ON THE OPTIMAL CONTROL OF IMPLICIT SYSTEMS
PHILIPPE PETIT
Abstract. In this paper we consider the following problem, known as implicit Lagrange problem: nd the trajectoryxargument of
min
Z
1
0
L(xx_)dt
where the constraints are dened by an implicit dierential equation F(xx_) = 0
with dimF=n;q <dimx=n. We dene the geometric framework of aq--submanifold in the tangent bundle of a surrounding manifold X, which is an extension of the-submanifold geometric framework dened by Rabier and Rheinboldt for control systems. With this geometric framework, we dene a class of well-posed implicit dierential equations for which we obtain locally a controlled vector eld on a submanifold W of the surrounding manifold X by means of a reduction procedure.
We then show that the implicit Lagrange problem leads locally to an explicit optimal control problem on the submanifold W, for which the Pontryagin maximum principle is naturally apply.
1. Introduction
We consider for the state
x
ofRnthe implicit dierential equationF
(x x
_) = 0:
(1.1)In this equation the control
u
does not appear explicitly, but only because there are less equations than unknowns, namelyF
: Rn Rn ! Rn;q, whereq < n
(see 6].) Here, the control variableu
belongs to Rq. The cost function is the LagrangianL
(x x
_) ofT
Rn. A process is a trajectoryx
() belonging toC
1(01]Rn) the set of continously dierentiable functions (resp.KC
1(01]Rn) the set of continuous and piecewise dierentiable functions,AC
(01]Rn) the set of absolutely continuous functions, see the footnotes of the subsection 5.2.) A trajectoryx
() is admissible ifx
(0) =a
,x
(1) =b
andF
(x
(t
)x
_(t
)) = 08t
201] (resp:
a:
e:
on 01]):
For any admissible trajectoryx
() the cost isJ
(x
()) =Z
1
0
L
(x
(t
)x
_(t
))dt:
Institut de mathematiques, UMR C9994 du CNRS, Universite Paris 7, UFR de math- ematiques, CASE 7012, 2, place Jussieu F-75251 Paris Cedex 05, France. E-mail:
petit@labomath.univ-orleans.fr.
Received by the journal April 29, 1997. Revised September 1997. Accepted for publi- cation January 5, 1998.
c
Societe de Mathematiques Appliquees et Industrielles. Typeset by LATEX.
An admissible trajectory
x
() belonging toC
1(01]Rn) (resp.KC
1(01]Rn),AC
(01]Rn)) is a weak minimum (resp. strong minimum) ofJ
ifJ
(x
())J
(x
())for any admissible trajectory
x
() belonging toC
1(01]Rn) (resp.KC
1(01]Rn),AC
(01]Rn)). An admissible trajectoryx
() belonging toC
1(01]Rn) (resp.KC
1(01]Rn),AC
(01]Rn)) is a weak local mini- mum (resp. strong local minimum) ofJ
if there exist an" >
0 such that for any trajectoryx
() belonging toC
1(01]Rn) (resp.KC
1(01]Rn),AC
(01]Rn)) such that jjx
();x
() jj1< "
(resp. jjx
();x
() jj0< "
) wherejj
x
()jj1= maxt201]maxfj
x
(t
)jjx
_(t
)jg (resp:
jjx
()jj0= maxt201]j
x
(t
)j) thenJ
(x
())J
(x
()):
Remark 1.1. a) An admissible trajectory
x
() belonging toC
1(01]Rn) which is a strong (local) minimum is also a weak (local) minimum, mean- while a trajectoryx
() belonging toC
1(01]Rn) can be a weak (local) minimum without to be a strong (local) minimum.b) The necessary conditions for the weak local minimum are also necessary conditions for the strong local minimum, and the suciency conditions for the strong local minimum are also the suciency conditions for the weak local minimum.
