THE CAR WITH N TRAILERS: CHARACTERISATION OF
THE SINGULAR CONFIGURATIONS
FREDERIC JEAN
Abstract. In this paper we study the problem of the car withntrailers.
It was proved in previous works (9], 12]) that when each trailer is perpendicular with the previous one the degree of nonholonomy isFn+3
(the (n+ 3)-th term of the Fibonacci's sequence) and that when no two consecutive trailers are perpendicular this degree isn+ 2. We compute here by induction the degree of non holonomy in every state and obtain a partition of the singular set by this degree of non-holonomy. We give also for each area a set of vector elds in the Lie Algebra of the control system wich makes a basis of the tangent space.
1. Introduction
A car with
n
trailers is a nonholonomic system it is, indeed, subject to non integrable constraints, the rolling without sliding of the wheels. The conguration of the system is given by two positions coordinates andn
+ 1 angles. There are only two inputs, namely one tangential velocity and one angular velocity which represent the action on the steering wheel and on the accelerator of the car.The problem of nding control laws was intensively treated in many pa- pers throughout the literature: for instance by using sinusoids (see the works of Murray, Sastry and alii 11], 15]) or from the point of view of dieren- tially at systems (introduced by Fliess and alii 3]).
In general the study of such systems (to prove controllability, to nd con- trol laws,
:::
) involves tools from nonlinear control theory and dierential geometry. In particular an important concept for such problems is the de- gree of nonholonomy, which expresses the level of Lie-bracketing needed to generate the tangent space at each conguration. This degree comes up for instance in estimation of the complexity required to steer the system from a point to another (see 7], 2]).Laumond (6]) has presented a kinematic model for the car with
n
trailers in 1991 and has proved the controllability for this model. He has also proved that the degree of nonholonomy of the system is bounded toward the top by 2n+1. Srdalen has afterwards proved in 12] that, when no two consec- utive trailers are perpendicular, this degree is equal ton
+ 2. The systemInstitut de Mathematiques, Universite Pierre et Marie Curie, Case 247, 4, place Jussieu, F-75 252 Paris Cedex 05. E-mail: jean@mathp6.jussieu.fr.
Received by the journal February 23, 1996. Accepted for publication August 5, 1996.
c Societe de Mathematiques Appliquees et Industrielles.
0
n-1
n n
vn
x y
(x,y)
ω θ θ
θ
Figure 1. Model of the car with
n
trailers.congurations corresponding to such cases are called the regular points of the conguration space (see 1], 8]).
More recently, it has been proved that the degree of nonholonomy is bounded by the
n
+ 3-th Fibonacci number (13]) and that this bound is a maximum (9],10]) which is reached if and only if each trailer (exept the last one) is perpendicular to the previous one.To close denitively the problem, we still have to study the non regular points for which the maximum degree of nonholonomy is not reached. For the car with 2, 3 and 4 trailers, a complete classication of the singularities has already been done in 4]. The goal of our paper is to extend this classi- cation to any number of trailers. Let us note that some results given here have already been presented without proof in 5].
In Section 2 of this report, we are going to write equations, give deni- tions and notations and construct an induction procedure. Section 3 groups together the main result of this paper, Theorem 3.1, and some conclusion on the form of the singular locus and on the degree of nonholonomy. Section 4 is devoted to the demonstration of Theorem 3.1, but the proof of some technical lemmas are relegated to the appendix.
2. Equations and notations 2.1. Control system
In this paper we are going to use the same representation as Fliess 3]
and Srdalen 12] for the car with
n
trailers. A car in this context will be represented by two driving wheels connected by an axle. The state is parametrised byq
= (xy
0:::
n)T where:(
xy
) are the coordinates of the last trailer, n is the orientation angle of the car with respect to thex
-axis, i, for 0i
n
;1, is the orientation angle of the trailer (n
;i
) with respect to thex
-axis.Esaim: Cocv, October 1996, Vol. 1, pp. 241-266
The kinematic model of a car with two degrees of freedom pulling
n
trailers can be given by:x
_ = cos0v
0y
_ = sin0v
0 _0 = R11sin(1;0)v
1 _ ... i = Ri1+1sin(i+1;i)v
i+1 _ ... n;1 = R1nsin(n;n;1)v
n_n =
!
