Ludovic Rifford
Sub-Riemannian
Geometry and
Optimal Transport
SpringerBriefs in Mathematics
Series editors
Krishnaswami Alladi, Gainesville, USA
Nicola Bellomo, Torino, Italy
Michele Benzi, Atlanta, USA
Tatsien Li, Shanghai, Peoples Republic of China
Matthias Neufang, Ottawa, Canada
Otmar Scherzer, Vienna, Austria
Dierk Schleicher, Bremen, Germany
Vladas Sidoravicius, Rio de Janeiro, Brazil
Benjamin Steinberg, New York, USA
Yuri Tschinkel, New York, USA
Loring W. Tu, Medford, USA
G. George Yin, Detroit, USA
Ping Zhang, Kalamazoo, USA
SpringerBriefs in Mathematics showcases expositions in all areas of mathematics and applied mathematics. Manuscripts presenting new results or a single new
result in a classical field, new field, or an emerging topic, applications, or bridges
between new results and already published works, are encouraged. The series is
intended for mathematicians and applied mathematicians.
For further volumes:
http://www.springer.com/series/10030
BCAM SpringerBriefs
Editorial Board
Enrique Zuazua
BCAM - Basque Center for Applied Mathematics & Ikerbasque
Bilbao, Basque Country, Spain
Irene Fonseca
Center for Nonlinear Analysis
Department of Mathematical Sciences
Carnegie Mellon University
Pittsburgh, USA
Juan J. Manfredi
Department of Mathematics
University of Pittsburgh
Pittsburgh, USA
Emmanuel Trelat
Laboratoire Jacques-Louis Lions
Institut Universitaire de France
Universite Pierre et Marie Curie
CNRS, UMR, Paris
Xu Zhang
School of Mathematics
Sichuan University
Chengdu, China
BCAM SpringerBriefs aims to publish contributions in the following disciplines:
Applied Mathematics, Finance, Statistics and Computer Science. BCAM has
appointed an Editorial Board that will evaluate and review proposals.
Typical topics include: a timely report of state-of-the-art analytical techniques,
bridge between new research results published in journal articles and a contextual
literature review, a snapshot of a hot or emerging topic, a presentation of core concepts that students must understand in order to make independent contributions.
Please submit your proposal to the Editorial Board or to Francesca Bonadei,
Executive Editor Mathematics, Statistics, and Engineering: francesca.bonadei@
springer.com
Ludovic Rifford
Sub-Riemannian
Geometry and
Optimal Transport
123
Ludovic Rifford
Laboratoire J.A. Dieudonn
Universit Nice Sophia Antipolis
Nice
France
ISSN 2191-8198
ISSN 2191-8201 (electronic)
ISBN 978-3-319-04803-1
ISBN 978-3-319-04804-8 (eBook)
DOI 10.1007/978-3-319-04804-8
Springer Cham Heidelberg New York Dordrecht London
Library of Congress Control Number: 2014933272
The Author(s) 2014
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief
excerpts in connection with reviews or scholarly analysis or material supplied specifically for the
purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the
work. Duplication of this publication or parts thereof is permitted only under the provisions of
the Copyright Law of the Publishers location, in its current version, and permission for use must
always be obtained from Springer. Permissions for use may be obtained through RightsLink at the
Copyright Clearance Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
Preface
Ludovic Rifford
vi
Preface
References
1. Agrachev, A., Barilari, D., Boscain, U.: Introduction to Riemannian and sub-Riemannian
geometry. To appear
2. Montgomery, R.: A tour of subriemannian geometries, their geodesics and applications.
In: Mathematical Surveys and Monographs, vol. 91. American Mathematical Society,
Providence, RI (2002)
3. Villani, C.: Optimal transport, Old and new. Springer-Verlag, Heidelberg (2008)
Contents
Sub-Riemannian Structures . . . . . . . . . . . . .
1.1 Totally Nonholonomic Distributions . . . .
1.2 Horizontal Paths and End-Point Mappings
1.3 Regular and Singular Horizontal Paths . . .
1.4 The Chow-Rashevsky Theorem . . . . . . . .
1.5 Sub-Riemannian Structures . . . . . . . . . . .
1.6 Notes and Comments . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
12
20
30
33
36
36
Sub-Riemannian Geodesics. . . . . . . . . . . . . . . .
2.1 Minimizing Horizontal Paths and Geodesics .
2.2 The Hamiltonian Geodesic Equation . . . . . .
2.3 The Sub-Riemannian Exponential Map . . . .
2.4 The Goh Condition . . . . . . . . . . . . . . . . . .
2.5 Examples of SR Geodesics . . . . . . . . . . . . .
2.6 Notes and Comments . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
37
37
44
52
59
70
75
76
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
77
77
81
91
97
111
116
118
121
125
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
139
vii
Chapter 1
Sub-Riemannian Structures
y Vx .
Such a family of smooth vector fields is called a local frame in Vx for the distribution
. All the distributions which will be considered later will be smooth with constant
rank m [1, n]. Thus, from now on, distribution always means smooth distribution with constant rank. A co-rank k distribution on M is a distribution of rank
m = n k and any smooth vector field X on M such that X(x) (x) for any x M
is called a section of .
Example 1.1 We call trivial distribution on M the rank n distribution defined by
(x) = Tx M for all x M. For topological reasons, such a distribution may not
admit non-vanishing sections (for example, by the hairy ball theorem, there is no
non-vanishing continuous vector fields on any even dimensional sphere).
Example 1.2 In R3 with coordinates (x, y, z), the distribution generated by the
vector fields X and Y , that is
(x, y, z) = Span X(x, y, z), Y (x, y, z)
(x, y, z) R3 ,
L. Rifford, Sub-Riemannian Geometry and Optimal Transport,
SpringerBriefs in Mathematics, DOI: 10.1007/978-3-319-04804-8_1,
The Author(s) 2014
1 Sub-Riemannian Structures
with
y
x
X = x z and Y = y + z ,
2
2
is a rank 2 (or co-rank 1) distribution on R3 .
Example 1.3 More generally, if x = (x1 , . . . , xn , y1 , . . . , yn , z) denotes the coordinates in R2n+1 and the 2n smooth vector fields X 1 , . . . , X n , Y 1 , . . . , Y n are defined by
X i = xi
yi
xi
z , Y i = yi + z
2
2
i = 1, . . . , n,
then the distribution generated by the above vector fields is a co-rank 1 distribution
on R2n+1 .
Example 1.4 Let be a smooth non-degenerate 1-form on M, that is a 1-form which
does not vanish (x = 0 for any x M). The distribution defined as
(x) = Ker (x )
x M,
is a co-rank 1 distribution on M.
We say that a given distribution on M admits a global frame if there are m
smooth vector fields X 1 , . . . , X m on M such that
(x) = Span X 1 (x), . . . , X m (x)
x M.
In general, distributions do not admit global frames (see Example 1.1). It is worth
noticing that in the particular case of Rn all distributions are trivial.
Proposition 1.1 Any distribution on Rn admits a global frame.
Proof Let us first show how to construct a non-vanishing section of a given distribution on Rn .
Lemma 1.2 Let be a distribution of rank m on Rn . Then there is a non-vanishing
smooth vector field X such that X(x) (x), for any x Rn .
Proof (Proof of Lemma 1.2) Define the multivalued mapping : Rn 2R by
n
(x) = v (x) | |v| = 1
x Rn .
for any x, y B(0n , 2) with |x y| < , and any v (x), there is w (y) such
that |v w| < 1. Let N 2 be an integer such that the increasing sequence of balls
B1 , . . . , BN defined by
Bi = B (0n , i)
i = 1, . . . , N,
x Bi+1 ,
(1.1)
n , 1) Rn as follows:
and let w (0) be fixed. We define the vector field X : B(0
We first set
x B1 .
x Bi+1 .
By construction (by (1.1) and the definition of ), Xi Pi (x) belongs to Ox for
n , 1) and satisfies
any x Bi+1 . In conclusion, X = XN is smooth on B(0
0n = X(x) (x) for any x B(0n , 1). Repeating the construction on the annuli
B(0n , 2) \ B(0n , 1), B(0n , 3) \ B(0n , 2), . . ., we obtain a non-vanishing section of
on Rn .
We now prove Proposition 1.1 by induction on m. Let be a rank (m + 1)
distribution on Rn . By Lemma 1.2, it admits a non-vanishing section X on Rn . The
n
multivalued mapping : Rn 2R defined by
(x)
= (x) X(x)
x Rn ,
is a smooth rank m distribution (here {X(x)} denotes the space which is orthogonal to X(x) with respect to the Euclidean scalar product). Thus by induction,
there are smooth vector fields X 1 , . . . , X m which generate on Rn . The family
{X 1 , . . . , X m , X} is a global frame for .
1 Sub-Riemannian Structures
y Vx .
Since M is paracompact, there is a locally finite covering V = {Vi }iI where each
open set Vi equals Vxi for some xi M.
Lemma 1.4 There are a locally finite open covering {Uj }jJ of M and a partition
n+1
l=1 Jl of J such that the following properties are satisfied:
(a) For every j J, there is i = i(j) I such that Uj Vi .
(b) For every l {1, . . . , n + 1} and any j = j Jl , Uj Uj = .
Proof (Proof of Lemma 1.4) Recall that every smooth manifold is triangulable. Let
T = {Tt }tT be a triangulation of M that refines the covering {Vi }iI , in the sense
that the closure of each face F of T is a subset of some Vi . For every {0, . . . , n},
denote by T = {Tt }tT the family of -dimensional faces in T . For every
{0, . . . , n}, we can construct easily a collection of open sets W = {Ws }sS
satisfying the following properties:
For that, it suffices to proceed by induction on and to make use of the properties
of a triangulation. We conclude easily.
Let us now show how to construct for every r {1, . . . , m} a family of sections
j
j
j
{X1 , . . . , Xn+1 | 1 j r} of such that Span{Xl (x) | 1 j r, 1 l n + 1}
has dimension r for any x M. We proceed by induction on r.
j
Span Xl (x) | 1 j r, 1 l n + 1
has dimension r for any x M (with r < m). For every j J, there is s = s(j)
{1, . . . , m} such that
j
Span Xxsi (j) (x), Xl (x) | 1 j r, 1 l n + 1
r+1
by
has dimension r + 1 for any x Uj . Define X1r+1 , . . . , Xn+1
Xlr+1 =
s(j)
Xi(j)
l = 1, . . . , n + 1.
jJl
s(j)
We leave the reader to check that by construction (modifying the Xxi (j) s if necessary
as above), the vector space
j
Span Xl (x) | 1 j r + 1, 1 l n + 1
has dimension r + 1 for any x M. The proof is complete.
The Hrmander condition. Recall that for any smooth vector fields X, Y on M
given by
X(x) =
n
i=1
ai (x)xi , Y (x) =
n
bi (x)xi ,
i=1
in local coordinates x = (x1 , . . . , xn ), the Lie bracket [X,Y ] is the smooth vector field
defined as
1 Sub-Riemannian Structures
Fig. 1.1 The Lie bracket [X, Y ](x) measures the extent to which X and Y do not commute from x
[X, Y ](x) =
n
ci (x)xi ,
i=1
n
xj bi aj xj ai bj
i = 1, . . . , n.
j=1
For the upcoming controllability results, it is important to keep in mind the following
dynamical characterization of the Lie bracket (see Fig. 1.1).
Proposition 1.5 Let X, Y be two smooth vector fields in an neighborhood of x Rn .
Then we have
[X, Y ](x) = Dx Y X(x) Dx X Y (x) = lim
etY etX etY etX (x) x
t2
t0
(1.2)
t.
= Y (h4 (t)) +
etY
x
(t,h3 (t))
h3 (t)
where h3 is defined by h3 (t) := etX etY etX (x). Then we have
h3 (t)
= X(h3 (t)) +
etX
x
(t,h2 (t))
h2 (t),
where h2 (t) := etY etX (x) and
h2 (t) = Y (h2 (t)) +
etY
x
(t,h1 (t))
h1 (t),
with h1 (t) := etX (x) and h1 (t) = X(etX (x)). Since partial derivatives of the form ex
at t = 0 are equal to Id, we get h1 (0) = X(x), h2 (0) = X(x) + Y (x), h3 (0) = Y (x)
and h4 (0) = 0. Therefore, the left-hand side of (1.2) is equal to 21 h4 (0). By derivating
the above formulas, we get
tX
h1 (0) = Dh1 (0) X h1 (0) = Dx X X(x),
tY
e
h2 (0) = Dh2 (0) Y h2 (0) + dtd
x
h1 (t)
(t,h1 (t))
.
t=0
etY
h (t)
x (t,h1 (t)) 1
t=0
tY
d etY
e
=
h1 (0) +
h (0)
dt x (t,h1 (t))
x (0,h1 (0)) 1
t=0
2 tY
2
tY
e
e
+
h (0) X(x) + Dx X X(x)
=
tx (0,x)
x 2 (0,x) 1
tY
e
=
X(x) + Dx X X(x) = Dx Y X(x) + Dx X X(x).
x
t
(0,x)
d
dt
We infer that h2 (0) = Dx Y (2X(x) + Y (x)) + Dx X X(x). In the same way, we have
h3 (0)
= Dh3 (0) X
h3 (0) +
d
dt
etX
x
(t,h2 (t))
h2 (t)
,
t=0
d
dt
etX
x
(t,h2 (t))
h2 (t)
d
h4 (0) = Dh4 (0) Y h4 (0) +
dt
etY
x
(t,h3 (t))
h3 (t)
t=0
1 Sub-Riemannian Structures
etY
x
etY
h3 (0)
x
(t,h3 (t)) t=0
(0,h3 (0))
= Dx Y Y (x) 2Dx X Y (x) + Dx Y 2X(x) + Y (x)
= 2 Dx Y X(x) Dx X Y (x) = 2[X, Y ](x),
d
dt
h3 (0) +
Remark 1.1 The Lie bracket is bilinear, skew-symmetric and satisfies the Jacobi
identity, that is given three smooth vector fields X, Y , Z, we have
X, [Y , Z] + Y , [Z, X] + Z, [X, Y ] = 0.
Remark 1.2 Given a smooth diffeomorphism from a smooth manifold U to a
smooth manifold V and X a smooth vector field on U , we recall that the pushforward (X) of X is defined by
(X)(y) := D 1 (y) X( 1 (y)
y V .
k 1.
Liek (F ).
k1
x Rn .
The non-zero Lie brackets of X and Y are always constant vector fields of the form
ad0X (Y ) := Y = b and adXk+1 (Y ) := X, adkX (Y ) = (1)k+1 Ak+1 b
k 0.
n1
i Ai b
x Rn ,
i=0
x R2 ,
where f is a smooth scalar function. Then, Lie(X, Y ) is the space of smooth vector
fields spanned by X and
adkY (X) = f (k) x2
for k 0.
Thus, Lie(X, Y ) is infinite-dimensional whenever the derivatives of f span an infinitedimensional space of functions.
For any point x M, Lie(F )(x) denotes the set of all tangent vectors X(x) with
X Lie(F ). It follows that Lie(F )(x) is always a linear subspace of Tx M, hence
finite-dimensional.
Example 1.7 Returning to Example 1.6 and denoting by (e1 , e2 ) the canonical basis
of R2 , we check that
x R2 .
Lie(X, Y )(x) = Span e1 , f (k) (x1 )e2 | k = 0, 1, 2, . . .
In particular, Lie(X, Y )(x) = Re1 if f (x) and all its derivatives at x vanish and
Lie(X, Y )(x) = R2 otherwise.
We say that the smooth vector fields X 1 , . . . , X m satisfy the Hrmander condition
on some open set O M if and only if
Lie X 1 , . . . , X m (x) = Tx M
x O.
10
1 Sub-Riemannian Structures
x O.
Proof It is sufficient to show that the left-hand side is included in the right-hand side
for any integer k 2. Since the Y j (x) are always linearly independent, there are
j
smooth functions i : O R with i, j = 1, . . . , m, such that
X (x) =
i
m
i (x)Y j (x)
x O, i = 1, . . . , m.
j=1
Then for every i = 1, . . . , m and every smooth vector field Z, there holds
[X , Z] =
i
m
j=1
j
i Y j , Z
m
j
i [Y j , Z]
j=1
m
di (Z)Y j .
j=1
Since Span X 1 (x), . . . , X m (x) Span Y 1 (x), . . . , Y m (x) for any x, this shows
that
x O.
Lie2 X 1 , . . . , X m (x) Lie2 Y 1 , . . . , Y m (x)
We conclude easily by an inductive argument.
Remark 1.3 Since for any smooth vector field X, there holds [X, X] = 0, a one
dimensional distribution cannot be totally nonholonomic.
Example 1.8 The distribution given in Example 1.2 is totally nonholonomic. We
check easily that
i, j = 1, . . . , n,
[X, Y ] = z
which means that has degree 2 everywhere.
Example 1.9 More generally, the distribution given in Example 1.3 is totally nonholonomic of degree 2. We check easily that
11
[X i , Y j ] = ij z
i, j = 1, . . . , n.
Example 1.10 The Martinet distribution in R3 (with coordinates (x, y, z)) is the
distribution generated by X and Y with
X = x , Y = y +
x2
z .
2
2p
ai dxi + dxn ,
i=1
i = 1, . . . , 2p.
i = 1, . . . , 2p,
12
1 Sub-Riemannian Structures
defines a local frame for = Ker() in V . On the one hand, the n = 2p + 1-form
(d)p at x reads
(d)p x
! aj
a
il
l
l=1,...,p
(1.3)
where P2p denotes the set of p-tuples of the form = (i1 , j1 ), . . . , (ip , jp ) with
{i1 , j1 , . . . , ip , jp } = {1, . . . , 2p} and il < jl for all l = 1, . . . , p. On the other hand,
we check easily that
X i , X j (x ) = xi aj xj ai xn (x )
i, j = 1, . . . , 2p.
Therefore, if there is i {1, . . . , 2p} such that [X i , X j ](x ) = 0 for any j, then
all the products appearing in (1.3) vanish, which implies that ( (d)p )x = 0,
contradiction. We deduce that for every i {1, . . . , n}, there holds
Span X 1 (x ), . . . , X 2p (x ), X i , X 1 (x ), . . . , X i , X 2p (x ) = Tx M.
(1.4)
13
m
u (0) = x)
(1.5)
i=1
is one-to-one.
Proof The set of controls u L 2 ([0, T ]; Rm ) such that the solution u of (1.5) is
well-defined on [0, T ] is a non-empty open set. Moreover, by construction, any path
u is absolutely continuous with square integrable derivative and almost everywhere
tangent to . This proves that the map under study is well-defined on some open set
x,T
L 2 ([0, T ]; Rm ). Let ,x,T be such that there are u, v L 2 ([0, T ]; Rm )
UF
with
m
m
ui (t)X i ( (t)) =
vi (t)X i ( (t))
a.e. t [0, T ].
(t) =
i=1
i=1
Since the tangent vectors X 1 ( (t)) , . . . , X m ( (t)) are always linearly independent
in T (t) M, we infer that u(t) = v(t) for almost every t [0, T ], which proves that
x,T
our map is injective. Furthermore, given
, for almost every t [0, T ],
the
path
is
differentiable
at
t
and
there
is
a
unique
u(t) Rm such that (t) =
"m
i ( (t)). By construction, the function u : [0, T ] Rm belongs to
u
(t)X
i=1 i
L 2 ([0, T ]; Rm ).
Remark 1.4 If M is compact, then solutions to (1.5) are defined for any u
x,T
L 2 ([0, T ]; Rm ), which means that UF
= L 2 ([0, T ]; Rk ).
