Flatness of nonlinear control systems : a Cartan-Kahler Approach
Paulo Sergio Pereira da Silva
Escola Politecnica da USP - PTC
Sao Paulo - SP - 05508-900, Brazil
paulo@lac.usp.br
Keywords : Nonlinear systems, differential flatness, feed-
back linearization, exterior differential systems, Cartan-
Kahler theorem.
Abstract
In this work we consider the concept of differential flatness
defined by Fliess, Levine, Martin and Rouchon. The struc-
tural property of k-flatness (the case in which flat outputs de-
pends up to the kth derivative of the inputs) is studied based
on the properties of the dynamic extension algorithm and the
Cartan-Kahler theorems on the existence of integral mani-
folds of exterior differential systems. In this context, it is
shown that there may exist two kind of flat systems the
nonsingular and the singular cases. Since the nonsingular
case contains properly the class of state-feedback lineariz-
able systems, this induces a classification of nonlinear flat
systems as state-feedback linearizable, nonsingular and sin-
gular systems, in an ascending order of complexity.
Necessary and sufficient conditions of k-flatness for the
nonsingular case are developed. Necessary and sufficient
conditions for the singular case are also presented, although
they are much more involved and are not in closed form. Ex-
amples are presented to illustrate the results.
1 Introduction and motivation
Feedback linearization is an important structural problem in
control systems theory. This problem was completely solved
in the static-state feedback case [20, 18] but necessary and
sufficient conditions for feedback linearizability by dynamic
state feedback are not yet known (see [3, 29, 4, 16, 32, 34,
30, 1, 27, 33, 26, 17, 35] for several results on this subject).
Fliess et al introduced the notion of differential flatness
for nonlinear control systems in [9, 11]. This structural con-
cept is strongly related to the problem of feedback lineariza-
tion, and corresponds to a complete and finite parametriza-
tion of all solutions of a control system by a differentially
independent family of functions. Differential flatness was
originally defined in a differential algebraic setting [9, 11]
and restated in a infinite dimensional geometric setting re-
cently introduced in control theory [10, 25, 12, 14, 13].
In this paper we consider analytic nonlinear affine control
systems of the form
m
X
x(t) = f (x(t)) + gj (x(t))uj (t) (1)
j=0
where x(t) evolves on an analytic manifold X of dimension
n, and the input u(t) is a vector of dimension m. We will
assume that this dynamics is well formed (see [28]) i.e., the
fields {gj (x), j bme} are linearly independent for every
x X.
Roughly speaking, a system (1) is said to be flat if there
exists a set of differentially independent functions y =
(y1 , . . . , ym ) called flat output such that every variable of the
system is a function of time and of y and its derivatives1 (see
[14]). The flat output is a set of functions that may depend
on the state x and on the input u and its derivatives u(s) for
s = 0, . . . , k. In this case, the system is said to be k-flat.
When the flat output depends only on x, the system is said to
be 0-flat.
Instrumental for our purposes are the structure algorithms,
like Singhs algorithm [31] and the Dynamic Extension Algo-
rithm (DEA) [6, 21, 7, 5, 24]. Our main results combine the
geometric properties of the DEA with Cartan-Kahler theory
[2, Chap.3] to parametrize that the jets of the flat output can-
didates and to obtain necessary and sufficient conditions of
k-flatness can be obtained. The results developed here indi-
cate that two kind of flatness exists. For the first one, called
nonsingular, we give a complete and effective characteriza-
tion; For the second one, called singular, a characterization
is also given, but the computations in this case may be related
to more sophisticated results like Cartan-Kahler theorem and
Cartan-Kuranishi prolongation theorem [2, Chaps.3, 6 and
8]. In the nonsingular case, one can check flatness of a given
system using only standard geometric operations like Lie-
derivations, exterior differentiations etc. Our results about
the singular case are inspired by the ideas of Fliess et al [15]
which show that checking flatness of a system is closely re-
lated to the problem of finding integral manifolds of exterior
differential systems. In this paper we show how to compute
this exterior differential system through some symbolic op-
erations.
If a system is flat, by definition it is k-flat for k big enough.
An important question that remains open is if there exists a
1 The definition of flatness given here is not rigourous. The reader may
refer the cited literature for a complete presentation.
bound on k depending on the number of the states and inputs
[15].
The paper is organized as follows. In section 2 we present
notation and some results about exterior differential systems.
In section 3 we present the notion of structure codistribu-
tions, instrumental for parametrizing the jet of the flat-ouput
candidate. In sections 4 and 5 we present the main results
respectively about the singular and the nonsingular case. In
section 6 some examples are discussed. Finally in section 7
we state some conclusions.
2 Preliminaries and notation
2.1 Notation
The field of real numbers will be denoted by IR. The sub-
set of natural numbers {1, . . . , k} will be denoted by bke. A
symmetric multiindex K of class s and length k is a set of el-
ements of the form (i1 , . . . , ik ), where ij bse for j bke,
and where all the permutations of (i1 , . . . , ik ) are identified
with each other. The set of all symmetric multiindeces of
class s is denoted by (s). The length of K (s) is de-
noted by kKk. Note that K (s), kKk = k is identified to
some (i1 , . . . , ik ) such that i1 i2 . . . ik s. Given
K (s), K = (i1 , . . . , ik ) and i bse then (Ki) stands
for K = (i1 , . . . , ik , i).
We will use the standard notations of differential geometry
and exterior algebra [36, 2]. Let P be a smooth manifold of
dimension p. Let F be a set of smooth functions defined on
P. Let Z = {z P | f (z) = 0, f F } be the set of
common zeroes of all f F. Then z is an ordinary zero
if : (i) There exists a subset F = {f1 , . . . , fr } F such
that the set df = (df1 , . . . , dfr ) is independent on z ; (ii)
There exists an open neighborhood U of z such that the set
of common zeros of F that are inside U coincides with ZU .
In particular, Z U is a submanifold of P.
Given a field f and a 1-form on P, we denote (f ) by
hf, i. The set of smooth k-forms on P will be denoted by
k (P) and (P) = kIN k (P).
Given two forms and in (P), then denotes their Let (I, ) be an exterior differential system with indepen-
wedge multiplication. The exterior derivative of (P)
will be denoted by d. Note that the graded algebra (P),
as well as its homogeneous elements k (P) of degree k,
have a structure of C (P)-module (see [36] for details).
Given a family = (1 , . . . , k ) of a C (P)-module, then
span {1 , . . . , k } stands for the span over C (P).
An ideal I is a C (P)-submodule of (P) such that,
given two forms and in I then I. Given a subset
S (P) then {S} stands for the least ideal that contains
S.
A differential ideal I is an ideal that is closed by the exte-
rior differentiation, i.e., dI I. A differential ideal is also
called an exterior differential system.
2.2 Exterior Differential Systems with independence
condition
We present some definitions and results about exterior differ-
ential systems. The reader may refer to the treatise [2] for
details.
Let M be an analytic manifold with dimension m and let I
be a differential ideal defined on M . Let = 1 . . .n , /
I be an n-form on M . Then the pair (I, ) is called Exterior
Differential System with independence condition.
An integral element E of I on x M is a subspace of
Tx M such that |E = 0 for all forms of I. An integral ele-
ment of (I, ) on x M is a subspace E of dimension n of
Tx M such that |E = 0 for all form of I and |E 6= 0 (this
last condition is called independence condition). An integral
manifold of (I, ) is an immersed manifold i : N M
dimension n of Tx M such that, for every point N , its
tangent space E = i (T N ) is an integral element of (I, ).
For every immersed manifold i : N M (not necessarily
an integral manifold) we define the restriction of (I, ) to N
by (I, ) = (i I, i ). We denote by Gn (T M ) the Grass-
mann bundle of all n-subspaces E Tx M . The bundle of all
integral elements E of I of dimension n is denoted by Vn (I)