We will subsequently turn our attention to the geometry of the implicit dierential equation (1.1). More precisely, we will extend the denitions of
-submanifold, reducible and completely reducible -submanifold in 16]to our situation (see also 15].) Let us consider the manifold
X
= Rn and its tangent bundleTX
=T
Rn=Rn Rn. Let us assume that the subsetM
=F
;1(0) is a submanifold ofTX
(it is the case whenF
is a submersion).A trajectory
x
() is admissible if (x
(t
)x
_(t
)) belongs toM
for anyt
201].The implicit Lagrange problem1 is
P
0 min
(x()x_())2M x(0)=a x(1)=b
Z
1
0
L
(x
(t
)x
_(t
))dt:
If
x
() is an admissible trajectory thenx
(t
) has to belong to the setW
= (M
) for anyt
201]. Let us assume thatW
is a submanifold ofX
, thenx
_(t
) has to belong to the subspaceT
x(t)W
ofT
x(t)X
for anyt
in 01] and thus (x
(t
)x
_(t
)) belongs to the setM
1 =TW
\M
for anyt
in 01]. In other1When q=n, the constraint F(xx_) = 0 being absent, this is the simple problem of the calculus of variations (see 3, 4, 5, 10, 11])
min
x(0)=a
x(1)=b Z
1
0
L(x(t)x_(t))dt:
ESAIM: Cocv, March 1998, Vol. 3, 49{81
words, any admissible trajectory for the implicit Lagrange problem (P0) is an admissible trajectory for the following implicit Lagrange problem
P
1 min
(x()x_())2M1 x(0)=a x(1)=b
Z
1
0
L
(x
(t
)x
_(t
))dt:
Thus, if
x
() is a solution of P0 then it is a solution of P1 and conversely.Moreover, the startpoint
a
and the endpointb
have to belong toW
. This replacement of the submanifoldM
ofTX
by the submanifoldM
1 ofTX
is the reduction procedure andM
1 is called the reduction ofM
. Let us assume that we are able to do withM
1, what we have done withM
, then we construct a submanifoldW
1 =(M
1) ofX
and a submanifoldM
2ofTX
. Ifx
() is an admissible trajectory for the problem P1 then it is an admissible trajectory for the problemP
2 min
(x()x_())2M2 x(0)=a x(1)=b
Z
1
0
L
(x
(t
)x
_(t
))dt:
Let us assume that we construct by induction a sequence of implicit Lagrange problem
Pk min
(x()x_())2Mk x(0)=a x(1)=b
Z
1
0
L
(x
(t
)x
_(t
))dt
such that
W
k = (M
k) is a submanifold ofX
andM
k+1 =M
k \TW
kis a submanifold of
TX
, then any admissible trajectoryx
() of Pk is an admissible trajectory ofPk+1. Therefore any admissible trajectory ofP0 is an admissible trajectory of Pk for anyk
. Thus any admissible trajectory ofP
0 is an admissible trajectory of the following implicit Lagrange problem
Pc min
(x()x_())2C(M) x(0)=a x(1)=b
Z
1
0
L
(x
(t
)x
_(t
))dt
with
C
(M
) =\k0M
k. Clearly, the strong (resp. weak) minimum ofP0 are the strong (resp. weak) minimum ofPc and the pointsa
etb
have to belong to(C
(M
)). Furthermore, if the sequencefM
kgk0is stationary thenC
(M
) is a submanifold ofTX
and the smallest integersuch thatM
k =M
8i
will be called the index. Then we can wonder if Pc is equivalent to an explicit control problem (see the subsection 5.2 of the Appendix). This will be the case for the class of well-posed implicit dierential equations.2. Definitions and Main results
Using the geometric framework of
q
--submanifold of the section 3, we are able to dene a well-posed implicit dierential equation.Definition 2.1. An implicit dierential equation (1.1) is well-posed if the
set
M
=F
;1(0)is a completely reducible
q
--submanifold ofT
Rnsuch that the coreC
(M
) is not empty.Definition 2.2. The index of a well-posed implicit dierential equation (1.1) is the maximum over the index of the non-empty connected component of the core
C
(M
).Now, we consider only well-posed implicit dierential equations. As in the introduction, for each
q
--submanifoldM
k ofT
Rnof the chain of reduction we consider the implicit Lagrange problemPk min