Given a family of smooth vector fields F = X 1 , . . . , X k on M and x M, T >
x,T
L 2 [0, T ]; Rk is called a control and the solution u :
0, a function u UF
[0, T ] M to the Cauchy problem
u (t) =
k
u (0) = x
(1.6)
i=1
is called the trajectory starting at x and associated with the control u. Since any horizontal path can be viewed as a trajectory associated to a control system like (1.6), we
restrict in the next paragraph our attention to End-Point mappings associated with
finite families of smooth vector fields.
End-Point mappings. Let F = X 1 , . . . , X k be a family of k 1 smooth vector
x,T
fields on M. As before, given x and T > 0, there is a maximal open subset UF
x,T
2
k
L [0, T ]; R such that for every u UF , there is a unique solution to the Cauchy
problem (1.6). The End-Point mapping associated to F at x in time T > 0 is defined
as follows,
x,T
x,T
: UF
M
EF
u u (T ).
14
1 Sub-Riemannian Structures
x,T
u the time-dependent vector field defined by
Given u UF
, we denote by XF
XF (t, x) :=
u
m
ui (t)X i (x)
a.e. t [0, T ], x M.
i=1
u (t, x) is well-defined and smooth on a neighbourhood of x; we denote by
Its flow F
u
Dx F (t, x) its differential at (t, x) with respect to the x variable. The following result
holds. (We refer the reader to Appendix A for reminders in differential equations and
to Appendix B for reminders in differential calculus in infinite dimension.)
x,T
x,T
Proposition 1.8 The End-Point mapping EF
is of class C 1 on UF
and for every
x,T
control u UF , its differential at u,
x,T
Du EF
: L 2 ([0, T ]; Rk ) TE x,T (u) M
F
is given by
#
x,T
Du EF (v) = Dx F (T , x)
T
u
Dx F
(t, x)
1
x,t
v
t, EF (u) dt
XF
(1.7)
u+v (T ) =
0
k
(ui (t) + vi (t)) X i u+v (t) dt,
(1.8)
i=1
with u+v (0) = x. For every i = 1, . . . , k and every t [0, T ], the Taylor expansion
of each X i at u (t) gives
X i u+v (t) = X i u (t) + Du (t) X i u+v (t) u (t) + |u+v (t) u (t)| o(1).
Setting x (t) := u+v (t) u (t) for any t, we may assume that x has size , then
(1.8) yields formally
x (T ) =
0
$ k
15
i=1
m
%
vi (t)X u (t) dt + o().
i
i=1
This suggests that the linear part in of the function t [0, T ] x (t) should be
solution to the Cauchy problem
(t) =
k
i
ui (t)Du (t) X
i=1
(t) +
k
vi (t)X u (t)
i
a.e. t [0, T ],
(1.9)
i=1
Su (t)1 Bu (t)v(t)dt.
x,T
is differentiable at u and (1.7) is satisfied. Using Gronwalls lemma again,
Thus EF
x,T
x,T
depends continuously on u on UF
.
we leave the reader to check that Du EF
x,T
Remark 1.5 If M = Rn , the derivative of EF
at u is given by
x,T
Du EF (v) = S(T )
S(t)1 B(t)v(t)dt,
(1.10)
16
1 Sub-Riemannian Structures
and where the matrices A(t) Mn (R), B(t) Mn,k (R) are defined by
A(t) :=
k
ui (t)JX i u (t)
a.e. t [0, T ]
(1.11)
i=1
x,t
(u) and JX i denotes the Jacobian matrix of X i at u (t)) and
(u (t) = EF
B(t) := X 1 (u (t)), . . . , X k (u (t))
t [0, T ].
(1.12)
x,T
, we set
Properties of End-Point mappings. Given u UF
x,T 2
k
Imx,T
F (u) := Du EF L ([0, T ]; R ) .
x,T
Defining y = EF
(u), we observe that Imx,T
F (u) is a vector space contained in Ty M,
x,T
x,T
hence of dimension n. We call rank of u UF
with respect to EF
, denoted by
x,T
x,T
rankF (u), the dimension of ImF (u).
Given u L 2 ([0, T ]; Rk ) and > 0, we define the controls u L 2 ([0; 1 T ];Rk )
and u L 2 [0, T ]; Rk by
a.e. t [0, 1 T ],
u (t) := u(t)
a.e. t [0, T ].
u u (t) =
u(t)
if 0 t T ;
u (t T ) if T < t T + T
a.e. t [0, T + T ].
x,T
x,T
Proposition 1.9 Let u UF
, u UF
and > 0 be fixed, then we have (we set
x,T
y := EF (u)):
1
x, T
x, T
(i) u UF
and rankx,T
(u ).
F (u) = rankF
y,T
y,T
x,T
(ii) u belongs to UF and rankF (u) = rankF (u).
x,T +T
+T
max rankx,T (u), ranky,T (u ) .
u
u
and rankx,T
(iii) u u UF
F
F
F
u, (t) = u (t) =
k
17
ui (t)X i (u (t)) =
k
i=1
i=1
1 T
x,
We infer that u belongs to UF
1 T
x,
Du EF
u (t)X i u, (t) .
and satisfies
v L 2 [0, T ]; Rk .
x,T
(v)
(v ) = Du EF
EFF
u = x
x,T
u UF
.
z,T
The mapping (z, v) EF
(v) is C 1 and its derivative with respect to the z variable
x,T
u (T , x)1 . Derivating the above equality at u
(u), u is given by Dx F
at y = EF
yields
x,T
(v) + Du EF
(Dx F )u (T , x)1 Du EF
y,T
v = 0
v L 2 [0, T ]; Rk .
E x,T (u),T
x,T +T
EF
(u u ) = EFF
(u ),
y,T
x,T
x,T
and u UF with y = EF
(u). Again, derivating yields
for any u UF
y,T
x,T +T
x,T
u
v v = Dx F
(T , y) Du EF
(v) + Du EF (v ),
Duu EF
(1.13)
As the next example shows, the inequality in Proposition 1.9 may be strict.
Example 1.13 Let F = {X 1 , X 2 } be the family of smooth vectors fields on R4 (with
coordinates x = (x1 , x2 , x3 , x4 ) and canonical basis (e1 , e2 , e3 , e4 ) ) defined by
X 1 = x1 and X 2 = x2 + x12 x3 + x1 x2 x4 .
Set x = (1, 0, 0, 0), y = (0, 0, 0, 0), and define the controls u, u L 2 [0, 1]; R2
by
u(t) = (1, 0) and u (t) = (0, 1)
t [0, 1].
x,1
. As a matter of fact, the trajectory
The control u has rank 3 with respect to EF
4
u : [0, 1] R starting at x and associated with u equals u (t) = (1 + t, 0, 0, 0)
for any t [0, 1] and using the representation formula given in Remark 1.5, we have
18
1 Sub-Riemannian Structures
#
Du EF (v) =
v L 2 [0, 1]; R2 ,
x,1
B(t)v(t)dt
0
where B(t) = X 1 (u (t)), X 2 (u (t)) for any t [0, 1]. Then,
&#
x,1
'
v1 (t)dt e1 | v1 L ([0, 1]; R)
2
&#
+ Span
v2 (t)dt e2 +
'
(1 t) v2 (t)dt e3 | v2 L ([0, 1]; R)
2
= Span {e1 , e2 , e3 } .
The trajectory u : [0, 1] R4 starting at y and associated with u equals u (t) =
(0, t, 0, 0) for any t [0, 1], and there holds
#
y,1
S(t)1 B(t)v(t)dt,
1 000
1
0
0 1 0 0
S(t) =
0 0 1 0 and B(t) = 0
t2
0
2 0 0 1
0
1
0
0
t [0, 1].
We infer that
y,1
ImF
(u ) = Span
*#
1
+ Span
+
v2 (t)dt e2 | v2 L 2 ([0, 1]; R)
*#
1
0
v1 (t)dt e1 +
# 1$
0
t2
1
2
%
v1 (t)dt e4 | v1 L 2 ([0, 1]; R)
= Span {e1 , e2 , e4 } .
u
(1, y)(e3 ) = e3 .
Dx F
4
Therefore, by (1.13), this implies Imx,2
F (u u ) = R , which means that the rank
y,1
x,1
rankx,2
F u u = 4 is strictly larger than the maximum of rankF (u) and rankF (u )
which is equal to 3.
The following proposition implies that the rank of a control is always larger or
equal than the dimension of the family {X 1 , . . . , X k } at the end-point.
19
x,T
Proposition 1.10 We have for every u UF
,
x,T
(u) Imx,T
X i EF
F (u)
i = 1, . . . , k.
Proof Let us first assume that we work in Rn . In this case (see Remark 1.5), the
x,T
at u is given by
derivative of EF
#
x,T
Du EF
(v) = S(T )
S(t)1 B(t)v(t)dt
v L 2 [0, T ]; Rk ,
where S() is the solution to the Cauchy problem (1.10) and where the matrices
A(t) Mn (R), B(t) Mn,k (R) are defined respectively by (1.11) and (1.12). Fix
k
of the
i {1, . . . , k} and denote by ei the i-th vector
canonical basis in R . Define,
2
k
for every (0, T ), the control v L [0, T ]; R by
&
v (t) =
0
if 0 t T ;
(1/)ei if T < t T .
We have
,
,
,
,
x,T
,Du EF (v ) X i (u (T )),
,
# T
#
,
1 i
,
= ,(1/)S(T )
S(t) X (u (t) dt (1/)S(T )
T
T
(1/) |S(T )|
#
(1/) |S(T )|
+ (1/) |S(T )|
T
T
T
# T
T
T
S(T )
,
,
X (u (T )) dt ,,
1 i
,
,
,
,
,S(t)1 X i (u (t)) S(T )1 X i (u (T )), dt
-,
,
-,
,
-S(t)1 - ,X i (u (t)) X i (u (T )), dt
,
-,
,
-,
-S(t)1 S(T )1 - ,X i (u (T )), dt.
x,T 2
k
k
Since Imx,T
F (u) = Du EF L ([0, T ]; R ) is a closed subset of R , we infer that
X i (u (T )) belongs to Imx,T
F (u).
20
1 Sub-Riemannian Structures
If we are now in M, then there exists a local chart around x and t (0, T ) such that
u (t ) O. Set T := T t and define u1 L 2 ([0, t ]; Rk ) and u2 : L 2 ([0, T ]; Rk ) by
u1 (t) = u(t)
t [0, T ].
x,
and F if and only if any control of the form u UF
(with = 0) is singular
y,T
x,T
with respect to x and F and if and only if u UF (with y = EF
) is singular
with respect to y and F . It also shows that if the concatenation of several controls is
singular then each of them is singular.
h i (x, p) =
hi
hi
(x, p),
(x, p) .
p
x
21
x,T
Proposition 1.11 The control u UF
is singular with respect to x and F if and
only if there exists an absolutely continuous arc : [0, T ] T M that never
intersects the zero section of T M, such that
k
ui (t) h i ((t))
a.e. t [0, T ]
(1.14)
hi ((t)) = 0, t [0, T ]
i = 1, . . . , k.
(1.15)
(t)
=
i=1
and
v L 2 [0, T ]; Rk .
v L 2 ([0, T ]; Rk ).
t [0, T ],
we deduce that
#
0
,
,2
,
,
, pS(T )S(t)1 B(t) , ds = 0,
which implies that pS(T )S(t)1 B(t) = 0 for any t [0, T ] (note that the function
t pS(T )S(t)1 B(t) is continuous). Let us now define, for each t [0, T ],
p(t) := pS(T )S(t)1 .
By construction, p : [0, T ] (Rn ) is an absolutely continuous arc. Since p = 0
and S(t) is invertible for all t [0, T ], p(t) does not vanish on [0, T ]. Moreover,
recalling that, by definition of S,
d
S(t)1 = S(t)1 A(t)
dt
a.e. t [0, T ],
22
1 Sub-Riemannian Structures
a.e. t [0, T ]
and
p(t)B(t) = 0
t [0, T ]
which shows that (1.14)(1.15) are satisfied with (t) = (u (t), p(t)) for any t
[0, T ). By the way, we note that by construction, we have for every t (0, T ],
x,t
(v) = 0
p(t) Dut EF
v L 2 [0, t]; Rk ,
(1.16)
t [0, T ].
t [0, T ],
This concludes the proof. Again, the above proof shows indeed that (1.16) holds for
any t (0, T ].
Assume now that we work on M. We can cut the path u : [0, T ] M associated
with u and u itself into a finite number of pieces 1 , . . . , l and u1 , . . . , ul such
that each control ul is singular and each path l is valued in a chart of M. Then
we can apply the previous arguments on each chart and thanks to (1.16) obtain a
non-vanishing absolutely continuous arc satisfying (1.14)(1.15) on [0, T ].
Remark 1.7 We keep in mind that if : [0, T ] T M is an absolutely continuous
arc satisfying (1.14)(1.15), then
23
v L 2 [0, t]; Rk ,
where (t) = (u (t), p(t)) and ut denotes the restriction of u to [0, t]. (In the sequel,
v or p v with = (x, p) in local coordinates denotes the evaluation of the form
at v Tx M.)
Remark 1.8 In local coordinates, Proposition
1.11 means that there exists an
absolutely continuous arc p : [0, T ] Rn \ {0} satisfying
p (t) =
k
a.e. t [0, T ]
(1.17)
t [0, T ], i = 1, . . . k.
(1.18)
i=1
and
p(t) X i u (t) = 0
x,T
is any control such that = u is solution to the Cauchy problem
where u UF
x,T
(1.6). We call rank of
, denoted by rank ( ), the dimension of Im (u).
We shall say that is singular (with respect to ) if rank ( ) < n and regular
otherwise.
Remark 1.9 By Remark 1.6, if has rank m = n then any non-trivial horizontal
path is regular.
Proposition 1.9 does apply to horizontal paths. The rank of an horizontal path
depends only on the curve drawn by the path in M, it does not depend upon its
parametrization. Moreover if an horizontal path which is the concatenation of several
paths (the concatenation of paths is defined in the same way as the concatenation of
controls, see Fig. 1.2) is singular then each piece is necessarily singular.
Example 1.14 Returning to Examples 1.2 and 1.8, we consider in R3 with coordinates x = (x1 , x2 , x3 ), the totally nonholonomic rank two distribution generated by
24
1 Sub-Riemannian Structures
X 1 = x1
x2
x
2 3
and
X 2 = x2 +
x1
x .
2 3
We claim that the singular horizontal paths are the constant curves or equivalently
that the only singular control with respect to F = {X 1 , X 2 } is the control u 0. Let
x,T
be a singular control.
us prove this claim. Let x R3 , T > 0 be fixed and u UF
3
Denote by x : [0, T ] R the solution to the Cauchy problem
x (t) = u1 (t)X 1 (x(t)) + u2 (t)X 2 (x(t)) a.e. t [0, T ], x(0) = x.
(1.19)
From Proposition 1.11, there exists an absolutely continuous arc p : [0, T ] R3 \
{0} such that
p (t) = u1 (t) p(t) Dx(t) X 1 u2 (t) p(t) Dx(t) X 2
(1.20)
t [0, T ].
(1.21)
a.e. t [0, T ], i = 1, 2
a.e. t [0, T ].
X, (x) := [X, Z](x) | Z section of .
(1.22)
25
The condition above being very restrictive, there are very few fat distributions. Fat
distributions on three-dimensional manifolds are the rank-two distributions satisfying
x V ,
Tx M = Span X 1 (x), X 2 (x), X 1 , X 2 (x)
where (X 1 , X 2 ) is a local frame for in V . Another example of co-rank one fat distributions in odd dimension is given by contact distributions which were introduced
in Example 1.12. In this case property (1.22) is an easy consequence of (1.4). Let us
now prove that fat distributions do not admit non-trivial singular horizontal paths.
By the property of singular concatenated horizontal paths, we just need to show
that non-constant short horizontal paths cannot be singular. Taking a local chart if
necessary we can work in Rn and assume that has a local frame X 1 , . . . , X m . Let
x,T
be a singular control. By Remark 1.8, there
x Rn , T > 0 be fixed and u UF
exists an absolutely continuous arc p : [0, T ] Rn \ {0} satisfying (1.17) and
(1.18). For almost every fixed t [0, T ] and every i = 1, . . . , m, derivating (1.18)
yields
m
j=1
m
uj (t) p(t) X j , X i u (t) = p(t)
uj (t)X j , X i u (t) = 0.
j=1
"
Setting the autonomous vector field X() := m
u (t)X j (), we deduce that p(t)
j=1
j
i
i
annihilates all the X u (t) s and all the X, X u (t) s. This contradicts (1.22).
Example 1.16 Returning to Example 1.10 (Martinet distribution), we consider in R3
with coordinates x = (x1 , x2 , x3 ), the totally nonholonomic rank two distribution
generated by
x2
X 1 = x1 and X 2 = x2 + 1 x3 .
2
We claim that the singular horizontal curves are exactly the traces of the distribution
on the so-called Martinet surface (see Fig. 1.3)
:= x R3 | x1 = 0 ,
which in other terms means that the singular horizontal paths are either constant
curves or are contained in a line lz of the form
lz = x = (x1 , x2 , x3 ) R3 | x1 = 0 and x3 = z
for some z R.
x,T
be a non-trivial
Let us prove this claim. Let x R3 , T > 0 be fixed and u UF
3
singular control. Denote by x : [0, T ] R the solution to the Cauchy problem
26
1 Sub-Riemannian Structures
Fig. 1.3 The singular horizontal curves are the traces of the distribution on the Martinet surface
(1.23)
t [0, T ].
(1.24)
We deduce that
2
|u(t)|2 p(t) [X 1 , X 2 ](x(t)) = 0
a.e. t [0, T ].
Since the three vectors X 1 (x), X 2 (x), [X 1 , X 2 ](x) span R3 for every x with x1 = 0,
this shows that x1 (t) = 0 for all t [0, T ], which in turn implies that u1 0. We
deduce that x has the form
# t
u2 (s)ds, 0, x3 (0) ,
x(t) = 0, x2 (0) +
0
x,T
which shows that it is contained in lx3 (0) . Conversely, if an horizontal path x
has the form
t [0, T ]
with z R, then any absolutely continuous arc p : [0, T ] R3 \ {0} of the form
p(t) = (0, 0, p3 )
t [0, T ]
27
with p3 = 0 satisfies (1.23) and (1.24). This shows that any horizontal path which is
contained in a line lz for some z R is singular.
Example 1.17 More generally, consider a totally nonholonomic distribution of
rank two in a manifold M of dimension three. We define the Martinet surface of
as the set defined by
:= x M | (x) + [, ](x) = Tx M ,
where
, (x) := [X, Y ](x) | X, Y sections of .
a.e. t [0, ].
i=1,2
t [0, ], i = 1, 2.
28
1 Sub-Riemannian Structures
But since (t) for every t, there are two continuous functions 1 , 2 : [0, ]
R such that
X 1 , X 2 ( (t)) = 1 (t)X 1 ( (t)) + 2 (t)X 2 ( (t))
t [0, ].
This implies that the pair (h1 , h2 ) is a solution of the linear differential system
X 2 , [X 1 , X 2 ] = x4 .
x 1 (t) =
x 2 (t) =
x 3 (t) =
x 4 (t) =
u1 (t)
u2 (t)
u2 (t)x1 (t)
u2 (t)x3 (t),
p 2 (t) = 0
p 3 (t) = u2 (t)p4 (t)
p 4 (t) = 0,
t [0, T ].
(1.25)
29
System (1.25) implies that p2 and p4 are constant on [0, T ]. If p4 = 0, then (1.25)
also implies that p3 is constant on [0, T ]. Hence we obtain that p2 + x1 (t)p3 = 0 for
every t [0, T ]. Which means that either x1 is constant or p2 = p3 = 0. Since p
does not vanish on [0, T ], we deduce that x1 is constant, which means that u1 0.
