**Corollary 2.6. Suppose Chow's condition holds. The topology defined by the sub-Riemannian distance d is the original topology of M**

**2.4. Direct, effective proofs of Chow's theorem**

So, by Ascoli-Arzela theorem, *P *is a compact subset of C([O, *T], M). *

*It follows that the length functional attains its infimum on P. ***In **other
words, there exists a path with length *T *= *d(q, q') *joining *q *to *q'. * *• *
**Remarks. 1. ***We do not assert, either that for q, q' belonging to B(p, c) *
*the geodesic joining q and q' is unique, or that it is contained in B (p, c). *

See the examples in §3.

2. Assuming Chow's condition, the conclusion of (ii) holds, in several
important cases: when *M is compact, when M * = ]Rn and the *Xi are *
bounded, and when M is a Lie group and the Xi are left-invariant vector
fields. Indeed, *M * is complete in these three cases.

**2.4. Direct, effective proofs of Chow's theorem **

At this point, it is impossible not to mention the existence of proofs of Chow's Theorem, more effective than the one we have given.

Here and in the sequel, it will be convenient to note on the right the
action of diffeomorphisms: the action of *etX * on point *p * will result in
*p **etx . This notation is consistent with our notation for the concatenation *
of paths, and complies with the fact that all diffeomorphisms we shall use
on Lie groups in the sequel come from flows of left-invariant vector fields,
and so are defined by right multiplications.

For the case n

### =

3, m### =

2, where one assumes that Xl,*X*

*2*and [Xl,

*X*

*2 ]*

span the tangent space at *p, *one proves that the mapping

where we write *t**l / 2 *for sgn(t)IW/2 , is tangent to the mapping
*(tl' **t**2, *^{t}*3) *^{f---7 }*peilXlet2X2et3[Xl,X2] *

(11)

at t = O. This shows that the end-point mapping *Ep * is open at the
origin. At the same time one gets local estimates for the sub-Riemannian
distance: given any Riemannian metric *b, *there exist a neighbourhood *U *
of *p *and constants C, C' such that

*Cb(q,q'):S d(q,q'):S C'b(q,q')1/2 *

2.5 TANGENT SPACE IN SUB-RIEMANNIAN GEOMETRY 21

*for any q, q' in U. *

The effective proof of Chow's theorem in case *n *

### =

3, m### =

2 is based on Campbell-Hausdorff formula (see Lobry [21]). It can be generalized to all cases where Chow's condition holds (see Gromov, this volume), and one gets similar estimates with 1/2 replaced by*l/r, where r is the smallest*integer for which the tangent space is spanned by brackets of length ::;

*r*of

*Xl, ... ,X*

*m*(the degree of nonholonomy, see §4.1). We shall prove more precise estimates in Section 7.

Observe that mappings such as (11) give a mean to construct a non-differentiable section, actually H6lderian of order 1/2, of the end-point map.

2.5. Accessibility does not depend on the class of controls used
The set of points accessible from a given point in M by means of control
functions belonging to any reasonable class of control functions, ranging
from piecewise constant to *L1, *is independent of the class of controls used.

More precisely, we have the following theorem.

Theorem 2.8. *Let * C be a *class * of *control functions such that *
C ([0, TJ, ~m) *is *a *dense subspace in L1 ([0, T], *~m). *Then *any *point *
*ac-cessible from p is acac-cessible from p *by means of *controls *of *class *C.

Proof. Let *q *be a point accessible from *p. *Using Lemma 2.2, choose
a *normal *control *u *E *L1([0, T], *~m) steering *p *to *q. * For simplicity, we
denote *Ep,T *by 7f. *Since u is normal, the linear mapping *

*d7ru : L1 *-+ *TqAp *

is surjective. Fix a sequence of finite-dimensional spaces

of *C([O, Tj, *~m), with strictly increasing dimension, such that

### U

*Hk is*dense in

*L1.*(The existence of such a sequence of subspaces stems from the separability of

*L1.)*For some integer

*ko,*we have

*d7ru(Hko) *

### =

*TqAp.*

Choose a linear subspace *V of Hko' of the same dimension as Ap, such *
that one has still

22 ANDRE BELLAICHE

Then there exists *E *

### >

0 such that (i)^{7f }is defined on B

*(0,*

^{V }*2E);*

§ 2

(ii) The mapping ¢ : *h *r--+ *7f(u *

### +

*is a diffeomorphism of*

^{h) }*13 =B*the closed ball with center 0 and radius

^{v }(O,E),*E,*onto

*¢(B)*= 7f(

*U*

### +

13).Now, let *Uk *E *Hk (k *= 1,2, ... ) a sequence of control functions converging
to u. For *k *large enough, the mapping

is defined on all of *13, *and it converges uniformly, in the C^{l } sense, to the
mapping ¢. Using Lemma 2.9 below, applied to the sequence of mappings