(x()x_())2Mk x(0)=a x(1)=b
Z
1
0
L
(x
(t
)x
_(t
))dt
and for
C
(M
) the core ofM
we consider the implicit Lagrange problemPc min
(x()x_())2C(M) x(0)=a x(1)=b
Z
1
0
L
(x
(t
)x
_(t
))dt:
The sequencefPkgk0 is called the chain of reduced implicit Lagrange prob- lems of the well-posed implicit dierential equation (1.1) and Pc is called the central implicit Lagrange problem. Evidently, the points
a
andb
have to belong to (M
k) for anyk
and thus to belong to(C
(M
)).Definition 2.3. Any point
x
ofRnis consistent with a well-posed implicit dierential equationF
if it belongs to the projection of the coreC
(M
).According to the denitions we can formulate the following theorems (proofs are given in the subsection 5.1 of the Appendix).
Theorem 2.4. Let
F
be a well-posed implicit dierential equation (q < n
) ofT
Rn,fP
kgk0 its chain of reduced implicit Lagrange problems and Pc its central implicit Lagange's problem. Then, any admissible trajectoryx
() ofP
0 is an admissible trajectory ofPk for any
k
. In particular, any admissible trajectoryx
() of P0 is an admissible trajectory ofPc and any strong (resp.weak) minimum
x
() ofPc is a strong (resp. weak) minimum of P0. Theorem 2.4 shows that the strong (resp. weak) minimum are living in the coreC
(M
). According to the theorem 3.31, theq
--submanifoldC
(M
) is locally the image of a controlled vector eld. Thus, we are able to show that locally the trajectories ofC
(M
) are in bijection with the trajectories of the controlled vector eld .Theorem 2.5. (Local equivalence) Let
F
be a well-posed implicit dier- ential equation,C
(M
) its core, (x
0p
0) a point belonging toC
(M
) andW
= (V
) a local projection ofC
(M
) at (x
0p
0),O
an open set of Rq, a controlled vector eld given by the theorem 3.31. Ifx
() is a local tra- jectory ofC
(M
) such that (x
(t
)x
_(t
)) belongs toV
for anyt
, then thereESAIM: Cocv, March 1998, Vol. 3, 49{81
exists an unique continuous (resp. piecewise continuous) control
u
() taking its value inO
such that (x
(t
)x
_(t
)) = (x
(t
)u
(t
))2 for anyt
. Conversely, for any initial conditionx
0 belonging toW
and for any continuous (resp.piecewise continuous) control
u
() taking its value inO
there exists a unique local trajectory ofC
(M
) such that (x
(t
)x
_()) belongs toV
for anyt
.Then, on the one hand, we have shown that the strong (resp. weak) minimum of the implicit Lagrange problem P0 are the strong (resp. weak) minimum of the central implicit Lagrange problem Pc (theorem 2.4), and on the other hand that (locally) the admissible trajectories ofPc are in bijection with the admissible trajectories of the controlled vector eld (theorem 2.5).
Now, let us consider a strong minimum
x
() of the central implicit Lagrange problem Pc. Let be a point in the setT
andW
=(V
) a local projection ofC
(M
) at (x
()x
_()),O
an open set of Rq, a controlled vector eld given by the theorem 3.31. There exists" >
0 such that for any pointt
belonging to the intervalI
" = ;"
+"
] then (x
(t
)x
_(t
)) belongs toV
. We naturally consider the following local implicit Lagrange problemPc" min
(x()x_())2V x(;")=x(;") x(+")=x(+")
Z +"
;"
L
(x
(t
)x
_(t
))dt:
Theorem 2.6. (Local optimality) Let
F
be a well-posed implicit dieren- tial equation. Ifx
() is a strong minimum of the central implicit Lagrange problem Pc then for any belonging toT
there exists" >
0 such that the trajectory _x
jI"() is a strong minimum of the implicit Lagrange problemPc"Let us also consider the following local explicit optimal control problem
Pe" min
(x()x_())=(x()u()) u()2O x(;")=x(;") x(+")=x(+")
Z +"
;"
L
( (x
(t
)u
(t
))dt:
Theorem 2.7.