But u2 (t)p3 = 0 for almost every t, hence p3 = 0 (remember that |u(t)| = 1 a.e.
t [0, T ]). We obtain a contradiction. Therefore, p4 = 0, hence we deduce easily that
p 3 (t)
p3 (t) = 0
a.e. t [0, T ].
0 = u2 (t)p3 (t) =
p4
Since p3 is absolutely continuous, this means that it is constant on [0, T ]. This implies
that u2 (t) = 0 for all t [0, T ]. Then, the curve x has the form
x(t) = x1 (t), x2 (0), x3 (0), x4 (0)
t [0, T ].
x M,
where
, [, ] (x) := X, [Y , Z] (x) | X, Y , Z sections of .
As above, we can work locally, so let us consider a frame {X 1 , X 2 } and a trajectory
2
2
x : [0, T ] R4 associated to some control
4 u L ([0, T ]; R ). If x is singular
\ {0} satisfying (1.17) and (1.18).
(with respect to ), there is p : [0, T ] R
Derivativing (1.18) two times yields for almost every t [0, T ] with u(t) = 0,
p(t) X 1 , X 2 x(t) = 0
and
u1 (t) p(t) X 1 , X 1 , X 2 x(t) + u2 (t) p(t) X 2 , X 1 , X 2 x(t) = 0.
(1.26)
(1.27)
Since M has dimension four and + [, ]] has dimension three, there is (locally)
a smooth non-vanishing 1-form whose kernel is equal to + [, ]. Then, by
(1.18) and (1.26) and p(t) are colinear along x, and in turn by (1.27) we have for
almost every t [0, T ] with u(t) = 0,
30
1 Sub-Riemannian Structures
u1 (t) x(t) X 1 , X 1 , X 2 x(t) + u2 (t) x(t) X 2 , X 1 , X 2 x(t) = 0.
By the above assumptions, for every x, the linear form
(1 , 2 ) R2 x X 1 , X 1 , X 2 (x) 1 + x X 2 , X 1 , X 2 (x) 2
has a kernel of dimension one. This shows that there is a smooth line field (a distribution of rank one) L on M such that the singular curves are exactly the integral
curves of L.
U
s.t.
u
2 < .
L
F
F
By Proposition 1.9 (iii), the function (0, +) d() is nondecreasing with
values in N. So, there is 0 and d0 N such that d() = d0 for any (0, 0 ).
Since F satisfies the Hrmander condition at x, the vector space spanned by
{X 1 (x), . . . , X k (x)} has dimension 1. Then, thanks to Proposition 1.10, there holds
d() = d0 1
[0, 0 ].
x,
L : Rd0 Tx M
x, "d0
j vj
= 1 , . . . , d0 Du EF
j=1
31
E : V M
" 0 j j
x,
EF
u + dj=1
v
u +
Imx,
F
d0
j vj = D E Rd0 = TE () N.
j=1
x S .
x S and t small.
From the above lemma it follows that all the brackets involving X 1 , . . . , X k at
y N belong to Ty N. Since F satisfies the Hrmander condition, this shows that
d0 = n and indeed that d() = n for any > 0.
x,T
, v O and > 0 to be chosen later. Since
Let O be an open subset of UF
x,
d() = n, there is u UF such that uL2 < and rankx,
F (u) = n. Define the
control v L 2 ([0, T ]; Rk ) by
v = u u v
T
T 2
The trajectory associated with v is the concatenation of the curve xu starting at x and
associated with u, xu starting at xu () and associated with u , and a reparametrization
of xv starting at x and associated with v (see Fig. 1.4).
x,T
, is regular and satisfies
By construction (see Proposition 1.9), v belongs to UF
x,T
x,T
EF
(v) = EF
(v). Then, as above, the image of a small ball centered at the origin
x,T
in L 2 ([0, T ]; Rk ) by the mapping w EF
(v + w) is an open neighboorhood of
-2
x,T
x,T
E (v) = E (v). Furthermore, the quantity -v v- 2 is equal to
F
32
1 Sub-Riemannian Structures
#
|u(t) v(t)|2 dt +
|u(2 t) v(t)|2 dt
,
,2
%
$
,
,
T t 2
T
,
,
v
v(t), dt
,
, T 2
,
T 2
|u(t)|2 dt +
# 2
0
|v(t)|2 dt 2
#
0
u(t), v(t)
dt + 2
# 2
,2
%
,2
2 # T ,, $
# T ,
,
,
,
T t 2
T
T
,
,
,
,
+
u(t), dt
,u
, T 2 u(t) u(t), dt + T 2
,
T 2
2
2 ,
0
%
# T / $
T t 2
T
T
u(t) u(t) dt.
u
+2
u(t),
T 2
T 2
T 2
2
T
u(t) u(t)
T 2
and
t [2, T ]
$
%
T t 2
u
u(t)
T 2
tend to zero in L 2 , we infer that v belong to O if is small enough. This shows that
x,T
x,T
x,T
(O) contains a neighborhood of EF
(v) = EF
(v).
EF
Statement and proof. The aim of the present section is to prove the following result.
Theorem 1.14 (Chow-Rashevskys Theorem) Let be a totally nonholonomic
distribution on M (assumed to be connected). Then, for every x, y M and every
x,T
T > 0, there is an horizontal path
such that (T ) = y.
Thanks to the above discussion, the Chow-Rashevsky Theorem will be a straightforward consequence of the following result.
Theorem 1.15 Let F = X 1 , . . . , X k be a family of smooth vector fields on M.
Assume that M is connected and that F satisfies the Hrmander condition on M.
x,T
such that the
Then for every x, y M and every T > 0, there is a control u UF
solution of (1.6) satisfies u (T ) = y.
33
Proof Let x and T > 0 be fixed. Denote by AF (x, T ) the set of points in M which
x,T
, that is
can be joined from x by a control in UF
x,T
x,T
UF
.
AF (x, T ) = EF
By Proposition 1.12, AF (x, T ) is an open set in M. Let us show that this set is
closed as well. Let {zk }k be a sequence of points in M converging to some
z M. By
z,1
z,1
z,1
z,1
is an
openness of the mapping EF and the fact that EF (0) = z, the set EF UF
neighborhood of z. Then, there is k large enough such that zk belongs to that set.
The concatenation of uk together with u steers x to z (see Fig. 1.5). This shows
that AF (x, T ) is closed in M. In conclusion AF (x, T ) is open, closed and nonempty
(it contains x). By connectedness of M, we infer that AF (x, T ) = M.
Remark 1.10 The Chow-Rashevsky may be of course obtained in different ways. For
instance, consider in R3 a totally nonholonomic rank two distribution generated
by two smooth vector fields X 1 , X 2 such that
Span X 1 (x), X 2 (x), X 1 , X 2 (x) = R3
x R3 .
34
1 Sub-Riemannian Structures
Example 1.20 The space R3 (with coordinates (x, y, z)) equipped with the rank
two distribution given in Example 1.2 and with the metric g = dx 2 + dy2 is
the most simple sub-Riemannian structure we can imagine. The length of a vector
v = (v1 , v2 , v3 ) (x, y, z) is given by
|v|g =
v12 + v22 .
Since v is an horizontal vector, the latter quantity does not vanishes unless v = 0.
1
m
If the distribution
admits
a frame X , . . . , X on an open set O M, then the
family F = X 1 , . . . , X m is called an orthonormal dgenerating family of vector
fields or an orthonormal frame for (, g) in O if there holds
gx X i (x), X j (x) = ij
i, j = 1, . . . , m, x O,
gx (v, v)
v (x).
x,T
is defined by
The length of an horizontal path
#
lengthg ( ) :=
0
,
, (t),g dt.
(t)
T,
Note that since any horizontal path is absolutely continuous with square integrable
derivative, the length of any horizontal path is finite.
Let (, g) be a sub-Riemannian structure on M, by the Chow-Rashevsky Theorem, for every x, y M, there is at least one horizontal path joining x to y in time
1. For every x, y M, the sub-Riemannian distance between x and y, denoted by
dSR (x, y), is defined as the infimum of lengths of horizontal paths joining x to y, that
is,
x,1
dSR (x, y) := inf lengthg ( ) |
s.t. (1) = y .
The function dSR defines a distance on M M (the triangular inequality is easy, the
fact that dSR (x, y) x = y follows from the proof of Proposition 1.11) and makes
M a metric space. Given x M and r 0, we call sub-Riemannian ball centered at
x with radius r the set defined as
BSR (x, r) = y M | dSR (x, y) < r .
35
The openness of End-Point mappings associated with totally nonholonomic distribution yields the following result.
Proposition 1.16 Let (, g) be a sub-Riemannian structure on M, then the topology
defined by dSR coincides with the original topology of M. In particular, the subRiemannian distance dSR is continuous on M M.
Proof We need to show that for every x M, the family of sub-Riemannian balls
{BSR (x, r)}r>0 is a basis of neighborhoods for x with respect to the original topology.
Let F = {X 1 , . . . , X m } be an orthonormal frame for on an open neighborhood
Vx of some x M Let V Vx be an open and relatively compact neighborhood of
x with respect to the initial topology. Let us show that there is r > 0 small enough
such that BSR (x, r) V . Let W be an open neighborhood of x such that W V .
Define the compact annulus A by
A =V \W.
Any continuous path joining x to a point outside V has to cross A . Hence since
X 1 , . . . , X m are bounded on A , there is > 0 such that any solution u : [0, 1] M
to the Cauchy problem
u (t) =
m
ui (t)X i (u (t))
i=1
x,1
and u (1)
/ V satisfies
with u UF
#
0
,
,
,
,
m
1 ,
i=1
,g
,
,
ui (t)X i (u (t)),
,
dt > .
u (t)
By Proposition 1.7, this means that the sub-Riemannian ball BSR (x, /2) is included in
V . Let us now show that any sub-Riemannian ball BSR (x, r) contains an open neigh-
x,1
borhood of x with respect to the initial topology. The set UF
is open in L 2 [0, 1]; Rm
and contains the control u 0. Thus there is > 0 such that the L 2 -ball BL2 (0, ) is
x,1
. Moreover, since F is orthonormal with respect to g, there holds
contained in UF
for every u BL2 (0, )
# 1
|u(t)|dt.
lengthg (u ) =
0
36
1 Sub-Riemannian Structures
References
1. Agrachev, A.A., Sachkov, Yu.L.: Control Theory from the Geometric Viewpoint. Encyclopaedia of Mathematical Sciences, vol. 87. Springer-Verlag, Heidelberg (2004)
2. Bellaiche, A.: The tangent space in sub-Riemannian geometry. In: Sub-Riemannian Geometry,
pp. 178. Birkhuser, Basel (1996)
3. Bismut, J.-M.: Large Deviations and the Malliavin Calculus. Progress in Mathematics, vol. 45.
Birkhuser, Boston (1984)
4. Chow, C.-L.: ber systeme von linearen partiellen differentialgleichungen ester ordnung. Math.
Ann. 117, 98105 (1939)
5. Gromov, M.: Carnot-Carathodory spaces seen from within. In: Sub-Riemannian Geometry,
pp. 79323. Birkhuser, Basel (1996)
6. Hrmander, L.: Hypoelliptic second order differential operators. Acta. Math. 119, 147171
(1967)
7. Jurdjevic, V.: Geometric Control Theory. Cambridge Studies in Advanced Mathematics, vol.
52. Cambridge University Press, Cambridge (1997)
8. Montgomery, R.: A tour of subriemannian geometries, their geodesics and applications. In:
Mathematical Surveys and Monographs, vol. 91. American Mathematical Society, Providence,
RI (2002)
9. Rashevsky, P.K.: About connecting two points of a completely nonholonomic space by admissible curve. Uch. Zapiski Ped. Inst. Libknechta 2, 8394 (1938)
10. Sussmann, H.J.: A cornucopia of four-dimensional abnormal sub-Riemannian minimizers. In:
Sub-Riemannian Geometry, pp. 341364. Birkhuser, Basel (1996)
11. Sussmann, H.J.: Smooth distributions are globally finitely spanned. In: Analysis and Design
of Nonlinear Control Systems, pp. 38. Springer-Verlag, Heidelberg (2008)
Chapter 2
Sub-Riemannian Geodesics
Throughout all the chapter, M denotes a smooth connected manifold without boundary of dimension n 2 equipped with a sub-Riemannian structure (, g) of rank
m n.
energy ( ) :=
g
1
2
(t)g
dt.
(t)
37
38
2 Sub-Riemannian Geodesics
Proof Let x, y M be fixed. First, we observe that, for every horizontal path :
[0, 1] M satisfying (0) = x and (1) = y, the Cauchy-Schwarz inequality yields
1
0
| (t)| (t) dt
1
| (t)| (t)
2
dt.
(2.1)
x,1
such that (1) = y yields dSR (x, y)2
Taking the infimum over the set of
x,1
eSR (x, y). On the other hand, for every > 0, there exists an horizontal path
,
with (1) = y, such that
length ( ) =
g
x,1
with (1) = y
Reparametrizing by arc-length, we get a new path
satisfying
(t)g
(t)
= lengthg ( )
Consequently,
1
eSR (x, y)
0
2
(t)g
dt = lengthg ( )2 (dSR (x, y) + )2 .
(t)
x,1
Given x, y M, we call minimizing geodesic between x and y any path
joining x to y such that
eSR (x, y) = energyg ( ).
Thanks to the above proof and the fact that equality holds in the Cauchy-Schwarz
g
inequality (2.1) if and only has constant speed (that is | (t)| (t) is constant), we
obtain the following result.
x,1
Proposition 2.2 Given x, y M, a path
is a minimizing geodesic between
x and y if and only if it is a minimizing horizontal path between x and y with constant
speed.
k+
lim k (1) = z,
k+
lim lengthg k = dSR (y, z), (2.2)
k+
39
r
2r
and lengthg ( ) ,
dSR (0), y <
24
3
we have for every t [0, 1],
ui (t)X i (t)
and
i=1
for a.e. t [0, 1]. Let y, z BSR (x, r/4) be fixed and { k }k : [0, 1] M be
a sequence of horizontal paths with constant speed verifying (2.2). By the above
discussion, we may assume without loss of generality that all the paths k : [0, 1]
M are valued in the compact set Vx with
derivatives
bounded by r and associated
with a sequence of controls {uk }k in L 2 [0, 1]; Rm such that uk L2 = lengthg ( k ).
Then by Arzela-Ascolis theorem taking a subsequence if necessary the sequence
k
{ k }k converges to some : [0, 1] M. Moreover, the sequence {u
}k is bounded
2
2
in L so it weakly converges up to a subsequence to some v L [0, 1]; Rm . We
obtain easily that (0) = y, (1) = z,
(t) =
v i (t)X i (t)
i=1
40
2 Sub-Riemannian Geodesics
energyg = v2L2 = dSR (y, z)2 = eSR (y, z).
Which shows that is a minimizing geodesic joining y to z.
Remark 2.1 The above proof shows indeed that up to taking a subsequence, the
sequence {uk }k converges strongly to v in L 2 ([0, 1]; Rm ). As a matter of fact, it
converges weakly to v and satisfies
lim uk
k+
L2
= vL2 .
The SR Hopf-Rinow Theorem. The following sub-Riemannian version of the classical Riemannian Hopf-Rinow Theorem holds.
Theorem 2.4 (Hopf-Rinow Theorem) Let (, g) be a sub-Riemannian structure
on M. Assume that (M, dSR ) is a complete metric space. Then the following properties
hold:
(i) The balls B SR (x, r) are compact (fort any r 0).
(ii) For every x, y M there exists at least one minimizing geodesic joining x to y.
Proof Let us first recall that thanks to Proposition 1.13, the metric space (M, dSR ) is
locally compact. That is for every x M, there is r > 0 such that the ball B SR (x, r) is
compact. Let x M be fixed. We first show that all the balls B SR (x, r) with r 0 are
compact. Denote by Ix the set of r 0 such that B SR (x, r) is compact. By inclusion
of the balls B SR (x, r ) B SR (x, r) if r r and local compactness of (M, dSR ), Ix
is an interval whose supremum Rx is strictly positive. We claim that I is both closed
and open in [0, +).
Lemma 2.5 The interval Ix is closed in [0, +).
Proof (Proof of Lemma 2.5) We need to show that Rx belongs to Ix , that is that
B SR (x, Rx ) is compact. Let {yk }k be a sequence of points in B SR (x, Rx ), we need to
show that it has a convergent subsequence. We construct a Cauchy subsequence of
{yk }k as follows. For every integer l 1, we set
K l = B SR x, Rx 1 2l .
By assumption, {K l }l is an increasing sequence of compact sets in B SR (x, Rx ). For
every k N, there is yk1 K 1 such that
41
R
x
.
dSR yk , yk1 = inf dSR (yk , z) | z K 1
2
By compactness of K 1 , there is a strictly increasing mapping 1 : N N such that
the sequence {y1 1 (k) }k converges to some y 1 K 1 . Thus there exists k1 0 such that
R
x
dSR y1 1 (k) , y 1
2
k k1 .
k k2 .
1 2 )(k ) , z2
1
2
2
2
Rx
Rx
Rx
Rx
+
+
+
2Rx .
2
2
2
2
k kl .
l.
42
2 Sub-Riemannian Geodesics
4Rx
dSR zl , zl+1 l .
2
Hence {zk }k is a Cauchy sequence in B SR (x, Rx ). Since (M, dSR ) is complete, it
converges to some z B SR (x, Rx ).
Lemma 2.6 The interval Ix is open in [0, +).
Proof (Proof of Lemma 2.6) We need to show that if R Ix , then there is > 0 such
that R + belongs to Ix . Let R > 0 in Ix be fixed. Denote by BSR (x, R) the boundary
of B SR (x, R), that is BSR (x, R) = B SR (x, R) \ BSR (x, R). Since B SR (x, R) is assumed
to be compact, its boundary is compact too. From Proposition 2.3, we know that for
every y BSR (x, R), there is y > 0 such that B SR (y, 2y ) is compact. Since
BSR (x, R) N
i=1 BSR yi , yi .
Set
yi
| i = 1, . . . , N .
= min
2
B SR (x, R + ) B SR (x, R) N
i=1 BSR yi , 2yi
which is a finite union of compact sets, hence compact as well. This shows that
B SR (x, R + ) is compact.
In conclusion, Ix is both open and closed in [0, +). Hence Ix = [0, +) which
concludes the proof of (i). Let us now prove assertion (ii). We note that since does
not necessarily admit a global orthonormal frame on M, we cannot repeat verbatim
the proof of Proposition 2.3. Let x, y M be fixed, set R := max{2dSR (x, y), 1}. By
(i), we know that B SR (x, R) is compact. Let { k }k be a sequence of horizontal paths
x,1
joining x to y such that
with constant speed in
dSR (x, y) = lim length( k ).
k+
43
k,
which means that all the curves k remain in B SR (x, R). By Proposition 2.3, for every
z B SR (x, R) there is z > 0 such that any minimizing sequence of horizontal paths
with constant speed contained in BSR (z, z ) converges uniformly (up to taking a
subsequence) to some minimizing geodesic. By compactness, there are z1 , . . . , zL
B SR (x, R) and an integer N > 1 with R/N < min{1 , . . . , L }/4 such that
BSR (x, R)
L
BSR zl , 1/N .
l=1
1
R
lengthg ( k )
< + < l ,
dSR k (t), zl dSR ( k (tj ), zl ) +
N
N
N
for every t Ij . This shows that each piece of horizontal path jk with length
lengthg ( k )/N is contained in some BSR (zl , l ). Therefore, up to taking a subsequence, the sequence {jk }k converges to some minimizing geodesic with length
dSR (x, y)/N. We deduce easily the existence of a subsequence of { k }k converging
to some minimizing geodesic between x and y.
