Dynamical systems

Oliver Knill
Harvard University, Spring semester, 2005

This course Math 118r was taught in the spring 2005 at Harvard University. The first lecture
took place February 2 2005, the last lecture on May 6, 2005. There were 13 weeks. Except of the
first week with an introduction and the last week with a final quiz and project presentation, the
course covered each week a different and independent topic.

Update, August, 2019. This document is placed in the public domain.

 INTRODUCTION Math118, O. Knill

ABSTRACT. We discuss the methodology and organization of the course.

The subject. Dynamical system theory has matured into an independent mathematical subject. It is linked to
many other areas of mathematics and has its own AMS classification which is 37-xx. The subject has grown so
fast, that already specific subareas of dynamical systems like one-dimensional dynamics or ergodic theory have
become independent research areas. As in other mathematical subjects, like topology, geometry or analysis which
have ”settled down”, the teaching of the subject from the bottom up needs a lot of time. It makes more sense to
study the subject by picking a few interesting subtopics.

The case method. This course is taught with an adaptation of the ’case method’. Each week, we pick a topic
and use it to discuss some aspect of dynamical systems theory. The advantage of the ’case method’ approach is
that one can start early with mentioning open research topics. Furthermore, there are frequent fresh starts. We
used this style for a course called ”Mathematical Chaos Theory” in 1994 at Caltech, where an integral part were
computer demonstrations using Mathematica and special software. It was also the first course, where I had used
a course web-site.

The ”case method” style has been used in Mathematics for a long time. Examples are the booklet pearls of
number theory by Khinchin or Bowen’s lectures in dynamical systems theory. The case method is a traditional
Russian presentation style which can be found in many books. It is also used in research summer schools, where
the breakup into different subjects and lectures comes naturally.

Systematic approach Case approach

The history of a mathematical subject. Each part of the course has its own theme and flavor and is labeled by
the name of a ”protector”, which is either a mathematician or physicist. We try to keep each subject independent
of the others but of course, we will cross reference and relate to older topics. We also aim to give a glimpse into
the history and gossip of the given subject. Because many different topics are covered, you will be able to get an
idea, what dynamical systems is about and pick your favorite theme for a final project. which can either be of
experimental or theoretical nature.

The level of difficulty. The course should be attractive for people who are interested in the applications of
dynamical systems theory as well as for students, who want to see more mathematics beyond calculus. Some of
the mathematical facts mentioned in class will be proven in full mathematical rigor and illustrated with live exper-
iments in class. Participants of the course will be provided tools to experiment using online applications, computer
algebra systems or using their own favorite programming language. No programming knowledge is required. More
theoretically inclined or application oriented students will be given the opportunity to read some hand-picked sur-
vey articles if they wish.

Other fields Many introductory books on dynamical systems theory give the impression that the subject is about
iterating maps on the interval, watching pictures of the Mandelbrot set or looking at phase portraits of some
nonlinear differential equations in the plane. This is far from the reality. The topic can be seen as an interdis-
ciplinary approach to many mathematical and nonmathematical areas. The field has matured and is successfully
used in other fields like game theory, it is used to approach difficult unsolved problems in topology, and helps to
see number theoretical problems with different eyes. There is hardly any mathematical field, which is not involved.
For example: iterating smooth map or evolving smooth flows on manifolds is rooted in geometry, a sequence of
independent random variables in probability theory can be modeled as a Bernoulli shift, the law of large numbers
a special case of the ergodic theorem, the learning process in artificial intelligence can be seen as a discretized
gradient flow. Dynamical systems are used heavily in number theory. For example, to understand the frequency
of decimal digits occurring in the real number π = 3.14159..., where a dynamical systems approach looks the most
promising one. The practical applications of the theory of dynamical systems are enormous: it ranges from medical
applications like bifurcations of heartbeat patterns to explain the synchronous rhythmic flashing of fireflies. And
then there are the obvious applications in population dynamics, fluid dynamics, quantum dynamics or statistical
mechanics.

Prerequisites. To follow this course, a one semester multi-variable calculus like math21a, applied math21a,
math23b, as well as a one semester of linear algebra course like math21b, applied math23b, math23b is required.

Exams. We plan to do several small quizzes. This, the homework, a final project and participation will make up
the grade.

Syllabus. The difficulty and pace of the course will somehow be adjusted according to the audience. For a modular
course like that, the theme structure allows an easy adaption of the pace.

• 1. Week: Introduction.

– What are dynamical systems?

– Organization of the course
– Examples of dynamical systems
• 2. Week: Feigenbaum: maps in one dimensions.

– Maps on the interval

– Periodic points and their stability.
– Bifurcation of periodic points
– The dynamical zeta function
– Invariant measures
– The Lyapunov exponent
• 3. Week: Henon: maps in two dimensions.

– Area preservation
– Periodic points and their nature
– Stable manifold theorem and homoclinic points
– Construction of stable manifolds
– Lyapunov exponents and random matrices
– Definitions of chaos
• 4. Week: Hilbert: Differential equations in two dimensions
– Differential equations in the plane and torus
– Poincare-Bendixon theorem
– Limit cycles
– Hopf bifurcations
– The Hilbert problem on limit cycles
• 5. Week: Lorentz: ODEs in higher dimensions
– Differential equations in space
– The attractor in the Lorentz system
– Forced oscillators
– Lyapunov functions for ODE’s
– Strange attractors
• 6. Week: Birkhoff: billiards
– Billiards
– The variational setup
– Existence of periodic points
– Polygonal billiards
– Chaos for the stadium billiard
• 7. Week: Hedlund: cellular automata
– Curtis-Hedlund-Lyndon theorem
– Topological entropy for CA
– Attractors
– Higher dimensional automata
– Special solutions
• 8. Week: Mandelbrot: maps in the complex plane
– Mandelbrot and Julia sets
– Basics of complex dynamics
– Some topological notions
– Connectivity of Mandelbrot set
• 9. Week: Bernoulli: subshifts of finite type
– Bernoulli shift
– Subshifts of finite type
– Sophic subshifts
– Normal numbers and randomness
– Normal numbers and randomness
• 10. Week: Weyl: dynamical systems in number theory
– Irrational rotation on the torus
– Dirichlet theorem
– Continued fractions
– Diophantine lattice point problems
– Unique and strict ergodicity
• 11. Week Poincare: many body problems
– The equations of the n-body problem
– Integrals and the solution of the 2 body problem
– The Sitnikov problem
– Changing into rotating coordinate system
– The planar restricted three body problem
– Non-collision singularities and special solutions
• 12. Week: Einstein: geodesic flows
– Geodesic flows examples
– Surfaces of revolutionn
– surface billiards
– Wave fronts and caustics
• 13. Week: Review
– Review
– Open problems in dynamical systems
 – Projects

The book. It is important to have a ’second opinion’ on things. We will not follow a book but the ”First course
in Dynamics, with a panorama of recent developments” by Boris Hasselblatt and Anatole Katok comes closest. It
is written by leading experts in the area of dynamical systems.

More literature suggestions can be found on the course web-site.

http : //www.math.harvard.edu/archive/118r spring 05/.

2/2/04 WHAT ARE DYNAMICAL SYSTEMS? Math118, O.Knill

ABSTRACT. We discuss in this lecture, what dynamical systems are and where the subject is
located within mathematics.

A FIRST DEFINITION.
The theory of dynamical systems deals with the evolution of systems. It describes processes
in motion, tries to predict the future of these systems or processes and understand the
limitations of these predictions.

RELEVANCE OF DYNAMICAL SYSTEMS.

To see that dynamical systems are relevant, one has just to look at a few news stories which
broke during the last few weeks:

• Tsunami damage prediction

• Roulette ball prediction
• Metor path computation
• Statistics of digits in π
• Currents in the sea
• Global earth temperature prediction
• Landing of the Cassini probe on Titan

A FANCY DEFINITION.
Mathematically, any semigroup G acting on a set is a dynamical system. A semigroup (G, ⋆)
is a set G on which we can add two elements together and where the associativity law
(x ⋆ y) ⋆ z = x ⋆ (y ⋆ z) holds. The action is defined by a collection of maps Tt on X. It is
assumed that Tt⋆s = Tt ◦ Ts , where ⋆ is the operation on G (usually addition) and ◦ is the
composition of maps.

CLASSES OF DYNAMICAL SYSTEMS:

Time G (semigroup) Action

Natural numbers (N, +) Maps
Integers (Z, +) Invertible maps
Positive real numbers (R+ , +) Semiflows (some PDE’s)
Real numbers (R, +) Flows (Differential equations)
Any group (G, ⋆) Representations
Lattice (Z n , +) Lattice gases, Spin systems
Euclidean space (Rn , +) Tiling dynamical systems
Free group (Fn , ◦) Iterated function systems

TWO IMPORTANT CASES OF ONE DIMENSIONAL TIME. We mention the general defini-
tion to stress that the ideas developed for one dimensional time generalize to other situations.
Because physical time is one dimensional, the important cases for us are definitely discrete
and continuous dynamical systems:

dynamics of maps defined by dynamics of flows defined by

transformations differential equations
DYNAMICAL SYSTEMS AND THE REST OF MATH. All areas of mathematics are linked
together in some way or an other. Intersections of fields like algebraic topology, geometric
measure theory, geometry of numbers or algebraic number theory can be considered full blown
independent subjects. The theory of dynamical systems has relations with all other main fields
and intersections typically form subfields of both.

Algebra Measure theory Analysis

Topology Probability theory Geometry
Logic Dynamics Number Theory

EXAMPLES OF INTERSECTIONS OF DYNAMICS WITH OTHER FIELDS:

• Link with algebra: group theorists often look at the action of the group on itself. The
action of the group on vector spaces defines a field called representation theory.

• Link with measure theory: in ergodic theory one studies a map T on a measure space
(X, µ). Measure theory is one foundation of ergodic theory.

• Link with analysis: the study of partial differential equations or functional anal-
ysis as well as complex analysis or potential theory.

• Link with topology: the Poincare conjecture states that every compact three dimen-
sional simply connected manifold is a sphere. The problem is currently attacked using a
dynamical system on the space of all surfaces which is called the Ricci flow.

• Link with geometry: Kleins Erlanger program attempted to classify geometries by

its symmetry groups. For example, the group of projective transformations on a projective
space. A concrete dynamical system in geometry is the geodesic flow. An other connection
is the relations of partial differential equations with intrinsic geometric properties of the
space.

• Link with probability theory: sequences of independent random variables can

be obtained using dynamical systems. For example, with T (x) = 2x mod 1 and with
the function f which is equal to 1 on [0, 1/2] and equal to 0 on [1/2, 1], f (T n (x)) are
independent random variables for most x.

• Link with logic: logical deductions in a proof or doing computations can be modeled as
dynamical systems. Because every computation by a Turing machine can be realized
as a dynamical system, there are fundamental limitations, what a dynamical system can
compute and what not.

• Link with number theory: some problems in the theory of Diophantine approxima-
tions can be seen as problems in dynamics. For example, if you take a curve in the plane
and look at the sequence of distances to nearest lattice points, this defines a dynamical
system.

• A final link: a category X of mathematical objects has a semigroup G of homomor-

phisms acting on it (topological spaces have continuous maps, sets have arbitrary maps,
groups, rings fields or algebras have homomorphisms, measure spaces have measurable
maps). We can view each of these categories as a dynamical system. One can even include
the category of dynamical systems with suitable homomorphisms. But this viewpoint is
not a very useful in itself.
2/4/04 EXAMPLES OF DYNAMICAL SYSTEMS Math118, O.Knill

ABSTRACT. In this lecture, we look at examples of dynamical systems. Most examples in this
zoo of systems belong to the ”hall of fame”. They are ”stars” in the world of all dynamical
systems and will appear later in this course.

THE LOGISTIC MAP. T (x) = cx(1 − x). This is an example of an interval map . The
parameter c is fixed in the interval [0, 4]. Lets look at some orbits. To compute an orbit
say for c = 3.0, start with some initial condition like x0 = 0.3, and iterate the map
x1 = T (x0 ) = 3x0 (1 − x0 ) = 0.63, x2 = T (x1 ) = 2x1 (1 − x1 ) = 0.6993 etc. Lets do this with
the computer. We show a few orbits for different parameters c. We always start with the ini-
tial condition x0 = 0.3. Time is the horizontal axes and the interval [0, 1] is on the vertical axes.

THE LORENTZ SYSTEM. The system of differential equa-

tions

ẋ = 10(y − x)
ẏ = −xz + 28x − y
8z
ż = xy −
3
is called the Lorentz system. We see a numerically inte-
grated orbit (x(t), y(t), z(t)) which is attracted by a set called
the Lorentz attractor. It is an example of what one calls
a strange attractor. Orbits behave chaotically on that set
in the sense that one observes sensitive dependence on ini-
tial conditions The set is also measured to be a fractal, of
dimension strictly between 1 and 2.

THE COLLATZ PROBLEM. Define a map T on the positive

integers as follows. If n is even, then define T (n) = n/2, if n 300

is odd, then define T (n) = 3n + 1. One believes that every 250

orbit n, T (n), T (T (n)) will end up at 1 but one does not have
a proof and there are people who think that mathematics is 200

not ready for this problem. Theoretically, it would be pos- 150

sible that an orbit escapes to infinity, or that there exists a

periodic orbit n, T (n), T 2 (n), ..., T k (n) = n. The problem 100

is also called Ulam problem or 3n + 1 problem. It is a 50

notorious open problem. The picture to the right shows how

long it takes to get from n to 1. 5000 10000 15000 20000 25000 30000
COMPUTING SQUARE ROOTS. Look at the map
2xy x + y
T (x, y) = ( , )
x+y 2
2

which assigns to two numbers a new pair, the harmonic

means as well as the algebraic mean. You can easily check 1.5

that the quantity F (x, y) = xy is preserved: F (T (x, y)) =

F (x, y). It is called an integral. A map in the plane with 1

such an integral is called an integrable system. All or-

bits converge to the line x = y which consists of fixed
0.5
points. Why is this useful? Start with (1, 5) for example.
The
√ sequence
√ (xn , yn ) will converge to the diagonal and so to
( 5, 5). Lets do it: we have (1, 5), ( 35 , 3), ( 15 , 7 ), ( 105 , 47 ). 0.5 1 1.5 2
√ 7 3 47 21
We know that 5 is in the interval [xn , yn ] for all n. For ex-
ample,
√ 47/21 = 2.238... is already a good approximation to
5 = 2.236....

CELLULAR AUTOMATA. Given a infinite sequence x of 0’s

and 1’s, define a new sequence y = T (x), where each en-
try yn depends only on xn−1 , xn , xn+1 . There are 256 dif-
ferent automata of this type. The picture below shows an
orbit of ”Rule 18”. One of the interesting features of this au-
tomaton is that its evolution is linear on parts of the phase
space. The nonlinear and interesting behavior is the motion
of the kinks, the boundaries between regions with linear mo-
tion. A sequence x has a kink at n, if for some k ≥ 0,
[xn−k , . . . , xn+k+1 ] = [1, 0, . . . 0, 1], like the pattern 10001.

DIFFERENTIAL EQUATIONS IN THE PLANE. Second or-

der differential equations can be written as differential equa-
tions in the plane. An example is the van der Pool oscilla-
tor
d
x = y
dt
d
y = −x − (x2 − 1)y;
dt
which shows a limit cycle. All orbits (except with the initial
condition (0, 0) converge to that limit cycle.

BILLIARDS. Let us take a table like the region x6 +y 6 ≤ 1. A

ball reflects at the boundary. What is the long time behavior
of this system? Is it possible that the angles a light ray makes
with the boundary of the table become arbitrarily close to 0
and arbitrarily close to 180 degrees? Are there paths which
come arbitrarily close to any point? The billiard flow defines
a smooth map on the annulus. The study of this system has
relations with elementary differential geometry. For example,
the curvature of the boundary plays a role. The study of
billiards is also part of a mathematical field called calculus
of variations which deals with finding extrema of functions.
STANDARD MAP. The map

T (x, y) = (2x + γ sin(x) − y, x)

on the plane is called the Standard map. Because T (x +

2π, y) = T (x, y + 2π) = T (x, y), one can take both variables
x, y modulo 2π and obtain a map on the torus. The real
number γ is a parameter. The map appeared around 1960 in
relation with the dynamics of electrons in microtrons and was
first studied numerically by Taylor in 1968 and by Chirikov
in 1969. The map can be completely analyzed for γ = 0.
The map exhibits more and more ”chaos”as γ increases. The
picture to the right shows a few orbits in the case γ = 1.3.

GEODESIC FLOW. Light on a surface takes the shortest pos-

sible path. These paths are called geodesics. On the plane,
the geodesics are lines, on the round sphere, the geodesics are
great circles, on a flat torus (see picture), the geodesics are
lines too, but they wind around the surface. On some surfaces
like surfaces of revolution or the ellipsoid, the geodesic flow
can be analyzed completely on. On other surfaces, the flow
can become very complicated. There are bumpy spheres on
which each geodesic path is dense in the sense that the curve
comes close to every point and also every direction at that
point.

THE HENON MAP. One of the simplest nonlinear nonlinear

maps on the plane is the Henon map

T (x, y) = (ax2 + 1 − by, x) .

For |b| = 1 the map is area-preserving. For |b| < 1, it con-

tracts area and produces attractors. The Henon attractor
is obtained for a = −1.4, b = −0.3. The Henon map is equiv-
alent to the nonlinear recursion xn+1 = ax2n−1 + 1 − bxn−1 .
While linear recursions like the Fibonacci recursion xn+1 =
xn + xn−1 can be solved explicitly using linear algebra, non-
linear recursions do no more lead to explicit formulas for xn .

SOLVING EQUATIONS. To solve the equation f (x) = 0 nu-

merically, one can start with an approximation x0 , then apply
the map the Newton iteration map T (x) = x− f (x)/f ′ (x).
If T (x) = x, then f (x) = 0. As long as the root y satisfies
f ′ (x) ̸= 0, this algorithm works for x0 near y. The method
also works in the complex. In the case of several roots, is
an interesting question to explore the basin of attraction
of a root. The picture to the right shows this in the case of
f (z) = z 3 −2, where one has 3 roots. Depending on the initial
point z0 , one ends up on one of the three roots. The Newton
map for polynomials f defines a rational map. Its study is
part of complex dynamics.
THREE BODY PROBLEM. Celestial mechanics determines
very much the timing of our lives. Our calendar is based on
it. While the motion of 2 bodies is understood well since Ke-
pler, the three body problem is very complicated. Part
of modern Mathematics, like topology have been developed
in order to understand it. The Sitnikov problem is a re-
stricted three body problem where the motion of a planet
moves with negligible mass in a binary star system. The two
suns circle each other on ellipses. The planet moves on the
line through the center of mass, perpendicular to the plane
in which the stars are located. For this system, there is a
mathematical proof of some chaotic motion.

EXTERIOR BILLIARS. A geometrically defined dynamical

system has been used to capture the main difficulties of the
three body problem. The system is defined by a convex table
as in billiards but this time, the a point outside the table is
reflected at the table boundary: take the tangent to the table
in the anti-clockwise direction and take the other point which
has the same distance to the touching point. The map defined
on the exterior of the table is area preserving and in general
very complicated. It is not known, whether there exists a
table for which there are unbounded orbits.

THE DIGITS OF PI. The digits of the number π =

3.14159265358979323846264338327950288419716939937510...
appear random. With T (x) = 10x mod1 and f (x) = [10x],
where [x] is the integer part of a number, the number
f (T n (x)) is the n’th digit of π. It appears that every digit
appears with the same frequency and also all combinations of
digit sequences. It is an open problem, whether this is true.
One would call π normal. The picture to the right shows the
sequence xn = f (T n (x)).

LATTICE POINTS NEAR GRAPHS. Given the graph of a

function f on the real line, one can look at the distances to
the nearest lattice points. This defines a sequence of numbers
which can be generated by a dynamical system. For polyno-
mials of degree n, the system is a map on the n dimensional
2
torus. For the parabola f (x) =[ ax] + bx[+ c we obtain
] which
x x+α
leads to a map of the type T ( )= on the two
y x+y
dimensional torus.
2/11/2005: ENTROPY AND CHAOS Math118, O. Knill

ABSTRACT. We look today at some notions of ”chaos”. One definition is the positivity of a number called the
positive entropy, an other is the positivity of the Lyapunov exponent for every orbit which is not eventually
periodic. An other definition is ”chaos in the sense of Devaney”. The Ulam map or the tent maps are examples
for which we know that this type of chaos happens.

A DEFINITION OF CHAOS.
A purely topological notion of chaos which applies also to map with no differentiability is the definition of
Devaney:

A map T : [0, 1] → [0, 1] is called chaotic, if there is a dense set

of periodic orbits and if there exists an orbit which is dense.

A set Y is dense in [0, 1] if there is no interval which has empty intersection with Y .
EXAMPLES. a) the set of rational numbers is dense in [0, 1].
b) the set of irrational numbers is dense in [0, 1].
c) The set of numbers {1/n | n = 1, 2, ...} is not dense in the interval.
d) Consider Champernown’s number x = 0.123456789101112131415161718192021222324... (do you see the
pattern?), and the map T (x) = 10x mod1. Then T (x) = 0.23456789101112131415161718192021222324... and
T 2 (x) = 0.3456789101112131415161718192021222324... etc. Can you see why the orbit of x under the map T
is dense in the interval [0, 1]?

THE ULAM MAP IS CHAOTIC. We only state this theorem now. We will put later in this course put together
some tools to prove it.

THEOREM. The map f4 (x) = 4x(1 − x) is chaotic in the sense of Devaney.

To have Devaney chaos, one needs to have an initial point, which visits
each interval as well as to find for each interval a periodic orbit which 0.8

visits that interval.

Because the Ulam map is conjugated to the tent map, we need only to 0.6

prove the claim for the tent map.

In the homework, you see the density of the periodic points by under- 0.4

standing the graphs of the iterates of the map.

0.2

The problem to construct a dense periodic point will be solved later.

0.2 0.4 0.6 0.8 1

TOPOLOGICAL ENTROPY. Let Pn (f ) be the number of periodic points of true period n. Define the topo-
logical entropy of the map as
1
p(f ) = lim sup log |Pn (f )| ,
n→∞ n
where the limits p(f ) = +∞ and p(f ) = −∞ are also allowed.
The topological entropy measures the growth of the number of periodic points. Similarly as the Lyapunov
exponent, it measures how ”complex” the map is.

EXAMPLES. The map defined by f˜(x) = 2x mod 1 has the topological entropy p(f ) = log(2) because Pn (f ) =
2n .

DYNAMICAL ZETA FUNCTION Related to topological entropy is the dynamical ζ function which is defined
as
∑∞
Pn (f ) n
log(ζf (z)) = z ,
n=1
n

where z is a real (or complex) variable. The series converges if Pn (f ) is finite for all n and for all complex
numbers |z| < e−p(f ) . If p(f ) = −∞ and Pn (f ) is finite for all n, then ζf (z) is defined for all z.
 ∑∞ n ∑∞
EXAMPLE. The dynamical
∑∞ zeta function satisfies log(ζf (z))
∑ = n=1 2n z n . Because n=1 xn = 1/(1 − x) − 1
∞
and integration gives n=2 xn /n = − log(1−x)−x we have n=1 xn /n = − log(1−x). We see that log(ζf (z)) =
− log(1 − 2z) and
ζf (z) = 1/(1 − 2z) .

EVENTUALLY PERIODIC ORBITS. If an orbit has the structure

x0 , x1 , x2 , ...., xm , xm+1 , xm+2 , ..., xm+n = xm , it is called eventually
periodic. Eventually periodic orbits appear often in dynamical systems
which are not invertible.

EXAMPLES:
1) The point x0 = 1 of the logistic map fc (x) = cx(1 − x) is eventually
periodic. It is actually eventually fixed. We have

x0 = 1, x1 = 0, x2 = 0, x3 = 0, x4 = 0

2) The point x0 = 7/10 is eventually periodic for T (x) = 1 − 2|x − 1/2|.

EVENTUALLY PERIODIC POINTS FOR THE TENT MAP.

THEOREM. x is eventually periodic if and only if x = p/q is

rational.

PROOF.
Since T (x) = 2x or T (x) = 2 − 2x we have

T (x) = integer + 2x

T 2 (x) = integer + 22 x
T n (x) = integer + 2n x
If T n (x) = T m (x) then k + 2n x = l + 2m x so that x = (k − l)/(2n − 2m ) and x is a rational number.

(ii) To see the other direction, lets assume now that x = p/q is rational. Then, T (x) = 2p/q or T (x) = 2−2p/q =
2(q − p)/q. In any case, T (x) is again of the form k/q. Repeating this argument shows that T n (x) is of the
form k/q. There are only finitely many fractions of the form k/q and x therefore has to be eventually periodic.

REMARK. It needs a bit of combinatorial thought to figure out, when an orbit is eventually periodic and when
it is actually periodic. Here is the answer (without a proof):

THEOREM. x = p/q is periodic for the tent map if and only p is

an even integer and q is an odd integer.

EXAMPLES:
1) x = 4/5 is a periodic point of period 2.
x0 = 4/5, x1 = 2/5, x2 = 4/5 etc

2) x = 5/7 is an eventually periodic point.

x0 = 5/7, x1 = 4/7, x2 = 6/7, x3 = 2/7, x4 = 4/7.

EVENTUALLY PERIODIC POINTS FOR THE ULAM MAP. The conjugation between the two maps T and
S matches periodic points of T to periodic points of S and ”eventually periodic points” of T with eventually
periodic points of S.

EXAMPLE: Because x0 = 5/7 is an eventually periodic point for the tent map, the point y0 = U −1 (5/7) =
(1 − cos(5π/7))/2 is the initial condition for an eventually periodic point for the Ulam map.

x0 = 5/7 x1 = 4/7 x2 = 6/7 x3 = 2/7 x4 = 4/7

↑U ↑U ↑U ↑U ↑U
y0 = 0.811745 y1 = 0.61126 y2 = 0.950 y3 = 0.1882 y4 = 0.61126
2/7/2005: THE LOGISTIC MAP Math118, O. Knill

ABSTRACT. Our first dynamical system is the logistic map f (x) = cx(1 − x), where 0 ≤ c ≤ 4 is a
parameter. It is an example of an interval map because it can be restricted to the interval [0, 1].

You can read about this dynamical system on pages 14-16, pages 57-60, pages 198-199 as well as from page 299
on in the book. On this lecture, we have a first look at interval maps. We will focus on the logistic map, study
periodic orbits, their stability as well as stability changes which are called bifurcations.

A FIRST POPULATION MODEL. In a simplest possible population

model, one assumes that the population growth is proportional to the 6000

population itself. If xn is the population size at time n, then xn+1 = 5000

T (xn ) = xn +axn = cxn with some constant a > 0. We can immediately

4000
give a closed formula for the population xn at time n:
3000

xn = T n (x) = cn x0 .
2000

We see that for c > 1, the population grows exponentially for c < 1, 1000

the population shrinks exponentially.

20 40 60 80 100

DERIVATION OF THE LOGISTIC POPULATION MODEL. If the

1
population gets large, food becomes sparse (or the members become too
shy to reproduce ...). In any case, the growth rate decreases. This can be 0.8
modeled with yn+1 = cyn − dyn2 . Using the new variable xn = (c/d)yn ,
this recursion becomes 0.6

xn+1 = cxn (1 − xn ) .
0.4

To the right, we see a few graphs of fc (x) = cx(1 − x) for different c’s.
0.2
The intersection of the graph with the diagonal reveals fixed points of
fc . You see that 0 is always a fixed point. The graph has there the slope
f ′ (0) = c. For c > 1, there exists a second fixed point x = 1 − 1c . 0.2 0.4 0.6 0.8 1

INTERVAL MAPS. If f : [0, 1] → [0, 1] is a map like T (x) = 3x(1 − x)

and x ∈ [0, 1] is a point, one can look at the successive iterates x0 =
x, x1 = T (x), x2 = f 2 (x) = T (T (x)), .... The sequence xn is called an
orbit. If xn = x0 , then x is called a periodic orbit of period n or
n-cycle. If there exists no smaller n > 0 with xn = x, the integer n
is called the true period. A fixed point of f is a point x such that
f (x) = x. Fixed points of f n are periodic points of period n. The fixed
points of f are obtained by intersecting the graph y = f n (x) with the
graph y = x. The iterates of an interval map can visualized with a
cobweb construction: connect (x, x) with (x, f (x)), then go back to
the diagonal (f (x), f (x)) and iterate the procedure.

STABILITY OF PERIODIC POINTS. If x0 is a fixed point of a dif-

ferentiable interval map f and |f ′ (x0 )| > 1, then x0 is unstable in
the following sense: a point close to x0 will move away from x0 at first
because linear approximation T (x) ∼ x0 + f ′ (x0 )(x − x0 ) shows that
|f (x) − x0 | ∼ |f ′ (x0 )||x − x0 | > |x − x0 | near x0 . On the other hand,
if |f ′ (x)| = 1, then x0 is stable. For periodic points of period n, the
stability is defined as the stability of the fixed point of f n . The picture
to the right shows situations, where f ′ (x0 ) < 1, f ′ (x0 ) = 1, f ′ (x0 > 1
at a fixed point. The parameter at which the stability changes will be
denoted a bifurcation.

REMEMBER THE IMPORTANT FACT: if f (x) = x is a fixed point of f and |f ′ (x)| < 1, then the fixed point
is stable. It attracts an entire neighborhood. If |f ′ (x)| > 1, then the fixed point is unstable. It repells points
in a neighborhood.
BIFURCATIONS. Let fc be a family of interval maps. Assume that
xc is a fixed point of fc . If |fc′ (xc )| = 1, then c0 is called a bifurcation
parameter. At such a parameter, the point xc can change from stable
to unstable or from unstable to stable if c changes. At such parameters,
it is also possible that new fixed points can appear. Different type of
bifurcations are known: saddle-node bifurcation (also called blue-
sky or tangent bifurcation). They can be seen in the picture to the
right). Flip bifurcations (also related to pitch-fork bifurcation)
lead to the period doubling bifurcation event seen below.

PERIOD DOUBLING BIFURCATION. Period doubling bifurcations

happen for parameters c for which (fcn )′ (xc ) = −1. The graph of fc
intersects the diagonal in one point, but the graph of fc2 which has slope
1 at x0 starts to have three intersections. Two of the intersections be-
long to newly formed periodic points, which have twice the period. To
the right, we see a simultaneous view of the graphs of fc and fc2 for
c = 2.7, c = 3.0, c = 3.3. You see that fc keeps having one fixed point
throughout the bifurcation. But fc2 , which has initially one fixed point
starts to have 3 fixed points! The middle one has minimal period 1, the
other two are periodic points with minimal period 2.

BIFURCATION DIAGRAM. The logistic map fc (x) = cx(1 − x) al-

ways has the fixed point 0. For c > 1, there is an additional fixed 0.9

point x = 1 − 1/c. Because f ′ (0) = c, the origin is stable for c < 1

0.8
and unstable for c > 1. At the other fixed point, f ′ (x) = c − 2cx =
c − 2c(1 − 1/c) = 2 − c. It is stable for 1 < c < 3 and unstable for c > 3. 0.7
The point c = 3 is a bifurcation. It is called a flip bifurcation. Because
a periodic point of period 2 is created, it is called a period doubling 0.6
bifurcation. To see what happens with the periodic point of period 2,
2 2
we look at
√ f (x) − x = c x(1 − x)(1 − cx(1 − x)) − x which has the roots 0.5

(c + 1 ± (c − 3)(c + 1))/(2c) which are real for c > 3. Its stability can
be determined with (f 2 )′ (x) = f ′ (x)f ′ (f (x)) 2
√ = 4 + 2c − c . This shows
√ 2.9 3.1 3.2 3.3 3.4 3.5
that the 2-cycle is stable for 3 < c < 1 + 6. At the parameter 1 + 6
it bifurcates and gives rise to a periodic orbit of period 4.

FEIGENBAUM UNIVERSALITY. √ We have computed the first bifur-

cation points c1 = 3, c2 = 1 + 6. The successive period doubling
ck+1 −ck
bifurcation parameters ck have the property that ck+2 −ck+1 converges to
a number δ = 4.69920166. It was Mitchell Feigenbaum, who realized
that this number is universal and conjectured how it could happen us-
ing a renormalization picture. In 1982, Oscar Lanford proved these
Feigenbaum conjectures: for a class of smooth interval maps with a sin-
gle quadratic maximum, the limit δ exists and is universal: that number
does not depend on the chosen family of maps. It works for example also
for the family fc (x) = c sin(πx). The proof demonstrates that there is a
fixed point g of a renormalization map Rf (x) = αf 2 (αx) in a class
of interval maps. The object which is mapped by R is a map!

HISTORY. Babylonians considered already the rotationf (x) = x +

α mod 1 on the circle. Since the 18th century, one knows the Newton-
Rapson method for solving equations. Already in the 19th century
Poincaré studied circle maps. Since the beginning of the 20th century,
there exists a systematic theory about the iteration of maps in the com-
plex plane (Julia and Fatou), a theory which applies also to maps in the
real. In population dynamics and finance growth models xn+1 = f (xn )
appereard since a long time. It had been popularized by theoretical bi-
ologists like Robert May in 1976. Periodic orbits of the logistic map
were studied for example by N. Metropolis, M.L. Stein and P.R. Stein
in 1973. Universaility was discovered numerically by Feigenbaum (1979)
and Coullet-Tresser (1978) and proven by Lanford in 1982.
 CHECKLIST Lecture of 2/9/2005 Math118, O.Knill

Checklist lecture of January 9th, 2005. (Some of the material might only be done on January
11’th depending on how much discussion we do on Wednesday). We introduce the Lyapunov
exponent of an orbit as the exponential growth rate of (f n )′ .
We can compute the Lyapunov exponent in the case of periodic orbits and in the case of the
tent map.
Then, we show that the tent map is conjugated to the Ulam map f4 (x) = 4x(1 − x).

• know the definition and the meaning of the Lyapunov exponent and how one computes
the Lyapunov exponent.

• understand the relation between Lyapunov exponents and stability in the case of periodic
orbits.

• see the conjugated maps have the same dynamics and the same Lyapunov exponent

• see the conjugation in the concrete situation of the Ulam and tent map.

Remark to a definition which we might not have much time to discuss this week:
The dynamical zeta function of f is defined as
∑
∞
Pn (f )
log(ζf (z)) = zn ,
n=1
n

where Pn is the number of periodic points of f with period n. One can compute
the dynamical zeta function of the Ulam map using the result shown in [1.5].
This leads to a closed formula.∑ ∑
Pn (f ) n−1 Pn (f ) n
(Note that d/dz log(ζf (z)) is ∑ ∞
n=1 z
which is ∞n=1 z
. Use the for-
∞ n
mula for the geometric sums n=1 a = 1/(1 − a).)

It might be a bit strange why one should consider this function. It was introduced
by Artin-Mazur in 1965 and Smale in 1967. For some dynamical systems like the
tent map one can explicitely compute the function. There are many reasons,
why one wants to study this object, algebraic reasons, there are connections with
statistical mechanics and the topological entropy which is defined as
1
p(f ) = lim log |Pn (f )| .
n→∞ n
 HOMEWORK 1. Week lecture 2/9/2005 Math118, O.Knill

This is the second part of the homework. All the five problems [1.1] until [1.5]
are due next Monday the 14th of february

1.4 We define the map f (x) = 5x + sin(πx) mod 1 on the interval [0, 1].

Sideremark: Because after identifying 0 and 1, the interval closes to a circle,

the map can be considered a smooth map on the circle. f is an example of
a circle map.

a) What is the Lyapunov exponent of the orbit of the map f wich an initial
condition x0 = 1/2?

b) Verify that the Lyapunov exponent of every orbit of f is positive.

1.5 We have shown that the Ulam map f4 (x) = 4x(1 − x) is conjugated to the
tent map T (x) = 1 − 2|x − 1/2|
a) Draw the graph of the the iterates T 2 (x), T 3 (x) of the tent map. Use the
fact that the tent map is piecewise linear.

b) Use the conjugation result to sketch the graphs of the second iterate
f42 (x) and third iterate f43 (x) of the Ulam map.

c) Conclude that f4n has 2n fixed points and therefore, that the Ulam map
f has 2n periodic points of period n.

d) What is the Lyapunov exponent of a periodic point of period n?

2/9/2005: LYAPUNOV EXPONENT Math118, O. Knill
ABSTRACT. We demonstrate that the logistic map f (x) = 4x(1−x) is chaotic in the sense that the Lyapunov
exponent, a measure for sensitive dependence on initial conditions is positive.

LYAPUNOV EXPONENT. For an orbit of f with starting point x, we

define the Lyapunov exponent as
1
λ(f, x) = limn→∞ n log |(f n )′ (x)|

where (f n )′ is the derivative of the n’th iterate (f n ). Remark. It

turns out that usually, the limit exists. If not, one should replace lim
with lim inf, the smallest accumulation point of the sequence. Choosing
lim inf instead of lim sup has nicer analytic properties.

A BETTER FORMULA. The function f n (x) becomes complicated already for small n. The following formula
is more convenient to compute the Lyapunov exponent of an orbit through x0 :

(f n )′ (x) = f ′ (xn−1 )...f ′ (x1 )f ′ (x0 )

PROOF. Use induction: for n = 1, the claim is obvious. If we differentiate f n (x0 ) = f n−1 (f (x0 )) we get
(f n−1 )′ (x1 )f ′ (x0 ), then use the induction assumption (f n−1 )′ (x1 ) = f ′ (xn−1 )...f ′ (x1 ). Therefore:

1
∑n−1
λ(f, x) = limn→∞ n k=0 log |f ′ (xk )|

EXAMPLE. For the logistic map, we compute the Lyapunov exponent by taking a large n and form
1
[log |c(1 − 2xn−1 )| + log |c(1 − 2xn−2 )| + · · · + log |c(1 − 2x0 )|] .
n

WHAT DOES THE LYAPUNOV EXPONENT MEASURE? If x and y are close, then |f (y)−f (y)| ∼ |f ′ (x)||x−
y|, if x and y are close. because Taylors formula assures f (y) = f (x) + f ′ (x)(y − x) plus something of the order
(y − x)2 . If xn is the orbit of x and yn is the orbit of y, then for a fixed n, we have |xn − yn | ∼ |(f n )′ (x)||x0 − y0 |
if x0 and y0 are close together.
The Lyapunov exponent is a quantitative number which indicates the sensitive dependence on initial
conditions. It measures the exponential rate at which errors grow. If the Lyapunov exponent is log |c| then
you can expect an error cn ϵ after n iterations, if ϵ was the initial error.

0.75

EXAMPLE. We will see below that the Lyapunov exponent of f (x) = 0.5
4x(1 − x) is log |2|. If your initial error is ϵ = 10−16 , then we have after
0.25
n iterations an error 2n ϵ which is of the order 1 for n = 53. To the right
we see the difference xn −yn between two orbits of the map f = f4 which 20 40 60 80

have an initial condition |x0 − y0 | = 10−16 . You see that after about 50 -0.25

iterations, the error has grown so much that it becomes visible. -0.5

-0.75

-1

LYAPUNOV EXPONENT OF PERIODIC ORBIT. If x0 , x1 , ..., xn = x0 is a periodic orbit of period n, then

1 n ′ 1
λ(f, x) = (f ) (x) = (log |f ′ (xn−1 )| + log |f ′ (xn−2 )| + ... + log |f ′ (x0 )|) .
n n
PROOF. We we have to show that the sequence sk = k1 (log |f ′ (x0 )| + log |f ′ (x1 )| + ... + log |f ′ (xk )|) converges
to the right hand side which is sn . If k is a multiple of n, then sk = sn . If M is the maximum of all the numbers
log |f ′ (xi )|, then |sj | ≤ jM for k = 1, ..., n and sk − λ(f, x) ≤ (nM )/k.

EXAMPLES.

• The Lyapunov exponent of the fixed point 0 is log(c). It is negative for c < 1 and positive for c > 1.
• The Lyapunov exponent of the fixed point 1 − 1/c of the logistic map fc is log |f ′ (1 − 1/c))| = log |2 − c|.
LYAPUNOV EXPONENT OF AN ATTRACTIVE PERIODIC ORBIT.
The Lyapunov exponent of an attractive periodic orbit is negative.

1
PROOF. We have λ(f, x) = n log |(f n )′ (x)|. We have seen that for an attractive periodic orbit, |(f n )′ | < 1.

It follows that the Lyapunov exponent of an orbit which is attracted to a periodic orbit is negative too.

LYAPUNOV EXPONENT AND BIFURCATION. A periodic point can only bifurcate if its Lyapunov exponent
is zero.

LYAPUNOV EXPONENT OF THE LOGISTIC MAP.

The picture to the right shows the Lyapunov exponent of an orbit
starting at x0 in dependence of c. You see that this graph looks very
complicated. If the Lyapunov exponent is negative, we typically have
an attractive periodic orbit.

It is difficult to say something about the Lyapunov exponent of a spe-

cific parameter. We know what happens for c = 4 and we know what
happens in case of an attractive periodic orbit. If an attractive periodic
orbit exists, there is an entire interval, where the Lyapunov exponent is
negative. It has only recently been shown that there is a dense set of
parameters for which the Lyapunov exponent is negative. This means,
we don’t find a single interval in [0, 4] on which the Lyapunov exponent
is positive.

CONJUGATION OF MAPS. Two interval maps T and S are conjugated, if there exists an invertible map U
from the interval onto itself such that T (U (x)) = U (S(x)). If both maps are differentiable maps, one usually
requires the map U to be smooth too.

COROLLARY: The Lyapunov exponents of corresponding orbits of two

conjugated interval maps are the same.

More precisely, if λ(f, x) is the Lyapunov exponent of the orbit of f through x, and λ(g, y) is the Lyapunov
exponent of the orbit of g through y = h(x), and gh(x) = hf (x)) is the conjugation, then λ(f, x) = λ(g, y).

Proof: This is an application the chain rule.

ULAM AND TENT ARE CONJUGATED MAPS: the Ulam map

T (x) = 4x(1−x) is conjugated to the tent map S(x) = 1−2|x−1/2| with 1

the conjugation U (x) = 21 − π1 arcsin(1−2x) and U −1 (x) = 21 − 12 cos(πx).

0.8

−1
PROOF. To check that U T U (x) = S(x), we show U T (x) = S(U (x).
0.6
One can get rid of the absolute value by distinguishing the cases x > 1/2
and x < 1/2. We have U (T (x)) = 12 − (arcsin(1 − 8(1 − x)x)))/π and
0.4
SU (x)) = 1 + 2(arcsin(1 − 2x))/π for x ∈ [1/2, 1]. To verify the identity,
d
we check that both sides are 1 for x = 1/2 and that dx arcsin(1 − 8x +
0.2
2 d
8x ) = − dx 2 arcsin(1−2x). The last identity is best checked by squaring
√
both sides and using arcsin′ (x) = 1/ 1 − x2 . The identity on [0, 1/2] is
0.2 0.4 0.6 0.8 1
solved in the same way.

LYAPUNOV EXPONENT OF THE ULAM MAP.

THEOREM. For all but a countable set of initial conditions x0 , the

Lyapunov exponent of f (x) = 4x(1 − x) with initial condition x0 is
equal to log(2).

The tent map S(x) = 1 − 2|x − 1/2| is piecewise linear. The derivative S ′ (x) is either 2 or −2. Since
log |S ′ (xk )| = log(2), the map has the Lyapunov exponent log(2c) for orbits, which do not hit one of the
discontinuities. Most initial points do not hit the discontinuity because there is only a countable set of initial
conditions for which this can happen.

Because the map is conjugated to T4 , the Lyapunov exponent of f4 is log(2) too by the corollary.
INTEGRABLE MAPS IN TWO DIMENSIONS Math118, O. Knill

ABSTRACT. A map T in the plane is called integrable, if there is a non-constant continuous function F (x, y)
which is invariant under T . We give examples of integrable maps.

INTEGRABILITY. A map T is called integrable, if there exists a real valued continuous function F (x, y)
called integral for which the level sets F = c are curves, or points and for which the identity

F (T (x, y)) = F (x, y)

holds for all (x, y).

EXAMPLES.
1. Let T (x, y) = (cos(α)x − sin(α)y, sin(α)x + cos(α)y) be a rotation in the
plane. It is integrable: the function F (x, y) = x2 + y 2 is an integral.

2. The map T (x, y) = (3x, y/3) is integrable with integral F (x, y) = xy.

3. The map T (x, y) = (x + sin(y), y) is integrable with integral F (x, y) = y.

4. The cat map T (x, y) = (2x + y, x + y) on the two dimensional torus is not
integrable as you will demonstrate in a homework.

AN EXAMPLE FROM PHYSICS.

THEOREM. For every smooth function F , we can find a map,

which has this function as an integral.

d d
Consider the system of differential equations dt x = Fy (x, y), dt y = −Fx (x, y).
By the chain rule, we have
d d d
F (x(t), y(t)) = Fx (x(t), y(t))
x(t) + Fy (x(t), y(t)) y(t)
dt dt dt
d d d d
= − y(t) x(t) + y(t) x(t)
dt dt dt dt
so that F does not change along a solution of the system. Define the map

T (x, y) = (x(1), y(1))

if x = x(0), y = y(0). This map has F as an integral.

In physics, the function F is often called the energy or Hamiltonian of the

system. The fact that F is an integral is then energy conservation. For example,
for F (x, y) = cos(x) + y 2 /2, one obtains the energy of the pendulum. The
d d
differential equations are then dt x = y, dt y = sin(x). They are equivalent to
2
d
the Newton equations dt2 x = sin(x). We will look at differential equations in
the plane in the next week.

BIRKHOFF ON INTEGRABILITY. Like ”Chaos”, ”Integrability” is a mathe-

matical term, which has many different definitions. One has to specify what one
means with ”integrable”. The fact that one has to deal with several different
definitions for integrability” expressed Birkhoff in the following way ”When,
however, one attempts to formulate a precise definition of integrability, many
possibilities appear, each with a certain intrinsic theoretical interest”. Birkhoff
suggested his own (as he admits not very precise) definition of integrability:
there exists a finite set of periodic orbits, around which the formal series de-
velopment converge and which allow to represent any solution of the system.”
This Birkhoff integrability is probably hard to check in a specific applications.
THE SBKP MAP. For |k| < 1, lets call the map

T (x, y) = (2x + 4 · arg(1 + k · e−ix ) − y, x)

on the torus the Suris-Bobenko-Kutz Pinkall map. It had been found

by Bobenko,Kutz, Pinkall and independently by Suris. Even so the map uses
complex numbers for its definition, it is real. The argument arg(z) of a complex
number z = x + iy = r cos(α) + ir sin(α) = reiα is defined as the angle α.

THEOREM. The SBKP map is integrable.

PROOF. The function

F (x, y) = 2(cos(x) + cos(y)) + k · cos(x + y) + k −1 · cos(x − y)

is an integral. It is not easy to verify that. Don’t ask how F was found!

THE COHEN-COLINE-DE VERDIERE MAP. The map

√
T (x, y) = ( x2 + ϵ2 − y, x)

in the plane is called the Cohen-Coline-de Verdiere map. By rescaling

coordinates in R2 , we can assume ϵ = 0 or ϵ = 1. For ϵ = 0, the map has the
form
T (x, y) = (|x| − y, x) .
We call it the Knuth map.

THEOREM (KNUTH) The Knuth map is integrable.

PROOF. We check that T 9 = Id. Note that the map is piecewise linear, we
only have to look at the orbits of the x axes to understand the entire picture.
Actually, every orbit is periodic with period 1, 3 or 9.

LEMMA. A map in the plane for which there exists n such that
T n (x, y) = (x, y), must be integrable.
∑n−1
PROOF. Take f (x, y) = y for example. Then F (x, y) = k=0 f (T k (x, y)) is an integral.

If we apply this lemma to the Knuth map, we get an explicit integral

F (x, y) = y + |y − |x|| + |x − |y − |x||| + |y − |x − |y||| + |x − |y| + |y − |x − |y|||| .

The level curves of this function are shown in the graphics above. For every value c > 0 the level set F (x, y) = c
is a closed gingerman shaped curve on which T is conjugated to a rotation by an angle 1/9.

Remark: The problem of proving periodicity of the map has been posed by Morton Brown in the American
Mathematical Monthly 90, 1983, p. 569. The Monthly published the elegant solution of Donald Knuth in the
volume 92, 1985 p. 218.