and is a subbundle of Gn (T M ). Similarly, Gn (T M, ) de-
notes the bundle of all n-subspaces E Tx M obeying the
independence condition |E = 6 0. The bundle of all integral
elements of (I, ) is denoted by Vn (I, ) and is a subbundle
of Gn (T M, ). Let E Gn (T M, ). Denote a basis of E
by e = (e1 , . . . , en ). It can be shown that
Vn (I) = {E Gn (T M ) | (e1 , . . . , en ) = 0,
I, of degree n}
If is an n-form on M , define a function on Vn (I, ) by
the formula |E = (E) |E . Let
F (I) = { | I, has degree n}. (2)
Then
Vn (I, ) is the subset of common zeroes of F (I) (3)
2.3 Prolongations
dence condition. The first prolongation is a Pfaffian sys-
tem (I (1) , (1) ) defined on a manifold M (1) . To define
the prolongation we define first a Pfaffian system (L, ) on
Gn (T M, ) in the following way. Let : Gn (T M, )
M be the canonical projection and for each (x, E)
Gn (T M, ) we define
I (1) (x, E) = E
J (1) (x, E) = Tx M
Then the filtration I (1) J (1) T Gn (T M, ) defines
a Pfaffian system with independence condition [2, p.104].
In particular, we may take L = {I (1) , dI (1) } and =
. To see what this gives in coordinates, assume that
(x1 , . . . , xn , y 1 , . . . , y s ) is a local coordinate system for M
such that = dx1 . . . dxn . Then, an n-plane E of
Gn (T M, ) is generated by e1 , . . . , en , where
X
ei = + pi
xi
y defined on N and let (J , ) be the exterior differential sys-
where (xi , y , pi ) is a local chart for Gn (T M, ). In other
words, E is defined by the equations (E) = 0, bse
where
X
= dy pi dxi (5)
i
It follows that I (1) is generated by the forms and we may
take = = dx1 . . . dxn .
Definition 2.1 Let M (1) = Vn (T M, ) and assume that
M (1) is a submanifold of Gn (T M ). Let : M (1)
Gn (T M ) be the canonical injection. Then the first prolon-
gation is the (linear) Pfaffian system ( L, ).
Note that one can always define the canonical map :
M (1) M as the composition of the canonical immer-
sion : M (1) Gn (T M ) with the canonical projection
: Gn (T M ) M .
Remark 2.1 Let X = span dx1 , . . . , dxn T M . As-
sume that the canonical map : M (1) M is a surjective
submersion. Let X = X, I(1) = I (1) and J(1) = J (1)
= T M .
Note that J(1) = I(1) X. Let : J(1) X be
the canonical projection. We can define a correspondence
T M 1 () X T M (1) by thePrule 1 () =
. In coordinates (x, y) for
s Let M = X N be a manifold and let = dx1 . . .dxn
P n i
PMn
, let =
Ps =1 dy + i be a volume form on X. Let (I, ) be an exterior differen-
i=1 i dx . Then 1 () = i=1 {i + =1 pi }dx .
Note that 1 () is the unique 1-form such that 1 ()
I(1) . Note also that X implies that 1 () X.
The next Lemma says that an integral manifold induces
canonically an integral manifold for the prolongations of any
order.
Lemma 2.1 Assume that the kth prolongation of a differen-
tial system with independence condition (I, ) is well de-
fined for all k IN . Assume that i : X M is an integral
manifold of (I, ). Then there is a unique integral manifold
i(k) : X M (k) associated to i : X M by a canonical
lifting.
Proof. Note that i : T X T M induces a map j :
X Vn (I, ) such that j(x) = i (Tx M ). Note also that
this map (canonical lifting) is an integral manifold for the
first prolongation (see [2, pp.147149]). The proof may be
completed by induction on k. 2
The next Lemma says that, under some natural assump-
tions, the prolongation of the pull-back is the pull-back of
the prolongation2 .
2 We abuse notation by using the same symbol for the independence
condition on all manifolds.
Lemma 2.2 Let Z, M, N be analytic manifolds and let be
a volume form on Z. Let M = Z M and N = Z N .
Let 1 : M Z, and 2 : N Z be the canonical
projections. Let i : M N be an immersion such that
(4) 2 i = 1 . Let (I, ) be an exterior differential system
tem defined on N by J = i I. Assume that both prolon-
gations (J (1) , ) and (I (1) , ) are well defined respectively
on M(1) and N (1) . Let j : M (1) N (1) be the map defined
by j(E ) = i E . Then J (1) = j I (1) .
Proof. The proof is an easy application of the definitions of
2.3. 2.
Remark 2.2 It is easy to show that the integral manifolds i
and i(1) of Lemma 2.1 and the map of Def. 2.1 obey the
relation i = i(1) .
2.4 Prolongation and Symbolical Calculus
In this section we will state some definitions that allow to per-
form symbolical calculations with functions that are the so-
lution of a partial differential equation3 . The idea is to define
symbolical operations in a way that, when specialized to an
integral manifold, these symbolical operations becomes the
usual manipulations of differential calculus. For instance, if
one denotes symbolically the partial derivatives y /xi of
a function y by pi , then one can perform all the differential
operations (of first order) with this function in a symbolical
fashion. Using the language of exterior differential systems
and their prolongations, we will consider this idea in a much
more abstract way.
tial system with independence condition defined on M . Let
(I (k) , ) be its prolongation defined on M (k) . Assume that
the canonical maps k : M (k+1) M (k) are surjective sub-
mersions4 for k IN and that M (k) = X N (k) . Note that
k (x, k ) = (x, k1 ) and hence the map (k ) () restricted
to T X T M (k) is an isomorphism.
Denote T M = T X T N . A symbolic field is a field
f : M T X T M such that, for every point on M ,
f is contained in Tx () X. Let be the canonical map of
Def. 2.1. Since ()|T X is an isomorphism onto T() X,
to every symbolic field we can associate a unique symbolic
field f (1) : M (1) T X T M (1) such that f (1) = f .
We may define the symbolic field f (k) on M (k) inductively
by the rule (k1 ) f (k) = f (k1) k1 .
Consider the map : J(1) X defined in Rem. 2.1.
One may define inductively the C (M )-linear map k :
J(k) T M (k) X (k) T M (k) . This map may
be extended in a natural way to a map from (M (k) ) to
{X (k) } (M (k) ). It follows that the composition
k = k k1 , k = 0, 1, 2, . . . (6)
3 Solutions of a partial differential equation are integral manifolds of ex-
terior differential systems, as will be shown in 2.6.
4 This assumption is needed for considering the map of Rem. 2.1.
1
defines a map from (M (k1) ) to {X (k) } (Mk ), k there exist some q dim n such that (q1) is
IN . involutive. In particular, (i(q1) ) (q1) is integrable as
a codistribution defined on X.
Definition 2.2 Given a function : M IR and a sym-
h fii = h1 d, f (1) i. Note that
bolical field we define hd,
h fii is a function defined on M (1) . If is a one
= hd, 2.5 Jet-spaces, contact-forms and prolongations
form on M we define h, h fii = h1 , f (1) i. Given a form We denote by J r (Z, Y ) the set of r-jets of all smooth maps
defined on M , denote = . Analogously, given a y : Z Y between smooth manifolds Z and Y . Then
codistribution on M we define = . Note that both J r (Z, Y ) has a structure of smooth manifold. For instance,
objects and are defined on M (1) . A symbolic
1-form is consider the manifold Jyr (Z, Y ), where Z has dimension t
a form defined on M such that span dx1 , . . . , dxn . and Y = IRs . If Z has local coordinates z = (z 1 , . . . , z t )
A codistribution
on M is a symbolic codistribution if then J r (Z, Y ) has local coordinates (z i , y , yK
: i bte,
span dx1 , . . . , dxn . For a symbolical codistribution bse K (t), kKk r), where y IR repre-
, define the flag (1) = { | d {I (1) + }}. We say sents the function evaluation y (z) and yK
represents, for a
that is symbolicaly involutive if (1) = . symmetricmultiindex K = (i1 . . . is ), the partial derivative
y
Define (0) = . As in the nonsymbolical case one z i1 ...z is z
. On the manifold J r (Z, Y ), we may define the
(1) contact forms
may compute the derived flag (k) = (k1) defined
on M (k)
for k IN . Let n = dim X and let n M (n) . Pn
dy j=1 yj dz j , bse
Denote by k1 = k (k ). We say that n is a regular point P n
dyk j=1 ykj
dz j , bse, k bte
of the derived flag {(k) , k {0, 1, . . . , n}, if k is a regular
...
point of (k) for k {0, 1, . . . , n}.
Pn
dyK j=1 yKj dz j , bse, K (t), kKk r 1
Definition 2.3 Let be a function and f be a sym- (7)
bolic field, both defined on M . Then LLf stands for Then the integral manifolds of the Pfaffian system I r gener-
ated by the contact-forms above with independence condition
h fi.
hd, i Define LLfk inductively by hk dLLfk1 , f (k) i.
= dz1 . . . dzt are the jets of functions y : Z IR,
Note that LLkf is a function defined on M (k) . Let bse (see [2, Theo. 3.2, p.26]). Furthermore it is easy to
be a symbolical 1-form on M . Define LLf = see that the prolongation of I r is I r+1 . It can be shown also