Remark 2.2 In fact, we proved a global version of Proposition 2.3. If (M, dSR ) is
a complete metric space, then for every x, y M and every minimizing sequence
x,1
joining x to y such that
{ k }k of horizontal paths with constant speed in
dSR (x, y) = lim length( k ),
k+
44
2 Sub-Riemannian Geodesics
x, y M.
Therefore, any Cauchy sequence with respect to dSR is a Cauchy sequence with
respect to dg . Hence it is convergent. Since both topology coincide, it is convergent
with respect to dSR as well.
ui (t)X i ( (t))
i=1
0
gu (t)
m
i=1
45
ui (t)X i (u (t)) dt =
0
i=1
m
1
ui (t)2 dt =: C(u),
i=1
among all controls u L 2 ([0, 1]; Rm ) such that the solution u : [0, 1] M of the
Cauchy problem
u (t) =
u (0) = x,
i=1
x,1
has been defined
is well-defined on [0, 1] and satisfies (the End-Point mapping EF
in Chap. 1)
x,1
(u) = y.
EF
In other terms, there is an open set U L 2 ([0, 1]; Rm ) such that u is solution to
the following optimization problem:
x,1
(u) = 1.
u minimizes C(u) among all u U with EF
v L 2 [0, 1]; Rm .
(2.3)
v L 2 [0, 1]; Rm .
x,1
This means that some nonzero linear form annihilates the image of EF
. Then u is
singular with respect to x and F or equivalently the path is singular with respect
to . By Proposition 1.11 and Remark 1.8, admits an abnormal extremal lift, that
is there is an absolutely continuous arc p : [0, 1] (Rn ) \ {0} with p(1) = p which
satisfies
p (t) =
i=1
a.e. t [0, 1]
46
2 Sub-Riemannian Geodesics
and
i = 1, , m.
1 2
i
i
H(x, p) :=
p X (x) = max
ui p X (x)
ui
2
2
uRm
i=1
i=1
(2.4)
i=1
(t) =
i
m
i
i=1 p(t) X ( (t)) X ( (t))
p (t) = H ( (t), p(t)) = m p(t) X i ( (t)) p(t) D X i
(t)
i=1
x
H
p ( (t), p(t))
(2.5)
v L 2 ([0, 1]; Rm ).
x,1
Moreover by Remark 1.5, the differential of EF
at u is given by
x,1
Du EF
(v) = S(1)
S(t)1 B(t)v(t)dt
v L 2 ([0, 1]; Rm ),
where the functions A, B, S were defined in Remark 1.5. Hence (2.3) yields
v L 2 ([0, 1]; Rm ).
Which implies
u (t) =
1
p S(1)S(t)1 B(t)
2
(2.6)
47
1
p S(1)S(t)1
2
t [0, 1].
By construction, for a.e. t [0, 1] we have u (t) = p(t) B(t), which means
that (2.6) is satisfied. Furthermore, as in the proof of Proposition 1.11, we have
p (t) = p(t) A(t) for a.e. t [0, 1]. This means that (2.5) is satisfied for a.e.
t [0, 1]. The pair ( , p) is solution to a smooth autonomous differential equation,
hence it is smooth.
The curve : [0, 1] T M given by (t) = ( (t), p(t)) for every t [0, 1]
is a normal extremal whose projection is and which satisfies (1) = (y, 2p ). We
say that is a normal extremal lift of . We also say that is a normal minimizing
geodesic.
Define the sub-Riemannian Hamiltonian H : T M R as follows. For every
x M, the restriction of H to the fiber Tx M is given by the nonnegative quadratic
form
p(v)2
1
| v (x) \ {0} .
(2.7)
p max
2
gx (v, v)
=
Let H denote the Hamiltonian vector field on T M associated to H, that is,
H
dH, or in local coordinates
H
H
(x, p),
(x, p) .
H (x, p) =
p
x
(t)
2H (t)
t [0, T ].
We check easily that the Hamiltonian defined by (2.7) reads as (2.4) in local coordinates. Then the previous study yields the following result.
Theorem 2.9 Let : [0, 1] M be a minimizing geodesic between x and y in M.
One of the two following non-exclusive cases occur:
is singular.
admits a normal extremal lift in T M.
Be careful, a minimizing geodesic could be both singular and the projection of a
normal extremal. In Sect. 2.5, we shall see several examples of minimizing geodesics,
48
2 Sub-Riemannian Geodesics
including the cases of singular normal minimizing geodesics and strictly abnormal
minimizing geodesic, that is abnormal geodesics admitting no normal extremal lift.
Remark 2.3 In the Riemannian case, that is if has rank m = n, any path is horizontal and regular (see Remark 1.9). As a consequence any minimizing geodesic is
normal.
Short normal geodesics are minimizing. Projections of normal extremals are minimizing for short times.
Proposition 2.10 Let x M and p Tx M with H(x , p ) = 0 be fixed. Then there is
a neighborhood W of p in Tx M and > 0 such that every normal extremal so that
(0) = (x , p) (in local coordinates) belongs to W minimizes the SR energy on the
interval [0, ]. That is if we set := : [0, ] M, then we have
i=1
1 2
ui X i (x)
ui
2
m
i=1
1
p, X i (x)2 ,
=
2
m
i=1
(t)
=
(x(t),
p(t))
=
p(t), X i (x(t))X i (x(t))
p
i=1
m
(2.8)
i=1
2H x , p0 =
p(t), X i (x(t))2 dt
i=1
i=1
ui (t)2 dt,
(2.9)
y (t) =
49
(2.10)
i=1
satisfies y() = x(). Let p0 Rn with H(x , p0 ) = 0 be fixed, we need the following
lemma.
Lemma 2.11 There exist a neighborhood W of p0 and > 0 such that, for every
p W , there exists a function S : B(x , ) R of class C 1 which satisfies
H(x, S(x)) = H(x , p), x B(x , ),
(2.11)
(2.12)
Proof (Proof of Lemma 2.11) The proof consists in applying the method of characteristics. Let be the linear hyperplane such that p0 , v = 0 for every v . We
first show how to construct locally S as the solution of the Hamilton-Jacobi Equation
(2.11) which vanishes on x + and such that S(x ) = p0 . Up to considering a
smaller neighborhood V , we assume that H(x, p0 ) = 0 for every x V . For every
x (x + ) V , set
H(x , p0 )
p0 .
p (x) :=
H(x, p0 )
Then, H(x, p (x)) = H(x , p0 ) and p (x) , for every x V . There exists > 0
such that, for every x (x + ) V , the solution (xx , px ) of (2.8), satisfying
xx (0) = x and px (0) = p (x), is defined on the interval (, ).
For every x (x + ) V and every t (, ), set (t, x) := xx (t). The
mapping (t, x)
(t, x) is smooth. Moreover, (0, x) = x for every x (x + )V
p(x), X i (x )X i (x ) does not belong to . Hence there exists
and (0, x ) = m
i=1
(0, ) with B(x , ) V such that the mapping is a smooth diffeomorphism from
(, ) ((x + ) B(x , )) into a neighborhood V of x . Denote by = (, )
the inverse function of , that is the function such that ( )(x) = ( (x), (x)) = x
for every x V . Define the two vector fields X and P by
X(x) := ( (x), (x)) and P(x) := p(x) ( (x)), x V .
Then,
50
2 Sub-Riemannian Geodesics
m
i=1
m
i=1
and
m
i=1
i=1
i=1
i (x)2 = 2H(x , p0 ),
i=1
d
d
d
x)v = X( (t, x))v = D(t,x) X(z(s)).
Dx (s, x)v = (t,
ds
dx
dx
Hence,
d
z(s), px (s) = z( s), px (s) + z(s), p x (s)
ds
= D(s,x) Xz(s), px (s)
z(s),
Since X(x) =
there holds
m
i=1 i (x)X
i (x)
and
m
i=1 i (x)
Dxx (s) X
51
m
m
px (s) =
i (xx (s)) Dxx (s) X i px (s) +
X i (xx (s)), px (s)i (xx (s))
i=1
=
=
i=1
m
i=1
px (s) +
i=1
i=1
We deduce that
d
ds z(s), px (s)
1
ui (t)2 dt
2
m
H(y(t), S(y(t))) +
= H(x , p) +
1
2
0
i=1
i=1
ui (t)2 dt.
52
2 Sub-Riemannian Geodesics
(x, p) T M.
(t)
= H ((t))
t [0, T ].
p x,p (1) .
53
t [0, 1],
x M.
dSR yk , yl V tk tl
k, l.
H(x, p) =
p X i (x)
2
m
i=1
(t) =
i
i
(t), pp (t) = m
i=1 pp (t) X (t) X (t)
p (t) = H
(t), p (t) = m p (t) X i
(t) p (t) D X i ,
p
p
p
(t)
i=1 p
x
H
p
i = 1, . . . , m,
pp (t) X i (t) mV
54
2 Sub-Riemannian Geodesics
and by compactness, the vector fields X1 , . . . , Xm and their differentials are bounded
in B SR (y, r). Thus there is a constant K > 0 such that
p p (t) K pp (t)
t [T , T ).
By Gronwalls Lemma (see Lemma A.1), we infer that both and pp are uniformly
bounded near T . This means that the extremal can be extended beyong T , which
gives a contradiction.
Remark 2.4 Let (x, p) T M such that p is singular be fixed. By Proposition 1.11,
p is the projection of an abnormal extremal : [0, 1] T M (written as (p , q)
in local coordinates). Then taking local coordinates, for every R the curve (here
x,p = (p , pp ) denotes the normal extremal starting at (x, p))
R.
55
2 (x, z) for any z M, the assumpProof (Proof of Lemma 2.15) Since eSR (x, z) = dSR
tion of the proposition implies that there is a neighborhood U of y in M such that
and
(2.13)
m
1
C(u) =
0
ui (t)2 dt,
i=1
among all the controls u L 2 ([0, 1]; Rm ) which are admissible with respect to x, F
x,1
(u) = y. Let u L 2 ([0, 1]; Rm ) be a
and V and which satisfy the constraint EF
x,1
control admissible with respect to x, F and V such that EF
(u) U . By (2.13)
one has
x,1
x,1
(u) EF
(u) .
C(u) eSR x, EF
Moreover
x,1
(u ) .
C(u ) = eSR (x, y) = (y) = EF
Hence u minimizes the functional D : L 2 ([0, 1]; Rm ) R defined as
x,1
D(u) := C(u) EF
(u) ,
x,1
over the set of controls u L 2 ([0, 1]; Rm ) such that EF
(u) U . This means that
56
2 Sub-Riemannian Geodesics
dSR (x, z) dSR x, z + z (z)
For every y Px , denote by rank(y) the rank of the minimizing horizontal path
y (see Sect. 1.3).
Lemma 2.17 The set of y Px with rank(y) = n is dense in M.
Proof (Proof of Lemma 2.17) We argue by contradiction. Assume that there is an
open set O M such that any point y Px O has rank < n. Set
r := max rank(y) | y Px O .
Fix y Px O such that rank(y) = r and set := y . For every y Px O denote
by y the affine subspace of Tx M such that p = y , that is the space of p Tx M
such that
t [0, 1].
p (t) = expx (tp) = x,p (t) = y (t)
Remembering Remark 2.4, we observe that the dimension of y is exactly equal
to n rank(y). As a matter of fact, given y Px O and an orthonormal family
F = {X 1 , . . . , X m } in a neighborhood V along y , remembering the arguments
given in Proposition 2.8 we check that p Tx M belongs to y if and only if
x,p (1) = (y, pp (1)) satisfies
x,1
(v) = Duy C(v)
2pp (1) Duy EF
v L 2 [0, 1]; Rm .
(2.14)
Span Du Ei1 , . . . , Du Eir = Span Du E1 , . . . , Du En .
57
(2.15)
j Duy Eij = uy ,
j=1
and more generally for every k there is k = (k1 , . . . , kr ) Tyk M (Rr ) such
that
r
kj Duk Eij = uk ,
j=1
x,1
x,1
Since {uk }k converges to u in L 2 ([0, 1]; Rm ), {Duk EF
}k converges to Du EF
and
k
58
2 Sub-Riemannian Geodesics
y Px O .
But since S has dimension strictly less than n, the set expx (S ) is a compact
set of measure zero in M. Then Px O cannot be dense in O . Which gives a
contradiction.
Returning to the proof of Theorem 2.14, we fix y Px with rank(y) = n. Given
an open set M, we call a function f : R Lipschitz in charts if it is Lipschitz
in a set of local coordinates in a neighborhood of any point of . This is equivalent
to saying that f is locally Lipschitz with respect to a Riemannian distance on M.
Lemma 2.18 There is an open set Oy of y in M such that the function
y Oy dSR x, y
is Lipschitz in charts.
Proof (Proof of Lemma 2.18) As before we fix an orthonormal family of vector
fields F in an open neighborhood V along := y which is associated with
u L 2 ([0, 1]; Rm ) through F . By a uniqueness-compactness argument, if {yk }k
converges to y and {k }k is a sequence of minimizing geodesics between x and yk
then it converges (up to a subsequence) to and is associated with a sequence of controls {uk }k which converges to u in L 2 ([0, 1]; Rm ) (see Proposition 2.3 and Remark
2.1). Then there is a neighborhood O of y such that for every y O every minimizing geodesic between x and y is contained in V with rank n. Let v1 , . . . vn in
L 2 ([0, 1], Rm ) be such that the linear operator
Rn Ty M
x,1
i
m
i=1 i Du EF v
x,1
is invertible. By continuity of u Du EF
, taking O smaller if necessary, we may
assume that for every y O and for every minimizing geodesic y from x to y
associated with a control uy , the linear operator
Rn Ty M
x,1
i
m
i=1 i Duy EF v
is invertible. For every y O, define F y : Rn M by
x,1
F () := EF
y
u +
y
i v
Rn .
i=1
59
and its differential at 0 is invertible. Hence by the Inverse Function Theorem, there are
an open neighborhood B y of y in M and a function G y : B y Rn with G y (y) = 0
such that
F y G y (z) = z
z B y .
From the definition of the sub-Riemannian distance between two points, we infer
that for any z B y we have
y
y i
G (z) i v
dSR (x, z) = eSR (x, z) u +
i=1
=: y (z).
L2
We conclude that, for every y O, there are a open set B y containing y and a C 1
function y : B y Rn such that
dSR (x, y) = y (y) and dSR (x, z) y (z) z B y .
The C 1 norms of the y s are uniformly bounded. This proves the lemma.
To conclude the proof of Theorem 2.14, we note that by the Rademacher Theorem,
the function y Oy dSR (x, y) is differentiable almost everywhere in Oy . By
Lemma 2.15, for every y Oy where the function is differentiable, there is py Tx M
such that
1
2
(x, ) .
y = expx py , dSR (x, y) = 2H x, py , x,py (1) = y, Dy dSR
2
Since dSR (x, ) is Lipschitz in Oy , there is some constant K > 0 such that all the
py s remain in a compact subset of Tx M. Now every y Oy can be approximated
by a sequence {yk }k of points in Oy where dSR (x, ) is differentiable. By compactness, up to taking a subsequence, the normal extremals starting at (x, pyk ) will converge to a normal extremal starting whose projection is a minimizing geodesic from
x to y.
Remark 2.6 We already know that the sub-Riemannian distance is continuous on
M M (see Proposition 1.13). The proof of Theorem 2.14 shows that if (, g) is
complete and x M be fixed, then the function y M dSR (x, y) R is locally
Lipschitz (in charts) on an open and dense subset of M.
60
2 Sub-Riemannian Geodesics
u L 2 [0, 1]; Rm .
u L 2 [0, 1]; Rm .
The Lagrange Multiplier Theorem asserts that if u minimizes C(u) under the conx,1
(u) = y, then there are (Rn ) and 0 {0, 1} with (, 0 ) = (0, 0)
straint EF
such that
x,1
= 0 Du C.
Du EF
In Sect. 2.2, we saw that whenever 0 = 0 we cannot deduce that satisfies the
geodesic equation, that is that it is the projection of a normal extremal. In the case
0 = 0, the control u UFx,1 is necessarily singular which means that it is a critical
x,1
point of EF
. Thus we have to study what happens at second order.
Let U be an open set in L 2 = L 2 ([0, 1]; Rm ) and F : U RN be a function of
class C 2 with respect to the L 2 -norm. We recall that we call critical point of F any
u U such that Du F : U RN is not surjective. Given a critical point u, we call
corank of u, the quantity
61
ind Du2 F
|Ker(Du F)
Im Du F
\ {0},
then the mapping F is locally open at u , that is the image of any neighborhood of u
is an neighborhood of F(u).
From Proposition 1.11 and Remark 1.8, we know that for every non-zero form
p (Rn ) with
x,1
= 0,
p Du EF
t [0, 1],
(2.16)
satisfies p (1) = p ,
p (t) =
a.e. t [0, T ],
i=1
and
p (t) X i u (t) = 0
t [0, T ], i = 1, . . . m,
S(t)
a.e. t [0, T ], S(0)
= In ,
(2.17)
Mn (R), B(t)
Mn,k (R) are defined by
and the matrices A(t)
:=
A(t)
u i (t)JX i u (t)
a.e. t [0, T ]
(2.18)
i=1
and
:= X 1 (u (t)), , X m (u (t))
B(t)
t [0, T ].
(2.19)
The following result combined with Theorem 2.19 will yield a necessary condition
for a minimizing horizontal path to be strictly
abnormal. It holds in the general case
of a control u which belongs to UFx,1 L [0, 1]; Rm . We do not need u to be
minimizing.
Theorem 2.20 Let u UFx,1 L [0, 1]; Rm and p (Rn ) \ {0} be such that
62
2 Sub-Riemannian Geodesics
x,1
p Du EF
= 0.
(2.20)
Assume that
x,1
ind p Du2 EF
x,1
Ker(Du EF
)
< +.
(2.21)
Then the absolutely continuous arc p : [0, 1] (Rn ) \{0} defined by (2.16) satisfies
p (t) X i , X j u (t) = 0
t [0, 1], i, j = 1, . . . , m.
(2.22)
Proof Let us first check that F is of class C 2 on UFx,1 (we refer the reader to Appendix
x,1
B for basics in differential calculus in infinite dimension). Given u UF
and
2
m
v L ([0, T ]; R ) we need to study the quantity
x,1
x,1
u + v EF
u ,
u+v (1) u (1) = EF
at second order when is small. We have
u+v (1) =
0
k
1
(2.23)
i=1
with u+v (0) = x. For every i = 1, . . . , m and every t [0, 1], the Taylor expansion
of each X i at u (t) at second order gives
1
+ D2u (t) X i u+v (t) u (t), u+v (t) u (t)
2
+ |u+v (t) u (t)|2 o(1).
Setting x (t) := u+v (t) u (t) for any t, (2.23) yields formally (x has size )
x (1) =
! m
i=1
+
0
"
vi (t)X i u (t) dt
i=1
i=1
2
+ v o(1).
"
m
1
2
i
vi (t)Du (t) X x (t) +
ui (t)Du (t) X (x (t), x (t)) dt
2
i
i=1
!
x1 (t)
63
"
ui (t)Du (t) X
x1 (t) +
i=1
"
vi (t)X u (t)
a.e. t [0, 1]
i=1
and
x2 (t)
! m
"
ui (t)Du (t) X
x2 (t) +
i=1
1
2
"
vi (t)Du (t) X
x1 (t)
i=1
ui (t)D2u (t) X i x1 (t), x1 (t)
i=1
u+v (1) = u (1) + x1 (1) + x2 (1) + o 2 ,
we obtain that (here we use the notations of the proof of Proposition 1.8)
x,1
Du2 EF
(v, v) = 2
v L 2 [0, 1]; Rm
where
C(t) =
(2.24)
i=1
and
1
ui (t)D2u (t) X i x1 (t), x1 (t) .
