)). []

We are not going to give the p r o o f of this Theorem, that can be

found in the literature. Nevertheless, some remarks are in order.

(1.5) Remark. An i n t u i t i v e u n d e r s t a n d i n g of the c o n s t r u c t i o n s that

are b e h i n d the s t a t e m e n t of S u s s m a n n ' s t h e o r e m m a y be o b t a i n e d in
this way.
Let TI,...,T k be a c o l l e c t i o n of v e c t o r fields of A and let
%1 Tk
@ t l , . . . , t k denote the c o r r e s p o n d i n g flows. It is clear that if p is

a p o i n t of N, and S is an i n t e g r a l m a n i f o l d of A p a s s i n g through p,
T.
then @ ti(p)
i should be a p o i n t of S for all v a l u e s of t i for w h i c h
T.
~tl(p)_ is defined. Thus, S s h o u l d include all points of N that can
l
be e x p r e s s e d in the form

rk T k - I m1
(I .6) o. (p)
@tk~_ 1 "'@t I

In particular, if m and 8 are v e c t o r fields of A, the smooth

curve

O : (-e,S) 'N

tl + ~T ~0 .T
~ t 1 ~ t ~ - t l (P)

p a s s i n g t h r o u g h p at t = 0, s h o u l d be c o n t a i n e d in S and its tangent

v e c t o r at p should be c o n t a i n e d in A(p). Computing this tangent
v e c t o r we o b t a i n

(@Ttl)~0(@~tl (p)) 6 A(p)

i.e. s e t t i n g q = @T (p)
-t I

@m
( tl).0(q ) e A ( ~ I (q))

and this m o t i v a t e s the n e c e s s i t y of S u s s m a n n ' s condition. O

 59

According to the s t a t e m e n t of T h e o r e m (1.4), in o r d e r to "test"

whether or n o t a given distribution A is integrable, one should check
that ( ~ ) ~ maps A(p) into A ( ~ ( p ) ) for all v e c t o r fields T in A.
Actually one c o u l d limit oneself to make this test o n l y on some sui-
table subset of v e c t o r fields in A, b e c a u s e the s t a t e m e n t of T h e o r e m
(1.4) can be given the following weaker version, also due to Sussmann.

(1.7) Theorem. A d i s t r i b u t i o n A has the m a x i m a l i n t e g r a l m a n i f o l d s

property if and only if there e x i s t s a set of v e c t o r fields T, w h i c h
spans A, w i t h the p r o p e r t y that for e v e r y T 6 I and e v e r y pair
(t,p) 6 ~ x N such that the flow ~(p) of Y is defined, the differ-
T (p))
ential (~), at p maps the s u b s p a c e A(p) into the s u b s p a c e A(~ t

(1.8) Remark. A t this p o i n t it is clear the p r o o f of the "if" part

of Theorem (1.4) comes directly from the "if" part of T h e o r e m (1.7),
because the set of all v e c t o r fields in A is i n d e e d a set of v e c t o r
fields w h i c h spans A. Conversely, the "only if" part Theorem (1.7)
comes from the "only if" p a r t of T h e o r e m (1.4). []

We have seen that i n v o l u t i v i t y is a n e c e s s a r y but not s u f f i c i e n t

condition for a d i s t r i b u t i o n A to have the m a x i m a l integral manifolds
property. However, the i n v o l u t i v i t y is s o m e t h i n g easier to test - in
principle - because it involves only the c o m p u t a t i o n of the Lie
bracket of v e c t o r fields in A w h e r e a s the test of the c o n d i t i o n stated
in the T h e o r e m (1.7) requires the k n o w l e d g e of the flows CtT a s s o c i a t e d
with all the v e c t o r fields T of the s u b s e t [ which spans A. Therefore,
one m i g h t w i s h to i d e n t i f y some special classes of d i s t r i b u t i o n s for
which the i n v o l u t i v i t y becomes a sufficient condition for them to have
the maximal integral manifolds property. Actually, this is p o s s i b l e
with a r e l a t i v e l y little effort.
A set [ of v e c t o r fields is locally finitely generated if, for
every p 6 N there exist a neighborhood U of p and a finite set
[Y1,...,Tk} of v e c t o r fields of [ w i t h the p r o p e r t y that e v e r y other
vector field b e l o n g i n g to T can be r e p r e s e n t e d on U in the form

k
(1.9) T = [ ciT i
i=I

where each c i is a r e a l - v a l u e d smooth function d e f i n e d on U.

The class of the d i s t r i b u t i o n s W h i c h are s p a n n e d by l o c a l l y fi-
nitely g e n e r a t e d sets of v e c t o r fields is a c t u a l l y one of the clas-
ses we were looking for, as it will be shown hereafter.
We prove first a slightly different result, w h i c h will be also
used independently.
 60

(1.10) L e m m a . Let I be a locally finitely g e n e r a t e d set of v e c t o r

fields w h i c h spans ~ and @ a n o t h e r v e c t o r field such that [8,T] E T
for all Y E T. Then, for every pair (t,p) 6 ~ N such that the flow
~t(p) is defined, the d i f f e r e n t i a l (~), at p m a p s the subspace &(p)
into the s u b s p a c e &(#~(p)).

Proof. The reader will have no d i f f i c u l t y in finding that the same

a r g u m e n t s u s e d for the s t a t e m e n t (ii) in the p r o o f of T h e o r e m I. (3.1)
can be used. []

N o t e that in the above s t a t e m e n t the v e c t o r field 0 may p o s s i b l y

not belong to T. If the set T is involutive, i.e. if the Lie b r a c k e t
[TI,T 2] of any two v e c t o r fields T 1 6 T, T2 6 T is again a vector
field in T, from the p r e v i o u s L e m m a and from S u s s m a n n ' s T h e o r e m we
derive i m m e d i a t e l y the f o l l o w i n g result.

(1.11) Theorem. A d i s t r i b u t i o n A spanned by an involutive and locally

f i n i t e l y g e n e r a t e d set of vector fields T has the m a x i m a l integral
m a n i f o l d s property. []

The e x i s t e n c e of an i n v o l u t i v e and locally f i n i t e l y g e n e r a t e d set

of v e c t o r fields appears to be s o m e t h i n g easier to prove, at least in
principle. In particular, there are some classes of d i s t r i b u t i o n s in
w h i c h the e x i s t e n c e on a l o c a l l y f i n i t e l y g e n e r a t e d set of vector
fields is a u t o m a t i c a l l y guaranteed. This yields the f o l l o w i n g corol-
laries of T h e o r e m (1.11).

(1.12) Corollary. A n o n s i n g u l a r d i s t r i b u t i o n has the m a x i m a l integral

m a n i f o l d s p r o p e r t y if and only if it is involutive.

Proof. In this case, the set of all vector fields w h i c h belong to the
d i s t r i b u t i o n is i n v o l u t i v e and, as a c o n s e q u e n c e of Lemma I. (2.7),
l o c a l l y finitely generated. []

(1.13) Corollary. An a n a l y t i c d i s t r i b u t i o n on a real a n a l y t i c m a n i f o l d

has the m a x i m a l integral m a n i f o l d s p r o p e r t y if and only if it is in-
volutive.

Proof. It depends on the fact that any set of a n a l y t i c v e c t o r fields

d e f i n e d on a real analytic m a n i f o l d is locally f i n i t e l y generated. []

We c o n c l u d e this section with another i n t e r e s t i n g c o n s e q u e n c e of

the p r e v i o u s results, w h i c h w i l l be u s e d later on.

(1.14) Lemma. Let & be a d i s t r i b u t i o n with the m a x i m a l integral mani-

folds p r o p e r t y and let S be a m a x i m a l integral s u b m a n i f o l d of 4. Then,
given any two points p and q in S, there e x i s t v e c t o r fields TI,...,Tk
TI T
in & and real n u m b e r s tl,. ,t k such that q = ~tl ..... ~ k(p).uk
 61

(1.15) Thcorem. Let A be an involutive d i s t r i b u t i o n i n v a r i a n t under a

complete v e c t o r field 8. Suppose the set of all v e c t o r fields in A is
locally f i n i t e l y generated. Let Pl and P2 be two points b e l o n g i n g to
8
the same m a x i m a l integral s u b m a n i f o l d of A. Then, for all T, }T(Pl )
8
and CT(P2 ) b e l o n g to the same m a x i m a l integral s u b m a n i f o l d of 4.

Proof. Observe, first of all, that A has the m a x i m a l i n t e g r a l sub-

manifold p r o p e r t y (see T h e o r e m (1.11)).
Let T be a v e c t o r field in A. Then, for e s u f f i c i e n t y small the
mapping

o : (-s,c) ~N

8 T
t P , ~TO~tO#~T(P)

defines a smooth curve on N w h i c h passes through p at t = 0. C o m p u t i n g

the tangent v e c t o r to this curve at t we get

T @
~* (~t)t = (~),T(#tO~_T(p)) =

= ( ~ ) , T (~@T(O (t)) )

But since T E A, we k n o w from L e m m a (1.10) that for all q

(~),T(#OT(q))~
_ C A(q) and therefore we get

~,(~t)t e A(o(t))

for all t E (-s,e). This shows that the smooth curve a lies on an in-
T
tegral s u b m a n i f o l d of A. Now,
let Pl = ~ T (p) and P2 = ~t(Pl )" Then
P2 and Pl are two points b e l o n g i n g to a m a x i m a l i n t e g r a l s u b m a n i f o l d
8 8
of 4, and the p r e v i o u s result shows that ~T(Pl ) and ~T(P2 ) again are
two points b e l o n g i n g to a m a x i m a l integral s u b m a n i f o l d of 4. Thus the
Theorem is p r o v e d for p o i n t s p1,p 2 such that P2 = ~Tt(Pl). If this is
not the case, u s i n g Lemma
(1.14) we can always find v e c t o r fields
T1 Tk
[1,...,Tk of A such that P2 = #t1'''#tk(Pl) and use the above re-
sult in o r d e r to prove the Theorem. []

2. The Control Lie A l g e b r a

The notions d e v e l o p e d in the p r e v i o u s section are useful in deal-

ing with the study of i n p u t - s t a t e i n t e r a c t i o n p r o p e r t i e s from a global
 62

point of view. As in c h a p t e r I, we c o n s i d e r here control systems de-

scribed by e q u a t i o n s of the form

m
(2.1) ~ = f(x) +
[ gi (x)u i
i=I

Recall that the local analysis of these properties was b a s e d upon

the consideration o f the smallest distribution, denoted R, invariant
under the v e c t o r fields f , g l , . . . , g m and w h i c h contains f'gl .... 'gm"
It w a s also shown that if this distribution is n o n s i n g u l a r , then it
is i n v o l u t i v e (Lemma I. (6.6)). This property makes it p o s s i b l e to use
immediately one of the results discussed in the p r e v i o u s section and
find a global decomposition of the state space N.

(2.2) Lemma. Suppose R is n o n s i n g u l a r , then R has the m a x i m a l integral

manifolds property.

Proof. J u s t use C o r o l l a r y (1.12). []

The decomposition of N i n t o m a x i m a l integral submanifolds of R

has the following interpretation from the p o i n t of v i e w of the study
of i n t e r a c t i o n between inputs and states. It is k n o w n that each of the
vector fields f,gl,...,g m is in R, a n d therefore tangent to e a c h m a x -
imal integral submanifold of R. L e t
S o be the m a x i m a l i n t e g r a l sub-
x
m a n i f o l d of R p a s s i n g t h r o u g h x O. F r o m w h a t we h a v e said b e f o r e we
m
k n o w t h a t a n y v e c t o r f i e l d of the f o r m T = ~ g i u i , w h e r e u l , . . . , u m
i=I
are real n u m b e r s , w i l l be t a n g e n t to S o and, t h e r e f o r e , t h a t the in-
x
t e g r a l c u r v e of T p a s s i n g t h r o u g h x at time t = 0 w i l l b e l o n g to S
o
x
o
W e c o n c l u d e t h a t any s t a t e t r a j e c t o r y e m a n a t i n g f r o m the p o i n t x ,
under the a c t i o n of a p i e c e w i s e constant control,will stay in S
o
x
Putting together this observation with the p a r t (b) of the sta-
tement of T h e o r e m .(6.15), one o b t a i n s the following result.

(2.3) Theorem. Suppose R in n o n s i n g u l a r . Then there exists a partition

of N i n t o m a x i m a l integral submanifolds of R, all w i t h the same di-
mension. Let S o d e n o t e the m a x i m a l integral submanifold of R p a s s i n g
x
through x . The set R ( x O) of s t a t e s reachable from x under piecewise
constant input functions

(a) is a s u b s e t of S
o
x
(b) c o n t a i n s an o p e n set of S o" []
X

The result might be interpreted as a g l o b a l version of T h e o r e m

I. (6.15). However, there are m o r e general versions, which do n o t re-
 83

quire the a s s u m p t i o n t h a t R is n o n s i n g u l a r . Of course, since one is

interested in h a v i n g global d e c o m p o s i t i o n s , it is n e c e s s a r y to w o r k
with d i s t r i b u t i o n s having the m a x i m a l integral manifolds property.
From the d i s c u s s i o n s of the p r e v i o u s section, we see t h a t a r e a -
sonable s i t u a t i o n is the one in w h i c h the d i s t r i b u t i o n s are spanned
by a set of v e c t o r fields which is i n v o l u t i v e and l o c a l l y finitely
generated. This motivates the i n t e r e s t in the f o l l o w i n g considera-
tions.
Let {Ti: I <i <q} be a f i n i t e s e t of v e c t o r fields and it,12 two
subalgebras o f V(N) which both contain the v e c t o r fields T 1,...,~q.
Clearly, the i n t e r s e c t i o n i I N i 2 is a g a i n a subalgebra o f V(N) and
contains T 1 , . . . , T q . T h u s we c o n c l u d e that there exists a unique sub-
algebra i of V(N) which contains TI, .... Tq a n d has the p r o p e r t y of
being c o n t a i n e d in all the s u b a l g e b r a s of V(N) which contain the
vector f i e l d s T I , . . . , ~ q. We r e f e r to t h i s as to the 8 m a Z l e s ~ sub-
algebra of V(N) w h i c h contains the v e c t o r fields T I , . . . , T q.

(2.4) Remark. One m a y g i v e a d e s c r i p t i o n of the s u b a l g e b r a i also in

the f o l l o w i n g terms. Consider the s e t

i o = {T 6 V(N) :T =[ T i k ' [ T i k - 1 ' .... [ Yi2'Til ]]] ; I ~ i k ~ q, I < k < ~}

and let LC(io) denote the set of all finite ~-linear combinations of
elements of i o. Then, it is p o s s i b l e to see t h a t i = LC(Lo). For, by
construction, every element of i o is an e l e m e n t of i b e c a u s e i,being
a subalgebra of V(N) which contains TI,...,T q , must contain every
vector f i e l d of the f o r m [ T ,k. [ T k_1' ,...,[ ~ i 2 , T i l ] ] ] . Therefore

LC(Lo) C i a n d a l s o Ti LC(i o) for I _< i _< q. To p r o v e that L=LC(Lo)

we only n e e d to s h o w t h a t LC(Lo) is a s u b a l g e b r a of V(N). This fol-
lows from the f a c t that the Lie b r a c k e t of a n y two v e c t o r fields in
iO is an ~ - l i n e a r combination of e l e m e n t s of L O. []

With the s u b a l g e b r a [ we m a y associate a distribution A i in a

natural way, by setting

&L = sp{T : ~ E L}

Clearly, A i n e e d n o t to be n o n s i n g u l a r . Thus, in o r d e r to be able to

operate w i t h ~i ' we h a v e to set e x p l i c i t l y some suitable assumptions.
In view of the r e s u l t s discussed at the e n d of the p r e v i o u s s e c t i o n we
shall a s s u m e t h a t the s u b a l g e b r a L is s p a n n e d by a l o c a l l y finitely
generated set of v e c t o r fields.
 64

An immediate consequence of this assumption is the following one.

(2.5) Lemma. If the subalgebra t is l o c a l l y finitely generated, the

distribution 4[ has the m a x i m a l integral manifolds property.

Proof. The set i is i n v o l u t i v e by construction (because is a sub-

algebra of V ( N ) ) . Then, using Theorem (1.11) we see that A i has the
maximal integral manifolds proDerty. []

When dealing with control systems of the form (2.1), we take

into consideration the smallest subalgebra of V(N) which contains the
vector fields f , g l , . . . , g m. This subalgebra will be d e n o t e d by C and
called the Control Lie Algebra. With C we associate the d i s t r i b u t i o n

A c = sp{T : T e C}

(2.6) Remark. It is n o t difficult to p r o v e that the c o d i s t r i b u t i o n AC

is i n v a r i a n t under the v e c t o r fields f , g l , . . . , g m. For, let T be any
i
vector field in C a n d ~ a c o v e c t o r field in A C. T h e n (~,T > = 0 and
<~,[f,T] > = 0 because [f,Tl is a g a i n a vector field in C. T h e r e f o r e ,
from the e q u a l i t y

(Lf~,T) = Lf(~,T ) -<~,[f,T] > = 0

we d e d u c e that Lf~ annihilates all v e c t o r fields in C. S i n c e A c is

spanned by v e c t o r fields in C, it f o l l o w s t h a t Lf~ is a c o v e c t o r
field in ~C ' i.e. that ~ is i n v a r i a n t under f. In the same w a y it
is p r o v e d that A C is i n v a r i a n t under gl .... 'gm"
If the c o d i s t r i b u t i o n ~ is smooth (e.g. when the distribution
A c is n o n s i n g u l a r ) , then using Lemma I.(4.8) one concludes that AC
itself is i n v a r i a n t under f, g l , . . . , g m.

(2.7) Remark. The distribution AC , and the distributions P a n d R in-

troduced in the p r e v i o u s chapter are related in the following way

(a) A C C p + sp{f} C R
(b) if x is a r e g u l a r point of A C , t h e n Ac(X)=(P+sp{f}) (x) = R(x).

We leave to the reader the p r o o f of this statement. []

The role of the C o n t r o l Lie Algebra C in the study of i n t e r a c -

tions between input and state depends on the following consideration.
Suppose A C has the m a x i m a l integral manifolds property and let S o be
x
the m a x i m a l integral submanifold of A c p a s s i n g through x . Since the
vector fields f,gl,...,g m , as w e l l as any v e c t o r field T of the form
 65

m
= f + ~ qiui with u I ..... u m real numbers, are in A C (and therefore
i=I
tangent to
S o ), then any state trajectory of the control system (2.1)
x
passing t h r o u g h x at t = 0, due to the a c t i o n of a p i e c e w i s e c o n s t a n t
control, will stay in S
o
x
As a c o n s e q u e n c e of this we see that, when studying the b e h a v i o r
of a c o n t r o l system intialized at x 6 N, we m a y regard as a n a t u r a l
state space the submanifold S o of N i n s t e a d of the w h o l e N. S i n c e for
x
all x e S o the t a n g e n t v e c t o r s f ( x ) , ~ ( x ) ..... gm(X) are e l e m e n t s o f
the
x
tangent s p a c e to S at x, by t a k i n g the restrictions to S of the
o o
x x
original v e c t o r f i e l d s f , g l , . . . , gm~'~ one m a y d e f i n e a s'et of v e c t o r f i e l d s
f,gl,...,g m on S o a n d a control system evolving on S
o
x x
1 m
(2.8) x = f(x) + [ gi(x)uq
i=I

which b e h a v e s exactly as the o r i g i n a l one.

By c o n s t r u c t i o n , the smallest subalgebra C of V(S ) which con-
x
tains f , g l , . . . , g m spans, at e a c h x E S o ' the w h o l e tangent space
x
T~S o" This m a y e a s i l y be seen u s i n g for C a n d C the description il-
x
lustrated in the R e m a r k (2.4).
Therefore, one m a y conclude that for the c o n t r o l system (2.8)
(which e v o l v e s on S ), the d i m e n s i o n of A~ is e q u a l to that of S o at
x x
each p o i n t or, also, that the s m a l l e s t d i s t r i b u t i o n R i n v a r i a n t u n d e r
f'gl ..... gm w h i c h contains f'gl ..... gm is n o n s i n g u l a r (see R e m a r k
(2.7)), w i t h a dimension equal to t h a t of S
o"
x
The c o n t r o l system (2.8) is such that the a s s u m p t i o n s of T h e o r e m
(2.3) are satisfied, and this m a k e s it p o s s i b l e to state the following
result.

(2.9) Theorem. Suppose the distribution A C has the m a x i m a l integral

manifolds property. Let S denote the m a x i m a l integral submanifold
o
x
of AC p a s s i n g through x . The set R(x ) of states reachable from x
under p i e c e w i s e constant input functions

(a) is a s u b s e t of S
o
x
(b) c o n t a i n s an o p e n set of S
o"
x
(2.10) Remark. Note that, if Ac h a s the maximal integral manifolds pro-
 66

pert y but is singular, then the d i m e n s i o n s of d i f f e r e n t maximal in-

tegral submanifolds of A c m a y be different. Thus, it m a y h a p p e n that
I 2
at two d i f f e r e n t initial states x and x one o b t a i n s two c o n t r o l
systems of the form (2.8) w h i c h evolves on two m a n i f o l d s Sxl and S
o
x
of d i f f e r e n t dimensions. We will see e x a m p l e s of this in s e c t i o n 4. 0

(2.11) Remark. Note that the a s s u m p t i o n "the d i s t r i b u t i o n ~C has the

maximal integral manifolds property"is implied by the a s s u m p t i o n "the
distribution ~C is n o n s i n g u l a r " . In this case, in fact, AC = R (see
Remark (2.7)) and R has the m a x i m a l integral manifolds property
(Lemma (2.2)). [}-

We c o n c l u d e this section by the illustration of some terminolo-

gy w h i c h is f r e q u e n t l y used. The control system (2.1) is said to sa-
tisfy the controllability rank aondition at x if

(2.12) d i m Ac(X) = n

Clearly, if this is the case, a n d if A C has the m a x i m a l integral

manifolds property, then the m a x i m a l integral submanifold of A C
passing through x has dimension n and, according to T h e o r e m (2.9),
the set of states reachable from x fill up at least an open set of
the state space N.
The following Corollary of T h e o r e m (2.9) describes the situation
w h i c h h o l d s w h e n one is free to choose arbitrarily the initial state
x O. A control s y s t e m of the f o r m (2.1) is said to be weakly control-
lable on N if for e v e r y initial state x 6 N the set of states reacha-
ble u n d e r p i e c e w i s e constant input functions contains at least an open
set of N.

(2.13) Corollary. A s u f f i c i e n t c o n d i t i o n for a control s y s t e m of'the

form (2.1) to be w e a k l y controllable on N is that

d i m Ac(X) = n

for all x 6 N. If the d i s t r i b u t i o n A C has the m a x i m a l integral

m a n i f o l d s p r o p e r t y then this c o n d i t i o n is also n e c e s s a r y .

Proof. If this c o n d i t i o n is satisfied, A C is n o n s i n g u l a r , involutive

and therefore, from the p r e v i o u s discussions, we c o n c l u d e that the
system is w e a k l y controllable. C o n v e r s e ly , if the d i s t r i b u t i o n A c has
the m a x i m a l integral manifolds property and d i m ~c(X ) < n at some
x 6 N then the set of states reachable from x b e l o n g s to a submani-
fold of N w h o s e dimension is s t r i c t l y less than n (Theorem (2.9)).So
 87

this set c a n n o t c o n t a i n an open s u b s e t of N. []

3. The O b s e r v a t i o n Space

In this s e c t i o n we study o u t p u t - s t a t e interaction properties from

a global p o i n t of view, for a s y s t e m d e s c r i b e d by e q u a t i o n s of the
form (2.1), together w i t h an o u t p u t m a p

(3.1) y = h(x)

The p r e s e n t a t i o n will be c l o s e l y analogue to the one g i v e n in the

previous section. First of all, recall that the local analysis carried
out in section 1.7 was b a s e d upon the c o n s i d e r a t i o n of the s m a l l e s t
codistribution invariant under the v e c t o r fields f , g l , . . . , g m and con-
taining the c o v e c t o r fields d h l , . . . , d h ~. If the a n n i h i l a t o r Q of this
codstribution is n o n s i n g u l a r , then it is also involutive (Lemma
I.(7.6)) and m a y be used to p e r f o r m a global decomposition of the
state space. Parallel to L e m m a (2.2) we have the following result.

(3.2) L e m m a . Suppose Q is n o n s i n g u l a r . Then Q has the m a x i m a l integral

manifolds property. []
The role of this d e c o m p o s i t i o n in e x p l a i n i n g the o u t p u t - s t a t e in-
teraction m a y be e x p l a i n e d as follows. Observe t h a t Q, b e i n g n o n s i n g u -
lar and involutive, satisfies the a s s u m p t i o n s of T h e o r e m (1.15) (be-
cause the set of all v e c t o r fields in a n o n s i n g u l a r distribution is
locally f i n i t e l y generated). Let S be any m a x i m a l integral submani-
fold of Q. Since Q is i n v a r i a n t under f , g l , . . . , g m and also under any
m
vector field of the form T = f + ~ giui , where u l , . . . , u m are real
i=I
numbers, using Theorem (1.15) we deduce that g i v e n any two points xa
m
and x b in S and any v e c t o r f i e l d of the form T = f + [ giui , the
i=I
points ~t(x ) and ~ (x b) for all t b e l o n g to the same m a x i m a l integral
submanifold of Q. In o t h e r words, we see that from any two initial
states on some m a x i m a l integral submanifold of Q, under the a c t i o n the
same p i e c e w i s e constant control one obtains two t r a j e c t o r i e s which,
at any time, pass through the same m a x i m a l integral submanifold of Q.
Moreover,it is e a s i l y seen that the functions h l , . . . , h are con-
stant on e a c h m a x i m a l integral submanifold of Q. For, let S be any of
these s u b m a n i f o l d s
and let h. denote the r e s t r i c t i o n of h. to S. A t
l 1
each p o i n t p of S the d e r i v a t i v e of hi along any v e c t o r v of TpS is
 68

zero, because Q C sp(dhi) , and therefore the function hi

A is a c o n -
stant.
As a conclusion, we immediately see that if x a a n d x b a r e two
initial states belonging to t h e same integral manifold of Q then under
the action o f the s a m e p i e c e w i s e constant control one obtains two
trajectories which, at any time, produce identical values on e a c h
component of the output, e.g. are indistinguishable.
These considerations enables us to s t a t e the following global
version of T h e o r e m I. (7.8).

(3.3) Theorem. S u p p o s e Q is n o n s i n g u l a r . Then there exists a parti-

tion of N into maximal integral submanifolds o f Q, all with the same
dimension. Let S denote the m a x i m a l integral submanifold of Q pas-
sing through x O. T h e n

(a) n o o t h e r point o f S c a n be d i s t i n g u i s h e d from x under piece-

wise constant input functions
(b) there exists an o p e n n e i g h b o r h o o d U o f x in N w i t h the p r o p e r t y
that any point x 6 U which cannot be distinguished from xO under
piecewise constant input functions necessarily belongs to U A S . []

Proof. T h e s t a t e m e n t (a) h a s already been proved. The statement (b)

requires some remark. Since Q is n o n s i n g u l a r , we know that around any
point x O we can find a neighborhood U and a partition of U into slices
each of w h i c h is c l e a r l y an i n t e g r a l submanifold o f Q. B u t also the
intersection of S with U, w h i c h is a n o n e m p t y open subset of S is an
integral submanifold of Q. T h e r e f o r e , since S is m a x i m a l , we deduce
that the s l i c e of U passing through x is c o n t a i n e d into U A S. From
the statement (b) o f T h e o r e m I.(7.7) we deduce that any other state
x of U which cannot be d i s t i n g u i s h e d from x under piecewise constant
inputs belongs t o the s l i c e of U passing through x , and therefore to
U ~ S. []

If the d i s t r i b u t i o n Q is s i n g u l a r , one m a y approach the p r o b l e m

on the b a s i s of the following considerations. Let {l i : I < i < } be a
finite set of r e a l - v a l u e d functions and {T i : I < i < q} be a f i n i t e set
of vector fields. Let S I and S 2 be two subspaces of C~(N) which both
contain the f u n c t i o n s 11,...,I Z a n d have the p r o p e r t y that, for all
6 S i and for all I ~ j ~ q, LT ~ ~ S i , i = 1,2. Clearly the in-
]
tersection S I N S 2 is a g a i n a subspaee of C~(N) which contains
11,...,~ i and is s u c h that, for all I e S I A S 2 and for a l l 1~j~q,
L T I E $I A $2. T h u s w e c o n c l u d e t h a t t h e r e e x i s t s a u n i q u e s u b s p a c e
3
S o f C~(N) w h i c h c o n t a i n s ~I .... ,~ a n d is s u c h that, f o r all I 6 S
 69

and for all I ~ j ~ q, L I e S. This is the smallest subspace of C~(N)

3
which contains ll,...,l Z and is closed under d i f f e r e n t i a t i o n along

TI,.-.,T q -
(3.4) Remark. The subspace S may be described as follows. Consider the
set

So= {X 6 C ~ ( N ) : I = l j or I = L r ...LT. lj; I < j < , I ~ i k ~ q, I ~ k < m}

ik 11
and let LC(S o) denote the set of all ~ - l i n e a r combinations of elements
of S O . Then, LC(So) = S. As a matter of fact, it is easily checked
that every element of LC(S o) is an e l e m e n t of S, so LC(S o) C S, that
Xj q LC(S o) for I ~ j ~ and that LC(S O) is closed under d i f f e r e n t i a -
tion along T I .... ,yq. []

W i t h the subspace S we may associate a codistribution ~S ' in a

natural way, by setting

~S = sp{dl : I e S}

The c o d i s t r i b u t i o n ~S is smooth by construction, but - as we

know - the d i s t r i b u t i o n ~S may fail to be so. Since we are interest-
ed in smooth d i s t r i b u t i o n s because we use them to partition the state
space into maximal integral submanifolds, we should rather be looking
at the d i s t r i b u t i o n smt(~) (see section 1.2).
The following result is important when looking at s m t ( ~ ) for
the purpose of finding global d e c o m p o s i t i o n s of N.

(3.5) Lemma. Suppose the set of all vector fields in s m t ( ~ ) is local-

ly finitely generated. Then s m t ( ~ ) has the m a x i m a l integral m a n i f o l d s
property.

Proof. In view of Theorem (1.11), we have only to show that s m t ( ~ ) is

involutive. Let T I and T 2 he vector fields in s m t ( ~ ) and I any func-
tion in S. Since <dl,T I ) = 0 and <dl,T 2 > = 0 we have

(dl,[T1,~ 2] > = LTI<dI,T 2 > - L T 2 < d i , ~ I > = 0

The vector field [YI,T2] is thus in ~ . But [TI,T2], being smooth, is

also in s m t ( ~ ) . []

In order to study o b s e r v a b i l i t y we consider the smallest sub-

space of C~(N) which contains the functions h l , . . . , h Z and is closed
under d i f f e r e n t i a t i o n along the vector fields f,gl,...,g m. This sub-
 70

space will be denoted by 0 and c a l l e d the o b s e r v a t i o n space. Moreover,

with @ we associate the c o d i s t r i b u t i o n

D0 = sp{dl : i 0}

(3.6) Remark. It is possible to prove that the d i s t r i b u t i o n ~0 is in-
v a r i a n t under the vector fields f, gl,...,g m. For, let I be any func-

tion in 0 and r a vector field in ~0" Then ( d l , ~ ) = 0 and {dLfl,Y)=0
because Lfl is again a function in 0. Therefore, from the equality

<dl,[f,T] )= Lf<dl,~ ) - < d L f l , T ) = 0

we deduce that [ f,T] a n n i h i l a t e s all functions in 0. Since ~0 is span-

ned by d i f f e r e n t i a l s of functions in 0, it follows that [f,Tl is a
vector field in ~ . In the same way one proves invariance under

gl,...,gm.
If the d i s t r i b u t i o n ~ is smooth (e.g. when the c o d i s t r i b u t i o n ~0
is nonsingular) then using Lemma I. (4.8) one concludes that ~0 itself
is invariant under f,gl,...,g m.

(3.7) Remamk. The distribution ~ and the d i s t r i b u t i o n Q introduced

in the previous chapter are related in the following way

(a) ~0
D Q

(b) if x is a regular point of ~0 ' then ~0(x) = Q(x).

We leave to the reader the proof of this statement. []

From the previous Remark and from Remark I. (4.6) it is deduced

that the d i s t r i b u t i o n smt(~) is i n v a r i a n t under the vector fields
f,gl,...,g m and so under any vector field T of the form
m
= f + ~ giui , where Ul,...,u m are real numbers. Now suppose that
i=I
the set of all vector fields in s m t ( ~ ) is locally finitely generated,
!

so that s m t ( ~ ) has the maximal integral manifolds property. Using

Theorem (1.15), as we did before in the case of n o n s i n g u l a r Q, we may
conclude that from any two states on the same integral submanifold of

smt(~0) , under the action of the same piecewise c o n s t a n t control one
obtains two t r a j e c t o r i e s that at any time lie on the same m a x i m a l in-
tegral submanifold of s m t ( ~ ) . Observe now that s m t ( ~ ) is also contain-
ed in sp{dhi }, I ~ i ~ , ~becauseevery tangent vector
~ in smt(Q~)
(x)-

is also in ~(x) and every tangent vector v in ~0(x) is such that
(dhi(x),v) = 0. Therefore one may deduce that the functions h i are
constant on each maximal integral submanifold of s m t ( ~ u) .
 71

This, together with the p r e v i o u s observations, shows that any two

initial states x a and x b on the same m a x i m a l integral submanifold of
smt(R~) are i n d i s t i n g u i s h a b l e under piecewise constant inputs. This
extends the s t a t e m e n t (a) of T h e o r e m (3.3). As for the s t a t e m e n t (b),
some r e g u l a r i t y is required, as it seen hereafter.

(3.8) Theorsm. Suppose the set of all v e c t o r fields contained in

smt(~) is l o c a l l y finitely generated. Let S denote the m a x i m a l i n -
tegral s u b m a n i f o l d of smt(O~) passing through x . Then

(a) any o t h e r p o i n t of S c a n n o t be d i s t i n g u i s h e d from x O u n d e r p i e c e -

wise constant inputs
(b) If x is a r e g u l a r point of ~0 ' then there'exists an open n e i g h -
borhood U of x in N with the p r o p e r t y that any p o i n t x 6 U which
c a n n o t be d i s t i n g u i s h e d from x O under p i e c e w i s e constant inputs
necessarily belongs to U N S.

Proof. The s t a t e m e n t (a) has a l r e a d y been proved. The statement (b)

is proved e s s e n t i a l l y in the same way as in the s t a t e m e n t (b) of The-
orem (3.3). []

The following example illustrates the need for the "regularity"

assumption in the statement (b) of the p r e v i o u s theorem.

(3.9) Example. Consider the following s y s t e m with N = ~ and

y = h(x)

where h(x) is d e f i n e d as

h(x) = exp(- ~)sin(~) for x ~ 0

x

h(0) = 0

For this system, two states x a and x b are indistinguishable if

and only if h(x a) = h(xb). In particular, the set of states w h i c h are
indistinguishable from the state x = 0 coincides with the set of the
roots of the e q u a t i o n h(x) = 0. E a c h p o i n t in this set is i s o l a t e d
but the p o i n t x = 0. Thus, no m a t t e r how small we choose an open neigh-
borhood U of x = 0, U c o n t a i n s points indistinguishable from x = 0.
It is also seen that
the c o d i s t r i b u t i o n D0 = sp{dh} has d i m e n s i o n
dh
I everywhere b u t at the points x in w h i c h ~-~ = 0, where its d i m e n s i o n
!

is 0. Thus, any smooth v e c t o r field b e l o n g i n g to ~ u m u s t v a n i s h ident-

 72

!
ically on ~ a n d smt(Q~) = {0}. The maximal integral submanifold of
smt(~) passing through x is the p o i n t x itself.
At the point x = 0, w h i c h is n o t a regular point of ~0 ' we have
that U N S = {0} for all U, w h e r e a s we k n o w there are other poinEs of
U indistinguishable f r o m x = 0. []

We c o n c l u d e this section with some global considerations. The

control system (2.1)-(3.1) is s a i d to s a t i s f y the observability rank
condition at x if

(3.10) d i m ~ 0 ( x ) = n

Clearly, if this is the c a s e then x is a r e g u l a r point of ~0 and

from the p r e v i o u s discussion it is seen that any point x in a s u i t a b l e
neighborhood U of x c a n be d i s t i n g u i s h e d under piecewise constant
inputs. A control system of the form (2.1)-(3.1) is s a i d to be locally
observable o n N if for e v e r y state x there is n e i g h b o r h o o d U of x
in w h i c h every point can be d i s t i n g u i s h e d from x under piecewise con-
stant inputs.

(3.11) Corollary. A s u f f i c i e n t condition for a c o n t r o l system of the

form (2.1)-(3.1) to be locally observable on N is t h a t

d i m ~0(x) = n

for all x E N.

4. L i n e a r Systems, Bilinear Systems a n d Some Examples

In t h i s section we describe some e l e m e n t a r y examples, in o r d e r

to m a k e the reader more familiar with the ideas introduced so far.
As a f i r s t application, we shall compute the Lie algebra C and
the d i s t r i b u t i o n ~C for a linear system

= Ax + Bu

y = CX

We m a y easily interpret this system as a s y s t e m of the form (2.1)-(3.1).

The m a n i f o l d N on w h i c h the system evolves i s the w h o l e of ~ n and,in
the standard (single) coordinate chart of ~ n , the v e c t o r fields f(x)
a n d g1(x) ..... gm(X) have the e x p r e s s i o n s
 73

(4.1) f(x) = Ax

(4.2) gi(x) = bi 1<i<mp

where b i is the i-th column of the m a t r i x B. The functions h I (x),...,

h(x) are e x p r e s s e d as

h. (x) = c . x 1 < i <

l 1 -- --

where c i is the i-th r o w of the m a t r i x C.

We want to p r o v e first that the Lie algebra C is the subspace of
V(N) consisting of all v e c t o r fields which are'l~-linear combinations
of the v e c t o r fields in the set

(4.3) {Ax} t) {Akb. : I < i < m, 0 < k < n-l}

1 . . . .

For, observe that this set c o n t a i n s the v e c t o r fields Ax and b 1,...,b m

(i.e. the v e c t o r fields f and gl,...,gm ) and also that this set is
contained in C, b e c a u s e any of its e l e m e n t s is a r e p e a t e d Lie bracket
of f and g l , . . . , g m. As a matter of fact,

Akbi = [[ [gi' f] .... ] ' f]

k-times

Moreover, it is e a s y to see that the set

(4.4) LC({Ax} U {Akb. : I < i < m, 0 < k < n-l})

1 . . . .

of all l~-linear combinations of v e c t o r fields in the set (4.3) is al-

ready a Lie subalgebra, i.e. is c l o s e d under Lie b r a c k e t i n g .
For, one easily sees that if T I (x) a n d ~2(x) are v e c t o r fields
of the f D r m

T I (x) = A k b i

T 2 (x) -- Ahbj

then [~I'T2] (x) = 0. On the o t h e r hand, if

~I (x) = A k b i

Y2 (x) = A x
 74

then

[ T I,T 2] = A k + I b
1

If k < n-l, this vector field is in the s e t (4.3) and, if k = n - l , t h i s

vector field is an A - l i n e a r combination of v e c t o r fields in the set
(4.3) (by C a y l e y - H a m i l t o n theorem) .
If ~I and ~2 are ]R-linear combinations of v e c t o r fields of (4.3),
then their Lie bracket is still an ~ - l i n e a r combination of v e c t o r
fields of (4.3), and this proves that the set (4.4) is a Lie sub-
algebra.
The set (4.4) is a L i e algebra which contains f'g1'" "''gm a n d is
contained in C, the s m a l l e s t Lie subalgebra which contains f'g1'''''gm"
Then, the set (4.4) coincides with C.
Evaluating the d i s t r i b u t i o n A C we get, at a p o i n t x 6 I~n,

(4.5) Ac(x) = s p a n { A x } +span{Akb : I < i < m, 0 < k < n-l}

l . . . .

n-1
= span{Ax}+ ~. Im(AkB)
k=0

We are also interested in the distributionP, the smallest di-

stribution which contains gl,...,g m and is i n v a r i a n t under f , g l , . . . , g m.
By m e a n s of a r g u m e n t s similar to the ones used before or, else, by
means of the r e c u r s i v e algorithm presented at the b e g i n n i n g of section
1.6, it is n o t difficult to d i s c o v e r that, at a n y p o i n t x 6 ~n,

(4.6) P(x) = span{Akbi : I .< i. < .m, .0 < k < n-l}

Thus, we see that

&C = sp{f} + P

The distribution A C is s p a n n e d by a set of v e c t o r fields which is

locally finitely generated (because any v e c t o r field in C is a n a Z y t i c
on A n ) , and therefore - by Lemma (2.5) - the distribution &C has the
maximal integral manifolds property. The distribution P is n o n -
singular and involutive and thus - by C o r o l l a r y (1.12) - it a l s o has
the m a x i m a l integral manifolds property.
The maximal integral submanifolds of P, all of the same dimension,
have the f o r m x+V, where
 75

V = Im(B)+Im(AB)+...+Im(An-IB)

(see E x a m p l e I.(3.4) and Remark I.(4.2)). The m a x i m a l integral submani-

folds of ~C m a y have different dimensions, because ~C m a y have sin-
gularities.
If, at some point x 6 ~n, f(x) E P(x), then the m a x i m a l integral
submanifold of ~C p a s s i n g through x coincides with the one of the
distribution P, i.e. is a s u b s e t of the form x+V. Otherwise, if s u c h
a condition is not verified, the m a x i m a l integral submanifold of ~C is
a submanifold whose dimension exceeds by I that of P a n d t h i s submani-
fold,in turn, is p a r t i t i o n e d into subsets of the form x'+V

(4.7) Example. The following simple example illustrates the case of

a singular ~C" Let the system described by

= 0 -I 0 x +

1
0 0 I

Then we e a s i l y see that

v = {x 6 ~ 3 x 2 = x 3 = 0}

and that

p = sp{~}

The tangent vector f(x) belongs to P o n l y at those x in w h i c h

x 2 = x 3 = 0, i.e. only on V. Thus, the m a x i m a l integral submanifolds
of A C w i l l have dimension 2 everywhere but on V. A d i r e c t computation
shows that these submanifolds may be d e s c r i b e d in the following way:

(i) if x O is s u c h that x~ = 0 (resp. x~ = 0) then the m a x i m a l sub-

manifold passing through x is the h a l f open plane

{x E ~ n : x2 = 0 a n d s g n ( x 3) = s g n ( x ~ ) }

(resp. {x e l R n : x 3 = 0 a n d sgn(x2) = sgn(x2)}

(ii) o
if x is such that both x 2 ~ 0 a n d x 3 ~ 0, then the m a x i m a l sub-
manifold passing through x is the surface
o o
{x e ]Rn : x 2 x 3 = x2x3}. []
 76

We turn now on the computation of the subspace 0 and the codi-

stribution ~0" It is easy to prove that 0 is the subspace of C ~(N)
consisting of all IR-linear combinations of functions of the form
ciAkx or ciAkbj , namely that

(4.8) 0 = LC{I C~(N):X(x) = ciAkx or ~(x) = ciAkb j

I < i < , I < j < m, 0 < k < n-l]

For, note that functions of the form CiA'~x~ or ciAkb j are such that

c'Akxl = Lf...Lfh i(x)

k-times

ciAkb j = LgjLf... Lfh i(~)

k-times

and this implies that the right-hand-side of (4.8) is contained in 0.

Moreover, the functions h1,...,h are elements of the right-hand-side
of ( 4 . 8 ) . T h e n , t h e p r o o f of (4.8) is completed as soon as we show that
its right-hand-side is closed under differentiation along f'g1' .... gm"
If ~(x) = ciAkx, then LfX = ciAk+Ix and Lgj~(x) = ciAkb j. If

X(x) = ciAkb j , then LfX(x) = Lgj%(x) = 0. Thus, using again Cayley-

Hamilton Theorem, it easily seen that the right-hand-side of (4.8) is

closed under differentiation along f'gl .... 'g m"
At each point x, the codistribution ~0 is given by
~0(x) = span{ciAk : I ~ i ~ , 0 ~ k ~ n-l} and therefore

n~(x) n-1
n ker(CA k)
k=0

The codistribution ~0 is nonsingular, and so is the distribution

e 0. Moreover, ~0 = smt(~0)'Frm Remark (3.7) we see that S 0 = Q and
so this distribution has the maximal integral manifolds property (Lemma
(3.2)). The maximal integral submanifolds of Q have now the form x+W
where

W = ker(C) N ker(CA)...N k e r ( C A n-l)

As a second application we consider a bilinear system, i.e. a

system described by equations of the form
 77

m
= Ax + [ (Nix)u i
i=1
y=Cx

Here a l s o the m a n i f o l d on w h i c h the s y s t e m e v o l v e s is the w h o l e of A n,

we set f a n d h l , . . . , h as b e f o r e , and

gi(x) = Nix I ~ i ~ m

In o r d e r to c o m p u t e the s u b a l g e b r a C we n o t e f i r s t t h a t any v e c t o r
field T in the set {f'gl ..... g m } has the f o r m T(x) = Tx, w h e r e T is an
nxn matrix. If we w a n t to take the Lie b r a c k e t of two v e c t o r fields
TI,T 2 of the f o r m

T 1(x) = T I x , T 2 (x) = T 2 x

we have

[ TI,Y 2] (x) = (T2TI-TITz)X = |TI,T2]x

where [ T I , T 2] = (T2TI-TIT2) is the commutator of T I a n d T 2.

On the b a s i s of this o b s e r v a t i o n , it is e a s y to s e t up a re-
cursive p r o c e d u r e yielding the s m a l l e s t Lie s u b a l g e b r a which contains
a set of v e c t o r fields of the f o r m T1(x) = T1x,...,Tr(X) = TrX.

(4,..9) Lemma. C o n s i d e r the n o n d e c r e a s i n g sequence of s u b s p a c e s of ~ n n ,

the ~ - v e c t o r s p a c e of: all nxn m a t r i c e s of r e a l n u m b e r s , d e f i n e d by
setting

Mo = span{T1,...,Tr}

Mk, = Mk_ I +span{[T.i,T] ..... [Tr,T] :T E ~k_1}

Then, there e x i s t s an i n t e g e n k such that

Mk = ~k*
for all k > k . The set of v e c t o r fields

L = {~ e v ~ n ) : T ( x ) =.Tx, T e Mk. }

is the s m a l l e s t Lie subalgebra of v e c t o r fields which contains

 78

~1(x) = T 1 x ..... Tr(X) = TrX.

Proof. The proof is rather simple and c o n s i s t s in the f o l l o w i n g steps.

A d i m e n s i o n a l i t y a r g u m e n t proves the e x i s t e n c e of the integer k ~ such
that M k = Mk~ for all k > k . Then, one checks that the subspace Mk,
c o n t a i n s T I , . . . , T r and any r e p e a t e d c o m m u t a t o r of the form
[Til,...,[Tih_1,Tih]] and is such that [P,Q] e Mk~ for all P e Mk~ and

Q E Mk~ . F r o m these p r o p e r t i e s , it is s t r a i g h t f o r w a r d to deduce that

i is the d e s i r e d Lie algebra. []

B a s e d on this result, it is easy to c o n s t r u c t the Lie a l g e b r a C

by s i m p l y i n i t i a l i z i n g the a l g o r i t h m d e s c r i b e d in the above L e m m a with
the m a t r i c e s A , N I , . . . , N m-
In this case, unlike the p r e v i o u s one, we cannot anymore give a
simple e x p r e s s i o n of &c(X) and/or its m a x i m a l integral submanifolds.
In some special situations, however, like the one i l l u s t r a t e d in the
f o l l o w i n g example, a rather s a t i s f a c t o r y analysis is possible.

(4.10) Example. C o n s i d e r the system

= Ax + N x u

where x E ~ 3 and

0 1 0 0 0 I

A = -I 0 0 N = 0 0 0

0 0 0 -1 0 0

An easy c o m p u t a t i o n shows that

0 0

[A,N] = 0 0

0 -1

[N,[A,N]] = A

[A,[A,N]] = -N

Therefore, we have

C = {T e V(I~3):T(x) = Tx, T span{A,N,[A,N] }}

 79

To compute the dimension of A C we e v a l u a t e the rank of the m a t r i x

x2 x3 0

(Ax,Nx, [ A,N] x) = -x I 0 x3

0 -x I -x 2

and we f i n d the following result

d i m Ac(X) = 0 if x = 0

d i m Ac(x) = 2 if x ~ 0

A direct computation shows that the m a x i m a l integral submanifold

of A C p a s s i n g through x is the set

{x E IR3 : x 21 + x 22+ x 32 = ( x ~ ) 2 + ( x ~ ) 2 + ( x ~ ) 2 }

i.e. the sphere centered at the o r i g i n passing through x .

Therefore, we can say that the state of the system is n o t free
to e v o l v e on the w h o l e of A n, but rather on the sphere centered at
the o r i g i n which passes through the initial state.
Around any p o i n t x ~ 0 the d i s t r i b u t i o n A C is n o n s i n g u l a r , so we
can o b t a i n locally a decomposition of the form I.(6.12), by means of
a suitable coordinates transformation.
To t h i s end, we m a y m a k e use of the c o n s t r u c t i o n introduced in
the p r o o f of T h e o r e m I.(3.3) and find a set of three vector fields

TI,T2tT3 with the p r o p e r t y that T I and T 2 belong to A C a n d TI(xO),

T2 (xO), T3(X ) are linearly independent. If we consider an i n i t i a l
point on the line

{x 6 ~ 3 : x I = x 2 = 0}

we m a y take the v e c t o r fields

~I (x) = (Nx)

T2(x) = ([A,N]x)

T 3 (x) = (0 0 I)'

Accordingly, we g e t
 8O

(COS t)x I + (sin t)x 3

1 (x) =
#t x2

-(sin t)x I + (cos t)x 3

x1

(cos t)x 2 + (sin t)x 3 1

2(X) =
~t
-(sin t) x 2 + (cos t)x 3 ]
xI

~3t(x) = x2

t+ x3

The local coordinate chart around the point x is given by the

inverse of the function

F : (Zl,Z2,Z 3) I ~ ~I ~2 ~3 (xO)
Z1 z2 z 3

For x~ = x O2 = 0 and x~ = a we have

(sin Zl) (cos z2) (z3+a)

F(Zl,Z2,Z 3) = (sin z2) (z3+a)

(cos zl) (cos z2) (z3+a)

The local representations of the vector fields f and g in the

new coordinate chart are given by

~(z) = (F~)-If(F(z)) = (F~)-IAF(z)

~(z) = (F~)-Ig(F(z)) = (F~)-INF(z)

A simple but tedious computation yields

 81

cos zltg z2

~(z)= -sin zI ; ~(z)=

We c o n c l u d e that around x the system, in the z coordinates, is

described by the e q u a t i o n s

Zl = COS zltq z2 + u

z2 = -sin zI

z3 = 0
[]

The study of the o b s e r v a b i l i t y of a b i l i n e a r system is m u c h

simpler. By means of a r g u m e n t s similar to t h o s e used in the case of
linear systems it is e a s y to p r o v e that 0 is g i v e n by

0 = LC{I e C~(N):I(X) = ciNjl...NjkX ;

1 .< i <. , I. < k . < n-l;

. 0 .< j l , . . . , j k < m}

(with N o = A). Therefore

I n-1 m
~(x) = n k e r (CNj I )
k=0 Jl ..... Jk =0 " " "Njk

The distribution ~0 = Q is n o n s i n g u l a r and its m a x i m a l integral
submanifolds have the f o r m x+W, where now

n-1 m
W = A N ker(CN~j ...Njk)
k=0 J1'" "''Jk =0 I

It m a y be w o r t h observing that the s u b s p a c e W thus defined is

invariant under A,NI,...,N m , is c o n t a i n e d in ker(C) and is the l a r g e s t
subspace of ~ n h a v i n g these properties. From linear algebra we k n o w
that by m a k i n g a suitable change of c o o r d i n a t e s in ~ n (see e.g. sec-
tion I.I) the m a t r i c e s A,NI,...,N m become block triangular and, the-
refore, the d y n a m i c s of the system becomes described by e q u a t i o n s of
the f o r m
 82

Xl = A11xl +A12x2 +i=I

~ (Ni'11xl + Ni,12x2)ui

m
x2 = A22x2 + ~ Ni,22x2ui
i=I

Moreover, the output y depends only on the x 2 coordinates,

y = C2x 2

The above equations are exactly of the form I. (5.7), this time ob-
tained by means of standard linear algebra arguments.
 CHAPTER III
INPUT-OUTPUT MAPS AND REALIZATION THEORY

I. F l i e s s Functional Expansions

The p u r p o s e of this section and of the following section is to

describe representations of the input-output behavior of a n o n l i n e a r
system. We consider, as usual, systems described by differential equa-
tions of the form

m
(1.1a) x = f(x) + ~ gi(x)ui
i=1

(1.15) yj = hj(x) j = I .....

Throughout the c h a p t e r , we systematically assume that the m a n i -

fold N on w h i c h the state evolves is an o p e n s e t of ~ n a n d that the
vector fields f , g l , . . . , g m are analytic vector fields defined on N.
Likewise, the o u t p u t functions h l , . . . , h Z are analytic functions de-
fined on N.
For the sake of n o t a t i o n a l convenience most of the times we re-
present the o u t p u t of the system as a v e c t o r - v a l u e d function

y -- h(x) = (h I (x) ...hi(x)) '

We r e q u i r e first some combinational notations. Consider the set

of m+1 indexes I = {0,1 ..... m} (we r e p r e s e n t here, as usual, indexes
with i n t e g e r numbers, but we could as w e l l represent the m+1 indexes
with e l e m e n t s of a n y s e t z with card(Z) = m+1). Let I k be the set of
all s e q u e n c e s (ik...il) of k e l e m e n t s i k , . . . , i I of I. An e l e m e n t of
this set Ik w i l l be c a l l e d a multiindex of lenght k. For c o n s i s t e n c y
we d e f i n e also a set 10 w h o s e unique element is the e m p t y sequence
(i.e. a m u l t i i n d e x of lenght 0), denoted @. F i n a l l y , let

I = U Ik
k>O

It is e a s i l y seen that the set I can be g i v e n a structure of

free m o n o i d , with composition rule

(ik...il) (jh...jl) J ~ ( i k . . . i l J h . . . j I)
 84

with neutral element ~.

A formal power s~ries in m+1 noncommutative indeterminates and
coefficients in ~ is a mapping

c : I i ,

In what follows we represent the value of c at some element ik...i 0

of I with the symbol c(ik...i0).
The second relevant object we have to introduce is called an
iterated integral of a given set of functions and is defined in the
following way. Let T be a fixed value of the time and suppose ul,...,um
are real-valued piecewise continuous functions defined on [ 0,T] . For
each multindex (ik...i0) the corresponding iterated integral is a
real-valued function of t

Eik-'-ili0(t) = I d~ik'''d~ild~i 0
0

defined for 0 ~ t ~ T by recurrence on the lenght, setting:

t0 (t) = t
t
~i(t) = lui(~)d~ for 1 < i < m
0
and
t t T

! d~i---d~i0
k
= ~d~i (~)Id~ik_1--'d~i0
!u k
0

The iterated integral corresponding to the multindex ~ is the

real number I.

(1.2) Example. Just for convenience, let us compute the first few
iterated integrals, in a case where m = I.

t t t
Id~o = t ; ]d~l = lu1(T)d~
0 0 0
t t t
Id~0d~0 - t2~ ; Id~0d~1 = I I~(@)d0dT
0 0 0 0
 85

t t t t

0 0 0 0 0

Given a formal power series in m+1 non-commutative indeterminates,

it is possible to associate with this series a functional of ~ , . . . , u
1 ILt

by taking the sum over I of all the products of the form

t
c(ik-.-i0)Id~ik--.d~i0
0

The convergence of a sum of this kind is guaranteed by some growth

condition on the "coefficients" c(ik...i0) , as stated below.

(1.3) Lemma. Suppose there exist real numbers K > 0, M > 0 such that

(1.4) Ic(ik...i0) I < K(k+I)!M k+l

for all k ~ 0 and all multiindexes ik...i 0.

Then, there exists a real number T > 0 such that, for each
0 ~ t ~ T and each set of piecewise continuous functions Ul,...,u m de-
fined on [0,T] and subject to the constraint

(1.5) max lui(T) I <I,

0<T<T

the series
t
(1.6) y(t) = c(~) + m~ c(ik...i 0) i d ~ i k . .0d ~"i
k=0 iD'''''ik=O 0

is absolutely and uniformly convergent.