INTEGRAGBLE OR NOT? Lets look at the case ϵ = 1, where

√
T (x, y) = ( x2 + 1 − y, x)

All orbits seem all to lie on invariant curves. The map looks integrable.
It had been communicated to me by M. Rychlik in 1998, that numerical ex-
periments by John Hubbard revealed a hyperbolic periodic orbit of period 14:
(x, y) = (u, u) with u = 1.54871181145059. The largest eigenvalue of dT 14 (x, y)
is λ = 1.012275907. The existence of a hyperbolic point of such a period makes
integrability unlikely since homoclinic points might exist, but it is not impossi-
ble. It is difficult to find other hyperbolic periodic points. An other indication
for non-integrability is a result of Rychlik and Torgenson who have shown that
this map has no integral given by algebraic functions.
What follows was added after the handout was distributed in class:

HOW TO FIND AN INTEGRAL?

If we know a map is integrable, we could recover the invari-

ant function F by taking f (x, y) = y and defining F (x, y) =
∑n−1
limn→∞ n1 k=0 f (T k (x, y)).

This invariant function is called the time average along the orbit. In the case of nonintegrability, this function
is constant on complicated sets or even be infinite on some part of the plane. If the map is integrable with a
nice analytic function, one could expect the integral to found using time averages.

[ ] [ 2kx ]
x (1+x 2) − y
THE McMILLAN MAP T = is an other example of an
y x
integrable map, where k is a parameter. It is called the McMillam map and
has the integral
F (x, y) = x2 + y 2 + x2 y 2 − 2kxy .
It is especially interesting to study because T is a rational function, a fraction
of two polynomials. I don’t think, one has a complete list of all integrable
rational maps in the plane.

WHAT HAPPENS CLOSE TO THE INTEGRABLE CASE? In general, in-

tegrability gets lost when making small changes to an integrable map. For
example, the Standard map T (x, y) = (2x − y + ϵ sin(x), x) can for small ϵ
be considered as a perturbation of the integrable map T (x, y) = (2x − y, x)
which has the integral F (x, y) = x − y. A study of the stable and unsta-
ble manifolds of the hyperbolic fixed point (0, 0) shows that they intersect
transversely for small ϵ. One usually studies the map in an other form.
Because H(x, y) = (−x, y − x) = H −1 (x, y) satisfies H(S(H(x, y))), where
T (x, y) = (2x − y + ϵ sin(x), x) and S(x, y) = (x + y + ϵ sin(x), y + ϵ sin(x)), we
can look also at the map S instead. This map has the integral F (x, y) = y for
ϵ = 0 and the invariant curves are horizontal.

KAM. Near integrable maps, remnants of integrability still exist. These traces of integrability persist in the form
of smooth invariant curves which are now called KAM curves. The acronym KAM stands for Kolmogorov-
Arnold-Moser. The proof that invariant curves persist after the perturbation is not easy. To find an invariant
curve on which the map is conjugated to an irrational rotation with angle α, we need to find a periodic function
q(x) such that qn = q(nα) satisfies the nonlinear recursion qn+1 − 2qn + qn−1 = ϵ sin(qn ). This means

q(x + α) − 2q(x) + q(x) = ϵ sin(q) .

Naively, one could try to find q using the implicit function theorem: if one could invert the linear map
L(q) = q(x + α) − 2q(x) + q(x).

inx
∑
SMALL ∑DIVISORS. Lets look at this inversion problem If q(x) = n cn e is the Fourier series of q, then
Lq(x) = n cn (eiα − 2 + e−iα )einx . If L(q) = p = n dn einx . then
∑

∑ 1 inx
∑ 2
q = L−1 p = dn e = dn einx .
n
einα − 2 + e−inα n
cos(nα) −1

2
You see the appearance of small divisors cos(nα)−1 . In order that the Fourier series of the inverse converges,
one needs α to be far away from rational numbers. Such numbers are called Diophantine numbers. Even so,
one is able to invert L in certain cases, the map L is not invertible as required for the implicit function theorem.
One needs a so called hard implicit function theorem.
 STABLE AND UNSTABLE MANIFOLDS Math118, O. Knill
ABSTRACT. Near a hyperbolic point, one can conjugate the map by its linearization. This conjugation defines
local curves through the origin which are invariant. These stable and unstable manifolds intersect in general to
form homoclinic points. We will not prove the linearization theorem in class.

STERNBERG-GROBMAN-HARTMAN LINEARIZATION
THEOREM. If T (x) is smooth map with a hyperbolic fixed ←H←
point x0 , then T is conjugated to its linearization DT near x0 .

Near the fixed point x0 , the dynamics can be computed by ↓ ↓

first going into a new coordinate system H −1 (x0 ), applying the T A
linear map A, and undoing the coordinate change by applying ↓ ↓
H.

More precisely, there exists a small disc D around x0 and a map

H in the plane such that in D the identity H ◦A(x) = T ◦H(x) ←H←
holds.

INVARIANT MANIFOLDS. The linear equation x 7→ Ax has two in-

variant curves, the lines spanned by the eigenvectors vi of A. The
conjugation defines two invariant curves ri (t) = H(tvi ) through a hy-
perbolic fixed point. These curves are called stable and unstable
manifolds of the hyperbolic fixed point. The picture shows the sta-
ble and invariant manifolds for one of the fixed points of the Henon
map. The unstable manifold lies in the attractor. Note that the un-
stable manifold of T (x, y) = (1 − ax2 + y, bx) is the stable manifold for
T −1 (x, y) = (y/b, (x − 1 + ay 2 /b2 ).
Here is the proof of the linearization theorem in its simplest case. The conjugation can actually be proven to be smooth too.
The theorem had first be proven by S. Sternberg in 1958 (smooth conjugation for smooth T ) and P. Hartman in 1960 (C 1
conjugation for C 2 maps T ). The proof (not done in class) is not so easy and requires the language of linear operators.

PROOF PART 0: Some notations and preparations.

The proof works in any dimension. So x is now a vector in n-dimensional space X = Rn . Write C(X, X) for the
linear space of all continuous maps from X to X. The norm on this space is defined as ||f || = supx∈X |f (x)|.
For example: || sin || = 1 The norm of a linear operator U from C(X, X) to C(X, X) is defined as ||U || =
sup||f ||=1 ||U (f )||. A linear map is called a contraction if ||U || < 1. If U is a contraction, then (1 − U ) is
∑∞
invertible: the inverse is given by a geometric series (1 − U )−1 = n=0 U n . For a hyperbolic matrix A, we write
X = E+ ⊕ E− , where E + is the linear space spanned by the eigenvectors of A belonging to the eigenvalues
|λi | < 1 and E − is the space spanned by the eigenvectors of A belonging to the eigenvalues |λi | > 1.

PROOF PART I: Reduction to a global conjugation problem.

Take first a smooth scalar function ϕϵ (x), which satisfies ϕϵ (x) = 1 for |x −
x0 | > 2ϵ and ϕϵ (x) = 0 for |x − x0 | < ϵ (see picture to the right). The map
S = T +ϕϵ ·(A−T ) is equal to T for |x−x0 | < ϵ and equal to A for |x−x0 | > 2ϵ.
If can write S(x) = Ax + f (x), where f is a smooth map satisfying ||f ′ ||∞ → 0
for ϵ → 0. Using this surgery, we can solve a global problem.

PROOF PART II: The conjugating equation and its linearization.

The aim is to show that S is conjugated by a map H(x) = x + h(x) to the linear map A if S = A + f if ||f ′ ||∞
is small enough. Remember that f ′ = Df is the Jacobean matrix of f . The condition H ◦ A(x) = S ◦ H(x) can
be rewritten with S(x) = Ax + f (x), H(x) = x + h(x) as

h(A(x)) − Ah(x) = f (x + h(x)) .

It is an equation for the unknown map h ∈ C(X, X). We first consider the linearized problem

(Lh)(x) := h(A(x)) − Ah(x) = f (x) .

PROOF PART III: Solving the linearized problem.
We can decompose the problem into two parts

h± (A(x)) − Ah± (x) = f± (x) ,

where h = h+ + h− , f = f+ + f− is the decomposition satisfying f± , h± ∈ E ± . The linear map on continuous

functions on the plane U : C(X) 7→ C(X), h 7→ h(A) as well as its inverse U −1 have norm ||U || = ||U −1 || = 1.
We write Af = A+ f+ + A− f− . Because

∑
∞
1
||(U − A+ )−1 || = ||U −1 An+ U −n || ≤
n=0
1−λ
∑
∞
λ 1
||(U − A− )−1 || = ||A−1
− A−n n
− U || ≤ <
n=0
1−λ 1−λ

with λ = max{||A+ ||, ||A−1

− ||} < 1, we can find h using the formula

h = h+ + h− = (U − A+ )−1 f+ + (U − A− )−1 f− .

PROOF PART IV: Solving the nonlinear problem.

Define Φ(h)(x) = f (x + h(x)) − f (x). We need to solve the equation

Lh = Φh + f

in for the unknown h in C(X). The solution to this equation (L−1 Φ − 1)h = L−1 f is

h = (1 − L−1 Φ)−1 L−1 f

if 1 − L−1 Φ is invertible. Sufficent to invertibility is that L−1 Φ is a contraction. This is indeed the case if ϵ is
small that is if ||f ′ ||∞ is small:
1 1
||(L−1 Φ)h1 − (L−1 Φ)h2 || ≤ · ||Φh1 − Φh2 ||∞ ≤ · ||f ′ ||∞ · ||h1 − h2 || .
1−λ 1−λ

COMPUTATION OF MANIFOLDS. The stable and unstable manifolds of a hyperbolic fixed point can be
computed using power series. This calculation is due to Francescini and Russo. To get one of the manifolds,
construct a curve r(t) = (x(t), y(t)) satisfying r(0) = (x0 , y0 ) and

T (x(t), y(t)) = (1 − ax(t)2 + y(t), bx(t)) = (x(λt), y(λt))

for all t ∈ R. Here λ is an eigenvalue of the Jacobean∑matrix at the fixed point. Because y(λt) = bx(t), it is
∞
enough to calculate x(t). With a Taylor series x(t) = n=0 an tn , the invariance condition 1 − ax(t)2 + y(t) =
2 −1
x(λt) or equivalently x(λt) + ax(t) − bx(λ t) = 1 becomes

∑
∞ ∑
n
n −n
[an λ − ban λ +a aj an−j ]tn = 1 .
n=0 j=0

This equation allows to calculate the Taylor coefficients an recursively. Comparing coefficients of tn gives
a(a0 an + a1 an−1 + ... + an−1 a1 + an a0 ) − bλ−n an = −λn an and so

a(a1 an−1 + ... + an−1 a1 )

an =
−λn − 2aa0 + bλ−n

once a0 , ..., an−1 are given. The first coefficient a0 is just x0 . Because a1 satisfies 2aa0 a1 − bλ−1 a1 = a1 λ,
it can be chosen arbitrary like a1 = 1. For the parameters a = 1.4, b = 0.3 the unstable manifold is r(t) =
(0.631354 + t − 0.25986t2 + ..., 0.189406 − 0.155946t − 0.0210654t2 + ...), the stable manifold is r(t) = (0.631354 +
t + 0.13278t2 + ..., 0.189406 + 1.92374t + 1.63796t2 + ...).

HOMOCLINIC POINTS. The intersection points of stable and unstable manifolds different from the fixed point
itself are called homoclinic points. It has been realized already by Poincaré that the existence of homoclinic
points produces a horrible mess. We will see why soon.
EXISTENCE OF SOLUTIONS TO ODE’s Math118, O. Knill

ABSTRACT. This is a proof of local existence of solutions of ordinary differential equations.

METRIC SPACES. Let X be a set on which a

distance d(x, y) between any two points x, y is EXAMPLES. 1) The plane R2 with the usual distance
defined. The function d must have the proper- d(x, y) = |x − y|. An other metric is the Manhattan or
ties d(y, x) = d(x, y) ≥ 0,d(x, x) = 0 and that taxi metric d(x, y) = |x1 − y1 | + |x2 − y2 |.
d(x, y) > 0 for two different points x, y. Fur-
thermore, one requires the triangle inequality 2) The set C([0, 1]) of all continuous functions x(t)
d(x, z) ≤ d(x, y) + d(y, z) to hold for all x, y, z. on the interval [0, 1] with the distance d(x, y) =
A pair (X, d) with these properties is called a maxt |x(t) − y(t)| is a metric space.
metric space.

CONTRACTION. A map ϕ : X → X is called EXAMPLES. 1) The map ϕ(x) = 21 x + (1, 0) is a con-

a contraction, if there exists λ < 1 such that traction on R2 .
d(ϕ(x), ϕ(y)) ≤ λ · d(x, y) for all x, y ∈ X. The 2) The map ϕ(x)(t) = sin(t)x(t) + t is a contraction on
map ϕ shrinks the distance of any two points C([0, 1]) because |ϕ(x)(t) − ϕ(y)(t)| = | sin(t)| · |x(t) −
by the contraction factor λ. y(t)| ≤ sin(1) · |x(t) − y(t)|.

EXAMPLES 1) (Rn , d(x, y) = |x − y|) is complete. The

CAUCHY SEQUENCE. A Cauchy sequence rational numbers (Q, d(x, y) = |x − y|) are not.
in a metric space (X, d) is defined to be a
sequence which has the property that for any 2) C[0, 1] is complete: given a Cauchy sequence xn , then
ϵ > 0, there exists n0 such that |xn − xm | ≤ ϵ xn (t) is a Cauchy sequence in R for all t. Therefore xn (t)
for n ≥ n0 , m ≥ n0 . converges point-wise to a function x(t). This function
is continuous: take ϵ > 0, then |x(t) − x(s)| ≤ |x(t) −
COMPLETENESS. A metric space in which xn (t)| + |xn (t) − yn (s)| + |yn (s) − y(s)| by the triangle
every Cauchy sequence converges to a limit is inequality. If s is close to t, the second term is smaller
called complete. than ϵ/3. For large n, |x(t) − xn (t)| ≤ ϵ/3 and |yn (s) −
y(s)| ≤ ϵ/3. So, |x(t) − x(s)| ≤ ϵ if |t − s| is small.

BANACH’s FIXED POINT THEOREM. A contraction ϕ in a

complete metric space (X, d) has exactly one fixed point in X.

PROOF.

(i) We first show by induction that

d(ϕn (x), ϕn (y)) ≤ λn · d(x, y)
for all n.
∑
(ii) Using the triangle inequality and k λk = (1 − λ)−1 , we get for all x ∈ X,

∑
n−1 ∑
n−1
1
d(x, ϕn x) ≤ d(ϕk x, ϕk+1 x) ≤ λk d(x, ϕ(x)) ≤ · d(x, ϕ(x)) .
1−λ
k=0 k=0

(iii) For all x ∈ X the sequence xn = ϕn (x) is Cauchy because by (i),(ii),

1
d(xn , xn+k ) ≤ λn · d(x, xk ) ≤ λn · · d(x, x1 ) .
1−λ
By completeness of X it has a limit x̃ which is a fixed point of ϕ.

(iv) There is only one fixed point. Assume, there were two fixed points x̃, ỹ of ϕ. Then

d(x̃, ỹ) = d(ϕ(x̃), ϕ(ỹ)) ≤ λ · d(x̃, ỹ) .

This is impossible unless x̃ = ỹ.

 THE CAUCHY-PICARD EXISTENCE THEOREM.
Assume f : Rn → Rn has a continuous derivative. For
every initial condition x0 there exists τ > 0 such that on
the time interval [0, τ ) there exists exactly one solution of
the initial value problem

ẋ(t) = f (x(t)), x(0) = x0 .

PROOF.

(i)
Consider for every τ > 0 and r > 0 the complete metric space

X = Xτ (r) = {x ∈ C[0, τ ] | max ||x(t) − x0 || ≤ r }

0≤t≤τ

with metric d(x, y) = max0≤t≤τ ||x(t) − y(t)||. With c(t) = x0 , we can write also X = {x | d(x, c) ≤ r}.

Define a map ϕ on C[0, τ ] by

∫ t
ϕ(y) : t 7→ x0 + f (y(s)) ds .
0

(ii) Define the constant

||f (u) − f (v)||

λ = max{ | ||u − x0 || ≤ 1, ||v − x0 || ≤ 1, u 6= v} .
||u − v||

For every x, y ∈ Xτ (r) and τ ≤ 1, one has then

||f (x(s)) − f (y(s))|| ≤ λ · ||x(s) − y(s)|| ≤ λ · d(x, y)

for every 0 ≤ s < τ . Therefore

∫ t ∫ t
d(ϕ(x), ϕ(y) = max || f (x(s)) − f (y(s)) ds|| ≤ ||f (x(s)) − f (y(s))|| ds ≤ λτ d(x, y) .
0<t<τ 0 0

We see that for small enough τ , the map ϕ is a contraction.

(iii) With M = max{||f (x(t))|| | 0 ≤ t ≤ 1, d(x, c) ≤ 1}, one has

∫ t ∫ t
||ϕ(c) − c|| = || f (x0 (s)) ds|| ≤ ||f (x0 (s))|| ds ≤ M · τ .
0 0

If τ ≤ 1 is small enough, then M · τ < (1 − λ)r. Using the triangle inequality, we obtain

d(ϕ(x), c) ≤ d(ϕ(x), ϕ(c)) + d(ϕ(c), c) ≤ λd(x, c) + M τ ≤ λr + (1 − λ)r = r

proving that ϕ maps X = {d(x, c) ≤ r} into itself.

(iv) The fixed point ϕ in X obtained by Banach’s fixed point theorem is a solution of the differential equation
ẋ = f (x) with initial value x(0) = x0 .

EXAMPLE WITH NO UNIQUE SOLUTION.

d √
The differential equation dt x = x with x(0) = 0 has the solution x(t) = 0 and x(t) = t2 /4. There are infinitely
many solutions with the initial condition x(0) = 0. Note that the function F (x) is not differentiable at t = 0.

EXAMPLE WITH NO GLOBAL SOLUTION.

d
The differential equation dt x = x2 with initial condition x(0) = 1 has the solution x(t) = 1/(1 − t). At t = 1,
the solution has escaped to infinity.

P.S. The photos show Stefan Banach (1892-1945), Emile Picard (1856-1941) and Augustin Cauchy (1789-1857).
DIDFFERENTIAL EQUATIONS IN TWO DIMENSIONS Math118, O. Knill

ABSTRACT. Differential equations in the plane do not show chaotic behavior. An interesting feature in two
dimensions are limit cycles and their bifurcation. We look at some examples of such differential equations.

DIFFERENTIAL EQUATIONS. Ordinary differential equations are equations for an unknown function
x(t) in which which the derivatives with respect to one variable t appears. If derivatives with respect to several
variables would occur, one would speak of partial differential equations. By introducing new variables for higher
derivatives and possibly for time t, one can always bring it into the form

d
x(t) = f (x(t))
dt

where x(t) is a vector.

2
d d
EXAMPLE. To write the second order inhomogeneous  equation dt2 x(t) +
 differential  dt x(t) = sin(t) in the
x(t) y(t)
d d 
above form, introduce y(t) = dt and z(t) = t. Then dt y(t)  =  sin(z(t)) − y(t) 
z(t) 1

DIFFERENTIAL EQUATIONS IN THE PLANE. A solution ⃗x(t) of a differ-

ential equation dt d
⃗x = F⃗ (⃗x) is a vector quantity changing in time. The vector
F (⃗x(t)) is the velocity vector. In two dimensions, we have

ẋ(t) = f (x, y)
ẏ(t) = g(x, y) .

The vector field is obtained by attaching a vector F⃗ (x, y) = (f (x, y), g(x, y))
to each point (x, y). Of special importance are equilibrium points. These
are points, where the velocity is zero.

EXAMPLE. COMPETING SPECIES. A population of two species, where both

compete for the same food can be modeled by the coupled logistic equations

ẋ = αx(1 − x/M ) − βxy

ẏ = γy(1 − y/M ) − δxy .

A specific example is

ẋ = 2x(1 − x/2) − xy
ẏ = 3y(1 − y/3) − 2xy

which has the equilibrium point (1, 1) because (f (1, 1), g(1, 1)) = 0. Addition-
ally, one has the equilibrium points (0, 3), (2, 0) and of course (0, 0).

EXAMPLE. PREDATOR-PREY. These systems of the form

ẋ = αx − βxy
ẏ = −γy + δxy

are also known under the name Volterra-Lodka systems. They can describe for
example a shark-tuna population. The tuna population x(t) becomes smaller
with more sharks. The shark population y(t) grows with more tuna. Histori-
cally, Volterra explained so the oscillation of fish populations in the Mediterrian
sea. Here is a specific example:

ẋ = 0.4x − 0.4xy
ẏ = −0.1y + 0.2xy
EXAMPLE. AIDS EPIDEMIC. The previous model can also model an epi-
demic as you can read in detail in Tom’s lecture notes. In the interpretation
of the epidemic, x(t) is the size of the susceptible population, while y(t) is the
size of the infected population. A specific example modeling AIDS is

ẋ = 0.2x − 0.1xy
ẏ = −y + 0.1xy

EXAMPLE. EBOLA EPIDEMIC. If the disease kills fast like in the case of
ebola, we get a different picture

ẋ = 0.2x − 0.5xy
ẏ = −y + 0.5xy

HARMONIC OSCILLATOR. The system ẋ = y, ẏ [= −x can ] in vector form

d 0 −1
⃗x = (x, y) be written as dt ⃗x(t) = A⃗x(t), with A = . The direction
1 0
field is always perpendicular to ⃗x so that by the product differentiation rule
d/dt⃗x · ⃗x = 2⃗x′ · ⃗x = 0 and |⃗x| is constant. The solution curves are circles. In
the homework, you look at a bit more general case. ẋ = y, ẏ = −cx, where c is
a constant.

HAMILTONIAN SYSTEMS. If H is a function of two variables, we can look

at the system

ẋ = ∂y H(x, y)
ẏ = −∂x H(x, y)

H is called the energy or Hamiltonian, x is called the position and y the

d
momentum. Hamiltonian systems preserve energy H(x, y): dt H(x(t), y(t)) =
∂x H(x, y)ẋ+∂y H(x, y)ẏ = ∂x H(x, y)∂y H(x, y)−∂y H(x, y)∂x H(x, y) = 0. The
level curves of H are solution curves of the system. The time T maps are inte-
grable. The illustration to the right shows the solution curves for the pendulum
H(x, y) = y 2 /2 − cos(x), where

ẋ = y
ẏ = − sin(x)

Here x is the angle between the pendulum and y-axes, y is the angular velocity,
sin(x) is the potential.

THE VAN DER POL EQUATION. ẍ + (x2 − 1)ẋ + x = 0 appears in electrical

engineering, biology or biochemistry. It is an example of a Lienhard system
differential equations of the form ẍ + ẋF ′ (x) + G′ (x) = 0, where F (x) = x3 /3 −
x, g(x) = x.

ẋ = y − (x3 /3 − x)
ẏ = −x

Lienhard systems often have limit cycles, closed solution curves on which
trajectories can be attracted to. Lienhard systems are useful for engineers,
who need oscillators which are stable under random noise.
BIFURCATIONS Math118, O. Knill

ABSTRACT. Equilibrium points can bifurcate. One distinguishes pitchfork bifurcation and blue-sky bi-
furcation, which were already knew in the one-dimensional setting. In two dimensions, where limit cycles can
occur, it can happen that an equilibrium point produces a limit cycle. This is called the Hopf bifurcation.

BIFURCATIONS OVERVIEW. If an eigenvalue of the Jacobean DF at an equilibrium point (x0 , y0 ) crosses the
y-axes, the stability of the equilibrium point changes. As in the discrete case, this is called a bifurcation. What
possibilties are there? Besides the pitch-fork and blue-sky bifurcations, we already already in one dimensions,
there are now possibilities which are not known in one dimensions. One is called Hopf bifurcation, which is
the birth of limit cycles.

1.5

PITCHFORK BIFURCATION. If both eigenvalues were initially in the left 1

half plane and one eigenvalue moves over to the right half plane, then a stable 0.5

sink becomes a hyperbolic point. This bifurcationis usually associated to the -2 -1.5 -1 -0.5 0.5 1 1.5 2

creation of two new stable equilibrium points. This is called a pitchfork -0.5

bifurcation. We know the one-dimensional version of this bifurcation already. -1

-1.5

-2

Example: c = 0 is a bifurca-
tion parameter for
d
x = y − 0.3 ∗ x
dt
d
y = cx − x3
dt

1.5

BLUE SKY BIFURCATION. It can happen that a hyperbolic equilibrium point 1

collides with a stable or unstable equilibrium point and disappears. The oppo- 0.5

site is also possible. Out of the blue, a parabolic equilibrium appears and splits -2 -1.5 -1 -0.5 0.5 1 1.5 2

into two equilibrium points. This is called the saddle node bifurcation or -0.5

blue-sky bifurcation. -1

-1.5

-2

Example: c = 0 is a bifurca-
tion parameter for
d
x = c + x2
dt
d
y = −y
dt

1.5

If both eigenvalues are on the left hand side and cross at the same 0.5

time, we have a Hopf bifurcation. In this case, a stable sink -2 -1.5 -1 -0.5 0.5 1 1.5 2

becomes an unstable source and a limit cycle will appear nearby. -0.5

-1

-1.5

-2

Example:
d
x = y
dt
d
y = −x − (x2 − c)y
dt
MORE BIFURCATIONS WITH LIMIT CYCLES (what follows will not be quizzed). These bifurcations above
started with critical points and led to limit cycles. With limit cycles, there are more possibilites:

PITCH-FORK BIFURCATION FOR LIMIT CYCLES.

A stable limit cycles can change stability, become unstable and produce two
limit cycles. This is called the saddle node bifurcation for limit cycles. An
example is given in polar coordinates by
d
r = r(r(1 − r)3 + c((r − 1)3 + (r − 1)))
dt
d
θ = α + r2
dt
(It can using the formula x = r cos(θ), y = r sin(θ) be rewritten as a system in
the x, y coordinates.)

SADDLE NODE BIFURCATION FOR LIMIT CYCLES.

The sudden appearance of limit cycles is called saddle node bifurcation for
limit cycles. An example is given in polar coordinates by
d
r = cr + r3 − r5
dt
d
θ = α + r2
dt
(It can using the formula x = r cos(θ), y = r sin(θ) be rewritten as a system in
the x, y coordinates.)

INFINITE PERIOD BIFURCATION.

A blue-sky bifurcation for equilibrium points can appear on a limit cycle. The
limit cycle will become the stable and invariant manifolds of the newly born 1

hyperbolic points. This bifurcation is called infinite period bifurcation be-

cause the limit cycle period will satisfy T → ∞. An example is given in polar 0.5

coordinates by
d -1 -0.5 0.5 1

r = r(1 − r2 )
dt -0.5
d
θ = c − sin(θ)
dt -1

The system has an invariant circle for all c but for c = 1, there is an equilibrium
point on the circle.

HOMOCLINIC BIFURCATION.
An equilibrium point can collide with a limit cycle and ”open” it up. This
bifurcation is called a homoclinic bifurcation. An example of a homoclinic
bifurcation happens for
d
x = y
dt
d
y = cy + x − x2 + xy
dt
with the parameter c = −0.86....

Reversed situations of ”supercritical” bifurcations (discussed above) are often called subcritical.

• A stable critical point can collide with two other hyperbolic critical points and become unstable. This is
d d
called subcritical pitch-fork bifurcation. An example is dt x = cx + x3 , dt y = −y. This example is
d 3 5 d
often associated to catastrophe like in the example dt x = cx + x − x , dt y = −y
• An unstable limit cycle collapses to a stable critical point and becomes an unstable critical point. This is
called a subcritical Hopf bifurcation.

These situations lead to ”catastrophes”. The stable equilibrium or cycle ”jumps” discontinuously.
LIENHARD SYSTEMS Math118, O. Knill
ABSTRACT. For a certain class of differential equations called Lienard systems, one can prove the existence of
a stable limit cycle. An example is the van der Pol oscillator.

LIENHARD SYSTEMS. A differential equation

d2 d
x + F ′ (x) x + G′ (x) = 0
dt2 dt
d
is called a Lineard system. With y = dt x + F (x), G′ (x) = g(x), this is equivalent to

d
x = y − F (x)
dt
d
y = −g(x) .
dt

VAN DER POL EQUATION. If F (x) = c(x3 /3 − x) and G(x) = x2 /2, we have
van der Pol equation

d2 d
2
x + c(x2 − 1) x + x = 0
dt dt

2
d
Physically, one has a harmonic oscillator dt 2 x + x = 0 for c = 0. For c > 0,
d
some velocity and space dependent force c(x2 − 1) dt x is added. This force is
accelerating the oscillator, if x < 1, it is slowing down the oscillator if x2 > 1.
2

For large c, one calls the oscillator a relaxation oscillator because the stress
accumulated during a slow buildup is relaxed during a sudden discharge.

THEOREM (Lienard) Assume F and g are smooth odd functions such that g(x) > 0 for x > 0 and such that
F has exactly three zeros 0, a, −a with F ′ (0) < 0 and F ′ (x) ≥ 0 for x > a and F (x) → ∞ for x → ∞. Then
the corresponding Lienard system has exacactly one limit cycle and this cycle is stable.

REMARK ON THE FIXED POINT (0, 0): Because g is odd with g(x) > 0 for g ≥ 0, we have g ′ (0) ≥ 0. The
Jacobean matrix [ ]
F ′ (x) 1
−g ′ (x) 0
√
has the eigenvalues λ1,2 = (−F ′ (x) ± F ′ (x)2 − 4g
√(x))/2. At the fixed point,
′ the real part of these eigenvalues
is positive because by assumption F (0) < 0 and | F ′ (x)2 − 4g ′ (x))| ≤ |F ′ (x)| since g ′ (0) ≥ 0. We see that the
′

fixed point 0 is repelling.

SOME REMARKS. Stable limit cycles appear in ecological, biological as well as mechanical systems. They are
relevant because they are in general stable under small changes of the system.

From 1920 to 1950, research on nonlinear oscillations florished. The work was initially motivated by the
development of radio and vacuum tube technology, where one realized that many oscillating circuits could
be modeled by Lienard systems. This has been applied to many other situations. For example, one has also
modeled the periodic firing of nerve cells driven by a constant current using van der Pol type differential
equations.

Balthasar Van der Pol (1889-1959) was a Dutch electrical engeneer. He started his investigation on the van
der Pol equation in 1926 and also studied versions with periodic forcing term, where chaotic motion can occur.

Lienards theorem was found and published in Russian by Lienard in 1958. For the proof of the Lienards
theorem, we followed the proof given in the book ”Differential equations and Dynamical systems” of Lawrence
Perko.
A nice discussion can also be found in the book ”Nonlinear dynamics and Chaos” by Steven Strogatz.
For historical facts mentioned in this section, we used ”Writing the History of Dynamical Systems: Longe Duree
and Revolution, Disciplines and Cultures” by David Aubin and Amy Dahan Dalmedico in Historia Mathematica
29, 2002. One should note also Mary Cartwright (1900-1998), who was making important contributions to
the theory of nonlinear oscillations and discovered many phenomena later known as chaos (when the oscillator
is driven, it becomes chaotic).
PROOF OF LIENHARDS THEOREM.
Draw in the xy− plane the graph of the function x → F (x). On this graph,
dy
the vector field is vertical. It is called a nullcline. For x > 0 we have dx < 0.
On the y-axes, the vector field is horizontal because g(0) = 0. The y-axes is
also a nullcline.
Consider an orbit which starts at (0, y0 ) on the positive y axes. It goes to the 4

right because g(x) > 0 for x ≥ 0. Because g(x) > 0 for x > 0, the orbit also 2

moves down. It has to hit the graph of F . It intersects that nullcline at a point
dy -2 -1 1 2
(x1 , 0) with positive vertical velocity and enters the region, where dx < 0. It
must then go to the left and hit again somewhere the y axes horizontally in -2

some point (0, y1 ) = (0, −S(y0 )). -4

Because the differential equations are invariant under the transformation (x, y) 7→ (−x, −y), we can analyze
the fate of the orbit on the left half plane in the same way as on the right plane.

A limit cycle exists if the map y0 → S(y0 ) has a fixed point. Alternatively, we can express this that the
”energy” H(x, y) = y 2 /2 + G(x) is the same at (0, y0 ) and (0, y1 ). The idea of the proof is to determine the
energy gain along the orbit and to see that only for one single orbit, the energy is conserved.

Compute
d d d
H(x, y) = y y + g(x) x = −F (x)g(x)
dt dt dt
If F (x(t)) were positive on the entire trajectory from (0, y0 ) to (0, y1 ), then H(0, y1 ) − H(0, y0 ) is positive. It
must therefore cross the graph of F at a point, where F (x) > 0. The theorem is proven if we can show the
following statement about the energy difference

∆(y0 ) = H(0, S(y0 )) − H(0, y0 )

depending on the intersection point (x1 , F (x1 )) with the null cline.

If x1 ≤ a, then ∆(y0 ) > 0. For y0 such that x1 > a, ∆(y0 ) is a

monotonically decreasing function for y0 . and ∆(y0 ) → −∞ for
y0 → ∞.

As a consequence, there exists then exactly one point y0 , where the energy gain is zero. This point y0 belongs
to a limit cycle. The rest of the proof is devoted to the verification of the above claim.

(i) ∆(y) > 0 if y0 is such that x1 ≤ a.

Note that F (x) is negative in the interval [0, a]. If x1 ≤ a, then x(t) ≤ a until we hit the y axes again. But
d
since then F (x(t)) < 0 and g(x) > 0 for x > 0, we have dt H(x, y) = −F (x)g(x) > 0. The energy gain is positive.

(ii) The monotonicity claim for x1 ≥ a.

d
Let A(y0 ) be the path (x(0), y(0)) = (0, y0 ) and (x(T ), y(T )) = (0, y1 ). From dt H(x, y) = −F (x)g(x) we obtain

F (x)g(x)
∫ ∫ ∫
∆(H)(y0 ) = −F (x(t))g(x(t)) dt = − F (x(y)) dy = − dx .
A A A y − F (x)

Split the path A into a path A1 from (0, y0 ) to x(t) = a, a path A2 which is the continuation until x(t) = a
again and into a path A3 until (0, y1 ). Along A1 and A3 , we can parametrize the curve by x instead of t, along
A2 , we can use the parameter y.
∫a (x)g(x)
We see that increasing y0 increases y(t) and so decreases the integral ∆1 (H)(y0 ) = 0 − Fy−F (x) dx along A1 .
∫ a F (x)g(x)
On A3 increasing y0 decreases y(t) which decreases the integral ∆3 (H)(y0 ) = 0 y−F (x) dx along along A3 .
Along A2 , use y as the variable.
∫ y Increasing y0 pushes the path A2 to the right so that F (x(t)) is increasing and
the integral ∆2 (H)(y0 ) = − y23 F (x(y)) dy is decreasing. The sum ∆(H)(y0 ) = ∆1 (y0 ) + ∆2 (y0 ) + ∆3 (y0 ) is
decreasing in y0 .

(iii) The limit y0 → ∞.

To see that ∆(y0 ) goes to −∞ for y0 → ∞, we split an orbit into paths B1 , B2 , B3 in the same way as A1 , A2 , A3
but where the value of a has been replaced by a + 1. The integrals along B1 and B3 are bounded by a constant
independent of y0 , while the integral along B2 is bigger or equal to F (a + 1) times the y differences of the two
points, where x(t) = a + 1. This difference goes to −∞ for y0 → ∞. So, the energy gain along the sum of the
paths B1 , B2 , B3 goes to −∞ for y0 → ∞.
THE POINCARE BENDIXON THEOREM Math118, O. Knill

ABSTRACT. The Poincaré-Bendixon theorem tells that the fate of any bounded solution of a differential
equation in the is to convergence either to an attractive fixed point or to a limit cycle. This theorem rules out
”chaos” for differential equations in the plane.

THEOREM (Poincare-Bendixon). Given a differential equation

d
dt x = F (x) in the plane. Assume x(t) is an solution curve which
stays in a bounded region. Then either x(t) converges for t → ∞
to an equilibrium point where F (x) = 0, or it converges to a single
periodic cycle.

PRELIMINARIES.

CYCLES, EQUILIBRIA AND CYCLES. Points x, where F (x) = 0 are called equilibrium points for the
differential equation dxdt x = F (x). If a solution starts at an equilibrium point, it stays at the equilibrium point
for ever. If x(t) is a solution curve and x(t + T ) = x(t) for some T > 0, then the curve is called a cycle. Note
that we do not include equilibrium points in this definition. The minimal time T for which x(t + T ) = x(T ) is
called the period of the cycle.

TRANSVERSE CURVES. A smooth curve γ(s) ∈ R2 is called transverse to

the vector field x 7→ F (x) if at every point x ∈ γ, the vector F (x) and at least
one tangent vector of γ passing through x are linearily independent.

OMEGA LIMIT SET. The omega limit set ω + (x0 ) of an orbit x(t) passing through x0 is the set of points x,
for which there exists∩a sequence of times tn such that x(tn ) converges to x. Equivalent is the mathematical
statement ω + (x0 ) = s≥0 {x(t) | t ≥ s }, where A is the closure of a set A. If the ω-limit set of an orbit is a
cycle, it is called a limit cycle.

JORDAN CURVE THEOREM.

A Jordan curve is a simple closed curve in the
plane. ”Simple” means that the curve should not
have selfintersections or be tangent to itself at any
point. The Jordan curve theorem assures that
such a curve devides the plane into two disjoint re-
gions, the ”inside” and the ”outside”. This seem-
ingly elmentary fact is surprisingly hard to prove.

EXAMPLE OF LIMIT CYCLE. The differential equation given in polar coor-

dinates as
dr dθ
= r(1 − r2 ), =1
dt dt
is with x = r cos(θ), y = r sin(θ) equivalent to

dx dr dθ
= cos(θ) − r sin(θ) = (1 − (x2 + y 2 ))x − y
dt dt dt
dy dr dθ
= sin(θ) + r cos(θ) = (1 − (x2 + y 2 ))y + x
dt dt dt
In this example, all initial conditions away from the origin will converge to the
limit cycle.
EXAMPLE OF ATTRACTIVE POINT. The differential equation given in
polar coordinates as
dr dθ
= r(r2 − 1), =1
dt dt
is with x = r cos(θ), y = r sin(θ) equivalent to

dx dr dθ
= cos(θ) − r sin(θ) = ((x2 + y 2 ) − 1)x − y
dt dt dt
dy dr dθ
= sin(θ) + r cos(θ) = ((x2 + y 2 ) − 1)y + x
dt dt dt
In this example, all initial conditions away from the limit cycle will converge
to the origin or to infinity.

PROOF OF THE POINCARE-BENDIXON THEOREM. The aim is to show that if the omega limit set
ω + (x0 ) is nonempty, then it either an equilibrium point or a closed periodic orbit.

(i) There are no equilibrium points on a transverse curve. The vector field f can therefore not reverse direction
along the curve.

(ii) Let γ be a transverse curve. If a solution x(t) crosses γ more than once,
the successive crossing points form a monotonic sequence on the arc γ.
Proof. Denote by x(t1 ) = γ(s1 ), x(t2 ) = γ(s2 ), the first two crossing times. We
can assume that s2 ≥ s1 because if this does not hold, one can reparametrize γ
by s′ = 1 − s if s1 < s2 . The union of the two smooth arcs {x(t) | t1 ≤ t ≤ t2 }
and {γ(s) | s1 ≤ s ≤ s2 } is a closed piecewise smooth curve. By Jordan’s
curve theorem, such a curve divides the plane into two different regions. For
t > t2 , the solution x(t) stays in one of these regions. For the next crossing
x(t3 ) = γ(s3 ) one has therefore s3 ≥ s2 .
(iii) It follows from (ii) that no more than one point of any transverse arc γ can belong to the ω limit set ω + (x0 ).

(iv) Given y0 ∈ ω + (x0 ). Because a solution y(t) with y(0) = y0 stays by assumption in a bounded region,
the solution y(t) is by the existence theorem for differential equations defined for all times. It stays in ω + (x0 )
because this set is invariant under the flow. Assume, there exists no stationary point in ω + (x0 ). There exists
then a transverse arc γ passing through y0 . Because ω + (x0 ) ∩ γ can have only one intersection and y(t) returns
arbitrary close to y0 , the orbit {y(t)} through y0 is a single periodic orbit.

DIFFERENT SURFACES. Does an anlogue of Poincare Bendixon hold also

on other two dimensional spaces? The answer depends on the space. On the
sphere, the answer is yes, on the torus, there are solutions which are neither
asymptotic to a limit cycle or equilibrium point. An example of such a curve
is (t, αt) mod 1 which is a solution of the differential equation

d d
x = 1, y=α.
dt dt
Differential equations of the form
d d
x = F (x, y), y = αF (x, y) .
dt dt
can even show some weak type of mixing. You explore the question a bit in a
homework problem.
BASICS FOR ODES Math118, O. Knill

ABSTRACT. This is an overview over the stability of equilibrium points of linear differential equations in the
plane.

LINEAR SYSTEMS. A linear differential equation in two dimensions has the form
d
x(t) = ax + by
dt
d
y(t) = cx + cy
dt
d
It can be written as x(t)
dt ⃗ = A⃗x(t) with a vector ⃗x and a matrix A. We denote the eigenvalues of A with λ1
and λ2 .
[ ]
λ1 0
If the eigenvalues are different, one can diagonalize A. In the eigenbasis of A, the matrix is B =
0 λ2
and the differential equation becomes
d
x(t) = λ1 x
dt
d
y(t) = λ2 y
dt
with explicit solution x(t) = eλ1 t x(0), y(t) = eλ2 t y(0).

PHASE-PORTRAITS. We plot some vector fields and typical orbits

2 2

1.5 1.5

1 1

0.5
λ1 < 0 0.5
λ1 < 0
λ2 < 0, [ ] λ2 > 0, [ ]
-2 -1.5 -1 -0.5 0.5 1 1.5 2
−2 0 -2 -1.5 -1 -0.5 0.5 1 1.5 2
−2 0
-0.5 i.e A = -0.5 i.e A =
0 −3 0 3
-1 -1

-1.5 -1.5

-2 -2
2 2

1.5 1.5

1 1

0.5
λ1 > 0 0.5
λ1 = 0
λ2 > 0, [ ] λ2 < 0, [ ]
-2 -1.5 -1 -0.5 0.5 1 1.5 2
2 0 -2 -1.5 -1 -0.5 0.5 1 1.5 2
0 0
-0.5 i.e A = -0.5 i.e A =
0 3 0 −3
-1 -1

-1.5 -1.5

-2 -2
2 2

1.5 1.5

1 1

0.5
λ1 = 0 0.5
λ1 = a + ib, a < 0
λ2 = 0, [ ] λ2 = a −[ ib, ]
-2 -1.5 -1 -0.5 0.5 1 1.5 2
0 0 -2 -1.5 -1 -0.5 0.5 1 1.5 2
−1 1
-0.5 i.e A = -0.5 i.e A =
1 0 −1 0
-1 -1

-1.5 -1.5

-2 -2
2 2

1.5 1.5

1 1

0.5
λ1 = a + ib, a > 0 0.5
λ1 = ib
λ2 = a −[ ib, ] λ2 = −ib,
[ ]
-2 -1.5 -1 -0.5 0.5 1 1.5 2
1 1 -2 -1.5 -1 -0.5 0.5 1 1.5 2
0 1
-0.5 i.e A = -0.5 i.e A =
−1 0 −1 0
-1 -1

-1.5 -1.5

-2 -2
AREA PRESERVATION. A differential equation for which we have solutions for all times defines for each t a
map Tt in the plane.

d
We say a differential equation dt x = F (x) is area-preserving if each of the time t maps Tt is area preserving.

DIVERGENCE. If F is a vector field, we denote by div(F ) the divergence of F . It is in two dimensions,

where F (x, y) = (f (x, y), g(x, y)) given by the formula div(F )(x, y) = fx (x, y) + gy (x, y).

d
A differential equation dt x = F (x) is area-preserving if and only
if div(F )(x, y) = 0 for all points in the plane.
∫∫ ∫∫
PROOF. By the change of variable formula Tt (A)
dA = A
|det(DT (x)| dA, where DT (x) is
2
the Jacobean matrix of the transformation T at x. Because Tt (x) [ = x + tF +] O(t ), one has
1 + ta tb
DTt = I2 + tDF + O(t2 ), where I2 is the identity matrix. We have DTt = + O(t2 ) we have
tc 1 + td
d
∫∫ d
det(DTt ) = 1 + (a + d)t + O(t2 ) = 1 + div(F )t + O(t2 ). Therefore dt
∫∫
T t (A)
dA = A dt
|det(DT (x)| dA =
∫∫ d ∫∫
A dt
(1 + tdiv(F )) dA = A
div(F )(x) dA. (We could get rid of the absolute value because 1 + tdiv(F ) is
positive for small t).

2. PROOF. Define G(x, y, t) = (f (x, y), g(x, y), 1) and a tube like region {(x(t), y(t), t) (x(0), y(0)) ∈ A, 0 ≤
t ≤ τ } in space-time. Applying the divergence theorem using div(G)(x, y, t) = div(F )(x(t), y(t)), using
the fact that the flux through the cylindrical walls is zero and the flux ∫throught the bottom is −area(A) and
τ ∫
the flux through the top is area(Tτ (A)) gives area(Tτ (A)) − area(A) = 0 Tt (A) div(F (x(t), y(t))) dAdt. This
elegant proof does not need the coordinate change formula.

DISSIPATIVE SYSTEMS. If div(F ) < 0 in a region, then area is shrinking. You will explore some of the
consequences of dissipation in the homework. Here just an example:

PROPOSITION. In a region with div(F ) < 0, there are no sources

or elliptic equilibrium points.

PROOF. If (x0 , y0 ) is the equilibrium point, then div(F ) = λ1 + λ2 . At sources, the real part of both λ1 and
λ2 are positive. At elliptic equilibrium points, λ1 and λ2 are purely imaginary and the sum is 0.

EQUILIBRIUM POINTS. Points, where F (x, y) = (0, 0) are called equilibrium points. An equilibrium point
is called hyperbolic, if no eigenvalue has a real part equal to 0. Bifurcations can happen, when an eigenvalue
passes through the axes Re(λ) = 0. In the hyperbolic case, one can conjugate the system near the equilibrium
point to a linear system. This is a continuous version of the Sternberg-Grobman-Hartman theorem.

NULLCLINES. In two dimensions, we can draw the vector field by hand: attaching a vector (f (x, y), g(x, y)) at
each point (x, y). To find the equilibrium points, it helps to draw the nullclines {f (x, y) = 0}, {g(x, y) = 0 }.
The equilibrium points are located on intersections of nullclines. The eigenvalues of the Jacobeans at equilibrium
points allow to draw the vector field near equilibrium points. This information is sometimes enough to draw
the vector field by hand.

EXAMPLE: COMPETING SPECIES. The system ẋ = x(6 − 2x − y), ẏ = y(4 − x − y) has the nullclines
x = 0, y = 0, 2x + y = 6, x + y = 5. There are [ 4 equilibrium points (0, 0), (3,
] 0), (0, 4), (2, 2). The Jacobian
6 − 4x0 − y0 −x0
matrix of the system at the point (x0 , y0 ) is . Without interaction, the two
−y0 4 − x0 − 2y0
systems would be logistic systems ẋ = x(6 − 2x), ẏ = y(4 − y). The additional −xy part is due to the
competition. If both x and y become large, then this produce resource problems for both species.
Equilibrium Jacobean
[ ] Eigenvalues Nature of equilibrium
6 0
(0,0) λ1 = 6, λ2 = 4 Unstable source
[ 0 4 ]
−6 −3
(3,0) λ1 = −6, λ2 = 1 Hyperbolic saddle
[ 0 1 ]
2 0
(0,4) λ1 = 2, λ2 = −4 Hyperbolic saddle
−4 −4
√
[ ]
−4 −2
(2,2) λi = −3 ± 5 Stable sink
−2 −2
DIFFERENTIAL EQUATIONS IN THREE DIMENSIONS Math118, O. Knill

ABSTRACT. Differential equations in space can exhibit more complicated behavior than in the plane. Higher-
dimensional systems occur naturally as we will see. Many systems can be studied using a Poincare map.

HOW DO SYSTEMS APPEAR IN THREE DIMENSIONS?