x
() is strong (local) minimum of the implicit Lagrange prob- lem Pc"if, and only if, the corresponding admissible process(x
()u
()) is a strong (local) minimum for the explicit control problem Pe".This leads to consider the following local implicit Lagrange problem. Let
W
=(V
) be a local projection such that there exists an open setO
in Rq and a controlled vector eld :W O
!TW
given by the theorem 3.31PV min
(x()x_())2V x(0)=a x(1)=b
Z
1
0
L
(x
(t
)x
_(t
))dt
where the points
a
andb
belong toW
. The admissible trajectories (resp.strong minimum) ofPV are in bijection with the admissible processes (resp.
2In a local coordinate systemxofW,(xu) takes the form (xf(xu)).
strong minimum) of the explicit optimal control problem
Pe min
(x()x_())=(x()u()) u()2O x(0)=a x(1)=b
Z
1
0
L
( (x
(t
)u
(t
))dt:
Finally, we choose a local coordinate system
x
= (x
1x
r) ofW
and then apply the Maximum Principle to the problem Pe with the pseudo- HamiltonianH
0 :T
?W O
!R (xu
)7!Xri=1
if
i(xu
) +0L
(xf
(xu
)) and the controlled vector eld;!
H
0 :T
?W O
!T
(T
?W
) (xu
)7!;!H
0(xu
) =Xri=1
@H
0@
i (xu
)@
@x
i;r
X
i=1
@H
0@x
i (xu
)@
@
iwhere
0 = 01.Remark 2.8. Obviously, the necessary conditions of optimality are invari- ant by bundle isomorphism
h
. Let us consider a bundle isomorphismh
x
~= ~X
(x
)u
~= ~U
(xu
) with inversex
=X
(~x
)u
=U
(~x u
~):
In the new coordinates (~
x u
~) the controlled vector eld isf
~(~x u
~) =@ X
~@x
(X
(~x
))f
(X
(~x
)U
(~x u
~))the Lagrangian ~
L
(~x u
~) =L
(X
(~x
)U
(~x u
~)) and the pseudo-HamiltonianH
~(~x
~ ~u
) =Pri=1 ~if
~i(~x u
~);~0L
~(~x
~u
)):
The extremals (~
x
()~()) are the projection of a triplet (~
x
()~()~u
()) such that(~a
): (~x
()~()~
u
()) is a trajectory of the controlled vector eld ( ~0 =0
1);!