### h

^{= }¢-l

*O¢k (k*= 1, 2, ... ), one shows that q is in the image of

*¢k.*Since

*¢k(B) *= *7f(Uk +13) *consists of images by ^{7f }of elements in

### U

*H k,*it results that

*q*is accessible by means of a control function in the class C. •

**Lemma 2.9.**

*LetB be the closed ball in lR*

^{n }

*of center q and radius*

^{E, }*and*

*let fk*:

*13*

*lR*

^{--t }^{n }

*(n*

### =

1, 2, ... ) be a*sequence of differentiable mappings*

*converging uniformly, in the C*

^{l }

*sense, to the identity map of B. Then,*

*for k large enough, the image h(B) contains q.*

**Proof. It **suffices to prove that, for any differentiable mapping *9 : **13 **--t *

lR *n *verifying

*Ilg(x) -*

### xii

^{:S; }

### 2'

*E*

for all *x inB, then the image of **9 *contains q.

For that purpose, consider the sequence in *13 *defined by
*Xo=q, Xi+l=q+Xi-g(Xi) * (i=1,2, ... ).

It is well defined, that is, one can prove inductively that *Xi *E *B. *Indeed,
we have

*E * *E * *E *

### Ilxi - qll

:S;### Ilxl - qll + IIx2 - xIII + ... + Ilxi - Xi-III

:S;### 2 + 4 + ... +

2i <*E.*

The same computation proves that the series

is convergent. In other words, * ^{Xi }*converges to some

*Clearly, we have*

^{Xoo. }### Ilxoo - qll

:S;*E*and

*g(xoo)*=

*q.*

*•*

3.1 TANGENT SPACE IN SUB-RIEMANNIAN GEOMETRY 23

The same result is still true, with a proof

### a

la Brouwer, if one sup-poses only that the*!k *

are continuous, and CO convergence holds. (The
use of such a fixed-point argument in the proof of theorem 2.8 has been
suggested to me by Hector Sussmann.)
Remark. The minimal notion of accessibility is obtained by using
only concatenations of integral curves of the vector fields *Xl,"" *

*X**m , * which amounts to use concatenations of controls of the form
(0, ... ,0, ±l, 0, ... ,0), the so-called bang-bang controls (in fact, one
should use controls (0, ... ,0,

*±t*

*i ,*0, ... ,0) if one wants to keep a fixed time interval). Bang-bang controls do not form a vector subspace of

*L1*but the same conclusion as in Theorem 2.8 holds, with a slightly modified proof.

3. Two examples

Examples of sub-Riemannian manifold include Riemannian
manifold-the case *D *

### =

*T M*and Riemannian foliations-the case where

*D*is inte-grable.

More genuine examples are the Grusin plane below-which is almost Riemannian, and really sub-Riemannian along some singular line only, and, most important, the Heisenberg group, where the role of nonholon-omy appears clearly, and which serves as a paradigm for the theory.

3.1. The Gru~iin plane G2

We take as underlying manifold of G2 the ]R2 plane (with coordinates
*x, y) *and consider the sub-Riemannian metric defined by the vector fields

These vector fields span the tangent space everywhere, except along the
line *x *= 0, where adding

is needed. So Chow's condition holds. Outside the line *x *= 0, the
sub-Riemannian metric is in fact sub-Riemannian, and is equal to

24 ANDRE BEL LAICHE § 3

Any path has finite length, provided its tangent is parallel to the x-axis when crossing the y-axis.

This example of sub-Riemannian manifold is named after Grusin, who
was the first to study the analytic properties of the operator *L *=

*Xl * *+ *

X~ =

*a; * +

*x*

^{2}*a;, *

and of its multidimensional generalizations [15,16].
Dilations and distance estimates. A very important feature is the
ex-istence of a one-parameter group of dilations for *G**2 : *if we set

then we have

(8,>,)*XI

### =

A-lXI,*(8,>,)*X2*

### =

A-^{1}X 2

for

*Ai-*

0. Therefore, the length of a controlled path is multiplied by IAI
under the action of 8,>,. It follows that
for all p, *q *E *G**2 *and *A *E JR.

It is easy to bound *d((O,O), (x,y)) *on the boundary of the square Ixl ::;

1, Iyl ::; 1: it is 2: 1 and::; 3. Using homogeneity under the action of 8,>, we get the estimates

sup(lxl, lyll/2) ::; *d((O, *0), *(x, y)) ::; *3sup(lxl, lyll/2). (12)
Instead of (12), one may prefer to use

Hlxl

### +

IYII/2) ::;*d((O,*0), (x,

*y)) ::;*3(lxl

### +

^{lyll/2). }

^{(13) }

In geometric terms, (12) means that balls *B(O, *c) are roughly of the shape
[-c, c] x [_c2 , c2 ].