• A second order differential equation ẍ = f (x, ẋ, t) can be written with (x, y, z) = (x, ẋ, t) as (ẋ, ẏ, ż) =
(y, f (x, z), 1). Such systems often appear in physcis. The time dependence allows to write the equation
in three dimensions.
• A mechanical system of two degrees of freedom defines a flow in four dimensional space. Every coordinate
has a position and velocity. Because energy is preserved, the dynamics takes place on a three dimensional
energy surface.

d
POINCARE MAP. Assume we have a differential equation dt x = F (x) in
space. Given a two-dimensional surface Σ in space, we can start at a point in
the plane, wait until the orbit returns back to the plane, hitting it transversely
and so define a map from a subset of the plane to the plane. For any surface
Σ in space, there is an open subset U , on which the return map T is defined
and smooth.

THE LORENTZ SYSTEM. The system has been suggested by Eduard Lorentz
in 1963. It is obtained by a truncation of the Navier Stokes equations. It
gives an approximate description of a horizontal fluid layer heated from below
which is itself a model for the earths atmosphere.

ẋ = a(y − x)
ẏ = cx − xz − y
ż = xy − bz

For a = 10, b = 8/3, c = 28, Lorentz observed a strange attractor.

THE ROESSLER SYSTEM. The following system of differential equations in

space was found by Otto Rössler in 1976. The system was designed as a model
for a strange attractor without any application in mind. It is theoretically
interesting because a return map resembles the one dimensional logistic map
fc (x) = cx(1 − x):

ẋ = −(y + z)
ẏ = x + 0.2y
ż = 0.2 + xz − cz

It is parametrized by a parameter c. The picture to the right shows an orbit for

c = 5.7. For parameters in the range 2.5 < c < 10 one observes a Feigenbaum
bifurcation scenario.

ẋ
THE DUFFING SYSTEM ẍ + b − x + x3 − c cos(t) = 0 can be written as

ẋ = y
ẏ = −by − x + x3 − c cos(z)
ż = 1

The Duffing system models a metallic plate between magnets. It is a harmonic

oscillator with an additional cubic force, some damping and an external periodic
driving force.
THE ABC FLOW. It is a flow with three parameters a, b, c, therefore its name
ABC flow. An other etymological explanation is that Arnold, Beltrami and
Childess worked on this system. Even so the system looks simple, its solutions
can be complicated.

ẋ = a sin(z) + c cos(y)
ẏ = b sin(x) + a cos(z)
ż = c sin(y) + b cos(x)

FORCED PENDULUM. The differential equation ẍ = cos(x) + g sin(t) de-

scribes a pendulum which is periodically shaken up and down. The equations

ẋ = y
ẏ = cos(x) + g sin(z)
ż = 1 .

in space have a natural Poincare section z = 0.

dyi
HENON-HEILS SYSTEM. The differential dx dt = Hyi (x, y), dt = −Hxi (x, y)
i

with
1 1
H(x, y) = (y12 + y22 + x21 + x22 ) + x21 x2 − x22
2 3
was studied first numerically by Henon and Heils in 1964. Energy surfaces
{H(x, y) = E } are invariant. For 0 ≤ E ≤ 1/6 the surface is bounded and
solutions stay bounded. The Poincaré section Σ = {x1 = 0} defines an area-
preserving map on a subset of the plane.

DOUBLE PENDULUM. The double pendulum is described by four variables.

Energy conservation defines a differential equation on a three dimensional
space. The return map x = 0 defines a map on the cylinder. If the gravi-
tational field is zero, the double pendulum is integrable. With gravity g > 0,
the system is complicated.

FALLING COIN. A falling coin defines a dynamical system which is often

used, to produce random events: you flip a coin or dice and let it hit the
ground, where it bounces. Flipping a coin and catching by the hand uses an
integrable system. Some people can throw, catch and predict the outcome. If
the stick moves in a gravitational field and if there are no impacts, then there is
besides energy conservation also momentum conservation: the system becomes
integrable. With impact, the system develops chaos.

3 BODY PROBLEM. The restricted three body problem in the plane is the 2

situation, where the third particle is assumed not to influence the two other 1

bodies. By Kepler, the two bodies moves on ellipses and produce a time periodic
force on the third body. Therefore, we obtain a differential equation of the form
-3 -2 -1 1 2 3

d2 -1

dt2 ⃗
x(t) = F (⃗x, t), where ⃗x = (x, y, ẋ, ẏ). Energy conservation defines a three
dimensional system. -2

-3

STOERMER PROBLEM. A charged particle in a magnetic dipole field has

rotational symmetry and so an angular momentum integral. This allows to
reduce the system to a differential equation with four variables. The energy
integral defines a flow on a three dimensional space. The system can be studied
using a return map. The relevance of the system is the motion of charged
particles in the van Allen belts and the explanation of the Aurora Borealis.
FRACTALS Math118, O. Knill

ABSTRACT. In order to define a strange attractor, we have to look at the notion of a ”fractal”, a set of
fractional ”dimension”. The term fractal had been introduced by Benoit Mandelbrot in the late 70ies. We will
see more about fractals later in this course, when we look at complex maps.

STRANGE ATTRACTOR. An attracting set of a differential equation ẋ = F (x) or map x → T (x), is called
a strange attractor, if it has fractal dimension (we will define that below), sensitive dependence on
initial conditions (positive Lyapunov exponent) and which has an indecomposable physical measure
∫t
which means that for almost all initial conditions x0 and all continuous functions f , the limit 1t 0 f (Ts (x0 )) ds
∑ n
rsp. n1 k=1 f (T k (x) exists and depends only on f and not x0 .

The Lorenz attractor: the dimension is numerically around 2.05 (Doering

Gibbon 1995), 2.0627160 (Viswanath, 2002), The in-decomposability (techni-
cally called ”SRB measure”) (Tucker, 2002).

The Henon attractor: the dimension is measures 1.36 (Grebogi, Ott, Yorke,
1987) The in-decomposability had been shown (Benedicks and Carlson, 1991).

The Solenoid: This is a toy attractor, for which all the properties can be
proven. It is a strange attractor for a map in space.

DIMENSION. Let X be a a set in Euclidean space. Define the s-volume of accuracy r of a set X as hs,r (X) = nrs ,
where n is the smallest number of cubes of side length r needed to cover X. The s-volume is the limit
hs (X) = limr→0 hs,r (X). The box counting dimension is defined as the limiting value s, where hs (X) jumps
from 0 to infinity.

LINE SEGMENT. A line segment of length 1 in the

SQUARE. A square X of a plane of area 1 in space
plane can be covered with n intervals of length 1/n
can be covered with n2 cubes of length 1/n and
and hs,r (X) = n(1/ns ). For s < 1 this converges to
hs,r (X) = n2 (1/ns ) which converges to 0 for s < 2
0, for s > 1, it converges to infinity. The dimension
and diverges for s > 2. The dimension is 2.
is 1.

CIRCLE. A circle or radius DISC. A disc of radius 1 in space

1 can be covered with with can be covered with πn2 /4 <
2πn squares of length 1/n and N < πn2 squares of length 1/n
hs,r (X) = 2πn(1/ns ). For s < 1 and π(n2 /4)/n2 ≤ hs,r (X) ≤
this converges to 0, for s > 1, it πn2 /ns which converges to 0 for
converges to infinity. The dimen- s < 2 and diverges for s > 2.
sion is 1. The dimension is 2.
THE CANTOR SET. The Cantor set is constructed recursively by dividing
the interval [0, 1] into 3 equal intervals and cutting away the middle one re-
peating this procedure with each of the remaining intervals etc. At the k’th
stop, we need 2k intervals of length 1/3k to cover the set. The s-volume
hs,3−k (X) of accuracy 1/3k is 2k /3sk . It goes to zero if s < 2/3 and diverges
for s > log(2)/ log(3).

SHIRPINSKI CARPET. The Shirpinski carpet is constructed recursively by

dividing a square in the plane into 9 equal squares and cutting away the middle
one, repeating this procedure with each of the remaining squares etc. At the
k’th step, we need 8k squares of length 1/3k to cover the carpet. The s-volume
hs,1/3k (X) of accuracy 1/3k is 8k (1/3k )s which goes to 0 for k approaching
infinity if s is smaller than d = log(8)/ log(3) and diverges for s bigger than d.
The dimension of the carpet is d = log(8)/ log(3) = 1.893 a number between 1
and 2. It is a fractal.

SHIRPINSKI GASKET The Shirpinski gasket is constructed recursively by

dividing a triangle in the plane into 4 equal triangles and cutting away the
middle one, repeating this procedure with each of the remaining squares etc.
At the k’th step, we need 3k triangles of side length 1/2k to cover the gasket.
The s-volume hs,1/2k (X) of accuracy 1/2k is 8k (1/2k )s which goes to 0 for k
approaching infinity if s is smaller than d = log(3)/ log(2) and diverges for s
bigger than d. The dimension of the gasket is d = log(3)/ log(2), a number
between 1 and 2.

MENGER SPONGE. The three-dimensional analogue of the Cantor set in one

dimensions and the Shirpinski carpet. One starts with a cube, divides it into
27 pieces, then cuts away the middle third along each axes. It is your task to
compute the dimension. Note that the faces of the Menger sponge are decorated
by Shirpinski Carpets.

THE PROBLEMS OF THE DEFINITION. If one takes the above definition, then the dimension of the set
of rational numbers in the interval [0, 1] is equal to 1. A better definition, the Hausdorff dimension is
needed. We include that definition below but it is a bit more complicated. The problem with the box counting
dimension is that the size of the cubes should be allowed to vary. This refinement is similar to the change from
the Riemann integral to the Lebesgue integral. c

HAUSDORFF MEASURE. Let (X, d) be a metric space. Denote by |A| = supx,y∈A d(x, y) the diameter of a
subset A. Define for ϵ > 0, s > 0 ∑
hsϵ (A) = inf |U |s ,
Uϵ
U ∈U ϵ

where U ϵ runs over all countable open covers of A with diameter < ϵ. Such covers are also called ϵ-covers. The
limit
hs (A) = lim hsϵ (A)
ϵ→0

is called the s− dimensional Hausdorff measure of the set A. Note that this limit exists in [0, ∞] (it can
be ∞), because ϵ 7→ hsϵ (A) is increasing for ϵ → 0.

LEMMA: If hs (A) < ∞, then ht (A) = 0 for all t > s. Take ϵ > 0 and assume {Uj }j∈N is an open ϵ-cover of A.
Then ∑ ∑
htϵ (A) ≤ |Uj |t ≤ ϵt−s · |Uj |s .
j j

Taking the infimum over all coverings gives

htϵ (A) ≤ ϵt−s · hsϵ (A) .

In the limit ϵ → 0, we obtain from hs (A) < ∞ that ht (A) = 0.

HAUSDORFF DIMENSION.
Either there exists a number dimH (A) ≥ 0 such that

s < dimH (A) ⇒ hs (A) = ∞ ,

s > dimH (A) ⇒ hs (A) = 0

or for all s ≥ 0, hs (A) = 0. In the later case, one defines dimH (A) = ∞.
The number dimH (A) ∈ [0, ∞] is called the Hausdorff dimension of A.

FRACTAL. A fractal is a subset of a metric space which has finite non-integer Hausdorff dimension.

The Hausdorff dimension

∑ is in general difficult to calculate numerically. The central difficulty is to determine
the infimum over i |Ui |t , where U = {Ui } is an ϵ-cover of A. The box-counting dimension simplifies this
problem by replacing arbitrary covers by sphere covers and so to replace the terms |Ui |t by ϵt . The prize one
has to pay is that one can no more measure all bounded sets like this. In general, the upper and lower limits
differ.

UPPER AND LOWER CAPACITY. Given a compact set A ⊂ X. Define for ϵ > 0, Nϵ (A) as the smallest
number of sets of diameter ϵ which cover A. By compactness, this is finite. Define the upper capacity

log(Nϵ (A))
dimB (A) = lim sup
ϵ→0 − log(ϵ)

and analogous the lower capacity dimB (A), where lim sup is replaced with lim inf. If the lower and upper
capacities coincide, the value dimB (A) is called box counting dimension of A.

CAPACITY DIMENSION. If the lower and upper capacity are the same, one calls it the capacity dimension.

BOX COUNTING DIMENSION. Cover Rn by closed square boxes of side length 2−k . and let Mk (A) be the
number of such boxes which intersect A. Define the box counting dimension

log(Mk (A))
dimB (A) = lim .
k→∞ log(2k )

If the capacity dimension exists, then it is equal to the box counting dimension.

PROOF: Any set of diameter 2−k can intersect at most 2n grid boxes. On the other hand, any box of side 2−k
has diameter smaller than 2−k+1 . There exists therefore a constant C such that

C −1 · Mk (A) ≤ N2−k (A) ≤ C · Mk (A) .

Therefore
log(Mk (A)) log(N2−k (A))
lim = lim .
k→∞ log(2k ) k→∞ log(2k )

SELF-SIMILARITY. The computation of the dimension in the example objects

was easy because they are self-similar. A part of the object is when suitably
scaled equivalent to the object. We we will see more about this when we look at
iterated function systems. To measure or estimate the dimension of an arbitrary
object, one has to count squares. As an illustration of fractals in nature, one
often takes coast lines. A rough estimate of the coast of Massachusetts leads
to a dimension 1.3.
HISTORY.

The Cantor set is named after George Cantor (1845-1918), who was putting
down the foundations of set theory. Ian Stewart writes in ”Does God Play
Dice”, 1989 p. 121:
”The appropriate object is known as the Cantor set, because it was discovered
by Henry Smith in 1875. The founder of set theory, George Cantor, used
Smith’s invention in 1883. Let’s fact it, ’Smith set’ isn’t very impressive, is it?

The Hausdorff dimension has been introduced in 1919 by Felix Hausdorff

(1868-1942).

Abram Besicovitch, around 1930, worked out an extensive theory for sets with
finite Hausdorff measure.

The name ”fractal” had been introduced only much later by Benoit Mandelbrot
(1924-) in 1975.

The Sierpinski carpet was studied by Waclaw Sierpinski in 1916. He proved

that it is universal for all one dimensional compact objects in the plane. This
means that if you draw a curve in the plane which is contained in some finite
box, however complicated it might be and with how many self-intersections
you want, there is always a part of the Shirpinski carpet which is topologically
equivalent to this curve.

This might not look so surprising but this result is not true for the Shirpinski
gasket. The Menger Sponge was studied by Klaus Menger in 1926. He
showed that it is universal for all one dimensional objects in space. This means
whatever complicated curve you draw in space, you find a part of the Menger
sponge, which is topologically equivalent to it.
THE LORENZ SYSTEM Math118, O. Knill

ABSTRACT. In this lecture, we have a closer look at the Lorenz system.

THE LORENZ SYSTEM. The differential equations

ẋ = σ(y − x)
ẏ = rx − y − xz
ż = xy − bz .

are called the Lorenz system. There are three parameters. For σ = 10, r =
28, b = 8/3, Lorenz discovered in 1963 an interesting long time behavior and an
aperiodic ”attractor”. The picture to the right shows a numerical integration
of an orbit for t ∈ [0, 40].

DERIVATION. Lorenz original derivation of these equations are from a model for fluid flow of the atmosphere:
a two-dimensional fluid cell is warmed from below and cooled from above and the resulting convective motion is
modeled by a partial differential equation. The variables are expanded into an infinite number of modes and all
except three of them are put to zero. One calls this a Galerkin approximation. The variable x is proportional
to the intensity of convective motion, y is proportional to the temperature difference between ascending and
descending currents and z is proportional to the distortion from linearity of the vertical temperature profile.
The parameters σ > 1, r > 0, b > 0 have a physical interpretation. σ is the Prandl number, the quotient of
viscosity and thermal conductivity, r is essentially the temperature difference of the heated layer and b depends
on the geometry of the fluid cell.

SYMMETRIES. The equations are invariant under the transformation S(x, y, z) = (−x, −y, z). That means
that if (x(t), y(t), z(t)) is a solution, then (−x(t), −y(t), z(t)) is a solution too.
If (x0 , y0 , z0 ) = (0, 0, z0 ), then the equations are ż = −bz. Therefore, we stay on the z axes and to the equilibrium
point (0, 0, 0).

VOLUME. The Lorenz flow is dissipative: indeed, the divergence of F is negative. The flow contracts volume.

div(F ) = −1 − σ − b

A TRAPPING REGION.

A region Y in space which has the property that if x(t) ∈ Y then for all s > t
also x(s) ∈ Y is called a trapping region. A function, which is nondecreasing
along the flow is also called a Lyapunov function. Don’t confuse this with
the Lyapunov exponent.

LEMMA. There exists a bounded ellipsoid E which is a trapping

region for the Lorenz flow. The time-one map T of the Lorenz
flow maps E into the interior of E.

PROOF. We show that the function V = rx2 + σy 2 + σ(z − 2r)2 is a Lyapunov function outside some ellipsoid.
Indeed, the time derivative satisfies

V̇ = −2σ(rx2 + y 2 + bz 2 − 2brz) .

Define D = {V̇ ≥ 0 }. This is a bounded region. If c the maximum of V in D and E = {V ≤ c + ϵ } for some
ϵ > 0. then E is a region containing D. Outside this ellipsoid E, we have V̇ ≤ −δ for some positive δ. With an
initial condition ⃗x0 outside E, the vlaue of V (x(t)) decreases and within finite time, the trajectory will enter
the ellipsoid E. All trajectories pass inwards through the boundary of E so that a trajectory which is once
within E, remains there forever.
GLOBAL EXISTENCE. Remember that nonlinear differential equations do not necessarily have global solutions
like d/dtx(t) = x2 (t). If solutions do not exist for all times, there is a finite τ such that |x(t)| → ∞ for t → τ .

LEMMA. The Lorenz system has a solution x(t) for all times.

Since we have a trapping region, the Lorenz differential equation exist for all times t > 0. If we run time
backwards, we have V̇ = 2σ(rx2 + y 2 + bz 2 − 2brz) ≤ cV for some constant c. Therefore V (t) ≤ V (0)ect .
∩
THE ATTRACTING SET. The set K = t>0 Tt (E) is invariant under the differential equation. It has zero
volume and is called the attracting set of the Lorenz equations. It contains the unstable manifold of O.

EQUILIBRIUM POINTS. Besides √ the origin√ O = (0, 0, 0, we have two other

equilibrium points. C ± = (± b(r − 1), ± b(r − 1), r − 1). For r < 1, all
solutions are attracted to the origin. At r = 1, the two equilibrium points
appear with a period doubling bifurcation. They are stable until some
parameter r∗ . The picture to the right shows the unstable manifold of the
origin for σ = 10, b = 8/3, r = 10 which end up as part of the stable manifold
of the two equilibrium points.

HYPERBOLICITY IN THREE DIMENSIONS. An equilibrium point is called

hyperbolic if there are no eigenvalues on the imaginary axes. This is quite a
wide notion and includes attractive or repelling equilibrium points as well as
the possibility to have a one dimensional stable and two dimensional unstable
direction or a two dimensional stable and a one dimensional unstable direction.

THE JACOBEAN. The Lorenz differential equations ẋ

 = F (x) has the Jacobean DF (x, y, z) =
−σ σ 0
 r − z −1 −x .
y x −b

THE ORIGIN. At the equilibrium point (0, 0, 0),

√ the Jacobean DF (0, 0, 0) is
−1−s± (1−s)2 +4rs
block diagonal. The eigenvalues are −b, 2 . For r < 1, where
√
(1 − s)2 + 4rs < (1 + s), all three eigenvalues are negative. For r > 1, we
have one positive eigenvalues and two negative eigenvalue. To the positive
eigenvalue belongs an unstable manifold which is part of the Lorenz attractor.

THE TWO OTHER POINTS. At the two other equilibrium points, the eigen-
values are the roots of a polynomial of degree 3. For σ > b + 1 and
1 < r < r∗ = (σ(σ + b + 3)/(σ − b − 1), all eigenvalues have negative a
real part and the two points C ± are stable. At r = r∗ , a Hopf bifurcation
happens. The two stable points C ± collide each with an unstable cycle and
become unstable. For σ = 10, b = 8/3 we have r∗ = 470/19 = 24.7.

PERIODIC ORBITS. For large r parameters, the attractor can be single periodic orbit. Known windows are
99.534 < r < 100.795, 145.96 < r < 166.07,214.364 < r < ∞. Some periodic solutions are knots.

LYAPUNOV EXPONENTS OF DIFFERENTIAL EQUATIONS. If Tt (x0 ) = xt is the time t map defined by

d
the differential equation dt x = F (x), then

1
λ(F, x) = lim log ||DTt (x)||
t→∞ t
is called the Lyapunov exponent of the orbit. It is always ≥ 0. The Lyapunov exponent is for non-periodic
orbits only accessible numerically.
THE LORENZ SYSTEM II Math118, O. Knill

ABSTRACT. This is a continuation of the discussion about the Lorenz system and especially on the r depen-
dence of the attractor.

OVERVIEW OVER BIFURCATIONS. We fix the parameter σ = 10, b = 8/3.

For 0 < r < 1, the origin is the only equilibrium point and all
points attracted to this point (you can find a proof in the book. At
r = 1, a pitchfork bifurcation takes place. The origin becomes
unstable and two stable equilibrium points appear. The picture
shows the case r = 1.5.

For 1 < r < 13.925, the unstable manifold of the origin connects
to the equilibrium points. The picture shows r = 10.

For r = r0 = 13.926, the unstable manifold becomes double

asymptotic to the origin.

At the parameter r0 , two unstable cycles appear. For 13.926 <

r < 24.06, these cycles come closer to the fixed points C ± . The
picture shows the parameter r = 20.

At the parameter r = r1 = 24.74 = 470/19, the unstable cy-

cles collide with the stable equilibrium points and render them
unstable. This is called a subcritical Hopf bifurcation.
At the parameter r = 28, one observes the Lorenz attractor.

Between r = 0.99524 and r = 100.795, one observes an infinite se-

ries of period doubling bifurcations of stable periodic points (one
has to start with the larger value and decrease r). These bifurca-
tions are analogue to the Feigenbaum scenario. The picture shows
the parameter r = 100.

Here we see the previous stable periodic cycle doubled. The pa-
rameter is r = 99.7. The period doubling scenario leads to the
same Feigenbaum constant as one can see in the one dimensional
logistic map family.

RETURN MAP. A good Poincare map is part of the

subplane z = r − 1. This plane contains the equilib-
rium point C ± . These points are fixed points of the
return map.

HISTORICAL. Lorenz carried out numerical investigations following work of Saltzman (1962). The Lorenz
equations can be found in virtually all books on dynamics. We consulted:
• C. Sparrow, ”The Lorenz equations: Bifurcations, chaos and strange attractors, Springer Verlag, 1982
• Strogatz, ”Nonlinear dynamics and Chaos”, Addison Wesley, 1994

• Dynamical systems X, Encyclopadia of Mathematics vol 66, Springer 1988

• Dennis Gulick, Encounters With chaos, Mc Graw-Hill, 1992
• Clark Robinson, Dynamical systems, Stability, Symbolic Dynamics and Chaos, CRC priss, 1995
BILLIARDS I Math118, O. Knill

ABSTRACT. The billiard dynamical system can be seen as a limiting case of a particle moving in the plane
under the influence of a potential V . In the limit, the ODE of three variables becomes a simple map, which
still has all the features of differential equations. We discribe the system as an extremization problem, show
the existence of periodic orbits and the area-preservation property. We also see that the ellipse is an integrable
billiard.

PARTICLE MOTION IN THE PLANE. The mo- Written as first order system, there are 4 variables
tion of a particle in the plane under the influence of x, y, u, v. Energy conservation H(x, y, u, v) = u2 /2 +
a force F (x, y) = (f (x, y), g(x, y) = −∇V (x, y) is v 2 /2 + V (x, y) = E reduces it to three variables:
described by the differential equations d
x = u
d1 dt
x(t) = f (x, y) d √ √
dt2 y = 2 E − V (x, y) − u2 /2
dt
d2 d
y(t) = g(x, y) u = f (x, y)
dt2 dt

EXAMPLE. For V (x, y) = x4 + y 4 , the differential equations are

d
x = u
dt
d √ √
y = 2 E − x4 − y 4 − u2 /2
dt
d
u = −4x3
dt
The picture shows an orbit close to a periodic orbit.

THE BILLIARD FLOW. Now, we take a particle in the plane and use a poten-
tial V which is zero inside a region G and which is infinite outside G. The mass
point will move freely on a straight line until it hits the ”wall”. There it will
reflect, bouncing off using the reflection law ”incoming angle”=”outgoing an-
gle”. The Birkhoff billiard is the dynamics of this billiard dynamical system,
if the table is convex.

THE BILLIARD MAP. With an initial position s on the boundary, and an

angle θ we have new initial position and a new angle. If the boundary of the
table is parametrized by x ∈ [0, 1] and the angle by θ ∈ [0, π], we obtain a map
(s, θ) → (s1 , θ1 ).

BETTER COORDINATES. If we scale the table such that the table has length
1 and reparametrize the boundary of the table such that x is the arc length
from some point 0 on the curve to s and take y = cos(θ), we obtain a map
1
T(x,y)
0.5
T : R/Z × [−1, 1], T (x, y) = (x1 , y1 )
1 2 3 4 5 6
Topologically R/Z × [−1, 1] is an annulus or a cylinder with boundary. -0.5 (x,y)

-1

MONOTONE TWIST MAP. One boundary R/Z × {−1} is fixed and the other boundary R/Z × {1} is rotated
once. Both boundaries, when the angle is 0 or π consist of fixed points. The map has the twist property:
d
dy x1 (x, y) > 0. We prefer the (x, y) coordinates over the (s, θ) coordinates, because T becomes so area-
preserving, as we will see below.
THE LENGTH FUNCTIONAL. Let h(xi , xi+1 ) denote the Euclidean distance between two points of the table
(this is the distance in the plane and not the distance along the boundary). If x1 , x2 , ..., xn are successive
impact points of the trajectory, then cos(θi ) = −hxi (xi , xi+1 ) = hxi (xi−1 , xi )

PROOF: You can see the relation

cos(θ) = dh/ds by watching the length
change dh = dh(xi , xi+1 ), when xi is
replaced by xi + ds (first picture). The dh
second formula is seen when observing ds
the length change dh = dh(xi−1 , xi ) ds dh
when xi is replaced with xi +ds (second
picture).

THE EULER EQUATIONS. The billiard map can be described by the equation

hxi (xi , xi+1 ) + hxi (xi−1 , xi ) = 0

This second order difference equation for the variables xi is called the Euler
equation of the billiard system. Given x0 , x1 , we can use these equations to
get x2 , then use these equations again to get x3 etc.

VARIATIONAL PRINCIPLE. If x1 , x2 , ..., xn is a sequence of impact points

of the billiard map and the initial point x0 and the final point xn+1 are fixed,
then x0 , x1 , x2 , ..., xn is a billiard orbit if and only if (x1 , x2 , .., xn−1 ) is a critical
point of the function
∑n
H(x1 , x2 , ..., xn−1 ) = i=0 h(xi , xi+1 ).

PROOF: just check that ∇H = 0 gives the Euler equations. In other words,
the billiard path extremizes the total length of the path. For n = 2, where we
extremize h(x0 , x1 ) + h(x1 , x2 ) we have to find the point x1 on the table such
that the path initiating at x0 and ending at x2 and which hits the table at a
point x1 is extremal.
This generalizes the Fermat principle: a light ray reflecting at a curve extremizes the distance to the curve only
if in- and out-going angles are the same.

PERIODIC POINTS. A sequence x1 , x2 , ..., xn , xn=1 = x1 is a periodic orbit if

and only if the total length of the polygon of the impact points is extremal. In
other words, we look for critical points of the total length of the closed polygon,
which is:
n
∑
H(x1 , ..., xn ) = h(xi , xi+1 )
i=1
= h(x1 , x2 ) + h(x2 , x3 ) + ... + h(xn−1 , xn ) + h(xn , x1 )

EXISTENCE OF PERIODIC POINTS. Since H is bounded, nonnegative and smooth, we have both a minimum
and a maximum. The global minimum is of course when x1 = ..., xn are all the same points. The maximum
leads to a true periodic point: we have shown

For a convex smooth billiard table, we find periodic

points of minimal period n if n is prime.

PROOF. A continuous function on a bounded and closed subset of Rn has a maximum. The period can not be
a factor of n because n was assumed to be prime. You show in a homework that the primality assumption is
not necessary.
Example: The long axes and short axes of a convex table are periodic orbits
of period 2.

Example: Triangles of maximal total length in the table are billiard orbits of
period 3.

BILLIARD IN A CIRCLE. The circle is an example of an integrable billiard.

The angle θ and so F (x, y) = y = cos(θ) is preserved. The billiard map T on
(R/Z) × [−1, 1] is given explicitly by

T (x, y) = (x + 2arccos(y)/(2π), y)

This is a shear map. On the first coordinate we have a rational or irrational

rotation.

KRONECKER SYSTEM. The dynamical system on the circle obtained by a translation T (x) = x + α mod 1
is called the Kronecker system. Let xn = [nα] = nα mod 1 be the orbit of T (x) = [x + α] on the circle R/Z.

LEMMA. The sequence xn = T n (x0 ) is dense on [0, 1] if α is

irrational.

PROOF. Given n divide [0, 1] into n equal intervals of length 1/n. Take an
orbit of length n + 1. By the pigeon hole principle, two of these points
0, α, ..., nα must be in the same interval and so have distance < 1/n. Therefore
δ = mα = (k − l)α < 1/n for some integer m. With an integer N larger than
1/δ, the set {mα = δ, 2mα = 2δ, ..., mN α = N δ} intersects every interval of
length δ at least once. The set {x0 , x1 , ...., xmN } intersects every interval of
length δ and so every interval of length 1/n.

Illustration: 6 pigeons and 5 holes. Two pigeons must be in the same hole.

COROLLARY. If (s, θ) is an initial point for the billiard in a circle, then the orbit is periodic if θ/(2π) is
rational. The ball will visit arbitrarily close to any given point of the table, if θ/(2π) is irrational.

CAUSTICS. For a billiard curve, one calls a curve a caustic, if the billiard
ball, once tangent to that curve, remains tangent after the reflection.

EXAMPLE: For a circular table, every concentric circle inside the table is a
caustic. For an ellipse, every confocal ellipse inside the table is a caustic.
EXAMPLE: given a convex curve, we can find a table which has this curve as
a caustic using the string construction.

GENERAL CAUSTICS IN OPTICS. Places, where families of light rays focus

are called caustics. If you take a family of parallel light and reflect it at a
circle, then the light rays will focus at a curve which is called the coffee cup
caustic. If the family of light rays is an orbit of a billiard ball in a table,
then caustics might exist or not. In the case of the circle, every orbit produces
caustics.
BILLIARD IN AN ELLIPSE.

The billiard in an ellipse is integrable.

PROOF. We find an invariant function F (x, y), which is the product

d1 (x, y), d2 (x, y)), where di (x, y) is the distance of the trajectory to the focal
point Fi . You will run a few lines of Mathematica to verify this in class.

BIRKHOFF-PORITSKY CONJECTURE: Is every integrable smooth convex

billiard an ellipse? A collaborator of Birkhoff at Harvard with name Hillel
Poritsky had worked on it and published a paper in 1950, where he made
some progress.
The picture shows Poritsky in 1936 at the 42. Summer Meeting of the Mathe-
matical Organizations of America in Cambridge, Massachusetts.

THEOREM. The billiard map is area-preserving.

PROOF. Let Y ⊂ T 1 × [−1, 1] be disc with boundary C. We show

∫∫ ∫
Y
dydx = Y dy ′ dx′ , where T (x, y) =
(x′ , y ′ ), T 2 (x, y) = (x′′ , y ′′ ) is the map. (We use primes here not as derivatives Using Greens formula, we get
∫ ∫ ∫ ∫
Area(T −1 (Y )) = dy dx = y dx = h1 (x, x′ ) dx
T −1 (Y ) T −1 (C) T −1 (C)
∫ ∫ ∫ ∫ ∫
= h1 (x′ , x′′ ) dx′ = −h2 (x, x′ ) dx′ = y ′ dx′ = dy ′ dx′ = Area(Y ) .
C C C Y

GENERALIZATION. Every map defined by the Euler equations h2 (x, x′ ) + h1 (x′ , x′′ ) of a smooth generating
function h(x, x′ ) is area-preserving in the coordinates (x, y) = (x, h1 (x, x′ ).

EXAMPLE. h(x, x′ ) = (x′ − x)2 /2 + V (x) leads to the Euler equation h1 (xi , xi+1 ) + h2 (xi−1 , xi ) = (xi+1 −
d
xi ) + dx V (xi ) − (xi − xi−1 ) = This is the second order difference equation xi+1 − 2xi + xi−1 + V ′ (xi ) = 0. Vor
V (x) = c cos(x), this recursion is the Standard map. For cubic V , it leads to the Henon map in the plane.

THE JACOBEAN MATRIX. An other proof to show that the map is area-preserving is to compute the Jacobean
matrix and to verify that the determinant is 1. We will write down the Jacobean later. An other proof of the
area-preservation property is given in proposition 6.4.2 of the textbook.

HISTORY.
Ludwig Boltzmann (1844-1906) studied the hard sphere gas. This is a
billiard system.
Emil Artin (1898-1962) looked in 1924 at billiard in the hyperbolic plane.
This is of interest in algebra.
Jacques Hadamard (1865-1963) Hedlund-Hopf studied the geodesic flow,
which is a generalization of billiards.
George Birkhoff (1884-1944) in 1927, proposed convex billiards as a model
for the 3-body problem
Hillel Poritsky in 1950 posed the integrability question.

WHY STUDY BILLIARDS?

It is a beautiful and simple dynamical system featuring all the complexities of
more complex systems. It is a limiting case of the geodesic flow and illustrates
theorems in topology, geometry or ergodic theory. It is related to Dirichlet
spectral problem ∆u = λu which can be considered the ”quantum version”
of the billiard problem, where the eigenfunctions describe a quantum particle
moving freely in the table with energy λ.
CHAOTIC BILLIARDS Math118, O. Knill

ABSTRACT. Billiards in tables with negative curvature as well as billiards like the Stadium are chaotic: The
Lyapunov exponent is positive. They are actually ergodic: every invariant set of positive measure will have
either area 0 or area 1.

POINCARES RECURRENCE THEOREM. Area preservation allows to make

a statement about recurrence of area-preserving map defined on a T invariant
subset in the plane. For example, X could be the annulus R/Z × [−1, 1] and
T could be a billiard map.

For every set Y of positive area |Y |, there exists n such that

T n (Y ) ∩ Y has positive area.

PROOF OF POINCARES THEOREM. Assume there exists a set Y of positive 0.75

area m(Y ) such that Yi = T i (Y ) satisfies m(Yi ∩ Y ) = 0 for all i > 0. Because 0.5

m(Yi ) = m(Y ) > 0 and the total space has finite area, there must exist 0 < i < 0.25

j such that m(Yi ∩Yj ) > 0. (This is a variant of the pigeon hole principle. If you 0.2 0.4 0.6 0.8 1

have a cage with finite room and each pigeon needs the same amount of space, -0.25

only a finite number of pigeons fit). But m(T −i (Yi ∩ Yj )) = m(Y ∩ Yj−i ) > 0 -0.5

contradicts that Y and Yk are disjoint. -0.75

-1

CONSEQUENCE FOR BILLIARDS. Does this mean that if you start shoot-
ing from a certain point in a certain direction, there will be times, when the
orbit will come back to a similar spot on the table with a similar angle? Not
necessarily. For example, if you are on the stable manifold of an unstable pe-
riodic point, then the orbit will converge to that periodic orbit. The Poincaré
statement is a statement about sets. It assures for example, that if you start
shooting from a certain interval on the table in a certain interval of directions,
you will come back to that range of initial conditions with probability 1.

0.75

ERGODICITY. Less obvious is the question, whether a given set ever reaches 0.5

an other set. If all ”measurable” invariant subset of the annulus have either area 0.25

1 or 0, then the map is ∫ ∫called ergodic. Measurable is a technical term which 0.2 0.4 0.6 0.8 1

assures that the area A

1 dxdy is defined. Any set which can be defined by -0.25

a (possibly infinitely) sequence of intersections or unions is measurable. -0.5

-0.75

-1

0.75

0.5
INVARIANT CURVES PREVENT ERGODICITY. If a billiard has an invari-
0.25
ant curve which is the graph of a function {y = f (x)}, then if (x0 , y0 ) is below
0.2 0.4 0.6 0.8 1
the graph, the entire orbit (xn , yn ) stays below the graph for all times. The -0.25

billiard can not be ergodic. -0.5

-0.75

-1

STRING CONSTRUCTION. It had been known since a long times, that if one
starts with a convex curve, winds a closed string around it and drags the string
around the curve which keeping the string tight, we obtain a table, which has
the original curve as a caustic. The picture shows some tables which have a
triangle as a caustic. These tables are not ergodic.
GLANCING BILLIARDS. An orbit (xj , yj ) of a billiard table for which yj
comes arbitrarily close to −1 and arbitrarily close to 1 is called a glancing
billiard orbit.

THEOREM. (Birkhoff) There are no invariant curves of T , if and only if there

exists a glancing orbit.

PROOF. If there is an invariant curve, there is trivially no glancing orbits because the
regions on both sides of the curve are left invariant. Assume now there is no glancing
orbit. This means there is an ϵ > 0 such that for all y0 < 1 − ϵ we have yn > −1 + ϵ.
Consider the region Y = {y < 1 − ϵ}. The set n T n (Y ) is a T -invariant set which
∪
does not intersect {y > 1 − ϵ}. The boundary of this curve is an invariant curve.
(One actually knows that such a curve must be the graph of a Lipshitz continuous
function).

THE JACOBEAN. Let κi denote the curvature at the impact point and angle θi the impact angle and let li
the length of the path from the impact point xi−1 to the impact point xi . The following formula is well known
in geometrical optics and used everywhere in the billiard literature like in the book of Kozlov-Treshchev.

LEMMA: There are coordinates for which the Jacobean DT (xi , yi ) of the bil-
liard map has the form
[
1 0
] [ ]
1 li
Bi = 2κi ·
− sin(θ i)
1 0 1

Remark: This is the composition of the Jacobean belonging to the translation and the Jacobean belonging to
the reflection at the wall. The value gi = sin(ϕ
2κi
i)
is the length of the billiard ball in the circle on the normal to
the reflection point which is tangent to the table and has radius 1/(2κi ).

PROOF OF THE JACOBEAN FORMULA. The formula can be derived geometrically. Instead, we find an
algebraic derivation from the Euler equations. It is still a bit messy.

We use the notation h1 , h11 for the first and second partial
[ derivative
] [ with]respect to the first variable and similar h12
xi xi+1
for the mixed partial derivative. The billiard map S : → is equivalent to the second order recursion
xi−1 xi
h1 (xi , xi+1 + h2 (xi−1 , xi ) = 0. Differentiation of these Euler equation with respect to xi , xi−1 gives ∂xi+1 /∂xi =
−bi /ai , ∂xi+1 /∂xi−1 = −ai−1 /ai , where
ai = h12 (xi , xi+1 )
and
bi = h11 (xi , xi+1 ) + h22 (xi−1 , xi ) .
The Jacobean of S is [ ]
−bi /ai −ai−1 /ai
dS = .
1 0
[ ]
a−1
i 0
With a first coordinate transformation Fi = we can achieve that the determinant is 1:
0 1
[ ]
−bi −a2i−1
Fi−1 dSFi−1 = Ai = (ai−1 )−1 .
1 0

Geometrically, we have
sin(θi ) sin(θi+1 ) 1 1
ai = , bi = sin2 (θi )( + ) − 2 sin(θi )κi ,
li li li−1
where li = h(xi , xi+1 ) are the lengths of the secants, θi[= θ(xi , xi+1 ) and κi = ] κ(xi ) are the curvatures at the reflection
0 − sin(θi )
points. Plugging this in the Jacobean gives with Gi = the new Jacobean
1/ sin(θi ) sin(θ)/li
[ ] [ ] [ ]
1 li 1 0 1 li
G−1
i · Ai · Gi−1 = 2κi 2κi = 2κi · .
− sin(θ i)
1− sin(θi )
− sin(θ i)
1 0 1
STABILITY OF PERIOD 2 ORBITS. Having the Jacobean given in geometric
terms allows to see, whether periodic orbits are stable or not. Inspection of the
trace of B2 B1 (a matrix which is similar to the Jacobean of T 2 and so has the
same trace) shows:

LEMMA. Assume ρi are the radii of curvature at the impact points. Assume
ρ1 < ρ2 . If l > ρ1 + ρ2 or ρ1 < l < ρ2 , then the periodic orbit of period 2 is
hyperbolic. If l > ρ2 or l < ρ1 + ρ2 , it is elliptic.

The fastest verification of th lemma is to run a line of Mathematica which gives the trace of the product of the
four matrices. For example, the long axis of a non-circular ellipsoid is a hyperbolic periodic point. The short
axis is an elliptic periodic point.

CURVATURE. If r(s) is a a curve in the plane parametrized by arc-length,

then the curvature κ(t) is |r′′ (s)|. If r(t) is the curve given by an arbitrary
parameterization, define the unit tangent vector T⃗ (t) = ⃗r ′ (t)|/|⃗r ′ (t)|. We get
the curvature κ(t) = |T⃗ ′ (t)|/|⃗r ′ (t)|. The function ρ(t) = 1/κ(t) is called the
radius of curvature. With the crossed product (a, b) × (c, d) = ad − bc in
′
r ′′ (t)|
two dimensions, we have a more convenient formula κ(t) = |⃗r (t)×⃗ r ′ (t)|3
|⃗ .

ROLE OF CURVATURE. The curvature of the table plays an important role

for the billiard dynamics. Here are some known results:

• Mather has shown that if the table has a flat point, this is a point at
which the curvature vanishes like at 4 points of x4 + y 4 = 1, then the
billiard map T has no invariant curve at all.
• Lazutkin and Douady have proven using KAM theory that for a smooth
billiard table with positive curvature everywhere, there always are ”whis-
per galleries” near the table boundary.

• From Andrea Hubacher (who had obtained this result as an undergradu-

ate student at ETH) is the result that a discontinuity in the curvature of
the table does not allow caustics near the boundary. For example, tables
obtained by the string construction at a triangle (see homework) do not
allow invariant curves near the boundary.
• It is easy to see that billiards for which the table has negative curvature
everywhere, the Lyapunov exponent is positive. The Matrices Bi have
then positive entries as we will just see.

POSITIVE MATRICES. If we multiply positive matrices with each other, the norm of the product grows
exponentially.

LEMMA. If det(A(x)) = 1 for all x and [A]ij (x) ≥ ϵ > 0, then the Lyapunov exponent λ(x) =
limn→∞ log ||A(T n−1 x)A(T k−2 x) · · · A(x)|| satisfies λ(A) ≥ 12 log(1 + 2ϵ2 ).

PROOF (Wojtkowski). Define the function F on pairs of vectors by v = (v1 , v2 ) 7→ F (v) = (v1 · v2 )1/2 . For a matrix B
with determinant 1 satisfying [B]ij (x) ≥ ϵ, define ρ(B) = inf F (v)=1 F (Bv).
[ ]
2 1/2 a b
(i) Given a 2 × 2-matrix A satisfying [A]ij ≥ ϵ. Then ρ(A) ≥ (1 + 2ϵ ) . Proof: If A = and w = (w1 , w2 )
c d
with F (w) = (w1 w2 )1/2 = 1, then F (Aw) = (aw1 + bw2 )1/2 (cw1 + dw2 )1/2 ≥ (ad − bc + 2bc)1/2 ≥ (1 + 2ϵ2 )1/2 .
|Av| F (Av)
(ii) ||B|| ≥ ρ(B). Proof: Take v = (1, 1). Then ||A|| ≥ |v|
≥ F (v)
≥ ρ(A).
F (ABv)
(iii) ρ(AB) = inf F (v)=1 F (ABv) ≥ inf F (Bv)=1 F (Bv)
· inf F (v)=1 F (Bv) = ρ(A) · ρ(B).

(iv) We get from (ii),(iii),(i) that 1

n
log ||An (x)|| ≥ 1
n
log(ρ(A(T n−1 x) . . . ρ(A(x))) ≥ 1
n
log((1 + 2ϵ2 )n/2 ).
 sin(θ)
CLASSES OF CHAOTIC BILLIARDS. Remember that g = 2κ , and l is the length of the trajectory.

THEOREM (Wojtkowski) Assume, a piecewise smooth convex table has the

property that for any pair of points x, x′ , on the non-flat parts of the curve
2g + 2g ′ ≤ l(x, x′ ), with strict inequality on a set of positive measure, then the
billiard map T has positive Lyapunov exponents on a set of positive measure.

PROOF. The Jacobian matrix is conjugated to B2 (x)B1 (x). A vector v = (1, f ) is mapped by the matrix B1 (x) to the
vector (1, f + l(x)). This vector is then mapped by B2 (x) to the vector

(1 − (f + l(x))/2g(T x), f + l(x))

which is after a rescaling of length equal to the vector

(f + l(x))g(T x)
(1, .
2g(T x) − f − l(x)

If we don’t care about the length of the vector, the map v 7→ B(x)v is determined by the map

1 (f + l)2g(T )
K : f 7→ f + l 7→ = .
1/(f + l) − 1/g(T ) 2g(T ) − f − l

At each point x ∈ X, we define a basis given by e2 (x) = (1, 0) and e1 (x) = (1, −g(x)).
Claim: Assume 2g(x) + 2g(T x) ≤ l(x) with inequality on a set of positive measure. In this basis, the matrix B(x) is
positive and there exists a set of positive measure, where B(x)ij ≥ ϵ > 0 for some ϵ > 0 so that we can apply the
previous lemma on positive matrices.

Proof. We have to show that the map K maps the interval [0, −2g(x)] into the interval [0, −2g(T x)] and into its interior
for a set of positive measure because:
l(x)2g(T x)
K(0) = ≥ −2g(T x) .
2g(T x) − l(x)
(−2g(x) + l(x))2g(T x)
K(−2g(x)) = ≤0.
2g(T x) + 2g(x) − l(x)

BUNIMOVICH STADIUM. A famous example is

the stadium, where two half circles are joined by
straight lines. An other example is the rounded
square.

For these billiards, one knows actually much more.

They are ergodic and chaotic in the sense of Devaney,
a notion we have met earlier in this course. The prove
of ergodicity is not so easy. One has to analyze some
stable and unstable manifolds and verify that they
are dense.

OPEN PROBLEMS. The following problems are open mathematical problems.

The first two problems probably go back to Poincaré. The third problem is
an old problem in smooth ergodic theory. The difficulty of that problem is
that for a smooth convex billiards, there are lots of invariant curves and also
lots of elliptic periodic orbits consequently, the chaotic regions are mingled
well with the stable regions and the techniques described in this handout do
not work.

1) Are periodic orbits dense in the annulus for a general smooth Birkhoff
billiard?

2) Is the total measure (”area”) of the periodic orbits always zero in the
annulus? One knows it for period 3 (Rychlik).

3) Does there exist a smooth convex billiards with positive Lyapunov exponents
on a set of positive measure (=”area”=”probability”)?
EXTERIOR BILLIARDS Math118, O. Knill

ABSTRACT. We look here briefly at the dynamical system called ”exterior billiard”. Affine equivalent tables
lead to conjugated dynamical systems. One does not know, whether there is a table for which an orbit can
escape to infinity nor does not know whether the ellipse is the only smooth convex exterior billiard table for
which the dynamics is integrable.

EXTERIOR BILLIARDS. Dual billiards or exterior billiards is played out-

side a convex table γ. Take a point (x, y) outside the table, form the tangent
at the table and reflect it at the tangent point (or the mid-point of the in-
terval of intersection). To have no ambiguity with the tangent, γ is oriented
counter clockwise. The positive tangent is the tangent at the curve in the same
direction.

EQUIVALENCE. Assume S(x) = Ax + v is an affine transformation in the

plane, where A is a linear transformation and v is a translation vector. Given
two tables γ1 , γ2 such that S(γ1 ) = γ2 , then the exterior billiard systems Tγ1
and Tγ2 are conjugated.

PROOF. Unlike angles, affine transformations preserve ratios and a trajectory

of the the exterior billiard at γ1 is mapped into a trajectory of the table γ2 .

EXAMPLE POLYGONS. Already the case of polygons can be complex. Ex-

terior billiard at a general quadrilateral (=four sided polygon) shows already
interesting dynamics. Note that the exterior billiard map is not continuous
for polygons. One already does not know whether orbits stay bounded for all
quadrilaterals. For regular pentagons, Tabatchnikov was able to compute the
Hausdorff dimensions of the closure of some orbits. They are fractals.