H
~~0 :T
?W
~O
~ !T
(T
?W
~) (~x u
~ )7!;!H
~~0(~x
~ ~u
) =Xri=1
@ H
~~0@
~i (~x
~u
~)@
@ x
~i ;n
X
i=1
@ H
~~0@ x
~i (~x
~u
~)@
@
~i(~
b
): for anyt
belonging to 01] (resp. a.e. on 01])H
~~0(~x
(t
)~u
(t
)~(
t
)) = maxu
~2O
~H
~~0(~x
(t
)u
~(t
)):
ESAIM: Cocv, March 1998, Vol. 3, 49{81
Clearly, the extremals (~
x
()~()) are in bijection with the extremals (x
()()) via the relationship
(~
x
()~()) = ( ~
X
(x
()t@X
@
~x
( ~X
(x
()))()):
For any triplet (
x
()()
u
()) such that (a
) and (b
) are satised, the triplet (~x
()~()~
u
()) = ( ~X
(x
())t@X
@
~x
( ~X
(x
()))()U
~(x
()u
())) satises (~a
) and (~b
).Example 2.9. The controlled rigid pendulum. A mass
m
is attached at the extremity of a rigid massless wire of lengthl
and xed at the origin. is the tension of the wire,g
the gravity constant and the controlu
= (u
1u
2) acts on the mass. The equations of the system arem x
1 = ;lx
1+u
1m x
2 = ;lx
2+mg
+u
20 =
x
21+x
22;l
2:
In order to return to an implicit dierential equation and to use the reduction procedure we consider the following mapping
F
0 :T
R7!R5F
0(xp
) =0
B
B
B
B
@
p
1;x
3p
2;x
4p
3+x
1p
7;p
5p
4+x
2p
7;p
6;g
x
21+x
22;l
21
C
C
C
C
A
p
= _x
where _
x
5 =u
1=m
, _x
6 =u
2=m
and _x
7==ml
and the submanifold, ofT
R7,M
0 =F
0;1(0) .M
0 has dimension 9. The equation of the setW
0 =(M
0)is
x
21+x
22;l
2= 0it is a submanifold of R7 of dimension 6.
TW
0 =G
;10 (0) whereG
0 is the mappingG
0(xp
) =
x
21+x
22;l
2p
1x
1+p
2x
2
:
Thus the reduction of
M
0 isM
1 =TW
0 \M
0 =F
1;1(0), whereF
1 is the mappingF
1(xp
) =0
B
B
B
B
B
B
@
p
1;x
3p
2;x
4p
3+x
1p
7;p
5p
4+x
2p
7;p
6;g
x
21+x
22;l
2x
1x
3+x
2x
41
C
C
C
C
C
C
A
it is a submanifold of dimension 8. The equations of the set
W
1 = (M
1)are
x
21+x
22;l
2 = 0x
1x
3+x
2x
4 = 0and it is a submanifold of R7 of dimension 5.
TW
1 =G
;11 (0), whereG
1 is the mappingG
1(xp
) =0
B
B
@
x
21+x
22;l
2x
1x
3+x
2x
4p
1x
1+p
2x
2p
1x
3+x
1p
3+p
2x
4+x
2p
41
C
C
A
:
ThusM
2 =TW
1\M
1 =F
2;1(0), whereF
2 is the mappingF
2(xp
) =0
B
B
B
B
B
B
B
B
@
p
1;x
3p
2;x
4p
3+x
1p
7;p
5p
4+x
2p
7;p
6;g
x
21+x
22;l
2x
1x
3+x
2x
4x
23+x
24+x
1p
5+x
2p
6+x
2g
;l
2p
71
C
C
C
C
C
C
C
C
A
is a submanifold of dimension 7. Finally the set
W
2 =(M
2) is in factW
1 and thusM
3 =TW
2\M
2 =TW
1 \M
2 =M
2 (sinceM
2TW
1), thenC
(M
0) is the submanifoldM
2 andW
=(C
(M
0)) is the submanifoldW
2. MoreoverC
(M
) = (W
R2) where (xv
) = (xf
(xv
)) is the vector eld of the statex
2W
depending onv
= (v
1v
2)2R2such thatf
(xv
) =0
B
B
B
B
B
B
B
B
B
B
@
x
3x
4;
x
1l
2(x
23+x
24) +x
2l
2(x
2v
1;x
1(v
2+g
));
x
2l
2(x
23+x
24);x
1l
2(x
2v
1;x
1(v
2+g
))v
1v
2l
12(x
23+x
24+x
1v
1+x
2(v
2+g
))1
C
C
C
C
C
C
C
C
C
C
A
and
v
1 =u
1=m
etv
2 =u
2=m
. On the other hand, from the relationx
1 =l
sin andx
2 =l
cos (2];) we obtain
x
3 =l
_cosx
4 = ;l
_sin8
>
<
>
:
x
_3 = ;l
_2sin+l
cosx
_4 = ;l
_2cos;l
sinx
_7 = _2+v
1sinl
+v
2cosl :
Therefore,
