Arnold - Ergodic Problems of Classical Mechanics

ERGODIC PROBLEMS
OF
CLASSICAL MECHANICS
THE MATHEMATICAL PHYSICS MONOGRAPH SERIES

A. S. Wightman,
EDITOR
Princeton University
Ralph Abraham, Princeton University

FOUNDATIONS OF MECHANICS
Vladimir I. Arnold, University of Moscow

Andre Avez, University of Paris
ERGODIC PROBLEMS OF CLASSICAL MECHANICS
Freeman .J. Dyson, The Institute for Advanced Study

SYMMETRY GROUPS IN NUCLEAR AND PARTICLE PHYSICS
Robert Hermann, Argonne National Laboratory

LIE GROUPS FOR PHYSICISTS
Rudolph C. Hwa, State University of New York at Stony Brook

Vigdor L. Teplitz, Massachusetts Institute of Technology
HOMOLOGY AND FEYNMAN INTEGRALS
.John R. Klauder and E. C. G. Sudarsban, Syracuse University

FUNDAMENTALS OF QUANTUM OPTICS
Andre Lichnerowicz, College de France

RELATIVISTIC HYDRODYNAMICS AND MAGNETOHYDRODYNAMICS
George W. Mackey, Harvard University

THE MA THEMA TICAL FOUNDATIONS OF QUANTUM MECHANICS
Roger G. Newton, Indiana University

THE COMPLEX j-PLANE
R. F. Streater, Imperial College of Science and Technology

A. S. Wightman, Princeton University
PCT, SPIN AND STATISTICS, AND ALL THAT
ERGODIC PROBLEMS
OF
CLASSICAL MECHANICS
V. I. ARNOLD
University of Moscow
and
A. AVEZ
University of Paris
Library of Congress Catalog Card Number 68-19936

Manufactured in the United States of America
The manuscript was put into production on October 10,1967;

this volume was published onJuly 5,1968
PREFACE
The fundamental problem of mechanics is computing, or studying

qualitatively, the evolution of a dynamical system with prescribed initial data.
Numerical methods allow one to compute the orbits for a finite time
interval, but they fail as the time increases indefinitely. The three-body
problem offers a typical example: Do there exist arbitrarily small perturbations of the initial data for which one of the bodieii moves to infinity? Mathematically speaking, the problem is the study of the orbits of a vector field on
phase-space. Far from being solved, such a problem involves areas as various
as probability and topology, number theory and differential geometry. Mr.
Nicholas Bourbaki may forgive us for mixing so many structures.
Maxwell, Boltzmann, Gibbs, and Poincare first proposed a statistical
study of complex dynamical systems, which is now known as ergodic theory.
[Ergodic theory was conceived for mechanics but applies to various other
branches, such as number theory. For example, how are the first digits 1, 2,
4, 8, 1, 3, 6, ... of the powers 2" distributed? (See Appendix 12.)] But the
mathematical definitions and the first important theorems are due to 1. von
Neumann, G. D. Birkhoff, E. Hopf, and P. R. Halmos, and they appeared
only in the thirties. During the past decade, a new step was taken, inspired by
Shannon's information theory. The main result, due to Kolmogorov, Rohlin,
Sinai, and Anosov, consists in a deep study of a strongly stochastic class of
dynamical systems. This class is wide enough to include all the sufficicntly
unstable classical systems. Among these systems figure the geodesic flows of
space with negativc t"tfrvature, as studied by Hadamard, Morse, Hcdlund,
E. Hopf, Gelfand, Fomin. On the other hand, Sinai proved that the Boltz-
vi
PREFACE
mann-Gibbs model, that is, a system of hard spheres with elastic collisions,
belongs also to this class; this proves the "ergodic conjecture."
This book is by no means a complete treatise on ergodic theory, and
references are not exhaustive.
The text presented here is based on lectures delivered during the spring
and fall of 1965 by one of the authors, who also wrote Chapter 4. The second
author is responsible for the proofs of Chapters 1,2, and 3.
We thank Professors Y. Choquet-Bruhat, H. Cabannes and P. Germain,
J. Kovalewsky, G. Reeb, L. Schwartz, R. Thorn, and M. Zerner, who welcomed the lecturer at their seminar. We also thank Professor S. Mandelbrojt,
who suggested that we write this book. The final manuscript was read by Y.
Sinai, who made a number of useful improvements for which we are sincerely
grateful.
The translator (A Avez) wishes to thank warmly Professors V. 1.
Arnold, S. Deser, and A S. Wightman, who prevented him from many
mistakes.
V. I. ARNOLD
A AVEz
CONTENTS
Preface
Chapter 1.
Chapter 2.
Dynamical Systems
1. Classical Systems
2. Abstract Dynamical Systems
3. Computations of Mean Values
4. Problems of Classification. Iwmorphism of
Abstract Dynamical Systems
5. Problems of Generic Cases
General References for Chapter 1
Ergodic Properties
6. Time Mean and Space Mean
7. Ergodicity
8. Mixing
9. Spectral InvaI:.iants
10. Lebesgue Spectrum
11. K -Systems
12. Entropy
vii
1
1
7
9
11
12
13
15
15
16
19
22
28
32
35
51
CONTENTS
viii
Chapter 3.
Chapter 4.
Appendixes.
Unstable Systems
13. C-Systems
14. Geodesic Flows on Compact Riemannian
Manifolds with Negative Curvature
15. The Two Foliations of a C-System
16. Structural Stability of C-Systems
17. Ergodic Properties of C-Systems
18. Boltzmann-Gibbs Conjecture
Stable Systems
19. The Swing and the Corresponding
Canonical Mapping
20. Fixed Points and Periodic Motions
21. Invariant Tori and Quasi-Periodic Motions
22. Perturbation Theory
23. Topological Instability and the Whiskered Tori
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
The Jacobi Theorem

Geodesic Flow of the Torus
The Euler-Poinsot Motion
Geodesic Flows of Lie Groups
The Pendulum
Measure Space
Isomorphism of the Ilaker's
Transformation and B (1/2, 112)
Lack of Coincidence Everywhere of
Space Mean and Time Mean
The Theorem of Equipartition Modulo
Some Applications of Ergodic Theory to
Differential Geometry
Ergodic Tran"slations of Tori
The Time Mean of Sojourn
The Mean Motion of the Perihelion
Example of a Mixing Endomorphism
Skew-Products
Discrete Spectrum of Classical Systems
Spectra of K-Systems
53
53
60
62
64
70
76
79
81
81
86
93
100
109
114
-tI5
117
119
120
121
123
125
127
129
131
132
134
138
143
145
147
153
CONTENTS
18. Conditional Entropy of a Partition a with

Respect to a Partition f3
19. Entropy of an Automorphism
20. Examples of Riemannian Manifolds with
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
Negative Curvature
Proof of the Lobatchewsky-Hadamard Theorem
Proof of the Sinai Theorem
Smale Construction of C-Diffeomorphisms
Smale's Example
Proof of the Lemmas of the Anosov Theorem
Iritegrable Systems
Symplectic Linear Mappings of Plane
Stability of the Fixed Points
Parametric Resonances
The Averaging Method for Periodic Systems
Surfaces of Section
The Generating Functions of Canonical Mappings
Global Canonical Mappings
Proof of the Theorem on the Conservation of
Invariant Tori under Small Perturbations of
the Canonical Mapping
ix
158
163
168
178
191
194
196
201
210
215
219
221
227
230
235
243
249
Bibliography
271
Index
283
CHAPTER 1
DYNAMICAL SYSTEMS
This chapter contains examples of dynamical systems and related problems.
l. Classical Systems
DEFINITION
1.1
Let M be a smooth manifold, /l. a measure on M defined by a continuous positive density, 1>t: M .... M a one-parameter group of measure-preserving diffeomorphisms. The collection (M, /l., 1>t) is called a classical
dynamical system.
The parameter t is a real number or an integer. If t ( R, the group
1>t
is usually defined in local coordinates by:
xi = fi(x 1 , ... , JIl),

If t ( Z,
1>t
feomorphism
1>
= 1, ... ,
= dimM.
is the discrete group generated by a measure-preserving dif-
1>
1> 1-
Then the system is merely denoted by (M, /l.,
and
is called the automorphism.
EXAMPLE
1.2. QUASI-PERIODIC MOTION
Let M be the torus I (x, y) mod
11-
The measure is dx dy, the group t
is a translation group:
x=l,
y=a
where a ( R, and dot denotes d/dt. Assume a

p, q
(Z,
p/q rational:
>0
and p and q relatively prime. In the covering plane (x, y), the orbit with
initial data x(O) = x o' y(O) = Yo has the form:

Y = Yo +
~(x-xo)
q
As x = Xo + q, y takes the value Yo + p and the corresponding point on

M coincides with the initial point (x o' Yo)' Thus, the torus is covered by
closed orbits. If a is irrational, each orbit is everywhere dense (Jacobi,

1835; see Appendix 1). More generally, let Tn = !(x 1, ... , xn) mod 11 be
the n-dimensional torus with the usual measure dx 1 .. dx n, and t the
I
I
-r-------I
I
I
I
-,--I
I
-T--------
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
Figure 1.3
one-parameter group of translations defined by:

Xi=W i ;
;=l, ... ,n; ,wlR n
Every orbit of t is everywhere dense if, and' only if,

k ( Z" and w k = 0 imply k = 0
DYNAMICM, SYSTEMS
EXAMPLE
1.4. GEODESIC FLOWS
Let V be a compact Riemannian manifold; M
=.
tary tangent bundle. Given a unit tangent vector ~
Tl V denotes its unif
Tl Vx to V at x,
there is one, and only one, geodesic y passing through x with initial velocity vector ~. We denote by y (~, s) the point of y obtained from x in
time s when moving along y with velocity 1. The unit tangent vector to
y at y(~, s) is
(1.5)
Formula (1.5) defines a one-parameter group of diffeornorphisms of M
Tl V.
DEFINITION 1.6
The group G t is called the geodesic flow of V. It can be proved that

G t preserves the measure Il induced on M by the Riemannian metric of V
(Liouville's theorem).
SOME MORE EXAMPLES 1.7
. Appendix 2 describes the geodesic flow of the usual torus immersed in

the Euclidean space 3. For the ellipsoid see Kagan [1], and for Lie
groups with a left-invariant metric see Appendixes 3 and 4. One more word,
in mechanics geodesic flow is called "movement of a material point on a
frictionless surface without external forces. "
Other mechanical systems
involve more general flows.

EXAMPLE
1.8. HAMILTONIAN FLOWS
Let PI"'" Pn; ql"'" qn (in short: p, q) be a coordinate system in
R2n , and
(1.9)
H(p, q) a smooth function.
The equations
dq
aH
dp
dt
ap
dt
define a one-parameter group of diffeomorphisms of R2n. This group is

called a Hamiltonian flow on R2n.
LIOUVILLE'S THEOREM
1.10
The Hamiltonian flow preserves the measure dP l
...
dP n . dq1 dqn
Proof:
The divergence of the vector field (1. 9) vanishes:
...P-.
aq
(aH ) +
(_ aH) - 0
ap,
ap
. aq
-
THEORil!M OF CONSERVATION OF ENERGY
(Q. E. D.)
1.11
The function H is a first integral of (1.9).

Proof:
dH
dt
aH aH + aH (_ aH) = 0
aq
ap
ap
aq
Lee us denote a subset H(p, q)
(Q. E. D.)
h ( R by M. For almost every h,
M is a manifold. This manifold is invariant under the flow.

COROLLARY
1.12
There exists an invariant measure on the manifold M.

Proof:
The invariant measure on M is defined by:

dp. =
da
I grad HII
length,
where a is the volume element of M induced by the metric of R2n. If(1.9)

has several first integrals, namely 11 ,12 "", Ik , then the equations (1.9)
determine a classical dynamical system on each (2n ..., k)-dimensional manifold: II
EXAMPLE
hI'"'' Ik = hkJ where the h's are constants.
1.13. LINEAR OSCILLATIONS IN DIMENSION 2
The Hamiltonian is:
Equation (1. 9) has two first integrals:

II =
p/ + q/,
The corresponding manifolds II
12 =
pi + qi
hI' 12 : h2 are two-dimensional tori.
DYNAMICAL SYSTEMS
The dynamical systems that are induced on these tori are isomorphic to
those of Example (1.2). Appendix 5 provides further examples.
REMARK 1.14. GLOBAL HAMILTONIAN FLOWS
More generally one may consider a symplectic
2n-dimensional mani-
fold M2n instead of H2n , and a closed one-form w 1 (= dH) instead of H.

Equation (1. 9) becomes
x=
where /: T* Mx
for any ~
/w 1 '
M2n
TM x is defined by
TM x' Let us now give some examples of the discrete case:
Z.
EXAMPLE 1.15. TRANSLATIONS OF THE TORUS

Let M be the torus
!(x, y) mod 1\
with the usual measure dx dy. The
automorphism is
(x,y) = (x+wl' y+( 2 )(modl),
w1
H.
Each orbit of is everywhere dense if, and only if, k w ( Z, for k

imply k
i';
0 (see Appendix 1).
EXAMPLE 1.16. AUTOMORPHISMS OF THE TORUS

Again M =
! (x, yY mod 1\
and dll = dx dy. The automorphism is de-
fined by:
(x, y) = (X+ y, x+ 2y) (mod 1) .
The mapping induces a linear mapping in the covering plane (x, y)
- (1 1)
As Det
1 2
1, is measure-preserving. A set A is transformed under
A symplectic manifold M 2n is a smooth manifold, together with a global
closed two-form
of rank n. Example;
dpl\dq
on R 2n .
. and then
as pictured in Figure (1.17). The linear mapping has
two real proper values A1 and A2 :
< A2 < 1 < Al
)I
Figure 1.17
DYNAMICAL SYSTEMS
nA
Then, for n large enough,
looks like a very long and very narrow rib-
bon of the plane. On M, this ribbon lies approximately in the neighborhood of an orbit of the system:
x=
According to Jacobi's
1,
the~rem
y=,\-l:
(Example 1.2), and because .\1-1 is irra-
tional, 1>n A converges to a dense helix of the torus as n ... + "".
2.
DEFINITION
Abstract O)'1lamicaJ Systems
2.1 2
An abstract dynamical system (M, Il' 1>/) is a measure-space (M, Il)

equipped with a one-parameter group 1>/ of automorphisms (mod 0) of
(M, Il), 1>/ depending measurably of t.
Thus, for any measurable sets A and B, 1l(1)/A nB) is a measurable

function of t, and 1l(1)/A)
= Il(A)
for any t. In the future (M, Il) will al-
ways be a nonatomic Lebesgue space, that is (M, fL) will be isomorphic

modulo 0 to [0, 1] with its usual Lebesgue measure. In particular Il (M)
1.
If
1>/
is the discrete group generated by an automorphism
merely denote the system by (M, Il'
1.
1>
1>1'
we
In the following we shall omit the
notation "mod 0." All of the preceding examples are abstract systems:
a compact Riemannian manifold M ~ith its canonical measure Il (Il(M)
= 1)
is isomorphic to [0, 1].

EXAMPLE 2.2. BERNOUILLI SCHEMES
The space M.
Let Zn
10, 1, ... ,n-11 be the first n nonnegative integers.
M is the Cartesian product M
Z nZ of a countable family of Zn 's. Thus,
the elements m of M are the bilateral infinite sequences of elements of

Zn:
m (M,
2 See Al'pendix 6 for these concepts.
The a-algebra of the measurable sets. It is the algebra generated by set!!,
of the form
i ( Zn
i (Z,
The measure p.. Define a normalized measure p.. on
Zn by setting:
p.(O) = PO'"'' p.(n-l) = Pn - 1 , L P j = 1.
We set p. (A II) = p.I for every i,
i.
The measure of a is the product-mea-
sure, denoted again by p.: if A!I

(iI'"'' i k all different) are kII '''., A!k
Ik
distinct generators, the measure of their intersection is the product oftheir
measures, that is,
p.lm I a j
iI''''' a jk
p .... p . .
ikl
II
lk
(M, p.)' is clearly a Lebesgue space.

The automorphism cp. It is the shift
m = (.",8 j ,,,.)
where a;= a j _
..
m'= (".,a;,,,.),
for every i,' cp is a bijection.
To prove cp is measlire-
preservi'ng, it is sufficient to take into account the generators:
Hence:
p.[cp(Aj)
p.[A!H)
Pj
= p.(Aj).
Notation. The above abstract dynamical system is called a Bernouilli
scheme and denoted by B(p o'"'' Pn- 1 ).

Remark. Tossing a coin involves the scheme B ('12,'12) .This fact was
first pointed out by
J.
Bernouilli. The elements of M
Z2Z ar,e indefinite
bilateral sequences of tosses: 0 means "head," 1 means "tail." The

set A~(resp.
A~) represents the set of the sequence in which "head"
I
~
(resp. "tail") appears at the
ith
toss. Thus, it is quite natural to set:
p.(AI> =prob (AI> =
EXAMPLE
2.3.
2' .
THE BAKER'S TRANSFORMATION
Let M be the torus I(x, y) mod
11
with its usual measure dxdy.
DYNAMICAL SYSTEMS
'J
1~
____________. -____
------~~
<p'A
Figure 2.4
The automorphism ' is defined as;
'" '( x, Y) _
j (2x,
- 'I
<~
mod 1, if Y2 S x < 1 .
mod 1, if 0 S x
Y2y)
Of'
f (2x, Y2(y+ 1))
To study ', it is convenient to introduce the induced mapping ' in the

covering plane (x, y); ' can be described as follows: Transform the unit
square by the affinity, making it twice as long as before in the direction
ox, twice as short in the direction oy. Then, cut off the right half of this
rectangle and move it, by translation, above the left half (see Figure 2.4).
As A C M and n converges to + "", 'n A is formed by a "very large
number" of segments which are parallel to ox.
Let us mention some problems on dynamical systems.
3. Computations of Mean Values

EXAMPLE
3.1
Lagrangei1], studying 'the problem of mean motion of the perihelion,

raised the following question: compute, if it exists
3
lim! Arg
/ ... "" t
l
k= 1
a k ' exp (icukt)
10
where the a k ,
(Uk
are constants and Arg z means "argument of the com-
plex number z."

EXAMPLE
3.2
To the sequence !2" In = 1,2, ... 1 we make correspond the sequence

of the first digits of the 2"'s:
1,2,4,8, i, 3, 6, ...
Let d7, N) be the number of 7's included in the first N terms of this new
sequence. Compute (if it exists)
.
r<7, N)
I 1m
---
N~oo
EXAMPLE
3.3
Let D be a region of a Riemannian space, and y (t) a geodesic. What

is the mean sojourn time of y (t) in D? In other words, we set
r<T)
= measureltlO~
t~ T,
y(t) (DI
and we ask for

r(T)
lim - T
T~oo
These three problems are special cases of a more ,general one: Let f be a
complex-valued, /l-summable function defined on the space M of an abstract system (M, /l, t). Compute (if it exists)
lim
T~oo
1...
T
IT
!(tm)dt,
m ( M.
In (3.3), ! is none other than the characteristic function of Tl D. Of course,

there exist more problems involving some computations of means:
EXAMPLE
3.4
Let (M, /l, t) be a dynamical system, and A and B two measurable

sets of M. Compute (if it exists)
lim /l [ tA nB]
t ... + 00
(see Figures 1.17 and 2.4). Physical intuition makes it plausible that for
11
DYNAMICAL SYSTEMS
a probabilistic "enough" system, the limit exists and is equal to
4. Probl('ms of ~Iassification
Isomorphism of Abstract Dynamical Systems
A natural way to classify the dynamical systems is to exhibit their invariants with respect to their corresponding group: ccnonical transformations for Hamiltonian flows, measure-preserving diffeomorphisms for general classical flows. Since abstract invariants are the deepest ones, we
give the definition:
DEFINITION
4.1
Two abstract dynamical systems (M, /1, ) and (M', /1" ') are isomorphic if there exists an isomorphism f: M .. MI (mod 0) of measurable
spaces making the following diagram commutative:
A similar definition holds in the continuous case.

EXAMPLE
4.2
The Bernouilli schemes B (112, 1/8, 1/8, 1/8, 1/8) and B (1/4, 1/4, 1/4,
1/4) are isomorphic (see Meshalkin
EXAMPLE
[1], Blum and Hanson [1]).
4.3
The translations of the tori (1.1S) are not isomorphic to the automorphism of the torus (1.16) (see Chapter 2. 12.40).
EXAMPLE
4.4
On the torus T2
I(x, y) mod
11
with the usual measure, let us con-
sider the automorphisms:

(x,y) = (3x+y,2x+y) (mod 1);
'(x,y)
(3x+ 2y, x+ y) (mod 1) .
12
Both of them are nonisomorphic to the automorphism (1.16) (see Corollary

12.30), but their isomorphism is still an open question.
EXAMPLE
4.5
The Bernouilli scheme B (Y2, 12) is isomorphic to the Baker's transform

(see Appendix 7 for a proof). One problem among the fundamental problems
of Ergodic Theory is to find necessary and sufficient conditions under
which two Bernouilli schemes are isomorphic.
5. Problems of Generic Cases

Faced with such a diversity of dynamical systems, it seems useful to
clarify the situation by neglecting the "exceptional cases." The word
"exceptional" becomes meaningful by putting a topology or a measure
over the group of the automorphisms. A class of dynamical systems can
be exceptional in the abstract frame, and generic in the classical one, or
conversely.
EXAMPLE
5.1
There exist abstract dynamical systems which are nonisomorphic to

any classical system (see 12.39).
EXAMPLE
5.2
In the abstract frame, the mixing is exceptional in the weak topology

(see Halmos [1], Rohlin [1]), By contrast, every diffeomorphism, C 1 _ close
enough to the automorphism ch: T2
->
T2 of Example (1.16), is mixing.
Thus, mixing can be generic in the classical frame.

EXAMPLE
5.3
In the abstract frame, ergodic systems are generic in the weak topology (see Halmos [1]). By contrast, every Hamiltonian systemcIose enough
to the geodesic flow of the torus T2 (see Appendix 2) is nonergodic. See
also the three-body system (Chapter 4). Thus, ergodicity can be nongeneric in the classical frame.
DYNAMICAL SYSTEMS
13

Abraham, R., Foundations of Mechanics, Benjamin (1967).
Birkhoff, G. D., Dynamical Systems, American Mathematical Society Colloquium Publications 9, New Y9rk (1927).
Godbillon, C., Geometrie differentielle et mecanique, Hermann, Paris
(1968).
Halmos, P. R., Measure Theory, Chelsea, New York (1958).
Halmos, P. R., Lectures on Ergodic Theory, Chelsea, New York (1959).
Whittaker, E. T., Analytical Dynamics, Dover, New York (1944).
CHAPTER 2
ERGODIC PROPERTIES
A series of concepts (ergodicity, mixing, spectrum, entropy, etc.) has
been introduced by the metric theory of dynamical systems to describe the
behavior of most of the orbits. This chapter is devoted to th'eir definition,
In Chapters 3 and 4 we shall make use of these concepts to describe classical systems.
6, Time Mf'an and Space Mean
DEFINITION
6.1. TIME MEAN
Let (M,Il, ) or (M, 11, t) be a dynamical system, f a complex-valued

function defined on M.
(6.2)
The time mean f of f, if it exists, is defined by:
*
[(x)
M,
in the discrete case, and by:
(6.2),
x ( M,
[(x)
in the continuous case.

DEFINITION
6.3. SPACE MEAN
It is defined, if it exists, as:
(Recall that 11 (M)
1.)
15
f(x)dll
16
THEOREM
6.4. (G. D. BIRKHOFF- A. J. KHINCHIN)l
Let (M, p., t) be an abstract dynamical system, I
L1 (M, p.) a com-
plex-valued p.-summable function on M. Then:
* exists almost everywhere (abbreviation a.e.), that is except, per(a) I(x)

haps, on a set of measure zero;
* is summable and invariant a.e., that is:

(b) I(x)
*
*
f(tx)
= I(x)
for every t,
except, perhaps, on a set of measure zero independent of t;
(c)
A proof will be found in P. Halmos [1] for the discrete case, in NemytskiiStepanov [1] for the continuous case.
REMARK
6.5
Examples of systems (M, p., t) and functions I can be found, in which
*
I(x)
does not exist, or is not equal to -I, on a dense subset of M; even if
I is analytic and (M, p., t) classical (see Appendix 8).

REMARK
6.6
There exist dynamical systems in which I* exists everywhere as soon

as I is continuous or even Riemannian-integrable. For instance, the translations of the torus (see Examples 1.2 and 1.15, Appendix 9).
7. Ergod;city
DEFINITION
7.1
An abstract dynamical system is ergodic il lor every complex-valued

p.-summable lunction I
L1 (M, p.) the. time mean is equal to the space
'mean a.e.:
(7.2)
1
I(x)
f(x), a.e.
Appendix 10 provides some applications of.this theorem to Di((erential Geome-
try.
17
ERGODIC PROPERTIES
Thus, for an ergodic system, the time mean does not depend on the initial
point x.
EXAMPLE
7.3
Let us assume that M is the disjoint union of two sets MI and M2 of

positive measure, each of which is invariant under (see Figure 7.4):
MI :
M1,
M2 :
(2)
M2
CD
M,
M2
Figure 7.4
Such a system (M, /1, ) is called decomposable. A decomposable system

is not ergodic. In fact, taking
f (x) :
*
the time mean f(x)
REMARK
{1o
if x ( MI
if x ( M2 '
i(x) depends on x.
7.5
Conversely, a nonergodic system (M, /1, ) is decomposable. In fact,

nonergodicity implies there exists a function f ( L1(M, /1), the time mean
of which is not constant a.e. (this function can be assumed real-valued:
take ~f or ~ f). Set:
*
M1=lxlf(x)<a!,
M2
I x I f*(x)
::: a! .
For a suitable a we get:
According to the Birkhoff theorem the time mean is invariant under ,

hence:
and the system is decomposable. Whence:
18
ERGODIC PROBLEM!? OF CLASSICAL MECHANICS
COROLLARY
7.6
An abstract dynamical system is ergodic if, and only if, it is indecomposable, that is if every invariant measurable'set has measure 0 or 1.
The preceding argument proves more: a system is ergodic if, and only
if, any.invariant measurable function f ( Ll (M, p.) is constant a.e.
EXAMPLE
7.7
Hamiltonian flows (Chapter 1, Theorem 1.11) are never ergodic since

the energy H is an invariant function. However, the geodesic flow on the
unitary tangent bundle can be ergodic (see Chapter 3, 17.12). But the geodesic flow on Tl V is not always ergodic: if V is the usual torus, the geo.
d'esic f1<;>w acts nonergodically on Tl V, since the function
cP(l + r cos 1/1)2
is invariant (see Appendix 2).2

EXAMPLE 7.8
The rotation : x
-+
x + a (mod 1) of the circle M = Ix {mod 1)1 is
ergodic if, and only if, a is not rational.

Proof:
p, q (Z.
First case, a is rational. We set a = p/q,
q> 0, p and
q relatively prime. Since i(x) = e 2TTi qx is a nonconstant measurable in.
V,ariant function, is not ergodic.

Second case, a is nonrational. Let A be an invariant set of positive
measure; we shall prove p.(M)
point
= ]
Xo
1. Since p.(A) > 0, A has a density
(Lebesgue), that is for any E with 0
<' E < 1 there is an arc
Xo -0, Xo + 0 [ of length at most E, such that

p.(A n[)
~ (1-
E)p.(I).
From the invariance of A and p., we obtain:
Thus, if n l
, ... , n k
are integers for which n 1[, ... ,
nk[
are disjoint, we
get:
2 Small analytic perturbations of this metric' of
the geodesic flow (see Chapter 4).
r2
preserve the nonergodicity of
ERGODIC PROPERTIES
19
IL(A)?
;= 1
On the other hand, the orbit of a given endpoint of [ is dense (Jacobi theorem, Appendix 1). Since IL([) S E, there exist integers n 1 ., n k such that
the.sets
cp"I[,oo.,cp"k[
are disjoint and cover M up to a set of measure2E:.
Consequently
and
Since E: is arbitrary, IL (A)
1 and the system is ergodic. A similar argu-
ment proves that the systems of Examples (1.2) and (1.15) are ergodic as
soon as their orbits are everywhere dense (see Appendix 11). Appendixes
12 and 13 provide more examples.
8. \lixing
Let M be a shaker fun'of an ins,olJlpressible fluid, which consists of
20% rum and BO% Coca Cola (see Figure B.l). If A is the region originally
occupied by the rum, then, for any part B of the shaker, the percentage of
rum in B, after n repetitions of the act of stirring, is
In such a situation, physicists expect that, after the liquid has been
stirred sufficiently often
(Il
"'-l. every part B of the sh;ll(pr will contain
approximately 20% rum. This leads to the following definition.
20
Figure 8.1
DEFINITION
8.2
An abstract dynamical system (M, /L, t) is mixing if:

(8.3)
lim
t->+
/L[tAnB]
/L(A) "/L(B)
00
for every pair of measurable sets A, B.

It is clear that a dynamical system isomorphic to a mixing one is mixing. Thus, mixing is an invariant property of the dynamical systems.
COROLLARY 8.4
Mixing implies ergodicity.

Proof:
Let A be an invariant measurable set. Take B
and from (8.3):

/L (A)
0 or 1.
A; we get:
21
ERGODIC PROPEPTIES
The following example proves the converse is false: ergodicity does not
imply mixing. 3
EXAMPLE
8.5
An isometry
images
,p" A
,p
of a Riemannian. manifold is never mixing because the
of a small set A are congruent to A, and so their intersec-
tion with another set B is sometimes empty, sometimes of positive measure. For instance, the ergodic translations of the tori (Examples 1.2 and
1.15) cannot be mixing.
EXAMPLE
8.(;
Comparison of Figures (1.17), (2.4), and (8.1) suggests a conjecture:

the automorphism (1.16) of the torus T2 and the Bernouilli schemes are
mixing. This will be proved in (10.5) and (10.6).
REMARK
8.7
Mixing can be defined for endomorph isms (Appendix 6) which are not
automorphisms (see Appendix 14).
REMARK
8.8
Between ergodicity and mixing there is another concept, which is also

an invariant of the dynamical systems, the concept of weak mixing (see
Halmos [1)). A dynamical system (M, p., ,pt) is, by defin~tion, weakly mixing if
lim
T->+oc
in the continuous case, and

N-l
~ ~
1p.(,p"AnB)-p.(A)p.(B)1
n=O
3 1n contrast, if
p. (M) = 00, A. 8.
Hajian (1) proved the following: let (M,
p., ,pt)
be
an ergodic system in which p.(M) = oc, A and B two measurable sets. For any
E
>
0 there exist arbitrarily large t such that p.[,ptA
nB)
< E.
22
in the discrete case, for every pair of measurable sets A, 8. R. V. Chacon

(forthcoming paper) proved that if (M, /1, I) is ergodic, then there exists a
measurable change of the modulus of the velocity which makes the system
weakly mixing.
V. A. Rohlin [1] has proposed the concept of n-fold mixing as a new invariant of the dynamical systems (see Halmos [1]): A dynamical system
(M, /1, I) is n-fold mixing, by definition, if

n
lim
inf
II; -1,1 ~+oo
/1[I I A i n12 A2 . , nlnAn]
i '1i
for every n-tuple Ai'"'' An of measurable sets. Mixing is a particular

case (n = 2). Whether there exist mixing systems which are not n-fold
mixing (n > 2) is an open question.
9. Spectral Invariants
Let (M, /1, ) be an abstract dynamical system. Let us denote the Hilbert space of the complex-valued functions defined on M with /1-summable
square by L 2 (M, /1). If f, g ( L 2 (M, /1); we set
where z is the complex conjugate of z, and
II f I
DEFINITION
Y<fTT>
9.1
We set
(9.2)
Uf(x) = ( (x)) ,
where f ( L 2 (M, /1). U is a mappi.ng that operates on functions, it is

caIled the operator induced by .
23
ERGODIC PROPERTIES
THEOREM
9.3 (KOOPMAN [1])
V is a unitary operator of" L 2 (M, /l).

Proof:
Since cp is measurable, V carri~s measurable, square summable functions into themselves. Hence, V maps L 2 (M,I1) into L 2 (M, /l).
(a) V i~ linear: For any a,b ( C; f, g ( L 2 (M, /l), we get:
V(af+bg) '"' (af+bg)ocp '"' a(focp)+b(gocp) '"' aVf+bVg.
(b) V is a bijection: Any g ( L 2(M, /l) can be written as some Vf. Explicitly f (x) '"' g (cp -1 x).
(c) V is isometric: Since cp is /l-measure preserving, setting cpy '"' x,
we get
IlV f l1 2
'"'
If(CPy)12 d/l(Y)
If(CPy)1 2d/l (cpy)
(Q. E. D.)
REMARK 9.4
In the continuous case (M, /l, CPt) we obtain a continuous unitary oneparameter group V t .
DEFINITION
9.S
It i~ clear that two dynamical systems (M, /l, cp) and (M', /l " cp '), which
are isomorphic under f (see Definition 4.1), induce operators V and V'
which are equivalent, that is, there exists an isomorphism F: L 2 (M, /l)
L 2 (M', /l ') such that V' '"' F V F- 1, according to the following diagram
-+
(see Figure 9.6). Thus, the invariants of V are certain invariants of the
dynamical system (M, /l, cp). They are called the spectral invariants. For
'instance, the spectrum of V is a spectral invariant. 4 A complete system
4
In
the continuous case Ut , it is the spectrum of the infinitesimal generator of
Ut ; or"again, the spectrum associated with the resolution of the identity E for
which, according to Stone's theorem: Ut '"' J~:: e277IAtdE(A)
24
of spectral invariants is known (see Halmos [3]), the spectral measures

an~
spectral multiplicities. Conversely, the equivalence of the operators
V and V' (the corresponding dynamical systems are called spectrally
equivalent) does not imply isomorphism (see Chapter 2, 12, Entropy; Appendix 15, Anzai Skew-products).
cp'
Figure 9.6
We now give some examples of ergodic properties that are reflected as

spectral properties.
THEOREM
9.7 (ERGODICITY)
(M, 11, ) is ergodic, if, and only if, 1 is a simple proper value 01 the
induced operator V.
Prool:
If I lies in L 2 (M; 11), then I is invariant if, and only if, Vf
f. Now
is ergodic if, and only if, the invariant functions are all a.e. constant.
Since the functions that are a.e. constant are scalar multiples of one another, is ergodic if, and only if, the subspace' of-solutions of Vf = f
has dimension 1. In the continuous case, the ergodicity of t is equivalent to A = 0 having multiplicity one in the spectrum of V t'
ERGODIC PROPERTIES
25
THEOREM 9.8 (MIXING)
The dynamical system (M, Il' t) is mixing if, and only if,
lim
(9.9)
T-+oo
< U/ I g >
< f 11 > . < 1 I g >
for every f, g ( L 2 (M, Il)

Proof:
If f and g are some characteristic functions, then (9.9) reduces tothe

very definition of mixing (8.2). The general case is derived easily by observing that the space of finite linear combinations of characteristic functions is dense in L 2 (M, Il). In spectral terms, (M, Il' t) is mixing if it is
ergodic and the spectrum of Ut (except for A = 0) is absolutely continuous
with respect to the Lebesgue measure. The converse is false. We say that
Ut has properly continuous spectrum if its only proper functions are constants. It can be proved (see Halmos [1]) that a dynamical system has
properly continuous spectrum if, and only if, it is weakly mixing (see 8.9).
We turn to the case in which the spectrum of Ut is discrete.
EXAMPLE 9.10
Let M be the circle \z I z ( C, Izl

translation (z)
= (). z,
()
e 2TTiW ,
tion zP, p ( Z:
11, Il its usual measure, the.
=
W
R . Let us consider the func-
sJ>
UzP
(Uz)P
Hence, the zP's are proper functions of U with corresponding proper values ()P. The set \zP I p.(
ZI.
which is called the discrete spectrum of U.
forms a complete orthonormal system of L 2 (M. Il); whence the definition:

DEFINITION 9.11
A dynamical system
tM. p.. ) has properly discrete spectrum if there
is a basis of L 2 (M. p.). each function of which is a proper function of the

induced operator U.
Let us turn back to (9.10). According to Theorem (9.7), the system is

ergodic if, and only if, 1 is a simple proper value,. that is, if, and only if,
pw
Z when p i 0, which means w is irrational. In otherwords, our sys-
26
tem is ergodic if, and only if, the orbits are dense on M (see Example 7.8
and Appendix 1). Observe that ergodicity implies that all the proper values (}P are distinct and simple. The system is not mixing; take f = g =
zP
in (9.8); we get:
< VnzP IzP > =

If p f, 0,
(Jpn
lim (JPn does not exist and (9.9) is not fulfilled. These results
n=oo
extend immediately to.the n-dimensional torus and suggest the following

theorem.
THEOREM
9.12
Let (M, Jl, ) be an ergodic dynamical system, V the induced operator.

Then:
(a) the absolute value of every proper function of V is constant a. e.;
(b) every proper value is simple;

(c) the set of all the proper values of V is a subgroup of the circle group
Izlz
(C,
Izl
1\;
(d) if (M, Jl, ) is mixing, the only one proper value is 1.
Proof:
Since V is unitary, every proper value A has absolute value 1. It
follows that if f is a corresponding proper function
f(x) = Al(x) a.e. implies
If(x) I =
If(x)1 a.e.
Hence, f(x)1 is invariant under , and ergodicity implies that If
I is con-
stant a.e. (Corollary 7.6). In particular, f f, 0 a.e.

Let h be another proper funetion with proper value A. Since f f, 0
a.e., hlf makes sense. We get:
V(!!)
Vh
f
Vf
=
I'
and hlf is an invariant function, so that h is a constant multiple of f.

This proves (b).
If A and Jl are proper values of V, with corresponding proper functions
f and g, we get
27
ERGODIC PROPERTIES
VCt)
g~
= ::
A~-1 Ct)
Hence II g is a proper function of V with proper value A~-1
This proves
(c).
Finally, if the system is mixing, take I = g equal to a proper function
with proper value in (9.9), we get:
lim < vn/l I>
n=oo
</11>'<11/>,
that is
lim
An
constant.
n=oo
Hence A = 1, and (d) is proved.

These properties of the discrete spectrum have been in some sense
extended to the continuous part of the spectrum by Sinai [2], [3] (see, however, the recent paper of Katok and Stepin [1] for an example of a system
whose 'maximal spectral measure does not dominate its convolution). Tht:
group of the proper values is obviously an invariant of the dynamical system. If the spectrum is discrete, this group forms a complete system of invariants. More precisely, we have:
DISCRETE SPECTRUM THEOREM
9.13
(VON NEUMANN, HALMOS)
(a) Two ergodic dynamical systems with discrete spectrum are isomorphic
ii, and only ii, the proper values 01 their induced operator coincide.
(b) Every countable subgroup 01 the circle group is the spectrum 01 an er-
godic dynamical system with discrete spectrum.

The proof will be found in Halmos [1]. It is based upon the construction of some compact abelian group (character group of the spectrum of
given ergodic dynamical systems with properly discrete spectrum). Then,
one proves the isomorphism of our given dynamical system with a translation of this abelian group.
To emphasize this result, we point out that the isomorphism problem
is solved as far as the discrete spectrum case and abstract frame are con-
28
cerned. In contrast, no characterization is known for the spectrum of a

classical system. For instance, does there exist a classical system whose
discrete spectrum is a prescribed subgroup of the circle group? Appendix
16 contains some information related to this question.

10. Lebesgue Spectrum
Let us begin with an example.
EXAMPLE
10.1
We again consider Example (1.16): M is the torus l(x, y) (mod 1)\ with
its usual measure;
4)
is the automorphism:
.(x, y)
(x + y, x + 2y) (mod 1);
U is the induced operator. It is well known thaUhe set
\e
p,q
(x, y) =
e 2TTi (pX+ qy),
p, q (
Z\
is an orthonormal basis of L 2 (M, p.). The set D can be identified with the
lattice Z2
= \(p,
q)! C R2. Since Uep,q
= ept q, p+2 q '
U induces an au-
tomorphism u on D:
Let us show that (0,0) is the unique finite orbit of u. Assume that (p, q)
( Z2 has a finite orbit. This orbit is a bounded subset of R2, invariant
under the linear operator of R2,
1>
has two proper values A1, A2, 0 < A2 < 1 < A1 Hence, 1> is "dilating" in the proper direction corresponding to A1 , and "contracting" in the
proper direction corresponding to A2 This implies that the only invariant
(under ) bounded subset of R2 is (0,0). (Q. E. 0.)

We conclude that Z2 -\0,0\ splits into a set I of orbits of u, and
each orbit is in an obvious one-to-one correspondence with Z.
29
ERGODIC PROPERTIES
Let us go back to D
Ie p, q I p,
q ( Z\. D -leo , 0\ splits into orbits
of V: C l , C 2 , ... , Ci,oo.; i (1. If fi,o is some element of C i , we may

write
C.,
where f.l,n
Ii.l,n In ( Z\ ,
V n f.1, o. To summarize, if H., is the space spanned by the
vectors of C i , then L 2 (M, /1) is the orthogonal sum of the H/s and ofthe
one-dimensional space of the constant functions. Each Hi is invariant
under V and has an orthonormal basis
Ii./,n In ( Z\
such that:
Situations such as this occur often enough to deserve a definition.

DEFINITION
10.2
Let (M, /1,

tor. (M, /1,
eM
eM
be an abstract dynamical system, V the induced opera-
has Lebesgue spectrum LI if there exists an orthonormal
basis of L 2 (M, r11) formed by the function 1 and functions f.1,J. (i (1, j ( Z)
such that:
Vf 1,]
. .
f1,]+
. . l' for every i, j.
The cardinality of 1 can be easily proved to be uniquely determined

and is called the multiplicity of the Lebesgue spectrum. If 1 is (countably)
infinite, we shall speak .of (countably) infinite Lebesgue spectrum. If 1
has only one element, the Lebesgue spectrum is called simple. An analogous definition holds in the continuous case. Let Vt be the one-parameter group of induced operators of a dynamical flow (M, /1, q, t)'
The flow is
said to have Lebesgue spectrum LI if every Vt (t .;, 0) has Lebesgue

spectrum L I.
REMARK
10.3
This terminology is derived from the following fact: Let

Vt
00
-00
e 2TTit A dE (A)
.'
be the spectral resolution of Vt . It can be proved that (M, /1, q,t) has
30
Lebesgue spectrum if. and only if. the measure
< E(,\)f II> is absolutely
continuous with respect to the Lebesgue measure. for every I ( L 2 (M. /L)
orthogonal to 1.
THEOREM 10.4
A dynamical system with
Lebes~ue
spectrum is
mixin~.
Prool:
From Theorem (9.8) we need to prove that:
lim
n"OO
< Un I I ~ > = . < 1 I ~ >
for every I. ~ ( L 2 (M. p.). This is equivalent to:

lim
n+ oo
for every I.
< Un I I ~ > = 0
orthogonal to 1. It is sufficient to prove this when I and
are basis vectors. for the general case follows by continuity and linearity
If 1= I j i ~ = Ik r then
<I.J,n+.l/
J . k ,r >
which is null for n large enough.
COROLLARY 10.5
The automorphism if> (x. y) = (x+ y. x+ 2y)(mod 1) of the torus

M =
I (x. y)(mod
1) I
(Example 1.16) has Lebesgue spectrum (Example 10.1). Then. it is mixing and ergodic (Corollary 8.4).
EXAMPLE 10.6
The Bernouillischemes have countable
Lebes~ue
spectrum.
In par-
ticular they are spectrally equivalent.