Proof. It is easy to see, from the definition of iterated integral,

that, if the functions u I ..... u m satisfy the constraint (1.5) then

t
f tk+1
d~ik''-d~i0 ! ~
0

If the growth condition is satisfied, then

t
I m~ c(ik---i 0) I d~ik-- "d~i 0 I _< K[M(m+1)t]k+1
. . .,i k = 0 0

As a consequence, if T is sufficiently small, the series (I .6) con-

 86

verges absolutely and uniformly on [0,T] . []

The expression (1.6) clearly defines a functional of U l , . . . , u m-

This functional is causal, in the sense that y(t) depends only on the
restrictions of U l , . . . , u m to the time interval [0,t].
A representation of the form (1.6) is u n i q u e .

(1.7) Lemma. Let c a a n d c b be two f o r m a l p o w e r series in m+1 noncom-

mutative indeterminates and let the associated functionals of the
form (1.6) be d e f i n e d on the same interval I 0,T] . T h e n the two func-
tionals coincides if a n d o n l y if c a = c b. D

Proof. L e t c a , c b be t w o f o r m a l p o w e r series and ya(t),yb(t) the as-

sociated functionals of the form (1.6). Note that

y(t) = ya(t)_yb(t)

is s t i l l a functional of the form (1.6) associated with a formal power

series c whose coefficients are defined as d i f f e r e n c e s between the
corresponding coefficients of c a a n d c b. To p r o v e the lemma, all w e
need is to show that if y(t) = 0 for all t e [0,T] and for all input
functions, all the coefficients of the series c vanish.
If, in p a r t i c u l a r , u1=...=Um=0 on [0,T], then y(t) = 0 for all
t 6 [ 0,T] implies

t2
c(g) + c(0)t + c(00)~T+ .... 0

for all t E [0,TI, i.e.

c(~) = 0

c(0...0) = 0 I < k <

k-times

Taking the d e r i v a t i v e of (1.6) with respect to time and e v a l u a t -

ing it at t = 0, one o b t a i n s

m
( dd--t)t=0 = ~' c ( i ) u i ( 0 )
i=I

Therefore, (~t) t=0 = 0 for all u I (0),...,Um(0) implies

c(i) = 0 I < i < m

Continuing this way, one m a y compute the second derivative of y(t) at

 87

t = 0 and g e t

2t__~2 m m (c (0i) +c (i0)u i (0)

= [ C ( i l i o ) u i (0) (0) +
( )t=O i0,ii= I I Uio i=I

If this is z e r o for all u 1 ( 0 ) , . . . , U m ( 0 ) , then

C ( l l i 0) = 0 I ! il,i 0 ! m

c(0i) = -c(i0) I < i < m

In the t h i r d d e r i v a t i v e , the c o n t r i b u t i o n of t e r m s

t t
(c(0i)Id~0d(i + c ( i 0 ) I d ~ i d ~ 0)
i=I
0 0

is

1 1 dui
i = I [ ~ c(0i) + ~ c(i0)] (-d~)t=0

du.
If this is zero for all l
(-~-)t=0 , then c(0i) = -2c(i0) which, toge-
ther w i t h the p r e v i o u s equality c(0i) = -c(i0) implies

c(0i) = 0 I < i < m

C o n t i n u i n g in the same way, one m a y c o m p l e t e the proof. []

We are n o w g o i n g to s h o w that the o u t p u t y(t) of the n o n l i n e a r

system (1.1) c a n be r e p r e s e n t e d as a f u n c t i o n a l of the i n p u t s Ul, .... um
in the f o r m (1.6). To this e n d we n e e d some p r e l i m i n a r y results.

(1.8) Lemma. Let g 0 , g l , . . . , g m be a set of a n a l y t i c vector fields and

I a real-valued analytic function defined on N. G i v e n a p o i n t x O 6 N,
consider the f o r m a l p o w e r series defined by

c (H) = ~ (x )
(I .9)
c ( i k . . . i l i 0) = L L . .L l(x )
gi 0 gi I gi k

Then, there e x i s t real n u m b e r s K > 0 and M > 0 s u c h t h a t the g r o w t h

condition (1.4) is s a t i s f i e d .

Proof. The r e a d e r is r e f e r r e d to the l i t e r a t u r e . []

 88

In view of this result and of Lemma (1.3), one may associate with
g 0 , g l , . . . , g m and I the functional

t
(1.10) v(t) = l(x ) + m[ L Lgil.. .Lg l(x) id~ik...d~ild~i0
k=0 io,...,ik=O gi 0 0
(1.11) Lemma. Let g 0 , g l , . . . , g m be as in the previous Lemma and let
11,...,I Z be r e a l - v a l u e d analytic functions defined on N. Moreover,
let y be a r e a l - v a l u e d analytic function defined on ~Z.
Let v1(t),...,vi(t) denote the functionals defined by'setting, in
(1.10), I = 11,...,i = I. The composition Y(v1(t),...,vi(t)) is again
a functional of the form (1.10), corresponding to the setting
I = y(ll,...,l~)-
Proof. We will only give a trace to the reader for the proof. Let
c1,c 2 denote the formal power series defined by setting, in (1.9),
I = 11 and r e s p e c t i v e l y I = 12 , and let v1(t),v2(t) denote the as-
sociated functionals (1.10). Then, it is immediately seen that with
the formal power series defined by setting I = ~i11 + ~212 , where ~I
and e2 are real numbers, there is a s s o c i a t e d the functional
~Iv1(t) + ~2v(t).
With a little work, it is also seen that with the formal power
series d e f i n e d by setting I = 1112 , there is associated the func-
tional v1(t)v2(t). We show only the very first computations needed
for that. For, consider the product
t t t
v I (t)v2(t) = (11+Lg0l I ; d~0+Lg I 1 IId~ 1+Lg0Lg01 lld~0d~0+" "" )
0 0 0

2+Lgo21do+glX2it d1*gogo21
0
t t
0 0

where, for simplicity, we have omitted specifying that the values of

all the functions of x are to be taken at x = x . M u l t i p l y i n g term-
b y - t e r m we have
t t
v1(t)v2(t)=1112+(lILg012+12Lg011);d%+(liLg112+12Lg111);d~1 +
0 0
t
t
(II+L L I.+I~L L 14) Jld~0d~0 +
go go z ~ go go !
0
t t

( 'go l I .go 21 (Iod o)

 89

The factors that multiply it d~0 and it d~1 are clearly Lg011 ~2 and re-
0 0
spectively Lg11112. For the other three, we have

Lg0Lg0lll 2 = lILg0Lg012+12Lg0Lg011+2(Lg011) (Lg012)

but also
t t t

0 0 0
so that the three terms in question give exactly

t
Lg0Lg01112 Id~0d~0
0
It is not difficult to set up a recursive formalism which makes it
possible to completely verify the claim.
If now ~ is any real-valued analytic function defined on ~, we
may take its Taylor series expansion at the origin and use recursively
the previous results in order to show that the composition
Y(v1(t),...,vz(t)) may be represented as a series like the (1.10)
with I replaced by the Taylor series expansion of (ll,...,IZ). []
At this point, it is easy to obtain the desired representation
of y(t) as a functional of the form (1.10).
(1.12) Theo~Gm. Suppose the inputs Ul,...,u m of the control system
(1.1) satisfy the constraint (1.5). If T is sufficiently small, then
for all 0 ~ t ~ T the j-th output yj (t) of the system (1.1) may be
expanded in following way
t
(1.13) yj (t)=hj(x ) + ~ m
~ L I
...Lg hj(x) d~ik.
k=0 i0'''''ik=0 gi0 0 ""d~i0

where g0 = f"
Proof. We first show that the j-th component of the solution of the
differential equation (1.1a) may be expressed as
t
m ...L x~ (x) Id~.
(1.14) xj (t)=xj (x)+ ~ ~ L Y~ J 0 ik ..d~i0
k=0 i0,...,ik=0 gi 0

where the function xj(x) stands for

 90

xj : (x I ..... Xn) t ' xj

Note that, by definition of iterated integral

t t
d-~
d !d~0d~ik_1 " "-'d~i0 = J0
[d~ik-1" "'d~i 0

and
t t
~t Id~id~ik_1"''d~io = ui(t)Id~ik_1'''d~io
0 0

for 1 < i < m. Then, taking the derivative of the right-hand-side of

(I .14) with respect to the time and rearranging the terms we have
t
~j (t)=Lfxj (x)+ ~ ~ L ...Lg Lfxj(x )Id~ik...d~i0 +
k=0 i0,...,ik=0 gi 0 ik 0
t
m ~ m I
[, L ...L L x. (xO) d~ik
+i~1[LgiXj (x)+k=~0 10 .... ik=0 gi0 gi k gi ] 0 "''d~i0]ui(t)

Now, let fj_ and gij denote the j-th components of f and gi '
I < j < n, I < i < m and observe that

Lfxj = fj (xI ..... x n)

Therefore, on the basis of the Lemma (1.11), we may write

Lfxj (xO) + ~ m L ...Lg Lfxj (xO) it d~ik...d~i0

k=0 i 0 .... ,ik=0 gi 0 ik 0

t
cx + L Lgikf j i aq0 =
k=0 i 0 , .... ik=O gi 0 ~ k

fj (x I (t) ..... x n(t))

A similar substitution can be performed on the other terms thus yield-

ing
m
xj (t) = fj(x I (t) ..... Xn(t))+ gij (Xl (t) ..... X n (t)) U i (t)
i=I
 91

Moreover, the xj(t) satisfy the c o n d i t i o n

xj(0) = x~

and therefore are the c o m p o n e n t s of the solution x(t) of the d i f f e r -

ential e q u a t i o n (1.1a).
A further a p p l i c a t i o n of L e m m a (1.11) shows that the output (1.1b)
can be e x p r e s s e d in the form (1.13). []

The d e v e l o p m e n t (1.13) will be from now on r e f e r r e d to as the

fundamental formula or Flies8 functional expansion of yj(t). Ob-
viously, one m a y deal d i r e c t l y w i t h the case of a v e c t o r - v a l u e d output
with the same formalism, by just r e p l a c i n g the s c a l a r - v a l u e d function
hi(x) with the v e c t o r - v a l u e d function h(x). We stress that, from Lemma
(1.3), it is known that the series (1.13) c o n v e r g e s a b s o l u t e l y and
uniformly on [0,T].

(1.15) Remark. The reader will i m m e d i a t e l y observe that the functions

hj(x) and Lg i ...L hj(x), with I ~ j ~ ~ and (ik...i0) (I*\I0) ,
0 gik
whose values at x c h a r a c t e r i z e the f u n c t i o n a l (1.13), span the ob-
servation space 0 d e f i n e d in section II.3. []

(1.16) Ezamples. In the case of a linear system, the formal p o w e r

series w h i c h c h a r a c t e r i z e s the functional (1.13) takes the form

c(~) = c x
3

cjAk+Ix O if i 0 = ... = ik = 0

c ( i k . . . i 0) = cjAkbio if i0 ~ iI = ... = i k = 0

0 elsewhere

In the case of a b i l i n e a r system, the formal p o w e r series w h i c h

characterizes the f u n c t i o n a l (1.13) takes the f o r m

c(9) = c x
3

o
e(i k. ..i 0) = c j N i k "Ni0x

where N O = A. []
 92

2. V o l t e r r a Series Expansions

The i n p u t - o u t p u t behavior of a n o n l i n e a r s y s t e m of the f o r m (1.1)

m a y a l s o be r e p r e s e n t e d by means of a s e r i e s of generalized convoZu-
tion integrals. A generalized convolution i n t e g r a l of o r d e r k is de-
f i n e d as follows. Let (ik...i I) be a m u l t i i n d e x of l e n g h t k, w i t h
ik,...,i I elements of the set {I, .... m}. With this m u l t i i n d e x there
is a s s o c i a t e d a r e a l - v a l u e d continuous function w. , defined on
the s u b s e t of ~ k + 1 Zk'''il

S k = {(t,T k ..... TI) e l ~ k + l : T > t > Tk... > T I > 0}

where T is a f i x e d n u m b e r . If u I ,... ,u m are real-valued piecewise con-

tinuous functions defined on [ 0,T], the g e n e r a l i z e d convolution in-
t e g r a l of o r d e r k of u l , . . . , u m w i t h k e r n e l W i k ...il is d e f i n e d as

t Tk T2

I I -.-I W i k . . . i l ( t , Tk ..... ~ l ) U i k ( T k ) . . . u i (TI)dTI...dT k

0 0 0

for 0 < t < T.

For consistency, if k = 0, r a t h e r than a generalized convolution
integral, one c o n s i d e r s simply a continuous real-valued function w 0
defined on the set

SO = {t e l ~ : T >_ t ! 0}

The s u m of a s e r i e s of g e n e r a l i z e d convolution integrals may de-

scribe a functional of ul, .... u m , under the c o n d i t i o n s s t a t e d below.

(2.1) Lemma. Suppose there exist real numbers K > 0, M > 0 s u c h t h a t

(2.2) lWik ...iI (t,T k ..... T I) I < K(k) !M k

for all k > 0, for all m u l t i i n d e x e s (ik...il) , a n d all (t,~k,...,~l) 6 ~ .

Then, there exists areal number T > 0 such that, for e a c h
0 ~ t ~ T and each set of p i e c e w i s e continuous functions ul,...,u m
defined on [0,T] and s u b j e c t to the c o n s t r a i n t

(2.3) max lui(T) I < I,

0<~ < T

the s e r i e s
 S3

t ~k T2
(2.4) ~, ... W. (t,Tk ..... TI)U~(Tk)...
y(t) =w0(t)+k~=1 ii .... ik=1 0 ~k'"il

Uil (T I )dT I ..d~ k

is absolutely and uniformly convergent.

Proof. It is similar to that of Lemma (1.3). []

The e x p r e s s i o n (2.4) clearly defines a functional of u I ..... Um,

which is causal, and is called a Volterra serges expansion.
As in the previous section, we are interested in the p o s s i b i l i t y
of using an expansion of the form (2.4) for the output of the non-
linear system (1.1). The existence of such an expansion and the ex-
pressions of the kernels may be d e s c r i b e d in the following way.

(2.5) Lemma. Let f,gl,...,g m be a set of analytic vector fields and

I a real-valued analytic function defined on N. Let ~tf denote the
flow of f. For each pair (t,x) E ~ N for which the flow ~(x) is
defined, let Qt(x) denote the function

(2.6) Qt(x) = lo~f(x)

It
and P (x),...,Ptre(x) the vector fields

c2.7) Pt(x) -- (~_ft), giocft(x)

I < i < m. Moreover, let

(2.8') w0(t) = Qt(x )

(2.8") Wik...il (t,T k .... ,T I) = (L i1(x ) ...L ik (x) Qt (x)) x=xO

PT I PT k

Then, there exist real numbers K > 0 and M > 0 such that the
condition (2.2) is satisfied. []

From this result it is easy to obtain the desired r e p r e s e n t a t i o n

of y(t) in the form of a V o l t e r r a series expansion.

(2.9) Th6orem. Suppose the inputs u I ..... u m of the control system (1.1)
satisfy the c o n s t r a i n t (2.3). If T is s u f f i c i e n t l y small, then for all
0 ~ t ~ T the output yj(t) of the system (1.1) may be e x p a n d e d in the
 94

form of a Volterra series, with kernels (2.8), where Qt(x) and P~(x)
are as in (2.6)-(2.7) and ~ = hj. []

This result may be proved either directly, by showing that the

Volterra series in question satisfies the equations (1.1), or indirec-
tly, after establishing a correspondence between the functional ex-
pansion described at the beginning of the previous section and the
Volterra series expansion. We take the second way.
For, observe that for all (ik...i I) the kernel w~...i1(t~Tk,...,~ I)
is analytic in a neighborhood of the origin, and consider the Taylor
series expansion of this kernel as a function of the variables
t-Tk,Tk-~k_1,...,T2-z1,~1. This expansion has clearly the form

nk nI n0
n0...n k (t-T k) ...(T2--T 1 ) T1
= Z C,
Wlk---il(t'tk'''''~l) n0 . .nk=0 Xk'''il nk! . ..nl,n
. . 0,

where

0
n0-..n k = Jl n +...+n k Wlk'''i I
ci k. . .i I nk nI n0
9(t-T k) . ..8(T2--T I) 8T I
t-xk=...=T2-~1=~1=0

If we substitute this expression in the convolution integral asso-

ciated with w. , we obtain an integral of the form
z k . .i I
nk nI n0
~ n 0 . .n k ('r2-T 1 ) "r1
Z C, . .
n0 .nk=0 Xk 11 nk! n1! ul I , n 0,
" " 0 0 0

The integral which appears in this expression is actually an

iterated integral of ul,...,u m , and precisely the integral
t
nk nO
(2.10) I(d~0) d~ik... (d~0)nld~il (d~ 0)
0

(where (dg0)n stands for n-times d~ 0) .

Thus, the expansion (2.4) may be replaced with the expansion

y(t) = ~ e0 (d~O)n
n=O
(2.11) 0
= t
n0""nk I ( d ~ 0 ) ~ d ~ " nl n
c. . . . (d(0) d~il (d(0) 0
+k-Z-1il...!k=l n0!..nk=0 lk'''11 0
 95

which is clearly an expansion of the form (1,6). Of course, one could

rearrange the terms and e s t a b l i s h a c o r r e s p o n d a n c e between the coef-
n n0"''nk
ficients c0,c i . (i.e. the values of the derivatives of w 0 and
k" "il
Wik...i I at t-Tk=...=T2-TI=TI=0) and the c o e f f i c i e n t s c ( ~ ) , c ( ~ . . . i 0)

of the expansion (1.6), but this is not needed at this point.

On the basis of these c o n s i d e r a t i o n s it is very easy to find
Taylor series expansions of the kernels which characterize the V o l t e r r a
series expansion of yj(t). We see from (2.11) that the c o e f f i c i e n t
n0...n k
eik...i I of the Taylor series e x p a n s i o n of Wik...i I coincides with

the coefficient of the iterated integral (2.10) in an e x p a n s i o n (1.6),

but we know also from (1.13), that the c o e f f i c i e n t of the iterated
integral (2.10) has the form

nO nI nk_ I nk
Lf LgiILf ...Lf LgikL f hj(x )

This makes it possible to write down i m m e d i a t e l y the expressions of

the Taylor series expansions of all the kernels which c h a r a c t e r i z e
the Volterra series expansion of y4 (t).
J
Oo

(2.1 2a) W 0(t) ~ Lfhj X O" tn

= ( ~HY'
n=0
nI nO
no n. (t-TI ) TI
(2.12b)
nl=0 n0=0 nl '- n0!
n2 n I no
~ ~ n n4 n~ (t-T2) (T2-TI) T1
(2.12c)
wi2iI (t,.~.,,-c~)=
L , n2=0
[ nl[O
=
}~ L~OL
n0=0 z
L'L
gil z
L~.(x3 )
gi2 z n2!n1!n0!

and so on.
The last step needed in order to prove Theorem (2.9) is to show
that the Taylor series expansions of the kernels (2.8), with Qt(x)
and Pt(x) defined as in (2.6), (2.7) for I = h i (t) coincide with the
expansions (2.1 2) .
This is only a routine computation, which may be carried out with
a little e f f o r t by keeping in m i n d the w e l l - k n o w n C a m p b e l l - B a k e r - H a u s -
dorff formula, which provides a Taylor series expansion of Pti (x) " A c -
cording to this formula it is possible to expand Pt(x) in the follow-
ing way
~n , ,tn
Pt(x) = (#ft), gio(~t(x)) = ~ aafgi~x'~.
n=0
 96

where, as usual, ad~g ~ [ f,ad~-Ig] and ad~g = g.

(2.13) ExampZe. In the case of bilinear systems, the flow ~[ may be

clearly given the following closed form expression

f(x) = (exp At)x

Ct

From this it is easy to find the expressions of the kernels of the

Volterra series expansion of Yi(t). In this case

Qt(x) = cj (exp At)x

P~(x) = (exp(-At))Ni(ex p At)x

and, therefore,

w0(t) = cj (exp At)x

wi(t,T I) = Cj (exp A(t-T1))Ni(ex p ATI)X

wz2il (t,T2,TI) =Cj (exp A(t-T2))Ni2(exp A(T2-TI))Nilexp(ATI)X

and so on.

3. Output Invariance

In this section we want to find the conditions under which the

output is not affected by the input. These conditions will be used
later on in the next chapter when dealing with the disturbance de-
coupling or with the noninteracting control.
Consider again a system of the form

m
= f{x) + [ gi(x)ui
i=I

yj = hi(x) (j = 1 ..... ~ )

and let

yj(t;x;u I ..... u m)

denote the value at time t of the j-th output, corresponding to an

 97

initial state x and to a set of input functions ul,...,u m. We say

that the output yj is unaf~ec%ed by (or ~nua~an~ under) the input ui,
if for every initial state x'6 N , f o r every set of input functions
ul,...,ui_1,ui+1,...,h m , and for all t

(3.1) yj (t;x;ul ..., ui_q,v a ,ui+ I ... ,u m) =

Yj (t;x;ul , . .. ,Ui_1,V b ,Ui+l,...,u m)

for every pair of functions v a and v b.

There is a simple test that identifies the systems having the
output yj u n a f f e c t e d by the input u i-

(3.2) Theorem. The o u t p u t yj is u n a f f e c t e d by the input u i if and

0nly if, for all r ~ I and for any choice of vector fields TI,...,T r
in the set {f'g1' .... gm }

Lgihj(x) = 0

(3.3) LgiLTI --LTrh j (x) = 0

for all x E N.

Proof. Suppose the above condition is satisfied. Then, one easily

sees that the function

(3.4) L I .-.LTrhj(x) = 0

is identically zero w h e n e v e r at least one of the vector fields

TI,...,T r coincides with gi" If we now look, for instance, at the
Fliess e x p a n s i o n of yj(t), we observe that under these c i r c u m s t a n c e s

C(ik...i0) = 0

whenever one of the indexes i0,...,i k is equal to i, and this, in

turn, implies that any iterated integral which involves the input
function u i is m u l t i p l i e d by a zero factor. Thus, the c o n d i t i o n (3.1)
is satisfied and the o u t p u t yj is d e c o u p l e d from the input u i-
Conversely, suppose the condition (3.1) is satisfied, for every
x 6 N, for every set of inputs u l , . . . , u i _ 1 , u i + 1 , . . . , u m and e v e r y
pair of functions v a and v b. Take in p a r t i c u l a r va(t) = 0 for all t.
Then in the Fliess expansion of y j ( t ; x ; u l , . . . , u i _ 1 , v a , u i + 1 , .... u m)
 98

an iterated integral of the form

I d~ik'''d~i0
0

will be zero whenever one of the indexes i0,...,i k is equal to i. All

other iterated integrals of this expansion (i.e. the ones in which
none of the indexes i0,...,i k is equal to i) will be equal to the cor-
responding iterated integrals in the expansion of yj(t;x;ul,...,ui_1,
b
v ,ui+1,...,u m) because the inputs U l , . . . , U i _ l , U i + l , . . . , u m are the
same. Therefore, we deduce that the difference between the right-hand-
side and left-hand-side of (3.1) is a series of the form

k=0
~ T c(ik'''i0 )I d~ik'''d~i0
i0'''''ik=0 0

in which the only nonzero coefficients are those with at least one of
the indexes i0,..,i k equal to i. The sum of this series is zero
b
for every input U l , . . . , u i _ 1 , v ,ui+1,...,u m. Therefore, according to
Lemma (1.7), all its coefficients must vanish, for all x 6 N. We con-
clude that (3.4) and, accordingly, (3.3) are satisfied for all x 6 N . []

The condition (3.3) can be given other formulations, in geometric

terms. Remember that, in section I, we have already observed that the
coefficients of the Fliess expansion of y(t) coincide with the values
at x of functions that span the observation space 0. The different-
ials of these functions span, by definition, the codistribution

~0 = sp(dX:X 6 0}

If we fix our attention only on the j-th output, we may in particular

define an observation space 0j as the smallest subspace of Ca(N) which
contains the function h and is closed under differentiation along
3
f,gl,...,g m. Therefore, the set of differentials dhj,dLg. ...Lg hj(x)
with ik,...,i 0 6 I and j fixed spans the codistribution 10 ik

~0A = sp{dX:X e 0j}

3
Now, observe that the condition (3.3) can be written as

(dhj,g i ) (x ~ = 0
 99

(dL ...Lg hj,g i )(x) = 0

gi k . i0

for all k ~ 0 and for all ik,...,i 0 6 I. From the above discussion we
conclude that the condition stated in Theorem (3.2) is equivalent to
the condition

(3.5) gi 6 n 0
3

Other formulations are possible. For, remember that we have shown

in section II.3 that the distribution ~0 is invariant under
the vector

fields f,gl,...,g m. For the same reasons, also the distribution ~0. is
invariant under f,gl,...,g m. 3
Now, let (f,gl,...,gmlsp{gi}) denote, as usual, the smallest di-
stribution invariant under f,gl,...,g m which contains sp{gi}. If (3.5)
is true, then, since ~0. is invariant under f,gl,...,g m , we must have
3
l
(3.6) (f'gl ..... gm}sp{gi } ) C ~0.
3

Moreover, since

~ C sp{dhj} i
3
we see also that if (3.6) is true, we must have

(3.7) (f'gl ..... gmlSp{gi } ) C (sp{dhj}) l

Thus, we have seen that (3.5) implies (3.6) and this, in turn,
implies (3.7). We will show now that (3.7) implies (3.5) thus proving
that the three conditions are in fact equivalent.
For, observe that any vector field of the form [T,g i] with
T 6 {f,gl,...,gm} is by definition in the left-hand-side of (3.7).
Therefore, if (3.7) is true,

0 = <dhj,[ Y,gi]) = L ~ L g i h j - L g i L T h j

But, again from (3.7), gi 6 (sp{dhj}) l so we can concluHe

LgiLTh j = 0
 100

i.e.

gi (sp{dLThj })

By i t e r a t i n g this a r g u m e n t it is e a s i l y seen that if Tk,...,T I is

any set of k vector fields b e l o n g i n g to the set {f,gl,...,gm }, then

(3.8) gi (sp{dLTk'''LTIhj})

From the Remark II.(3.4), we know that 0 consists of m - l i n e a r

3
combinations of functions of the form hj or L k .L with
- ~lhj '
Ti E { f'g1'''" ,gm } , I _< i ~ k, I ~ k < ~. Thus, from (3.8) we deduce
that gi annihilates the d i f f e r e n t i a l of any function in 0j , i.e. that
(3.5) is satisfied.
Summing up we may state following result

(3.9) Theorem. The output yj is u n a f f e c t e d by the input u i if and only

if any one of the following (equivalent) conditions is satisfied

i) gi 6 ~ 0
3

ii (f'gl ..... gm]Sp{gi } > C (sp{dhj}) .

iii) (f'gl ..... gmlSp{gi } > C n0.
J

(3.10) Remark. It is clear that the statement of T h e o r e m (3.2) can be

slightly m o d i f i e d (and weakened) by asking that

L h(x) = 0
gi 3

L L ...L h.(x) = 0
gi T1 ~r 3

for all r ~ I and any choice of vector fields ~I' .... T r in the set

{f'g1' .... gi-1'gi+1'''''gm}"

Accordingly, the s t a t e m e n t of T h e o r e m (3.9) could be m o d i f i e d by
taking into consideration, instead of < f,gl,...,gmlsp{gi} ) , the smal-
lest d i s t r i b u t i o n containing gi and invariant under the vector fields
f,gl,...,gi_1,gi+1,.-.,gm . Consistently, instead of 0j , one should
c o n s i d e r the s m a l l e s t subspace of C~(N) c o n t a i n i n g h and closed under
J
d i f f e r e n t i a t i o n along the v e c t o r fields f , g l , . . . , g i _ 1 , g i + 1 , . . . , g m.

(3.11) Remark. Suppose (f'gl ..... g m l s p { g i } ) and ~0j are nonsingular.
 101

Then both d i s t r i b u t i o n s are also involutive (see Lemmas I.(6.6),I.(7.6)

and Remark I I . ( 3 . 7 ) ) . I f the condition (iii) of Theorem (3.9) is sa-
tisfied, then a r o u n d each point x N it is possible to find a coord-
inate n e i g h b o r h o o d U on which the n o n l i n e a r system is locally repre-
sented by e q u a t i o n s of the form

xl = f 1 ( x l , x 2 ) + ~ g k l ( X l , X 2 ) U k + g i ( x l , x 2 ) u i
k=1
k~i
m
x2 = f2(x2 ) + ~ gk2(X2)Uk
k=1
k~i

Yj = hj(x 2)

from which one sees that the input u i has no influence on the output
yj. D

Suppose there is a d i s t r i b u t i o n A which is invariant under the

vector fields f'g1'''''gm ' contains the vector field gi and is con-

rained in the d i s t r i b u t i o n (sp{dhj})
. Then

(f'gl ..... gmlSp{gi } > C A C (sp{dhj})
We conclude from the above inequality that the c o n d i t i o n (ii) of
Theorem (3.9) is satisfied. Conversely, if condition (i) of T h e o r e m
(3.9) is satisfied, we have a distribution, ~0. ' which is invariant
under the vector fields f,gl,...,g m , c o n t a i n s 3 g i and is contained in
(sp{dhj}) . Therefore we may give another d i f f e r e n t and useful formu-
lation to the invariance condition.

(3.12) Theorem. The output yj is u n a f f e c t e d by the input u i if and

only if there exists a d i s t r i b u t i o n ~ with the following p r o p e r t i e s

(i) A is invariant under f,gl,...,g m

(ii) gi A C (sp{dhj}) []

(3.13) Remark. Again the c o n d i t i o n (i) may be w e a k e n e d by simply asking

that

(i') ~ is i n v a r i a n t under f'gl' .... g i - 1 ' g i + 1 ' ' ' ' ' g m

Note that this implies that if there exists a distribution A with the
properties (i') and (ii) there exists another distribution ~ with the
properties (i) and (ii). []

We leave to the reader the task of e x t e n d i n g the previous result

to the situation in which it is required that a specified set of out-
 102

puts Y J I ' ' ' ' ' Y J r has to be u n a f f e c t e d by a given set of inputs

u ,...,u. . The c o n d i t i o n s stated in T h e o r e m (3.2) remain formally

zI zs
the same, w h i l e the ones s t a t e d in T h e o r e m s ( 3 . 9 ) and (3.12) require
appropriate modifications.
In c o n c l u d i n g this section it m a y be w o r t h o b s e r v i n g that in
case the s y s t e m in q u e s t i o n reduces to a linear s y s t e m of the form

m
~ = A x i=[Ibiu Z

yj = cjx j = I ..... Z

then the c o n d i t i o n (3.3) b e c o m e s

cjAkbi = 0 for all k _> 0

The c o n d i t i o n s (i), (ii), (iii) of T h e o r e m (3.9) become respectively

n-1
N ker (cjA k)
bi k=0

n-1 n-1
Im(Akbi ) C N k e r ( e j A k)
k=0 k=0
n-1
im(Akbi ) C ker(cj)
k=0

These c l e a r l y imply and are implied by the e x i s t e n c e of a subspace V

i n v a r i a n t u n d e r A and such that

b i C V C ker(cj) .

4. L e f t - I n v e r t i b i l i t y

In this section we c o n s i d e r the p r o b l e m of finding c o n d i t i o n s

w h i c h ensures that, in a given system, d i f f e r e n t input f u n c t i o n s pro-
duce d i f f e r e n t o u t p u t functions. If this is the case then the input-
o u t p u t map is i n v e r t i b l e from the left and it is p o s s i b l e to recon-
struct u n i q u e l y the input a c t i n g on the s y s t e m from the k n o w l e d g e of
the c o r r e s p o n d i n g output. Since, as we know, the i n p u t - o u t p u t m a p of
a nonlinear system d e p e n d s on the initial state x , one has to in-
corporate the d e p e n d e n c e on the initial state into a p r e c i s e defini-
 103

tion of i n v e r t i b i l i t y .
A system is left-invertible at x if w h e n e v e r u a and u b are two
different input functions

y ( t ; x ; u a) ~ y ( t ; x ; u b)

for at l e a s t a value of t > 0.

k

We r e s t r i c t our attention to s y s t e m s with a scalar-valued input

(but p o s s i b l y vector-valued output) because this case can be d e a l t
with r e l a t i v e ease. Thus our system will be d e s c r i b e d by the e q u a t i o n s

= f(x) + g(x)u
(4.1)
yj = hj(x) I < j <

A simple sufficient condition for invertibility at x is the fol-

lowing one.

(4.2) Lemma. The s y s t e m (4.1) is l e f t - i n v e r t i b l e at x O if for some in-

teger k o _> 0 a n d some I _< j _< Z

k
(4.3) L Lfh. (x ) ~ 0
g 3

(4.4) L L';h. (x) = 0

g ~ 3

for all x 6 N a n d for all 0 < k < k

-- o

Proof. Suppose that u a and u b are two different analytic input func-
tions. Then, there exists an i n t e g e r r such that

dru a dru b
(4.5)
(d--~ -) t=o ~ (d-~- ) t:o

NOW, let r d e n o t e the smallest integer such that (4.4) is s a t i s f i e d .

We will show that the (k + r + 1 ) - t h derivatives of Y i ( t ; x ; u a ) and
of Y i ( t ; x ; u b ) with respect to the time t are different at t = 0, so
that we m a y conclude that the two o u t p u t functions, which are analytic,
are d i f f e r e n t .
For, remember that the c o e f f i c i e n t s of the F l i e s s expansion of
y j ( t ; x ; u a) h a v e the e x p r e s s i o n
 104

h ( x )
3

L L ...Lg hj(x )
gJ0 gJl Jk
where, in this case, 0 ~ j 0 ..... Jk ~ I and g0 = f' gl = g" From (4.4)
we have that the only possibly nonzero coefficients in the series are
those in which:

- either J0 = "'" = Jk = 0

- or k > k and J k-k +I = ... = jk = 0

o

These coefficients multiply iterated integrals which either do

not contain the input function, or have the form

Id~ 0...d~0d~jk_k ...d~j0

0 o
~r-~

ko-times

Let's now take the k-th derivatives of the function yj(t;x;u)

with respect to t and evaluate them at t = 0. It is clear from the
structure of the iterated integrals that only those terms of Fliess
series whose index has a lenght smaller than or equal k will contri-
bute, because all terms whose index has a lenght greater than k vanish
at t = 0. Thus we have

ko+ro+ 1
d yj =
k +r +1
dt o o t=O

ko i k Iro1!t ]
= ~ [ L ...Lg Lfhj(x ) d~jk'"d~j0 0
k=0 J0 ..... Jk =0 gJ0 Jk IdtrO+1 t=

At this point, we observe that

Idrol!tI Id t1
r+------~
dt d~1 t=0 [ dt o t=0

and that all other (ro+1)-th derivatives of the iterated in-

 105

tegrals d e p e n d only on u(0),u(0), up to the (ro-1)-th d e r i v a t i v e of

u(t) at t = 0.
Therefore, since
dku a dku b
(d--~-) t=0 = (dt
~ ) t=0

for all 0 < k < r -I, we conclude

-- -- O

k +r +I k +r +I
d o . o b, I

dtko+ro+ I
it=0 dtko+ro+l Jt= 0
k r Idrou b] ]
= LgLfhj ( x ) I I ~ ] - m ~ 0
[fat jt--0 [dt it=0J
This c o m p l e t e s the proof. []

The c o n d i t i o n of T h e o r e m (4.3) may fail to be n e c e s s a r y for left

o
invertibility at a given x , but it h a p p e n s to be n e c e s s a r y and suf-
ficient for a s t r o n g e r n o t i o n of i n v e r t i b i l i t y . For, suppose there
exists an integer k such that the c o n d i t i o n s (4.3) and (4.4) are sa-
tisfied for some x . Then, there e x i s t s a n e i g h b o r h o o d U of x such
that
k
LgLfhj (~) ~ 0

for all x 6 U and this t o g e t h e r w i t h (4.4) implies - a c c o r d i n g to our

previous t h e o r e m - that the system is left i n v e r t i b l e at all points
of U. Conversely, suppose we c a n n o t find an integer k such that
(4.3) is s a t i s f i e d for some x . This implies that

LgLkhj (x) = 0

for all k > 0 and for all I < j < . This in turn implies that all
the c o e f f i c i e n t s of Fliess e x p a n s i o n of y(t) vanish but the ones in
which o n l y d i f f e r e n t i a t i o n s along the v e c t o r field f occurr. Under
these c i T c u m s t a n c e s we have

= k
y(t) =k~0Lkh(x)~.l
" =

and there is no x for w h i c h the s y s t e m is l e f t - i n v e r t i b l e .

Thus, we ~may state the f o l l o w i n g r e s u l t
 106

(4.6) Theorem. There exists an open subset U of N w i t h the p r o p e r t y

that the s y s t e m is left invertible at all p o i n t s x of U if and only
if there exists an i n t e g e r k > 0 such that

L g L ~ h j ( x ) ~ 0

for some x 6 N and some I < j < Z. []

Of course, the s y s t e m b e i n g analytic, if the c o n d i t i o n (4.3) is
satisfied at some x , then it is s a t i s f i e d on an open subset U of N
which contains x O and is dense in N. T h e r e f o r e the e x i s t e n c e of an
integer k such that the c o n d i t i o n (4.3) is s a t i s f i e d for some x E N
and some I ~ j ~ Z is a c t u a l l y necessary and s u f f i c i e n t for the ex-
istence of an open subset U dense in N w i t h the p r o p e r t y that the
system is i n v e r t i b l e at all x E U.

5. R e a l i z a t i o n Theory

The p r o b l e m of "realizing" a given input-output behavior is

generally known as the p r o b l e m of f i n d i n g a dynamical system with
inputs and o u t p u t s able to reproduce, when initialized in a suitable
state, the g i v e n input-output behavior. The dynamical system is thus
said to "realize", from the c h o s e n initial state, the p r e s c r i b e d
input-output map.
Usually, the s e a r c h for d y n a m i c a l systems which realize the
input-output map is r e s t r i c t e d to special classes in the u n i v e r s e of
all dynamical systems,depending on the structure and/or properties
of the g i v e n input-output map. For example, when this m a p m a y be re-
presented as a c o n v o l u t i o n integral of the form
t
y(t) = Iw(t--T)U(T)dT
0

where w is a p r e s c r i b e d function of t d e f i n e d for t > 0, then one

usually looks for a linear dynamical system

= Ax + Bu

y=Cx

able to reproduce, when initialized in x = 0, the g i v e n behavior.

For this to be true, the m a t r i c e s A , B , C m u s t be such that

C exp(At)B = w(t)
 107

We w i l l now describe the fundamentals of the r e a l i z a t i o n theory

for the (rather general) class of i n p u t - o u t p u t maps which can be re-
presented like functionals of the form (1.6): In v i e w of the r e s u l t s
of the p r e v i o u s sections, the search for "realizations" of this kind
of maps w i l l be r e s t r i c t e d to the c l a s s of d y n a m i c a l system of the
form (1.1).
From a formal point of view, the p r o b l e m is s t a t e d in the fol-
lowing way. Given a formal power series in m+1 noncommutative in-
determinates w i t h c o e f f i c i e n t s in ~ , f i n d an i n t e g e r n, an e l e m e n t
o
x of lqn , m+1 a n a l y t i c v e c t o r f i e l d s g 0 , g I .... 'gm a n d an a n a l y t i c
-vector valued function h defined on a n e i g h b o r h o o d U of x s u c h that

h ( x O) = c(~)

L L ...L h(x ) = c ( i k . . . i l i 0)
gi 0 gi 1 gi k
If t h e s e conditions are satisfied, then it is c l e a r that the
dynamical system

m
x = g0(x) + ~ gi(x)ui
i=I

y = h (x)

initialized in x O 6 ~ p r o d u c e s an i n p u t - o u t p u t behavior of the form

y(t) = c(~) + ~ m~ c(Jk...j0) Id ~ k . . . d ~ 0

k=0 j0...Jk=0 0

In v i e w of this, the set {g0,g I ..... g m , h , x } will be c a l l e d a

realization of the f o r m a l power series c.
In o r d e r to p r e s e n t the b a s i c results of the realization theory,
we n e e d first to d e v e l o p some notations and d e s c r i b e some simple al-
gebraic concepts related to the formal power series. In v i e w of the
need of d e a l i n g with sets of s e r i e s and d e f i n i n g certain operations on
these sets it is u s e f u l to r e p r e s e n t each series as a formal infinite
sum of "m o n o m l "a l s " . Let z 0 , z I ,...,z m d e n o t e a set of m+1 abstract non
commutative indeterminates and let Z = {z0,zl,...,Zm}. With each multi-
index (ik...i 0) we associate the m o n o m i a l (Zik.,.zi0) and we r e p r e s e n t
the s e r i e s in the form

m
(5.1) c = c(~) + [ X C (ik. i~) Z . . . . Z
"" v ik i0
k=0 i0...ik=0
 108

The set of all p o w e r series in m+1 n o n c o m m u t a t i v e indeterminates

(or, in other words, in the n o n c o m m u t a t i v e indeterminates z 0, .... zm)
and c o e f f i c i e n t s in ~ Z is d e n o t e d with the symbol RZ({ Z )}. A special
subset of ~ < ( Z }} is the set of all those series in w h i c h the number
of n o n z e r o c o e f f i c i e n t s (i.e. the n u m b e r of n o n z e r o terms in the sum
(5.1)) is finite. A series of this type is a p o l y n o m i a l in m+1 non-
commutative indeterminates and the set of all such p o l y n o m i a l s is
d e n o t e d w i t h the symbol ~ Z < Z ) . In p a r t i c u l a r ~ ( Z }is the set of all
p o l y n o m i a l s in the m+1 n o n c o m m u t a t i v e indeterminates z0,...,z m and
coefficients in R .
An e l e m e n t of ~ ( Z ) m a y be r e p r e s e n t e d in the form

d m
(5.2) p = p(@) + ~ [ P(ik...i0)zik...zi0
k=0 i 0 . . . i k = 0

where d is an integer w h i c h d e p e n d s on p and p ( @ ) , P ( i k . . . i 0) are real

numbers.
The sets ~ ( Z > and RZ<(Z>) may be given d i f f e r e n t a l g e b r a i c struc-
tures. They can c l e a r l y be r e g a r d e d as ~ - v e c t o r spaces, by letting
R-linear c o m b i n a t i o n s of p o l y n o m i a l s a n d / o r series be d e f i n e d coef-
ficient-wise. The set ~< Z > m a y also be given a ring structure, by
l e t t i n g the o p e r a t i o n of sum of p o l y n o m i a l s be d e f i n e d c o e f f i c i e n t -
wise (with the n e u t r a l e l e m e n t given by the p o l y n o m i a l whose coeffi-
cients are all zero) and the o p e r a t i o n of p r o d u c t of p o l y n o m i a l s de-
fined t h r o u g h the c u s t o m a r y p r o d u c t of the c o r r e s p o n d i n g r e p r e s e n t a -
tions (5.2) (in w h i c h case the n e u t r a l e l e m e n t is the p o l y n o m i a l whose
coefficients are all zeros but p(@) w h i c h is equal to I). Later-on, in
the p r o o f of T h e o r e m (5.8), we shall also e n d o w ~ ( Z ) and ~ i < ( Z )) with
s t r u c t u r e s of m o d u l e s o v e r the ring ~ < Z } but, for the moment, those
additional s t r u c t u r e s are not required.
What is i m p o r t a n t at this p o i n t is to k n o w that the s e t ~< Z ) can
also be given a structure of a Lie algebra, by taking the a b o v e - m e n -
tioned R - v e c t o r space s t r u c t u r e and d e f i n i n g a Lie b r a c k e t of two poly-
n o m i a l s p 1 , p 2 by s e t t i n g ~pl,P2| = p2Pl - plP2. The s m a l l e s t sub-
a l g e b r a of ~ < Z ) w h i c h c o n t a i n s the m o n o m i a l s z 0, ....,;zm will be de-
n o t e d by i(Z) . Clearly, i(Z) may be v i e w e d as a subspace of the ~ -
v e c t o r space ~ ( Z ) , w h i c h contains z0,...,z m and is c l o s e d u n d e r Lie
bracketing with z0,...,z m. Actually, it is not d i f f i c u l t to see that
L(Z) is the s m a l l e s t subspace of ~ ( Z > w h i c h has these p r o p e r t i e s .
N o w we return to the p r o b l e m of r e a l i z i n g an i n p u t - o u t p u t m a p
r e p r e s e n t e d by a f u n c t i o n a l ~f the form (1.6). As expected, the ex-
 109

istence of r e a l i z a t i o n s w i l l be c h a r a c t e r i z e d as a p r o p e r t y of the
formal p o w e r series w h i c h s p e c i f i e s the functional. We a s s o c i a t e w i t h
the formal p o w e r series c two integers, w h i c h will be called, fol-
lowing Fliess, the Hanke~ rank and the Lie rank of c. This is done
in the f o l l o w i n g manner. We use the given formal p o w e r series c to
define a m a p p i n g

F c : ~ ( Z ) ~ ~ Z < < Z ))

in the f o l l o w i n g way:

a) the image under F c of any p o l y n o m i a l in the set Z =

= {zj ...zj 6 ~ ( Z ): (jk...jo) ~ I } (by definition, the polyno-
mial k
a s s o c i a0t e d w i t h the m u l t i i n d e x ~ 6 I* will be the p o l y n o m i a l
in w h i c h all coef/_%cients are zero but p(@) w h i c h is equal to I,
i.e. the u n i t of ~ ( Z )) is a formal power series d e f i n e d by s e t t i n g

[Fc(Z3k
. . . . Zjo)] (ir-.-i 0) = c(i r...i 0 Jk...jo )

for all Jk'''JO 6 I .

b) the map F c is an E - v e c t o r space m o r p h i s m of ~( z ) into ~Z({ Z )) .

Note that any p o l y n o m i a l in ~ ( Z ) may be e x p r e s s e d as an k - l i n e a r

combination Qf e l e m e n t s of Z and, therefore, the p r e s c r i p t i o n s (a)
and (b) c o m p l e t e l y specify the m a p p i n g F c-
L o o k i n g at F c as a m o r p h i s m of ~ - v e c t o r spaces, we define the
Hankel rank PH(C) of c as the rank of F c , i.e. the d i m e n s i o n of the
subspace

F c~(Z )) C ~R(( Z ))

Moreover, we define the Lie rank PL(C) of c as the d i m e n s i o n of

the subspace

Fc(L(Z)) Cm~<{Z ))

i.~. the rank of the m a p p i n g FclL(Z) .

(5.3) Remark. It is easy to get a m a t r i x r e p r e s e n t a t i o n of the m a p p i n g

F c. For, suppose we r e p r e s e n t an e l e m e n t p of ~ ( Z ) w i t h an infinite
column v e c t o r of real n u m b e r s w h o s e e n t r i e s are i n d e x e d by the e l e m e n t s
o I and the e n t r y indexed by jk...jo is e x a c t l y p ( j k . . . j o ) . Of course,
p beiag a polynomial, only f i n i t e l y many e l e m e n t s of this v e c t o r are
 11o

nonzero. In t h e s a m e way, we may represent an e l e m e n t c o f ~Z<< Z ))

with an infinite column vector whose entries are Z-vectors of real
numbers, indexed by the e l e m e n t s of I and such that the entry in-
dexed by i r . . . i 0 is C ( i r . . . i 0 ) . Then, any ~-vector space morphism de-
f i n e d o n ~( Z ) w i t h values in ~Z<(Z )) will be represented b y a n in-
finite matrix, whose columns are indexed by elements of I and in
which each block of Z rows is a g a i n indexed by elements of I . In
particular, the m a p p i n g Fc will be r e p r e s e n t e d by a matrix, denoted
Hc , in w h i c h the b l o c k of Z rows of index ( i r . . . i 0) o n t h e c o l u m n of
index (jk...j0) is e x a c t l y the c o e f f i c i e n t

C ( i r . . . i 0 J k . . . J 0)

o f c. W e leave to the r e a d e r the e l e m e n t a r y check of this statement.

The matrix H c is c a l l e d the Hankel matrix o f the series c. It
is c l e a r from the above definitions that the r a n k of the m a t r i x H
c
coincides with the Hankel rank of F . []
c
(5.4) Example. If the set I consists of only one element, then it is
easily seen that I can be identified with the s e t Z + of t h e n o n -
negative integers numbers. A formal power series in one indeterminate
with coefficients in ~ , i.e. a mapping

c : Z + ~I~

may be represented, like in (5.1), as a n i n f i n i t e sum

os

C ---- ~ ckzk
k=0

and the Hankel matrix associated with the mapping F c coincides with
the c l a s s i c a l Hankel matrix associated with the sequence c0,cl...

cO cI c2

cI c2 c3
S ~-
c
c2 c3 c4 O O I

I Q D
[]

The importance of the H a n k e l and Lie ranks o f the m a p p i n g F

c
depends on the following basic results.

(5.5) Lemma. L e t f , g l , . . . , g m , h and a point x E A n be g i v e n . Let ~C

 111

be the distribution associated with the control Lie algebra C and ~0

the codistribution associated with the observation space 0. Let K(x )
denote the subset of vectors of ~C (x) which annihilate ~0(x ) i.e.
the subspace of T oI~n defined by
x

K(x ) = ~c(X ) n ~i(x) ={v E ~c(X): <dl(x),v }= 0 Vl 0}

Finally, let c be the formal power series defined by

(5.6a) c(~) = h(x )

(5.6b) c(ik...i 0) = L ...Lg h(x )

gi 0 ik
with g0 = f. Then the Lie rank of c has the value

ac(X )
PL(C) = dim Ac(X)-dim K(x ) = dim
ac(x ) n ~0i(x)

Proof. Define a m o r p h i s m of Lie algebras

: L(Z) -~ V~R n)

by setting

~(zi) = gi 0 < i < m

Then, it is easy to check that if p is a polynomial in L(Z) the

{ik...i0)-th coefficient of Fc(p) is L (p)Lgi0...Lg ik h(x). Thus, the

series F c(p) has the expression

m
Fc(p) = L (p) h(x) ...L h(x)z . . . . z.
+k~ [ L (p)Lgi0 gi k ik 10
0 io,...,ik=O ~

If we let v denote the value of the vector field ~(p) at x , the above
can be rewritten as

m
FC(p) = ( d h ( x ) , v )+ (dL ...L h(x ) ,v)z . . . . z.

k=0 i0,...,ik=0 gi 0 gi k zk 10

When p ranges over i(Z), the tangent vector v takes any value in
Ac(x). Moreover, the covectors dh(x ) . . dL ...Lg h(x),...
, . , gi0 ik
 112

span ~0(x). This implies that the n u m b e r of A - l i n e a r l y independent

power series in F c ( A ( Z ) ) is e x a c t l y equal to

o
d i m A C(x ) - d i m A C ( x ) N ~0(x )

and this, in v i e w of the definition of Lie rank of F c , proves the

claim. []

We immediately see from this that if an i n p u t - o u t p u t functional

of the form (1.6) is r e a l i z e d by a dynamical system of d i m e n s i o n n,
then necessarily the Lie rank of the formal power series which specify
the functional is b o u n d e d by n. In o t h e r words, the finiteness of the
Lie rank OL(C) is a n e c e s s a r y condition for the existence of finite-
dimensional realizations. We shall see l a t e r on that this condition is
also sufficient. For the m o m e n t , we w i s h to i n v e s t i g a t e the role of
the finiteness of the other rank associated with F c i.e. the H a n k e l
rank. It c o m e s from the definition that

PL(C) < PH(C)

so the H a n k e l rank may be infinite when the L i e rank is finite. How-

ever, there are special cases in w h i c h PH(C) is finite.

(5.7) Lemma. Suppose f'g1' .... g m , h are linear in x, i.e. that

f(x) = ~x, gl (x) = N I X ..... gm(X) = NmX , h(x) = Cx

for suitable matrices A , N I , . . . , N m , C. Let x O be a p o i n t of A n . Let V

denote the smallest subspace of A n w h i c h contains x and is i n v a r i a n t
under A , N I , . . . , N m. Let W denote the largest subspace of A n w h i c h is
contained in ker(C) and is i n v a r i a n t under A , N I , . . . , N m. The Hankel
rank of the formal power series (5.7) has the value

V
p_(c) = dim V-dim W N V = dim
W O----V

Proof. We h a v e already seen, in section II.4, that the subspace W may

be e x p r e s s e d in the following way

m
W = (ker C) N | n N k e r ( C N i r . ..Ni0 ~
r=0 i 0...ir=0

with N O = A. With the same kind of a r g u m e n t s one proves that the sub-
space V may be e x p r e s s e d as
 113

V =
span{xO] + ~, m[ s p a n { N j k - - - N j x O}
k=0 j 0 . . . J k = 0 0

In the p r e s e n t case the Hankel m a t r i x of F c is such that the block of

rows of index (ir...i 0) on the c o l u m n of index (jk...j0), i.e. the
coefficient C ( i r . . . i 0 J k . . . j 0 ) of c has the e x p r e s s i o n

o
CN . . . . N N . . . . N. x
ir 10 ]k 30

By factoring out this e x p r e s s i o n in the form

(CNir'''Ni0) (Njk'''Nj0x)

it is seen that the Hankel m a t r i x can be f a c t o r e d out as the p r o d u c t

of two matrices, of which the one on the l e f t - h a n d - s i d e has a kernel
equal to the subspace W, while the one on the r i g h t - h a n d - s i d e has an
image equal to the subspace V. F r o m this the c l a i m e d r e s u l t follows
immediately. []

Thus, it is seen from this Lemma that if an input o u t p u t func-

tional of the form (1.3) is r e a l i z e d by a d y n a m i c a l system of d i m e n s i o n
n described by e q u a t i o n s of the form

m
= Ax + ~ N xu
i=I x m

y = Cx

i.e. by a b i l i n e a r d y n a m i c a l system of d i m e n s i o n n, then the Hankel

rank of the formal p o w e r series w h i c h s p e c i f i e s the functional is
bounded by n. The f i n i t e n e s s of the Hankel rank PH(C) is a n e c e s s a r y
condition for the e x i s t e n c e of b i l i n e a r r e a l i z a t i o n s .
We turn n o w to the p r o b l e m of showing the s u f f i c i e n c y of the above
two conditions. We treat first the case of b i l i n e a r r e a l i z a t i o n s , which
is simpler. In analogy w i t h the d e f i n i t i o n given at the b e g i n n i n g of
the section, we say that the set {N O ; N I r o" .,Nm,C r x O} # where x O 6 ~ n #

N E ~nn for 0 < i < m and C C ~ n is a bilinear realization of the

formal p o w e r series c if the set { g 0 , g l , . . . , g m , h , x } d e f i n e d by

g0(x) = N0x, g1(x) = NI(X) ..... gm(X) = NmX

h (x) = Cx
 114

is a r e a l i z a t i o n of c.

(5.8) Theorem. Let c be a formal power series in m+1 n o n c o m m u t a t i v e

i n d e t e r m i n a t e s and c o e f f i c i e n t s in ~. There exists a b i l i n e a r realiza-
tion of c if and only if the Hankel rank of c is finite.

Proof. We n e e d only to prove the "if" part. For, c o n s i d e r again the

m a p p i n g F c. The sets R< Z ) and ~ Z ( < Z )} will now be e n d o w e d with
structures of modules. The ring ~( Z ) is r e g a r d e d as a m o d u l e over
itself. ~i(( Z )} is given an ~ ( Z } - m o d u l e s t r u c t u r e by letting the
o p e r a t i o n of sum of p o w e r series be d e f i n e d c o e f f i c i e n t - w i s e and the
p r o d u c t p.s of a p o l y n o m i a l p ~ ~ ( Z ) by a series s E ~ Z { < Z )} be
d e f i n e d in the f o l l o w i n g way

a) 1.s = s

b) for all 0 < i < m the series z.'s is given by

(zi's) (i r...i 0) = s(i r...i0i)

c) for all p1,p 2 6 ~ < Z } and e1,~2 6

(elp1+e2P2).s = el (P1"S)+e2(P2 "s)

Note that from (a) and (b) we have that for all __ik-.-J0 @ I

(Zjk .. "z30
-s) (ir...i 0) = s(i r...i0jk...j 0)

Note also that since the ring ~< Z } is not commutative, the order in
which the p r o d u c t s are p e r f o r m e d is essential.
We leave to the reader the simple proof that the m a p F c previously
d e f i n e d becomes an ~( Z }-module m o r p h i s m when ~ ( ( Z ) } is e n d o w e d with
this k i n d of I~(Z > - m o d u l e structure. As a m a t t e r of fact, it is trivial
to check that F c(p) = p.c.
Now, c o n s i d e r the c a n o n i c a l f a c t o r i z a t i o n of F c
F

~<Z }
Ker F
c

in which, as usual, Pc d e n o t e s the c a n o n i c a l p r o j e c t i o n p ~--+(p+ker F c)

and Qc the i n j e c t i o n (p+ker F c ) ~ F c(p). Pc and Qc are l~-vector space
morphisms, but there is also a c a n o n i c a l I~( Z } - m o d u l e s t r u c t u r e on
 115

]9( Z )/ k e r F c which makes Pc and Qc ]R( Z ) - m o d u l e morphisms.