More precisely, we have

### !

^{[-c,c]x[-c}

^{2}

^{,c}

^{2] }

^{C }

^{B(O,c) }

^{C }

^{[-c,c]x[-c}

^{2}

^{,c}

^{2]. }

Similar estimates hold around *(xo, Yo) *when *Xo * = 0, but not around
regular points, when *Xo *

### i-

0. In this case, there are no dilations centered at*(xo, Yo),*and one has only local estimates: balls centered at regular points p

### =

*(x, y),*that is

*x*

### i-

0, have the overall form[-c,c]x[-c,c]

for small c, since the metric is Riemannian near those points.

3.2 TANGENT SPACE IN SUB-RIEMANNIAN GEOMETRY 25

3.2. The Heisenberg group

Consider now JR3, with the sub-Riemannian metric defined by

The vector fields Xl, *X *2 and

span JR3 everywhere. Here, JR3 can be identified with the Heisenberg group
*H**3 , *so every point can be reached from any other point. Setting ^{X3 }*= *

[Xl, *X**2], *we see that

Therefore, the Lie algebra generated by Xl and *X**2 *is isomorphic to the
Heisenberg Lie algebra ~3' We can actually identify JR3 with the
Heisen-berg group *H3 * by letting *(x, y, z) map to **ezX3eyX2eXX1. *The product
operation on JR3 is given by

*(x, y, z)(x', y', z') *= *(x *

### +

*x', y*

### +

*y', z*

### +

*z'*

### +

*xy').*

*(14)*In this picture, the distance defined by Xl and X

*2*becomes left-invariant, i.e., we have

*d(gg', gg")*=

*d(g', g").*These statements may be easily proved by using the identity

*eaeb*=

*ebeae[a,bl,*holding in

*H*

*3 •*

Dilations and distance estimates. Here also there exists a one-parameter group of dilations

*b),. : (x, y, z) *^{1--* }*(Ax, AY, A*^{2}*z), *

so, as in the Grusin case, one can prove estimates of the form

*C(lxl *

### + IYI + Izll/2)

~*d(O, *

*(x, y, z»)*~

*C'*

### (Ixl + Iyl + Izll/2).

^{(15) }

On the set

### Ixl + Iyl + Izll/2

^{= }1 the function

*d(O, *

*(x, y, z»)*is positive and finite. Since the set is compact and

*d is continuous on JR3, there exists*positive, finite constants

*C, C'*such that

*C*~

*d(O, *

*(x, y, z»)*

### <

*C'*for

### Ixl + Iyl + Izll/2 =

1. Using dilations, we get (15).It follows that balls *B(O, *e) look roughly like
[-e, e]X [-e,e] x [_e2, e2].

26 ANDRE BELLAlCHE § 3

**Exact distance estimates. **We can give precise bounds for d. First, from
the formula

for *z *~ 0, and from a similar formula for *z *~ 0, one constructs a
concate-nation of integral curves of Xl and X*2 *of total length

### Ixl + Iyl + 4Izll/2,

leading from the origin to *(x, y, z). *This gives an upper bound for *d. *To
get a lower bound, we observe that finite length paths starting at the
origin are obtained by integrating the system

*p *

^{= }

*UI(t)XI*

### +

*U2(t)X2,*

*p(O)*

### =

0, that is,Integrating gives

{

*x(O) *

### =

0*y(O)*

### =

0*z(O)*

### =

O.*x(T) *

### = *lT *

*UI(t) dt, y(T)*

### = *lT *

*U2(t) dt, z(T)*

### = *IT(lt *

*UI(T) dT )U2(t) dt.*

If we choose controls such that *UI(t)2+ U2 (t)2 *= 1, we obtain the estimates

### Ix(T)1

^{~ }

^{T, }### ly(T)1

^{~ }

^{T, }### Iz(T)1

^{~ }

^{T2 }Since *d(O, p) *is the infimum of *T *such that there exists a path with velocity
1, parameterized by [0, Tj, *and joining 0 to p, it follows that *

### Ixl

~d(O,(x,y,z)),### Iyl

~d(O,(x,y,z)),### Izl

~d(0,(x,y,z))2.Summing up, we obtain

### Hlxl + Iyl + Izll/2)

~*d(O,(x,y,z))*~

### 4(lxl + IYI + IzII/2).

^{(16) }

Observe that, because of group invariance, all points of *H3 play the *
same role. So, every point of H3 is the center of a I-parameter group of
dilations. Estimates similar to (16) hold for

*d( *

*(x, y, z), (x', y', Zl)).*See §7, Eq. (52).

**3.3. The Heisenberg group using exponential coordinates **