Prool:
We prove it for B(~. ~); the same statement holds for B(Pl' .. Pn) UI
to minor modifications. Let us recall (see Example 2.2) that M = Z2 Z
the space of the infinite bilateral sequences:
m = ... m_ l mo' m l .; mj (
\0.11.
31
ERGODIC PROPERTIES
The function 1 and the function

j -1
Yn(x) =
if x = 0
if x = 1
1+ 1
form an orthonormal basis of the space L 2(Z2. 1l ) associated to the n-th

factor of M. From the product structure of M. we get an orthonormal basis of L 2 (M. Il) which consists of the function 1 and all the finite pro~
ucts
nl
..... y
fIJc
of the y 's with distinct indices n 1 .. n k

n
U be the induced operator of the shift
Now. let
..
. Call two elements of the above
basis equivalent if some integer power of U carries one onto the other.
The function 1 constitutes its own equivalence class; the other basis
functions split into countably many equivalence classes. Each such equivalent class is in a one-to-one correspondence with Z: the action of U on
the class is to replace the element corresponding to n
Z by the element
corresponding to n+ 1.
To ~ummarize. there exists an orthonormal basis of L 2 (M. Il) consisting of the function 1 and of functions I.I,]. (i = 1. 2 ... ; j l Z) such that
UI 1,]
. . = l.I, J.+1
for every i. j. The number i is the number of the equivalence class. the
number j is the number of its element which corresponds, as described
above. to j
Z. Thus. B(Yl. Yl) has countable Lebesgue spectrum.
Let (M 1.1l1' 1) and (M 2 .1l2' 2) be two Bemouilli schemes. There

exist. from the above. orthonormal bases
I 1, I;~j I in
11, I;,j I
..
10
L 2 (M 1.1l1) and
L 2(M 2 , 1l2) such that:
I~I, }+1
for every i, j. The isometry of L 2(M l' Ill) onto L 2(M 2 , 1l2) defined by
1 ... 1,
carries the spectral type of the first scheme into that of the other.
32
ll. K-Systems
In this section we define a class of abstract dynamical systems with
strongly stochastic properties.
DEFINITION
11.1
An abstract dynamical system (M, IL' ) is caIled a K-system 6 if there

exists a subalgebra ff of the algebra of the measurable sets satisfying:
(j' C ff.
(a)
n
00
(b)
0,
nff =
n=-oo
where
0 is
the algebra of the sets of measure 0 or 1,

00
(c)
i,
n (j' =
n=-oo
where , by abuse of language, is the automorphism of
induced by .
The above conditions become, in the continuous case:

(a ')
00
n /i
(b ')
1.
t=-oo
00
(c ')
V /f.
t~-OQ
From the very definition, the isomorphic image of a K-system is a K-system.

EXAMPLE
11.2
BERNOUILLI SCHEMES (SEE
2.2)
The BernouiIli schemes are K-automorphisms.

Proof:
Let B(PI"'" Pn) be a Bernouilli scheme. The algebra 1 is generatec
5 See~ Appen~dix 17 for notations as C, /\, V,.... The standard notations (Rohli,
are.
I =~1ll, 0 = ;ll.
6 A. N. Kolmogorov [2) introduced this class under the name of quasi-regular sys
terns.
33
ERGODIC PROPERTIES
by the:
Let ct be the algebra generated by the A Ij,s, i
< O. We know that:
(Aj) = Aj+l
where is the shift. Hence ct is the algebra generated by the
k
At's,
1, and
ct c ct,
proving the property (a). On the other hand, every generator
q(A/)
Aj+q , i ~ 0, for q
A! of
1 is a
r-i. Hence we get the property (b):

00
V nct
Let us now prove the property (c). Let
93
i.
be the subalgebra of 1, each
element of which belongs to some subalgebra generated by a finite number

of AI- To every A
= fl(A) " fl(B)
93
for any B
there corresponds an N
f
-nct, n
fl(A) "fl(B) holds for every B
still holds for any A
Z such that fl(A
~ N (exercise). Hence fl (A
(n;; -ncf.., Since ~ 1. this relation

n,; -net. Especially:
=
and B (
fl (B) = fl (B
n B)
that is fl(B) = 0 or 1, for every B. (
= [fl (8)]2 ,
n;; -net.
We conclude:
hence:
(Q. E. D.)
COROLLARY
n B)
n B)
11.3
The Baker's Transformation is a K-system.
34
Prool:
This system is isomorphic to B (V2, V2) (see Appendix 7).
1104
EXAMPLE
Chapter 3 will be devoted to a wide class of classical K-systems. This

class contains the automorphisms of the tori, the geodesic flows on compact Riemannian manifolds with negative curvature, and the Boltzmann-Gibbs
model of particles colliding elastically.
11.5
THEOREM
A K-system has a denumerably multiple Lebesgue spectrum. In particular, it is mixing and ergodic (Theorem lOA). This theorem is due to Kolmogorov [2] for K-automorphisms and to Sinai [6] for K-flows. We sketch a
proof for K-automorphisms, the complete proof will be found in Appendix 17.
H8UH
U-1 H8H
.........
.'
o. o.
0.00
000
000.00
:UH
U-1H
....................
x Uh 1
x hl
x U-1h 1
..........................
Hl)
U-1hj .................. ;" ....
Hd
x h2
000
00
oX
Uhj
x hj
II
Figure 11.6
Let (f be the subalgebra of Definition (11.1). We d~note the subspace

of L 2 (M, f.L) generated by the characteristic functions of the elements of (f
by H. If U is the induced operator, the properties (11.1) of (f are translated as follows:
n
~
unH C," C UH C H C U- 1H C," C
U
n=-oo
n=-oo
= L 2 (M, f.L),
where Ho is the one-dimensional space of the constants.
UnH
35
ERGODIC PROPERTIES
Let us select an orthononnal basis
Ih.1
I
on the'orthocomplement H 9VH
of VH in H. H. is the space spanned by the sequence ... , V-lh., h.,

I
Vh.,.... The H's are invariant under V, and their orthogonal sum is
I
L 2 (M, p.) 9 Ho' Hence, if we set

e ..
1,J
Vjh.,
I
i(Z+,
j(Z,.
the e 1,1
. .'s and the function 1 constitute an orthononnal basis of L2(M,~)
such that:
for every i, j. We conclude that V has Lebesgue spectrum. The proof
~il1
be complete after it has been shown that the dimension of H 9 VH is infinite (see Appendix 17): V has a spectrum of infinite multiplicity.
12. Entropy
This section is devoted to the definition and the study of a non spectral invariant of dynamical systems introduced by A. N. Kolmogorov [4].
Throughout, z (t) denotes the function on [0, 1] defined by:
z(t) = {-tLOgt if O<t:;1
o if t = 0
where Log denotes the base-2 logarithm. We use the following properties
of z: z is nonnegative, continuous, strictly concave
< 0) ,
(z"(t) = - Lotge
and z(t) = 0 is equivalent to t = 0 or 1. Let a = IA 1.1.I (
measurable partition of M (see Appendix 18): I is finite and

p.
(M -.U Ai)
1(1
if i f, j.
0,
II(A.nA.)=O
I
I
,..
be a finite
36
DEFINITION
12.17
By definition, the entropy of the partition a

h(a)
EXAMPLE
IS:
~ Z(Il(A))
jll
12.2. PARTITION INTO N ELEMENTS OF EQUA L MEASURI
When Il(A j ) = liN, then h(a) =fLogN. Observe that if f3 is some

other partition into N elements, then:
h (f3) ~ LogN
with equality if, and only if, each element has measure liN. In fact, Jensen's inequality applied to the concave function z gives
Hence h({3) is nothing but the weighted logarithm of the number of the
elements of f3.
Observe that if two partitions are equivalent, that is, if their elements
coincide up to some sets of measure zero, their entropies are equal. In
particular, the elements of measure zero can be removed. Finally, h(a) =0
I'
e
Figure 12.3
For a probabilistic treatment see A. M. yaglom -
I. A. Yaglom
[d.
37
ERGODIC PROPERTIES
means that a
v (mod 0) where v is the trivial partition, the unique ele-
ment of which is the space M.

DEFINITION
12.4. CONDITIONAL ENTROPY OF A PARTITION a WITH
RESPECT TO A PARTITION
{3
Let a = !Aili = 1, ... ,r\ and {3 = !Bjlj = 1, ... ,sl be two finite measurable partitions (in the future, for short: partitions). The entropy of a
with respect to (3 is defined by:
where
Il(A.nB.)
I}
Il(A./B. )
I}
1l(B.)
}
a induces on each B. a
is the conditional measure of A. relative to B.;

.
finite measurable partition a B ' the elements of which are B. n AI' ... , B.
i
J }
n A.r After a suitable renormalization, B.} can be considered as a space
of measure 1 on which aBo has entropy:
]
h(aB ) = ~ z(Il(A./B.)).
i
I}
Hence, h(a/{3) is the weighted sum of the h(aB)'s:

]
h(a/{3) = ~ Il(B.)h(a B )
}
The following theorem is proved in Appendix 18:

THEOREM
Let a.
(12.6)
12.5
/3.
y be finite measurable partitions. Then:
h(a/{3).:: 0 with equalityil. and only if
a~{3
(12.7)
h(a V f3/y) = h(a/y) + h({3/a V y).
(12.8)
a ~ (1 (mod 0) "~,'> h (a/y) ~ h (1/1');
(modO).
that is. conditional entropy is nondecreasing in its first argument.
(12.9)
38
that is, conditional entropy is non increasing in its second argument.

h (a V f3/y)
(12.10)
s h (aly) + h (f3/y) ,.
that is, conditional entropy is subadditive in its first argument.

Let v be the trivial partition \Ml and let us take y
relations. Since h(alv)
(12.11)
=v
in the above
h(a), we get:
h (a V (3) '" h (a) + h (f3/ a) ,
(12.12)
> h (a) S h (f3) ,
a S f3 (mod 0)
h (a/ (3)
(12.13)
s h (a)
h(a V (3) S h(a) + h(f3).
(12.14)
Finally, if is an automorphism of the measure space (M, p.) and a =
\A l , ... , Anl is a partition, a is a partition, namely:
One verifies at once that:
(12.15)
(a V (3)
(12.16)
h (a/ (3)
DEFINITION
a V f3 ,
h (aI m .
12.17. ENTROPY OF A PARTITION WITH RESPECT TO
AN AUTOMORPHISM
Let (M, p., ) be a dynamical system, and a a finite measurable parti

tion of M. By definition, the entropy of a relative to is:
n (
Z+
Of course, we need to prove that this limit exists. Let n be a positive i

teger. We set:
LEMMA
12.18
39
ERGODIC PROPERTIES
Proof:
- According to (12.12). a V cpa V .. V cpn-l a :::: a V V cpn a implies

(Q. E. D.)
LEMMA 12.19
I sn I
is a nonincreasing sequence.
Proof:
According to (12.11):
(12.20)
sn
h(a V"'V cpnal - h(a V"'V cpn- l al
= h(cpnal a V"'V
cpn- l al .
Hence:
sn_l
h(cpn-lalaV"'Vcpn-2al;
using (12.15) and (12.16) we get:

sn_l = h(cpn a ICPaVVcpn-l a l.
Since chn V V cpn-l a ~ a V V cpn-l a, (12.9) implies:
THEOREM 12.21
h(a. cpl exists and is equal to
lim h(al cp- l a V"'V cp-nal.

n->oo
Proof:
sn is a nonincreasing sequence of positive numbers: sn has a limit s.
Observe that h n = h (al + sl + ... + sn' thus. Cesaro's mean convergence

theo,rem implies:
lim
n = oc 'n
s.
Theorem (12.21) follows at once from (12.20) and the very definition of
h (a, cp).
Let 8(p t ... Pk ) be a Bernouilli scheme (see Example 2.2). cp the

shift. We consider the finite partition a. the k elements of which are the
40
A6
= Imlmo = il.
1, ... ,k.
1=
We are going to prove that:

k
-~ Pi Log Pi .
Since n ..16
A~
are the:
Imlmn=il, the eleme!lt of aVaVVn-l a

Aio nAil
n ... n Ain-I
n-l
'
whose measure is p. . .. p. . Therefore:

'0
'n
p .... p.
'0
'n-l
Log(p .... p.
'0
To carry out the summation over i o' observe that

p ... p.
'I
1.
'n-I
We get:
and by induction:
whence:
- ~ Pi Log Pi .
DEFINITION
12.23. ENTROPY OF AN AUTOMORPHISM
The entropy h () of an automorphism is:

h(ch) = sup h(a,<h)
8 Notion due to Kolmogorov [4]. See also Sinai
bl.
,
[8].
'n-
) .
1
41
ERGODIC PROPERTIES
where the supremum extends over all finite measurable partitions a. It is
clear that h ()
O.
THEOREM 12.24
h () is an invariant of the dynamical system (M, /l, ).

Proof:
Let (M', /l', ') be a system isomorph.ic to (M, /1, ). (see Definition
-+ M', that is '= fe l . If
4.1). There exists an isomorphism f: M
'a is ,a partition of M, fa is a partition of M '. We get from (12.15) and
(12.16):
h((l,')
h(fa,fe l )
lim h(aV ... Vfo-lf-lfa)

n
0=00
lim h[(aV"'Vo-la)]
n~oo
lim h(aV ... Vo-la)
n:::oo
h(a,).
On the other hand, when a runs over all the partitions of M, fa runs over
all the partitions of M '. We deduce at OIlce:
sup h(a', ')
sup h(a, ) .
We now turn to a theorem which enables one, in many cases, to compute

the entropy.
DEFINITION 12.25. GENERATOR 'WITH RESPECT TO AN AUTOMOR,PHISM
~et a be'a finite measurable partition, m(a) the measure subalgebra
generated by a:
a is called a generator with respect to if:
00
n=-oo
,.
K'oLMOGOROV'S THEOREM 9 12.26
If a is a generator-relative to , then h() = h(a, ). The proof will

9 See Kolmogorov [2], [4]; Sinai [7], [8].
42
ERGODIC PROBLEMS OF CLASSICAL MECHANics
be found in Appendix 19. Thus if possesses a generator, its entropy

may be computed by the above formula. Let us give some examples:
The entropy of the shift of B(Pl' ... , Pk) is:

k
l:
h()
Pi Log Pi
i=1
Proof:
Let a' be the partition of 12.22, the elements of which are the
A~
Since
n A~
lmlmo
iI,
A~, n ( Z, the algebra
erators of the algebra
1.
1, ... ,k.
V..:ooo nm(a)
contains all the gen-
Consequently a is a generator relative to .
The above formula is a direct consequence of (12.22) and (12.26).

CONSEQUENCES 12.28
(1) Given an arbitrary nonnegative number a, there exists an abstract dy-
namical system, namely a Bernouilli scheme, the entropy of which is equal

to a.
(2) We proved (Example 10.6) that the Bernouilli schemes are spectrally
equivalent. But B('I2, 'h) and B('I" '1" "1,) differ in their entropy and, according to Theorem (12.24) are not'isomorphic. Hence, there exist non-isomor-
phic abstract dynamical systems which are spectraIIy equivalent.

It is conjectured that two K-systems are isomorphic if they possess the
same entropy. Particular cases have been examined which indicate that
this may be correct. We mention a result due to Meshalkin. [1]: B(Pl' ... )
and B(ql' ... ) are isomorphic if they possess the same entropy and if the
P;'s, q;'s are negative integrC;lI powers olone and the same positive integer n. For instance, B('I" 'la, 'la, '10, '10) and B(';., ';., ';., ';.) are isomorphic
(here n
2). Blum and Hanson [1]. improved this result. In this direction
we mention a theorem due to Sinai [9]: . Two K-systems with the same en- .
tropy are weakly-isomorphic, that is, each system is an homomorphic image

of the other (see Appendix 6).
43
ERGODIC PROPERTIES
EXAMPLE 12.29. AUTOMORPHISM OF THE TORUS
If
rp
is an ergodic automorphism of the torus I(x, y)(mod 1)1:
rp (x,
y) = (ax + by, cx + dy! (mod 1),
ad - bc
1,
Sinai [7] proved that:
where A1 is the proper value, whose modulus is greater than 1, of the matrix:
This result extends to the r-dimensional torus T r = Rr /z r. Let
ergodic automorphism of Tr. If the matrix of
rp
rp
be an
has r distinct proper val-
ues A1, ... ,A r , then:
h(rp)
I
Log
1\1
Ail>1
(See Genis [1] and Abramov [1] for a correction of the proof.)
COROLLARY 12.30 (SEE EXAMPLE 4.4)
The dynamical systems defined on T2 by:
; ) and (;
are nonisomorphic.
Those defined by:
(; D
and
(~
i)
possess the same spectral type and the same entropy. Whether they are
isomorphic is still an open problem. It is only known that they are weakly
isomorphic (Sinai [9])'.
THEOREM 12.31 10 ENTROPY OF K-SYSTEMS
The entropy of a K-automorphism is positive.

10 Due to Kolmogorov [4].
44
Proof:
According to Definition (11.1) there exists a subalgebra d' of
such
that:
6=
n nd' c c d' c d' c c
1.
nd'
-00
We first prove there exists a finite subalgebra
(12.32)
k 93)
k$.n
93 c
k P,
d' such that:
for n > n';
n, n' (
Z.
k$.n'
Assume that for every finite subalgebra
93 c
d' there exists n, n' ( Z such
that n > n' and
k<n'
Since (M, fl) is a Lebesgue space, there exists an increasing sequence of

finite shbalgebras
93.:
I
satisfying:
From our assumption it follows that:
for every i. As
V is associative and commutes with
, we obtain:
which leads to a contradiction.

Now, let {3 be a finite partition genera ting the algebra
dix 18.5). From (12.21) we deduce:
93
(see I\ppen-
45
ERGODIC PROPERTIES
n=oo
The sequence -1{3, -1{3 V -2{3, ... , is nondecreasing, therefore, from

(12.9) we get,
h () 2: h ({31
V k
k<-1
(3) .
Now, it is sufficient to show thath ({31 Vk <-1 k(3) > O. Assume that
h({3!Vk <_1 k(3)
0; this means (12.6);
{3 ~
k{3 (mod 0) .
k~-1
Consequently:
that is
which. contradicts (12.32).

REMARK
(Q. E. D.)
12.33
Guirsanov [1] has constructed a nonclassical dynamical system with

zero entropy and denumerably infinite Lebesgue spectrum. Hence it i~ not
a K-system. Gourevitch [1] proved that the horocyc1ic flow on a compact
surface with constant negative curvature has denumerably infinite Lebesgue spectrum and zero entropy. Hence, it is a classical system with Lebesgue spectrum, but it is not a K-flow.
REMARK
12.34
By defin~~ion, the entropy of a flow (M, fl' /) is h (1). If (M, fl' /)

is a K-flow (see Definition 11.1) then (M, fl' 1) is a K-automorphism.
Consequently (Theorem 12.31) the entropy h'(1) of the K-flow is positive.
46
KOUCHNIRENKO'S THEOREM
12.35. ENTROPY OF CLASSICAL SYSTEMS 11
Classical systems have finite entropy.

Proof:
Let (M, p., ) be a cla.ssical system. Since M is smooth, it carries
some smooth Riemannian metric g. By a suitable conformal deformation
g -+ e 2p g, it is possible to assume that the volume element of g is precisely the measure element dp..
The area of a submanifold will be the
area in the sense of the metric g. By definition, a classical partition of

M is a partition into a finite number of complexes with piecewise differen-
tiable boundary. Since M is compact and smooth, such a partition always

exists (Cairns [1]).
Let us begin with two obvious remarks:
(1) The classical partitions are dense in the sense of the entropy metric
(see Appendix 19) in the set of the finite measurable partitions. Since
h (a, ) is continuous for the entropy metric (Appendix 19), we get:
h(a, )
sup
a classical
(2) Let a be a (dim M-1)-dimensional submanifold, the area of which is

S (a). Since M is compact, there exists a constant
(independent of a)
satisfying:
S(a)
(12.36)
S(a)
< A
Let a be a classical partition; AI' ... ' An the elements of a V V k-l a.

According to the isoperimetric inequality, there exists a constant C, which
depends only on the manifold M, such that:
N = dimM.
Hence, we obtain:
11
See Kouchnirenko
[11.
[2].
47
ERGODIC PROPERTIES
i = 1
i= 1
If we denote by S (a) the sum of the areas of the boundaries of the elements of a (each boundary is counted twice), clearly:
n
~ S(aA) = S(a) + '" + S(q}a) ,

i =1
and then:
n
~ [/L(A i )](N-1)IN
:S
C[S(a) + ... + S(q/-1 a)]
i= 1
From (12.36) we deduce:
Log l
[/L(A)](N-1)IN S k Log'\ + constant.
But Log t is concave, /L(A) :: 0, l/L(A)
1. Hence, Jensen's in-
equality applied to the left member gives:

n
~ /L (A) Log /L (A )-1 IN S k Log ,\ + constant,

i= 1
n
i= 1
/L(A.)Log/L(A.)
1
< N. Log ,\ +
If k .... + 00, we get:
h (a,
S N Log ,\ ,
and from remark 1:
(12:38)'
constan~
h (1)) 'S. (dim M)..- Log ,\ ~
48
where the constant A is given by relation (12.36).
(Q. E. D)
COROLLARY 12.39
There exist abstract dynamical systems non isomorphic to a classical

one, for example, the infinite Bernouilli scheme:
8(1/2.1/4.1/16.1/16... 1/22n ... 1/22n ... )
~/.--.........../
22n-n:-1 hmes
COROLLARY 12.40
If an abstract system has discrete spectrum whose rank is finite, then

the system has zero entropy.
Proof:
In fact the system is isomorphic to a compact, finite dimensional abelian group on which a translation
h (1))
1>
acts. Hence A = 1, and from (12.38):
O.
A PROBLEM 12.41
Whether the entropy h (1)) of a classical system depends continuously

on
1>
is an open question.
REMARK 12.42
Kouchnirenko's theorem is connected to recent results of M. Artin and

B. Mazur [1]: Let M be a smooth compact manifold, then for a dense set
of C 1 -diffeomorphisms the number N (n) of isolated fixed points of 1>n,
n
1,2, ... , is exponentially bounded from above:
='
C (1)),
= 1..(1))
REMARK 12.43
Recently, Kouchnirenko 12 introduced some new nontrivial invariants

. of abstract dynamical systems: A-entropies. Let A be a monotone sequence of integers
12 See his report at the lnt. Math. Congr., Moscow, 1966.
ERGODIC PROPERTIES
49
Then the A-entropy of an automorphism with respect to a partition a is

defined as:
As in Definition (12.23), A-entropy is:
One obtains the usual entropy if A
10, 1,2, ... 1. The A-entropies can
distinguish some systems with usual entropy O. Let us give an example.

Let A
.3) is 0
<h <
12n I. Then the A-entropy of the horocyclic flow (see Chapter

00.
A-entropy is 0
Consider the direct product of this flow onto itself.
< 2h <
00.
~ h,
Since 2h
Its
the product is not isomorphic to
the horocyclic flow; however, they have both zero usual entropy and
countable Lebesgue spectrum.
REMARK
12.44
Recently, Katok and Stepin [1]
13
introduced some new nontrivial invari-
ants of abstract dynamical systems: periodical approximations speeds.

Let (M, Jl, ) be an abstract dynamical system, let ~n be a partition of
(M'Jl) into sets C~ of measure l/n (i= 1, ... ,n). Anautomorph-ism Snof
(M, Jl) will be called cyclic with respect to the partition ~n if:
(a)
Sn~n = ~n;
(b)
(S)n
n
E (identity), (S )k ~ E for k
n
< n.
We shall say that admits approximation by cyclic transformations at the

rate 0 [f(q)]
if, for an increasing sequence of natural integers qn , there
n
exists a sequence of partitions ~ ...
qn
and a sequence of automorphisms
Sqn' cyclic with respect to ~qn' such that

qn
i = 1
Katok and Stepin have proved some important theorems which connect the
l3
See also their report at the lnt. Math. Congr.
Moscow, 1966.
50
notion of rate of approximation by periodic transformations with entropy

and spectra. The importance of their results is related to the fact that, in
many cases, it is possible to obtain some information on the approximation
speeds of concrete systems, even if the explicit computation of spectra is
impossible. Some of their theorems are the following:
(1) If the automorphism admits approximation by cyclic
at the rate 0(1/ln 2 q ), then h(J
n
transf~rmations
= O.
(2) An automorphism which admits approximation by cyclic transformations

at the rate O(l/qn) is ergodic. Furthermore, strong convergence U qn
>
E occurs, where U is the unitary operator in L 2(M, 11) induced by . As

a corollary, is not mixing and the maximal spectral type of U is singular.
One can find more theorems on approximations and their applications
to the study of concrete dynamical systems, such as, for example, the mapping
~l
~3
~3
~l
or the flow on the torus

dx
-
dt
dy
F(x, y)
dt
F (x, y)
in the papers of Katok and Stepin, in Doklady, in Founktzionalnyi analys i

ego prilojeniia (1967), and in Uspehi (1967)
Hero npHJlOlKeHHH, MocKBa
1967.
Il>YHKl.\HOHaJlbHblH aHaJIH3
ERGODIC PROPERTIES
51
Halmos, P. R., Lectures on Ergodic Theory, Chelsea (New York).

Halmos, P. R., Entropy in Ergodic Theory: Lecture Notes, University of
Chicago (1959).
Hopf, E., Ergodentheorie, Springer, Berlin (1937).
Neumljl?n,
J.
von, Zur operatoren Methode in der klassischen Mechanik,
.Ann. Math. 33 (1932) pp. 587-642.

Rohlin, V. A., New Progress in the Theory of Transformations with Invariant Measure, Russian Math. Surveys 15, No.4 (1960) pp. 1-21.
Sinai, Ya., Probabilistic Ideas in Ergodic Theory, Trans!. Math. Soc.
Series 2, 31 (1963) pp. 62-84.
CUAPTER
lI\STABLE SYSTE\IS
This chapter contains the study of classical systems with strongly stochastic properties, the so-called C-systems.
The orbits of a C-system are highly unstable: two orbits with close
initial data are exponentially divergent. This property turns out to imply
the asymptotic independence of past and future: C-automorphisms are ergodic, mixing, have Lebesgue spectrum, have positive entropy, and, in general, are K-automorphisms. The set of the C-systems defined on a prescribed manifold M is an open set in the space of the classical systems
defined. on M. Consequently, ev,ery classical system, close enough to a Csystem, is
a C-system.
Geodesic flows on Riemannian manifolds with negative curvature are

first examples of C-systems.
13. C-Systems
EXAMPLE
13.1
Let us consider the torus M
!(x, y) mod 11 and the diffeomorphism
: M ... M:
The length
IIXII
of a tangent vector
X is
referred to the usual Riemannian
metric dx 2 + di of M. Let *: TMm'" TMm be the differential of .

1
UsualIy calIed U-systems in the English literature for, in Russian:"y-cItCTeMbl:'
Our terminology was introduced by Anosov; C is used because these systems satisfy a "condition C", in Russian J.J YCAOBHC y. 'H
.
53
54
ERGODIC PROBLEMS OF .CLASSICAL MECHANICS
In the chart (x, y), * is the linear mapping:
(~ ~)
Hence,
has two real proper values Al and A2 , (0 < A2 < 1 < AI)' with
corresponding proper directions X and Y. The differential * is dilating

in the direction X and is contracting in the direction Y. To be precise,
let Xm and Ym be the subs paces of TM m, respectively parallel to X and
Y. Then:
II*(II
A1
'1I(1I
if (( X m ( \ > 1) ;
II"(II ::;
A2
'11(11
if (( Ym , (0 < A2 < 1)
This is a characteristic example of the C-systems that we define next.
Figure 13.2
55
UNSTABLE SYSTEMS
DEFINITION 13.3
Let be a C 2 -differentiable diffeomorphism of a compact, connected,

smooth manifold M. We denote the differential of by *.
(M, ) is called a C-diffeomorphism if there exist two
fi~/ds
of tangent
planes Xm and Ym 'such that:

(1) TM m falls into the direct sum of X m and Y :
m
TM m =Xm mY,
dimX m =kIO, dimYm
m
(2)
110.
For every positiVe integer and for some Riemannian metric g:
1I(,n)\~'11 2 a' i,n Iltll,
II(-n)*tll
11(n)*tll
II(-n)* til ~ a' e.\n Iltll,
:S b e-.\nlltll,
:S b e-.\nlltll, if
if
t ( X m,
t ( Ym
The constants a, b, .\ are positive and independent for n and

b
depend 2
but a and
on the metric g. Xm is called the dilating space, and Ym is the
contracting space.
Example (13.1) is a C-diffeomorphism:
a = b =
1,
.\.
= "1'
-.\
= "2
This definition extends to the continuous case (t (R): Let t be a oneparameter group of C 2 -differentiable diffeomorphisms of a compact, connected, smooth manifold M.
d~
(M, t) is called a C-flow if:
(0)
the velocity vector
tmlt=o does not vanish;
(1)
TM m splits into a direct sum:

TMm=XmmYmmz~,
where Zm is the one-dimensional space spanned by the velocity vector at

m, and dim X m
2
k .;, 0,
dim Ym
I ,;, 0 i
And hence for every metric: Let gl and g2 be two Riemannian metrics on M.
Due to the
compactness
of M, there exist two positive constants a,
ail til 2:S I till :S f3lltll2
t(
f3
such that:
for every
TM . . Thus, if the inequalities 2 hold for
~ I with constants a and b, they still hold for ~2 with corresponding constants
(al (3) a
and
metric g.
(f3/ a) b.
This proves the independence of the definition from any
56
(2) for any positive real number t and for some Riemannian metric g:
11(1)/tll ~ aoeAtlltll, 11(1)_/t!1 ~ boe-Atlltll, if t (X m ;

11(1) /tll ~ boe-Atllt!!, 11(1)_/tll ~ aoeAtlltll, if t( Ym .
The constants
a, b, A are positive and independent
for t and
t,
but
and b depend on the metric g. Condition (D) means that the system has
no equilibrium position. Condition (2) describes the behavior of the orbits. A C-diffeomorphism or a C-flow will be called, for short, a C-system
REMARK
13.4
It is easy to show that:

(1) the subspaces Xm and Ym are uniquely determined (they are, respec-
tively, the "most dilating" and the "most contracting" subs paces of
TM m );
(2) dim X m = k and dim. Ym = I do not depend on m (k is a continuous

function of m, with integer values, on the connected space
M)~
(3) Xm and. Ym depend continuously on m.

Finally, observe that a C-system is not a classical system (see Definition 1.1) since we do not postulate the existence of an invariant measure.
Now let us show how to construct certain C-flows from C-diffeom:>rphisms
EXAMPLE
13.5 (SMALE)
The Space M.
Let T2
Iu I 0
= I(x, y) mod 1 I be the two-dimensional torus and [0, 1] =
~ u ~ 110 We construct the cylinder T2 x [0, 1], and after we identi-
fy T 2 x
101
and T2 x 11 I according to:
x, y), 1) ;:
where
1>
(1) (x, y), 0)
X+ y,
x+ 2y), 0) (mod 1) ,
is the diffeomorphism (13.1):
1> : ( ; ) ->
nD(;)
(mod 1) .
57
UNSTABLE SYSTEMS
il
Figure 13.6
We obtain a compact manifold M. Let (x, y, u) be a point of M. The mapping p: M ~ 51
= ! u (mod
Hence, M is a fibre
bundle 3
1) l, p (x, y, u)
=u
has rank 1 everyvihere.
with basis 51 and fibre T2
The flow /.
We define a flow / by its infinitesimal generator:
x = 0,
(13.7)
= 0,
u=
1 .
An Auxiliary Riemannian Metric

Let Al and A2 , (0
<
A2
< 1 < Al ),be the proper values of:
(~ ~).
We define a Riemannian metric on T2 x [0, 1] by:
3
A differentiable fibre bundle (M, B, p) over B consists of the following: (i) a
compact, connected, smooth (n+ q) -dimensional manifold M; (ii) a smooth, n-dimensional manifold
called the base; (iii) a C 2-differentiable mapping p:
whose rank is n everywhere and called the projection. The p-l(b) 's,
p,i-> B
b ( B, are
called the fibres. They are q-dimensional manifolds diffeomorphic one to another.
58
ds 2 =,\2u[A 1dx + (1-A 1 )dy]2 + Ai U [A 2dx + ~dy]2 + du 2 .
(13.8)
c)::-,vt" )
It is readily proved that this metric is invariant under the substitution;
x
->
x + y,
x + 2y,
->
->
u-1 .
I
In other words, this metric is compatible with our i~entification of T2 x 101

and T2 x
111.
Thus (13.8) can be considered as a metric of M.
(M, t) is a C-f1ow.
Take a look at conditions 0, 1,2 of (13.3).

(0) From (13.3) the velocity vector is nonvanishing.
(1) If m = (x, y, u) { M, we define three subs paces of TMm :Xm (resp.
Ym ) is tangent to the fibre T2
x I ul
and is parallel to the proper direction
of :
: T2 x lui
(~)
(b
->
->
T2
x lui ,
n(~)
(mod 1)
with oorresponding proper value A1 (resp. A2 ). Zm is collinear to the

velocity vector (13.7). Conditions (1) are fulfilled:
(2) In the chart (~, y, u) the components of ~ { Xm are of the form:

(s, S(A 1 -1),0),
R.
On the other hand, according to (13.7), the matrix of t* reduces to the

identity. We deduce from (13.8):
11(;)~112 =~~(u+t)[Als + (l-A 1)(A 1 -1)s]2 = A/ t 11~112

.
Hence, II(;)~II
Ai I ~II. This proves the first group of conditions (2) of
(13.3):
a
= 1,
The second group is proved ift the same way.
59
UNSTABLE SYSTEMS
The field of two-planes Xm E9 Zm (resp. Ym E9 Zm) is clearly smooth and

completely integrable: it defines a foliation 4 on M. Each sheet is the
union of orbits of -00
(or +
00;
CPt m
Figure 13.9
This property will be proved general for C-systems.

REMARK
13.10
The previous construction is quite general. Let (V, <Po) be a C-diffeo"

morphism. We construct a compact manifold M by identification of (v, 0)
and (<pov, 1) in the topological product V x [0, 1].
We define a C-flow over M by setting:
where v ( V,
4
[0, 1], <!,nd [a] means the integral part of a.
M, we mean a completely integrable field of k-planes

M. The connected complete integral manifolds are called the sheets. They
By a foliated manifold
over
are k-dimensional submanifolds.
60
14. Geodesic Flows on Compact Riemannian

Manifolds with Negative Curvature
We turn next to an important example of C-system.

DEFINITION 14.1. MANIFOLDS WITH NEGA TIVE CURVATURE 5
Let v be a point of a Riemannian manifold V and TVv the tangent

vector space at v. Two noncollinear vectors e 1 , e 2 of TVv define a twoplane (e 1 , e 2 ). The geodesics of V emanating from v and tangent to (e 1 ,
e 2 ) generate a surface
~.
The Gaussian curvature of
is a Riemannian submanifold of V.
~
at
is called the sectional curvature
p(e 1 , e 2 ) of V in the two-plane (e 1 , e 2 ). If the sectional curvature is
negative for every (e 1 , e 2 ), V is called a manifold of negative curvature.

Then, if V is compact, the continuity of (e 1 , e 2 ) implies there exists a
constant _k2 which bounds the sectional curvature from above. Appendix
20 provides an example of a manifold with negative curvature.
INSTABILITY OF THE GEODESICS 14.2
The geodesic flow t of a Riemannian manifold V describes the movement of a material point on the frictionless manifold II without external
forces [(see (1. 7)], If V has negative curvature the geodesics are very un-
stable: if v, Vo ( Tl V , the distance ltV, tVol increases exponentially with t. To be precise, we. have the following result:
LOBA TCHEWSK Y - HADAMARD THEOREM 7 14.3
Let V be a compact, connected Riemannian manifold of negative curvature. Then, the geodesic flow on the unitary tangent bundle M
Tl V
is a C-flow.
~ppendix
21 details the proof we sketch here (see Figure 14.4).
For further infonnation see S. Helgason[I], Chapter 1.

6 .
V=
universal coyering of
V.
This theorem is "due to Lobatchewsky for surfac:.es of constant negative curvature. Hadamard
[1]
extended it to surfaces of arbitrary negative curvature.
61
UNSTABLE SYSTEMS'
y'
positive
asymptotes
negative
asymptotes
Figure 14.4
Let y (u, t)
= y (t) = y
etrized by arc length t.
be a geodesic emanating from u ( T1 V and paramLet x be a point of V. There exists a geodesic
Y1 . issuing from x and passing through- y (t1)' As t1 .... + 00, Y1 converges

to a limit which is a geodesic y '(u', t) emanating from u' ( TVx '
For a suitable choice of the origin of y, it can be proved that:
(14.5)
distance (y (t), y '(t
5. b e->'"
t ,
t ::::
0,
where the constants b and A are positive and independent for y, y', t.
Geodesics such as y' are known as the positive asymptotes to y. They

can be proved to be orthogonal trajectories to (n -1) -dimensional submanifolds (n = dim V): the so-called positive horospheres S+.
Let us denote by S+ (u) the horosphere emanating from the origin of
u ( T1 V and which is orthogonal to the positive asymptotes of y (u, t).
This horosphere can be interpreted as an (n -1) -dimensional submanifold
of T1 V: S+(u) is the union of its. orthogonal unitary vectors oriented as u.
The tangent plane at u of S+(u) C T1 V is an (n -1) -plane Yu of T( T1 V).
62
Exchanging the role of t and - t, we define negative asymptotes and

negative horospheres S- in the same way. The tangent plane at u of
S-(u) C Tl V is an (n -1) -plane Xu of T (T1 Vu). From the very definition,
we have:
where' Zu is the one-dimensional space generated by the velocity-vector

of the geodesic flow. That is the condition (1) of C-flows (Definition 13.3).
Condition (2) comes from (14.5).
Observe that the fields Xu and
u are completely integrable. Their
integral manifolds are the horospheres S+ and S-. Both of these foliations are invariant under the geodesic flow, for the horospheres are orthogonal trajectories of (n-l) -parameter families of geodesics of V .
. We turn to prove that general C-systems admit two invariant foliations.
15. The Two Foliations of a C-System

Let(M, ) be a C-diffeomorphfsm; Xm (resp. Ym ) denotes the k-di-
mensional dilating space at m ( M (resp. the IGdimensional contracting

space). A Riemannian. metric on M is definitively selected. Hence, Xm
and Ym are:Euclidean su~spaces of TMm'
SINA(THEOREM 8
15.1
Let (M, ) be a C-diffeomorphism, then:

(1) There exist two foliations
X and
~ that are invariant under
and that are respectively tangent to the dilating field Xm and the contracting field Ym. Hence, these fields are always integrable.
(2) Every diffeomorphism ': M
-+
M, C 2 -cIose enough to , is a
C-dilleomorphism. The dilating and contracting foliations
X, and
~' of '
depend (;Ontinu6usly on '.

8 This w~s proved essentially in the paper by.V. I. Arnold and Y. Sinai [6]; al~hough
their discussion was concerned with the particular case of small perturba-
tions of automorphisms of a two-dimensional torus, the proof extend to the general

case.
63
UNSTABLE SYSTEMS
Appendix 22 completes the proof we sketch here.

CONSTRUCTION
15.2
The. space K of the fields p of the tangent k-planes to M inherits a

natural metric Ip1 - P2 1 which makes it into a complete metric space. Let
p be such a field, and p (m) the k-plane of TM m' The diffeomorphism cp

induces a mapping cp**: K ~ K:
cp**(m)
cp*p(cp- 1 m).
where cp * is the differential of cp which maps a k-plane of TM m onto a

k-plane of TM cp(m)
The dilating and contracting fields X and Y of cp are clearly fixed
points of cp**. It can be proved (see Appendix 22) that the axioms of Csystems imply that cp ** (or a positive integer power (cp **)n) is contracting in a neighborhood of the dilating field X:
(15.3)
for
IX -P 1 1 < 0, IX -P 2 1 < 0,
small enough. Of course, (15.3)
and
still holds for any diffeomorphism cp'. C 2 -c1ose enough to cp, since cp'*
is C 1-close to cp *. We deduce from the contracting mapping theorem that
a mapping verifying (15.3) admits a fixed point. The fixed point of the
mapping cp is X, but for cp' it is another field p'. Clearly:
p'
(cp ,**)n X ,
lim
n=~
cp'*p'(m)
p'(cp'm) ,
and the field p' is dilating for cp'. A similar study of (M, cp -1) leads to
the contracting field of cp' that is close to Y. Thus cp' is a C.diffeomorphism.
THE INVARIANT FOLIATIONS
15:4
First assume that (M, cp) possesses two invariant foliations
and
tangent, respectively, to X and Y. Then, the same property holds for

(M, cp '). In fact, the invariant field p' of cp' is obtained as:
p'
lim (cp ,**)n X

n=oo
y,
64
But ( ,**)n X is the field of the i<Lplanes tangent to the foliation ,n:x: .
Thus, the limit field is. integrable, that is tangent to a certain foliation
:X:'.
This completes the proof of the part (2) of Theorem (15.1). As a matter of
fact, the previous argument proves that there exists a foliation
:X:.
Let us
cover the compact manifolds M by a finite number of local charts (C j , !/Jj);

each C j is a neighborhood of a point mj and !/J j is the mapping !/Jj: C j
Rn
(n
->
dim M).
Consider a foliation
:X:?1
of C.1 such that 'fJ

./..I CX?)
be the foliation of
I
!/J.(
C.);
which consists in planes parailel to !/J~I (X ml ). If the local charts
I
I
are selected small enough, then the tangent plane XO(m)

of
I
:X:?I
at m
C.I
is close enough to the dilating plane Xm ' for all' m. Obviously there are
se~eral planes.
X?(m)
passing through m if m belongs to several C I..
I
Consider the foliations
:X:7
n:x: jO of nC i' These foliations cover
M and (15.3) implies that their fields of tangent planes con verge to X as
->
+ 00. We deduce readily that there exists a limit foliation :x: tangent to
X. This concludes the proof.