Since, by d e f i n i t i o n , ~R< Z ) / k e r F c is i s o m o r p h i c to the image
of F c , we h a v e that the dimension of 1R( Z ) / k e r F c as an lR-vector
space is e q u a l to the Hankel rank PH(C) of the formal power series c.
Let, for s i m p l i c i t y , denote

m(Z >
X =
ker F
c
But X is also an ]R( Z ) - m o d u l e , so to e a c h of the indeterminates
z0,...,z m we m a y associate mappings

Mi:X~X

X ~--+Zi'X

The m a p p i n g s M i are c l e a r l y JR-vector space morphisms. We also

define an ]R-vector space morphism

H : X~]R

by taking

H X = [Qc(X)I (@)

With the n o t a t i o n on the right-hand-side we m e a n the c o e f f i c i e n t with

empty i n d e x in the series Qc(X).
Finally, l e t x be the e l e m e n t of X

o
x = P (I)
C

where I is the u n i t polynomial i n ]~( Z > .

We claim that

(5.9a) c(~) = Hx O

o
(5.9b) c(i k...i 0) = H M i k . . . M i 0 x

For, it is seen immediately that

o
(5.10a) c = Fc(1) = QcPc(1) = Qc x
 116

Moreover, suppose that

o
(5.10b) Fc(zik...zi0) = QcMik...Mi0x

t h e n we h a v e

Fc(Z 1 z i.k. . . z.l 0 ) = zi.F c (zi k ... zi 0) = z i ( Q c M i k "''Mi 0 x O)

x O) ...M. x O
Qc(Zi-Mik...Mi0 = QcMiMik 10

for 0 ~ i ~ m. T h u s (5.10b) is true for all (ik...i 0) I .

Now, keeping in m i n d the d e f i n i t i o n of F c , one h a s

[ F c ( Z i k . . . z i 0 ) ] (~) = c ( z i k . . . z i 0 )

and therefore, in v i e w of the d e f i n i t i o n of the m a p p i n g H, (5.9) are

proved.
Take now a basis in the D H ( C ) - d i m e n s i o n a l vector s p a c e X. The
mappings M 0 , . . . , M m a n d H w i l l be r e p r e s e n t e d by m a t r i c e s N0,NI,...,N m
a n d C; x w i l l be r e p r e s e n t e d by a vector ~o. These quantities are such
that
-o
c(i k . . . i 0) = C N i k . . . N i 0 x

for all (ik...i0) E I . This shows t h a t the s e t {C,N0,...,Nm,X} is a

bilinear realization for our series. []

The r e s u l t w h i c h follows presents a necessary and sufficient con-

dition for the e x i s t e n c e of r e a l i z a t i o n s of an i n p u t - o u t p u t functional
of the f o r m (1.6), provided that the c o e f f i c i e n t s of the p o w e r series
which characterize the f u n c t i o n a l are suitably bounded.

(5.11) Theorem. L e t c be a f o r m a l p o w e r series whose coefficients sa-

tisfy the c o n d i t i o n

(k+1)
(5.12) 11c(ik...i0)11 1 <_ C(k+1) !r

for all (ik...i 0) E I , for some p a i r of real n u m b e r s C > 0 a n d r > 0.

Then there exists a realization of c if and o n l y if the Lie rank of c
is finite.

Proof. Some m o r e m a c h i n e r y is required. For each polynomial p 6 ~(Z >

we d e f i n e a mapping S :~((Z >> ~ ~(< Z >> in the f o l l o w i n g w a y
P
 117

a) if p e z % = { Z j k . . . z j 0 e ~< Z ) : (Jk'''J0) I~} then Sp(C) is a

formal p o w e r series defined by setting

[Sz . . . . z. (c)] (ir...i 0) = c ( J k . . . J 0 i r. ..i 0)

3k 30

b) if ~i,e2 E }{ a n d p 1 , P 2 E ~(Z > then

Sp(C) = aiSpl (c) + a 2 S p 2 ( C )

Moreover, suppose that, given a formal power series s I 6 2(< Z ))

and a f o r m a l p o w e r series s 2 6 }{( ( Z ) ), the sum of the n u m e r i c a l series

m
(5.13) sl (@)s2(~) + ~ ~ sl ( i k . . . i 0 ) s 2 ( i k . . . i 0)
k=0 i0,...,ik=0

exists. If this is the case, the s u m of this s e r i e s w i l l be d e n o t e d

by (Sl,S 2 ) .
We n o w t u r n o u r a t t e n t i o n to the p r o b l e m of finding a realization
of c. In o r d e r to s i m p l i f y the n o t a t i o n , we a s s u m e = I (i.e. we c o n -
sider the c a s e of a s i n g l e - o u t p u t system). By d e f i n i t i o n , there e x i s t
n polynomials in L(Z), denoted pl,...,pn, with the p r o p e r t y t h a t the
formal p o w e r series F c(pl ) ' ' ' ' ' F c ( p n ) are } { - l i n e a r l y independent.
W i t h the p o l y n o m i a l s PI'"" ''Pn we a s s o c i a t e a formal power series

n ~ ~
cs14) w = exp xiP i) = 1 xiPi Ik
i=1 1 " i=I

where X l , . . . , x n are r e a l v a r i a b l e s .
The s e r i e s c which is to be r e a l i z e d a n d the s e r i e s w thus de-
fined are u s e d in o r d e r to c o n s t r u c t a set of a n a l y t i c functions of
Xl,.i.,Xn, defined in a n e i g h b o r h o o d of 0 a n d i n d e x e d by the e l e m e n t s
of I , in the f o l l o w i n g way

h(x) = {c,w >

h. (x) = ( S (c),w)
ik...i 0 zx.k. . . . zl 0

The g r o w t h c o n d i t i o n (5.12) guarantees the c o n v e r g e n c e of the

series on the r i g h t - h a n d - s i d e for all x in a n e i g h b o r h o o d of x = 0.
It w i l l be s h o w n n o w t h a t there e x i s t m + I vector fields,
g0(x),...,gm(X), defined in a n e i g h b o r h o o d of 0, w i t h the p r o p e r t y
that
 118

(5.15) Lgihik...i0(x) = hik...i0i(x)

for all (ik...i 0) 6 I . This will be actually enough to prove the

Theorem because, at x = 0, the functions hik...i0(x) by construction
are such that

h(0) = c(g)

hik...i0(0) = c(ik...i0)

and this shows that the set {h,g0,...,g m} together with the initial
state x = 0 is a realization of c.
To find the vector fields g0,...,g m one proceeds as follows.Since
the n series Fc(Pl),...,Fc(Pn) are ~ - l i n e a r independent, it is easily
seen that there exist n monomials ml,..~,m n in the set Z ~ with the
property that the (nxn) matrix of real numbers

[Fc(Pl)] (m I) .-.[Fc(Pn)] (m I)
(5.16)

[Fc(Pl ) ] (ran) "" "[Fc(pn )] (mn)

has rank n. It is easy to see that

[Fc(Pi )] (mj) = (~i (Smj (c),w > )x=0

For, if Pi 6 Z , then by definition

[Fc(Pi)] (mj) = c(mjp i) = [Smj (c)] (pi) = (~i (Smj ( c ) , w ) ) x = 0

From this, using linearity, one concludes that the above expression is
true also in the (general) case where Pi is an ~ - l i n e a r combination
of elements of Z .
Using this property, we conclude that the j-th row of the matrix
(5.16) coincides with the value at 0 of the differential of one of
the functions hik...i 0 , the one whose multiindex corresponds to the

monomial m..
3
Consider now the system of linear equations
 119

{Sm1(C),W > ( Sml Zk (c) ,w >

gk (x) =

V~ <Smn(c)'w ) (Smnzk(c),w)

in the unknown n - v e c t o r gk(x)_. The c o e f f i c i e n t m a t r i x is n o n s i n g u l a r

for all x in a n e i g h b o r h o o d of 0 (because at x = 0 it c o i n c i d e s - as
we have seen - w i t h the m a t r i x (5.16)). Thus, in a n e i g h b o r h o o d of 0
it is p o s s i b l e to find a v e c t o r field gk(x) such that

Lg k (S m i ( c ) , w ) = (Smizk(c),w)

and this p r o v e s that (5.15) can be satisfied, at least for these

hik...i 0 w h o s e m u l t i i n d e x e s c o r r e s p o n d to the m o n o m i a l s m l , . . . , m n.

The proof that (5.15) holds for all the o t h e r f u n c t i o n s hi~...i0(x)

depends on the fact that e v e r y formal p o w e r series in Fc(i(Z)) is an

R - l i n e a r c o m b i n a t i o n of F c ( P l ) , . . . , F c ( P n ) , and is left for the reader.[]

It is seen from the above T h e o r e m that if a formal p o w e r series

c has a finite Lie rank, and its c o e f f i c i e n t s satisfy the growth con-
dition (5.12), then it is p o s s i b l e to find a d y n a m i c a l system of dimen-
sion 6L(C) w h i c h realizes the series.
This fact, t o g e t h e r w i t h the result stated before in Lemma (5.5)
induces to some f u r t h e r remarks. A realization { f , g l , . . . , g m , X } of a
formal power series c is minimal if its dimension, i.e. the d i m e n s i o n
of the u n d e r l y i n g m a n i f o l d on w h i c h f , g l , . . . , g m are defined, is less
then or e q u a l to the d i m e n s i o n of any o t h e r r e a l i z a t i o n of c. Thus,
from L e m m a (5.5) we i m m e d i a t e l y deduce the f o l l o w i n g corollaries.

(5.17) CoroZlary. A r e a l i z a t i o n { f , g l , . . . , g m , X } of a formal power

series c is m i n i m a l if and only if its d i m e n s i o n is equal to the Lie
rank ~L(C).

(5.18) Corollary. A r e a l i z a t i o n {f'g1' .... gm,X } of a formal p o w e r

series c is m i n i m a l if and only if

dim ~c(X O) = dim ~0(x ) = n

or, w h i c h is the same, the r e a l i z a t i o n satisfies the c o n t r o l l a b i l i t y

 120

rank condition and the observability rank condition at x .

6. U nniqueness of M i n i m a l Realizations

In t h i s section we p r o v e an i n t e r e s t i n g uniqueness result, by

showing that any two m i n i m a l realizations of a f o r m a l power series
are locally "diffeomorphic"

(6.1) Theorem. L e t c be a f o r m a l power s e r i e s and let n d e n o t e s its

a a a a~ b b b
Lie rank. Let {g~'g1'''''v g m , n ,x ~ a n d {g0,gl, . . . , g m , h b , x b} be two
minimal, i.e. n-dimensional realizations of c. L e t g~ , 0 < i m,
a n d h a be d e f i n e d on a n e i g h b o r h o o d U a of x a in ~ n and
g~,--0- < i < m ,
and h b be d e f i n e d on a n e i g h b o r h o o d U b of x b in ~ n Then, there exist
open subsets V a C U a and V b C U b a n d a d i f f e o m o r p h i s m F:V a ~ ~ such
that

(6.2) gb(x) = F , g iao F -1(x) 0 ! i ! m

(6.3) hb(x) = h a o F -I (x)

for all x 6 V b.

Proof. W e b r e a k up the p r o o f in s e v e r a l steps.

(i) Recall that a minimal realization {f'gl .... , g m , X O} of c satisfies

the o b s e r v a b i l i t y rank condition at x (Corollary (5.18)). From the
definitions of 0 a n d ~0 ' one deduces that there exist n real-valued
functions 11,...,I n , defined in a n e i g h b o r h o o d U of x O, h a v i n g the
form

li(x) = Lv . . . L v l h j (x)
r

with Vl,...,v r vector fields in the set { f , g l , . . . , g m }, r (possibly)

depending on i and I ~ j ~ ~ such that the c o v e c t o r s d l l ( X ) , . . . , d l n ( X )
are linearly independent (i.e. span the cotangent space T* U). F r o m
o
x
this property, using the inverse function theorem, it is d e d u c e d that
there exists a neighborhood U H C U of x O s u c h that the mapping

H : x ~-~(ll(X) ..... In(X))

is a d i f f e o m o r p h i s m of U H o n t o its image H(UH).

From any two m i n i m a l realizations, labeled "a" and "b", we will
 121

construct two of such m a p p i n g s , denoted H a and respectively H b.

(ii) Let 01 ..... O n be a s e t of v e c t o r fields, defined in a n e i g h b o r -

hood U of x , h a v i n g the f o r m

m
0i = f + [ -i
j=1 gjuj

with u 3 E IR for I < j < m. L e t Cti d e n o t e the f l o w of 81 a n d G d e n o t e

the m a p p i n g

G : (t I ..... tn ) F--+~n ..... ~I (xO)

defined on a n e i g h b o r h o o d (-e,c) n of 0.
F r o m a n y two m i n i m a l realizations,labeled "a" a n d "b" we w i l l
construct two of s u c h m a p p i n g s , denoted G a and G b (the same set of
u~'s b e i n g u s e d in b o t h G a a n d Gb).
3 a a
Recall that a minimal realization {fa,g~,...,gm,X } satisfies the
controllability rank condition at x a (Corollary (5.18)). From the p r o -
perties of 6 C a n d R (see R e m a r k II.(2.7), one deduces t h a t the d i s t r i -
bution R is n o n s i n g u l a r and n - d i m e n s i o n a l a r o u n d x a. Then, using the
same a r g u m e n t s as the o n e s u s e d in the p r o o f of T h e o r e m I. (6.15), it
is p o s s i b l e to see t h a t there exist a choice of u ~ ' s and an o p e n sub-
3
set W of (0,E) n such t h a t the r e s t r i c t i o n of G a to W is a d i f f e o m o r -
phism of W o n t o its image Ga(w).
a a a .a a~
(iii) It is n o t d i f f i c u l t to p r o v e t h a t if {f , g l , . . . , g m , n ,x ~ a n d
{fb, g ~ , . . . , g mb, h b ,x b } are t w o r e a l i z a t i o n s of the s a m e f o r m a l p o w e r
series c, then, for all 0 < t. < e, I < i < n, w i t h sufficiently small
6,

(6.4) H a o G a ( t l , . . . , t n ) = H b o G b ( t 1 , . . . , t n)

AS a m a t t e r of fact, if c is small t h e n G ( t l , . . . , t n) is a p o i n t of
UH , r e a c h e d from x under the p i e c e w i s e constant control defined by

uj (t) = u i for t ~ [ t i % .+t. 1 t1+ " "+ti)

3 "" l-- ' "

Moreover, the v a l u e s of the c o m p o n e n t s of H (i.e. the v a l u e s of f u n c -

tions 1 1 , . . . , I n) at a p o i n t w e r e shown to c o i n c i d e with the v a l u e s
of certain derivatives, at time t = 0, of some c o m p o n e n t s of an
output f u n c t i o n y(t) obtained under suitable piecewise constant con-
trols (see p r o o f of T h e o r e m I. (7.8)). So, one m a y interpret the com-
 122

ponents of H o G ( t l , . . . , t n) as the v a l u e s at time t = t 1 + . . . + t n of

certain derivatives of an o u t p u t function y(t) obtained under suitable
piecewise constant controls.
Two minimal realizations of the same p o w e r series c characterize
two s y s t e m s w h i c h by d e f i n i t i o n display the same input-output beha-
vior. These two s y s t e m s , initialized respectively in x a and x b, under
any p i e c e w i s e constant control produce two i d e n t i c a l output functions.
Thus, the two s i d e s of (6.4) m u s t coincide.

(iv) Recall that, if the r e a l i z a t i o n "a" is m i n i m a l , if (t I, .... tn)6W

and ~ is s u f f i c i e n t l y small, the m a p p i n g H a o G a is a c o m p o s i t i o n of
diffeomorphisms. If a l s o the r e a l i z a t i o n "b" is m i n i m a l , H b is indeed
a diffeomorphism, b u t a l s o G b m u s t be a d i f f e o m o r p h i s m of W o n t o its
image, because of the e q u a l i t y (6.4) a n d of the f a c t t h a t the left-
hand-side is i t s e l f a diffeomorphism The f o l l o w i n g diagram

where V a = Ga(w), V b = Gb(w), V a C U Ha , V b C ~H a n d W = H a o G a ( W ) =

= HboGb(w), is a c o m m u t a t i v e diagram of d i f f e o m o r p h i s m s . Thus, we m a y
define a diffeomorphism

F : Va ~ Vb
as

(6.5a) F = (Hb)-1 oH a

whose i n v e r s e m a y a l s o be e x p r e s s e d as

(6.5b) F -I = G a o ( G b) -I

(v) By m e a n s of the same a r g u m e n t s as the o n e s a l r e a d y used in (iii)

one m a y e a s i l y p r o v e a more general version of (6.4). M o r e p r e c i s e l y ,
setting

m givia 0b fb ~ b
8a = fa + i=I
~ = +i I givi

one m a y d e d u c e that, for s u f f i c i e n t l y small t

 123

eb b
Hao~aoGa(t1,...,t n) = Hbo~t oG (t 1,...,t n)

Differentiating this one with respect to t and setting t = 0 one ob-

tains

(Ha),~aoGa(tl ..... t n) = (H b),eboGb(t I, .... t n)

Because of the arbitrariness of Vl,...,v m one has then

a a a b b b
(H),gioG (tl, .... t n) = ( H ) , g i G (t I ..... t n)

for all 0 < i < m. But these ones, in view of the definitions (6.5),
may be rewritten as

b
gi(x) = F ~gia oF-I (X) 0 < i _< m

for all x 6 ~ , thus proving (6.2)

(vi) Again, using the same arguments already used in (iii) one may
easily see that

haoG a(t 1,...,t n) = hboG b(t 1,...,t n)

i.e. that

hb(x) = ha.F -l(x)

for all x e V b, thus proving also (6.3). []

 CHAPTER IV
DISTURBANCE DECOUPLING AND NON INTERACTING CONTROL

I. N o n l i n e a r Feedback and Controlled Invariant Distributions

In this a n d in the f o l l o w i n g chapters, we a s s u m e t h a t in the con-

trol s y s t e m

m
(1 .I) = f(x) + [ gi(x)ui
i=I

it is p o s s i b l e to a s s i g n the v a l u e s of the i n p u t s u I, .... u m at e a c h

time t as f u n c t i o n s of the v a l u e at t of the s t a t e x and, possibly, of
some o t h e r real-valued functions V l , . . . , v m. T h i s c o n t r o l m o d e is cal-
led a s t a t i c state-feedback control. In o r d e r to p r e s e r v e the s t r u c -
ture of (1.1), we l e t u i d e p e n d on x a n d V l , . . . , v m in the following
form

m
(1.2) u i = Gi(X) +j=~l~iJ (X)Vj

where ai(x) a n d 6ij(x) , I ~ i,j ~ m, are r e a l - v a l u e d smooth func-

tions defined on the same o p e n s u b s e t N of ~ n on w h i c h (1.1) is de-
fined
In d o i n g t h i s we m o d i f y the o r i g i n a l dynamics (1.1) and obtain
the c o n t r o l system

m
(I 3) ~ = ~(x) + [ ~g i (x) vi
i=I

in w h i c h

% m
(1.4a) f(x) = f(x) + [ gi(x}~i(x)
i=I
m
(1.4b) gi(x) - j [ i g j (x) 6ji(x)

For reasons of n o t a t i o n a l simplicity, m o s t of the t i m e s we c o n -

s i d e r ~i(x) as the i - t h e n t r y of an m - d i m e n s i o n a l vector ~(x),
~ij (x) as the (i,j)-th entry of an mxm-dimensional~ matrix ~(x) a n d we
c o n s i d e r the v e c t o r f i e l d s g~(x) a n d g (x) as j - t h c o l u m n s of
3 ]
n x m - d i m e n s i o n a l m a t r i c e s g(x) a n d g(x).
 125

In this w a y we m a y replace (I .4) w i t h the shorter expressions

~u
(1.5a) f(x) = f(x) + g(x)~(x)

(1.55) ~(x) = g(x) B(x)

we a l s o systematically assume that the m m m a t r i x 8(x) is in-

vertible for all x. T h i s makes it p o s s i b l e to i n v e r t the transforma-
tion (1.5), and to o b t a i n

I
(1.6a) f(x) = f(x)-~(x)~- (x)a(x)

(1.6b) g(x) = g(x) 8-I (x)

(1.8) Remark. Strictly speaking, only (1.5a) may be r e g a r d e d as a

"feedback", while (1.5b) should be r e g a r d e d as a c h a n g e of c o o r d i n a t e s
in the space of input values, depending on x. []

The p u r p o s e for w h i c h feedback is i n t r o d u c e d is to o b t a i n a dy-

namics w i t h some nice properties that the o r i g i n a l dynamics does not
have. As we shall see later on, a typical situation is the one in
which a modification is r e q u i r e d in o r d e r to o b t a i n the invariance of
a given distribution A under the v e c t o r fields which characterize the
new d y n a m i c s . This kind of p r o b l e m is u s u a l l y dealt with in the fol-
lowing way.
A distribution A is said to be controlled inuariant on N if there
exists a feedback pair (~,B) defined on N w i t h the p r o p e r t y that A is
invariant under the vector fields f,gl,...,g m (see (1.4)), i.e. if

(1.9a) [ f,A] (x) c A(x)

(1.9b) [gi,A] (x) C A(x) for I ~ i ~ m

for all x E N.
A distribution A is said to be ZocalZy controZZed invari~nt if
for e a c h x 6 N there exists a neighborhood U of x w i t h the p r o p e r t y
that A is c o n t r o l l e d invariant on U. In v i e w of the p r e v i o u s defini-
tion, this requires the e x i s t e n c e of a f e e d b a c k pair (~,8) defined on
U such that (1.9) is true for all x 6 U.
The n o t i o n of l o c a l controlled invariance lends itself to a s i m p l e
geometric test. If w e set
 126

G = s p { g I ..... gm }

we may express the t e s t in q u e s t i o n in the f o l l o w i n g terms.

(1.10) Lemma. L e t A be an i n v o l u t i v e distribution. Suppose A, G and

A + G are n o n s i n g u l a r on N. T h e n A is l o c a l l y c o n t r o l l e d invariant if
a n d o n l y if

(1.11a) [f,A] C A + G

(1.11b) [ gi,A] C A + G for I < i < m

Proof. N e c e s s i t y . Suppose A is l o c a l l y controlled invariant. L e t x 6 N,

U a neighborhood of x a n d (e,8) a feedback pair defined on U w h i c h
makes (1.9) satisfied on U. L e t T be any v e c t o r f i e l d of A. T h e n we
have

m m
[ f,T] =[ f+ge,T] =[ f,T] + [ [ gj,T]~j + ~ (LT~ j)gj
j=1 j=1

m m m

l i+J % iEgj ji %Ltg +J+ji

for I < i < m.
Since 8 is i n v e r t i b l e , one m a y solve the last m equalities for
[ gj, T] , o b t a i n i n g

m
[gj,Tl e [ [gi,A] + G
i=I

for I < j < m. T h e r e f o r e , from 1.9b) we d e d u c e (1.11b). Moreover,

since

m
[f,T] e [f,A] + ~ [gi,A] + G
i=I

again from (1.9) and (1.11b) we d e d u c e (1.11a). []

In o r d e r to p r o v e the s u f f i c i e n c y , we f i r s t n e e d the following in-

terestlng result, which is a c o n s e q u e n c e of F r o b e n i u s Theorem.

(1.12) Theorem. L e t U a n d V be o p e n sets in ~ m and ~ n respectively. Let

Xl,...,x m denote coordinates of a p o i n t x in ~ m a n d y l , . . . , y n c o o r d i n a -
tes of a p o i n t y in ~ n . Let FI,...,F TM be s m o o t h functions

P i : U -~ ~ n n
 127

Consider the set of partial differential equations

(I . 1 3 ) -By(x)
Sx = Fi(x)y(x) I -< i -< m
l

where y denotes a function

y : U~V

Given a point (x,y ) 6 U x v there exist a n e i g h b o r h o o d U o of x in U

and a unique smooth function

y : Uo~V

which satisfies the equations (1.13) and is such that y(x ) = yO if

and only if the functions FI,...,F n satisfy the c o n d i t i o n s

(1.14) ~i 3k + Fik - FkFi = 0 I < i,k < m

Bx k Bx i -- _

for all x 6 U.

Proof. Necessity. Suppose that for all (x,y ) there is a function y

which satisfies (1.13). Then from the p r o p e r t y

~2y = ~2y
~xi3x k ~XkDX i

one has

~x i (Fk (x)y ( x ) ) = ~k (Fi(x) y(x))

o
Expanding the d e r i v a t i v e s on both sides and e v a l u a t i n g them at x = x
one obtains

[ t"sFk"
~i;xO + F k ( x ) F i ( x ) ] Y =[ ( ~ ) x o +Fi(xO)Fk(xO)]yO

which, due to a r b i t r a r i n e s s of x,y O, yields the c o n d i t i o n (1.14).

Sufficiency. The proof of this part consists of the following
steps.

(i) It is shown that the f u l f i l l m e n t of (1.14) enables us to define

on UxV a certain involutive distribution &, of dimension m.

(ii) Using Frobenius Theorem, one can find a n e i g h b o r h o o d U' V' of

 128

(x,y ) and a local coordinates tranformation

F : (x,y) ~-~

defined on U'X V' , with the p r o p e r t y that

A(x,y) = span{ (~-~i)(x,y) ..... (7~m) (x,y) }

for all (x,y) 6 U' V'

(iii) From the t r a n s f o r m a t i o n F one constructs a solution of (1.13).

As for the step (i), the d i s t r i b u t i o n A is defined, at each

(x,y) ~ U xV, by

n n
A(x,y) = span{ (8 + ) + ~ ~ Fhk
i (x)Yk(~h), 1 _< i _< m}
i h=0 k=0

In other words, A (x,y) is spanned by m tangent vectors whose

coordinates with respect to the canonical basis { ( ~ i ) ..... ( ~ m ) ,
( ~ I ) ..... (~yn) } of the tangent space to UXV at (x,y) have the form

1 0 0
0 I 0

0 0 I
[,1 (x)y F 2 (x) y Fm(x)y

These m vectors are linearly independent at all (x,y) and so the

distribution A is n o n s i n g u l a r and of dimension m. Moreover, it is an
easy c o m p u t a t i o n to check that if the "integrability" condition (1.14)
is satisfied, then A is involutive.
The p o s s i b i l i t y of c o n s t r u c t i n g the coordinate transformation
described in (ii) is a s t r a i g h t f o r w a r d consequence of Frobenius theorem.
The function F thus defined is such that if v is a vector in A, the last
n components of F~v are vanishing. Since, moreover, the tangent vectors
(~) ..... ( ~ ) span a subspace which is c o m p l e m e n t a r y to A(x,y) at

all (x,y) and F is nonsingular, one may easily conclude that the func-
tion
= F(x,y)

is such that the jacobian m a t r i x

 129

$~m+I ~m+l
8YI "'" $Yn
(1.15) , , .

~m+n ~m+n
~Yl "'" ~Yn

is n o n s i n g u l a r at all (x,y) 6 U' V'.

Without loss of g e n e r a l i t y we m a y assume that

~ i ( x , y O) = 0

for all m+1 < i < m+n. As a c o n s e q u e n c e , the i n t e g r a l submanifold of

A passing t h r o u g h (x,y ) is d e f i n e d by the set of e q u a t i o n s

(x,y) = 0 I < i < n

%m+i -- --

Since the m a t r i x (1.15) is n o n s i n g u l a r , thanks to the implicit

function t h e o r e m the a b o v e equations m a y be s o l v e d for y, y i e l d i n g a
set of f u n c t i o n s

(1.16) Yi = Hi(x) I < i < n

defined in a n e i g h b o r h o o d U c U' of x . M o r e o v e r

q i ( x o) = y O I < i < n

The functions (1.16) satisfy the d i f f e r e n t i a l equations (1.13)

and t h e r e f o r e , are the r e q u i r e d solutions. As a m a t t e r of fact, the
functions

~i (x'y) = Yi - qi (x) I _< i < n

are c o n s t a n t on the i n t e g r a l submanifold of ~ p a s s i n g through (x,y )

and, t h e r e f o r e , if v is a v e c t o r in A,

d~iv = 0 I < i <_ n

at all p a i r s (x,q(x)). These equations, taking for v e a c h one of the

m vectors used to d e f i n e A, y i e l d e x a c t l y
 130

~x. = (Fj (x) i"I(x)) i I .< i . < n,

. .I < j < m. [3
3

P~oof. (of L e m m a 1.10). Sufficiency. Recall that, by a s s u m p t i o n , A,G

a n d A + G are n o n s i n g u l a r ; let d d e n o t e the d i m e n s i o n of A a n d l e t

p = dim G - dim A N G

Given a n y x C N it is p o s s i b l e to f i n d a n e i g h b o r h o o d U of x a n d an
mm nonsingular m a t r i x B, w h o s e (i,j)-th element b.. is a s m o o t h real-
z3
valued function defined on U, s u c h that, for

gi =j=~igjbji I _< i < m

the f o l l o w i n g is true

sp{gp+l ..... gm } C A

(I .17) (A+G) = A (9 s p { g 1 . . . . . gp}

The tangent vectors g1(x),...,gp(X) are c l e a r l y linearly inde-

pendent at all x 6 U.
Now, observe that if the a s s u m p t i o n (1.11b) is s a t i s f i e d , then
also

(1.18) [gi,A] C A + G

and l e t T I , . . . , ~ d be a set of v e c t o r fields which locally span A

a r o u n d x . F r o m (1.17) and (1.18) we d e d u c e the e x i s t e n c e of a u n i q u e
k o
set of s m o o t h r e a l - v a l u e d functions cji , d e f i n e d l o c a l l y around x ,
and a vector field 6k E A d e f i n e d locally around x such that
i
P
(I .11b' ) [ gi' Tk] =]~I'=ejkigj- + 6ki

for all I < i < m and I < k < d. U s i n g the same a r g u m e n t s and setting

g0 = f

from (1.11a) and (1.18) we d e d u c e the e x i s t e n c e of a u n i q u e set of

real-valued smooth functions k
cj0 and a vector f i e l d 6 k0 E A, d e f i n e d
locally a r o u n d x O, such t h a t
 131

P
(1,11a')
[ go' ~k] j 0

Now, suppose there exists a nonsingular mm matrix B, whose

(i,j)-th element bij is a smooth real-valued function defined locally
around x, such that

(I .19)

for I < k < d, I < h < p, I < i < m. Then, it is easy to see that

m
(1.20) [ [ ghbhi,Tk] 6 A
h=1

for I < i < m, I < k < d. For,

m m m
[h=1
~ ghbhi'Tk] = -h=1
~ (LTkbhi)gh +j~1 bji[~j'Tk]

P m . P k - ~k = ~k
=- " "
~ (LTkbhi)gh + ~ b.. ~Ichjgh + i l
h=1 j=1 ] l h

where ~ is a vector field in A. Since TI,...,T k locally span A, (1.20)

l
implies that

m A
[ [ ghbhi,A] c A
h=l

Therefore, the matrix

B = BB

is such that (1.9b) is satisfied.

Using similar arguments, one can see that if there exists an m1
vector ~, whose i-th element ai is a smooth real-valued function de-
fined locally around x , such that

(1.21) k = 0
-Lrkah" +j ICkjaj + Ch0

for I < k < d, I < h < p, then

m
(I .22)
[ go +h~ighah'~k ] ~ A
 132

for I < k < d. For,

m P m P k P
[g0 + h=~ l g h a h " r k I = -h =['1 ( L k a.h ) g h +.j ! l a .j h ~ i. C h j ~ h. + h ~ i c ~ 0 g h + 6 k : ~k

where ~k is a v e c t o r field in A. F r o m this one deduces that the vector

a = Bfi

is such that (1.9a) is satisfied.

Thus, we h a v e seen that the p o s s i b i l i t y of f i n d i n g B and a which
satisfy (1.19) and (1.21) enables us to c o n s t r u c t a pair of feedback
functions that m a k e s (1.9) satisfied. In o r d e r to c o m p l e t e the proof,
we have to s h o w that (1.19) and (1.21) can be s o l v e d for B a n d a.
Since A is n o n s i n g u l a r and involutive, we m a y assume, without loss
of g e n e r a l i t y , that our choice of local coordinates is such that

a
Tk = - ~ k 1 <_ k < d.

The equations (1.19) and (1.21) may be r e w r i t t e n as a set of part-

ial d i f f e r e n t i a l equations of the form (1.13) by simply setting

k k k
Cll .-. C l m c10
. . .

Fk = k k ck 1 < k < d
Cpl "'" C p m pO

0 ... 0 0

0 ... 0 0

AS a m a t t e r of fact, for e a c h fixed i, the e q u a t i o n s (1.19) correspond

to an e q u a t i o n for the i-th column of B, of the form

~x k = I _<kid

(where bi stands for the i-th column of B) and the e q u a t i o n s (1 .21)

correspond to

(I .24) a~ k I I I < k < d

Both these equations have exactly the form

 133

(1.25) ~Y = Fky I < k < d

8x k -- _

the unknown vector y being m+1 dimensional. Since now the functions Fk
depend also on the c o o r d i n a t e s Xd+1,...,x n (with respect to which no
derivative of y is considered), in order to achieve uniqueness, the
o o
value of y must be specified, for a given x l , . . . , x d, at each Xd+1,...,xn.
For consistency, the last c o m p o n e n t of the initial value of the solu-
tion sought for the equations (1.23) must be set equal to zero, whereas
the last c o m p o n e n t of the initial value of the s o l u t i o n sought for the
equation (1.24) m u s t be set equal to I. In addition, the first m com-
ponents of the initial values of the solutions sought for each of the
equations (1.23) must be columns of a n o n s i n g u l a r mxm matrix, in order
to let B be nonsingular.
The s o l v a b i l i t y of an equation of the form (1.25) depends, as we
have seen, on the f u l f i l l m e n t of the i n t e g r a b i l i t y conditions (1.14).
This, in turn, is implied by (1.11). Consider the Jacobi identity

-[[gi,Tk],Th]+[[gi,Th],T k] =[ g i , [ T h , Tk]]

for any 0 <__ i <_ m. Using for [gi,Tk] and [gi,Th] the e x p r e s s i o n s given
by (1.11a') or (1.11b') and taking Tk = ~x-~ ' Th = ~ one easily ob-
tains

P k - 5k 3 ] P h - + 6h ~ ]
[j~ICjigj + i,~ h - [ X C4ig4 i,~ k = 0
j=1 ~

This yields

P ~cki P P h - ~k 3 .
- ~. ~x h gj + ~ c k +6 h) +|
j=1 j=1 Ji (~I cjgi i' ~ h l

P ~c~'i P h P k - 6h 3_~_]
j=1 j=1 3x i'gx k

Now, recall that 9 and .~ are both vector fields of A, which is

0x h ~x k
involutive. Therefore, also [~ik " ~-Xh
~ ] and [ ~ h
i , ~~x k ] are in A. Since &

and sP{~1 ..... gp> are direct summands and gl ..... ~p are linearly in-
dependent, the previous equality implies

ck . ~c h . P P
31 + 31 h k k h
3xh ~ +~icj~c~i= -~Ic9~c~i= = 0
 134

for I ~ j ~ p, 0 < i < m, I < h,k < d, w h i c h is e a s i l y seen to be

identical to the c o n d i t i o n (1.14). O

We see from this L e m m a that, under reasonable assumptions (na-

mely, the n o n s i n g u l a r i t y of A, G and ~+G) an i n v o l u t i v e distribution
is l o c a l l y controlled invariant if and only if the c o n d i t i o n s (1.11)
are satisfied. These conditions are of special interest because they
don't invoke the e x i s t e n c e of f e e d b a c k functions e and B, as the de-
fini t i o n does, but are e x p r e s s e d only in terms of the v e c t o r fields
f,gl,...,g m which characterize the g i v e n c o n t r o l system and of the
distribution itself. The fulfillment of c o n d i t i o n s (1.11) implies the
existence of a pair of f e e d b a c k functions w h i c h make A invariant under
the n e w d y n a m i c s but the actual construction of such a f e e d b a c k pair
generally involves the s o l u t i o n of a set of p a r t i a l differential equa-
tions, as we have seen in the p r o o f of L e m m a (1.10). There are cases,
however, in w h i c h the s o l u t i o n of p a r t i a l differential equations may
be a v o i d e d and these, luckyly enough, include some situations of great
importance in control theory. These will be e x a m i n e d later on in this
chapter.

2. The D i s t u r b a n c e Decouplin 9 Problem

The n o t i o n of l o c a l l y controlled invariance will now be u s e d in

orde r to solve the f o l l o w i n g control problem. Consider a control system

m
(2.1a) x = f(x) + ~ gi(x)u i + p(x)w
i=I

(2.1b) y = h(x)

wher e the a d d i t i o n a l input w r e p r e s e n t s an u n d e s i r e d perturbation,which

influences the b e h a v i o r of the s y s t e m through the v e c t o r f i e l d p. The
system is to be modified, via static state-feedback control on the in-
puts Ul,...,Um, in such a way that the d i s t u r b a n c e w has no influence
on the o u t p u t y.
In v i e w of some e a r l i e r results (Theorem III.(3.12) and Remark
III.(3.13)) this p r o b l e m consists in f i n d i n g a feedback pair (~,8) and
a distribution A which is i n v a r i a n t under f = f+ge and gi = (g~)i '
I ~ i ~ m, contains the v e c t o r field p and is c o n t a i n e d in (sp{dhj})
for all I < j < .
According to the t e r m i n o l o g y introduced in the p r e v i o u s section,
a distribution ~ which is i n v a r i a n t under f = f+g~ and ~
gi = (g6)i '
 135

I < i < m, for some f e e d b a c k (e,8) is controlled invariant. If we set

H = N (sp{dhj}) ~ = (sp{dh I ..... dh}) 1

J=1

we may e x p r e s s the p r o b l e m in q u e s t i o n in the f o l l o w i n g terms.

Disturbance decoupling problem. F i n d a d i s t r i b u t i o n A w h i c h

(i) is c o n t r o l l e d i n v a r i a n t
(ii) is such that p E A C H. []

As we have seen in the p r e v i o u s section, the notion of local con-

trolled i n v a r i a n c e is s o m e t i m e s e a s i e r to deal w i t h than (global) con-
trolled invariance. This m o t i v a t e s the c o n s i d e r a t i o n of the f o l l o w i n g
problem.

Local disturbance decoupling problem. Find a d i s t r i b u t i o n ~ w h i c h

(i) is locally c o n t r o l l e d i n v a r i a n t
(ii) is such that p 6 A C H.

(2.2) Remark. Note that the d i s t r i b u t i o n ~ is not r e q u i r e d to be non-

singular, n e i t h e r involutive. However, n o n s i n g u l a r i t y and i n v o l u t i v i t y
may be n e e d e d in order to c o n s t r u c t the p a i r of feedback functions
(~,~) w h i c h make it p o s s i b l e to i m p l e m e n t the d i s t u r b a n c e - d e c o u p l i n g
control mode. This t y p i c a l l y happens when one has found a d i s t r i b u t i o n
which s a t i s f i e s (ii) and, i n s t e a d of (i), satisfies the c o n d i t i o n

(i') [ f,~] C ~ + G

[ gi,A] C A + G I < i < m

In this case, we k n o w from L e m m a (1.10) that n o n s i n g u l a r i t y of A,G

and A+G helps in finding at least locally a p a i r of f e e d b a c k functions
(~,B) w i t h the d e s i r e d properties.
If A is n o n s i n g u l a r and involutive, i n v a r i a n t u n d e r f and gi '
I ~ i ~ m, and s a t i s f i e s (ii), then it is k n o w n from the a n a l y s i s de-
veloped in c h a p t e r I that there e x i s t local c o o r d i n a t e transformations
which put the c l o s e d - l o o p s y s t e m into the form

Xl = f1(x1'x2) + i=igil (Xl 'x2)ui + Pl (xl 'x 2 ) w

m
(2.3) x2 = f2(x2 ) + g i 2 ( x 2 ) u i
i=I

y = h(x2)
 136

Here, o n c e again, one sees that the d i s t u r b a n c e w has no i n f l u e n c e on

the o u t p u t y. []

A systematic way to d e a l w i t h the D i s t u r b a n c e Decoupling Problem

is to e x a m i n e first whether or n o t the f a m i l y of all c o n t r o l l e d inva-
riant distributions contained in H has a " m a x i m a l " element (an element
which contains all other members of the f a m i l y ) . For, if this is true,
then the p r o b l e m is s o l v e d if a n d o n l y if this m a x i m a l element contains
the v e c t o r f i e l d p.
If, r a t h e r than controlled invariant distributions, we l o o k at
locally c o n t r o l l e d invariant distributions, then the e x i s t e n c e of such
a maximal e l e m e n t m a y be s h o w n u n d e r rather mild assumptions. T o this
end, we i n t r o d u c e a notation a n d an a l g o r i t h m . Let X(f,g;K) denote the
collection of all s m o o t h distributions which are c o n t a i n e d in a given
distribution K and satisfy the c o n d i t i o n s (1.11). In v i e w of L e m m a
(1.10), the m a x i m a l element of ~ ( f , g ; K ) is the n a t u r a l candidate for
the m a x i m a l locally controlled invariant distribution in K. As a mat-
ter of fact, the m a x i m a l element of X ( f , g ; K ) m a y be f o u n d by m e a n s of
the f o l l o w i n g algorithm.

(2.4) Lemma ( C o n t r o l l e d I n v a r i a n t D i s t r i b u t i o n Algorithm). Let

~0 = KI
(2.5)
m
Dk = ~k-1 + Lf(G L n~k_l ) + [ L (Gi n ~k_1)
i=1 gi

Suppose there e x i s t s an i n t e g e r k such t h a t ~k~ = ~ k ~ + 1 " T h e n ~k = ~ k *

for all k > k .
If ~k~ N G a n d ~ ~ are smooth, t h e n ~i~
! is the m a x i m a l element
of X(f,g;K).

Proof. T h e f i r s t p a r t of the s t a t e m e n t is a t r i v i a l c o n s e q u e n c e of
the d e f i n i t i o n s . A s for the other, note first that f r o m the e q u a l i t y
~ k * + 1 = a k e we d e d u c e

L g i ( G i n ~k *) c ~ k *

for I ~ i _< m a n d a l s o for i = 0 if w e s e t f = go ' as s o m e t i m e s we

did before. L e t ~ be a o n e - f o r m in G ~ n ~k~ , and Y a v e c t o r f i e l d in
I In the e x p r e s s i o n
~k* "

<Lgi~,w } = Lgi(~,T ) - (~,[gi,T])

 137

we have

<L ~,T ) = 0
gi

because Lgi~ @ ~k* and

(~,T) = 0

because T 6 ~k,
+ G. T h u s

(~,[gi,Y]) = 0

Since G f] is s m o o t h by a s s u m p t i o n , [gi'T] annihilates every co-

vector in Gi~k~ flk~ , i.e.

[ gi,] E ~k* + G

for 0 ~ I ~ m. Thus, ilk* is a m e m b e r of X ( f , g ; K ) . Let ~ be any o t h e r

element of this collection. We w i l l prove that A c ilk*"
F i r s t of all,
note that if ~ is a o n e - f o r m in ~i N G i and T a vector field in ~ we
have

<L ~,~ ) = 0
gi

so that (recall that ~ is a s m o o t h distribution)

L (~L n ~ ) c ~l
gi
Suppose

~A D ~k

for some k > 0. Then

m ~ ~
~k+1 C ~k + Lf (~ ~ G) + ~ n ( A G I) C
i=I gi

Thus, since ~D C , we d e d u c e that

/
C ~k*

and ~k% is the m a x i m a l element of I ( f , g ; K ) . []
 138

For convenience, we introduce a terminology which is u s e f u l to

remind both the c o n v e r g e n c e of the sequence (2.5) in a f i n i t e number
of stages and the d e p e n d e n c e of its final element on the distribution
K. W e set

L
(2.6) 3(K) = (~0+~1+...+~k+...)

a n d we s a y that J(K) is ~in~teZy oompu~abZe if t h e r e e x i s t s an integer

k such that, in the sequence (2.5), ~k. = ~k* + I" If this is the
case, then obviously J(K) = ~
k ~"
In the L e m m a (2.4) we h a v e seen that if J(K) is f i n i t e l y
computable and if J(K) i ~ G ~ and J(K) are smooth, then J(K) is the
maximal element of X ( f , g ; K ) . In o r d e r to let this distribution be
locally controlled invariant all we need are the assumptions of Lemma
(1.10), as stated below.

(2.7) Lemma. S u p p o s e J(K) is f i n i t e l y computable. Suppose K is an in-

volutive distribution a n d G, J(K), J(K)+G are n o n s i n g u l a r . Then J(K)
is i n v o l u t i v e and is the largest locally controlled invariant dis-
stribution contained in K.

Proof. First, observe that the assumption of n o n s i n g u l a r i t y on G,J(K),

J(K)+G indeed implies the smoothness of J(K) i N G a n d ](K). So, in
view of L e m m a (1.10) we need only to s h o w that J(K) is i n v o l u t i v e .
For, let d denote the dimension of J(K). At any point x one may
find a neighborhood U of x a n d v e c t o r fields TI,...,T d such that

J(K) = sp{T I ..... T d}

on U. C o n s i d e r the distribution

D = sP{Ti: I < i < d} + s p { [ T i , T j ] : I < i,j < d}

and suppose, for the m o m e n t , that D is n o n s i n g u l a r on U. Then, every

vector field T in D can be e x p r e s s e d as the s u m of a v e c t o r field T'
in J(K) and a vector field T" of the form

d d
T" = [ ~ cij[ Yi,Tjl
i=I j=1

where cij , I ~ i,j ~ d, are smooth real-valued functions defined on U.

We w a n t to s h o w that
 139

[gk,D] C D + G

for all 0 < k < m. In v i e w of the a b o v e decomposition of a n y v e c t o r

field ~ in D, this amounts to s h o w that

[gk,[Ti,Tj]] C D + G

The e x p r e s s i o n of the v e c t o r field on the left-hand-side via Jacobi

identity yields

[gk,[Ti,Tj]] = [~i,[gk,Tjl] -[Tj,[gk,Ti]]

The v e c t o r field [gk,Tj] is in J(K) + G and therefore, because of the

nonsingularity of J(K) a n d J(K) + G, it c a n be w r i t t e n as the s u m of
a vector field T in J(K) and a vector field g in G. Since,
[Ti,g ] 6 J(K) + G for a n y g 6 G, we h a v e

[~i,[gk,Tjl] = [ Ti,~ + g ] 6 D + J(K) + G = D + G

and we c o n c l u d e that D is such that

[gk,D] C D + G

for all 0 < k < m.

Now, recall that K is i n v o l u t i v e by a s s u m p t i o n , and therefore that

D CK

From this and from the p r e v i o u s inclusions we deduce that D is an

element of X ( f , g ; K ) . Since D D J(K) by c o n s t r u c t i o n and J(K) is the
maximal element of X(f,g;K), we see that

D = J (K)

Thus, any Lie bracket of v e c t o r fields o f J(K), which is in D by

construction, is s t i l l in J(K) and the l a t t e r is an i n v o l u t i v e distri-
bution.
If we drop the assumption that D has constant dimension on U, we
can s t i l l conclude that D coincides with 3(K) on the subset U C U con-
sisting of all regular points of D. Then, using Lemma I.(2.11), we can
as well prove that D = J(K) on the w h o l e of U. []
 140

In the Local Disturbance Decoupling Problem one is interested in

the largest locally controlled invariant distribution contained in H.
Since this latter is involutive (see chapter I), in order to be able
to use the previous Lemma, we need to assume that the distribution
J(H) is finitely computable and that G, J(H), J(H) + G are nonsin-
gular. If this is the case, then, as we said before, the Local Dis-
turbance Decoupling Problem is solvable if and only if

p E J (H)

We conclude the section with a remark about the invariance of

the a l g o r i t h m (2.5) under feedback transformation.
n., ,~, eu
(2.81 Lemma. Let f,gl,...,g m be any set of vector fields deduced from
f'gl .... gm by setting ~ = f + g~, gi = (g~)i ' I < i < m; then each
c o d i s t r i b u t i o n ~k of the sequence (2.5) is such that

m
~k = ~k-1 + L ~ (Gi ~ ~k-1 ) + ~ L~ (G N nk_1)
i=I gi

Proof. Recall that, given a covector field ~, a vector field and a

scalar function ,

L(Ty)~ = (Ly~)y+<~,T >dy

If ~ is a covector field in G n ~k-1 , then

m m
L%~ = Lfe + ~ (Lgl~)~i + ~ (~'gi dei
f i=I " i=I
m m
L% ~ = ~. (Lgj~)Sji + ~ <~,gj )d~ji
gi j=1 j =I

But (~,gj > = 0 because ~ E G ~ and therefore

m m

L~(G
If n ~k-1 ) +~i=IL~gi (G~ n ~k_1 ) C Lf(G ~ n ~k_1)+~i=iLgi (G N '~k-1)

Since S is invertible, one may also write f = ~-gs-le and

gi = (~B-1)i and, using the same arguments, prove the reverse inclu-
sion. The two sides of inclusion are thus equal and the Lemma is
proved. []
 141

3. Some Useful A l g o r i t h m s

In this section we describe a practical implementation of the algo-

rithm y i e l d i n g the largest locally c o n t r o l l e d invariant distribution
contained in H. Moreover, we show that in some p a r t i c u l a r cases the
construction of this d i s t r i b u t i o n may be o b t a i n e d with simpler methods.
We begin with the easiest situation, first. For each output func-
tion hi(x) we define an integer Pi , called the c h a r a c t e r i s t i c number
of Yi " as the integer i d e n t i f i e d by the conditions

(3.1a) LgjL~hi(x) = 0

for all ~ < Pi ' all I ~ j ~ m, all x 6 N and

Pi
(3.1b) LgjLf hi(x) ~ 0

for some j and x.

Note that if for some output Yi the c h a r a c t e r i s t i c number is not
defined (i.e. (3.1a) holds for all k, all j and all x), then the out-
put yj is in no way affected by any of the inputs ul,...,u m. The ex-
pansions d e s c r i b e d in chapter III show that' if this is the case

tk (~(xO))
Yi(t) = ~ L~h i(x)~T. = h i
k=0

Thus, it seems reasonable to assume that our control system is

suc~ that the c h a r a c t e r i s t i c numbers are defined for each output.
Once the c h a r a c t e r i s t i c numbers are known, we may define an Zxm
matrix A(x) whose e l e m e n t aij(x) on the i-th row and j-th column is

Pi Pi
(3.2) aij (x) = LgjLf hi(x) = {dLf hi(x),gj (x))

and an i-vector b(x)i whose e l e m e n t bi(x ) on the i-th row is

Pi +I Pi
(3.3) bi(x) = Lf hi(x) = <dLf hi(x),f(x))

We point, out first of all an i n t e r e s t i n g property of the objects

defined so far

(3.4) Lemma. Let (~,~) be any pair of feedback functions and let
= ~ + g~' gi (g~)i" Then
 142

Lkh. (x) = Lkhi (x)

f 1

for all k <_ P i a n d a l l x E N. M o r e o v e r , let A(x) be the `gxm m a t r i x

whose (i,j)-th element aij is

% Pi
aij(x) = L,~ L_ h i ( x )
gj

and b(x) the ,g-vector whose i-th element b .1 i s

bi(x) = L~i+lhi(x)

Then

A(x) = A(x) 8(x)

~(x) = A(x)e(x)+b(x)

Proof. The first equality is easily proved by induction. It is true

for k = 0 and, if true for some 0 < k < Pi ' yields

m _k+l
L~k+1.n i (x) = L L hi (x) ~ = Lk+lhf i (x) + ~ L n ~ h (x)dj (x) = bf hi(x)
f j&1 gj r i

The other equalities are straightforward consequences of the first

one.

(3.5) Remark. Note that the invertibility of ~ implies the invariance

of the integers pl,...,p`g as well as that of rank of A(x) under feed-
back transformations. []

From this one can deduce the following interesting result.

(3.6) Lemma Every locally controlled invariant distribution contained

in H is also contained in the distribution A defined by
sup

Z Pi
(3.7) = N n (sp{dL h i} )x
&sup i=I k=0

Suppose A is a smooth distribution. A pair of feedback functions

sup
(~,B) is such that

(3.8a) [f + g~,Asup] C Asu p

(3.8b) [(gS)i,&sup] C &sup 1<i<m

 143

if and only if the d i f f e r e n t i a l s of each entry of the column vector

A(x)a(x) + b(x) and those of each entry of the m a t r i x A(x)8(x) belong
to the c o d i s t r i b u t i o n A
sup"
Proof. Let A be a locally controlled invariant d i s t r i b u t i o n c o n t a i n e d
in H. Then, by definition, A C (sp{dhi}) ~ for all I < i _< .
Moreover, for some local feedback a, [~,A] C A. Suppose A C (sp{dL~i])
~ I "
for some k < Pi; then using Lemma (3.4) we have for any vector field
~EA

= = = dLf n i, T >

i.e. A C (sp{dLk+lhi}) i. This proves that

A c n
i=I k=O
P~nZ(sp{dL~hi))

and therefore the d i s t r i b u t i o n (3.7) contains every locally c o n t r o l l e d

invariant distribution.
Now, suppose there exists a pair of feedback functions that makes
(3.8) satisfied. Let T be a vector field in Asu P. Then

(3.9a) (dL~hi,T~ > = 0

(3.9b) (dLkhi,[~,T]> = 0

(3.9c) ( dLkhi ,[gj,T])

~ = 0

for all I ~ i ~ , 0 ~ k _< Pi ' I _< j _< m. From (3.9b) w r i t t e n for

k = Pi ' we deduce, using Lemma (3.4),

Pi Pi Pi +I
0 = L~(dLf hi,T )-<dL~Lff hi,T ) = <dL~ hi,r> = ( d b i , T >

Similarly , for (3.9c) w r i t t e n for k = Pi we deduce that

0 = (d~ij,T)

Therefore, the d i f f e r e n t i a l s of b i and ~ij belong to the codi-

stribution Asup" Conversely, if the d i f f e r e n t i a l s of ~z, a n d ~ .lj b e l o n g
to the c o d i s t r i b u t i o n A~ , we have that (3.9b) and (3.9c) hold for
sup
k = Pi" For values k < Pi (3.9b) and (3.9c) hold for any feedback
(a,~) because of Lemma (3.4) and, therefore, we deduce that A is
% sup
invariant under ~ and gi' []
 144

F r o m this r e s u l t we see t h a t t h e r e are c a s e s in w h i c h the computa-

tion o f the l a r g e s t controlled invariant distribution contained in H
is not t e r r i b l y difficult. An interesting special case is the one in
which the m a t r i x A(x) has a rank equal to the n u m b e r of its r o w s (i.e.
the n u m b e r of the o u t p u t channels); this is e x p l a i n e d in the following
results.

(3.10) Lemma. Suppose that the m a t r i x A(x) has rank at x . T h e n the

covectors

dh I (x ) , ..., d L f h I ix ) , . . . ,dh i (x ) , ... ,dL Zh (x )

are linearly independent. As a c o n s e q u e n c e , the d i s t r i b u t i o n A is

sup
nonsingular in a n e i g h b o r h o o d U of x a n d

(3.11) d i m A~up(X) = Pl + "'" + P Z + ~ ~ n

O
Proof. Suppose t h a t the d i f f e r e n t i a l s are l i n e a r l y dependent at x
Then there exist real n u m b e r s Cik , 1 ~ i ~ Z, 0 --< k --< P i s u c h that

(3.12) ~ ~ CikdL~hi(x) = 0
i=I k = 0

Now consider the f u n c t i o n

Pi
C i k L f h i ix)
i=0 k=0

According to the d e f i n i t i o n of p l , . . . , p m , this f u n c t i o n is s u c h that

Z Pi ~ C aij (x)
< dl,gj > ix) =i=I~ c.lpi<dLf h i , g j > (x) =i=I zPi

But, on the o t h e r hand, i3.12) shows t h a t d l ( x O) = 0 a n d t h e r e f o r e

the a b o v e equality implies the l i n e a r dependence of the r o w s of the
matrix A(x),i.e. a contradiction. Therefore we c o n c l u d e that if (3.12)
holds, we m u s t h a v e CIp I . .. = c p ~ 0.