n:
where
R
i is the distance from the trailer (n
;i
) to the trailer (n
;i
+ 1),!
n is the angular velocity of the car andv
n is the tangential velocity of the car.v
n and!
n are the two inputs of the system.The tangential velocity
v
i of trailern
;i
is given by:v
i= Ynj=i+1cos(
j;j;1)v
n:
Let us denote:f
ni=Qnj=i+1cos(j;j;1)v
i =f
niv
ni
= 0n
;1:
The motion of the system is then characterized by the equation:
q
_=!
nX
1n(q
) +v
nX
2n(q
) with8
>
>
<
>
>
:
X
1n = @@nX
2n = cos0f
0n @@x+ sin0f
0n@@y + sin(R11;0)f
1n@@0 ++ sin(nR;nn;1)@n@;1
(2.1) We will suppose that the distance
R
i doesn't depend oni
and to simplify we shall, from now on, consider it equal to 1 (we will come back to this hypothesis in Subsection 3.4).2.2. Characterization of the singular locus
We are going now to dene the singular locus of the control system
f
X
1nX
2ng, and give a characterization of this locus easy to use.In this section
n
is xed and we writeX
1 andX
2 instead ofX
1nandX
2n. LetL1(X
1X
2) be the set of linear combinations with real coecients ofX
1 andX
2. We dene recursively the distribution Lk =Lk(X
1X
2) by:Lk =Lk;1+ X
i+j=kLi
Lj]where Li
Lj] denotes the set of all bracketsVW
] forV
2Li andW
2Lj.Esaim: Cocv, October 1996, Vol. 1, pp. 241-266
Thus Lk is the set of linear combinations of iterated Lie brackets of
X
1 andX
2 of lengthk
. The union L of all Lk is a Lie subalgebra of the Lie algebra of vector elds onR2(S1)n+1. It is called the Control Lie Algebra of the systemfX
1X
2g.Let us introduce some notations. For
s
1 denote byi
= (i
1:::i
s) a sequence ofs
elements in f12gand by As the sets of these sequences such thati
1= 1 ifs >
1, that is:A
1 =f(1)
(2)gAs=f
i
= (1i
2:::i
s) ji
j = 1 or 2g ifs >
1:
(2.2) The functionsu
(i
) andd
(i
) indicate respectively the number of occurences of 1 and the number of occurences of 2 in the sequencei
= (i
1:::i
s), and the length of the sequence isji
j=s
. Obviously we haveu
(i
) +d
(i
) =ji
j.The vector eld
::: X
i1X
i2]:::X
is;1]X
is] will be denoted byX
i] orX
i1:::X
is] and its value inq
byX
i]q. By using the Jacobi identity, we can write a bracket of lengthk
(i.e. belonging to Lk) as a sum ofX
i] withji
jk
.Moreover, because of skew symetry of Lie bracket,
X
i] = 0 ifi
1 =i
2 andX
2X
i1:::X
is] = ;X
i1X
2:::X
is]. Then Lk is generated by the bracketsX
i] such thati
2As, for 1s
k
.For a given state
q
, letL
k(q
) be the subspace ofT
q(R2(S1)n+1) wich consists of the values atq
taken by the vector elds belonging to Lk. We have an increasing sequence of dimensions:2 = dim
L
1(q
)dimL
k(q
)n
+ 3:
(2.3) If this sequence stays the same in an open neighbourhood ofq
, the stateq
is called a regular point of the control system otherwise,q
is called a sin- gular point of the control system (see 1]). Thus the sequence (2.3) at any stateq
allows to characterize regular and singular points, i.e., the singular locus.To determinate the sequence (2.3), we dene, for
i
2f1n
+ 3g:ni(
q
) = minfk
j dimL
k(q
)i
gd
ni(q
) = minfd
j dimspanhX
j]q jj
jni(q
)d
(j
)d
ii
g (2.4) In other words, the fact thatk