REMARK
15.5
Each sheet of :x: is C1-differentiable. But the foliation :x: is not required to be smooth: the normal derivative to t!le sheets may not exist. If
k = I = 1, the field X is C1-differentiable (see Arnold and Sinai [6]). It
seems probable that :x: is not smooth in the general case. Anosov concocted an example where :x: and 'Yare not C 2 -differentiable.
REMARK
15.6
The preceding proof extends to C-flows.

r
16. Structural Stability of C-Systems

In this section we prove that C-systems are structurally stable.
DEFINITION
16.1'.
STRUCTURAL STABILITY
(A) Diffeomorphisms
Let M be a compact, smooth manifold and : M

able diffeomorphism...
-+
M a CT-differenti_
6S
UNSTABLE SYSTEMS
By definition, is structurally stable if given a neighborhood V (Id M )
of the identity Id M (in the CO-topology 9) there exists a neighborhood W()

of (in the Cr-topology) such that if
t/I (
W(), then there exists an ho-
meomorphism k ( V(IdM ) making the following diagram commutative:
that is: k.
onto the orbits

(B) Flows
t/I. k. In
of It/ln I n
other words, k maps the orbits of

(
I" In ( ZI
ZI.
\
Let M be a compact smooth manifold and X a Cr-differentiable vector

field which generates a flow t:.
X(m) =
d.,l., m I
at
'f't
t=
m (M.
By definition, t is structurally stable if given a neighborhood V(IdM ) of

the identity IdM (CO-topology) there exists a neighborhood W(X) of X (C r_
topology) such that if Y (W(X), then there exists an homeomorphism k (
V(IdM) that maps each orbit of X onto an orbit of Y. In the future we assume r S 2.
REMARK
16.2
One might restrict k to be a diffeomorphisjJlJather than an homecimor~

phism. Then, consider in R2 the system:
x = y,
y = -x-Ky,
where K is a positive constant. The orbits are spirals focusing to the

singular point (0,0) (see Figure 16.3). But, according to the Poincare
theory 10 of proper values, K is a continuous invariant of diffeomorphisms.
Two mappings (or two vector fields) are C'-close if their derivatives of order
inferior to r+ 1 are close.
9
10 See, for instance, K. Coddington and Levinson
[1].
66
Consequently, two systems with distinct values of K would not be topologically conjugate and the focus (0,0) would not be structurally stable.
y
focus
Figure 16.3
In the continuous case, one could be tempted to propose a definition

quite similar to that of the discrete case: there exists an homeomorphism
k ( V (Id M) that makes the following diagram commutative for all t:
where rjJ t is the flow generated by Y. But, in such a definition, a limit

cycle (see Figure 15.4) would not be structurally stable since its period
is a continuous invariant (see footnote 10, p. 65).
Two problems arise naturally:
(1) What are the structurally stable phase-portraits?
(2) Given a manifold, are the structurally stable vector fields generic (or
dense) in the space of the vector fields?
UNSTABLE SYSTEMS
67
Figure 16.4
Andronov and Pontrjagin [1] gave an affirmative answer if M is the

sphere S2. The other two-dimensional manifolds were studied by Peixoto
[1].
1 the dimension is greater than two, the situation is involved. For instance, the system (13.1) is structurally stable but very complicated 11 (ergodic, the cycles are everywhere dense, and so on).
On the other hand, Smale [2] constructed an example (M, ) such that:
dim M ~ 3, every diffeomorphism close enough to is not structurally stable (see Appendix 24). This invalidates the genericity of structurally stable systems.
ANOSOV'S THEOREM 12
16.5
C-Systems are structurally stable.

Sketch of the proof (see Appendix 25).
Let (M, ) be a C-diffeomorphism, and W() a neighborhood of
(C 2 -topology) in the space of the diffeomorphisms of M. We prove next
that given '
W(), there exists a small and well-defined homeomor-
phism k: M .... M such that:
11
See S. Smale [1].
12
See Anos ov
[11.
68
'==k..r 1
sup d[k {m), m] < ,

'mM
where d ( , ) is the Riemannian distance. We prove next that , which

depends on ', is bOlmded from above:
sup
W()
< 1 .
Hence the structural stability holds good. In fact, we make correspond

to 1 a neighborhood W() such that given '
W() there exists a
homeomorphism k: M ... M satisfying:
sup d[k (m), m] < f:: 1
mM
Let '
W() be a diffeomorphism, C 2 -close to
. We already know that
, is a C-diffeomorphism and and 'have invariant dilating foliations X.

X' and invariant contracting foliations Y. Y' (Sinai's theorem, 15.1). If
there exists an -homeomorphism k: M ... M such that ' ==k . k- 1 , setting m' == km, we have:
(16.6)
for any n
Z.
From the very definition of C-systems we see that there exists at most
one point m' satisfying (16.6). In fact, for any
in one or the other case n ... +
00
~.f,
0, we have:
(or -(0). The uniqueness of the homeo-
morphism k Qeing proved, if k exists we obtain it by making correspond

to each point m the only point m' that satisfies (16.6) for all n. Theproblem is reduced toproving that such a point m' exists. The construction
of m' is as follows.
The images 'n f3 (n > 0) of each dilating sheet
f3
X', close to m,
clearly have some points neighboring n m (see Lemma A, Appendix 25
69
UNSTABLE SYSTEMS
for the precise meaning). It can be proved (Lemma B, Appendix 25) that
there exists, among these sheets, a unique sheet (j(m) such that its images
cp '" (j(m) are still close t~ cp"(m) for all n
<O.
The same device proves that there exists a unique contracting sheet
8(m) (
'Y'
whose images cp '"8(m) are close to cp"(m) for all n (' Z. The
foliations ~' and
'Y'
are transverse; thus (3(m) and 8(m) intersect in a
unique point m' in the neighborhood of m. It is eas y to see that cp '(m ')
=
(cp(m)) , and that m' varies continuously with m; thus the mapping k:
m ... m' is the desired homeomorphism.
C-FLOWS 16.7
Anosov's theorem extends to C-flow.

~(m)
Figure 16.8
A similar construction (Figure 16.8) gives two sheets, (3(m) and 8(m).
They are formed with orbits of cp; (respectively for t
which are
a~ymptotiC
->
+ 00 and t ... -(0)
to an orbit of cp; that is their intersection. As a
geometric (nonparametrized) curve, this orbit is close to the orbit cp/m.

The desired homeomorphism k is obtained by making correspond to m the
nearest point 0f (3(m) n 8(m) in the. Riemannian sense.
70
17. Ergodic Properties of C-Systems

Throughout this section we study C-systems that possess a positive
and invariant measure 11. Thus, in contrast with the preceding sections,
we 8!e concerned with classical dynamical systems (see Definition 1.1).
Let us begin with ergodic properties of the automorphism (13.1): M is
the torus I(x, y) (mod 1) I; the invariant measure is dll = dx dy; the automorphism is:
: ( ; ) -THEOREM
G~ )(;)
(mod 1) .
17.1
(M, 11, ) is a K-system
With a view to proving this theorem, we construct a subalgebra (f of

THE SUBALGEBRA
(1'. 17.2
The matrix:
(~ ~)
y
x
Figure 17.3
f.
71
UNSTABLE SYSTEMS
has two real proper values A1, A2 :

< A2 < 1 < A1 with corresponding
proper directions X and Y. We split the unit square of (x, y) into a square
(3) and four triangles (1, 2, 1 " 2 '), with sides parallel to X and Y (see
Figure 17.3). By pairwise identification of the opposite sides of the unit
square, 1 and 1', 2 and 2', and 3 give rise to three disjoint parallelograms
P1,P2, P3 on the torus M. Again split P1, P2 , and P3 into parallelograms
the sides of which are parallel toX and Y and whose lengths are inferior
to a constant A. We get a partition {3 of M and we set:
V (p-n{3.
n~O
The subalgebra
ct
is the closure of the algebra m(a) generated by a (see
Appendix 18).
17.4
LEMMA
If the constant A is small enough, then:

(1) The atoms of
ct
are segments that are parallel to the contract-
ing direction Y.
(2) Let I be a given positive number. The measure (in the sense
of dxdy) of the union of those atoms whose length is inferior to I isbound-
ed from above by C I, where C is an absolute constant. In particular, almost every atom of
ct.. is a
segment and does not reduce to a point.
Proof:
We argue in the covering plane (x, y). An element of {3 V -1 {3 is of

the form:
B1 n-l B2 ;
B 1, B2
'f
{3.
-1 B 2 is a paraUelogram with sides parallel to X and Y and whose

lengths are, respectively, inferior to
sections B1
-l B2
Ai- 1 . A
and A;l. A. On M, the inter-
n -1 B2 results from the intersection on
with six parallelograms B 1, TB 1, l
'that are deduced ,from B I C
(x, y) ~
M of
1B 1, tB1' tTB1' tT- 1B I ,
[0: 1] x [0, 1] by the translations
t ~ (-1,0). The projections of these six parallelograms 1nto
T ~ (1,0),
72
it
c (X, y) =
are six segments, the lengths of which are inferior to A and whose mutual
distances are greater than K - A, where:
K
as Ik 1 + 1 k'i
I<
inf 1 k cos e-k' sin e 1
'
0; k = -1,0,1; k' = - 2, -1, 0, 1, 2; and
e is the angle
of X with Oy. Consequently, if 2A is inferior to K, -1 B 2 intp.rsects

at most one of those six parallelograms. Thus, on M, Bin -1 B2 is a
connected set: it is a parallelogram with sides parallel to X and y, the
lengths of which are, respectively, inferior to A~1 A and A.
Exchanging {3 and (3 V -1{3, the above argument applies for
Max(A
1 1 A, A)
::; A.
TJlUs, the elements of:
are parallelograms with sides parallel to X and Y
and the lengths of
which are, respectively, inferior to A~2 A and A.

By induction on n, we see that the c~lements of {3V-I{3VV-n{3
are parallelograms with sides parallel to X and Y and the lengths of which
are, respectively, inferior to A1- n A and A. As n
-+
+ 00, we obtain the first
part of Lemma (17.4), Al > 1. Let L be the sum of the lengths of the
sides of the elements of {3 that are parallel to X. The analogous sum for
P{3, p ( Z, is
At L.
Thus, the analogous sum for a is bounded from
above by.
def
I-A -1
C.
Some of the elements of a have a side parallel to Y with length inferior

to 1. Consider the set of these elements and dellote by m the measure of
their union. From the preceding formula we deduce at once that m < C 1 ;
this proves part (2).
73
UNSTABLE SYSTEMS
THE CONTRACTING FOLIATION 17.5

Consider a nonnull constant vector field on M whose vectors are parallel to the contra(.ting direction Y. The integral lines are called the con-
tracting sheets (see 15). This foliation is called ergodic if any union of
sheets, with positive measure, coincides with M up to a set of measure
zero.
LEMMA
17.6
The contracting foliation is invariant under and ergodic.

Proof:
The invariance follows from the invariance of Y. The contracting
sheets are the orbits of a translation group of M. To prove ergodicity it
is sufficient to prove that each sheet is everywhere dense (Appendix 11)
cr, equivalently, is not closed (Jacobi theorem; Appendix 1).
Assume
there exists a closed contracting sheet F with length f. The invariance

of the foliation implies that n F is again a closed sheet with length
A; .f.
Since 0 < 1..2 < 1, we have
A;' f
.... 0 as n ... + 00.
But this con-
tradicts the obvious fact that the length of any sheet has to be greater
than 1.
Proof of Theorem 17.1:

According to the very definition of U': U' C U' this is the condition
(a) of Definition (11.1). Let us prove the condition (c):
00
V nU'
1.
n~O
It is sufficient to prove that the atoms of

00
V na
n~O
have arbitrarily small length. According to Lemma (17.4) the atoms of
a V .. V na are of the form:

AonA1n . nnA n ;
AO, ... ,An (a,
74
where the A;'s are segments parallel to Y and with length inferior to A.
Since A is small enough, then
Ao n Al n ... n n An
is a connected
segment (see argument of Lemma (17.4)). The length of this segment is

inferior to A2 A, for is contracting in the direction Y
As n
-+
+ 00, we have A2 A
-+
(0 < A2 < 1).
0; this proves the condition (c).
We now turn to the condition (b). Let H be an element of positive

measure of
-00
H is the union of atoms of (j', the union of atoms of -1(j', -2(j', and so
on. Let E be an arbitrary positive number inferior to 1. Take a number I

such that:
CI
<
E p.(H),
where C is the constant of Lemma (17.4). Up to a set E of measure inferior to E p.(H), the atoms of (j' have a side parallel to Y whose length
is greater than I. Since -1 is measure-preserving and dilating in the
direction Y with ratio A;l, one concludes that, up to a set -nE of measure inferior to . p.(H), the atoms of -n(j' have a side parallel to Y
whose length is greater than
~-n
. 1. As n
-+
+ 00, we obtain-up to a set
of measure inferior to E p.(H) -H is the union of contracting sheets.

From the arbitrariness of E and from the ergodicity of the contracting foliation (Lemma (17.6)), one deduces at once that p.(H)
1. Whence:
00
(Q. E. D.)
The above arguments extend to general C-diffeomorphisms. The delicate

point consists in constructing a partition similar to
f3.
A theorem due to
Anosov [2] overcomes this difficulty. First we need a definition.

Consider a foliation of an n-dimensional Riemannian manifold Minto
p-dimensional sheets.
Let us take any small element of an (n-p)-
75
UNSTABLE SYSTEMS
dimensional mc....,ifold II transversal to the sheets. Assume that near II

there is another such manifold II' and that each sheet passing through m
f
II intersects II' in a point m' near m in the induced Riemannian metric
of the sheet. Consider the map f: II.... II' taking m to m '.

DEFINITION
17.7
If for any such pair of neighboring manifolds the map f has a continuous generalized Jacobian, and if for smaIl deformations of II' this Jacobian varies continuously, then the foliation is caIled absolutely continuous.
For both of the two foliations of a C-system (see 15), Anosov proved:
THEOREM
17.8
The foliations ~ and Yare absolutely continuous.
We cannot give here the proof of Theorem (17.8), which is too long. We
only announce the final results obtained with its help. The statement of
Theorem (17.1) extends and leads to the following theorems (Anosov [2]).
THEOREM
17.9
Every C-system (C-diffeomorphism or C-fIow) is ergodic.

THEOREM
17.10,
(SEE SINAI
[11])
Every C-diffeomorphism is a K-system.
For C-flows the situation is more complicated. The C-flow

ample (15.5) has nonconstant proper functions. Consequently,
1>t of Ex1>t does not
have Lebesgue spectrum (see 1O) and cannot be a K-flow (Theorem 11.5).
The following theorem proves that, in some sense, Example (15.5) is the
only one exception.
THEOREM
Let
1>t
17. 11
be a C-fIow on a compact, n-dimensional manifold M, then:
(1) either
(2)
1>t
1>t
is a K-system; or
has nonconstant proper functions.
In case (2), such a proper function is continuous and there exists a compact, (n -1) -dimensional submanifold V of M and a C-diffeomorphism
76
1>:
->
V such that
1>/
is the one obtained from
1>
by means of the con-
struction described in (13.10), up to a change of the time scale
t (t ...
C t, C = constant).
COROLLARY 17.12
The geodesic flow on the unitary tangent bundle of a Riemannian manifold 'V of negative curvature is a K-system.
Proof:
According to Theorem (14.3), the geodesic flow
from Theorem (17.9),
1>/
1>/
is a C-flow. Hence,
is ergodic.
On the other hand, by passing to the double covering, we may assume

V orientable. If V is two-dimensional, it is not diffeomorphic to the torus
T2 since the Gauss-Bonnet formula implies that the E;uler-Poincare characteristic is negative. Consequently, according to Corollary (A 16.10) of
Appendix 16, the only continuous proper functions of
1>/
are constants.
The second case of Theorem (17.11) cannot occur and this theorem proves
that
1>/
is a K-flow. 13
Thus,
1>/
has positive entropy (see (12.31)), has denumerably multiple
Lebesgue spectrum 14 (see (11.5)), is mixing 15 (see (10.4)), and ergodic 16

(see (8.4)).
18. Boltzmann-Gibbs Conjecture

The methods and ideas of the preceding sections are applicable to cere
tain problems of Classical Mechanics, for ex-ample to the Boltzmann-Gibbs
model of a gas. This model consists of hard spheres contained in a parallelepipedic box with rigid walls. Collisions between spheres, or spheres
13 This result is due to Smai
[10] for compact surfaces of negative curvature.
14 This result is due to Gelfand and Fomin [2] for arb~trary dimension and constant negative curvature.
15 This result
1S
due to E. Hopf
16 This result is due to Hedlund
[1l in the case of constant negative curvature.

[1l and E. Hopf h] for surfaces of negative
curvature and manifolds of constant negative curvature.
77
UNSTABLE SYSTEMS
and wall, are supposed to be perfectly elastic. Sinai has proved ergodicity17 of this system on each manifold T
constant ~ O.
Ergodicity derives from the collisions. As the simplest model let us

consider the motion of two perfectly elastic circular particles on the surface of a
two~dimensional
torus having a Euclidean metric. For simplicity
we consider one of the particles as fixed. The second particle (which can
now be regarded as a point) moves on a "torus billiard table" (Figure 18.1),
being reflected from
t~e
fixed circumference according to the law "the
angle of incidence a is equal to the an gle of reflection
f3."
Let us, at the same time, consider an elliptic billiard table (Figure
18.2). The ellipse can be regarded as an oblate ellipsoid on which the

I
I
I
I
I
I
I
-;------------I
Figure 18.1
17
This result was conjectured as the "quasi-ergodic hypothesis." At the beginning of Ergodic Theory this conjecture occasioned sharp discussions. But the
problem was, perhaps, overvalued: statistical mechanics deals with asymptotic
behavior as N -> +00 (N = number of parhcles) and not as I'" +00 for fixed N.
78
ERGODIC PROBLEMS OF CLASSIC' AL MECHANICS
point moves along a geodesic, passing at each reflection from one side to
the other. In the same way a torus billiard table can be regarded as a twosided torus with a hole on which the point moves along a geodesic. But if
the two-sided ellipse is a'n oblate ellipsoid, the two-sided torus with a hole
will be an oblate surface of genus 2. Thus, motion on our tor~s billiard
table is a limiting case of geodesic flow on a surface of genus 2.
Figure 18.2
We now turn to our billiard tables and consider the curvature to which
Figures 18.1 and 18.2 correspond. The ellipsoid has a positive curvature
whose integral is equal to 411 (Gauss-Bonnet formula). On flattening the
ellipsoid, all the curvature is accumuiated along the boundary of an ellipse.
For a surface of genus 2 the integral of the t:urvature is equal to -411. Thus,
a two-sided torus billiard table can be regarded as an oblate surface with
negative curvature everywhere: on flattening, all the curvature is accumulated along the circumference.
The preceding arguments (Arnold [4] p. 184) are not, of course, a proof
of ergodicity, not even in the simplest cases we considered. Nevertheless,
using methods and notions dealing with C-systems (asymptotic orbits, transverse foliations), Sinai [4], [5] succeeded in proving that the BoltzmannGibbs model is ergodic on each submanifold T
more, is a K-system.
constant
f.
0 and, even
UNSTABLE SYSTEMS
79
The proofs of these results require hundreds of pages and contain a

theory of generalized C-systems with discontinuous foliations. We only
mention that the general case reduces to a "billiard table" problem in the
configuration space: the colliSion-hypersurface of the configuration space
plays the part of our circumference.

Anosov, D. V., Geodesic Flows on Compact Riemannian Manifolds of
Negative Curvature, Trudy Instituta Steklova 90 (1967).
Hopf, E., Statistik der geodatischen Lin~en in Mannigfaltigkeiten negativer
Kriimmung, Ber. Verh. Sachs. Akad. Wiss., Leipzig 91 (1939) pp. 261-
304.
Sinai, Va, Dynamical Systems with Countably Multiple Lebesgue Spectra
II. lsvestia Math. Nauk.30 No.1 (1966) pp. 15-68.
CHAPTER 4
STABLE SYSTEMS
Many dynamical systems are known, the orbits of which, with remarkable stability, fail to fill up the "energy level"
H =, Ct -ergodically
and remain (to the end) in their particular corner of phase-space. The systems that are close to an "integrable" s'ystem, and the systems to which
the Theory of Perturbations of Celestial Mechanics applies, belong to this
class. Let us mention, for instance, the three-body problem, th: fast rotations of a heavy-rigid body, the motion of a free point in a geodesic on a
convex surface, and the adiabatic invariants.
Only in recent times, "beginning with the work of C. L. Siegel [1] (1942)
and A. N. Kolmogorov [5] (1954), has some progress been made in studying
these systems. This chapter reports on the present status of this problem.
We refer to V. Arnold [4], [5] for further details.
Let us begin with an example.
19. The SWing and the Corresponding
Canonical Mapping
EQUATIONS OF THE MOTION 19.1
Pendulum equations are:

(19.2)
q = p,
2
P = -w Sin q,
where w is the "proper frequency" that depends on the length of the penduillm.
A swing is a pendulum whose length I periodically varies (Figure 19.3).
Its equations can be written as
81
82
Figure 19.3
(19.4)
q = p,
_w 2 (t) sin q,
where w(t+ r) '" ,w(t). Appendix 5 studies (19.2) by using the phase-plan.e.
Equations (19.4) contain the time variable t explicitly.
Thus, we deal
with the study of a vector field in the three-dimensional space p, q, t (see

Figure 19.5).
Figure 19.5
83
STABLE SYSTEMS
THE MAPPING T 19.6
The initial data p(O) = Po, q(0) = qo define an orbit p = p(t), q
q(t). The time periodicity of Equations (19.4) allows one to identify the
surfaces t
=0
and t
the space R1 x Sl x Sl
T of the surface
r. Thus, (19.4) can be regarded as equations in

Ip, q mod217, t mod
~ (t= 0)
rI. This defines a mapping
into itself:
Clearly:
p (nr), q (nr) = Tn(po' qo) ,
and the study of p (t), q (t) as t ... + oc reduces to the study of the iterates Tn, n ( Z. The mapping T is canonical, for the Equations (19.4)
are canonical. Therefore, T preserves the area dp 1\ dq.
The equilibrium positions p
= 0,
k 17 (k= 0, 1) are solutions of
(19.4). Thus, the points p = 0, q = krr are fixed points of T.

19.7
THE INTEGRABLE CASE
Let us
be~in
"inte~rable
with the
case" (,)
constant, to get acquaint-
ed with the mapping T. In this case, the system is conservative; thus, the
energy is invariant. In other words, on the surface ~,the curves (Figure
AS.l, Appendix S):
r:
'hp2 -
(,)2
cos q
are invariant under T.

L~t us consider the part of ~ that is interior to the separatrix (h
< (,)2).
We use "action-angle variables 1, " to study T. It can be proved

(Appendix 26) that there exists a canonical transformation p, q ... 1, such
that the equation 1
Ct defines an invariant curve
nate (mod 217) is the angular coordinate of
r[
r[.
The coordi-
and, in coordinates 1, '
the mapping T becomes:

T: 1, ... 1, + AU)
~his
means that the curve
r[
rotates through an angle A(I) that is con-
84
stant along each curve 11 (if is selected as parameter), but depends on

the curve. It is readily seen that lim A(I)
to the separatrix V2p2 - cu 2 cos q
=0
as the curve 11 converges
cu 2 , and that 1 lim A([)
CUT
as 11
converges to (0,0). Hence, some curves I~ are rotated through an angle
A([) commensurate with 277, and others through an angle incommensurate

with 277. Let us consider the iterates of T. If x
curve 1 such that
A=
277 ~ , then: Tnx
belongs to a
= (p, q)
x .. Thus, each point o~ 1 is
a fixed point of Tn, and the orbit of x consists in a finite number of points
(Figure 19.8). If the angle A (I) that corresponds to 11 is
incommen~urate
Figure 19.8
with 277, then the points
Finally, observe that the equilibrium position p
= q =
IXol = jp~ + q~
IT" xol remains small for any
n ( Z.
is small enough, then
T" x are everywhere dense on 11. (Appendix 1).
The linear part of T at zero (p, q
1): It
= p,
fJ
= _CU 2 q
0 is stable: if
is a rotation with
frequency CU (see Appendix 5). This gives rise to a rotation with angle II)T after
time
T.
85
STABLE SYSTEMS
NONINTEGRABLE CASES 19.9
Suppose that
is a nonconstant periodic function. We assume addi-
tionally that w(t) is close to a constant w O' for instance:
o<
The mapping TE , which corresponds to
T. This mapping
WE'
1,
v = 27T/r.
is close to the above mapping
preserves the area dp /\ dq and the point (0,0), but
neither the energy nor the curves
rl
are preserved. The main purpose of
the Theory of Perturbations is to study the behavior of the iterates
-+
+ 00 and E
(1) E
I,
T; as
1. There are two ways to consider the problem:

-00
Inl
(2) E I,
<n <
< E- k
00
(Theory of Stability);
(Asymptotic Theory of the k-th
approximation).
The main result of the Theory of Stability is due to Kolmogorov (see

21). In our example, it reads as follows:
THEOREM 19.10
If E is small enough,
curves
t~an
the mapping TE has invariant analytical
that are close to the invariant curves
small enough, these curves

trices (lhp2 -
wg
cos q
:S
wg),
of T. Besides, for E
fill up the domain interior to the separaup to a set whose Lebesgue measure is
small with E . Roughly speaking, for E small enough, the invariant

curves
r I , whose angles
1I.(I) are "incommensurate enough" with
27T, do
not collapse but are only slightly deformed. The images TEnx of x
are all contained in
t ~
~.
To an invariant curve ~ of ~ corresponds a torus 2 ,of the space p,

q, t (q mod 27T, t mod r) that is invariant under (19.4). These tori divide
the space p, q, t (see Figure 19.11). Therefore a trajectory of (19.4) peginning in a gap between two invariant tori cannot pass out of this gap.
Thus, Theorem (19.10) allows us to reach conclusions regarding the stability of motion.
2 "That is a conditionally periodic motion of the swing.
86
Figure 19.11
Theorem (19.10) is a direct corollary of Theorein (21.11) whose proof

can be found in Appendix 34.
20. Fixed Points and Periodic Motions

Let us study the fixed points of the mapping
and its iterates, to
understand better the structure of the gaps between two invariant curves
r;.
These points correspond to the periodic motions of the swing.
THE ELLIPTIC AND HYPERBOLIC POINTS 20.1
In the neighborhood of a fixed point, a mapping reduces to its linear

part, that is, its differential, up to terms of ~econd order. The differential
of a canonical mapping is a linear canonical mapping. Linear canonical
mappings are studied in Appendix 27. If this linear mapping is hyperbolic
(resp. hyperbolic with reflection, resp. elliptic), then the fixed point is
called hyperbolic (resp. hyperbolic with reflection, resp. elliptic).
Hyperbolic fixed points are easily proved to be unstable, not only for
the linear mapping, but also for the nonlinear mapping (Hadamard).
The
problem of stability of elliptic points is known as the Birkhoff problem.

Points of elliptic type are, in general, stable in the two-dimensional case
(see Appendix 28).
STABLE SYSTEMS
STABILITY OF THE LOWEST POSITION OF EQUILIBRIUM 20.2
Now, consider the fixed point p
If E
0, then the mapping Te
=T
= q =
0 of the mapping ~ of 19.
is elliptic at the origin; in fact, the
differential of T is an elliptic rotation with angle A = CUT

Thus, for E small enough and A f, krr (k
= 21T(CU/V).
1,2, ... ), the mapping Te' is
elliptic" too. In other words, stability can change only for:

(20.3)
2cu
2cu 2cu
= -1, 2' 3 ' ....
A sharper computation proves that, for these values of v (called values of

"parametric resonance"; see Appendix 29), the mapping Te is hyperbolic
at p
= q =
0 (see Figure 20.4). In other words, the equilibrium of the
swing alters (and the swing starts oscillating) if one deflects during an integral number of half-periods of the proper oscillations. This result!s wellknown empirically. 3
values of parametric resonance
2w
20)
,Figure 20.4
3 Besides, observe that the amplitude of the oscillations of system (19.4) remains
small if e is small enough. For, according to Theorem (19.10), the orbit remains
between two curves ~. This follows from the nonlinearity of our system: the frequency A(1) depends on the amphtude; condition (20.3) is no longer fulfilled since
the oscillations do not have tIme enough to amphfy.
88
ERGuDIC PROBLEMS OF CLASSICAL MECHANICS
Te: EXISTENCE 20.5
THE FIXED POINTS OF THE ITERA TES OF
Now, consider the fixed points of 1~n. Let
be the invariant curve of
T formed with the fixed points of T" (see 19). We set:

A([)
dA /
dT r O.
217 E!.. ,
n
On an n-fold iteration of T each point of
returns to its original position.
This property of T is not retained for a small perturbation (T ... Te ). But

Poincare proved, for E small enough, that Ten has 2kn fixed points close
to the curve
r.
In fact, let us consider two curves that are invariant under
T and close to r: namely, the curves r+ and r-, with angles of rota-
tion A+ > A> A- (Figure 20.6). Thus, the mapping T" rotates r+ posi-
r-
tivelyand
negatively.
Figure 20.6
This property still holds for Ten if E is small enough. Thus, on each
radius
constant, there exists a point d, E) which moves under Ten
along the radius
Moreover, if E is small enough, the points r(, E) (O:S

clos~d
1'" n
analytical curve Re close to
r.
:s 277)
form a
Now, remember that the mapping
is canonical and so is area-preserving.
Consequently, the image
89
STABLE SYSTEMS
TenRe cannot be surrounded by Re and, conversely, Re cannot be surrounded by TenRe' Thus, Re and ~nRe intersect (Figure 20.7).
The
points of intersection are fixed points of Ten, for Ten moves each point of
Re along its /adius

~n
Ct. This proves the existence of fixed points of
in the neighborliood of
r.
Figure 20.7
FIXED POINTS OF THE ITERATES OF
Te: CLASSIFICATION 20.8
What is the type of these fixed points? Are they elliptic or hyperbolic?
If = 0, then they are all parabolic with proper value '\12 = 1. Consequently, for small enough,. '\12 "" 1 and the hyperbolic case with reflection is impossible. On the other hand, consider the "radial displacement":
90
ERGODIC PROBLEMS'OF CLASSICAL MECHANICS
~n.
The function D.() vanishes at the fixed points of

case" these zeros have multiplicity one (~'
zeros at which
= d~/d
In the "generic
f, 0). Hence, the
'> 0 separate the others, and the number of fixed points,
is even.
Let us make correspond to any point x the vector whose extremity is
Te n x (see Figure 20.7). It is readily seen that the index (Appendix 27) of
this vector field at a fixed point is:
Ind
sign of ( dA d~ \
dl
d)
Thus, half of the fixed points have index + 1 and the others have index -1This means that half of these points are elliptic and half of hyperbolic
type (an elliptie point has index + 1, an hyperbolic point has index -1).
Elliptic and hyperbolic points are illustrated in Figure 20.7.
Now, consider an elliptic~ fixed point: Ten x
x. The orbit of x is
x, Te x, ... , Ten-lx, therE!'fore:
All points of the orbit of x are fixed points of Ten and are elliptic since
they l1ave the same proper value. Hence, the set of the elliptic fixed points
splits into orbits consisting of
bits, then there are
.0
points. Let 1c be the number of such or-
elliptic points. This gives us 2kn fixed points
(elliptic and hyperbolic), as promised in 20.5.
ZONES OF INSTABILITY 20.9
We now turn to the neighborhoods of the above elliptic and hyperbolic

fixed points. According to V. Arnold [7] (see also Appendix 28), each
"generic" elliptic point is surrounded by closed curves that are invariant
under Ten. These curves form" islands" (see Figure 20.10). Each island
repeats in miniature the whole structure, with its curves C'. islands be-
91
STABLE SYSTEMS
tween these curves, and so on. Between these islands and curves
, re-
main zones around the hyperbolic points. In fact,4 the separatrices of hyperbolic fixed points of the Ten,s intersecting each other create an intricate network, as depicted in Figure 20.10. On discovering this, Poincart:
wrote ([2], V.3, Chap. 33, p. 389):
"One is struck by the complexity of thIs figure that I am not even
attempting to draw. N othmg can give us a better idea of the complexity of the three-body problem and of all problems of dynamICS
where there is no holomorphlc integral and Bohlin's series diverge."
Figure 20.10
The ergodic properties of the motion in zones of instability are unknown.

There probably exist systems with singular spectrum and K-systems among
the ergodic components.
4 See Poincare [21. V. Melnikov [1].
92
REMARK
20.11
The above argument does not prove there exist infinitely many elliptic
islands for a given 1. Poincare's last geometrical theoremS proves
that there exist infinitely many fixed points of ~n (n
->
+ 00) with index
0.1
o
-0.1
-0.4 -0.3 -0.2-0.1
0 0.1
0.2 0.3 0.4 0.5 0.6 Y
Figure 20.12
Yr-~-.-r-.--~~~-r-.--r-~-r~-.r-~
.5
.4
"-,3
..
"
...: "
-.5-.4-.3-.2-.1
,/
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 Y
Figure 20.13.
5 See Poincare [3], G. D.'Birkhoff [1].
93
STABLE SYSTEMS
+ 1 inside the annulus located between the invariant curves
(Theorem
19.10). Perhaps some of these points are not elliptic but hyperbolic with
reflection. Numerical computations 6 seem to support this conclusion.
y
0.8
0.6
0.2
o
-0.2
-0.04
-0.6
-0.8
-0.8 -0.6 -0.04-0.2
0.2 0.04 Q6 0.8 X
Flgure 20.14
We have borrowed Figures 20.12-14 from the work of M. Henon and
C. Heiles [1]. They depict the orbits of a mapping of type ~ computed

with an electronic computer. All the points, exterior to the curves, belong
to one and the same orbit!
21. Invariant Tori and Quasi-Periodic Motions
The example we considered in 19 and 20 is a particular case of a situation which occurs for each system close enough to an "integrable" system.
6
See Gelfand, Graev. Sueva. Mlchai1ova. Morosov

M. Henon. C. Helles [1].
b);
[3];
Ochozimskl. Sarychev ...
94
(A) INTEGRABLE SYSYEMS 21.1
If one takes a look at the "integrable" problems of Classical Mechanics,7 one finds that, for all of these problems, bounded orbits are either periodic or quasi-periodic. In other words, the phase-space is stratified
i~to
invariant tori supporting quasi-periodic motions.
E'XAMPLE 21.2
Assume the phase-space Q is the product of a bounded domain Bn of
Rn by the torus 1"'. Let P
(P l '
, .. , Pn )
be coordinates on Bn and q
(ql' "" qn) (mod 277) coordinates on Tn. The Hamiltonian equations, with
Hamiltonian function H = Ho(p), read:

(21.3)
The motion is quasi-periodic on the invariant tori P = Ct, with frequencies w (p). Frequenies depend on the torus: if
f 0,
"
then, on each neighborhood o(the torus p = Ct , there are invariant tpri

un which frequencies are independent and orbits everywhere dense (see
Appendix 1). There exist other tori on which frequencies are commensurate; they are exceptional, that is they form a set of measure zero. Coordinates (p, q) of B n x 1'" are called "action-angle" coordinates.
For all integrable systems, it can be shown (Appendix 26) that a certain (2n - 1)-dimensional hypersurface divides the phase-space into invariant domains each of which is stratified into invariant n-dimensional manifolds. If the domain is bounded, these manifolds are tori supporting quasiperiodic motions. The action-angle coordinates can be introduced into such
a domain, thus, the system can be described by (21.3).
7 For instance, the motion of a free point along a geodesic on the surface of a triaxial ellipsoid or a torus (see (1.7) and Appendix 2), a heavy solid body (Euler,
Lagrange, and Kovalewskaia cases),
95
STABLE SYSTEMS
(B) SYSTEMS CLOSE TO INTEGRABLE SYSTEMS 21.4

Now, we assume that the Hamiltonian function is perturbed:
the "perturbation" HI being" small enough." The Hamiltonian equations

are then:
(21.5)
oH I
oq
---,
For most initial data, A. N. Kolmogorov [6] proved that the motion remains
quasi-periodic (see Theorem 21. 7). Consequently, (21.5) is not ergodic on
the "energy surface" H = Ct and, among the ergodic components, there
are components with discrete spectrum, the complement of which has small
Lebesgue measure as H I is small.
Assume that the function H (p) is anal ytic in a c0cnplex domain [0] of
the phase-space:
O(~p,
I~ pi
'R q, (0,
< p,
I~ ql
< p) .
Assume also that the unperturbed system is nondegenerate:

Det
(21.6)
(U
I~
0
lo2H
Det __0_
op
op2
Select an incommensurate 8 frequency-vector w
oJ- O.
w*. The equations of
the invariant torus To(w*) of the unperturbed system (21.3) are p = p*,
where wo(P*) = w*. Thus, the system (21.3) has frequencies w* on
To(w*).
THEOR.EM
21.7
If HI is small enough, then for almost 9 all w*, there exists an invari-
ant torus T(w*) of the perturbed system (21.5) and T(w*) is close to
To (w*). To be precise:
That is, (w, k) ,;, 0 for all integers k.
All, except for a set of Lebesgue measure zero.
96
For all K> 0 there exist E> 0 and a mapping p = p(Q), q = q(Q) of
an abstract torus T =
I Q (mod
21T)} into
n such that,
according to (21.5)
Q = w*,
and:
Ip(Q)-p*1 <K,
Iq(Q)-QI <K,
if:
holds in [n].
Moreover, the tori T (w*) form a set of positive measure whose complement set has a measure that tends to zero as
IH 11
->
O. Proof of Theorem
(21. 7) is found in Arnold [5].

The behavior of the orbits emanating from this complement is not well
known. If our system has two degrees of freedom (n = 2) then the phasespace
has dimension four, and the two-dimensional invariant tori that
are found divide the three-dimensional manifold H = constant.
The do-
mains of the complement set are toric annuli betwee~ these invariant tori
(see Figure 19.11).
For n > 2, the n-dimensional invariant tori do not di-
vide the (2n -1) -dimensional manifold H
constant and those orbits that
do not belong to the tori T (w*) can travel very far along H
h (see 23).
(C) ApPLICATIONS AND GENERALIZATIONS 21.8

Theorem (21. 7) applies to the motion of a free point along a geodesic
on a convex surface close to an ellipsoid or '8 surface of revolution. This
theorem allows one to prove the stability in the plane restricted circular
three-body problem. 10
One can also deduce the stability of the fast rota-
tions of a heavy asymmetric solid body. 11

But this theorem does not apply if the unperturbed motion has fewer frequencies than the perturbed motion (degenerate case) for, in this case,
10 A. N. Kolmogorov [7].
11 V. I. Arnold [5].
97
STABLE SYSTEMS
condition (21.6) does not hold:

Det
- O.
The cases of "limiting degeneracy" of the oscillation theory (points of

equilibrium, periodic motions) also require a particular study. In that direction we mention some results that generalize Theorem (21.7).
V. I. Arnold [7] proved the stability of positions of equilibrium and of

periodic motions of systems with two degrees of freedom in the general elliptic case. As a corollary, A. M. Leontovich [1] deduced the stability of
the Lagrange periodic solutions for the reduced problem of the three-body
(plane and circular).
V. Arnold [8], [9], [10], studied the generation ~f new frequencies from
the perturbation of degenerate systems. As a corollary, one obtains the
perpetual adiabatic in variance of the action for a slow periodic variation of

the parameters of a nonlinear oscillatory system with one degree of freedom, and also that a "magnetic trap" with an axial-symmetric magnetic
field can perpetually retain charged particles.

Finally, quasi-periodic motions in the n-body problem have been found.
If the masses of n planets..are small enough in comparison with the mass
of the central body, the motion is quasi-periodic for the majority of initial
conditions for which the eccentricities and inclinations of the Kepler ellipses are small. Further, the major semiaxes perpetually remain close to
their original values, and the eccentricities and inclinations remain small
(see V. Arnold [4]).
On the other hand,
J. Moser [1], [5] generalized Theorem (21.7). Moser
abandons the requirement of analyticity of the Hamiltonian and substitutes
instead the requirement that several hundred derivatives exist.
For in-
stance, for systems with two degrees of freedom, it is sufficient (hat 333
derivatives exist!
(D) INVARIANT TORI OF CANONICAL MAPPINGS 21.9
Theorem (21.7) can be reformulated by using the construction of the
98
"surfaces of section" of Poincare-Birkhoff. Assume that, in Equations

(21.3), the first component w l of w is nonvanishing.
Consider a sub-
manifold ~2n-2 of the phase-space n2n whose equations are: ql
0,
H = h = constant. The orbit x (t) of (21.5) through a point x on ~2n-2

will, as t increases from zero, return to ~2n-2 and will cut ~2n-2 in a
uniquely determined point Ax (Figure 19.11). If tile perturbation H 1 is
small enough and w 1 (p)
0, the mapping A: ~2n-2
f.
-->
~2n-2 is well
defined in a neighborhood of the (n -1) -dimensional torus: p = Ct,
ql = O.
Since
then P2' ... , Pn; q2' ... , qn (mod 211) are "action-angle" coordinates in this
neighborhood. The mapping A is canonical (see Appendix 31).
Now, consider the unperturbed system (H 1
0). According to (21.3),
the map A may be written as follows:

(k = 2, ... , n).
(21.10)
In other words, each torus p
Ct is invariant and rotates through w(p)
under the mapping A.
If the perturbation H 1 is small, then the corresponding canonical map1,2n-2
-->
~2n-2 is close to (21. l-O). The (n -1) -dimensional
invariant tori of
d'
are, obviously, similar to the n-dimensional invariant
ping
d':
tori of (21.5) and there is a theorem, similar to Theorem (21.7), for mappings (see Theorem 21.11).
Let
again be the phase-space p, q:
Assume that B: p, q
-->
p '(p, q), q '(p, q) is a global canonical mapping,
that is:
Pdq=~Pdq,
Y
for any closed curve y of
By
(see Appendix 33). Then, assume that the
99
STABLE SYSTEMS
functions p '(p, q), q'(p, q) - q are analytic in the complex neighborhood
[a] of a:
Let A: p, q
-+
p, q + w (p) be the canonical mapping defined by an analyt-
ic function w(p) in [a], and To(w*) the torus p = p*, w(p*) = w* that
is invariant under A..
THEOREM
21.11
If B is close enough to the identity, then, for almost
12
all w *, there
exists a torus T (w*) that is invariant under BA and close to To(w*).

To be precise, to any K> 0 corresponds an S> 0 and a mapping 0:
-+
A, p = p(Q), q = q(Q), of an abstract torus T = !Q (mod 27T)1 into
A, such that:
O(Q+w*)
8A0(Q),
01-1
Q - Q+w*,
and:
Ip(Q)
-p*1 < K,
Iq(Q)-QI<K,
provided that:
holds in [01. Moreover the tori T (w*) form a set of positive measure
whose complement set has measure tending to zero as
Ip' - pi
+ Iq' -
ql
-+
O.
Theorem (19.10) is a direct corollary of Theorem (21.11): n = 1. Theorem (21.11) has been known since 1954, though its proof was never published.
J.
Moser [1] gave a proof for mappings of the plane (n = 1). This
proof makes use of the topology H2. Appendix 34 gives a proof for arbitrary n: the topological part is reducl'd to the technique of generating
functions of global canonical transformations (see Appendix 33).
12 All, except for a set of Lebesgue measure zero.
100
(E) COMPARISON OF THEOREM (21.7) WITH THEOREM (21.11) 21.12

Whether each analytic canonical transformation close to A can be constructed by using a section of a suitable Hamiltonian system is unknown.
Thus, Theorem (21.11) cannot be deduced from Theorem (21. 7).
Even if one restricts to the Hamiltonian system (21.5), Theorems ('U.7)
and (21.11) are nonequivalent. In fact, according to (21.6) and (21.10), the
nondegeneracy conditions of Theorems (21.7) and (21.11):
Det
I ~; I
Det
ji 0 ,
I aw I
ji 0
ap
may be written in terms of the unperturbed Hamiltonian function H 0 as follows:
(21.13)
Det
a2H 0
ap2
ji 0,
Det
a2H0
aHo
ap2
ap
aHo
ap
.;, 0
These conditions (21.13) are clearly independent. Each of them is sufficient to emmre that there exist invariant tori. The second one ensures,
additionally, that there exists such tori on each "energy surface"; this
implies the stability (see Figure 19.11) for systems with two degrees of
freedom (n = 2 for Theorem 21.7, n = 1 for Theorem 21.11).
In most applications, both of the conditions (21.13) are simultaneously
fulfilled or invalid.
22. Perturbation Theory

We turn next to the asymptotic theory, that is we restrict our study to
the behavior of orbits for 0
< t < liE,
E being the magnitude of the per-
turbation. In contrast, non-Hamiltonian systems can be considered.
101
STABLE SYSTEMS
(A) AVERAGING METHOD 13 22.1
Let Tk = ! = (1' ... , ) (mod 277) 1 be the k-dimensional t~rus and

B1 =
II
)1
= ([1' .. ,1 1
a bounded domain of Rl. In the phase-space
Tk x B1, consider the unperturbed system:
(22.2)
It is, obviously, a generalization of system (21.3): each torus I = con-
stant is invariant and, if the frequencies ware incommensurate on this

torus T, then the orbits (t) are everywhere dense on T. In such a case,
the motion (22.2) is called quasi-periodic on the torus T. If the frequencies are commensurate, then the closure of an orbit is a k-dimensional
torus (k < n) (resonance).
Consider next the perturbed system that generalizes (21.5):
(22.3)
{~
I
w([)+ Et(I,)
EF(I, )
where {
f(I, + 277) := f([, )

F([,
1> + 277)
Of course, for t '" 1, the eVQluti,on \Ht) -1(0)\ '" E
:=
F([, ),
E 1.
1. Notable ef-
fects, of order 1, of the evolution appear only after a long enough time:
t '" 1/ E .
Perturbation Theory proceeds to study the perturbed system as follows.
Let
fi(n
be the mean:
One considers the "averaged system," or "system of evolution":
j = Efi(]).
(22.4)
For E
1, one expects that:
(22.5)
IHi) -j(t)1 1 for 0 < t < 1
where [(t), (t) is a solution of (22.3) and j (t) is the solution of

(22.4) with initial data: j (0) =
Ho).
13 Thls method goes back to Lagrange, Laplace, and Gauss, who used It in Celestia I Mechanics.
102
Now the problem arises as follows: what relations exist, for 0

between the perturbed motion [(t) and the "motion of evolution"
< t < 1.,
I (t)
Does the inequality (22.5) 'hold?