Now consider the f u n c t i o n

Pi -1
k
~(x) = I [ cikLfh i (x)
i=0 k=0
 145

(with the u n d e r s t a n d i n g that the above sum is exendend over all non-
negative k's) and observe that, if 0 ~ k ~ Pi-1, then (*)

-(dL~hi'[r f'gj] ) = ( d L k + l h )
f i'gj

Now, by the d e f i n i t i o n of pl,...,pm and from this formula, we have

i Pi -I Z
(dy,[f,gj]> (x) = - ~ ~ . - k+1
Cik%dLf ,gj ) = - [ c aij(x )
i=0 k=0 i=I i'Pi-1

But since in the (3.12) the c o e f f i c i e n t s ClPl, - ., Czp Z have already

been proved being equal to 0, the function X(x) is such that dy(x)=0
and the above e q u a l i t y implies again the linear dependence of the rows
of the m a t r i x A(x), i.e. a contradiction. Therefore Ci,Pi_I =

= ... = Cz,p_1 = 0 (for all Ci,Pi_1 defined, i.e. such that Pi ~ I).

By r e p e a t i n g the procedure one c o m p l e t e s the proof.

(3.13) Remark. As a c o n s e q u e n c e of this Lemma, if the m a t r i x A(x ) has

rank , the functions L~hi(x) , I ~ i ~ Z, 0 ~ k ! Pi are part of a
coordinate system in a n e i g h b o r h o o d U of x . This fact will be ex-
tensively used in the sequel. []

The a s s u m p t i o n on the rank of A(x) identifies a special case in

which the c o m p u t a t i o n of the largest c o n t r o l l e d invariant distribution
contained in H is p a r t i c u l a r l y simple.

(3.14) Corollary. Suppose the m a t r i x A(x) has rank at x . Then in

a neighborhood U of x the d i s t r i b u t i o n A coincides with the
sup
largest locally c o n t r o l l e d invariant d i s t r i b u t i o n contained in H.

Proof. If A(x) has rank Z at x~ in a n e i g h b o r h o o d U' of x O the distri-

bution A is n o n s i n g u l a r and therefore smooth. Moreover, in a neigh-
sup
borhood U c U' of x the e q u a t i o n s

(3.15a) A(x)~(x) + b(x) = y(x)

(3.15b) A(x)~(x) = 6(x)

where y (x) and 6(x) are an a r b i t r a r y k-vector and r e s p e c t i v e l y an

(*)
k k )
- { dLfhi, [ f'gj ] > = < dLk+lh i' g j> --Lf< dLfh i 'gj
and the last term is zero because k < 01-1.
 146

arbitrary xm m a t r i x , have smooth solutions. If the e n t r i e s of ~ and

6 are such that their differentials belong to & , t h e n the feedback
sup
(@,8) is s u c h that (3.8) are satisfied on U. In p a r t i c u l a r this is
true if the e n t r i e s of a a n d ~ are constants. Note that the m a t r i x
must have rank in o r d e r to let 8 be n o n s i n g u l a r .

(3.16) Remark. R e c a l l that any p a i r of feedback functions ~ and

which makes & i n v a r i a n t is a s o l u t i o n of (3.15), p r o v i d e d t h a t y
sup
and 6 have entries with differentials in Al (see L e m m a (3.6)). []
sup
The p r o c e d u r e s o u t l i n e d so far are n o t always usable, because A(x)
may fail to h a v e rank or, more in g e n e r a l , A may n o t be a locally
sup
controlled invariant distribution. In this case one may still use the
general algorithm (2.4). A practical implementation of this algorithm
can be o b t a i n e d in the following way.

(3.17) Algorithm (Construction of the largest locally controlled in-

variant distribution contained in H).
Suppose that in a n e i g h b o r h o o d of the p o i n t x the c o d i s t r i b u t i o n
s p { d h l , . . . , d h ~} has constant dimension, say s O . Let 10(x) be an
s0-vector whose entries 1 0 1 , . . . , 1 0 s 0 are e n t r i e s of h, w i t h the pro-

perty that d101,...,d10s 0 are linearly independent at all x in a neigh-

borhood o f x .
The algorithm consists of a f i n i t e number of i t e r a t i o n s , each one
defined as follows.
Iteration (k). Consider the sk xm matrix Ak(X) whose (i,j)-th
entry is ( d l k i ( x ) , g j ( x ) ) . Suppose that in a n e i g h b o r h o o d U k of the
point x the r a n k of Ak(X) is c o n s t a n t and equal to r k. Then it is
possible to f i n d rk rows of Ak(X) which, for all x in a n e i g h b o r h o o d
U'k C U k of x , are linearly independent. Let

Pkl
Pk =
Pk2

be a s k x sk permutation matrix, chosen in such a way that the r k rows

of P k I A k ( X ) are linearly independent at all x E U~. L e t Bk(X) be an
Sk-vector whose i-th element is < d l k i , f ) ( x ) . As a consequence of pre-
vious positions, the e q u a t i o n s

(3.18a) PklAk(X)~(x) = -PklBk(X)

 147

(a.18b) PkIAk (x) ~ (x) = K

(where K is a m a t r i x of real numbers, of rank r k) may be solved for

and 8, an m - v e c t o r and an m m invertible m a t r i x w h o s e entries are
real-valued smooth functions defined in a n e i g h b o r h o o d U" of x
k
Set 0 = f + ge and gi = , I ! i ! m.
Consider the set of functions

A k = {I = L~Ikj : I <_ j < s k , 0 <_ i < m}

gi
and the c o d i s t r i b u t i o n s
sk
= sp{dlkj}
~kl j~1

~k2 = sp{dl: I E Ak }

Suppose the c o d i s t r i b u t i o n ~kl + ~k2 has c o n s t a n t dimension,

say Sk+ 1 , in a n e i g h b o r h o o d U~' C U"k of x . This integer Sk+ ! is ne-
cessarily larger than or equal to r k because the r k entries of Pkllk
have linearly independent differentials at all x 6 U~ , otherwise
Ak(X) would not have rank r k. Let lk+1,1,...,Ik+1,Sk+1 be entries of

Ik and/or elements of A k with the p r o p e r t y that the d i f f e r e n t i a l s

dlk+1,1'''''dlk+1,Sk+1 are linearly independent at all x in n e i g h b o r -

hood U~' C U~' of x O. Thus

Sk+ 1
~kl + ~k2 = ~ sP{ dlk+1 }
j=1 'J

Define the S k + 1 - v e c t o r Ik+ 1 whose i-th e n t r y is the function

Ik+1,i"
This concludes the d e s c r i p t i o n of the algorithm. []

As a m a t t e r of fact, it is possible to show that the o p e r a t i o n s

thus described are e x a c t l y the ones required in order to compute the
codistribution ~k from codistribution ~k-1 and therefore that,under
suitable assumptions, the a l g o r i t h m ends at a certain stage, yielding
the required distribution. Since the p o s s i b i l i t y of c o m p l e t i n g the
operations d e f i n e d at the k-th stage depends on assumptions on the
rank of A k and on the dimension of ~kl + ~k2 ' we set for c o n v e n i e n c e
all these assumptions in a suitable definition. We say that x is a
regular point for the a l g o r i t h m (3.17) if, for all k > 0, the matrix
 148

A k has c o n s t a n t rank in a n e i g h b o r h o o d of x and the c o d i s t r i b u t i o n

~kl + ~k2 has c o n s t a n t dimension in a n e i g h b o r h o o d of x.
In t h i s c a s e rk, the r a n k o f A k, a n d Sk+1, the d i m e n s i o n of
~kl + Sk2 are w e l l - d e f i n e d quantities in a n e i g h b o r h o o d of x . Note,
however, that a r o u n d a regular point x I other than x , r k and Sk+ I
might be d i f f e r e n t .
The following statement shows that the a l g o r i t h m in q u e s t i o n
provides the largest locally controlled invariant distribution con-
tained in H.

(3.19) Proposition. S u p p o s e x O is a r e g u l a r p o i n t for the a l g o r i t h m

(3.17). Then, there exists an i n t e g e r k with the p r o p e r t y that
Sk~+1 = Sk, and, therefore, the a l g o r i t h m terminates at the (k~)-th
iteration. Suppose a l s o G is n o n s i n g u l a r . T h e n on a s u i t a b l e neigh-
borhood U of x O d i s t r i b u t i o n

Sk*
A = n (sp{dlk,,i
i=O

coincides with the l a r g e s t locally controlled invariant distribution

contained in H. The p a i r of f e e d b a c k functions that solve (3.18) for
k = k * is s u c h t h a t

[f + ge,A*] C 4"

[ (g~)i,A ] C ~ I < i < m

Proof. W e s h a l l p r o v e b y i n d u c t i o n that the a s s u m p t i o n s of L e m m a

(2.7) are s a t i s f i e d a n d that

sk
~k = ~ sp{dlkj}
j=1

This is true for k = 0, by d e f i n i t i o n .

Suppose it is true for some k. To c o m p u t e Sk+1 we n e e d to com-
pute f i r s t Sk N G i. N o t e that ~k is n o n s i n g u l a r around x because
the d i f f e r e n t i a l s dlkj , I < j _< s k , are l i n e a r l y independent at
all x 6 U~ . The intersection ~k N G ~ a t x is d e f i n e d as the set of
all l i n e a r combinations of the form

sk
cidlkj(X)
i=I
 149

which annihilates gl (x),...,gin(x). Therefore, it is easily seen that

the coefficients ci,...,csk of this combination must be solutions of
the equation

(Cl...csk)Ak(X) = 0

Since A k(x) has constant rank r k in a neighborhood of x , ~k n G k is

n0nsingular around x , has dimension sk-r k and is spanned by covector
fields which may be expressed as

(3.20) = (1(X)Pk2 + Y2(X)Pkl)dlk

Y1(x) being an arbitrary (sk-rk)-row vector of smooth functions. With

dlk we denote an Sk-COlumn whose i-th entry is the covector field
d~ki.
In computing ~k+1 , we make also use of the fact that, if (e,~)
is any feedback pair, then (see Lemma (2.8))

m m
Dk + [ L (~k A G ) = ~k + ~ L~ (~k A G L)
i=0 g i i=O g i

NOw, take the Lie derivative of (3.20) along gi ' with ~ and 8
solutions of (3.18). As a result one obtains

L ~ = ((L%YI)Pk2 + (L~Y2)Pkl)dl k + y i P k 2 d L Ik + Y 2 P k l d L ~ l k
gi gi gi gi gi

But the way the gi are defined is such that

PkiL~Ik = Pk1(dlk,g0 )= 0
go
PkIL~Ik = Pk1(dlk,~i ) = constant
gi
for all I <_ i < m. Thus, in the above expression we may replace Y2
with any arbitrary rk-row vector 72 of smooth functions. This makes
it possible to express L ~ in the form
gi

(3.21) L%~ = Y3dlk + Y4dL~l k

gi gi
 150

where Y3 is some sk-row vector and Y4 is an a r b i t r a r y sk-row vector of

smooth functions. The first term of this sum is already an element of
~k ' by assumption, while the second, due to the a r b i t r a r i n e s s of Y4 '
spans the c o d i s t r i b u t i o n ~k2 (see above). Thus, we may conclude that

m Sk+1
~k+1 = ~k + [ L~ (~k n G i) = ~ k + ~ k 2 =~kl + ~ k 2 = 3~I sp{dlk+1,j}
i=0 gi '=

By assumption, the c o d i s t r i b u t i o n s ~k ' k ~ 0, are nonsingular

around x (their dimension is Sk). Thus, there exists an integer k
such that

~k* = ~k*+1

This clearly implies the termination of the a l g o r i t h m at t h e k -th step.

We have also assumed that ~k n G ~ are n o n s i n g u l a r around x (their
dimension is rk). So in p a r t i c u l a r ~k* + G is nonsingular. If G is
also n o n s i n g u l a r all the assumptions of Lemma (2.7) are satisfied
and ~ . is the r e q u i r e d distribution.
In order to complete the proof, we have to show that the feedback
pair which solves (3.18) for k = k is such as to make ~ . invariant
under the new dynamics. To this end, consider again the e x p r e s s i o n
(3.21) of the Lie derivative along gi of a covector field ~ of ~k DG"
If the a l g o r i t h m terminates at k , then

L~ (~k* N G i) C ~k*
gi

and, therefore, we see from (3.21) that every entry of dL Ik. (due
gi
to the a r b i t r a r y n e s s of y4 ) is a c o v e c t o r field of ~k* . But, since
the entries of dlk. span ~k* ' this implies

L ~k. C ~k* 0 ~ i ~ m
gi
and thus ~k* is invariant under g 0 , g l , . . . , g m . ~k* being nonsingular
and therefore smooth, we may conclude that ~k* is invariant under the
new dynamics. []

This result is v e r y important because it shows that, under suit-

able r e g u l a r i t y assumptions,it is possible to find the largest local-
ly c o n t r o l l e d invariant distribution contained in H, and also a (local-
ly defined) feedback pair ~ and 8 which makes it invariant under the
 151

new dynamics. The latter is p a r t i c u l a r l y useful because we see that,

as far as one is c o n c e r n e d w i t h the m a x i m a l l o c a l l y c o n t r o l l e d in-
variant d i s t r i b u t i o n c o n t a i n e d in H, the c o m p u t a t i o n of a such a
feedback pair does not require solving p a r t i a l d i f f e r e n t i a l equa-
tions (like in the general case, as seen from L e m m a (1.10)) but may
be carried out e s s e n t i a l l y solving x - d e p e n d e n t linear a l g e b r a i c
equations.
We c o n c l u d e the section w i t h some a d d i t i o n a l c o n s i d e r a t i o n s
about the p r o p e r t i e s of the a l g o r i t h m (3.17). It is o b s e r v e d that,
if the a l g o r i t h m m a y be c a r r i e d out until its final stage (i.e. if
o
x zs a r e g u l a r p o i n t for the algorithm), as a b y - p r o d u c t one ob-
tains, for all k ~ 0, not only the d i m e n s i o n sk of each c o d i s t r i b u -
tion ~k of the s e q u e n c e (2.5) but also the d i m e n s i o n sk-r k of the
eodistribution ~k N G ~.
Thus the rank r k of A k may be i n t e r p r e t e d as

~k
(3.22) r k = dim
~k N G ~

The integers r 0 , r l , . . . , r k ~ are rather i m p o r t a n t also for rea-

sons not d i r e c t l y r e l a t e d to the c o n s t r u c t i o n of the d i s t r i b u t i o n
6 . We will see in the n e x t chapter, for instance, that the se-
quence of integers d e f i n e d by setting

61 = r 0

62 = rl-rO
(3.23)

6k~+I = rk~ - rk~_1

may be, in a special case, d i r e c t l y e v a l u a t e d s t a r t i n g from the

coefficients of the functional e x p a n s i o n of the i n p u t - o u t p u t b e h a v i o r
and plays an e s s e n t i a l role in the p r o b l e m of m a t c h i n g linear models.
It is also p o s s i b l e to relate the integers r i , 0 ~ i ~ k , to
the characteristic n u m b e r s Pi ' I ~ i ~ , as s t a t e d below.

(3.24) Proposition. Suppose that the outputs have been r e n u m b e r e d in

such a way that the sequence of the c h a r a c t e r i s t i c n u m b e r s pl,...,p
is increasing. Let x be a r e g u l a r p o i n t for the A l g o r i t h m (3.17).
If rank A(x ) = ~, then the integers 51,...,6k~+I d e f i n e d by (3.23)
are such that 6 i is equal to the number of outputs w h o s e c h a r a c t e r -
 152

istic n u m b e r is (i-I) and, m o r e o v e r , 61+...+6k~+i = 4. []

The p r o o f of this p r o p o s i t i o n is left as an e x e r c i s e to the reader.

4. N o n i n t e r a c t i n g C o n t r o l

Consider again a control s y s t e m of the form

m
= f(x) + ~ gi(x)ui
i=I

Yi = hi(x) I <_ i <

and suppose ~ < m.

It is r e q u i r e d to m o d i f y the system, via static s t a t e - f e e d b a c k
control, in o r d e r to obtain a c l o s e d loop s y s t e m

= ~(x) + [ gi(x)vi
i=1

Yi = hi (x) 1 _< i <

in which, for some s u i t a b l e p a r t i t i o n of the inputs V l , . . . , v m into

d i s j o i n t sets, the i-th o u t p u t is i n f l u e n c e d o n l y by the i-th set
of inputs.
This control p r o b l e m may e a s i l y be dealt w i t h on the b a s i s of
the results d i s c u s s e d in c h a p t e r III (Theorem III. (3.12)) and its
s o l u t i o n has i n t e r e s t i n g c o n n e c t i o n s w i t h the a n a l y s i s d e v e l o p e d so
far in this chapter. In the p r e s e n t case, in order to e n s u r e the in-
d e p e n d e n c e of Yi from a set of inputs v ,...,v we have to find a
Jl~ ~k
distribution A i which is i n v a r i a n t under f and gj , I < j < m, is
c o n t a i n e d in (sp{dhi}) , and contains the v e c t o r fields gjl,...,gjk.

Since this is r e q u i r e d to h o l d for e a c h i n d i v i d u a l output, one has

to find d i s t r i b u t i o n s A I , . . . , A Z all i n v a r i a n t with r e s p e c t to the
same set of v e c t o r fields ~'~I'
"" "'gm"
A set of d i s t r i b u t i o n s A I , . . . , A with the p r o p e r t y that

(4.1a) [~,Ai] C Ai

(4.1b) [gj,Ai] C Ai I _< j _< m

for all I < i < i is c a l l e d a set of compatible c o n t r o l l e d invariant

distributions.The f e e d b a c k pair w h i c h m a k e s (4.1) s a t i s f i e d is called
 153

a compatible feedback. Obviously, in the v e r y same w a y one c a n in-

troduce the n o t i o n of c o m p a t i b l e local controlled invariance.
Thus, the p r o b l e m we face is the following one.

(Loeal) single-outputs noninteracting control problem. F i n d a set of

distributions &l .... ,& which:

(i) are c o m p a t i b l y (locally) controlled invariant

(iia) satisfy the c o n d i t i o n s Ai C (sp{dhi}) x
(lib) for some p a r t i t i o n 1 1 U I 2 U ... U I z of the i n d e x set {1,...,m}
and for some c o m p a t i b l e feedback, satisfy the c o n d i t i o n s

(g~)j e &i

for all j ~ I i. []

The e x i s t e n c e of a s o l u t i o n to t h i s p r o b l e m is c h a r a c t e r i z e d as
follows

(4.2) Theorem. The L o c a l Single-Outputs Noninteracting Control Problem

is solvable if a n d o n l y if the m a t r i x A(x) has r a n k for all x.

Proof. ( N e c e s s i t y ) . Suppose there exists a p a i r of f e e d b a c k functions

which s o l v e s the S i n g l e - O u t p u t s Noninteracting Control Problem. Then,
we know f r o m the a n a l y s i s of c h a p t e r III, section 3,that,in particular,
for all k a n d all I < i <

L~ Lkhi(x) = 0
gj

whenever j ~ I i. W i t h o u t loss of g e n e r a l i t y we may assume the inputs

Vl,...,v m b e i n g renumbered in s u c h a w a y that

I i = {mi_ I + I ..... m i} I ~ i ~

with m 0 = I and m = m. The a b o v e c o n d i t i o n , written for k = Pi s h o w s

that the m a t r i x A(x) has a block-diagonal structure: on the i - t h r o w
0nly the e l e m e n t s whose indexes belong to the set I i are n o n z e r o . But
we have a l s o t h a t

~(x) = A(x)~(x)

Thus, since the m a t r i x ~ is n o n s i n g u l a r and each r o w of A(x) is n o n -

zero by c o n s t r u c t i o n (we a s s u m e d t h a t all Pi'S are d e f i n e d ) , each
row of A(x) is n o n z e r o . A(x) being block-diagonal, this implies that
 154

the rows of ~(x) are linearly independent and so are the Z rows of
A(x).

(Sufficiency). It is known from the analysis given in the previous

section that if the i-th row of A(x) is nonzero for all x, the largest
locally controlled invariant distribution contained in (sp{dhi}) l ks
nonsingular and given by

Qi
Ai* = k=0n (sp{dL~hi}) I

A pair of feedback functions (~,8) such that

(4.3a) [f + g~,A~] C A~.

(4.3b) [ (g~)j ,A~] C A*i < j --

1 -- < m

is a solution of equations of the form

(4.4a) Ai(x)~(x) + bi(x) = Yi(x)

(4.4b) Ai(x) B(x ) = ~i(x)

where Ai(x) and bi(x) denote the i-th rows of A(x) and b(x). The
scalar Yi(X) and the 1m,row vector 6i(x) are functions whose dif-
ferentials belong to (Ai): in particular, real numbers.
Considering the equations (4.4) all together, for all I ~ i ~ ,
one sees immediately that, thanks to the assumption on the rank of
A(x), there exists a pair of feedback functions (~,8) that makes
(4.3) satisfied simultaneously for all A i , i.e. that AI,...,A are
compatible locally controlled invariant distributions. In particular
if the right-hand-side of (4.4b) is chosen to be the i-th row of a
block diagonal matrix, one has that in the i-th row of A(x) 8(x),i.e.
in the i-th row of A(x), the only elements whose indexes belong to
the set I. are nonzero. This proves that a compatible B exists with
l
the property that

Pi
L~ L~ h i = 0
gj f

for all j ~ I i. But this, in view of Lemma (3.4) is equivalent to

 155

Pi
L ~ Lf h i = 0
gj

i.e., b e c a u s e by d e f i n i t i o n L L f h i = 0 for 0 < k < Pi '

gj

gj 6 A i

for all j ~ I i.
This p r o v e s that the L o c a l Single-Outputs Noninteracting Control
Problem is solved []

It m a y be i n t e r e s t i n g to look at the i n t e r n a l structure of the

decoupled s y s t e m o b t a i n e d in the p r o o f of this t h e o r e m . Suppose again
that A(x) has rank Z on some n e i g h b o r h o o d U and let ~ a n d ~ be so-
lutions of the e q u a t i o n s (4.4) on U. O n e k n o w s from Lemma (3.10) (see
also R e m a r k (3.13)) that the f u n c t i o n s L~hi(x), I ~ i ~ Z, 0 < k < Pi '
are part of a l o c a l c o o r d i n a t e system. Without l o s s of g e n e r a l i t y we
may assume that t h e y are c o o r d i n a t e functions exactly on the n e i g h -
borhood U. W e w a n t to e x a m i n e the s p e c i a l structure of the c o n t r o l
system in the n e w c o o r d i n a t e s , a f t e r the i n t r o d u c t i o n of the d e c o u p l i n g
feedback
To this end, we set the n e w c o o r d i n a t e s in the f o l l o w i n g way. Let

zi0 h i (x)

Zil L f h i (x)

~i(x) =

Pi
Z.
Lf h i (x)
iP i

for I ~ i ~ . If Pl + "" ~ PZ + is s t r i c t l y less than n, an e x t r a

set of c o o r d i n a t e s , say ~+I , is n e e d e d
The c o m p u t a t i o n of the f o r m t a k e n by the d i f f e r e n t i a l equations
describing the system in the n e w c o o r d i n a t e s is r a t h e r easy. For
I < i < Z a n d k < Pi

~Zik (~ + m
Zik - ~x j=1 gjvj) = L ~ zik + j=1
[ L~gj ZikV j
(4.5) m
: + kh.v.
j=1 gj f i 3 = Lf h i = Zi,k+ I
 156

Whereas, for k = Pi (see Lemma (3.4))

m
~ik = i (x) +j~16iJ (x)vj

where Yi(x) is the right-hand-side of (4.4a) and dij(x) is the j-th

element of the right-hand-side of (4.4b). If this latter is chosen
to be as the i-th row of a block-diagonal matrix,as in the proof of
Theorem (4.2), then the above equation reduces to

(4.6) zik = Yi + ~ 6ijVj

j e ii

Again from the proof of Theorem (4.2), it is seen that Yi and ~i4J
(*)
depend only on zi0'''''Z&Pi As a matter of fact, Yi and ~ij may
be simply real numbers.
Finally, by definition, for all 1 < i <

(4.7) Yi = zi0

As a result, we see that in the new coordinates the closed loop

system may be described in the form

~i = fi(~i ) + [ gij(~i)Vj I < i <

(4.8) J CI i
m

~Z+I = f+1((I ..... (.%+1) + ~ g~,+l j((1 ..... 5~,+l)Vj

j=l

Yi = hi(~i )

wi th
Zil

fi(~i ) = gij(~i ) =
Z. 6
~Pi

Yi(~i ) ij(~i )
hi(~i ) = zi0

(~) Let yi(z) = Yiox(z). Then

P.
i ~z
~i ~x is ~x
3zj-----~= dY i 3Zjk Cis ~x ~Zjk
s=O
because i ~ j.
 157

These equations clearly stress the d e c o u p l e d structure of the

closed loop system.

(4.9) Remark. The choice of Xi(~i) linear in ~i ' i.e. the choice

Pi
i(x) = ai0hi(x) + aiILfhi(x) + ... + a i p Lf hi(x)
i

with ai0, ...,a. real numbers, is admissible, because dYi(x) in this

IQ i Pi
case belongs to sp{dh i,...,dLf hi}. It is also possible to choose 6ij
constant, p r o v i d e d that, for some j I i , ~ij is nonzero because this
is required for the solution 8 of the (4.4b) be nonsingular. The two
facts show that a suitable choice of d e c o u p l i n g feedback makes linear
the first subsystems of (4.8).

(4.10) Remark. Note that A i , the largest locally c o n t r o l l e d invariant

distribution c o n t a i n e d in (sp{dhi}) ~, in the c o o r d i n a t e s is e x p r e s s e d
as

= - - : j ~ i, 0 < k~Pi } + sp{. ~ : 1 <k < d}

sP{~zjk -- ~zz+I, k

where d denotes the dimension of ~Z+I (see chapter I, section 3). []

At the b e g i n n i n g of this section, we have formulated the Nonin-

teracting Control Problem looking at the existence of a set of com-
patible c o n t r o l l e d invariant distributions, each one c o n t a i n e d in
(sp{ dh i } )~ and c o n t a i n i n g the v e c t o r fields gj = (gL)j for all j ~ I i,
0ne can also c o n s i d e r a complementary formulation in the following terms.

Local single-outputs noninteracting control problem. Find a set of

distributions ~I,...,~ which
(i) are c o m p a t i b l y locally c o n t r o l l e d invariant
(iia) satisfy the c o n d i t i o n s ~i C (sp{dhj}) ~ for all j ~ i
(iib) for some p a r t i t i o n 11 U 12 U... U I z of the index set {1,...,m}
and for some compatible feedback, satisfy the conditions

(g~)j ~i

for all j I i. []

Also in this case, in fact, the o u t p u t Yi of the c l o s e d - l o o p

system will be a f f e c t e d only by the inputs whose index belongs to the
set I i .
Clearly the c o n d i t i o n that the rank of A(x) is equal to re-
mains n e c e s s a r y and s u f f i c i e n t for the e x i s t e n c e of a solution to the
problem. If desired, one could directly prove the sufficiency in terms
 158

of the c o m p l e m e n t a r y formulation discussed above. As in T h e o r e m (4.2),

it is e a s y to p r o v e t h a t the a s s u m p t i o n on A(x) makes it p o s s i b l e
to e x p r e s s the l a r g e s t locally controlled invariant distribution con-
tained in N (sp{dhj}) ~ as

~j
N n (sp{dL~hj}) k
Ki = j~i k=0

The d i s t r i b u t i o n s K 1 , . . . , K are c o m p a t i b l e and a compatible feedback

is e x a c t l y the one t h a t m a k e s AI,...,A Z compatible.

(4.11) Remark. N o t e that in the n e w c o o r d i n a t e system

i = sp{ ~ : 0 ~ k ~ pi } + s p { D z i + l , k :1 ~ k _< d}

(4.12) Remark. Summarizing some of the a b o v e results, one m a y observe

t h a t if A(x) has r a n k ~, there is a set of d i s t r i b u t i o n s D1,...,Dg+I,
namely

D i = sp{ ~ ~ : 0 < k < pi } I < i < Z

= sp{-~-~7
~ : I < k < d}
DZ+I --+I,k

which are independent, i.e. such that

Di A ( [ Dj) = 0
j/i

a n d span the t a n g e n t space, i.e. are s u c h that

D I + D 2 + ... + D + I = T M

Moreover,
h. = ~ D.
1 ji 3

Ki = Di + D~,+I
 159

5. C o n t r o l l a b i l i t y Distributions

The a p p r o a c h to the n o n i n t e r a c t i n g c o n t r o l d i s c u s s e d at the end

of the p r e v i o u s section, was the one of looking at a set of c o m p a t i b l e
locally c o n t r o l l e d i n v a r i a n t d i s t r i b u t i o n s ~ i , . . . , ~ , such that

rU
(5.1) sp{gj: j e Ii } C A i C A (sp{dhj}) i
9~i

with gj o b t a i n e d by m e a n s of a c o m p a t i b l e feedback. It was shown that

if A(x) has rank (~), the l a r g e s t l o c a l l y c o n t r o l l e d i n v a r i a n t di-
stributions c o n t a i n e d in
N (sp{dhj}) , d e n o t e d K~,...,K. , are such
j~i
as to satisfy these requirements. This a p p r o a c h e s s e n t i a l l y looks at
the "maximal" A i w h i c h satisfy (5.1); however, one c o u l d as well look
at the "minimal" A i w h i c h satisfy these conditions. This k i n d of ap-
proach yields the n o t i o n of a c o n t r o l l a b i l i t y distribution.
A d i s t r i b u t i o n A is said to be a controllability distribution on
N if it is i n v o l u t i v e and there e x i s t a feedback pair (~,8) d e f i n e d on
N and a subset I of the index set {1,...,m} w i t h the p r o p e r t y that
%
A N G = sp{gi: i I}, and A is the s m a l l e s t d i s t r i b u t i o n w h i c h is in-
variant under the v e c t o r fields ~, 1,...,g m and c o n t a i n s gi for all
i 6I.
A d i s t r i b u t i o n A is said to be a local controllability distribu-
tion if for each x 6 N there exists a n e i g h b o r h o o d U of x w i t h the pro-
perty that A is a c o n t r o l l a b i l i t y d i s t r i b u t i o n on U.
It is clear that, by definition, a (local) c o n t r o l l a b i l i t y distri-
bution is (locally) c o n t r o l l e d invariant. Therefore, a c c o r d i n g to the
result of L e m m a (1.10), such a d i s t r i b u t i o n m u s t s a t i s f y (1.11) (note
that the n e c e s s i t y of (1.11) is not d e p e n d e n t on the a s s u m p t i o n s made
in Lemma (1.10) but only on the c o n t r o l l e d i n v a r i a n c e and the nonsin-
gularity of B). T h e r e f o r e it is i n t e r e s t i n g to look for the e x t r a con-
dition to be added to (1.11) in o r d e r to let a given c o n t r o l l e d inva-
riant d i s t r i b u t i o n become a local c o n t r o l l a b i l i t y distribution. To this
purpose, it is u s e f u l to i n t r o d u c e the f o l l o w i n g algorithm.

(5.2) Lemma (Controllability D i s t r i b u t i o n A l g o r i t h m ) . Let A be a fixed

distribution. Define a sequence of d i s h r i b u t i o n s S i s e t t i n g

SO = A N G
(5.3) m
Sk = A ([ f,Sk_ I] + [ [ gj,Sk_ I] + G)
9=I
(~) This condition is indeed necessary in the Single-Outputs Noninteracting Control
Problem if B is nonsingular and all Oi! s are defined.
 160

This sequence is n o n d e c r e a s i n g . If t h e r e exists an i n t e g e r k ~ such that

Sk, = Sk,+l , t h e n S k = Sk, for all k > k~

Proof. We n e e d o n l y to p r o v e that S k D Sk_1. T h i s is c l e a r l y true for

k = I. If true for some k, then

m m
([ f,S k ] + j=l
[ [ gj' Ski) D ([ f , S k _ 1 ] + j=1
[ { gJ 'Sk-1 ])

and t h e r e f o r e

Sk+l o Sk

(5.4) Remark. N o t e t h a t we m a y as w e l l r e p r e s e n t S k as

m
Sk = A A ([ f,Sk_1] + [ [gj,SkJ1 ] +G) + Sk_ I
3=I

or as

m
Sk = A ([ f,Sk_11 + [ g j , S k - I] + Sk_ I + G)
j=l

The l a s t one comes f r o m the f i r s t a n d f r o m the m o d u l a r distributive

rule, which holds because S k _ I C 4. []

As we d i d for the a l g o r i t h m (2.5) we introduce now a terminology

which will be u s e d in o r d e r to r e m i n d both the c o n v e r g e n c e of the
sequence (5.3) in a f i n i t e n u m b e r of s t a g e s and the d e p e n d e n c e of its
final e l e m e n t on the d i s t r i b u t i o n A. W e set

(5.6) S(A) = (S O + S 1 + . . . + S k + . . . )

and we say that S(A) is finitely computable if there e x i s t s an in-

teger k* such that, in the sequence (5.3), Sk, = Sk,+l. If this is
the case, then obviously S(A) = Sk, .
An interesting property of the algorithm (5.3) is the following
one.

(5.7) Lemma. Let ~,gl,...,g m be any

set of vector fields deduced from
% .
f , g l , . . . , g m by s e t t i n g ~ = f+g@ a n d gi = (g~)i ' I ~ i ~ m; then
each distribution Sk of the sequence (5.3) is such that

m
Sk = A A ([ f,Sk_1] + ~ [ g j , S k _ I] + G)
j=l
 161

Proof. Let T be a vector field of Sk_ I. Then, we have

m
[~,T] = [ f+ge,T] = [ f,T] + ~. ([ gj,~]ej - (LT~ j)gj)
j=l
m

[gi,T]= [(gB)i , T] = ([gj,~]6ji- (LT6ji)g j)

j=l

Therefore

m m

[~'Sk-1] +j=1
[ [gj'Sk-1] + G C [ f,Sk_1] +j=1
[ [ gj' Sk-1] + G

But, since 8 is invertible, then f = ~-gB-le and.g i = (gS-1)i so that,

by doing the same computations, it is found that the reverse inclusion
holds. The two sides are thus equal and the lemma is proved. []

From this it is now possible to deduce the desired "intrinsic"

characterization of a local c o n t r o l l a b i l i t y distribution.

(5.8) Lemma. Let A be an involutive distribution. Suppose A, G, A + G

are nonsingular and that S(A) is finitely computable. Then A is a
local c o n t r o l l a b i l i t y distribution if and only if

(5.9a) [ f,A] C A +G

(5.9b) [gi,A] c A + G I < i < m

(5.10) S(a) = A

Proof. Necessity. Suppose A is a local c o n t r o l l a b i l i t y distribution.

Then it is locally controlled invariant and (5.9) are satisfied. Mo-
reover, locally around each x there exists a feedback (e,6) with the
property that A N G = sp{gi , i ~ I}, where I is a subset of {1,...,m},
and A is the smallest distribution which is invariant under f,g1,...,g
~''
and contains gi for all i E I. Consider the sequence of distributions
defined by setting

(5.11a) A0 = A A G
m
(5.11b) A k = [ ? , A k _ 11 + ~ [ ~ i , A k _ l ] + A0
i=1

It is easily seen, by induction, that

 162

Ak C A

for all k. This is true for k = 0 and, if true for some k = 0, the in-
v a r i a n c e of A u n d e r ~, 71' "" "'gm shows that Ak+ I C A. Therefore, one
has

m
= ~
Ak A ~ ([~,Ak_ I] + [ [gi' k-1 ] + G)
i=I

i.e., from Lemma (5.7)

(5.12) A k = Sk

It is also seen that, by definition, A 0 = sp{gi: i 6 I} and that, by

construction, Ak_ I c A k for all 1 ~ k. Thus, the sequence of distribu-
tions g e n e r a t e d by the a l g o r i t h m (5.11) is e x a c t l y the same as the one
y i e l d i n g < ~ , ~ 1 , . . . , g~ m ~ sp{gi: i I}> ,the s m a l l e s t d i s t r i b u t i o n inva-
riant u n d e r ~ , g
~ l , . . . , g m and c o n t a i n i n g sp{gi: i I}. F r o m (5.12) and
from the a s s u m p t i o n that S(A) is f i n i t e l y c o m p u t a b l e we k n o w that there
is an i n t e g e r k such that Ak% = Ak,+1. Therefore, in v i e w of Lemma
I. (6.3) , the l a r g e s t d i s t r i b u t i o n in the sequence (5.11) is exactly
.. gm[ sp<i= i I}) . F r o m this, one c o n c l u d e s that the largest
distribution in the s e q u e n c e (5.11) m u s t c o i n c i d e w i t h A, i.e. again
from (5.12), that the c o n d i t i o n (5.10) is satisfied.

Sufficiency. We k n o w from L e m m a (1.10) that if A is involutive, if

G,A and G+A are n o n s i n g u l a r and if the c o n d i t i o n s (5.9) are satisfied,
then l o c a l l y a r o u n d e a c h x there e x i s t s a pair of feedback functions
(e,S) w i t h the p r o p e r t y that A is i n v a r i a n t u n d e r ~ , ~ 1 , . . . , g m. From
this fact one m a y deduce that

m
A n ([~,Sk_ll + X [gi' S k - 1 ] + G) + Sk-1 =
i=I
m

= [ ~,Sk_ I] + ~ [gi,Sk_1] + A N G + Sk_ I =

i=I
m

= [~'Sk-ll +i=I
~ [gi'Sk-l] + Sk-1

In v i e w of L e K ~ a (5.7) and Remark (5.4), this shows that

m
(5.13) Sk = [~,Sk_1] + X [~i,Sk_1] + Sk_ I
i=I

W i t h o u t loss of generality, we m a y assume that g l , . . . , g m are

 163

such that A N G = sP{~i: i 6 I} for some index set I. In fact, A N G

is nonzero b e c a u s e , otherwise S(A) w o u l d be zero, thus contradicting
(5.10). S i n c e A A G is n o n s i n g u l a r , one m a y find a new feedback func-
tion and construct new vector fields gi = ( ~ ) i ' I --
< i --
< m, s u c h
~at, for some index s e t I, sp{gi: i 6 I} = A A G and gi = gi for
i ~ I. This n e w set of v e c t o r fields still k e e p s A invariant because
gi 6 A for i 6 I a n d A is i n v o l u t i v e .
So S 0 = G N A = sP{~i: i 6 I}, a n d the s e q u e n c e of d i s t r i b u t i o n s
Sk c o i n c i d e s with the sequence of d i s t r i b u t i o n s yielding

f g , . . . , g~ m l S p { g~i : i 6 I}) . Since, by a s s u m p t i o n , for some k * ,Sk~ =

(~'~1
= Sk,+1 we d e d u c e f r o m L e m m a I. (6.3) t h a t Sk~ is the s m a l l e s t d i -
stribution w h i c h is i n v a r i a n t u n d e r ~ "gl
~ .... 'gm and c o n t a i n s
sp{gi: i E I}. B u t (5.10) says t h a t Sk, c o i n c i d e s with A and this
completes the proof. []

In v i e w of the use of the n o t i o n of l o c a l c o n t r o l l a b i l i t y distri-

bution in p r o b l e m s of d e c o u p l i n g or n o n i n t e r a c t i n g control, it is
useful to be a b l e to c o n s t r u c t a "maximal" local controllability di-
stribution c o n t a i n e d in a g i v e n distribution K. T o this e n d one m a y
use the f o l l o w i n g result.

{5.14) L~mma. L e t A be an i n v o l u t i v e distribution. Suppose G,A,G+A

are n o n s i n g u l a r and

[ f,A] C A + G

[gi,A] C A + G I _< i <_ m

Moreover, suppose S(A) is f i n i t e l y computable and nonsingular. Then

S(A) is the l a r g e s t local controllability distribution contained in A.

Proof. As in the p r o o f of L e m m a (5.8) (sufficiency) it is e a s i l y seen

that the a s s u m p t i o n s imply that locally around each x there exists a
pair of f e e d b a c k functions with the p r o p e r t y that A ~ G = s p { g i : i 6 i }
and S(A) is the smallest distribution which is i n v a r i a n t under
~'~I- ..... -~m a n d c o n t a i n s sp{gi: i 6 I}. M o r e o v e r , since

q,
sp{gi: i 6 I} C S(A) C A

and A G = sp{gi: i 6 I}, it is seen that

S(A) N G = sp{gi: i 6 I}
 164

Thus S(A) is a local c o n t r o l l a b i l i t y distribution.

Let ~ be another local controllability distribution contained in
A. Then, by definition, in a n e i g h b o r h o o d U of each x there exists a
feedback (~,~) with the property that ~ A G = sp{gi:i e I} for some
subset I of {I ..... m}, and ~ is invariant under ~'gl ..... gm' where
= f+g~ and gi = (g~)i for I ~ i ~ m. C o n s i d e r the sequence of di-
stributions

~0 = sp{gi: i I}
m

Ak = [f'Ak-1] +i:I
~ [gi'Ak-1] + Ak-1-

Note that ~k C A C A. Thus

Ak C A n ([ f,Ak_1] + [ [gi,Ak_1] + ~k-1 + G)

i=I

Since ~0 = ~ A G C A i% G = S O , it is easy to show, by induction, by

means of Lemma (5.7) and Remark (5.4) that ~k C Sk for all k ~ 0, i.e.

Ak C S (A)

Now recall (see Lemma I. (6.4)) that there exists a dense subset U with
the p r o p e r t y that at each x 6 U, ~(x) = ~k(X) for some integer k. Thus,
we have that

~(x) c S(A)(x)

for all x in a dense subset. Since ~ is smooth and S(A) is nonsingular,

this implies A C S(A). []

If the d i s t r i b u t i o n K in which one seeks the maximal controlla-

bility distribution does not satisfy the above conditions, one may
proceed finding first the largest locally c o n t r o l l e d invariant distri-
bution c o n t a i n e d in K. From Lemma (2.7) we know that this one is given
by J(K), provided that this d i s t r i b u t i o n is finitely computable, K is
involutive and G , J ( K ) , J ( K ) + G are nonsingular. If S(J(K)) is finitely
computable and nonsingular, then S(J(K)) itself is the required di-
stribution. In fact, we know from Lemma (5.14) that S(J(K)) is not
only the largest local c o n t r o l l a b i l i t y distribution contained in J(K),
but also the largest local controllability distribution contained in
K, because any c o n t r o l l a b i l i t y distribution contained in K, being
locally c o n t r o l l e d invariant, m u s t be also c o n t a i n e d in J(K).
 165

(5.15) Remark. From (5.13) it is a l s o seen that the distribution S(A)

is left i n v a r i a n t b y any s e t of v e c t o r fields ~ ,~g l , . . . , g m w h i c h
leaves A i n v a r i a n t . As a matter of fact, the c o n d i t i o n

Sk~ = Sk~+1

implies [ f,Sk,] C Sk, and [~i,Sk,] C Sk, , 1 < i < m.

6. More On N o n i n t e r a c t i n ~ Control

In this section we shall see that the n o t i o n of c o n t r o l l a b i l i t y

distribution makes it p o s s i b l e to a n a l y z e under a different per-
spective the k i n d of p r o b l e m s dealt with in the section 4. C o n s i d e r
again the L o c a l Single-Outputs Noninteracting Control Problem, that
we know is s o l v a b l e if a n d o n l y if the m a t r i x A(x) has rank i. In
order to a v o i d unessential notational complications, we m a y assume
~at the n u m b e r of i n p u t channels is e q u a l to t h a t of the o u t p u t
channels, i.e. Z = m, so t h a t each decoupled channel is s i n g l e - i n p u t
and s i n g l e - o u t p u t . In s e c t i o n 4 we h a v e seen that a pair of f e e d b a c k
functions which solves the p r o b l e m may be found as a s o l u t i o n of the
equations (4.4) (where, in p a r t i c u l a r , Xi m a y be zero and 6i the i-th
row of the identity matrix). We h a v e also observed that this solution
provides a feedback which makes the following simultaneously inva-
riant:

- A , the largest locally controlled invariant distribution contained

in H,
- A i , the ]argest locally controlled invariant distribution contained
(sp[dhi}) , I ~ i ! ~,
- K i , the largest locally controlled invariant distribution contained
n (sp{dhj}) i, I < i < .
j~i -- --
We have also investigated the internal structure of the system
~us obtained, and found a local state-space description of the form
(4.8), i.e.

~i = ~ i ( ~ i ) + ~ i ( ~ i ) v i

(6.1) ~#,+1 = f~,+l (~1 . . . . ' (9,+1) + [ gZ+l,j((1 ..... (+1)vj
j=1

Yi = h i ( ~ i )
 166

The approach to the n o n i n t e r a c t i n g control p r o b l e m via the solu-

tion of (4.4) makes simultaneously invariant a set of distributions
which generally are not independent. For instance, the set KI,...,K
is indeed a set of c o m p a t i b l y locally c o n t r o l l e d invariant distribu-
tions w h i c h satisfy the conditions

C Ki C n (sp{dhj}) i
(gS) i

for some compatible feedback but, as we have seen before (Remark

(4.12)), if

d = n - (pl+P2+...+p~+) > 0

then for any pair I < i <

K i* n (k~i
~ K k) = Kj n (k~jKk) = K I n K 2 . . . A KZ ~ {0}

The existence of such a n o n z e r o intersection corresponds to the

presence of the set of c o o r d i n a t e s ~i+I = (zi+1,1'''''zz+1,d) which
characterizes the (Z+1)-th s u b s y s t e m of (6.1).
M o t o v a t e d by this consideration, we want to investigate in this
section a slightly different version of the n o n i n t e r a c t i n g control
problem, d e f i n e d as follows.

Local, single-outputs, strong noninteracting control problem. Find a

set of d i s t r i b u t i o n s ~I,...,A which:
(i) are c o m p a t i b l y locally c o n t r o l l e d invariant
(ii) for some compatible feedback satisfy the c o n d i t i o n s

C &i C n (sp{dhj}) , I ~ i ~
(gS)i
j~i
(iii) are nonsingular, independent and span the tangent space
(iv) are s i m u l t a n e o u s l y integrable. []

In view of Theorem I.(3.12), one may replace the r e q u i r e m e n t (iv)

with the r e q u i r e m e n t
(iv'
) for each i = I,...,Z the d i s t r i b u t i o n D i = j~i
[ 4 3 is involutive.

Note that, for instance, the d i s t r i b u t i o n s K~,...,K~ were already

nonsingular and spanned the tangent space, so that the real new con-
straint added in (iii) is the one of the i n d e p e n d e n c e of the set of
distributions in question. On the other hand, simultaneous integra-
bility, introduced in (iv), is useful because it makes it possible
 167

to find l o c a l coordinates in w h i c h the system, once decoupled, appears

as the a g g r e g a t e of i n d e p e n d e n t single-input single-output sub-
systems. We discover such a decomposition as an intermediate step in
the proof of the following result.

(6.2) Lemma. Let = m. The Local Single-Outputs Strong Noninteracting

Control P r o b l e m has a solution if a n d o n l y if there exists a s e t of
distributions A I, ...,A Z which:

(i) are locally controlled invariant

(ii) satisfy the c o n d i t i o n s A i C j~i
n (sp{dhj}) , I --
< i --
<

(iii) are n o n s i n g u l a r , independent and span the t a n g e n t space

(iv) are simultaneously integrable
(v) are such that 4 i A G is n o n s i n g u l a r and one-dimensional, for all
I < i <

(6.3) Remark. In o t h e r words, this Lemma shows that the simpler state-
ment "A. n G is n o n s i n g u l a r and one-dimensional for all I < i < Z"
essentially replaces the statement "41,...,4 are c o m p a t i b l e and, for
some c o m p a t i b l e feedback, (gS)i 6 A i for all I < i _< ".

Proof. N e c e s s i t y . All we h a v e to s h o w is t h a t (v) is true. Recall that

the m a t r i x A(x) has necessarily rank Z for all x. Since

Pl
d L f h I (x)

A(x) = (g1(x)...gm(X))

PZ
d L f h (x)

~d = m, w e deduce that dim G(x) = for all x.

On the o t h e r hand, from the c o n d i t i o n (g6)i E Ai we h a v e also

(A i n G) C G = sp{(g~) I ..... (g~)z} C ~ (A i n G)

i=I i=I

ie.
G = [ (~i n G)
i=I

Since the d i s t r i b u t i o n s A I , . . . , A Z are independent the latter is a

direct s u m a n d this yields

Z = [ dim(A i n G)
i=I
 168

i.e. the c o n d i t i o n (v) b e c a u s e dim(A i N G) > 0.

Sufficiency. Suppose there exists involutive distributions AI,...,A i

which are n o n s i n g u l a r , independent, span the t a n g e n t space and are si-
multaneously integrable. So, a r o u n d every point, there e x i s t local
coordinates of the form

= col(~ I ..... C~)

with
~i = cl({il ..... ~in. )
l

such that

(6.4) A i = sp{~ : I < j < ni} , I <_ i _<

From (v) one a l s o d e d u c e s the e x i s t e n c e of a (locally defined) ZZ non-

singular matrix ~ of s m o o t h functions with the p r o p e r t y that

~i = (g~)i

spans the o n e - d i m e n s i o n a l distribution A i ~ G.

Moreover, from the fact that the d i s t r i b u t i o n s A i are l o c a l l y con-
trolled invariant and f r o m L e m m a (1.10) (necessity) we have

[f,Aj] c A j + sp{~ I ..... ~ }

[gi,Aj ] C Aj + sp{g I ..... gz} , I < i < i

for all I < j < ~. F r o m these we get, in p a r t i c u l a r

[ f,A2+...+Az] C h 2 + . . . + A g + s p { g I ..... ~Z} =

= A 2 + . . . + A ~ + sp{g I}

[gl,A2+...+A] C A 2 + . . . + A + sp{gl}

These two c o n d i t i o n s have the form (1.11). Thus, since A2+...+A Z

is i n v o l u t i v e (see T h e o r e m I.(3.12)) and c o n s t a n t dimensional,
d i m sp{gl} = I and sp{gl} N (A2+...+AZ) = 0, f r o m L e m m a (1.10) (suf-
ficiency) we deduce the e x i s t e n c e of a l o c a l l y defined pair of scalar
functions, ~I and $I such that
 169

[f + ~ 1 ~ 1 , A 2 + . . . + A ] C A2+...+A

[g161,A2+...+A~] C A2+...+A ~

One can p r o c e e d in the same w a y a n d f i n d o t h e r p a i r s of f u n c t i o n s ~2

~d 82,... , up to ~ a n d 8 w h i c h m a k e s conditions like

(6.5a) [ f + ~iai , @ A.] C @ A

j#i ] j~i ]

(6.5b) [giBi , ~iAj] C e 6.

j j~i ]

satisfied for I < i < Z.

F r o m ~, al .... ,e , 61,...,6 Z we construct an o v e r a l l feedback
pair ~ and 6 setting

F o~1 61 0 ...0

c~2 0 62...0
(6.6) o~ =

~9.' 0 0 .6,%

This f e e d b a c k is c l e a r l y such that

= f + g~ = f + g l e l + . . . + g z ~ z

gi = gi6i I < i _<

We s h o w n o w that t h i s is a c o m p a t i b l e feedback for A I , . . . , A Z.

The check of t h i s p r o p e r t y is p a r t i c u l a r l y easy in the l o c a l c o o r d i n a t e
chosen to s a t i s f y (6.4). Since gi E A i , we d e d u c e t h a t the i - t h g r o u p
of c o m p o n e n t s of ~ c o i n c i d e s with the i - t h of c o m p o n e n t s of f + g i e i .
Moreover, from (6.5a), using the same k i n d of a r g u m e n t s employed in the
proof of L e m m a I. (4.3), it is e a s i l y seen t h a t the i - t h g r o u p of c o m -
ponents of f + g i ~ i depends o n l y on the l o c a l coordinates ~i" For s i m i l a r
reasons it is also seen t h a t in g i ~ i the o n l y n o n z e r o g r o u p of c o m p o -
nents is the i - t h one, w h i c h depends o n l y on the l o c a l coordinates ~i"
Thus, in the l o c a l coordinates ~ = c o l ( ~ I .... ,~), the v e c t o r fields
and gi ' I ~ i ~ , are r e p r e s e n t e d in the f o r m
 170

:u

gI(~i)
0
~2(2 )
~(~) = ..... ~(~) =

~(~) g~(~)

This c l e a r l y shows that the feedback (6.6) is a c o m p a t i b l e feedback

and c o m p l e t e s the proof.

(6.7) Remark. In the c o o r d i n a t e s ~ = coI(~I,...,~), the i-th output

depends only on ~i (because of (ii)). Therefore, the d e c o u p l e d system
is d e s c r i b e d as a set of i n d e p e n d e n t s i n g l e - i n p u t s i n g l e - o u t p u t sub-
systems of the form

~i = fi(~i ) + gi(~i)vi
(6 .s)
Yi = hi(~i)

(6.9) Remark. The d i s t r i b u t i o n s K I , . . . , K ~ satisfy all the requirements

(i) to (v) if and only if p 1 + . . . + p ~ + Z = n, i.e. if and only if A
has d i m e n s i o n 0. If this is not the case, then, in order to be able to
solve the Local S i n g l e - O u t p u t s S t r o n g N o n i n t e r a c t i n g Control Problem,
one has to try w i t h smaller c o n t r o l l e d i n v a r i a n t d i s t r i b u t i o n s . []

If the set K I , . . . , K is not suited, a reasonable a l t e r n a t i v e for

the solution of this control p r o b l e m is the set S ( K I ) , . . . , S ( K ) . As a
m a t t e r of fact, it is p o s s i b l e to prove that, if the m a t r i x A(x) has
rank (a c o n d i t i o n w h i c h is indeed n e c e s s a r y for the s o l v a b i l i t y of
the problem), the only extra c o n d i t i o n n e e d e d to let this set of di-
stributions solve the p r o b l e m in q u e s t i o n is simply the c o n d i t i o n (iii)
of Lemma (6.2).

(6.10) Theorem. Let = m. Suppose the Local S i n g l e - O u t p u t s Noninter-

a c t i n g Control P r o b l e m is solvable. Suppose also that, for each
I ~ i ~ , S(K i) is finitely c o m p u t a b l e and n o n s i n g u l a r . If the set
S(K~) ..... S(K~) is i n d e p e n d e n t a n d spans the t a n g e n t space then the
Local S i n g l e - O u t p u t s Strong N o n i n t e r a c t i n g Control P r o b l e m is also
solvable.

Proof. If the m a t r i x A(x) has rank for all x, then G also has rank
for all x (see p r o o f of L e m m a
(6.3)), K. is n o n s i n g u l a r (see Remark
(4.11)) l
and K~ N G also is n o n s i n g u l a r . For, the i n t e r s e c t i o n K ~
i A G
 171

at x is g i v e n b y the s e t of all l i n e a r c o m b i n a t i o n s of the f o r m

gi (x) c.
i=1 Z
Pj
which a n n i h i l a t e dhj (x) .... ,dLf hj (x), for all j ~ i. T h e c o e f f i c i e n t s
ci, .... c of this c o m b i n a t i o n m u s t be s o l u t i o n of the e q u a t i o n

all (x) ...alz(X )

.. o . cI
ai-l, 1 (x)...ai_ 1 , l (x)

a i + l , 1 (x) . . . a i + 1 , Z (x)
c

az1 (x) . . .a9.9.(x)

The m a t r i x on the l e f t - h a n d - s i d e of this e q u a t i o n has rank -1 and

therefore, at e a c h x, the set of v e c t o r s in G w h i c h are a l s o in K0 is
l
exactly one-dimensional.
From these properties, using Lemma ( 5 . 1 4 ) , we d e d u c e that if
S(K i) is f i n i t e l y computable and nonsingular, then it is the l a r g e s t
local c o n t r o l l a b i l i t y distribution c o n t a i n e d in K..
~ 1
Moreover, it is k n o w n t h a t K I , . . . , K are c o m p a t i b l e , i.e. in-
variant u n d e r the same set of v e c t o r g , .."'gi"
f i e l d s ~ ,~I Therefore,
from R e m a r k (5.15), it is d e d u c e d that a l s o S(K$),. ..,S(K)~ are in-
variant u n d e r ~'gl ..... g. W i t h o u t loss of g e n e r a l i t y , one m a y as-
sume that

gi Ki I _< i < Z

so that

(6.11) K i N G = sp{ i}

because K. n G is o n e - d i m e n s i o n a l .
l
By d e f i n i t i o n

G n Ki C S ( K i) c K i

so that

G n S..(K
i) = G N K i
 172

and

(6.12) G n S(Ki) = sp{gi}

Consider now the distribution

Di = S (Kj)
j~i

It is easy to see that this distribution is also invariant under

~'~I ..... ~i and that

%
D i O sp{gj : j ~ i}

Therefore,

%
Di D ( ~ ' g l ..... gzlsp{gj : j / i} )

We will show now that also the reverse inclusion holds, so that
D i is actually the smallest distribution invariant under ~, 1,...,g
which contains sp{gj: j ~ i}. As a matter of fact, consider the se-
quence of distributions

'b

Si0 = sp{g i}
Z
Sik = [?,S i k_l ]+ [ lgj,Si,k- I] + Si,k_ I
' j=1
From (5.13), and (6.11 , we deduce that for some k

Sl,k*+1 = Sik* = S(K i)

and therefore, from Lemma I. (6.3), that

* ~ >
S(Ki) = (~'gl .... ,g}sp{g i }

This shows that

D l C < ~',c~1 , -- .,%g1 sp{gj : j ~ i}>

Using Lemma I. (6.6) we have that D. is involutive and this, in

view of Theorem I.(3.12), shows that the set S(K1),...,S(K) is simulta-
neously integrable, i.e. that the condition (iv) of Lemma (6.2) is sa-
 173

tlsfied. Conditions (i) and (li) are satisfied by definition and con-
dition (iii) by assumption. Moreover, the fulfillment of (v) derives
from (6.12). This completes the proof. []

(6.13) Remark. The interest in the set S(KI),...,S(K) is also motiva-

ted by the fact that there exists a well defined algorithm which pro-
duces each S(Ki).
(6.14) Remark. Unfortunately, the condition expressed by the above
Theorem is not generally necessary for the solution of this noninterac-
ring control problem.