(f (1) )(1 d) + d(h , f (1) i) , where ()() denotes that every integral element of I r at (z, y , yK
: bse,
the interior product (see [36, 2.25 (d)]) and 1 is defined K (t), kKk r) is generated by the set {e1 , . . . , et }
by (6). If isa symbolical form defined on M (k1) define given by
LLf (k) = k (f (k) )(k d) + d(hk1 , f (k) i) . Then de- Xs X

fine inductively LLkf = LLf (k) (LLk1 ). ei = i + yJi (8)
f z =1
yJ
J(n)
Pn kJkr
Note that, in coordinates we have LLf i=1 i dxi =
Pn
Lf (i ) dxi + i dhdxi , f (1) i . where yJ1 i1 = yJ2 i2if the symmetric multiindeces (J1 i1 )
i=1 L
The next proposition shows that, when specialized to in- and (J2 i2 ) are identified by a convenient permutation.
tegral manifolds the symbolical calculations defined above Using Cartans test [2, Thm. 1.11, pp. 74] one may show
coincides to the standard differential calculus. that every integral element of I 1 is involutive and, by the
Cartan-Kahler Theorem[2, Cor. 2.3, p.86] it admits local in-
Proposition 2.1 Let i : X M be a local integral manifold tegral manifolds around every point. Using [2, Theo. 2.1,
of I and let i(k) : M (k) be the the local integral manifold p.248], that shows that the prolongation of an involutive sys-
of I (k) induced by i(k) (see Lemma 2.1). Note that there tem is also involutive, one may show that I r is involutive for
exists a unique field f defined on X such that hi dxi , fi = all r.
hdxi , f ii. Let : M IR be a function, let f be a
symbolic field, let be a symbolic 1-form, and let be a
2.6 Involutive codistributions and restricted Jet-spaces
symbolic distribution. Then
h fii i(1) = hd( i), fi
(i) hd, In the sequel we abuse notation and J r (Z, Y1 ) J r (Z, Y2 )
(ii) LLkf i(k) = Lkf) i stands for J r (Z, Y1 Y2 ). Denote by (z, i ) the points
of J ri (Z, Yi ), i = 1, 2. In a similar vein, J r1 (Z, Y1 )
(iii) (i(k) ) LLfk = Lfk(i ).
J r2 (Z, Y2 ) will stands for the set of points (z, 1 , 2 ) (with
(iv) Assume that the symbolic distribution is symbolically a common z).
involutive. Then i is involutive as a codistribution de- Given involutive codistribution =
a nonsingular
fined on X. span 1 , . . . , s defined on Z, with dim = s, then one
(v) Let n be a regular point of the derived flag obtained from may define a particular-class of contact forms on a manifold
. Assume that n is contained in the image of i(n) . Then R = Z IRs0 IRs with local coordinates (z i , y , j : i
bne, bs0 e, j bse) given by :
Ps
dy j=1 j j , bs0 e (9)
Denote the Pfaffian system generated by these forms by
I. Since is involutive, by the Frobenius theorem we
may assume without loss of generality that the local coor-
dinate system (z1 , . . . , zs , zs+1 , . . . , zt ) is such that =
span {dz1 , . . . , dzs }. Hence, in a new local coordinate sys-
tem (z i , y , j : i bte, bs0 e, j bse) for R, these
forms can be written as
Ps
= dy j=1 j dz j , bs0 e (10)
Consider the independence condition = dz1 . . . dzn .
One can show that every n-integral element of R and its pro-
longations of any order admit corresponding integral mani-
folds (they are all involutive with respect to the independence
condition = dz1 . . . dzn ). In particular, one can show
the following result
Proposition 2.2 Let R = Z IRs0 IRs0 s be a manifold
with local coordinates (z i , y , j : i bte, j bse,
bs0 e). Let I be the Pfaffian system generated by the forms
(9). There exist integral manifolds of (I (k) , ) at every point
of the prolongation R(k) for every k IN .
Note that the construction above may be considered as a
restricted jet space in the sense that the partial-derivatives
are subject to some relations. The following definition gen-
eralizes this situation.
Definition 2.4 Let Z be an analytic manifold and let be a
volume form on Z. Let N = J r1 (Z, Y1 ) . . . J rp (Z, Yp ),
where Yi = IR, i bpe. Let I be the Pfaffian system gen-
erated on N by the contact forms associated to all the jet
spaces J r1 (Z, Y1 ), . . . , J rp (Z, Yp ). Let M be an analyti-
cal manifold of the form Z M . A restricted jet space is an
analytic immersion i : M N such that i = 1 where
1 : M Z and : N Z are the canonical projections.
The pull-back J = i (I) defines the restricted exterior dif-
ferential system generated by the restricted contact forms.
Note that solving any partial differential equations with p-
unknown functions (y 1 , . . . , y p ) (depending on the variables
z1 , . . . , zt , with order r on y for bpe), is equivalent to
finding integral manifolds of a convenient restricted exterior
differential system.
The next proposition implies that the prolongation of a re-
stricted jet-space is a restricted jet-space of greater order.
Proposition 2.3 Consider the same notation stated in the
last definition. Assume that the prolongation J (1) is well
defined on M(1) . Let j : M(1) N (1) be the map defined
by j(E ) = i E . Then J (1) = j I (1) .
Proof. Note that the prolongation M (k) of M is always
well defined and is given by
J r1 +k (Z, Y1 ) . . . J rp +k (Z, Yp ).
Hence, the proof is a straightforward application of Lemma
2.2. 2
2.7 Extension of Restricted Jets-Spaces
Let M = Z N be a restricted jet space. Let I be the codis-
tribution generated on N by the restricted-contact forms and
let I = {I, dI}. Let = dz 1 . . . dz t be a volume form
on Z. Assume that the prolongations I (k) are well defined
on M (k) and are involutive for all k IN . We will define
an extension of this space by adjoining the restricted 1-jet of
new functions. For this, let span dz 1 , . . . , dz t be
a nonsingular symbolical codistribution on M generated by
a basis { 1 , . . . , s }. Assume that is symbolically involu-
tive, i.e., d mod {I+} 05 . Let Y = Y1 . . .Y and let
A = A1 . . .A , where Yi = IR, i bs0 e, and Aj = IRs ,
j bse. Consider the manifold Y A with global coordi-
nates (y b , jb : b be, j bse), and let M1 = M Y A.
Define the forms on M1 given by
s
X
0b = dhb jb j , b be (11)
j=1
Define the codistribution I1 = span 0b : b be + I and
let I1 = {I1 , dI1 }. Abusing notation, let stands for its