For the simplest periodic motions (k = 1) it is readily proved -(see
Appendix 30 and Bogolubov and Mitropolski [1]) that if
< CE, for 0 <
I[(t) 71(t)1
But the situation
be~mes
(U
~ 0 then:
< 1.E
more complicated as the number k of frequen-
cies increases, even for k = 2.
(B) A COUNTER- EXAMPLE 22.6

Assume k = 1= 2, a> 1 and consider the system:
1 = 11 ,
2 = 12 ,
11 = E,
i2
= Eacos(1-2)
Of course, the system of evolution is:
11 = E,
12 = 0
(corresponding to small arrows on Figure 22.7). Consider the following

initial data:
11 =12 =1 1 "'1 2 = 1,
12 .J2
0,
1
2 = arcos a
1 (t)
-----
--....,.
~
~
~
--....,.
--....,.
---+
---+
~
(t)
-+
~
11 , J 1
Figure 22.7
STABLE SYSTEMS
103
Then:
1.
Thus,
I[(l/s) - ](l/s)1
In other words, after the interval of time
lis,
1.
the averaged motion loses
any relation with the real motion which remains locked in by the resonance
wI = w 2
(C) MATHEMATICAL FOUNDATIONS OF THE AVERAGING METHOD 22.8
There exist, at least, four distinct approaches to the problem of the

mathematical foundations of averaging method. All four lead to rather
modest results.
(1) The neighborhoods of particular solutions (for example, positions
of equilibrium
0) of the averaged system can be fairly well studied.
For instance, there exist attracting tori of (22.3) which correspond to the
< t < (0) obviously

holds in the neighborhood of such a torus. N. N. Bogolubov [2], J. Moser
attracting points of system (22.4). Stability (for 0
[2], [5], and 1._ Kupka [1] proved that attracting tori still exist for perturbed
systems.
This approach does not apply to Hamiltonian systems because attracting points do not exist according to the Liouville theorem (see 1.10).'
(2) One can study the relations between I (t) and] (t) for most (in
the sense of measure theory) initial data, neglecting points that correspond to resonance. For instance, Anosov [3] and Kasuga [1] proved
theorems of the following type:
Let R ( s, p) be the subset of
of the initial data such that
I[(t) -](t)1
for certain 0
>P
< t < lis. Then, lim measure R (S, p)

e
-+
0 for all p> O.
This approach allows one to obtain similar results for systems much
more general than (22.3); whence its weakness: estimates of the measure
104
of R (f;., p) are not realistic and one has no information concerning the motion in R(E,p).
(3) One can study passages through states of resonance,
(.4) One restricts oneself to Hamiltonian systems to obtain more information.
(D) PASSAGE THROUGH STATES OF RESONANCE 22.9
Let us begin with an example:

1 = II + 12,
1>2 = 12,
il
= E,
i2
= E cos(1 -2)
The averaged equations are (see Figure 22.10):
j2=0.
J 1 =E,
-- -- ---- -
-~
,~
----.
--
-----
Figure
--
2(O)
22.10
Consider the initial data that correspond to resonance wI

1(O)
I (t)
= 11 (0)
12 (0)-1
= w2:
O.
The system is easily integrated:

lI(t)-j(t}1
1/2(t}-11
fiEjTcosx2.dX,
o
T=..jE/2t.
l?
105
STABLE SYSTEMS
For t
liE, obviously,
I/( t) - J( t) I =
C YE.
Thus, the passage through the resonance w 1 = w 2 disperses the

bundle of orbits I(t), cp (t), which in the beginning differ only by phases
cp(O). The scattering of 12 after going through the resonance is of the order of
Vc (see
quencies (k
Figure 22.10). For a general system (22.3) with two fre-
2), one obtains 14 the following theorem:
If the quantity:
does not vanish in {l, then we have the estimate:

for all 0
(22.11)
< t < !.
E
Condition A ;, 0 means that the system cannot remain locked in at

any resonance: (22.3) implies
In example (22.6) condition A
changes sign at 11
A
t.
t.
0 is violated:
12 if a> 1. This example shows that condition
0 cannot be replaced by an analogous condition for the averaged sys-
tem.
The idea used in proving (22.11) is that the scattering produced by
each resonance is of the order C YEand that, among the infinitely many
resonances
w/ w2
min, only the greatest

(m, n
produces notable effects.

14 V. I. Arnold [12].
< In !. )
E
106
Passage through resonance!': for systems with more than two frequencies (k > 2) has not been studied.
(E) EVOLUTION OF HAMILTONIAN SYSTEMS 22.12
Next, apply the. averaging

. method to Hamiltonian systems (21.5). If
condition (21. 6) of nondegeneracy holds, then most of the unperturbed orbits are ergodic on tori p = constant. Thus, it is satisfactory to write
this system in the form (22.3), with I
= p, = q, k
= n:
aH 1
al
w(I) +' E - -
~, i
- E--
aH 1
a
where
w
The averaged system is ] = 0, for
o.
In other words, there is no evo}ution for nondegenerate Hamiltonian systerns: ] = constant.
Theorem (21.7) of conservation of quasi-periodic motions rigorously
establishes this conclusion. In fact, Theorem (21. 7) implies that:
IHt) -](t}1 <

(for all initial data if n
K for all t ( Rand
2, and
a2 H o
aP
E<
Eo(K)
107
STABLE SYSTEMS
and for most initial data in the general case). This illuminates the part
that conditions of conservation play in Theorem (21.7): they prevent evo1ution. ls In the same way, evolution is prohibited in Theorem (21.11), for
the mapping is globally canonical.
On the other hand, one also understands the part that the condition of
nondegeneracy plays. In fact, in case of degeneracy, the generic orbit of
the unperturbed system is ergodic on k-dimensiona1 tori (k strictly inferior
to n). In such a case, the algorithm of perturbation theory allows one to
predict the averaging on the k-dimensiona1 torus. Hence evolution becomes
possible, even for canonical systems.
IS This particular feature of canonical systems already comes out from simple examples. Let us consider the following perturbations of a center (Figure 22.13):
{t
-x - Ey
x.
Figure 22.13
The first perturbation, which is canonical, moves the orbits along a direction
which is orthogonal to the perturbation, and does not give rise to evolution. The
second perturbation, which is not canonical, gives rise to an evolution to zero.
Volume-preserving examples with evolution can b~ constructed in the four-dimens~onal
space x, y, z, u:
x = y,
-X-E Y,
.i
u,
ti
-Z+EU.
108
EXAMPLE
22.14
Consider the Hamiltonian system:
d 277)
J)
where
If this averaged system is either integrable (e.g. the plane three-body

problem) or close to an integrable system (e.g. the planetary many-body
problem) the!! there exist 16 quasi-periodic solutions corresponding to the
initial system. These quasi-periodic motions have k "fast" frequencies
(w 1' ... ,
W k) ,....,
1 that come from the unperturbed system, and I = n - k
"slow" frequencies (w k+ l' ... , Wn ) ,...., E that arise from the averaging system.
In the general case, when the averaged system is not integrable, the
relation between the solutions of the perturbed and the averaged systems
is still unknown even for 0 < t
< liE. The only known results arise from
approaches 2 and 3 (22.8). Moreover, observe that, even for nondegenerate systems, we need a study of the motion in the zones of instabilify
16 V. Arnold.[101. [4].
109
STABLE SYSTEMS
(complement set of the invariant tori) for n> 2, at least for t '" 1/E: (or
t '" liE: m).
In such a zone, one can probably find} 7 (n -1) -dimensional
invariant tori of "elliptic" or "hyperbolic" type that generalize, in arbitrary dimension, periodic motions of 20. If n > 2, recall that n~dimen
sional invariant tori do not divide the (2n -1) -dimensional energy level
H
constant. Consequently, the "separatrices" of the "hyperbolic"
tori can travel very far along H
constant, producing instability.
The
next section is devoted to the study of a similar mechanism of instability.
23. Topological Instability and Whiskered Tori

We give next an example 18 (see 23.10) of an Hamiltonian system that
satisfies conditions of Theorems (21. 7) and (21.11), but that is topologically unstable: I[(t) - ] (t) I is unbounded for
-00
<t <
00.
According to
Theorems (21.7) and (21.11), this system is stable for most initial data
(the corresponding motions are quasi-periodic). The secular changes of
I (t) have the velocity exp (-1/ {E.) and consequently cannot be dealt
with by any approximation of the classical theory of perturbations.

We first introduce some definitions.
(A)
THE WHISKERED. TORI
23.1
Assume that in the phase-space of the dynamical system there is an

invariant torus T and on it a quasi-periodic motion with everywhere dense
orbits.
We shall call T a whiskered torus if T is a connected component of
the intersection of two invariant open manifolds: T
= M+
n M-,
where
17 One can find motivation in V. Arnold [I4]. Since this was written, the proof
was given independently by V. K. Melnikov [2],
J.
Moser [5], and G. A. Krasinskii
[1].
18 Example (23.10) is rather artificial, but we believe that the mechanism of
"transition chains" which guarantees that instability in our example is also
applicable to the generic case (for instance, to the three-body problem).
110
all the orbits on arriving whisker M- approach T as t
departing whisker M+ all the orbits approach T as t

lim
t ......
Ix{t) - TI
-+
-+
+ 00, and on the
-00:
0 for x(O) ( M+
-00
lim Ix{t) - TI
+ 00
0 for x (0) ( M-.
t ......
For instance, the torus Tk: x=y=z=
x=
(23.2)
A x,
in the system:
z =
= - f.L y,
if> =
0,
(U
(A, f.L > 0, cp(mod 211) ( Tk, (U incommensurate) defined in the space R/+
x R/- x RIo x Tk has a (/+ + k)-dimensional whisker
(y = z = 0) and
a (/_ +'k)-dimensional whisker M- (x = z = 0).

(B) THE TRANSITION TORI 23.3
Let M be a smooth submanifold of space X. We shall say that the subset
C X obstructs the manifold M at the point x ( M if every manifold
n.
N that is transverse to M at x is intersected by
n which winds onto a limit cycle
spiral
For instance, 19 a
M obstructs M (see Figure 16.4,
Chapter 3). If the whisker~ torus T has the property that the images of
an arbitrary neighborhood V of an arbitrary point ~ of one of its arriving
whiskers M- obstruct the departing ~hisker M + at an arbitrary point TJ of
M+, then the torus will be saip to be a transition torus (see Figure 23.4).
LEMMA 23.5
The torus x
= y =
0 in system (23.2) is a trafsition torus.
Proof:
We set ~ = (0, Yo, 0, cpo),
TJ = (Xl' 0, 0, CPl).
mensurate, there exists a sequence ti' ti

from
CPo + (Uti
-+
The (U's being incom-
+00, such that the distance
to CPl tends to zero.
Consider the part V of V whose equation is y = yo By
19 Articles by Sitnikov
[1]
and A. Leontovich
[11 are
based on this fact.
U
00
V(t)
111
STABLE SYSTEMS
we denote the set of all points of all the orbits emanating from U. Then
contains the set of all the images gtlV, where gt is the transformations
I-------....,..T)
Figure 23.4
group' defined by (23.2). For t; large enough, these images gt, V intersect
the neighborhood of TJ (because ,\ > 0). The intersections have equations:
Y = Y,
J
Thus
e -Ilt; Yo ... 0
n contains the set of all the surfaces
g tl V that are parallel to M +
and converge to M+. These surfaces already obstruct M+ at TJ; this

proves Lemma (23.S).
(C) THE TRANSITION CHAINS 23.6
Assume that the dynamical system has transition tori T1, T2, ... , Ts'
These tori will be said to form a transition chain if the departing whisker
M7 of every preceding torus T; is transverse to the arriving whisker M;-::'l

of the following torus Ti+ 1
at some point of their intersection (see
112
Figure 23.7):
Mi nM;
;, 0, M; nM3- ;, 0, ... , M;_l nM; ;, 0.
M; M;_l
Figure 23.7
LEMMA 23.8
Let T1 , T2 ,
.. ,
Ts be a transition chain. Then, an arbitrary neighbor-
hood V of an arbitrary point ~
M;- is connected with an orbit (t) to
an arbitrary neighborhood V of an arbitrary point TJ
(0)
V,
(t)
M; :
V for a certain t.
Proof:
Consider the future U ~
torus, then
Uobstructs Mi
Ut > 0 V(t)
at ~1 ~
the open set U. Let ~; be a point of
of V. Since Tl is a transiti~n
Mi nM;.
Thus, M; intersects
M; nu, then there exists a neigh-
borhood V 1 of ~; that belongs to U. The future of V 1 belongs to U

and it is sufficient to perform the same argument s times to prove that U
obstructs
M;
at TJ.
(Q. E. D.)
(D) AN UNSTABLE SYSTEM 23.9
Let U ~ R2 x T3 be the five-dimensional space 20 11' 12 ; <P1' <P2' t .

(<p 1' <P2' t taken mod 21T). The Hamiltonian, depending on the parameters
E. /l will have the form:
20
It is easy to construct a conservative system with the Hamiltonian (23.10).
113
STABLE SYSTEMS
Yi/~ + Ii) + ErOS (<p~- ~[1 + ~],
sin <P2 + cos t
In other words, we consider the system of differential equations:

(23.10)
where J.L E
THEOREM
1.
23.11
< A < B. For every E> 0 there exists a J.Lo = J.Lo(A, B, E,)
> 0 such that for 0 < J.L < J.Lo the system (23.10) has a solution satisfying:
12 (0) < A, 12 (t) > B for a certair) t.
Assume 0
To prove Theorem (23.11) it is sufficient, in view of Lemma (23.8), to

find a transition chain T1, ... , Ts such that: 12
< A on T1 and 12 > B on
Ts'
LEMMA
23.12
Each manifold Tw defined by the equations 11 = <P1 = 12 - w = 0,

where w is irrational, is a two-dimensional whiskered torus of the system
(23.10).
In fact:
(1) Tw is clearly an invariant torus of (23.10);
(2) for J.L = 0, the three-dimensional whiskers have equations:
11
(3) for J.L
t-
=+
-
2"jEsin .p1 ,
2
0 and small enough, the whiskers still exist and can be found
by the Hadamard method (see Chapter 3, 15). The argument of Lemma

(23.5) proves that the tori Tw are also transition tori.
finally, making use of variation formulas of the whiskers for J.L small
enough, 21 the following lemma is proved:
21 See Poincare [2] and V. K. Mel;',kov
[tl.
114
LEMMA 23.13
Assume A < w < B.
Then the departing whisker M~ of the torus Tw
intersects with the arriving whiskers M -, of all tori
ficiently close (provided that
Iw - w'l < K,
w ,which are suf-
where K = K(s, 11, A, B)).
Proof of this lemma requires certain computations that will be found in

V. Arnold [13]. These computations show also that K ~ 11 exp (
-1/ Vs).
Lemmas (23.12) and (23.13) imply that the whiskered tori T , ... , T
Wi
(w. irrational,
I
Iw. 1
(J.).
1+
11
< K,
w1 < A, ws > B)
Ws
form a transition. chain.
Application of Lemma (23.8) to this chain implies Theorem (23.11).

Arnold, V., Small Denominators I, Izvestia Akad. Nauk., Math. Series 25
1 (1961) pp. 21-86. [Trans!. Am. Math. Soc. 46(1965) pp. 213-284.]
Small Denominators II, Usp. Math. Nauk. No.5 (1963) pp. 13 -40.
[Russian Math. Surveys no. 5 (1963) pp. 9-36.]
Small Denominators III, Usp. Math. Nauk. No. 6(1963) pp. 89-192.
[Russian Math. Surveys No. 6(1963) pp. 85-193.]
Birkhoff, G. D., Dynamical Systems, New York (1927).
Moser,
J., On Invariant Curves of Area-Preserving Mappings of an Annulus,
Gottingen Nachr. No.1 (1962).

Poincare, H., Les methodes nouvelles de la mecanique celeste, I, II, III,
Gauthier-Villars, Paris (1892, 1893, 1899).
Siegel, C. L., Vorlesungen tiber Himmelsmechanik, Springer, Berlin (1956).
APPENDIX]
THE JACOBI THEOREM

(See Example 1.2)
Let S 1 =
(mod 1),
W f
! x (mod
1)1 be the circle and be the translation: x ... x + w
R. Each orbit of is ever:ywhere dense if, and only if, w
is irrational.
Proof: Assume w rational:
Let w = pi q where p, q .f Z, g > O. Here q is the identity transformation. Since every point of our circle is left fixed by q, every orbit
is closed and consists of a finite set of points.
Assume w irrational:
Let x be an arbitrary point of 51. The n x's are distinct for
implies that (n -.m)w
Z and then n = m. Thus each orbit consists in an
infinite number of distinct points. Since S1 is compact, this orbit has a

limit point. Consequently, for any E > 0 there are distinct integers n
and m such that:
Setting
In - ml
p and observing that is length-preserving, we have:
Thus, Px, 2 p x, ... , kp x, ... partition S1 into segments of length less

than E. Our theorem is proved, for E is arbitrary. An N-dimensional extension of this theorem reads as follows:
115
116
Let Tn
x
->
Rn /Zn
be the n-dimensional torus and
x';' w (mod 1), w ( Rn. Each orbit of
 x+ tw

(mod 1), t ( R. w (R n . Each orbit of <P t is everywhere dense if, and
only if, w' k = 0 and k ( Zn imply k = O.
APPENDIX 2
GEODESIC FLOW OF THE TORUS

(See Example 1. 7)
V is the two-dimensional torus, that is, the surface of revolution of a
circle of radius r about a line Oz in the plane of the circle at the distance
1 (> r) from the center of the circle. Equations of V in geographical coordinates are:
x ~ (1 + r cos
t/J)
cos
if>
y ~ (1 + r cos
t/J)
sin
if>
z
where
t/J,
r sin
if> is the latitude and t/J is the longitude.
-",
\
\
\
\
Flgure A2.1
117
118
Figure A2.2
Conservation of energy and conservation of angular momentum about

Oz give equations of the geodesics:
r2
, 2 + (1 + r cos rjI)2 " 2 =
h = constant.
" . (1 + r cos rjI) 2 = k = constant.

If h = 1 we obtain the geodesic flow on M = Tl V. This flow is invariant
under rotations
-+
+ constant. Thus, in order to study the geodesics
it is sufficient to study those of them with initial point on a prescribed

meridian. Figures A2.1 and A2. 2 depict the generic cases Yl and Y3 that
are separated by
Y2"
APPENDIX 3
THE EULER-POINSOT MOTION

(See Example 1. 7)
It is the motion of a heavy rigid body fixed at its center of gravity.
The rigid body is a system with three degrees of freedom and a six-dimensional phase space. There exist four independent single-valued first integrals: the energy T and three components of the angular momentum vector m. These four functions of position in the phase space do not change
their values for a given motion. The points of the six-dimensional phase
space for which the four functions have given values form, in general, a
two-dimensional manifold M. These manifolds M(T, m) are tori. In fact,
M is invariant, and so the phase-velocity vector at each point of M is tangent to M; consequently M admits a vector field without singular point.
It is evident that M is orientable and compact. The only compact two-
dimensional orient able manifold admitting a tangential vector field without

Singular point is, as is well-known, the torus. On the other hand, M being
invariant under the dynamical flow t' admits an invariant measure /1
(Liouville theorem). Thus (M, /1, t) is a classical system.
A canonical transformation makes (M, /1, t) into a system of the form:
x
= 1, ;. = -a (Example 1.2; see Appendix 26). Thus the Euler-Poinsot
motion is, in general, quasi-periodic and the orbits are dense on M. Its
two periodic components are called, respectively, precession and rotation
(Euler [1]).
119
APPENDIX 4
GEODESIC FLOWS OF LIE GROUPS
(See Example 1. 7)
The geodesic flow of a Lie group carrying a left-invariant (or rightinvariant) metric has important applications:
The geodesic flow of the connected component SO(3} of the group of
the rotations of the three-dimensional Euclidean space corresponds to the
rotations of a heavy solid moving around a fixed point. Each orbit corresponds to a motion.
The homotheties of positive ratio and the translations of the n-dimensional space Rn form a group that generates the geodesic flow of the
(n+ l}-dimensional space of constant negative curvature.
Then, let us consider the group S Diff(~) of the measure-preserving

diffeomorphisms of a compact Riemannian domain~. T~e corresp.onding
algebra consists in divergence-free vector fields V or. ~: div V = O. The
energy
< V,
>
Ii) v2 dx
is a positive definite quadratic form on the
space of such vector fields, and defines some right-invariant Riemannian

metric on the group S Diff(~).
The geodesics of this metric are just. flows of perfect (incompressible
and unviscous) fluid in~. The Rief!1annian curvature of this infinite dimensional "manifold" S Diff(~} can be computed. For example, if ~
T2 is the torus lx, y mod 2171 with its usual f!1etric, then the sectional
curvature is nonpositive in every section containing the lamihar flow:

Yx
cosy,!... Vy
O.
See ~lso V, I. Arnold [11 and L. Auslander, L. Green, F. Hahn [1].
120
APPENDIX 5
THE PENDULUM
(See 1.13)
The equation of the pendulum is:
q+ k
sin q
0, where k is a posi-
tive constant; it is equivalent to the system:
q ~ p
p=
-k sin q.
The Hamiltonian function is H = (p2/2) -k cos q and Figure (AS. 1) depicts the orbits. The system is invariant under the symmetry p'" - p and
the translations:
(q, p) ... (q + 2K7T, p),
Z.
Figure AS.!
121
122
The points (kl7,O) are critical points: points (217k, 0) are centers (stable
equilibrium) and points ((2k + 1)17,0) are saddle points (unstablE! equilibrium). The orbits split into three types: the orbits 1 (small oscillations);
the separatrices 2 joining two saddle points; and the orbits 3 (complete
rotation around the hanging point).
The natural phase space is no longer the plane (p, q) but it is rather
the cylinder (q mod 17, pl. This is an example of global Hamiltonian flow.
APPENDIX 6
MEASURE SPACE
(See Chapter 1, Section 2)
An a-algebra
93
defined on M is a class of subsets of M that is closed
under the formation of complement and countable union. A nonnegative

(possibly infinite) countably additive set function /1 defined on
a measure (M,
93, 11)
is called a measure space;
93
'B
is called
is the family of the
measurable subsets of M and 11 is the measure. (See Halmos [2] and Rohlin
[3j for these concepts.)
In fact, the object of interest is not (M,
93, 11)
but equivalence classes
(M,I1) of measure spaces. Let us make this point clearer. Let A and B
be elements of
93,
we set A ~ B (mod 0) if I1(A U B - A
n B)
~ O. This
;.
relation is an equivalence relation and the equivalence class of the empty

set consists in the sets of measure zero. We denote th~ quotient of
der this equivalence relation by
93
93
un-
(mod 0); it is a Boolean algebra, for
the following properties are readily proved:

B2 (mod 0)
imply
M - A1
M - A2 (mod 0) .
If A ~ B (mod 0), then I1(A) ~ I1(B) and 11 can be regarded as a function

defined on
93
(mod 0).
To study abstract dynamical systems one neglects sets of measure

zero. This means that one replaces the study of (M,
123
93, 11)
by the study of
124
(M, ~ (mod 0), p.), which we still denote unambiguously by (M, p.), because
the function p. defines completely ~ (mod 0) (but not ~). In other words,
we identify (M,~, p.) and (M, ~ " p.') if their measurable a-algebras ~
(mod 0) and ~' (mod 0) coincide.
Let (M, p.) and (M', p.') be two measure spaces. A mapping : M
-+
M'
is called an homomorphism modulo zero if:

(a) is defined on M -I, where / is a set of measure zero (possiblyempty);
(b) /'
M' - (M -I) has measure zero, that is is onto, up to a
subset I' of M' of measure zero;

(c) is measure-preserving, that is each equivalence class of ~'
(mod 0) contains a representative, say A', such that -l(A') exists and
belongs to ~ with:
Thus, induces an homomorphism of Boolean measurable algebras: -l:
:B ' (mod
0)
-+
:B
(mod 0).
If (M, p.) = (M', p.') then is called an endomorphism (mod 0). If both
of the mappings and -l are homomorphisms (mod 0), then is called
an isomorphism (mod 0); if, additionally, (M, p.) and (M', p.') coincide,
then is called an automorphism (mod 0).
APPENDIX 7
ISOMORPHISM OF THE BAKER'S TRANSFORMATION

AND B(Y:z, Y:z)
(See Example 4.5)
We need to construct an isomorphism (mod 0) 1 making the following
diagram commutative:
De/inition 01 I. Let m
f(m)
a_ 1 , aO' a 1 ,
...
be a point of Z2Z. We set
(x, y), where

00
(A7.1)
= ... ,
x=
i= 0
a -,.
00
y=
2i+1
~
i= 1
The mapping 1 is a bijection, except on the elements (x, y) of T2 for

which x or y is a dyadic fraction. Such elements are denumerable and so
constitute a set of measure zero.
1 is measure-preserving. It is sufficient to prove it for a generator A!,

1m I a i = iI of the measure ~lgebra of Z2Z : the set
f(Aj)
~( k;O 2:~ k~1 ;~ )1 a,

125
126
consists in 21il rectangles, the sides of which are 1 and 1/2I i l+1. Thus,
we get:
The diagram is commutative. Let x and y be given by formula (A 7.1),

we have:
where
ai
al _ 1,
00
k=O
that is to say:
f (2x, 'hY)
if a o
0, i.e., 0 ~ x
(2x, Wy+ 1)) if a o = 1, i.e., 'h
Consequently: I 1- 1 = '.
< 'h
~ x < 1.
(Q. E.. D)
APPENDIX 8
LACK OF COINCIDENCE EVERYWHERE OF SPACE

MEAN AND TIME MEAN
(See Remark 6.5)
Consider again the dynamical system of Example (1.16): M is the torus
{(x, y) mod 11. the measure is dx dy and the automorphism
is:
(x,y) = (x + y, x+ 2y) (mod 1).
This automorphism induces a linear mapping
M=
of
YH.
{(x,
of the covering plane
The matux
has two proper values: 0

X =
<
"2 < 1 < \.
The line
y = ("2 - 1)s,
s (
projects onto a curve y of M under the natural projection
rii
->
M.
This
curve y is invariant under and y is dense on M, for .:\2 - 1 is irrational

Qacobi's theorem, Appendix 1). Let m
(x, y) be a point of y. Of course,
we have:
and 0
<
"2 < 1 implies:

lim "(m) = (0, 0) .
n=oo
127
128
Consider the analytic function f(x, y) = e 2TTix We have

N-l
N-l
1N ~
f(rpnm) =
~ ~
n=O
e 2TTixA,f.
n=O
Usual convergence implies Cesaro convergence and lim x A;
0, there-
n=oo
fore:
N-l
lim
f(m)
N-++oo N
f(rpn m)
1 .
n= 0
On the other hand:
Thus, whatever the point m of the dense subset y be, we have: [*(m)
though f is analytic and
rp
is classical.
f T,
APPENDIX 9
THE THEOREM OF EQUIPARTITION MODULO I

(See 6.6)
We prove here 1 the Theorem of Equipartition Modulo 1 due to Bohl,
Sierpinskii, and Weyl: II is a rotation 01 the circle M through an angle
incommensurate with 21T:

M
I z ( C, \z\
I\, (z)
8 z, 8 =
e2lTiW ,
w is irrational
and I is a Riemannian integrable lunction, then the time mean 01 I exists

everywhere and coincides with the space mean.
Prool.
lsi ca~e: i(z) = zP, P (
N-I
~ ~
N-I
~ ~
I(n z ) ,=
n=O
z.
We get:
(8 n z)P =
1
1
N-
n=O
Since w is irrational'we have eP -1 ~ 0 and
2nd case:
'" I(z)
zP
eN P - l
8p -1
\8PN -1\ < 2,
if p =
if p ~ 0
I is a trigonometrical polynomial, that is

l(z) '" ~ a zP,
P
{I
if p = 0
p (Z,
Compare to G. Polya and G. Szego [I] p. 73.
129
Z (
M,
ifp~O.
so we get:
~RGODIC PROBLEMS OF CLASSICAL MECHANICS
130
in which
0, except for a finite number of them. From the first case
one deduces at once:
*
i(z)
BO
r.
3rd case: f is realovalued and Riemannian-integrable. To every

E > 0 correspond two trigonometrical polynomials ~- and ~+ such that:
Pe -(z)
< fez) < Pe + (z) for every z ( M
and
j(Pe+(z)-Pe-(Z))d ll
<
E.
From the second case we deduce:

(A 9.1)
N-l
~ lim sup 1
N ~
N~oo
f(n z)
~ + dll
< E. Since E is arbitrary we have

fez), which exists everywhere. Relation (A9.1)
Consequently, lim sup - lim inf

lim sup = lim inf = lim =
implies that
fez) is constant, whence:
*
i(z)
= f
(Q. E. D.)
Extension to translations of the torus Tn is obvious: the time mean and

the space mean of a Riemannil\n-integrable function coincide everywhere
if, and only if, the orbits are everywhere dense.
APPENDIX 10
SOME APPLICATIONS OF ERGODIC THEORY

TO DIFFERENTIAL GEOMETRY
The Birkhoff theorem was used by A. Avez [1] to prove the fOliowlng:
Let V be a compact n-dimensional Riemannian manifold without conjugate point, then the proper values of the operator
tl-~
n-1
are nonnegative (tl is the Laplacian _Va Va' R is the scalar curvature).
In particular (L. W. Green):
131
APPENDIX 11
ERGODIC TRANSLATIONS OF TORI

(See Example 7.8)
We prove that translations of tori (Examples 1.2 and 1.15) are ergodic
if and only if, their orbits are everywhere dense (or if. and only if, the time
mean and the space mean of a continuous function coincide everywhere).
Let M be the n-dimensional torus le 277ix l x ( Rn \, where x = (xl' ... ,
X )
and e 277ix means (e277ixl, ... , e277ixn). The measure of M is the usual
product measure /l. The translation is:
THEOREM.
(M, /l, ep) is ergodic it, and only il k 0
Z and k ( Zn imply k = O.
Prool:
Let ( be a measurable invariant function. Its Fourier coefficients are:
ak =
e- 277ik
x I(x)d/l
M-
The Fourier coefficients of I(cf>x) are:

bk =
e- 277ik (x -0), f(x)d/l = e 277ik ' 0 . a k .
The invariance of I is equivalent to b k = a k for any k, that is a k = 0

or k
Z.
132
133
APPENDIX 11
If the
w;'s
are integrally independent, the second case occurs only for
k = O. Thus a o is the only Fourier coefficient possibly nonzero, f is

constant, and (M, /L, ) is ergodic (see 7.2).
If a k f- 0 exists such that k w ( Z, then f (x)
e 21Tik x is a non-
constant invariant function and (M, /L, ) is not ergodic.

Remark.
In the continuous case (M, /L, t)' where
t:
e 21Tix
->
e 21Ti (X + tw) ,
we have a similar result: (M, /L, t) is ergodic, if, and only if, k ( Zn
and kw = 0 imply k = 0 (or if, and only if, the orbits are everywhere
dense; see Jacobi's theorem).
APPENDIX 12
THE TIME MEAN OF SOJOURN

(See Chapter 2, Section 7)
TH~OREM
A12.1
, An abstract dynamical systeJll (M, /l, /) is ergodic if, and only if, the
sojourn time reT) in an arbitrary measurable set A of an orbit
I /x I 0 $. t s; TI
is asymptotically proportional to the measure of A:
(A12.2)
dT)
T->oo
11m - -
It (A),
for all measurable A and almost every initial point x ( M.

Proof:
*
Assume (M, /l, /) is ergodic and A is measurable. We have f(x)
for every f ( Ll (M, It) and for almost every x ( M (see 7.1). Take f
T
XA
(characteristic function of the set A), we obtain:
for almost every x.

The converse is derived at once: (A12.2) implies ergodicity. It is sufficient to observe that the characteristic functions
XA
generate Ll (M, /l).
Theorem (A12 .1) clearly holds in the discrete case (M, Il. ).
134
135
APPENDIX 12
EXAMPLES AI2.3: TRANSLATIONS OF TORI

Let M be the n-dimensional torus
I e 27Tix I x
Rn I, Il the usual mea-
sure, and the translation:
If k ( Zn artd k (U
Z imply k
0, then (M, Il, ) is ergodic (see Appen-
dix 11). Thus, relation (AI2.2) holds for almost every initial point. This
can be rephrased as follows: denote by r(N, A) the number of elements of
the sequence
that belong to A, then:

r(N, A)
(A)
11m
- - - = Il
(AI2.4)
N~oo
for almost every initial point e 27Tix . If A is Jordan-measurable, that is,

if ~A is Riemannian-integrable, then (AI2.4) holds for every initial poiRt.
To prove it, it is sufficient to use the theorem of Appendix 9 and to take
f
= ~A'
Extension tothe continuous case holds good. This result i!l
known as the theorem of equipartition modulo 11 and is due to P. Boh1 [1],

W_Sierpinskii, and H. Wey1 [1], [2], [3] ~ It is one of the first ergodic theorems. Historically, it originated from an attempt to solve the Lagrange
problem of the mean motion of the perihelion (see Example 3.1 and Appendix 13).
Here follow some applications. 2
ApPLICATION AI2.5: DISTRIBUTION OF THE FIRST DIGITS OF 2"
(see Example 3.2)
The first digit of 2n is equal to k if, and only if:
k lOT $. 2n
<
(k +
P . lOT
1 F. P. Callahan [11 gave an elementary proof.

2 The reader will find further applications to various fields in: Compositio Mathe-
matica. V 16. fascicles 1. 2.
136
that is to say:
r + Log10k S n Log 10 2 ::; r + Log10(k+ 1) .
Set a = Log 10 2 and (n a) = na - [n a], where [ ] means the integer
part. The above inequality may be written:
Log10k S (na)
< Log10(k+
1)
Now, we turn to the dynamical system consisting of the one-dimensional

torus M =
e 21Tix
->
le 21Tix I x
e 21Ti (X+ a).
RI.
the usual measure /l, and the translation :
(M, /l, ) is ergodic, for a is irrational (see Exam-
ple 7.8). Thus, the sequence 1(na) I n ( N I is equidistributed. In particular, take A
[Log 1ok, Log10(k+ 1)] in relation (A12.4), we have:

lim
r(N, A)
-N-
N->oo
/l(A)
= Log 10(1-+; k-).
But r(N, A) is nothing but the number of elements of the sequence 1, 2, ... ,
2N - 1 , the first digit of which is k. Thus, if we go back to the notation of

ExamRle (3.2), we have:
Consequently, the proportion of 7's is greater than the proportion of 8's in

the sequel!,ce of thEt~fi,,~t digits of
12n In
1,2, ... 1. This is not what one
expects from an inspection of the first terms: 1,2,4,8, 1,3, 6, 1, 2,5, ...
This is due to the fact that a = 0, 30103 '"
REMARK
is very close to 3/10.
A12.6
Since the sojourn time in a domain A of a point of an ergodic system

is proportional to the measure of A, it is natural to ask about the dispersion. Let us mention some results due to Sinai [1]. Let t be the geodesic flow of the unitary tangent bundle Tl Y of a surface .y of constant
negative curvature. If A is a domain of Tl Y with piecewise differentiable
boundary, then the mean sojourn time of a geodesic tX in this domain has
a Gaussian distribution and verifies the central limit theorem:
137
APPENDIX 12
lim !l
T-+oo
where
~ x ITT(X)
T
TT(x) =
measure
- !l(A)
< ;;
yT
I t I tX
A, 0
$nl
217
:s. t :s. Tl
Ca e- u2/2 du
-00
and C is a constant.
APPENDIX 13
THE MEAN MOTION OF THE PERIHELION

(See Example 3.1 and Appendix 12)
The problem of mean motion arises from the theory of the secular perturbations of the planetary orbits (Lagrange [1]). One asks for the existence and estimate of:
(A13.1)
1 Arg
iwkt ,
~ ak . e
t-++oo
k= 1
= lim
where w k ' t ( Rand a k f, 0, ak ( C.

In other words, if one considers a plane linkage AoAl ... An consisting of the links A k _ 1 Ak of fixed lengths lakl, moving with constant rotation-velocity w k ' we are interested in the mean rotation-velocity of the
vector AO An (see Figure A13.1).
THEOREM A13.2 (See H. Weyl, [1]- [5].)
Assume that the wk's
(A13.3)
are
integrally independent, that is:
k = 0 and k ( Zn imply k = O.
Then the mean motion 0 exists and is expressed as:
(A13.4)
The Pk's depend on the lakl only. Ifl p(ak ; al' ... , ~k' ... , an) is the
probability that an (n-l) -linkage with prescribed sides al' ... , ~k' ... , an
spans a distance inferior to ak (see formula A13.12), then:
1
means cancellation.
138
139
APPENDIX 13
o =AO
Figure A13.1.
Case n '" 3: initial position of the linkage .
(A13.S)
In particular, for n '" 3, if there exists a triangle the sides of which
&re
lall, la2 1, la3 1 and the
angles of which are Al , A 2 , A 3 , then (Bohl's
formula):
(A13.6)
n '"
_A_l_w..;.1_+_A...;:2:....w.....;2'--.+_A_3:....w_3;:....
17
The case in which no triangle can be constructed was investigated by

Lagrange [1]. In the general case A. Wintner [1] found the expression:
in terms of the Bessel functions ]0 and ]1 Relation ~ Pk

an "addition" theorem for these functions.
1 provides
140
Proof of Theorem A13.2

THE CORRESPONDING DYNAMICAL SYSTEM
A13.7
Let us consider the dynamical system (M, p., >t)' where

M
Tn
= Izlz = (zl, ... ,zn)\'
is the n-dimensional torus, p. is the usual measure, and >t is the translation group:
The phase space of the n-linkage is M, and >t depicts the movement. Let
us define a function a on M by:
n
(A13.B)
a(z) = Arg
lakl
Zk'
O:s a < .211
k=l
This function is discontinuous over the slit l:
= I z I a(z) = 01,
and is not
defined on the so-called singular manifold S = Izll:~lak~zk = 01 which

consists of all possible states of a closed n-linkage with the prescribed
sides lakl. Nevertheless, the function:
(A13.9)
is analytic outside of S. The limit (k13.1), if it exists, is nothing but the
time mean ~ of {3:
(A13.10)
where
THE SPACE MEAN
A13.11
The system (M, p., >t) is ergodic, for the
(Uk's
are integrally indepen-
dent (Appendix 11). If the function {3 were Riemannian-integrable, then,
141
APPENDIX 13
according to the theorem of Equipartition modulo 1 (Appendix 9), the time

mean
f1*
n ~ i3
the Birkhoff theorem implies that
f1
is Lebes-gue-integrable. Thus,
for almost every initial phase.
This suggests the study of the space mean

shows that
f1
f1.
would be equal to the space mean
We only know (see A. Wintner [1]) that
i3.
Relation (A13.9)
depends linearly on the (U/s. Therefore,
f1
depends lin-
early on the (Uk's:
To compute Pl (for instance) we set:

(Ul
217,
(U2
(Un
We have:
1 -
217
f1 (217,0, ... ,0)
where
Relation (AI3.8) allows one to carry out the integration over (Jl:
if Ila2Ie217i(J2 + + Ian Ie 217i (Jn I < lall

if Ila:ile217i(J2 + + Iani e 217i (Jn I > lall
Thus we obtain:
where
This proves relation (AI3.S).

Relation ~ Pk ~ 1 is derived easily by setting
(Ul
(Un
2"
142
A13.13
EXISTENCE OF THE TIME MEAN
Hence, formulas (A13.4) and (A13.s) are proved for almost every initial
phase Arg a k . To prove them for all initial phases we use a special device inaugurated by Bohl [1] for n
n
3 and improved by Weyl [4], [5] for
> 3. We define a function on the torus M by:

N (z)
algebraic number of the points of intersection of the curve
I/z, -271 < t < 01 with the slit ~.

We count + 1 a point of intersection z. for which (3 (z.) > 0 and -1 if
J
(3(z.)
J
< O. (See Figure AI3.1s.) It can be proved that
N(z) is bounded.
N=O
Fii:Ure
~13.1S
Thus, according to (AI3.10), the following relation holds uniformly over
m:
(~13.14)
Since the function N is piecewise

nian-integrable, the time mean
N*
continuo~s
and, in particular, Rieman-
exists everywhere and is equal to N
(Appendix 9). From (AI3.14) one deduces that {3*

where and is constant.
*=
N exists every-
(Q. E. D.)
APPENDIX 14
EXAMPLE OF A MIXING ENDOMORPHISM

Let us consider the transformation: 1
cf>: (x, y) ... (2x, 2y) (mod 1)

of the torus M
= l(x, y) mod 11
carrying the usual measure dx dy.

cp-' A
I'
~ ~
@ ~
Figure A14.l
~--------------------~
Which is called "multiplication of loaves" since Figure (A14.1) shows the solution of a well-known historical problem.
143
144
To be more explicit, we write:

( (2x, 2y)
(x, y)
~ (2x,2y-l)
(2x-1,2y)
if 0 $ x,
y<\6
if 0 $ x < \6, \6$ y < 1

if \6 $ x < 1,
O$y<Y2
, (2x-1,2y-l) if \6 $ x, y < 1.
The application is not one-to-one, in fact it is everywhere four-to-one.
If E is a square with dyadically rational vertices, then -1 E istheunion

of four similar squares (see Figure). Thus, in such a case, p.(-1E) =
p. (E) and from there it follows easily that
is measure-preserving. We
have here an example of a measvre-preserving mapping that is not invertible.

The transformation is mixing, that is:
(A14.2)
for any measurable sets A and B. To prove it, it is sufficient to consider
the souares B:
I, m ( Z+ .
If N::: p, then B contains 4 N -
an inverse image has measure
4- N p.(A),
inverse images of A under -No Such

therefore:
Thus (A14.2) holds good because p is arbitrary.