(6.15) Remark. From the proof of Theorem (6.10) it is seen that, when
rank A(x) = and S(KI),...,S(K i) are independent and span the tangent
space, then any feedback solving the Local Single-Outputs Noninterac-
ring Control Problem also solves the strong version of this problem. []

We conclude the section with an example which illustrates the dif-

ference between the approach taken in section 4 and the one discussed
here.

(6.16) Example. Suppose

I x2 + XlX 3 ] XlX 3 I XlX3

2 2
f(x) = XlX2 + x2x3 ' gl (x) = x~x 3 , g2(x) x2x 3
2
xI - x3 1 x2

h I (x) = x I

h 2 (x) = x 3

and consider first the Local Single-Outputs Noninteracting Control

Problem.
Since

dh I = (I 0 0)

dh 2 = (0 0 I)

we have
 174

L g l h I = < d h l , g I } = XlX 3

L g 2 h I = XlX 3

Lglh 2 = I

Lg2h 2 = x 2

Then Pl = 0, P2 = 0 and

A(x) = I XlX3 XlX3

[ I x2

Since

det A(x) = X l X 3 ( X 2 - I)

is n o n z e r o at all x in the dense subset of ~ 3

U = {x 6 ~3: Xl ~ 0, x 2 ~ 0, x 2 ~ I}

the p r o b l e m in q u e s t i o n can be s o l v e d on U.
A feedback solving the p r o b l e m is f o u n d via the e q u a t i o n s (4.4).
Taking Yi = 0 and 6i = i-th row of the 2x2 i d e n t i t y matrix, these be-
c ome

A(x)~(x) = -b(x)

A(x) 6(x) = I

where

b(x) = ILfhlx
IIx+XlX31
Lfh 2 (x)
=
x~ - x 3
This yields
2 2 3
-I [ x2 + X l X 2 X 3 + X l X 3 - X l X 3

~(x) x1~3(x2_1)

t
:

3 2
-x 2 - x Ix 3 + x Ix 3 - x Ix 3
 175

x2 -x I x 3
I
6(x) =

x I x 3 (x2-I)

-I XlX 3

One may w i s h to e x a m i n e the f o r m taken by ~'~I 'g2 in the n e w

local c o o r d i n a t e s ~1 . . . . ,5%+ 1 . I n this case we h a v e

51 = h I (x) = x I

52 = h 2 (x) = x 3

and 5 3 m a y be c h o s e n as x 2. The e q u a t i o n s (4.8) became

~1 = V l

~2 = v2
2
~3 = f(~) + ~ gi(~)vi
i=I

Yl = 51

Y2 = ~2

Let us see n o w h o w K I , K 2 , S ( K I ) , S ( K 2) look, like.

Computation of S(K1). We n e e d first K 1, the largest locally con-
trolled invariant distribution contained in (sp{dh2}) . B y C o r o l l a r y
(3.14), since A(x) has rank 2 (on U),

KI
=
(sp{dh 2}
)L = s p { ~ ~}
, ~x 2

and we may p r o c e e d to compute S(K 1) via the algorithm (5.2). In this

case, in o r d e r to f i n d S O = K I n G we have to solve a set of e q u a t i o n s
of the f o r m

XlX 3 XlX 3
2 2
x2x 3 x2x 3 cI I =
c2
I x2

for cl,c 2 . F r o m this it is seen that K I n G is a o n e - d i m e n s i o n a l di-

stribution, s p a n n e d by the v e c t o r f i e l d
 176

xI
2
T = x2

Now, note that

[g1,T] 6 G, [g2,T] G

So that

[f,S0]+ [g1,S0]+ [g2,S0]+G = [ f,S0]+ G = s p { [ f , T ] , g l , g 2}

Since

x 2 (I -x 2 )

[ f, T] = x2x I (x2-I)

-2x~

then

sp{[ f,~] 'gl 'g2 } = Tx~3

From this, it is seen that on U

s(xl) = KI = sP{~--7 ' ~2

Computation of S(K2). In this case K 2 , the l a r g e s t locally con-

trolled invariant distribution contained in (sp{dhl}) ~, is given by

)L _~}
K2 = (sp{dh I} = s p { - ~ 2 , ~x 3

The algorithm (5.2) now yields

SO = K2 N G = sp{ }
3

Moreover,

[f, ~ x 3 ] G, [ g 1 , ~-x- ~3 ] E G, [ g 3 ' ~ 3 ] G

 177

so that

$I = K 2 n ([ f , S 0 ] + [g1,s0]+ [ g 2 , s 0 ] + G) = K 2 A G = S

From this, it is s e e n that o n U

S ( K 2) = K 2 n G = sp{ }

The distributions thus f o u n d are such that

KIn K2 = s P { ~ x 2 }

whereas

S(KI) N S(K2) = 0.
 CHAPTER V
EXACT LINEARIZATION METHODS

I. ~ i n e a r i z a t i o n of the Input-Output Response

Throughout this chapter we c o n s i d e r again a control s y s t e m de-

scribed by e q u a t i o n s of the form

m
(I .la) = f(x) + [ gi(x)ui
i=I

(1 .lb) y = h(x)

and we w a n t to e x a m i n e to w h a t e x t e n t the b e h a v i o r of such a system

c o u l d be m a d e "linear" under the e f f e c t of an a p p r o p r i a t e feedback
control law. In the first five sections we c o n c e n t r a t e our a n a l y s i s on
the input-output response, whereas in the last two ones the input-state
and the state-output behavior will be c o n s i d e r e d . We shall refer to all
of these subjects as to "exact" linearization problems, as o p p o s i t e to
"approximate" linearization, which generally indicates the approxima-
tion of the b e h a v i o r of a n o n l i n e a r s y s t e m by m e a n s of its first-order
truncated power series expansion.
The first p r o b l e m we deal w i t h is the one of f i n d i n g a static
state-feedback, i.e. a feedback of the form

m
(1.2) u i = ~i(X) + ~ ~ij(X)Vj
j=1

unde r w h i c h the input-output behavior of the s y s t e m (1.1) becomes the

same as the one of a linear system. To this end, we shall first deduce
a set of c o n d i t i o n s which express in simple terms the property, for a
nonlinear s y s t e m of the form (1.1), of d i s p l a y i n g an e s s e n t i a l l y linear
input-output response.
Consider the V o l t e r r a series expansion of the input-output re-
sponse of (1.1) (see III.(2.4), where the i n d i v i d u a l kernels have,e.g.,
the e x p r e s s i o n s III.(2.8)). Suppose the first order k e r n e l s wi(t, T1,x) ,
I ~ i ~ m, depend only on the d i f f e r e n c e (t-T I) and do not d e p e n d on
x, in a n e i g h b o r h o o d U of the initial point x . If this is the case
we see from III.(2.8") that, because of the independence of wi(t, T1,x)
of x, all k e r n e l s of o r d e r higher than one are v a n i s h i n g on U. Thus
the w h o l e expansion III.(2.4) reduces to an e x p a n s i o n of the form
 179

(I .3) y(t) = Q ( t , x O) + [
mft ki(t-T)ui(T)d~
i=I
0

wi th

ki(t-T ) = wi(t,T,x)

The response (1.3) is v e r y much close to the one of a linear

system. Indeed, it is e x a c t l y the one of a l i n e a r system if one
neglects the e f f e c t of the zero-input t e r m Q ( t , x ) . Anyhow the
input-dependent p a r t of the response (1.3) is l i n e a r in the input.
Since in m o s t practical situations one is e s s e n t i a l l y interested in
getting l i n e a r i t y only between input and output, the achievement of
a response of the form (1.3) will be c o n s i d e r e d as s a t i s f a c t o r y .
o
(1.4) Remark. Suppose, for instance, that the initial state x is an
equilibrium state. In this case, it is r e a d i l y seen from III.(2.6)
that Q ( t , x O) = h ( x ) and, therefore, by subtracting f r o m y(t) the
constant t e r m h(x), one obtains in (1.3) exactly the zero-state be-
havior of a l i n e a r system. []

Note that, if a V o l t e r r a series expansion takes the p a r t i c u l a r

form (1.3), then necessarily the first order kernels wi(t, T1,x) are
independent of x a n d d e p e n d only on the difference t-T I , so that
this p a r t i c u l a r property of the first order kernels becomes a neces-
sary a n d sufficient condition for (1.3) to hold.
If, instead of the e x p r e s s i o n III.(2.8), one considers the
Taylor s e r i e s expansion III.(2.12b) of w i ( t , Y 1 , x ) , it is f o u n d that
a necessary and sufficient condition for this kernel to be independent
of x a n d d e p e n d e n t only on t-T I , or - in o t h e r words - for (1.3) to
h01d is t h a t

(1.5) L g i L ~ h j (x) = i n d e p e n d e n t of x

for all k > 0 a n d all I < j < , I < i < m. We m a y summarize this by
saying t h a t the input-dependent part of the r e s p o n s e of a n o n l i n e a r
system of the form (1.1) is l i n e a r in the input if and only if the
conditions (1.5) are satisfied.
In g e n e r a l , the c o n d i t i o n s (1.5) will n o t be satisfied for a
specific nonlinear system. If t h i s is the case, we m a y w i s h to h a v e
them s a t i s f i e d via feedback, thus setting a rather interesting syn-
thesis p r o b l e m . As usual, we c o u l d look at a g l o b a l problem, in w h i c h
a globally defined feedback is sought which solves the p r o b l e m for
 180

all x E N or, more simply, a local problem in w h i c h a point x O is

given a n d one w i s h e s to f i n d a f e e d b a c k defined in a n e i g h b o r h o o d U
of x . The latter, which is e a s i e r , will be dealt with in the sequel.
For the sake of c o m p l e t e n e s s we state this as follows.
Inp~t-Outpu~ Lineariz~tion Problem. G i v e n (f,g,h) and an i n i t i a l state
o
x , find (if p o s s i b l e ) a neighborhood U of x and a pair feedback
functions e and B, w i t h invertible B, d e f i n e d on U, such that for all
k > 0 and for all I < j < , I < i < m (~)

(I .6) L~ Lkh.(x) = independent of x on U []

gi ~ ]

The possibility of solving this problem may be e x p r e s s e d as a

property of the functions LgjL~hi(x) which characterize the T a y l o r

series expansions of the k e r n e l s wj(t,O,x) around t = 0. For conve-

nience, we arrange these data into m m a t r i c e s and l e t Tk(X) denote
the u
matrix whose entry tij(x) on the i-th row and j-th column is
LgjL~hi(x). As a matter of fact, the p o s s i b i l i t y of s o l v i n g the p r o -

blem in q u e s t i o n may be e x p r e s s e d in d i f f e r e n t forms, each one being

related to a d i f f e r e n t way in w h i c h the data Tk(X), k ~ 0, are ar-
ranged.
One w a y of a r r a n g i n g these data is to c o n s i d e r a formal power
series T(s,x) in the indeterminate s, d e f i n e d as

-k-1
(I .7) T(s,x) =k~0Tk= (x)s

We w i l l see b e l o w that the p r o b l e m in q u e s t i o n may be solved if and

only if T(s,x) satisfies a suitable separation condition.
Another equivalent condition for the existence of solutions is
based on the construction of a s e q u e n c e of T o e p l i t z matrices, denoted
Mk(X), k ~ 0, and defined as

T O (x) T I (x) ... T k(x)

0 T0(x) ... Tk_1(x)

(1.8) Mk(X) =
o .

0 0 ... T o (x)

(~) Recall that f= f+g~ and gi = (g6)i (see IV.(l.4) and IV.(l.5)).
 181

In this case, one is i n t e r e s t e d in the special situation in w h i c h

linear dependence between rows m a y be t e s t e d by t a k i n g linear combi-
nations with constant coefficients only.
In v i e w of the r e l e v a n c e of this p a r t i c u l a r property throughout
all the s u b s e q u e n t analysis, we d i s c u s s the p o i n t w i t h a little more
detail. Let M(x) be an Zm m a t r i x whose entries are smooth real-valued
functions. We say that x is a reguZar point of M if there exists a
neighborhood U of x w i t h the p r o p e r t y that

(I .9) rank M(x] = rank M ( x O)

for all x 6 U. If this is the case, the integer rank M ( x ) is d e n o t e d

r~(M) ; c l e a r l y r~(M) depends on the p o i n t x O, b e c a u s e on a n e i g h b o r -
hood V of a n o t h e r point x I, rank M ( x I) m a y be different.
With the m a t r i x M we will associate another notion of "rank", in
the f o l l o w i n g way. L e t x be a r e g u l a r point of M, U an open set on
which (1.9) holds, and M a matrix whose entries are the r e s t r i c t i o n s
to U of the c o r r e s p o n d i n g entries of M. We c o n s i d e r the v e c t o r space
defined by taking linear combinations of rows of M over the field ~,
the set of real numbers, and denote rE(M) its d i m e n s i o n (note that
again ~R(M) may depend on x). Clearly, the two integers ~(M) and
r~(M) are such that

(1.10) r.p(M) _> rK(M)

The e q u a l i t y of these two i n t e g e r s may easily be tested in the

following way. Note that both remain unchanged if M is m u l t i p l i e d on
the left by a n o n s i n g u l a r matrix of real numbers. L e t us call a
row-reduction of M the p r o c e s s of m u l t i p l y i n g M on the left b y a
nonsingular m a t r i x V of real n u m b e r s with the p u r p o s e of a n n i h i l a t i n g
the m a x i m a l number of rows in VM (here also the r o w - r e d u c t i o n process
may d e p e n d on the p o i n t x). Then, it is t r i v i a l l y seen that the
two-sides of (1.10) are e q u a l if and only if any p r o c e s s of row-
reduction of M leaves a number of n o n z e r o rows in V M w h i c h is equal
to rz (M).
We m a y n o w r e t u r n to the o r i g i n a l synthesis problem and prove
the m a i n result.

(1.11) Theorem. There exists a solution at x to the I n p u t - O u t p u t

Linearization Problem if and only if e i t h e r one of the following equi-
valent c o n d i t i o n s is satisfied:
 182

(a) there e x i s t a formal p o w e r series

0o

K(s) = ~ Kk s-k-1
k=0

whose coefficients are xm m a t r i c e s of real numbers, and a formal

power series
co

R(s,x) = R_I (x) + ~ Rk(X) S -k-1

k=0

whose c o e f f i c i e n t s are m x m m a t r i c e s of smooth f u n c t i o n s d e f i n e d on a

n e i g h b o r h o o d U of x , w i t h invertible R_1(x) , w h i c h factorize the
formal p o w e r series T(s,x) as follows:

(1 .12) T(s,x) = K(s)'R(s,x)

(b) for all i > 0, the p o i n t x is a r e g u l a r p o i n t of the T o e p l i t z

m a t r i x M.1 and

(I .13) rlR(Mi) = r~(Mi). []

The proof of this T h e o r e m c o n s i s t s in the f o l l o w i n g steps. First

we i n t r o d u c e a r e c u r s i v e algorithm, k n o w n as the S t r u c t u r e Algorithm,
w h i c h o p e r a t e s on the sequence of m a t r i c e s Tk(X), k ~ 0. Then, we
prove the s u f f i c i e n c y of (b), e s s e n t i a l l y by showing that this as-
sumption makes it p o s s i b l e to c o n t i n u e the S t r u c t u r e A l g o r i t h m at
each stage and that from the data thus e x t r a c t e d one m a y c o n s t r u c t a
feedback solving the problem. Then, we c o m p l e t e the proof that (a) is
n e c e s s a r y and that (a) implies (b).

(1.14) Remark. For the sake of n o t a t i o n a l compactness, from this point

on we make s y s t e m a t i c use of the f o l l o w i n g notation. Let X be an s1
v e c t o r of smooth functions and { g l , . . . , g m } a set of v e c t o r fields.
We let Lgy denote the sXm m a t r i x w h o s e i-th column is the v e c t o r
L ~, i.e.
gi

Lgy = [ Lg 1 X --- Lg mY]- []

(1.15) Algorithm (Structure A l g o r i t h m ).

S t e p I. Let x be a r e g u l a r p o i n t of T O and suppose ~R(T0) = r ~ ( T 0 ) .
Then, there exists a n o n s i n g u l a r m a t r i x of real numbers, d e n o t e d by
 183

Vl = I PII
KI

where PI performs row permutations, such that

VITo(X) = I S10(x) I

where St(x) is an roXm matrix and rank St(x) = r O. Set

61 = r 0

YI (x) = P1h(x)

{I (x) = K11h(x)

and note that

LgY1(x) = SI(X )

Lg~1(x) = 0

If To(X) = O, then PI must be considered as a matrix with no rows and

K II is the identity matrix.
Step i. Consider the matrix

LgY1 (x)
iSi_1(x
)
LgYi_1 (x) LgLfYi_1 (x)

LgLfYi_1(x)

and let x be a regular point of this matrix. Suppose

(1.16)
LgLfYi-I LgLfYi-1
 184

Then, there exists a nonsingular matrix of real numbers, denoted by

I61 ... 0 0

oo,
V.
1

0 0
"- I6i_I

0 ... 0 Pi

i Ki K~
K1 "'" i-I z

where Pi performs row permutations, such that

Lg1 (x)
Ii Si (x)
V, =
l

LgYi_ i (x) 0

LgLfYi_ I (x)

where Si(x) is an ri_ I x m matrix and rank Si(x ) = ri_ I. Set

6 i = ri_ I - r i _ 2

Yi(x) = PiLfYi_l(X)

~i(x ) = KiY
i I (x)+...+Ki_IYi_
i i - I (x)
1(x)+KiLfYi_

and note that

LgY1 (x)

= S i (x)

h Lg Yi (x)

LgYi (x) = 0

If the condition (I .16) is satisfied but the last -ri_ 2 rows of

the matrix depend on the first ri_ 2 , then the step degenerates, Pi
 185

must be c o n s i d e r e d as a m a t r i x with no rows, K i is the identity matrix,

6 i = 0 and Si(x) = Si_1(x). []
As we said before, this a l g o r i t h m may be c o n t i n u e d at each stage
if and only if the assumption (b) is satisfied, because of the fol-
lowing fact.

(1.17) Lemma. Let x O be a regular point of T 0 and suppose r~(T0)=~(T0).

Then x is a regular point of

- Si_1

I_ LgLfYi-1

o
and the condition (I.16 holds for all 2 < i < k if and only if x is
a regular point of T i and the condition (1.13) holds for all I < i < k-1.

Proof. We sketch the proof for the case k = 2. Recall that

TO T1 I=[ Lgh LgLfh

MI =
0 TO 0 Lgh

Moreover, let Vl, Y1 and ~I be defined as in the first step of the

algorithm. Multiply M I on the left by

VI 0
V =
0 VI J
As a result, one obtains

LgP I h LgLfP1h
VILgh LgV I Lfh
0 LgLfK~h
VM I =
0 LgP1h
0 ViLgh
0 0

S1 LgLfX I
0 LgL fX I
0 SI
0 0
 186

Note that ~ ( S I ) = ~($I) . Thus, because of the special structure of

VM~ , x is a regular point of M I and the condition r~(M I ) = ~K(MI)
is satisfied if and only if x O is a regular point of

LgL fX I

and
i SI

rR
-I
I LgLf1

SI
=5<
LgLf I

SI

i.e. the condition (1.16) holds for i = 2. For higher values of k one
may proceed by induction. []

From this, we see that the Structure Algorithm may be continued

up to the k-th step if and only if the condition (1.13) holds for
all i up to k-1. The Structure Algorithm may be indefinitely con-
tinued if and only if the assumption (b) is satisfied.

Proof of Theorem (1.11). Sufficiency of (b): construction of the line-

arizing feedback. If the Structure Algorithm may be continued inde-
finitely, two possibilities may occur. Either there is a step q such
that the matrix

Ii LgY1(x)

LgXq[1(x)

LgLfyq_1(x)

has rank at x . Then the algorithm terminates. Formally, one can

still set P q = identity, V q = identity

Yq = PqLfXq_1(x)
and

i
- Sq_ I (x) I = Sq(X)
_ LgYq (x)
 187

and c o n s i d e r K~,...,K qq as m a t r i c e s w i t h n o rows. O r , e l s e , from a

i
certain s t e p on all f u r t h e r s t e p s are d e g e n e r a t e In t h i s case, let
q denote the i n d e x of the l a s t n o n d e g e n e r a t e step. Then, for all
J > q, P-J w i l l be a m a t r i x w i t h no rows, K~3 the i d e n t i t y a n d ~-3 = 0
F r o m the f u n c t i o n s Y1" .... yq g e n e r a t e d by the S t r u c t u r e A l g o -
rithm, one m a y construct a linearizing feedback in the f o l l o w i n g way.
Set

I Yq (x)

that S q = LgF
and r e c a l l is an rq_ 1 m matrix, of r a n k rq_ I at x O .
Then the e q u a t i o n s

(I .18a) [LgF(X)]~(x) = -LfF(x)

(1.18b) [LgF(X)]B(x) = [I 0]
rq-1
on a s u i t a b l e neighborhood U of x are s o l v e d by a p a i r of s m o o t h
functions ~ a n d B.
S u f f i c i e n c y of (b): p r o o f t h a t the a b o v e feedback solves the p r o b l e m .
We show f i r s t t h a t

(1.19a) PIL~L~h(x) = independent of x

p Ki-1 I k
(I 19b) i i-I "'" K I L g L ^ h ( x ) = independent of x

for all 2 < i < q a n d that

(1.19c) q - l " " "K II L ~ L ~k h(x)

KqKq-lq = independent of x

To this end, note that (1.18) imply

(1.20a) L~y i = 0

(1.20b) L~y i = independent of x

g

for all I ~ i _< q. M o r e o v e r , s i n c e L g ~ i = 0 for all i _> I, a l s o

 188

(I .20c) L~y i = Lfy i

I

(1.20d) L~ i = 0
g

for all I < i. Using (I .20) repeatedly, it is easy to see that, if

k > i

(1 .21) K~ 1 k Ki 2 k-
...KIL%h = i...K2L~7I
f
= Ki 3 k-1-
i'''K3L~ Y2 = "'"

= KiL k-i+2 -
i ~ i-I

= -i1

If k < i

11 2 2 1 Ki =
i'" i''" k+l ~Yk

These e x p r e s s i o n s hold for every i > I (recall that, if i > q, K~ is

-- 1

an identity matrix).
Thus, if i < q and k > i-1 we get from (1.21)

Pi Ki-li-1 ...K~L~L ~ h = L P.L k-i+2 - k-i+1

Yi

which is either i n d e p e n d e n t of x (if k = i-1) or zero, while for

i < q and k < i-1 we get from (I .22)

p iKi-1 I
i-I"''KIL~L ~ h = Pi'''Kk+2L~(Yk+Ig
k+2 - _3~i_k+1
.= ~j Yj;.

The r i g h t - h a n d - s i d e of this e x p r e s s i o n is again independent of x and

this complete the proof of (1.19b).
Moreover, if k > q, (1.21) yields

q
= ~_ %Y~_k+1
kk+ IgL ~~ - -
= L~ (~k+1 _311
'= K k3+ 1 j )

and this, together with (1.22) w r i t t e n for i = q, which holds for

k < q, shows that also (1.19c) is true. Finally, (1.19a) is also true,
 189

because the l a t t e r is e i t h e r i n d e p e n d e n t of x

(if k = 0) o r zero.
In o r d e r to c o m p l e t e the p r o o f of the s u f f i c i e n c y of (b), we n e e d
0nly to p r o v e t h a t the m a t r i x

PI

1
P2K1

(I .23) H =

Pq q-Kql
KqKq- I I
q q-1 "" "KI

is s q u a r e and n o n s i n g u l a r . This, together with the (1.19) already

proved, shows in f a c t t h a t

L~L~h(x) = independent of x

for all k ~ 0. B u t the n o n s i n g u l a r i t y of (1.23) is a s t r a i g h t f o r w a r d

consequ6nce of the f a c t t h a t this m a t r i x m a y be d e d u c e d f r o m the
matrix V q . . . V 2 V I by m e a n s of e l e m e n t a r y row operations.
Necessity of (a). L e t

(x) = B -I (x)

(x) = -8 -I (X)~(x)

and let ~
T k(x) = L~L~h(x)

~u
If the f e e d b a c k p a i r e a n d (3 is such as to m a k e Tk(X) independent
of x for all k (i.e. to s o l v e the p r o b l e m ) , then

k ~ ~ A ~ - % k-1
(I .24) L h = L h + Tk_1(~ + T k _ 2 L f e + . . . + T o L f

This e x p r e s s i o n m a y be e a s i l y p r o v e d by i n d u c t i o n . In fact, one has

k+1 ~ ~ - ~ k-1
nf h = n ( ~ + ~ & ) n h + L f ( T k _ I ~ + . . . + T o L f ~) =

= L~
 190

From (1.24) one then deduces

L Lkh = (L L k h ) ~ + %
g x ~ '1~ Tk-lLga- + %Tk-2LgLfC~+""
- ~ k-1-
"+ToLgLf a

orr

(I .25) Tk(X) = ~k~(X) + T k _ I L g ~-( X ) + %T k _ 2 L g L f e-( x ) + . . .+~0LgL~-1~(x)

Now, consider the formal power series

~. ~ -k-1
K(s) = TkS
k=0

k=0

and note that the latter is invertible (i.e. the c o e f f i c i e n t of the

0-th power of s is an i n v e r t i b l e matrix). At this point, the e x p r e s -
sion (1.25) tells us e x a c t l y that the Cauchy product of the two series
thus defined is e q u a l to the series (1.7), thus proving the n e c e s s i t y
of (a)

(a) ~ (b). If (1.7) is true, we m a y w r i t e

K0 KI .-. K k R_I (x) R0(x) R I (x) . . . R k _ I (x)

0 K 0 . .. K k _ I 0 R_I (x) R 0(x) ...Rk_ 2(x)

M k (x) =

0 0 ... K 0 0 0 0 ...R_I (x)

The factor on the left of this matrix is a m a t r i x of real numbers,

whereas the factor on the r i g h t is n o n s i n g u l a r at x , as a c o n s e q u e n c e
of the n o n s i n g u l a r i t y of R_1(x). Thus x is a r e g u l a r point o f M k and
the condition (1.13) holds. []

(1.26) Rsmark. We stress again the importance of the S t r u c t u r e Algo-

rithm as a t e s t for the fulfillment of the conditions (a) (or (b)) as
well as a p r o c e d u r e for the c o n s t r u c t i o n of a l i n e a r i z i n g feedback.

(1.27) Remark. An obvious sufficient condition for the e x i s t e n c e of a

solution to the Input-Output Linearization Problem is that the rank of
the m a t r i x A(x) is e q u a l to ~, i.e. that there exists a solution to
the L o c a l Single-Outputs Noninteracting Control Problem. If this con-
 191

dition holds, the S t r u c t u r e Algorithms terminates at a f i n i t e s t a g e q,

yielding S (x) = A(x).
q

2. The I n t e r n a l Structure of the L i n e a r i z e d System

In t h i s s e c t i o n we a n a l y z e some i n t e r e s t i n g features of the l i n e -

arization procedure discussed so far. F i r s t of all, we e x a m i n e some
simple p r o p e r t i e s relating the S t r u c t u r e Algorithm with the A l g o r i t h m
IV.(2.5), the one y i e l d i n g the l a r g e s t locally controlled invariant
distribution contained in H.
We b e g i n w i t h a simple remark, which w i l l be ; e c a l l e d several
times l a t e r on, a n d t h e n we g i v e two l e m m a s w h i c h establish the re-
quired r e l a t i o n between the A l g o r i t h m (1.15) and Algorithm IV.(2.5)

(2.1) Remark. The submatrix

Pi

Ki
i

of the m a t r i x V i introduced at the i - t h s t a g e of the a l g o r i t h m (1.15)

is n o n s i n g u l a r by definition. This makes it p o s s i b l e to e x p r e s s
Lf~i_ I as a l i n e a r c o m b i n a t i o n of XI .... 'Yi a n d ~i" For, let Q[ a n d
Qi be the two m a t r i c e s of real n u m b e r s defined by

(Q[ Qi" ) [Pll

K
= Q i'P i + Q i,,K ii = I

Then, o n e has

Lf~i-1 = QiYi
, + Q i,,( ~ i - K iiY I - . . . - K i_IXi_1
i )

If the i - t h s t a g e is trivial, Qi! is a m a t r i x w i t h no c o l u m n s and

Qi. = I. If the a l g o r i t h m terminates at the q - t h stage, then Q'~ is a
matrix w i t h no c o l u m n s a n d Qqg =
I. []

In w h a t follows, in o r d e r to s i m p l i f y the n o t a t i o n , whenever we

have an sxl v e c t o r y of r e a l - v a l u e d functions and we want to c o n s i d e r
the c o d i s t r i b u t i o n sp{d1 , .... dXs}, we d e n o t e the l a t t e r by s p { d y } .

(2.2) Lemma. S u p p o s e the I n p u t - O u t p u t Linearization Problem is sol-

 192

vable at x O. S u p p o s e G is n o n s i n g u l a r around x O, and the c o d i s t r i b u -

tions ~k g e n e r a t e d via the A l g o r i t h m IV.(2.5), initialized with
~0 = sp{dh}, are n o n s i n g u l a r around x . Then, for all k ~ 0

k+1 k+1
(2.3) ~k = ( ~ sp{dYi} + ~ sp{d~i})
i=1 i=1

k+1 k+1
(2.4) ( ~ sp{dYi } n X sp{dYi}) = 0
i=I i=I

k+l
(2.5) ak n G ~ = X sp{d~ i}
i=I

and ~k n G ~ is n o n s i n g u l a r around x O.

Proof. The p r o o f p r o c e e d s by induction. For k = 0, (2.3) reduces to

~0 = sp{d1} + sp{d~1}

which is c l e a r l y true b e c a u s e

Y1 = Vlh

iI
and V lis a n o n s i n g u l a r matrix. Moreover, by d e f i n i t i o n LgY1 = 0, i.e.

sp{d~1 } C G ~

which implies

~0 n G L = (sp{d1 } + sp{d~1]) N G i = sp{dY1 } n G ~ + s p { d ~ 1 }

But
sp{dY1 } n G ~ = 0

because, if this w e r e not true at x , then there w o u l d exist a I6 I row

vector of real numbers I such that

ILgY1 (x ) = IS 1(x ) = 0

thus c o n t r a d i c t i n g the linear i n d e p e n d e n c e of the 61 rows of S I (x) at

o
x . T h e r e f o r e , we conclude, that
 193

~0 N G i = sp{d~1 }

i.e. (2.5) for k = 0. Moreover, this argument also shows that

sp{dY1 } n sp{d{1 } = 0

i.e. (2.4) for k = 0.

The codistribution sp{dY1} has constant dimension 6 1 (because,
otherwise, the matrix St(x) would not have rank 6 1 at each x in a
neighborhood of x). n 0 has constant dimension by assumption and the-
refore also sp{d~1] , i.e. ~0 n G ~, has constant dimension.
Suppose now (2.3),(2.4),(2.5) are true for some k and ~k N G ~
has constant dimension around x O. From (2.5) we see that

m k+1
i=~ILgi(~k n G ~) C ~ sp{d~j} C ~k
3=I
(because Lgi7 j = 0) and, therefore, that

(2.6) 2k+1 = 2k + Lf(~k O G )

This, in turn, yields

k+1 k+1 k+1

ak+ I = ~ sp{dYi} + [. sp{d~i} + Lf( ~ sp{d~ i}
i=I i=1 i=I

k+1 k+1 k+1

= [ sp{d i} + [ sp{d~ i} + [ sp{dLfy i}
i=I i=I i=I

k+1 k+1
= ~ sp{dy i} + ~ sp{d~ i} + sp{dYk+ 2} + sp{d~k+ 2}
i=I i=I

(the last equality being a consequence of the Remark (2.1)), and this
proves (2.3) for k+1.
Moreover, it is easily seen that

k+2 L
sp{d~i} C G
i=1

(because Lg~ i = 0) and that

k+2
[ sp{d7 i} r? G i = 0
i=I
 194

(because o t h e r w i s e the linear independence of the rows of Sk+ 2 would

be contradicted). The two c o n d i t i o n s together prove (2.4) for k+1 and
also that

k+2 k+2 G k+2

~k+l 0 Gi= ( ~ sp{dYi } + ~ sp{d~i}) N = [ sp{d~i }
i=I i=1 i=I

i.e. (2.5) for k+1.

k+2
The c o d i s t r i b u t i o n ~ sp{d~i} has c o n s t a n t dimension 61+...+6k+ 2
i=I
(because otherwise the linear independence of the rows of Sk+ 2 would
be contradicted) and this, together with the a s s u m p t i o n that 2k+I has
c o n s t a n t dimension, proves that ~k+1 N G i has c o n s t a n t dimension
around x O. []

(2.7) Remark. Note, from the proof of Lemma (2.2), that an obvious
necessary condition for the existence of a solution to the Input-
Output Linearization Problem is that the sequence of codistributions
~k g e n e r a t e d by m e a n s of the A l g o r i t h m IV.(2.5) coincides with the
one generated by means of the (simpler) algorithm (2.6).

(2.8) Lemma. For all k > 0

k+l k+l
(2.9) dim ~ = dim( ~ sp{dYi} ) = ~ 6.
i=1 i=I l

Proof. The first equality follows directly from (2.3),(2.4) and (2.5).
The second one is a c o n s e q u e n c e of the fact that the r k = 61+...+6k+i
rows of Sk+1(x) are linearly independent at each x in a n e i g h b o r h o o d
of x O. []

From these Lemmas one may deduce a series of i n t e r e s t i n g con-

clusions. First of all, the c o m p a r i s o n of (29) with IV.(3.22) shows
the c o i n c i d e n c e of the 61~ s d e f i n e d by means of IV. (3.23) with the
6~ s defined by means of the S t r u c t u r e Algorithm. Since the latter
o p e r a t e s on data a s s o c i a t e d with the i n p u t - o u t p u t behavior (the
matrices Tk(X)) , it follows that at least in the case of systems in
which the I n p u t - O u t p u t Linearization P r o b l e m has solutions, the in-
tegers IV.(3.23) have an i n t e r p r e t a t i o n in terms of i n p u t - o u t p u t data.
AS a m a t t e r of fact, there is an e x p l i c i t formula relating the 6'k s
to the m a t r i c e s Tk(X)'S. Following a procedure similar to the one
suggested in the proof of Lemma (1.17), one may arrive at the con-
clusion that

~(Mk) = (k+I)61 + k 6 2 + " "'+~k+1

 195

or, in other words, that

8 1 = rK(M0)

k+1
~i = ~ ( M k ) - ~(Mk-1) k ~ I
i=I

Since by d e f i n i t i o n 6 i = 0 for i > q (the last n o n d e g e n e r a t e

stage of the S t r u c t u r e Algorithm) and 8 ~ 0, one d e d u c e s from (2.9)
q ,
that ~q-1 _
~ ~q-2 and, therefore, that the integer k (which c h a r a c t e -
rizes the last m e a n i n g f u l stage of the A l g o r i t h m IV. (2.5)) is r e l a t e d
to q by the i n e q u a l i t y

(2.10) k > q-1

A s u f f i c i e n t c o n d i t i o n for (2.10) to become an e q u a l i t y is the

following one.

(2.11) Lemma. If the n u m b e r of rows of S is equal to , then k = q-1.

q
Proof. Suppose the n u m b e r of rows of S is equal to . Then the algo-
q
rithm (1.15) t e r m i n a t e s at the q - t h stage. F r o m L e m m a (2.2) we deduce
that

~q = ~q-1
%
i.e. that k = q-1. []

The case in w h i c h the a s s u m p t i o n of this Lemma holds (namely,

the case in w h i c h the a l g o r i t h m (Io15) t e r m i n a t e s at a finite stage)
deserves a special attention, b e c a u s e of some i n t e r e s t i n g p r o p e r t i e s
that will be p o i n t e d out hereafter.

(2.12) L~mma. If the n u m b e r of rows of S is equal to ~, then, in a

q
neighborhood of x , the d i s t r i b u t i o n

q-1 - I
= q
N sp{ dxi }J- A n sp{dxi }-
i=I i=I

coincides w i t h the l a r g e s t locally c o n t r o l l e d i n v a r i a n t d i s t r i b u t i o n

contained in H and any pair of f e e d b a c k f u n c t i o n s ~ and ~ w h i c h solves
the e q u a t i o n s (1.18) is such as to m a k e A invariant.

Proof. The first p a r t of the s t a t e m e n t is a c o n s e q u e n c e of (2.3) and

Lemma (2.11). The second part may be p r o v e d e x a c t l y as done in the
last p a r t of the proof of P r o p o s i t i o n IV. (3.19). []
 196

(2.13) Lemma. If the n u m b e r of rows of S q is equal to Z, then the dif-

ferentials of the e n t r i e s of the v e c t o r s Yi ' I ~ i ! q, and Yi '
I < i < q-l, are l i n e a r l y independent at x .

Proof. Let n i = -ri_ I d e n o t e the n u m b e r of e n t r i e s of Yi" We p r o v e

that if

P P
(2.14) dim( ~ sp{d~i(x) }) < ~ n i
i=I i=I

for some p, then all Yi

-' s w i t h i ~ p+1 are n o n t r i v i a l .
We k n o w from L e m m a (2.2) that the c o d i s t r i b u t i o n

P
~p-1 F] G i = [ sp{d{i }
i=I

has constant dimension around x - Thus y if (2 14) holds

t then there
exist k < p row v e c t o r s 1 1 , . . . , 1 k of s m o o t h real-valued functions
(whose dimensions are r e s p e c t i v e l y I X n l , . . . , 1 n k) d e f i n e d in a n e i g h -
borhood of x , w i t h Ik # 0, such that

(2.15) lk(X)d~k(X) = 11 (x)d{1 ( x ) + . . . + l k _ I (x)d~k_ I (x)

O
for all x a r o u n d x .
Differentiating (2.15) along f yields

k-1
(Lflk)d~ k + l k ( d L f ~ k) = ~ ((Lfli)d~ i + h i(dLf~i))
i=I

and also (see R e m a r k (2.1))

, + ,, - ,, k+1- ,, k+1
(Lflk)d{ k + lk(Qk+IdYk+1 Qk+IdYk+l-Qk+IK1 ~Y1-'''-Qk+IKk dYk) =

k-1
=i=I
[ ((Lf~ild~i+~iCQi+Id~i+1Q[Id~i1-Q~IK~IdC..-Q'~1~iIdYill
! i - , " "

This may be r e w r i t t e n as

k
(2.16) k(Qk+ldYk+l +Qk+ldYk+l ) = [ (PidYi + PidYi )
i=1

for suitable ~[s and ~ls.

We will see n o w that ~k+1 is n o n t r i v i a l and that there exist k+1
row v e c t o r s I ~ , . . . , I ~ + I , of s m o o t h real-valued functions defined in
a neighborhood of x , w i t h I~+ I ~ 0, such that
 197

(2.17) I~+ 1 (x)d~k+1 (x) = I~ (x)d{1 ( x ) + . . . + l ~ ( x ) d ~ k ( X )

To this end note that, bearing in m i n d (2.4) and (2.9),(2.16) yields

' = 0
IkQk+ I

pi = 0 I _< i _< k

I n"
k~k+l dYk+1 = PldYl+...+~kdYk

If Yk+1 were trivial, then Qk+1 = I and I k 0, i.e a contradiction.

Thus ~k+1 is n o n t r i v i a l and, also, I k ~n"k + I ~ 0 b e c a u s e otherwise the
equality

Ik (Qk+1 Qk+1 ) = 0

would contradict the n o n s i n g u l a r i t y of (Qk+1 Q{+I ) . T h i s shows that

(2.17) holds, with I~+ I = IkQk+ I and l! = Pi for I < i < k
l -- --

We can iterate this a r g u m e n t and c o n c l u d e that all ~s w i t h

i > k+1 are n o n t r i v i a l . If the a l g o r i t h m terminates at some step q,
then (2.14) is c o n t r a d i c t e d and the d i f f e r e n t i a l s of the e n t r i e s of
~i,...,~q_i are l i n e a r l y independent at x O. []

The above results enable us to i n v e s t i g a t e the e f f e c t of the

linearizing feedback on the state-space description of the system.
From the last L e m m a it is seen that the e n t r i e s of i(x) and ~i(x)
are p a r t of a local coordinate system. Thus, one m a y set

~i = Yi(x) I < i < q

~i = ~i (x) I _< i < q-1

and find a s u i t a b l e vector-valued function q with the p r o p e r t y that

the m a p p i n g

x I > (1 . . . . . q'~1 . . . . . ~q-1 ' q )

is a local coordinate transformation.

The description of the s y s t e m

i = ~(x) + ~(x)v
y = h(x)
 198

in the new coordinates may be easily obtained in the following way.

Consider the right-hand-side of (1.18b) and let EI,..,E q be ~im
matrices which partition [I 0] as

(2.18) [I~ 0] =
IiEIo
Eq

Then, if ~(x) and ~(x) are solutions of (1.18), one has

L%y i = 0
I

L~y i = E i
g

for I ~ i ~ q. These yield for ~I,...,~ q the equations

(2.19a) ~i = ~i = L~Yi + L%7iv = Eiv

g

Moreover (see Remark (2.1)),

L~i = Lf~i = Qi+1Yi+1' +Qi+1(~i+1" -~I'i+II "''-~i~i+Iyi )"

L~y i = 0
g

for I < i < q-2. For i = q-1

L~yq_1 = q

L~yq_I = 0

From these one gets

(2.19b) ~i = Qi+1~i+1-i+1~i+1- QlIK~+1~1-...-O'iiKi+li~i

for I < i < q-2 and

(2.19c)
~q-1 = ~q"

The output y is related to ~I and ~I in the following way

 199

(2.19d) Y = QI~1 + Q~'~I

combining the (2.19) 's, one finds in the new coordinates a state
space description of the form

~. = Fz + Gv

fi = f ( z , q ) + g(z,n)v

y =Hz

with
Z = C o I ( C I ..... C q ' ~ l ..... ~q-1 )

and

0 0 0 ... 0 0 0 0 0 ... 0

. ~ . o . .

0 0 0 ... 0 0 0 0 0 ... 0

"K21 Q~_ 0 " " 0 0 0 Q~ 0 " " " 0

~Q 2
F =
. 3 _,,K 3 Q~...
-Q3K1 -(23 2 Q~ . .. 0 0 0 0 0

. , . . , o . .

-Q~_~?-~ -Q~_~-~- Q ~ _ ~ ,, ~-~

... ,
Q~_~ o o o o... ,,
%_~

o o o . . . o ~o o o..o

EI

E
q
G =
0

:~Q; o ... o o" o . . . o]

 200

The p a r t i c u l a r choice of f e e d b a c k makes A invariant (see L e m m a

(2.11)) and this is the reason for the p r e s e n c e of an u n o b s e r v a b l e
(and nonlinear) subsystem. The other subsystem, which is the only one
contributing to the i n p u t - o u t p u t response, is fully linear.
We c o n c l u d e the section w i t h two remarks, which are c o n s e q u e n c e s
of the above result.

(2.20) Remark. From the above equations, we see that if the A l g o r i t h m

(1.15) terminates at the q-th stage (i.e. if the number of rows of Sq
is equal t o the n u m b e r ~ of outputs), the r e s p o n s e of the c l o s e d loop
system becomes
t
y(t) = HeFtz + I HeF(t-T)Gu(T)d~
0

The input-dependent part is linear in the input, as expected,

but also the zero-input term is linear in the initial state z O.

(2.21) Remark. The s t r u c t u r e of the m a t r i x F which characterizes the

linear p a r t of the c l o s e d - l o o p system shows that all its e i g e n v a l u e s
are vanishing. Thus, one m i g h t w i s h to add an a d d i t i o n a l feedback in
orde r to achieve not only a linear input-output behavior, but a
linear and stable i n p u t - o u t p u t behavior, if possible. As a m a t t e r of
fact, the p a i r of m a t r i c e s (F,G) turns out to be a reachable pair and
so a m a t r i x K may always be found w h i c h assigns the s p e c t r u m to F+GK.
In o r d e r to o b t a i n a linear input-output behavior with prescribed
spec t r a l properties, instead of the feedback a(x) and B(x) proposed
so far, one has to c o n s i d e r the feedback

a' (x) : a(x) + $(x)Kz(x)

S'(x) = S(x)

The r e a c h a b i l i t y of the pair (F,G) may be c h e c k e d by d i r e c t com-

putation of the rank of (G FG F2G ...). At e a c h stage, it is sug-
gested to take a d v a n t a g e of the n o n s i n g u l a r i t y of [Qi' O'~]
- l in o r d e r
to p r o v e that n e w linearly independent columns are added. []

3. Some A l g e b r a i c Properties

In this section we analyze the s t r u c t u r e of the formal power

series (1.7) with some detail, and show that the i n t e g e r s 61,...,6 q
are r e l a t e d to the b e h a v i o r of T(s,x) for s ~ ~.
 201

In the proof of Theorem (1.11), we have shown that the existence

of a solution of the I n p u t - O u t p u t Linearization Problem at x m a k e s
it possible to separate T(s,x) as a product of two formal power series
as in (1.12). In particular, it was shown that, if ~ and ~ are a line-
arizing feedback, then

K(s) = ~ (L L~h)s -k-1

k=0

(L~L~h being i n d e p e n d e n t of x for all k >_ 0 ) and

R(s,x) = ~-1(x) -k=~0 (LgLk~-1~(x))s-k-1

Clearly, (1.12) holds in the n e i g h b o r h o o d U of x O where the feedback

and 6 is defined.
k
An e x p l i c i t expression for L~L~h, that is for K(s), is not dif-
g x
ficult to obtain. For, consider again the m a t r i x H defined in (I .23)
and let ~ and 6 be any solution of (I .18). Simple c o m p u t a t i o n s , b a s e d
on appropriate use of the p r o p e r t i e s (I .20), yield

P2YI
2-
Hh = P3K2YI

q 1

2
p - 2
L Hh = (2-KIY1)

Kq 3 - 2
q "'" K3(Y2-KIY I)

0
0
L~Hh = Y3

Kq 4 - 3 3
q "'" K 4 (Y3-KIYI-K2Y 2)
 202

and so on, until

Lq-IHh =
f Yq

. . q_IYq_1

and

~ +iHh = 0
0
- _K~+i+1 q+i+1
Yq+i+1 YI-'''-Kq Yq

which holds for all i > O.

Differentiation of these along gl,...,g m enables us to obtain the
expression o f H(L~L~h) for all k ~ 0. Use o f t h e . p a r t i t i o n (2.18) makes

it p o s s i b l e to get

E1
0

HL~h = 0
g

E2

HL L h= 2
-P3KIEI

_Kq 3 2
q "'K3KIEI

and so on, until

 203

q-1
0

Eq

-K~EI-... -Kq_ iEq_ I

and, for all i > 0,

q+i
0

0
-K? +i+ 1E I -"" "-Kq+i+ IE
q q

Since El, .... Eq are rows of the matrix [ Iq 0], one easily under-
stands that the formal power series
co
(3.1) W(s) = H ~ (L Lkh)s -k-1
k=0 f

displays the following pattern of elements

- -I
I61s 0 ... 0 0

16 s -2 ... 0 0
2

w31(s) 0 ... 0 0

(3.2) W(s)= w41(s) W42(S) ... 0 0

o,

Wq, I (s) Wq,2(s) . .. I8 s--q 0

q

Wq+1,1 (s) Wq+1,2(S) ... Wq+1,q_1(s) Wq+1,q(S) 0

We recall that the partition for the rows corresponds to a

partition of the output vector into q+1 blocks of dimensions
 204

~1,...,dq,(%-rq-1) , while the one for the colununs c o r r e s p o n d s to a parti-

tion of the input vector into q+1 blocks of dimensions ~1, ...,6q,(m-rq-1).

(3.3) Remark. Note that, if the A l g o r i t h m (1.15) terminates at the

q-th stage (i.e. if the number of rows of S is equal to the number
q
of outputs), the (q+1)-th b l o c k - r o w of the m a t r i x W(s) does not exist,
and the m a t r i x itself is right-invertible. []

From the previous expression for HL~L~h,~ one also sees that in

the j-th block column of (3.1), I < j < q, the largest power of s
appearing in any o f f - d i a g o n a l element is -(j+1). As a c o n s e q u e n c e , o n e
may conclude that Wij(s)sJ is a strictly proper formal power series.
This p r o p e r t y will be i m m e d i a t e l y used in the following way. Set

I 0 ... 0

0 I ... 0

-W31 (s)s 0 ... 0

(3.4) PI (s) =
-W41 (s)s 0 ... 0

-Wq+1,1 (s)s 0 ... I

and note that

I I~ s -I 0 I
P1(s)W(s ) = I
0 W22(s)

W22(s) being the l o w e r - r i g h t - h a n d (~-61)(m-61) submatrix of W(s).

The power series P1(s) is proper (because of the a f o r e m e n t i o n e d
property of -W31 (s)s,...,-Wq+1,1 (s)s) and its inverse too. A proper
formal power series whose inverse is also proper is called a bipro-
per power series. Thus, we have that the power series (3.4) is bi-
proper.
Continuing this process, one c a n find b i p ~ o p e r formal power
series P2(s),...,Pq(S) which reduce W(s) to a p u r e l y diagonal form,
and prove the following interesting result.
 205

(3.5) Theorem. Suppose the s y s t e m (1.1) is such that the I n p u t - O u t p u t

L i n e a r i z a t i o n P r o b l e m has a s o l u t i o n at x . Then there e x i s t a bi-
proper formal p o w e r series

-k-1
R(s) = R_1(x) +kk0Rk(X)S

w h o s e c o e f f i c i e n t s are m x m m a t r i c e s of smooth f u n c t i o n s d e f i n e d on
a n e i g h b o r h o o d U of x , and a b i p r o p e r formal power series

L(s) = L_I +k~0Lks-k-1

whose c o e f f i c i e n t s are Z m a t r i c e s of real n u m b e r s such that

(3.6) T(s,x) = L(s)A(s)R(s,x)

where

13.7) A(s) = diag I l ' 2 7"'1 .,I q

Proof. The formal p o w e r series

P(s) = Pq(S)...PI(S)H

is biproper, b e c a u s e each P. (s) is and H is invertible. On the o t h e r

l
hand,

Pq(S)...Pl (s)W(s) = A(s)

and thus (3.6) follows from

L(s) = p-1 s). []

A f a c t o r i z a t i o n of the form (3.6) r e v e a l s the b e h a v i o r of T(s,x)

as s ~ ~. As a m a t t e r of fact, the limits of L(s) and R(s,x) for s ~
are n o n s i n g u l a r ; in i-th set of diagonal e l e m e n t s of A(s) , each func-
tion has a zero of m u Z t i p l i c i t y i at the infinity. For this reason,
the s t r i n g {61,62,...} is k n o w n as the structure at the infinity of
the formal power series T(s,x), or of the system (1.1). Note that
the s t r i n g {61,62,... } is u n i q u e l y a s s o c i a t e d w i t h T(s,x) and does
not d e p e n d on the p a r t i c u l a r p r o c e d u r e chosen to o b t a i n a f a c t o r i z a -
 206

tion of the form (3.6).

(3.8) Remark. W e have seen b e f o r e that the integers 61,~2,... are

related to the d i m e n s i o n s of the c o d i s t r i b u t i o n s ~0,~i,... generated
by m e a n s of the C o n t r o l l e d I n v a r i a n t D i s t r i b u t i o n A l g o r i t h m (see
Lemma (2.8)). In o t h e r words, we have

n0
(3.9a) 61 = dim ]_
n o AG

n. hi_ 1
(3.9b) 6i+ I = dim l dim i > I
~i n G i ni_l n G~

Since ~0,nl,... and G are i n v a r i a n t under f e e d b a c k t r a n s f o r m a t i o n s

(see L e m m a IV. (2.8)), it turns out that the s t r u c t u r e at the i n f i n i t y
of a system is i n v a r i a n t u n d e r f e e d b a c k t r a n s f o r m a t i o n s .

4. L i n e a r m o d e l m a t c h i n g

In the first section of this C h a p t e r we have seen that, under

s u i t a b l e conditions, it is p o s s i b l e to s y n t h e s i z e a feedback under
w h i c h the i n p u t - d e p e n d e n t p a r t of the response of a given n o n l i n e a r
s y s t e m b e c o m e s the same as that of a linear system. Our aim was the
one of a c h i e v i n g a response of the form (1.3), w i t h o u t any p a r t i c u l a r
p r e s c r i p t i o n on the first order k e r n e l s ki(t) , I ~ i _< m. As a m a t t e r
of fact, the t r a n s f e r function K(s) o b t a i n e d for the l i n e a r i z e d part
of the response, whose form was a n a l i z e d in the p r e v i o u s section,
happens to d e p e n d on the p a r t i c u l a r choice of feedback, i.e. on the
p a r t i c u l a r m a t r i c e s Pi,K %,...,Ki_
Ii s e l e c t e d at each stage of the
S t r u c t u r e Algorithm.
The p u r p o s e of the p r e s e n t s e c t i o n is to discuss a m o r e d e m a n d i n g
problem, the one in w h i c h a prescribed linear i n p u t - o u t p u t b e h a v i o r
rather than some linear i n p u t - o u t p u t b e h a v i o r is sought. We tackle
this new s y n t h e s i s p r o b l e m in a more general s e t t i n g than before,
letting the s t a t e - f e e d b a c k to be dynamic rather than static. This
means that we let u. to be r e l a t e d to the state x and, possibly, to
l
o t h e r i n p u t v a r i a b l e s V l , . . . , v ~ by m e a n s of e q u a t i o n s of the form

(4.1a) % = a(~,x) + [ bj (~,x)vj

j=1
P
(4.1b) u i = ci(~,x) + ~ d (~,x)vj
i=I z3
 207

These e q u a t i o n s characterize a new dynamical system, whose state

evolves on an open subset of ~ . As usual, we assume that all func-
tions which c h a r a c t e r i z e these equations are smooth functions, de-
fined now on a subset of ~ 9 x ~ n. Most of the times, we shall con-
sider bj(~,x) as the j-th column of a ~ ~ m a t r i x b(~,x),ci(~,x) as
the i-th row of an m1 vector c(~,x) and dii (~,x) the (i,j)-th entry
of a m a t r i x d(C,x). Note that the number p of new inputs may be dif-
ferent from m.
The c o m p o s i t i o n of (4.1) with (1.1) defines a new dynamical
system, with input v = c o l ( v 1 , . . . , v ), output y = col(y I ..... y)
d e s c r i b e d by equations of the form

(4.2a} =f(~,x) + ~ gi(~,x)v i

i=I

4.2b) Yi = hi (~'x)

in which

f(~,x) =
I a(~,x)m
f(x) + [ gi(x)ci(~,x)
i=I

bi(~,x)
gi (~,x) = m
j~igj(x)dji (~'x)

hi(~,x) = hi(x)

The integer ~, which c h a r a c t e r i z e s the dimension of the dynamic-

al system (4.1),and the q u a d r u p l e t (a,b,c,d) are to be chosen in such
a way as to obtain, for the closed loop system (4.2), an i n p u t - o u t p u t
response of the form (see (1.3))

C
t
(4.3) y(t) = Q[t,(~,x)) + j WM(t-T)V(T)dT

0
WM(t) being a ~ixed %p m a t r i x of functions of t, the impulse-re-
sponse m a t r i x of a p r e s p e c i f i e d linear model. As before, we seek
local solutions defined in a n e i g h b o r h o o d of the initial state. In
view of our earlier discussions, this yields the following formal
 208

statement.