pull-back from Z to M1 . We have the following result :
Proposition 2.4 Consider the construction of M1 and I1
above. Assume that the canonical maps k : M (k)
M (k1) are surjective submersions. Then
(i) The exterior differential system with independence con-
dition (I1 , ) defined on M1 is is involutive.
(k)
(ii) The prolongations (I1 , ) are well defined on the re-
(k)
stricted jet-spaces M1 and are involutive for all k
IN .
(k+1) (k)
(iii) The canonical maps k : M1 M1 are surjective
submersions.
3 Structure codistributions
In this section we discuss some structural properties of a non-
linear control system of the form
m
X
x(t) = f (x(t)) + gj (x(t))uj (t) (12a)
j=0
y = h(x(t)) (12b)
3.1 Dynamic extension algorithm (DEA)
Let us recall the main aspects of the dynamic extension al-
gorithm (in the version of [7]). Given an analytic system
(12a)(12b), the dynamic extension algorithm is a sequence
of applications of regular static-state feedbacks and exten-
sions of the state by integrators.
It is well known (see [7]) that the dynamic extension al-
gorithm has an intrinsic interpretation based on the algebraic
structure at infinity {1 , . . . , n }. The integer = n is
5 Note that, by Prop. 2.1 part (iv), this implies that restricted to an
integral manifold of I is involutive.
called output rank [8, 7]. The following result is well known6
but is restated here in the differential geometric setting of
[14] :
Theorem 3.1 [7, 5, 24] Assume that system (12a) is ana-
lytic. Let S be the system associated to (12a) in the sense of
[14]. Consider the filtrations (defined on S)
Yk = span dx, dy, . . . , dy (k) , k IN
Yk = span{dy, . . . , dy (k) }, k IN
Let Sk S be the open and dense set of regular points of the
codistributions Yi and Yi for i = 0, . . . , k. In the kth step of
the dynamic extension algorithm, one may construct around
Sk1 a partition (and a reordering) of the output y =
(yk , ybk ) with card yk = k , a new local state representation
(1) (k)
(xk , uk ) of the system S with state xk = (x, y1 , . . . , yk )
(k)
and input uk = (y k , uk) such that
(i) span {dxk } = span dx, dy, . . . , dy (k) .
(ii) span {dxk , duk } = span dx, dy, . . . , dy (k+1) , du .
(k) (k+1)
(iii)y k yk+1
(iv) Let D(C) denote the generic dimension of a codistribu-
tion C generated by the differentials of a finite set of ana-
lytic functions. The sequence k = D(Yk ) D(Yk1 ) is
nondecreasing, the sequence k = D(Yk ) D(Yk1 ) is
nonincreasing, and both sequences converge to the same
integer , called the output rank, for some k n =
dim x. In particular, we have k n for k =
0, . . . , n and the the sequence k = D(Yk span {dx})
converges for k = k 1.
Proof. The reader may refer to [24] for a complete proof
which an adaptation of known results to the geometric setting
of [14]. 2
3.2 Structure codistributions
Consider the system (12a)(12b), defined in the sense of
[14]. Define the codistributions
k calculated from Yk =
span dx, dy, . . . , dy (k) in the following way :
Definition 3.1
0 = span
{dx}
k = span k1 | (k) Yk , k bne
The involutive closure k = k for k bne are system
invariants called structure codistributions of (1).
This codistributions may be computed based on the same
ideas presented in [1].
Proposition 3.1 Consider system (12a) with output y. As-
sume that the structure at infinity of this system is
{1 , . . . , n }. Denote by y = (yk , ybk ) the partition of the
output corresponding to the application k step of the DEA
6 The approaches of [7, 5] are algebraic. Some other properties of the
DEA considered in the geometric approach of [19] can be found in [22].
(see the notation of 3.1), where yk are, up to a convenient
reordering, the first k components of y. Then span {db yk }
k .
(1) (k)
Proof. By Theo. 3.1 it follows that {dx, dy1 , . . . , dyk }
is a basis of Yk and we may take ybk ybk1 . In particular,
(j)
dbyk Yk for j = 0, . . . , k. The desired result follows from
Definition 3.1. 2
The following proposition characterizes 0-flatness.
Proposition 3.2 Consider an analytic system S defined by
(1) in the sense of [14] and assume that the system is well
formed. i.e., span {du} span {dx, dx}. Consider the out-
put yi = hi (x), i bme. Let S be a regular point
of Yk and Yk for k = 0, 1, . . . , n. Then S is (locally) 0-flat
around with (local) flat output y if and only if there exist
k n such that one of the following equivalent conditions
are satisfied :
(i) The algebraic structure at infinity obeys the following
condition
kX 1
n+ i = mk (14a)
i=1
k = m (14b)
(ii) span {dx} Yk 1 .
4 The nonsingular case
In order to develop sufficient conditions of k-flatness we will
define a symbolic version of the dynamic extension algo-
rithm (SDEA). Choosing a nondecreasing integer solution
(1 , . . . , n ) of equation (14), this algorithm constructs a re-
stricted jet-space that parametrizes the flat-output candidates
y1 = h1 (x), . . . , ym = hm (x). The corresponding exte-
rior differential system (J , ) (generated by the restricted
contact forms, see Definition 2.4) is constructed in a way
that these restrictions correspond to the properties stated in
Proposition 3.1. In each step of the algorithm, the restricted
jet space is extended by the procedure of 2.7. So by Prop.
2.4, the exterior differential system constructed in step k is
involutive. If some further dimensional properties are ful-
filled, the existence of an integral manifold of (J , ) will
imply, by Proposition 3.2, the existence of a 0-flat output of
(13)
the system.
Remark 4.1 It is important to point out the meaning of the
word nonsingular employed here. This means that we will
consider that we are working in a neighborhood points where
our (symbolic) objects are nonsingular, i.e., we consider only
generic points7 .
4.1 SDEA
Consider the analytic system (1) defined on an open set
X with global coordinates x = (x1 , . . . , xn ). De-
note (f0 , g0 ) = (f, g). Let s0 = n, 0 =
7 This remark refers to step (S3) of SDEA, described in 4.3.