=ni(q
) is equivalent to:dim
L
k(q
)i
dim
L
k;1(q
)< i
(2.5) The sequence (2.3) can be deduced from theni(q
)'s,i
= 1:::n
+3, by:if9
i
2f1n
+3gsuch thatk
=ni(q
), then dimL
k(q
) is strictly greater than dimL
k;1(q
) and equal to the greatestj
such that nj(q
) =k
, otherwise dimL
k(q
) = dimL
k;1(q
).Thus the functions
ni(q
),i
= 1:::n
+ 3, characterize completely the singular locus. Hence the paper is devoted to the calculation of these func- tions. Notice that the functionsd
ni(q
) are not useful in the characterizationEsaim: Cocv, October 1996, Vol. 1, pp. 241-266
of the singular locus, they are only a tool for the computation of the func- tions
ni(q
).To simplify we will omitt the dependence on
q
inni(q
) andd
ni(q
). Accord- ing to its denition,niincreases with respect toi
, fori
lesser than dimL
(q
) (wheni
is strictly greater than this dimension, ni is equal to ;1). We will prove in this paper (in Theorem 3.1) that this sequence is strictly in- creasing with respect toi
for 2i
n
+ 3, which means that, for anyk
, dimL
k(q
);dimL
k;1(q
) 1. In other words, we will prove that, for 2i
n
+ 3,k
=ni is equivalent to (compare with (2.5)):dim
L
k(q
) =i
dim
L
k;1(q
) =i
;1 (2.6)We can yet calculate the rst values of these sequences.
-
L
1(q
) is two dimensionnal for allq
, and 1n and n2 are equal to 1. To span a two dimensionnal linear space we need bothX
1andX
2whereas for a one dimensionnal linear spaceX
1 is sucient. Thend
n1 = 0 andd
n2 = 1.-
L
2(q
) is generated by the familyX
1X
2X
1X
2] which is three di- mensional for allq
(it is clear from Formula (2.1)), so3n= 2. Moreover it is not possible to nd another three dimensionnal family of vector elds which contains \a fewer number of 2", thend
n3 = 1.Finally, for all state
q
: 1n= 1d
n1 = 0 2n= 1d
n2 = 1 3n= 2d
n3 = 1:
(2.7)2.3. Induction procedure
For
q
2 R2(S1)n+1 and 1p < n
, we will denote byq
p the projec- tion ofq
on the rst (n
+3;p
) coordinates, that isq
p = (xy
0:::
n;p)T. Let us consider now the system of a car withn
;p
trailers. The statesq
0 belong toR2(S1)n;p+1 and the control systemfX
1n;pX
2n;pgis given by Formulas (2.1) withn
;p
instead ofn
, that is:(
X
1n;p = @@n;pX
2n;p = cos0f
0n;p@x@ + sin0f
0n;p@y@ +s
1f
1n;p@@0 ++s
n;p @@n;p;1where
c
m = cos(m;m;1),s
m = sin(m;m;1) andf
in;p =c
i+1c
n;p. Hence for anyq
0 we have the sequences jn;p(q
0) andd
nj;p(q
0),j
= 1:::n
;p
+ 3. The dimensions of the spacesL
k(X
1n;pX
2n;p)(q
0),k
1, are characterized by the sequencejn;p(q
0).On the other hand,
X
1n;p andX
2n;p can be seen as vector elds on R2 (S1)n+1 whose lastp
coordinates are zero and which values atq
depends only on the projectionq
p. We can then consider L(X
1n;pX
2n;p) as a subalgebra ofL(X
1nX
2n).Esaim: Cocv, October 1996, Vol. 1, pp. 241-266
Remark 2.1. The motion of a point
q
0 2R2 (S1)n;p+1 is characterized by an equation _q
0=u
1X
1n;p(q
0) +u
2X
2n;p(q
0), and so depends only on the stateq
0. On the other hand, the motion of the pointq
p is given by the projection on the rstn
;p
coordinates of the motion equation ofq