INTEGRABLE PARALLELEPIPED. The exterior billiard at a parallelepiped

is integrable.

PROOF. By affine equivalence, it is enough to show this for squares. Check

that every orbit is periodic.

INTEGRABLE ELLIPSE. The exterior billiard at an ellipse is integrable.

PROOF. By affine equivalence, it is enough to show this for circles.

INTEGRABLE TRIANGULAR BILLIARD. The exterior billiard at any

triangle is integrable.

PROOF. By affine equivalence, it is enough to show integrability for equilateral

triangles. Since every orbit is periodic, we have integrability by a lemma proven
earlier. For the hexagon, we have also the property that every orbit is periodic.
INTEGRABLE HEXAGONAL BILLIARD. The exterior at a regular hexagon
is integrable.

PROOF. The key is to see that the successive reflections of the sides of the
polygon at the corners of the polygon produces a regular tessellation of the
plane.

GENERATING FUNCTION. Similar as for billiards, there is a generating

function h(x, x′ ) for the exterior billiard. Given two polar angles φ, φ′ , draw
the tangents with this angle. The function h(φ, φ′ ) is the area of the region
enclosed by these lines and the curve. We can check that the partial derivative
∂ ′ 2
∂x h(φ, φ ) = −r /2, where r is the distance from the point to the point of
tangency. The exterior billiard is area-preserving.

PERIODIC POINTS. By maximizing the functional

∑
n
H(x1 , ...., xn ) = h(xi , xi+1 )
k=1

one obtains periodic orbits of the exterior billiard. To say it in words: among
all closed polygons for which all sides are tangent to the table, the ones which
maximizes the sum of the areas h(xi , xi+1 form a periodic orbit of the dual
billiard.

INVARIANT CURVES. For smooth tables, every orbit is bounded. This is a

consequence of KAM (Kolmogorov-Arnold-Moser) theory. In that case, there
are invariant curves far from the table which enclose the table. A point on this
curve will remain on this curve for all times and the dynamics is conjugated to
a Kronecker system. A proof of the ”invariant curve theorem” is not easy: it
requires heavy analytic artillery, modifications of the Newton method or ”hard”
implicit function theorems. One has to find a smooth invertible map on the
circle such that h(q(x − α), x) + h(x, q(x + α)) = 0 is satisfied. The irrational
rotation number α has to be ”far away from rational numbers”, one calls this
Diophantine. For the story of dual billiards, the proof is even more tricky and
has been done by R. Douady.

AN UNSOLVED PROBLEM. Is the ellipse the only smooth convex table for
which exterior billiard is integrable?

AN UNSOLVED PROBLEM. Is there a table with an unbounded orbit?

An example of where one does not know the answer, is a semicircle. Tabatch-
nikov states numerical evidence that for this billiard, there is an unbounded
orbit.

HISTORY.
1960. The problem is suggested by B.H. Neumann
1963 The problem is posed by P.Hammer in a list of unsolved problems
1973 In Moser’s book ”Stable and Random Motion”, the stability problem is
raised. Some people call exterior billiard also the Moser billiard.
1978 The exterior billiard is also featured in Mosers Intelligencer article ”Is the
solar system stable”.
The photo of Moser to the right had been taken by J. Pöschel in the year 1999,
when Moser was lecturing in Edinburgh about twist maps. Moser died in the
same year.
FIXED POINT THEOREMS Math118, O. Knill

ABSTRACT. Fixed point theorems are important in dynamics.

BANACHS FIXED POINT THEOREM.

A contraction T in a complete metric space X has a fixed point.

This theorem can be used for example to prove the existence of solutions to differential equations.

BROWERS FIXED POINT THEOREM.

Every continuous map T from the unit ball Dn = {x ∈ Rn | ||x|| ≤

1 } onto itself has a fixed point.

SKETCH OF PROOF FOR n = 2. If T (x) ̸= x for all x ∈ Dn , one can find a continuous map g from Dn to its
boundary S n−1 : the point g(x) is the intersection of the line through x and T (x) with S n−1 . This map is the
identity on the boundary. If such a map existed, one could smooth it. We would have a smooth map from the
interior of D to S n−1 . For most y ∈ S n−1 the set S −1 (y) is a curve in D which begins and ends at y. The re-
gion it contains must by continuity also be mapped to y and S −1 (y) would contain a disc and can not be a curve.

REMARK TO 1D: The Brower fixed point theorem in one dimensions, (D1 is an interval [a, b]) follows from
the intermediate value theorem: Since T (a) ≥ a, T (b) ≤ b, the function g(x) = T (x) − x satisfies g(a) > 0 and
g(b) < 0. It must have a root. This root is the fixed point. theorem.

KAKUTANI FIXED POINT THEOREM.

A continuous map T on a compact convex set D in locally convex

space X has a fixed point.

(One can relax the condition that T must be a map: it can also be a correspondence for which T (x) is a convex
subset of X.) A locally convex set is a vector space in which the topology is given by a sequence of seminorms.
An example is C ∞ (R), the space of all infinitely many times differentiable functions.) While von Neumann
used Browers fixed point theorem, John Nash was among the first to use Kakutani’s Fixed Point Theorem in
game theory, where fixed points can lead to equilibria.

POINCARE BIRKHOFF THEOREM.

An area-preserving transformation on the annulus, which moves

boundary circles in the opposite directions has at least two distinct
fixed points.

Poincare had conjectured this but could no more prove it. The conjecture was therefore called Poincarés last
theorem. It was George Birkhoff who proved it in 1917.

APPLICATION TO BILLIARDS.

COROLLARY. For a billiard in a smooth convex table, there are

at least 2 periodic orbits of type 0 < p/q < 1 meaning that T q
winds around the table p times.

PROOF. The map T q leaves one boundary of the annulus X = T 1 × [−1, 1] fixed, the other boundary is turned
around q times. Now define S(x, y) = (x − 1, y) which rotates every point once around. Now, T q S −p rotates
one side of the boundary by −2πp and the other side of the boundary by 2π(q − p). Since the boundary is
now turned into different directions, there are fixed points of T q S −p . For such a fixed point T q (x, y) = S p (x, y)
which is what we call orbit of type 0 < p/q < 1.
APPLICATIONS TO DUAL BILLIARDS.

COROLLARY. For exterior billiard at a smooth convex table,

there are at least 2 periodic orbits of type 0 < p/q < 1/2 meaning
that T q winds around the table p times.

Periodic orbits with small rotation numbers p/q are close to the table, periodic orbits with rotation number
close to 1/2 are far away from the table.
POLYGONAL BILLIARDS Math118, O. Knill

ABSTRACT. Billiards in polygons are integrable in the case of rectangles, regular triangles or hexagons.

INTEGRABLE SQUARE. The square and the rectangle are example of an integrable billiard. If θ is the impact
angle, then F (s, θ) = sin(2θ) is an integral.

INTEGRABLE POLYGONAL BILLIARDS. If unfolding the polygon produces a tessllation of the plane, the
corresponding billiard is integrable.

TRIANGULAR BILLIARDS. Even for triangles, the billiard dynamics is com-

plicated. There are many open questions, one of the most astonishing ones is
the open problem:

Does every triangular billiard have a periodic orbit?

One can solve the problem for a triangle with a right angle? The answer is
easy - if you see it. One can also solve the problem for acute triangles, where
the Fagnano trajectory connecting the footpoints of the triangles altitudes
is a periodic orbit.
LETS PLAY SOME GAMES: Lets mention without proof that the Lyapunov exponent of a polygonal billiard
is always zero. The chaos, you obtain with these systems is ”weak”.

LYAPUNOV
[ EXPONENTS. [ Because ]the[ Jacobean matrix of a billiard is conjugated to
1
]
1 0
]
li 1 li
2κi 2κi = 2κi · and the curvature in a polygonal billiard is zero,
− sin(θ i)
1 − sin(θ i)
− sin(θ i)
1 0 1
we have

All Lyapunov exponents are zero in polygonal billiards.

CONNECTIONS WITH OTHER FIELDS. The mathematics of billiards in polygons has relations with other
fields like Riemann surfaces, Teichmuller theory and leads to interesting ergodic theory. One knows for example
that for a ”generic” polygon, the billiard map is ergodic.
CELLULAR AUTOMATA Math118, O. Knill

ABSTRACT. A shift invariant continuous map on the sequence space AZ over a finite alphabet A is called
a cellular automaton or short a CA. These dynamical systems can be considered as discretized cousins of
differential equations, for which time, space, as well as the configuration space are discretized.

THE NAME CELLULAR AUTOMATON. Interactions between different sci-

entific fields is always productive. Historically, it seems that cellular automata
were introduced in the late 40ies while some applied Mathematicians were
dealing with problems from biology. The etymology of the name ”CA” could
confirm a ”bonmot” of Stan Ulam:

Ask not what mathematics can do for biology.

Ask what biology can do for Mathematics.

Source: cited from David Campbell, who received his B.A. in chemistry and physics
from Harvard in 1966 and worked in nonlinear science. Ulam himself was at Har-
vard from 1936-1939, eating at Adams house where ”the lunches were particularly
agreeable” and was also teaching the Math1A here (Source: Ulam: Adventures of a
mathematician).
Anyway, it would not surprise if ”cellular automaton” had been derived from
”cellular spaces” because of mathematical research on biological problems.

SEQUENCE SPACES. Let A be a finite set called the alphabet and let AZ denote the set of all sequences and
σ(x)n = xn+1 the shift on X. A distance between two sequences is given by d(x, y) = 1/(n + 1), where n is the
largest number such that xi = yi for |i| ≤ n. Example: Let A = {1, 2, 3, 4 }. For
... x−3 x−2 x−1 x0 x1 x2 x3 ... ... y−3 y−2 y−1 y0 y1 y2 y3 ...
... 1 1 4 3 2 1 1 ... ... 1 2 3 3 4 1 1 ...
we have d(x, y) = 1/3, because xi = yi if |i| ≥ 3 but x−2 ̸= y−2 .

LEMMA: X is a compact metric space (X, d).

PROOF. To have a metric space, show d(x, x) = 0, d(x, z) ≤ d(x, y) + d(y, z), d(x, y) = d(y, x). To have com-
pactness, every sequence x(k) in X must have an accumulation point. That is, there must exist a subsequence
x(kl ) in X which converges for k → ∞. See homework.

1D-CELLULAR AUTOMATA. A continuous map T on X which commutes

with σ is called a cellular automaton. A theorem of Curtis, Hedlund and
Lyndon, which we will prove later implies that there is a function φ from
A2R+1 → A such that T (x)i = φ(xi−R , xi−k+1 , ..., xi+R ). The integer R is
called the radius of the CA. It is assumed that R is the smallest number for
which the CA still can be defined like that. One can visualize the dynamics
of one dimensional CA by coding each letter in a sequence with a color. The
first row is the initial condition. Applying the map gives the second row, etc.
Drawing a few iterates produces a phase space diagram. The example shows
the automaton over the alphabet {0, 1}, where xn = xn +xn−1 mod 2 and where
0 is black. If initially xn (0) = 0 for n ̸= 0 and x0 (0)
( = 1, we) have an explicit
n+t
solution formula with binomial coefficients xn (t) = mod(2).
n

CANTORS DIAGONAL ARGUMENT.

THEOREM (Cantor) The set X = AZ is uncountable.

PROOF. If X were countable, one could enumerate all sequences x(k) using
integer indices k. Define the ”Diagonal” sequence yn = (1 + xn (|n|)) (here a + 1 is
the next in the alphabet A, or the first element in A, if a was the last). The sequence y is different

from any of the sequences x(k) because y and x(k) differ at the k’th entry. The
assumption about the enumerability was not possible.
WOLFRAMS NUMBERING OF 1D CA. Any one-dimensional cellular
automata with radius 1 and alphabet {0, 1} can be labeled by a rule number.
Because there are 23 = 8 possible∑maps φ, we have 28 = 256 possible rules.
8 k
The Wolfram number is w = k=1 f (k)2 , where y0 = f (k) is the new
color for k = 4x−1 + 2x0 + x1 .

For example, let φ(a, b, c) = a, then the new middle cell is 1 for the neighbor-
hoods 111, 110, 101, 100 which code the integers 7, 6, 5, 4. So, f (7) = f (6) =
f (5) = f (4) = 1, and f (k) = 0 otherwise. The rule of the automaton is
w = 27 + 26 + 25 + 24 = 240. Indeed, rule 240 is the shift automaton. Let us
look at an other example.

EXAMPLES. The binominal CA discussed above has rule 90. One of the most
studied CA is rule 18. Since 18 = 24 + 21 , which is 10010 to the base 2, we
obtain the following function φ:

neighborhood (dec) neighborhood (bin) new middle cell factor

7 111 0 128
6 110 0 64
5 101 0 32
4 100 1 16
3 011 0 8
2 010 0 4
1 001 1 2
0 000 0 1

SPEED OF A CA. Every CA has a maximal speed c with which signals can
propagate. This means if we take an initial conditions x which is constant
outside an interval I, then then T k (x) will still be constant outside an interval
Ik of size |Ik | ≤ |I||2c

LEMMA. The speed of a CA is bounded above by the radius R.

PROOF. Each timestep can change only cells maximally R units to the left or
to the right.
Example: The ”Takahashi-Susama Soliton automaton” is defined on
points x ∈ {0, 1}Z for which only finitely many cells are 1. The rule for T
is to start from the left and move each 1 to the next 0 position. Since a pack
of n adjecent 1′ s moves with speed n, the map T is not a cellular automaton.

EXAMPLES.
a) The cellular automaton T = σ c shifting c ∈ N entries to the right has the
speed c. Since c is also the radius, this shows that the speed can not be faster
then the radius R. The speed ratio c/R satisfies c/R ≤ 1.
b) The CA T (x)n = (..., a, a, a, a, ...) is obtained by a function φ which is
constant. Every orbit of this automaton is attracted to the fixed point. The
speed is zero. The picture to the right shows rule-100 cellular automaton.

POSSIBLE SPEEDS. Note that we can enumerate the set of cellular automata:
it is a countable set. Because the set of real numbers in the interval [0, 1] is
uncountable, we can not obtain all the speeds.

PROPOSITION. Fix A. For every 0 < a < b < 1, there is a CA with radius R
over the alphabet A for which the speed c satisfies a ≤ c/R ≤ b.

You explore this fact a bit in a homework. The idea is first to use a larger
alphabet in order to slow down the motion using internal ”color swapping”. For
different alphabets A, B, a A-automaton can be simulated by a B automaton,
possibly changing the radius.
CELLULAR AUTOMATA II Math118, O. Knill

ABSTRACT. This page contains three mathematical results: the Curtis-Hedlund-Lyndon theorem which says
that every continuous, translational invariant map on X is a CA, the proof that σ is chaotic in the sense of
Devaney and on a rather technical proof that the topological entropy which we define for CA agrees with the
classical topological entropy for general topological dynamical systems.

THE CURTIS-HEDLUND-LYNDON THEOREM.

For every continuous map T on X = AZ which commutes with σ, there is a

finite set F = {−R, ..., R} and a map ϕ such that T (x)n = ϕ(xn−R , ..., xn+R ).

PROOF.
(i) We claim that the map f from X to A defined by f (x) = T (x)0 depends only on {xi , i ∈ F (x) }, where
F (x) is some finite set.
Proof: If this were not true, there existed a sequence x(nk ) in X with nk → ∞ such that xl = x(nk )l for
l ̸= nk and xl = ̸ x(nk )l for l = nk and T (x(nk )) ̸= T (x). Because x(nk ) → x for k → ∞, the continuity of T
implies that T (x(nk )) = T (x) eventually because of the finiteness of the alphabet. This is a contradiction to
T (x(nk )) ̸= T (x) for all k.

(ii) The set F (x) is independent of x.

Proof. First of all, x → m(x), where m(x) = min(F (x)) and x → M (x), where M (x) = max(F (x)) are
continuous. This implies that xn → x implies F (xn ) = F (x) if d(xn , x) is close enough. The set F (x) is
invariant under the shifts σ by assumption. Assume, there exist two points x, y, where F (x) ̸= F (y). We can
find z and sequence of translations σ nj such that σ nj (z) → y and a sequence of translations mk such that
σ mk (z) → y. We have F (z) = F (σ nj z) and F (z) = F (σ mk z) and so F (x) = F (y).

ISOMORPHIC AUTOMATA. Some of the elemen-

tary automata are isomorphic. For example, the par-
ity transformation P (x)n = x−n , then P −1 T P is a
new elementary automaton with a different number.
Also C(x)k = (1 − xk ) which changes 0 and 1 brings
a new automata C −1 T C. Many of the 256 differ-
ent rules lead to isomorphic systems. Counting the
equivalence classes reduces the number 256 to 88.
The pictures to the right show rule 170 and rule 240,
the left and right shift.

THE ”CHAOTIC” SHIFT. The shift map σ is also CA with rule 240.

CA is chaotic in the sense of Devaney: it has a dense set of periodic points and
has a dense orbit.

PROOF. To get a dense orbit, enumerate all finite words wk and concatenate
them together to an infinite sequence y, for k > 0. Define xk = y|k| . T n (x) is
dense.
For every x, and every ϵ, there exists a N -periodic sequence y such that
d(x, y) < ϵ.

PARTICLES INTERACTIONS. Automata with nearest neighbor interaction

and larger alphabets can exhibit already quite interesting behavior. Physicists
are intrigued by the similarity to particle physics. Certain configurations travel
with some speed, interact and destroy each other like real particles. The picture
to the right shows the automaton over the alphabet Zp , p = 9 with ϕ(a, b, c) =
a ∗ b ∗ c + 1. If the CA rule is the ”physics” of the ”CA micro world”, one
calls particles elements in X which are constant outside some interval and
which satisfy T n (x) = σ m (x). They have speed v = m/n. If you are lucky, the
interaction of particles produces new particles.
SUBSHIFTS. A closed σ-invariant subset X of AZ is called a subshift. If a subshift X is invariant under a
CA map T , we can look at the system (X, T ). Examples:

a) If x = (..., 0, 1, 1, 0, 1, 1, 0, 1, 1, ...), then X = {x, σ(x), σ 2 (x)} is a subshift. More generally, the set of all
M -periodic sequences forms a subshift. Restricting a CA map T onto X means simulating the CA with
periodic boundary conditions.

b) Take all sequences with alphabet {a, b, c}, so that transitions a− > b− > c− > a and b− > b are possi-
ble. The space X with words like (, ..., abcabcabbcabbbcabcabbbbc, ...). is an example of a subshift of finite type.

c) If T is a cellular automaton map and X is a subshift, then T (X) is a subshift. It is called a factor of the
original subshift. That is how CA were first introduced by Hedlund.

d
ATTRACTOR. The image X1 = T (X0 ) of the set of all configurations X0 = AZ is a T invariant subshift. The
image∩X2 = T (X1 ) is invariant too etc. We obtain a nested sequence of subsets X0 ⊃ X1 ⊃ X2 .... The limit
X = k Xk is called the attractor of the cellular automaton. It is a T -invariant subshift.
EXAMPLES. For the shift σ, the attractor is the entire set AZ . For the rule 0-automaton, the attractor is a
single point.

TOPOLOGICAL ENTROPY OF 1D CA. The topological entropy of a 1D CA

is defined as
log(R(N, K))
h(T ) = lim lim .
K→∞ N →∞ N
where R(N, K) be the number of distinct rectangles of width K and height N
which occur in a space-time diagram of T .
The picture to the right shows a rectangle R(N, K) for an automaton, where
the attractor is a point. Here R(N, K) depends on K but stays bounded in N .
The entropy is zero.

EXAMPLE. The shift T = σ has the maximal possible entropy log(|A|). Take a random sequence x, then T n (x)
will be random sequences too. We have R(N, K) = |A|N .

TOPOLOGICAL ENTROPY IS DIFFICULT TO COMPUTE:

THEOREM (Hurd,Kari and Culik) Given ϵ > 0. There is no computer al-

gorithm which when given as an input the rule of the CA, the output is the
topological entropy up to accuracy ϵ.

The strategy of the proof is to relate the problem of calculating the entropy to the ”stopping problem of Turing
machines, which is a undecidable problem: there exists no algorithm which takes a Turing machine and decides
whether it halts or not.

BOUNDARY CONDITION. If an initial sequence x is periodic, satisfying

xi+N = xi for all i, then T (x) is periodic. We can then watch x1 , ..., xN
and know the entire sequence. In this case, the possible configurations are fi-
nite, namely |A|N , where |A| is the cardinality of the alphabet A. The cellular
automata map is a map on a finite set XN .
We can also take fixed boundary conditions, assuming that x0 = xN = 0. In
analogy to PDE’s (and CA are in a sense discrete PDE’s), one could call this
Dirichlet boundary conditions.

GROWTH OF LARGEST ATTRACTOR. For a fixed automaton we can look at the size s(N ) of the largest
attractor on the subshift X = XN set N periodic sequences. Define the growth rate
1
0 ≤ lim sup log(s(N )) ≤ log |A|
N N

This growth rate is different from the topological entropy in general: the growth rate of the shift σ is 0, while
the topological entropy is log |A|.
GENERAL DEFINITION OF TOPOLOGICAL ENTROPY. The topological entropy of a continuous map T
on a compact space X is in general defined as h(T ) = limϵ→0 limn→∞ log(M (n, ϵ))/n, where M (n, ϵ) is the
minimal number of ϵ-balls in the metric dn (x, y) = max0≤i≤n−1 d(T i x, T i y) which cover X.

The topological entropy of the CA agrees with the general topological entropy:

PROOF. Given two (N, 2K + 1)-rectangles A, B in the space-time diagram. Enumerate the rows of A and B
starting from the bottom with A1 , . . . , AN and B1 , . . . , BN and take two elements x, y ∈ X such that

Aj = (T j (x)−K , . . . T j (x)−1 , T j (x)0 , T j (x)1 , . . . , T j (x)K ) ,

Bj = (T j (y)−K , . . . T j (y)−1 , T j (y)0 , T j (y)1 , . . . , T j (y)K ) .

Because Aj = Bj if and only if d(T j (x), T j (y)) < 2−K , we know that A = B implies dN (x, y) < 2−K . On the
other hand, if x, y ∈ X satisfy dN (x, y) ≥ 2−K , we have two different rectangles. With

M (N, 4 · 2−k ) ≤ R(N, 2K + 1) ≤ M (N, 2−k /4) .

(i) Left inequality. Take for each R(N, 2K + 1) rectangles A a point x such that

A1 = (x−K , . . . x−1 , x0 , x1 , . . . , xK ) .

This gives a finite set Y ⊂ X with R(N, 2K + 1) points. Every point x ∈ X has distance ≤ 2 · 2−K to one of
the points in Y . The R(N, 2K + 1) balls of radius 4 · 2−k with midpoints in Y cover X. This proves a).

(ii) Right inequality: two different points in Y have distance ≥ 2−K /2. We need therefore at least R(N, 2K + 1)
balls of radius 2−K /4 to cover X.

The two inequalities together give R(N, 2(K + 4) + 1)) ≥ M (N, 2−(K+2) ) ≥ R(N, 2K + 1) so that

log(R(N, 2(K + 4) + 1)) log(M (N, 2−(K+2) )) log(R(N, 2K + 1))

lim ≤ lim ≤ lim .
N →∞ N N →∞ N N →∞ N
For K → ∞, the left and right limits converge to the same number. The limit in the middle is the topological
entropy.
HIGHER DIMENSIONAL CA Math118, O. Knill

ABSTRACT. We look at some higher dimensional automata like the game of life or lattice gas automata. Note
that 2 hours after this lecture, unix time is 1111111111 = Fri, 18 Mar 2005 01:58:31.

HIGHER DIMENSIONAL AUTOMATA. Everything said before can be generalized to higher dimensions. Lets
2
restrict to two dimensions. The space is X = AZ . It consists of elements xn,m , where (n, m) are the coordinates.
Define the shifts σ1 (x)n,m = xn+1,m , σ2 (x)n,m = xn,m+1 . A continuous map on X which commutes with both
σi is called a Cellular automaton. We have T (x)n = φ(xm ) with n − m in some finite set F . The composition
of two CA is a CA. A distance is defined as d(x, y) = 1/(n + 1) if xk = yk for |k| ≤ n and xl ̸= yl for some
|l| = n, where |(i, j)| = |i| + |j|.

GAME OF LIFE. One of the most famous automaton is Conways

game of life. A dead cell comes alive if and only if it has three
neighbors. A live cell dies if it has less then 2 ore more than 3 neighbors.

SPECIAL SOLUTIONS. A configuration x has compact support if

there are only finitely many cells which are alive. Examples of solutions
with compact support are gliders, stones and blinkers.

The picture to the right shows life after a random initial condition, after
having iterated for 500 iterations.

GLIDERS. Solutions which satisfy T n (x) = σ v (x) for integer n and v = (v1 , v2 ) are called gliders. Gliders
travel with velocity v/n. If x is a glider, then T n (x) converges to 0.

PERIODIC SOLUTIONS. If T n (x) = x, then x is called a periodic solution of T . The left two configurations
below show fixed points called ”stones”. We also see a periodic two orbits called ”blinker”.

THE HPP MODEL. is a simple deterministic two-dimensional cellular au-

tomata designed by Hardy,Pazzis and Pomeau in 1972. Its aim to have a
simple toy model to simulate the Navier Stokes equations. The automaton
has a color for each of the possible particle configurations. There can be max-
imally 4 particles at the same spot. One assigns a letter to each of the 16
configurations.
Particles always point away from the origin. Either there is a particle in one of
the four directions, or there is not. Once can code each color with a code like
(n, w, s, e) = (1, 1, 0, 1) The rules are designed such that particles move freely.
For example, if if xn,m = (0, 0, 0, 1) and all other nodes satisfy xi,j = (0, 0, 0, 0),
then xn+1,m = (0, 0, 0, 1). A particle has moved from node (n, m) to node
(n + 1, m). If particles collide with a right angle, they will scatter as if they
would pass through each other. If they hit head on, both directions change by
90 degrees.
HEXAGONAL LATTICE GAS CA. Designed by
Frisch,Hasslacher and Pomeau in 1985. The rules
are designed to conserve particle number and mo-
mentum at each vertex. Additionally, there is a ran-
dom number generator, when particles collide head
on. The possible directions in which the particle pair
can scatter is chosen randomly. Also this lattice gas
automaton conserves particle numbers as well as mo-
mentum of the particles.

d
ATTRACTOR. The image X1 = T (X0 ) of the set of all configurations X0 = AZ is a T invariant subset. The
image∩X2 = T (X1 ) is invariant too etc. We obtain a nested sequence of subsets X0 ⊃ X1 ⊃ X2 .... The limit
X = k Xk is called the attractor of the cellular automaton. It is a closed T -invariant subset and T (X) = X.

WHERE DO CA BELONG?

Space Time States Object

Continuous Continuous Continuous Partial differential equations
Continuous Discrete Continuous Maps on function spaces
Discrete Continuous Continuous Coupled differential equations
Discrete Discrete Continuous Coupled map lattices
Discrete Discrete Discrete Cellular automata
d d
PDE Example: dt u(x, t) = f (u(x, t), dx u(x, t)).
Maps on functions: u(x, t + 1) = f (u(x, t)).
d
Coupled differential equations: dt u(n, t) = f (u(n − 1, t), u(n, t), u(n + 1, t))
Coupled map lattices: u(n, t + 1) = f (u(n − 1, t), u(n, t), u(n + 1, t)) with u(t, x) real.
Cellular automata: u(n, t + 1) = f (u(n − 1, t), u(n, t), u(n + 1, t)) with u(t, x) finite.

HISTORY. Numerical treatments of ODE’s and PDE’s leads to CA: Example: the heat equation ut = uxx leads to a difference equation u(t + 1, x) − u(t) =

cu(t, x + 1) − 2u(t, x) + u(t, x − 1) which becomes a CA, when u(t, x) takes finitely many values only. If the PDE is translational invariant, the discretisation

is a CA with an alphabet of 1/ϵ elements, if the computing accuracty is ϵ. Difference methods for PDEs were used since a long time, at least since 1920 (L.F.

Richardson), and research on it exploded during WW2 and when the first computers appeared (i.e. the first electronic computer ENIAC in 1945). John von

Neumann seemed have introduced CA in these years. Ulam claims to have found CAs first in ”Adventures of a Mathematician” p.285: ”my own simple minded

model”. 1936 Turing machines are shown to be able to do all computations. A Turing machine with n states and a tape alphabet of k symbols is a special cellular

automaton with an alphabet of n + k letters.

1950 Idealized models of biological systems were studied using CA. Ulam and von Neuman called this ”nearest neighbor-connected cellular spaces”. Source: From

Cardinals to Chaos, Ed: Necia Grant Cooper, Cambridge University Press.

1969 Gustav Hedlund considered in the mid 50ies ”shift commuting block maps”. see ”Endomorphisms and automorphisms of the shift dynamical systems” Math.

Systems Theory 3, p.320-375, (1969). Hedlund got his PhD at Harvard in 1930.

1970 Conway article on the ”game of life” in the Scientific American 223, (October 1970): 120-123. The name CA had already been coined, like in ”Essays on

cellular automata Ed. Arthur W. Burks, 1970.

2004 MathSciNet shows 3328 papers authored on Cellular automata.

MORE ON CA Math118, O. Knill

ABSTRACT. We add some additional remarks about CA and an open problem.

AUTOMATA ON GRAPHS. Cellular automata can be defined in any dimensions and even on any homogenous
graph, where each node looks the same. A popular ”two dimensional” example different from the square lattice
is the hexagonal lattice. Setting up the CA story on more general graphs is nothing more than changing notation.

A general class of graphs, for which most of the theory goes over are Cayley graphs Γ of finitely presented
groups like G = {a, b | a2 b = ba2 }. The graph has nodes for each word in the generators a, b and two nodes
v, w are connected, if va = w or av = w or w = va or w = vb.
As a metric, one first introduces the geodesic distance in the graph Γ which is the shortest number of steps
(applying one of the generators of the group G) to get from one point to the other. Write |k| for the distance
to the origin. The distance between two configurations in X = AΓ is still defined as d(x, y) = 1/(n + 1), where
xk = yk for |k| ≤ n and xk ̸= yk for some k satisfying |k| = n.

Hedlunds theorem still applies: a continous map on X = AΓ which is invariant under translations (applying
the group G on the Cayley graph) is defined by a local law φ.

The proof we have given before applies almost word by word: the continuity of the map T forces a local law.
The translational invariance and the fact that the action of the group G on the graph is transitive, implies that
the law is the same at every node.

PROBLEMS WITH CA. The discretisation distroys rotational symmetry. In the plane, one can make CA
more symmetric by using a hexagonal lattice but still, there is no rotational symmetry. Even in the limit when
the cells become infinitesimally small, their stucture can be seen from the propagation of solutions.

SURJECTIVITY. Which automata are T are invertible maps on X and so homeomorphisms (every bijective
map on a compact space has a continuous inverse). It is also known that an injective CA is surjective. To check
injectivity, one actually can restrict to finite configurations. These results had been obtained in the 60ies. Thew
fact that injectivity implies surjectivity is called a ”Garden of Eden theorem”. The from E.F. Moore coined
expression ”Garden of Eden patterns” is a picturesque name for points in X, which are not in the image of T .

AN OPEN PROBLEM. An automaton T is called transitive, if it has a dense orbit in X. We have seen that
the shift is transitive. We also have seen that the shift has a dense set of periodic points. F. Blanchard asks:

Does every transitive automaton have a dense set of periodic points?

Francois Blanchard writes: ”The answer, positive or negative, is a necessary step before on understands the
meaning of chaos in the field. ” Source: This problem can be found in Michael Misiurewicz list of open problems
in dynamical systems (http://www.math.iupui.edu/ mmisiure/open)

THE SEMIGROUP OF CA. If you have a CA T and a CA S defined on the same space X, then T ◦ S is a new
CA. So, the set of all CA is a semigroup. Historically, this was one of the original ways how CA were introducd
because according to Hedlund, cellular automata are just the homomorphism on the category of subshifts. Note that
the semigroup of all cellular automata is not commutative.
If you look at the set of all CA which are invertible, then the set of all these cellular automata forms a group.
The identity in this group is the trivial CA, where T (x) = x.

A CLASS OF REVERSIBLE AUTOMATA. Given an alphabet A and an elementary automaton T defined by

a function φ : A3 → A we can define an automaton

T (x, y)i = (yi + φ(xi−1 , xi , xi+1 ), xi )

The map T is now invertible with the inverse T −1 (x, y)i = (yi , xi − φ(yi−1 , yi , yi+1 ). It suffices to look the first
coordinate because y(t) = x(t − 1).

This automaton on can actually be written as an automaton on X = B Z , where B is the alphabet A × A. For
example, for A = {0, 1}, the new alphabet B is {(0, 0), (1, 0), (0, 1), (1, 1)}. The translation if xk = (0, 1), then
this would correspond to (xk , yk ) = (0, 1) in the original picture.
CA AS MAPS ON SUBSHIFTS. If X is a subshift that is a shift invariant subset of AZ , and T is a CA map,
then T (X) is again a subshift. It is called a factor of X. There are some properties of subshifts which stay the
same after applying CA maps.

• Topological transitive: there exists a dense orbit.

• Almost periodic = minimal: every orbit is dense.
• Uniquely ergodic: there exists exactly one invariant measure.
• Strictly ergodic: minimal and uniquely ergodic.

• Dense set of periodic orbits: x periodic orbit: T n (x) = x.

• Prime: Every factor of (X, T ) is either trivial or isomorphic to X.
• Totally minimal: No factor is a finite permutation.

• Completly positive entropy: all non trivial factors have positive directional entropy.
• Zero directional topological entropy:
• Topologically strongly mixing: U, V open. Exists n ∈ Z such that U ∩ T n V ̸= ∅.
• Topologically weakly mixing: X × X is topologically transitive.

• Uniquely ergodic, strong mixing: µ(U ∩ T n V ) → µ(U ) · µ(V ).

• Uniquely ergodic, weakly mixing subshifts: X × X is ergodic.
• Sophic: a factor of a subshift of finite type.
• Chaotic in the sense of Devaney: topological transitive and dense set of periodic orbits .

(If one requires additionally that the shift is not periodic, then this property is not invariant. There are shifts
which have periodic factors).

Cellular automata maps can be used to generate new subshifts with given
dynamical properties!

Is this useful? It can be. If you have a complex subshift to analyze and if you can show that it is obtained by
applying CA maps from a simpler shift, then you have proven that the subshift inherits the properties of the
initial subshift.

ABOUT COMPLEXITY. The shift acting on all periodic sequences is not very spectacular. It just rotates
a sequence. Every orbit is n periodic. Other cellular automata like rule 30 have complexer behavior when
restricted to periodic sequences in the sense that there are longer periodic orbits in that space X of 2n possible
configurations. Note that T can never be transitive on X in the periodic setup because if you start with a
constant sequence x, then T (x) is a constant sequence. But orbits can get long.

The complexity of a dynamical system can depend

dramatically on the space, on which it is defined.

REMINDER: A linear map A like the cat map on R2 behaves differently then the same map
on the torus R2 /Z 2 . The map on the torus is complex. However, when restricting the map on
the set of rational points (x, y) ∈ X, the map is not complex at all: every orbit is eventually periodic.

REMINDER: The free motion of a particle in the plane is trivial. But when confined to a finite
region (a billiard table), the motion can become complex. Then again, restricting this complex
motion to some subset can be completely understandable like restrictiion to the invariant curve on
which the dynamics is just a translation.

Talking about the complexity of a map or differential equation does not make sense per se. The set X on which
one wants to understand the system is important. Complexity is often mentioned in discussions about CA. Like
other buz words, the word is loaded with many different meanings. One precise mathematical definition is the
”computational complexity of a problem” which is a measure on how the number of computations grows with
a parameter of the problem.
TURING MACHINES Math118, O. Knill

ABSTRACT. This is an excursion into a class of dynamical systems called Turing machines. They are remarkable
because any computation can be done by Turing machines. Because Turing machines can be realized as subshifts
and subshifts are abundant in dynamical systems theory, most dynamical systems like the Henon map would
be capable to do any possible computation.

TURING MACHINES. A Turing machine is a dynamical system (Y, T ) defined as follows. Define Y = X × S =
{0, 1}Z × S, where S is a finite set of states. The set S contains an element 0, which is called the halting state.
The set {(. . . , 0, . . .)} × S is called the empty tape. The set X is the space of 0, 1 sequences for which only
finitely many 1 are called data. The Turing machine is defined by three maps from finite sets to finite sets.

f : {0, 1} × S → {0, 1} defines the new letter

g : {0, 1} × S → S defines the new state
h : {0, 1} × S → {−1, 0, 1} decides whether to move the tape to left, right or stay

one can define now a continuous map on the compact metric space Y by

T (x, s) = (σ h(x0 ,s) (· · · , x−2 , x−1 , f (x0 , s), x1 , x2 · · ·), g(x0 , s)) .

This dynamical system is called a Turing machine. Note that this is not a CA, since the map does not commute
with the shift. But alread John von Neumann noticed that one can find for every Turing machine a CA, which
simulates the Turing machine. Note that the set Y is not compact but it is a subset of a compact set.

HALTING STATE The description of a Turing machine is given by a finite amount of information, because the
three involved functions map finite sets into finite sets. The set X × {0} is called the halte set. One step of a
Turing machine can be described as follows: the Turing machine with tape x and state s moves the tape h(x, s)
steps goes into the state s and then writes the entry f (x, s) at the position 0.

CHURCH THESES. Turing showed, that every computation which can be done by known computations can
be done by Turing machines. The question of what actually can be computed is probably beyond the scope of
mathematics. There is a widely accepted statement called the Church thesis (1934) which tells that everything
which can be computed can be computed with a partial recursive function. Such functions can be computed by
Turing machines. Everything we know to compute can be computed with partial recursive functions.

TURING MACHINES AS DATA. The set of pairs (T, x) where T is a Turing machine and x ∈ X is an input
data, is countable. We can encode therefore the set of such pairs into data X. Let T M ⊂ X be the set of all
the so obtained pairs (T, x). Denote by H the subset of T M , which consists of halting Turing machines.

DECIDABLE SETS IN TM. A subset Z of T M is called decidable, if there exists a Turing machine, which
tells after finitely many steps, whether a given x ∈ T M is in Z or not.

THE HALTE PROBLEM IS NOT DECIDABLE.

THEOREM (Turing) The subset H ⊂ T M of all halting Turing machines is not decidable.

PROOF. Assume the halting problem is decidable. Then there exists a Turing machine HALT which returns
from the input (T, x) the output HALT(T, x) = true, if T halts with the input x and otherwise returns
HALT(T, x) = false. Turing constructs a Turing machine DIAGONAL, which has as an input an input x and
does the following

1) Read[x]
2) Define Stop=HALT[(x,x)];
3) While Stop==True repeat Stop:=True.
4) Print[Stop]
Now, either DIAGONAL is in the set H or it is not.

(i) Assume first DIAGONAL is in H. Then the variable Stop was T rue, which means that the program
DIAGONAL runs for ever. So, Halt[(DIAGONAL,DIAGONAL)]=False, and DIAGONAL is not in H.

(ii) Assume now DIAGONAL is not in H Then, the variable Stop becomes F alse, which means that
Halt[(DIAGONAL,DIAGONAL)]=true, which implies DIAGONAL is in H.

Since the assumption of the existence of a Turing machine HALT leads to a contradiction, a machine DIAGONAL
can not exist. This argument of Turing is very similar to Cantors diagonal argument.

UNIVERSAL TURING MACHINE. Turing also showed the existence of a universal Turing machine. This is a
machine which can simulate all Turing machines. The universal Turing machine takes a Turing machine with
input (T, x) as input and returns as output, the output of the machine x. What Turing showed 1936 means
translated into the dynamical systems language:

The universal Turing machine can be realized as a dynamical system.

Indeed, there exists a compact set X and a continuous transformation T on X, such that for a subset Z of X,
(Z, T ) can do any computation in Mathematics. This tells us also that there are fundamental limitations, what
can be said about dynamical systems in general. There are dynamical systems, so that we can not decide for
a given set U and a point x, whether T n (x) will every enter U or not. Note that all said here about Turing
machines is just rephrasing of what Turing knew 70 years ago already in an other language. This has to be said
because there is literature which can give the impression that such statements are a new discovery.

BUSY BEAVER. The busy beaver problem is the task to construct a Turing machine which has n states not
counting the halting state 0 and satisfies the following: The machine starts on the empty tape and should write
as many 1 onto the tape as possible before it stops. For n = 1, 2, 3, 4, the optimal solutions are known. For
n = 5, Heiner Marxen has built a Turing machine in 1989 which produces 4098 marks. Its orbit is has length
11’798’826. You find a Mathematica program which simulates this Turing machine on the website.

REDDY’S THEOREM. A topological dynamical system is a pair (X, T ), where X is a compact metric
space and T is a homeomorphism of X. (A homeomorphism is a map which is continuous and invertible and
for which the inverse is continuous too). A topological dynamical system is called expansive, if there exists
ϵ > 0 such that for all x ̸= y ∈ X, there exists n such that d(T n x, T n y) ≥ ϵ. A dynamical system is called zero
dimensional if X is zero dimensional, that is if there is a basis for X which consists of sets which are both
open and closed. (A basis is is a set B of subsets such that (i) the empty set is in B, arbitrary unions of sets
in B are in B, the intersection of two sets in B is a union of sets in B.)

THEOREM (Reddy) A zero dimensional expansive dynamical system is isomorphic to a subshift.

PROOF (sketch) partition X into n sets Xi which are both closed and open, such that each of the sets has
diameter ≤ ϵ. An orbit T n (x) defines a code y ∈ AZ , where A labels the partition. The expansiveness assures
that the encoding is injective.

THE TURING MACHINE AS A SUBSHIFT. We first change the Turing dynamical system to make it expansive.
This can be done by a topological trick. The zero-dimensionality is assured already. The abstract theorem of
Reddy shows that

COROLLARY. There is a subshift which can simulate the universal Turing machine.

Because a subshift is a subset of the shift and the shift can be realized in a dynamical system with a horse shoe,
one obtains

COROLLARY. The map T (x, y) = (−1.5x2 − 0.3y, x) can simulate any computation.

Proof. An iterate T m of T contains a horse shoe, on which the dynamics is conjugated to a shift of 2 symbols.
The map T mk is on this set conjugated to a shift of 2k symbols.

Again, it is important to state that such corollaries are nothing more than climbing onto the shoulders of Turing
and other mathematicians working in topological dynamics. While there is nothing original in such statements,
it is amusing. It also illustrates that dynamical systems have relations with the foundations of mathematics or
what one sometimes calls the ”theory of computation”.
COMPLEX DYNAMICS Math118, O. Knill

ABSTRACT. When maps are iterated in the complex plane it leads to interesting dynamics. An example is the
Newton method in the complex. We look at some examples and especially show finally that the Ulam map is
chaotic. Actually, the interval on which the Ulam map is defined is the Julia set of the corresponding quadratic
map.

0.6

THE NEWTON METHOD IN THE REAL. The Newton method to find a

0.4
root of f (x) = 0, is to start with a point x0 and apply the map T (x) = x −
f (x)/f ′ (x). If T (x) = x, then f (x) = 0. Because T ′ (x) = f (x)f ′′ (x)/(f ′ (x))2 0.2

is small near f (x) = 0, T is a contraction in an interval [x0 − ϵ, x0 + ϵ] and 0.25 0.5 0.75 1 1.25 1.5 1.75

has a fixed point. The basin of attraction of a root xi are all the points for -0.2

which T n (x) → xi . -0.4

THE NEWTON METHOD IN THE COMPLEX. The Newton method to find

a root f (z) = 0 can also be done in the complex plane. We start with a point
z0 and apply the map T (z) = z −f (z)/f ′ (z). If T (z) = z, then f (z) = 0. Again
T ′ (z) = f (z)f ′′ (z)/(f ′ (z))2 is small near f (z) = 0, the map T is a contraction.
The basin of attraction of a root xi are all the points for which T n (x) → xi .
The picture to the right shows the basins of attractions for each fixed point.
Each of this region is the ”stable manifold” of the fixed point. The rest is called
the Julia set of T .

QUADRATIC MAP. The quadratic map

fc : z 7→ z 2 + c

with a complex parameter c defines s discrete dynamical system on the complex

plane. fc leaves a set Jc ⊂ C called Julia set and its complement Fc , called the
Fatou set invariant. The parameter space C is divided into a Mandelbrot
set M , parameters, where Jc is connected and its complement, where Jc is
disconnected.

PARAMETRIZING ALL QUADRATIC MAPS. The quadratic family fx is not as special as one might think:

LEMMA. A quadratic polynomial T (z) = az 2 +2bz +d is conjugated by S(z) =

az + b to
fc (z) = z 2 + c
where c = ad + b − b2 .

Proof. Just verify S −1 fc S(z) = T (z).

Remark. You show in the homework taht every cubic polynomial T (z) can be conjugated to fa,b (z) = z 3 −
3a2 z + b. The parametrization is chosen so that −a, a are critical points of fa,b .
When dealing with maps on the real line, we could also chose the normal form

z 7→ az(1 − z) .

Parametrized like this, the quadratic map is also called the logistic map. It maps the interval [0, 1] onto itself.
The linear map S(z) = −az + a/2 conjugates z 7→ az(1 − z) to z 7→ z 2 + c, when c = a/2 − a2 /4. Especially,
the Ulam map is conjugated to f−2 .

EXAMPLE. THE SQUARING MAP. Let us look at the map f (z) = z 2 . If

n n
z = reiθ with r = |z|, then f n (z) = r2 ei2 θ . If r > 1, then f n (z) → ∞. If
n
|r| < 1, then f n (z) → 0. If r = 1, then f n (z) = ei2 θ . On |z| = 1, the map is
T (x) = 2x mod 1.

There is a set J on which f is chaotic and the complement F where f is

attracted to some attracting fixed point.
EXAMPLE. THE ULAM MAP AS A QUADRATIC MAP. What happens with
the Ulam map f (z) = 4z(1 − z) in the complex plane? We have seen that it is
conjugated to f2 (z) = z 2 − 2. The conjugating map S(z) = 2 − 4z maps the
interval [0, 1] to the interval [−2, 2]. This interval is invariant and the map T
restricted to this interval is the Ulam map.

√
FIXED POINTS. The fixed points of the quadratic map are z± = (1 ± 1 − 4c)/2. The value of f ′ (z)
determines the stability. If |f ′ (z)| < 1, then the fixed point is stable, if |f ′ (z)| > 1, it is unstable.

Note that when a complex map is written as a real map, then it is not possible that T has a hyperbolic fixed point.

EXAMPLE. f (z) = z 2 + z + 1 has the fixed points i, −i. Since f ′ (i) = 2i and f ′ (−i) = −2i, we have |f ′ (i)| = 2
and both fixed points are unstable.

JULIA SETS. Let f be a polynomial. Let Pc denote the set of all points for which f n (z) stays bounded. This
is called the prisoner set K (or filled in Julia set). The boundary of K is called the Julia set J. The
complement of J is an open set called the Fatou set F of f . It is known that the Julia set is the closure of
all repelling periodic points. For the quadratic family, the Julia set is totally disconnected if c is outside the
Mandelbrot set and connected, if c is inside the Mandelbrot set.

CHEBYSHEV POLYNOMIALS.
1

Let f (z) = 2z 2 − 1. Because f (cos(z)) = 2 cos2 (z) − 1 = cos(2z) we have 0.5

f n (cos(z)) = cos(2n z). Actually, the map S(z) = (z + 1/z)/2 satisfies Sz 2 =

f S(z). In other words, the map S semiconjugates f to the map g(z) = z 2 -1 -0.5 0.5 1

which we have seen above. The conjugating map S maps the unit circle to the
interval [−1, 1]. This can be used to conjugate the Ulam map to a shift. One -0.5

generalize this example to the case, where Tk (z) is the Chebychev polynomial -1

cos(kz) = Tk (cos(z)). (See Homework).

THE ULAM MAP AND THE SHIFT.

The Ulam map T (x) = 4x(1 − x) is chaotic in the sense of Devaney.

Proof. The Ulam map is conjugated to the Chebychev map C(z) = 2z 2 − 1. The idea is to use the
semiconjugation of the later to f (z) = z 2 which is semiconjugated to the shift on {0, 1}N . That the later is
chaotic in the sense of Devaney had been shown last week in the CA week.