span 0i : i = 1, . . . , n , where 0i = dxi . Define I0 = chosen. Otherwise, up to a reordering of the input compo-
{0}. The SDEA can be summarized as follows. nents, assume that bk = b11 b12 where b11 is generi-
cally nonsingular (with rank k ). Define
(S0) Let 0 = 0. For every nondecreasing integer so-
lution (1 , . . . , k ) of equation (14), consider the mani- 1
fold Z = X IR1 . . . IRk , with local coordinates b11 b12 ak (x)
k = ; k = k
(x, v1 , . . . , vk ). Let M0 = Z. 0 I 0
Then, for k = 1 to k , execute:
and let
(S1) (a) Extend the restricted jet-space as described in 2.7 uk1 = k + k vk
by adding k k1 more functions to the the re-
stricted jet-space Mk1 constructed in the previous be an analytic regular static state feedback, where vk =
T T
step. Note that Mk1 is immersed in J r1 (Z, H1 ) vk vbkT Note that and are respectively matrices
. . . J rk (Z, Hk ), where Hj = IRj , j bke. De- (k)
m 1 and m m of functions defined on Tk . Add a dy-
fine a restricted jet space Tk = Mk1 Lk , where namic extension :
Lk = IRk , k = sk1 (k k1 ) and Lk has canon- uk = v k
ical coordinates (hi , ji : i {k1 + 1, . . . , k }, j u
bk = vbk
{1, . . . , sk1 }.
T
and let uk = ukT , u
bTk .
(S1) (b) Let Jk be a Pfaffian system defined on Tk generated
by Jk1 and the forms :
P j
4.3 Step (S3)
i = dhi jsk1 ji k1 , i {k1 + 1, . . . , k }
This step corresponds to the computation of the symbolical
where the symbolic distribution k1 = 1
span{k1 , structure codistribution k (see [1] for similar computations).
sk1
. . . , k1 } span {dx} was computed in the previous In iteration k 1 we have computed k1 =
1 sk1 d k1 j
step. span k1 , . . . , k1 and their derivatives dt k1
span {dx, dv1 , . . . , dvk1 }. In this step we have to find func-
(S2) Compute (symbolically) the kth step of the Dynamic tions j such that
Extension Algorithm. This will produce the (symbol-
ical) system (fk , gk ). X dk j
= j Zk = span {dx, dv1 , . . . , dvk } .
(S3) Compute the (symbolical) structure codistribution j
dt k1
(16)
k = span k1 , . . . , ksk span {dx} . dk j d k1 j
where dt k1 = LLfk +gk uk dt k1
. If j are so,
then k is spanned by all symbolic forms that obey condi-
The steps (S2) and (S3) will be described in detail in the se- tion (16).
quel8 . d k1 j
Let j = dt k1 (computed in the iteration k 1).
Compute = LLfk +gk uk j . From (16) one may find a
j