Linear Model M a t c h i n g Problem. Given (f,g,h), an initial state x , and

a linear model (A,B,C), find (if possible) an integer w, an initial
state ~o E ~ , a quadruplet of smooth functions (a,b,c,d) defin&d in a
neighborhood U of (~,x) such that for all k > 0

(4.4) L~L~h(~,x) = cAkB. []

If the system (1.1) is such that a solution to the Input-Output

Linearization Problem exists, then it is quite simple to find the
extra conditions needed for the existence of a solution to the Linear
Model Matching Problem and to construct such a solution. The main tool
is again the Structure Algorithm described in the first section.
The data of a Linear Model Matching Problem are, besides the
initial point x O, the triplet (f,g,h) which characterizes the system
to be controlled and the triplet (A,B,C) which characterizes the model
to be reproduced. These data will be used in order to define an ex-
tended system, described by the following set differential equations

= f(x) + g(x)u

(4.5) z = Az + Bv

w = h(x) - Cz

The output w of this system is actually the difference between the

output of the system (1.1) and that of the model. For convenience,
we represent (4.5) in the form

~E = fE(xE ) + gE(xE)uE

w = hE(x E)

letting x E= col(x,z), u E =col(u,v) and

g (x, z) = I g(x) 0 I
(4.6a) E
fE(x'z) = I Azf(X)
0 B

(4.6b) hE(x,z) = h(x) - Cz

 209

The c o n d i t i o n s for the e x i s t e n c e of a s o l u t i o n to the Linear Model

M a t c h i n g P r o b l e m may easily be e x p r e s s e d in terms of p r o p e r t i e s of the
system thus defined, as we will see hereafter.
S u p p o s e the s y s t e m (1.1) is such that the I n p u t - O u t p u t L i n e a r i z a -
tion P r o b l e m has a s o l u t i o n at x . Then, the triplet (f,g,h) is such as
to fulfill the c o n d i t i o n (a) of T h e o r e m (1.11). It is easily seen that,
for any z , also the triplet (fE,gE,hE) is such as to fulfill a similar
c o n d i t i o n at (x,z). For, let

k cAkB]
T~(x,z) = L E L k E h E (x, z) = [LgLfh(x)
g f

Then

oo co

T kE(x,z)s-k-1 = [T(s,x) [ cAkBs -k-l] =

k=0 k=0

= [K(s)

where

co

WM(S) = ~. cAkBs -k-1

k=O

denotes the transfer function of the model.

As a c o n s e q u e n c e of this, the A l g o r i t h m (1.15) may also be per-
formed on the triplet (fE,gE,hE), around the p o i n t (x,z), and one may
define on the formal power series TE(s,x) a s t r u c t u r e at the infinity,
E E
c h a r a c t e r i z e d by a s t r i n g of integers {~1'~2'''" }.
The c o i n c i d e n c e b e t w e e n the structure at i n f i n i t y of the formal
p o w e r series T(s,x) and that of the formal p o w e r series TE(s,x) is
e x a c t l y the c o n d i t i o n that c h a r a c t e r i z e s the p o s s i b i l i t y of s o l v i n g
a L i n e a r Model M a t c h i n g Problem. In order to be able to prove this
r e s u l t and give an e x p l i c i t c o n s t r u c t i o n of the r e q u i r e d feedback, we
n e e d a little more notation.
i
Let Pi,KI,. .. 'K i
i be the set of m a t r i c e s d e t e r m i n e d at the i-th
stage of the S t r u c t u r e A l g o r i t h m , w h e n o p e r a t i n g on the triplet (f,g,h).
Let the triplet (A,B,C) characterize the model to be followed. We set

(4.7a) C I = PI C
 210

(4.7b) CI = K~C

and, for i > 2

(4.7c) C i = PiCi_1A

(4.7d) Ci = K ~ C I + " "" +Ki-1 C i - I + K ii-

Ci-IA

W i t h the f u n c t i o n s X 1 ( x ) , . . . , y q ( X ) d e t e r m i n e d at e a c h n o n d e g e n e -
rate stage of the A l g o r i t h m we associate, as before, a matrix

r(x) = -71(x)

7q(X)

and w i t h the m a t r i c e s C I , . . . , C q d e f i n e d above we a s s o c i a t e the m a t r i x

-C I

(4.8) D =

The c o n s t r u c t i o n s d e f i n e d above are helpful in f i n d i n g a solution

tothe p r o b l e m in question.

(4.9) Theorem. Suppose the s y s t e m (1.1) is such that the I n p u t - O u t p u t

L i n e a r i z a t i o n P r o b l e m is solvable at x O. The Linear Model M a t c h i n g
P r o b l e m is solvable at x if and only if e i t h e r one of the f o l l o w i n g
equivalent conditions is s a t i s f i e d

(a) C B = 0 for all i > I

(b) the s y s t e m (1.1) and the e x t e n d e d s y s t e m (4.5) are c h a r a c t e r i z e d
by the same s t r u c t u r e at the infinity.

A dynamical s t a t e - f e e d b a c k w h i c h solves the p r o b l e m is the one

d e s c r i b e d by the f o l l o w i n g e q u a t i o n s

(4.10a) ~ = A~ + Bv

(4.10b) u = ~(X)-S(x)DA~-8(x)DBv
 211

in which ~(x) and B (x) are solutions of

(4.11a) LgF(X)~(x) = -LfF(x)

(4.11b) LgE(X)~(x) = Ir
q-1

with D defined as in (4.8). The initial state C of (4.10) may be set

arbitrarily.
Proof. (a) ~ (b). Consider the sequence of functions thus defined

y~(x) = PIhE(x)

-E I E
Y1 (x) = K1h (x)

and, for i > 2,

y~(x) = PiLfE i-I

~E (x)

~E(x) i E(x)+. i E (x) i -E

=KiY I . .+Ki_iYi_ I +KiLfEYi_1 (x)

Note, also, that for all i > I

E(x) = Yi(x) + Ciz

~(x) = ~i(x) + 6i z

Suppose that, for all I < k < i,

Ck B = 0

Then, for all I < k < i,

-E
L EYk(X) = 0
g
and the matrix
E
L EY1(x) Ii LgYI (x)
g

L EYk
g
E(x) LgYk (x)
cIB1
CkB
 212

has a rank equal to the number 61+...+6 k of its rows, at (x,z O) (note
that z O is irrelevant) As a consequence, one may conclude that the
first i steps of the Structure Algorithm on the triplet(fE,gE,h E) may
be performed exactly in the same way as on the triplet(f,g,h). At each
of these steps, the same set of matrices Pk,K~ ..... K~ makes it pos-
sible to perform the required operations. In particular, since the
integers which characterize the structure at the infinity do not de-
pend on the choice of matrices in the Structure Algorithm, we see that
the first i entries in the structure at the infinity of (f,g,h) and
(fE,gE,hE) coincide. Now, let Vi+ I be the matrix determined at the
(i+1)-th stage of the A l g o r i t h m (1.15) and observe that

E E
L EY1(x ) LgY1(x) CIB Lgy I (x) CIB
g
"E(xE )
Vi+1 L Ey i = Vi+ I Lg~i(x) CiB = LgYi+ I Ci+IB
g
L _L _ ~ (xE) ngLf~ i (x) CiAB 0 Ci+IB
g~ f~ I

From this we see that the (i+1)-th entries in the structure at the
infinity of (f,g,h) and (fE,gE,hE) coincide if and only if

Ci+1B = 0

Sufficiency
one ends up with

FE (xE) =
of

E
(b). At the last nondegenerate
a matrix

YI (xE) ~1 (x)

ClZ

step of the algorithm

= (r(x) Dz)
YqE (xE ) yq (x) CqZ

of rank rq_i=61+...+6 q at (x,z). Consider now the two equations

(4.12a) L E FE(x E) E ( x E) = _LfEFE(x E)

g

(4.12b) L EFE(xE)BE(x E) = (I r 0)
g q-1

which correspond to the equations (1.18). If ~(x) and ~(x) are solu-
tions of (4.11), solutions ~E(xE) and ~E(xE) of (4.12) may be found as
 213

~(x)-~(x)DAz
aE(x,z) =

~(x) -8(x)DB
8E(x,z) =
0

Note that B is m r q _ I a n d that ~E is (m+~)(rq_1+~).

Now, suppose the functions E and B E thus defined are used in a
static state-feedback loop on the extended system (4.5). As a conse-
quence of all previous discussions, we get

(4.13) L EsE Lk h E (x E) = independent of x E

g fE+gEE

for all k > 0 around (x,z), for any z. In p a r t i c u l a r (see e.g. the
structure of (3.2)), the last ~ columns of these matrices vanish
for all k > 0.
The extended system (4.5) subject to the feedback thus defined
is d e s c r i b e d by equations of the form

= f (x) +g (x) ~ (x) -g (x) $ (x) D A z + g (x) 8 (x) u - g (x) 13(x) D B v

(4.14) z = Az + B~

w = h(x) - Cz

where u and v represent new inputs. The response of this system con-
sists of a " z e r o - i n p u t " t e r m w0(t, (x,z)) and of a linear term in
alone, because, as w e observed, the last ~ columns of (4.13) are
vanishing. This means that, if u = 0, the response of such a system
consists of w0(t, (x,z)) alone.Equations (4.14) with u = 0 may be
interpreted as the composition of the original system (1.1) with the
dynamic feedback

-- Az + B ~

u = ~(x)-B(x)D(Az+Bv)
 214

together with the a n e w o u t p u t map

w = y - Cz

For all i n i t i a l states around (x,z ) and all inputs

w(t) = w0(t, (x,z))

and, therefore,

t
y(t) = w0(t , (x,z))+ceAtz + I ceA(t-T)Bv(T)dT
0

This shows that the r e s p o n s e of the s y s t e m (1.1) under the feed-

back (4.10) has the d e s i r e d form (4.3).
Necessity of (a). T h i s part of the p r o o f consists in a r e p e a t e d
use of the e x p r e s s i o n s which define Yi(x) and ~i(x), in o r d e r to show
that

(4.15) L~Lkh(~,x) = cAkB

for all k > 0, imply

C B = 0
1

for all i > 0. One p r o v e s first that (4.15) implies

(4.16a) [LgLf~i(x)]d(~,x) = ~iAB

and t h a t this, in turn, implies

(4.16b) [LgYi(x)]d(~,x) = CiB

for all i > 0. Then, (4.16) imply the d e s i r e d result, because

i
0 = (Lg~i)d = ( L g [ K i y 1 + . .+ K l _ i Y i _ 1 + K iiL f Y-i _ I] ) d =

i ) d+. i d+K i - )d =
= K I (Lgy I - .+Ki_ I (LgYi_ I ) (LgLfYi_ I

= (K C I + . . . + K i _ I C i _ I + K i C i A ) B = Ci B

This completes the proof. []

 215

(4.17) Remark. F r o m the above statement, we see that a d y n a m i c state-

feedback which solves a Linear Model Matching Problem may easily be
found in terms of data related to the s o l u t i o n of an I n p u t - O u t p u t
L i n e a r i z a t i o n Problem. As a m a t t e r of fact, the a v a i l a b i l i t y of
Pi'KIi ''''' K i i' 1 ~ i ~ q makes it p o s s i b l e to c o n s t r u c t the m a t r i c e s
Ci,C i , I ~ i ~ q and, then, to c h e c k the e x i s t e n c e condition (a). If
this is satisfied, one takes any s o l u t i o n ~(x) and ~(x) of (4.11)
(i.e. any f e e d b a c k solving the L i n e a r i z a t i o n Problem) and c o n s t r u c t s
a solution of the Model Matching in the form (4.10).

(4.18) Remark. In the p r e v i o u s procedure, no special attention was

paid to the p r o p e r t i e s of the zero-input term Q(t, (~,x)), which
represents the e f f e c t of the initial states on the r e s p o n s e of the
closed loop system (4.2). If an a s y m p t o t i c a l l y decreasing zero-input
response is required, one should m o d i f y the o u t l i n e d construction and
use, instead of a s o l u t i o n ~(x) and 8(x) of (4.11), a feedback which
makes linear and asymptotically stable the i n p u t - o u t p u t behavior of
the e x t e n d e d system (4.5). This m a y be a c c o m p l i s h e d on the basis of
the ideas discussed in the Remark (2.21). []

5. More on L i n e a r Model Matching, Output Reproducibility and N o n i n t e r a c -

tion

In this section we will see that it is p o s s i b l e to solve a Linear

Model M a t c h i n g P r o b l e m even though the I n p u t - O u t p u t Linearization Pro-
blem is not solvable. In p a r t i c u l a r , we will see that the c o n d i t i o n
(b) of T h e o r e m (4.9) still implies the s o l v a b i l i t y of the L i n e a r Model
Matching problem, even in case the a s s u m p t i o n of s o l v a b i l i t y of the
Input-Output Linearization Problem does not hold.
To this end, note first of all that the so-called structure at
the i n f i n i t y can be a s s o c i a t e d with any s y s t e m of the form (1.1) and
not only with input-output-linearizable systems. This is b e c a u s e the
string of i n t e g e r s {61,62,...}, that we i n t r o d u c e d by m e a n s of the
Structure Algorithm, can also be i n d e p e n d e n t l y defined in terms of
dimensions of the c o d i s t r i b u t i o n s generated by m e a n s of the C o n t r o l l e d
Invariant Distribution Algorithm (see e.g. (3.9)). For this to be pos-
sible, it is only required that G, ~k and ~k N G have constant di-
mension, for all k > 0, a r o u n d the p o i n t x , i.e. that x is a regular
point for this a l g o r i t h m (see c h a p t e r IV, section 3).
If this is the case, one m a y a s s o c i a t e with the triplet (f,g,h)
the s e q u e n c e of integers
 216

~k
(5.1) r k = dim G~ k h 0
~k n

Given also a linear m o d e l (A,B,C), we may a s s o c i a t e w i t h the e x t e n d e d

triplet (fE,gE,hE) (see (4.5)), a similar sequence of i n t e g e r s

E
E ~k
(5.2) r k = dim E r] G EL k >_ 0
k

w h e r e now the s u p e r s c r i p t "E" d e n o t e s o b j e c t s p e r t i n e n t to the e x t e n d e d

system, namely

GE E E E E ]
= s p { g 1 ' ' ' ' ' g m ' g m + 1 ..... gm+p

QO
=E (sp{dh~ ..... dh~])

m+~
E E LfE(~ K A GEL)" [ L E(ak_1 A G El)
9k = ~k-1 + -I +
i=I gi

The s t r u c t u r e at the i n f i n i t y {61,62,...} of a s y s t e m of the form

(1.1) is u n i q u e l y r e l a t e d to the sequence {rl,r2,...} and, therefore,
the e q u a l i t y b e t w e e n the s t r u c t u r e at the i n f i n i t y of the system (1.1)
and that of the e x t e n d e d s y s t e m (4.5) (i.e. the c o n d i t i o n (b) of
Theorem (4.9)) is e q u i v a l e n t to the e q u a l i t y

E
rk = r k

for all k > 0.

We prove now that this is still s u f f i c i e n t for the s o l v a b i l i t y
of the p r o b l e m in q u e s t i o n .

(5.3) Theorem. Suppose x O is a r e g u l a r p o i n t of the A l g o r i t h m IV.(3.17)

for the triplet (f,g,h) and (x,z ) is a regular p o i n t of the A l g o r i t h m
IV. (3.17) for the triplet (fE,gE,hE). Then the L i n e a r Model M a t c h i n g
P r o b l e m is solvable at x if

(5.4) rk = r k

for all k > 0.

Proof. We first e s t a b l i s h some n o t a t i o n s . Let u denote the d i m e n s i o n

of the l i n e a r model (A,B,C). T h r o u g h o u t this proof we will be inter-
 217

ested in some d i s t r i b u t i o n s and/or codistributions defined around

the p o i n t (x,z ) of ~ n ~ u in the f o l l o w i n g way. We set

E E}
Gu = sp{gl ..... gm

Gv s { E E }
= p gm+1,---,gm+u

and we n o t e that

(5.5) GE = Gu Gv

Moreover, we d e f i n e a sequence of c o d i s t r i b u t i o n s ~k ' k _> 0, as

~k(X,Z) = ~k(X) X{O] C T x* lgn x T z*jR~

It is easy to v e r i f y that the s e q u e n c e of c o d i s t r i b u t i o n s thus

defined is such that

m+ U -i
(5.6) ~k+l = ~k + ~ T. ~.(~k n Gu)
i=0 gi

(with g~ = rE) and also that

~k ~k
(5.7) rk = dim ~ = dim _~
~k n G ~k ~ Gu

Finally, we d e f i n e another codistribution F as

r(x,z) = {o} T ~Z ~

We p r o c e e d now with the proof, which is d i v i d e d into three steps.

(i) It w i l l be shown that the a s s u m p t i o n (5.4) implies

(5.8) -~ D ~kE n ~u
Gv

for all k > 0. To this end, note first that the a s s u m p t i o n (5.4),
because of (5.7), m a y be r e w r i t t e n as
E GEL
~k + S~ ~k +
(5.9) dim( T ) = dim(--)
= G E~
U

Suppose now that

 218

E
(5.10) ~k + F = ~k +

for some k. Then, we m a y d e d u c e the f o l l o w i n g implications

E -k E + -i -L
~k + F = ~2k + F ~ ~k + G u = e k Gu (because r C Gu)
E -L
~k + G u ~kE + G E L
dim - - = dim (by (5.9))
-i
G GEL
U

d i m ~kE n G ~u = d i m ~kE n G E X

E ~ ~i = E n G E i (because G EL C ~ )
~k u ~k u "

The condition (5.10) also implies (because F C G~)

(5.11) -L
Gu n +r=
au n +r) = S Lu n (~ +r) =
E
n ~k+F

Thus, we h a v e

_ m+~ ~ L n ~k) + F by (5.6)

~k+1 + r = ~k + ~' L E ( u
i=0 gi

m+u
= ~k + X L E(Cu~ n ~k + r) + r
i=0 gi

m+p (-L N E
= ~k + ~ L E G u ~k + F) + F by (5.11)
i=0 gi

m+p
(G EL n ~kE + ) + (see above)
=~k +~. ~E
i=0 gi

= ~k + m +I~ L E (GEL n ~k) + F

i=0 gi

E m+p
E + F
= ~k + [ L E {GE L N ~2k ) + r = ~k+1 by (5.10)
i=0 gi

This shows that (5.10) holds also for k+1. Since (5.10) is true
for k = 0, the p r e v i o u s argument shows that it is true for all k > 0.
As a c o n s e q u e n c e , we h a v e also
 219

E n -~ E n G EL
~k G u = ~k

for all k > 0 and this, since G El C ~ k implies (5.8) Note that (5.8)
-- V "

m a y be r e w r i t t e n as

EL
(5.8') G V C Rk + Gu

(ii) S i n c e (x,z O) is a r e g u l a r point of the A l g o r i t h m IV.(3.17) for

the t r i p l e t (fE,gE,hE), there e x i s t s an i n t e g e r k* such that, in a
E E *
neighborhood of --(x,z
O) ~k = ~k* for all k ~ k . M o r e o v e r (see L e m m a
IV.(2.4)) the d i s t r i b u t i o n

AE, = ~
EL,
k

is s u c h that

[fE,AE*] C A E* + GE

[ gE, AE*] C AE* + GE 1 < i <_ m+u

From these, using (5.8'), we deduce that

(5.12a) [ fE,AE*] C b E* +
u

(5.12b) [ E ,E*I C b E* + G I < i < m

gi 'a ] u -- --

Since E
~k* n Gu
- = ~k*
E N GEL is n o n s i n g u l a r around (x,z ) so is
bE* + Gu" A l s o b E* and Gu are n o n s i n g u l a r and therefore one m a y use
Lemma IV.(I.10) and d e d u c e the e x i s t e n c e of an m1 vector a(x,z) of
smooth functions, defined locally around (x,z), such that

(5.13a) [ fE + E
gi~i , A E*] C A E*
i=I

Moreover, f r o m the c o n d i t i o n (5.8') one m a y d e d u c e the e x i s t e n c e

of an mxH m a t r i x (x,z) of s m o o t h functions, defined locally around
(x O z O) such t h a t

m
E E AE*
(5.13b) gm+i + ~ E I < i < P
j=1 gjYji -- --

Finally, note that, because of the i n v o l u t i v i t y of A E*, the a b o v e

 220

c o n d i t i o n implies

m
(5.13c) [ gm+i
E + [ E
gjxj i,A E*] C A E* 1 < i <
j=1

and recall that

(5. I 3d) ~E, c (sp{dh~. . . . . dh~})L

(iii) C o n s i d e r the d y n a m i c a l system

m p m
.E fE E E E
x = + ~. giai + ~ (gm+i + I g j i )vi
i=1 i=I j=l

w -- h E (xE)

This s y s t e m is such that the c o n d i t i o n s (5.13) hold. Thus,thanks

to T h e o r e m III.(3.12), we deduce that the inputs V l , . . . , v p have no
i n f l u e n c e on the o u t p u t w, i.e. that for all initial states (in the
n e i g h b o r h o o d w h e r e a(x,z) and X(x,z) are defined) the r e s p o n s e of this
s y s t e m c o n s i s t s of a z e r o - i n p u t t e r m w0(t, (x,z)) alone. Thus, by
m e a n s of the same a r g u m e n t s as the ones used at the end of the proof
of T h e o r e m (4.9), we c o n c l u d e that the c o m p o s i t i o n of the o r i g i n a l
system (1.1) w i t h the dynamic feedback

= Az + B Y

u = a(x,z) + y(x,z)v

has a r e s p o n s e of the form

t
y(t) = w0(t, (x,z)) + ceAtz O + I c e A ( t - T ) B v ( T ) d ~
0

This c o n c l u d e s the proof.

(5..16) Remark. The r e a d e r may e a s i l y check that the value of z is

i r r e l e v a n t in the p r e v i o u s discussions.

(5.17) Remark. We stress that the proof of the p r e v i o u s T h e o r e m is

constructive. The f u l f i l l m e n t of the c o n d i t i o n s (5.4) makes it pos-
sible to find, locally around (x,z), a v e c t o r a(x,z) such that
(5.13a) holds and a m a t r i x X(x,z) such that (5.13b) holds. A dynamical
s t a t e - f e e d b a c k w h i c h solves the p r o b l e m in q u e s t i o n is the one de-
 221

scribed by the equations

(5.18a) ~ = A~ + Bv

(5.18b) u = ~(x,~) + (x,~)v []

As an a p p l i c a t i o n of this Theorem, we deduce now an i n t e r e s t i n g

result which is r a t h e r useful in c o n n e c t i o n with problems of output
reproducibility and noninteraction (via d y n a m i c feedback).

(5.19) Corollary. Suppose rk~ = . T h e n there exists an integer 6 > 0

such that the L i n e a r Model Matching Problem is s o l v a b l e for a linear
model (A,B,C) with transfer function

I
0 ... 0 0 ... 0

1
0 ~ ... 0 0 ... 0
(5.20) W M(s)=C(sI-A)-IB =
... 0 ... 0

I
0 0 ... --~ 0 ... 0
s o

Proof. It is l e f t as an e x e r c i c e to the r e a d e r .

(5.21) Remark. Note that the transfer function (5.20) is r i g h t - i n v e r t -

ible. Thus, given any smooth i-vector-valued function y, defined on ~,
and such that

= (dt)0 = .. = (dtS_l) 0 = 0

there exists a smooth w-vector-valued function ~, defined on ~,such that

(t
y(t) = ] WM(t-T)5(T)dw. []

Now, suppose rk8 = i and suppose we have solved the p r o b l e m of

matching a linear model with transfer function (5.20). This means
that we have found an appropriate dynamic state-feedback compensator
(e.g. the o n e described in the p r o o f of Theorem (5.3), which has the
form (5.18)) under which the input-output behavior of the system (1.1)
be c o m e s
 222

t
(5.22) y(t) = Q(t, ({,x)) + I WM(t-~)v(T)dT
0

Let y be any smooth - v e c t o r - v a l u e d function, d e f i n e d on ~, such

that

(dkye(t) = (dkQ(t, (~,x)))0

(5.23)
dt k )0 dt k

for 0 < k < 6-I.

From the R e m a r k (5.21) we e a s i l y deduce that there exists an input
v under w h i c h the r i g h t - h a n d - s i d e of (5.22) becomes e x a c t l y y . Thus,
the c o m p o s i t i o n of (I .I) w i t h the dynamic s t a t e - f e e d b a c k c o m p e n s a t o r
w h i c h solves the p r o b l e m of m a t c h i n g the t r a n s f e r function (5.20) is
a s y s t e m that, in the i n i t i a l state (~,x), can reproduce any output
function w h i c h s a t i s f i e s the c o n d i t i o n s (5.23).
Moreover, we note that in a linear s y s t e m with t r a n s f e r f u n c t i o n
(5.20) each o u t p u t c o m p o n e n t is i n f l u e n c e d only by the c o r r e s p o n d i n g
c o m p o n e n t of the input. Thus, we also see that if the c o n d i t i o n rk~ =
holds, we can achieve non-interaction via dynamic state-feedback.

6. S t a t e - s p a c e linearization

In the first section of this chapter, we e x a m i n e d the p r o b l e m of

achieving, via feedback, a linear input-output response. The sub-
sequent a n a l y s i s d e v e l o p e d in the s e c o n d section showed that, from
the p o i n t of v i e w of a s t a t e - s p a c e description, in suitable local
coordinates, the s y s t e m thus l i n e a r i z e d assumes (at least in the
special case where rq_ I = ) the form

= Fz + Gv

= f(z,~) + g(z,~)v

y = Hz

In other words , the input-output-wise linear system one o b t a i n s

by means of the techniques in q u e s t i o n may be interpreted, at a
s t a t e - s p a c e level, as the i n t e r c o n n e c t i o n of a (possibly) nonlinear
unobservable s u b s y s t e m w i t h a s y s t e m that, in s u i t a b l e local c o o r d i n a -
tes, is s t a t e - s p a c e - w i s e linear. Moreover, the latter s u b s y s t e m was
also shown b e i n g both r e a c h a b l e and o b s e r v a b l e (Remark (2.21)).
 223

In o t h e r words again, we may say that the techniques developed at

the b e g i n n i n g of this chapter modify the b e h a v i o r of the o r i g i n a l
system in a w a y such as to m a k e a part of it (i.e. the o b s e r v a b l e one)
locally diffeomorphic to a r e a c h a b l e linear system.
Motivated by these considerations, we want to e x a m i n e now the
problem of m o d i f y i n g , via feedback, a given nonlinear system in a w a y
such that not simply a part, but the w h o l e of it, is locally diffeo-
morphic to a r e a c h a b l e linear system. In f o r m a l terms, the p r o b l e m
thus introduced may be c h a r a c t e r i z e d as follows.

State-Space Linearization Problem, Given a collection of v e c t o r fields

f'g1' .... gm a n d an i n i t i a l s t a t e x O, f i n d (if p o s s i b l e ) a n e i g h b o r h o o d
o
U o f x , a p a i r of f e e d b a c k f u n c t i o n s ~ a n d 8 (with I n v e r t i b l e 8)
defined on U, a coordinates transformation z = F(x) defined on U,a
matrix A 6 ~nn and a set of v e c t o r s b I @ ~n,...,b m E ~n such that

(6.1) F,(f+g~)oF-1(z) = Az

(6.2) F,(gS)ioF-1(z) = bi I ~ i ~ m

for all z 6 F(U), and

n-1 m
(6.3) [ ~ Im(Akbi ) : ~n
k=0i=l

(6.4) Remark. L e t x(t) denote a state trajectory of the system

m
= (f+g~) (x) + [ (gSi) (x)u i
i=1

and suppose x(t) 6 U for all t 6 [0,T] for some T > 0. If (6.1) and
(6.2) hold, then for all t 6 [0,T]

z(t) = F(x(t))

is a s t a t e trajectory of the linear system

m
~ = Az + ~. b.u.
i=I i

Moreover, if (6.3) also holds, the latter is a r e a c h a b l e linear

system. []

We shall describe first the solution of this problem in the

special case of a s y s t e m with a single input, which is r a t h e r easy.
Then, we m a k e some remarks about the usefulness of this linearization
 224

technique in p r o b l e m s of asymptotic stabilization. Finally, we con-

clude the section with the analysis of the (general) multi-input
systems.
For the sake of simplicity, we state some intermediate results
which may have their own independent interest

(6.5) Lemma. Suppose m = I and let g = g1" The State-Space Lineariza-

tion Problem is s o l v a b l e if a n d o n l y if t h e r e exists a neighborhood
V of x O and a function ~ :V ~ such that

(6.6) Lg~(X) = LgLf~(x) = ... = L g L n-2

f ~(x) = 0

for all x E V, and

(6.7) L g L n-1
f ~(x ) ~ 0

Proof. N e c e s s i t y . Let (A,b) a reachable pair. Then, it is w e l l known

from the theory of linear system that there exist a nonsingular nn
matrix T and a 1n r o w v e c t o r k such that

0 I 0 ... 0

0 0 1 .. 0

(6.8) T ( A + b k ) T -I = Tb=

0 0 0 ... 1

0 0 0 ... 0

Suppose (6.1) and (6.2) hold, and set

= ~(x) = TF(x)

a(x) = e(x) + $(x)kF(x)

(x) = ~ (x)

Then, i t is e a s i l y seen that

F=~(q~)oE '-1(~) =
 225

0 1 0 ... 0

0 0 I ... 0

#,(f+g~)oF-l(z) = . . , o .

0 0 0 ... 1

0 0 0 ... 0

From this, we deduce that there is no loss of g e n e r a l i t y in as-

suming that the p a i r A,b which makes (6.1) and (6.2) satisfied has
directly the form specified in the right-hand-sides of (6.8).
Now, set

z = F(X) = c o l ( z 1 ( x ) ..... Zn(X))

If (6.1) holds (with A a n d b in the f o r m of the right-hand-sides

of (6.8)), we have for all x 6 U,

F,(f(x)+g(x)a(x)) = AF(x)

that is

~z 1
~x ( f ( x ) + g ( x ) a ( x ) ) = z2(x)

SZn_ I
8--~(f(x)+g(x)a(x)) = Zn(X)

~Z
n
x-~--~--(f(x)+g(x)a(x)) = 0

If a l s o (6.2) holds we have

F%g(x) B(x) = b
that is
~z I
~x g(x)S(x) = 0
, ..

~Zn_ I
- x-~-~-- q ( x ) 6 ( x ) = 0

~z
~X n g(x)(3 (X ] = I
 226

Since ~(x) is n o n z e r o for all x E U, the second set of c o n d i t i o n s

imply

~z.
(6.9) ~xl g(x) = LgZi(x) = 0 I _< i _< n-1

8Zn I
(6.10) ~x g(x) = LgZn(X) =

for all x E U. These, in turn, together with the first set of c o n d i -

tions imply

(6.11) Lfzi(x) = zi+ I (x) I <i<n-1

~(X)
(6.12) L f Z n (x) = ~(x)

for a l l x E U.
If one sets

(6.13a) ~(x) = z I (x)

the c o n d i t i o n s (6.11) yield

(6.13b) zi+ I (x) = Lf~(x) 0 < i < n-1

Thus, from (6.9) one obtains

Lg~(X) = LgLf~(x) = ... = L g L fn-2 ~(x) = 0

for all x 6 U and, from (6.10),

Ln-1
Lg f ~(x) ~ 0

for all x 6 U. T h i s completes the p r o o f of the n e c e s s i t y .

Sufficiency. Suppose (6.6) a n d (6.7) are true and let U C V be a
n-1
neighborhood of x s u c h that LgLf ~(x) ~ 0 for all x 6 U. U s e (6.13)
in o r d e r to d e f i n e a set of f u n c t i o n s z l , . . . , z n on U. T h e functions
thus defined are clearly such that (6.9) and (6.11) hold. Moreover,
since LgZn(X) is n o n z e r o on U, one can define a nonzero function ~(x)
and a function e(x) by m e a n s of (6.10) and (6.12). This pair of func-
tions ~ and ~ and the m a p p i n g
 227

F ; x I ~ col(z1(x) ..... Z n ( X ) )

are c l e a r l y such t h a t

F,(f(x) + g(x)a(x)) = AF(z)

F,(f(x) 8(x)) = b

w i t h A a n d b in the f o r m of the r i g h t - h a n d - s i d e s of (6.8). Thus, in

o r d e r to c o m p l e t e the proof, we only have to s h o w t h a t F q u a l i f i e s as
a local coordinates transformation a r o u n d x O, i.e. t h a t its d i f f e r -
ential F, is n o n s i n g u l a r at x .
For, observe that the v e c t o r fields ~ = f+g~ a n d ~ = g~ are such
that

F,~(X) = AF(x)
rU
F,g(x) = b

or, in o t h e r words, that f is F - r e l a t e d to the v e c t o r field f' de-

f i n e d by

f'(z) = Az

and t h a t ~ is F - r e l a t e d to the v e c t o r f i e l d g' d e f i n e d by

g'(z) = b

As a c o n s e q u e n c e , we have that the Lie b r a c k e t [~,~] is F - r e l a t e d

to the Lie b r a c k e t [ f',g'] . U s i n g this fact repeatedly, one m a y c h e c k
that

i~ i
F , ( a d ~ g ) (x) = (adf,g')o F(x)

for all 0 < i < n-1. The special f o r m of f' a n d g' is s u c h that

(ad~,g' = (-1)iAib

All together, these yield

F,(g ad ~ .. a a_n-1~,
~ g} =

= (b -Ab ... ( - 1 ) n - I A n-1 b)

 228

The matrix on the right-hand-side is n o n s i n g u l a r , because (A,b)

is a r e a c h a b l e pair, and so is F#. This completes the p r o o f of the
sufficiency. []

(6.14) Lemma. Let ~ be a real-valued function defined on an o p e n set

V. T h e n the c o n d i t i o n s (6.6) and (6.7) hold if and only if

(6.6') Lg~(X) = L[ f,g] ~(x) = ... = L n-2 ~(x) = 0

(adf g)

for all x E V, and

(6.7') L ~(x ) ~ 0
(adf -I g)

Proof. We show, by induction, that the set of c o n d i t i o n s

0
(6.15a) LgLf~ = .. . = L g L ~ = 0

is e q u i v a l e n t to the set of c o n d i t i o n s

(6.15b) L ~ = ... = L k ~ = 0
(ad~g) (adfg)
and b o t h imply

(6.16) L(adfg)LJ~ = (-1)iLgLf+J#

for all i,j such that i+j = k+1.

This is c l e a r l y true for k = 0. In this case (6.15a) and (6.15b)
reduces to Lg~ = 0 and

L[ f,g] m = L f L g m - L g L f ~ = -LgLf~

Suppose (6.15a) and (6.15b) true for some k and (6.16) true for
all i,j such that i+j = k+1. The latter yields, in p a r t i c u l a r ,

k+l k+l
L ik+1 ~ = (-I) LgLf
(aaf g)
So t h a t L g L k+1
f ~ = 0 if a n d o n l y if L. Ik+I , = 0. A s s u m e either one
of these conditions holds. Then [aef gj
 229

L (aaf~k+2g)'@ = LfL.~aaf.k+lg)~ - L (adk+lf g)Lf~ =

= (-1)LfL k Lf~ + (_1)2L (adkg) Lf~

2
(adfg)
.k+1 k+1 2L g)L2
= (-I) LfLgLf ~ + (-I) (adk

= (-I) 2L (adkg) L2~

2 2 + (_i) 3L 3
= (-I) LfL, ~k-1 ,Lf~ k Lf~
~aaf g) (ad_-Ig)

k+IL L L k+1 3L 3
= (-I) f g f ~ + (-I) (adk+1 L f ~
f g;
3
= (-1)3L(adk_Ig)Lf~ = - . .

We see in this way that for all 0 < j < k+2

g)LJ~ (-I" k+2L L k+2

L(ad'+2g)~f = (-1)JL(=~f~k+2-j = J g f

and therefore that (6.16) is true for all i,j such that i+j = k+2.
From (6.15) and (6.16) the statement follows immediately. []
(6.17) Remark. We have proved, by the way, that either one of the two
equivalent sets of conditions (6.15) imply

L L~ = (-1)JL
(ad~g) r (ad~+Jg)~

for all i,j such that i+j < k+1. This fact will be used in the sequel.D

(6.18) T~eorern. Suppose m = I and let g = g1" The State-Space Lineariza-

tion Problem is solvable if and only if:
.n-1 , o,
(i) dim(span{g(x),adfg(x ) ..... aaf gtx )}) = n

(ii) the distribution

n-2
(6.19) A = sp{g, adfg,...,adf g}

is involutive in a neighborhood U of x .
 230

(6.20) Bemark. Note that the c o n d i t i o n (i) i m p l i e s that the tangent

n-1
vectors g(x),adfg(x),...,adf g(x) are l i n e a r l y i n d e p e n d e n t for all x
in a s u i t a b l e neighborhood of x . T h e r e f o r e the d i s t r i b u t i o n (6.19) is
nonsingular around x and has dimension (n-l).

Proof. We k n o w from the p r e v i o u s Lemmas that the p r o b l e m is s o l v a b l e

if and only if there e x i s t s a real-valued function ~ defined in a
neighborhood V of x such that the c o n d i t i o n s (6.6') and (6.7') hold.
These may be r e w r i t t e n as

(6.6") i ) (x) = 0
( d9, adfg

for all 0 < i < n-2 and all x E V, and

n-1
(6.7") (d~,adf g } (x ) / 0

If both these c o n d i t i o n s hold, then n e c e s s a r i l y the t a n g e n t

o ~n-1 . o,
vectors g(x ) , a d f g ( x ) .... ,aaf g[x ) are l i n e a r l y i n d e p e n d e n t . For,
we see from R e m a r k (6.17) that (6.6") implies

i > = L g) L J9 =
<dL~, adfg (adf

= (-1)JL(adf+Jg)9 = (-1)J(dg,adf+Jg }

for all i+j < n-1. Therefore, using again (6.6") and (6.7") we h a v e

ad g>(x) = 0

for all i,j such that i+j < n-2 and all x E V and

< d L fj~ , a d fig ) (x ) ~ 0

for all i,j such that i+j = n-1.

The above conditions, all together, show that the m a t r i x

d~(x )

dLfg(x )
(6.21) (g(x O) adfg(x)...ad~-Ig(x)) =

dL~-l~(x )
 231

d~, g ) (x ) (d~,adfg) (xO) ... ( d ~ , a d f - l g ) (x O)

dLf~,g ) (x ) < d L f ~ , a d f g ) (x ) ... ( d L f ~ , a d ~ - I g ) (x )

d L ~ - 1 ,g )(x o) <dLn-1
f ~,adfg ) (x) ... ( d L ~ - 1 9 , a d fn-1 g )(x o)

has rank n and, therefore, that the vectors g(x),adfg(x),...,a~f-lg(x O)

are linearly independent.
This proves the n e c e s s i t y of (i). If (i) holds then the distribu-
tion (6.19) has dimension n-1 around x and (6.6") tell us that the
exact covector field d~ spans A i around x O. So, because of Frobenius
theorem (see Remark I.(3.7)) we conclude that A is c o m p l e t e l y in-
tegrable and thus involutive, i.e. the n e c e s s i t y of (ii).
Conversely, suppose (i) holds. Then the d i s t r i b u t i o n (6.19) is
nonsingular around x . If also (ii) holds, A is c o m p l e t e l y integrable
around x and there exists a real-valued function ~, defined in a
neighborhood V of x , such that d~ spans A on V, i.e. such that
(6.6") are satisfied. Moreover, the covector field d~ is such that
(6.7") also is satisfied, because otherwise d~ would be a n n i h i l a t e d
by a set of n linearly independent vectors. This, in view of the pre-
vious Lemmas, completes the proof of the sufficiency. []

For the sake of convenience, we summarize now the procedure

leading to the c o n s t r u c t i o n of the feedback e and 8 which solves the
State-Space Linearization Problem in the case of a single input
channel.
Suppose (i) and (ii) hold. Then, using Frobenius T h e o r e m one
constructs a function ~, d e f i n e d in a n e i g h b o r h o o d V of x O, such that
(6.6") and (6.7") hold. Then, one sets

I
(6.22a) 6 (x)
L L n-1 (x)
g f

and
n
-Lf~(x)
(6.22b) e(x) =
LgL~-1~(x)

for all x E V. This pair of feedback functions, together with the

local c o o r d i n a t e s transformation defined by
 232

z (X)l = Lfi-1~(x)

for I ~ i ~ n, is s u c h as to s a t i s f y (6.1) and (6.2) with A and b I =b

in the f o r m of the r i g h t - h a n d - s i d e s of (6.8).

(6.23) Remark. T h e r e is a s u r p r i s i n g affinity between some results de-

scribed in this section and the o n e s described in the sections IV.3
and IV.4. For instance one may rephrase Lemma (6.5) by saying that
the S t a t e - S p a c e Linearization Problem is s o l v a b l e if and o n l y if one
may define, for the system

= f(x) + g(x)u

a (dummy) output function

y = ~(x)

whose characteristic number is exactly n-1. Of course, this w i l l be

possible if a n d o n l y if the c o n d i t i o n s (i) and (ii) are satisfied.
Once such a dummy output function has been found, then the solu-
tion of a S t a t e - S p a c e Linearization Problem proceeds like a solution
of a (degenerate, because both ~ a n d m are equal to I) n o n i n t e r a c t i n g
control problem. As a matter of fact, we h a v e from Lemma IV.(3.10)
t h a t the d i f f e r e n t i a l s d~,dLf~,...,dL~-1~ are linearly independent at
o
x a n d thus t h a t the m a p p i n g

n-1
F : x I )col(~(x),Lf~(x) ..... L f ~(x))

qualifies as a l o c a l c o o r d i n a t e s transformation. Then, from Corollary

IV. (3.14) we learn that

n-1 i i
A = 0 (sp{dLfg}) = 0
i=O

is the largest locally controlled invariant distribution contained in

(sp{d~}) .
The feedback (6.22) coincides with a solution of IV. (3.15) (with
7(x) = 0 and d(x) = I). Under this feedback the system becomes linear
in the n e w coordinates, as it is seen from the constructions given in
the section IV.4 (see Remark IV.(4.9)). []

We n o t e that the formal statement of the State-Space Lineariza-

tion Problem, given at the b e g i n n i n g of the section, does not in-
 233

corporate any requirement about the i m a g e F(U) of the c o o r d i n a t e s

transformation that makes it p o s s i b l e (6.1) and (6.2) to h o l d . H o w e v e r ,
one m a y w i s h to i m p o s e the a d d i t i o n a l requirement t h a t the i m a g e F(U)
contains the o r i g i n of ~ n . In this case, the c o n d i t i o n of T h e o r e m
(6.18) m u s t be s t r e n g h t e n e d a l i t t l e bit.
Suppose the c o o r d i n a t e s transformation z = F(x) solving the
State-Space Linearization Problem is s u c h t h a t

(6.24) z = F ( x ) = 0

Then, from (6.1) we d e d u c e that necessarily

(6.25) f(x ) + g ( x ) ~ ( x O) = 0

If f(x ) = 0, t h e n the c o n s t r u c t i o n already proposed for the

s o l u t i o n of the p r o b l e m m a y be a d a p t e d to m a k e (6.24) and (6.25)
satisfied. As a m a t t e r of fact, one m a y a l w a y s c h o o s e a function
satisfying (6.6") and (6.7") in such a w a y that ~(x ) = 0 (see, e.g.,
the c o n s t r u c t i o n proposed along the p r o o f of T h e o r e m I.(3.3)). If this
is the case, then

z 1 ( x ) = ~(x ) = 0

and also, for 2 < i < n,

zi(x O) = Lfi-I ~(x O) = (dL $-2 (xO),f(xO)) = 0

b e c a u s e we have a s s u m e d f(x ) = 0. T h u s the p r o p o s e d coordinates trans-

formation satisfies (6.24). Moreover,

L fn ~ ( x ) (dL f -I ~(x o ) , f ( x o) )
(x ) . . . . . 0
L L n - 1 ~ ( x ) L L n-1
g f g f ~(x)

and a l s o (6.25) holds.

One m a y thus a s s e r t that if f(x ) = 0, i.e. if the i n i t i a l state
o
x is an e q u i l i b r i u m state for the a u t o n o m o u s system

= f(x)

and if the S t a t e - S p a c e Linearization Problem is s o l v a b l e , one m a y a l -

ways find a s o l u t i o n such t h a t ~ ( x ) = 0 a n d F ( x ) = 0.
 234

If f(x ) ~ 0, the c o n d i t i o n (6.25) may be rewritten as

(6.26) f(x ) = c g ( x )

where c is a n o n z e r o real n u m b e r . Again, if the S t a t e - S p a c e Lineariza-

tion P r o b l e m is s o l v a b l e one m a y find a function ~ such that z1(x ) =
= ~(x O) = 0. B u t also, for 2 ~ i ~ n, (6.26) ensures that

zi(x) = < d L ~ - 2 ~ , f ( x ) ) = C L g L f-2 ~(x o) = 0

a n d thus (6.24) s t i l l holds. Moreover, the p r o p o s e d ~ is such that

( d L ~ - 1 ~ ( x ) , f ( x ) )
~ ( x ) . . . . c
L g L ~ - I ? ( x )

as e x p e c t e d .
In this case, the i n i t i a l state x is n o t an e q u i l i b r i u m state
for the o r i g i n a l system, b u t an a m a y be f o u n d such that x O is an
equilibrium state for the s y s t e m

= f(x) + g(x)a(x)

In s u m m a r y , we h a v e the f o l l o w i n g result.

(6.27) Corollary. S u p p o s e m = I and let g = g1" S u p p o s e the S t a t e -

Space Linearization Problem is s o l v a b l e . Then, a solution with
F(x ) = 0 e x i s t s if a n d o n l y if

f(x ) 6 s p { g ( x ) } []

(6.28) Remark. W h e n F ( x ) = 0, one m a y use the s o l u t i o n of the State-

Space Linearization Problem for l o c a l stabilization purposes. Indeed,
since (A,b) is a r e a c h a b l e pair, one m a y a r b i t r a r i l y assign the eigen-
values to the m a t r i x (A + bk), via suitable choice of the 1n r o w
vector k. If this is the case, the feedback control law

u = a(x) + S(x)kF(x) + S(x)v

makes the s y s t e m l o c a l l y diffeomorphic, On U, to the a s y m p t o t i c a l l y

stable system

~ = (A + bk) z + bv []
 235

We now describe the e x t e n s i o n of the p r e v i o u s discussion to the

case of m a n y inputs. This requires the i n t r o d u c t i o n of some further
notations, but the s u b s t a n c e o f the p r o c e d u r e is e s s e n t i a l l y the same
as the one e x a m i n e d so far.
Given a set of v e c t o r fields f , g l , . . . , g m we d e f i n e a sequence of
distributions as f o l l o w s

G O = s p { g I ..... gm }

G i = Gi_ I + [ f , G i _ I]

The following Lemma describes the p o s s i b i l i t y of c o m p u t i n g all

Gls
l
in a s i m p l e way.

(6.29) Lemma. Suppose all Gis are n o n s i n g u l a r . Then

(6.30) G i = sp{ad~gj : 0 < k _< i, I _< j _< m}

k
Proof. S u p p o s e G i = sp{adfg. : 0 < k < i, I < j < m}. S u p p o s e that at
o k 1 ] kC -- --
x some v e c t o r s aaf g J 1 ' ' ' ' ' a d f gJr are l i n e a r l y i n d e p e n d e n t and

span G i(x ). T h e n on a n e i g h b o r h o o d U of x a n y v e c t o r f i e l d T in G.
r k i
may be w r i t t e n as =~=I [ c adfagj , w i t h c a e C ~ ( U ) . T h e n [ f,T] =

r k +I k k +I
=~=I
[ (c~adf~ gJa+(Lfc )adf~gj ) . Therefore, on U, G i + I = s p { a d f ~ gj ,

k
a k
adf g_j~: I <_ ~ <_ r}. Since, by c o n s t r u c t i o n , all adfgj , 0 _< k _< i+I

and I < j _< m are in G i + 4, , t h i s p r o v e s t h a t Gi+1=sp{adkgj~ :0 < k < i+I f

I < j < m}. D

Since G i C Gi+1 by d e f i n i t i o n , if the G~s are n o n s i n g u l a r we have

that

G i + I (x)
dim ~ - - - independent of x
l

Thus we may define a sequence of i n t e g e r s v0,Vl,.., by setting

(6.31a) ~0 = d i m G O

S.
(6.31b) V i = dim ~ i > I
Gi_ I
 236

The integers thus defined have the following property

(6.32) Lemma. The following condition holds

ui --> ~i+I

for all i > 0. Let v., denote the last nonzero element in the sequence
l
{~i : i ~ 0}. If

dim G = n
l

then

v 0 + v1+...+vi~ = n

Proof. Consider G i a n d Gi_ I. B y d e f i n i t i o n

dim Gi(x) = d i m Gi_1(x) + ~i

From (6.30), we deduce that, given a point x , there will be v. v e c t o r s

1
adfgjli ( x O ) , . . . , a d ~ g j ~ . (x ) l i n e a r l y i n d e p e n d e n t and with the p r o p e r t y
1
that all vector fields in G. m a y be w r i t t e n as linear combinations,with
1
i
smooth coefficients, of v e c t o r s of Gi_ I and of sp{ad=g. :1<s<u }. T h u s
~ 3s i
i+I
Gi+1 = Gi+sp{adf gJs : I < s < ~i }

and

v i + I _< v i . []

From the sequence {w. : 0 < i < i } w e d e f i n e another sequence of

integers mO,m 1, . . . , m i , , setting

m 0 = vi

m o + m I = ~i~_I

(6.33) mo+m1+m2 = vi~ 2

m o + m 1 + ' ' ' + m i * = ~0

 237

(6.34) Lemma. The following conditions hold

m0 > 0

m. > 0 I < i < i

Moreover, if d i m G., = n, then

1

(6.35) d i m G.*
i -i = i m 0 + . . . + 2 m i _ 2 + m i _ 1

for I < i < i . []

There is a n e e d for a third sequence of i n t e g e r s {<i:I~i~90} re-

lated to the p r e v i o u s ones by the following relations

~. = i +I if 1 < i < v_,

1

<'z = i if m I > 0 and w.,+11 --

< i --< 9i*-I

(6.36) K. = i -I if m 2 > 0 and ui, 1+I ~ i ! ~ i * - 2

1

<i = I if mi, > 0 and vi+I ! i ! ~0

With the h e l p of t h e s e notations it is r a t h e r simple to s t a t e the

necessary and sufficient conditions for the e x i s t e n c e of a solution to
the S t a t e - S p a c e Linearization Problem in the general case where m > I.

(6.37) Theorem. The State-Space Linearization Problem is solvable if

and only if

(i) x O is a r e g u l a r point of the distribution Gi, for all i ~ 0

(ii) d i m G . ~ ( x ) = n
l
(iii) the d i s t r i b u t i o n G i is i n v o l u t i v e , for all i such that mi,_i_1~0.

Proof. We restrict ourselves to the p r o o f of the sufficiency, which is

constructive. Without loss of generality, we m a y assume that

~0 = m

For, if this is n o t the case, since G O by a s s u m p t i o n is n o n s i n g u l a r

around x , we m a y always find a nonsingular mm matrix ~(x), defined
in a n e i g h b o r h o o d U of x , such that
 238

Go(X) = s p a n { g 1 ( x ) ..... gv (x)}

0

and

g~o+1 (X) = ... = gin(x) = 0

for all x 6 U, w h e r e

gi(x) = (g(X) 6(X)) i

for I < i < m. If a f e e d b a c k (~,B) solves the State-Space Linearization

Problem for the set f,gl,...,gv0, then it is e a s i l y seen that a feed-
back of the form

~, =~(~ ~ o)
o ) 6'=6(

solves the p r o b l e m for the o r i g i n a l set f , g l , . . . , g m.

For the sake of s i m p l i c i t y , we b r e a k up the construction in two
stages.

(i) Recursive construction of a c o o r d i n a t e s t r a n s f o r m a t i o n around the

point xO .

Step (I): B y a s s u m p t i o n

d i m G. ~ = n
1

|
and dim G~_I_ = m 0 > 0. M o r e o v e r , Gi,_1 is a s s u m e d to be involutive.
Then, by Frobenius theorem, we k n o w that there exist a neighborhood
U I of x a n d m 0 f u n c t i o n s h01,...,h0m defined on U I, w h o s e differ-
0
entials span Gi~_1(x) at all x U I. In p a r t i c u l a r ,

(6.38) (dh0i, adfgj ) (x) = 0

for all I <__ j .< m ., I .< i . < m .0, 0 < e < i -I and all x U I. M o r e o v e r ,
the d i f f e r e n t i a l s dh01 (x), .... d h 0 m 0 ( X ) are linearly independent at all
xEU I.
We claim that the m 0 m m a t r i x

M0 = {m ij
(0) (x) } = {< d h 0 i , a d fig j (x)}

has rank m 0 at all x E UI. For, suppose it is f a l s e at some x 6 UI .

 239

Then, there exist real n u m b e r s Cl,...,Cm0 such that

m0
i
< ~ e i d h 0 i , a d f g j )(x) = 0
i=I

for all 1 < j < m. This, together with (6.38), implies

m0
(6.39) < Z cidh0i,adfg j >(~) = o
i=I

for all I < j < m, 0 < e < i , and this in turn i m p l i e s

m0
( y~ c i d h 0 i ( x ) , v } = 0
i=I
m 0
for all v 6 Gi~(x). S i n c e d i m Gi~ = n, then Z c i d h 0 i ( 2 ) m u s t be a zero
i=I
covector, but since dh01 ( x ) , . . . , d h 0 m 0 ( X ) are i n d e p e n d e n t , then
= 0.
c I = ... = Cm0

Step (2): C o n s i d e r the d i s t r i b u t i o n Gi, 2 , which is s u c h that

dim G~,_2 = 2m 0 + m I

We c l a i m t h a t d L f h 0 1 , . . . , d L f h 0 m 0 are s u c h that

( dLfh0i, ad;gj )(x) = 0

for all I .< j .< m,. I .< i .< m 0,

. 0 < ~ < i -2 and all x 6 U I.
This comes f r o m the p r o p e r t y

-(dLfh0i,ad; gj)=(an0i,aa f gj ) - L f ( d h 0 i adfgj

in w h i c h both the t e r m s are zero on U I b e c a u s e ~ ~ i -2.

We c l a i m also t h a t the 2m 0 d i f f e r e n t i a l s

(6.40) {dh01(x),...,dh0m0(X),dLfh01(x),...,dLfh0m0(X)}

are l i n e a r l y independent all x E U I. For, suppose t h i s is false; then,

for s u i t a b l e reals Cli,C2i, we had
m0 m0
(6.41) Z e I i d h 0 i (x)
i=I + i =Z I c 2 i d L f h 0 i (x) -- o
 240

at some x E U. This w o u l d imply

m 0 ,

(i=I
~ (clidh0i + c 2 i d L f h 0 i ) ' a d ~ -Igj )(~) = 0

for all I < j < m. This in turn implies (because of (6.38))

m 0 , m0 ,

< [ c2idLfh0i,ad ~ -Igj >(~) = _< [ c2idh0i, ad if gj )(x) = 0

i=I i=I

i.e. a contradiction, like in (6.39). Therefore c21 = ... = C2m = 0

A

in (6.41), and also Cil = ... = C l m 0 = 0 b e c a u s e dh01 (x),...,dh;m0(X)

are l i n e a r l y i n d e p e n d e n t .
If m I = 0, the 2m 0 c o v e c t o r s (6.40) span Gt,_2. If m I > 0, u s i n g
again Frobenius theorem (because Gi,_2 is i n v o l u t i v e ) , we may find m I
more functions h11(x),...,hlm (x), d e f i n e d in a n e i g h b o r h o o d U2 C UI
of x , such that the 2m0+m I dlfferentials

(6.43) {dh01 (x) ..... dh0m0 (x) ,dL~01 (x) .... ;dL~0m0 (x) ,d~ I (x) ..... ~hm I (x) }

are linearly independent and

( dhl i' adfgj >(x) = o

for all I <_ j .< m,. I .< i .< m .I, 0 < ~ < i -2 and all x E U 2.
We c l a i m that the (m0+m I ) xm m a t r i x

I:i)
where M 0 is as b e f o r e and M I d e f i n e d as

MI = tmijr
(I) (x)} = { ( d h l i , a d f -Igj >(x)}

has rank m 0 + m I at all x ~ U 2.