2
m
D(t) =
(2.25)
i=1
x,1
is C 2 on UFx,1 .
Using Gronwalls lemma, we leave the reader to check that EF
Let us now fix u UFx,1 L [0, 1]; Rm and p (Rn ) \ {0} such that (2.20)
and (2.21) are satisfied and prove that (2.22) holds. Note that we have for every
v L 2 ([0, 1]; Rm ) ,
x,1
(v, v) = 2
Du2 EF
S(t)
1 C(t)
+ D(t)
S(1)
dt,
(2.26)
D
are obtained by replacing u by u in (2.24)(2.25) and the definitions of
where C,
A,
B (see (2.17)(2.19)).
S,
Lemma 2.21 There is K > 0 such that
for any t , > 0 with [t , t + ] [0, 1],
x,1
there holds for every v Ker Du EF with Supp(v) [t , t + ],
64
2 Sub-Riemannian Geodesics
2 x,1
t , (v) K v2 2 2 ,
Du EF (v, v) Q
L
t , : L 2 ([0, 1]; Rm ) Rn is defined by
where Q
t +
t
m
m
i
j
t , (v) :=
p t
vi (t)D (t ) X
vj (s)X (t ) ds dt,
Q
t
t
j=1
i=1
(2.27)
S(1)
1 B(t)v(t)
S(t)
dt = 0.
x,1
p Du2 EF
|Ker
x,1
(du EF
)
(v) = 2
+ D(t)
dt.
p (t) C(t)
Setting
x1 (t) := S(t)
1 B(s)v(s)
S(s)
ds
t [0, 1],
we have x1 (t) = 0 for every t 0, t t + , 1 and by Cauchy-Schwarzs
inequality, we have for every t t , t + ,
1
S(s)
1 B(s)
t t vL2 K1 vL2 ,
x (t) sup S(t)
s[0,1]
S 1 , B in a neighborhood
where K1 is a constant depending only upon the sizes of S,
of the curve u ([0, 1]). Then we have
D(t)
=0
and
t t 0, t t + , 1 ,
D(t)
K3 v2 2 u
L
which gives
1
0
t t , t + ,
65
t
m
t
j=1
= x1 (t)
=
=
t
0
vj (s)X j (t ) ds
t
m
t
j=1
vj (s)X j (s) ds +
t
m
t
j=1
S(s)
1 B(s)v(s)
S(t)
B(s)v(s)
ds +
t
1 B(s)v(s)
S(s)
S(s)
S(t)
ds +
vj (s) X j (s) X j (t ) ds
t
m
t
j=1
t
m
t
j=1
vj (s) X j (s) X j (t ) ds
j
K2 vL2 23 , t t t , t + ,
(t)
v
(s)X
(
t
)
ds
j
x
t
j=1
(2.28)
t +
t
m
m
m
i 1
i
j
=
p (t)
vi (t)D (t) X x (t)
vi (t)D (t ) X
vj (s)X (t ) ds dt
=
i=1
t +
p (t)
i=1
j=1
t
m
i=1
j=1
1
0
66
2 Sub-Riemannian Geodesics
Ni,j (t ) := p t X i , X j (t ) > 0
= p t D (t ) X j X i (t ) D (t ) X i X j (t ) .
t , : L 2 ([0, 1] Rn be the mapping
Let > 0 such that [t , t + ] [0, 1] and Q
defined by (2.27). We observe that there holds for every v L 2 [0, 1]; Rm ,
t , (v) =
Q
t
=
t +
t
t +
vi (t)vj (s) p (t ) D (t ) X i X j (t ) ds dt
i,j=1
t
v(s), Mv(t)
ds dt =
t
t +
w(t), Mv(t)
dt,
i,j = p (t ) D (t ) X i X j (t ) ,
M
and
w(t) :=
v(s) ds
t [0, 1].
v L \ {0}.
|L
: L Rn
67
Let N an integer strictly larger than n be fixed and > 0 with t , t + [0, 1]
to be chosen later. Denote by L = Lt ,,N the vector space in L 2 [0, 1]; Rm of all the
controls v such that there is a sequence {a1 , . . . , aN } such that
v (t) = N ak cos k (tt )2
k=1
i
v (t) = N ak sin k (tt )2
k=1
j
t t , t + ,
t
/ t , t + ,
vi (t) = vj (t) = 0
and
vi (t) = 0, i = i, j
Let v L \ {0}, taking as before w(t) :=
wi (t) =
w (t) =
N
+t
t
t [0, 1].
v(s)ds, we have
sin k (tt )2
ak
(tt )2
,
k 1 cos k
ak
k=1 k
N
k=1
wi (t) = wj (t) = 0
t t , t + ,
t
/ t , t + ,
and
wi (t) = 0, i = i, j
t [0, 1].
Then we have
0
wi (t)vj (t) dt =
+ 2 2
a
k
k=1
4 k
and
0
wj (t)vi (t) dt =
+ 2 2
a
k
k=1
4 k
v (t) + w (t)M
v (t) + w (t)M
v (t) + w (t)M
v (t).
w(t), Mv(t)
= wi (t)M
ii i
i
ij j
j
ji i
j
jj j
But
68
2 Sub-Riemannian Geodesics
1
0
v (t) dt = M
wi (t)M
ii i
ii
0
wi (t)w i (t) dt = 0 =
v (t) dt.
wj (t)M
jj j
In conclusion, we have
t , (v) =
Q
1
0
N
N
2 Ni,j t
2 ak2
ak2
=
.
w(t), Mv(t) dt = Ni,j t
4 k
4
k
k=1
k=1
ak2 ,
k=1
which yields
Q
t , (v)
v2L2 2
N a 2
k
Ni,j t
1 Ni,j t
k=1 k
=
.
2
4 N
4
k=1 ak
p (t) X i , X j (t) = 0
t [0, 1], i, j = 1, . . . , m.
The latter property is called the Goh condition and is a Goh path.
Proof According to the previous notations, we define the mapping F : UFx,1
Rn R by
x,1
u UFx,1 .
(u), u2L2
F(u) := EF
This function, which is of class C 2 , cannot be open at u . As a matter of fact, if the
image of a neighborhood of u contains a neighborhood of F(u) then it contains a
control u UFx,1 with
2
x,1
EF
(u) = y and u2L2 u L2 ,
which contradicts the minimality of u from x to y. Therefore by Theorem 2.19 we
69
ind Du2 F
|Ker(du F)
< r := n rank(u).
Since the control u is strictly abnormal, the last coordinates of is zero. Denote by
p the dual of the first n coordinates of . Then we have
v L 2 [0, 1]; Rm .
x,1
p Du EF
(v) = 0
Since u is minimizing, |u(t)| is constant and u belongs to L [0, 1]; Rm . Theorem 2.20 concludes the proof.
Example 2.1 A distribution is called medium-fat if, for every x M and every
section X of with X(x) = 0, there holds
Tx M = (x) + [, ](x) + X, [, ] (x),
(2.29)
where
, (x) := [X, Y ](x) | X, Y sections of
and
X, [, ] (x) := X, [Y , Z] (x) | Y , Z sections of .
Any two-generating distribution is medium-fat. An example of medium-fat distribution which is not two-generating is given by the rank-three distribution in R4 with
coordinates x = (x1 , x2 , x3 , x4 ) generated by the vector fields defined by
X 1 = x1 , X 2 = x2 , X 3 = x3 + (x1 + x2 + x3 )2 x4 .
Medium-fat distribution do not admit non-trivial Goh paths. As a matter of fact, if
: [0, T ] M is an horizontal path which admits an abnormal lift = ( , p) :
[0, T ] T M satisfying the Goh condition, then we have
p(t) X i , X j (t) = 0
i, j = 1, . . . , m,
(2.30)
for every t in a small interval I [0, T ] such that (t) is in a local chart of M and
is parametrized by a family {X 1 , . . . , X m } of smooth vector fields. Then if we denote
by u the control which is associated to through F , derivating the previous equality
yields for any i, j = 1, . . . , m,
p(t)
! m
"
(t) = 0
uk (t)X , X , X
k
t I.
(2.31)
k=1
70
2 Sub-Riemannian Geodesics
and
x
Y = y + z ,
2
and g is the metric making the family {X, Y } orthonormal, that is defined by
g = dx 2 + dy2 .
The above structure can be shown to be left-invariant under the group law
1
(x, y, z) x , y , z = x + x , y + y , z + z + xy x y .
2
Thanks to Proposition 1.7, any horizontal path on [0, T ] has the form u = (x, y, z) :
[0, T ] R3 where
x (t) = u1 (t)
y (t) = u2 (t)
1
(x(t)y(t) y(t)x (t)) dt =
2
c
1
(xdy ydx) ,
2
where (t) = (x(t), y(t)) is the projection of the curve to the plane. According to
the Stockes Theorem, we have
1
1
dx dy +
(xdy ydx) =
(xdy ydx) ,
2
2
D
c
where D denotes the domain which is enclosed by the curve and the segment
c := Q1 , Q2 := (x(0, y(0)), (x(T ), y(T )
from Q1 to Q2 (see Fig. 2.3).
Therefore, given two points P1 = (x1 , y1 , z1 ), P2 = (x2 , y2 , z2 ) in R3 , the horizontal paths which minimizes the length from P1 to P2 are the curves : [0, 1] R3
whose signed area of D satisfies
71
D
dx dy = (z2 z1 )
c
1
(xdy ydx) ,
2
with minimal length. According to the isoperimetric inequality, the curves in the
plane sweeping the same area and which minimize the length are given by circles.
This fact can be easily recovered by Theorem 2.9 and Proposition 2.8 (we saw
in Example 1.14 that admits no non-trivial singular horizontal paths). Assume
that u = (x, y, z) : [0, 1] R3 is a minimizing geodesic from P1 := u (0)
to P2 := u (1) = P1 . Then according to Proposition 2.8, there is a smooth arc
p = (p1 , p2 , p3 ) : [0, 1] (R3 ) such that the following system of differential
equations holds
y
x = px 2 pz
y = py
+ 2x pz
z = 21 py + 2x pz x px 2y pz y ,
x
pz p2z
p x =
2
py +
p = px 2y pz p2z
y
p z = 0.
72
2 Sub-Riemannian Geodesics
X = x1 , Y = (1 + x1 (x)) x2 + x12 x3 ,
where is a smooth function and equipped with a metric g making {X, Y } an
orthonormal family. In a sufficiently small neighborhood of the origin V , singular curves are given by the horizontal paths which are contained in the Martinet set
= x1 = 0 ,
that is of the form
t
x(t) = 0, x2 (0) +
u2 (s)ds, 0, x3 (0) ,
0
(t) = 0, t, 0
t [0, ],
minimizes the length among all horizontal paths joining (0, 0, 0) to (0, , 0).
Proof We need to show that among all controls u = (u1 , u2 ) : [0, ] R2 with
u12 + u22 1 steering the origin to P := (0, , 0), we have < . There is r > 0 such
that B SR (0, r) is included in V . If (0, r), then any minimizing geodesic joining
0 to (0, , 0) is contained in B SR (0, r). As a matter of fact, we know that dSR (0, P)
< r. Let C > 0 be upper bounds for on B SR (0, r) Let u = x : [0, ] R3 be
a competitor for . We get easily
+
x1 ( ) = +0 u1 (s) ds
=0
x3 ( ) = 0 u2 (s)x1 (s)2 ds = 0.
(2.32)
Set
:= max x1 (s) | s [0, ] .
(2.33)
x1 (s)2 1 u2 (s) ds +
0
u2 (s) ds .
2
x1 (s)2 ds =
x1 (s)2 u2 (s) ds
(2.34)
73
Let s [0, ] be such that |x1 (s)| = . Since |x1 (s)| 1 for almost every s [0, ]
and x1 (0) = x1 ( ) = 0, we have s , s . Which means that the interval
[s /2, s + /2] is included in [0, ] and
x1 (s)
2
s s /2, s + /2 .
Therefore we have
x1 (s)2 ds
s +/2
s /2
x1 (s)2 ds
3
.
4
u2 (s) ds
.
4
(2.35)
Then by the second line in (2.32) and the definitions of and C, we have
0
0
u2 (s) ds +
u2 (s) ds +
0
u2 (s) x1 (s) x(s) ds
u2 (s) ds + C.
x 1 = p1
x 2 = (p Y (x)) (1 + x1 (x))
x 3 = (p Y (x)) x12 ,
(2.36)
74
2 Sub-Riemannian Geodesics
p
(x)
+
x
=
Y
(x))
(x)
+
2p
x
(p
1
2
1
3
1
(x)
p 2 = p Y (x) p2 x1 x
2
p 3 = p Y (x) p2 x1 (x) ,
x3
(2.37)
with
(2.38)
2
2
Note that
if
0, that is whenever g = dx1 + dx2 , then the horizontal path given by
(t) = 0, t, 0 for any t [0, ] is the projection of the normal extremal starting at
(0, p ) with p = (0, 1, 0). Then it is a singular normal minimizing geodesic between
its end-points (see Example 1.16 and Theorem 2.9). Different choices of metrics can
provide examples of strictly abnormal minimizing geodesics.
1
1
(p X(x))2 + (p Y )2 .
2
2
Since H is constant along its extremals and x1 (t) = x3 (t) = 0 for any t [0, ], we
have
2
p(t) X (t)
+ p(t) Y (t)
= p1 (t)2 + p2 (t)2 = p 21 + p 22
t [0, 1].
On the other hand, since x1 (t) = 0 for every t [0, 1], the second and third equations
in (2.37) yield
p 2 = p 3 = 0 = p2 (t) = p 2 t [0, 1].
Moreover, (2.36) also gives x 2 = p 2 that is p 2 = 0 ( has constant speed). Since p1
is smooth and both p2 and p21 + p22 are constant, p1 is necessarily constant. The first
equation in (2.36) and x1 = 0 give x 1 = p1 . Hence p1 = p 1 = 0. Then, using that
p1 = 0, p2 = p 2 = 0, the first equation in (2.37) gives
75
p 2 (t) = 0
t [0, 1].
76
2 Sub-Riemannian Geodesics
characteristic lines in the Martinet surface (Theorem 2.23) is taken from the
monograph by Liu and Sussmann [6] which provide a more general class of counterexamples. Note that the Montgomery counter-example as well as all other known
counter-examples exhibit smooth singular minimizing curves. The existence of nonsmooth sub-Riemannian geodesics is open.
In the first example of Sect. 2.5, we briefly explained that the sub-Riemannian
structure under study was indeed left-invariant under some group law. This additional
structure makes H1 a Carnot group. We refer the reader to the Montgomery textbook
[8] or to the Jean monograph [5] for further details on Carnot groups.
References
1. Agrachev, A.: Any sub-Riemannian metric has points of smoothness. Dokl. Akad. Nauk. 424(3),
295298 (2009), translation in. Dokl. Math. 79(1), 4547 (2009)
2. Agrachev, A.A., Sachkov, Y.L.: Control Theory from the Geometric Viewpoint. Encyclopaedia
of Mathematical Sciences, vol. 87, Springer, Heidelberg (2004) NULL
3. Rifford, L., Trlat, E.: Morse-Sard type results in sub-riemannian geometry. Math. Ann. 332(1),
145159 (2005)
4. Clarke, F.H.: Optimization and Nonsmooth Analysis. Wiley-Interscience, New York (1983)
Republished as vol. 5 of Classics in Applied Mathematics, SIAM (1990)
5. Jean, F.: Control of Nonholonomic Systems and Sub-Riemannian geometry. Lectures given at
the CIMPA School Gomtrie sous-riemannienne, Beirut (2012)
6. Montgomery, R.: Abnormal minimizers. SIAM J. Control Optim. 32(6), 16051620 (1994)
7. Agrachev, A., Sarychev, A.: Sub-Riemannian metrics: minimality of singular geodesics versus
subanalyticity. ESAIM Control Optim. Calc. Var. 4, 377403 (1999)
8. Montgomery, R.: A tour of subriemannian geometries, their geodesics and applications. Mathematical Surveys and Monographs, vol. 91. American Mathematical Society, Providence, RI
(2002)
9. Pontryagin, L., Boltyanskii, V., Gamkrelidze, R., Mischenko, E.: The Mathematical Theory of
Optimal Processes. Wiley-Interscience, New-York (1962)
10. Liu, W., Sussmann, H.J.: Shortest paths for sub-Riemannian metrics on rank-2 distributions.
Mem. Amer. Math. Soc. 118, 564 (1995)
11. Vinter, R.B.: Optimal Control. Birkhuser, Boston (2000)
Chapter 3
Abstract This Chapter is concerned with the study of optimal transport maps in the
sub-Riemannian setting. We first provide a course in optimal transport theory. Then
we study the well-posedness of the Monge problem for sub-Riemannian quadratic
costs.
Throughout all the chapter, M denotes a smooth connected manifold without boundary of dimension n 2.
c x, T (x) d(x),
M
B measurable set in M.
77
78
CM (, ) := inf
c x, T (x) d | T = ,
(3.1)
(T (x)) d(x) =
(y) d(y),
for all -integrable function . If M = Rn and and are absolutely continuous with respect to the Lebesgue measure respectively with densities f and g in
L 1 (Rn ; [0, +)), the latter property can be written as
(T (x))f (x) dx =
M
(y)g(y) dy,
M
f (x)
g(T (x))
a.e. x Rn .
Example 3.1 Transport maps may not exist. For example, consider in Rn the probability measures , given by
= x and =
1
1
y 1 + y 2 ,
2
2
79
Example 3.2 Minimizers of Monges problem may not be unique. On the real line
R, consider the probability measures and given by
= 1[0, 1] L 1 and = 1[1, 2] L 1 ,
where L 1 denotes the Lebesgue measure in R. In other terms, and are respectively
the restriction of the Lebesgue measure on the intervals [0, 1] and [1, 2]. The two
maps T1 , T2 : R R given by
T1 (x) = x + 1 and T2 (x) = 2 x
x R,
x, y R.
We check easily that the transportation cost for T1 and T2 are given by
c x, Ti (x) d(x) =
|Ti (x) x| dx = 1
i = 1, 2.
c x, T (x) d(x) =
|T (x) x| dx
=
[T (x) x] dx =
=
y dy
0
1
T (x) dx
x dx
0
x dx = 1.
This shows that the infimum in the definition of CM (, ) is attained by all transport
maps from to . So, it is not unique.
The constraint T = being highly non-linear, the Monge optimal transport
problem is quite difficult from the viewpoint of optimization. That is why we will
study a notion of weak solution for this problem.
The Kantorovitch relaxation. Given two probability measures , on M, we denote
by (, ) the set of probability measures in the product M M with first and
second marginals and , that is such that
1 = and 2 = ,
(3.2)
80
CK (, ) := inf C() | (, ) .
Remark 3.2 The property (3.2) is equivalent to
(B) = B M and (B) = M B ,
for any measurable set B in M, which is also equivalent to
MM
1 (x) d(x) +
M
2 (y) d(y),
M
for all -integrable function 1 and -integrable function 2 . In particular, the set
(, ) is a convex set which always contains the product measure .
Remark 3.3 If T : M M is a transport map from to then the measure on
M M given by
:= (Id T ) ,
is a transport plan between and . This means that the Kantorovitch optimization
problem is more general than the Monge optimization problem, or
CK (, ) CM (, ),
for all probability measures , on M.
Example 3.3 Returning to Example 3.1, we note that the product measure
=
1
1
(x, y1 ) + (x, y2 ) ,
2
2
81
C()
= CK (, ) := inf C() | (, ) .