satisfy the following second order implicit dierential equationl
=v
1cos;(v
2+g
)sin (2.1) Thus, we take forW
the parameterizationx
=X
(z
) = (l
sinl
cosl#
cos;l#
siny
5y
6y
7) wherez
= (#y
5y
6y
7)2]; R4the controlled vector eld(
zv
) = (zg
(zv
)) =# @@
+ (v
1cosl
;(v
2+g
)sinl
)@@#
+v
1@y @
5 +v
2@y @
6 + (#
2+v
1sinl
+ (v
2+g
)cosl
)@y @
7ESAIM: Cocv, March 1998, Vol. 3, 49{81
and the Lagrangian ~
L
(zv
) =L
(X
(z
)@X @z
(z
)g
(zv
)):
Then the problemPcis equivalent to the explicit optimal control problem min
(z(t)z_(t))=(z(t)v(t))8t201]
v(t)2R2 z(0)2fagR3 z(1)2fbgR3
Z
1
0
L
~(z
(t
)v
(t
))dt
for which we obtain the necessary conditions of optimality with the pseudo- Hamiltonian
H
0(zv
) = 1#
+2(v
1cosl
;(v
2+g
)sinl
) +3v
1+4v
2 +5(#
2+v
1sinl
+ (v
2+g
)cosl
);0L
~(zv
):
Remark 2.10. For this system, the kinetic energy is
T
(_) = 12
ml
2_2, the potential energy isV
() =;mgl
cos and the Lagrangian isL=
T
;V
= 12ml
2_2+mgl
cos:
The virtual work of the control
u
isW
u =Q
= (u
1l
cos;u
2l
sin) and for the tension it is zero. The Lagrange equationdt d @
L@
_;@
L@
=Q
gives the second order dierential equation 2.1.3. Geometry of Implicit Differential Equations For the problem P,
M
is a submanifold ofT
Rn it is obvious that the reduction procedure that we present in the introduction is not applicable to any submanifoldM
ofT
Rn, especially the submanifoldsM
for which jMadmits singularities. In this section we will dene the class of submanifolds of
T
Rnthat will be allowed for the problem P. For this class of submanifolds we will be able to apply locally the reduction procedure. First of all, let us make some comparisons with the denition of -submanifolds given in 16]. The authors' concern is to answer to the problem of the existence and uniqueness of solutions, namely to putM
in the formM
=(Y
) (3.1)for a section
:Y
!TY
of a connected submanifoldY
ofRnwith a dimen- sion equal to that ofM
. In this situation,M
is equivalent to an ordinary dierential equation and, thus, the problem of existence and uniqueness is solved. Here, it is not our purpose to obtain the existence and uniqueness of the solutions (since in this case the optimal control problem admits an obvious solution, namely the trajectory which (possibly) goes froma
tob
) but to have the existence and uniqueness of a family of solutions, in other words to nd a submanifoldY
of Rn, an open setU
of Rq and a mapping:
Y U
!TY
such that the diagramY U
!TY
T
RnPr
& #Y
Rn switches and such thatM
= (Y U
):
(3.2)This occurs for the submanifold
C
(M
) in the example of the controlled rigid pendulum. Even though in 16] the equality (3.1) suggest that the submanifoldM
is locally embedded in a tangent bundleTY
of a submanifoldY
of Rnwith the same dimension as thatM
, in our situation the equality (3.2) suggests that the submanifoldM
is locally embedded in a tangent bundleTY
of a submanifoldY
of Rnof dimension less than or equal to the dimension ofM
.3.1.