: then the motion ofq
p depends onq
and not only onq
p.According to Formulas (2.1), we can write
X
2nas:X
2n=s
nX
1n;1+c
ns
n;1X
1n;2++c
nc
n;p+2s
n;p+1X
1n;p+ (2.8) +c
nc
n;p+2c
n;p+1X
2n;p wherec
m= cos(m;m;1) ands
m= sin(m;m;1).With this relation, for 1
p
n
;1, we will be able to express a vector eld in L(X
1nX
2n) in function ofX
1n:::X
1n;p+1 and of vector elds inL(
X
1n;pX
2n;p).For instance, for
p
= 1, we have:X
2n =s
nX
1n;1+c
nX
2n;1X
1nX
2n] =c
nX
1n;1;s
nX
2n;1X
1nX
2nX
2n] = ;X
1n;1+X
1n;1X
2n;1]:
(2.9) HenceL(X
1nX
2n) is equal to L(X
1n;1X
2n;1)hX
1ni, where hX
1ni is the subalgebra generated byX
1n. Formula (2.8) allows to describe the projec- tion of L(X
1nX
2n) on L(X
1n;1X
2n;1). We will see for instance that the projection ofLk(X
1nX
2n) isLk;1(X
1n;1X
2n;1).The induction will be done in the following way: we will assume that the functions
jn;p(q
p) (j
= 1:::n
;p
+3) are known for anyp < n
and, by us- ing the relation (2.8), we will calculate the dimensions of theL
k(X
1nX
2n)(q
), and so theni(q
)'s (i
= 1:::n
+ 3), in function of thejn;p(q
p)'s.From now on the dependence on
q
orq
p will be omitted if there is no possible confusion for example we will write ni instead of ni(q
) and jn;p instead ofjn;p(q
p).3. Singular configurations 3.1. Exposition of the result
In this chapter we present the main result of this paper, Theorem 3.1, which gives the recursion formulas satised by the sequence of functions
ni. The proof of the theorem is given in Section 4.Let us introduce a sequence
a
p by:
a
1 = 2a
p = arctansina
p;1:
(3.1) This sequence is clearly positive and decreasing, that is: 0< a
p<
2 forp >
1. Notice that the recursion relationship is odd, that is if we dene anEsaim: Cocv, October 1996, Vol. 1, pp. 241-266
other sequence
a
p by the same recursion relationship and the initial valuea
;1 =;2, we havea
;p =;a
p.Let us dene also some brackets
A
n(d
)] (> d
or =d
= 1) by:8
>
>
<
>
>
:
A
n(10)] =X
1nA
n(11)] =X
2nA
n(d
)] =X
1nX
2n:::X
2n| {z }
d
X
1n:::X
1n| {z }
;d;1 ]
for >
2 (3.2) From this denition we can see that the bracketA
n(d
)] is of length and thatX
2n occursd
times in it.With these notations, we have:
Theorem 3.1.
8
q
2 R2 (S1)n+1, for 2i
n
+ 3, ni is streactly increasing with respect toi
, and we haved
ni=ni;1;1.We can calculate the functions
ni(q
) by the following induction formulas, fori
2f3n
+ 3g:1. If
n;n;1 =2, then: ni=in;1;1+d
ni;1;1:
2. If 9
p
2 1n
;2] and = 1 such that k ;k;1 =a
k;p for everyk
2fp
+ 1n
g, then: ni= 2in;1;1;d
ni;1;1:
3. Otherwise, ni=in;1;1+ 1:
Moreover, a basis Bn = f
B
nii
= 1:::n
+ 3g ofT
q(R2 (S1)n+1) is given by:B
ni=A
n(nid
ni)]q (3.3) 3.2. Form of the singular locusLet us study the sequence
n = (ni)i=2:::n+3 (we remove n1 because it is always equal to 2n). The level sets of this sequence give a partition of the conguration space. For example, Figure 2 shows us the partition obtained forn
= 3. Since each area is a cylinder with respect to the di- rection1;0, we have just shown the projection of these cells on a plane 1;0 =constant