For suppose for some reals c01,...,C0m0,C11,...,Clml we had

m0 , m I

<i=I
[ c0idh0i(x)'adf g j ( x ) > + ( i = i[ c l i d h l i ( x ) ' a d f -Igj(x) > = 0

at some ~ E U2 . T h e n (recall Remark (6.17))

 241

m0 mI ,
(6.44) < ~ c0idLfh0i(~) + ~ Clidhli(~),ad~ -Igj(~)) = 0
i=I i=I
m0 mI
The c o v e c t o r [ c0idLfh0i(x) + [ Clidhli(x) annihilates, as we have
i=I i=I
seen before, all ad~gj(x), e < i~-2, I ~ j ~ m, but (6.44) tells us
that it also annihilates " i ~ I gj (x) . Thus,
aaf this covector annihilates
all vectors in Gi*_1.
From the previous discussion, we conclude that this covector must
belong to s p a n { d h 0 1 ( x ) , . . . , d h 0 m 0 ( X ) } , but this is a contradiction,
because the covectors (6.43) are linearly independent Therefore, the
c0i's and c11.'s of (6.44) must be zero.
Eventually, with this procedure we end up with a set of functions

i - i -I h
h01,...,h0m0,Lfh01,...,Lfh0m0,...,Lf lh01,...,L f 0m 0

i i -2 h
h11,...,hlm1,...,Lf -2h11,...,L f lm I
(6.45)

hi*-l,1,''',hi~-1,mi, i

(of course, some of these lines may be m i s s i n g if some m i is zero)

with the following p r o p e r t i e s :
- the total number of functions is

i m 0 + (i - 1 ) m 1 + . . . + 2 m i , _ 2 + m i , _ l = n-m

- the n-m d i f f e r e n t i a l s of these functions are independent at all

x E U, a n e i g h b o r h o o d of x ,

- the vlxm m a t r i x
M0

Mi#_ I

where M. is m . x m and
l l

mj
(i) (x) = ( dhi,ad f -igj) (x)

has rank v I at all x E U.

 242

If mi~ > 0, one m a y s t i l l f i n d mi, m o r e functions hi~1,...,hi%m.~,

l
that, together with the functions (6.45) and with the a d d i t i o n a l func-

tions Lfl h o 1 , . . . , L fi hOmo ,... 'Lfhi* - I ,I''''' L f h i * - I ,m.~ A , give r i s e

IO--]
to a set of n l i n e a r l y i n d e p e n d e n t d i f f e r e n t i a l s at x .
For c o n v e n i e n c e , l e t us r e l a b e l the f u n c t i o n s hij a n d set

~i = h0i if I _< i _< ~i ~

~i = h l , i - u . % if mI > 0 and v.~+ll --

< i --< 9i*-I
1

~i = h i ~ , i - ~ 1 if mi, > 0 and ~I+I ~ i ~ v 0 = m

The previous constructions t e l l us t h a t the m a p p i n g

F : x ~-+ col (61 (x) ..... ~m(X)) , w h e r e

~i(x)

~i(x) = Lf~ i(x)

Lf I ~i(x)

o
qualifies as a local diffeomorphism around x .
Moreover, by c o n s t r u c t i o n ,

(6.46) ( d L f ~ i , g j >Cx) = 0

for all 0 ~ ~ ~ Ki-2, 1 ! i,j ~ m, at all x a r o u n d x , a n d the m m

matrix

(6.47) A(x) = {aij(x)} = {<dLf I ~i,gj >(x)]

is n o n s i n g u l a r at x = x O.

(ii) Construction of the l i n e a r i z i n g feedback. F r o m the c o n d i t i o n s

(6.46) and (6.47), we see t h a t the c o n t r o l system

m
(6.48a) = f(x) + ~ gi(x)ui
i=1
 243

with (dummy) outputs

6.48b) Y i = ~i (x) I ~ i ~ m

is s u c h that:

- the c h a r a c t e r i s t i c number Pi a s s o c i a t e d with the i-th output channel

is e x a c t l y equal to K.-1,
1
- the s i n g l e - o u t p u t s noninteracting control problem is s o l v a b l e around
o
x .

Choose a feedback ~ and 8 as a s o l u t i o n of the e q u a t i o n s

<1
Lf ~l(X)
A(x)a(x) = -

Lfm~m(X)

A(x) B(x) = I

(they c o r r e s p o n d to the e q u a t i o n s IV. (4.4a) with Yi = 0 a n d IV.(4.4b)

with 6. the i-th r o w o f an m m identity matrix) Under this feedback,
1

the system (6.48) splits into m noninteracting single-input single-

output channels. In p a r t i c u l a r , in the n e w c o o r d i n a t e s defined at the
previous stage, each subsystem is d e s c r i b e d by equations of the form
(see IV. (4.8))

0 1 0 ... 0 0 0

0 0 1 ... 0 0 0

... .
~i = vi

0 0 0 ... 0 I 0

0 0 0 ... 0 0 I

Yi = ( I 0 0 ... 0 0)~ i

This completes the p r o o f . O

 244

7. O b s e r v e r with Linear Error Dynamics

We consider in t h i s s e c t i o n a problem which is in s o m e sense dual

of that considered in the p r e v i o u s one. We have seen that the solva-
bility o f the S t a t e - S p a c e Linearization Problem enables us to d e s i g n
a feedback under which the system becames locally diffeomorphic to a
linear system with prescribed eigenvalues. In the c a s e of linear system,
the d u a l notion of spectral assegnability via static state-fedback is
the e x i s t e n c e of state-obsevers with prescribed eigenvalues. Moreover,
i t is k n o w n that the dynamics of a state-observer and that of the
observation error (i.e. the d i f f e r e n c e between the u n k n o w n state and
the estimated state) are t h e same. In v i e w of this, if w e w i s h to
dualize the r e s u l t s developed so far, we are led to the p r o b l e m of
the synthesis of (nonlinear) observers yielding an e r r o r dynamics
that, possibly after some suitable coordinates transformation,becames
linear and spectrally assignable.
For the sake of simplicity, we restrict ourselves t o the considera-
tion of systems without inputs and with scalar output, i.e. systems
described by equations of the form

= f (x)

y = h(x)

with y 6 ~.
Suppose there exists a coordinates transformation z = F(x) under
which the v e c t o r field f and the output map h become respectively

F ~ f o F -I (z) = A z + K(cz)

h o F -I (z) = c z

where (A,c) is an o b s e r v a b l e pair a n d K is an n - v e c t o r valued function

of a real variable.
If t h i s is the c a s e , then an o b s e r v e r of the form

= (A+kc) ~ - k y + K (y)

yields an observation error (in t h e z coordinates)

e = ~ - z = ~ - F(x)

governed by the d i f f e r e n t i a l equation

 245

6 = (A + k c ) e

which is l i n e a r and spectrally assignable (via the nxl c o l u m n v e c t o r k).

Motivated by these consideration, we examine the following pro-
blem.

Observer Linear~zation ProbLem. Given a vector field f, a r e a l - v a l u e d

function h a n d an i n i t i a l state x find (if p o s s i b l e ) a neighborhood
U o f x , a coordinates transformation z = F(x) defined on U, a matrix
A C ~nxn and a row vector c E ~Ixn, a mapping K:h(U) ~ n such that

(7.1) hoF-1(z) = cz

(7.2) F ~ f o F -I (z) - A z = K ( c z )

f o r all z 6 F(U), and

n-1
(7.3) N k e r ( c A i) = {0}. []
i=0

The conditions for the solvability of this problem c a n be d e -

scribed as follows

(7.4) Lemma. The Observer Linearization Problem is s o l v a b l e only if

(7.5) dim(span{dh(x),dLfh(x ) ..... d L ~ - l h ( x ) } ) = n

Proof. The condition (7.3) says that the p a i r (A,c) is o b s e r v a b l e . Then,

it is k n o w n from the theory of linear systems that there exist a non-
singular nxn matrix T and a n1 column vector k such that

0 0 ... 0 0

1 0 ... 0 0
(7.6) T (A+kc) T -I = c T -I = (0 0 . .. 0 1)

0 0 ... I 0

Suppose (7.1) and (7.2) hold, and set

z = F(x) = TF(x)

K(y) = T(K(y)-ky)

where y E h(U).
 246

Then, it is e a s i l y seen that

ho~ -I(~) = (o o ... o 1)z

0 0 ... 0 0

1 0 ... 0 0

~,fo~-1(~)_ 0 0 ... 0 0 z = K((0 0 ... 0 I)~)

. o ... .

0 0 ... 1 0

From this w e d e d u c e t h a t there is no loss of g e n e r a l i t y in as-

suming t h a t the p a i r (A,c) that makes (7.1) and (7.2) satisfied has
directly the f o r m specified in the r i g h t - h a n d - s i d e s of (7.6).
Now, set

z = F(x) = col(z I (x) . . . . . Zn(X))

If (7.1) and (7.2) hold, we have, for all x E U

(7.7) h(x) = Zn(X)

~z I
3 ~ f(x) = k I (z n(x))

~z 2
~x f(x) = z1(x)+k2(Zn(X))
(7.8)

~Z
n

~x f(x) = Zn_ 1 ( x ) + k n ( z n ( x ) )

where kl,...,k n denote the n c o m p o e n t s of K.

Observe that
3z
L f h (x) _ 3x n f(x) = Zn-1 (x) + kn(Zn (x))

~Zn_ I ~k n ~z n
L h(x) - 3x f(x) + ((~--~--)y=Zn)~-~ - f(x)

~k ~z
= Zn_2(x) + ( ( ~ - n) y = Z n ) ~ - X -n- f ( x ) + kn_ I (Zn (x))
 247

= Zn-2(x) + kn-1 (Zn(X)'Zn-1(x))

where

~k 3k
- _ n + n
kn-1(Zn'Zn-1) ~z Zn-1 ~ z - kn(Zn) + kn-1(Zn)
n n

Proceeding in this way one obtains for each Lfi(x), for 2 < i < n-l,
an expression of the form

L~h(x) = Zn-i(x) + kn-i+1 (Zn(X) ..... Zn_i+ I (x))

Differentianting with respect to x and a r r a n g i n g all these ex-

pressions together, one obtains

~h ~h
0 0 ... 0 I
~x
DLfh ~Lfh
O 0 ...
~x Sz
~~z
= 0 Q I F,(X)

~L~-lh aL~-lh
1 * ... *
~x ~z

This, because of the n o n s i n g u l a r i t y of the m a t r i x on the right-

hand-side, proves the claim. D

If the c o n d i t i o n (7.5) is satisfied, then it is possible to define,

in a n e i g h b o r h o o d U of x O, a unique vector field T which satisfies the
conditions

= L -n-2h(x) = 0
LTh(X) = L~[Lfh(x) = ... TLf

LTLf-lh(x) = I

for all x 6 U.
As a matter of fact, one only needs to solve the set of equations

dh (x)
dLfh (x)
(7.9) ... T(X) =
n-2
dLf h(x)

dLf- lh (x)
 248

The c o n s t r u c t i o n of this vector field T is useful in order to

find n e c e s s a r y and sufficient conditions for the solution of our pro-
blem.

(7.10) Lemma. The Observer Linearization Problem is solvable if and

only if

(i) d i m ( s p a n { d h ( x ) , d L f h ( x ) ..... dL~-lh(x)}) = n

(ii) There exists a m a p p i n g { of some open set V of ~ n onto a neigh-

borhood U of x that satisfies the equation

(7.11) ~z = (I:
~-~ -adfi " "" (-1)
n-1 ~n-1~)o(Z)
aaf

for all z 6 V, where T is the unique vector field solution of (7.9).

Proof. Necessity. We already know that (i) is necessary. Suppose (7.1)

and (7.2) are satisfied and set {(z) = F-1(z) for all z = F(U).
Moreover, let e be the (unique) vector field ~-related to ~ , i.e.
]
let

O(x) = ,(~1 )o-I(x)

We claim that

(7.12) adke (x) = (-1)k, (5Z~-k+-----~)-I(x)

for all 0 ~ k ~ n-1. To show this, we proceed by induction (because

(7.12) is true by definition for k = 0), and we use the fact, deduced
from (7.2),(see also (7.8)), that

-I
f(x) = (F,) (Az + K(cz))oF(x) =

= ~*(kl( Z n ) ~ 1 + (z1+k2(Zn))~z--~ +'''+(zn-1+kn(zn)~}~-z)n o-I (x)

Suppose (7.12) is true for some k < n-1. Then

lk+I 0 = [f,(-1)k#,(
aaf )o-I] =

= (-I) k , [ k i ~ ~ I + (z1+k2)Tz2+. . .+ (z n _ 1 + k n ) ~ n o ~-I

' ~__~___]~Zk+1

= (-I)k+1~,(~Z-~k+----~)
0 - I
 249

Collecting all (7.12) together one obtains

(e -adfe ... (-1)n-lad~-le)o =Q,(~z I ~z2~ ... ~zn~ ) =

If we show that @ necessarily coincides with the unique solution

of (7.9) the proof is completed, because the p.d.e. (7.11) will
coincide with the one just found.
To this end, observe that

(-1)kL aafu
.k~h(x) ~h ~*(~z
= ~-x _ _ k ~ 1 ) o -1 (x) =

,~ho , o~-I (x)

= t~z--~+IJ

but, since ho~(Z) = Zn, we have

L =k~h(x) = 0
~t~/fU

for all 0 < k < n-2 and

(-1)n-IL ~n-1~h(x) = I
~f

Using Lemma (6.14) we deduce that

LeLkfh (x) = 0

for all 0 < k < n-2 and (see also Remark (6.17)

LoL~-Ih(x) = I

Thus, the vector field e necessarily coincides with the unique

solution of (7.9).
Sufficiency. Suppose (i) holds and let T denote the solution of (7.9).
Using Remark (6.17) one may immediately note (see (6.21)) that the
matrix
dh(x )

dLfh (x) n-1

(T(X) adfT(x ) ... adf ~(xO))

dLf-lh (x )
 250

has r a n k n. T h e r e f o r e , the v e c t o r fields T( x ) , a d f ~ ( x ) .... , a d ; - I T ( x )

O
are l i n e a r l y independent at x .
Let ~ denote a solution of the p . d . e . (7.11) a n d let z be a
point s u c h t h a t %(z ) = x . F r o m the l i n e a r independence of the v e c t o r
fields on the r i g h t - h a n d - s i d e of (7.11) we d e d u c e that % has rank n
at z , i.e. that # is a d i f f e o m o r p h i s m of a n e i g h b o r h o o d of z o n t o
a neighborhood o f x ,
Set F : ~-I a n d

(7.13) F m f o F -1 (z) = fl ~ I + f2 ~ 2 + "'" + fn 8z n

By d e f i n i t i o n , the m a p p i n g # is such t h a t

so t h a t

(7.14) F , a d ~ T o F -1 (z) = (-1)k

Zk+ I

for all 0 < k < n-1.

Using (7.13) and (7.14), one o b t a i n s , for all 0 < k < n-2

(_i)k+1 ~ _ ~k+l
~Zk+ 2 = # , a a f ToF -I (z) = F~[ f,ad k T] oF -I (z) =

= (-1)k[f 1 -~1 +... + fn z ~;)n 8 I ] =

' ~z---k+

= (-1) k+l( _ _~I

_ _ + . . .~ ~fn
+ (~_~+l)_~n)
(aZk+l) 8z I

that, because of the l i n e a r independence of ~ ..... ~ , implies

~n

1
for i ~ k+2
z~TT+l =

3fk+2
= 1
3Zk+ 1

F r o m these, one d e d u c e s t h a t fl d e p e n d s o n l y on z n a n d t h a t fi'

for 2 ~ i ~ n, is s u c h t h a t fi - z i - 1 depends o n l y on z n. In o t h e r
terms, o n e has
 251

k 1 ( z n)

z I + k 2 ( z n)
F ~ f o F -I (X) =

Zn_ I + k n ( Z n)

where k l , . . . , k n are s u i t a b l e f u n c t i o n s of z n a l o n e , a n d this shows that

the c o n d i t i o n (7.2) holds.
Moreover, since

L .k h = 0
aafT

for all 0 < k < n-l, and

n-1
L _n-1 h = (-1)
aaf T

we d e d u c e that

B h o F -I ~hoF -I
- 0
~z I -'- 8Zn_ I

and t h a t

~hoF -I
- I
~z
n

This shows that also (7.1) holds. []

The integrability of the p.d.e. (7.11) m a y be e x p r e s s e d in t e r m s

of a p r o p e r t y of the v e c t o r fields T,adfT,...,ad~-IT. TO t h i s e n d , o n e
m a y use the following consequence of F r o b e n l u s Theorem.

(7.15) Theorem. Let T I , . . . , T n be v e c t o r fields on ~ n . Consider the

set of p a r t i a l differential equations

~x
(7.16) ~z i ~i(x(z))

where x denotes a mapping f r o m an o p e n set of ~ n to an o p e n set o f ~ n .

Let (z,x ) be a p o i n t in ~ n ~ n and suppose T I ( x O ) , . . . , T n ( X O) are
linearly independent. There exist neighborhoods U of x a n d V of z
and a d i f f e o m o r p h i s m x:V ~ U solving the e q u a t i o n (7.16), and such
 252

that x(z ) = x , if a n d o n l y if

(7.17) [ T i , T j] = 0

for all 1 < i,j < n.

Proof. We l i m i t o u r s e l v e s to give a s c k e t c h of the p r o o f of the suf-

ficiency. To this end, set

A i = sp{Y i}

Then, the c o l l e c t i o n of d i s t r i b u t i o n s A 1 , . . . , A n is i n d e p e n d e n t , s p a n s
the t a n g e n t s p a c e a n d is s i m u l t a n e o u s l y i n t e g r a b l e b e c a u s e of (7.17)
(see T h e o r e m I. ( 3 . 1 2 ) ) . A s a c o n s e q u e n c e , we m a y f i n d a c o o r d i n a t e c h a r t
(U,~), such that z O = ~(x ) a n d

}
~i(x) = sp{ ( ~ i ) x

for all x 6 U. The a b o v e m a y be r e w r i t t e n as

~Ti(x) = ci({) (7~i)~o~ (x)

for all x E U , where c. is a s m o o t h real-valued function, and

l
ci(z) ~ 0. The c o n d i t i o n (7.17) m a y be u s e d a g a i n to s h o w that c i
depends o n l y on ~i" Thus, there exist functions z i = ~i(~i ) s u c h that
o
z i = ~i(z ) and

~9i
~--~. ci(~i) = I

The c o m p o s e d function

z = %(x) = (~1 ..... ~n )~(x)

is c l e a r l y such t h a t ~(x ) = z a n d

~*~i(x) = ( ~ 7 ) o~(x)
l

-1
Thus x = ~ (z) s o l v e s the p . d . e . (7.16). []

Merging Lemma (7.10) with Theorem (7.15) yields the d e s i r e d re-

sult.
 253

(7.18) Theorem. The O b s e r v e r Linearization Problem is s o l v a b l e if a n d

only if

(i) d i m (span {dh (x) , d L f h ( x ) ..... d L ~ - l h ( x O) }) = n

(ii) the u n i q u e vector field T solution of (7.9) is s u c h t h a t

(7.19) [adfT,adJT] = 0

for all 0 < i,j < n-1.

(7.20) Remark. Using the J a c o b i identity repeatedly, one c a n e a s i l y

show that the c o n d i t i o n (7.19) m a y be r e p l a c e d by the c o n d i t i o n

k
[ T,adfT] = 0

for all k = 1 , 3 , . . . , 2 n - I . []

In s u m m a r y one m a y p r o c e e d as f o l l o w s in o r d e r to o b t a i n an ob-
server with linear (and s p e c t r a l l y assignable) error dynamics. If
condition (i) holds, one finds first the v e c t o r field solving the
equation (7.9). If a l s o c o n d i t i o n (ii) holds, one solves the p . d . e .
(7.11) and finds a function , d e f i n e d in a n e i g h b o r h o o d V of z ,
s u c h t h a t ~(z ) = x . T h e n one sets F = -I. E v e n t u a l l y , one c o m p u t e s
the m a p p i n g K as

k 1 ( z n) 0

k2(z n ) zI
K(Zn) = = F,foF-I(z) -

k n ( Z n) Zn- I

At this point, the o b s e r v e r

= (A+kc) ~ - k y + K (y)

with (A,c) in the f o r m of the r i g h t - h a n d - s i d e s of (7.6) y i e l d s the

desired result
 APPENDIX
BACKGROUND MATERIAL IN D I F F E R E N T I A L GEOMETRY

I. S o m e facts from advanced calculus

L e t A be an o p e n subset of ~n and f: A ~ a function. The value

of f at x = ( X l , . . . , x n) is d e n o t e d f(x) = f(xl,...,Xn). The function f
is s a i d to be a function of class C~ (or, simply, C ~ or also, a smooth
function) if its p a r t i a l derivatives of any order with respect to
Xl,...,x n exist and are c o n t i n u o u s . A function f is s a i d to b e a n a l y t i c
(sometimes noted as C ~) if i t is C ~ a n d f o r e a c h p o i n t x E A t h e r e
o
exists a neighborhood U of x , such that the Taylor series expansion
of f at x c o n v e r g e s to f ( x ) for all x E U.

Example. A t y p i c a l e x a m p l e o f a f u n c t i o n w h i c h is C ~ b u t n o t a n a l y t i c
is the function f: ~ ~ defined by

f(x) = 0 if x < 0

I
f(x) = exp(- ~) if x > 0 []

A mapping F: A ~ m is a c o l l e c t i o n (fl,...~fm) of f u n c t i o n s
fi: A ~. The mapping F is C if a l l fl~ s are C
Let U C ~n and V C ~n be o p e n sets. A mapping F: U ~ V is a d i f -
feomorphism if is b i j e c t i v e (i.e. one-to-one a n d onto) and both F and
F -I are o f c l a s s C ~. The jaaobian matrix o f F at a p o i n t x is the
matrix

~fl ~fl
~Xl "--
3F
~x

3f 3f
n n
~x I "'" ~x n

The value 8F a t a p o i n t
of ~-~ x = x o is sometimes denoted 8F
(~-x) o
x
Theorem. (Inverse function theorem). L e t A b e an o p e n set of ~n and
F: A ~ n a C ~ m a p p i n g If (~F)
~ xO is n o n s i n g u l a r at some x o C A, then
there exists an open neighborhood U of x in A s u c h t h a t V = F(U) is
open in ~ n and the r e s t r i c t i o n o f F to U is a d i f f e o m o r p h i s m o n t o V.
 255

Theorem. (Rank theorem). Let A C ~n and B C ~m be o p e n sets, F:A ~ B

a C ~ mapping. Suppose ~F x has
(~-x) rank k for all x 6 A. For each point
x 6 A there exist a neighborhood A 0 of x in A a n d an o p e n neighbor-
h o o d B 0 of F ( x 0) in B, two open sets U 6 ~n and V C ~ m , and two dif-
feomorphisms G: U ~ A o a n d H: B o ~ V such that HoFoG(U) C V and such
that for all (xl,...,Xn) E U

(HoFoG) (x I ..... x n) = (Xl, .... Xk,O ..... O)

Remark. Let Pk d e n o t e the m a p p i n g Pk: ~ n ~ ~m defined by

Pk(Xl .... ,x n) = (Xl, .... Xk,O ..... O)

Then, since H a n d G are invertible, one m a y restate the previous ex-

pression as

F = H - I o P k o G -I

which holds at all points of A o-

Theorem. (Implicit function theorem). Let A C ~m and B C ~n be o p e n

sets. Let F: A x B ~n be a C~ mapping. Let (x,y)=(x I t..o, Xm,Yl, ..or yn )
denote a point of AB. Suppose that for some (x,y O) 6 A B

F ( x , y ) = 0

and that the m a t r i x

~fl 5fl
~Yl " " "

9F ,oQ
~y
9f ~f
n n
~Yl "'" ~Yn

is nonsingular at (x,y). Then, there exists open neighborhoods Ao

of x in A and B 0 of yO in B and a unique C~ mapping G: A ~ B 0 such
that
F(x,G(x)) = 0

for all x E A

Remark. As an a p p l i c a t i o n of the implicit function theorem, consider

 256

the following corollary. Let A be an open set in ~ m, let M be a kxn

matrix whose entries are real-valued C~ functions defined on A and b
a k-vector whose entries are also real-valued C~ functions defined on
A. Suppose that for some x E A

rank M ( x O) = k

Then, there exist an open neighborhood U of x and a C ~ mapping

G : U ~ ~n such that

M(x)G(x) = b(x)

for all x E U.
In o t h e r words, the equation

M(x)y = b(x)

has at least a solution which is a C function of x in a n e i g h b o r h o o d

of x . If k = n this solution is unique.

2. Some elementa~ notions of topology

This section is a r e v i e w of the most elementary topological con-

cepts that w11 be encountered later on.
Let S be a set. A topological structure, or a topology, on S is
a collection of subsets of S, called open sets, satisfying the axioms

(i) the union of any number of open sets is open

(ii) the intersection of any finite number of o p e n sets is o p e n
(iii) the set S and the empty set @ are open

A set S with a topology is c a l l e d a topological space.

A basis for a topology is a collection of open sets, called basic
open sets, w i t h the following properties

(i) S is the union of basic open sets

(ii) a nonempty intersection of two basic open sets is a union of
basic open sets.

A neighborhood of a point p of a topological space is a n y open

set which contains p.
Let S I and S 2 be topological spaces and F a mapping F: S I ~ S 2.
The mapping F is continuous if the inverse image of every open set of
S 2 is an o p e n set o f S I. T h e mapping F is open i f t h e i m a g e of an o p e n
 257

set o f S I is an o p e n s e t o f S 2. T h e m a p p i n g F is an homeomorphism ~f
is a b i j e c t i o n and both continuous and open.
If F is a n h o m e o m o r p h i s m , the inverse mapping F -I is a l s o an
homeomorphism.
Two topological spaces S1,S 2 such that there is a n h o m e o m o r p h i s m
F:S I ~ S 2 a r e s a i d to be homeomorphic.
A subset U of a topological space is s a i d to b e closed if its
complement U in S is open. It is e a s y to s e e that the intersection of
any number of c l o s e d s e t s is c l o s e d , the union of any finite number of
closed sets is c l o s e d , and both S and ~ are closed.
If S O is a s u b s e t of a topological space S, there is a u n i q u e
open set, noted int(So) and called the interior o f S o , W h i c h is c o n -
tained in S O a n d c o n t a i n s any other open set contained in S o . A s a
matter of fact, int(So) is the u n i o n of all open sets c o n t a i n e d in S o .
Likewise, there is a u n i q u e closed set, noted cl(So) and called the
closure o f So, w h i c h c o n t a i n s S O a n d is c o n t a i n e d in a n y o t h e r closed
set which contains So. Actually, c l ( S o) is the intersection o f all
closed sets which contain So.
A subset o f S is s a i d to b e dense in S i f its c l o s u r e c o i n c i d e s
with S.
If S I and S 2 are topological spaces, then the c a r t e s i a n product
SIS 2 can be given a topology taking as a basis the c o l l e c t i o n of
all subsets of the form UIxU 2 , with U I a basic open set of SI and U2
a basic open set o f S 2. T h i s topology o n S I S 2 is s o m e t i m e s called
the product topology.
I f S is a t o p o l o g i c a l space and S 1 a subset o f S, then S I can
be given a topology taking as o p e n sets the subsets of the form
$I A U, w i t h U any open set in S. T h i s topology o n S I is s o m e t i m e s
called the subset topology.
L e t F: S I ~ S 2 be a continuous mapping of topological spaces,
and l e t F ( S I) d e n o t e the image of F. C l e a r l y , F ( S I) w i t h the subset
topology is a t o p o l o g i c a l space . Since F is c o n t i n u o u s , the inverse
image of a n y open s e t o f F ( S I) is an o p e n s e t o f S I. H o w e v e r , not all
open sets of S I are taken onto open sets ofF(St). In o t h e r w o r d s , t h e
mapping F': S I ~ F ( S I) d e f i n e d b y F' (p) = F(p) is c o n t i n u o u s but not
necessarily open. The s e t F ( S I) c a n b e g i v e n another topology, taking
as o p e n sets in F ( S I) the images of open sets in S I. It is e a s i l y
seen that this new topology, sometimes called the induced topology,
contains the subset topology (i.e. any set which is o p e n in the subset
topology is o p e n also in the induced topology), and that the m a p p i n g
F' is n o w o p e n . If F is an i n j e c t i o n , t h e n S I a n d F(SI) endowed with
 258

the induced topology are homeomorphic.

A topological space S is s a i d to s a t i s f y the Hausdorff separation
axiom (or, briefly, to be an Hausdorff space) if any two d i f f e r e n t
points Pl and P2 h a v e disjoint neighborhoods.

3. S m o o t h manifolds

Definition. A locally Euclidean space X of d i m e n s i o n n is a t o p o -

logical space such that, for e a c h p E X, there exists a homeomorphism
mapping some open neighborhood of p o n t o an o p e n set in ~ n . []

Definition. A ManifoZd N of d i m e n s i o n n is a t o p o l o g i c a l space

which is l o c a l l y Euclidean of d i m e n s i o n n, is Hausdorff a n d has a
countable basis. []

It is n o t p o s s i b l e that an o p e n subset U of ~ n be h o m e o m o r p h i c
to an o p e n subset V of ~ m , if n ~ m (Brouwer's theorem on invariance
of d o m a i n ) . Therefore, the dimension of a l o c a l l y Euclidean space is
a well-defined object.
A coordinate chart on a m a n i f o l d N is a p a i r ( U , ~ ) , where U is an
open set of N a n d ~ a homeomorohism of U o n t o an o p e n s e t of ~ n . s o m e -
times ~ is r e p r e s e n t e d as a set (~1;...,~n) , and ~i: U ~ is c a l l e d
the i - t h coordinate function. If p E U, the n - t u p l e of real numbers
(~l(p),...,~n(p)) is c a l l e d the set of local coordinates of p in the
coordinate chart (U,~). A coordinate chart (U,~) is c a l l e d a cubic
coordinate chart if ~(U) is an o p e n cube about the origin in ~ n . If
p E U and ~(p) = 0, then the coordinate chart is s a i d to be centered
at p.
Let (U,~) and (V,~) be two c o o r d i n a t e charts on a m a n i f o l d N,
with U N V ~ @. L e t (~I .... ,~n ) be the s e t of c o o r d i n a t e functions
associated with the m a p p i n g ~. The homeomorphism

-I
+o~ : ~(U nv) ~ ~(u n v)

taking, for e a c h p E U A V, the s e t of l o c a l coordinates

(~1(p) .... ,~n(p)) into the s e t of local coordinates (%1 (p) ..... %n(p)),
is c a l l e d a coordinates transformation on U N V. C l e a r l y , ~o~ -1 gives
the inverse mapping, which expresses (~l(p),...,~n(p)) in t e r m s of

(~I (p) ..... %n (p)) "

Frequently, the set (~1(p), .... ~ n ( P ) ) i s represented as an
n-vector x = col(xl,...,Xn) , and the set (1(p),...,%n(p)) as an
n-vector y = col(yl,...,yn). Consistently, the c o o r d i n a t e transforma-
 259

-I
tion ~o~ c a n be r e p r e s e n t e d in the form

Yl Yl (Xl ..... Xn)

.
y = =
= y (x)

Yn Y n ( X l .... ,x n)

-I
and the inverse transformation ~o~ in the form

x = x(y)

X2
Jill :

:
3 ~ 4 1 ", _

. it--- -_
Y1
TwO coordinate charts (U,9) and (V,%) a r e C -Compatible if, w h e -
-I
never U n V ~ ~, the c o o r d i n a t e transformation o~ is a d i f f e o m o r -
phism, i.e. i f y(x) a n d x(y) are both C~ maps.
A C ~ atlas on a m a n i f o l d N is a c o l l e c t i o n A = {(Ui,~i)}iE I of
pairwise C~-compatible coordinate charts, with the p r o p e r t y that
U U. = N. A n a t l a s is complete if n o t properly contained in a n y o t h e r
i6I i
atlas

Definition. A smooth or C m a n i f o l d is a m a n i f o l d equipped with

a complete C~ atlas. []

Remark. If A is a n y C atlas on a manifold N, there exists a

unique c o m p l e t e C~ atlas A ~ containing A. The latter is d e f i n e d as
the s e t o f all coordinate charts (U,~) which are c o m p a t i b l e with every
coordinate chart (Ui,~i) o f A. This set contains A, is a C~ atlas w and
is c o m p l e t e by construction. []

Some elementary examples of smooth manifolds are the o n e s de-

scribed below.
 260

Example. A n y o p e n s e t U of ~ n is a s m o o t h m a n i f o l d , of dimension n.
For, consider the a t l a s A consisting of the (single) coordinate chart
(u, i d e n t i t y map on U) and let A ~ denote the unique complete atlas
containing A. In p a r t i c u l a r , A n is a s m o o t h manifold.

Remark. O n e m a y d e f i n e d i f f e r e n t complete C~ a t l a s e s on the same

manifold, as the following example shows. Let N = ~, and consider the
coordinate charts ~,~) and ~,), with

(x) = x

3
(x) = x

Since -I (x) = x a n d ~-1(x) = x I/3 ,

~o~-I (x) = x 1/3

and the two charts are not compatible. Therefore the u n i q u e complete
atlas A which includes (R,~) and the unique complete atlas ~ which
includes dR,~) are different. This means that the same manifold N may
be considered as a s u b s t r a t e of two different objects (two smooth
manifolds), one arising with the a t l a s A and the o t h e r with the atlas
A~ . []
Example. L e t U b e a n o p e n s e t o f ~ m a n d l e t 1 1 , . . . , I m _ n be r e a l - v a l u e d
C~ functions defined on U. L e t N d e n o t e the (closed) subset of U on
which all functions 11,...,Im_ n vanish, i.e. let

N = {x 6 U : I (x) = 0, I < i < m-n}

Suppose the rank of the jacobian matrix

211 211
~x I "'" ~x m

m-n m-n
~x I "'" ~x m

is m - n a t all x E N. Then N is a s m o o t h manifold of dimension n.

The proof of this essentially depends on the I m p l i c i t F u n c t i o n
o o o o ,xo)
Theorem, and uses the following arguments. L e t x =(xl, ....Xn,Xn+1,...
be a point of N and assume, without loss of generality, that the m a t r i x
 261

al I aX I
aXn+ 1 "'" ax m

~l ax
m-n m-n
~Xn+ 1 "'" ~Xm

is n o n s i n g u l a r a t x 0. Then, there exist neighborhoods A 0 of (x~ ..... Xn

O)
i n IRn a n d B o f 0 I . . .,x 0)
(Xn+ in IRm - n and a C ~ mapping G: A o ~ B o s u c h
that

h i { x I .... , X n , g 1 ( x I .... ,x n) ..... g m _ n ( X l ..... Xn)) = 0

for all I < i < m-n. This makes it p o s s i b l e to d e s c r i b e points of N

around x 0 as m - t u p l e s ( X l , . . . , x m) such that Xn+ i =gi(xl,...,Xn) for
I < i < m-n. In t h i s w a y one can construct a coordinate chart around
each point x o f N a n d the c o o r d i n a t e charts thus defined form a C~
atlas.
A manifold of this type is s o m e t i m e s called a smooth hypersurface
in ~ m . An important example of hypersurface is the sphere sm-l,defined
by taking n = m-1 and

x I = X l+x~ +... +X2-I

The set of points of ~Rm o n w h i c h fl (x) = 0 c o n s i s t s o f all the p o i n t s

on a s p h e r e of radius I centered a t the o r i g i n . Since

ax I ax I

never vanishes on this set, the r e q u i r e d conditions are satisfied and

the s e t is a s m o o t h manifold, of dimension m-1.

Example. An open subset N' of a smooth manifold N is i t s e l f a smooth

manifold. The topology of N' is the subset topology. If (U,~) is a
coordinate chart of a complete C~ atlas o f N, such that U ~ N' ~ ~,
then the p a i r (U',~') defined as

U' = U C~ N'

~' = r e s t r i c t i o n of ~ to U'
 262

is a c o o r d i n a t e chart o f N'. In t h i s way, one m a y define a complete

C atlas o f N'. The dimension o f N' is the same as t h a t o f N.

Example. L e t M a n d N b e s m o o t h m a n i f o l d s , of d i m e n s i o n m a n d n. Then
the cartesian product MN is a s m o o t h manifold. The topology of MN is
the p r o d u c t topology. If (U,9) and (V,) are coordinate charts of M
a n d N, the p a i r (UxV, (~,)) is a c o o r d i n a t e chart o f MxN. The dimension
of MN is c l e a r l y m+n.
An important example of this type of m a n i f o l d is the torus
T 2 = S I x S I, the cartesian product of two circles. []

Let h be a real-valued function defined on a m a n i f o l d N. If (U,~)

is a c o o r d i n a t e chart o n N, the c o m p o s e d function

= ~o -I : ~(U) -*:~

taking, for e a c h p E U, the set of local coordinates (Xl,...,Xn) of p

into the real number l(p), is c a l l e d an expression of ~ in local
coordinates.
In p r a c t i c e , whenever no confusion arises, one often uses the
-I
same symbol I to d e n o t e 1o4 , and write l ( X l , . . . , x n) to d e n o t e the
value of X at a p o i n t p of local coordinates (xl,...,Xn).
If N a n d M a r e m a n i f o l d s , of dimension n a n d m, F : N ~ M is a
mapping, (U,~) a coordinate chart on N a n d (V,~) a coordinate chart
o n M, the composed mapping

= o F o ~ -I

is c a l l e d an e x p r e s s i o n o f F in l o c a l coordinates. Note that this

definition make sense only if F(U) N V ~ @. If this is the case, then
is w e l l d e f i n e d for a l l n-tuples ( X l , . . . , x n) w h o s e image under
-I
Fo9 is a p o i n t in V.
-I
Here again, one often uses F to d e n o t e $~Fo~ , writes Yi =
= fi(xl,...,Xn) to d e n o t e the v a l u e of the i-th coordinate of F(p),
p being a point of local coordinates (x I .... ,Xn), and also

Yl f1(xl ..... x n )

y = = F(x)

Ym fm (Xl ..... Xn)

Definition. Let N and M be smooth manifolds. A mapping F: N ~ M is a

smooth m a p p i n g if f o r e a c h p 6 N there exists coordinate charts (U,~)
 263

of N a n d (V,~) o f M, w i t h p 6 U a n d F(p) 6 V, such that the e x p r e s s i o n

of F in l o c a l coordinates is C .

Remark. N o t e t h a t the p r o p e r t y of being smooth is i n d e p e n d e n t of the

choice of the c o o r d i n a t e charts o n N a n d M. Different coordinate charts
(U',~') and (V',~') are b y d e f i n i t i o n C~ c o m p a t i b l e with the former and

~, = ~,oFo~ ,-1 =

= ,o-1ooFo -1 o~o~' -1 =

= (,, o,-1) oFo ( ~ , o ~ - 1 ) - 1

being a composition of C= functions is still C ~. []

Definition. L e t N a n d M b e s m o o t h m a n i f o l d s , b o t h of d i m e n s i o n n.
A m a p p i n g F : N ~ M is & diffeomorphsim if F is b i j e c t i v e a n d b o t h F
a n d F -I are s m o o t h m a p p i n g s . T w o m a n i f o l d s N a n d M are diffeomorphic
if t h e r e exists a diffeomorphism F : N ~ M. []

The rank o f a m a p p i n g F : N ~ M at a point p E N is the rank of

the jacobian matrix

~fl ~fl
Sx I "-- Sx n

~f ~f
m m
~x I "'" ~x n

at x = ~(p). It m u s t be stressed that, although apparently dependent

on t h e c h o i c e of local coordinates, the n o t i o n of rank thus defined is
actually coordinate-independent. The reader may easily verify that the
r a n k s o f the jacobian matrices of two d i f f e r e n t expressions of F in
local coordinates are equal.

Theorem. L e t N a n d M b e s m o o t h m a n i f o l d s both of dimension n. A m a p -

ping F : N ~ M is a d i f f e o m o r p h i s m if a n d o n l y if F is b i j e c t i v e , F
is s m o o t h and rank(F) = n at a l l points o f N.

Remark. I n s o m e c a s e s , the a s s u m p t i o n that functions, mappings, etc.

a r e C ~, m a y be replaced by the s t r o n g e r assumption that functions,
mappings,etc, are a n a l y t i c . I n this way one m a y define the n o t i o n of
analytic manifold, analytic mappings of manifolds, and so on. We shall
 264

make this assumption explicitly whenever needed.

4. S u b m a n i f o l d s

Definitions. Let F : N ~ M be a s m o o t h mapping of manifolds.

(i) F is an immersion i f r a n k ( F ) = dim(N) for all p C N.

(ii) F is a univalent immersion if F is an i m m e r s i o n and is i n j e c t i v e .
(iii) F is an embedding if F is a u n i v a l e n t immersion and the topology
induced o n F(N) by the o n e o f N c o i n c i d e s with the topology of
F(N) as a s u b s e t o f M. []

Remark. The mapping F, b e i n g smooth, is in p a r t i c u l a r a continuous

mapping of topological spaces. Therefore (see s e c t i o n 2) the topology
induced o n F(N) by the o n e of N may properly contain the topology of
F(N) as a s u b s e t of M. This motivates the d e f i n i t i o n (iii). []

The difference between (i), (ii) and (iii) is c l a r i f i e d in the fol-

lowing examples.

Examples. Let N = ~ and M = ~2. Let t denote a point in N a n d (Xl,X 2)

a point in M. T h e m a p p i n g F is d e f i n e d by

x1(t) = at-sin t
x2(t) = cos t

and, then,

i a - cos t
rank(F) = rank
- sin t

If a = I t h i s m a p p i n g is not a n i m m e r s i o n because rank(F) = 0 at

t = 2k~ (for a n y integer k).
I f 0 < a < I the m a p p i n g is a n i m m e r s i o n , because rank(F) = I for
all t, b u t no~ a u n i v a l e n t immersion, because F ( t I) = F ( t 2) f o r all
tl,t 2 such that t I = 2k~-T, t 2 = 2k~+T and sin T = aT.
As a second example we consider the so-called "figure-eight".Let
N be the o p e n interval (0,2~) of the real line and M = ~2. Let t
denote a point in N a n d (xl,x 2) a p o i n t in M. The mapping F is d e f i n e d
by
x1(t) = sin 2t

x2(t) = sin t
 265

X2 X
2

\ -<~--~'~~X 1
a=! O<a<l

This mapping is a n i m m e r s i o n because

dx 1 2 cos2t

rank(F) = rank =
= I

dx 2
q-6 cos t

for a l l 0 < t < 27. It is a l s o univalent because

F(tl) = F ( t 2) ~ t I = t 2

However, the mapping is not an e m b e d d i n g . For, consider the image o f F.

The mapping F takes the open set
(~-e,~+) of N onto a subset U' of F(N)
which is o p e n by definition in the to-
pology induced b y the o n e o f N, but is
not an open set in t h e topology of F(N)
as a s u b s e t o f M. This is b e c a u s e U'
cannot be seen as the intersection of
F(N) with an o p e n set of ~ 2 .
As a third example one may consider
the m a p p i n g F : ~ ~3 given by

xl (t) = C O S 2~t

x2(t) = sin 2~t

x3(t) = t

whose image is an " h e l i x " winding o n an infinite cylinder whose axis

 266

is the x 3 a x i s . The reader may easily check that is an e m b e d d i n g . []

The following theorem shows that the e v e r y immersion locally is an
embedding.

Theorem. Let F : N ~ M be an immersion. For each p 6 N there exists a

neighborhood U of p with the p r o p e r t y that the r e s t r i c t i o n o f F to U
is an e m b e d d i n g .

ExampZe. Consider again the "figure eight" discussed above. If U is

any interval o f the type (6,2~-6), then the c r i t i c a l situation we had
before disappears and the image U' of (z-e,n+E) is n o w o p e n also in
the topology o f F(N) as a s u b s e t of ~ 2 . []

The notions of univalent immersion and of embedding are u s e d in

the following way.

Definition. T h e i m a g e F(N) o f a u n i v a l e n t i m m e r s i o n is called an ~m-

mersed submanifold o f M. T h e i m a g e F(N) o f an e m b e d d i n g is c a l l e d an
embedded submanifold o f M.
Remark. Conversely, one may say t h a t a s u b s e t M' o f M is an immersed
(respectively, embedded) submanifold o f M if t h e r e is a n o t h e r manifold
N and a univalent immersion (respectively, embedding) F: N ~ M s u c h
that F(N) = M'. []

The use of the w o r d "submanifold" in the a b o v e definition clearly

indicates the p o s s i b i l i t y of giving F(N) the structure of a smooth ma-
nifold, and this may actually be done in the following way. Let M'=F(N)
a n d F': N ~ M' denote the m a p p i n g defined by
F' (p) = F(p)
for all p 6 N. Clearly, F' is a b i j e c t i o n . I f the topology of M' is
the o n e induced by that of N (i.e. open sets o f M' are the images
under F' of open sets o f N), F' is a h o m e o m o r p h i s m . Consequently, any
coordinate chart (U,~) of N induces a coordinate chart (V,~) o f M',
defined as

-I
V = F' (U) , % = ~o (F')

C~-compatible charts of N induce C~-compatible charts o f M' and so

complete C~-atlases induce complete C~-atlases. This gives M' the
structure of a smooth manifold.
The smooth manifold M' thus defined is diffeomorphic to the
smooth manifold N. A d i f f e o m o r p h i s m between M' and N is i n d e e d F'
itself, which is b i j e c t i v e , smooth, and has rank equal to the dimension
of N at each p 6 N.
 267

Embedded submanifolds can also be characterized in a different

way, based on the following considerations.
Let M be a smooth manifold of dimension m and (U,~) a cubic
coordinate chart. Let n be an integer, 0 < n < m, and p a point of U.
The subset of U

Sp = {q E U: xi(q) = xi(P), i = n+1, .... m}

is c a l l e d an n-dimensional slice o f U p a s s i n g through p. In other

words a slice of U is the locus of all points of U for which some
coordinates (e.g. the last m-n) are constant.

Theorem. Let M be a smooth manifold of dimension m. A subset M' of M

is an e m b e d d e d submanifold of dimension n < m if a n d only if for each
p E M' there exists a cubic coordinate chart (U,~) of M, with p 6 U,
such that U ~ M' coincides with an n - d i m e n s i o n a l slice of U passing
through p.

This theorem provides a more "intrinsic" characterization of the

notion of an e m b e d d e d submanifold (of a m a n i f o l d M), directly related
to the existence of special coordinate charts (of M ) . Note that, if
(U,~) is a coordinate chart of M such that U N M' is an n - d i m e n s i o n a l
slice of U, the pair (U',~') defined as

U' = U N M'

~' (p) = (x I (P) ..... X n ( P ) )

is a coordinate chart of M'. This is illustrated in the following

figure (where M = ~3 and n = 2).

Remark. Note that an o p e n

subset M' of M is indeed an
embedded submanifold of M,
of the same dimension m.
Thus, a submanifold M' of
M may be a proper subset of
M, although being a manifold
of the same dimension.

Remark. It can be proven that any smooth hypersurface in ~ m is an

embedded submanifold of ~m Moreover it h a s also been shown that if N
is an n-dimensional smooth manifold, there exist an integer m > n and
a mapping F :N ~ m which is a n embedding (Whitney's embedding theo-
 268

rem). In o t h e r words, any m a n i f o l d is d i f f e o m o r p h i c to an e m b e d d e d

submanifold of ~ m, for a s u i t a b l y large m.

Remark. Let V be an n - d i m e n s i o n a l subspace o f ~ m. A n y subset of~ TM of

the form

x+V = {x e ~ m : x = x'+x; x' e V},

where x is some fixed point of ~ m, is indeed a smooth hypersurface

and so an e m b e d d e d submanifold of ~ m , of d i m e n s i o n n. This is some-
times called a flat submanifold of ~ n.

5. T a n g e n t vectors

Let N be a smooth manifold of d i m e n s i o n n. A r e a l - v a l u e d func-

tion I is s a i d to be smooth in a neighborhood of p, if the d o m a i n of
h includes an o p e n set U of N c o n t a i n i n g p and the restriction of l
to U is a s m o o t h function. The set of all smooth functions in a n e i g h -
borhood of p is d e n o t e d C~(p). Note that C~(p) forms a vector space
over the f i e l d ~. For, if I, X are functions in C~(p) and a,b are real
numbers, the function ah+bx defined as

(al+bx) (q) = a l ( q ) + b x ( q )

for all q in a n e i g h b o r h o o d of p, is a g a i n a function in C ~ ( p ) . Note

also that two functions l,X E C~(p) may be m u l t i p l i e d to g i v e another
element of C~(p), written lX and defined as

(hX) (q) = h ( q ) ' x ( q )

for all q in a n e i g h b o r h o o d o f p.

Definition. A tangent vector v at p is a m a p v: C~(p) ~ with the

following properties:

(i) (linearity) : v(al+bx) = av(1)+bv(y) for all l,X E C~(p) and

a,b E
(ii) (Leibnitz rule): v(Ix) = X(p)v(/)+l(p)v(~) for all I,X e C~(p).

Definition. L e t N be a smooth manifold. The tangent space to N at p,

written TpN, is the set of all tangent vectors at p.

Remark. A map which satisfies the properties (i) a n d (ii) is a l s o

called a derivation.
 269

Remark. The set TpN forms a v e c t o r space over the field ~ under the
rules of scalar m u l t i p l i c a t i o n and a d d i t i o n d e f i n e d in the f o l l o w i n g
way. If v l , v 2 are t a n g e n t vectors and Cl,C 2 real numbers, clv1+c2v2
is a new tangent v e c t o r w h i c h takes the f u n c t i o n I E Ca(p) into the
real n u m b e r

(ClV1+C2V2) (I) = clv I(I) + c2v 2(I)

R e m a r k . We shall see later on that, if the m a n i f o l d N is a smooth

hypersurface in~ TM, the o b j e c t p r e v i o u s l y d e f i n e d may be n a t u r a l l y
i d e n t i f i e d w i t h the i n t u i t i v e notion of " t a n g e n t h y p e r p l a n e " at a
point. []

Let (U,9) be a (fixed) c o o r d i n a t e c h a r t a r o u n d p. W i t h this

c o o r d i n a t e c h a r t one may a s s o c i a t e n t a n g e n t v e c t o r s at p, d e n o t e d

d e f i n e d in the f o l l o w i n g way

(V~ai)p (x) = (a(Xo

~x i -I )x=~ (p)

for I ~ i ~ n. The r i g h t - h a n d - s i d e is the value taken at x = (Xl,...,Xn)=

= ~(p) of the partial d e r i v a t i v e of the function Io~-l(xl .... ,x n) w i t h
--I
r e s p e c t to x i (recall that the f u n c t i o n Io~ is an e x p r e s s i o n of 1
in local c o o r d i n a t e s ) .

Theorem. Let N be a smooth m a n i f o l d of d i m e n s i o n n. Let p be any p o i n t

of N. The tangent space TpN to N at p is an n - d i m e n s i o n a l vector space
o v e r the field ~. If (U,~) is a c o o r d i n a t e c h a r t a r o u n d p, then the
tangent v e c t o r s (~)pd~ ..... (~-~-)~ form a basis of TpN. []
I n p
The basis {(~)p ..... ( ~ ) p } of TpN is s o m e t i m e s c a l l e d the
I I

natural basis induced by the c o o r d i n a t e c h a r t (U,~).

Let v be a t a n g e n t vector at p. F r o m the above t h e o r e m it is seen
that

n
v : [ v i(a--~)
i=I i p

where V l , . . . , v n are real numbers. One may compute the v~s e x p l i c i t l y

in the f o l l o w i n g way. Let ~i be the i-th c o o r d i n a t e function. Clearly
 270

6 Ca(p), and then

n n 3 (~io~ -I)
v(gi) = ) ~ivj(~'~j
j p(~i ) = ( [IVj
j 3xj )x=~(p) = v.l

because ~ i ~
-I(Xl,...,x n) = x i. Thus the real number v i coincides with
the value of v at ~i ' the i-th coordinate function.
A change of coordinates around p clearly induces a change of basis
in TpN. The computations involved are the following ones. Let (U,~)
and (V,) be coordinate charts around p. Let {( )p, .... ( ~3 }p} denote
the natural basis of T N induced by the coordinate chart (V,). Then
P

~) (I) = (3(ho-I~ = (3(1o~ -Io~o-I~ =

(3i P ~i "Y= (p) 3Yi Y= (p)

(3(ho _1)) (3(~j o-I)

j=1 "~xj x=~(p) 3y i )y=%(p)

n 3 3 (~jo-1)
j=1 3Yi )Y= (P)

In other words

) = n 3(~j0 -I) (_~4)

j=1 3Yi Y=% (p) j p

Note that the quantity

3(~jo -1)
3Y i

is the element on the j-th row and i-th column of the jacobian matrix
of the coordinate transformation

x = x{y)

So the elements of the columns of the jacobian matrix of x = x(y)

are the coefficients which express the vectors of the "new" basis as
linear combinations of the vectors of the "old" basis.
 271

If V is a t a n g e n t vector, and (Vl, .... Vn), (Wl, .... Wn) the n - t u p l e s

of real n u m b e r s which express v in the form

n n
! )p
v = v i =
i=I ~i p i=I

then

ax 1 ~x I
v1 wI
3Yl " " aYn

~x ax
vn n n wn
3y I " " aYn

Definition. Let N and M be s m o o t h m a n i f o l d s . Let F : N ~ M be a smooth

mapping. The differential of F at p E N is the m a p

F~ : TpN -+ T F ( p ) M

defined as follows. For v C TpN and I 6 C ~(F(p)),

(F~(v))(~) = V(~oF)

Remark. F, is a m a p of the t a n g e n t space of N at a p o i n t p into the

tangent space of M at the p o i n t F(p). If v 6 TpN, the value F%(v) of
F~ at v is a t a n g e n t vector in TF(p)M. So one has to e x p r e s s the w a y
in w h i c h F,(v) maps the set C~(F(p)), of all functions which are smooth
a neighborhood of F(p), into ~. This is a c t u a l l y what the d e f i n i t i o n
specifies. Note that there is one of such m a p s for each p o i n t p o f N.

Theorem. The d i f f e r e n t i a l F~ is a l i n e a r map. []

Since F% is a l i n e a r map, given a basis for TpN and a b a s i s for

TF(p)M one m a y w i s h to find its m a t r i x representation. Let (U,~) be a
coordinate chart around p, (V,%) a coordinate chart around q = F(p),
{(~-q--)p ..... ( the n a t u r a l b a s i s of T N and { ( .... (
~la ~ ! )np } P ~1,q, _~m)q }
the n a t u r a l b a s i s of T q M In o r d e r to find a m a t r i x r e p r e s e n t a t i o n
of F~ , one has simply to see how F, m a p s (~-~--~)P~*ifor e a c h I ~ i ~ n.

p(loF) = (~ (IF~-I))
(F~(~i). p) (I) = (~.)i x=~(P )
 272

(toq-looF= -1)) m (~o+-I)) (

~(%joF~ -I)
)
ax i x=~(p) ~ ( - ~yj
= j=l - y=~ (q) ~x i x=~ (p)

m ~ ~(~joFo~ -I)
j=1 ((~ )q (i)) ( ~xi )x=? (p)

In o t h e r w o r d s

m a(+joFo~ -I)
F. (~--~i)P = j=1 axi )x=~(p) a~j)q

-I
Now, recall that %oFo~ is an e x p r e s s i o n of F in local coordinates
Then, the q u a n t i t y

a(%joFo~ -I)
~x1

is the e l e m e n t on the j-th row and i-th column of the jacobian matrix
of the m a p p i n g expressing F in local coordinates. Using again

F I (X I ..... x n)

F(X) = F(xl, .... x n)

Fm(X I , .. ,x n)

-I
to d e n o t e ~oFo~ , one has simply

F% ( ~ i ) P ~ aF.
j=1 i

If v 6 TpN and w = F~ (v) C T F ( p ) M are e x p r e s s e d as

n m
v = v i w = w
i=I i p i=I i
then
aF I ~F I
wI vI
~Xl --- ~Xn
=

~F m ~F m
W m Vn
~x I " .. ~x n
 273

Remark. The m a t r i x r e p r e s e n t a t i o n of F~ is e x a c t l y the j a c o b i a n of its

expression in local c o o r d i n a t e s . F r o m this, it is seen that the rank
of a m a p p i n g c o i n c i d e s w i t h the rank of the c o r r e s p o n d i n g d i f f e r e n t i a l .