We can find C(z), by forming θ = arccos(z) and then get y = cos(2θ). if arccos(z)/π = 0.x1 x2 x3 ... in binary
expansion, then C(z) = cos(π · 0.x2 x3 x4 ...).

To find a dense set of periodic points, take a periodic sequence x ∈ {0, 1}N then z = cos(π · 0.x1 x2 ...) is a
periodic point of the Ulam map. The map x → z is continuous and surjectiv. We can find so periodic orbits
intersecting each interval [a, b]. To show transitivity, take z = cos(π0.x1 x2 ...), a sequence x ∈∈ {0, 1}N which
is transitive (concat an enumeration of all finite words onto each other)
THE MANDELBROT SET Math118, O. Knill

ABSTRACT. This is a proof a theorem of Douady and Hubbard assuring that the Mandelbrot set is connected.
The proof needs some concepts from topology and complex analysis and topology.

BÖTTCHER-FATOU LEMMA.

Assume f (z) = z k + ak+1 z k+1 + . . . with k ≥ 2 is analytic near 0. Define

n
ϕn (z) = (f n (z))1/k = z + a1 z 2 + .... In a neighborhood U of z = 0 ϕ =
limn→∞ ϕn (z) : U → Br (0) satisfies ϕ ◦ f ◦ ϕ−1 (z) = z k and ϕ(0) = 0 and
ϕ′ (0) = 1.

PROOF. We show that ϕn converges uniformly. The properties ϕ(f (z)) = ϕ(z)k as well as ϕ(0) = 0 and
ϕ′ (0) = 1 follow from the assumptions. The function

f (z)1/k
h(z) := log( )
z
with the chosen root f (z))1/k = z + O(z 2 ) is analytic in a neighborhood U of 0 and there exists a constant C
such that |h(z)| ≤ C|z| for z ∈ U . U can be chosen so small that f (U ) ⊂ U and |f (z)| ≤ |z|. We can write ϕ(z)
as an infinite product
ϕ1 (z) ϕ2 (z) ϕ3 (z)
ϕ(z) = z · · · ... .
z ϕ1 (z) ϕ2 (z)
This product converges, because n=0 log φφn+1 (z)
∑∞
n (z)
converges absolutely and uniformly for z ∈ U :

]1/kn
(f ◦ f n (z))1/k
[
ϕn+1 (z) 1 1 C · |z|
| log | = | log | = n · |h(f n (z))| ≤ n C · |f n (z)| ≤ .
ϕn (z) f n (z) k k kn

COROLLARY (*). If c 7→ fc (z) is a family of analytic maps such that c 7→ fc (z) is analytic for fixed z, and c
is in a compact subset of C, then the map (c, z) 7→ ϕc (z) is analytic in two variables.

PROOF. Use the same estimates as in the previous proof: the maps (c, z) 7→ ϕn (c, z) are analytic and the
infinite product converges absolutely and uniformly on a neighborhood U of 0.

PROPOSITION The Julia set Jc is a compact nonempty set.

PROOF.
(i) The Julia set is bounded: the Lemma of Boettcher-Fatou implies that every point z with large enough |z|
converges to ∞. This means that a whole neighborhood U of z escapes to ∞. In other words, the family
F = {fcn }n∈N is normal, because every sequence in F converges to the constant function ∞.

(ii) The Julia set is closed: this follows from the definition, because the Fatou set Fc is open.

(iii) Assume the Julia set were empty. The family F = {fcn } would be normal on C. This means that for any
sequence fn in F, there is a subsequence fnk converging to an analytic function f : C → C. Because such
a function can have only finitely many zeros and poles, it must be a rational function P/Q, where P, Q are
polynomials. If fnk → f , there are eventually the same number of zeros of fnk and f . But the number of zeros
of fnk (counted with multiplicity) grows monotonically. This contradiction makes Jc = ∅ impossible.

COROLLARY. The Julia set Jc is contained in the filled in Julia set Kc , the union of Jc and the bounded
components of the Fatou set Fc .

PROOF. Because Jc is bounded and f -invariant, every orbit starting in Jc is bounded and belongs by definition
to the filled-in Julia set. If a point is in a bounded component of Fc , its forward orbit stays bounded and it
belongs to the filled in Julia set. On the other hand, if a point is not in the Julia set or a bounded component
of Fc , then it belongs to an unbounded component of the Fatou set Fc .
GREEN FUNCTION. A continuous function G : C 7→ R is called the potential theoretical Green function of
a compact set K ⊂ C, if G is harmonic outside K, vanishing on K and has the property that G(z) − log(z) is
bounded near z = ∞.

The Green function Gc exists for the filled-in Julia set Kc of the polynomial
fc . The map (z, c) 7→ Gc (z) is continuous.

PROOF. The Boettcher-Fatou lemma assures the existence of the function ϕc conjugating fc with z 7→ z 2 in a
neighborhood Uc of ∞. Define for z ∈ Uc

Gc (z) = log |ϕc (z)| .

n
This function is harmonic in Uc and growing like log |z| because by Boettcher satisfies |fcn (z)| ≥ C|z|2 for some
constant C and so
1
Gc (z) = lim n log |fcn (z)| .
n→∞ 2

Although Gc is only defined in Uc , there is one and only one extension to all of C which is continuous and
satisfies
Gc (z) = Gc (fc (z))/2 . (1)
In fact, we define Gc (z) = 0 for z ∈ Kc , and Gc (z) = G(fcn (z))/2n otherwise, where n is large enough
so that fcn (z) ∈ U . We know from this extension that Gc is a smooth real analytic function outside
Kc . From the maximum principle, we know that Gc (z) > 0 for z ∈ C \ Kc . We have still to show that
Gc is continuous in order to see that it is the Green function. The continuity follows from the stronger statement:

(z, c) 7→ Gc (z) is jointly continuous.

2
G−1
c ([0, ϵ)) is open in C for all ϵ > 0 if and only if there exists n such that

An := {(c, z) | Gc (fcn (z)) ≥ 2n ϵ}

is closed ∀ϵ > 0. Given r > 0. There exists a ball of radius b which contains all the sets Kc for |c| ≤ r. For
R ≥ Gr (b), all the solutions ξ of Gc (ξ) ≥ R satisfy |ξ| ≥ b if |c| ≤ r. The set B = {(c, ξ) | Gc (ξ ≥ R} ∩ {|c| ≤ r}
is closed. For n large enough, also An ∩ {|c| ≤ r} is closed and An is closed.

THEOREM (DOUADY-HUBBARD). The Mandelbrot set M is connected.

CORE OF THE PROOF. The Böttcher function ϕc (z) can be extended to

Sc := {z | Gc (z) > Gc (0)} .

√
Continue defining ϕc (z) := ϕc (z 2 + c) to get ϕc having defined in larger and larger regions. This can be done
as long as the region ϕ−1
c ({r}) is connected (this assures that the derivative of ϕc is not vanishing). Because
Equation (1) gives Gc (c) = 2Gc (0) > Gc (0), every c is contained in the set Sc and the map

Φ : c 7→ Gc (c)

is well defined. It is analytic outside M and can be written as

n
Φ(z) = lim [fcn (c)]1/2 .
n→∞

Claim:
Φ:C \M →C \D
is an analytic diffeomorphism, where C = C ∪ {∞} is the Riemann
sphere. (This implies that the complement of M is simply connected in
C, which is equivalent to the fact that M is connected). The picture to
the right shows the level curves of the function ϕ6 (c) = [fc6 (c)]1/64 . The
function (ϕ6 (z) is already close to the map Φ(z) in the sense that the
level sets give a hint about the shape of the Mandelbrot set.
(1) Φ is analytic outside M . This follows from the Corollary.

(2) For cn → M , we have |Φ(cn )| → 1. Proof. Continuity of the Green function.

(3) The map Φ is proper. (A map is called proper if the inverse of any compact set is compact).
Given a compact set K ⊂ C \ D. The two compact sets D and and K have positive distance. Assume ϕ−1 (K)
is not compact. Then, there exists a sequence cn ∈ Φ−1 (K) with cn → c0 ∈ M so that |Φ(cn )| → 1. This is not
possible because Φ(cn ) ∈ K is bounded away from D.

(4) The map Φ is open (it maps open sets into open sets). This follows from the fact that Φ is analytic. (This
fact is called open mapping theorem (see Conway p. 95))

(5) The map Φ maps closed sets into closed sets.

A proper, continuous map Φ : X → Y between two locally compact metric spaces X, Y has this property.
Proof. Given a closed set A ⊂ X. Take a sequence Φ(an ) in Φ(A) which converges to b ∈ Y . Take a compact
neighborhood K of b (use local compactness of Y ). Then Φ−1 (K ∩ ϕ(A)) is compact and contains almost all an .
The sequence an contains therefore an accumulation point a ∈ X. The continuity implies Φ(an ) → Φ(a) = b
for a subsequence so that b ∈ Φ(K). Consequently Φ(K) is closed.

(6) Φ is surjective.
The image of Φ(C \ M ) is an open subset of set C \ D because Φ is open. The image of the boundary of M is
(use (5)) a closed subset of C \ D which coincides with the boundary of D because the boxed statement about
the the Green function showed Gc (c) → 0 as c → M .

(7) Φ is injective.
Because the map Φ is proper, the inverse image ϕ−1 (s) of a point s is finite. There exists therefore a curve Γ
enclosing all points of Φ−1 (s). Let ♯A denote the number of elements in A. By the argument principle (see
Alfors p. 152), we have
1 Φ′ (z)
∫
−1
♯(ϕ (s)) = dz
2πi Γ Φ(z) − s
and this number is locally constant. Given M > 0, we can find a curve Γ which works simultaneously for all
|s| ≤ M . Because Φ is surjective and ♯(ϕ−1 (∞)) = 1, we get that ♯(ϕ−1 (s)) = 1 for all z ∈ C\D and ϕ is injective.

(8) The map Φ−1 exists on C \ D and is analytic.

Because an injective, differentiable and open map has a differentiable inverse, (this is called Goursat’s theo-
rem), the inverse is analytic.

NOTATIONS.

• f (z) is analytic in a set U if the derivative f ′ (z) = limw→0 (f (z + w) − f (z))/w of f exists at every point
in U . This means that for f (z) = f (x + iy) = u(x + iy) + iv(x + iy) the partial derivatives ∂u ∂u ∂v ∂v
∂x , ∂y , ∂x , ∂y
are all continuous real-valued functions on U . In that case u(x, y), v(x, y) are harmonic: uxx + uyy = 0.
• Br (z) = {w | |z − w| < r } is a neighborhood of z called an open ball.

• A sequence of analytic maps fn converges uniformly to f on a compact set K ⊂ U , if fn → f in C(K),

which means maxx∈K |fn (x) − f (x)| → 0.
• A family of analytic maps F on U is called normal, if every sequence fn ∈ F has a subsequence which
converges uniformly on any compact subset of U . The limit function f does not need to be in F. With
respect to the topology of convergence on compact subsets normality is precompactness in this topology:
F is normal, if and only if its closure is compact. The theorem of Arzela-Ascoli (see Alfors p. 224)
states says that normality of F is equivalent to the requirement that each f is equicontinuous on every
compact set K ⊂ U and if for every z ∈ U , the set {f (z) | f ∈ F} is bounded. z is part of the Fatou
set of f , {f n }n∈N is normal in some neighborhood of z. The Julia set is the complement of the Fatou set.

• A set is called locally compact, if every point has a compact neighborhood. In the plane, a set is
compact if and only if it is bounded and closed. A subset is closed, if and only if its complement is open.
A subset U is open, if for every point x in U there is a ball Br (x) which still belongs to U .
SOME HISTORY:

In 1879, Arthur Cayley poses the problem to study the regions in the plane,
where the Newton iteration converges to some root.

Gaston Julia (1893-1978) and Pierre Fatou

(1879-1929) both worked already 90 years ago on the
iteration of analytic maps. Julia and Fatou sets are
called after them. Julia and Fatou were both com-
peted for the 1918 ’grand priz’ of the academie of
sciences and produced similar results. This produced
a priority dispute. Julia lost his nose in world war I
and had since to wear a leather strap across his face.
He had continued with his research in the hospital.
Robert Brooks and Peter Matelski produce in 1978 the first picture of
the Mandelbrot set in the context of Kleinian groups. Their paper had the
title ”The dynamics of 2-generator subgroups of PSL(2, C)”. The defined
M̃ = {c | fc has a stable periodic orbit }. This set is now called Brooks-
Matelski set and is now believed to be the interior of the Mandelbrot set M .
If the later were locally connected, this would be true: int(M ) = M̃ .
John Hubbard made better pictures of a quite different parameter space
arising from Newton’s method for cubics. Hubbard was inspired by a question
from a calculus student. Benoit Mandelbrot, perhaps inspired by Hubbard,
made corresponding pictures in 1980 for quadratic polynomials. He conjectured
the set M is disconnected because his computer pictures showed ”dust” with
no connections to the main body of M . It is amusing that the journals editorial
staff removed that dust, assuming it was a problem of the printer.
John Milnor writes in his book of 1991: ”Although Mandelbrot’s statements
in this first paper were not completely right, he deserves a great deal of credit for
being the first to point out the extremely complicated geometry associated with the
parameter space for quadratic maps. His major achievement has been to demonstrate
to a very wide audience that such complicated fractal objects play an important role
in a number of mathematical sciences.”
Adrien Douady and John Hubbard prove the
connectivity of M in 1982. This was a mathematical
breakthrough. In that paper the name ”Mandelbrot
set” was introduced. The paper provided a firm foun-
dation for its mathematical study. We followed on
this handout their proof. Note that the Mandelbrot
set is also simply connected, but this is easier to
show. Both statments use that a subset of the plane
is connected if and only if the complement is simply
connected.
Evenso one of the first things which comes in mind, when talking about fractals
is the Mandelbrot set. It is not a ”fractal”: in 1998, Mitsuhiro Shishikura
has shown that its Hausdorff dimension of M is 2. (M. Shishikura, ”The
Hausdorff dimension of the boundary of the Mandelbrot set and Julia sets,
Annals of Mathematics 147 (1998), 225-267.)
Also for higher dimensional polynomials, one can define Julia and Mandelbrot
sets. For cubic polynomials fa,b (z) = z 3 − 3a2 z + b, define the cubic locus
set {(a, b) ∈ C 2 | Ka,b is connected }, where Ka,b is the prisoner set Ka,b =
n
{z | fa,b (z) stays bounded. }. Bodil Branner showed around 1985, that the
cubic locus set is connected. This generalizes the main result discussed in this
handout.

OPEN PROBLEMS. The major open problem is whether the Mandelbrot set is locally connected or not. A
subset M of the plane is called locally connected, if at every point x ∈ M if every neighborhood of x contains
a neighborhood, in which M is connected. A locally connected set does not need to be connected (two disjoint
disks in the plane are locally connected but not connected). A connected set does not need to be locally
connected. An example is the union of the graph of sin(1/x) and the y-axes.
NOTIONS IN COMPLEX DYNAMICS Math118, O. Knill
ABSTRACT. This page summarizes some definitions in complex dynamics and gives a brief jumpstart to some
notions in complex analysis and topology.

MANDELBROT SET. fc (z) = z 2 + c is called the quadratic map. It is

parametrized by a constant c. The set M of parameter values c for which fcn (c)
stays bounded. In the homework you seea that M = {c, |fcn | ≤ 2 for all n }.
With G(c) = limn→∞ log |(fcn (c))1/2 | one can also say M = {c | G(c) = 0 }.
n

The level curves of G are equipotential curves: if you would charge the
Mandelbrot set with a positive charge, G(z) = c is the set of points where
the attractive force of an electron to the set is the same. By definition, M
is closed. Douady-Hubbard theorem tells it is connected. That M is simply
connected is much easier to see: it follows from the maximum principle
that the complement of M is connected.

JULIA SET. The set of complex numbers z for which fcn (z) stays bounded is
called the filled in Julia set Kc . It is the set of z for which the function
Gc (z) = limn→∞ log |(fcn (z))1/2 | is zero. Its boundary is called the Julia set.
n

The Julia set can be a smooth curve like in the case c = 0 or for c = −2 but
it is in general a complicated fractal. It is known that the Julia set Jc is the
closure of the repelling periodic points of fc . It is also known that fc restricted
to Jc is chaotic in the sense of Devaney. The complement of Jc is called the
Fatou set Fc . The bounded components of Fc are called Fatou components.

COMPLEX MAPS. A complex map f can be written as a map in the real

plane f (x + iy) = u(x, y) + iv(x, y). The derivative at a point z0 is defined as
the complex number

a = f ′ (z) = lim (f (z + w) − f (z))/w .

w→0

If the derivative exists at each point in a region U and f ′ is a continuous

function in U , the map f is called analytic in U .

CAUCHY-RIEMANN. Since the linearization of f at z0 is the map z → az

which is a rotation dilation and the linearization of f is the Jacobean
[ ]
ux uy
A= ,
vx vy

we must have ux = vy , uy = −vx (A rotation matrix has identical diagonals and

antidiagonals of opposite signs and this property is preserved after multiplying the
matrix with a constant) . These two equations for u, v are called Cauchy-
Riemann differential equations.

CONFORMALITY. If a ̸= 0, then angles are preserved because both rotations and dilations preserve angles.
Therefore the rotation dilation z → az preserves angles. If f ′ (z) is never zero in a region U , the map f is called
conformal in U . In that case, it maps U bijectively to f (V ) and preserves angles. Angle preservation is useful
in cartography or computer graphics.

HARMONICITY. From the Cauchy-Riemann equations follows uxx + uyy = 0

and vxx + vyy = 0. Therefore, the real and imaginary
∫ part of f are har-
monic functions. The mean value property |w−z|=r u(w(t)) dt = u(z)
∫
and |w−z|=r v(w(t)) dt = v(z) for harmonic functions can be written as
∫
|w−z|=r
f (w(t)) dt = f (z).
TAYLOR FORMULA. Because df (w(t))/dt = f (x + r cos(t) + i(y +
r sin(t)) = f ′ (w)(r cos(t) + ir sin(t)) = f ′ (w)(z − w), this can be rewritten
as |w−z|=r f ′ (w(t))dt/(z − w) = f (z). This is the Cauchy integral formula.
∫

Since we can differentiate the left hand side arbitrarily often with respect to
z, this proves that an analytic function is arbitrarily often differentiable and
f (w)n!
f (w)/(z − w) has the n’th z-derivative (z−w) n+1 , we get

∑ f (n) (z)(w − z)n

f (w) =
n
n!

which is the familiar Taylor formula if f, z, w are real.

CAUCHY THEOREM. The Cauchy Riemann equations also prove the Cauchy
formula. If C is a closed curve in simply connected region U in which f is
analytic, then

f (z)dz = f (z(t))z ′ (t) dt = 0

∫ ∫
C

because the later is the line integral of F (x, y) = (−v(x, y), u(x, y)) and Greens
theorem in multi-variable calculus shows that curl(F ) = curl((−v, u)) = (ux −
vy ) = 0. In other words, the vector-field F (x, y) = (−v(x + iy), u(x + iy)) is
conservative.

FIXED POINTS. Because the eigenvalues of the rotation dilation A come in

complex conjugate pairs, the fixed points or periodic points can not be hyper-
bolic. Fixed points are either stable sinks, or unstable sources elliptic, con-
2
jugated
√ to a rotation. For example, the fixed points of f (z) = z√ + c are
(1 ± 1 − 4c)/2 and the linearization at those points is df (z) = (1 ± 1 − 4c)z

TOPOLOGY. Here are some topological notions occuring in complex dynamics:

OPEN. A set U in the plane is called open if for every point z, there exists r > 0 such that Br (z) = {w | |w−z| <
r} is contained in U . One assumes the empty set to be open. The entire plane is open too.
CLOSED. A set U in the plane is closed, if the complement of U is open. The entries plane is closed.
INTERIOR. The interior of a set U is the subset of all points z in U for which there exists r > 0 such that
Br (z) ⊂ U . If a set is open, then it is equal to its interior.
CLOSURE. The closure of a set U is the set of all points which are limit points of sequences in U . It is the
complement of the interior of the complement of U . If a set is closed, then U is equal to its closure.
BOUNDARY. The boundary of a set U is the closure of U minus the interior of U . The boundary of a closed
set without interior is the set itself.
SIMPLY CONNECTED. A set A is simply connected, if every closed curve contained in A can be deformed
to a point within A. A simply connected subset of the plane has no ”holes”.
CONNECTED. A set A is called connected if one can not find two discjoint open sets U, V such that A∩U ̸= ∅,
A ∩ V ̸= ∅.

A set A is connected if and only if the complement is simply connected.

To verify that the complement of M is simply connected, one finds a smooth bijection of the complement of the
unit disc with the complement of M . The bijection is given by Φ(c) = limn→∞ (fcn (c))1/2 . The Mandelbrot
n

set M is connected as well as simply connected. The Julia sets Jc are connected, if c is in M .
COMPACT. A subset of the complex plane is called compact if it is closed and bounded. A sequence in a
compact set always has accumulation points. The Mandelbrot set as well as the Julia sets are examples of
compact sets.
PERFECT SETS. A subset J in the complex plane is perfect if it is closed and every point z in J is accumulation
point of points in S \ z. Perfect sets contain no isolated points.
NOWHERE DENSE. A subset J in the complex plane is nowhere dense if the interior of its closure is empty.
A Julia set Jc is nowhere dense if c is outside the Mandelbrot set.
CANTOR SET. A perfect nowhere dense set is also called a Cantor set. An example is the Cantor middle
set. A Julia set Jc is a Cantor set if c is outside the Mandelbrot set.
THE BERNOULLI SHIFT Math118, O. Knill

ABSTRACT. When equipped with an invariant measure, which is the area measure when representing it as the
Baker map, the shift is called the Bernoulli shift. It produces independent random variables.

A SHIFT INVARIANT MEASURE. We have defined a map S from the unit square Y = [0, 1) × [0, 1) to the
sequence space X = {0, 1}Z by {
0 un < 1/2
S(u, v)n =
1 un ≥ 1/2
if T n (u, v) = (un , vn ) is the orbit of the Baker map. This was called symbolic dynamics. We can use the map
S to measure subsets in X by requiring that it preserves the measure: the left half of the square of area 1/2 is
mapped into the set of sequences x which satisfy x0 = 0, the right half of the square of area 1/2 is mapped into
{x | x0 = 1}. The set {x0 = 0, x1 = 1} in X corresponds to the lower left quarter of the square which has area
1/4.

THE BERNOULLI MEASURE. The space X can be equipped with a shift invariant probability measure P .
In that case, we say P [U ] is the measure or the probability of U . We can define P [U ] as the area of S −1 (U ) in
the square. We know then that
P [xn+1 = f1 , ..., xn+m = fm ] = 2−m .
This measure is called a Bernoulli measure. It is invariant under the shift. for any subset U of X, then
P [σ(U )] = P [U ].

If U is a subset of the square and S is the map conjugating the Baker map to
the shift, then P [S(U )] is the area of U .

RANDOM VARIABLE. A random variable is a (continuous) function from X to R. Examples of random

variables are Xk (x) = xk . Two random variables are called independent if P[{Y = a, Z = b}] = P[{Y =
a}]P[{Z = b}] for any choice a, b.

The random variables Xk = xk in the Bernoulli shift are independent.

PROOF. P[Xk = a, Xl = b] = 1/4 for any choice of a, b. This is the same as P[Y = a]P[Z = b] = (1/2) · (1/2).

In other words, one can use the Bernoulli shift or the Baker map to produce random numbers.
This is not a very practical way to produce random numbers: lets look at the first coordinates, when applying the Baker
map, we have T n (x) = 2n x mod 1. If we start with a a rational number, then T n (x) will be attracted by a periodic orbit
like for example 1/3, 2/3, 1/3, .... For a practical generation of random numbers other maps are better suited.

EXPECTATION. The expectation of a random variable which takes finitely many values f1 , ..., fm is

E[Y ] = P [Y = f1 ]f1 + ... + P [Y = fn ]fn

Two random variables Y, Z are called uncorrelated if E[Y Z] = E[Y ]E[Z]. Two independent random variables
are automatically uncorrelated.

EXAMPLE. A = {head, tail } models throwing a coin The random variable

[
3 x0 = head
X(x) =
5 x1 = tail

has the expectation

E[X] = P [X = 3]3 + P [X = 5]5 = 3/2 + 5/2 = 4 .

EXAMPLE. Consider the shift over the alphabet A = {1, 2, 3, ..., 6 }. The ran-
dom variables X1 , X2 , ... simulate the outcomes of a dice event. If X3 = 5, then
the third dice rolling produced a 5. These random variables are uncorrelated
and independent.

THE LAW OF LARGE NUMBERS. The law of large numbers tells that if Xk are independent random variables
with the same distribution, then
n
1∑
Xk
n
k=1

converges to the common expectation E[Xk ] for almost all experiments.

1
∑n
EXAMPLES. In the dice case, we have for almost all sequences x, that n k=1 Xk → 7/2.

OTHER MEASURES. The set X can be equipped with other measures. Assume the letter xk = 1 should have
probability
( ) p and xk = 0 should have probability 1 − p. In that case, the probability P[x1 = a1 , ...., xn = an ]
n
is pk (1 − p)n−k , where k is the number of times, ai = 1. Knowing the probability of all these events
k
defines the invariant measure. All theses measures are called Bernoulli measures.

MARKOV CHAINS. Often, one does not know the invariant measure, but one knows the conditional prob-
abilities: P [xn+1 = a | xn = b] = Mab . In words, the probability that xn+1 = a under the condition xn = b
is Pab . The matrix Mab is called a Markov matrix. It has the property that the sum of coefficients in each
column is equal to 1. The matrix M is a n×n matrix, if the alphabet A has n elements. You have seen examples
of the following fact in linear algebra:

The eigenvector p = (p1 , ..., pn ) to the eigenvalue q of the matrix M normalized

so that the p1 + ... + pn = 1 defines a Bernoulli probability measure on X.

EXAMPLES.
a) If Mij =[1/2 for all i,]j, we have the Bernoulli shift.
1/2 2/3
b) If M = , we can read off the probabilities p that xn = 1 and 1 − p that xn = 0 by computing
1/2 1/3
the eigenvector
[ v of M
] to the eigenvalue 1 and normalizing it, so that the sum of its entries is 1.
1/3 1
c) If M = , we obtain a measure supported on the Fibonnacci shift introduced above. The transitions
2/3 0
11 is not possible.

MEASURES ON SUBSHIFTS OF FINITE TYPE. If we use a Markov matrix for which Mab = 0 if ab is a
forbidden word, then we obtain an invariant measure for the subhift of finite type by inducitively determining the
probability of the cylinder sets P [{x0 = a0 , ..., xn = an }] using the Bayes formula P [A|B] = P [A ∩ B]/P [B].
Subshifts of finite type have a lot of invariant measures. Markov matrices provide a possibility to define such
measures.

MEASURES ON SUBSHIFTS. Every subshift X has an invariant measure. It can be obtained by averaging
along an orbit. This averaging
∑n does not converge in general, but there is a subsequence, along which the limit
sets P [A] = limn→∞ n1 k=1 1T k (x)∈A

UNIQUELY ERGODIC SUBSHIFTS. If there is only one shift-invariant measure then the subshift is called
uniquely ergodic. An example are Sturmean sequences, which are obtained by doing symbolic dynamics on
using a half open I and an irrational rotation on the circle. The There is only one invariant measure, because
also the irrational rotation on the circle has only one invariant measure.

ERGODIC THEORY. The part of dynamical systems, which deals with invariant measures of a map or dynam-
ical system is called ergodic theory. It has close relations to probability theory. The law of large numbers we
mentioned here has a generalization which is called Birkhoffs ergodic theorem.
SHIFTS IN QUADRATIC AND STANDARD MAP Math118, O. Knill

ABSTRACT. We look on this page at an analytic proof that there is an invariant shift embedded in some Hénon
maps, Standard maps or quadratic maps. The proof uses the implicit function theorem and is based on an
idea of Aubry and Abramovici called anti-integrable limit.

THEOREM OF DEVANEY-NITECKI. Fix b 6= 0. For large enough c, the

Hénon map H : (x, y) 7→ (x2 − c − by, x) has an invariant set K such that T
restricted to K is conjugated to the shift

S = (..., x−1 , x0 , x1 , x2 , ...) → (..., x0 , x1 , x2 , x3 , ...)

on all sequences with two symbols.

√
PROOF. With the new parameter a = 1/ c and the new coordinates q = x · a, p = y · a, the map becomes

q2 − 1
T (q, p) 7→ ( − bp, q)
a
and is equivalent to the recurrence
a · qn+1 + a · b · qn−1 = qn2 − 1 .
We look for sequences qn = q(S n x), where S is the shift on the space of all sequence X = {−1, 1}Z and where
q is a continuous map from X to R. We have to solve

a · q(Sx) + a · b · q(S −1 x) − (q(x)2 − 1) = 0 .

With the map F : R × C(X) → C(X) defined by

F (a, q)(x) = a · q(Sx) + a · b · q(S −1 x) − (q(x)2 − 1)

this equation can be rewritten as F (a, q) = 0. The partial derivative Fq (a, q) is

Fq (a, q)u = a(u(S) + b · u(S −1 )) − 2q · u .

The map F (0, q) : C(X) → C(X) has the property that every function q ∈ C(X) with values in {−1, 1} is a
solution of F (0, q) = 0. We take for such a solution the map q(x) = x0 .

The derivative Fq (0, q) is the linear map

(Fq (0, q)u = −2q · u
which is invertible because q is bounded away from 0.

By the implicit function theorem, there exists a solution a 7→ qa = G(a) satisfying F (a, qa ) = 0 for small a.
Define ϕa : X → R2 by
ϕa (x) = (q(x), q(S −1 x)) .
The map ϕa is continuous, because q and T are continuous.
Using F (a, q) = 0, we check that

(q(x)2 − 1)
ϕa ◦ T (x) = (q(Sx), q(x))) = ( − b · q(S −1 x), q(x))
a
= T (q(x), q(S −1 x)) = T ◦ ϕa (x)

for all x ∈ X.

The map is injective because if two points x, y are mapped into the same point in R2 then the fact that qa (x)
is near q0 (x) = x0 implies x0 = y0 . The conjugation ϕa ◦ S n (x) = T n ◦ ϕa (x) gives us T n (x) = T n (y) and so
xn = yn for all n.

ϕ has a continuous inverse because every bijective map from a compact space to a compact space has a continuous
inverse. The map is indeed a homeomorphism from X to a closed subset K = ϕ(X) ⊂ R2 .
THE IMPLICIT FUNCTION THEOREM. Given a family q → F (a, q) of
maps, parametrized by a parameter a. If F (0, q0 ) = 0 and F ′ (0, q0 ) 6= 0, then
there exists a continuous function q in some interval I such that F (a, q(a)) = 0
for a ∈ I.
0.6

PROOF. The Newton map Ta (q) = q − F (a, q)/F ′ (a, q) has as a stable fixed 0.4

point which is the root q(a). This fixed point exists for small a and changes 0.2

continuously with a.
This proof works also in infinite dimensional spaces, in which it is possible 0.2 0.4 0.6 0.8 1 1.2 1.4

to differentiate. An example is the space C(X) of continuous functions on -0.2

a compact set X. Example: let F (f ) = f 3 + 5f . The function F maps a -0.4

continuous function to a continuous unction. One has F ′ (f )g = (3f 2 + 5)g.

Example: let F (f ) = f (x2 ). Because this is a a linear map in f , we have
F ′ (f )g(x) = f (x2 )g(x).

HORSE SHOES IN THE STANDARD MAP. For large enough c, the Standard
map T : (x, y) 7→ (2x + c sin(x) − y, x) has an invariant set K such that T
restricted to K is conjugated to the shift

S = (..., x−1 , x0 , x1 , x2 , ...) → (..., x0 , x1 , x2 , x3 , ...)

on all sequences with two symbols.

PROOF. If T n (q, p) = (qn , pn ) is an orbit of the Standard map, then pn = qn−1 and so qn+1 − 2qn + qn−1 +
c sin(xn ) = 0. With ϵ = 1/c, this means

ϵ(qn+1 − 2qn + qn−1 ) + sin(xn ) = 0

Let X be all {0, 1} sequences. Consider the space of all continuous functions q from X to [0, 2π].
If we find a solution q to the equation

F (ϵ, q) = ϵ(q(σx) − 2q(x) + q(σ −1 x)) + sin(q(x)) = 0

then q is a conjugation from (X, σ) to (q(X), T ) showing that we can find a shift similar as the horse shoe
construction does.

(i) There is a solution for ϵ = 0: Just take q(x) = πx0 . Because sin(0) = sin(π) = 0, the equation sin(q(x)) = 0
is satisfied.

(ii) In order to have a solution for small ϵ, we compute the derivative of L = Fq (0, q) = cos(q) and see whether
it is invertible. Indeed, since L = cos(q(x)) = ±1, we can invert L, the inverse is actually equal to L. (Note
that F has as an argument a function q and the the derivative Fq (a, q) = limu→q (F (a, q + u) − F (a, q))/u is defined with
respect to the function q. It was computed in the same way as derivatives with respect to real parameters.)
(iii) The implicit function theorem now assures that we can find for small ϵ a function qϵ which satisfies
F (ϵ, qϵ ) = 0. This function qϵ conjugates the shift with the standard map Tc restricted to the set K = qϵ (X).
Since ϵ = 1/c, this conjugation works for large enough c.

JULIA SETS. The same construction works also for the map f (z) = a(z 2 − 1).
We look for a function q ∈ C(X, C) such that q(σ) − a(z 2 + 1) = 0. With
ϵ = 1/a, this is
F (ϵ, q) = ϵq(σ) − (z 2 − 1) = 0 .
For ϵ = 0, the function q(x) = (2x0 − 1) is a solution. The derivative L =
Fq (0, q) = 2q is invertible. We have solutions for small ϵ, which corresponds to
large a. Actually, the image q(X) is just the Julia set of f .

SUMMARY. The anti-integrable limit construction allows to get embedded shifts in a purely analytic way
using the implicit function theorem. In comparison, the construction of a horse shoe is a geometric
construction. Finding a generating partition is a more combinatorial task. The shift brings different areas
of mathematics together.
SYMBOLIC DYNAMICS Math118, O. Knill

ABSTRACT. We have seen shifts as cellular automata, in a horse-shoes or in Julia set. We look at this
dynamical system a bit closer.

THE SHIFT. Given a finite alphabet A, define X = AN and σ(x)n = xn+1 . This dynamical system is called
the one sided shift. The shift on AZ is called the two sided shift. While the later is invertible, the first is
not.

SUBSHIFTS. The shift restricted to a closed shift-invariant subset X of AZ is called a subshift.

EXAMPLE. Let T (x) = x + α mod 1 and Y = [0, a) and { interval. Look at all sequences obtained by taking a
1 x∈Y
point x and defining xn = 1Y (x + nα), where 1Y (x) = . That is
0 x∈ /Y
{
1 (x0 + nα) mod 1 ∈ Y
xn = .
0 (x0 + nα) mod 1 ∈ /Y
√
Lets assume for example, Y = [0, 1/2) and α = 2. With the starting point x = 0, we obtain the sequence
{x0 , x1 , x2 , ...} = {1, 1, 0, 1, 0, 1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 1, 0, 1, 1, 0, 1, ...}. The image X of the map S is a closed
subset of the sequences. Every orbit of the shift σ in X is dense.
√
A particular interesting case is α = ( 5 − 1)/2 and F = [0, α). If x0 = 1,
then x1 = 0. If x0 = 0, then x1 = 1, x2 = 0. One can obtain the sequence
also by applying the substitution rule 1 → 0, 0 → 10. A sequence ob-
tained like this is called a Fibonacci sequence. Here is part of the sequence:
..., 0, 1, 0, 1, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 1, ...

EXAMPLE: SUBSHIFTS OF FINITE TYPE. Given a finite set of words K over an alphabet A. The set X of
all sequences, in which the words of K do not appear, is called a subshift of finite type.
The language of X is the set of all words which occur in sequences of X. There is a finite set of words which
can build up any sequence x ∈ X and such that the forbidden words determine which words can be adjacent
and which not. We can define a directed graph (V, E), which has as vertices these words and where an arrow
goes from one word to an other if these words can be glued together. One says that the graph represents the
subshift.

EXAMPLE. Assume K = {00, 111} are the forbidden words, then a sequence can be
...010110101101011010101011011.... We can get any sequence by gluing together words w1 = 01, w2 = 11 and
w3 = 10. The combinations w1 → w1 , w1 → w2 , w2 → w1 , w3 → w1 , w3 → w2 , w3 → w3 are possible.

EXAMPLE. Let K = {11}, then X consists of all se-

quences, where no double 11 occur. The language of X is
{0, 1, 00, 10, 01, 10, 000, 001, 010, 100, 101, 0000, 0001, ...}. With the set
V = {00, 01, 10} of words one can build any sequence. The gluing
00 → 01, 01 → 01, 00 → 00. 10 → 10, 10 → 01 are possible, while the
gluing 01 → 10 is not possible. 0 1

SOFIC SHIFTS. If X is a subshift of finite type and T is a cellular automaton map, then T (X) is called a
sofic shift.

Sophic shifts produce regular languages, languages accepted by finite state automata, but they are in general
no more of finite type. The next example shows this.

EXAMPLE. The even shift is the set of all x ∈ {0, 1}Z so that between any two 1, there is an even number of
0’s. The even shift is not a subshift of finite type, but it is a sofic shift. Start with the subshift of finite type,
with the forbidden word 00. Take the elementary CA which gives only 0 for 1, 0, 1 and 0, 1, 0 and 0, 1, 1. For
example, x = ...01111010111101101110111111011111111111... is mapped to y = ...0000111001001100.... The
image of this cellular automaton consists of all sequences for which 0 occurs only in blocks containing an even
number.
IRREDUCIBLE SHIFTS. A subshift is called irreducible if the language B(X) has the property if v, w are
words in B(X), then there is a word u in B(X) such that vuw is also in B(X).

PROPOSITION. A subshift (X, T ) is irreducible if and only if it is transitive.

PROOF. Assume the subshift is irreducible. We show that for every n, there is an orbit which comes 1/n
close to any point in X. To do so, make a list of all words w0 , ..., wL of length 2n + 1 which appear in X. By
assumption we can fill in words v1 , ..., vL such that w0 v1 w1 v1 ....vL wL is part of a sequence x ∈ X. Now, T n (x)
comes 1/n close to any point in X.
On the other hand, if (X, T ) is transitive, there is a point x such that T n (x) is dense. Given two words u, w which
in the language of X, there exists n such that (T n (x)1 ....T n (x)k ) = u and m such that (T m (x)1 ....T m (x)l ) = w.
The word v between u and w in the sequence x is the one we need to prove irreducibility.

OVERVIEW. class of all subshifts ⊃ class of all sofic shifts ⊃ class of all shifts of finite type
CA leave the class of sofic shifts invariant because the composition of two CA is again a cellular automaton.

MINIMAL SHIFTS. A subshift (X, σ) is called minimal, if every orbit is dense. Note that minimal shifts can
not have periodic points unless it is periodic itself.

EXAMPLE: STURMEAN SEQUENCES. Sturmean sequences xn = 1A (t + nα), where α is irrational and A

is an interval on the circle are minimal because the irrational rotation on the circle is minimal and the symbolic
map S is continuous and invertible. Because every orbit T n (x) of the irrational rotation is dense, also the
corresponding orbit S(T n (x)) is dense.

EXAMPLE: The full shift as well as subshifts of finite type are not minimal. They have many periodic orbits.

SYMBOLIC DYNAMICS. The basic construction of symbolic dynamics for a

given dynamical system (Y, T ) is to find a partition of the set X into subsets
A0 , ..., An−1 . Every point x is then assigned a sequence where xk = a if T k (x) ∈
Aa . This generating partition defines a map S from Y to X = {0, ..., n−1}N
if T is not invertible, or to X = {1, ..., n}Z if T is invertible. The map S
conjugates (Y, T ) to the subshift (S(Y ), σ). The map S is continuous, but it is
in general neither injective nor surjective. In the homework, you deal with a a
partition in case of the cat map. It is called Markov partition.

EXAMPLE. Let T (y) = y + α a rotation on the circle Y = R/Z. With A0 = [0, 1/2), A1 = [1/2, 1), the sequence
x = S(y) is called a Sturmean sequence. The map S is a continuous map from the circle to the sequence
space. But the image is not the entire space. For example, it does not contain any periodic sequences.

THE BAKER TRANSFORMATION. The baker transformation is a map on the square Y = [0, 1) × [0, 1):
The map preserves area and is invertible
{
(2u, v/2) , 0 ≤ u < 1/2 {
(u/2, 2v) , 0 ≤ v < 1/2
T (u, v) = T −1
(u, v) =
(2u − 1, (v + 1)/2) , 1/2 ≤ u ≤ 1 ((u + 1)/2, 2v − 1) , 1/2 ≤ v ≤ 1

The inverse is obtained by switching u and v, applying T and switching u

and v again. Now take the generating partition A0 = {u ∈ [0, 1/2)}, A1 =
{u ∈ [1/2, 1]}. The symbolic dynamics of a point (u, v) defines a sequence
x ∈ {0, 1}Z .

THEOREM. The map S is an invertible map from ...x−2 x−1 , x0 x1 x2 x3 ... (u, v)
the square Y to X = S(Y ) and σ ◦ S(u, v) = ...0000, 10000.. (1/2, 0)
S ◦T (u, v). For any given sequence x in the image ...0001, 00000.. (0, 1/2)
S(Y ) , we can get back (u, v) = S −1 x with ...0000, 01110.. (7/16, 0)
EXAMPLES: ...0000, 11100.. (7/8, 0)
∞
∑ −1
∑ ...0001, 11000.. (3/4, 1/2)
u= xk 2−k−1 , v = x k 2k . ...0011, 10000.. (1/2, 3/4)
k=0 k=−∞ ...0111, 00000.. (0, 7/8)
...1110, 00000.. (0, 7/16)

Remark: While S is injective, it is not surjective. (The point (....0000000, 0111111...) is not reached, but
represented by (....0000000, 1000000...) ∼ (1/2, 0) While the map S −1 is continuous, T and S are both not.
 I: The binary expansion of u is u = 0.x0 x1 x2 ....
∞
∑
u= xk 2−k−1 .
i=0

x0 = 0 means that u ∈
x0 = 1 means that u ∈
[0, 1/2). Note that u =
[1/2, 1).
1/2 gives x0 = 1.

x0 = 0, x1 = 0 means u ∈ x0 = 0, x1 = 1 means u ∈
[0, 1/2) and 2u ∈ [0, 1/2) [0, 1/2) and 2u ∈ [1/2, 1)
which is equivalent to u ∈ which is equivalent to u ∈
[0, 1/4). [1/4, 1/2).

x0 = 1, x1 = 0 means u ∈ x0 = 1, x1 = 1 means u ∈
[0, 1/2) and 2u ∈ [0, 1/2) [0, 1/2) and 2u ∈ [1/2, 1)
which is equivalent to u ∈ which is equivalent to u ∈
[1/2, 3/4). [3/4, 1).
In general, fixing x0 , ..., xn−1 determines in which of the 2n intervals [(k − 1)/2n , k/2n ) the coordinate u is.

II: The binary expansion of v is v = 0.x−1 x−2 x−3 ....

−1
∑
v= xk 2k .
k=−∞

x−1 = 0 means that v ∈

[0, 1/2). T maps the left
half of the square to the x−1 = 1 means that v ∈
lower half of the square so [1/2, 1).
that T −1 maps the lower
half of the square to the
left half.

x−1 = 0, x−2 = 0 means x−1 = 0, x−2 = 1 means

v ∈ [0, 1/2) and 2v ∈ v ∈ [0, 1/2) and 2v ∈
[0, 1/2) which is equiva- [1/2, 1) which is equiva-
lent to v ∈ [0, 1/4). lent to v ∈ [1/4, 1/2).

x−1 = 1, x−2 = 0 means x−1 = 1, x−2 = 1 means

v ∈ [0, 1/2) and 2v ∈ v ∈ [0, 1/2) and 2v ∈
[0, 1/2) which is equiva- [1/2, 1) which is equiva-
lent to v ∈ [1/2, 3/4). lent to v ∈ [3/4, 1).
Fixing x−1 , ..., x−n determines in which of the 2n intervals [(k − 1)/2n , k/2n ) the coordinate v is.
 III: Combination of part I and Part II

x−1 = 0, x0 = 0 means x−1 = 0, x0 = 1 means

u ∈ [0, 1/2) and v ∈ u ∈ [0, 1/2) and v ∈
[0, 1/2). [1/2, 1).

x−1 = 1, x0 = 0 means x−1 = 1, x0 = 1 means

u ∈ [1/2, 1) and v ∈ u ∈ [1/2, 1) and v ∈
[0, 1/2). [1/2, 1).

Fixing x−m , ..x0 , ..xn determines in which of the 2n+m+1 rectangles [(k − 1)/2n , k/2n ) × [(l − 1)/2m−1 , l/2m−1 )
the coordinate (u, v) is.

IV: Symmetry

We know u = 0.x0 x1 x2 x3 x4 .... Because replacing T and T −1 corresponds to switching u with v and replacing
the partition A0 , A − 1 with B0 = {v < 1/2}, B1 = {v ≥ 1/2}, the itinery y with respect to the new partition
gives v = 0.y0 y1 y2 y3 y4 ... Because T (A0 ) = B0 , we have v = 0.x−1 x−2 x−3 ....

BAKER MAP. In the baker map, the second rectangle is translated straight onto the first rectangle.

FAT HORSE SHOE MAP. The symbolic dynamics of the horse shoe is similar except that the second rectangle
is turn around by 180 degrees. In the horse shoe, the stretching is stronger. There was a set K which never
leaves the rectangle (the horse shoe is kind of a ”Julia set”).

TWO REMARKS. The baker map can also be conjugated to the right∑shift σxn = xn−1 . If we ∑ntake the same
0
generating partition A0 , A1 , then the formulas for S −1 become u = k=−∞ xn 2−n−1 , v = k=1 xn 2−n . In
many treatments of symbolic dynamics of the Baker transformations, one neglects things of area zero. In that
case, it does not matter, what boundary we take for the generating partition. If we want the symbolic dynamics
to work for every point in the square Y , then we remove the right and upper boundaries in all rectangles which
appear as we have done that here.
APPROXIMATION OF NUMBERS Math118, O. Knill

ABSTRACT. The approximation of real numbers by rational numbers is a special and solvable case of solving
the logarithm problem in dynamical systems.

DIRICHLET THEOREM. Let x ∈ [0, 1] be a real number in and

n > 1 be an integer. There exist integers p and 1 ≤ q ≤ n
1 ≤ p ≤ n such that
p 1
|x − | ≤ .
q qn

PROOF. The Pigeonhole principle shows that at least one of the n intervals [k/n, (k + 1)/n] in [0, 1] contains
two elements of the set {0, {x}, {2x}, ...{nx}}, where {kx} is the fractional part of kx. So |(kx − lx) + p| ≤ 1/n
for some integer p and q = k − l < n. Division through q gives |x + p/q| ≤ 1/(nq).

APPROXIMATION. For any irrational x, there are infinitely many p/q such that |x − p/q| ≤ 1/q 2 .

PROOF. If x is rational, q = 0 is possible and the result is not true. If x is irrational, then k − l = 0 is not
possible and q > 1. Now |x + p/q| ≤ 1/(nq) ≤ 1/q 2 .

CONTINUED FRACTION EXPANSION. We have seen the same result using continued fraction expansion
pn /qn because
|x + pn /qn | ≤ 1/(qn−1 qn ) ≤ 1/qn2
There is a huge difference between this result and the above result

The pigeonhole principle is not constructive. It does not tell you what p/q is. The
continued fraction expansion is constructive. You can determine p/q efficiently.
The Dirichlet method needs n computations to determine the approximation, the
continued fraction method essentially log(n).

THE LOG PROBLEM IN DYNAMICAL SYSTEMS.

Given a point x and a set I. At which time does the orbit of x enter I. For differential
equations, we want to solve T t (x) = y up to some error, for maps, we want to solve
T n (x) = y up to some error.