Proof We first note that CK (, ) is finite. As a matter of fact, since the product measure belongs to (, ) and c is bounded on Supp() Supp()
(by assumption c is continuous and Supp(), Supp() are compact), we have
CK (, ) C( ) < +. In fact, the supports of all transport plans between
and are contained in the set Supp() Supp() M M which is compact by
assumption on Supp() and Supp(). Then we can assume without loss of generality
that M is compact. Denote by P(M M) the set of probability measures on M M
and define F : P(M M) R by
F() :=
P(M M).
c(x, y) d(x, y)
MM
(x, y) dk (x, y) k+
MM
k+
82
The supports of optimal transport plans have specific properties. Let us introduce
the notion of c-cyclically monotone sets.
Definition 3.2 A subset S M M is called c-cyclically monotone if for any
finite number of points (xj , yj ) S, j = 1, . . . , J, and a permutation on the set
{1, . . . , J},
J
J
c(xj , yj )
c x (j) , yj .
j=1
j=1
Remark 3.4 The definition given above is equivalent to the following one: for any
finite number of points (xj , yj ) S, j = 1, . . . , J,
J
c(xj , yj )
j=1
J
c xj , yj+1 ,
j=1
with yJ+1 = y1 . The equivalence is a straightforward consequence of the decomposition of a permutation into disjoint commuting cycles.
Remark 3.5 If c is assumed to be continuous, the c-cyclical monotonocity is stable
under closure. The closure of a c-cyclically monotone set is c-cyclically monotone.
Given two probability measures , on M, we call optimal transport plan between
and any (, ) satisfying CK (, ) = C(). Optimal transport plans
always have c-cyclically monotone supports.
Theorem 3.3 Let , be two probability measures on M. Assume that c is continuous and that Supp() and Supp() are compact. Then there is a c-cyclically monotone
compact set S Supp() Supp() such that the support of any optimal transport
plan between and is contained in S .
Proof Let us first show that the supports of optimal transport plans are always
c-cyclically monotone. We argue by contradiction and assume that there is an optimal
transport plan (, ) whose support is not c-cyclically monotone. Then there
is an integer J > 1, J points (x1 , y1 ), . . . , (xJ , yJ ) in Supp() and a permutation
on the set {1, . . . , J} such that
J
c(xj , yj ) >
j=1
J
c x (j) , yj .
j=1
By continuity of c, there are open sets Uj , Vj for j = 1, . . . , J which contain respectively xj , yj such that
J
j=1
c(uj , vj ) >
J
c u (j) , vj
J
uj , vj j=1,...,J j=1
Uj Vj .
j=1
(3.3)
83
J
j=1
1
1Uj Vj .
Uj Vj
m
:= min Uj Vj | j = 1, . . . , J ,
J
denote by Uj (resp. Vj ) the projection from j=1
Uj Vj to Uj (resp. to Vj ) and
define the measure on M M by
J
U V
m
U (j) Vj
, P j , j P .
= +
J
j=1
We have
m
m
Uj Vj
1
1Uj Vj
, P =
J
J
V
j
j
j=1
j=1
J
1
1Uj Vj = 0.
J
J
j=1
Moreover
J
J
J
J
U
U
Uj , Vj P =
U (j) , Vj P
1
j P =
(j) P = 1
j=1
and
j=1
j=1
j=1
J
J
J
V
Uj , Vj P =
U (j) , Vj P .
j P = 2
2
j=1
j=1
j=1
c(x, y) d (x,
y) =
MM
c(x, y) d(x, y)
MM
J
c u (j) , vj c uj , vj dP u1 , v1 , , . . . , uJ , vJ ,
j=1
84
and the last term is negative [by (3.3)]. This means that cannot be optimal and
gives a contradiction. Then we know that the supports of any optimal transport plan
between and is c-cyclically monotone. Denote by opt (, ) the set of optimal
transport plans in (, ) and set
S :=
Supp().
opt (,)
:=
j=1
J
c(xj , yj )
c x (j) , yj .
j=1
Example 3.4 Returning to Example 3.2, we can show that the set provided by
Theorem 3.3 has to be S = [0, 1] [1, 2] = Supp() Supp(). As a matter of
fact, for every (x, y) [0, 1] [1, 2] there is a bijective function T : [0, 1] [1, 2]
which is lower semicontinuous, increasing and piecewise affine with slope 1, and
whose graph contains (x, y) (see Fig. 3.2). Thanks to the observation we did in Example 3.2, such a function is a transport map from = 1[0, 1] L 1 to = 1[1, 2] L 1 ,
hence it is optimal.
Kantorovitch potentials. The aim of this section is to characterize c-cyclically
monotone sets in a more analytic way.
Definition 3.4 A function : M R {+}, not identically +, is said to be
c-convex if there is a non-empty set A M R such that
x M.
(3.4)
85
y M.
x M.
(x) = sup c (y) c(x, y) | y M
Proof By definition of c we have
c (y) c(x, y) (x)
x M, y M.
Which implies that (x) supyM { c (y) c(x, y)} for any x M. Let us show
that (x) supyM { c (y) c(x, y)} for any x M. Argue by contradiction and
assume that there is x M such that
x > sup c (y) c x , y y M .
Since is c-convex, there are a set A M R, (y, ) A and > 0 such that
c x , y + x sup c (y) c x , y | y M + 3.
Then we get c y 2, which by definition of c (y) implies that there is
x M such that
(x) + c x, y .
This contradicts (3.4).
86
y Rn ,
|y|2
x, y | (y, ) A
(x) = sup
2
2
x Rn .
|y x|2
|y|2
| (y, ) B ,
+
2
2
87
c := (x, y) M M | y c (x) .
Remark 3.6 By the above definitions, a pair (x, y) in M M belongs to c if and
only if
(x) + c(x, y) (z) + c(z, y)
z M,
which is also equivalent to
c (y) c(x, y) c (z) c(x, z)
z M.
x M,
(3.5)
y M.
(3.6)
j=1
j=1
j=1
J
j=1
c x (j) , yj .
j=1
88
(x) := sup c x , y c x1 , y
+
J1
c xj , yj c xj+1 , yj + c xJ , yJ c x, yJ
j=1
| J N, J 2, xj , yj S, j = 1, . . . , J ,
for every x M. We claim that is a c-convex function whose contact set contains S.
First taking J = 2, x = x1 = x2 = x and y1 , y2 = y , we check easily that (x ) 0.
Furthermore, by c-cyclical monotonicity of S, we have
J1
c x , y c x1 , y +
c xj , yj c xj+1 , yj + c xJ , yJ c x , yJ 0,
j=1
c xj , yj c xj+1 , yj + c xJ , y
(y) := sup c x , y c x1 , y +
j=1
| J N, J 2, xj , yj S, j = 1, . . . , J 1, (xJ , y) S
y M.
Note that if y M is such that there are no x M with (x, y) S, then (y) = .
However, as above we check easily that (y) = 0 which shows that is not identically
. Therefore, by construction we have for every x M,
(3.7)
z M.
c xj , yj c xj+1 , yj
(z) sup c x , y c x1 , y +
j=1
+ c(x, y) c(z, y) | J N, J 2, xj , yj S,
89
j = 1, . . . , J 1, xJ = x
x M,
(3.8)
y M.
(3.9)
y 2 (S).
(3.10)
x M,
x M.
y M.
z 1 (S).
90
91
(i) is optimal.
(ii) Supp() S .
Proof By Theorem 3.3, there is a c-cyclically monotone compact set S Supp()
Supp() such that the support of any optimal transport in (, ) is contained in
S . Let us show that S satisfies the equivalence given in the statement of the theorem. First, by construction we have (i) (ii). By Theorem 3.7, there is a c-pair of
potentials with S c . Then we have
c (y) (x) = c(x, y)
(x, y) S .
(3.13)
Furthermore we have
c (y) (x) c(x, y)
c(x, y)
x, y M.
(3.14)
Let us show that (ii) (i). Let (, ) be such that Supp() S . On the one
hand, by (3.13), we have
(y) d(y)
(x) d(x) =
c
(y) (x) d(x, y)
MM
c(x, y) d(x, y) = C().
=
MM
c (y) d(y)
M
c
(y) (x) d (x, y)
(x) d(x) =
M
MM
c(x, y) d(x, y) = C( ).
MM
92
x M,
(3.15)
y M,
(3.16)
and
S c .
(3.17)
To prove the existence and uniqueness of an optimal transport map, we will show
that S is concentrated on a graph. More precisely, we will prove that for every x
outside a -negligible set N M, the set c (x) is a singleton.
Theorem 3.9 Let , be two probability measures on M. Assume that c is continuous and that Supp() and Supp() are compact. Let S and (, c ) given by
Theorems 3.3 and 3.7 as above. Moreover assume that for -a.e. x M, the set
c (x) is a singleton. Then there is a unique optimal transport map from to . It
satisfies
c (x) = T (x)
a.e. x M.
(3.18)
y .
y .
93
y1 = y2 M, x M,
94
Proof Argue by contradiction and assume that y1 = y2 both belong to c (x). Then
we have
z M.
c (yi ) = (x) + c(x, yi ) (z) + c(z, yi )
Thus, for every i = 1, 2,
c(z, yi ) (z) + (x) + c(x, yi ),
with equality at z = x. Since is super-differentiable at x, we infer that both
functions z c(z, y1 ) and z c(z, y2 ) share a common sub-differentiable at x.
This contradicts the sub-TWIST condition.
By Theorem 3.9 and Lemma 3.11, in order to prove the existence and uniqueness
of optimal transport maps from a compactly supported probability measure to
another one , it is sufficient to show that the super-differential of the potential
is non-empty for -almost every point in M. Such a property can be obtained
thanks to Rademachers Theorem. We recall that a function defined on a smooth
manifold is called Lipschitz in charts if it is Lipschitz in a set of local coordinates
in a neighborhood of any point. The Rademacher Theorem asserts that any function
which is Lipschitz in charts on an open subset of M is differentiable almost
everywhere in .
Theorem 3.12 Let c : M M [0, +) be a cost which is Lipschitz in charts
and satisfies the sub-TWIST condition. Let , be two probability measures with
compact support on M. Assume that is absolutely continuous with respect to the
Lebesgue measure. Then there is existence and uniqueness of an optimal transport
map from to . In fact, there is a c-convex function : M R which is Lipschitz
in charts such that
c (x) = T (x)
a.e. x M.
(3.19)
Proof By Theorems 3.3 and 3.7 there is a c-cyclically monotone compact set S
Supp()Supp() which contains the supports of all optimal plans between and
together with a c-pair of real-valued continuous potentials (, c ) such that (3.15)
(3.17) are satisfied. In a neighborhood of each x M, the function is the maximum
of a family of functions x 2 (S) c (y) c(x, y) with y 2 (S ) which are
uniformly Lipschitz (in charts) in the x variable . Therefore, is Lipschitz in charts
on M. Since is assumed to be absolutely continuous with respect to the Lebesgue
measure, Rademachers Theorem implies that is differentiable and a fortiori superdifferentiable -a.e. We conclude easily by Theorem 3.9 and Lemma 3.11.
95
x, y Rn .
a.e x Rn .
In other terms, the unique optimal transport map from to is given by the gradient
of a convex function.
Example 3.8 Let M = Rn , note that the Monge cost c : Rn Rn [0, +) given
by c(x, y) = |y x| (cf. Examples 3.2, 3.5) is Lipschitz but does not satisfy the
sub-TWIST condition. As a matter of fact, we have
xy
c
(x, y) =
x
|x y|
x = y Rn .
This means that Dx c(, y1 ) = Dx c(, y2 ) for any y1 , y2 such that y1 x and y2 x
are positively colinear. Hence Theorem 3.15 do not apply. In fact, we already saw
through Example 3.2 that uniqueness of optimal transport maps does not hold in this
context.
Example 3.9 (McCanns Theorem) Let (M, g) be a complete Riemannian manifold.
The geodesic distance dg is Lipschitz in charts on M M. Define the quadratic
geodesic cost or McCanns cost c : M M [0, +) by
c(x, y) :=
1 2
d (x, y)
2 g
x, y M.
z M.
96
z M,
with equality at z = x. Then arguing as above, we deduce that for every pair of
compactly supported probability measures , on M with absolutely continuous
with respect to the Lebesgue measure, there is a unique optimal transport map T
from to satisfying (3.26) where : M R is a c-convex function which is
Lipschitz in charts. By the above discussion, we have
T (x) = expx Dx
a.e x M
(3.20)
1
dSR (x, y)2
2
x, y M.
As we saw before, in order to obtain existence and uniqueness results for optimal
transport maps, it is convenient to be able to show that super-differentials of potentials
are non-empty almost everywhere and that some sub-TWIST condition is satisfied
by the cost function. The sub-TWIST condition follows immediately from Lemma
2.15. So we just have to deal with regularity issues of c-convex functions. In the case
of compactly supported probability measures, regularity properties of Kantorovitch
potentials can be obtained from the regularity of the cost. We develop this approach
97
in the Sect. 3.4 by showing that under additional assumptions the sub-Riemannian
distance is Lipschitz and even locally semiconcave outside the diagonal.
Remark 3.11 As explained above, if M equipped with a SR structure for which the
2 is Lipschitz on M M, then for every pair of compactly supported
cost c = dSR
probability measures , on M with absolutely continuous with respect to the
Lebesgue measure, there is a unique optimal transport map T from to which can
be expressed as
T (x) = expx Dx
a.e x M,
(3.21)
[0, 1], x, y B x , .
x B x , ,
with f (x) C|x|2 concave, that is as the sum of a concave function and a smooth
function. Note that every locally semiconcave function is locally Lipschitz on its
domain, and thus, by Rademachers Theorem, it is differentiable almost everywhere
on its domain.
98
The following result is useful to prove the local semiconcavity of a given function
(Fig. 3.5).
Lemma 3.13 Let f : R be a function defined on an open set Rn . Assume
that for every x there exist a neighborhood V of x and a positive real
number such that, for every x V , there is px Rn such that
u(y) u(x) + px , y x + |y x|2
y V .
z B.
Remark 3.12 Thanks to Lemma 3.13, a way to prove that a given function f :
R is locally semiconcave on is to show that for every x we can put a C 2
support function on the graph of u at x with a uniform control of the C 2 norm of .
Outside the diagonal, sub-Riemannian distances of ideal SR structures enjoy the
same kind of regularity as Riemannian distances.
Theorem 3.14 Let (, g) be an ideal sub-Riemannian structure on M. Then the
SR distance is continuous on M M and locally semiconcave on M M \ D. In
particular, dSR is Lipschitz in charts on M M \ D.
99
Proof The continuity of dSR follows from Proposition 1.13. To prove the local semiconcavity, we proceed as explained in Remark 3.12. Let us fix (x, y) M M \D and
x,1
be a minimizing geodesic joining x to y. There is an open neighborhood
V of ([0, 1]) in M and an orthonormal family F (with respect to the metric g) of
m smooth vector fields X 1 , . . . , X m such that
z V .
(z) = Span X 1 (z), . . . , X m (z)
Taking a change of coordinates if necessary, we may assume that V is an open subset
of Rn . Furthermore, there is a control u L 2 ([0, 1]; Rm ) such that
(t) =
m
ui (t)X i ( (t))dt
i=1
Since u is regular, there are v1 , . . . vn in L 2 ([0, 1]; Rm ) such that the linear operator
Rn Rn
m
x,1
vi
i Du EF
i=1
z,1
(z, ) z, EF
u +
m
i vi
i=1
(z, w) B.
Denote by 1 the second component of G . From the definition of the subRiemannian energy between two points, we infer that for any (z, w) B we have
m
1
(z, w) vi .
eSR (z, w) u +
i
i=1
L2
100
Set
m
1 (z, w)
x,y (z, w) := u +
i
i=1
(z, w) B.
L2
We conclude that, there is a function x,y of class C 2 such that dSR (z, w) x,y (z, w)
for any (z, w) in a neighborhood of (x, y) in M M, and dSR (x, y) = x,y (x, y). By
compactness, the C 2 norms of the functions x,y are uniformly bounded. As a matter
of fact, from Remark 2.2 we know that the set of minimizing geodesics from x to y is
compact with respect to the uniform topology; any sequence of minimizing geodesics
{k }k from xk to yk converges uniformly to a minimizing geodesic from x to y. We
also know (see Remark 2.1) that if we cover the set of minimizing curves from x to
y by a finite number of open tubes admitting orthonormal frames, then minimizing
p(t) X i , X j (t) = 0
t [0, 1], i, j = 1, . . . , m.
Of course, the above definition does not depend upon the parametrization.
Theorem 3.15 Let (, g) be a complete sub-Riemannian structure on M, assume
that any sub-Riemannian minimizing geodesic joining two distinct points in M is not
a Goh path. Then, the SR structure (, g) is Lipschitz.
x,1
Proof Let us fix (x, y) M M \D and
a minimizing geodesic joining x to
y. As before, denote by F = {X 1 , . . . , X m } an orthonormal family of vector fieldsd
101
along ([0, 1]) and by u the control associated with . Two cases may appear:
First case: u := u is not singular.
Then by the arguments given in the proofs of Lemma 2.18 and Theorem 3.14, there
are , K > 0 such that
eSR (x, z) eSR (x, y) + K|z y|
z B(y, ).
u u
L2
x,1
< , eSR x, EF
(u) = uL2 ,
there holds
x,1
x,1
(u) + 2K z EF
(u) ,
eSR (x, z) eSR x, EF
(3.22)
x,1
for every z B EF
(u), /2 .
Second case: u = u is singular.
By Theorem 2.20, we have necessarily
x,1
ind Du2 EF
x,1
|Ker(Du EF
)
= +,
(3.23)
x,1
for all Im Du EF
\ {0}. Recall that C : L 2 ([0, 1]; Rm ) is defined by
u L 2 [0, 1]; Rm .
C(u) := u2L2
x,1
x,1
Ker (Du C) and F := EF
E := E0 Ker Du EF
|{u}+E
x,1
By construction, Du EF
and Du F have the same image in Rn and E0 has finite
dimension. Then by (3.23), we have
ind Du2 F
|Ker(Du F)
= +,
102
for all Im (Du F) \ {0}. We can apply Theorem B.4 to the function
F.Hence
there are c > 0, (0, 1) such that for every (0, ) and every z B F u , c2 ,
there are w1 , w2 L 2 ([0, 1]; Rm ) such that
z = F u + w1 + w2
(3.24)
w1 Ker Du F , w1 L2 < , w2 L2 < 2 .
(3.25)
and
Let z B(y, c2 ) with |z y| = c2 /2. Then there are w1 , w2 L 2 [0, 1]; Rm such
that (3.24)(3.25) are satisfied. Set u := u + w1 + w2 . Then we have
x,1
(u),
z = EF
2
= eSR (x, y) + Du C w2 + w1 + w2 L2
2
eSR (x, y) + 2u L2 2 + + 2
4u L2 + 8
eSR (x, y) +
|z y|.
c
Proceeding as in the proof of Theorem B.4, we can show that the above estimate
holds in a neighborhood of u , that is (taking c > 0, (0, 1) smaller if necessary)
for every (0, ), for every u L 2 [0, 1]; Rm , and every z Rn with
u u
L2
x,1
< , z EF
(u) < c 2 ,
u u
L2
x,1
< , eSR x, EF
(u) = uL2 ,
103
there holds
4u 2 + 8
x,1
x,1
L
(u) +
eSR (x, z) eSR x, EF
z EF (u) ,
c
(3.26)
x,1
for every z B EF
(u), c /4 .