4.2 Step (S2) basis {k1 , . . . , rkk } of k by solving the linear equations.
This step corresponds to the symbolical computation of X
the kth step of the dynamical extension algorithm. In the h j , gklii = 0, l bme
jh
(k 1)th iteration we have computed a symbolical system j

(fk1 , gk1 ), with state xk1 = (xT , v1T , . . . , vk1

T
), input
uk1 and output y i.e., fk1 and gk1 are symbolical fields We stress the standard assumption that we are working
defined on Mk1 (see 2.4). Let hk = (h1 , . . . , hk )T . around (generic) nonsingular points, i.e., points where the
Compute rank of the matrices of analytical functions, codistributions,
etc. are maximal (see Rem 4.1).
(k)
yk = LLkfk1 hk + (LLgk1 LLfk1 h )uk1 Using the symbolic derived flag (see 2.4), we may
k1 k (15) obtain the greater involutive symbolic codistribution k =
= ak + bk uk1 dk j
span k1 , . . . , ksk contained
in k . Then compute dt k
(k) d k1 j
Note that ak and bk are matrices of functions defined on Tk . =L
Lfk +gk uk dt k (note that one may use the expres-
(k) dr j
Assume that the generical rank of bk on Tk is k . If this is sions of for r = 1, . . . , k 1 computed in the previous
dt r
not true, no solution is possible with the structure at infinity iterations).
(p ) (p )
8 The step (S2) includes a finite number of choices represented by a re- Denote by Mk = Tk k and Jk = Jk k , where pk is big
ordering of the inputs. enough for performing all these symbolic calculations.
4.4 A sufficient condition for k-flatness Then (I, ) is an exterior differential system with indepen-
dence condition defined on M .
Theorem 4.1 Let = (1 , . . . , n ) be a candidate for the
Now let {e1 , . . . , en } be defined by (8) replacing Z by
structure at infinity obeying condition (14). Assume that the
X, z j by xj , t by n, r by n and s by m. Define the
Symbolic Dynamic Extension Algorithm for system (12a), as- Pn Pm (0)
field f on M by f = i=1 (fi (x) + j=1 gi (x)uj )ei +
sociated to , constructs in the kth-iteration the symbolic Pm Pn1 (k+1)
structure codistributions k and the symbolic system (fk , gk ) j=1 k=0 uj uj (k) . Let k n be an integer. Con-
such that rank bk is generically equal to k . Then the system sider the following restriction :
is 0-flat
spann dxi : i bne mod o I
Proof. By construction, in the k-th iteration of the DEA k (17)
span dLf y : bme mod I
we have a manifold Mk1 and a differential system Jk1
and we contruct a manifold Tk , and a differential system
It can be shown that (17) is equivalent to {i = 0, i ble}
Jk as described in 2.7. Note also that (Jk , ) is a pk -
where i are analytic functions defined on M . So assume
prolongation of (Jk , ) for a convenient pk . By proposition
that this restriction defines an immersion : N M (in the
2.4 it follows that this exterior differential system is involu-
general case the analytical manifold can be decomposed in a
tive. In particular it admits local integral manifolds passing
strata). Then let I = I and = . Let be a 1-form
through every point. These integral manifolds induce func-
on M and let be the unique 1-form in T Z T M such
tions (h1 (x), . . . , hm (x)) defined on an open neighborhood
that mod I 0. We have the following result :
of X. By Prop. 3.1, Prop. 3.2, Prop. 2.1 and from the fact
the SDEA is a symbolic version of the DEA it follows that Theorem 5.1 The system (12a) is 0 flat around a point of
(h1 (x), . . . , hm (x)) is a local flat output of system (1). 2 Z if and only if :
(i) (I, ) admits a local integral manifold that passes
Definition 4.1 A 0-flat system is called nonsingular if the
through N such that i() =
conditions of Theo. 4.1 holds.
(ii) The codistribution
Proposition 4.1 A system (12a) that is static-feedback lin- n o
earizable is nonsingular in the sense of Def. 4.1. span dLrfy : r = 0, 1, . . . , k , bme

5 Singular Case is nonsingular in i().

For the proof of a version of Theo. 5.1 and the details

In this section we give necessary and sufficient conditions for
about the constructions above see [23].
0-flatness of a system (12a) in the general case. We construct
a restricted jet-space N and an exterior differential system Remark 5.1 Note that necessary and sufficient conditions of
(I, ), generated by restricted contact-forms, such that (lo- k-flatnes can be obtained by putting k integrators in series
cal) 0-flatness of system (12a) is equivalent to the existence with the inputs of the system (see [23]).
of (local) integral manifolds of (I, ).
Let X = IRn be the state space of system (12a). Let
U (k)
= IRm , k {0, . . . , n}. Consider the manifold 6 Examples
Z = X U (0) . . . U (n) with global coordinates
(k) We begin with a very simple academic example :
(xi , uj : i bne, j bme, k {0, . . . , n}). Con-
Pn Pm (0)
sider the field f = i=1 (fi (x) + j=1 gi (x)uj ) x i
+ Example 1 Consider the system with input u = (u1 , u2 )
Pm Pn1 (k+1) evolving on X = IR5 , given by
j=1 k=0 uj uj (k) . Let y = h(x) be an output for
system (12a). Then it is easy to show that y (k) = Lkfh(x), x1 = x4 x3
k {0, . . . , n} (see [22]). x2 = x4
Now let Y = IRm with global coordinates (y 1 , . . . , y m ). x3 = x5
Consider the jet-space J n+1 (X, Y ) and let I n+1 be the x4 = u1
codistribution generated by the corresponding contact-forms x5 = u2
(the contact forms are given by (7) replacing Z by X, z j by
xj , t by n, r by n and s by m). It is easy to verify that this system is not linearizable by
Now let M = U (0) . . . U (n) J n+1 (X, Y ) with static-state feedback because the conditions of [20, 18] are
(k)
global coordinates (xi , uj , y , yK