. The complement of the four lines are the regular points of the system and corresponds to the values (12345) of the sequence 3. Forn
3, letq
2be the projection ofq
on the rstn
+1 coordinates. Theo- rem 3.1 allows us to calculate the values ofn(q
) in function ofn;2(q
2). We illustrate it in Figure 3, where we represent the set of pointsq
= (q
2n;1n) wich have the same projectionq
2.Esaim: Cocv, October 1996, Vol. 1, pp. 241-266
Let us consider now a point
q
such thatk;k;1 6=2 fork
= 2:::n
. It is clear (for instance by looking at Figures 2 and 3) that there exists a neighbourhood ofq
in which the sequencenis constant. Conversaly, if there is an integerk
, 2k
n
, such thatk;k;1 =2, such a neighbourhood doesn't exist. Then we have determined the regular and singular points (see also 12] and 1]).π/2
π/2 π/
- 2
π/
- 2
π/4
π/4 -
θ3- θ2
(1,2,3,5,7) (1,2,3,5,8)
(1,2,3,4,5) (1,2,3,4,7)
(1,2,3,4,6)
0 θ θ2- 1
Figure 2. Partition of the conguration space (
n
= 3) by the values of the sequence (3233435363).θn-2
θn-1-
θ θn- n-1
π/2 π/2
π/2 θn-1- θn-2
π/
- 2 - π/2
π/
- 2 - π/2
π/4
- - π/4
π/4 π/4
π/2 B
0 E
D C D C
E G B
A
A
0 F
H a
a
p+2 p+1
θ θn- n-1
q
2 in case 3q
2 in case 1 or 2A :
ni=in;2;2+ 2 B :ni=in;2;2+d
ni;2;2+ 1 C :ni= 2in;2;2+ 1 D :ni= 2in;2;2+d
ni;2;2 E :ni=in;2;2+ 2d
ni;2;2 F :ni= 2in;2;2;d
ni;2;2+ 1 G :ni= 3in;2;2;d
ni;2;2 H :ni= 3ni;2;2;2d
ni;2;2Figure 3. Cells of the subset (
q
2n;1n).Esaim: Cocv, October 1996, Vol. 1, pp. 241-266
Theorem 3.2. The singular locus of the system is the set of the points for which there exists
k
22n
] such that k;k;1 =2.3.3. Application to the degree of nonholonomy
The degree of nonholonomy of the system at a point
q
is the degree from wich the sequence (2.3) is constant, that is the degreer
such that:L
r;1(q
)6=
L
r(q
) =L
r+1(q
) = =L
(q
):
With our notations, this degree
r
is given by the greatestni, fori
n
+3.Thus Theorem 3.1 implies that the degree of nonholonomy of the system is equal to
nn+3 at any pointq
, and then the rank of the Control Lie Algebra at any point isn
+ 3.Let us recall the Chow theorem (also called the Lie Algebra Rank Con- dition): if the rank of the Control Lie Algebra at any point
q
of the cong- uration space is equal to the dimension of the tangent space in this point, the system is controllable (see for instance 14]).This condition is satised here, therefore the system is controllable. We are meeting a classic result, wich was rst proved by Laumond in 1990 (6]).
By using Theorem 3.1, we can study the function
nn+3 and nd some other results about the degree of nonholonomy (these results were already proved in 12],13], and 10]):Theorem 3.3.
(i) At a regular point, that is a point such that
k ;k;1 6=2 for everyk
22n
], the degree of nonholonomy of the system isn
+ 2.(ii) The maximum of the degree of nonholonomy is the (
n
+3)-th Fibonacci numberF
n+3 (recall that the Fibonacci sequence is dened byF
0 = 0,F
1 = 1,F
n+2 =F
n+1 +F
n), and this maximum is obtained if and only if all the trailers are perpendicular (except the last one), that is if k ;k;1 =2 for everyk
22n
].Proof.