Let us explain how to conclude by compactness. Let x M and B a compact set
in M such that {x} B D = be fixed. Denote by S the set of all y B such
that there is at least one singular minimizing geodesic between x and y. The set S is
a compact subset of B, and the set of singular minimizing geodesic between x and
a point in S is compact with respect to the uniform topology. Then by the previous
observation (second case) together with a compactness argument (see Remarks 2.1,
2.2), we infer that an inequality of the form (3.26) holds for any minimizing control u
which is close enough to a control corresponding to a singular minimizing geodesic
joining x to a point in S . Denote by S the set of y in B corresponding to such
controls. By construction, any minimizing geodesic from x to a point in B \ S is
regular. Actually, it is far from being singular. Then by the arguments given in the first
case together with compactness arguments, an inequality of the form (3.22) holds
x,1
(u)) in B \ S . In that way, we prove that eSR (x, ) (or equivalently
for any y (= EF
dSR (x, )) is locally Lipschitz in M \ {x}. The same proof shows that eSR is indeed
uniformly locally Lipschitz with respect to one variable. We conclude easily.
Remark 3.15 The above arguments can be used to prove the following result. Let
(, g) be a sub-Riemannian structure of rank m < n on M. Assume that it is complete
and that there is an open set M M such that for every (x, y) with x = y,
no minimizing geodesic between x and y is a Goh path. Then dSR is Lipschitz in
charts on \ D.
Remark 3.16 Note that if the path is constant on [0, 1], it is a Goh path if and only
if there is a differential form p T(0) M satisfying
p X i ( (0)) = p X i , X j ( (0)) = 0
i, j = 1, . . . , m,
104
M := x M | x c (x) ,
S := M \ M = x M | x c (x) .
As shown by the following result, under classical assumptions on the measures
and Lipschitzness of the sub-Riemannian structure, static points do not move while
moving points obey a transportation law of the form (3.20)(3.21).
Theorem 3.17 Let (, g) be a Lipschitz sub-Riemannian structure on M and ,
be two compactly supported probability measures on M. Assume that is absolutely
continuous with respect to the Lebesgue measure. Then there is existence and
uniqueness of an optimal transport map from to for the SR quadratic cost
c : M M [0, +) defined by
c(x, y) :=
1 2
d (x, y)
2 SR
x, y M.
expx (Dx ) if x M ,
x
if x S ,
and for -a.e. x M there exists a unique minimizing geodesic between x and T (x).
Proof Let S Supp() Supp() and (, c ) be respectively the c-cyclically
monotone set and the c-pair of potentials satisfying (3.15)(3.17). Since the sets
Supp(), Supp() are assumed to be compact, both , c are indeed continuous
and the supremum and infimum in (3.15)(3.16) are attained. We check easily that
x M belongs to S if and only if (x) = c (x). Then M coincides with the
set
105
S ), there is r > 0 such that dSR (x, y) > 2r for any y c (x). In addition, since
the set c is closed in M M (again by continuity of , c and compactness of
S ), there exists a neighborhood Vx of x which is included in M such that
dSR (z, w) r
z Vx , w c (z).
1 2
x,r (z) := sup c (y) dSR
(z, y) | y 2 (S ), dSR (z, y) r .
2
By construction, coincides with x,r on Vx . By assumption, dSR is Lipschitz
in charts outside the diagonal, then by compactness of S we deduce that x,r is
Lipschitz in charts. In conclusion is Lipschitz in charts on M and (i) is proved.
To prove (ii), we observe that it suffices to prove the result for x belonging to
an open set V M on which the horizontal distribution (x) is parametrized by a
orthonormal family a smooth vector fields F = {X 1 , . . . , X m }. In fact, up to working
in charts, we can assume that V is a convex subset of Rn where the C 2 -norms of
the X i s are bounded. Let us fix a compact ball B in V and show that (ii) holds for
-a.e. x B.
Recall that the Hamiltonian H : V (Rn ) R which is associated to our
sub-Riemannian structure is defined by (see Chap. 2)
2
1
p X i (x)
2
m
H(x, p) :=
(x, p) V Rn .
i=1
p := v Rn | p v = 0 .
From Lemma 2.11 and its proof, for every x V and every p (Rn ) with H (x , p ) =
0, there is > 0 such that the Dirichlet problem
H(x, Dx S(x)) = H x , p ,
S|x+p = 0,
(3.27)
admits a solution of class C 1 on the ball B(x , ). We leave the reader to check that
the radius depends continuously on x , H(x , p) and |p| (|p| denotes the Euclidean
norm of p). Then, by compactness of B there is a function
: (0, +) (0, +) (0, )
which is decreasing in the first variable and increasing in the second variable such
that for every x B and every p (Rn ) with H (x , p ) = 0, the solution to (3.27) is
106
defined on the open ball B (x , (H(x , p ), |p|)). For any x , p satisfying the previous
assumptions, we denote by
Sx ,p : B x , H(x , p ), |p| R
the solution to the Dirichlet problem (3.27), with x ,p := (H(x , p ), |p|). The functions Sx ,p being constructed by the method of characteristics (see Proof of Lemma
2.11), the following result holds (note that the parametrization of characteristics that
we use in the statement of Lemma 3.18 differs from the one which is used to construct
Sx ,p , see last statement).
Lemma 3.18 There is a function
: (0, +) (0, +) (0, +)
which is increasing in the first variable and decreasing in the second variable such
that the following property holds (see Fig. 3.6):
n
For every x B, for every p (R ) with H (x , p ) = 0, and every x
B x , x ,p /2 there are
zx ,p (x) x + p B x , x ,p
and tx ,p (x) H(x , p ), |p| , H(x , p ), |p|
such that
x = x ,p tx ,p (x); zx ,p (x)
where (we set x , p := H(x , p ), |p| )
107
x ,p ; zx ,p (x) , px ,p ; zx ,p (x) : x ,p , x ,p V Rn
is the solution to the Hamiltonian system
H
x ,p t; zx ,p (x) =
x ,p t; zx ,p (x) , px ,p t; zx ,p (x)
p
t x ,p , x ,p .
For every x V , we denote by (x) the set of p (Rn ) such that H(x, p) = 0.
Pick a sequence {(xk , pk )}k of B (Rn ) which is a dense subset of
(x, p) B Rn | p (x) .
and set for every k,
k := xk ,pk , k := xk ,pk , tk () := txk ,pk (),
zk () := zxk ,pk (), k (, ) := xk ,pk (, ), pk (, ) := pxk ,pk (, ).
The following result is a consequence of the Lipschitz regularity of the subRiemannian distance along horizontal paths together with Rademachers theorem.
Lemma 3.19 There is a set N of Lebesgue measure zero in V such that for every
x B \ N and any k, the following property holds:
x B xk , k /2 and x = k t; zk (x)
= s k s; zk (x) is differentiable at t.
Proof (Proof
3.19) Let k be fixed. By construction, all the curves k (; z)
of Lemma
(with z x + pk B (xk , k )) are horizontal with respect to the distribution (we
may assume
withoutloss of generality that the curves k (; z) are defined on (k , k )
for all z x + pk B (xk , k )). The potential is expressed as
1 2
(x, y) | y 2 (S )
(x) = max c (y) dSR
2
x M,
108
:=
,
p
x,p
x,p
x,p (; x), px,p (; x) , the solution of the Hamiltonian system starting at (x, p); it is
defined on the interval ( (H(x, p), |p|), (H(x, p), |p|)).
Lemma 3.20 There is a function
C : (0, +) (0, +) (0, +)
which is decreasing in the first variable and increasing in the second variable such
that the following property holds:
For every h, R > 0, every k, and every (x, p) B (Rn ) satisfying
H xk , pk , H(x, p) > h, pk , |p| < R, x B xk , k /2 ,
(3.28)
k tk (x) + s; zk (x) x,p (s) C(h, R) pk tk (x); zk (x) p s,
(3.29)
one has
for every s ( (h, R), (h, R)) (tk (x) (h, R), tk (x) + (h, R)).
Proof (Proof of Lemma 3.20) Since the C 1 -norms of the X i s are bounded on V ,
there is an increasing function P : (0, +) (0, +) such that the solutions to
our Hamiltonian system starting from a pair (x, p) with x B, H(x, p) > h and
109
|p| < R remains in the set V B(0, P(R)) on the interval ( (h, R), (h, R)) (note
that since H is constant along the Hamiltonian trajectories, the solutions remains
in the set {H(x, p) > h}). Now, considering Lipschitz constants of the Hamiltonian
vector field on the cylinder V B(0, P(R)) (the C 2 -norms of the X i s are bounded
on V ) and using Gronwalls Lemma (see Appendix A), we prove easily the existence
of an increasing function C : [0, +) [0, +) such that
k tk (x) + s; zk (x) x,p (s) + pk tk (x) + s; zk (x) px,p (s)
C(R) pk tk (x); zk (x) p ,
(3.30)
for every h, R > 0, every k, and every (x, p) B (Rn ) satisfying (3.28), and
every s ( (h, R), (h, R))(tk (x) (h, R), tk (x) + (h, R)). Let us denote
by I the latter interval and set
u(s) := k tk (x) + s; zk (x) x,p (s)
s I.
Considering again the Lipschitz constants of the Hamiltonian vector field that we
always denote by K, we obtain formally for every s,
u(s) =
H
H
k tk (x) + r; zk (x) , pk tk (x) + r; zk (x)
x,p (r), px,p (r) dr
p
0 p
s
s
pk tk (x) + r; zk (x) px,p (r) dr,
K
u(r) dr + K
s
u(r) dr + K
C(R) pk tk (x); zk (x) p dr.
We are now ready to prove that for every x B \ N, we have c (x) = {x}. Fix
x B \ N and argue by contradiction, that is assume that there is y = x such that
y c (x) \ {x}. Then we have (remembering Remark 3.6)
(x) + c x, y (z) + c z, y
z M,
1 2 1 2
d z, y dSR x, y
2 SR
2
z M.
(3.31)
Since dSR is Lipschitz in charts outside the diagonal, there is a normal minimizing
geodesic joining x to y (see Remark 3.17), that is there is p Tx M such that
expx (p) = y and dSR (x, y)2 = 2H(x, p) = 0. Note that since x belongs to c (x),
we have
110
z V .
k (tk (x) + s; zk (x)) of length s 2H (zk (x), pk ). Since x does not belong to N, the
function
s k tk (x) + s; zk (x)
is differentiable at s = 0. Then Lemma 3.19 together with the previous inequality
allows us to write
d
k tk (x) + s; zk (x) |s=0 = 0.
ds
(3.32)
Since dSR is Lipschitz outside the diagonal and y = x, there are , K > 0 such that
2
2
z , y K|z z|
dSR z, y dSR
z, z B(x, ).
Then applying (3.31) with z = k tk (x) + s; zk (x) and s small and using (3.29)
yields (as in Lemma 3.20, x,p denotes the geodesic starting at x with initial covector
p, note that x,p (s) belongs to V for small s)
(x) k tk (x) + s; zk (x)
1 2
1 2
dSR
k tk (x) + s; zk (x) , y dSR
(x, y )
2
2
1 2
1 2
K
k tk (x) + s; zk (x) x,p (s) + dSR
x,p (s), y dSR
(x, y )
2
2
2
1 2
1
KC(h, R)
2
pk tk (x); zk (x) p s + (1 s)2 dSR
x, y dSR x, y
2
2
2
2 (x, y)
d
KC(h, R)
2
s2 .
pk tk (x); zk (x) p dSR
=
x, y s + SR
2
2
111
The quantity
KC(h, R)
pk tk (x); zk (x) p
2
tends to 0 as (xk , pk ) tends to (x, p). We infer
) close enough to (x, p),
that for (xk , pk
the derivative of the function s k tk (x) + s; zk (x) cannot be zero. This
contradicts (3.32).
It remains to prove the formula for T (x) and the uniqueness of minimizing geodesic between x and T (x) -almost everywhere. We need to show that
1
Dx
2
for all x M Supp() where is differentiable, which is the case for -almost
every x M by assertion (i) and Rademachers theorem. This is a consequence
of Lemma 2.15 applied to the function z (z) + c (y) at the point x with
y c (x). Moreover, again by Lemma 2.15, the geodesic from x to T (x) is unique
for -a.e. x M Supp(). Since T (x) = x for x S Supp(), the geodesic
is clearly unique also in this case.
Remark 3.18 If the sub-Riemannian structure is assumed to be ideal, then the potential can be shown to be locally semiconcave on the moving set.
Remark 3.19 The above arguments show that Theorem 3.17 remains true under
more general assumptions. Let (, g) be a complete sub-Riemannian structure on M
and , be two compactly supported probability measures in M with absolutely
continuous with respect to the Lebesgue measure. Assume that there are two open
sets 1 , 2 M with
(M \ 1 ) = 0
and
Supp() 2
1 2
d (x, y)
2 SR
x, y M.
112
Given a cost function, we shall say that the Monge problem is well-posed, if we have
existence and uniqueness of optimal transport maps from an absolutely continuous
compactly supported measure to a compactly supported measure. All the examples
that we review below have already been encoutenred within the text.
Fat distributions. Recall (see Example 1.15) that a distribution on M is called fat
if, for every x M and every section X of with X(x) = 0, there holds
Tx M = (x) + X, (x),
where
We saw that fat distributions do not admit non-trivial singular horizontal paths. This
means that any complete sub-Riemannian structure associated with a fat distribution
is ideal. In conclusion, by Theorem 3.17, the Monge problem for any sub-Riemannian
structure associated with a fat distributions is well-posed.
Two-generating distributions. A distribution is called two-generating if
Tx M = (x) + [, ](x)
x M.
:= x M | (x) + [, ](x) = R3 .
The set is called the singular set or the Martinet set of .
Proposition 3.21 Let be a totally nonholonomic distribution on a threedimensional manifold. Then, the set is a closed subset of M which is countably
2-rectifiable. Moreover, a non-trivial horizontal path : [0, 1] M is singular if
and only if it is included in .
Proof The first part will follow from Proposition 3.22 while the second part has
already been proved in Example 1.17.
Proposition 3.21 implies that for any pair (x, y) M M (with x = y) such that
x or y does not belong to , any sub-Riemannian minimizing geodesic between x
and y is nonsingular. Moreover has Lebesgue measure zero. As a consequence,
113
:= x M | (x) + [, ](x) = Tx M .
The following result holds.
Proposition 3.22 If is a nonholonomic distribution of rank n 1, then the set
is a closed subset of M which is countably (n 1)-rectifiable. Moreover, any Goh
path is contained in .
Proof The fact that is a closed subset of M is obvious. Let us prove that it is
countably (n 1)-rectifiable. Since it suffices to prove the result locally, we can
assume that we have
x V ,
(x) = Span X 1 (x), . . . , X n1 (x)
where V is an open neighborhood of the origin in Rn . Moreover, doing a change of
coordinates if necessary, we can also assume that (with coordinates (x1 , . . . , xn ))
X i = xi + i (x) xn
i = 1, . . . , n 1,
j
i
xi
xj
j
i
i
j
xn
xn
&
xn ,
114
j
j
i
i
+
= x V |
i
j = 0
xi
xj
xn
xn
i, j {1, . . . , n 1}} .
For every tuple I = (i1 , . . . , ik ) {1, . . . , n1}k we denote by X I the smooth vector
field constructed by Lie brackets of X 1 , X 2 , . . . , X n1 as follows,
'
X I = X i1 , X i2 , . . . , X ik1 , X ik . . . .
We call k = length(I) the length of the Lie bracket X I . Since is totally nonholonomic, there is some positive integer r such that
x V .
It is easy to see that, for every I such that length(I) 2, there is a smooth function
gI : V R such that
x V .
X I (x) = gI (x)xn
Defining the sets Ak as
(Ak \ Ak+1 ) .
k=2
By the Implicit Function Theorem, it is easy to see that each set Ak \ Ak+1 can be
covered by a countable union of smooth hypersurfaces. Indeed assume that some
given x belongs to Ak \ Ak+1 . This implies that there is some J = (j1 , . . . , jk+1 ) of
length k + 1 such that gJ (x) = 0. Set I = (j2 , . . . , jk+1 ). Since gI (x) = 0, we have
gJ (x) =
Hence, either
gI
xj1 (x)
= 0 or
gI
gI
(x) +
(x)j1 (x) = 0.
xj1
xn
gI
xn (x)
= 0.
k+1
length(I)=k
gI
x V | i {1, . . . , n} such that
(x) = 0 .
xi
115
We conclude easily.
The fact that any Goh path is contained in is obvious.
As a consequence by Remarks 3.13 and 3.19, the Monge problem for subRiemannian structures involving codimension one distributions is well-posed.
Rank-two distributions in dimension four. Let (M, , g) be a complete subRiemannian manifold of dimension four, and let be a regular rank-two distribution,
that is satisfying
Tx M
'
'
:= (x, y) M M | y
/ O(x) .
According to Remark 3.13, the following result holds:
Proposition 3.23 Under the assumption above, the function dSR is locally semiconcave in the interior of .
The above result allow us to obtain existence and uniqueness of optimal transport
maps in certain cases. Let us consider the distribution given in Example 1.18, that is
the distribution in R4 spanned by the vector fields
X 1 = x1 ,
X 2 = x2 + x1 x3 + x3 x4 .
t [0, 1].
By the above proposition, we deduce that, for any complete metric g on R4 , the
sub-Riemannian distance function dSR is locally semiconcave on the set
/ Span{e1 } ,
= (x, y) R4 R4 | (y x)
116
where e1 denotes the first vector in the canonical basis of R4 . Consequently, for any
pair of compactly supported probability measures , on M such that is absolutely
continuous with respect to the Lebesgue measure and
Supp ,
the Monge problem is well-posed.
117
118
is Lipschitz in charts on a dense subset of M. We do not know if this set has necessarily
full Lebesgue measure in M (note that the Sard Conjecture that we mentioned in
Sect. 2.6 would imply such a result). Anyway, such a result would not be sufficient to
prove the well-posedness of Monge problem for general sub-Riemannian structures.
References
1. Agrachev, A., Lee, P.: Optimal transportation under nonholonomic constraints. Trans. Amer.
Math. Soc. 361(11), 60196047 (2009)
2. Ambrosio, L., Rigot, S.: Optimal transportation in the Heisenberg group. J. Funct. Anal. 208(2),
261301 (2004)
3. Bernard, P., Buffoni, B.: Optimal mass transportation and Mather theory. J. Eur. Math. Soc.
9(1), 85121 (2007)
4. Billingsley, P.: Convergence of Probability Measures, 2nd edn. John Wiley & Sons Inc., New
York (1999)
5. Brenier, Y.: Polar factorization and monotone rearrangement of vector-valued functions. Comm.
Pure Appl. Math. 44, 375417 (1991)
6. Cannarsa, P., Rifford, L.: Semiconcavity results for optimal control problems admitting no
singular minimizing controls. Ann. Inst. H. Poincar Anal. Non Linaire 25(4), 773802 (2008)
7. Cannarsa, P., Sinestrari, C.: Semiconcave functions, Hamilton-Jacobi equations, and optimal control. Progress in Nonlinear Differential Equations and their Applications, vol. 58.