: i bne, j bme, not fulfilled.
k {0, . . . , n}, bme, K (n), kKk n + 1). De- A possible solution of condition (14) is k = 3, 1 = 0,
note by I the pull-back of I n+1 from J n+1 (X, Y ) to M . Let 2 = 1 and 3 = 2. We will execute the steps of SDEA for
I = {I, dI} and let this choice of structure at infinity.
S0. Let 0 = 0, Z = XIR2IR3 with global coordinates
(0) (n)
= dx1 . . .dxn du1 . . .du(0) (n)
m . . .du1 . . .dum . (x, v2 , v3 ). Let 0 = span dx1 , dx2 , dx3 , dx4 , dx5 and

Set f0 = f , g0 = g, M0 = Z and let I0 = {0}. where 1 span dx3 , dx2 and 2
Iteration 1. span dx1 , dx2 , dx3 , dx4 , dx5 .
(S1) Since 1 0 = 0, we have T1 = M0 and J1 = {0}.
(S2) Since 1 = 0, we have f1 = f0 , g1 = g0 , x1 = x, Iteration 3. (S1) We have 3 2 = 1. Hence T3 = M3 L3
u1 = u. where L2 = IR IR3 with coordinates (h2 , j2 : j b2e).
P3
(S3) Note that 1 = { 0 | Lf1 +g1u1 span{dx1 }}. Let I2 = {dh1 j=1 j1 j2 }.
Simple computations give 1 = span 11 , 12 , 13 , where (S2) Let h3 = (h1 . . . h3 ). After tedious computations it
i = dxi : i b3e Compute 1j = LLf1 +g1 u1 ij for i b3e, follows that
obtaining 11 = x3 du1 + u1 dx3 , 12 = dx4 , 13 = dx5 . Note (3)
that J1 = {0} implies that symbolical involutivity reduces y3 Lf21 h3 )u1
= LLf31 h3 + (LLg1 L
to the involutivity of 1 . In particular, 1 = 1 , M1 = M0 = a3 + b3 u1
and J1 = {0}. where

Iteration 2. 1 0
b3 = .
(S1) We have 2 1 = 1. Hence T2 = X L2 where (x2 ) 12 (x5 31 + x4 ) + 22
L2 = IR IR3 with coordinates (h1 , j1 : j b2e). Let
P3 It follows that b3 is generically nonsingular. By Theorem 4.1
I2 = {dh1 j=1 j1 dxj }. is (y 1 , y 2 )
1
1 is 0-flat.
it follows that the system
2 3
The
flat output
2
(S2) Let h2 = (h1 . . . h2 ) = h1 Note that where dy span dx , dx , dx and dy 2 is such
that b3 is nonsingular.
(2)
y2 = LL2f1 h2 + (LLg1 LLf1 h2 )u1
The next example shows that there exist flat systems that
= a2 + b2 u1
are singular with respect to Def. 4.1 :
where Example 2 Consider the system with input u = (u1 , u2 )
b2 = 11 x3 + 21 31 . evolving on X = IR5 , given by
Hence we may take x1 = x4
! x2 = x5
1 1
11 x3 +12
1 x33+1 a2 x3 = x4 x5
2 = 1 2 2 = 2
0 1 0 x4 = u1
x5 = u2