q
--submanifoldFor our geometric framework we will consider separable, Haussdorf man- ifold
X
with nite dimension and, for reasons of convenience, they are as- sumed to be smooth (although they could be of classC
k,k
2). Let us recall some elements of dierential geometry. The dimension of a manifoldM
is the maximal dimension among the dimension of the connected compo- nents ofM
. A pure manifoldM
is a manifold such that all the connected components have the same dimension. For any manifoldX
, the points belonging to the tangent bundleTX
are denoted by (xp
) withx
belonging toX
andp
belonging toT
xX
. The canonical projection:TX
!X
is the mapping such that (xp
) =x
. For the manifold Rn, the tangent bundle is identied withRn Rnand the projectionis identied with the projection onto the rst factor. Moreover, for any submanifoldY
ofX
and any point belonging toY
, the subspaceT
xY
is identied with a subspace ofT
xX
and, thus,TY
is identied with a submanifold ofTX
. Subsequently, the follow- ing notationf
: (Xa
)!(Yb
) means that the mappingf
is dened in an open neighborhoodU
ofa
inX
andb
=f
(a
). As in the case of manifolds, all the mappings are assumed to be smooth (once again they could be of classC
k,k
2). For any mappingf
:X
!Y
and any pointx
belonging toX
the linear tangent mapping is denoted byT
xf
. Now let us give the denition of subimmersion and the subimmersion theoremDefinition 3.1. (subimmersion) Let
X
,Y
be manifolds and a mappingf
:X
!Y
.(a):
f
is a subimmersion atx
2X
ifr
= rankT
xf
is constant in an open neighborhood ofx
inX
.(b):
f
is a subimmersion onX
if it is a subimmersion atx
for all pointsx
ofX
. In particular, for each connected component ofX
the rankr
has a constant value on , we shall call it the rank off
on . Theorem 3.2. (subimmersion theorem) LetX
be connected manifolds andf
:X
!Y
a subimmersion with rankr
. Then, the following statements holdESAIM: Cocv, March 1998, Vol. 3, 49{81(a): for any
y
belonging tof
(X
) the setM
=f
(y
) is a submanifold of dimensionm
;r
andT
xM
=KerT
xf
.(b): for any point
x
belonging toX
there exists an open neighborhoodV
ofx
inX
such that the setW
=f
(V
) is a submanifold of dimensionr
andT
yW
=ImT
xf
for any pointy
belonging toW
. Moreover, ifN
is any submanifold ofX
of dimensionr
such thatx
belongs toN
andT
xN
\kerT
xf
=f0g then the mappingf
jN is a local dieomorphism of some open neighborhood ofx
inN
onto an open neighborhood ofy
inf
(V
).Proof. see 9]
For an implicit dierential equation (1.1) the following proposition gives a criterion for the projection
jM to be a subimmersion.Proposition 3.3. Let
G
:Rn Rn!Rn;q be a mapping with 0q
n
such thatDG
(xp
) has full rankn
;q
in an open neighborhood of a point (x
0p
0) belonging toG
;1(0) andU
an open neighborhood of this point inRn Rn such that the set
M
=U
\G
;1(0) is a submanifold of Rn Rnof dimensionn
+q
. Then, the mappingjM where is the projection onto the rst factor is a subimmersion at (x
0p
0) of rankr
=+q
if, and only if, rankD
pG
(xp
) =n
;q
is constant in an open neighborhood of (x
0p
0) inM
.Proof. On the one hand, for any point (
xp
) belonging toM
, the tangent spaceT
(xp)M
is equal to kerDG
(xp
)and his dimension is equal ton
+q
. On the other hand, the linear tangent mappingT
(xp)(jM) :T
(xp)M
!T
xRnis the restriction of the canonical projection to the subspaceT
(xp)M
. Namely, the mapping(
xp
)2T
(xp)M
7!x:
Then, for any point (
xp
) belonging toM