Birkhuser, Boston (2004)
8. Champion, T., De Pascale, L.: The Monge problem in Rd . Duke Math. J. 157(3), 551572
(2011)
9. Chitour, Y., Jean, F., Trlat, E.: Genericity results for singular curves. J. Differ. Geom. 73(1),
4573 (2006)
10. De Pascale, L., Rigot, S.: Monges transport problem in the Heisenberg group. Adv. Calc. Var.
4(2), 195227 (2011)
11. Evans, L.C., Gariepy, R.: Measure Theory and Fine Properties of Functions. CRC Press, Boca
Raton, FL (1992)
12. Figalli, A., Rifford, L.: Mass transportation on sub-Riemannian manifolds. Geom. Funct. Anal.
20(1), 124159 (2010)
13. Gangbo, W., McCann, R.J.: The geometry of optimal transportation. Acta Math. 177(2), 113
161 (1996)
14. Jean, F.: Control of nonholonomic systems and sub-Riemannian geometry. Lectures given at
the CIMPA School Gomtrie sous-riemannienne, Beirut, Lebanon (2012)
15. Kantorovitch, L.: On the translocation of masses. C.R. (Doklady) Acad. Sci. URSS 37, 199201
(1942)
16. McCann, R.J.: Polar factorization of maps on Riemannian manifolds. Geom. Funct. Anal.
11(3), 589608 (2001)
17. Mitchell, J.: On Carnot-Carathodory metrics. J. Differ. Geom. 21(1), 3545 (1985)
18. Monge, G.: Mmoire sur la thorie des dblais et des remblais. Histoire de lAcadmie Royale
des Sciences de Paris, pp. 666704 (1781)
19. Montgomery, R.: A tour of subriemannian geometries, their geodesics and applications. Mathematical Surveys and Monographs, vol. 91. American Mathematical Society, Providence, RI
(2002)
20. Pratelli, A.: On the equality between Monges infimum and Kantorovitchs minimum in optimal
mass transportation. Ann. Inst. H. Poincar Probab. Statist. 43(1), 113 (2007)
21. Rifford, L.: propos des sphres sous-riemanniennes. Bull. Belg. Math. Soc. Simon Stevin
13(3), 521526 (2006)
References
119
22. Rifford, L., Trlat, E.: On the stabilization problem for nonholonomic distributions. J. Eur.
Math. Soc. 11(2), 223255 (2009)
23. Rockafellar, R.T.: Characterization of the subdifferentials of convex functions. Pacific J. Math.
17, 497510 (1966)
24. Villani, C.: Optimal Transport, Old and New. Springer-Verlag, Heidelberg (2008)
Appendix A
We recall here without proofs basic facts on ordinary differential equations. For
further details, we refer the reader to the textbook [1].
A function f : [a, b] Rn is said to be absolutely continuous, if for each > 0,
there exists > 0 such that for each family of disjoints intervals {]ai , bi [}iN included
in [a, b], and satisfying
bi ai < ,
iN
we have
iN
a
d
f (s)ds t [a, b].
dt
121
122
square integrable derivative) in charts. Such a notion does not depend on the atlas
chosen to cover M.
The Gronwall lemma is a key tool to obtain estimates involving solutions of
differential equations.
Lemma A.1 (Gronwalls Lemma) Let > 0, : [0, ] : R be a continuous
function, and L 1 ([0; ], R). Assume that u : [0, ] R is a continuous function
satisfying
t
u(t) (t) +
(s)u(s)ds t [0, ].
t
0
(s)ds
s
0
(r )dr
(s)(s) ds t [0, ].
t
0
(s)ds
t [0, ].
a.e. t, x (t0 ) = x0 .
(A.1)
f (x) :=
123
|x| x R.
t2
t R.
4
124
a.e. t I, S(t0 ) = In ,
has a unique solution which is defined on I . In the same way, the Cauchy problem
Y (t) = Y (t)A(t), a.e. t I, Y (t0 ) = In ,
admits a solution defined on I . Hence, the function Z : I Mn (R) defined as
Z (t) := Y (t)S(t) for every t I , satisfies for almost every t I ,
(t) = S(t)0 + S(t)
f or a.e. t I, (t0 ) = 0
(A.2)
S(s)1 C(s)ds, t I.
(A.3)
t0
Appendix B
We recall here basic facts of first order calculus in normed vector spaces and less
basic facts of second order calculus. We refer the reader to textbook [2] for further
details on differential calculus in normed spaces. The results of second order calculus
are taken from the textbook [3].
T
= sup
T (u)
Y | u X,
u
X = 1 .
Let (X,
X ) and (Y,
Y ) be two normed vector spaces, U be an open subset
of X and let F : U X Y be a given mapping. Let u U . We say that F is
differentiable at u provided there is a continuous linear map Du F : X Y such that
for every > 0, there is > 0 such that
0 <
u u
X <
F(u) F u Du F u u
< .
u u
F(u) F u Du F u u
lim
= 0,
u u
uu
X
or
F(u) = F(u)
+ Du F u u +
u u
X o(1).
125
126
X ) be a normed vector
space, U be an open subset of X , and E : U Rn and C : U R two mappings
of class C 1 on U . Assume that u U satisfies the following property:
C u C(u) for every u U such that E(u) = E u .
Then there exist 0 R and Rn with (0 , ) = (0, 0) such that
Du E = 0 Du C.
Proof Define the mapping : U X R Rn by
(u) := (C(u), E(u)) , u U .
The mapping is of class C 1 on U . We claim that u is necessarily a critical point
of , that is Du is singular. We argue by contradiction. If u is not a critical point,
the continuous linear map Du : X R Rn is surjective. Then there exists a
linear subspace Y of X of dimension n + 1 such that the restriction of Du to Y is
an isomorphism. Let y1 , . . . , yn+1 be a basis of Y and B be an open neighborhood
of 0 in Rn+1 such that
u +
n+1
i yi U
= 1 , . . . , n+1 B.
i=1
The mapping
: B Rn+1
n+1
= 1 , . . . , n+1 u + i=1
i yi
127
T
= sup
T (u 1 , u 2 )
Y | u 1 , u 2 X,
u 1
X =
u 2
X = 1 .
Given an open set U X and a mapping F : U X Y , we define
D 2 F := D D F : U X L 2 (X, Y )
if it exists (where we identify L (X, L (X, Y )) with L 2 (X, Y ). If D 2 F exists and
is continuous on U , we say that F is of class C 2 on U . In this case, the second
derivative Du2 F is symmetric at any point, that is
Du2 F (v, w) = Du2 F (w, v) v, w X, u U .
If F : U X Y is a function of class C 2 then we have for every u U the
second order Taylor formula
F(u + h) = F(u) + Du F(h) +
1 2
D F (h, h) +
h
2X o(1),
2 u
F(v) F(u) Du F (v u) 1 D 2 F (v u, v u)
2 u
Y
lim
= 0.
vu
v u
2X
128
By the Inverse Function Theorem, any function of class C 1 is locally open around any
point with an invertible derivative. We are going to provide a second-order sufficient
condition for local openness around critical points.
Let (X,
X ) be a normed vector space, N be a positive integer, U be an open
subset of X and F : U R N be a mapping of class C 2 on U . Given a critical
point u U , we call corank of u, the quantity
corank F (u) := N dim Im Du F .
We also recall that if Q : X R is a quadratic form (that is Q is defined by
Q(v) := B(v, v) with B : X X R a symmetric bilinear form), we define its
negative index by
ind (Q) := max dim(L) | Q |L\{0} < 0 ,
where Q |L\{0} < 0 means
Q(u) < 0 u L \ {0}.
The following result provides a sufficient condition for local openness around a
critical point at second order.
Theorem B.3 Let F : U R N be a mapping of class C 2 in an open set U X
and u U be a critical point of F of corank r . If
ind Du2 F
|Ker(Du F)
r Im Du F
\ {0},
(B.1)
In the above statement, Du2 F |Ker(D F) refers to the quadratic mapping from
Ker(Du F) to R N defined by
Du2 F
|Ker(Du F)
u u
129
w1 Ker (Du F) ,
w1
X < ,
w2
X < 2 .
The proof of Theorems B.2 and B.3 that we give in the next sections are taken
from the Agrachev-Sachkov textbook [3] and the Agrachev-Lee article [4].
We need two preliminary lemmas.
Lemma B.5 Let G : Rk Rl be a mapping of class C 2 with G(0) = 0. Assume
that there is
v Ker(D0 G) with D02 G v , v Im D0 G ,
such that the linear mapping
w Ker(D0 G) ProjK
D02 G v , w K
(B.2)
is surjective, where K := Im(D0 G) and ProjK : Rl K denotes the orthogonal projection onto K . Then there is a sequence {u i }i converging to 0 in Rk such
that G(u i ) = 0 and Du i G is surjective for any i.
Proof Let E avector space in Rk such that Rk = E Ker(D0 G). Since D02 G v , v
belongs to Im D0 G there is v E such that
1
D0 G v = D02 G v , v .
2
Define the family of mappings { }>0 : E Ker(D0 G) Rl by
(z, t) :=
1 2
3
4
5
G
t
+
z
(z, t) E Ker(D0 G), > 0.
5
1
D 2 4 G(T ),
2 v + v
for any (Z , T ) E Ker(D0 G). For every (Z , T ) E Ker(D0 G), the first term
of the right-hand side D2 v +4 v G(Z ) tends to D0 G(Z ) as tends to 0 and since
130
2
2
1
1
2
4
4
o(1)
D
D
G(T
)
=
G(T
)
+
D
G
,
T
+
v
2
4
0
0
2 v + v
2
1
= 2 D02 G 2 v + 4 v , T + 2 v + 4 v o(1) ,
the second term tends to D02 G(v, T ) as tends to 0. By (B.2), the linear mapping
(Z , T ) E Ker(D0 G) D0 G(Z ) + D02 G v , T Rl
is surjective. Then there is > 0 such that D0 is surjective for all (0, ).
Therefore for every (0, ) the set
(z, t) E Ker(D0 G) | (z, t) = 0
is a submanifold of class C 2 of dimension k l > 0 which contains the origin. Then
there is a sequence {(z i , ti )}i converging to the origin such that 1/i (z i , ti ) = 0 and
D(zi ,ti ) 1/i is surjective for all i large enough. Thus setting
u i :=
1
1
1
1
v + 3 ti + 4 v + 5 z i i,
2
i
i
i
i
we get G(u i ) = 0 and Du i G surjective for all i large enough. This proves the
lemma.
Lemma B.6 Let Q : Rk Rm be a quadratic mapping such that
ind Q m, Rm \ {0}.
(B.3)
Then the mapping Q has a regular zero, that is there is v Rk such that Q(v) = 0
and Dv Q is surjective.
Proof Since Q is a quadratic mapping, there is a symmetric bilinear map B : Rk
Rk Rm such that
Q(v) = B(v, v) v Rk .
The kernel of Q, denoted by Ker(Q) is the set of v Rk such that
B(v, w) = 0 w Rk .
It is a vector subpace of Rk . Up to considering the restriction of Q to a vector space
E satisfying E Ker(Q) = Rk , we may assume that Ker(Q) = 0. We now prove
the result by induction on m.
In the case m = 1, we need to prove that there is v Rk with Q(v) = 0 and
Dv Q = 0. By (B.3), we know that ind (Q) 1 and ind (Q) 1, which
means that there are two vector lines L + , L in Rk such that Q |L + \{0} < 0 and
131
Q(w)
:= Proj F Q(w) w E,
= Q(w).
Q(w)
We claim that ind ( Q) m r , for every F \ {0}. As a matter of fact, by
assumption, for every F \ {0} there is a vector space L of dimension m such that
( Q)|L\{0} < 0. The space L E has dimension at least m k as the intersection
of L of dimension m and E of dimension k r in Rk . By induction, we infer that Q
Im(Dv Q) and
has a regular zero w E = Ker(Dv Q), that is Q(w)
w F
w E = Ker(Dv Q) Proj F B( w,
is surjective. Define F : Rk Rm by
F(u) := Q v + u u Rk .
The function F is of class C 2 verifies D0 F = Dv Q, D02 F = B and the assumptions
of Lemma B.5 are satisfied with v = w.
We deduce that Q has a regular zero as well.
Second case: Q 1 (0) = {0}.
In fact, we are going to prove that this case cannot appear. First we claim that Q is
surjective. Since Q is homogeneous (Q(r v) = r 2 Q(v) for all v Rk and r R),
we have
Q(Rk ) = r Q(v) | r 0, v Sk1 .
132
The set Q(Sk1 ) is compact, hence Q(Rk ) is closed. Assume that Q(Rk ) = Rm
and take x = Q(v) on the boundary of Q(Rk ). Then x is necessarily a critical point
for Q. Proceeding as in the first case, we infer that x = Q(w) for some non-critical
point. This gives a contradiction. Then we have Q(Rk ) = Rm . Consequently the
mapping
Q
: Sk1 Sm1
Q := |Q|
Q(v)
v |Q(v)|
is surjective. By Sards Theorem, it has a regular value x, that is x Sm1 such that
Dv Q is surjective for all v Sk1 satisfying Q(v) = x for all v Sk1 . Among the
set of v Sk1 such that Q(v) = x take v for which |Q(v)| is minimal, that is such
that
Q(v) = ax
and
a > 0, v Sk1 ,
Q(v) = ax = a a.
x = 0 .
(B.4)
1
Dv Q(h) + a [Dv |Q|(h)] x.
a
133
k1 of dimension k m in a neighborhood
is a smooth
submanifold of (0, +) S
of a,
v . Then for every (h a , h v ) Ker(Da,
v ), which is equivalent to h a = 0 and
Dv Q(h v ) = 0 with h v Tv Sk1 , there is a smooth curve = (a , v ) : (, )
C such that (0) = (a,
v ) and (0) = (h a , h v ). Then differentiating two times the
2
equality ( (t)) = 0 and using that a2 = 0 and
v ) = Dv Q = 0, we
v (a,
get
2
a,
v = (0) x = (0).
2 a,
v = (0)
v
a
Q(h) 0 h Ker Dv Q Tv Sk1 .
Since Q(v) = a > 0 we have indeed
Q(h) 0 h Ker Dv Q Tv Sk1 Rv =: L .
Let us compute the dimension of the non-negative subspace L of the quadratic form
Q. Since Dv Q is surjective, we have
dim Im Dv Q = m 1.
Which means (remember (B.4)) that Im Dv Q |Sk1 has dimension m or m 1. But
Dv Q = 0 with = 0, thus we have necessarily
dim Im Dv Q |Sk1 = m 1
and
dim Ker Dv Q Tv Sk1 = dim Ker Dv Q |Sk1 = k 1 (m 1)
= k m.
Consequently, dim(L) = k m + 1, thus ind ( Q) has to be m 1, which
contradicts the hypothesis of the lemma. This shows that Q 1 (0) = {0} is impossible
and concludes the proof of the lemma.
134
By continuity of the mapping Du2 F |E , there is an open set O S such
that
Du2 F |E \{0} < 0 O .
I
Oi
i=1
|Ker(Du F)
with
r = corank F u := N dim Im Du F = N dim Im Du F .
K
and define the quadratic mapping Q : Ker(Du F)
Set K := Im Du F
by
Du2 F (v, v) v Ker Du F ,
Q(v) := ProjK
where ProjK : R N K denotes the orthogonal projection onto K . The assumption (B.3) of Lemma B.6 is satisfied. Then by Lemmas B.6, Q has a regular zero,
such that
that is v Ker(Du F)
Q v = 0
Du2 F v , v K = Im Du F
and
Dv Q surjective
w Ker Du F ProjK Du2 F v , w K surjective.
135
u + v) F(
u)
Setting G(v) := F(
and applying Lemma B.5, we get a sequence
and Du i F is surjective for any i. By
{u i }i converging to u such that F(u i ) = F(u)
the Inverse Function Theorem, this implies that F is locally open at u.
D02 F (v, v) v Ker (D0 F) ,
1 2
D0 F (t, t) (z, t) E Ker(D0 F).
2
Then assumptions of Lemma B.5 are satisfied and there is a sequence {(z i , ti )}i
converging to 0 such that G(z i , ti ) = 0 and D(zi ,ti ) G is surjective for all i.
Lemma B.7 There are , c > 0 such that the image of any continuous mapping
G : B(0, 1) R N with
G(u) | u = (z, t) B X (0, 1)
sup G(u)
(B.5)
c).
G (y) B X u i , /4 y B(0,
There is > 0 such that any continuous mapping G : B X (0, 1) R N verifying
(B.5) satisfies
G(u)
G |V (B) u B X (u i , /2) V
136
u B X (u i , /2) V.
G G (u) u
4
and
G(u)
= y.
1 2
F
z
+
t
(z, t) E Ker(D0 F), > 0.
2
z = w1 + w2 , w1 Ker(Du F),
w1
X < ,
w2
X < 2 .
Let us now show that the above result holds uniformly for u close to u = 0. Since F
Moreover, again
is C 1 , the vector space Ker(Du F) is transverse to E for u close to u.
),
by C 1 regularity, for every > 0, there is > 0 such that for every u B X (u,
Ker(Du F) B(0, 1) y + z X | y Ker(Du F) B(0, 1),
z
X < .
137
G(z,
t) := 2 F u + 2 z + 2 1 (t) + 2 (t) F(u) ,
for every (z, t) E Ker(D0 F) B(0, 1). Taking and > 0 smaller if necessary,
by Taylors formula for F at u at second order, by the above construction and by the
fact that Du F and Du2 are respectively close to D0 F and D02 F, we may assume that
(B.5) is satisfied. We conclude easily.
References
1. Hirsch, M.W., Smale, S.: Differential Equations, Dynamical Systems, and Linear Algebra. Pure
and Applied Mathematics, vol. 60. Academic, New York (1974)
2. Abraham, R., Marsden, J.E., Ratiu, T.: Manifolds, tensor analysis, and applications. Global
Analysis Pure and Applied: Series B, 2. Addison-Wesley Publishing Co., Reading (1983)
3. Agrachev, A.A., Sachkov, Y.L.: Control Theory from the Geometric Viewpoint. Encyclopaedia
of Mathematical Sciences, vol. 87, Springer, Heidelberg (2004)
4. Agrachev, A., Lee, P.: Optimal transportation under nonholonomic constraints. Trans. Amer.
Math. Soc. 361(11), 60196047 (2009)
5. Bredon, G.E.: Topology and Geometry. Graduate Texts in Mathematics, vol. 139. Springer, New
York (1993)
Index
Symbols
c-convex, 84
c-cyclically monotone, 82
c-subdifferential, 86
c-transform, 84
C
Concatenation
of controls, 16
of paths, 23
Control, 13
regular, 20
singular, 20
strictly abnormal, 68
Critical point, 60, 126
D
Degree of nonholonomy, 10
Distribution, 1
codimension-one nonholonomic, 113
contact (and contact form), 11, 25
fat, 24, 112
Martinet, 11, 25
medium-fat, 69, 113
totally nonholonomic, 9
trivial, 1
two-generating, 112
E
End-Point mapping, 13
Extremal
abnormal, 21
normal, 47, 52
F
Frame
global, 2, 12
local, 1
orthonormal, 34
Function
absolutely continuous, 121
absolutely continuous with
integrable derivative, 121
cost, 77
Lipschitz in charts, 58, 94
locally semiconcave, 97
square
G
Generating family, 4
orthonormal, 34
Goh condition, 68
path, 68
Gronwalls Lemma, 122
H
Hrmander condition, 9
Heisenberg group, 70
Horizontal path, 12
length (of a), 34
minimizing, 37
regular, 23
singular, 23
K
Kantorovitch potential, 85
L
Lie
139
140
Lie (cont.)
algebra, 8
bracket, 5, 6
M
Martinet
set, 112
surface, 25, 27
Method of characteristics, 49
Minimizing geodesic, 38
singular, 46
strictly abnormal, 68
N
Negative index (of a quadratic form), 60, 128
O
Optimal transport
plan, 82
Optimal transport problem
Kantorovitch, 80
Monge, 77
R
Rank
of a control, 16
of an horizontal path, 23
S
Set
contact, 86
Index
moving, 104
singular, 112
static, 104
Singular, 20
Sub-differential, 92
Sub-Riemannian
ball, 34
distance, 34
energy, 37
exponential map, 52
Hamiltonian, 47, 52
norm, 34
Sub-Riemannian structures, 33
complete, 43
ideal, 97
Lipschitz, 100
Sub-TWIST condition, 93
super-differential, 92
Support (of a measure), 81
T
Theorem
Brenier, 94
Cauchy-Carathodorys, 123
Cauchy-Peano, 122
Chow-Rashevsky, 32
Inverse Function, 126
Lagranges Multipliers, 126
McCann, 95
SR Hopf-Rinow, 40
Trajectory, 13
Transport
map, 77
plan, 80