Denoting by 2 = 2 2 where 2 is the first column and Note that this example corresponds to the one of [15] with
2 is the second column of 2 , then (f2 , g2 ) is a system with 1 = 2 = 1 where one has extended the state by inte-
state x1 = (x, v2 ) and input u2 = (u2 , u2 ) given by : grators in series to the inputs. There exists two solutions of
equation (14) given by : {k = 4, 1 = 2 = 3 = 1,
x = f1 + g1 2 + g1 2 v2 + g2 2 u2 4 = 2} and {k = 3, 1 = 0, 2 = 1, 3 = 2}. After
v 2 = u2 some tedious computations one can show that the condition
of Theo. 4.1 is not satisfied by either of these structures at
(S3) To compute 2 , let infinity because rank bk < 2 in both cases. It is shown in
[15] that this system is 0-flat (the original system is 1-flat9 ).
2 = span 1 |LLf2 +g2 u2 span {dx, dv2 } Hence this example shows that there exist 0-flat systems that
P3 P3 are singular with respect to Def. 4.1.
Let = i=1 i 1i Then = i=1 {i 1i + i 1i }. It is
easy to show that LLf2 +g2 u2 span {dx, dv2 } if and only
P3 7 Conclusions
if h i=1 d1i , g i = 0. Hence
In this paper it is shown that there may exist two kind of 0-flat
1 1 systems the nonsingular and the singular cases. The non-
1 x 1 3 3 1 + 2 1 3 3 1 + 3 = 0.
3
1 x + 2 1 x + 2 singular case contains properly the class of state-feedback
linearizable systems. Hence 0-flat systems may be classified
It follows that 2 = span 12 , 22 , where as state-feedback linearizable, nonsingular and singular sys-
tems, in an ascending order of complexity.
21 = dx1 x3 dx2 , 22 = (11 x3 + 21 )dx2 + 31 dx3
Necessary and sufficient conditions of k-flatness for the
Further computations show that 2 is symbolically involu- nonsingular case are developed. For the singular case, nec-
tive and so 2 = 2 and that : essary and sufficient conditions of k-flatness relies on a con-
venient application of Cartan-Kahler and Cartan-Kuranishi
d2 1
dt 2 = 1 x5 dx4 x4 dx5 theorems.
d2 2 9 In
dt 2 = 2 + dv2 [23] it is shown that the example of [15] is not 0-flat.
Acknowledgments
This work was partially supported by Conselho Nacional
de Desenvolvimento Cientifico e Tecnologico (CNPq) and
Fundacao de Amparo a Pesquisa do Estado de Sao Paulo
(FAPESP), under grants 300492/95-2 and 97/04668-1.
References
[1] E. Aranda-Bricaire, C. H. Moog, and J. B. Pomet. A
linear algebraic framework for dynamic feedback lin-
earization. IEEE Trans. Automat. Control, 40:127132,
1995.
[2] R. Briant, S. Chern, R. Gardner, H. Goldschmidt, and
P. Griffiths. Exterior Differential Systems. Springer
Verlag, 1991.
[3] B. Charlet, J. Levine, and R. Marino. On dynamic feed-
back linearization. Systems Control Lett., 13:143151,
1989.
[4] B. Charlet, J. Levine, and R. Marino. Sufficient condi-
tions for dynamic state feedback linearization. SIAM J.
Control Optim., 29:3857, 1991.
[5] E. Delaleau and P. S. Pereira da Silva. Filtrations in
feedback synthesis: Part I systems and feedbacks.
Forum Math., 10(2):147174, 1998.
[6] J. Descusse and C. H. Moog. Decoupling with dy-
namic compensation for strong invertible affine non-
linear systems. Internat. J. Control, 42:13871398,
1985.
[7] M. D. Di Benedetto, J. W. Grizzle, and C. H. Moog.
Rank invariants of nonlinear systems. SIAM J. Control
Optim., 27:658672, 1989.
[8] M. Fliess. Automatique et corps differentiels. Forum
Math., 1:227238, 1989.
[9] M. Fliess, J. Levine, P. Martin, and P. Rouchon. Sur les
systemes non lineaires differentiellement plats. C. R.
Acad. Sci. Paris Ser. I Math., 315:619624, 1992.
[10] M. Fliess, J. Levine, P. Martin, and P. Rouchon.
Linearisation par bouclage dynamique et transforma-
tions de Lie-Backlund. C. R. Acad. Sci. Paris Ser. I
Math., 317:981986, 1993.
[11] M. Fliess, J. Levine, P. Martin, and P. Rouchon. Flat-
ness and defect of non-linear systems: introductory the-
ory and examples. Internat. J. Control, 61:13271361,
1995.
[12] M. Fliess, J. Levine, P. Martin, and P. Rouchon. Deux
applications de la geometrie locale des diffietes. Ann.
Inst. H. Poincare Phys. Theor., 66:275292, 1997.
[13] M. Fliess, J. Levine, P. Martin, and P. Rouchon. Non-
linear control and diffieties, with an application to
physics. In J. Krasilshchik M. Henneaux and A. Vino-
gradov, editors, Secondary Calculus and Cohomologi-
cal Physics, volume 219 of Contemporary Math., pages
8192, 1998.
[14] M. Fliess, J. Levine, P. Martin, and P. Rouchon. A
Lie-Backlund approach to equivalence and flatness of
nonlinear systems. IEEE Trans. Automat. Control,
44(5):922937, 1999.
[15] M. Fliess, J. Levine, P. Martin, and P. Rouchon. Some
open question related to flat nonlinear systems. In V.D.
Blondel, E. Sontag, M. Vidyasagar, and J.C. Willems,
editors, Open Problems in Mathematical Systems and
Control Theory, pages 99103, London, 1999. Springer
Verlag.
[16] R. B. Gardner and W. F. Shadwick. The GS algorithm
for exact linearization to Brunovsky normal form. IEEE
Trans. Automat. Control, 37:224230, 1992.
[17] M. Guay, P. J. McLellan, and D. W. Bacon. A condition
for dynamic feedback linearization of control-affine
nonlinear systems. Internat. J. Control, 68(1):87106,
1997.
[18] L. R. Hunt, R. Su, and G. Meyer. Design for multi-input
nonlinear systems. In R. Brocket, R. Millmann, and
H. J. Sussmann, editors, Differential Geometric Meth-
ods in Nonlinear Control Theory, pages 268298, 1983.
[19] A. Isidori. Nonlinear Control Systems. Springer
Verlag, 3nd edition, 1995.
[20] B. Jakubczyk and W. Respondek. On linearization of
control systems. Bull. Acad. Pol. Sc., Ser. Sci. Math.,
28:517522, 1980.
[21] H. Nijmeijer and W. Respondek. Dynamic input-output
decoupling of nonlinear control systems. IEEE Trans.
Automat. Control, 33:10651070, 1988.
[22] P. S. Pereira da Silva. On the nonlinear dynamic dis-
turbance decoupling problem. J. Math. Systems Estim.
Control, 6:126, 1996.
[23] P. S. Pereira da Silva. Flatness of nonlinear control sys-
tems and exterior differential systems. In Second Non-
linear Control Network (NCN) Workshop, Paris, June,
58, 2000. Submitted.
[24] P. S. Pereira da Silva. Geometric Properties of
the Dynamic Extension Algorithm. Internal Report
BT/PTC/00xx, Escola Politecnica da Universidade
de Sao Paulo http://www.lac.usp.br/paulo/down.html,
2000.
[25] J.-B. Pomet. A differential geometric setting for
dynamic equivalence and dynamic linearization. In
B. Jackubczyk, W. respondek, and T. Rzezuchowski,
editors, Geometry in Nonlinear Control and Differen-
tial Inclusions, pages 319339, Warsaw, 1995. Banach
Center Publications.
[26] J.-B. Pomet. On dynamic feedback linearization of
four-dimensional affine control systems with two in-
puts. ESAIM Control Optim. Calc. Var., 2:151230
(electronic), 1997.
[27] P. Rouchon. Necessary condition and genericity of dy-
namic feedback linearization. J. Math. Systems Estim.
Control, 5(3):345358, 1995.

[28] J. Rudolph. Well-formed dynamics under quasi-static

state feedback. In B. Jackubczyk, W. Respondek, and
T. Rzezuchowski, editors, Geometry in Nonlinear Con-
trol and Differential Inclusions, pages 349360, War-
saw, 1995. Banach Center Publications.

[29] W. F. Shadwick. Absolute equivalence and dynamic

feedback linearization. Systems Control Lett., 15:35
39, 1990.

[30] William F. Shadwick and Willem M. Sluis. Dynamic

feedback for classical geometries. In Differential geom-
etry and mathematical physics (Vancouver, BC, 1993),
volume 170 of Contemp. Math., pages 207213. Amer.
Math. Soc., Providence, RI, 1994.

[31] S. N. Singh. A modified algorithm for invertibility in

nonlinear systems. IEEE Trans. Automat. Control, AC
26:595598, 1981.

[32] W. M. Sluis. A necessary condition for dynamic feed-

back linearization. Systems Control Lett., 21:277283,
1993.
[33] W. M. Sluis and D. M. Tilbury. A bound on the num-
ber of integrators needed to linearize a control system.
Systems Control Lett., 29(1):4350, 1996.

[34] D. Tilbury, R. M. Murray, and S. R. Sastry. Trajectory

generation for the n-trailer problem using Goursat nor-
mal form. IEEE Trans. Automat. Control, 40:802819,
1995.

[35] M. van Nieuwstadt, M. Rathinam, and R. M. Mur-

ray. Differential flatness and absolute equivalence of
nonlinear control systems. SIAM J. Control Optim.,
36(4):12251239 (electronic), 1998.
[36] F. W. Warner. Foundations of differentiable manifolds
and Lie Groups. Scott, Foresman and Company, Glen-
view, Illinois, 1971.