This theorem is obtained by applying the recursion formulas of Theorem 3.1 and by using the values for
n
= 0 given by Formula (2.7): 3n= 2 =F
3 andd
n3 = 1 =F
2.Hence we see that
n
+ 2nn+3F
n+3. Moreover, forn
4,nn+3 can take all the values fromn
+2 toF
n+3, but this property is no more true forn >
4.3.4. Case where the distances between the trailers are not all equals
We have assumed (see Subsection 2.1) that the distance
R
i between the trailer (n
;i
) and the trailer (n
;i
+ 1) is independent oni
and equal to 1.If we remove this hypothesis, the result is the same as Theorem 3.1, except
Esaim: Cocv, October 1996, Vol. 1, pp. 241-266
that we have to replace the case 2 by:
2 bis.If 9
p
2 1n
;1] such that k ;k;1 =a
k;p(p
) for everyk
2f
p
+ 1n
g, then: ni= 2in;1;1;d
ni;1;1 where the sequencea
k;p(p
) is dened by:(
a
1(p
) = 2a
k;p(p
) = arctan(RRk;1k sina
k;p;1(p
))for k > p
+ 1:
The proof of this result is similar than the one of Theorem 3.1, but requires more notations. Therefore we don't give it in this paper. The dierence with the case
R
i = 1 is that the sequence of angles wich gives singularities de- pends onp
. For instance, ifR
p+2 6=R
p+1 =R
p,a
2(p
;1) is equal to 4 whereasa
2(p
) is not.4. Proof of Theorem 3.1
In this chapter,
n
1 is xed. The proof is organized as follows: in a rst time, we study the relationships between the Lie Algebra for then
-trailers system and the Lie Algebras for the systems with less thann
trailers. In a second time we use these relationships to establish the induction formulas for the functions ni(q
). The main point of the proof is the rst part, that is Lemma 4.1. This kind of proof is inspired by 10].4.1. Preliminary result
We have seen in Subsection 2.3 that, for 2
< m < n
, a vector eld inL(
X
1nX
2n) can be decomposed in a linear combination (with functions as coecients) ofX
1n:::X
1m+1 and of vector elds in L(X
1mX
2m). Lemma 4.1 gives such a decomposition and allows to conclude, in some particular cases, on the nullity or non nullity of the decomposition coecients.We will denote, for
m
n
:'
m = m;m;1c
m = cos'
ms
m = sin'
mt
m = sin'
m;cos'
msin'
m;1:
(4.1)Lemma 4.1. Let
p
, 1p
n
;1 andi
2 Ajij,i
6= (1) (the sets As are dened in Formula 2.2). Then there exist functionsh
k('
n;p+1:::'
n),n
;p
+ 1k
n
;1 andf
l('
n;p+1:::'
n) inC
1(Sp) depending oni
such that:X
ni] = nX;1k=n;p
h
kX
1k+Xds=1
X
l2As
f
lX
ln;p] whered
= maxf1d
(i
);p
+ 1g.Moreover, ifj
i
jp
+ 1, then:Esaim: Cocv, October 1996, Vol. 1, pp. 241-266
1. we have:
f
l =Xb2Il(
c
n;p+1)b0;b1(t
n;p+2)b1;b2(t
n)bp;1;bpg
bl (4.2) where the functionsg
bl('
n;p+1:::'
n)2C
1(Sp) depend onbl
andi
and the setI
lZp+1 satises:I
lfb
= (b
0:::b
p)jd
(l
) =b
0b
p0 Xpj=1
b
j ji
j;jl
jg 2. ifb
2I
l is such thatb
p= 0, then Ppj=1;1b
j< d
(i
);jl
j3. if we denote by
I
l+ the following subset ofI
l:I
l+=fb
2I
lj Xpj=1
b
j=ji
j;jl
jgfor every
b
2I
l+, there exist an integerb>
0 and a functionG
('
n;p+1,:::'
n) which depends only on ji
j, jl
j,d
(i
),d
(l
) andb
such that:g
bl=bG
bjijjljd(i)d(l)4. if
X
ni] =A
n(+ (p
;1)+r
+ (p
;1))], with>
r
1, and ifX
ln;p] is such thatl
2 A andd
(l
) = , the sequence(:::r
) belongs toI
l+(the denition of the bracketA
n(d
)] is given by (3.2)) 5. ifp
= 1,ji
j3, andd
(i
) = ji
j;1, then the coecient ofX
ln;1] suchthat
l
2Ad(i) andd
(l
) =d
(i
);1 is:f
l('
n) = (c
n)d(i);2:
Proof. The proof is quite long and technical, so it is done in the appendix (where the lemma is divided in four parts: Lemmae 5.1, 5.2, 5.3 and 5.4).