Remark (Chain rule). It is easily seen that, if F and G are smooth

mappings, then

(GoF). = G,F.

The f o l l o w i n g examples may clarify the n o t i o n of tangent space

and the one of differential.

Example. The tangent vectors on I~n. Let ]Rn be e q u i p p e d with the "natur-
al" complete atlas already c o n s i d e r e d in p r e v i o u s e x a m p l e s (i.e. the
one i n c l u d i n g the chart ~n, i d e n t i t y map on ~ n ) ) . Then, if v is a
tangent v e c t o r at a p o i n t x and I a s m o o t h f u n c t i o n

n ~ n 3t
v(l) = i-~-Iv i ( ~ i ) x ( l ) = i=I[ (~
X ) V.l

SO, V(1) is just the value of the derivative of I along the direction
of the vector

col(v I ..... v n )

at the p o i n t x.

Remark. Let F : N ~ M be a u n i v a l e n t immersion. Let n = d i m ( N ) and

m = dim(M). By definition, F~ has rank n at each point. Therefore the
image F~(TpN) of F~ , at each point p, is a subspace of T F ( p ) M isomor-
phic to TpN. The subspace F~(TpN) can a c t u a l l y be identified w i t h the
tangent space at F(p) to the s u b m a n i f o l d M' = F(N). In order to un-
d e r s t a n d this point, let F' denote the f u n c t i o n F': N ~ M' d e f i n e d as

F' (p) = F(p)

for all p C N. F' is a d i f f e o m o r p h i s m and so F~ is an isomorphism.

Therefore the image FI(TpN) is e x a c t l y the t a n g e n t space at F' (p) to
M'. Any tangent v e c t o r in TF(p)M' is the image F~(v) of a (unique)
v e c t o r v 6 TpN and can be i d e n t i f i e d w i t h the (unique) v e c t o r F~(v)
of F,(TpN).
In other words, the tangent space at p to a submanifoZd M' of M
can be i d e n t i f i e d with a subspac~ of the tangent space at p to M.
The same c o n s i d e r a t i o n s can be r e p e a t e d in local c o o r d i n a t e s . It
 274

is k n o w n that an i m m e r s i o n is l o c a l l y an e m b e d d i n g . Therefore, around

every point p 6 M' it is p o s s i b l e to find a coordinate chart (U,~) of
M, w i t h the p r o p e r t y that the p a i r (U',~') defined by

U' = {q 6 U : ~ i ( q ) = ~i(p), i =n+1,...,m}

~' = (~1,...,~n)

is a c o o r d i n a t e chart of M'. According to this choice, the tangent

space to M' at p is identified with the n - d i m e n s i o n a l subspace of
TpM spanned by the tangent vectors { (~)p ..... ( ~ ) p } . []
I

ExampZe. The tangent vector to a s m o o t h curve in ~ n . We define first

the notion of a s m o o t h curve in ~ n Let N = (tl,t 2) be an o p e n interval
on the real line. A smooth curve in ~ n is the image of a u n i v a l e n t
immersion a : N ~ ~n. Thus, a smooth curve is an i m m e r s e d submanifold
of ~ n . In N a n d ~ n one m a y choose natural local coordinates as usual
and, letting t denote an e l e m e n t of N, express ~ by means o f an n - t u p l e
of s c a l a r - v a l u e d functions o1,...,On of t.
A smooth curve is a l - d i m e n s i o n a l immersed submanifold of ~ n . At
a point a(to) , the tangent space to the c u r v e is a l - d i m e n s i o n a l vector
space which, as w e h a v e s e e n , m a y be identified with a subspace of the
tangent space to ~ n at this point. A basis of the
t a n g e n t s p a c e to the
d
curve at O(to) is g i v e n by the i m a g e u n d e r o, of ( ~ ) t , a tangent
o
vector at t o to N. This image is c o m p u t e d as f o l l o w s

n do.
o,(~t) t = ~ (~) (DTi)O(to)
o i=1 tO '

Thinking of t @ N as time and o(t) as a p o i n t moving in ~ n , we m a y in-

terpret the v e c t o r

do I do n )
col ((-~-) to ..... (-a~) to

as the velocity along the curve, evaluated at the p o i n t o(t). So, we

have that the v e l o c i t y vector at a p o i n t of the c u r v e spans the tangent
space to the curve at this point. From this point of view, we see that
the n o t i o n of tangent space to a l - d i m e n s i o n a l manifold may be identi-
fied with the geometric notion of tangent line to a c u r v e in a E u c l i d e a n
space.

Example. L e t h be a s m o o t h function h :~2 ~ and F :~2 ~3 a mapping

defined by
 275

F(xl,x2) = (xl,x2,h(xl,x2))

This m a p p i n g is an e m b e d d i n g and therefore F ~ 2 ) , a surface in ~2, is

an e m b e d d e d s u b m a n i f o l d of ~3. At each point F(x) of this surface, the
tangent space,ldentified as a subspace of the tangent space to ~ 3 at
this point, may be c o m p u t e d as

span{F, (~xl) x , F, (~x2) x}

Now,

3 3F 1 3 (3h ,
F , ( T X 1)x =1~i.=(-~i) (~-~i)F(x) = (Txl)F(x) + ~ i ) (-~3)F(x)

) = 3 + ~h (~x3)

o o o
This tangent space to F(~ 2) at some point (x~,x2,h(Xl,X2)) is the set
of tangent vectors whose expressions in local c o o r d i n a t e s are of the
form

V = B

3h 8h

~h ~h
~,~ being real numbers and ~ ,~ being e v a l u a t e d at x I = x~ and
o
x 2 = x 2. F r o m this point of view, we see that the notion of tangent
space to a 2-dimensional m a n i f o l d may be identified with the geometric
notion of tangent plane to a surface in a E u c l i d e a n space. []
One may define objects dual to the ones c o n s i d e r e d so far.

Definition. Let N be a smooth manifold. The cotangent space to N at p,

written T'N, is the dual space of T N Elements of the c o t a g e n t
P P
space are c a l l e d tangent covectors.

Remark. Recall that a dual space V of a vector space V is the space

of all linear functions from V to ~. If v E V , then v : V ~ and
the value of v at v 6 V is w r i t t e n as ( v * , v ) . V forms a v e c t o r space
over the field ~, with rules of scalar m u l t i p l i c a t i o n and addition
which define c l v 1 + c 2 v 2 in the following terms
 276

(clv1+c2v2,v)= c I ( v I ,v ) + c 2 (v 2 ,v )

If e l , . . . , e n is a b a s i s of V , the u n i q u e basis e ~n, . ]

..,e of V which
satisfies

.,e > = ~..

z 3 z3

is c a l l e d a dual basis,
If V a n d W are v e c t o r spaces, F : V ~ W a linear mapping and wEW,
v E V, the m a p p i n g F : W ~ V defined by

< F * ( w * ) , v > = <w , F ( v ) >

is c a l l e d the dual m a p p i n g (of F). []

Let I be a smooth function I : N ~. There is a n a t u r a l way of
identifying the differential I, of I at p w i t h an e l e m e n t of TpN. For,
observe that I~ is a l i n e a r mapping

I, : TpN ~ Tl(p)~

and that Tl(p)~ is i s o m o r p h i c to ~ . The n a t u r a l isomorphism between

and Tl(p)~ is the one in w h i c h the e l e m e n t c of ~ corresponds to the
tangent vector c ( ~ t ) t" If c (~t) t is the v a l u e at v of the d i f f e r e n t i a l
I~ at p, then c must depend linearly on v, i.e. there must exist a co-
vector, denoted (dl)p , such that

l.(V) < (dl)p,V) d

= (~E)t

Given a basis of TpN, the c o v e c t o r (dl) (like any other covector),

may be r e p r e s e n t e d in m a t r i x form. Let { (~ )~p ,P. . . , (~ )p } be the

natural basis
of T N i n d u c e d by the coordinate chart (U,~). The image
P
under l~ of a v e c t o r

n
v = ~ vi( )
i=I ~ p
is the v e c t o r
n
~, (v) = ( 2 ~ .~ vi) (~t) t
i=I 1

and this shows that

 277

{ (dl)p,V)=
21
( 8x I "'"
81
~n )
vl)
vn

Remark. N o t e a l s o t h a t the v a l u e at I of a t a n g e n t v e c t o r v is e q u a l to
the v a l u e at v of the t a n g e n t c o v e c t o r (dl)p , i.e.

v(1) = ( (dl)p,V } []

The dual b a s i s of {(~-~1)p .... ,( )p} is c o m p u t e d as f o l l o w s .

F r o m the e q u a l i t y v(1) = ((dl)p,V) we deduce that

8 ( ~ i o ~ -I) 8x.
< (d~i)p' ( 3 )p > = ( )p ( ~i) 8x . 8x.3 i3

so t h a t the d e s i r e d dual basis is e x a c t l y provided by the set of tan-

g e n t covectors, {(d~1 )p, .... (d~n)p}.
If v is any t a n g e n t eovector, expressed as

, n ,
v = ~ vi(d~i) p ,
i=I
,
the real numbers v.],...,v n are such that

v i = <v ,( )p

Note also that, if v is any t a n g e n t vector expressed as

n
v = [ v i(~-~i )p
i=1

the real n u m b e r s V l , . . . , v n are s u c h that

v i = < (d~i)p,V) .

6. V e c t o r fields

Definition. L e t N be a s m o o t h m a n i f o l d , of d i m e n s i o n n. A vector field

f on N is a m a p p i n g assigning to e a c h p o i n t p E N a t a n g e n t v e c t o r f(p)
in TpN. A vector field f is smooth if for e a c h p E N t h e r e e x i s t s a
 278

coordinate chart (U,~) about p and n r e a l - v a l u e d smooth functions

fl,...,fn defined on U, such that, for all q ~ U,

n
f(q) = ~ fi(q) ( ~ ) q
i=1

Remark. Because of C ~ - c o m p a t i b i l i t y of coordinate charts, given any

coordinate chart (V,~) about p other than (U,9), one may find a neigh-
b o r h o o d V' C V of p and n r e a l - v a l u e d smooth functions f~'" "''f'n de-
fined on V' , such that, for all q E V'

n
,2)
f(q) = [ f~ (q) 'D--~i q
i=I l

Thus, the notion of smooth vector field is independent of the coord-

inates used.

Remark. If (U,@) is a coordinate chart of N, on the submanifo~d U of

N one may define a special set of smooth vector fields, denoted
.... ~ n ' in the following way
~I'

(~i): P ~-+(~.)p

It m u s t be stressed, however,that such a set of vector fields

is an object defined only on U. []

For any fixed c o o r d i n a t e chart (U,~), the set of tangent v e c t o r s

{ (~-~)q ..... (~-~)q} is a basis of TqN at each q 6 U, and therefore

there is a unique set of smooth functions {f1' .... fn } that makes it

possible to express the value of a vector field f at q in the form

n
f(q) = [ fi (q) ~-~-~)
i=I (~i q

Expressing each f. in local coordinates, as

1

provides an e x p r e s s i o n in local c o o r d i n a t e s of the vector field f

itself. So, if p is a point of coordinates (Xl,...,x n) in the chart
(U,~),f(p) is a tangent vector of c o e f f i c i e n t s (fl (Xl, .... Xn),...
.... fn(Xl ..... Xl)) in the natural basis {(~) ,.. , ( ~ ) } of T N
I P " a~n p P
 27g

induced by (U,~). M o s t of ~he times, w h e n e v e r p o s s i b l e , the s y m b o l f.

-I l
replaces f o~ a n d the e x p r e s s i o n o f f in l o c a l c o o r d i n a t e s is g i v e n
1
a f o r m of an n - v e c t o r f = c o l ( f l , .... fn ).

Remark. Let f be a s m o o t h v e c t o r field, (U,9) and (V,%) two c o o r d i n a t e

charts about p a n d f(x) = f(xl,...,Xn), f' (y) = f' (yl,...,yn) the cor-
responding expressions of f in l o c a l c o o r d i n a t e s . Then

f' (y) = (~xf(X))x=x(y) []

The n o t i o n of v e c t o r field makes it p o s s i b l e to i n t r o d u c e the

concept of a differential equation on a m a n i f o l d N. For, let f be a
smooth vector field. A smooth curve a: (tl,t2) ~ N is an ~ntegral curve
of f if

d
o~(~-{) t = f(o(t))

for all t 6 (tl,t2). The left-hand-side is a t a n g e n t vector to the sub-

manifold o((tl,t2)) at the p o i n t o(t); the r i g h t - h a n d - s i d e is a t a n g e n t
vector to N at ~(t). As usual, we i d e n t i f y the t a n g e n t space to a sub-
manifold of N at a p o i n t w i t h a subspace of the t a n g e n t space to N at
this point.
In l o c a l coordinates, o(t) is e x p r e s s e d as an n - t u p l e
(~1(t), .... ~n(t)), and f(~(t)) as

f(o(t)) = ~ fi(o1(t) ..... On(t)) ( ~ i ) o ( t )

i=]

Moreover

d i
o~(~t)t = ~ ~(~i)~(t)
i=I

Therefore, the e x p r e s s i o n of o in local coordinates is s u c h that

do
1
dt - fi(ol (t} ..... On(t))

for all 1 < i < n. T h i s shows t h a t the n o t i o n of i n t e g r a l curve of a

vector field corresponds to the n o t i o n of s o l u t i o n of a set of n
ordinary differential equations of the f i r s t order.
F o r this reason, one o f t e n u s e s the n o t a t i o n
 280

d
to i n d i c a t e the image of (~)t under the d i f f e r e n t i a l ~, at t.
The following theorem contains all relevant informations about
the p r o p e r t i e s of integral curves of v e c t o r fields.

Theorem. Let f be a s m o o t h vector field on a m a n i f o l d N. For each

p 6 N, there exist an o p e n interval - depending on p and w r i t t e n Ip -
of ~ such that 0 E I and a smooth mapping
P

%:W~N

defined on the subset W of ~ N

W = {(t,p) ~xN: t Ip}

with the following properties:

(i) ~(0,p) = p,
(ii) for e a c h p the m a p p i n g Op:Ip ~ N defined by

Op(t) = #(t,p)

is an i n t e g r a l curve of f,
(iii) if ~ : (tl,t 2) ~ N is a n o t h e r integral curve of f satisfying the
condition ~(0) = p, then (tl,t2) C Ip and the restriction of Op
to (tl,t 2) coincides with ~,
(iv) ~(S,~(t,p)) = (s+t,p) whenever both sides are defined,
(v) whenever #(t,p) is d e f i n e d , there exists an o p e n neighborhood U
of p s u c h that the m a p p i n g ~t: U ~ N defined by

Ct(q) = #(t,q)

is a d i f f e o m o r p h i s m onto its image, and

~1 = ~-t

Remark. Properties (i) and (ii) say that ~p is an i n t e g r a l curve of f

passing through p at t = 0. Property (iii) says that this curve is
unique and that the domain I on w h i c h ~ d e f i n e d is m a x i m a l . P r o p e r t y
P P
(iv) and (v) say that the f a m i l y of m a p p i n g s {t } is a o n e - p a r a m e t e r
(namely, the p a r a m e t e r t) group of local diffeomorphisms, under the
operation of c o m p o s i t i o n . []
 281

Example. Let N = ~ and use x to d e n o t e a point in N. C o n s i d e r the v e c t o r

field

f(x) = (x 2 + I ) ( ~ x ) x

An integral curve ~ of f must be such that

(t) = do (~x) x =
(~-~) (02(t)+i) (~x) x

so

d~ - o 2 + I
dt

A solution of this equation has the form

0(t) = t g ( t + t g - 1 (x))

with x being indeed the v a l u e of ~ at t = 0. C l e a r l y , for e a c h x the

solution is d e f i n e d for

_ ~ < t + t g -I (x ) < ~_
2 2

Thus W is t h e set

W = { ( t , x O) : t e (-~-tg-1(x), ~ tg-1(x))}

which has the form indicated below. []

The mapping ~ is c a l l e d the flow of
f. O f t e n , for p r a c t i c a l purposes, the
notation ~t r e p l a c e s ~, w i t h the under- IN
standing that t is a variable. To stress
the d e p e n d e n c e o n f, s o m e t i m e s Ct is
f
w r i t t e n as ~t"

Definition. A vector field f is complete

if, for all p ~ N, the interval I
P
coincides with ~, i.e. - in other
words -if the flow # of f is d e f i n e d on
the w h o l e cartesian product ~XN. []

The integral curves of a c o m p l e t e

vector field are thus defined, whatever
the i n i t i a l point p is, for a l l t C ~.
 282

Definition. Let f be a s m o o t h vector field on N and I a smooth real-

valued function on N. The Lie derivative of I along f is a f u n c t i o n
N ~, written Lfl and defined by

(Lfl) (p) = (f(p)) (I)

(i.e. (Lfl) (p) is the value at I of the tangent vector f(p) at p). []

The function Lfl is a s m o o t h function. In l o c a l coordinates, Lfl

is r e p r e s e n t e d by

fl
) -
(Lfl) (x I .... ,Xn) = ~x I ... ~x n

fn
If fl,f2 are v e c t o r fields and I a real-valued function, we de-
note

LflLf 2 I = Lf 1 (Lf21)
The set of all smooth vector fields on a m a n i f o l d N is d e n o t e d by
the symbol V(N). This set is a Vector space o v e r ~ since if f, g are
vector fields and a,b are real numbers, their linear combination af+bg
is a v e c t o r field defined by

(af+bg) (p) = a f ( p ) + b g ( p )

If a , b are smooth real-valued functions on N, one m a y still define a

linear combination af+bg by

(af+bg) (p) = a ( p ) f ( p ) + b ( p ) f ( p )

and this gives V(N) the structure of a module o v e r the ring, denoted
C~(N), of all smooth real-valued functions defined on N. The set V(N)
c a n be q i v e n , h o w e v e r , a more interesting algebraic structure in t h i s way.

Definition. A v e c t o r space V over ~ is a Lie algebra if in a d d i t i o n to

its v e c t o r space structure it is p o s s i b l e to d e f i n e a binary operation
V x V ~ V, called a product and written [ .,.], which has the following
properties

(i) it is skew commutative, i.e.

Iv,w! = -[w,v]
 283

(ii) it is bilinear over ~, i.e.

[alVl+e2v2,w] = a1[v 1,w]+~2 |v2,w]

(where a1,~ 2 are real numbers)

(iii) it satisfies the so called Jacob~ identity, i.e.

[v,[w,z]]+ [w,[z,v]]+ [z,[v,w]] = 0. []

The set V(N) forms a Lie algebra with the vector space structure
already discussed and a product [ .,.] defined in the following way.
If f and g are vector fields, [f,g] is a new vector field whose value
at p, a tangent vector in TpN, maps C~(p) into ~ according to the rule

([ f,g~ (p)) (I) = (LfLgl) (p)-(LgLfl) (p)

In other words, [ f,g] (p) takes I into the real number (LfLgl) (p)-
+(LgLfl) (p). Note that one may write more simply

L[ f,g] I = LfLgl-LgLfl

Theorem. V(M) with the product [ f,g] thus defined is a Lie algebra. []

The product [ f,g] is called the Lie bracket of the two vector
fields f and g.

Remark. If f,g are smooth vector fields and I,X smooth real-valued
functions, then

[ lf,xg] = l.y-[ f,g] + l'Lfx'g - X'Lgl'f

Note that /,x,Lfy,Lgl are elements of C~(N), and g,f,[ f,g] elements of
V(N). []

The reader may easily find that the expression of [ f,g] in local
coordinates is given by the n - v e c t o r

~gl 8gl fl ~)fl

~Xl "'" ~n

~gn
~x I "'"
~gn
~x n ill ~f
n
i

~f
n
gl1

gn
=~f _
~g
~f
 284

If, in p a r t i c u l a r , N = ~n and

f(x) = Ax , g(x) = Bx

then

[ f,g] (x) = (BA-AB)x

The m a t r i x [A,B] = (BA-AB) is c a l l e d the commutator of A,B.

The importance of the n o t i o n of L i e bracket of v e c t o r fields is
very much related to its a p p l i c a t i o n s in the study of n o n l i n e a r con-
trol systems. For the m o m e n t , we give hereafter two interesting pro-
perties.

Theorem. Let N' be an e m b e d d e d submanifold of N. L e t U' be an o p e n

s e t of N' and f,g two smooth vector fields of N s u c h that for all
p 6 U'

f(p) 6 T p N' and g(p) E T pN ' .

Then also

[ f,g] (p) 6 TpN'

for all p E U'. []

In o t h e r words, the L i e b r a c k e t of two v e c t o r fields "tangent"

to a f i x e d submanifold is s t i l l tangent to t h a t submanifold.

Theorem. Let f,g be two s m o o t h vector fields on N. L e t ~tf d e n o t e the

flow of f. F o r e a c h p E N.

lira ~[I ( # f t ) ~ g ( ~ f ( p ) ) _ g ( p ) ] = [ f,g] (p)

t~O

Remark. The first t e r m of the expression under bracket is a t a n g e n t

vector at p, o b t a i n e d in the following way. With p, the m a p p i n g ~
(always defined for sufficiently small t) associates a point q=k(p).
The v e c t o r field g is e v a l u a t e d at q, and the v a l u e g(q) ~ T N is
q
taken back to T _ N v i a the differential (f~)~ (which m a p s the t a n g e n t
space at q o n t o the tangent space at p = ~ _ t ( q ) ) Thus, the m a p p i n g
p ~-+ ( ~ft)
_ ,g(~(p)) defines a vector field, on the d o m a i n of f
~t"

Remark. Let f be a smooth vector field on N, g a smooth vector field

on M a n d F : N ~ M a s m o o t h function. The vector fields f,g a r e said to
 285

be F-related if

Fef = goF

Note that the v e c t o r field (#ft)~g({(p))_

5 considered in the a b o v e
Remark is ~ft-related_ to g.

Remark. If f is F - r e l a t e d to f a n d g is F - r e l a t e d to g, then |f,g] is

F-related to [ f,g].

D~f~nition. L e t f,g be two s m o o t h v e c t o r fields on N. T h e Lie derivative

of g along f is a v e c t o r f i e l d on N, w r i t t e n Lfg and defined by

(Lfg) (p) = l i m l[
t ( ~ -f t ) ~ g (#~(p)) - g ( P )| " []
t~0

Thus, by d e f i n i t i o n , the L i e derivative L f g of g a l o n g f coinci-

des w i t h the L i e b r a c k e t [ f,g] . T h e r e is a l s o a t h i r d n o t a t i o n often
used, which expreses the L i e derivative of q a l o n g f as

Lfg = adfg

Both notations m a y be u s e d r e c u r r e n t l y , taking

0 k k-1
Lfg = g and Lfg = Lf(Lf g)

or

ad~g = g and a d fk g = adf (ad~-1 g)

Remark. T h e Lie d e r i v a t i v e of g a l o n g f m a y be interpreted as the

value at t = 0 of the d e r i v a t i v e with respect to t of a f u n c t i o n de-
f i n e d as

W(t) = (~ft)_ . g ( ~ ( p ) )

Moreover, it is e a s i l y seen t h a t for any k > 0

, d ~ (t), k k
- dt
- k-
~ t=0 = Lfg(p)
= a d f g(p)

If W(t) is a n a l y t i c in a n e i g h b o r h o o d of t = 0, then W(t) c a n be ex-

panded in the f o r m

tk
k=O
 286

known as Campell-Baker-Hausdorff formula. D

One may define an o b j e c t w h i c h d u a l i z e s the n o t i o n of a v e c t o r

field.

Definition. L e t N be a s m o o t h m a n i f o l d of d i m e n s i o n n. A covector field

(also c a l l e d one-form) m on N is a m a p p i n g a s s i g n i n g to e a c h p o i n t
%
p 6 N a t a n g e n t c o v e c t o r e ( p ) in T N. A c o v e c t o r f i e l d f is ~mooth if
P
for e a c h p E N there exists a coordinate chart (U,~) about p and n
real-valued smooth functions el,...,en defined on U, such that, for
all q E U

(q) = [ ~i (q) (d~i) []

i=1 q

The n o t i o n of s m o o t h covector field is c l e a r l y independent of the

coordinate used. The expression of a c o v e c t o r field in l o c a l coord-
inates is o f t e n given the f o r m of a r o w v e c t o r e = r o w ( ~ 1 , . . . , e n) in
!
which the ~i s are real-valued functions of x 1 , . . , x n.
If m is a c o v e c t o r field and f is a v e c t o r field, {~,f ) denotes
the smooth real-valued function defined by

{ ~ , f } (p) = ( ~ ( p ) , f ( p ) )

With any smooth function h:N ~ ~ one may associate a covector

field by taking at e a c h p the c o t a n g e n t vector (dl) . The covector
P
field thus defined is u s u a l l y still represented by the symbol dl.
However, the c o n v e r s e is not always true.

Definition. A c o v e c t o r field m is exact if there exists a smooth real

valued function k:N ~ such tha%

= dl []

The set of all smooth covector fields on a m a n i f o l d N is d e n o t e d

by the symbol V (N).
One may also define the n o t i o n of Lie derivative of a c o v e c t o r
field ~ along a vector field f. In o r d e r to do this, one has to in-
troduce first the n o t i o n of a covector field ~-related to a g i v e n
covector_ field ~. L e t p be a p o i n t of the domain of ~ .f R e c a l l that
(~)~ :T N ~ T f N is a l i n e a r mapping and let (~)~:T~f N ~ T~N
P ~t(p) ~t(P) P
denote the d u a l mapping. With e and %~ we associate a new covector
f
field, whose value at a p o i n t p in the domain of Ct is d e f i n e d by
 287

(~ft)*~(#tf(p))

The covector field thus defined is said to be ~ - r e l a t e d to ~.

Theorem. Let f be a smooth vector field and ~ a smooth covector field

on N. For each p E N the limit

1 f * {~(p))_m ]
lim ~[ ({t) ~ ( (p)
t~ 0

exists.

Definition. The Lie derivative of ~ along f is a c o v e c t o r field on N,

w r i t t e n Lfe, whose value at p is set equal to the value of the limit

lim 1[ (~f)* ( f []
t ~t (p)) -m (p) ]
t~0

The e x p r e s s i o n of Lf~ in local c o o r d i n a t e s is given by the (row)

n-vector

~)o~ 1 ~05n ~fl ~fl

(fl...fn) 7 (~l...~n) x~1 "'" ~Xn

+ T ~f
q-----~xf) +m

~f ~f
B~I ~n n n
~Xn "'" ~n ~x I ..

where the superscript "T" denotes "transpose".

Remark. The three types of Lie derivatives Lfl,Lfg,Lf~ defined above

are related by the following L e i b n i t z - t y p e relation

L f ( ~ , g )= ( L f ~ , g ) + (~,Lfg)

Remark. If ~ is an exact covector field, i.e. if e = dl for some I,

(dl,f > = Lfl

and
Lfdl = d(Lfl)

Remark. If e is a covector field, f a vector field, I and y r e a l - v a l u e d

functions, then
 288

Llfy~ = l.y.Lf~o + l.Lfy.~o + y {o~,f } dl

o~
Note that l,y,Lfy,<~,f > are elements of C (N) and ~,Lf~,dl elements
of V (N).
 B I B L I O G R A P H I C A L NOTES

Chapter I

The d e f i n i t i o n of d i s t r i b u t i o n u s e d here is taken from S u s s m a n n

(1973); in m o s t of the r e f e r e n c e s in D i f f e r e n t i a l G e o m e t r y q u o t e d in
the Appendix, the term "distribution" w i t h o u t any further s p e c i f i c a -
tion is used to denote w h a t we m e a n here for " n o n s i n g u l a r d i s t r i b u -
tion". D i f f e r e n t proofs of Frobenius' T h e o r e m are available. The one
used here is m u t u a t e d from L o b r y (1970) and S u s s m a n n (1973). T h e o r e m s
on s i m u l t a n e o u s i n t e g r a b i l i t y of d i s t r i b u t i o n s are due to J a k u b c z y k -
R e s p o n d e k (1980) and R e s p o n d e k (1982).
The i m p o r t a n c e in c o n t r o l theory of the n o t i o n of i n v a r i a n c e of
a d i s t r i b u t i o n under a v e c t o r field was p o i n t e d out i n d e p e n d e n t l y by
H i r s c h o r n (1981) and Isidori et al. (1981a). A m o r e general n o t i o n of
invariance, u n d e r a group of local d i f f e o m o r p h i s m s , was given by
S u s s m a n n (1973). The local d e c o m p o s i t i o n s d e s c r i b e d in section 5 are
c o n s e q u e n c e s of ideas of K r e n e r (1977).
T h e o r e m (6.15) and (6.20) were first p r o v e d by S u s s m a n n - J u r d j e v i c
(1972). The proof d e s c r i b e d here is due to K r e n e r (1974). An e a r l i e r
v e r s i o n of T h e o r e m (6.15), d e a l i n g w i t h "reachability" along trajec-
tories t r a v e r s e d in e i t h e r time direction, was given by Chow (1939).
C o n t r o l l a b i l i t y of systems e v o l v i n g on Lie groups was s t u d i e d by
B r o c k e t t (1972a). T h e o r e m (7.8), a l t h o u g h in a slightly d i f f e r e n t
version, is due to H e r m a n n - K r e n e r (1977).

C h a p t e r II

The proof of Theorems (1.4) and (I.7) may be found in S u s s m a n n

(1973). An i n d e p e n d e n t p r o o f of T h e o r e m (1.11) was given e a r l i e r by
H e r m a n n (1962) and an i n d e p e n d e n t p r o o f of C o r o l l a r y (1.13) by N a g a n o
(1966).
The r e l e v a n c e of the s o - c a l l e d "control Lie algebra" in the ana-
lysis of global r e a c h a b i l i t y derives from the w o r k of Chow (1939) and
was s u b s e q u e n t l y e l u c i d a t e d by L o b r y (1970), H a y n e s - H e r m e s (1970),
E l l i o t t (1971) and S u s s m a n n - J u r d j e v i c (1972), among the others. The
p r o p e r t i e s of the " o b s e r v a t i o n space" were studied by H e r m a n n - K r e n e r
(1977), and, in the case of d i s c r e t e - t i m e systems, by S o n t a g (1979).
Reachability, o b s e r v a b i l i t y and d e c o m p o s i t i o n s of b i l i n e a r systems
were s t u d i e d by B r o c k e t t (1972b), Goka et al. (1973) and d ' A l e s s a n d r o
et al. (1974).

C h a p t e r III

The f u n c t i o n a l e x p a n s i o n s i l l u s t r a t e d in the first section were

i n t r o d u c e d by Fliess since 1973. A c o m p r e h e n s i v e e x p o s i t i o n of the
subject, t o g e t h e r w i t h several a d d i t i o n a l results, may be found in
Fliess (1981). The e x p r e s s i o n s of the K e r n e l s of the V o l t e r r a series
e x p a n s i o n w e r e d i s c o v e r e d by L e s j a k - K r e n e r (1978); the e x p a n s i o n s
(2.12) are due to Fliess et al. (1983). The s t r u c t u r e of the V o l t e r r a
kernels was e a r l i e r a n a l y z e d by B r o c k e t t (1976), w h o p r o v e d that any
i n d i v i d u a l kernel can always be i n t e r p r e t e d as a kernel of a s u i t a b l e
b i l i n e a r system, and r e l a t e d results may also be found in G i l b e r t
(1977). The e x p r e s s i o n s of the kernels of a b i l i n e a r system were
first c a l c u l a t e d by Bruni et al. (1971). M u l t i v a r i a b l e L a p l a c e trans-
forms of V o l t e r r a k e r n e l s and their p r o p e r t i e s are e x t e n s i v e l y s t u d i e d
by Rugh (1981).
 290

F u n c t i o n a l e x p a n s i o n s for n o n l i n e a r d i s c r e t e - t i m e systems have

been s t u d i e d by S o n t a g (1979) and N o r m a n d Cyrot (1983).
The w a y the invariance analysis is dealt w i t h r e f l e c t s jointly
ideas d e v e l o p e d by Isidori et al. (1981a) and C l a u d e (1982); the former
contains, in particular, a d i f f e r e n t p r o o f of T h e o r e m (3.9). L e f t in-
v e r t i b i l i t y of n o n l i n e a r systems was m o s t l y s t u d i e d by H i r s c h o r n
(1979a) (1979b); our p r e s e n t a t i o n follows an idea of N i j m e i j e r (1982b).
D e f i n i t i o n s and p r o p e r t i e s of g e n e r a l i z e d Hankel m a t r i c e s were
d e v e l o p e d by Fliess (1974). T h e o r e m (5.8) was p r o v e d i n d e p e n d e n t l y
by Isidori (1973) and Fliess. The n o t i o n of Lie rank and T h e o r e m
(5.11) are due to Fliess (1983). E q u i v a l e n c e of m i n i m a l r e a l i z a t i o n s
was e x t e n s i v e l y s t u d i e d by S u s s m a n n (1977); the v e r s i o n given here of
the u n i q u e n e s s T h e o r e m e s s e n t i a l l y develops an idea of H e r m a n n - K r e n e r
(1977); r e l a t e d results may also be found in Fliess (1983).
An i n d e p e n d e n t a p p r o a c h to r e a l i z a t i o n theory was f o l l o w e d by
J a k u b c z y k (1980). R e a l i z a t i o n of finite V o l t e r r a series was studied
by Crouch (1981). C o n s t r u c t i v e r e a l i z a t i o n m e t h o d s from the L a p l a c e
t r a n s f o r m of a V o l t e r r a kernel may be found in the work of Rugh (1983).
R e a l i z a t i o n theory of d i s c r e t e - t i m e response maps was e x t e n s i v e l y
s t u d i e d by S o n t a g (1979).

C h a p t e r IV

C o n t r o l l e d i n v a r i a n t d i s t r i b u t i o n is the n o n l i n e a r v e r s i o n of
the n o t i o n of c o n t r o l l e d i n v a r i a n t subspace, i n t r o d u c e d i n d e p e n d e n t l y
by B a s i l e - M a r r o (1969) and W o n h a m - M o r s e (1970). For a c o m p r e h e n s i v e
p r e s e n t a t i o n of the theory of m u l t i v a r i b a l e linear control systems,
the reader is r e f e r r e d to the c l a s s i c a l treatise of W o n h a m (1979).
C o n t r o l l e d i n v a r i a n t d i s t r i b u t i o n s were i n t r o d u c e d i n d e p e n d e n t l y by
H i r s c h o r n (1981) and, in the more general form d e s c r i b e d here, by
Isidori et al. (1981a). The p r o o f of Lemma (1.10) may be found in
H i r s c h o r n (1981), Isidori et al. (1981b) and ~ i j m e i j e r (1981).
Lemma (3.6) is due to Claude (1982); the special case where the
m a t r i x A(x) has rank was dealt w i t h in Isidori et al. (1981a). The
a l g o r i t h m (3.17) has been s u g g e s t e d by K r e n e r (1985).
E a r l y results on n o n l i n e a r d e c o u p l i n g and n o n i n t e r a c t i n g con-
trol were given by S i n g h - R u g h (1972) and F r e u n d (1975). The p o s s i b i -
lity of solving d e c o u p l i n g p r o b l e m s in a d i f f e r e n t i a l - g e o m e t r i c
s e t t i n g was d e s c r i b e d by H i r s c h o r n (1981) and Isidori et al. (1981a).
The n o t i o n of c o n t r o l l a b i l i t y d i s t r i b u t i o n , the n o n l i n e a r version of
the one of c o n t r o l l a b i l i t y subspace, was i n t r o d u c e d by I s i d o r i - K r e n e r
(1982) and N i j m e i j e r (1982a). The solution of n o n i n t e r a c t i n g control
p r o b l e m s via c o n t r o l l a b i l i t y d i s t r i b u t i o n s is d e s c r i b e d in N i j m e i j e r
(1983).
C o n t r o l l e d i n v a r i a n c e for general n o n l i n e a r control systems (i.e.
systems where the control does not enter linearly) is s t u d i e d in
N i j m e i j e r - V a n der S c h a f t (1983). C o n t r o l l e d i n v a r i a n c e for d i s c r e t e -
time n o n l i n e a r s y s t e m is s t u d i e d in Grizzle (1985) and M o n a c o - N o r m a n d
Cyrot (1985).

Chapter V

The i n p u t - o u t p u t l i n e a r i z a t i o n p r o b l e m was treated by I s i d o r i -

Ruberti (1984). A s l i g h t l y d i f f e r e n t v e r s i o n of this p r o b l e m was
e a r l i e r s t u d i e d by C l a u d e - F l i e s s - I s i d o r i (1983) and, in the case of
d i s c r e t e - t i m e systems, by M o n a c o - N o r m a n d C y r o t (1983). The s o - c a l l e d
"structure algorithm" was i n t r o d u c e d by S i l v e r m a n (1969) and its
importance in c o n n e c t i o n with the c o m p u t a t i o n of the "zero structure
at infinity" was o u t l i n e d by Van D o o r e n et al. (1979). The p o s s i b i l i t y
 291

of c o m p u t i n g a "zero s t r u c t u r e at infinity" on the c o e f f i c i e n t s of the

formal p o w e r series a s s o c i a t e d with the e x t e r n a l b e h a v i o r of a non-
linear s y s t e m was p o i n t e d out by Isidori (1983); a g e o m e t r i c a p p r o a c h
to the d e f i n i t i o n of a "zero structure at infinity" is f o l l o w e d by
N i j m e i j e r - S c h u m a c h e r (1985). The p r o b l e m of m a t c h i n g a linear model
via dynamic state feedback was s t u d i e d by Isidori (1985) and Di Be-
n e d e t t o - I s i d o r i (1985). The proof of T h e o r e m (5.9) is the n o n l i n e a r
v e r s i o n of a proof of M a l a b r e (1984).
The s t a t e - s p a c e l i n e a r i z a t i o n p r o b l e m was p r o p o s e d and solved
for s i n g l e - i n p u t systems by B r o c k e t t (1978). C o m p l e t e s o l u t i o n for
m u l t i - i n p u t systems was found by J a k u b c z y k - R e s p o n d e k (1980). I n d e p e n d -
ent work of Su (1982) and H u n t et al. (1983a) lead to a slightly
w e a k e r formulation, t o g e t h e r w i t h a c o n s t r u c t i v e a l g o r i t h m for the
solution; The p o s s i b i l i t y of using n o n i n t e r a c t i n g control t e c h n i q u e s
for the solution of such a p r o b l e m was p o i n t e d out in Isidori et al.
(1981a); the c o n s t r u c t i o n s u g g e s t e d here e s s e n t i a l l y r e c a p t u r e s an
idea of M a r i n o (1982). A d d i t i o n a l results on this s u b j e c t may be
found in S o m m e r (1980), H u n t et al. (1983b), B o o t h b y et al. (1985),
and C h e n g et al. (1985).
The o b s e r v e r l i n e a r i z a t i o n p r o b l e m was s t u d i e d i n d e p e n d e n t l y by
B e s t l e - Z e i t z (1983) and K r e n e r - I s i d o r i (1983), for s i n g l e - o u t p u t
systems, and by K r e n e r - R e s p o n d e k (1985) for m u l t i - o u t p u t systems.

Appendix

For a c o m p r e h e n s i v e i n t r o d u c t i o n to the subjects dealt with in

this appendix, the reader is r e f e r r e d to B o o t h b y (1975), W a r n e r
(1970), S i n g e r - T h o r p e (1967).
REFERENCES

G. BASILE, G. M A R R O
(1969) C o n t r o l l e d and c o n d i t i o n e d i n v a r i a n t subspaces in linear systems
theory, J. Optimiz. Th. & Appl. 3, pp. 306-315.

D. BESTLE, M. ZEITZ
(1983) C a n o n i c a l form o b s e r v e r design for n o n - l i n e a r t i m e - v a r i a b l e
system, Int. J. Contr. 38, pp. 419-431.

W.A. BOOTHBY
(1975) "An I n t r o d u c t i o n to D i f f e r e n t i a b l e M a n i f o l d s and R i e m a n i a n n
Geometry", A c a d e m i c Press: N e w York.

F.R. BRICKELL, R.S. C L A R K

(1970) "Differentiable Manifolds", Van Nostrand: New York.

R.W. B R O C K E T T
(1972a) S y s t e m theory on group m a n i f o l d s and c o s e t spaces, S I A M J.
Contr. 10, pp. 265-284.
(1972b) On the a l g e b r a i c s t r u c t u r e of b i l i n e a r systems, in "Theory
and A p p l i c a t i o n s of V a r i a b l e S t r u c t u r e Systems", R. M o h l e r and A.
R u b e r t i Eds., A c a d e m i c Press: New York, pp. 153-168.
(1976) V o l t e r r a series and g e o m e t r i c control theory, A u t o m a t i c a 12,
pp. 167-176.
(1978) F e e d b a c k i n v a r i a n t s for n o n l i n e a r systems, Preprints 6th IFAC
Congress, Helsinki, pp. 1115-1120.

C. BRUNI, G. DI PILLO, G. KOCH

(1971) On the m a t h e m a t i c a l m o d e l s of b i l i n e a r systems, Ricerche di
A u t o m a t i c a 2, pp. 11-26.

P. B R U N O V S K Y
(1970) A c l a s s i f i c a t i o n of l i n e a r c o n t r o l l a b l e systems, Kybernetika
6, pp. 173-188.

D. CHENG, T.J. TARN, A. ISIDORI

(1985) Global e x t e r n a l l i n e a r i z a t i o n of n o n l i n e a r systems via feedback,
IEEE Trans. Aut. Contr. AC-30, to appear.

W.L. CHOW
(1939) Uber systeme von l i n e a r e n p a r t i e l l e n d i f f e r e n t i a l g l e i c h u n g e n
ester ordnung, Math. Ann. 117, pp. 98-105.

D. CLAUDE
(1982) D e c o u p l i n g of n o n l i n e a r systems, Syst. & Contr. Lett. I,
pp. 242-248.

D. CLAUDE, M. FLIESS, A. ISIDORI

(1983) Immersion, directe et par bouclage, d'un systeme non lineaire
dans un lineaire, C.R. Acad. Sci. Paris 296, pp. 237-240.
 293

P. C R O U C H
(1981) D y n a m i c a l r e a l i z a t i o n s of finite V o l t e r r a Series, S I A M J.
Contr. Optimiz. 19, pp. 177-202.

P. d'ALESSANDRO, A. ISIDORI, A. RUBERTI

(1974) R e a l i z a t i o n and s t r u c t u r e theory of b i l i n e a r d y n a m i c a l systems,
S I A M J. Contr. 12, pp. 517-535.

M.D. DI BENEDETTO, A. ISIDORI

(1985) The m a t c h i n g of n o n l i n e a r m o d e l s via dynamic state-feedback,
to be published.

D.L. E L L I O T T
(1970) A c o n s e q u e n c e of c o n t r o l l a b i l i t y , J. Diff. Eqs. 10, pp. 364-370.

M. FLIESS
(1974) M a t r i c e s de Hankel, J. Math. pures et appl 53, pp. 197-224.
(1981) F o n c t i o n n e l l e s c a u s a l e s non l i n e a i r e s et i n d e t e r m i n ~ e s non
c o m m u t a t i v e s , Bull. Soc. math. France 109, pp. 3-40.
(1983) R e a l i s a t i o n locale des s y s t e m e s non lineaires, a l g e b r e s de
Lie filtrees t r a n s i t i v e s et series g e n e r a t r i c e s non c o m m u t a t i v e s ,
Invent. math. 71, pp. 521-537.

M. FLIESS, M. LAMNABHI, F. L A M N A B H I - L A G A R R I G U E
(1983) An a l g e b r a i c a p p r o a c h to n o n l i n e a r f u n c t i o n a l expansions, IEEE
Trans. Circ. Syst. CAS-30, pp. 554-570.

E. F R E U N D
(1975) The structure of d e c o u p l e d n o n l i n e a r systems, Int. J. Contr.
21, pp. 651-659.

E.G. GILBERT
(1977) F u n c t i o n a l e x p a n s i o n for the r e s p o n s e of n o n l i n e a r d i f f e r e n t i a l
systems, IEEE Trans. Aut. Contr. AC-22, pp. 909-921.

T. GOKA, T.J. TARN, J. Z A B O R S Z K Y

(1973) On the c o n t r o l l a b i l i t y of a class of d i s c r e t e b i l i n e a r systems,
A u t o m a t i c a 9, pp. 615-622.

J.W. GRIZZLE
(1985) C o n t r o l l e d i n v a r i a n c e for d i s c r e t e time n o n l i n e a r systems w i t h
an a p p l i c a t i o n to the d i s t u r b a n c e d e c o u p l i n g problem, to be published.

G.W. HAYNES, H. H E R M E S
(1970) N o n l i n e a r c o n t r o l l a b i l i t y via Lie theory, S I A M J. Contr. 8,
pp. 450-460.

R. H E R M A N N
(1962) The d i f f e r e n t i a l g e o m e t r y of foliations, J. Math. and Mech. 11,
pp. 302-316.

R. HERMANN, A.J. K R E N E R
(1977) N o n l i n e a r c o n t r o l l a b i l i t y and o b s e r v a b i l i t y , IEEE Trans. Aut.
Contr. AC-22, pp. 728-740.
 294

R.M. H I R S C H O R N
(1979a) I n v e r t i b i l i t y of n o n l i n e a r c o n t r o l systems, S I A M J. Contr.
& Optimiz. 17, pp. 289-297.
(1979b) I n v e r t i b i l i t y of m u l t i v a r i a b l e n o n l i n e a r c o n t r o l systems,
IEEE Trans. Aut. Contr. AC-24, pp. 855-865.
(1981) ( A , B ) - i n v a r i a n t d i s t r i b u t i o n s and d i s t u r b a n c e d e c o u p l i n g of
n o n l i n e a r systems, S I A M J. Contr. & Optimiz. 19, pp. 1-19.

L.R. HUNT, R. SU, G. M E Y E R

(1983a) Design for m u l t i - i n p u t n o n l i n e a r systems, in " D i f f e r e n t i a l
g e o m e t r i c c o n t r o l theory", R.W. Brockett, R.S. M i l l m a n and H. S u s s m a n n
Eds., Birkhauser: Boston, pp. 268-298.
(1983b) Global t r a n s f o r m a t i o n s of n o n l i n e a r systems, IEEE Trans. Aut.
Contr. AC-28, pp. 24-31.

A. ISIDORI
(1973) D i r e c t c o n s t r u c t i o n of m i n i m a l b i l i n e a r r e a l i z a t i o n s from non-
linear i n p u t - o u t p u t maps, IEEE Trans. Aut. Contr. AC-18, pp. 626-631.
(1983) N o n l i n e a r feedback, structure at infinity and the i n p u t - o u t p u t
l i n e a r i z a t i o n problem, in " M a t h e m a t i c a l theorv of n e t w o r k s and
systems", P.A. F u h r m a n n Ed., S p r i n g e r Verlag: Berlin, DD. 473-493.
(1985) The m a t c h i n g of a p r e s c r i b e d linear i n p u t - o u t p u t b e h a v i o r in
a n o n l i n e a r system, IEEE Trans. Aut. Contr. AC-30, to appear.

A. ISIDORI, A.J. KRENER, C. GORI GIORGI, S. M O N A C O

(1981a) N o n l i n e a r d e c o u p l i n g via feedback: a d i f f e r e n t i a l geometric
approach, IEEE Trans. Aut. Contr. AC-26, pp. 331-345.
(1981b) L o c a l l y (f,g)-invariant d i s t r i b u t i o n s , Syst. & Contr. Lett. I,
pp. 12-15.

A. ISIDORI, A. RUBERTI
(1984) On the s y n t h e s i s of linear i n p u t - o u t p u t r e s p o n s e s for n o n l i n e a r
systems, Syst. & Contr. Lett. 4, pp. 17-22.

B. J A K U B C Z Y K
(1980) E x i s t e n c e and u n i q u e n e s s of r e a l i z a t i o n s of n o n l i n e a r systems,
S I A M J. Contr. & Optimiz. 18, pp. 455-471.

B. JAKUBCZYK, W. R E S P O N D E K
(1980) On l i n e a r i z a t i o n of control systems, Bull. Acad. Polonaise
Sci. Set. Sci. Math. 28, pp. 517-522.

R.E. K A L M A N
(1972) K r o n e c k e r i n v a r i a n t s and feedback, in " O r d i n a r y d i f f e r e n t i a l
equations", C. W e i s s Ed., A c a d e m i c Press: New York, pp. 459-471.

S.R. KOU, D.L. ELLIOTT, T.J. TARN

(1973] O b s e r v a b i l i t y of n o n l i n e a r systems, Inform. Contr. 22,
pp. 89-99.

A.J. KRENER
(1974) A g e n e r a l i z a t i o n of C h o w ' s T h e o r e m and the b a n g - b a n g T h e o r e m
to n o n l i n e a r control systems, S I A M J. Contr. 12, pp. 43-52.
 295

(1977) A d e c o m p o s i t i o n theory for d i f f e r e n t i a b l e systems, SIAM J.

Contr. & Optimiz. 15, pp. 2S9-297.
(1985) (Adf,g), (adf,g) and l o c a l l y (adf,g) i n v a r i a n t and c o n t r o l -
l a b i l i t y d i s t r i b u t i o n s , S I A M J. Contr. & Optimiz., to appear.

A.J. KRENER, A. ISIDORI

(~982) (Adf,G) i n v a r i a n t and c o n t r o l l a b i l i t y distributions, in
"Feedback Control of L i n e a r and N o n l i n e a r Systems", D. H i n r i c h s e n
and A. Isidori, Eds. S p r i n g e r Verlag: Berlin, pp. 157-164.
(1983) L i n e a r i z a t i o n by o u t p u t i n j e c t i o n and n o n l i n e a r observers,
Syst. & Contr. Lett. 3, pp. 47-52.

A.J. KRENER, W. RESPONDEK

(1985) N o n l i n e a r observers w i t h l i n e a r i z a b l e error dynamics, SIAM
J. Contr. & Optimiz., to appear.

C. LESJAK, A.J. K R E N E R
(1978) The e x i s t e n c e and u n i q u e n e s s of V o l t e r r a series for n o n l i n e a r
systems, IEEE Trans. Aut. Contr. AC-23, pp. 1090-1095.

C. L O B R Y
(1979) C o n t r S l a b i l i t ~ des syst~mes non lineaires, S I A M J. Contr. 8,
pp. 573-605.

M. M A L A B R E
(1984) S t r u c t u r e a l ' i n f i n i des triplets invariants: a p p l i c a t i o n a
la p u r s u i t e p a r f a i t e de modele, in "Analysis and O p t i m i z a t i o n of
Systems", A B e n s o u s s a n and J.L. Lions Eds., S p r i n g e r - V e r l a g : Berlin,
pp. 43-53.

R. M A R I N O
(1982) F e e d b a c k e q u i v a l e n c e of n o n l i n e a r systems with a p p l i c a t i o n s to
p o w e r systems equations, D.Sc. D i s s e r t a t i o n , W a s h i n g t o n University,
St. Louis.

R. MARINO, W.M. BOOTHBY, D.L. E L L I O T T

(1985) G e o m e t r i c p r o p e r t i e s of l i n e a r i z a b l e control systems, Math.
Syst. Theory, to appear.

S.H. MIKHAIL, M.H. W O N H A M

(1978) Local d e c o m p o s a b i l i t y and the d i s t u r b a n c e d e c o u p l i n g p r o b l e m
in n o n l i n e a r a u t o n o m o u s systems, Proc. 16th A l l e r t o n Conf.,
pp. 664-669.

S. MONACO, D. N O R M A N D - C Y R O T
(1983) The i m m e r s i o n under feedback of a m u l t i d i m e n s i o n a l d i s c r e t e -
time n o n l i n e a r s y s t e m into a linear system, Int. J. Contr. 38,
pp. 245-261.
(1985) I n v a r i a n t d i s t r i b u t i o n s for d i s c r e t e - t i m e n o n l i n e a r systems,
Syst. & Contr. Lett. 5, pp. 191-196.

T. N A G A N O
(1966) L i n e a r d i f f e r e n t i a l systems w i t h s i n g u l a r i t i e s and a p p l i c a t i o n s
to t r a n s i t i v e Lie algebras, J. Math. Soc. Japan 18, pp. 398-404.
 296

H. N I J M E I J E R
(1981) C o n t r o l l e d i n v a r i a n c e for affine control systems, Int. J. Contr.
34, pp. 824-833.
(1982a) C o n t r o l l a b i l i t y d i s t r i b u t i o n s for n o n l i n e a r systems. Syst.
& Contr. Lett. 2, pp. 122-129.
(1982b) I n v e r t i b i l i t y of affine n o n l i n e a r control systems: a geometric
approach, Syst. & Contr. Lett. 2, pp. 163-168.
(1983) F e e d b a c k d e c o m p o s i t i o n of n o n l i n e a r control systems, IEEE Trans.
Aut. Contr. AC-28, pp. 861-862.
(1985) Right i n v e r t i b i l i t y of n o n l i n e a r control systems: a geometric
approach, to be published.

H. NIJMEIJER, J.M. S C H U M A C H E R
(1985) Zeros at i n f i n i t y for affine n o n l i n e a r control systems, IEEE
Trans. Aut. Contr. AC-30, to appear.

H. NIJMEIJER, A.J. VAN DER S C H A F T

(1982) C o n t r o l l e d i n v a r i a n c e for n o n l i n e a r systems, IEEE Trans. Aut.
Contr. AC-27, pp. 904-914.

D. NORMAND-CYROT
(1983) Theorie et p r a t i q u e des syst~mes n o n l i n e a i r e s en temps discret,
Th6se d'Etat, U n i v e r s i t 6 de Paris Sud.

W. R E S P O N D E K
(1982) On d e c o m p o s i t i o n of n o n l i n e a r control systems, Syst. & Contr.
Lett. I, pp. 301-308.

W.J. RUGH
(1981) " N o n l i n e a r S y s t e m Theory: the V o l t e r r a - W i e n e r Approach", Johns
H o p k i n s Press: Baltimore.
(1983) A m e t h o d for c o n s t r u c t i n g m i n i m a l l i n e a r - a n a l y t i c r e a l i z a t i o n s
for p o l y n o m i a l systems, IEEE Trans. Aut. Contr. AC-28, pp. 1036-1043.

L.M. S I L V E R M A N
(1969) I n v e r s i o n of m u l t i v a r i a b l e linear systems, IEEE Trans. Aut.
Contr. AC-14, pp. 270-276.

L.M. SINGER, J.A. THORPE

(1967) "Lecture Notes on E l e m e n t a r y T o p o l o g y and Geometry", Scott,
Foresman: Glenview.

S.N. SINGH, W.J. RUGH

(1972) D e c o u p l i n g in a class of n o n l i n e a r systems by state v a r i a b l e
feedback, T r a n s A S M E J. Dyn. Syst. Meas. Contr. 94, pp. 323-329.

R. S O M M E R
(1980) Control design for m u l t i v a r i a b l e n o n l i n e a r t i m e - v a r y i n g systems,
Int. J. Contr. 31, pp. 883-891.

E. S O N T A G
(1979) "Polynomial R e s p o n s e Maps", S p r i n g e r Verlag: Berlin,
 297

R. SU
(1982) On the linear e q u i v a l e n t s of n o n l i n e a r systems, Syst. & Contr.
Lett. 2, pp. 48-52.

H. SuSSMANN
(1973) Orbits of families of v e c t o r fields and i n t e g r a b i l i t y of di-
stributions, Trans. A m e r i c a n Math. Soc. 180, pp. 171-188.
(1977) E x i s t e n c e and u n i q u e n e s s of m i n i m a l r e a l i z a t i o n s of n o n l i n e a r
systems, Math. Syst. Theory 10, pp. 263-284.

H. SUSSMANN, V. J U R D J E V I C
(1972) C o n t r o l l a b i l i t y of n o n l i n e a r systems, J. Diff. Eqs. 12,pp. 95-
116.

P.M. VAN DOOREN, P. DEWILDE, J. W A N D E W A L L E

(1979) On the d e t e r m i n a t i o n of the S m i t h - M a c M i l l a n form of a rational
m a t r i x from its L a u r e n t expansion, IEEE Trans. Circ. Syst. CT-26,
pp. 180-189.

F.W. WARNER
(1970) "Foundations of d i f f e r e n t i a b l e m a n i f o l d s and Lie groups",Scott,
Foresman: Glenview.

M.H. W O N H A M
(1979) "Linear M u l t i v a r i a b l e Control: a G e o m e t r i c Approach", Berlin:
S p r i n g e r Verlag.

M.H. WONHAM, AS. MORSE

(1970) D e c o u p l i n g and pole a s s i g n m e n t in linear m u l t i v a r i a b l e systems:
a g e o m e t r i c approach, S I A M J. Contr. 8, pp. 1-18.