Similar arguments show that the mappings:
k: (x, y)
->
(kx, ky)(mod 1),
k ( Z+
are measure-preserving and mixing. They satisfy:

kor = kr for k,r (Z+, i.e., !k1k (Z+! .
is a mixing s~igroup under composition. 2
2 This semigroup can be interpreted in terms of Tchebyschev polynomials

(R. Adler and T. Rivlin [1])
APPENDIX I))
SKEW-PRODUCTS
(See Definition 9.5)
Let (M, p.) and (M', p.')
b~.two
Lebesgue spaces. We denote by (M x M',
p.x p.') their direct product: the measure algebra
1M , A' ( ~;
1M , A' ( 1M "
erated by the AxA', A (

p.(A)p.(A') for any A
I Mx M'
of Mx M' is gen-
and we set (p.xp.')(AxA')
Assume that (M, p., ) is a dynamical system, and make correspond to

every m
M an automorphism .pm: M' .. M' such that (m, m') ... .pm(m')
!s measurable for every m (M, m' ( M'. Then
( x l.pl): M x M' ... M x M',

defined by
is measurable and measure-preserving.

In fact, for every measurable set F
IMxM '
whose characteristic
function is XF' we have:

(p.xp.') x I.pD-IF) =
X F [( x I.pD(m, m')]d(p.xp.')
MxM'
[,x
F(m,
rp mm ')dp.
Since .pm is p. '-measure preserving, this may be written

(p.x p. ') x l.pl)-I F)
[
145
{,X F(m, m ')dp.
dp.
dp.
146
Apply Fubini's theorem and observe that is measure-preserving:

(Il X Il ')x !t/JO-1 F) =
tX~ XF(m, m ')dll] dll'

~{LXF(m,m')djdll '
(Il X Il')F
The dynamical system (MxM', IlXIl', x!t/JI) ~s called the skew-product

of (M, Il' ) and (M', Il" IjJ m).
A 15.1
EXAMPLE
If t/J m
' is constant, that is, if x !t/J 1 is defined by

( x ')(m, m ') = (m, 'm ') ,
then the skew-product is the direct product of (M, Il' ) and (M', Il', ').
EXAMPLE
A 15.2
Tflke M
51
! x (mod 1) 1 with the usual measure dx and : 51
->
51
an ergod'ic translation:
x = x + w (mod 1),
w irrational. Now, let n be an integer.
with the usual measure dy and we make correspond to each x ( M

translation t/J x,n : M'
->
1)1
We take M' = 51 = !y (mod

=
51 a
M' defined by:
t/J x,n (y) = y +
(lX
(mod 1)
Thus, to any integer n corresponds the skew-product (51 x 6 1, dxx dy,

x!t/J x
nO.
Anzai [1] proved that the ergodic measure-preserving automorphisms

x!t/J K,n 1 and x!t/J K,p l, with w irrational and
Inl p Ipl,
are not isomor-
phic although they have the same spectral type and the same vanishing entropy.
APPENDIX 16
DISCRETE SPECTRUM OF CLASSICAL SYSTEMS

(See 9.13)
Let (M, 11, t) be a classical flow and Ut the one-parameter group of
unitary transformations induced by t
The discrete component of the spec-
trum of Ut is called the discrete spectrrim.

The dynamical systems constructed in the second part of the discrete
spectrum theorem (see 9.13) are classical systems if the rank 1 of the abelian group of the proper values of Ut is finite.
All known examples of ergodic classical systems have discrete spectrum with finite rank and this rank is bounded from above by the uimension
of the space. 2 Thus, the finiteness of this rank is a natural conjecture.
The proper functions of a classical system can be everywhere discontinuous (see an example due to Kolmogorov [1]), but if all of them are continuous, then the rank of the discrete spectrum is bounded from above by the
first Betti number b l
THEOREM
dim H 1 (M, R).
A 16.1
u,t
If the proper functions of the induced unitary group
of an ergodic
classical system (M, 11, t) are continuous, then:

rank of the discrete spectrum
S. b l
This theorem is an obvious corollary of the following one:
I. e., the maximum number of independent generators.
2 This fact is general for systems with continuous proper functions (Avez [2]).
147
148
THEOREM A 16.2
Let (M, p., t)
be an ergodic classical system. Denote by C the sub-
group of the discrete spectrum corresponding to continuous proper functions,

then:
rank
C :s
b1
Before proving this theorem we first introduce the winding n':lmbers.

WINDING NUMBERS A
16.3
Let (M, p., t) be an ergodic classical system.
The first homology
group H 1(M, R) has finite integral base: Y1"'" Yb l ' Each Yk is a closed
curve which can be assumed differentiable. Let a
:s T I
Itx\ x ( M,
O:s t
be an arc of an orbit of t. We join the endpoints Tx and x with
a geodesic arc {3 (geodesic in the sense of some Riemannian metric).

Thus, y(T) = a{3 is a piecewise differentiable closed curve (see Figure
A 16.3 ') and there exist integers nk(T) such that:
y(T) = n 1 (T) Y1 + +.nb1(T). Yb 1
Fieure A 16.3'
Let (w k ) be the dual base of '(Yk) in the first cohomology group H1(M, Z),
that is the closed one-forms satisfying:
if i = k
ifi,tk.
149
APPENDIX 16
We obtain:
which can be rewritten:
(A 16.4)
(y, w k ) is the value of the one-form w k for the infinitesimal generator y of t at point tx. Since length {3 < diameter M, we have:
where
(A 16.5)
!. (Wi<
lim
1{3
T-+oo
On the other hand, ergodicity implies (see 7.1):

(A 16.6)
for almost every initial point x, and the limit does not depend on x. Finally, from (A 16.4), (AI6.5), and (AI6.6) we deduce that:
.
(A 16.7)
nk(T)
hm - T
T-+oo
f.
(y, wk)dJL
de!
JLk
exists for almost every x and does not depend on x.

The numbers a k
= e2TTil-'lc,
1, ... , b l , ~enerate a sub~roup ~ of
the circle ~roup; ~ is called the ~roup of the windin~ numbers. It is read-
ily seen that ~ does not depend on the base (Yk). Thus, t defines a
r:al homology class:
the winding numbers JLk define the "homological position" of a generic orbit. In other words, they define how an "average" orbit w-anders around M.
'This concept was first introduced by H. Poincare [1] for flows on the
torus T2. Further investi~ations of systems:
x=
F(x, y),
G(x, y)
150
on the torus T2 are due to A. Denjoy [1] and C. L. Siegel.

Proof of Theorem A 16.2 3
From its very construction, the rank of the winding numbers group is
bounded from above by b 1 , Therefore, Theorem (A 16.2) is a corollary of
the following lemma.
LEMMA A
16.8
The subgroup C of the discrete spectrum corresponding to continuous

proper functions is a subgroup of the winding numbers group.
Proof:
Let i(x) be a nonvanishing continuous proper function of I:

f ( IX) = e217iAI. f (x)
The function f is continuously differentiable with respect to the flow I'

that is:
217iA' e217iAI i(x) ,
which may be written:
(y,
(A 16.9)
for t
df(x))
217iAf(x)
O. But ergodicity implies (see Theorem 9.12)

If(x)1
constant
a.e.,
that is to say
li(x)1 = constant';' 0 everywhere,
because f is continuous. Thus, up to a constant, we have:

f(x) = e217iljl (X) ,
where 1jI: M
-+
S1 is continuous and (A 16.9) reads:
(A 16.10)
Therefore dljl defines a closed current of degree 1 in De Rham's sense. 4
3 See also I. M. Gelfand and Shapiro-Piatetski
[1].
S. Schwartzman
4 G. De Rham. Varieles differenliables. Hermann (Paris) 1955.
[1].
151
APPENDIX 16
As it is well-known, such a current is homologous to a smooth closed oneform [dy,]:

dy,
[dy,] + dh .
Assume that M is metrized with a Riemannian metric whose volume element is dfL. Since "if> t is fL-measure-preserving, the infinitesimal generator
yof
; is co-closed (oy = 0)
Then, relation (A 16.10) implies:
(y, [dy,]) dfL =

(y, dy,) dfL = A
M
M
Now it is sufficient, according to (A 16.3), to prove that [dy,] has integral
periods. Let u: [0, 1] .... M be an arbitrary smooth loop of M. Since dy,
and [dy,] are homologous, we have:
f
u
[dy,]
dy,
= Y, [u (1)] -
y, [u (0)] ( Z .
(Q. E. D.)
COROLLAR-V' A 16.11 (Arnold [2], [3]).
Let V be a compact orientable Riemannian manifo: { which is not. a torus and the dimension of which is greater than !'ne. If the geodesic flow on
the unitary tangent bundle M
Tl V is ergodic, then the continuous proper
functions reduce to constants.

Proof;
According to Lemma (A 16.8) it is sufficient to prove that every winding
number vanishes. Under our topological assumptions, Gysin [1] proved that
every closed one-form w, which is not homologous to zero, is the lift of a
closed one-form of V, say again w. On the other hand, a winding number
of the flow has the form (see A 16.7):
where 71 (resp. a) is the volume element of V (resp. the fiber Sn-l).
152
From
we deduce:
f (1
V
S"- 1
(y, w) a)1J =
Jrv
(1
S"-1
ya, w)1J
o
(0. E. D.)
APPENDIX 17
SPECTRA OF K-SYSTEMS
(See Theorem 11.5)
A. Subalgebras of Measurable Sets
Let (M, /L) be a measure space. We denote by 1 the algebra of all the
measurable sets and by
DEFINITION
6 the
algebra of the subsets of measure 0 or 1.
A 17.1
A subalgebra (f of measurable sets is a subset of
which is closed
under the formation of complement and the denumerable union, and which
contains M.
INCLUSION A
17.2
If (1'0 and (fl are subalgebras of
1,
then (fo C (fl means that (fo is
a subalgebra of (tl' that is, that every element of (to is an element of (fl.
The relation C is a reflexive partial ordering on the family of sub algebras
of 1.
INTERSECTION A 17.3
Let
(1'); ([
be a family of subalgebras of 1. We denote by
(f
;(/
the largest subalgebra of

SUM
which belongs to each (fr
A 17.4
Likewise, we denote by
153
154
V G';
; ( I
the sum of the (1.'s,

that is, the smallest subalgebra of
1
which contains
every (J;.
THE SPACE
L 2 (G'). A17.5
Let (J be a subalgebra of 1. We denote by L 2 (G') the subspace of
G'.
L 2 (M, /L) generated by the characteristic funCtions of the elements of
The following preperties are readily verified:
G' c 93
L2(
implies L 2 (G') C L 2 (93) ,
n G';)
;(1
L 2 (6)
".
n L 2(G';)
;(1
". H 0' one-dimensional space of the constants.
B. Spectra of K-Systems
We next prove the following theorem (see Theorem 11.5).

THEOREM
A 17.6
A K-system (M, /L, ep) has denumerably multiple Lebesaue spectrum.

Recall that there exists (see Definition 11.1) a subalgebra
G'
of
such that:
(A 17.7)
6 '=
nG' . c -1G' c G' c G' c c
V nG'
1.
n=-oo
The proof breaks up into several lemmas.

LEMMA A
17.8
Let U be the unitary operator induced by . If H ". L 2 (t) then:
155
APPENDIX 17
n""
HO =
UnH
cc UH
Let us denote by H
C H CC
n=-oo
UnH = L 2(M, /l).
0=-00
e H0
H' the orthocomplement of H 0 in H; this may
be written again:
101
""
n
UnH'C"'CUH'CH'CU-1H'C'"
0==-00
""
U
UnH' = L; = L 2 (M, /l)
e Ho
0=-00
Proof:
Let A be .an element of
(1, and
~ A its characteristic function. Then:
if (x)
=
if (x) ( A ,
that is,
{
U~A(x)
Thus,
uXA
0 if x
I -lA
1 if x ( -lA .
== ~-lA and, according to the definition of L 2 f), we have:
UL 2(f) = L2(-1(f)
Now, the lemma is a direct corollary of properties (A 17.5).

L~MMA
A17.9
U has Lebesgue spectrum, the multiplicity of which is equal to
dim (H
U H) .
Proof:
Let
Ih.1I be a complete orthonormal basis of
H' e UH '. Denote by J{.I
the closure of the subspace spanned by hi' Uh i ,.... From their very construction, the Uih I.'s, and so the J{.'s
are orthogonal to each other. From
I
Lemma (A 17.8) we have
plete system of H ':
n""
n=-~
UnH' =
101.
therefore lUih.!
is a comI
156
(A 17.10t
On the other hand, the relation:
can be rewritten:
oJ
v-nH'
L~.
n= 0
This relation and (A 17.10) yield:

00
EB
~
i
V-n}{i
n=O
Let us set:
(A 17.11)
H,
U V- n }{.,
,
n=O
this can be rewritten

L~ =
(A 17.12)
EB
~ Hi
i
According to (A 17.11), from the basis Ih i' Vh i' ... 1 of }{ i we obtain a complete orthonormal basis of Hi' namely:
Furthermore:
Ve ,,]
..
for every i and j.
Together with (A 17.12), this proves that V has Lebesgue spectrum,
the multiplicity of which is equal to the cardinality of IH ii, that is
dim (H' e VH') = dim (H
uR)
157
APPENDIX 17
LEMMA
A 17.13
dim(H
uR)
00
Proal:
Lemma (A 17.8). implies:
$. dim UH $. dim H $. dim U-1H $.
This proves dim H
dimUnH
dimU n+1H
Similarly, dim UH
function I ( H
and UH ,;, H, since dim H < 00 would imply
00
00.
dimlOI for n large enough, that is H
Since UH ~ H, there exists some non-vanishing
UH. Let us denote by F the support of I:
F =
1m I m (
We have F ( (1 and p. (F) > 0 for

L =
Ig:tFlg
Lo
i(
01
L2 (1), and the space
00.
F)
Similarly, as p. (F) > 0 and dim UH
19:tFI g ( UHI has infinite dimension. Let us set

L e L1 and take g.:t F ( Lo and h ( UH, then we have:
the space L1
=
M,/(m) ,;,
(:t F : characteristic function of
(HI
has infinite dimension for dim H

00,
= O.
<g:tFlh> = <g:tFlhX F > = O.
Thus Lo C H
e UH,
and it is sufficient to prove that dim Lo
AS,[,l is infinite-dimensional and
00.
p. (F) > 0, there exists a sequence
of bounded real-valued functions: h 1, h2' ... ( UH such that the XF~h1'
:t ~2'
... ( L1 are linearly independent. Since
I does not vanish in F,
we find that the Ih 1, Ih2' ... ( L are linearly independent. But they belong to Lo because h ( UH implies:
(h k h ( UH is ortJlOgonal to
n.
Thus, there exist infinitely many linear-
ly independent functions in Lo.
(Q. E. D.)
The preceding lemmas prove Theorem (A 17.6).
APPENDIX 18
CONDITIONAL ENTROPY OF A PARTITION a

WITH RESPECT TO A PARTITION {3
(See Section 12, Chapter 2)
A. Measurable Partitions
DEFINITION
A 18.1
Let (M, /1) be a measure space. A partition a : IA)j ([ of M is a col-
lection of nonempty, nonintersecting measurable sets that cover M:

,,(A.I nA.):
0 if i 1= j, ,,(M-U A.)
:
J r .
I
O.
A partition a is said to be measurable if there exists a countable system
lB.J LJ l J
of measurable sets such that:
(1) each B j is a sum of elements of a;

(2) for any two elements A., A., of a there exists a Bk such that eiJ
ther Aj C B k , Aj
rt.
B k , or Aj
rt.
B k , Aj C B k
A finite or countable partition is clearly measurable.

DEFINITION A 18.2
From the very definition, it follows that we can remove the elements of
measure zero from a partition.
Mo~e
generally, two partitions a and {3 will
be identified: a: {3 (mod 0) if their elements coincide up to some sets of

measure zero.
In the future we delete (mod 0).

DEFINITION A 18.3
A partition {3 is said to be a refinement of a partition a: a:S {3 if

158
159
APPENDIX 18
every element B of {3 is a subset of some element A of a: p. (B - B n A)

=
o.
DEFINITION A 18.4
Let
la;\iEI
be a family of measurable partitions. We define their sum:

a =
ai
iff
as the smallest partztion which contains every ai In other words:
n AIJ A.J
a ={
a. for all
jff
The operation
i}.
V is commutative and associative and

a :S. a',
{3:S. {3' imply a V {3 :S. a' V (3'
DEFINITION A 18.5
Given an arbitrary measurable partition a, we denote by m(a) the sub-
al~ebra of the algebra
f that consists of the sets
that are sums of elements
of a. The al~ebra m(a) is called the algebra ~enerated by a.

It turns out that for every subalgebra (j of
i, there exists a measur-
able partition a such that:
One verifies at once:

a
= (3 <
a :S. (3
'l]j (
>m(a)
= m({3)
< >m(a)
C ')J{(f3}
ai )
iEI
m(a)
iEI
(See A. N. Kolmogorov [3] and V. Rohlin [3]).
8. Entropy of a Given {3
Let a
IAi I i
= 1, ... , d and {3 =
IB j 1 i
= 1, ... , 81 be- two finite mea-
surable partitions. We can assume, without losing generality, that any

element Ai or B j has positive measure (see A 18.2).
160
DEFINITION
A 18.6
Let z (t) be the function over [0, 1] defined by:
z(t)
- t log t if 0 < t
:s 1
ift=O.
f3
The conditional entropy of a with respect to
is:
h(alf3} = ~ Il(B j ) ~ z(Il(A/B j )),

i
where
Il(A./B.)
1
Il(A. nB.)
1
Il(B.)
]
is the conditional measure of Ai relative to Bj"

We tum next to the proof of Theorem (12.5) which we reformulate.
THEOREM
Let a
12.5
f3 = IBjl,
1Ai\'
=-ICkl
be finite measurable partitions.
Then:
h (al f3) ~ 0 with equality if, and only if a
(12.6)
:s f3;
h(aVf3.(y) = h(aly) + h(f3/aVy);
(12.7)
:s f3
(12.8)
(12.9)
f3:s
(12.10)
> Maly)
:s h (f3/y) ;
> h(a/y)
~ h(alf3};
h(aVf3/y)
:s h(aly) + h(f3/y)
Proof:
Proof of (12.6) is left to the reader as an easy exercIse.
The elements of aV f3 and aVy are, respectively, of the form: Ai
B j and Ai
n Ck
Therefore
h (aV f3/y)
But we have
161
APPENDIX 18
jl(A,nBjnC k )
jl(Aj~Ck)
jl(AjnBjnC k )
jl(C k )
jl(A,nC k )
jl(C k )
= jl(A/CkhdB/ A( n'C k )
and we deduce relation (12.7):

h(aV{3/y)
= -
jl(AjnBjnC k ) Log jl(A/C k )
i,j,k
= -
jl(A j
n C k ) Log
jl(A/C k ) -
i,k
h(aly) + h({3/aVy)
Let us prove relation (12.8): If a :::. {3, then aV {3
(3 and relations
(12.6) and (12.7) imply:

h({3/y)
h(a/y) + h({3/aVy) ~ h(aly)
Let us prove relation (12.9): Since lk jl(CkIB j ) = 1 and jl(C/B j )

~
0, the concavity of z (t) implies:
Since {3 :::. y, each B j is the disjoint union of some Ck's; therefore we

have:
where the sum extends over those C k,'s be longing to B j
We deduce:
162
z(p.(A/Ck))p.(CkIB j ) $. z[p.(A/B;)].
Multiplying both members by p. (B j) and summing over i and j yields

(12.9).
Finally, (12.10) is a consequence of (12.7) and (12.9): a V y ::: y implies
h (3/ aVy) $. h (3/y)
and
h(aV(3/y)
h(a/y) + h(3/aVy) $. h(aly) + h(3/y).
The preceding definitions and properties extend to denumerable measurable partitions (see Rohlin and Sinai [5]).
APPENDIX 19
ENTROPY OF AN AUTOMORPHISM
(See Theorem 12.26)
The purpose of this appendix is to prove the following theorem dl.e to
Kolmogorov.
THEOREM
A 19.1
If possesses a generator a, then h () = h (a, ):
The proof breaks into several lemmas. Denote by F the set d all finite measurable partitions. Given a, {3
F we write
Ia, {31
= h (a
I (3) +
h ({31 a).
LEMMA
19.2
Ia, {31 is a distance on F.

Proof:
It is clear that
Ia, {31
O. From formula (12.6) of Chapter 2 we de-
duce:
Ia, {31
> h (a I (3)
It is also evident that
= h ({3
la, (31
I a)
='
>a $. {3 and {3 $. a
>a = {3.
1{3, al. According to (12.11), (12.12), and
(12.9) we have:
h(a/y) = h(aVy) -h(y) $. h(aV{3Vy) -h({3Vy) + h(j3vy) -h(y)
= h(aI{3vy)+h({3/y) ~ h(a/{3)+h({3!y)
1 The proof follows Rohlin [4].
163
164
and symmetrically:
h(y/ a)
5 h({3/a) + h(y/{3)
Addition yields:
la, yl ~ la, {31 + 1{3, yl

LEMMA A
19.3
Given ' h (a, ) is a continuous function on F in' its argument a.

More precisely:
Proof:
Given a, {3 ( F, we set:
an = aVa . Vn-l a ; {3n = (3V Vn-l{3

From (12.11) of Chapter 2 follows:
h({3n/an) -h(an /{3n) = [h(an V{3n) -h(an)]-[h(an V{3n)-h({3n)]
= h ({3n) -h (an)
Since h (
I) ~
0, we deduce:
On the other hand, from (12.10) of Chapter 2 follows:

.h(an
/(3 )n
= h(aVV
n- 1 al{3n
) < h(a/{3 n ) + + h(n-l a/{3 n ).
.
Similarly, from (12.9) and because {3, ... , n-l {3
Symmetrically:
Addition yields:
513
fT ,
'we have:
165
APPENDIX 19 _
Dividing both sides of this inequality by n and passing to the limit as n

->
00,
we obtain Lemma (A 19.3).
LEMMA A 19.4
If aI' a 2 , ... is a sequence of finite partitions such that
00
V
n~l
m(an )
1 .
then the set B of partitions {3 ( F, such that {3
~ an
for at least one
value of n, is everywhere dense in F.

Proof:
We-rreed to prove that for every finite partition a and every 0 > 0 there
exist an n and a {3 ( B such that:
{3
s.
an'
la, {31 < 0 .
Let AI" .. , Am be the elements of a.
is dense in
( m(an)
1,
Since
for every 0' > 0 there eXist an n and subsets AI',, A~_l
such that:
i~1 .... ,m-1.
Let us denote by {3 the partition of M into sets B 1
It is clear that {3
s.
an. On the other hand:
la. (31 ~
r hi (3) + h ({31 a)
...
Bm defined by:
166
IL(A)
!.
Il(B k nA j )
Il(B k )
- !
k
Il(A j n B k )
Il(A) !
- !
Log [ Il (A I. n B k)
Il (A j)
Log
[1l(B k nA.h
I
Il(B k )
Il (B k)
- 2 ! IL(AjnB k ) Log IL(AjnB k ) + ! IL(A) Log IL(A j )

i
i,k
These formulas show that

and vanishes when
Al
la, ~I
depends continuously on
A l , ... , A~_l
Ai, ... , A~_l
A m _ 1 . Therefore, if 8' is small
enough, then la,~1 < 8.

Proof of Kolmogorov theorem:
Assume that possesses a generator a. We set, for i\
F and q =
0,1, ... :
We have:
From Lemma (A 19.4) it follows that the set B' of partitions

that
~
~ l
F such
$. an for at least one value of n is everywhere .dense in F. Let
be an element of B '. Clearly:
Therefore, from (12.12) of Chapter 2 follows:

h (~m) $. h (iin+m_ l
167
APPENDIX 19
n + m -1
m
Now, observe that:
h(>.V V 2 q -2>..)
2q-1
2q-1
Thus, passing to
~he
... 2h (A, ), as q ...
00
limit as m ... + 00, we obtain

h ({3, )
h (a, ) .
Recall that B' is everywhere dense in F and that h ({3, ) is continuous

in {3 (Lemma A 19.3), then:
h (a, )
that is:
sup h ({3, )
B'
sup h({3, )
F
~ h() ,
\PPENDIX 20
EXA\IPLES OF RIEMANNIAN \IANIFOLDS
\"\ITII NEGATIVE CURVATURE

(See 14.1, Chapter 3)
Consider the proper affine group G of the real line
I tit
( R I. An ele-
ment g' of G has the form:
g:
-+
yt +
x, y (R,
x,
y > 0,
and can be denoted by (x, y).

Given g'
we obtain:
= (x', y'),
g'(g(t))
y'(yt+x)+x'
y'yt+y'x+x'.
Therefore, if we denote the group operation by .L this may be written:

(x', y').L (x, y) = (y'x + x', y'y)
The neutral element is e = (0, 1) an~ the inverse of (x, y)

Both .L and g
-+
IS
(_xy-l, y-l).
g-l are smooth operations. Thus, G is a L"ie group that
is diffeomo.rphic to. the upper half-plane I (x, y) I y >
01.
to. a Riemannian manifo.ld.

THEOREM A 20.1.
THE RIEMANNIAN METRIC OF G
The leFt-invariant metric of G which reduces to
at the neutral element e
(0, 1) is:
dx 2 + di
y2
168
No.W we turn G in-
169
APPENDIX 20
Proof:
To any element X = (x, y) of G corresponds the left translation LX
L x(U)
x 1. U,
where U
(u, v)
G .
We have:
!:) ,
( -u-x
y-' y
the tangent mapping of which is:
(A 20.2)
Define a metric over the Lie algebra TG e by setting:
This defines a left-invariant metric at each point X:
Therefore, if X = (x, y), (A 20.2) implies:
(~1)2 + (~2)2
In other words, the mejric is:

(A20.3)
DEFINITION A 20.,+
The upper half-plane G endowed with the metric (A 20.3) is called the
Lobatchewsky-Poincare plane.
It can be useful to represent a point (x, y) of G by the complex number
z = x + iy.
THEOREM A 20.S.
THE ISOMETRIES OF G
The symmetry (x, y)
-+
(-x, y) and the homographies:
170
(A20.6)
az + b
cz + d ;
z ... z
a, b, c, d (
R. ad-bc
preserve the metric (A 20.3).
Proof:
Proof is purely computational and easy if one observes that:
ds 2
-4dzdz
h
- - - - , were z
(z_z)2
X-ly.
THEOREM A20.7. ANGLES
The angles of metric (A20.3) coincide with Euclidean angles.

Consequently, words such as "orthogonal," and so on can be used unambiguously.
Proof:
i's proportional to dx 2 + di .
THEOREM A 20.8. GEODESICS
The geodesics of (A 20.3) are the straight lines: x
constant, y > 0
and the upper half-circles centered on ox. In particular, there exists one,
and only one, geodesic passing through two given distinct points.
Proof:
Let ab be a segment of x
0, y > O. For any arc y joining a and
b we have:
ds 2
ab
This proves that x = 0, y > 0 is a geodesic.

An image of this geodesic under any isometry (A20.6) is still a geodesic. We obtain so all the upper half-circles centered on ox and the halfstraight lines x
constant, y > O. In fact we obtained all the geodesies
for, given a vector u ( Tl G, there exists a half-circle centered on ox (ora

parallel to oy) which is tangent to u.
171
APPENDIX 20
THEOREM
A20.9.
CURVATURE
The Gaussian curvature of (A 20.3) is equal to -1.

Proof:
The Gaussian curvature K is constant,
tor the metric is invariant
under a transitive group of isometries. The Gauss-Bonnet formula applied

to a geodesic triangle L\
=0
A+ B+ C ':"
ABC gives:
1T
ff
K da
1T
+ K area L\
t"
~
The particular case of Figure (A 20.10) gives A
=0
O. As the
element of area is da = (dxdy)/y2, we obtain:

area L\
We conclude that K
1T
-1.
----O~=-B---------/------~~==~C~--~x
r
Fi~re
A20.10
172

;:
THEOREM A20.11. ASYMPTOTIC GEODESICS
Let y (u, t) = y (t) be a geodesic parametrized by arc length t, and

g
G. The geodesic passing through g and
as t1
->
}J!J.- has a
limit position
+ 00 (resp. -00). This limit position is the geodesic passing
through g and the intersection y(+oo) (resp. y(-oo)) of y with ox.

Geodesics emanating from y(+oo) (resp. y(-oo)) are called the positive (resp. negative) asymptotes to y.
Proof:
Let y(t 1) be a point of y. The geodesic passing through g and y(t 1)
is a circle centered on ox, possibly reduced to a straight line (A 20.8).
From the very definition of the metric (1\20.3), y(tl) runs to-ox as
tI
-+
+ 00 (resp. - 00), that is y (t 1) converges to the, intersection y (+ 00)
(resp. y(-oo)) of y with ox (see Figure A20.12). Thus our geodesic has a
limit position, namely the upper half-circle centered on ox and passing
through g and y (+ 00) (resp. y (-00)). Consequently, this limit position is
a geodesic.
y
'Y(-m)
Figure A 20.12
173
APPENDIX 20
DEFINITION A 20.13. HOROCYCLES
The orthogonal trajectories of the positive (resp. negative) asymptotes

to yare called the positive (resp. negative) horocycles of y.
THEOREM A20.14.
The positive (resp. negative) horc.::ycles of yare the Euclidean circle

of G which are tangent to y
=0
at y(+oo) (resp. y(-oo)). In particular,
the straight lines y = C > 0 are horocycles. They are positive horocycl/
of the axis oy (y
-+
00).
Proof:
The 'positive (resp. negative) asymptotes to y form the upper part of

the pencil of circles that are orthogonal to y
0 at y(+oo) (resp. y(-oo)).
Theorem (A 20.14) follows at once from the elementary properties of conjuii!,ate pencils of circles.
The points y(+oo) and y(-oo), which do not belong to G, have to be
removed.
THEOREM A20.1S. RIEMANNIAN CIRCLES
The Riemannian circles of (A 20.1) centered at m form the upper part

of the pencil of circles whose radical axis is ox and whose Poncelet
points consist in m and the symmetric m' of m with respect to ox.
Proof:
The Riemannian circles centered at m are the orthogonal trajectories

of the geodesics emanating from m. This family of geodesics is
nothi~g
but the upper part of the pencil of circles passing through m and m'.
(Q. E. D.)
In particular, the power of any point d of ox with respect to one of these
Riemannian circles centered at m is:
(see Figure A 20.17).

1 Notion due to Lobatchewsky (in Greek, "horos" = horizon).
174
THEOREM
A 20.16
Horocycles are Riemannian circles the radii of which are infinite and
the centers of which are at infinity (on y
0). '/
Proof:
Consider the Riemannian circle passing through a fixed point n of a
geodesic y and centered at m ( y (see Figure A 20.17). If m moves to
infinity along y, that is, if m converges to ox, then mm' ... O. Therefore,
the power of any point of ox with respect to our circle tends to zero. Thus
our circle has a limit position which is the circle tangent to ox at y (+ 00)
and which passes through n. Theorem (A 20.16) shows that this limit position is an horocycle. Conversely, any horocycle is obtained from the
(Q. E. D.)
above construction.
'Y(+OO )
m'
Fi\t:ure A20.17
175
APPENDIX 20
THEOREM
A 20.18
Let y(u, t) and y'(u', t) be two geodesics which are positively (to
fix the idea) asymptotic one to the other. We denote their arc length counted from their origins nand n' by t. Then, after a suitable selection of
nand n " we have:
d (y (t), y '(t)) $. nn 'e- t ,
t 2. 0,
where d means the Rierrftmnian distance, and nn' is the arc-length of the
horocycIe.
Proof:
Origins nand n' are selected on the same horocycle 1 (Figure A 20.19).
Denote by m and m' the intersections of y and y' with another horocycle
2. Arcs
nm
and
'm' are equal, for 1 and 2 are parallel curves:

n 'm' =
nm =
Let us compute the arc
mm'
t.
that belongs to 2. Horocycle 2 has the equa-
tion:
x
r sin u, y
= r + r cos
u.
Thus, with obvious notations:

m' -l-+-d
-Cu-o-s-u
Symmetrically, on horocycle 1:
nn
un'
tg -
- tg ~
A straightforward computation with y and y' leads to:

t =
om =
n'm'
Log 1tg
Log tg
~n I -
Log I tg u;
Uri -
Log tg
I'
U; 'I .
176
Figure A 20.19
Consequently:
tg
et
u
_n_
2
u
tg .2!..
2
tg
2--u
tg .2!..
2
mm
tg ~ - tg
2
tg
- tg
I-t
nn e
Theorem (A 20.18) follows from d(m, m ')
:s.
mm' .
-!!.
u
.2!!_
nn
mm
177
APPENDIX 20
GENERALIZATION A 20.20
The manifold V is the upper space xn > 0 of Rn endowed with the

metric:
(dx )2 + .. + (dx )2
1
(x )2
n
V is the Lobatchewsky space of constant curvature - 1. The horocycles
are (n -1) -dimensional manifolds, namely the planes xn
constant and
the Euclidean spheres of V which are tangent to the plane xn
o.
APPENDIX 21
PROOF OF TilE
LOBATCHEWSKY-IIADAMARD THEOREM
(See 14.3, Chapter 3)
A. Manifolds of Neglltive Curvature
Fa:>,
of negative
,t us recall some classical properties of Riemannian manifolds

'-M
'me.
THEOREM A 21.1
Let V be a complete, simply connected Riemannian manifold of negative curvature. Then:

(1) There exists one, and only one, geodesic passing through two
dis'tinct given points;

(2) V is diffeomorphic to the Euclidean space;
(3) let ABC be a geodesic triangle whose angles are A, B, C 'and

whose sides are a, b, c. Then:
Proof will be found in S. Helgason [1],

A direct consequence is the following corollary;
COROLLARY A 21.2
Under the above assumptions, Riemannian spheres of V are convex,

that is, a geodesic has at most two common points with a sphere.
178
179
APPENDIX 21
B. Asymptotes to a Given Geodesic

As usual, y (x,
II,
t)
= y (t) = y
denotes a geodesic emanating from x,
with initial velocity-vector u and arc length t. The point of y corresponding to t is denoted also by y (t). The Riemannian distance of two points
a and b is denoted by
la, bl.
Denote a complete, simply connected Rie-
mannian manifold of negative curvature by V.

THEOREM A
21.3
Let v' be a point of V. The geodesic joini!lg v' to a point y (t) ( y

converges to a limit as t
->
+00 (resp. t
->
-(0).
This limit is a geodesic.
froof: (See Figure A 21.4.)
v'
Y(v',u',t)
----------------~------------~----~--~Y
Fi~re
A21.4
The points v' and y (t 1 ) define one, and only one, geodesic y (v', u l' t ).
We set
s1 =
Iv: y(t 1)1.
Take t2 > t1 and apply relation (3) of Theorem
(A21.1) to the geodesic triangle v.', y(t 1), y(t 2 ); With obvious notations
we have:
On the other hand, the triangular inequality applied to v, v' y (t 1) gives:
t1 whence:
lv, v'l
~ s1 ~ t1 +
lv, v'l '
180
Similarly:
We deduce:
1,
that is to say:
Thus, according to Cauchy, u 1 converges to a limit u' as t1
->
00.,
The geodesic y (v', u ',t) is the limit position of y (v', u l' t), for the
exponential mapping Expv' is continuous.
y(v', u', t) is called a posi-
tive asymptote to y. Negative asymptotes are defined in the same way

(t 1
->
-(0).
REMARK
A 21.5
It is readily proved that the positive asymptote to y emanating from a
given point of the positive asymptote y (v', u', t) is nothing but y (geometrically). Therefore, we may speak of a positive asymptote to y without referring to a definite point v'. Furthermore, the set of the positive
asymptotes to y is a (dim V -1)- parameter family of geodesics.
C. The Horospheres 1 of V
The Riemannian manifold V is again complete, simply connected, and

of negative curvature. Let y (v,
11,
t )., =' y (t) be a geodesic and v' an ar-
bitrary point of V.
LEMMA
A21.6
converges to a finite limit L (v': y, v) as t

differentiable function of v' and v.
1
See A. Grant
[1].
->
+ 00, and this limit is a C 1 .
181
APPENDIX 21
Proof:
Take t2 > t l' The triangular inequality applied to v, Y (t l)' Y(t 2 )
gives:
(t2)
lv', y(t2)1 -Iv, y(t2)1 ::: lv', y(tl)1 + Iy(tl)' y(t2)1 -Iv, y(t2)1
=
Iv,' yUl)1 -Iv, y(tl)1
(tl) .
Therefore, (t) decreases monotonicall::::. On the other hand, (t) is

bounded, for the triangular inequality applied to v, v', y (t) gives:
This proves the existence of:

(t) = L(v'; y, v) .
lim
t ..... +
00
The second assertion follows from the inequality:
that is
Obviously:
L(v'; y, v) - L(v'; y, Vl) = vVl
(A 21. 7)
where
vVl
is the algebraic measure of
vVl
'
on the oriented geodesic y .
DEFINITION A 21.8
The locus of the points x for which L (x; y,O) = 0 is called the
pos~
itive horosphere through 0 of y and will be denoted by H+(y, 0).

According to Lemma (A21.7), H+(y,O) is a Cl-differentiable submanifold. of dimension (dim V-I). Let vt be an arbitrary point of y. Relation
(A 21. 7) shows that H +(y, 0) has equation:
L(x; y,vt )
OVl
Now we obtain the horospheres as spheres with center at infinity and radius
182
infinite. The Riemannian sphere, the center of which is a and passing

through b, will be denoted by L (a, b).
LEMMA
A21.9
L (y(t),O) converges to H+(y,O)
as t -+ +00.
Proof:
Let x be a point of H+(y, 0), we have:

(t) '" Ix,y(t)I-IO,y(t)1
On the other hand, (t)
-+
0 as t
-+
+00.
O. Therefore, L (y (t), 0) intersects the geo-
desic segment xy (t) at a point b(t) (see Figure A 21.10).
-----+--------~--------+-------~y
Figure A21.10
We have:
lx, b(t)1 = Ix,y(t)I-ly(t), b(t)1
Ix,y(t)I-IO,y(t)1
-+
0 as t
-+
+00.
This means that every point of H+(y, 0) is a limit point of the spheres
L(y(t),O) as t
-+
+00. Conversely, we prove that SUGh a limit point be-
longs to H+(y,O). Let b(t) be a point of L (y(t), 0) and x
lim b(t).
t -+ +00
The triangular inequality gives:
Ilx, y(t)1 -10, y(t)11 < I lx, y(t)1 -Ib(t), y(t)11 + Ilb(t), y(t)1 -10, y(t) II
Ix, b( t) I
-+
0 as
-+
+ 00
183
APPENDIX 21
Therefore: L (x; y, 0)
0, that is,
x
H+(y,O).
COROLLARY A21.11
Horospheres are convex, and strictly convex if the curvature of V
IS
boundoo from above by a negative constant.

Proof:
H+(y,O) is the limit of the balls passing through 0 and the center of
which goes to infinity along y, and these balls are convex (see A 21.2).
LEMMA
A21.12
Let H+(y, 0) and H+(y, 0') be two horospheres of y. If a (H+(y, 0)

and a' ( H+(y, 0'), then la, a'i :: 10,0'1.
Proof:
Assume la, a' I
< 10, 0' I. From (A 21. 9) we conclude that to each
corresponds a point a (t)
and a point a'(t)
~ (y (t),
~ir(t),
0) such that:
lim a(t) ~ a,
t ... + 00
0') such that:
lim
t ... + 00
a'(t)
a'.
Thus, for t large enough, we have:

la(t),a'(t)1
< 10,0'1.
To fix the ideas assume that the point 0' lies between the points 0 and
y (t). We obtain the following contradiction:
laCt), y(t)1 ::; laCt), a'(t)1 + la'(t), y(t)1
~ 10,y(t)1
LEMMA
< 10,0'1 +
la(t),y(t)I
la'(t), y(t)1
(Q. E. D.)
A21.13
Two positive horospheres H+(y, 0) and H+(y, 0 ') cut off an arc of
length 10,0' I on every positive asymptote to y.
184
Proof:
Y(a',Uo,t)
-----+--------~------------~-----------Y
0
FIgure A 21.14
Let y(a', u, t) be a positive asymptote to y that intersects H+(y, 0')

at a '. The points y (t) and a' define a geodesic on which we select a
point a(t) such that la(t), a'i = -L(a'; y,O) = 10,0'1 and a' lies between a (t) and y (t) (see Figure A 21.14). Since the exponential mappinl
Exp a' is continuous, we obtain: .
a(t) = a
lim
y(a',u',t) and la,a'i
-L(a'; y,O).
We deduce:
I la,y(t)I-IO,y(t)1 I :; Ila,a(t)1 + la(t),y(t)I-IO,y(t)11
=
I Ia, a (t) I + Ia', y (t ) I -I
"
y (t) II
->
0 as t ... + 00
Thus, a ( H+(y, 0).

(Q. E. D.)
THEOREM A
21.15
The positive asymptotes to yare the orthogonal trajectories of the

positive horospheres of y.
Proof:
DireCt consequence of (A 21.12) and (A 21.13).
Finally, observe that negative horospheres H-(y, O) can be defined as
above from the negative asymptotes (t
->
-00).
185
APPENDIX 21
D. The Horospheres of Tl V
The unitary tangent bundle of V is denoted by Tl V and p: Tl V
-+
is the canonical projection.