The point 3 implies that functions
G
('
n;p+1:::'
n) doesn't depend on the sequencesi
andl
but only on the length ji
j,jl
jand on the \number of 2" in these sequences (namelyd
(i
) andd
(l
)). The form of the sequencesi
andl
acts only on the integerb, and then not on the sign ofg
bl. The exact form ofG
('
n;p+1:::'
n) is given in Lemma 5.3 but it is not useful here.Remark 4.2. It appears from this lemma that the terms
c
n;p+1,t
n;p+2,:::
,t
nhave a particular part in the decomposition (4.2) (it will be conrmed in what follows). Thus it is interesting to notice that all of these quantities are zero if there exists =1 such thatk;k;1=a
k;p fork
=p
+1:::n
. In this case, the functionf
l can be non zero only if there existsb
2I
l such thatb
0 ==b
p (we set 00= 1).4.2. Proof
4.2.1. Plan of the proof. We are not going to prove directly Theorem 3.1 but the following proposition, which implies the theorem.
Proposition 4.3. Let
n
1. Then, for every stateq
we have: ni increases strictly with respect toi
(fori >
1),d
ni=in;1;1,Esaim: Cocv, October 1996, Vol. 1, pp. 241-266
- if91
p
n
;1and
=1 such that'
n;p+1 =a
1:::'
n=a
p(
a
p is dened by (3.1)), then:ni=ni;1;1+ 1
for i
= 3:::p
+ 2 ni=ni;;pp+p d
ni;;ppfor i
=p
+ 2:::n
+ 3 - otherwise, ni=ni;1;1+ 1 fori
= 3:::n
+ 3,f
B
ni =A
n(nid
ni)]qi
= 1:::n
+ 3g is a basis ofT
q(R2(S1)n+1) (the denition of the bracketsA
n(d
)] is given by (3.2)),every vector
X
ni] such thati
2 Ani andd
(i
) =d
ni has a positive coordinate on the basis vectorB
ni (recall that, ifi
2 As ands >
1, theni
1= 1 (see (2.2)).The rst four points of this proposition are equivalent to Theorem 3.1 (the induction formulas for
ni are the same but expressed in a dierent way). The last point of the proposition is an induction hypothesis required for the proof and then is omitted in the theorem.The proof will be done by induction on
n
. We assume that Proposition 4.3 is true for everym < n
, and we will prove that it is true also forn
by proceeding as follows:- for each
q
andi
we have "candidates" for ni andd
ni (given by the proposition)- in Lemma 4.5 and Corollary 4.8, we prove that these "candidates" are less than
ni andd
ni- with Lemma 4.9, we establish that there exists a basis of
T
q(R2 (S1)n+1) formed by vectors the length (and number ofX
2n) of which are equal to the "candidates", fori
= 1 ton
+ 3- by using Lemma 4.7 we prove that
ni andd
ni are indeed equal to the"candidates"
- Lemma 4.10 allows us to establish the last point of Proposition 4.3 and so to conclude.
The form of Proposition 4.3 implies that we have to distinguish several possibilities for the state
q
:9 =1 such that
'
n=a
1,9
p
2and
= 1 such that'
n;p+1 =a
1:::'
n;1 =a
p;1 and'
n=a
p,9
p
2and
= 1 such that'
n;p+1 =a
1:::'
n;1 =a
p;1 and'
n6=a
p and 6=a
1,such a
p
2 doesn't exist and'
n6=a
1. We can resume these possibilities in two cases:q
2(a
) if 9p
2f1:::n
;1gand, ifp >
1 9 =1 such that'
n6=a
1 and, ifp >
1'
n;p+1=a
1:::'
n;1=a
p;1'
n6=a
pq
2(b
) if 9p
1p
n
;1 and =1 such that'
n;p+1=a
1:::'
n=a
p:
(4.3) For instance the generic case is
q
2(a
) andp
= 1. In this case,'
n6=2 (sincea
1 = 2 from (3.1)) and there is no sequence'
n;p+1:::'
n;1 equalEsaim: Cocv, October 1996, Vol. 1, pp. 241-266