EXAMPLES.
• If T (x) = x + α mod1 and x = 0 is a real and y = 0. Determining T q (t) = 0 is the problem to find n such
that |qα − p| = y for some integer p. In other words, we want to find close solutions of |α − p/q| = 0. The
continued fraction expansion gives such values.
• The differential equation ẋ = ax has the solution T t (x) = et x(0). To solve T t (x) = at = y, we have
t = loga (y). Computation of the real logarithm is a special case of the dynamical logarithm problem.
• Given an prime number p and an integer a, we have a map T (x) = ax mod p on the set X = {1, ...p − 1}.
For given x and y, to compute n such that T n (x) = y is called the discrete logarithm problem in
number theory. Logarithms are called indices in number theory. For a composite n = pq, if you could
solve ak = 1 mod n we could find p. For example 54 = 1 mod 15 so that gcd(4 + 1, 15) = 5 is a factor.
The discrete log problem is harder then factoring.
• If T t (x) is the evolution of the weather and x is the current meteorological condition and y is a severe
storm, determining t such that T t (x) is close to y is a an example of a dynamical logarithm problem.
• If T t (x) is the position of an asteroid relatively to the earth and y = 0, then T t (x) = y determines the
time it takes until the asteroid has an impact. It is an example of a dynamical logarithm problem.
• If T is the cellular automaton realization of a Turing machine, x is the initial condition with the empty
tape and y is the ”halt” state, then T n (x) = y determines how long it takes until the Turing machine
halts. It is an example of a dynamical logarithm problem.
HURWITZ THEOREM. For any irrational x, there are in-
finitely many p/q such that |x − pq | ≤ √5q
1
2
.

PROOF (Borel) One of the consecutive continued fraction

convergent pn−1 /qn−1 , pn /qn , pn+1 /qn+1 satisfies this bound.
This is not so difficult to prove but could be part of a project.

This result can not be improved. The golden ratio satisfies this
bound. There is an interesting story attached. If one takes away
the bad example (the golden ratio) and all numbers which can
be obtained by applying a modular transformation T (x) = (ax +
b)/(cx +
√ d) with integers a, b, c,
√d satisfying ad − bc = 1, then the
bound 5 can be improved to √ 8 which is the best possible bound
attained by the silver ratio 2 + 1.

SOLVING THE LOG PROBLEM FOR IRRATIONAL ROTATION. The following theorem solves the dynami-
cal log problem for irrational rotations on the circle. Given two points on the circle, we can construct integers
qn such hat T n (x) = x + qn α is close to y.

TCHEBYCHEV THEOREM. Assume x is irrational with peri-

odic approximation pn /qn . Assume y is real. For every n, there
exists k ≤ qn such that {y + kx} < 3/qn .

PROOF. Because |x − pn /qn | ≤ 1/(qn qn−1 ), we can write x = pn /qn + δ/(qn2 ) with |δ| < 1, where pn /qn are the
periodic approximations of α.

Choose an integer t with |qn x−t| ≤ 1/2 so that y = t/qn +δ ′ /(2qn ), |δ ′ | ≤ 1. Find k, l satisfying qn /2 ≤ k ≤ 3qn /2
with pn k − qn l = t. Then |xk − l − y| = pn k/qn + δk/(qn2 ) − l − t/qn − δ ′ /(2qn )| = |kδ/qn2 − δ ′ /(2qn )| <
k/(qn qn ) + 1/(2qn ). Because k < 3qn /2, the right hand side is ≤ 3/qn .

ECLIPSES AND PERIODIC APPROXIMATION. A synodic

month is defined as the period of time between two new moons.
It is α = 29.530588853 days. The draconic month is the pe-
riod of time of the moon to return to the same node. It is
β = 27.212220817 days. Intersections between the path of the
moon and the sun are called ascending and descending nodes.
Such an intersection is called a solar eclipse. This appears in a
period of a bit more then 18 years = 6580 days which is called one
Saros cycle). This cycle and others are obtained from the contin-
ued fraction expansion of α/β. It is said that Thales using the
Saros cycle to predict the solar eclipse of 585 B.C. The next big
eclipse will happen May 26, 2021. Source: http://www.websters-
online-dictionary.org/definition/english/mo/month.html
The Eclipse cycles can be explained using the continued fraction expansion (see homework).

cycle eclipse synodic draconic

fortnight 14.77 0.5 0.543
month 29.53 1 1.085
semester 177.18 6 6.511
lunar year 354.37 12 13.022
octon 1387.94 47 51.004
tritos 3986.63 135 146.501
saros 6585.32 223 241.999
Metonic cycle 6939.69 235 255.021
inex 10571.95 358 388.500
exeligmos 19755.96 669 725.996
Hipparchos 126007.02 4267 4630.531
Babylonian 161177.95 5458 5922.999

See http://www.phys.uu.nl/ vgent/calendar/eclipsecycles.htm for more details.

LATTICE POINT PROBLEMS Math118, O. Knill

ABSTRACT. Finding lattice points close to curves leads to problems in dynamical systems theory.

CURVES AND DYNAMICAL SYSTEMS. A curve

r(t) = (t, f (t)) in the plane defines a sequence of points
xn = f (n) mod1 = f (n) − [f (n)] on the circle T = R/Z and so a
dynamical system T : X → X, where X is the closure of all the
translates of sequences x = xn and T is the shift.

More generally, with the vectors ⃗xn = (xn , xn−1 , ..., xn−d ), we can
define a map T (⃗x) = (xn+1 , xn , ..., xn−d+1 ) on the d-dimensional
torus T d = Rd /Z d . (For curves in space, there is a map on a higher
dimensional torus, for two dimensional surfaces, time becomes two
dimensional).

EXAMPLE STURMIAN SEQUENCES.

If r(t) = (t, αt) is a line in the 2
plane with slope α, then xn =
αn mod1 and ⃗x = (..., nα, ...) 5
is called a Sturmian sequence.
The map T is a rotation on the 4
0
circle. It is a prototype of what
one calls an integrable system,
systems in which one can for ex-
1
ample solve the dynamical loga- 3
rithm problem. 6

EXAMPLE: PARABOLIC SEQUENCES.

For the parabola r(t) = (t, γ + αt + βt2 ) we obtain the se-

quence xn = γ + nα + n2 β mod1. it leads to a measure [ ]pre-
x
serving transformation on the two dimensional torus T ( )=
y
[ ][ ] [ ]
1 0 x 2α
+ = A⃗x + ⃗b
1 1 y 0

POLYNOMIALS If p(x) is a polynomial of degree n, define pn (x) = p(x), pn−1 (x) = pn (x + 1) − pn (x), pn−2 =
pn−1 (x + 1) − pn−1 (x), .., p0 (x) = α. Each pi is a polynomial of degree i. If T (x1 , x2 , ..., xn ) =
(x1 + α, x + 2 + x1 , ..., xn + xn−1 ), then T (p1 (n), p2 (n), ..., pd (n)) = (p1 (n + 1), ..., pd (n + 1)) =
(p1 (n) + α, p2 (n) + p1 (n), ..., pd (n) + pd−1 (n).

QUADATIC CASE: p2 (x) = γ+βx+αx2 , p1 (x) = p2 (x+1)−p2 (x) = α+β+2αx, p0 (x) = p1 (x+1)−p1 (x) = 2α.
We have a map T (x, y) = (x + 2α, x + y).

WEAK CHAOS IN PARABOLIC SEQUENCES.

[ ] [ ]
x x + 2α
The map T = has zero Lyapunov exponent
y x+y
1 n
n lim(log ||dT ||). There is no sensitive dependence on initial con-
ditions. If α is irrational, then the map has only one invariant
measure, the area. The map is also minimal: every orbit is dense.
It is not chaotic in the sense of Devaney. It does not have even one
single periodic orbit. The map T is an example of a system ex-
hibiting a ”weak type of chaos”. There is no hyperbolicity present
like in the cat map. Still, a single orbit covers the torus densely.
THE INTEGRABLE FACTOR IN PARABOLIC SEQUENCES.

If we look at the lines y = const, then these lines are tossed around
in a regular way by the dynamics.

SOME DECAY OF CORRELATIONS.

The system also has mild chaotic behavior. A curve y =

const experiences a shear. Lets take a random variable
f (x, y) = f (x) which is independent of y. The random
n
variables f, f (T ), ...f (T ), ... show some decay of correlations
n 2
∫
T2
(f (T (x, y))f (x, y) − f (x, y) ) dxdy → 0 as time progresses.

WHY CONSTRUCT LATTICE POINTS CLOSE TO CURVES?

0) The problem is relevant in cryptology.
1) Estimating points close to curves is a problem in the matric
theory of Diophantine approximation. √
2) Finding points close enough to algebraic curves like z = p(x)
lead to actual rational points on the manifold solving Diophan-
tine equations.
3) Estimating lattice points in regions is a problem in the geom-
etry of numbers, a field founded by Hermann Minkoswki.
4) It relates to recurrence problems for classes of dynamical sys-
tems. It is a source for new type of dynamical systems.

CRYPTOLOGICAL APPLICATION: FACTORING INTEGERS.

Given an integer n = pq which is the product of two prime factors p, q, we want to find numbers √ y such that
y 2 = O(nα ) mod(n), with α as small as possible. One way to do that is to look at numbers [ nx]2 mod n.
More generally, one can look at integer points (x, y) close to the curve y 2 = np(x). As closer we are to the
curve, as smaller y 2 − np(x) = a is. Any algorithm which would find a of the order O(nα ) would with α < 1/2
improve the speed of the current factorization algorithms.

FACTORING ALGORITHMS. Some of the best factoring algorithms for a composite number n = pq are based
on an idea of Fermat: find x such x2 mod n is a small square y 2 , then x2 − y 2 is a multiple of n and gcd(x − y, n)
likely a factor of n. Example of algorithms are the Morrison Brillard algorithm, the quadratic √ sieve or
the number field sieve. These methods allow to construct x for which y is of the order n. A method to
construct numbers x with x2 mod n of the order n1/2−ϵ for some ϵ > 0 would improve factorization methods.

EXAMPLE: PELL’S EQUATION. √

With p(x, y) = x2 , the curve y 2 −nx2 = 1 is a hyperbola with asymptotes y = ± nx. The equation
√ y 2 = 1+nx2
is called Pells equation or Brouncker equation.√ Integer points close to the line y = nx can be found
using the continued fraction algorithm: if n ∼ yj /xj , then yj2 − nx2j = a and yj2 = a mod n. Because
√ √ √ √ √
n = y/x + C/x2 we have x n − y = C/x and x2 n − y 2 = (x n + y)C/x = C n + Cy/x ∼ 2C n. Here
θ = 1/2.

EXAMPLE
√ PARABOLA. pn (x, y) = 2n√+ x. The curve y 2 = npn (x) is a parabola. The tangent at (x, y) =
(0, 2n + 1) to the curve has slope n/ 8n2 + 1. The Diophantine error is O(1/x). The nonlinearity error
2

y ′′ (0)x2 ∼ x2 n2 /y 3 . We have y = O(n). In order that 1/x = n2 x2 /y 3 , we must have x = n1/3 . The error is
then y/x = n2/3 so that α = 2/3. If we could get rid of the quadratic or cubic erros, α would get smaller.
POINTS CLOSE TO A CURVE. The following result is a
contribution to the geometry of numbers.

THEOREM. For every 0 ≤ δ < 1/3 and every three times differ-
entiable curve of finite length, there exists a positive constant C
depending only on the curve, such that the number M (n, δ) of
1/n-lattice points in a 1/n1+δ neighborhood of the curve satisfies
M (n, δ)/n1−δ → C for n → ∞.

Remarks: if the curve is not a line, the constant C is positive. The constant can change under rotations of the
curve, but does not change under translation of the curve.

OUTLINE OF THE PROOF.

Cut the curve to have line seg-

Cut the curve so that each piece is Approximate the curve by a poly-
ments or curves with nonzero cur-
a graph gon with Diophantine slopes
vature
Remark. The polygon pieces have to be large enough so that continued fraction algorithm finds lattice points.
On the other hand, the pieces have to be small enough to get a small nonlinearity error. A compromise is possible
for δ < 1/3. This bound 1/3 is a limitation of the method. Results in the metric theory of Diophantine
approximation indicate that δ < 1/2 should be possible. Numerical experiments suggest that one can go even
higher. An approximation by polynomials of higher degree could also put the bound higher. But then the proof
no more be constructive.

After cutting the curve into pieces, we can reformulate the

12.5
theorem as follows:
10
THEOREM (Same result reduced to graph) Given a curve which
is the graph of a smooth function f (t) such that f ′′ (t) ≥ ϵ > 0 7.5

on [0, 1]. If M (n, δ) is the number of 1/n-lattice points between

f (t) − 1/n1+δ and f (t) + 1/n1+δ . Then there is a constant C such
5

that 2.5
M (n, δ)
→C.
n1−δ -4 -2 2 4

12.5
PROOF part (0). Let [a, b] = f ′ [0, 1] be the interval of possible
slopes f ′ (t) of f on [0, 1]. Choose and fix a number δ < θ < 1/3 10

and call ϵ = 1/3 − θ. Let K be the maximum of f ′′ (x) on the

interval [0, 1]. For every n, divide the interval [a, b] into r(n, θ) = 7.5

[n1−θ ] intervals Ik , called small intervals. The number of 1/n-

5
intervals in each of these intervals Ik is [nθ ]. Call Mk (n, δ) the
number of 1/n lattice points in the parallelepiped Jk above the 2.5

interval Ik between f (t) − 1/n1+δ and f (t) + 1/n1+δ .

-4 -2 2 4
 4

RROOF part (i) (Nonlinear error) On one of the small inter-

vals, the discrepancy of the curve to a tangent line is bounded 2
above by K/n2−2θ < K/n1+θ+ϵ . This uses Taylors formula
f (x + s) ∈ [f (x) + f ′ (x)s − Ks2 , f (x) + f ′ (x)s + Ks2 ] . It follows
that if Mk,x (n, δ) denotes the number of lattice points in a n−(1+δ) -4 -2 2 4

neighborhood Jk of a line segment at x above the interval Ik , then

(Mk,x (n, δ) − Mk (n, δ))/n1−δ → 0. -2

-4

PROOF part (ii) (Sufficiently many strongly Diophantine slopes).

Let h(n, δ) denote the number of intervals Ik , in which we can
find xk for which the slope f ′ (xk ) = [a0 ; a1 , a2 , ...] satisfies ai ≤
√
r(n, δ). Then h(n, δ)/r(n, δ) → 1 for n → ∞.
Reformulation: the set of all numbers y = [u, v, a1 , a2 , ....] with
u, v ≤ M is 1/M 2 dense on a set YM ⊂ [0, 1] with |YM | → 1.
A new reformulation: the set {f (u, v, x) = 1/u + 1/(v + x) =
(v + x)/(u(v + x) + 1) | u, v ≤ M } for x ∈ [0, 1] is 1/M 2 dense
on a set YM which has asymptotically full√measure 1. This is a
multivariable calculus problem: for u, v ≥ M , the distance from
one point to the next is of the order 1/M 2 because fv (u, v, x) =
1/(1 + u(v + x))2 .

PROOF part iii) (Reformulation for a line segment). Each of the

h(n, θ) parallelograms Jk above Ik has slope αk , thickness n−1−δ
and contains [nθ ] lattice units. In a scale, where the lattice size is
1, we have the following problem: Jk

Estimate the number of lattice points in a parallelogram

Jk of length [nθ ] and thickness n−δ for which the contin-
ued fraction expansion of the slope αk = f ′ (xk ) = αk =
[a1 , a2 , ...], with ai < nδ .
x
k Ik
The answer is that that there are nϵ lattice points.

PROOF part iv) (Number of lattice points in a Diophantine parallelogram Jk ). There exists ck (n), dk (n)
such that the line segment Jk contains at least [ck (n)nϵ ] lattice points and maximally [ck (n)nϵ ] lattice points.
Furthermore, ck (n) → 1 and dk (n) → 1 uniformly in k. There is a more general result of Schmidt and which even
gives the error term.

PROOF part v) (Putting things together) The total number Mk,xk (n, δ) of lattice points is between c(n)h(n, δ)nϵ
and d(n)h(n, δ)nϵ . Because of ii), we know it is between c(n)r(n, δ)nϵ = c(n)n1−δ and d(n)r(n, δ)nϵ = c(n)n1−δ .
Dividing by n1−δ and using c(n), d(n) → 1, we get the result.

AN OPEN PROBLEM. There is an efficient method to solve the dynamical logarithm problem for the map
T (x) = x + α: the continued fraction expansion gave an efficient method to find lattice points close to a line.

Is there an efficient way to solve the dynamical logarithm problem for

[ ] [ ]
x x + 2α
T = .
y x+y

on the torus. A concrete problem: for α = π, find n such that T n (0.5, 0.5) is
within distance 10−1000 of (0, 0).

Geometrically, we look for an efficient method to find lattice points close to the parabola y = αx2 + βx + γ with irrational
α. Of course, we could just list all numbers [αn2 + βn + γ] and see which one is close, this is not practical. While we can
find in a few thousand computation steps an integer n such that [αn] is smaller than 10−1000 (it is a [P] problem) more
then 10100 computations seem needed in the parabolic case (is it a [NP] problem?). Note that the big bang happend
about 1017 seconds ago.
STRICT ERGODICITY (* not treated in class) Math118, O. Knill

ABSTRACT. The irrational rotation on the circle is a minimal uniquely ergodic system. Other systems occuring
in number theory have the same property.

inx 2
∑ ∑
ERGODICITY. A map T is ergodic if for every function f (x) = n∈Z cn e with finite n cn , the condition
f (T ) = f implies f = const.

THEOREM. For irrational α, the map T (x) = x + α is ergodic.

PROOF. Comparing Fourier coefficients of f (T ) and f gives einα cn = cn so that cn = 0 unless n = 0.

UNIQUE ERGODICITY. A continuous transformation T on a compact topological space is called uniquely

ergodic if there is only one invariant measure µ of T .

KRONECKER-WEYL THEOREM. The only measure which is invariant under

an irrational rotation is the length measure dx.

∫PROOF. A measure µ is a linear map from the space of all continuous functions ∑n C(X) to R given by µ(f ) =
f (x) dµ(x). If µ is T invariant, then µ(f (T )) = µ(f ) and by linearity µ( n1 k=1 f (T k )) = µ(f ). Because for
f (x) = eikx , we have
n n
1∑ k 1 ∑ ij(x+kα) eijx (1 − eijnα )
f (T ) = e = →0
n n n (1 − eiα )
k=1 k=1
∑n
we have µ(f ) = µ( n1 k=1 f (T k )) → c0 for n → ∞ which implies µ(f ) = c0 =
∑ ikx
∫also for any f = ke
f (x) dx.

MINIMALITY. A map T is called minimal, if every orbit of T is dense.

THEOREM. The irrational rotation on the circle is minimal.

PROOF. This follows in a constructive way from Chebychevs theorem. For every x and y and ϵ > 0, there
exists n such that |x + nα − y| < ϵ.

STRICT ERGODICITY. A map is called strictly ergodic, if it is both minimal and uniquely ergodic.

COROLLARY. The irrational rotation on the circle is strictly ergodic.

HIGHER DIMENSIONAL GENERALIZATION. The above statements go through word by word for a rotation
T (x) = x + α with vectors α = (α1 , ..., αd ) for which n · α = n1 α1 + ... + nd αd ) = 0 implies
∑∞ n = 0. We call such
vectors irrational. Functions of several variables have a Fourier expansion too: f (x) = n=−∞ cn ein·x , where
n = (n1 , ..., nd ) runs over all lattice points in Z d .

COROLLARY. The irrational translation on the torus T d = Rd /Z d is strictly

ergodic.

PROOF. We have shown both minimality as well as unique ergodicity.

THEOREM (FURSTENBERG) If α is irrational and bij ∈ Z, 1 ≤ j < i ≤ d real with bi,i−1 6= 0. Then
d
T (x1 , . . . , xd ) = (x1 + α, x2 + b21 x1 , ..., xd + bd1 x1 + ... + bd,d−1 xd−1
 ) defines a uniquely ergodic system on T .
1 0 . 0
 b21 1 . 0 
 
It can be written as ⃗x 7→ A⃗x + e1 α, where A =   b31 b32 . . .

 . . . . 
bd1 . . 1
PROOF. Fourier theory shows that T is ergodic: f (T ) = n cn ein·T (x) with n · T (x) = (n1 , ..., nd ) · (x1 + α, x2 +
∑
b21 x1 , ..., xd + bd1 x1 + ... + bd,d−1 xd−1 ) = n1 α + An · x. Comparing Fourier coefficients gives cAn = cn e2πin1 α
which implies n = (n1 , 0, ..., 0) and therefore cn = cn e2πin1 α which implies that cn = 0 unless n = 0.

Unique ergodicity is shown with induction to d. We know it for d = 1, where the system is an irrational
rotation. To prove the result in dimension d, write T (⃗x, xd ) = (S(⃗x), xd + A · ⃗x). Note that S does not depend
on xd . By induction, S is uniquely ergodic on T d−1 . Given invariant measure µ for T , the projection of µ on
Td−1 is S-invariant. and by induction assumption the volume measure dx1 ...dxd−1 .

Because T commutes with R(⃗x, y) = (⃗x, y + β), a T invariant measure must also be Rβ invariant for every
β. By Birkhoffs ergodic theorem, we know that µ almost all points x = (⃗x, xd ) are generic in the sens that
∑n−1
µ(f ) = limn→∞ n1 k=0 f (T k x). Assume x = (⃗x, y) is generic. Then also (⃗x, y + β) is generic.

A uniquely ergodic system on the torus which preserves the volume measure dx1 ...dxn is automatically minimal:
if there were an orbit x which were not dense, then its closure Y would be a T invariant set which is not the
entire torus. This set would carry an other invariant measure.

ILLUSTRATION. Lets see this in the case T (x, y) → (x, x + y) → (x + α, x + y). When projecting onto the
first coordinate, we have the uniquely ergodic map x → x + α. The key is that the map T commutes with
R(x, y) = (x, y + β):

T (R(x, y)) = T (x, y + β) = (x + α, x + y + β), R(T (x, y)) = R(x + α, x + y) = (x + α, x + y + β) .

If (xn , yn ) is an orbit, then the distribution of xn on the first coordinate is the measure dx. Assume two different
points (x, y), (x, y + β) with irrational β produce measures µ(x, y), µ(x, y + β) which must coincide.

APPLICATION: Let p(x) be polynomial of degree n. Define pn (x) = p(x), pn−1 = pn (x + 1) − pn (x), pn−2 =
pn−1 (x + 1) − pn−1 (x), .., p0 (x) = α. Each pi is a polynomial of degree i. With
   
x1 x1 + α
 x2   x2 + x1 
 ...  = 
Tp =    
... 
xn xn + xn−1

we have Tp (p1 (n), p2 (n), ..., pd (n)) = (p1 (n + 1), ..., pd (n + 1)).

CORROLLARY. If p = an xn + ... + a1 x + a0 is a polynomial of degree n and

assume an is irrational, then Tp is a uniquely ergodic transformation on the n
dimensional torus which preserves the volume µ = dx1 ...dxn .

QUESTION. Are polynomials the only functions f for which one can describe
f (n) mod 1 by a finite dimensional system?

EXAMPLES. √ √
1) For f (x) = x. What dynamical system does n mod 1 generate?
2) Does f (x) = exp(x) generate an infinite dimensional system?
3) If f (x) is a k-periodic function, then f (n) is periodic too For f (x) = sin(2πxα) with irrational α, then f (n)
is an almost periodic sequence. The system on the torus (x, y) → (x + α), sin(2πx)) allows to read of f (n) in
one coordinate.
4) For rational functions like f (x) = x/(1 + x2 ), the system has a fixed point which attracts all points.

OTHER STRICTLY ERGODIC SYSTEMS. Any factor of a strictly ergodic system is strictly ergodic. This
applies to symbolic dynamics.

Doing symbolic dynamics with a strictly ergodic system produces strictly ergodic subshifts. Let A1 , A2 , .., An
be a partition of T d into subsets, define S(x)n = k if T n (x) ∈ Ak . This defines a subshift which is strictly
ergodic.

EXAMPLES. Sturmean sequences xn = 1A (x + nα) and especially the Fibonacci sequence are uniquely ergodic
subshifts. Applying cellular automata maps on such subshifts generates new subshifts which are strictly ergodic.
CA maps preserve both minimality as well as unique ergodicity.
THE SQUARE ROOT OF 1 Math118, O. Knill

ABSTRACT. The problem of Factoring large integers n has many connections with dynamical systems theory.
Knowing to take square root of 1 modulo n is an example, which would allow factorization.

FACTORING PROBLEM. Some approaches to factor large integers rely on finding solutions to polynomials like
p(x) modn. An example is the Fermat method which takes p(x) = x2 − a2 mod(n). Kraichick introduced in
1926 the idea of factor bases: Since roots of p are hard to be found directly, one can find close roots p(xk ) = ak
and hope to factor p(xk ) − ak . For example, if x2 is a small square a2 modn, then x2 − a2 can be factored.
In general, sieving methods with respect to a factor base are needed. A specific problem is to find the roots
of p(x) = x2 − 1 mod(n) which is the problem to find the square root of 1. An analytic √ problem is to find q
√
for which qn + 1 is close to an integer or more generally to find q for which f (q) = an2 + qn + 1 is close
to an integer. For a 6= 0, the later polynomial has an appealing √ property that increasing q is essentially a
quasiperiodic motion on the circle because f ′ (q) → α = 1/(2 2) for n → ∞. If it were truly quasiperiodic,
then good solutions could be found with the continued fraction algorithm: if α − p/q = O(1/q 2 ) then qα is
O(1/q) close to an integer. The geometric interpretation is to find close integer points to the graph of the
function f (x, y) = y 2 − (an2 − nx). In the case of the Pell equation f (x, y) = y 2 − (an2 − knx2 , hyperbola have
asymptotes and linear CF method works, usually with a = 0.

THE SQUARE ROOT OF 1.

LEMMA. Given two primes p, q, there are exactly four square roots of 1 modulo
n = pq. If p = a−r, q = a+r are known, then (a−r)(a+r) = a2 −r2 = 0mod(n)
which means that (a/r) = 1 mod(n) is the nontrivial square root.

PROOF. The multiplicative group Zn is the product of two cyclic groups Zp−1 × Zq−1 by the isomorphism
x = a mod(p), x = b mod(q), x 7→ (a, b) if n = pq with prime p, q, every element k in that group can be written
as uk v l , where up−1 = 1 and v q−1 = 1. Now x = u(p−1)/2 and y = v (q−1)/2 are nontrivial square roots. From
x2 − y 2 = (x − y)(x + y) and x 6= y, we see that x = n − y.

Finding a nontrivial square root x of 1 modulo n = pq gives a factor gcd(x − 1, n) of n because x2 = 1 modn.
The factor is nontrivial because otherwise x − 1 would have no common divisor with n and x + 1 would be
divisible by n. Fermat’s little theorem assures that aλ(n)/2 is a square root, where λ(n) = lcm((p − 1), (q − 1)).
If p, q are odd primes, the square roots of 1 are ±1, ±aλ(n)/2 .
If x is a nontrivial square root, we can solve x2 = qn + 1 for q with q < n or x2 = 2n2 + qn + 1 for q and q < 2n.
In other words, the function g(x, q) = 2n2 + qn + 1 − x2 = 0 has nontrivial integer solutions (x, q).
An other possibility is Pell’s equation
√ x2 − nq 2 = 1 leads to a solution in time depending on √ the length of
continued fraction expansion of n gives. This can be bad: William 1981 gives a heuristic bound√ n log log(n)
which is not effectiv. Other possibilities are to hit for solutions of x2 = (an/b)y 2 or x/y = an/b. If we are ϵ
close, then bx2 − any 2 = bϵ and bx2 = bϵ (modn) so that bx2 = bϵ (modn).

DISTRIBUTION. How is x/n distributed, if x = x(n) is the

2
nontrivial square root of n? The first graph shows the distribution
in the case n = p(k) p(k + 1), where p(k) is the k’th prime.
The distribution is not uniform. Note that for prime twins 1.5

p = a − 1, q = a + 1 we have p ∗ q = a2 − 1 so that a = q − 1
is the square root and a/n = (q − 1)/(pq) is close to 1/p. (This 1

is a special case that for two odd primes, a = (p + q)/2 is the

squareroot of b, because p = a − b, q = a + b with b = (q − p)/2 0.5

gives a2 − b2 = n.) Taking prime numbers close to each other

makes it easy to find a factor. 25 50 75 100 125 150 175

The devil stair case for primes close to each other is easy to explain. For every√
prime twin pk , pk+1 we know the
squareroot pk + 1 and (pk + 1)/n = 1/pk+1 + 1/n. So, we have many values of 1/n near 0 and 1. If p − q = 2r,
then p = a − r, q = a + r, n = a2 − r2 so that a2 = r2 mod(n). and a/r is the square root mod n.

√
A QUASIPERIODIC MOTION. The reason for the representation If x(q) = 2n2 + qn + 1 of the square root
is that for q = O(n), the root is x = O(n) and x′ (q) = n/(2x) = O(1). This means that x(q) 7→ x(q + 1)mod(1)
is to first order an irrational rotation x 7→ x + α modn.
CONTINUED FRACTION EXPANSION. Let α be a real number and let a0 = [α] the integer part of α.
Define recursively αn+1 = 1/(αn − an ), an+1 = [αn+1 ]. The sequence [a0 , a1 , · · · , an ...] is called the continued
fraction expansion of α. The rational number [a1 , . . . , an ] = pn /qn = a0 + 1/(a1 + 1/a2 + ...) is called the
n’th convergent. Example: π = [3, 7, 15, ...] means 3 + 1/(7 + 1/15).

FINDING SMALL SQUARES. If we had a linear map x(q) 7→ x(q + 1), then, to find the a square root of 1,
we have to find q such that β + qα − p is close to an integer. This is a well known problem: if p/q = pl /ql are
continued fraction approximation of α satisfying α = p/q + δ/q 2 and t is an integer such that |qx − t| < 1/2,
then there are integers x, y with px − qy = t and x ∈ [q/2, 3q/2] and (see Hua p. 266)

xp xδ t δ′ xδ δ′ x 1
|αx − y − β| = | + 2 −y− − |=| 2 − |< 2 +
q q q 2q q 2q q 2q

We need an integer root x approximation error of 1/n √to get an x with x2 = O(1). √ Because
x′′ (q) = −n2 /(4x
√
3
) = O(1/n), there is an error O(1)√for q = n. If q = O(n 1/4
, then x ′′ 2
q = O(1/ n) and
xx′′ (q 2 ) = O( n). This leads to squares of the order n.

MORRISON-BRILLHARD. Finding small squareroots is central to factorisation algorithms based on Fermat’s

method (find x, y with x2 = y 2 + qn). Examples are √ the quadratic sieve or Morrison-Brillhard. The continued
2
fraction method used by Morrison-Brillhard takes an = pl /ql + ϵ with √ ϵ = O(1/q l ). This gives an = p2l /ql2 +
2 2 2 2
ϵ + 2pl /ql ϵ so that anql = pl + 2pl ql ϵ and pl is a square of order O( n).

NEWTON TYPE ITERATION. We could try to make jumps of the form x 7→ x + qk α(x), where qk /pk is a
continued fraction approximation of the irrational number α(x) = n/(2x). For small qk , the linear approximation
is good and we can expect to get a number x in a distance 1/qk to an integer [x]. This gives us a square [x]2 of
the order n/qk . It would be desirable if a nonlinear iteration could be found which takes care of the quadratic,
cubic etc parts.

√
A DISCRETE LOG PROBLEM. Finding a q such that x(q) = 2n∑2 + qn + 1 is an integer is a discrete
m
logarithm problem for a dynamical system. The Taylor series x(q) = i=1 αi q i can be approximated by a
polynomial, which leads to a linear map in m dimensions. Already taking care of the quadratic part would
improve the ability to find small square roots: we have to solve the following problem: consider the area
preserving map T (x, y) = (x + y + α, y + β) on T2 . Find a sequence of qn → ∞ such that T q (x, y) = (xq , yq )
satisfies |xq | < 1/q. For every measure-preserving system on the torus there exists good numbers q but it is
hard to find if the dynamics has no discrete spectrum.

Side Note: factorisation is a priori equivalent to a discrete log problem. Find k with (ak )2 = 1) so that we
know ϕ(n). Because the map T (x) = ax has sensitive dependence on initial conditions, it is again difficult to
solve the discrete log problem.

√
A DIOPHANTINE PROBLEM. We need to find integer points (x,y) close to the graph of f (q) = 2n2 + qn + 1.
Near (q, f (q)), we can form a linear approximation g(x) = αx + β. The CFA gives good integer choices (qk , pk )
near that linear graph. In order that we are close to the graph, we need qk to be close to q. Unfortunately, we
need to take large qk + q to get a good approximation, but then we have a large error also from the nonlinear
corrections. We are forced to take the nonlinear terms into account. But then, we run into computational
problems because the underlying dynamics is nonintegrable. If it were integrable, we could solve the problem
by approximating the dynamics by a quasiperiodic motion on the first few eigenfunctions and solving the discrete
log problem there.

POLYNOMIAL MAPS. If p(x) is a polynomial with at least two rationally independent coefficients, then
p(n) mod(n) is uniformly distributed. In modern terminology, the sequence forms a uniquely ergodic dynamical
system (see Furstenberg, Queffelec, Cornfeld for proofs). Already (x, y) 7→ (x + y, y + β), an Anzai product has
some discrete spectrum mixed with absolutely continuous spectrum (see Cornfeld, Fomin, Sinai). The presence
of absolutely continuous spectrum, makes it unlikely that we can find a fast method to solve the discrete
logarithm problem efficiently like the CFA does in the case of the irrational rotation. We could nevertheless
try: exp(iy · n). √
Finding integer points close to an analytic curve is probably a difficult
√ problem in general. Example nmod(1):
this forms a uniquely ergodic system for any approximation of 1 + n by polynomials. The unique √ ergodicity
probably survives in the limit. What is the spectrum of the shift on the closure of the sequence n?
NUMBERS AND DYNAMICAL SYSTEMS Math118, O. Knill

ABSTRACT. Numbers can be represented in various ways. In many cases, the representation of real numbers
can be seen as a construction in symbolic dynamics.

REPRESENTATIONS OF REAL NUMBERS.

Given a finite generating partition A0 , A1 , ..., An of the interval [0, 1), define f (y) = i if y ∈ Ai and a map
T : [0, 1) → [0, 1) we can look at the orbit of a point y and define the sequence xn = f (T n y).
We are interested in cases, where the sequence xn determines x for all x. If T is a piecewise smooth expanding
map, then this is the case.

Many representations of numbers as sequences of a finite symbols is described

by symbolic dynamics.

DECIMAL EXPANSION. Let T (x) = 10x and f (x) = [10x] where [r] is the 0.8

integer part of r. Let A0 , A1 , ...A9 be the intervals defined by Ak = {f (x) = 0.6

k}.
0.4
This∑is the decimal expansion of x. From the sequence aj , we can reconstruct
∞
x = j=1 aj 10−j . 0.2

0.2 0.4 0.6 0.8 1

CONTINUED FRACTION EXPANSION. Take T (y) = 1/y mod 1 and f (y) = [1/y]. For a point y, define the
sequence an = f (T n (y)). It is called the continued fraction expansion of y. if y is a rational number, then
1
y = [a0 ; a1 , ...an ] = a0 + 1 = pn /qn
a1 + 1
a2 +... 1+a n

If y is an irrational number, then

1
y = [a0 ; a1 , ...an , ...] = a0 + 1
a1 + 1
a2 +... 1+a +....
n

√ √
EXAMPLES.
√ 2 = [1; 2, 2, 2, 2, 2, ...]. Since 1/(2 + x) = x has the solution
√ 2 − 1.
( 5 − 1)/2 = [1; 1, 1, 1, 1, 1, ...]. Since 1/(1 + x) = x has the solution ( 5 − 1)/2.
5/7 = [0; 1, 2, 2]

PARTIAL QUOTIENTS. The partial quotients pn /qn satisfy the recursion pn = an pn−1 + pn−2 , qn =
an qn−1 + qn−2 with the initial conditions p−1 = 1, p0 = a0 , q−1 = 0, q0 = 1 so that p0 /q0 = a0 , p1 /q1 =
a0 + 1/a1 = (a0 a1 + 1)/a1 .

[ ]
pn
CONVERGENCE ESTIMATES. One can write the second order recursion as a first order recursion =
[ ][ ] [ ] [pn−1 ]
an 1 pn−1 pn qn ak 1
. In the product of matrices An = An ...A0 = each matrix Ak =
1 0 pn−2 pn−1 qn−1 1 0
has determinant (−1). The product has therefore the determinant (−1)n . This gives the important identity

pn−1 qn − pn qn−1 = (−1)n

which implies pn−1 /qn−1 − pn /qn = (−1)n /(qn qn−1 ). Since qn ≥ qn−1 = 1, we have qn ≥ n and |pn−1 /qn−1 −
pn /qn | ≤ (−1)n /n2 so that pn /qn is a Cauchy sequence. Because pn /qn is alternatively below and above x (look
at the images of the basis vectors of Ak ), we have even the bound
pn 1 1
|x − qn | < qn qn+1 < 2
qn
SOLVING LINEAR EQUATIONS. Given a, b, c, how do we solve ax + by = c for integers x, y?
Solution: we can solve pn−1 qn − pn qn−1 = (−1)n by making the continued fraction expansion of pn /qn then
multiply the result with (−1)n c.

EXPANSION OF PI. To find the continued fraction expansion of x = π: π = 3 + 1/(7 + ...), look at the orbit
of x = 0.141592653... under the map T (x) = {1/x} and see in which intervals they fall.

T[x ] := Mod[1/x, 1]; S = NestList[T, Pi − 3, 10]; f[x ] := Floor[1/x]; Map[f, S]

Mathematica has already built in the continued fraction expansion as a basic function:

ContinuedFraction[Pi, 10]

The result is π = [3; 7, 15, 1, 292, 1, 1, 1, 2, 1, 3, 1, 14, 2, 1, 1, 2, 2, 2, 2, 1, 84, 2, 1, 1, 15, 3, 13, 1, 4, ...]. Continued frac-
tion expansion of π has been computed up to 108 terms. One can use partial quotients like

π ∼ [3; 7] = 22/7 = 3.14286

π ∼ [3; 7, 15] = 333/106 = 3.14151

π ∼ [3; 7, 15, 1] = 355/113 = 3.14159
to approximate π with rational numbers. Mathematica has the reconstruction of a number from the continued
fraction built in too:
FromContinuedFraction[3, 2, 1]

KHINCHIN CONSTANT. If [a0 ; a1 , a2 , ...] is the continued fraction expansion of a number, then the limit
(a1 a2 a3 ...an )1/n exists for almost all irrational numbers. The limit is called Khinchins constant. Numerical
experiments indicate that this limit is obtained for π but one does not know.

β-EXPANSION. A generalization of the decimal or expansion with respect to an integer base is the beta
expansion.
∑∞ For any given real number β > 1, define the map T (x) = βx and f (x) = [βx]. One has still
x = i=1 ai β
−i
however, the transformation is no more so easy to understand as in the integer case. For
example, Tβ does not preserve the length measure dx any more in general.

PERIODIC POINTS. As in any dynamical system, also for dynamical systems which define number, periodic
points are important. Examples:

• Rational points are eventually periodic points of the decimal expansion.

• quadratic irrationals are eventually periodic points of the continued fraction expansion.
• Numbers which lead to eventually periodic orbits of the β-expansion are called beta numbers.

The determination whether an orbit is eventually periodic or not is nontrivial. For example it is unknown
whether π + e is rational. In other words, one does not know whether the shift on Xπ+e,10 is eventually
periodic.

BETA NUMBERS. An interesting question is for which real numbers β and x = 1, the attractor is a periodic
orbit. If this is the case, then β is called a beta number. Examples are Pisot numbers, algebraic integers
β > 1 for which all conjugates β σ have norm |β σ < 1 besides the identity. The positive root of x3 − x − 1 = 0 is
known to be the smallest Pisot number. If |β σ | ≤ 1 for any embedding and β is not a Pisot number, it is called
a Salem number.

NORMALITY. If every word of length k in the decimal expansion of π appears with probability 10−k , then
π is normal. One does not know whether this is true. Normality results are hard to get. And normality
with respect to one base does not mean normality with respect to an other base. Normality is a statement
with respect a specific shift invariant measure and If a number is normal with respect to all bases is called
absolutely normal. A well studied open problem is

Is π normal with respect to any base or

even absolutely normal?
STRANGE SINGULARITIES AND ORBITS Math118, O. Knill

ABSTRACT. Non-collision singularities are possible in the Newtonian n-body problem by careful construction.
Also the construction of special solutions to the n-body problems is an art.

PAINLEVES CONJECTURE: Painlevé asked in his Stockholm lectures of

1895: for n > 3, do there exist solutions of the Newtonian n-body problem
with singularities that are not due to collisions?

HISTORY. Zeipels theorem showed that singularities of the Newtonian n-body problem are either collisions or
configurations for which particles escape to infinity in finite time. Poincaré seems have considered this question
already, even so he never wrote it down. Painlevé gave Poincarée credit for having asked that some xi (t) might
go to infinity or oscillate wildly like sin(1/(t − τ ) as t converges to the singularity. Painlevé himself proved that
non-collision singularities do not exist for the three body problem. Painlevés question whether non-collision
singularities can occur, stayed open until Jeff Xia constructed non-collision singularities in 1992. (By the way,
Xia was at Harvard from 1988-1990, so some of the final polishing of this paper could have been done here). An
other mathematician, Joseph Gerver, had also been in the race but considered a planar approach, where the
number of particles is large. John Mather and Richard Mc Gehee had already in 1974 shown that particles can
escape to infinity but their construction on the one dimensional line and binary collisions were allowed. While
it is known that for four bodies, non-collision singularities have measure zero, one does not know whether they
exist. There is a construction of a planar 4 body situation of Gerver from 2003 which suggests that the answer
could be yes.

A THEOREM OF PAINLEVE.

THEOREM (1897) There are no non-collision singularities

in the three body problem.
∑3
PROOF. The Lagrange-Jacobi equation is I¨ = U + 2H, where I = j=1 mj rj2 is the moment of inertia. I
is a measure of diameter of the triangle defined by the positions of the three particles. These equations imply
that whenever two particles come close, the triangle they span has to become large. By the triangle inequality,
two sides of the triangle are then large. The Sundman-van Zeipel lemma assured that I(t) → I ∗ for t → τ with
I ∗ = ∞ if there is a non-collision singularity. Assuming I(t) → ∞ for t → τ we have I(t¨k ) → ∞ for some
sequence of times tk → τ which means U (tk ) → ∞. This implies that two of the three particles must come
close to each other. In the same time, the third ”lonely” particle has to be far away from these two particles
because I(t) → ∞. Because the acceleration of the lonely particle and the center of mass of the binary both
stay bounded for t → τ , these positions converge to a definite finite value for t → τ . The collision assumption
means that the binary system collides for t = τ but at a finite distance from the third particle. Consequently
I(τ ) = I ∗ < ∞, which is in direct contradiction to the assumption I ∗ = ∞.

THEOREM OF XIA.

Non-collision singularities exist in the Newtonian 5 body

problem. There are initial conditions for the Newtonian
5 body problem in which the bodies escape to infinity in
finite time.

BASIC IDEA. The setup is to add a second binary solar system to the Sit-
nikov system. The planet moving on the z-axes visits alternatively the two
binary systems. The timing is done in such a way that the planet will bounce
back accelerated after visiting one of the systems. The energy is drawn from
the potential energy of the two binary systems which move closer and closer
together. The four suns have all the same mass. The upper and lower ”so-
lar systems” have opposite angular momentum and their ”Kepler orbits” are
highly eccentric.
THEOREM OF GERVER. Joseph Gerver proved a theorem for the planar
case:

THEOREM. For large n, non-collision singularities exist for

the planar n-body problem

BASIC IDEA. There are 3N bodies in the plane. The configurations are sym-
metric with respect to rotations by 2π/N . There are N binary systems in which
all suns have the same mass. There are N planets which move from one pair to
the other. The successive time spans, which the planets need to jump
∑ from one
to the next system forms a sequence ∆k with the property that k ∆k < ∞.

GERVERS SUGGESTION: Are there planar four body configurations in

which particles escape to infinity in finite time?

Gervers model contains two planetary systems: there are two suns S1 , S2 with
large mass and two planets P1 , P2 with small mass. Planet P2 circles sun S2
in an elliptical orbit. Planet P1 circles around Sun S1 and visits the second
planetary system, where it alternatively gains angular momentum and energy.

SPECIAL SOLUTIONS. An interesting research topic is the search for special solutions of the 3 body problem.

EQUILIBRIUM SOLUTIONS. Whenever we studied differential equations, we

were interested in equilibrium solutions, stationary solutions. Are there
equilibrium solutions for the Newtonian n body problem? The answer is no:
from ẍk = 0 would imply that∑nUxk = 0 and from Eulers theorem on homoge-
neous functions that −U = j=1 xk Uxk = 0. But the potential U is clearly
positive everywhere.
EULERS SOLUTIONS (1767) Euler was the first who found special solutions
to the three body problem. In these solutions, the three bodies rotate on circles
but remain on a line. The Euler solution and the Lagrange solution below are
the only solutions for which the particles move uniformly along circular orbits
in a fixed plane.

LAGRANGIAN SOLUTIONS (1772). The three bodies are on an equilateral

triangle. This system appears in nature: the Trojan asteroids together with
Jupiter and the sun essentially move according to this. Lagrange, who found
this solution did not think this has any significance in astronomy.

HILLS SOLUTIONS. These are configurations resembling the Earth-Moon-Sun

system. Two bodies move closely around each other while both of them circle
a third body.

MOORE CHOREOGRAPHIES. Three bodies of equal mass follow each other

on a figure eight type orbit. These solutions have been discovered by Cristopher
Moore in 1993 through computer calculations.

LITERATURE. A vivid account on the history of non-collision singularities also containing many annectotes
about the discovery is the book ”Celestial Encounters” by Florin Diaco and Philp Holmes. The article ”Off
to infinity in Finite Time” by Donald Saari and Jeff Xia gives a nice summary. For a suggestion, how a four
body noncollision singularity might work, see Joseph Gervers article ”Non collision Singularities: Do four bodies
Suffice?”.
RESTRICTED 3 BODY PROBLEMS Math118, O. Knill

ABSTRACT. The Newtonian 3 body problem can exhibit chaos. The simplest situation is when the third body
moves in the time dependent potential of a binary system but itself does not influence the motion of the binary
system. A first example is the Sitnikov problem, where one can establish the existence of a horse shoe which
leads to a in general chaotic calendar for inhabitants of the Sitnikov planet. An other example is the circular
planar restricted three body problem which leads to cases, where one has an area preserving map on a region
with finite area. It is also a historically important example because some results in ergodic theory like Poincare
recurrence and topology like fixed point theorems were developed with the three body problem in mind.

RESTRICTED THREE BODY PROBLEMS. The restricted 3-

body problem deals with the situation, where one of the three
bodies has a neglectable mass, and moves under the influence of
the two other bodies which evolve according to Keplers law. Lets
call here the two heavy bodies the double star binary system
and the third body the planet.

ASTEROID 2004 MN4 IMPACT RISK? In December 2004, As-

teroid 2004 MN4 was given a 1/233 chance, then a 1/38 chance to
hit the earth in April 13, 2029. Despite numerological support for
bad luck like 2+0+2+9=13 and 1+3=4=shi also means ”death”
in Japanese, subsequent observations have shown that there will
be no impact in 2029. It will pass by the Earth at a distance of
between 15’000 and 25’000 miles, about a tenth of the distance
between the Earth and the Moon and be so close that it can e
seen with the naked eye. The change of orbit might put 2004 of
a collision course in 2034, 2035 or 2036. One will know more in
2029.

SITNIKOV PROBLEM. The Sitnikov problem deals with the

situation, where the double star system moves in the xy-plane
and the planet is on the z-axes. Both stars have equal mass m
normalized to m = 1/2 and move on elliptic orbits, where the
center of mass is at rest. The third body has no mass. Its z
coordinate satisfies the Sitnikov differential equation

d2 z z
=− 2 ,
dt2 (z + r(t)2 )3/2

where r(t) is the distance of a sun to the origin at time t. By

normalizing time, we can assume that r(t) has period 2π. For
small values of the eccentricity ϵ of the ellipse, one has r(t) =
1 2
2 (1 − ϵ cos(t)) + O(ϵ ).

SITNIKOV YEAR. A Sitnikov year is the time it takes to return

to the xy-plane, the summer position on Sitnikov planet. Winter
is when the planet has the maximal distance to the stars. The
inhabitants on ”Sitnikov” know to measure time and count the
number of Sitnikov days in one Sitnikov year k by

sk = [(tk+1 − tk)/2π] .