Let u be a point of Tl V; u defines a geodesic y (pu, u, t)
= y (u, t)
y (t) .the lift of which, in Tl V, is denoted again by y (t). From B we

know there exist two horospheres H +(y, pu)
H +(u) and H -(y, pu) =
H -(u) passing through pu. The set of the unitary vectors orthogonal to
H+(u) (resp. H-(u)) along H+(u) (resp. H-(u)) and oriented like u is a
(dim V-I)-dimensional submanifold }(+(u) (resp. }(-(u)) of Tl V. The }{'s

are called the horospheres of Tl V.
THEOREM
A 21.16
(1) The y(u, t)'s and the }{+(u)'s, }{-(u)'s are the sheets of three
foliations of Tl v.
(2) At each point u ( Tl V these foliations are transverse, that is:
T(T1 V)u
where
X:
(resp.
X;;,
X;
E9
X;; Zu
Zu) is the tangent space of }(+(u) (resp. }(-~u),
y (u, t) ) at u.
(3) These foliations are invariant under the geodesic flow t:
Proof:
(1) Follows from the very construction of the sheets.
(2) Follows from the strict convexity of H + (resp. H -, see A 21.11).
(3) Follows from Theorem (A 21.15).
The invariance of the foliations reduces the study of the differential
; to the study of its restriction to }(+(u) (resp. }(-(u) ) and y(u).

we assume definitively that V is the universal covering
Iv
Now,
of a compact
. Riemannian manifold W of negative curvature. In particular, 'the curvature

of V is bounded from above by a negative constant _ k 2 .
186
LEMMA
A21.17
Let rs(t) be a one-parameter family (s> 0) of numerical, C"-differentiable functions. Assume that:
2
r. s >
- k r s
(k = constant
> 0)
for every s,t 2: 0, and rs(O) > 0, rs(s) = O. Then:
<
r (t)
S
cosh [k (s -~, for 0 S t S s .

cosh [ks]
r (0)
s
Assume, additionally, that:

lim
s-++oo
Then, for s large enough:

[i)t)[<kors(O).
sinh [k (s - t)]
,forOStS4.
cosh (ks)
Proof:
The function:
r s(O)
rs(t) -
cosh[k(s-t)]
cosh (ks)
verifies:
i'
.s
(t)
>
k21 (t),
s
Thus, Isis concave between 0 and s and vanishes for t
0, s, conse-
quently I s(t) S 0 for 0 S t S s. This proves the first part. This proves
also that is increases between 0 and s; thus: i s(t) S is(s) for 0 S t
s. On the other hand,
in particular,
is (t) <is (s)

The second part follows easily.
rs(s)
... 0 as s ... + 00
187
APPENDIX 21
THEOREM
A 21.18
Let >t be the geodesic flow of Tl V. Then, for any positive number
1I>;(il S..b'e-ktII(II,
II>; (;1
a'ektll(ll,
The positive constants
IlcP~t(11 ~a.ektll(11 if(lX:,
11>~lll S
a and
b'e-ktll(11
are independent of
if ( l X~
t and
~,
and
!I
denotes the length of a vector of Tl V equipped with its natural Riemannian metric.
Proof:
We p.rove the first inequality, the others can be proved in the same way.
Let y (0, u, t)
y (t)
y be a geodesic of V, and let x be a point of
H+(y,O), close enough to O. There is a well-defined geodesic y s(x, us' t)

= y s(t) passing through x and y (s)
the Riemannian distance of y(t) and
y. Our first purpos"e is to compute
ys(t), regarded as ~lements of
Tl V.
Let r s(t) be the Riemannian distance of their projections y(t) and y s(t)
on V. To compute r s(t) 'we consider a Jacobi field 2 1/1 (t) -along y, 'that is
orthogonal to y and vanishes for t
s. By definition:
where R ( , ) is the curvature tensor and V is the covariant derivative

along y. By definition the sectional curvature in the two-plane (y,l/I) is:
< R (y, I/I)y, 1/1 >

:11/1 112
We know that p(y,l/I)
On the other hand,
2 See
J.
Milnor
[1].
_k 2 , consequently:
188
V <VI/I,I/I>
YlV 2 \11/I112
Yl d2211 1/1112 ,
dt
I\VI/I 112 ? (~ 111/1 11)2

Therefore, the length ls(t). of I/I(t) verifies:
that is
rs
>
k 2 I s , and I s (0) > 0,
-
ls(s)
0 .
Lemma (A 21.17) and the classical possibility to select the Jacobi field
1/1 such that:

r s(t) = I s(t) + 0(1)
if x is close enough to y imply:

r s(t) < r s(O)
(A21.19)
cosh[k(s-t)]
, for 0 ::;, t ::;, s .
cosh (ks)
Now it is readily seen that the angle of y and y s at y (s) converges

zero as s
->
+ 00. Thus,
i)s)
->
0 as s
->
~o
+ 00, and Lemma (A 21.17) im-
plies again:
I; s(t)1 <
(A 21.20)
As s
->
k r (0) sinh [k (s - t)]

s
, for 0 ::;, t ::;, s .
sinh (ks)
+00, ys(t) converges to a point y'(t) of the positive asymptote
y'(x, u', t) to y, and
ys(t)
converges to V(t). If r(t) denotes the dis-
tance of y (t) to y '(t), then the inequalities (A 21.19) and (A 21.20) imply (s
->
+(0):
IHOI < kr(O) e- kt ,
for t > 0 .
189
APPENDIX 21
Thus, the Riemannian distance of y(t), V(t)
Tl V verifies:
We easily deduce th.e first inequality of Theorem (A 21.18).
Due to this theorem, the sheets J{+(u) (resp. J{ -(u) ) are called the
~
"contracting" (resp. "dilating") sheets of Tl V.
E. Proof of the Lobatchewsky-Hadamard theorem 3

THEOREM
A 21.21
Let W be a compact, connected Riemannian manifold of negative curvature, then the geodesic flow on Tl W is a C-flow.
Proof:
Let V =
IV
be the universal covering of W equipped with the inverse
image of the Riemannian metric of W under the canonical projection":
IV . . W. V
satisfies the assumptions of preceding sections. Thus the geo-
desic flow on Tl V verifies the conditions of C-flows: condition (0) is

trivially fulfilled; condition (1) follows from Theorem (A 21.16); condition
(2) follows from Theorem (A 21.18). We finish the proof by proving that"
is compatible with the three foliations of V
IV
and Tl IV. The first ho-
motopy group "l (W) is isomorphic to a group of automorphisms of
IV,
for
W is connected. The group "l(W) acts also as a group of automorphisms
of Tl IV: if u',
J{ .(u
U" f
Tl
IV
are congruent mod" 1 (W), then J{ .(u ') and
") are themselves congruent mod" 1 (W).
REMARK
A 21.22
The horospheres of a compact, n-dimensional manifold Ware diffeomorphic to Rn-l. In fact, let us consider the horosphere J{ +.
It is a
See J. Hadamard [l]. Proofs of SectlOns Band C are mainly due to H. Busemann: Metric Methods in Finsler Spaces and Geometry. Ann. Math. Study. No.8.
Princeton University Press.
190
paracompact manifold. Let S be a compact subset of J{+. Then 1>tS is

covered by a disk D of
1> t J{ +
(take t large enough). The counterimage
1>; 1D i~
a disk which covers S in J{ +. Therefore, J{ + is diffeomorphic
to Rn -
according to the following lemma of Brown (Proc. Amer. Math.
1,
Soc., 12(1961), 812-814) and Stallings (Proc. Cambridge Philos. Soc., 58

(1962), 481-488): Let M be a paracompact manifold such that every compact subset is contained in an open set diffeomorphic to Euclidean space.
Then M itself is diffeomorphic to Euclidean space.
This result
~oes
not hold for noncompact manifolds.
Consider the
space I(x, y) I y > 0, x-(mod 1)1 endowed with the metric:
The Gaussian curvature is equal to - 1 and the universal covering space

is the Lobatchewsky plane (see Appendix 20). The curve y
horocycle homeomorphic to S1.
FIgure A 21.23
1 is an
APPENnIX 22
PROOF OF THE SINAI THEOREM
Let (M, ) be a C-diffeomorphism and X m {resp. Ym ) the k-dimensional dilating space at m Crespo the I-dimensional contracting space). A Rie-
mannian metric is definitively selected on M. Thus, Xm and Ym are Euclidean subspaces of TMm'
THE METRIC SPACE OF THE FIELDS OF TANGENT k-PLANES
A22.1
The tangent space TMm is the direct sum Xm
III
Ym . Therefore, the
equation of a k-plane V m C TM m' transverse to fm' is:
where x ( Xm , y (fm , and P(Um): X m ... fm is a linear mapping. We

define a metric in accordance with the norm of the linear mappings P(V):
if V m and U:n are two k-planes of TM m ' then we set:
IVm -V'[
m
IIP(U m )-P(U')II
m
sup
X(Xm
.lx[<l
IP(v)x-P(U')xl
m
We turn the set K of the fIelds of the tangent k-planes transverse to Ym

into a metric space by setting:
IV-V'I
= sup
m(M
IV m -
V'
I.
V, V' ( K.
The distance of Inequality (15.3) of Section 15 has to be understood in

this sense. Since M is compact, then K is a compact and complete metric space.
191
192
LEMMA A22.2
Let RI and R2 be two n-dimensional Euclidean spaces (n
Assume that R j (i
Rj =
Let A: RI
->
dim M).
1,2) is the direct sum of two subspaces Xj and Yj:
Xj
III
lj,
dim Xj
k,
dim Y.I
I:
R2 be a linear mapping such that:

X2 ,
AX I
Y2 ,
Ixl
for x ( Xl
IIAyl < alxl
for y ( YI
IIAxl
(A 22.3)
AYI =
~ fL
where fL and a are constants.

Let us denote by
Ci'
the operator induced by A, which makes corre-
spond to the k-planes of RI the k-planes of R 2 . If V and V' are transverse to YI , then:
Proof:
By definition:
lCi'v - Ci'v'l
sup IP(Ci'U)x - P(Ci'v')x1
Ixl < I
X(X 2
sup
Ixl
IA[P(U)A-Ix] - A'[P(U')A-Ix]1
<1
X(X 2
According to (A 22.3), Ixl < 1 and x ( X2 imply A-Ix ( Xl and IA-Ixl
:s.
fL- l . Therefore:
:s.
sup
Ixl<l
sup
Izl<rl
X(X 2
X(X I
and
:Ci'v-Ci'v'l:s.
fL- l
sup
!z < 1
Z(
XI
IA[P(U)-P(U')]z!
193
APPENDIX 22
Since [p(U) - p(U ')]z ( Yl , (A 22.3) Implies:

IA[P(U) - p(U ')]z
I s:.
a I[p(U) - p(U ')]z
~ a'
sup I[P(U)-P(U')]zl
Izl < 1
a'IU-U'1 .
Z(X I
Finally we have:
IAU-AU'I ~ /I-laIU-U'1
(Q. E. D.)
INEQUALITY
(15.3) OF SECTION 15. A 22.4
The mapping I ** (or a positive integer power un is contracting

in a neighborhood of the dilating field X:
for
that is, for Ul and U2 transverse to Y.

Proof:
We apply the preceding lemma and we set:
The linear mapping A is the differential (n) * of n. Since the dilating

and the contracting fields X and Yare invariant under , AX 1 = X 2 and
AYI = Y2 are verified. Inequalities (A 22.3) are consequences of the axioms of C-systems:
For n large enough we have
e=
I
In fact
**
a- l be- 2An < 1.
is the extension of the mappIng
**
to the k-planes.
(Q. E. D.)
APPENDIX 23
SMALE CONSTRUCTION OF c-OIFFEo\10RPHISMS

(See Section 12.3, Chapter 3)
Smale [3] has proved that there do exist nontoral C-diffeomorphisms.
We give next an example of his construction.
The Space M.
Let G be the nilpotent Lie group of the 6x 6 matrices:
C: ;J
f1
where x, y, z, X, y, Z ( R. The group G is diffeomorphic to R6.

Let us denote by Q(v'3)
Ip + qy'3l p, q (
zi the number field of YJ

x = p - qv'3 the nontriv-
adjoined to the rationals, and by x = p + qy'3 ~
ial Galois automorphism. We consider the subgroup
of G, the elements
of which satisfy
x, y, z ( Q( v'3) ,
X
It is readily proved that
x,
y,
=Z
is discrete and that the right coset space M =
lar! = G/r is compact.

Of course, the first homotopy group of M is isomorphic to
and, con-
sequently, is a nonabelian nilpotent group. Therefore, M is nontoral.

The Diffeomorphism : M
M.
Let us identify an element
a(G
mapping : G G by:
194
with (x, y, z, X, y,Z). We define a
APPENDIX 23
1> (x,
195
y, z, X, Y, Z) = (Ax, p.y, vz,
AX, ii Y, i7 Z)
where:
1>
1>
2 + V3,
v = (2-V3)2,
P. = Av = 2-y3,
is an automorphism of G, because p. = Av. Therefore, ;f
f, and
defines a diffeomorphism of M by:
(M, ) is a C-Diffeomorphismo
An element of the Lie algebra TG e of G is of the form
C:iO;g)
The metric
of TG e defines a right mvariant metric on G and, consequently, a Riemannian metric on M = Golf. The Lie algebra TG e splits into the sum X + Y,
where the elements of X (respo Y) are of the form:
(respo)
Next, by right translations, the splittmg

TG ~ of every point
IS
imposed on the tangent space
g of G:
TG~ = X~ + ~
Thus, the tangent space TMm at m ( M splits into:
m
= Xm+ Ym
It is ea,sily checked that the linear tangent mapping" d9

and contracting on Ymo
IS
dilating on X m
APPENDIX 24
SMALE '5 EXAMPLE

Smale [2] proved the following theorem, which gives a negative answer
to the "problem of structural stability": are the structurally stable diffeomorphisms dense in the C 1-topology?
THEOREM
A24.1
There exists a diffeomorphism

phism
1/1',
C 1-c1ose to
1/1,
1/1:
T3 ... T3 such that no diffeomor-
is structurally stable.
We tum to the construction of
1/1.
A. The Auxiliary Diffeomorphism >

Let T2 be the torus I(x, y) mod 1\. We define a diffeomorphism >1 of
T2 x Iz 1-1 ~ z ~ 1\ onto itself by setting:
>.
1
(p . . q P (P
(mod 1)
z ... ~z
Let By, be the ball (Figure A 24.2) of T2 x R with center (0, 0, 2) and ra-
dius
1/2:
x2
We defioe a diffeomo",hism
+;+ (z_2)2
~;
>1:
{Of
:v: i~~
~ ~ .
T2 x Iz 1
y ... ~y
z ... 2z-2 .
196
~ z ~
3\ by setting
APPENDIX 24
197
Now, the torus T3 is T2 x S1, where S1 is [-3,3] with endpoints identified.
x: dilating direction
Y: contracting direction
t
Figure A 24.2
The following lemma is easily proved.
198
LEMMA
A24.3
There exists a diffeomorphism : T3 ... T3 such that:

(1) its restriction to T2 x Iz 1-1 S; z S; 1\ is l'(2) its restriction to By, is
i ,-
(3) leaves {(O, 0, z) 10 < z S; 2\ invariant with no fixed point.

PROPERTIES OF
. A 24.4
T2 x 10 \ is obviously an invariant torus under . The restriction of

(or l) to T 2 x 10\ is nothing but the diffeomorphism of Example (13.1):
(~)
(A 24.5)
. . 0 0(;)
(mod 1).
Let us recall some properties of this diffeomorphism: There exist two foliations
X and
~ on T2 x 10\. They correspond, respectively, to the dilat-
ing and the contracting eigenspaces Xm and Ym of (A 24.5). Every sheet

of
X (or
~) is everywhere dense in T2 x 10 \. The periodic points 1 of
are dense in T2 x 10\. This fact can be proved by observing that every rational point (pi q, p 'I q) ( T2 is periodic.
Now we pass to the diffeomorphism : T3 ...
that the periodic points of
cp,
in T2 x I z
I-
those of (A 24.5), as do also the foliations
1 S; z
r.
s:.
It is easy to see
1 \, coincide with
and ~ in T2 x 10\. The foli-
ation ~ generates an invariant contracting foliation of T2 x 1z

S; 1\, whose sheets are the "planes" of the form Y x Izl-1
I-
1 S; z
s:. z S; 11,
where
Y is some sheet of ~.
B. The Diffeomorphism ifJ
The diffeomorphism ifJ is obtained by perturbing. cp_ Let Go be the

ball of T3 with radius d and center (0, 0, %):
x 2 + ; + (z _ %)2
We set G =
d2 .
cp-l Go' tx, y, z) = (x', y: z '),
1 That is, the points'; (

integer N.
s:.
T2x
10\
such that
N';
and we observe that d can
= .; for some nonvanishing
199
APPENDIX 24
be chosen small enough for

G
nG =0 .
We define. our desired diffeomorphism

tjJ(x, y, z)
(x, y, z)
r/J by setting:
= (x', y', z ')
outside G
(x' + 1/1I>(x, y, z), y', z ') on G,
where II> is a nonnegative COO function with compact support in G and

nondegenerate maximum value + 1 at -1(0, 0, %), and finally 1/
>
is
small enough so that tjJ is a diffeomorphism.

PROPERTIES OF tjJ. A24.6
" z)\
Now the curve tjJI(O,.O,
s.
This bump lies in the region T2 x I z
S. 21 has a bump B (see Figure A24.2).

\-1 $. z S. 11 where tjJ coincides with
l' This region is foliated into contracting planes (see A 24. 4). Let
~xla) + (ylb) =
1 be the equation of such a contracting plane in the chart
(x, y, z). Among these planes intersecting the bump B, we select the plane
~, for which a is maximum (see Figure A 24.2). Either j= contains a peri-
odic point of tjJ, or it does not. In the first case, the bump is called periodic and in the second case, nonperiodic.
LEMMA A24.7
tjJ is not structurally stable.
This follows from two remarks:

(1) An arbitrarily small change of 1/ in the definition of tjJ gives
rise to a diffeomorphism tjJ" arbitrarily C1-close to tjJ and similar to tjJ.

The density of the periodic points (see A 24.4) implies that we can suppose the bump of tjJ is periodic and the bump of tjJ" is nonperiodic, and
vice versa.
(2) If tjJ and tjJ" are in the opposite cases there is no homeomorphism h: T3
->
T3 close to the identity" such that tjJ".
h establishes a one-to-one correspondence

planes, and periodic points of tjJ and tjJ ".
b~tween
h=
h tjJ. In fact,
bumps, contracting
200
LEMMA
A24.8
Every diffeomorphism
!/J',
Ciclose to
!/J,
possesses an invariant tom.
similar to T2 x 101, a bump, a contracting sheet similar to
1,
and so on:
Complete. proof of this lemma is announced by Smale [2]. Now Theorem

(A 24.1) follows readily from Lemma (A 24.7) and (A 24.8).
APPENDIX 25
PROOF-OF THE LEMMAS OF THE ANOSOV THEOREM

Lemma A
Let (M, ) be a C-diffeomorphism. We select definitively a Riemannian metric on M. Since M is compact, there exists a number d > 0 such
that, whatever be the ball B (p ; d) C TM
with radius d and center p ( M,
.
p
the restriction
Exppl
B(p; d)
of the exponential mapping at p is a diffeomorphism. Let lnm I n ( Z!

be an orbit of ' A chart of a neighborhood of this orbit is (B, 1jJ-1),
where B is the sum of the balls
B
= B(nm, d)
C TM,J..n
o..p
m.
and the restriction IjJ I-B is Exp ,J..n' . Let us denote by X the dilating
n
"fJ m
n
k-space X (n m) of TMnm' and by Yn the contracting I-space y(nm),
Tke invariant dilating and contracting foliations rand
new foliations on B:
Finally, induces a mapping:
201
of M induce
202
such that the restriction 1 I
Bn
maps B
into B
n+
l' with obvious restric.
tions concerning the range of 1' Assume d small enough, then the sheets
of the foliations
X1
and '!:J 1 can be regarded as sheets of the Euclidean
space XnED Yn of origin 0 = nm, In which their equations are, respectively:

y
= y(O)
+ in(x, y(O)) and
= x(O)
+ gn(Y' x(O))
where x ( X n , Y ( Yn and where in' gn' and their first derivatives can bl:
made arbitrarily small by a suitable choice of d.
a n of '!:J n which passes through the center 0 of
Consider the sheet
The mapping:
e:
whose restriction
eIY
n:
I Yn I n
Yn
->
( ZI
an
->
I an In (
is defined for
ZI
y ( Yn n B n
by
is a diffeomorphism. Therefore, y ( Yn can be regarded as coordinates on
Figure A2S.1
203
APPENDIX 25
The diffeomorphism 1 maps an into a n+ 1 (see Figure A 25.1).

the coordinates y, this defines a mapping 2:
In
and
ASSERTION A 25.2
It follows from the very definition of C-systems that the restriction
is contracting:
(A 25.3)
where
e is a constant.
REMARK A 25.4
To be precise, (A25.3) holds for a certain iteration ~ of 2: we
must "kill" the constant b in the definition of C-systems. For simplicity,
we assume that (A 25.3) already holds for v = 1. Now let ' be a diffeo
morphism C 2.close to ' Then ' is a C-diffeomorphism (Sinai theorem,
Section 15) and the foliations :t~ = 1j1- 1:t.' 'lJ; =rljl-1'lJ' and the mapping
; induced by ':
;: B ... B.
; = 1j1-1'1jI,
IBn:
Bn ... B n+ 1 ,
are defined as above. If ' is C 2close enough to . then the sheets of
:t~ are close to those of :t1 and transverse to the sheet an' Therefore,
there exists a projection IT:
which makes correspond to each point a
Bn' the intersection ITa of an with
the sheet of:t; passing through a (see Fig~re A 25.5).

Now let us consider the mapping (Figure A 25.5):
e- 1 IT;e,
204
cp'
Figure A 25.5
ASSERTION A25.6
If ' is C 2-close enough to ,

If
;. is C-c1ose to 2: to any
> 0 corresponds a positive 0 su~: Lnat

11;'y-:yll <
(A 25.7)
where
I II c 2
If
II'- I c 2 < 0
for any y ( Yn
implies:
Bn ,
is the C 2 -norm.
Proof:
{ is close to 1' II: ; an -+ an+ 1 is small for { an '" 1 an =
a n + 1 , and the sheets of !; are transverse to an +l' Now denote the sheet
of!; passing through the center 0 '" m of Bo by {3,
LEMMA A. A 25.8
If ' is C 2 -close enough to , then the sheet 'n{3 is close to nm
for any n
(A25.9)
where
0,- To be precise:
IW;mll
e, is defined at (A 25.3).
<
If
1-e
205
APPENDIX 25
Proof:
According to (A 25.6), given ce > 0 there exists 0 > 0 such that
11'-ll c 2 < 0
implies:
From (A25.3) and (A25.7) we deduce:
II; y I < 8 Ily II

Set
IIYII <
ce/(l- 8); therefore, if
+ ce .
, then
II; y I <
and
Iwi y II
< , and so on. But since 11m II < , inequality (A 25.9) is proved.
REMARK
A25.10
Ily!1 <
From (A 25.3) and (A 25.7) one deduces also that
IW2n y I ::;
In fact
IIY I <
c and c
if c ? 1 ~ 8
c implies:
? ce/(1- 8) imply:
II; y II <
Ily II
+ ce
< 8c + ce < c .
(Q. E. D.)
Lemma B
Now we consider the sheet Yn of
Yn1
:x:
C B n be the corresponding sheet of
Yn
which passes through n m. Let
:x: 1:
',1,-1
'f'
Yn
'
The equation of Yn1 is y = fn(x,O). We define a mapping
11,
similar to
e,
by setting:
11: lXnln
11
Z!
lYn1 1n
->
is a diffeomorphism and x
11l xn :
ZI,
Xn
->
Yn1
X n can be regarded as coordinates on Yn1.
The diffeomorphism 1 maps Yn1 into Y;+1 ~ In the coordinates x, this

defines a mapping:
3 = ~-11:D:
lXnln
Obviously 3 (0) = 0,
Z!
->
lXnln
ZI,
206
ASSERTION A 25.11
It follows from the very definition of C-systems that 31x : X
n
Xn+ J
is dilating:
(A25.12)
REMARK
A 25.13
In fact, (A 25.12) holds for a certain iteration of 3. For simplicity

we assume that (A 25.12) already holds for 3. Now let ' be a diffeo-
f3 n =
morphism C 2 -close to . Consider the sheet

foliation
'In
f3
C Bn
X; = rjJ-J'X' which passes through e';(O) (n::. 0).
of the
(See Figure
A25.14.) According to Lemma A, this sheet is close to the center 0 of

Bn.
Let y = hn(x), (x ( X n ), be the equation of
enough to ' then we can choose x ( Xn
n Bn
f3 n .
If ' is close
as local coordinates of
f3 n :
the mapping
E
which is defined by x ~ (x, hn(x)) for x ( Xn
Ixn :
..
FIgure A 25.14
f3 n '
n Bn is a diffeomorphism.
Yn+l
q>'
Xn ~
207
APPENDIX 25
From the very construction of the (3n's, we see that ;
maps (3n into
(3n+1' Therefore, this defines a diffeomorphism:
ASSERTION A 25.15
If and ' are C 2 -close enough, then 3 and ; are C 1-close:

To any ce > 0 corresponds a positive 0 such that 11-'ll c 2 < 0 implies:
(A25.16)
n 8 n , n ~ O. This is a direct consequence of the

construction of the Yn , (3n (see Sinai theorem, Section 15), and of the
for any x, xl' x2 ( Xn
fact that the (3n's are C 1-close to the Yn's.
LEMMA B. A25.17
II ' is C 2 -close enough to ' then there exists a well-defined sheet
o(
'Y'
such that ,n
o is close to
nm for any n ~ O. To be precise,
there exists one and only one point Xo ( Xo such that 11;n xoll
<
E for
any n-::: O.
First we need a sublemma.
LEMMA A 25.18
Let R be a union of equi-dimensional Euclidean spaces R n , n
Let T
K + L: R
->
R, T IR n : R n
->
O.
Rn+ 1 be diffeomorphisms such that:
(1) K(O) = 0,
IIK(x) -K(y)11 > 0 Ilx - y II,
(2) IlL II $. E,
IIL(x) - L(y)11 <
Eilx - y II,
0 > 1,
0 - E > 1,
for any x, y (R. Then, there exists one and only one point x ( Ro such
that' the sequence Tnx is bounded, and
(A 25.19)
IiTnx II $. _ E _
0-1
for any n ~ O.
208
Proof:
The mapping T-11 R n : R n
->
R n- 1 (n > 1) is obviously a diffeomor-
phism. On the other hand:

II (Kx - Ky) + (Lx - Ly) II
ilTx- Tyll
2:. IIKx-Kyll-IILx-Lyll 2:. (8- E)llx-yll;

therefore:
(A 25.20)
Let bn(c) be the ball Ilx II ::; c of Rn' Since
IITxl1 = IIKx + Lxii 2:. IIKxll-IILxll > 811xll- E ,
we have:
(A25.21)
As:;ume c large enough, that is:
8c - E 2:. c.
(A25.22)
Then (A25.21) implies Tbn(c) ) bn+1(c), therefore

T-1bn+1(c) C bn(c) ,
consequently
T-1b1(c) ) T- 2 b 2(c) ) .. ) T-nbn(c) ) ....
But, according to (A 25 .20), we have:

diam~er
T-nb n (c) ::; 2c(8'- E)-n
->
0 as n
->
00
Therefore nn> 0 T-nbn(C) reduces to a unique point x (bo(c)'
This
finishes the proof if one observes that c = E/(8 -1) verifies (A 25.22) .
. Proof of Lemma B. A2S.23
According to (A 25.11) and (A 25.15), the mapping ; verifies the conditions of the preceding lemma. It is sufficient to set K
3' L
~ - 3' and change E into ce in condition (2). If we take ce = E(8 -1)
in (A25.19), we obtain 11;nxoll < E. This proves Lemma B.
209
APPENDIX 25
To summarize, we found a contracting sheet 0 (
'l:J' which remains
close to the orbit nm for n ::: 0 (in the sense of Lemma 8). If ' is
close enough to '
is close to nm, even for n
< O. To prove it, it
is sufficient to apply Lemma A t~ -1. The foliation
'!:I'
is the dilating
foliation of ,-1 and the sheet 0 is close to m. Consequently, according to Remark (A2S.10), the sheets mo, (n
< 0) stay in the neighborhood
of the orbit nm (in the sense of Lemma A):
Therefore, Lemmas A and 8 imply the following assertion:

ASSERTION A2S.24
If ' is C 2-close enough to ' then there exists a sheet ;5 C Bo of
the foliation '!:I~ such that the sheets
t 8 C Bn'
(-00
< n < 00), stay
inside an E-neighborhood of the center of Bn. Using the same argument

for ,-1, we find a sheet
f3
8
erties. Since the sheets
C Bo of the foliation ~; with similar prop-
and
fj are transverse in B o' there exists one

8 n fj in an E-neighborhood of the
and only one point of intersection z =
center of Bo. The desired homeomorphism k of the Anosov theorem is defined by setting k (m)
y,z. One easily checks that all the preceding con-
structions depend continuously on m. This proves that k is an homeomor:

phism. The relation 'k
to the identity.
= k
is obvious, as is the fact that k is E-close
APPENDIX 26
INTEGRABLE SYSTEMS
J. Liouville
(A26.1)
proved that if, in the system with n degrees of freepom:
aH
aq
= --,
q "" ap
aH ,
p = (Pl" .. , p ), q = (ql'' q ),
n
n first integrals in involution 1
(A26.2)
are known, then the system is integrable by quadratures.
Many examples of integrable problems of classical mechanics are
known. In all these examples the integrals (A 26.2) can be found. It was
pointed out long ago that, in these examples, the manifolds specified by
the equations F,
f; "" constant turn out to be tori, and motion along
them is quasi-periodic (compare with Example 1.2). We shall prove that

such a situation is unavoidable in any problem admitting single-valued
integrals (A 26.2). The proof is based on simple topological arguments.
THEOREM
A 26.3
Assume that the equations F;
f;
constant, i
1, ... , n, define an
n-dimensional compact, connected manifold M = Mf such that:

(1) at each point of M the gradients grad F; (i
1, ... , n) are linearly
independent;
1 Two funchons F(p, q) and G(p, q) are in involution if their Poisson bracket
vanishes idenhcally:
(F, G) = aF aG
ap aq
_ aF aG
aq ap
210
== 0
211
APPENDIX 26
(2) a Jacobian Det lal/af I, which is defined below (A 26.7) does not
vanish identicaIIy.
Then:
(1) M is an n-dimensional torus and the neighborhood of M is the direct product Tn
Rn;
(2) this neighborhood admits action-angle coordinates (I, , (I ( B n

C Rn , > (mod 2") ( Tn), such that the mapping I, > ... p, q is canonical 2
and Fi
F/I).
Thus, Equations (A 26.1) may be written:

I
1>
0,
w (I), where w (I)
aH ,
aI
and the motion on M is quasiperiodic since H = Fl = H(I) and Equations (A 26.1), in action-angle coordinates, are Hamiltonian equations 2
with corresponding Hamiltonian function H(J).
Proof:
NOTATIONS A
26.4
We use the following notations. Let x
= (p, q)
be a point of the phase
space R2n; we shall denote by grad F the vector gradient F
Xl
, ... , F
x2n
of
a function F(x). The Hamiltonian equations (A 26.1) then take the form:
(A26.5)
x=
1=(O-E)
E-O
I grad H,
where E is the unit matrix of order n. We introduce in R2n the skewscalar product of two vectors x, y:
[x, y] = (Ix, y) = -[y, x],

where ( , ) is the usual scalar product. As can be easily verified, [x, y]
expresses the sum of the areas of the projections of the parallelogram with
sides x, y onto the coordinate planes Piqi (i = 1, ... , n).
Linear transformations S, which preserve the skew-scalar product
[Sx, Sy}
--------------------See Appendix 32.
[x,
Jd
for all x, y,
212
are called symplectic. For instance. the transformation with matrix I is

symplectic. The skew-scalar product of the gradients [grad F, grad G] is
called the Poisson bracket (F, G) of the functions F, G. Obviously. F is
a first integral of the system (A 26.5) if and only if its Poisson bracket
(F, H) with the Hamiltonian vanishes identically. If the Poisson bracket
of two functions vanishes identically. the functions are said to be in invo-
lution.
THE CONSTRUCTION. A 26.6
Consider the n vector fields: ~,~ I grad ~, (i ~ 1 ... n). On account

of the nondegeneracy of I and the linear independence of the grad F;'s.
the vectors
~;
are linearly independent at each point of M.
Let us consider the system (A 26.5) with Hamiltonian F;. Since (F;,
F.) ~ O. all the functions F. are first integrals. and every orbit lies wholJ
lyon M. Therefore the velocity field

Finally. the fields
fJ
and
f,
~; ~ I
grad F, is tangent to M.
commute, for their Lie bracket is nothin~
but 3 the velocity field of the system (A 26.5) with Hamiltonian (F.,
F.)
~ O.
,
J
Thus, M is a connected. compact orbit of the group Rn acting smooth
ly and transitively; therefore we proved that M ~ Tn. Besides. M being
specified by the equations F;
f; ~ constant. the fields grad F; define a
structure of direct product in the neighborhood of M.

Now. let us choose the torus Mf: F
f in the neighborhood of M and
consider the n integrals

(A26.7)
over the basic cycles y /f) of the torus Mf" Since the l;<f) are functionally independent. the equation can be solved in the neighborhood of M and
yields the torus M([) ~ Mf(I)' which corresponds to a given I. Let us set:
The Lie bracket of the Hamiltonian vector fields I grad F
and I grad G
IS
an
Hamiltonian vector field with Hamiltonian function - (F, G). We shift the proof to
Appendix 32.
APPENDIX 26
S([, q)
(A26.8)
jq
213
pdq
'lo
where the path of integration lies on M([) (therefore, p
p ([, q) ).
The
many-valued function S is the generating function (see Appendix 32) of the

canonical transformation f,
1> ->
p, q, which defines action-angle coordi-
nates:
p
(A26.9)
LEMMA
as
as
oq
af
A26.10
The one-form pdq of M([) is closed.

Proof:
It is sufficient to prove that the integral of pdq along infinitely small
parallelograms lying in M([) vanishes. If D is a parallelogram with sides

~,.,." then ~ pdq (i.e., the sum of the areas of the projections of D onto
the coordinate planes P; q;, i = 1, ... , n) is the skew-product [~,.,.,] of ~

and .,.,. Suppose now that ~ and .,., touch M([) at a certain point. In accordance with (A 26.6) any vector tangential to M ([) is a linear combination of the n vectors f grad F;. But these vectors are skew-orthogonal
since, in accordance with (A 26.2),
[grad F , grad F.]
0 ,
and thus, since f is symplectic,
[! grad F;, f grad Fj J
O.
Therefore [~,.,.,] ~O, as required. The integral (A 26.8) can therefore be

regarded as a many-valued function S and Equations (A 26.9) define, locally, a canonical transformation f,
ACTION-ANGLE VARIABLES A
1>
->
p, q.
26.11
In fact formulas (A 26.9) define a global canonical mapping in which p

and q have period 211 with respect to
1>.
To prove it, we observe that, for
214
every I, the differential of S([, q) is a global one-form on M([).

fore d, as defined by (A 26.9), is also a global
There-
one~form.
Let us compute the periods of the one-forms d, over the basic cycles
of the torus Mf' According to (A 26.7) we have:
Yi
di
f d(aS)
al.
Yj'
~
d/.
'Yj
dS
~(2771.)
= 2770 ..
d / . } '}
,
Therefore the variables , are angular coordinates on the torus M(l) and
our theorem is proved.
APPENDIX 27
SYMPLECTIC LINEAR MAPPINGS OF PLANE

Let A be a symplectic linear mapping of the plane (p, q). The mapping
A preserves the area-element dp 1\ dq, therefore Det A
1. Consequent-
ly, the product of the proper values Al and A2 of A is equal to 1. Besides, Al and A2 are roots of the real polynomial Det (A -AE). Therefore,
either Al and A2 are both real, or they are complex conjugate: A2 = Xl .
In the first case, we have:
(A27.1)
Figure A 27.3
215
216
In the second case, we have:

(A 27.2)
and the roots belong to the unit circle (see Figure A 27.3). The third and
last possible proper value coniiAuration is:
hyperbolic rotation
hyperbolic rotation
with reflection
Figure A27.5
217
APPENDIX 27
EXAMPLE
A27.4
The hyperbolic rotation:

p, q
2p, Yz q ,
-+
or the hyperbolic rotation with reElection:

p, q
In both cases the orbit Tnx of x
2p, - Yz q (see Figure A 27.5).
-+ -
(p, q) belongs to the hyperbola pq
constant. Of course, the fixed point 0 is unstable. From classical theorems of linear algebra it follows that every mapping A of the first type
(Ai of, A2 ; Ai' A2 ( R) is an hyperbolic rotafion, possibly with reflection.
This means that, up to a suitable change of variables, A may be written
under the form:

P, Q
EXAMPLE
->
1 Q
AP, X
A27.6
A r9tation through an angle a belongs to the second class (Ai

A2 = ei~:
p, q
-+
p cos a- q sin a, p sin a + q cos a .
x
p
Figure A27.7
e- ia,
218
This rotation transforms into an "elliptic rotation" (see Figure A27.7) under a linear change of ,,:ariables. In this case the orbit r'x of x = (p, q)
belongs to an ellipse centered at O. The fixed point 0 is obviously stable. Classical theorems of linear algebra show that every mapping A of
the second type (1'\11 = 1'\21 = 1, '\1';' '\2) is an elliptic rotation.
In the first case (A 27.1), the fixed point 0 is called an hyperbolic
point and one says that A is hyperbolic at O. In the second case (A 27.2)
the fixed point 0 is called an eIIiptic point and one says that A is eIIiptic at O. Finally, the third case (,\ 2 = 1) is called the parabolic case.
REMARK
A27.8
Every canonical mapping A ~ close enough to an eIIiptic mapping A, is

eIIiptic. In fact, the roots '\1 and '\2 depend continuously on A and are
restricted to lie either on the real axis or on the unit circle (see Figure
A 27.3). Therefore, these roots cannot leave the unit circle, except at
points ,\ = :!::. 1 which correspond to the parabolic case. Finally, we define
the topological index of a vector field at a fixed point. Let us consider a
vector field ';(x) of the plane p, q, with an isolated fixed point ';(0)
This defines a mapping of the unit circle x2
= p2
O.
= 1 onto itself:
If E is small enough, then the topological degree of this mapping does not
depend on E and is called the index of .; at 0, or the index of O.
Now, consider the vector field ';(x) = Ax - x.
If the mapping 1\ is
nonparabolic, then 0 is an isolated fixed point of ';(x).

THEOREM
A27.9
An eIIiptic point, or an hyperbolic point with reflection, has index + 1.

An hyperbolic point has index - 1.
Proof consists in a mere inspection of Figures (A27.S) and CA27.7).
APPENDIX 28
STABILITY OF THE FIXED POINTS

Consider an analytical canonical mapping A of the plane p, q, with
(0,0). Assume that 0 is elliptic, that is, that the differential of A at zero has proper values A1 = e- 1a , A2 ~ e ia . It has been
fixed point 0
known since G. D. Birkhoff's time 1 that, if al2rr is irrational, to every

s > 0 corresponds a canonical mapping B = B(s) of a neighborhood of 0
B: p, q
->
P, Q,
B (0)
0,
which reduces A to a "normal form":
that is as follows. Let I, be the canonical polar coordinates:

21 = p2 + Q2,
2/'= p'2 + Q'2,
arctg(P/Q)
'
arctg(P'/Q')
then:
I'-I
(A28.1)
'-
The coefficients a, a 1 ,
= O(ls+1)
a+
a1 i+ ai 2 + .. + a/ s +
...
do not depend on the mapping R (~) by which A
is reduced to the form A '. If a
~ 2rrm/n
0([S+1).
and if there is a nonvanishing
coefficient a 1 , a 2 , ... , Birkhoff says that A is of "generic elliptic type."

1 Dynamical Systems, Chapter 3.
719
220
THEOREM A 28.2
(See Arnold [7]).
The fixed point of a
~eneric
elliptic mapping is stable.
The proof consists in applying the construction of Theorem (21.11) of

Chapter 4 (see Appendix 34) to the mapping (A 28.1): for 1 I, 0(18+1)
is regarded as a perturbation of the mapping
l'
I,
Similar theorems are obtained concerning the stability of equilibrium positions and elliptic periodic solutions of Hamiltonian systems with two degrees of freedom (see Arnold [7]).
J. Moser [1]
obtained the strongest re-
sult in that way:

MOSER'S THEOREM A28.3
The fixed point of an elliptic canonical mapping A of the plane is stable provided that:
(1) a f, 217
j!- ,
217 :
(2) a 1 f, 0 ;
(3) A is C333 -differentiable. (As it is pointed out in a recent paper

of Moser [6], this number of derivatives can be fairly well reduced.) A complete proof will be found in
J. Moser [1].
REMARK A 28.4
If a
Civita [1].
217m/3, th~n the fixed point can be unstable, as shown by Levi-
APPENDIX 29
PARAMETRIC RESONANCES
The analysis of the stability of the fixed point (0,0) of a linear mapping of the plane is due to Poincare and Lyapounov. Only in recent times
(1950), were these results extended by M. G. Krein [11. [2] to systems with
many degrees of freedom. Krein's investigations have been enlarged by
Jacoubovich [11. Gelfand and Lidskii [41. and so on. J. Moser [3] published
a report of Krein's theorem.
Let A be a linear symplectic mapping 1 of the canonical space R2n.
We say that A is stable if the sequence An is bvunded. We say that A is
parametrically stable if every symplectic mapping, close to A, is
st~ble.
We proved in Appendix 27 (and used it in Section 20, Chapter 4) that every

elliptic mapping of R2 is parametrically stable. M. G. Krein displayed all
the parametrically stable mappings of R2n.
LEMMA A 29.1 (Poincare-Lyapounov)
Suppose A is a symplectic mapping and ,\, is a proper value of A.
Then 1/,\"
X, and I/X
are proper values of A.
A preserves the skew-scalar product
product and 1 =
(~=~), E
[e-,." 1 = (1
e-, .,,), where (
, ) is the inner
= unit matrix of order n. Therefore, we have:
[Ae-. A."l
[e-. ."l
221
and
A iA
= 1.
222
ERGODIC PROBLEMS OF CLASSICA.L MECHANICS
Prool:
It is sufficient to prove that the characteristic polynomial of' A is real
and reciprocal. In fact, we have:

p(..\)
= Det(A-..\E) =
Det(-lA,-1 1 + ..\p)
= Ded-A '-1 + ..\E) = Det(-A- 1 + ..\E)

= Ded-E + ..\A) = ..\2n. Det (A_..\-1 E) = ..\2n. p(..\-I)
From thi~ lemma the following corollary is readily deduced;

COROLLARY A29.2
The proper values of A divide into couples and "quadruples." Couples

are lormed by ..\
~nd
..\-1, ..\ belonging to the real axis or the unit circle:
1..\1 = 1. Quadruples are formed by ..\,X,..\-I, and X-I (see Figure A 29.3) .
1.3
I
I
\ I
\I
~.l
\I'
,.
'~'I
;\3
\
\
Flgure A 29.3
223
APPENDIX 29
COROLLAR Y
A 29.4
If the proper values are simple and lie on the unit circle IAI = 1, then
A is parametricalIy stable, because if all the proper values are simple and
lie on IAI
= 1,
then:
(1) A is stable (for obvious reasons of normal form);

(2) all the proper values of a symplectic mapping A', close enough
to A, lie on IAI = 1. In fact, assume the contrary, then A' would have'
two proper values A and X-I close to a unique isolated proper value of A.
Let us now assume definitively that :i 1 are not proper values of A.
Krein classified the proper values belonging to the unit circle IAI = 1:
they split into positive and negative proper values. First assume that all
the proper values are simple; we prove the following lemma:
LEMMA A29.5
Let ~1 and ~2 be the proper vectors with corresponding proper values
Al and A2 Then, either AIA2
1, or [~1' ~2] = O.
Proof:
Since A ~1
Al ~1 and A~2 = A2~2' we have:
[A~I' A~2]
Al A2 [~1' ~2]
[~1' ~2]
(Q. E. D.)
COROLLAR y A
Let
(J
29.6
be a plane, invariant under A and correspollding to conjugate
proper values AI' A2, IAll = IA21 = 1. Then:

(1) (J is skew-orthogonal to every proper vector ~3 corresponding to
another proper value A3 ;
(2) the skew-product [~,
1)]
of noncolinear vectors ~ and
1)
of
(J
IS
nonvanishing.
Assertion (1) is a direct consequence of AIA3 ~. 1, A/\3 ~ 1: in view
224
of Lemma (A 29.5) we have [~l' ~3]

We have also [~1' ~1]
= 0,
[~2' ~3] = O. Suppose [~1' ~2] = O.
and assertion (1) implies [~1' ~3]
= 0 for
every
~3' Therefore [~1' 71] = 0 for every 71, which is impossible. Consequent-
ly [~1' ~2] f, 0 and assertion (2) holds good.

DEFINITION
A29.7
A proper value A, such that IAI
1, A2 f, 1, is called a positive (resp.
negative) proper value of A if: [A~,~] > 0 (resp.
< 0) for every
~ of the
real invariant plane a corresponding to the proper values A and X

This definition is correct. Indeed, the vectors
A~
and
of a are non-
co:near because A2 f. O. Therefore, in view of Corollary (A 29.6), [A~,~]
f. 0 on
Consequently [A~,~] has constant sign for every ~ (a.
a.
REMARK A
29.8
The sign of a proper value lias a simple geometrical meaning. The plane
a
admits a canonical orientation, for [~, 71] f. 0 if ~ is nonparallel to 71'
Therefore, oile may speak of positive (or negative) rotations. The restriction of A to a is an elliptic rotation through an angle a, 0
< lal < 7T. The
proper value A is positive (resp. negative) if A rotates a through a posi-
tive (resp. negative) angle.