ker
T
(xp)(jM) =f(xp
)2T
(xp)M=x
= 0g=f0g kerD
pG
(xp
):
(3.3) Clearly, the mapping jM has constant rankr
in an open neighborhood of (x
0p
0) inM
if, and only if, dim ImT
(xp)(jM) =r
in an open neighborhood of (x
0p
0) inM
therefore if, and only if,dimker
T
(xp)(jM) = dimT
(xp)M
;dimImT
(xp)(jM) =n
+q
;r
(3.4) in an open neighborhood of (x
0p
0) inM
. According to (3.3), (3.4) holds if, and only if, dimkerD
pG
(xp
) =n
+q
;r
=n
;in an open neighborhood of (x
0p
0) inM
that is if, and only if, dimImD
pG
(xp
) = in an open neighborhood of (x
0p
0) inM
.Now we give the denition of a
q
--submanifoldM
ofTX
.Definition 3.4. (
q
--submanifold) LetX
be a manifold,q
a xed integer less than or equal to the dimension ofX
,M
a submanifold ofTX
and (xp
) a point ofM
.M
is aq
--submanifold ofTX
at (xp
) (in an neighborhood of (xp
) inM
) if the following conditions hold(a): there exists a connected open neighborhood
U
of (xp
) inM
and a submanifoldY
ofX
such that dimY
+q
= dimU
andU
is a subset ofTY
.(b): the mapping
jU :U
!X
is a subimmersion in the neighborhood of (xp
).M
is aq
--submanifold ofTX
if for any point (xp
) belonging toM
,M
is aq
--submanifold at (xp
).Remark 3.5. a) If
M
is aq
--submanifold at a point (xp
) ofM
, then we can assume that the mapping jU :U
!X
is a subimmersion onU
, even if this means shrinkingU
. Moreover, for any point (xp
) belonging toU
, the rst condition of the denition holds (It is enough to takeU
andY
).Thus, for any point (
xp
) belonging toU
,U
is aq
--submanifold at (xp
) in other wordsU
is aq
--submanifold ofTX
.b) When
M
is not aq
--submanifold ofTX
we can consider the set, possibly empty, of points (xp
) ofM
such thatM
isq
--submanifold ofTX
at (xp
).If it is a non-empty set, according to a), it is an open set of
M
and aq
-- submanifold ofTX
.The denition of a
q
--submanifold can be formulated in the following wayDefinition 3.6. (bis) LetX
be a manifold andq
an integer less than or equal to the dimension ofX
. A submanifoldM
ofTX
is aq
--submanifold ofTX
if for any connected component ofM
the following conditions hold.(a): for any point (
xp
) of there exists an open neighborhoodU
in of (xp
) and a submanifoldY
ofX
such that dimY
+q
= dim andU
is a subset ofTY
.(b): the mapping
j : !X
is a subimmersion in a neighborhood of any point (xp
) of .Remark 3.7. This denition extends the denition of a
-submanifold in 16], which is the caseq
= 0 of our denition. The rst condition means exactly thatM
is locally embedded in the tangent bundleTY
of a subman- ifoldY
ofX
of dimension less than or equal to the dimension ofM
. For the second, according to the subimmersion theorem, for any point (xp
) ofM
there exists an open neighborhoodV
of (xp
) inM
such thatW
=(V
) is a submanifold ofX
of dimension the rank of the mapping j at (xp
).This is the local analogous of the condition,
W
= (M
) is a submanifold ofX
, supposed in the global reduction procedure. For any connected com- ponent ofM
the inequality 2q
dim is satised (dim 2dimY
, dimY
+q
= dim). We shall use this inequality to prove a property of the index.Example 3.8. Let :
X U
!TX
be a smooth controlled vector eld with dimU
=q
. Let us assume that(a): the mapping @@u(
xu
) has full rankq
.(b): for any (
xp
) belonging toTX
either the equation (xu
) = (xp
) has a unique solution or it does not have any solution.(c): the mapping (
xu
)7! (xu
) is proper.ESAIM: Cocv, March 1998, Vol. 3, 49{81