Far away from the double star system, a winter day could look as
in the picture to the right.
A CHAOTIC CALENDAR.
THEOREM (Sitnikov-Moser) For sufficiently small eccentricity ϵ > 0, there
exists an integer m such that for any sequence s1 , s2 , ... of of integers sk ≥ m,
there exists a solution of the Sitnikov differential equation for which year k has
sk days.

REMARKS. One can also allow sk = ∞ in which case, the planet would escape for ever, or the solar binary
system could capture an orbit which stays bounded for ever. The proof of the theorem relies on the horse shoe
construction and is robust. The result therefore holds also for planets with small positive mass. The result can
be shown to be true for all 0 < ϵ < 1 except a discrete set of values.
Most orbits in this dynamical system go to infinity. It is not
quite clear what the filled in Julia set is, the points which stay
bounded for all times. Sitnikov-Moser theorem constructs a Can-
tor set of points which stay bounded for ever. It is not excluded
that there are some stable elliptic periodic points. Numerical ex-
periments suggest that such stable periodic points exist but I have
not seen a proof. The stability problem is in nature similar to the
one for the quadratic Henon map in the plane and depends on
subtle Diophantine properties which have to be satisfied for the
periodic points. We expect for most parameter values ϵ a set of
positive area stays bounded. This could be good news for Sitnikov
inhabitants.
The bad news is that these regions might be very small and a small disturbance - for example by an asteroid -
could free the Sitnikov planet and send its inhabitants to a deadly eternal winter ride. One of the last pictures
taken from that escaping planet could look as the picture above.

TO THE PROOF (Moser 1973).

Look at the Poincare return map to the plane with polar coordinates (r, ϕ) = (|v|, t), where v is the velocity
of the planet and t mod 2π is the time given by the suns clock. t = 0 corresponds to the moments, when the
suns are closest to the z axes. The return map is defined in a simple closed region D0 . Outside this region, the
orbit escapes. Here is an outline of the proof. The details are quite technical and can be found in Mosers book.

(0) The return map Tϵ maps D0 into D1 = ρ(D0 ), where ρ is the reflection (v, t) → (v, −t). The map Tϵ is area
preserving: the area element 2vdvdt = dE dt is preserved.

(i) For small enough ϵ, the boundaries of D0 and D1 are smooth curves which intersect transversely. The proof
of this fact is done by writing the right hand side of the Sitnikov equations as a power series in ϵ and neglecting
ϵ2 and larger terms. This computation from perturbation theory allows to establish that the angle between the
boundary curves becomes nonzero.

(ii) For ϵ = 0, the map T0 is integrable and of the form

[ ] [ ]
v v
T0 =
t t + f (v)

where f (v) → ∞ if v → 2. The differential equation is in this case

−z
z̈ = .
(z 2 + 1/4)3/2

This is an integrable system: indeed, the energy

1 2 1
E= ż − √ ≥ −2
2 2
z + 1/4

is conserved and the map leaves its level curves of E invariant. The origin is a fixed point, each circle gets
rotated and the rotation becomes faster and faster until the boundary E = 0 is reached. In physical terms, this
means that if we start with a larger initial velocity, it takes longer to return.

(ii) There are horse shoes arbitrarily close to the boundary of D0 . This is a consequence of i and ii and will be
explained in class. (needs a good picture)
PLANAR CIRCULAR THREE BODY PROBLEM. The planar restricted 3-body problem deals the sit-
uation, where one of the three bodies has neglectable mass, but moves under the influence of two other bodies
which evolve along circles according to Keplers law. An example is the motion of the moon in the influence of
the earth and sun. A second example is the motion of an asteroid under the influence of the sun and Jupiter,
the second largest body in our solar system. An other example is the motion of a planet in a binary star system.

ROTATING COORDINATE SYSTEM. Assume [ ⃗y = ]R(ωt)⃗x, where R(α) is a rotation in the plane with angle
0 −1
α. We can write R(ωt) = eAωt , where A = .
1 0

LEMMA. In the rotating coordinate system

d2 d2 d
2
⃗y = R 2 ⃗x + 2AωR ⃗x − Rω 2 ⃗x
dt dt dt
one observes additionally to the rotated forces also a centrifugal force and
a velocity dependent Coriolis forces.

PROOF. Differentiating twice the identity y = Rx using Ṙ = ωAR gives ẏ = Ṙx + Rẋ = ωAR⃗x and ÿ = ω[2 A2 R⃗]x +
2 ẏ1
2ARẋ + Rẍ. Because A = −1, this gives the equation in the lemma. The same calculation in coordinates: =
ẏ2
[ ] [ ] [ ] [ ]
ẋ1 − ωx2 ÿ1 ẍ1 − ω 2 x1 − 2ωx2 cos(ωt) − sin(ωt)
R and =R , where R = . Remark. The same
ẋ2 + ωx1 ÿ2 ẍ2 − ω 2 x2 + 2ωx1 sin(ωt) cos(ωt)
computation can be done in three dimensions, where both the centrifugal and Coriolis forces can be expressed
using cross products.

THE EQUATIONS OF THE PLANAR CIRCULAR 3-BODY PROBLEM. Two stars of mass m1 = µ, m2 =
1 − µ move on circular orbits along their center of mass. Going into a rotating inertial coordinate system
(Keplers 3. law implies from zero eccentricity uniform rotation), in which the stars are fixed at the points
(1 − µ, 0), (−µ, 0), the equations of motion become

d d
xk = Eyk , yk = −Exk ,
dt dt

where E = 21 (ẏ12 + ẏ22 ) + 2x2 y1 − 2x1 y2 − rµ1 − (1−µ)

√
r2 is the Hamilton function. Here r = x21 + x22 is the
√ √
distance of the planet to the origin, r1 = (x1 + µ − 1)2 + x22 and r2 = (x1 − µ)2 + x22 are the distances
from the planet to the two stars. We can decompose E = (ẋ21 + ẋ22 )/2 − U (x1 , x2 ) with U = 21 r2 + rµ1 + (1−µ) r2 .
The function E is called the Jacobi integral. It contains 21 r2 called centrifugal potential and ẋ21 + ẋ22 ,
the Coriolis potential. How did we get that? The Newton equations in the rotating coordinate system are
according to the previous lemma:
After multiplying the first equation with ẋ1 and the second with ẋ2 ,
∂ ∂ ∂
ẍ1 − 2ẋ2 = ∂x 1
U addition gives ẋ1 ẍ1 + ẋ2 ẍ2 = ∂x 1
U ẋ1 + ∂x 2
U = U̇ so that E = (ẋ21 +
∂ 2
ẍ2 + 2ẋ1 = ∂x2 U ẋ2 )/2 − U is conserved. Introducing y1 = ẋ1 − x2 , y2 = x1 + ẋ2 leads to
the Hamilton equations at the top of this box.
What is the deal? We started with the Newton equations ÿi = ∂x∂ 1 W and ended up with a system looking more
complicated. But it is not! In the original coordinates, the potential W is time dependent! Especially, there was
no energy conservation. Going into the rotating coordinate system led us to a Hamiltonian system with a preserved
quantity, the Jacobi integral.

HILLS REGION. Assume E = c1 and c < c1 . The regions

U (x1 , x2 ) = c bound regions in the (x1 , x2 ) plane called Hills
regions.

LEMMA. If (x1 , x2 ) is in a Hills region U ≥ c, then

(x1 (t), x2 (t) is in the Hills region for all times.

For large c, these regions consist of three parts. Two in the neighbor-
hood of the two stars (satellite bound by one of the bodies) and one
far away (asteriod encircling both). They define an allowed region in
which the planet can stay. A large c corresponds to the case, where one
is either close to one of the stars with large gravitational potential or
very far away, with large centrifugal potential.
RECURRENCE. The energy surfaces E = c are invariant as are the sets {(x1 , x2 ) | a ≤ −E(x1 , x2 ) ≤ b} for
a < b. If c < c1 , then
1
G = (ẋ21 + ẋ22 ) − E > c1 > c .
2
So, (x1 , x2 ) stays in a bounded region. Also (x1 , x2 , y1 , y2 ) stays in a bounded set. The differential equation
preserves the four dimensional volume. When normalizing the volume to 1, we obtain a probability space. The
time 1 map is a measure preserving map on that space and Poincares recurrence theorem applies.
There is a subtlety with this argument which has to be mentioned: Not all solutions in the finite region have
a global solution. There are initial conditions, in which the planet crashes into one of the suns but these cases
can be shown to have zero volume.

CHAOS IN THE SOLAR SYSTEM. Chaos in the solar system has been measured at different places:

1) The solar system itself is weakly chaotic. The Lyapunov expo-

nent has been measured to be very small 2.8 · 10−15 . For Pluto
the Lyapunov exponent had been measured 7 · 10−16 . Numeri-
cal experiments have also been done with other parameters. The
heliocentric distance for outer planets would behave much more
erratically, if the sun would have 1/3 less of its current mass, sug-
gesting that some of the outer planets like Neptune or Uranus
would escape in such a case. For our solar system, it looks as if
one can not predict the trajectory of the earth for time periods ex-
ceeding 100 Million years. More precisely, the uncertainty of 1 km
in the initial condition could lead to an uncertainty of the order of
1 astronomical unit in 100 Million years. Numerical simulations
of the solar system have been done for time intervals reaching 35
billion years.
2) Many comets and asteroids in the solar system have irregular
orbits. Numerical experiments have been done for example in the
case of the asteroid Chiron. To measure sensitive dependence on
initial conditions, one starts integrating with various close initial
conditions and looks at the outcome. Chiron will undergo several
close approaches to planets. One estimates a 1/8 chance that
Chiron will eventually leave the solar system. Other objects have
an other fate. The comet Shoemaker-Levy 9 had a spectacular
impact with Jupiter in July 1994 after having been disrupted by
a close Jupiter approach in 1992.

3) The tumbling of Saturns little moon Hyperion. Most satellites

in the solar system are in synchronous rotation, keeping one face
towards the planet. Hyperion has an irregular shape and is known
to tumble erratically in its orbit. The Cassini spacecraft will
fly past this moon later this year, on September 26, 2005. The
Lyapunov exponent of the irregular tumbling motion has been
been measured to be of the order 10−7 .

4) The motion of charged particles in a magnetic dipole field has

been shown to be chaotic. Brown has constructed a horse shoe
for the return map. The dynamics can be reduced to a relatively
simple Hamiltonian system
1 2 1 1 q1
H(q1 , q2 , p1 , p2 ) = (p1 + p22 ) + ( − 2 )2
2 2 q1 (q1 + q22 )3/2

called the Stoermer problem. The dynamics of charged parti-

cles in the van Allen belts can explain the aurora Borealis.
For the Lyapunov exponent data on this box, we the sources:
P. Gaspard: ”Chaos Scattering and Statistical mechanics”, 1998
I. Peterson: ”Newtons Clock: Chaos in the solar system”, 1993
C.D. Murray and S.F. Dermott: Solar system dynamics”, 2001
D. Goroff: Editorial introduction article in ”New Methods of Celestial Mechanics by H. Poincare”.
K. Zyczkowski ”On the stability of the Solar system”.
For the planar 3 body problem, we followed Siegel-Moser. Sitnikovs problem is treated in detail in Mosers 1973 book.
SINGULARITIES Math118, O. Knill

ABSTRACT. Singularities for the n-body problem can occur when bodies collide or when bodies escape to
infinity. A theorem of van Zeipel shows that these are the two only possibilities.

OPEN PROBLEM. Lets start with a major open problem in

celestial mechanics.

Is it true that the Newtonian n-body problem has a full

3d-dimensional Lebesgue measure set of initial conditions,
for which the solutions exist for all times? In other words,
can the singularity set have positive measure?

COLLISIONS.

If x(t) → ∆ for t → τ , then x(τ ) is called a col-

lision singularity. Collisions can already occur in Our galaxy and M31, the Andromeda galaxy, form a
the 2-body problem, if the total angular momentum relatively isolated system known as the local group.
of the two bodies is zero. Analysing collision sin- The center of mass of M31 approaches the center of
gularities involving more than two bodies helps to mass our galaxy with a velocity of 119 km/s. In
understand what happens when particles move close about 1010 years, these galaxies are likely to collide.
to such collision configurations. It is known that ini- Such a collision would have dramatic consequences
tial conditions leading to collisions are rare in the for both systems. Nevertheless, even a direct en-
n-body problem. Noncollision singularities in which counter would probably not lead to any collision of
particles escape to infinity in finite time exist already stars.
for the 5-body problem.

EXISTENCE OF SOLUTIONS.

For every point (x, y) in phase space, there exists τ = τ (x, y) such that for t ∈ [0, τ (x)] the Newtonian
n-body equations have a unique solution (xt , y t ). Moreover, if K is a closed and bounded subset in the
phase space, then there exists δ > 0 such that (xt , y t ) is outside K for t ∈ [τ − δ, τ ].

(i) The first statement follows from a general existence theorem for differential equation ẋ = f (x) on a subset
M of Euclidean space. The function ẋ = f (x) is Lipshitz continous on a bounded open set in M .

(ii) For any compact (closed and bounded) set K, there is a time τK = minx∈K τ (x) > 0 such that for all
initial conditions x ∈ K, a solution exists in the time interval [0, τK ).
Therefore, if x(t) exists in the interval [0, τ ) and the solution can not be extended beyond τ , then for t ∈
(τ − τK , τ ], x(t) is outside K.

The existence theorem shows that if

SINGULARITIES. A point (x, y) ∈ (T ∗ M )n is called a singu-
a singularity is approached, then the
larity if τ (x, y) < ∞. A singularity is called a collision if there
some velocities become unbounded. It
exists x ∈ ∆ such that xt → x. A singularity which is not a
is not possible that posititions become
collision is a pseudo collision or a non collision singularity.
unbounded but velocities stay bounded.

PAINLEVE THEOREM. If (x, y) is a singularity, then |U (xt )| →

∞ for t → τ (x, y). In other words, the minimal distance between
two particles goes to zero. This result holds in any dimensions
and for any potential U = u(|x|) satisfying u(r) → ∞ for r → 0
and such that u ∈ C 2 ([ϵ, ∞)) for every ϵ > 0.
PROOF. Assume the contrary: there exists δ > 0 such that mini̸=j |xti − xtj | ≥ δ for t ∈ [0, τ ). We want to show
that τ is not maximal.

(i) The differential equation ẋ = f (x) with |f | ≤ M in Br (x0 ) and f ∈ C 1 has a solution xt with x0 = x0 , as
long as |t| ≤ r/M . The piece of orbit {xt }t∈[0,r/M ] is contained in Br (x0 ).

Proof. See the proof of the Cauchy-Piccard existence theorem.

/ Br (x0 ).
(ii) There exists M such that |∇x U | ≤ M for x ∈

Proof. We have 0 ≤ −U ≤ C/ρ, where C is a constant depending only on n and the masses mj . Therefore, we
have |∇x U | ≤ C/ρ2 .

(iii) There exists M such that |yj | ≤ M .

Proof. This follows from the decomposition of the energy H = K + U and the boundedness of U .
∑ d 2 2
j=1 yj /2mj ≤ H + 2M /d.

(iv) For t arbitrarily close to τ (x, y), we can extend the solution for the time interval [0, r/2M ].
Proof. Using (ii), (iii), we can apply (i).
∑n
MOMENT OF INERTIA. The number I(x) = i=1 mi |xi |2 the moment of inertia of the configuration.

LAGRANGE-JACOBI FORMULA. 12 I(x ¨ t ) = U (xt ) + 2H(xt , y t ) = T (y t ) + H(xt , y t ), where H(x, y) = T (y) −

U (x) is decomposition of the energy into kinetic and potential energy.
∑d
PROOF. From 21 I˙ = j=1 mj (xj , ẋj ), we get

1¨ ∑
d ∑
d
I = mj (xj , ẍj ) + 2T = (xj , −∇xj U (x)) + 2T
2 j=1 j=1
= U + 2T = −U + 2H = T + H .

We have used that U is homogeneous of degree −1: U (λx) = λ−1 U (x) which gives with the Euler identity
(x, ∇x U ) = −U .

REMARK TO 4D. Interesting is the analoguous case in n = 4, where U is homogeneous of degree −2. Then
1¨
2 I = 2H is constant. This shows that we have in the case of a negative initial energy H < 0 always collapse
in finite time and that solutions can stay bounded only on the energy surface H = 0. You have this fact in the
case of the Kepler problem in four dimension.

SUNDMAN-VAN ZEIPEL LEMMA If (x, y) is a singularity, there exists I ∗ = I(xτ (x,y) ) ∈ [0, ∞] such that
I(xt ) → I ∗ for t → τ (x, y). The same relation holds for potentials for which x · ∇x U (x) + U (x) is globally
bounded.

PROOF. From the Lagrange formula and the theorem of Painlevé, we see that I¨ > 0 for t near τ (x, y). This
implies that I˙ is monotonically increasing and one can assume that I˙ is always positive or always negative in
the interval [t, τ ] because one could else, if it changes sign, make the interval smaller. The positive function I
is therefore monotonic and has a limit.

VAN ZEIPEL’s THEOREM. This is a heavy theorem. Even so

the proof had been simplified considerably by McGehee, its not
possible to hide that this is a relatively deep result:

THEOREM. If (x, y) is a singularity, then I(xτ (x,y) ) < ∞ if

and only if (x, y) is a collision. In other words, I(xτ (x,y) ) = ∞
if and only if (x, y) is a pseudo-collision.

The proof follows closely McGehee’s 1986 paper.

PROOF (i) Clusters. Denote by ω a partition ∩ of the set N = {1, . . . , n}. For µ ⊂ N , define
∆µ = {x ∈ R3n | i, j ∈ µ ⇒ xi = xj } and ∆ω = η∈ω ∆µ .
 ∑
PROOF (ii) New scalar product. Consider the scalar product in R3n by < x, x′ >= j mj (xj , x′j ), where
(·, ·) is the standard dot product in R3 . The norm ||x|| of x in this scalar product allows to rewrite the moment
of inertia as I(x) = ||x||2 .

PROOF (iii) Orthogonal decomposition. Define for µ ⊂ N the linear map R3n → R3
∑ ∑
x 7→ πω x = cµ x = mi xi / mi
i∈µ i∈µ

3n 3n
and the linear map πω from R ∑ to R . We have (πω x)i = πµ x if i ∈ µ. This is an orthogonal projection with
range ∆ω and kernal Γω = { j∈µ mj xj = 0 ∀µ ∈ ω}. Denote by Πω = Id − πω the orthogonal projection onto
Γω . Write x = πω x + Πω (x) = z + w.

∑ ∑
PROOF (iv) Moment of inertia. Define Iω (x) = ||πω x||2 = µ∈ω∑ ( j∈µ mj )|cµ x|2 . Denote by Jµ the
moment of inertia of a subsystem having particles j ∈ µ and by Jω = µ∈ω Jµ the sum of these moment of
inertias. The equation
||x||2 = ||πω x||2 + ||Πω x||2 = Iω (x) + Jω (x)
means that the total moment of inertia is the sum of the moment of inertias of the subsystems and the fictious
system obtained from the center of masses of the subsystems.

m m
RROOF (v) Potential energy. Define Uij (x) = 12 |xi i−xjj | for i 6= j and Uij (x) = 0 if i = j. Let Vµ (x) =
∑ ∑
i,j∈µ Uij be the potential energy of the subsystem µ and∑ Vω (x) = µ∈ω Vη (x) the sum of the potential
energies of the subsystems of a partition ω. Define Uµν (x) = i∈µ,j∈ν ∑Uij (x) if µ ∩ ν = ∅ and Uµν (x) = 0 else.
The potential energy due to the interaction of the subsystems is Uω = µ,ν∈ω Uµν . The total potential energy
U (x) can be written as
U (x) = Uω (x) + Vω (x) .

PROOF (vi) Dynamics. For z ∈ ∆µ , we have Vω (x + z) = Vω (x) which gives Vω (x + πω y) = Vω (x) for
all y ∈ R2n . Differentiation of this with respect to y and putting y = 0 gives ∇Vω (x)πω = 0. Because πω is
orthogonal, we have therefore πω ∇Vω (x) = 0. Applying the projection πω on ẍ = ∇U (x) = ∇Uω + ∇Vω gives

ẅ = πω ẍ = πω ∇Uω ,

from which we derive

d2 d2
2
Iω (x) = 2 < πω x, πω x >= 2 < πω ẋ, πω ẋ > +2 < πω x, πω ∇Uω (x) > .
dt dt

PROOF (vii) Statement of the goal: We assume that I(xt ) → I ∗ < ∞ and show that xt converges.

RROOF (viii) The collision set ∆∗ . By assumption on the theorem, the set
∩
∆∗ = O(t, t∗ ) ⊂ ∆
t<τ (x,y)

with O(a, b) = {xt }t∈(a,b) is nonempty and compact. For each partition ω define ∆∗ω = ∆∗ ∩ ∆ω . From the
partitions ω with ∆∗ω we choose a partition with minimal cardinality and fix this partition for the rest of the
proof.

PROOF (ix) Bound the force in a neighborhood G of ∆∗ . Since ∆∗ is compact we can find an open
neighborhood G of ∆∗ and a constant M such that

||∇Uω ||, | < πω x, ∇Uω (x) > | ≤ M .

PROOF (x) If ∆∗ is a subset of ∆ω , then xt converges.

If ∆∗ ⊂ ∆ω , then z t = πω xt converges for t ∈ τ (x, y). There exists t2 such that xt ∈ G for t ∈ (t2 , τ (x, y)).
From ẅ = πω ∇Uω (x) and the bound in (ix), we get ||ẅ|| ≤ M for t ∈ (t2 , τ (x, y)). It follows that wt approaches
a limit w∗ for t → τ (x, y). Hence xt = wt + z t → w∗ + 0 converges.
PROOF (xi) The situation that ∆∗ is a not a subset of ∆ω is not possible. Assume ∆∗ is a not a
subset of ∆ω . In claim (ix) below, we will derive a contradiction and so finish the proof of the theorem.

PROOF (xii) Definition of a compact set Kσ ⊂ R3n . Choose a bounded open subset B of ∆ω such that
∆∗ω ⊂ B ⊂ B ⊂ ∆ω ⊂ G. Let Dσ denote an open ball of radius σ in the linear space Γω . Define the compact
set
Kσ = B × D σ .
Since the boundary δB of B is compact and B does not intersect ∆∗ , there exists σ0 and t0 < τ such that
O([t0 , τ )) ∩ Dσ0 × δB = ∅. We can choose σ0 so small that additionally Kσ0 ⊂ G.
Since by our assumption, ∆∗ is not a subset of ∆ω , there exists 0 < σ < σ0 such that for infinitely many values
of t close to t∗ , we have xt ∈
/ Kσ . Choose and fix σ with this property.

PROOF (xiii) Definition of a time t1 . Chose t1 so small that

σ2
|I(xt ) − I ∗ | ≤ , ∀t1 ≤ t < t∗ .
12

PROOF (xiv) Definition of a time interval I = [a, b] with some properties. There exists an interval
I = [a, b] such that
1) O(I) ⊂ Kσ 2) ||Πω xa || = √
||Πω xb || = σ 2
t 2
3) mint∈[a,b] ||Πω x || < σ /2 4) b − a < σ/ 3M .
Proof. Because xt comes arbitrarily often arbitrarily close to ∆∗ω , xt must enter and leave Kσ infinitely many
often. 1) is therefore no problem for intervals arbitrarily close to τ . 3) can be met for intervals arbitrarily close
to τ because xt comes arbitrarily often arbitrarily close to ∆∗ω , where ||Πω xt || = 0. 2) is clear because if xt
enters Kσ it can not enter through Dσ × δB and must therefore enter through δDσ × B.

PROOF (xv) Let s ∈ (a, b) be such that Iω (xt ) is maximal. Remember that I(xt ) = Iω (xt ) + Jω (xt )
converges for t → τ so that a maximum exists from (vii) and (vi).

σ2
PROOF (xvi) Derive a contradiction. From mint∈[a,b] ||Πω xt || < 2 and |I(xt ) − I ∗ | < σ 2 /12, t1 ≤ t < t∗
we obtain
7σ 2
Iω (xs ) > I ∗ − .
12
From ||Πω xa || = ||Πω xb || = σ 2 and |I(xt ) − I ∗ | < σ 2 /12, t1 ≤ t < t∗ we obtain

11σ 2
Iω (xb ) < I ∗ −
12
so that
σ2
Iω (xs ) − Iω (xb ) > .
3
On the other hand we have for t ∈ [a, b]

d2
Iω (x) = 2 < πω ẋ, πω ẋ > +2 < πω x, πω ∇Uω (x) > ≥ 2 < πω x, πω ∇Uω (x) >≥ 2M .
dt2
Because s is a local maxium of Iω (xt ), we know that

σ2
Iω (xs ) − Iω (xb ) ≤ M (b − s)2 < ,
3
where the last inequality uses 4) from (ivx).

REMARK. Edvard Hugo von Zeipel (1873-1959) was a Swedish astronomer. Van Zeipel’s result holds for
every potential U (x) for which one can prove a Sundman-Van Zeipel lemma. For the Newton potential in four
dimensions, where I¨ = const, we know trivially that I ∗ exists in (0, ∞). It follows that for the graviatational
Newton potential in four dimensions, there are only collision singularities. For negative energy they have full
measure.
SHORTEST PATHS IN THE PLANE Math118, O. Knill

ABSTRACT. The minimization of the arc-length while connecting two points in the plane has been studied by
Archimedes already. It can also be solved, if the arc length is generalized. It leads to differential equations.

PLANE. Given two points P, Q in the plane. What is the path connecting P with Q which minimizes the
length? While everybody knows that the straight line solves this problem, how does one prove this?

CONNECTING POINTS IN THE PLANE. Let f (x) be a graph over the interval [a, b] such that P = (a, f (a))
and Q = (b, f (b)). The length of this graph is
∫ b √
I(f ) = 1 + f ′ (x)2 dx .
a

Which function f minimizes that? We could look at paths connecting points (xi , yi ) with (x0 , y0 ) = P and
(xn , yn ) = Q using fi (x) = f (a)(x − xi ) + (x − xi )(yi+1 − yi )/(xi+1 − xi ), to connect neighboring points. The
∑n−1 √ ∑n−1
length of such a graph is by Pythagoras I(y1 , ..., yn−1 ) = i=0 (xi+1 − xi )2 + (yi+1 − yi )2 = i=0 li . To
minimize this, the gradient of I must vanish. Because the partial derivative with respect to yi is (yi − yi−1 )/li −
(yi+1 − yi )/li = sin(αi ) − sin(αi+1 , all the slopes of the polygonal graph must agree and the line has to be a
straight line. We have verified

LEMMA. Among all polygonal graphs connecting P and Q, the straight line
has minimal length.

One can also see by the triangle inequality that any corner in the
graph can be shortened. A polygon which is not a straight line can
be shortened by a definite amount. For any given differentiable
function f , we can approximate the graph of f by piecewise linear
graphs of gn so that the length differences ϵn of the f and g
graphs goes to zero. If there was a f for which the length were by
an amount δ > 0 smaller than the length of the straight line, we
could approximate that function f with a polygon gn for which
ϵn < δ and have a polygon with smaller length contradicting the
lemma. We have now shown:

THEOREM (Archimedes). Among all differentiable functions whose graph

connects two points P and Q in the plane, the straight line minimizes the
length.

Remark: this proof seems oblivious, since it can be shot down with mathematical cannon called ”calculus of
variations”. Besides the fact that it is always nice to avoid heavy artillery, if not needed, the Archimedes proof
has an advantage: it goes through also in a larger class of rectifiable functions which do not need to be differ-
entiable like Snells refraction example below. The discretization approach also generalizes to inhomogeneous
media, where it gives a numerical method. Remarkably, the proof does not need the notion of ”derivative” at
all, if one defines ”rectifiable curves”, as curves for which the lengths of the polygonal approximations converges
and replaces ∇I = 0 with the triangle inequality.

INHOMOGENEOUS MEDIUM. Lets assume that we are in a medium, where it is difficult to travel at some
places and hard at others. If we replace the length by the work
∫ b √
Ig (f ) = g(x, f (x)) 1 + f ′ (x)2 dx
a

and again ask for the problem to find the most efficient path connecting two points P and Q, the result will
critically depend on the function g(x, y). There will be no more unique solutions. Lets discretize the problem
∑n−1
again: we have to minimize I(y1 , ..., yn−1 ) = i=0 g(xi , yi )li . Setting the partial derivatives with respect to yi
equal to zero shows that gy (xi , yi )li + g(xi , yi )(yi+1 − yi )/li = C is constant. This allows to compute recursively
the slope
sin(αi ) = (C − gy (xi , yi ))/g(xi , yi ) .
The constant C is obtained by the requirement that P and Q are connected.
EXAMPLES. The following examples were obtained by numerically solving for the shortest path connecting
two given points.

Rippled medium. The The inhomogeneity is ver- A more general optimiza-

Flat medium. The short-
path prefers to stay in the tical. Again, the parti- tion problem: again the
est connection between
bright regions, where trav- cle prefers to stay in the path tries to avoid staying
two points is a line.
eling is easy. bright regions. too long in the dark area.

EULER-LAGRANGE EQUATIONS. Let F (t, x, p) be a function of three variables. We look at the variational
problem to extremize
∫ b
I(γ) = F (t, x(t), ẋ(t)) dt
a

among all smooth paths γ connecting x(a) with x(b). If t 7→ h(t) is an other path, then (I(γ + h) − I(γ)) =
Dh Ih+O(h2 ) for h → 0 defines a ”directional derivative” Dh I called here the first variation. By linearizing F ,
∫b ∫b
we know that I(γ +h)−I(γ) = a Fx (t, x, ẋ)+Fẋ (t, x, ẋ) dth+O(h2 ) = a Fx (t, x, ẋ)− dt
d
Fẋ (t, x, ẋ) dth+O(h2 ).
d
The first variation is zero if Fx (t, x, p) = dt Fp (t, x, p) for all t. These are the Euler-Lagrange equations.

∫b
INHOMOGENOUS PLANE. If γ : t 7→ (t, x(t)) is a curve in the plane, we can look at a F (t, x, ẋ) dt =
∫b√ √
a
1 + ẋ(t)2 dt. The Euler equations show that ẋ/ 1 + ẋ(t)2 is time independent. Therefore ẍ is constant
and consequently, the √ optimal curve isd a straight line. In the inhomogeneous case, the Euler-Lagrange equations
2 ẋ(t)g(t)
for F (t, x, ẋ) = g(t) 1 + ẋ are 0 = dt ( √ 2
). This proves
1+ẋ(t)

√
SNELLS THEOREM. g(t)ẋ/ 1 + ẋ2 = g(t) sin(α(x)) is constant, where α(x)
is the angle the curve makes with the x axes.

SNELLS LAW. A limiting situation is when the medium has two densities like
air and water. In this situation, the Euler-Lagrange equations do not help. But
the Archimedes approach still works. If g = u on the left hand side and g = v
on the right hand side, then sin(αi ) = sin(αi+1 ) as before in the left or the
right region and u(yi − yi−1 )/li − v(yi+1 − yi )/li = u sin(αi ) − v sin(αi+1 = 0
at the boundary. Therefore, the shortest path is a line with angle α on the left
hand side and angle β on the right hand side and

u sin(α) = v sin(β).

This is called Snells law named after Willebrord Snel, who had discovered
this refraction law. Descartes and Fermats thought about this too. Their
dispute about this is described in Nahins book ”When least is best”. For a more
general density distribution Archimedes proof also gives that g(t) sin(α(x)) is
constant. Archimedes proof is more powerful: it leads to a result for
nonsmooth g(t).

AN INITIAL VALUE PROBLEM. With the assumption that a particle moves without an influence of an external
force and minimize the action, we are lead to a dynamical system. Start at a point P and a direction v. The
extremization requirement leads to a Newton law, which is a differential equation of the form ẍ = f (x, ẋ, t).
One can actually derive all of Newtons law from a minimization principle. Extremization of action is one of the
most important principles in physics: Newton equations, Maxwell equations, Einstein equations can be derived
like this.
WAVE FRONTS AND CAUSTICS Math118, O. Knill

ABSTRACT. Wave fronts which start at a point evolve and break at caustics. Given a metric in Euclidean
space, the wave fronts form a one-parameter family of piecewise smooth surfaces.

WAVE FRONTS AND CAUSTICS. The set of points reached at

time t from a given point x form the wave front Kx (t) of x. If the
geodesics starts with an initial velocity (cos(ϕ), sin(ϕ), it reaches
at time t the point Kx (t, ϕ). A conjugate point of x is a point
Kx (t, ϕ), for which DKx (t, ϕ) has zero determinant. The set Cx
of all conjugate points K(t, ϕ) form the caustic of x. The caustic
of a curve ϕ 7→ r(ϕ) in the plane is defined as the set Kγ of points
for which DKγ (t, ϕ) has zero determinant, where Kγ (t, ϕ) is the
point reached when we start at r(ϕ) in the normal direction n(ϕ).
Given a closed compact surface and a point P . How does the wave
front K( t) look like? Does it become dense on the surface?

EXAMPLES.

FLAT TORUS. On the flat torus, the wave front Kx (t) becomes
dense on the surface for every point x. The caustic is empty. The
picture to the right shows the wave front on the flat torus at time
3.

ROUND SPHERE. The wave front Kx (t) is a circle or a point at

all times. In the case of the flat torus, the caustic is empty, in the
case of the sphere, the caustic Cx consists of two points, x and
the antipole S(x) of x.

CAUSTIC FLAT CASE. Let γ : r(ϕ) = (x(t), y(t) be a curve in

the flat plane and let n(ϕ) = (−y ′ (ϕ), x′ (ϕ) be the normal vector
to the curve and ρ(ϕ) = 1/||n(ϕ)|| = 1/||r′ ||. Then Kγ (t, ϕ) =
r(ϕ) + tn(ϕ)ρ(ϕ) [ = (x(ϕ) − ′ty(ϕ)ρ(ϕ,′ y(ϕ) + tx(ϕ)ρ(ϕ)′ so that ]
DK γ (t, ϕ) = n(ϕ)ρ(ϕ) r (ϕ) + tn (ϕ)ρ(ϕ) +[tn(ϕ)ρ (ϕ) =]
n(ϕ)ρ(ϕ) r′ (ϕ) + tn′ (ϕ)ρ(ϕ) = 1/ρ + tρ2 (ϕ) n(ϕ) n′ (ϕ)
[ ]

using det(⃗a, ⃗b +⃗a) = det(⃗a, ⃗b). The caustic of the curve γ is called
the evolute of the curve.

EXAMPLE: Locally, we can represent a plane curve as a

graph (x, f (x)).
√ The wave front W (t, x) = (x, f (x)) +
t(−f ′ (x), 1)/ 1 + f ′ (x)2 has the caustic

{(t, x) = (1 + f ′ (x)2 )3/2 /f ′′ (x), x) } .

For example, for f (x) = x2 , we have {W ((1 + 4x2 )3/2 /2, x)} =
{(−4x3 , 1/2 + 3x2 ) } which is essentially the graph of y = x2/3 .
For f (x) = x4 , we have {W ((1 + 16x6 )3/2 /(12x2 ), x) } = {(2x/3 −
16x7 /3, 7x4 /3 + x−2 /12) }.
THE MIRROR EQUATION. If P and Q are successive points on P
a caustic for a geodesic ray which is reflected at the boundary
point M with curvature κ and impact angle θ, then f = |P − M | Q
and e = |Q − M | satisfy
1 1 2κ
f + e = sin(θ) f
e
PROOF. The change of the incoming angle dθ1 a and the outgoing
ray dθ2 is related by dθ2 = 2dθ − dθ1 . The claim follows from
dθ = 1/ρ = κ, dθ1 = sin(θ)/f, dθ2 = sin(θ)/e. M
Interpretation: If P = x is a point, then Q is a point of the differential geometrical caustic Cx of the point x.
If you light a flashlight at P , then the point Q will be a focal point, where the light density is strong.

THE COFFEE CUP CAUSTIC. If r(t) = (− sin(t), cos(t)) is the boundary of the cup and light enters in the
direction (−1, 0), then the impact angle θ is just t. The curvature κ(t) is 1. Parallel light coming from the
right focuses at infinity so that 1/f = 0. The light which leaves into the direction (cos(2t), sin(2t)) focuses after
reflection at a distance e = sin(θ)/(2κ) = sin(θ)/2. The caustic is therefore parameterized by (− sin(t), cos(t)) −
(cos(2t), sin(2t)) sin(t)/2 = (− sin(t) + cos(2t) sin(t)/2, cos(t) + sin(2t) sin(t)/2). Image credit for the picture to
the right: Henrik Wann Jensen 1996.

CAUSTICS OF BILLIARDS. The word ”caustic” has different

meaning in billiards and in differential geometry. Caustics can be
defined for any family of light rays. In differential geometry, one
looks at all the light rays which are emitted at one spot or all light
rays emitted orthogonally to a given curve. If we look at all the
light rays emitted from a point x in a billiard table, we will see
caustics too. The differential geometrical Cx will be dense however
in general. In billiards, we have looked at the caustic of a family
of rays which correspond to billiard trajectories on an invariant
curve. However, there are some cases, where there is a direct
connection between differential geometrical caustics and caustics
of billiards. We can deform a sphere in such a way that the caustic
of a point on the sphere is the caustic of a special billiard table.
We have used this construction once to find metrics on spheres for
which the caustics is nowhere differentiable.

CAUSTICS OF BILLIARDS. Caustics

of billiards can be quite complicated.
To the right, we see some examples for
billiards in tables of equal thickness.
GEODESICS Math118, O. Knill

ABSTRACT. Light moves on shortest paths. The corresponding dynamical system is called the geodesic flow.
We will see examples of geodesic flows which are integrable like the flow on a surface of revolution. This is an
introduction to geodesic flows without Riemannian geometry which allows to go straight to the essential math
without too much formalism.

ARCHIMEDES THEOREM. We have seen that the shortest distance between two points in Euclidean space is
the line. We have proven this in the case of the plane without use of derivatives. This ”Archimedes proof” can
be generalized to higher dimensional Euclidean spaces too.

DEFINITION. Given a smooth surface in space, a point P on the surface and

initial tangent velocity vector v. Define a path on the surface by letting a
particle move freely in space under the influence of a force perpendicular to
the surface in such a way that the particle stays on the surface. This defines
a path on the surface called geodesic flow. This dynamical system can be
described using differential equations too. However, for many of the examples
considered here, we can work with the intuitive notion. If the surface has a
boundary, then we have a surface billiard. In that case, we assume the mass
point bounces off the boundary according to the usual billiard law.

The force F (x, v) perpendicular to the surface at the point x to the direction v
can be computed by intersecting the plane spanned by the unit normal vector
⃗n and the vector v with the surface, leading to a curve with a radius of
curvature r. Applying the centrifugal force F (x, v) = |v|2 n/r assures that
the particle stays on the surface. The number κ(x, v) = 1/r(x, v) is called the
sectional curvature at the point in the direction v.

MOTIVATION. The numerical method, we used to compute the geodesic flow

on some of the pictures on this page is a mechanical one. We constrain the
free motion onto the surface. Given a surface X in space we look at the free
evolution of the particle subject to a strong force which pulls the particle to the
surface. That force is always perpendicular to the surface and so perpendicular
to the velocity of the particle. Especially, it does not accelerate the particle.
Do a free evolution in space for some time dt, then projection the vector back
onto the surface. X(u, v) → X(u, v) + V → X(u1 , v1 ) This method is so
efficient and simple, that we have let the ray-tracing program (Povray) do all
the computation for the pictures.

EXAMPLE: GEODESICS ON THE SPHERE.

On a sphere, the mass-point is at any time subject to a force which goes through
d d
the center of the sphere. Angular momentum conservation dt L = dt x×v =0
implies that the particle stays on a plane spanned by the normal vector and
the initial vector v. The geodesic curve is the intersection of the plane with
the sphere: it is a grand circle. The plane can be seen as a limiting case of the
sphere, when the radius goes to infinity. A particle which initially is on a plane
and has a velocity tangent to the plane stays on the plane without any need of
constraint. The geodesic curves consist of lines.
EXAMPLE: GEODESICS ON SURFACE OF REVOLUTION. If
ϕ is the angle between a longitudinal line and the geodesic curve
and r is the distance from the axes of rotation, then the angular
momentum L = r sin(ϕ) is conserved. It is called the Clairaut
integral. Examples of surfaces of revolution are the cylinder, the
cone or the torus. If we write the torus as part of the plane with
a space dependent metric which depends only on one coordinate,
we have a geodesic flow on a surface of revolution. The Clairaut
integral r sin(ϕ) is the analogue of Snells integral g(x) sin(α) we
have seen before.

METRIC AND DISTANCE. Consider a two-dimensional parametrized surface (u, v) 7→ r(u, √ v). At a point
(u, v, r(u, v), we have the tangent vectors dx = ru du, dy = rv dv The distance element ds = [ dx · dx + dy · dy ]
ru · r u r u · r v
satisfies ds2 = (ru du+rv dv)2 = ru ·ru dudu+ru ·rv dudv +rv ·ru dvdu+rv ·rv dvdv. With g = ,
rv · ru rv · rv
√
this can be written as ds2 = (du, dv) · g(du, dv). A new dot product < a, b >= a · gb and length ||a|| = < a, a >
∫b
allows to write the length of a curve as a ||r′ (t)|| dt. Riemanns view is to start with a two dimensional surface
M and a symmetric matrix at each point gij (x, y) defined so that both eigenvalues of g are positive everywhere.
The pair (M, g) defines a Riemannian manifold. One can measure distances on it without referring to the
ambient space in which the surface is embedded.

EXAMPLE: GEODESICS ON THE FLAT TORUS. Because a

region in a flat torus can be seen as a region in the plane, geodesics
on the flat torus are made of lines. With gij = 1 if i = j and
gij = 0 if i 6= j as in the case of the plane, the differential equations
for the geodesics are ẍk = Γkij ẋi ẋj = 0. There is no acceleration.
The fact that the shortest connections between two points A, B
on the flat plane are straight lines can be seen in different ways.
The straight line gives a distance between the two points as we
have seen before in the plane.

EXAMPLE: HILLY[REGION. Let r(u, ] v)[ = (u, v, f (u, v)) ]

be a parameterization of the graph of f . The
r u · r u r u · rv 1 + fu2 fu fv
metric is g(u, v) = = . So, if r(u(t), v(t)) is a curve on the surface,
rv · r u r v · r v fu fv 1 + fv2
we can calculate its length. We should get the same result as if we would compute the length of the curve
r(t) = (u(t), v(t), f (u(t), v(t))) in three dimensional flat space. But with the internal formalism, it is possible
to compute the length without using the third dimension.

CONNECTION. When minimizing the length of a curve, we have to find the Euler Lagrange equations. This
involves differentiating the metric g further. The Christoffel symbols are defined as
1 ∂ ∂ ∂
Γijk = [ gjk (x) + gik (x) − gij (x)] .
2 ∂xi ∂xj ∂xk
For a parametrized surface, this is
Γ111 = ruu · ru , Γ112 = ruu · rv Γ211 = rvu · ru , Γ212 = rvu · rv
Γ121 = ruv · ru , Γ122 = ruv · rv Γ221 = rvv · ru , Γ222 = rvv · rv
FREE MOTION ON A SURFACE. A particle of momentum p has the Lagrangian F (t, x, p) = 12 gij (x)pi pj .
We use Einstein summation convention to automatically sum over pairs of lower and upper indices. We
∫b ∫t
want to minimize I(x) = a F (t, x, ẋ)dt = t12 gij (x)ẋi ẋj dt With Fpk = gki pi and Fxk = 21 ∂x∂ k gij (x)pi pj and
∂ i j 1 ∂ i j i i j ij
the identities 12 ∂x j gik (x)ẋ ẋ = 2 ∂xi gjk (x)ẋ ẋ gki ẍ = −Γijk ẋ ẋ and the definitions g
−1
= gij , Γkij := g lk Γijl
this can be written as

ẍk = −Γkij ẋi ẋj

Because F is time independent, H(p) = pk Fpk − F = pk gki pi − F = 2F − F = F (p) is constant along the orbit.

√ √ ∫t √
GEODESICS ON A SURFACE With G(t, x, p) = gij (x)pi pj = 2F , the functional I(γ) = t12 gij (x)ẋi ẋj dt
d
is the arc length of γ. The Euler-Lagrange equations dt Gpi = Gxi can using the previous function F be written
F
d √p i xiF d d
as dt 2F
= √2F Which means dt Fpi = Fxi because dt F = 0. Even so we got the same equations as for the free
motion, they are not equivalent: a reparametrization of time t 7→ τ (t) leaves only the first equation invariant
and not the second. The distinguished parameterization for the extremal solution is proportional to the arc
length. The relation between the two variational problems for energy and arc length is a special case of the
Maupertius principle.

EXAMPLE: GEODESICS ON THE HYPERBOLIC PLANE. This is an example, where the surface is not given
as an embedded surface in R3 . Instead, we assume that the distance on the upper half plane H is given by the
formula ∫ b√
ẋ(t)2 + ẏ(t)2
L(γ) = dt .
a y(t)

THEOREM. On the hyperbolic plane, geodesics between two points P, Q is the

circle through P, Q which hits the x axes in right angles.

PROOF. For points P = (x, a), Q = (x, b) with the same x coordi- 8
∫b
nate, the distance is d(P, Q) = a y ′ (t)/y(t) dt = | log(b/a)|. The
geodesic connection is a line. Now see H as part of the complex 7

plane and note that Moebius transformation 6

(az + b) 5
T (z) =
(cz + d)
4

with ad − bc = 1 maps circles to circles or lines is an isometry: 3

d(P, Q) = d(T (P ), T (Q)). Indeed, the two formulas Im(T (z)) =
Im(z)/|cz + d|2 and d/dtT (z(t)) = z ′ (t)/|cz + d|2 imply 2

b b 1
d/dtT (z(t)) z ′ (t)
∫ ∫
dt = dt .
a Im(T (z(t)) a Im(z(t)) -4 -3 -2 -1 1 2 3 4

To see that a Moebius transformation preserves circles, note that one can write T as a composition T = T2 IT1 ,
where T1 (z) = cz +d, T2 (z) = a/c+(ad−bc)z/c and where I(z) = 1/z is the inversion at the unit circle. Because
all three transformations preserve circles also A circle through the origin is maped into a line. If a, b, c, d are
real, then T maps the upper half plane onto itself.
CHAOTIC GEODESIC FLOW. We have seen that the cat map T (x, y) = (2x + y, x + y) is integrable and
harmless on the plane. You have computed in a homework an integral, a function F (x, y) which is invariant
under T . When projecting the map onto the torus R2 /Z 2 , then chaos happens. We have seen that the map
allows a description by a symbolic dynamical system. Especially, it is chaotic in the sense of Devaney. A similar
thing happens when we look at the geodesic flow on the upper half plane H. The orbits are circles. Even so
you have sensitive dependence on initial conditions (as you can see in the picture above that if you start with
different direction from the same point, the trajectories separate fast). We can do the analogue of the torus
construction on the hyperbolic plane: take a discrete subgroup Γ of the group of all Möbius transformations.

For example Γ could be the subgroup of Möbius transformations

with integer entries. It is called the modular group. An other
subgroup is the modular group lambda Λ of all transformation
T (z) = (az + b)/(cz + d), where a, d are odd integers and b, d are
even integers. The equivalent region to the square in the case of
the torus is the fundamental region H/Λ which is displayed to
the right. Billiard trajectories move on circles, when hitting the
the boundary z of the region they enter at an other place γ(z)
similar than Pacman does for the torus. The corresponding flow
is chaotic for any known notion of chaos.
-1 1

THE DOUGHNUT. The rotationally symmetric torus in space is

parameterized by

r(u, v) = ((a + b cos(2πv)) cos(2πu), (a + b cos(2πv)) sin(2πu), b sin(2πv)) ,

where 0 < b < a. The metric is

g11 = 4π 2 (a + b cos(2πv))2 = 4π 2 r2
g22 = 4π 2 b2
g12 = g21 = 0

so that length of a curve is measured with the formula

∫ b
4π 2 (r(u(t), v(t)))2 u̇2 + b2 v̇ 2 ) dt .
a

The circles v = 0, v = 1/2 are geodesics as are all the circles

u = u0 . The surface is rotationally symmetric and one has the
Clairaut integral.