Krein's main result is: collision of two proper values with identical
signs on the unit circle IAI
1 does not provoke instability. In contrast,
two proper values with opposite signs can leave the unit circle after they
have collided, so forming a "quadruple" with their conjugates (see Figure
A 29.3).
To be p~ecise, let A(t) be a symplectic mapping -.yliich depends continuouslyon a parameter t, and the proper values of which are different
from :. 1 if It I <
A are simple
collide for t
THEOREM
an~
T.
Suppose that, for t < 0, all the proper values Ak of
lie on the unit circle, while certain of these proper values
O.
A 29.9
If all the proper values that collide have identical sign, then they re-
225
APPENDIX 29
main on the unit circle after the collision and the mappinl1 A remains sta-
ble for t < E, E > O.

We shall prove the theorem in the simplest case in which all the proper
values A, 1A > 0 collide. The general case can be reduced to this c'ase
by selecting a canonical subspace R21(t) corresponding to th'e I colliding
proper values and their conjugates. 1'0 fix the ideas, suppose that the proper values, Ak are positive:
[Ae-. e-] > 0 for e- ( Uk '

where Uk is the plane generated by e-k , ~k (Ae-k = .Ake-k).
Proof of the Theorem. A29.10

Consider the quadratic form [A e-. e-]; its polar bilinear form is nondegenerate. We have, indeed:
[Ae-. 1/] + [A1/. e-]

Suppose [(A-A- 1 )e-. 1/]
(A 2 - E) Ae-
[Ae-. 1/] - [A -1 e-. 1/]
[(A - A-I) g. 1/]
0 for every 1/, then (A-A- 1 )e-
0 and
O. Thus, 1 would be a propervalue of A2. which contr"dicts
the condition of Theorem (A 29.9) (A
del1enerate for It
I<T
i' :t 1). Therefore [A(t} e-, e-] is non-
and, in particular, for t
O. On the other hand,
this form is positive definite for t < O. In fact, every vector 1/ is equal
to the sum of its projections 1/k into the invariant planes Uk corresponding to the proper val,ues Ak
Xk
According to Lemma (A 29. 6) these planes
Uk are skew-orthogonal, therefore: .
[A1/. 1/]
=I
[A1/k' 1/1]
k,l
because
But [A1/k' 1/k] > 0 for Ak is a positive proper value, hence:
[A1/. 1/] >
t) .-
So, the form [A (t ) e-, e-] is positive definite for t < 0 and nondel1enerate
226
for t = O. Therefore, this form is positive definite for t

ly, for t
< E, E> O. But [A An~,
An~] = [A~. ~l for
Thus the orbit An ~ lJelongs to the ellipsoid [A~.~]

A(t) is stable for t
0 and, consequent.
An is symplectic.
=
< E.
constant, that is
(Q. E. D.)
RE:'IARK A 29.11
The above argument proves the criterion of parametric stability:
The symplectic mapping A is parametrically stable if and only if all the

proper values
>'k lie
on the unit circle
1.\1
1,
>.;
1. Besides, the
quadratic form [A~.~] is definite on every invariant subspace correspg.nd.

ing ." the multiple proper values t.. k
Xk
APPENDIX 30
THE AVERAGING METHOD FOR PERIODIC SYSTEMS

Let
n=
8 1 )( SI be the phase space, where 8 1 =
an open bounded subset of RI and SI

sider -periodic smooth functions
F:
n ...
RI,
f:
CI)
II =
([1' ... ,11)\ is
I (mod 21T)} is a circle. We con-
(J), F (I, ), HI, ):
n ...
RI,
CI):
8 1 ... RI ,
and finally, E I denotes a small parameter.

THEOREM
A30.I
n:
We consider the systems defined in

(A 30.2)
, ~ =
CI) ( [ )
1I
E F ([, )
+ Et([, )
and
(A30.3)
= E
F(j), where F(j)
---.L
21T
If
CI) ( [ )
,;,
0 in
n,
21TF (j, ) d .
then the solutions I (t) and ] (t) of (A 30.2) and (A 30.3),
with equal initial data [(0)

1[(t)-](t)1
= ]
(0), satisfy 1 :
< CE for every t, O:s: t ::: liE,
where C is a constant which does not depend on E,
We suppose that J(I) ( 8' for every t, 0::: t
227
:S
l/f .
228
Proof:
Let us improve (A 30.2) by using a new variable:

(A 30.4)
P ~ P(I, ) ~ 1+ Eg(J, ),
From (A 30.2) and (A 30.4) follows:

(A 30.5)
In order to cancel the terms of order E, we set:

(A30.6)
g(I,) ~
F(P)-F(P,) d;
cu (P)
this expression is well-defined since

cu(p) .;, 0,
2TT
(ii - F)d
0,
and therefore g( + 2TT) ~ g(). Now, our system (A30.S) may be written:
(A 30.7)
Let P(t) be the solution of (A 30.7) with initial data P(O) ~ J(O)
(A30.8)
= /(O):
p(t) " P([(t), (t.
Obviously, (A 30.7) implies:

(A 30.9)
IP(t)-J(t)1
< Cl'
for every t, 0 $. t $. liE.
Finally, from (A 30.4), (A30.6) and (A 30.8) follows:

(A30.10)
IP(t}-J(t}!
< C2
for every t.
Inequalities (A30.9) and (A 30.10) conclude the proof. They prove also
that the motion decomposes into the averaged motion and fast small oscillations (see Figure A30.U).
229
APPENDIX 30
I (t )
"-="_- J(t)
t=o
Figure A30.1l
APPENDIX 31
SURFACES OF SECTION
(See Section 21. 9, Chapter 4)
Let H(p, q) be the Hamiltonian function of a system with n degrees of
freedom (therefore the phase space is 2n-dimensional). Let L: H = h,
q1 =
be a (2n - 2) -dimensional submanifold of the "level of energy"
H = h. If, in a certain domain Lo of L, P = (P2' ... , Pn)' Q = (q2' ... , qn)

forrt: a local chart and q1 10 0, L is called a surface of section (see Figure A31.1). Assume that an orbit of the Hamiltonian system, through a
------;...... y
,I
,
I
----__\r-"
Figure A 31.1
point x of
a point
L o' returns to Lo' Then, in view of q1 10,0, the orbit through
x: on Lo sufficiently close to x, will, as

230
t increases, return .to
APPENDIX 31
231
Lo and will cut Lo in a uniquely determined point Ax '. In this manner,

we define a mapping A:
THEOREM 1
A 31.2
The mapping A is canonical, that is for every closed curve y of L l ,

we have:
(A 31.3)
PdQ =
where PdQ
PdQ,
Ay
P2 dq2 + ... + Pn dqn .
Proof:
Consider the orbits emanating from y in the (2n + 1) -dimensional space
I(p, q, t)l. The curves y and Ay of the space I(p, q)l are the projections of
two closed curves y'and Ay'of !(p, q,Ol which are formed respectively,
q
p
Figure A 31.4
by the initial points (t
0) and the end points of tbe above orbits (see Fig-
ure A31.4). Therefore, we have by the

1
Poinc~re-Cartan
theorem:
A proof of this well-known theorem has apparently never been published.
232
L pdq- Hdt
(A 31.5)
~,
where pdq
pdq- Hdt ,
Ay'
PI dql + . + Pn dqn But H
constant along y' and Ay'
then:
.! Hdt
~y'
o.
Besides. we have:
pdq
Ay'
pdq.
Ay
constant on 1. we also have:
In view of ql
Thus. finally.
,pdq-Hdt = jPd Q,
pdq-Hdt
Ay'
1.
PdQ.
Ay
and (A 31.5) implies (A 31.3). This proves the theorem.

EXAMPLE A
31.6
Consider the problem of the "convex billiard table" (Birkhoff). Let
be a closed convex curve of the plane E2. Suppose that a material point
N
Figure A31.7
233
APPENDIX 31
M moves inside r and collides with r according to the law "the angle of
incidence is equal to the angle of reflection" (see Figure A 31. 7).
The
states of M, immediately before aAd immediately after a reflectim, are determined by the angle of incidence a, 0 :::; a :::; 217, and the point of incidence. The point of incidence A is defined by the algebraic length q2 of
the arc OA of r (0 is an arbitrary origin). In other words, the set of the
states of M, immediately before and immediately after a reflection, form a
torus
r2
la(mod 217), q2 (mod L)! in the phase space (L is the length
r2
of r). We obtain naturally a mapping A of a subset of
info another
one: the state which immediately follows a reflection is transformed into

the state immediately preceding the next reflection.
THEOREM A 31.8 (G. D. Birkhoff)
sin a d q2 /\ da is invariant under the mapping A .
Proof:
Between two reflections, the motion of M is determined by Hamiltonian

equations in the corresponding four-dimensional phase-space. In the neighborhood of the above torus
r2, let us select special coordinates. Our point
M is well-defined by coordinates (ql' q2)' where ql
MN is the distance
from M to rand q2 is the algebraic length of the arc ON.

Coordinates ql and q2 (mod L) .are clearly Lagrangian coordinates in
a neighborhood of r. Let Pl and P2 be the corresponding momenta (the
mass of M is supposed to be 1). On r, P l and P2 coincide obviously
with the components of the velocity-vector v:
Pl
= Ivl
sin a,
P2
= Ivl
cos a.
The Hamiltonian function H is the kinetic energy:

H
v2
=""2
In the four-dimensional space I(P l , P2 ' ql' q2)~' consider the surface ~
whose equation is:
1, M ( r) :
234
From one reflection to the next, the motion defines a mapping :\: L
->
L.
The coordinates P2' q2 are local coordinates of L (a 1= 0) which is a Surface of section. In view of Theorem (A 31.2), the mapping t\ is canonical
and therefore preserves the two-form:
dP 2 /\ dq2 = sin a' dq2 /\ da .
(Q. E. D.)
An elementary proof of this theorem, due to G. D. Birkhoff [1], requires ex~
tensive computations.
APPENDIX 32
FUNCTIO~S
TilE GENERATI:\G
OF
"APPI~GS
CANONICAL

The following results are due to Hamilton and Jacobi.
A. Finite Canonical Mappings
Let x ~ (p, q), (p ~ (Pi' ... , Pn)' q ~ (ql' ... , qn))' be a point of the
.canonical space R2n. The differentiable mapping:
is
call~d
canonical if A preserves the Poincare integral-invariant:
(A 32.1)
pdq
1.
pdq,
Ay
fOf- any closed curve y. Let a be an arbitrary two-chain. Relation (A 32.1)

lmplies that A preserves the sum of the areas of the projections of a 'into
the coordinate planes Pj' qj:
(A 32.2)
[(a)
ff
dp 1\ dq
ff
dp 1\ dq
[(Aa) .
Aa
In other words, the two forms dp 1\ dq and dP 1\ dQ coincide:

(A 32.3)
dp 1\ dq = dP 1\ dQ, where P = P(p, q), Q = Q(p, q) .
If the domain of A is simply connected, then conditions (A 32.1) and

(A 32.2) are equivalent. Relation (A 32.3) shows that:
235
236
pdq + QdP, where P
P(p, q),
is a closed form of R2n (since dp 1\ dq + dQ 1\ dP
A(x)
Q(p, q),
=
0).
Therefore:
x pdq + QdP, where P = P(p, q), Q = Q(p, q),

Xo
defines, locally, a function on R2n. Suppose that ql' .. , qn; PI'' Pn ,

form a local chart in some neighborhood of the point x, that is:
Det(
~:) ~
O.
Then, ~(x) can be regarded as a function of P, q, defined in the neighborhood of the point P, q:
(A 32.4)
A(P, q)
(P, q)
pdq + QdP, where p
p(P, q),
Q(P, q).
DEFINITION A 32.5
The function A(P, q) is called the generating function of the canonical

mapping A.
Of course, A is only defined locally and up to a constant. From (A32.4)
follows:
(A32.6)
LEMMA
Q,
aA
aq
= p.
A32.7
Let A (P, q) be a function such that:
in the
nei~hborhood
of a point (P, q). Then, Equations (A 32.6) can be
solved locally with respect to P and Q:
= P(p,
q), Q
= Q(p, q),
and the fuactions P, Q determine a canonical
mappin~
A.
237
APPENDIX 32
In fact, pdq + QdP is a closed form on R2n; thus

dp 1\ dq = dP 1\ dQ .
(Q. E. D.)
Unfortunately, the generating fur-:-tion A is not a geometric object:
A not only depends on the mapping A, but also on the coordinates p, q of
R2n.
According to (A 32.6), the generating function of the identity 1 is Pq.

Thus, every canonical mapping, close enough to the identity, has a generating function close to Pq.
R. Infinitesimal Canonieal Mappings
Consider a family of canonical mappings SE' the generating functions

Pq + ES(P, q; E) of which depend smoothly on a parameter E 1. The
mapping
Se is close to the identity, if E is small. According to (A 32. 6),
the Taylor expansions of P(p, q) and Q(p, q) with respect to E are:

(A 32.8)
where
S
S(p, q; E) ,
By definition, the infinitesimal canonical mapping Se is a class of equivalent families Se: two families Se and SE' are equivalent if ISE- Se' I
O(E 2 ),
DEFINITION
A32.9
The function S(p, q) on the phase-space is called the generating function of the infinitesimal mapping SE (or Hamiltonian function).
Of course, S is defined up to a constant. Now we prove that the function S is a geometric object: S neither depends on the canonical coordinates p, q, nor on the choice of a representant SE in the class of equivalence: it is a mapping S: R 2n ... Rl. In fact, let y be a curve of R2n
1 ThiS is a way to memorize (A 32.6).
238
joining x and y:
ay = y - x. We set
Ye = Se Y ,
and we denote by a(E) the strip form-
ed by the curves Ye" 0 < E' < E, and

oriented in such a way that
aoe
Y - Ye + (see Figure A32.10).
Figure A32.10
Let us set:
(A 32.11)
![a(E)]
11
dp/\dq.
a(e)
According to (A 32.2), this integral does not depend on the canonical coordinates and, according to (A 32.1), does not depend on the curve Y, but
only on x and y.
LEMMA A32.12
The generating function S. of the infinitesimal canonical mapping Se
is given by:
(A32.13)
S(y) - S(x) =
![a(E)]1
dE
'
e=O
and does not depend on the choice of the canonical coordinates p, q.

Proof:
Let us set Se x (A32.14)
ox
= (op,oq).
E(S(y) -S(x)) = E
According to (A 32.8), we have:
f~
yap
dp +
dq
aq
(oqdp-opdq) + 9(E2)
On the other hand, according to (A32.11), the integral of dp /\ dq along

a(E) is;
(A32.1S)
239
APPENDIX 32
Formulas (A32.14) and (A32.1S) imply (A32.13).
(Q. E. D.)
One can express the invariance of the generating function S in another

form. Let -\ be a finite canonical mapping and SE an infinitesimal canonical mapping. The canonical mapping T E = AS E A-I is clearly infinitesimal.
LEMMA
A32.16
The generating functions Sand T of the infinitesimal mappings SE

and TE are related by:
T(Ax)
(A 32.17)
S (x) + constant.
Proof;
Let YE and a (E) be the curve and the surface of Lemma (A 32.12).
The curve y' = Ay joins the points Ax and Ay. Besides, the curves
TE ,y', 0
~ E' ~ E,
form a strip T(E), which is nothir.g but:
(A 32.18)
From (A 32.13) follows:
(A32.19)
S(y)-S(x) = ~ J[a(E)),
dE
T(Ay)-T(Ax) = ~[[r(E)].
dE
But the mapping A is canonical. Thus, according to (A32.2) and (A 32.18),

we have:
[[o(E)] = [[r(E)].
Comparison with (A 32.19) yields (A 32.17).
(Q. E. D.)
COROLLARY A32.20
Let '\ and CE be infinitesimal canonical mappings with corresponding generating functions Band C, and let ,\ be a finite canonical mapping . Then, the infinitesimal canonical mapping:
(A 32.21)
has the following generating function:
240
(A 32.22)
B '(x) =
c (x)
+ B (x) - C (A -1 X) + constant.
In fact, (A 32.8) implies that the generating function of the product of two
infinitesimal mappings is the sum of their generating functions, and also
that the generating function of the inverse mapping C;-1 is - C. Relation
(A32.22) is easily derived from these remarks and from Lemma (A 32.16).
C. Lie Commutators and Poisson Brackets
Given two infinitesimal canonical mappings Ae and Be' there exists
one and only one infinitesimal canonical mapping Ce such that:
(A32.23)
AaBbA_aB_b
Cab + O(a 2 ) + O(b 2 ); a, b ... O.
The mapping Ce is called the Lie commutator of Ae and Be .

LEMMA
A 32.24.
The generating function
of
Ce
is equal, up to a sign, to the Pois-
son bracket of the generating functions A and B of Ae and Be :
(A32.2S)
'lie
[V A, V Bl.
gradient.
We use the notation [x, y] = (lx, y) as in Appendixes 26 and 27.
Proof:
Again let Y be a curve joining x and y:
ay
y - x. We consider the
five-sided prism (see Figure A 32.26) formed by four strips:

a1
BeY' -b < E < 0,
aa 1
Y - Y1 + ''',
a2
\y~. -a<E<O,
oa2
Y1 - Y2 + "',
a3
BeY2'
0 < E < b,
aa3
Y2 - Y3 + ''',
a4
.'\ Y3'
0< E < a,
aa4
Y3- Y4+"',
and closed by a fifth strip as' formed by the
~egments
joining the corre-
sponding points of Y and Y4' aas = Y4 - Y ... Finally, we denote by

T x and T y the bases of our prism, such that the two-chain
a 1 +a2+a3+a4+Us+Ty+Tx
241
APPENDIX 32
Figure A 32.26
forms a two-cycle homologous to zero. Since dp /\ dq is a closed form we

have, with notations of (A 32.2):
[(Ul) + [(u 2 ) + [(u3 ) + [(u 4 ) + [(us)
(A 32.27)
+ [(7 y) + [(7)
[(~)
0 .
But Lemma (A32.12) implies:
(A32.28)
1(ul )
- b[B(y) - B(x)] + 0(b 2 ),
[(u2 )
[(u 3 )
- a[A(Yl)-A(x l )] + 0(a 2 ) ,
b[B(Y2)-B(x2 )] + 0(b 2 ),
[(u 4 )
a[A(Y3)-A(x 3 )] + 0(a2 ) ,
[(us)
-ab[C(y)-C(x)] + 0(a 2 ) + 0(b 2 )
On the other hand, 1y - y 41

along the surfaces
(A 32.29)
0 (ab); therefore, the integrals of ~ dp " dq
are given by:
[(7 )
--ab[V B(y), VA(y)] + 0(a2 ) + 0(b 2 ),
[(7) = + ab[V B(x), V A (x)] + 0(a2 ) + 0(b 2 )
Finally, from (A 32.8) itO follows that the vector fields corresponding to Ae
and Be are 1 V A and 1 V B. Consequently, up to terms of order 0 (a 2 + b2 ),
we Have:
A(Y3)-A(Yl) = (VA, Y3- Yl) = (VA, Y3- Y2) + (VA, Y2- Yl).
=
{V A, 1 V B)b - (V A, 1 VA)a =
= {VB, VA]b -
[VA, VA]a = [VB, VA]b.
242
Similarly:
- [V A, V B]a .
Thus, we deduce from (A 32.28):
(A 32.30)
Comparison of (A 32.27), (A 32.29), and (A 32.30) yields (A 32.25).
(Q. E. D.)
APPENDIX 33
GLOBAL
CANO~ICAL
\lAPPINGS

This appendix gives the topological reasons for the existence of periodic orbits for Hamiltonian systems with n degrees of freedom.
A. Generating Functions
n = B n x Tn be the canonical space, p: n B n C Rn

and q: n
T" = j(ql, ... ,qn )(mod21T)! thecoordi~ates p
Let
p)1
n
->
->
and q
= q(x)
(n.
of the point x
/Pl' ... ,
= p(x)
By definition, the mapping A:
->
is globally canonical if it is homotopic to the identity and satisfies
(A33.1)
pdq =
Y
pdq
Ay
for anyone-cycle y (even nonhomologous to zero). In conformity to Appendix 32, the mapping A is locally determined by a generating function
Pq + A (P, q), provided that
Det (~: )
0,
(A33.2)
p
=P
aA ,
aq
aA
= q+ - ,
ap
Thus, locally, the function A (P, q) verifies:

(A33.3)
A (P, q) =
(P, q)
(Q- q)dP + (p-P)dq.
243
E~GOD/C PROBLEMS OF CLASSICAL MECHANICS
244
Let us set:
A(x) = A(P(x), q(x)) ,
where
x
LEMMA
(p(x), q(x)) (
n.
A33.4
The mapping (A 33.2) is globally canonicai if and only if the function

A (x), defined by (A 33.3), is single-valued on
n.
Proof:
Let
y be a
closed curve of
(A33.S)
n.
Let us prove that:
f<Q-q)dP+(P-P)d q =
y
o.
In fact, (A 33.1) is equivalent to:
(A33.6)
pdq =
PdQ:
Thus, we obtain:
(Q- q)dP + (p-P)dq
QdP + PdQ-(qdP + Pdq)
fd
[P(Q-:- q)].
But the increment of P (Q - q) along y is equal to zero, because A is

homotopic to zero:
(A33.7)
d[P'(Q-q)] = O.
Y
Conversely, (A33.S) and (A33.7) yield (A33.6).
(Q. E. D.)
B. A Topological Lemma
Now, let A be a globally canonical diffeomorphism, T the torus, p =

0, and AT the image of T by A.
APPENDIX 33
LEMMA
245
A33.8
The tori T and AT have at least 2n common points (counted with their
order 01 multiplicity) provided that the equation 01 AT
(A33.9)
IS:
p = p(q),
Besides. the number 01 geometrically different points 01 intersection is, at

feast, n + 1.
Prool:
Consider the following function on AT:
I(x) =
(A33.10)
p(x)dq(x) ,
Kij
where x
(p(x), q(x))
AT, and where the path of integration lies on AT.
Function I(x) is well-defia.ed since the integral (A33.10) does not depend
on the path. In fact, let y be a closed curve of AT. then:
pdq =
1.
pdq = 0,
A-'y
because A- 1 'is globallycanonical, A- 1 y C T, and p
0 on T. The
function f(x) is a smooth function on the n-dimensional torus Tn. Thus,

by the Morse' inequalities, the number of critical points is at least 2n
(and, following the Lusternic-Schnirelman theorem, the number of geometrically distinct points of intersection is at least the Lusternic-Schnirelman
category of Tn, which is n + 1).
From (A 33.10) follows dl = p (x) . dq(x) on AT. The function p (x)
vanishes at the intersection of AT and T. Thus, the points of intersection of T and AT are critical points of I on AT. Conversely, in view of
condition (1\ 33.9), we have pdq
of I on AT; therefore p(x)
0, for any dq, at every critical point x
0 and the critical point x belongs to the in-
tersection of AT and T.
,
See Milnor [1]; 2 n
= ~o
(Q. E. D.)
bl' b,
= i-th
Beth number of 'r'.
246
COROLLARY
A33.11
The tori T and AT have at least 2 n common points, provided that

their equations are:
p = p'(q), p =
(A33.12)
p"(q)(I~:'1
<
00, I~'I
<
00)
and that dp 1\ dq vanishes on T.
In fact, if dp 1\ dq
:=
0 on T, then the mapping

p, q
-+
p- p'(q), q
is a canonical diffeomorphism. This diffeomorphism reduces (A 33.12) to

(A 33.9) with
p(q)
REMARK
If n
p"(q).-p'(q).
A33.13
1 (mappings of annuli), Lemma (A 33.8) still holds without con-
dition (A33.9). The proof makes use of Jordan's theorem and does not extend for n > 1. Whether T and AT intersect, for n> 1, if condition (A33.9)
is not fulfilled, is an open question.
If condition (A 33.9) can be relaxed from Lemma (A 33.8), we ob1ain

many "recurrence theorems" of the following type:
Assume that the initial values a"
b i of the axis of the Kepler ellipses,
in the plane many-body problem, are such that the ellipses do not intersect.
Then, whatever
be, there exist initial phases 2 1., g. such that the axis
"
of the ellipses return to their initial values after a time

REMARK
T.
A 33.14
If we drop condition (A 33.9), Lemma (A 33.8) cannot hold without assuming that i\ is a diffeomorphism, because (even for n
1) regular and
globally canonical mappings can be constructed such that T and AT do

not intersect.
2 Phases 1/. ~/ are angles (mod 271); ~/ determines the position of the ellipses
and 1/ determmes the position of the planets on these ellipses.
247
APPENDIX 33
C. Fixed Points
Now, let A be a global canonical mapping of the following particular

type:
(A33.1S)
A:
p., q -+ p, q <+
(p)
Assum~ that; on the torus p = Po' all the frequencie~ are commensurate:
(A33.16)
and that the twisting is nondegenerate:
(A33.17)
THEOREM
Det (
aw ) ~. 0
ap Po
A33.18
Every globally canonical mapping B, C1.close enough to A, has at

least 2" points 3 with period N in the neighborhood of the torus p = Po:
BNx =
x.
Proof:
In view of (A33.1S), (A33.16), and (A 33.17) the mapping AN can be
written under the form:
(A33.19) A": p, q
-+
p, q + a(p), where a(po)
in the neighborhood of the torus p
a(p)
0,
Det
(aa)
ap
Po
Po' It is sufficient to put:
N[w(p)-w(po)]
The neighboring mapping B" can be written under the form:

(A33.20)
BN : (p, q
-+
p+
f3 1(p, q),
3 Counted with their multiplicity.
q + a(p) +
f3 2 (p, q))
(P, Q)
~ 0
2<tS
Let us consider the points which move along the radii: Q
= q,
that is:
(A33.21)
From the implicit function theorem we deduce:
(1) equation (A 33.21) determines a torus T, close to p
wh~ch
Po and
moves along the radii;

(2) the tori T and BN T have equations of the form:
(A 33.22)
I I
= P '(q), P = P "(q), where d:

d ' <
00
IddPq " I < -,~ .
The mapping BN , being globally canonical, is given by a generating function of the form Pq + B (P, q). In view of Lemma (A 33.4), the function
B(x) = B(P(x), q(x)) is s-ingle valued in
a.
Now, consider the restriction
of B (x) to the torus T. This is a smooth function on an n-dimensionaltorus, thus B has at least 2" critical points (compare to Lemma A 33.S).
Let us prove that these critical points belong to the intersection of T
and BNT. Formula (A33.3) yields:
dB
(Q-q)dP + (p-P)dq, where B: x
(p, q)
->
(P(x), Q(x)).
But, according to (A 33.20) and (A 33.21), we have Q - q

T. Therefore (p - P)dq
0 on our torus
0 at the critical points of B on T. But, in view
of (A 33.22), this implies P
p.
(Q. E. D.)
REMARK
A 33.23
This theorem is not a corollary of Lemma (A 33.S): the manifold Q- q
o need
not verify dp 1\ dq
JhJapping proves:
;r
0, as the following example of canonical
APPENDIX 34
PROOF OF THE THEOREM ON THE CONSERVATION

OF INVARlA"T TORI UNDER SMALL PERTURBATIONS
OF THE CANONICAL \It\PPING
(See Theorem 21.11. Section 21, Chapter 4)
The construction of the ipvariant tori is pertormed in Sections E - H of
this appendix. Lemmas of Sections B - D are used. Proof is based upon a
method of successive approximations of Newtonian type suggested by A. N
Kolmogorov [6].
A. Newton s Method
Newton's approximation for the construction. of a zero x of a function
f, with a prescribed approximation, consists in replacing the curve y =
f(x) by its tangent at ~xo' f(x o))' where Xo is an approximation of x. If
Ix-xol < E . the substitution gives an error of order E2. Hence, the linearized equation
f(x o ) + f'(xo)(x ~ x o) = 0
admlts a solution Xl whose deviation from x will be of the order E2 (see

Figure A 34.1).
Iteration of this process leads to an accelerated
conver~
gent approximation:
(A 34.2)
I' x n+ l-xl<Clx-x
n
n
n- 112
I
Hence, the deviation of the n tl :approximatlOn from the solution will be:
I
Ix -
x n 1 .....JE
249
2n-1
250
ERGODIC PROBLEMS OF CLASSICAL MECHA,NICS
Figure A 34.1
This method extends easily to equations in Banach spaces 1. But, in analysis, one is most often concerned with poly-Banachic spaces. and the operator ['(xO)-l maps a space into another one. This situation is illustrated
by'the following lemma, which offers a characteristic example, and where
(A 34.4) plays the role of (A 34.2).
LEMM.A
A34.3 2
Let L be an operator which maps the functions (z), which are analytic in each complex domain G, into the functions Lf(z) which are analytic
in the domain 3 G-o and such that, for every 0
(A344)
< 0 <00 :
ILfIG_o < Ifl5 . o-v ,
where v > 0, 0 0 >

are absolute constants. Then, for every 0' > 0, the
series ~sIT,Sf! converges in G-o', provided that IflG < M = M(o') is
small enough.
1
See KantoroVlch [ I
J.
2 This lemma was.already known by H. Cartan [1]. whose' paper [I] is one of
the first appearances of Newton's approximatlOns m theoretical analysis
3 Here and further,
G - 5 denotes the set of the pomts of G which are at a dis-
tance > 5 from the complement of G. Example of an operator which verifies

(A 34.4): Lf = f:Ii..
dz
251
APPENDIX 34
Prool:
Let
~
2v + 1
and u~
2
~ 3/2
= u1 ,
M13/2 . .... Ms+1
M2
... , U~ s+1
Ms3/2,
hence
M
= 02v+l
s
Then, if
we have:
Now, let
G 1 = G,
Then,
IIIG < Ml
IIIG s < M
G2=Gl-ol.'Gs+l~Gs-oS:
implies
ILsII Gs <
Ms because, in view of (A 34.4),
implies:
But G s => G-o'forany s,because LOs< 0', Therefore,wehave:
inG-o',
(Q. E. D.)
B. Small Denominators
Let I(q) be a function on a torus Tn,

I(q) =
~
k~O
q = (ql' .... , qn) (mod 217):

'ei(k,q)
'
where (k, q) = k 1 ql +'''+ knqn' and let W = (wI' ... ,w n ) be a vector

with irrational components, such that (k, w) ,;, ko for nonvanishing integers
252
k, k o ' Consider the equation

g(q+ w) -g(q) = f(q) ,
(A34.5)
where g is an unknown 271 periodic function. This equation admits a "formal" solution:
(A 34.6)
g(q) =
gk' e,(k,q),
ei(k,c. _ 1
k~O
The following lemma asserts the convergence of (A 34.6):

LEMMA A34.7
Assume that the function f(q) is analytic and If(q)1
for 11m q I
<
<
M always holds
p. Then, for almost every vector w (except for a set of Lebes-
gue measure zero), the function g (q), defined by (A 34.6), is analytic and,
for Ilmql
Here 00
< p-o,
we have: Ig(q)1
< Mo-II, II
2n+4, if 0<0<00.
> 0 is an absolute constant (independent of n).
Proof of this lemma 4 is based on elementary results from th~ Theory'of

Diophantine approximations. In fact, Lemma (A 34.7) holds, with 00
0o(n, K) and for some K
> 0, for any element w of the set no(K) (defined
below). Let us denote by
n (K)
(A34.8)
lei(k,c. - 11
for any w, Iw'-wol
< K~II, II
the set of the wo satisfying:
> KN-l/
n+2, and for any k, Ikl
<
N. Let us de-
II
= n+2.
note by no(K) the set of the Wo verifying:

(A34.9)
Of course,
le,(k,c.>o) - 11
n (K) c
> K
Ik I-II,
no(K)
LEMMA A34.10
Almost every (in the Lebesgue measure sense) point Wo belongs to
n (K)
for some K> 0 (hence, Wo ( no(K)).
4 The technique of evaluating small denominators was extensively worked out by

C. L. Siegel [2],
[3],
in connectIOn with similar problems.
APPENDIX 34
253
Proof:
Let
be a bounded domain of the space
< d for some
lk,d = Iwollei(k,Cil) - 11
Then, clearly:
only on.
n.
mea~ (lk, d
n m :s.
Iwol.
Let:
w, Iw-wol
< dl .
C d, where the constant C depends
Relation (A 34.8) holds outside of Uk Ik,Klkl- v ' But we have
since
!k I-v <
for v
00,
n+2 .
Therefore:
meas
n n (K)
o.
k"'O
(Q. E. D.)
Now, let
~ fk
f(q)
ei(k,q)
be an analytic function.
LEMMA
A34.11
(A) Iffor IImql < p we have If(q)1 < M, then Ifkl < Me-plkl.
(B) If Ifk(q)! < Me-P !kl, then for 11m ql < p-o (where 0 < 0 < 0 0 ):
(C) If for 11m ql
N
254
To prove (A) we have to
s~ift
the contour of integration in the formula:
fk =
to
f e-i(k, q) dq
. ip.
Proofs of (B) and (C) consist in mere summations of geometrical series.
Lemma (A34.7) follows at once from Lemmas (A34.10) and (A 34.11):
take Wo ( Do(K) and take into account (A34.6), (A34.9), (A) and (B) from
Lemma (A34.11), and the elementary inequality:
<
e-lkI5'lklv
C(v)o-V
Then, to obtain Lemma (A 34.7) it is sufficient to take 00
<
K/C(v). We
refer to Arnold [11] for further details.

REMARK
A34.12
Suppose that Wo ( D(K) and f k
0 for Ik
>
N. Then, (A 34.6) re-
duces to a finite sum, which depends continuously on w. Besides, Lemma

(A34.7) still holds, with the same 00
Iw-w'l
<
0o(K, n), for any w' such that
KN- v Because if w ( D(K), then, in view of Definition (A 34.8),
every w', Iw-w'l
<
KN- v verifies (A34.9) for Ik 1 ~ N. But the proof
of Lemma (A 34.7) makes use of (A 34.9) only for Ik
N if [k = 0 for
Ikl > N.
c. Sketch of the Proof
Now, let us recall notations of Theorem (21.11) (see Chapter 4). The
set
~ Bn
x Tn is a domain of the canonical space, a point x of
is
denoted by x = (p, q), where p = (Pl' ... , Pn ) is a point of the Euclidean

ball 8 n and q = (ql' ... , qn) (mod 217) is a point of the t<?rus Tn. The mapping A: p, q .. p, q + w(p) is {he "unperturbed" mapping. The mapping ./
B, the generating function (see Appendix 32) of which. is Pq + B (P, q),
,is a small analytic globally canonical perturbation. We look for the
ant tori of the mapping B A .
invari~
255
APPENDIX 34
Following the idea of the Perturbation Theory (Appendix 30), we try to

"kill" the perturbation B by an appropriate canonical coordinate transformation C, with generating function Pq + C(P, q). In coordinate Cx, the mapping B A can be written:
C(BA)C- 1
where B'
C B A C- 1 A-I.
B'A
Therefore, taking into account Corollary
(A 32.20), the generating function of B' is:
Hence, killing B reduces to solving, with respect to C,

C(x) + B(x) - C(A- 1 x) = O.
Now, observe that this equation, for each fixed p, has precisely the form
of Equation (A34.5) (with w
= w(p),
f = -B,
C). Now, construct-
ing successive approximations to the invariant tori reduces to obtaining

the inequalities:
(A34.13)
in a domain
n' c n
(but not "too small"). Then, the convergence is car-
ried out as in Lemma (A 34.3); Inequalities (A 34.13) are proved with the
help of Lemma (A 34.7): they hold "far from resonances," that is for w(p)
( n(K).
To perform, along Sections E - H, the above program, we shall make

use of several devices. First of all, instead of "killing" B, we only kill
a truncation of its Fourier series: the remainder term RNB can be regarded as part of the "terms of higher order" 0(B 2 ) and 0(C 2 ). Hence, every
app.roximation deals with only a finite number of resonances and small denominators in (A 34.6)5.
On the other hand, to use Lemma (A 34.7), we
5 This process was already used by Bogolubov. Mltropolski [1]. One could do
without it by replacing the frequencIes cu(p) with constants cu ( Q K in the small
denominators.
256
need first to eliminate the constant term B 0 (averaged value of B with

respect to q): we add Bo to the "unperturbed" mapping A (vahation of
the frequencies, Section E).
J. Moser [4], the I\olmogorov
Finally, observe that, as pointed out by
method still works in the case of differentiable mappings: in this case,

one can devise a smoothing process like a suitable truncation of the Fourier series at the Nth frequency.
(21.11) in the case n
Indeed, Moser [1] improved Theorem
1 (mappings of the plane) by abandoning the re-
quirement of analyticity and substituting instead the requirement that 333

derivatives exist! Recently Moser [6] gave a proof which requires only a
temperate number of derivatives.
O. Canonical Mappings Close to the Identity
In this section we use the following notation: Let F(o, M) be an
arbi~
trary function and a a proposition involving 0, M, F. We shall say that
"a is true and
IF I h M"
if there exist absolute constants 6 VI' v 2 > 0
and 00 > 0, such that:

a is true and
(A 34.14)
provided that 0 <
LEMMA
<
0,
M<
IF I :s.
MO- v 2
oV I
A34.15
Let G be a complex domain and f(z) an analytic function in G, which

satisfies Il(z)1 < M. ..Then, for z ( G-o, the k-th derivative of f
lytic and
IS
ana-
<
n M.
6 That is that these constants depend on the d,mens,ons of the domains dealing
with the proposition a, on the number of depivatives, etc., but neither on the functions nor on the domains.
257
APPENDIX 34
Proof:
From the Cauchy formula
dkf
k!
F()d(
277i . Y (_z)k+l
we deduce:
.E kf
\ dz k
:s.
1
k! MB- k
Hence, we obtain (A34.14) with:
1
k'
o.
(Q.~. D.)
Here:
F (0, M) =
sup
f.G.z (G-O
dkf I
,
Idz
k
where sup ranges over all the bounded domains G, all the analytic
tions f,
If I
func~
< M in G, and all the points z of G - B.
Observe that: CM B- v ~ M. and that the inequalities F ~ M and

G ~ M imply F + G ~ M, FG ~ M2
Then. if
F(B, M) ~ M.
we have F (C 0, M) ~ M (here, C and
f.J
are absolute positive constants).
Now, let us make clearer how the global canonical mappings S, close
to the identity. are related to their generating functions Pq + S (P, q). Let
B n x Tn, B n = !p lip I < y, p ( Rn
Tn
= !q (mod 277), q = (ql' ... , qn)l ,
I,
and [0) be the complex domain of 0, given by

Ipi < y, IImql < p, where 0 < Y < 1.0 < P < 1.
LEMMA
A34.16
Let S (p, q) be an analytic function in [0], satisfying:

IS(p, q)1
< M.
258
Then; the formulas:
(A34.17)
as ,
q+-
ap
S (p, q)
determine a global canonical diffeomorphism S:
P (p, q),
Q(p, q) ;
S: [0] - 20
->
[0] - 8 ,
and the folJowing inequalities hold in [0] - 28:
I(P LEMMA
p) - as (p, q)
ap
I <
q) I <
I(Q- q) + as(p,
aq
n
M2
n'
M2.
A34.18
= P (p, q), Q = Q(p, q) is a global analytic canonical mapping,

which satisfies IP - pi < M, IQ - ql < M in [0], then this mapping is deIf P
fined in [0] -8 by formulas

IS I
LEMMA
M,
(A 34.17),
IS(p, q) -
where S is analytic and verifies:
(p,q)
(Q- q)dp - (P-p)dql
M2
A34.19-
Let S (p, q) an.d T(p, q) be two analytic functions, which satisfy IS I
< M, IT I < M in [0]. Then the product R
ST of the corresponding
canonical mappings is a global canonical diffeomorphism, which maps

[0] -38 infO [0] - 28, and which is defined in [0] - 8
by an analytic gen"
erating function R satisfyinB:
Now let. 'A: [0]
->
[0 '] be an analytic global canonical diffeomorphism
([0
is 'given by Ip I < y', 11m q I < p', 0 < y', p' < 1). Let a-1ly - xi
< lA<r)-A(x)1 < aly-xl, and S be an analytic function in [OJ, S the canonical diffeomorphism corresponding to S by virtue of Lemma (A 34,16).
259
APPENDIX 34
LEMMA
A 34.20
The formula T
of [n'] - 38 into
A S A-I defines a global canonical diffeomorphism
to'] - 28.
which is given in [n'] - 8 by an analytic gen-
erating function T satisfying 7 :
Proofs of the preceding lemmas reproduce those of Appendix 32. making

use of Lemma (A 34.15) to compute the terms of order 0 (c: 2). For further
details we refer to Arnold [4]. [5].
E. Variation of the Frequencies
Let us now begin the construction of the invariant tori of the mapping
B A (see Theorem 21.11. Chapter 4).
To understand the estimates performed in Sections E - F. it is useful
to keep in mind that the positive numbers {3. y. 8. M and p. K. 0 satisfy:
0,< M 8
and that vi'
Ci
are absolute
(3
constants~ v
p. K. 0- 1
<1
> 1 > c.
CONSTRUCTION OF THE VARIATION OF THE FREQUENCIES.
A 34.21
Let A and B be two global canonical mappings:
!\: P. q
-+
P. q + w( q)
r. f
and B have generating function Pq + B(P. q). We put:
ii(p)
= (21T)-n
w 1(p)
B (P. q)dq l '" dqn
aii
ap
= w(p) + -
and we consider the following canonical mappings (see
Figu~
7 In thiS lemma, the constant 50' which enters mto the defmihon of
also on a.
A 34.22):
, depends
260
Figure A 34.22
Of course, BA = B'A 1
LEMMA CONCERNING THE VARIATION OF THE FREQUENCIES. A34.23
Suppose that:
O-l\dp\
< \dw\ < O\dp\ and \B(p, q)\ < M
hold in the domain \p - p*\ < y, \Im q \ < p. Then the global canonical
mapping B' possesses a generating function Pq + B '(P, q) which satisHes 8 :
where
in the domain \p-p*\ <
y-o, \Imq\
<
p-o.
Proof:
Application of Lemma (A34.19) to the mappings Band AA1l: p, q
-->
p,q-(aB/ap) yields \B'-(B-B)\ ~ M2

(Q. E. D.)
F. Fundamental Lemma
We now make use of the estimates of the small denominators (Section B)
to obtain inequallties of the form \B 1\ ~ M2, after a suitable change of
coordinates C.
8 The constant 50' m the relation
eventually depends on
e.
261
APPENDIX 34
A34.24
FUNDAMENTAL CONSTRUCTION.
Let A and B be global canonical mappings:
A: p, q
-+
p, q + wI (p) ,
and B defined by a generating function Pq + B (P, q),

B (P, q)
~ B k(P)ei(k, q)
B(P) +
k,t 0
We put: -
C (P, q)
Ck(P)ei(k,q) ,
O<lkl~N
(A 34.25)
Bk(P)
e-i(k,Ull (P)) _
where N is a positive integer. We denote by C the global canonical mapping, whose generating function is Pq + C (P, q), and we consider the
global canonical mapping Bl
C B A C- 1 A-I. Of course:
BIA
C(BA)C- 1
Observe also that:
B ""'C
A
B1 .
B
cl
----
Bl
F,gure A34.26
FUNDAMENTAL LEMMA
We suppose that
wI (p),
in
A34.27
the domain
Ip - p*1 < y, 11m qI <
B (p, q) are analytic and satisfy:

IB(p,
q)1 <
M,
p, the functions
IB(p)1 < Ii
262
We suppose also that w * ~ wI (p *) belongs to the set U(K) of Lemma

(A34.10), then:
(1) in the domain IP-p*1
< y, IImql < p-5 the function C(P,q)
is analytic and verifies:
(2) in the domain, IP - p*1 < y -5, 11m q I < p -{3, the generating function Pq + B 1 (P, q) of B1 satisfies:
(A34.28)
provided that [for (1) and (2) l:
Proof:
From Idwl < OldPI and y < C4 N-(n+2), C 4 < K/O, it follows that
all the points w 1 (p) of the domain Ip-p*1 < y ,:erify Iw-w*1 < KN-(n+2).
Consequently, according to the last remark of Section 8", the assertion of
Lemma (A 34.7) holds"for the trigonometric sum C. Let 50 ~ 5 0(n, K) be
t~e constant of Lem~a (A 34.7). If C 3 < 00' then we have 5 < 50; therefore Lemma (A 34.7) implies that IC I < M5- vI for jIm q I < p -8. This is
precisely the (1) of our lemma.

Observe next that if C 2 (in y < C 2 (3) is small enough, then we have
11m w(p) I < Oy < (3. Under this condition, the mappings A and A-I are
diffeomorphism.s:
!p, q\lp-p*1 < y', IImql < p'l .. !p, q\lp-p*1 < y", IImql < p"l,
where' p" < p' +
Or < p.
This allows one to apply Lemma (A 34.20) to
9 That is. tpey depend only on the dimension n and the constants K,
e.
The fun-
damental lemma states that there exist constants vI' v2 (large enough) and C l'
C 2 , C 3 , C 4 (small en0':1gh) such that (1) and (2) hold.
263
APPENDIX 34
A C- 1 A-1. Besides, if v 1 and v 2 (in M < 8 V2 ) are large enough and C 3

(in 8 < (3 < C 3 ) small enough, the relations ~ of Lemmas (A 34.16),
(A34.18), (A34.19), and (A34.20) are of the form: < M8- vl .

Hence, we deduce from these lemmas that, under the hypotheses of the
fundamental lemma, with suitable constants C, v, the mapping B1
C BAC- 1 A- 1 has a generating function Pq + B 1 (P, q) such that in the

domain IP-p*1 < y-8, IImql < p-(3, the function B 1(P, q) is analytic
and:
(A34.30)
But from (A34.2S) follows:
(A34.31)
If 11m q I < P - (3 < P - 8, then assertion (C) of Lemma (A 34.11) implies:
IRNB
(A34.32)
Since 181
I < Me- 13N . (3-v1
< ii, formulas (A 34.30), (A34.31), and (A 34.32) imply (A 34.28).

(Q. E. D.)
G. The Inductive Lemma
The construction of the invariant tori makes use of an iterative process,

each step being based on the following construction.
THE INDUCTIVE CONSTRUCTION.
A 34.33
Let A and B be two canonical mappings:

,,; p, q
-+
p, q + w(p) ,
B given by a generating function Pq + B (P, q), and N a positive integer.

Performing the variation of the frequencies (Section E) we obtain:
"1; p, q
-+
p, q + w 1 (p) ,
and B', which is given by a generating function Pq + B '(P, q) and such

that B"
B'
"1'
We now apply the fundamental construction of Section F
to the mappings "1' B '. We obtain a canonical mapping C, and B1 =
264
C B'A 1C- 1A 1 ; hence (see Figure A34.34) we have:
B1A1
C(BA) C- 1
c ._ _ _ _ _ _ ---.:}
Figure A 34.34
THE INDUCTIVE LEMMA
A 34.35
We suppose that in the domain Ip - p*1
IB(p, q)1 < M,

Let us define
w(p*)
< y, 11m q I < P, we have:
w*
O(K) .
P:
by w 1(P:) = w*, and let Pq+ B 1(P, q) and Pq+ C(P, q)

be the generating functions of B, and C. Then:
(1) in the domain IP-p*1 < y, IImql < p-o, the function C(P, q)
is analytic and
IC I <
M o-Vt ;
(2) the domain IP - p:1 < Y1IIm q I < PI
P -{3 belongs to the do-
main IP - p*1 < y, 11m q I < p, and in this domain the function B1 is analytic and
provided that:
where the constants
1/1,1/2,1/3' 1/4
> 1 > C 1 , C 2 , C 3 , C4 , C s > 0 are ab-
solute constants, that is depend only on the dimension n and on the constants
eo'
K, but not on B, w,
e,
M, and so on.
265
APPENDIX 34
Proof:
Proof of the inductive lemma follows at once from the two preceding
lemmas. The only two novelties are:
P: :
(1) The existence of
This existence and the inequality
Ip: I <
CsY follow from the inequal-
ities:
IdWI
dp
(see Section E) and from M8- v
<
<
_ dw
dp
Ip - p*1 < Y-8
belongs to
<
M8- v
C 6 y (which follows from M8- v
C 6 Y for v 2 large enough). Taking into account Yl
Ip - p:1 < Y1
I
<
< 8 v2- v
C4 y, the domain
because 8 < Cly
(2) The estimate of B 1 :
From Section E it follows that averaging B' gives:
~ ow
replace M in relation (A 34.28) of the fundamental lemma, we obtain:

M28-V1 +
M<
M28- V4
(Q. E. D.)
H. Proof of Theorem (21.11)
(see Chapter 4)
CONSTRUCTION A34.36
The invariant torus T (w *) of the mapping B A, which corresponds to

the frequencies w *, is constructed by making use of the process of the
preceding section. This process depends on a sequence 0 < N 1 < N 2 < ...
< Ns <
00',
Ns
-+
+ 00 which will be specified later.
After the sequence N s is selected, the construction goes as follows.

We put Al ~ A. B1 = B. N 1 = N, and we make use of the inductive con~
struction of Section G. This construction determines canonical mappings,
namely C 1 A2
H2
we have:
B2A2
C1(R1 A1)C 1- 1 .
266
We again apply the sam~ construction with A2 = A. B2 = B. N2 = N;

we obtain A3 B3 C2 , and so on. If As' Bs are constructed, the construction of Section G, with A = As' B = Bs' N = N s leads to Cs ' As+ 1 '
Bs+l (see Figure A 34.37):

1
Cs (B s A s )Cs
000
As
Figure A 34.37
This
construction also determines the points p*,
ws(p*)
s
s
Ts(w*) denotes the torus p
p:. The mapping
torus, is the translation defined by w *. We put
T:
be the torus p
P:,
Now
As' restricted to this
p:
lim
s' and let
s~""
A"" the mapping:
A : T*
""
= w*.
""
~ T* ,
""
A""q
q +w
Finally, we put:
os --
C 1-1 C2-1
...
C s-1
-1
11m
Us
s~""
The invariant torus of B A is given by:

T(w*) = DT*
""
We now prove that the above limits exist and that D:\""
B -\ D on T (w * ).
(Q. E. D.)
267
APPENDIX 34
A 34.38
CONVERGENCE,
In view of dw/dp
~ 0, we can suppose that:
0-11 dp I
< Idw(p)1 < Oldp I
holds in [n]. Let us suppose that w * (O(K). According to Lemma (A

34.10), almost every w* belongs to O(K) for some K> O. We put ()o
2(),
and we define a sequence of constants by:
Let us put:
oY,s '
f3 s
Y '/.o(n+2)
s
-Y,(n+2)
Ys
'
then we have:
f3 s+ 1 = f3 s3/2 '
N s+l
To define all these numbers, we just need to choose
o.
Ys+l
y3/2
s'
Ns3 / 2
Denote by a a
positive constant, if a is taken large enough we have:
Let C, 0 < C < 1/10, be an absolute constant, that is, C depends only
on n, K, 00' a and on the constants
tive lemma. If 0 < C,' then,
\!k'
C k which enter into the
induc~
obviou~:;ly:
(1)
(A34.39)
(2)
(A 34.40)
(3) for any s
1,2, ... , we have:
(A 34.41)
where C 1 C 2 , C 3 , C 4 , C s are the constants of the inductive lemma, which
268
depend on the constants K and
eo
we introduced above;
(4)
(A 34.42)
Our number 8 is nowoselected to satisfy: 0 < 8 < C. Hence, the numbers
N s' on which our construction depends, are determined.
Suppose now that our mapping B1
B has a generating function
Pq+ B 1 (P, q} which verifies IB11 < M1 = 8 1a in Ip-p~1 < Y1' 11m ql <
P1 = Y:zp. According to Lemma (A 34.18) this inequality holds if E, under
the conditions of Theorem (21.11), is small enough. Taking into account

inequalities (A 34.41), the hypotheses -of the inductive lemma are satisfied
A1 , B = B1 (because a > 1J 2 ) Therefore we obtain .'\2' 8 2 , C 1
and so on from the inductive lemma.
by A
p;,
Let us prove that the mappings B2 A2
again satisfy all the condi-
tions of the inductive lemma in the domain Ip - p;1
< Y2' 11m q I < P2
P1 - (31' In fact, (A34.39~ implies P2 > 0; (A 34.40) and the third part
of the inductive lemma show that "2 satisfies the inequalities:
Finally, the second part of the inductive lemma, (A34.42), and a>
a >
1J 1
21J 4
+ 1,
+2, imply:
IB21 < Mf8 1Y4 +
M 1 e- iV'I1
8- vl < 8;a-v4 + 813/2 (at-1) < 813012
In other words:
Hence, B2 and A2 fulfill all the conditions of the inductive lemma.

Iteration of this argument leads to:
IB s I < Ms
8 a , for Ip-p:1 < Ys '

s
IImql < Ps '
Let G s be the domain Ip-p!1 < Y s ' IImql < P s ' Then,thediffeomor
1 map G 1 into G and, in the C 1.norm, we have:
phisms Cs
s+
s
APPENDIX 34
269
E = identity,
(A 34.43)
in G s+I' The point
p:
is the intersection of the balls:

S
...
00
'.
From the estimate (A 34.43), follows at once the convergence of the ns
on
the torus
r*
00
nG
s>1
The inequalities
on the torus
IB sl < 8~
imply;
r*.
00
Finally, from IWs+l -
wsl < 8 s
follows the convergence of the map-
pings As:
Thus'
'\
00'
on
r*
00
(Q. E. D.)
BIBLIOGRAPHY
Abramov. L. M.
'[l] Concerning a note of Genis, R Z Math. 8 (1963) p. 439.
[2] Metric Automorphisms with Quasi-Discrete Spectrum, lsvestia Math.
Series 26 (1962) pp. 513-531.
Adler. R. L.. and Rivlin. T. V.

[1] Ergodic and Mixing Properties of Chebyschev Polynomials, Proc.
Amer. Math. Soc. 15 No.5 (1964) pp. 794-796.
Andronov. A. A. and Pontrjagin. L. S.

[1] Rough Systems, Dokl. Akad. Nauk. 14 (1937) p. 247-250 .
Anosov. [). V.
[1] Roughness of Geodesic Flows on Compact Riemannian Manifolds
of Negative Curvature, Dokl. Akad. Nauk. 145 (1962) p. 707-709.
[Sov. Math. Dokl. 3 No.4 (1962) pp. 1068-1069.]
[2] Ergodic Properties of Geodesic Flows on Closed Riemannian Manifolds of Negative Curvature, DokI. Akad. Nauk. 151 (1963) pp.
1250-1253. [SOy. Math. Dokl. 4 No.4 (1963) pp. 1153-1156.]
[3] A,feraging in Systems of Ordinary Differential Equations with Rapidly Oscillating Solutions, lsvestia, Math Series 24 (1960) pp. 721742.
Anzai. H.
[1] Ergodic Sltew Product Transformations on the Torus, Osaka Math.
J. (1951) pp. 83-99.
Arnold. V. I.
[1] Sur la geomtHrie des groupes de difeomorphismes et ses applications en hydrodynamique des fluides parfaits, Ann. lnst. Fourier
(1966).
[2] Remarks on Rotation Numbers, Sibirski Math. Zh. 2 No.6 (1961)
pp. 807-813.
[3] Some Remarks on Flows of Line Elements and Frames, Dokl. Akad.
Nauk. 138 No.2 (1961) pp. 255-257. [SOy. Math. Dokl. 2 (1961) pp.
562-564.]
271
272
[4] Small Denominators and Problems of Stability of Motion in Classical and Celestial Mechanics, Usp. Math. Nauk. 18 No.6 (1963) pp.
91-196. [Russian Math. Surveys 18 No.6 (1963) pp. 85-193.]
[5] Proof of a Theorem of A. N. Kolmogorov on the Invariance of QuasiPeriodic Motions under Small Perturbations of the Hamiltonian,
Usp. Math. Nauk. 18 No.5 (1963) pp. 13-40. [Russian Math. Surveys 18 No: 5 (1963) pp. 9-36.]
Arnold. V. I.. and Sinai. Y.

[6] Small Perturbations of the Automorphisms of a Torus, Dokl. Akad.
Nauk. 144 No.4 (1962) pp. 695-698. [SOy. Math. Dokl. 3 (1962) pp.
783-786.1
Arnold. V. I.
[7] On the Stability of Positions of Equilibrium of a Hamiltonian Sys. tern of Ordinary Differential Equations in the General Elliptic Case,
Dokl. Akad. Nauk. 137 No.2 (1961) pp. 255-257. [SOy. Math Dokl.
2 pp. 247-279.]
.
[8] On the Generation of a Quasi-Periodic Motion from a Set of Periodic
Motions, Dokl. Akad. Nauk. 138 No.1 (1961) pp. 13-15. [SOy. Math.
Dokl. No.2 pp. 501-503.]
[9] On the Behavior of an Adiabatic Invariant under a Slow Periodic
Change of the Hamiltonian, Dokl. Akad. Nauk. 142 No.4 (1962) pp.
758-761. [SOy. Math. Dokl. No.3, pp. 136-139.]
[10] On the Classical Theory of Perturbations and the Problem of Stability of Planetary Systems, Dokl. Akad. Nauk. 145 No.3 (1962)
pp. 487-490. [SOy. Math. Dokl. No.3 (1962) pp. 1008~101l]
[ll] Small Denominators I, on the -Mapping of a Circle into Itself, Izvestia Akad. Nauk. Math. Series 25, 1 (1961) pp. 21-86. [Transl.
Amer. Math. Soc., Series 2,46 (1965) pp. 213-284.]
[12] Conditions for the Applicability, and Estimate of the Error, of an
Averaging Method for Systems which Pass through States of Resonance in the Course of Their Evolution, Dokl. Akad. Nauk. 161
No.1 (1965) pp. 9-12. [SOy. Math. Dokl. 6 No.2 (1965) pp. 331337.]
[13] Instability of Dynamical Systems with Many Degrees of Freedom,
Dokl. Akad. Nauk. 156 No.1 (1964) pp. 9~12. [SOy. Math. Dokl. 5
No.3 (1964) pp. 581-585.]
[14] Stability and Instability in Classical Mechanics, Second Summer
Math. School, 1964 2, Kiev (1965) p. 110.
.
Arlin. M.. and Mazur. B.

[1] On Periodic Points, Ann. Math. 81 No.1 (1965) pp. 82-99.
BIBLIOGRAPHY
273
Auslander, L., Green, L., and Hahn, F.

[1] Flows on Homogeneous Spaces, Ann. Math. Studies 53 (1963).
Avez, A.
[1] Quelques inegalites de. geometrie diferentielle glob ale deduites
de la theorie ergodique, C. R. Acad. Sci. Paris t.261 (1965) pp.
2274-2277.
[2] Spectre discret des systemes ergodiques classiques, C. R. Acad.
Sci. Parist. 264 (1967) pp. 49-52.
Birkhoff, G. O.
[1] Dynamical Systems, New York (1927).
J. R. and Hanson O. L.
[1] On the Isomorphism Problem for Bernouilli Schemes, Bull. Am.
Math. Soc. 69 No.2 (1963) pp. 221-223.
Blum,
Bogolubov, N. N. and l\1itropolski, Y. A.

[1] Les methodes asymptotiques en theorie des oscillations non lineaires, Gauthier-Villars, Paris (1962).
Bogolubov; N. N.
[2] Proceedings of a Summer Math. School at Kanev (1963) Editions
"Naukova Dumka,". Kiev (1964).
Bohl, P.
[1] Uber ein in der Theorie der Sakularen storungen vorkommendes
Problem, }. Re~ne u. Angew. Math. 135 (1909) pp. 189-283.
Cairns, S. S.
[1] On the Triangulation of Regular Loci, Ann. Math. 35 No.2 (1934)
pp. 579-587.
Callahan, F. P.
[1] Density and Uniform Density, Proc. Am. Math. Soc. 15 No.5 (1964)
pp. 841~843.
Cartan.
n.
[1] Sur les matrices holomorphes de n variables complexes, }. Math.

pures appli. 19 (194 0) pp. 1-26.
Chacon, R. V.
[1] Change of Velocity in Flows, }. Math. Mech. 16 No.5 (1966).
274
Coddington. K. and L('vinson. I'll.

[1] Theory of Ordinary Differential Equations, McGraw-Hill, New York
(1955).
()(' Baggis. II. F.
[1] Dynamical Systems with Stable Structure. Contribution to the Theory of Nonlinear Oscillations, Vol. 2, Ann. Math. Studies 29 (1952)
pp. 37-59.
Denjoy. A.
[1] Sur des courbes definjes par des equations differentielles
face du tore, }. de Math. 9 (1932) pp. 333 375.
a la sur-
Eul('r. L.
[1] Theoria Motus Corporum Solidorum Seu Rigidorum (1765).
Ge Ifand. I. M. and Shapiro-Piatct.lski. I. I.
[1] A Theorem of Poincare, Dok!. Akad. Nauk. 127 No.3 (1959) pp.
490 493. [Math. Review (1960) No. 6460.]
0
Gelfand. I. M. and Fomin. S. V.

[2] Geodesic Flow on Manifold of Constant Negative Curvature, U sp.
Math. Nauk. 47 No.1 (1952) pp. 118-137. [Trans!. Amer. Math.
Soc. 2 No.1 (1955) pp. 49067.]
Gelfan(l. I. \1 . Gra('v. 1\1. I.. Zueva. H. ~I . l\1ichailova. M. S .. an'd Morosov
.\. I.
[3] All Example of a Toroidal Magnetic Field Not Having Magnetic
Surfaces, Dok!. Akad. Nauk. L-rJ No.1 (1962) pp. 81083. [SOy.
Phys. Dok!. 7 pp. 223 724.]
0
Gelfand. I. \I. and Li(lsky. V. B.

[4] On the Structure of Regions of Stability of Linear Canonical Syso
tems of Differential Equations with Periodic Coefficients, U sp.
Math. Nauk. 10 No.1 (63) (1955) pp. 3040. [Trans!. Amer. Math.
Soc. 2 (8) (1958) pp. 143-181.J
Genis. '\. L.
[1] Metric Properties of the Endomorphisms of the n-Dimensional
Torus, Dok!. Akad. Nauk. 138 (1961) pp. 991 993. [SOy. Math.
Dok!. 2 (1961) pp. 750752.]
0
Gourcvitch. B. M.
[1] The Entropy of Horocycle Flows. Dok!. Akad. Nauk. I % No.4
(1961) pp. 768 770. [SOy. Math. Dok!. 2 (1961)pp. 124-130.
0
BIBLIOGRAPHY
275
Grant, A.
[1] Surfaces of Negative Curvature and Permanent Regional Transitivity, Duke Math. J. No.5 (1939) pp. 207-229.
Guirsanov, I. V.
[1] On the Spectra of Dynamical Systems Which Arise from Stationary
Gaussian Process, DokI. Akad. Nauk. 119 No.5 (1958) pp. 851853.
Gysin, W.
[1] Zur Homologietheorie der Abbildungen und Faserungen von Mannigfaltigkeiten, Comment. Math. Helv. 11 (1941) pp. 61-122.
Hadamard. J.
[1] Les surfa~es a courbures opposees et leurs lignes geodesiques,
J. Math. pures appl. (1898) pp. 27-73.
Hajian. A. B.
[1] On Ergodic Measure Preserving Transformations Defined on an
Infinite Measure Space, Proc. Am. Math. Soc. 16 (1965) pp. 45-48.
Halmos. P. R.
[1] Lectures on Ergodic Theory, Chelsea, New York (1958).
[2] Measure Theory, New York (1951).
[3] Introduction to Hilbert Spaces, Chelsea, New York (1 q57).
Hedlund. G.
[1] The Dynamics of Geodesic Flows, BuIl. Am. Math. Soc. 45 (1939)
pp. 241-246.
Helgason. S.
[1] Differential Geometry and Symmetric Spaces, Academic Press,
New York (1962).
Henon. M. and Heiles. C.
[1] The Applicability of the Third Integral of Motion: Some Numerical
Experiments, Astron. J. 6q No.1 (1964) pp. 73-79.
Hopf, E.
[1] Ergodentheorie, Springer, Berlin (1937).
lusvinski i. S. A.
[1] On Metrical Automorphisms with Simple Spectrum, DokI. Akad.
Nauk. 172 No.5 (1967).
276
Jacoubovitch, V. I.
[1] Questions of the Stability of Solutions of a System of Linear Differential Equations of Canonical Form with Periodic Coefficients,
Mat. Sbornik. 37,79 (1955) pp. 21-68. [Transl. Am. Math. Soc. 2
10 (1958) pp. 125-175.]
Kagan, V. F.
[1] Foundations of the Theory of Surfaces, Osnovy Teorii Poverchnoster I, Moscow (1947).
Kantorovitch, L. B.
[1] Functional Analysis and Applied Mathematics, Usp. Math. Nauk.
3 No.6 (1948) pp. 89-185.
Kasuga, T.
[1] On the Adiabatic Theorem for the Hamiltonian System of Differential Equations in the Classical Mechanics, I. II. III., Proc. Japan.
Acad. 37 No.7 (1961) pp. 366-382.
Katok, A. B.
[1] Entropy and Approximation of Dynamical Systems by Periodical
Mappings, Funkzionalnyi Analys i ego Prilojenija, Moscow, I,
No.1 (1967) pp. 75-85.
[2] On Dynamical Systems with Integral Invariants on the Torus,
Funkz. Anal. i ego Prilojenija, Moscow I No.3 (1967).
Katok, A. B. and Stepin, A. M.
[1] On the Approximations of Ergodic Dynamical Systems by Periodical Mappings, Dokl. Akad. Nauk. 171 No.6 (1966) pp. 1268-1271.
Kolmogorov, A. N.
[1] Sur les systemes dynamiques avec un invariant integral la surface du tore, Dok!. Akad. Nauk. 93 No.5 (1953) pp. 763-766.
[2] A New Metric Invariant of Transitive Systems and AutomorphisJ.T1s
of Lebesgue Spaces, Dokl. Akad. Nauk. 119 (1958) pp. 861-864.
[Math. Review 21 No.2035a.]
[3] Foundations of Probability Theory, Chelsea, New York (1956).
[4] On the Entropy per Time Unit as a Metric Ihvariant of Automor- .
phisms, Dokl. Akad. Nauk. 124 (1959) pp. 754-755.
[5] La theorie genera Ie des systemes dynamiques et la mecanique
classique, Amsterdam Congress 1(1954) pp. 315-333. [Math. Review 20 No. 4066.]
[6] On the Conservation of Quasi-Periodic Motions for a Small Change
in the Hamiltonian Function, Dokl. Akad. Nauk. 98 No.4 (1954)
277
BIBLIOGRAPHY
pp. 527-530. [Math. Review 16 No. 924.]

[7] Lectures given in Paris (1956).
Koopman, R. O.
[1] Hamiltonian Systems and Transformations in Hilbert Spaces, Proc.
Nat!. Acad. Sci. U. S. 17 (1931) pp. 315-318.
Kouchnirenko, A. G.
[1] An Estimate from Above for the Entropy of a Classical System,
Dokl. Akad. Nauk. 161 No.1 (1965) pp. 37-38. [Sov. Math. Dokl.
6 No.2 (1965) pp. 360-362.]
[2] Every Analytical Action of a Semi-Simple Lie Group in the Neighborhood of a Fixed Point is Equivalent to the Linear Action, Funkz.
Analys i ego Prilojenija, Moscow, 1 No.1 (967) pp. 103-104.
[3] Sur les invariants metriques du type entropie (Russian) Int. Congress, Moscow VIII (1966).
Krasinskii, G. A.
[1] Normalization of a System of Canonical Differential Equations Near
a Quasi-Periodic Motion, Bull. Inst. Theor. Astr., Leningrad (1967).

Krein, ~. G.
[1] A Generalization of Several Investigations of A. M. Lyapounov,
Dokl. Akad. Nauk. 73 (1950) pp. 445-448. [Math. Review 12 No. 100.]
[2] The Basic Propositions of the Theory of ,\ -Zones of Stability of a
Canonical System of Linear Differential Equations with Periodic
Coefficients, Pamyati A. A. Andronova, Izvestia Akad. Nauk. (1955)
pp. 413-498. [Math Review 17 No. 738.]
Kupka, I.
[1] Stabilite des 'varietes invariantes d 'un champ de vecteurs pour les
petites perturbations, C. R. Acad. Sci. Paris 258 (1964) pp. 4197~
4200.
Lagrange, R.
[1] Oeuvres t. 5, pp. 123-344.
Leontovitch, A. M.
[1] On the Stability of the Lagrange Periodic Solutions for the Reduced
Problem of the Three Bodies, Dokl. Akad. Nauk. 141 No.3 (1962)
pp. 525-528. [Sov. Math. Dokl. 3 No.2 (1962) pp. 425 430.]
e
Levi-Civita, T.
[1] Sopra alcuni criteri di instabilita, Ann. Mat. pura appl. 3,5 (1901)
pp. 221-307.
278
Margoulis, A.
[1] Sur quelques questions Liees la theorie des C-systemes d'Anosov
(Russian) Int. Congress, Moscow VIII (1966).
Melnikov. V. K.
[1] On the Stability of a Center for Time Periodic Perturbations, Trudy
Moskovskogo Math. Obschestva 12 No.3 (1963) pp. 3-53. [Math. Review No. 5981 (1964).]
[2] On Some Case of C~nservation of Conditionally Periodic' Motions
under a Small Change of the Hamiltonian Function, Dokl. Akad.
Nauk. 165 (1965) pp. 1245-1248.
Meshalkin, L. D.
[I] A Case of Isomorphism of Bernouilli Schemes, Dokl.
128 (1959) pp. 41-44.
A~ad.
Nauk.
Milnor, J.
[I] Morse Theory, Ann. Math. Studies, Princeton (1961).
Mosel', J.
[1] On Invariant Curves of Area-Preserving Mappings of ail Annulus,
Nachr. Akad. Wiss. Gottingen No.1 (1962).
[2] On Invariant Surfaces and Almost Periodic Solutions for Ordinary
Differential Equations, Am. Math. Soc. Notices, 12 No.1 p. 124.
[3] New Aspect in the Theory of Stability of Hamiltonian Systems,
Commun. Pure Appl. Math. II (1958) pp. 81-114.
[4] A New Technique for the Construction of Solutions of Non-Linear
Differential Equations, Proc. Nat!. A cad. Sci. U. S. 47 (1961)'pp .
I
. 1824-1831.
[5] . On Theory of QuaSi-Periodic Motion, Siam Review (1966).
[6] A Rapidly Convergent Iteration Method and Non-Linear Partial Differential Equations 'I, II, Annali Della Scuola Norma/e Superiore di
Pisa, Serie III 20 Fasc. II, III (1966) pp. 265-315; pp. 499-535.
[7] Convergent Series Expansions for Quasi-Periodic Motions, Math.
Annalen, 169 (1967) pp. 136-176.
[8] On a Theorem of V. Anosov (in press).
Moser, J. and Jeffrey>, W. II.
[9] QuaSi-Periodic Solutions for the Three':'Body Problem, Astron. J.
71 No.7 (1966) pp. 568-578.
Nemytskii, V. and Stepanov, V. V.
[I] Qualitative theory of Differential Equations, Princeton (1960).
BIBLIOGRAPHY
279
Ochozimski, O. E., Sarychev, V. A., Ziatoustov, V. A., and Torzevsky,
A. P.
[1] Etude des oscillations d'urt satellite dans Ie plan d'une orbite elliptique, Kosmicheskie Issledovania 2 No.5 (1964) pp. 657-666.
Peixoto, M. H.
[1] On Structural Stability, Ann. Math., Series 2, t. 69 (1959) pp. 199-
222.
Poincare If.
[1] Oeuvres completes, 1 pp. 3-221.
[2] Les methodes nouvelles de la mecanique celeste, 3. Gauthier-Villars,
Paris (1899).
[3] Sur un theor~me de geometrie, Rendiconti circolo Mathematica di
Palermo, t. 33 (191.2) pp. 375-407.
Polya, G. and Szego, G.
[1] Aufgaben und Lehrsaetze ausder Analysis, 1, 2nd ed. Springer, Berlin (1954).
Rohlin, V. A.
[1] In General a Measure-Pr~serving Transformation Is Not Mixing,
Dokl. Akad. Nauk. 13 (1949) pp. 329-340.
.
. [2] .On Endomorphisms of Compact Commutative Groups, Izvestia Math.
Nauk. 13 (1949) pp. 329-340.
[3] On the Fundamental Ideas of Measure Theory, Mat. Sbomik (N. S.)
25,67 (1949) Pi>. 107-150. [Transl. Am. Math. Soc. 1,10 (1962) pp.
1-54.]
[4] Exact Endomorphisms of Lebesgue Spaces, Izvestia Akad. Nauk.
25 (1961) pp. 499-530. [Transl. Amer. Math. Soc. 2, 39(1964)pp.1-36.l
Rohlin, V. A. and Sinai, Ya;
[5] Construction afld Properties of Invariant Measurable Partitions,
Dokl. Akad. Nauk. 141 No.6 (1961) pp. 1038--1041.
[Sov. Math. Dokl. 2 No.6 (1961) pp. 1611-1614.]
.
Schwartzman, S.
[1] Asymptot'ic Cycles, Ann. Math. 66 No.2 (1957) pp. 270-284.
Siegel,
[1]
[2]
[3]
C. L.
Iterations of Analytical Function, Ann. Math. 43 (1942) pp.607-612.
Vorlesungen iiber Himmelsmechanik, Springer, Berlin (1956).
Uber die Exisfenz einer Normalform analytisclter Hamiltonsche
Differentialgleichungen in der Nahe einer GleichgewichtslOsung,
Math. Annalen. 128 (1954) pp. 144-170.
280
Sinai. Ya.
[1] The Central Limit Theorem for Geodesic Flows on Manifolds of
Constant Negative Curvature, Dok!. Akad. Nauk. 133 (1960) pp.
1303-1306. [Sov. Math. Dok!. I No.4 (1961) pp. 983-987.]
[2] Properties of Spectra of Ergodic Dynamical Systems, Dokl. Akad.
Nauk. 150 (1963) pp. 1235-1237. [SOy. Math. Dok!. 4 No.3 (1963)
pp. 875-877.]
[3] Some Remarks 'on the Spectral Properties of Ergodic Dynamical
Systems, Usp. Mat. Nauk. 18 No.5 (1963) pp. 41-54. [Russian Math.
.
SurVeys 18 No.5 (1963) pp. 37-51.]
[4] On the Foundations of the Ergodic Hypothesis for a Dynamical
System of Statistical Mechanics, Dokl. Akad. Nauk. 153 No.6
(1963). [SOy. Math. Dokl. 4 No.6 (1963) pp. 1818-1822.]
.
[5] Vestnik Moscovskovo Gosudrastvennovo Universitata. Series Math .
. No.5 (1962).
[6] Dynamical Systems with Countably Lebesgue Spectra, Izvestia
Math. Nauk. 25 (1961) pp. 899-924. [Transl. Amer Math. Soc.,
Series 2, 39 (1961) pp. 83-110.]
[7] On the Concept of Entropy of a Dynamical System, Dokl. Akad.
Nauk. 124 (1959) pp. 768-771. [Math. Review 21 No.2036a.]
[8] Letter to the Editor, Usp. Math. Nauk. 20 No.4 (124) (1965) p. 232.
[9] A Weak Isomorphism of Transformations with an Invariant Measure,
Dok!. Akad. Nauk. 147 (1962) pp. 797-800. [SOy. Math. Dok!. 3
(1962) pp. 1725-1729.]
[10] Geodesic Flows on Compact Surfaces of Negative Curvature, Dokl.
Akad. Nauk. 136 (1961) p. 549. [SOy. Math: Dok!. 2 No.1 (1961)
pp. 106-109.]
[ll] Dynamical Systems with Countably Multiple Lebesgue Spectra II.
Izvestia Math. Nauk. 30 No.1 (1966) pp. 15-68.
Sitnikov, K.
[1] The Existence of Oscillatory Motions in the Three-Body Problem,
Dokl. Akad. Nauk. 133 No.2 (1960) pp. 303-306. [SOy. Phys. Dokl.
5 (1961) pp. 647-650.]
Slater, N. B.
[1] Distribution Problems and Physical Applications, Compositio
Mathematica, 16 fasc. 1.2 (1963) pp. 17&-183.
Smale, S.
[1] DynamiCal Systems and the Topological Conjugacy Problem for
Diffeomorphisms, Proc. Int. Congress Math. (1962) pp. 490-496.
BIBLIOGRAPHY
281
[2] Structurally Stable Systems Are Not Dense, Am. ]. Math. 88 (1966)
pp. 491-496.
[3] Differentiable Dynamical Systems, Rull. Am. Math. Soc. n (1967)
pp. 747-817.
Slepin, A. M.
[1] O~ th~ Approximation of Dynamical Systems by Periodic Mappings
and Its Spectrum, Funkz. Analys i ego Prilojenija, Moscos 1 No.2
(1967).
Weyl, H.
[I] Sur une application de la theorie des nombres fI la mecanique statistique, Enseignement Math. 16 (1914) pp. 455-467.
[2] Uber die Gleichveirteiligung von Zahlen mod. 1, Math. Annalen 77
(1916) pp. 313-352.
[3] Selecta Hermann Weyl, Bassel-Stuttgart (1956) pp. 11-147.
[4] Mean Motion I, Am. ]. Math. 60 (1938) pp. 889-896.
[5] Mean Motion II, Am. ]. Math. 61 (1939) pp. 143-148.
Winlner, A.
[I] Upon a Statistical Method in the Theory of Diophantine Approximations, Am. ]. Math. 55 (1933) pp. 309-331.
Yaglom, A. M. and Yaglom. I. M.

[1] Probabilites et informations, Dunod (Paris).
INDEX
Bernoulli scheme, 7, 11, 21, 30, 32,

39,42
Birkhoff problem, 86
Birkhoff theorem, 16
Boltzmann-Gibbs model, 34, 76
A- Entropy, 48
Abstract dynamical system, 7
Action, adiabatic invariance, 97

Action-angle coordinates, 94, 211,
213
Algebra of measurable sets, lS3
inclusion, 153
intersection, 153
subalgebra, 153
sum, 153
Anosov theorem, 67, 201
Asymptote, 172, 179
negative, 61, 180
positive, 61, 180
Asymptotic theory of the k-th approximation, 85
Automorphism, 124
Averaged system (see System of
evolution)
Averaging method, 101, 103, 227
C-Flow, 55, 69, 189

C-Systems, 53, 191
cr-Topology, 65
Canonical mapping, 231, 235
global, 99, 243
infinitesimal, 237
Classical systems, 1
Compact manifold of negative curvature, 60, 168, 178
Conservation of energy, 4
Contracting space, 55
Convex billiard table, 232
Decomposable system, 17
Density point, 18
Di1ati~g space, 55
Discrete spectrum, 25, 48, 95, 147
Baker's transformation, 8, 12, 33,

125
283
284
INDEX
Discrete spectrum
theorem, 27
Geodesic flow
of the torus, 3, 12, 18, 94, 117
Dynamical system, 1
Elliptic point, 218
Endomorphism, 124, 143
Entropy, 24, 35
of an automorphism, 40, 163
of a classical system, 46
of a flow, 45
of partition, 36
of a partition with respect to an
automorphism, 38
of a partition a with respect to a
partition (3, 37, 158
Equipartition, 129, 135
Ergodic system, 12, 26, 134
classical, 147
Ergodicity, 16, 20, 24
EuleroPoinsot motion, 119
Fibre bundle, 57
Foliation, 59, 185
absolutely continuous, 75
co.ntracting, 73
dilating, 73
ergodic, 73
Generating function, 99, 235, 243
Generator, 41, 163
Generic elliptic type, 219
Generic property, 12
Geodesic flow, 3, 18, 151
of the ellipsoid, 3, 94, 96
of the Lie groups, 3, 120
of the manifolds of negative curo
vature, 34, 60, 76, 136
Hadamard method, 113

Hamiltonian flow, 3, 12, 18, 104,
106
global,S, 122
Hamiltonian function (see Generating function)
Homomorphism, 124
Horocycle, 173
Horocyclic flow, 45, 49
Horosphere
negative, 6~, 181
positive, 61, 180
Hyperbolic point, 218
Hyperbolic rotation, 216
with reflection, 216
Induced operator, 22, 147
Instability (zones of), 90, 108
Integrable system, 94, 108, 210
Invariant torus, 93, 97
Involution (integrals in), 210, 212
Isometry, 21
Isomorphism, 11, 27, 124
Jacobi field, 187
Jacobi theorem, 2, 115
KoSystem, 32, 43, 70, 91

Kolmogorov theorem, 41, 166
Kouchnirenko theorem, 6
Krein theorem, 221
Lagrange problem, 9, 135, 138
Lebesgue space, 7
285
INDEX
Lebesgue spectrum, 28, 34, 154

multiplicity, 29
simple, 29
Periodic point, 198

Periodical apporximation speed, 49
Perturbation (theory of), 85, 100
Levi-Civita theorem, 220
Poincare invariant, 235

Poincare-Lyapounov lemma, 221
Poisson bracket, 210, 212
Proper function, 26, 147
Proper value, 24, 26
Properly continuous spectrum, 25
Properly discrete spectrum, 25
Lie commutator, 240

Linear oscillation, 4
Linear symplectic mapping,
negative proper value of, 224
parametrically stable, 221
positive proper value of, 224
stable, 221
Liouville theorem, 3, 4, 119
(see also Integrable Systems)
Lobatchewsky-Hadamard theorem,
60, 178
LobatchewskY-Poincare plane, 169
Mean motion (see Lagrange problem)

Mean sojourn time, 10
Measure, 123
Measure space, 123
Mixing, 12, 19, 25, 26, 30
nfold mixing, 22
semigroup, 144
Moser theorem, 97, 220, 256
Quasi-periodic motion, 1, 93, 97,

119, 210
Resonance, 101, 104

Rotation of a solid, 96, 120
(see also Euler-Poinsot motion)
Sheets, 59, 185

contracting, 189
dilating, 189
Sinai theorem, 62, 191, 207
Singular spectrum, 91
Skew-product, 24, 145
Skew-scalar product, 211, 223
Smale example, 56
Parabolic point, 218
Smale nontoral C-system, 194

Smale theorem, 196
Small denominator, 251
Space mean, 15, 127
Parametric resonance, 87, 221

Partition, 158
algebra generated, 159
refinement, 158
sum, 158
Pendulum, 81, 121
Spectral invariant, 22
Spectral measure, 24
Spectral multiplicity, 24
Stability (theory of), 85
Stability of fixed point, 219
Stable system, 81
Newton method, 249
286
INDEX
Structural stability, 64
of C-system, 64
problem (see Smale theorem)
Surface of section, 98, 230
Swing, 81
Symplectic
manifold,S
mapping, 212, 215, 221
(see also Linear symplectic
mapping)
System of evolution, 101, 108
Three-body problem, 12, 96, 97

108
TimE' mean, 15, 127,134
Topological index, 218
Topological instability, 109
Torus
automorphism,S, 11, 21, 28, 30,
34, 43, 53, 62, 70, 127
Torus
translation, I, 5, 11, 16, 18, 21,
115, 130, 132, 135
Transition chain, 111
Transition torus, 110
Unsta~le
system, 53
Variation of frequency, 259
Weak mixing, 21, 25

Weyl theorem (see Equipartition)
Whisker
arriving, 110
departing, 110
Whiskered torus, 109
Winding number, 148
group of, 149

Arnold - Ergodic Problems of Classical Mechanics

Caricato da

Informazioni sul documento

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Arnold - Ergodic Problems of Classical Mechanics

Caricato da

Copyright:

Formati disponibili

ERGODIC PROBLEMS

THE MATHEMATICAL PHYSICS MONOGRAPH SERIES

Ralph Abraham, Princeton University

Vladimir I. Arnold, University of Moscow

Freeman .J. Dyson, The Institute for Advanced Study

Robert Hermann, Argonne National Laboratory

Rudolph C. Hwa, State University of New York at Stony Brook

.John R. Klauder and E. C. G. Sudarsban, Syracuse University

Andre Lichnerowicz, College de France

George W. Mackey, Harvard University

Roger G. Newton, Indiana University

R. F. Streater, Imperial College of Science and Technology

ERGODIC PROBLEMS OF CLASSICAL MECHANICS

Library of Congress Catalog Card Number 68-19936

The manuscript was put into production on October 10,1967;

The fundamental problem of mechanics is computing, or studying

The Jacobi Theorem

18. Conditional Entropy of a Partition a with

is usually defined in local coordinates by:

xi = fi(x 1 , ... , JIl),

is the discrete group generated by a measure-preserving dif-

Then the system is merely denoted by (M, /l.,

is called the automorphism.

1.2. QUASI-PERIODIC MOTION

Let M be the torus I (x, y) mod

The measure is dx dy, the group t

where a ( R, and dot denotes d/dt. Assume a

ERGODIC PROBLEMS OF CLASSICAL MECHANICS

initial data x(O) = x o' y(O) = Yo has the form:

As x = Xo + q, y takes the value Yo + p and the corresponding point on

closed orbits. If a is irrational, each orbit is everywhere dense (Jacobi,

one-parameter group of translations defined by:

;=l, ... ,n; ,wlR n

Every orbit of t is everywhere dense if, and' only if,

1.4. GEODESIC FLOWS

Let V be a compact Riemannian manifold; M

tary tangent bundle. Given a unit tangent vector ~

Tl V denotes its unif

The group G t is called the geodesic flow of V. It can be proved that

SOME MORE EXAMPLES 1.7

. Appendix 2 describes the geodesic flow of the usual torus immersed in

Other mechanical systems

involve more general flows.

1.8. HAMILTONIAN FLOWS

Let PI"'" Pn; ql"'" qn (in short: p, q) be a coordinate system in

H(p, q) a smooth function.

define a one-parameter group of diffeomorphisms of R2n. This group is

ERGODIC PROBLEMS OF CLASSICAL MECHANICS

The Hamiltonian flow preserves the measure dP l

The divergence of the vector field (1. 9) vanishes:

THEORil!M OF CONSERVATION OF ENERGY

The function H is a first integral of (1.9).

Lee us denote a subset H(p, q)

h ( R by M. For almost every h,

M is a manifold. This manifold is invariant under the flow.

There exists an invariant measure on the manifold M.

The invariant measure on M is defined by:

where a is the volume element of M induced by the metric of R2n. If(1.9)

hI'"'' Ik = hkJ where the h's are constants.

1.13. LINEAR OSCILLATIONS IN DIMENSION 2

The Hamiltonian is:

Equation (1. 9) has two first integrals:

The corresponding manifolds II

hI' 12 : h2 are two-dimensional tori.