HOPF-RYNOV THEOREM ETC. The geodesic flow is defined for all times for closed complete surfaces without
boundary. On every point on the surface and in any direction, there exists exactly one geodesic curve. Every
geodesic subsegment of a geodesic curve is a geodesic curve. The shortest path between two points on the
surface is a geodesic. But as the sphere shows, not every geodesic is the shortest path (you might go into the
wrong direction on the grand circle). If two points are close enough, then the shortest geodesic connecting the
two points is the shortest curve.

REMARKS. It is not custom to define the geodesic flow by constraining the free flow to the surface. But it is a
useful fact and used for proving the integrability of the geodesic flow on the ellipsoid. The construction works in
general: the Nash embedding theorem assures that any Riemannian surface can be embedded isometrically
in an Euclidean space.
CONCLUSIONS Math118, O. Knill
ABSTRACT. We summarize the main points of this course and add some didactical comments.

DYNAMICAL SYSTEMS. While the notion of dynamical systems can be defined in much greater generally,
all dynamical systems considered here were either given by a map T on space X or by a differential equation
ẋ = f (x).

MATHEMATICAL STRUCTURES. The space X can carry different structures. It can be topological, mea-
sure theoretical, combinatorial, geometrical or analytical. Stressing the topological structure leads to
topological dynamics, using an invariant measure reaches out to probability theory or ergodic theory, the
geometrical structure is involved when dealing with differentiable functions and subject to differential geom-
etry. Combinatorial structures come into play, when doing symbolic dynamics, when dealing with complexity
or counting issues. The analytic structure is involved when the map can be extended to the complex, crossing
the boundary to complex analysis, algebraic geometry or potential theory.

Topic Examples Key points

the subject has relations with
dynamical systems semigroup action virtually any field of mathemat-
ics
periodic points and their bifurca-
1D dynamics quadratic map, Ulam map tions, conjugation, Lyapunov ex-
ponents
horse shoe construction, stability
2D dynamics Henon map, Standard map of periodic points, stable and un-
stable manifolds, Jacobean
van der Pool equation, linear sys-
2D differential equations Poincare-Bendixon
tems
Poincare return map, Hopf bifur-
3D differential equations Lorentz system cation, Lyapunov function, frac-
tals
variational principle to construct
billiards polygons, ellipse, stadium
orbits, effect for chaos
elementary 1D automata, life, topology of sequence space, at-
cellular automata
lattice gases tractor special solutions
Newton method, stability of pe-
complex dynamics quadratic maps riodic points, conjugation to nor-
mal form
graphs from forbidden words,
Baker map, full shift, Fibonacci
symbolic dynamics symbolic dynamics in general
shift, even shift
system
continued fraction expansion,
irrational rotation, maps on fi-
dynamics in number theory dynamic logarithm problem, dy-
nite sets
namical systems from curves
Sitnikov, restricted planar 3- integrals, horse shoe construc-
celestial mechanics Kepler,
body problem tion, rotating coordinate systems
plane, sphere, surfaces of revolu- surface billiards, integrals, caus-
geodesic flow
tion tics, calculus of variations
Some of main points I wanted to make in this course:

• Even deterministic systems lead to unpredictable or uncomputable situations.

• Some systems allow explicit solutions, other systems remain mysterious.
• The history of dynamical systems often sits at the heart of the history of mathematics or science.

• The subject has connections with many other fields of mathematics.

• Dynamical systems theory has many applications.
• There are many open problems left in the area of dynamical systems.

DIDACTICS. We have covered a lot of different topics. One could teach this course with the material from the
first or second week, but in more depth. That would make sense too. I personally think that in a time where
knowledge is accumulated at a tremendous speed, it makes sense to be trained also in the process of acquiring
a lot of knowledge in a short time. Equally important is the ability to solve not so straightforward problems
and to find creative solutions.
WHAT DID WE LEAVE OUT? Each of the topics could be extended to a full course. Important fields, which
have not been touched at all: partial differential equations and systems in fluid dynamics in particular, systems
with higher dimensional time as they appear in statistical mechanics, dynamical systems of algebraic origin. A
large area for dynamics is also game theory or the theory of neural networks. Then there are problems
of statistical flavor which deals with the problem to find the laws of the dynamical system from data. A
particular case in statistics is to recover the space X the transformation T as well as the measure µ which
produces the data. Finally, there are quantum versions of many dynamical systems considered so far. For
billiards or surface billiards, the quantum problem is the study of the Laplacian on the surface with Dirichlet
boundary conditions. The eigenvectors of the Laplacian vn in the limit n → ∞ have connections with the
billiard or geodesic flow on the surface. Quantum dynamical systems can be obtained reformulating things
first on a function space. For a topological dynamical system (X, T ), consider the space X̃ = C(X) of all
continuous functions on X. The map T induces a linear map T̃ on X̃ by ˜(T )(f )(x) = f (T (x)). Allowing more
general spaces X̃ C ∗ algebras allows the study of quantum versions. Also measure theoretical systems (X, T, µ)
can be reformulated in function space. Instead ofT̃ (f ) = f (T ) on all bounded measurable functions, consider
the dynamics of a general unitary operator or more generally an automorphism on a von Neuman algebra. Also
geometric structures have been ”quantized” leading to a subject called ”noncommutative geometry”. The topic
of perturbation theory, which is used for example to prove the persistence of stable motion (KAM) or the
existence of homoclinic points (Melnikov theory). Finally, there is spectral theory, the study of the unitary
operator Ut f = f (Tt ) on L2 (X, µ) for a map or flow T preserving a measure µ.

KNOWLEDGE VERSUS CREATIVITY. Even special areas of dynamical system theory have fragmented. It
is relatively easy to be creative, when ignoring knowledge. It is much harder to find new results in the context
of what is known. The right balance has to be found. In a first stage of research, avoiding the literature might
be a good idea since too much information can be deadly for creative work. But after having figured out a way
to solve the problem, looking up the literature is a necessity to face the possibility that a result has been proven
already, maybe a special case of a much more general result. In that ”library stage”, a lot of information has to
be processed in a short time. In a time, where patent offices pass sometimes requests which have a long time
been ”prior art” and in the public domain, some effort to pass some of the information which is available in
books, in databases or papers to the brain has been made. Fortunately, technology softens some of the need to
know vast amount of information. Still, most information is not online, nor in text books, not even in recent
papers. The challenge is to balance two different but equally important things:

Acquire: process, absorb and learn information Inquire: question, generate new ideas and solutions

HOMEWORK: most homework questions seemed have been just at the right level of difficulty a few were a
notch too hard. I think most of the homework problems could not be solved without spending a few hours each
week.

QUIZZES: the weekly quizzes tested knowledge and presence in the classroom. They also served as a tool to
gauge, how the information have been absorbed during lecture.

PROJECTS: Most papers had a high standard. The topics ranged from history of dynamical systems to
summaries. There was one project, which dealt with an unsolved problme: The exterior billiard project
confirmed empirical evidence that the semicircle is an unstable billiard. Here are some project titles chosen for
the end project:

The evolution of a universal grammar in the case of super symmetry differential equations
HIV and Immune Respose Dynamics differential equations
A Survey of Results Concerning the Collatz 3x + 1 conjecture discrete dynamics
Examples of Experimental Mathematics Number theory
Exterior billiards on the Semicircle 2D maps
Continuous Newtonis Method differential equations
Gravity in One Dimension n-body problems
Stability in lineary stochastic systems summary
A Survey of the History of Celestial Mechanics from Aristotle to Poincare 3 body problem
A dynamical systems view of the leaders dilemma game theory
SOME OPEN PROBLEMS Math118, O. Knill
ABSTRACT. We summarize some open problems in dynamics. Most have been mentioned
during this course. One possibility for a project is to write down a short essay about such a
problem.

STABILITY OF EXTERIOR BILLIARDS. Does there exist a billiard table, for which the
exterior billiard features an unbounded orbit?

Discussion. A semicircle had been mentioned as a candidate for unstability. One doesn’t
know the answer also for general polygonal billiards. Smooth strictly convex tables produce
bounded orbits (KAM). Some ”rational polygons” produce bounded orbits too.

INTEGRABLE EXTERIOR BILLIARDS. Does there exist an integrable smooth convex

exterior billiard map different from the ellipse? Integrability means that there should exist a
smooth function F for which each set {F = c} outside the table is a curve.

Discussion. The problem could be changed by allowing F to be continuous only or requiring

F to be realanalytic. For some polygons like the sqare, the billiard is integrable.

MEASURE OF PERIODIC POINTS OF BILLIARDS. Can a smooth billiard table have a set
of periodic points which have positive area?

Discussion. It is known that periodic points of period 2 or 3 have zero area. See Rychliks
paper of 1998.

MEASURE OF PERIODIC POINTS OF EXTERIOR BILLIARDS. Can a smooth strictly

convex exterior billiard table have a set of periodic points which have positive area?

Discussion. I have not seen that question asked and the answer could be easier to find than
in the billiard case. Some polygons like the square have a full set of periodic orbits.

MEASURE OF NONCOLLISION SINGULARITIES. what is the measure of noncollision

singularities of the Newtonian n-body problem.

Discussion. This is an old problem appearing for example in Simons list of unsolved prob-
lems for the 21’st century: The difficulty to figure out all possible configurations leading to
noncollision singularities suggest that a completely different approach is necessary.

INTEGRABIILITY AND DYNAMICAL LOGARITHM PROBLEM. Does nonintegrability (in

the sense that no smooth invariant function F exists) imply that the dynamical logarithm
problem can not be solved efficiently? Specifically, for the nonintegrable map
[ ] [ ]
x x + 2α
T = .
y x+y

on the torus. Assume α = π, find n such that T n (0.5, 0.5) is within distance 10−1000 of (0, 0).

LATTICE POINTS NEAR PARABOLA. For every 0 ≤ δ < 1, there exists a positive constant
C such that the number M (n, δ) of 1/n-lattice points in a 1/n1+δ neighborhood of a parabola
satisfies M (n, δ)/n1−δ → C for n → ∞.

Discussion: we know that the result is true for δ < 1/3. We suspect it is true for δ < 1.
WAVE FRONTS AND CAUSTICS ON THE TORUS. Is it true that for a general metric on
the two dimensional torus (a bumpy doughnut), the wave front of a point becomes dense on
the torus? Is it true that for a nonflat metric, the caustic of a point is dense on the torus?

Discussion. This seems completely unexplored and might be not too difficult. It would
be enough to verify that the wavefront Kt contains longer and longer pieces with shrinking
curvature.

HILBERTS 16th PROBLEM. Find upper bounds on the number of limit cycles for a differential
equation on the plane. A concrete problem: verify that a differential equation
d
x = p(x, y)
dt
d
y = q(x, y)
dt
with polynomials p and q of degree n has only finitely many limit cycles. Find a bound for
their number H(n).

Discussion: this is an old and probably very difficult problem.

CELLULAR AUTOMATA. Find closed formulas for the topological entropy of each of the 256
elementary CA in one dimension.

Discussion: there are some automata, for which we know the answer, like for rule 90, the
shift. For subshifts of finite type, the entropy is log(κ), where κ is the largest eigenvalue of an
integer matrix.

SPEED SPECTRUM OF CA. Given a CA in the plane like life, one can define the speed
spectrum as the set of possible speeds, gliders can have. If T n (x) = σ m (x), then the velocity
is v = m/n and the speed |m|/v. The speed spectrum of a cellular automaton is a discrete
subset of [0, d], where d is the diameter of the CA rule. Find a higher dimensional automaton,
where the speed spectrum can be proven to be dense in [0, d].

Discussion: I don’t think, this is known in the specific example of ”Life”. There are many
gliders known. ”Life” should be rich enough to have a desne speed spectrum.

ENTROPY OF STANDARD MAP. Does the Standard map have a positive average Lyapunov
exponent for large λ. More specifically, is it true that for all n,
1 ∫ 2π ∫ 2π
log ||A(T n−1 (x, y) · · · A(x, y)|| dxdy ≥ log(c/2)
n 0 0
where T (x, y) = (2x + c sin(x) − y, x) and A(x, y) = dT (x, y).

Discussion. The problem had been posed in the sixties by Sinai. One knows that the Lyapunov
exponent is positive on horse-shoes which form Cantor sets, but these sets have zero measure.

BOUNDED ORBITS IN SITNIKOV PROBLEM. What is the measure of the set of initial
conditions for which the Sitnikov planet stays bounded?

Discussion: there is a continuum of bounded solutions constructed by the horse-shoe con-

struction. The problem is already difficult for maps in the plane, which are explicitely given
like the area preserving Henon map.
STABILITY OF SOLAR SYSTEM An ancient problem. Is the solar system stable? More
precisely: is the n-body problem with the initial condition given by the position leading to a
solution in which orbits stay in a bounded region for all times? How big is the set of initial
conditions, for which one has stability?

Discussion. One knows that the measure of initial conditions leading to stability is positive,
but one knows no estimates for the size of the stability region.

NONCOLLISION SINGULARITIES WITH 4 BODIES. Does Gervers construction lead to

noncollision singularity with four bodies?

Discussion. Gervers construction works in the plane. Xias construction needed a lot of
mathematical resources. It is likely that also this problem is very difficult.

CONSTRUCTING LATTICE POINTS CLOSE TO PARABOLA. Can one solve the dynami-
cal logarithm problem efficiently for for maps obtained from polynomials.

Discussion: if it were, one could factor integers fast.

LYAPUNOV EXPONENT FOR DOUBLE PENDULUM. Is the Lyapunov exponent positive

for the double pendulum (equal mass and equal length of the legs). The same question can be
asked for other Hamiltonian systems for which the flow leaves a bounded region invariant like
the Henon-Heils system.

Discussion. There are many Hamiltonian systems of two degree of freedom, where chaos is
observed. While it is often possible to construct a Cantor set on which the dynamics is chaotic,
the question is whether one can establish this kind of motion on a set of positive measure.

LOCAL CONNECTIVITY OF MANDELBROT SET. is the Mandelbrot set locally connected?

Discussion. This is the holy grail in complex dynamics.

CHAOTIC SMOOTH CONVEX BILLIARD. Find a smooth convex billiard table for which
the Lyapunov exponent is positive on a set of positive area.

Discussion. One could try to smoothen out the stadium but any of these attempts produces
stable periodic orbits which distroy ergodicity. Also invariant curves near the boundary will
prevent ergodicity. One has to prove coexistence of stable and unstable behavior.

CHAOTIC THREE BODY PROBLEM. Is the Lyapunov exponent positive on a set of positive
area for some restricted three body problem.

Discussion. By constructing horse shoes, one can get Cantor sets of zero area, for which the
Lypapunov exponent is positive.

NORMALITY OF PI. Are the digits of Pi uniformly distributed?

√ In other words, is π normal.
The same question can be answered for other numbers like 2.

Discussion. Some doubts have been raised that π produces good random numbers:
http://news.uns.purdue.edu/UNS/html4ever/2005/050426.Fischbach.pi.html
 {
x/2, x even
3N+1 PROBLEM. Is 1 → 4 → 2 the only attractor of the Collatz map T (x) = .
3x + 1, x odd
Discussion. The mathematical tools seem not to catch. The best hope for success is probably
to find a periodic orbit different from the trivial one. Heuristic arguments show however that
the ”chance” for success is small. As higher periodic orbits you look for, they are more and
more ”unlikely”.

A LATTICE POINT PROBLEM. The map T (x) = (3/2)x mod 1 is related to a lattice point
problem for a function on the real line. It is conjectured that the orbits of T are uniformly
distributed modulo 1.

Discussion. For T (x) = 2x mod 1, the distribution is uniform for most initial points x.

HOMOCLINIC POINTS AND INTEGRABILITY. Assume that a smooth map T on the

plane has a hyperbolic fixed point with a transverse homoclinic point. Prove that there is no
smooth function F which is invariant for which {F = c} is either a curve or a finite set.

Discussion. This might be relatively easy to settle. The existence of a transverse homoclinic
point produces a horse shoe as we have seen. The function is constant on that horse shoe as
well as on the union of the invariant stable and unstable manifolds. While the horse shoe has
dimension bigger than 0, it could be part of a complicated level set which is a union of arcs.
But unlike realanalytic functions, smooth function can have complicated level sets.

BLANCHARD PROBLEM. Does every transitive automaton have a dense set of periodic
points?

Discussion: An automaton T is called transitive, if it has a dense orbit in X. We have seen

that the shift is transitive. We also have seen that the shift has a dense set of periodic points.
Francois Blanchard writes: ”The answer, positive or negative, is a necessary step before on
understands the meaning of chaos in the field. ” Source: This problem can be found in Michael
Misiurewicz list of open problems in dynamical systems (http://www.math.iupui.edu/ mmisi-
ure/open)

COHEN MAP. Does the hyperbolic point of period 14 found by Hubbard in the map
√
T (x, y) = ( x2 + 1 − y, x)

have stable and unstable manifolds which intersect transversly?

Discussion: I had been asked as an undergraduate to make experiments with this map to
see whether it is integrable or not - and did not find any signs of non-integrability. Marek
Rychlik told me in 1998, that numerical experiments by John Hubbard revealed a hyperbolic
periodic orbit of period 14: (x, y) = (u, u) with u = 1.54871181145059. The largest eigenvalue
of dT 14 (x, y) is λ = 1.012275907. The existence of a hyperbolic point of such a period makes
integrability unlikely since homoclinic points might exist, but it is not impossible. It is difficult
to find other hyperbolic periodic points. An other indication for non-integrability is a result
of Rychlik and Torgenson who have shown that this map has no integral given by algebraic
functions.
PROJECTS Math118, O. Knill

ABSTRACT. We give you project guidelines and show a list of projects. The format is quite
free. An important part of the project to choose a topic, to choose a format, to gauge whether
it is reasonable or not. The project has to do with one of the topics we covered in class.

1 Project guidelines
PROJECT FLAVOURS:

• Read an article and summarize its content. Check first, whether the mathematics is
manageable. We have articles of all difficulty levels.

• Explore some dynamical system with the computer. You probably should be familiar
with the programming language, or computer algebra system, you use.

• Review a book on dynamical systems of your choice. You probably should have looked
at the book already by now.

• Understand and reformulate a proof of a difficult theorem. Examples are given below.

• Work on an unsolved problem and solve it (;-). More seriously: it is enough to write an
exposition about the problem, or do some experiments which make it plausible to the
reader that there is something interesting going on.

SCOPE. Plan 2 days of intense work on the project. Chose your topic so that you can finish it.
Gauging whether the project is realistic is part of the task. So, choosing a computer project
only makes sense, if you have experience with the software, you want to use. Choosing a book
review only is realistic if you have read part of the book. If you choosing a paper to summarize,
it is good if it connects to some of your interests or other course work. Beside the projects
presented here, you are free to choose anything on your own, as long as it is tied to dynamical
systems.

DEADLINE. The project is due on May 21 2005. It can be handed in earlier. It is advisable
to show me the project before handing it in. Start to write early! Directly type in your notes
like a diary and polish at the end. Time management can be one of the biggest challenges in
any project.

FORM. I strongly recommend LATEX, especially for mathematical content. There is a template
on the course web-site. The use of Latex might slow you down for a few hours at first, but you
will make up any minute later on.
All our notes for this course have been written in LATEX.

LATEXalso makes it easy to structure the paper. References and referals are taken care of
automatically, the structure with title, abstract etc are all prewired.

GRADING. There will be 4 grades for each paper: originality, correctness, style and presenta-
tion. Unlike in SAT essays (NYT article of today), length is pretty irrelevant. The originality
does not mean that the paper has to contain an original result. An original thought, an original
experiment, an original question can give full score. The presentation and style components
not only looks at the form of the paper, illustrations etc. but also how easy it is to read and
how attractive the paper is overall.
LENGTH. Anything between 2-10 pages is ok. One strategy is to write a lot, then compress
and trimm things. Mark Twain once said:

”I didn’t have time to write a short letter, so I wrote a

long one instead.”

2 Project suggestions: a bazaar

CONTINUED FRACTIONS. Continued fractions and the Gauss map. Give a proof that
dx/(1 + x) is the invariant measure of the Gauss map on the unit interval. Then prove Propo-
sition 15.3.3 - 15.3.5 in the book of Katok-Hasselblatt.

QUATERNIONIC JULIA SETS. State definitions of the analogue of the Mandelbrot set or Julia
set when iterating maps on the quaternions instead of the complex numbers. The quaternionic
fractals are objects in 4D and usually presented as three dimensional slices. Explain the problem
and make some pictures.

DIFFERENTIAL NEWTON METHOD. The differential Newton method. Read research paper
of John Neuberger which had appered in the Intelligencer. Unlike in the discrete case, the basins
of attraction have smooth boundaries.

INTEGRABILITY IN DYNAMICAL SYSTEMS. Look at Audins article on integrability in

dynamical systems.

IRRATIONAL GROWTH IN LIFE. Study irrational life. Read the article of William Geller
and Michael Misiurewicz on irrational life.

THE FEIGENBAUM STORY. The Feigenbaum fixed point. What is renormalization. What
does the theorem say? There is Mathematica code by Marek Rychlik.

DISCRETE APPROXIMATIONS OF CHAOTIC MAPS. Look at Lanfords article on orbit

structure of discrete approximations to chaotic maps. This could lead to your own experiments
with discretizations.

THE LATEST ABOUT THE LORTENTZ ATTRACTOR. What is known about the Lorentz
attractor. Read and understand the article of Viana.

VIRAL DYNAMICS. Read some papers on Viral infections or a chapter of Novaks book.

CELLULAR AUTOMATA WITH RULES DEPENDING ON MACROSCOPIC VARIABLES.

Investigate experimentally a cellular automata, where the CA rule changes depending on global
properties. Take for example life with three different rules and switch rule if the global density
changes.

PERIODIC POINTS OF BILLIARDS. The problem of periodic points in billiards appears in

different places. The case of triangular billiards where one does not know whether there are
always periodic points or the case of smooth billiards where one does not know whether there
can be sets of positive measure. The project could also include exterior billiards.

STABILITY OF EXTERIOR BILLIARD. Exterior billiards in semi circle. Make some experi-
ments and form an opinion whether the table is stable.
INDECOMPOSABLE CONTINUA IN DYNAMICAL SYSTEMS. Read a paper Judy
Kennedy, ”How indecomposable continue arise in dynamical systems” and how it ties in with
our course.

WOMEN IN DYNAMICAL SYSTEMS. Write an expository article about contributions of

women in dynamical systems theory. In each example, the relevant mathematics should be
described. Examples: Sonja Kovalevskaya, Mary Cartwright, Krystina Kuperberg, Emmy
Noether, Lai-Sang Young, Bodil Branner, Erika Jen, Linda Keen, Jane Cronin Scanlon, Michele
Audin.

WAVE DYNAMICS IN RELATIVISTIC SETUP. General relativity in 1+1 dimensions. How

can wave fronts look like in an inhomogenous medium, where the distance is given by a metric
with signature +-.

THE CODING OF THE CAT MAP. Write a careful exposition about the coding in the cat
map T (x, y) = (2x + y, x + y). Can one do the encoding faithfully with three sets as we have
seen in the homework?

CONTINUED FRACTIONS AND MUSIC. Continued fractions in music. The article had been
distributed in class.

CHAOS IN THE SOLAR SYSTEM. Read and summarize the paper of Kirchgraber and Stoffer:
Chaotic motion proof of comets. There are also some books about chaos in the solar systems
which could be part of the project.

THE 3n+1 PROBLEM. 3n+1 problem. Exclude certain periodic patterns using Mathematica.
Example: (3x + 1)/4 = x implies x = 1. The problem to relate cycle length with where the
cycle is is related to continued fraction expansion of log(3)/ log(2).

A LATTICE POINT PROBLEM. Study the map T (x) = (3/2)x mod 1. This problem is a a
lattice point problem for a function on the real line. It is conjectured that the orbits of T are
uniformly distributed modulo 1. Run some statistical experiments to check this and find out
what is known about it.

THE KAM THEORY. KAM theorem. Track down what the KAM theorem is and where it is
used. A possibility is to focus on the twist map theorem and see what it means for maps like
the Standard map, billiards or exterior billiards or some elliptic periodic points.

WAVE FRONTS IN 3D. Visualize wave fronts in 3D emanating from a two dimensional surface.
Find examples, where one knows the caustic as in 2D. To plot surfaces, ttart with a parametrized
surface, draw the normals and plot all points in a distance d from the surface. This can be a
computer graphics project. A more mathematical task would be to find the equation for the
caustics and visualize the caustic.

POINCARES BLUNDER. Formulate the mathematics of the prize problem which King Oscar
II of Sweden had posed. There is a Book of June Barrow-Gree, Poincare and the three body
problem, AMS 1996 you can borrow.

ARNOLDS THEOREM IN INTEGRABLE SYSTEMS. Arnolds theorem on integrability.

Write down a proof of Arnolds theorem telling that if a Hamiltonian system of n degrees
of freedom has n independent involutive integrals, bounded solutions must lie on tori.
REAL QUADRATIC MAP SURVEY. Summarize Lubichs survey article on real quadratic
maps.

DYNAMICAL SYSTEMS SURVEY. Summarize one of Lai-Sang Youngs survey articles on

dynamical systems.

SHARKOVSKI THEOREM. What does it say? There is a nice proof in the book of Brin-Stuck.

HOMOCLINIC TANGLE IN THE STANDARD MAP. Visualize the homoclinic tangle in the
Standard map. Adapt the code for the Henon map to the Standard map.

DOUBLE PENDULUM SYSTEM. Experiment with the double pendulum system. Mathemat-
ica code is available. Plot some Poincare sections which are area preserving maps. Alternatively,
one could look at nonlinarly coupled penduli.

MANDELBROT SET. HISTORY AND OTHER SETS. Write down the story of the discovery
of the Mandelbrot set. Alternatively, look at other Mandelbrot sets like the map fc (z) =
z 2 sin(z) + c.

BETA EXPANSION. Explore the invariant measures of the beta expansion. (Book of Dajani
and Kraaikamp).

NEWTON PROBLEM WITH CONSTANT INTERACTION FORCE. The Newton problem

in one dimensions. Simulate 3 particles on the line with constant force interaction. This is a
model for three galaxies.

IS PI NORMAL. Find out whats about this recent indication that π is not normal. Alterna-
tively, there is an article of Bailey,Borwein,Borwein and Plouffe to read. The paper should be
availble now Journal of Modern Physics C, vol. 16, no. 2.

FIXED POINTS IN DYNAMICAL SYSTEMS. Write an essay on ”fixed points in dynamical

systems”. Particularly interesting is Poincares last theorem which is mentioned in the book as
well as Browers fixed point theorem or the Newton method which is used to prove the existence
of the Feigenbaum fixed point.

COMPLEX HENON MAPS. Read an article on Complex Henon maps and state some questions
one can ask when iterating maps in the two dimensional complex space C 2 .

LYAPUNOV EXPONENTS OF HENON MAP. Lyapunov exponents of Henon map

http://alamos.math.arizona.edu/ rychlik/notebooks.html

LOZI MAPS. Lozi maps are Henon type maps, which are piecewise linear. They are easier
accessible.
VORTEX DYNAMICS Math118, O. Knill

ABSTRACT. The dynamics of vortices is a Hamiltonian system which is relevant in fluid dynamics

VORTEX SYSTEMS. The motion of finitely many vortices is an interesting Hamiltonian system showing in-
tegrable and chaotic motion. It is an example of a Hamiltonian system which is not derived by the Legendre
transformation of a Lagrange system. Point vortices exist in superfluid helium. Vortex motion can simulate
the Euler equations of a two-dimensional incompressible fluid. They describe singular solutions of the two
dimensional Euler equations. The chaotic motion of vortices should shed some light to turbulence in fluid
dynamics. The question, what is the measure of the set of initial conditions leading to a collapse in R2 seems
to be unsolved. It is also unknown if for an arbitrary domain the motion of two vortices is already chaotic. For
most, it probably is.

HISTORY Vortex theory was introduced by Helmholtz one century ago. Kirchoff and Poincaré outlined the
Hamiltonian structure of the system. Kirchhoff 1876 invents the model of finitely many vortices. The integrabil-
ity of the 3-vortex problem was known to Gröbli in 1877 Poincare (1893) and Synge (1949). Often, the system
is also called Kirchhoff problem. He remarks that the equations of a point-vortex motion in an unbounded plane
defines a Hamiltonian system. Lin at Caltech investigates in 1943 the motion of vortices in two dimensions
in arbitrary domain. There is much numerical work on finitely many vortices in the plane giving numerical
evidence that the motion of four vortices is chaotic. Ziglin 1980 argues with a Melnikov perturbation argument
that the problem of four vortices has a horse shoe and is therefore not integrable. Marsden, Weinstein remark
in 1983 that the proof of Ziglin has some gaps and some discussion followed. For an other 4 vortex distribution
show nonintegrability. Recent work [?] give new chaotic regions. The statistical mechanics of a gas of vortices
in the thermodynamic limit has been studied by Ruelle and Fröhlich. Using KAM theory Khanin showed that
in the phase speac of any system with an arbitrary number of vortices,s there exists a set of initial conditions
of positive measure for which the motion of vortices is quasi-periodic. Celletti and Falcolini [?] have (using
computer assisted proofs) quantitative results for the break-down of invariant tori in the four vortex problem.

VORTEX DYNAMICS. The motion of N vortices zi with vorticity ωi in the complex plane is given by the
differential equations

d 1 ∑ ωj
ωk · zk =
dt 2πi zk − zj
j̸=k

which is a Hamiltonian system

∂H ∂H
ωk ẋk = , ωk ẏk = −
∂yk ∂xk
for the variables xi + iyi = zi with the Hamiltonian
1 ∑ 1 ∑
H(x, y) = − ωi ωj log(|zi − zj |) = − ωi ωj log(|zi − zj |) .
4π 2π i<j
i̸=j

One can also write shorter

d 2 ∂H
ωk · zk =
dt i ∂zk
with
∂ 1 ∂ ∂
= ( − ).
∂zk 2 ∂xk ∂yk
The phase space of the system is
D = CN \ ∆ ,
∪
where ∆ = 1≤i≤j≤N {zi = zj } is the set of collisions.
INTEGRALS. The N vortex system has 4 classical integrals of motion:
∑
∑= i ω2i zi
a) the center of mass Z = Q + iP
b) the angular momentum I = i ωi |zi | .
c) the energy H.

These integrals come from the invariance of the differential equations by space translations, space rotations
and time translations.
∑
PROOF. a) Every term in i ωi żi appears twice with opposite sign.
b) Define ∑
S= ωk zk ż k
k

Because I and I˙ are real, we have I˙ = S + S.

1 ∑ ω k ω j zk 1 ∑
Ṡ = = ωk ωj
2πi z k − zj 2πi
k̸=j k̸=j

is purely∑imaginary and so I˙ ∑
= S + S = 0.
c) Ḣ = i Hxi ẋi + Hyi ẏi = i Hxi Hyi ωi−1 − Hyi Hxi ωi−1 = 0.

POISSON BRACKET. Define the Poisson bracket

N
∑ 1 ∂f ∂g ∂f ∂g
{f, g} = ( − )
ω
i=1 i
∂x i ∂y i ∂y i ∂xi

∂ 1
for two smooth functions f, g : C N → C. Using the notation ∂z = ∂z = 2 (∂x − i∂y ) and the scalar product
< u, v >= Re(u · v) on C, we can also write
∑ 4
{f, g} = < ∂zi f, i∂zi g > .
i
ωi

The equations of motion are then żi = {zi , H} or

ẋi = {xi , H}, ẏi = {yi , H} .

Because the integral are invariant under the motion, one has

{I, H} = {Z, H} = 0 .

Of course one can build also other integrals by combining the known ones. So, also
1 2 1
J := |Z| = (P 2 + Q2 )
2 2
is an integral.

One can show that H, I, J have pairwise vanishing Poisson bracket. Indeed, this implies that a system of three
vortices is integrable. The explicit integration can be found for example in [?]. The threshold for chaotic
behavior in a point-vortex system starts with N = 4. Indeed already the restricted four vortex system, where
the vortex strength of the forth vortex is zero is already is not integrable [?]. The fourth vortex is in this case
just a particle of the fluid.

GLOBAL EXISTENCE.
SYNGE THEOREM. If all the vorticities ωi > 0, and z ∈ D, then the solution of the vortex flow exists for all
times.
In the case when all ωi > 0, all the solutions in the phase space are bounded, because
∑
I= ωi |zi |2 = const
i

defines a compact set. The flow is then defined for all times since H = const prevents collisions.
SELF SIMILAR SOLUTIONS. In general, when ωi can take both signs, the existence and uniqueness problem
is nontrivial since the logarithmic divergence of the Green function may generate catastrophes. There are
self-similar initial conditions of a three vortex situation leading to a collapse in finite time: √
Let z1 = (−1, 0), z2 = (1, 0) be the initial position of two vortices with vorticity ω1 = ω2 = 2 and z3 = (1, 2)
be the initial position of a vortex with vorticity ω3 = −1. Write ai for the length of the sides of the triangle
formed by three vortices and denote with A its area. One has
d 2 2ωk A 1 1
ak = ( 2 − 2 ).
dt π ak−1 ak+1

The ratios of the sides are conserved and d 2

dt ak = − 3√12π a2k (0). This gives explicit Ely
√
t
ak (t) = ak (0) 1− √
3 2π
√
and a collaps at time t = 3 2π.

Using KAM theory, Khanin showed that in the phase space of any system with an arbitrary number of vortices,
there exists a set of initial conditions of positive measure for which the motion of vortices is quasi-periodic.
Celletti and Falcolini have even quantitive results for the break down of invariant tori in the four vortex problem.

RESTRICTED FOUR VORTEX PROBLEM. The three vortex problem is integrable and one expects chaos for
the motion of four vortices. The restricted four vortex problem is obtained by taking an integrable three
vortex motion and letting evolve a fourth vortex of zero vorticity in the stream lines of the others. The next
lemma part c) is due to Novikov-Sedov 1978,Aref-Pomphrey:
a) {I, H} = {P, H} = {Q, H} = 0
b) {I, Q} = −P, {I, P } = Q}
c) {I, J} = 0

PROOF.
d 1 ∑ ωj
(P + iQ) =
dt 2πi zk − zj
j̸=k

changes sign when replacing j with k. Because the sum does not change by this permutation, it must vanish.
To the Hamiltonian I belongs the differential equation
d ∂
zk = I = zk
dt ∂zk

which induces a rotation zk (t) = eit zk . Because the Hamiltonian H is invariant under this flow, we have
{H, I} = 0. This gives also {I, H} = 0 which means that I is an integral.
b) We get ∂zi Z = ωi , ∂zi I = ω2i z i and so
∑ 4
< ωi , iωi zi >= −iZ
i
ωi

giving {I, P } = Q, {I, Q} = −P .

ARBITRARY DOMAINS. The motion of vortices in arbitrary domains. Let D be a 2-dimensional manifold
which can have a boundary. Let g(x, y) be the Green function of the Laplacian with Dirichlet boundary
conditions. This means that g(x, y) solves the Poisson equation ∆x g(x, y) = δ(y) and g(x, y) = 0 for x ∈ δD or
y ∈ δD and where ∆x is the Laplacian on the manifold. Write zi = xi + iyi for the local coordinates of a point
vortex on D and write J∇i = 2∂zi = ∂xi − i∂yi .
The motion of the vortices is defined by the differential equation
d ∑
ωi z i = J∇i H = J∇i ωi ωj g(zi , zj ) .
dt
i̸=j
∑
With ω = ω1 , ω2 , . . . ωN ), z = (z1 , . . . , zN ) and ∇ = (∇1 , . . . , ∇N ) and H(z) = i̸=j ωi ωj g(zi , zj ), this can be
written shortly as ω ż = J∇H(z).
EXAMPLES.

(1) On the plane D = R2

1
g(x, y) = − log |x − y|
2π
(2) On the torus D = T 2 = R2 /(2L)2 one has
1 ∑ 2πi
g(x, y) = − e L k·(x−y) .
(2πL)2
k∈Z 2 \{0}
∑
One must require i ωi = 0.

(3) On the cylinder D = R × T with T = R/(2L) one has

1 π
g(x, y) = − log sin( (x − y)) .
2π L
∑
Also here, one requires i ωi = 0. Famous is the van Karman vortex street which a 2 vortices system of
opposite vorticity on the cylinder.

(4) On the sphere D = S 2 , one gets

g(x, y) = log(1 − x · y) ,
where x, y ∈ S 2 ⊂ R3 have the dot product x · y.

(5) On the two dimensional disc D = {|z| ≤ R} one gets

1 |x − y|R
g(x, y) = − log( )
2π |y||x − y|

where y = R2 y/|y|2 is the conjugate point of y.

For the Hamiltonian one has then
1 ∑ |zi − zj |R ∑∑
H=− ( log( zj ) + log(1 − |zi |2 ) ,
4π |zi ||zi − |zj |2 |
i̸=j i

where the second sum is the energy of the selfinteraction of the vortices with its mirror vortices.

For general vortex strengths, vortices can collide also in regions. Dürr and Pulvirenti proved in 1982 that the
set of initial conditions leading to the collapse of two or more vortices has Lebesgue measure zero provided the
vortices are in a compact domain or located on the torus.

LITERATURE:

Paul K. Newton, The N-Vortex Problem, Analytical Techniques, Springer, New York, 2001.
PARTICLE IN DIPOL FIELD (*) Math118, O. Knill

ABSTRACT. A single particle in a magnetic dipol field leads to a Hamiltonian system of two degrees of freedem.
For many energies, the energy surface has compact components which allow a Poincare section leading to an
area preserving map on a compact region in the plane. This map is the composition of two noncommuting
integrable maps. The motion of a charged particle in a magnetic dipol field is so shown to be an example of
a chaotic dynamical system. It describes particles moving in the Van Allen belts of the earths magnetic field.
Their interaction with the atmosphere is responsible for the northern lights (aurora borealis) or southern lights
(aurora australis).

A RELATIVISTIC SYSTEM. A = (0, 0, Az ) = (0, 0, M ρr−3 ) is a vector potential in cylinder coordinates

(ρ, φ, z). Here ρ is the distance from the z-axes, r = r(ρ, z) is the distance from the dipole center 0 and M is
the magnetic dipole moment of the field.

This vector potential generates a rotational symmetric magnetic field B = curl(A) = (0, −∂Az /∂ρ, 0). The
Hamiltonian system describing the motion of a single charged particle of charge q, mass m and velocity v in the
magnetic dipole field B is
)1/2
K = m2 c4 + c2 p2z + c2 p2ρ + c2 (pφ /ρ − qAz )2 = γmc2 ,
(

where c is the speed of light, γ = (1 − v 2 /c2 )1/2 , and (pz , pρ , pφ ) = γm(ż, ρ̇, φ̇ρ2 + qρAz ).

A NONRELATIVISTIC HAMILTONIAN. Because v 2 and so γ are constants of motion, one can introduce a
new Hamilton function
1 ( 2 ) 1
H= pz + p2ρ + (pφ − qAz )2 = γmv 2 ,
2γm 2
where (pz , pρ , pφ ) = γm(ż, ρ̇, φ̇ρ) − qρ(0, 0, Az ). The equivalence of the two Hamilton function K and H can be
seen by observing that the partial derivatives with respect to all dynamical variables agree.

Because H is invariant under the one-parameter group of rotations along the z axes one has φ̇ = −∂H/∂φ = 0
and pφ = qM Γ, where Γ is a constant having as dimension an inverse length.
After elimination of φ, we describe the motion of a particle in the ρ − z plane in the potential V :

q2 M 2 Γ
( )
ρ
V (ρ, z) = − 3 .
2γm ρ r

A TWO DEGREE OF FREEDOM SYSTEM. After introducing dimensionless variables z ′ = Γz, ρ′ = Γρ, φ′ =
φ, t′ = Γ3 qM γm−1 t and leaving away the apostrophes, the Hamilton function becomes
1 2 1 1 ρ 1
H= (p + p2ρ ) + ( − 3 )2 = ,
2 z 2 ρ r 32γ14
1 qM 2 4
where γ1 is called the Störmer constant γ14 = 16 ( vγm ) Γ .

THE POTENTIAL. The potential V vanishes on the path

ρ2 = r3 . This path is the minimum of the elsewhere positive
potential.

For γ1 > 1, the allowed region of the particle is not connected. We

are especially interested in the motion of the compact connected
region. The motion on the non-compact region is integrable by
an observation of Moser, the asymptotic velocities of the escaping
particles being integrals of motion.
THE STÖRMER PROBLEM. The Störmer problem is to analyse the two degree of freedom Hamiltonian system
with Hamiltonian
1 1 1 q1
H(q1 , q2 , p1 , p2 ) = (p21 + p22 ) + ( − 2 )2 .
2 2 q1 (q1 + q22 )3/2
The energy surface H(q, p) = E with 0 < E < 1/32 contains a compact component on which the flow is area
preserving. The two dimensional hypersurface q1 = 0 in the energy surface is a Poincaré surface and the ruturn
map is a symplectic map.

STUDIING A MAP. The second iterate of this return map can be written as a composition of integrable twist
maps. The first map is to shoot the particle from q2 = 0 to the north (with p2 > 0) and wait until it comes
back. The second map is to shoot the particle from q2 = 0 to the south and wait until it comes back to the
equator q2 = 0. Both maps are twist maps in the plane which have a single fixed point which is the initial
condition which shoots the particle into the dipole. The two fixed points of the two maps do not agree, the two
maps don’t commute.

When shooting from the south pole towards the north pole, the particle will not bounce back to the south pole.

AURORA BOREALIS.
The photos of the north-
ern lights to the right are
done by Jan Curtis.

LITERATURE. For the description of the Hamiltonian system we followed [?].

LINEAR MAPS ON DISCRETE TORI (*) Math118, O. Knill

ABSTRACT. When studing linear maps or curves on finite tori Y = Zm1 × · · · × Zmn , we are in the ream of
elementary number theory.

a) Iteration of modular linear maps.

The map T (⃗x) = A⃗x mod m ⃗ defines a dynamical system on the finite group Zm1 × ... × Zmn . Since the discrete
torus Y does not match with the torus X , orbits on this finite set behave rather irregular. The system can be
extended to the real torus R/(m1 Z) × R/(mn Z), where it is in general a hyperbolic map. The orbits behave
differently, if A is very singular, for example if A has only one column.

EXAMPLE. The map [ ] [ ] [ ]

x 31x + 34y 7
T = mod
y 3x + 38y 17
has 6 different orbits on Y with a maximal orbit length of 49. It seems difficult to find ergodic examples with
different moduli where ergodic means that there is only one orbit besides the trivial orbit of ⃗0 = (0, 0) a case
which appears for example in [ ] [ ] [ ]
x 18x + 5y 37
T = mod
y 7x + 14y 37
Homework 1. Week Math118, O.Knill

1.1 We consider the interval map f (x) = f4 (x) = 4x(1 − x). For this particular
value, the logistic map is also called the Ulam map. We have met it in the
first lecture, when we saw that a computer does not ”know” the distribu-
tative law.

a) Find all the fixed points of the map T (x) = f (x).

b) Analyze the stability of these fixed points. For each point, just tell,
whether it is stable or unstable.

c) Draw a graph of this map and start iterating the map using the cobweb
construction with the initial value 0.3. Do at least 5 iterates.

1.2 Consider the map Qc (x) = x2 + c. It is called the quadratic map. Again,
c is a constant parameter.

a) Verify that this map undergoes a saddle node bifurcation (which

is also called blue-sky bifurcation because of obvious reasons, periodic
appear or disappear out of the blue sky. It is also called tangent bifurca-
tion). For which value of c does this happen?

b) Analyze the stability of the periodic orbits near the bifurcation value.

c) What happens with the orbits for parameter values c for which we have
no fixed point?

1.3 a) Look at the fixed points of the map Qc (x) = x2 + c for c = −0.5 and
determine their stability.

b) Look at the fixed point of the map for c = −1 and determine its stability.

c) Verify that the map undergoes a flip bifurcation at the parameter

c = −3/4.
1.4 We define the map f (x) = 5x + sin(πx) mod 1 on the interval [0, 1].

Sideremark: Because after identifying 0 and 1, the interval closes to a circle,

the map can be considered a smooth map on the circle. f is an example of
a circle map.

a) What is the Lyapunov exponent of the orbit of the map f wich an initial
condition x0 = 1/2?

b) Verify that the Lyapunov exponent of every orbit of f is positive.

1.5 We have shown that the Ulam map f4 (x) = 4x(1 − x) is conjugated to the
tent map T (x) = 1 − 2|x − 1/2|
a) Draw the graph of the the iterates T 2 (x), T 3 (x) of the tent map. Use the
fact that the tent map is piecewise linear.

b) Use the conjugation result to sketch the graphs of the second iterate
f42 (x) and third iterate f43 (x) of the Ulam map.

c) Conclude that f4n has 2n fixed points and therefore, that the Ulam map
f has 2n periodic points of period n.

d) What is the Lyapunov exponent of a periodic point of period n?

2. Homework set Math118, O.Knill

2.1 a) Realize the Henon map

T (x, y) = (y + 1 − ax2 , bx)

as a second order difference equation. A second order difference equation

is a recursion of the form xn+1 = F (xn , xn−1 ).

b) An orbit x0 , x1 , x2 , ... of a difference equation is called periodic, if there

exists an integer n such that xk+n = xk for all k.
Verify that periodic points of the difference equation define periodic points
of T .

c) Find a periodic point of prime period 2 of the Henon map in the case
a = 1, b = 1. The map is then

T (x, y) = (1 − x2 + y, x)

. The notion prime period 2 means that it should not be a fixed point.

d) (This is optional. Do it only if you have time and access to a CAS.)

Can you find formulas for period 2 orbits for general a, b. You might need
a computer algebra system. If you use a computer algebra system, find
also all periodic orbits of the Henon map with prime period 3 and 4. The
formulas can get messy.
2.2 a) Analyze the stability and nature of all the fixed points of the cubic Henon
map T (x, y) = (cx − x3 − y, x) depending on the parameter c.

b) Find the bifurcation points, which are parameter values, where the
stability of one of these fixed points changes.

2.3 Consider the map

T (x, y) = (2x + 3y, x + 2y)
on the torus.

a) Is T is area preserving?

b) Verify that the fixed point (0, 0) of T is hyperbolic. What are the stable
and unstable manifolds of this fixed point?
 c) Find the Lyapunov exponents of each orbit of T as well as the entropy,
which is the average of the Lyapunov exponent.

d) Argue, why T (x, y) = (2x + 3y + ϵ sin(x), x + 2y) has homoclinic points

for small ϵ.

Remark. No formal proof is required in d). Just explain in words.

2.4 a) Compute the Lyapunov exponent of fixed points of the Henon map
T (x, y) = (1 − x2 + y, x).

b) Compute the Lyapunov exponent of the periodic orbit you found in [2.1c].

c) What is the Lyapunov exponent of an initial point on the stable manifold

to the periodic point you found in b)?

Remark. No computation is necessary in c).

2.5 a) Prove that the cat map T (x, y) = (2x + y, x + y) on the torus is not
integrable.

b) Show that the cat map T (x, y) = (2x + y, x + y) defined on the plane is
integrable.

Remark: It is possible to give an explicit integral for T but it is also possible

just give arguments.
3. Homework set Math118, O.Knill

3.1 a) Is the flow generated by the differential equation

d d
x = −2 sin(x + 2y) − x sin(xy), y = sin(x + 2y) + y sin(xy)
dt dt
area-preserving?

b) Find a function H(x, y), such that the differential equation can be rewritten as
d d
x = Hy (x, y), y = −Hx (x, y)
dt dt

c) Is the system integrable?

3.2 a) Assume that ẋ = F (x) is a differential equation defined in an annulus A = {1 <

(x2 + y 2 ) < 4 } and assume that A is left invariant under the differential equation.
Assume that div(F )(x, y) < 0 everywhere in the annulus. Prove that there can exist
maximally one cycle in A.

b) Assume that ẋ = F (x) is a differential equation defined in the disk D = {x2 + y 2 <
1 }. Assume that this disk is left invariant under the differential equation. Assume that
div(F ) < 0 everywhere in the disk. Prove that there can not exist any limit cycle in D.

3.3 a) Verify Dulacs criterion: assume ẋ = F (x) is a differential equation in a region D

of the plane. If there exists a smooth function g, such that div(gF (x)) has no zeros in
D, then there are no closed cycles in D.

b) Use Dulacs criterion to show that

d
x = x(2 − x − y)
dt
d
y = y(4x − x2 − 3)
dt
has no closed cycles in the region D = {x > 0, y > 0}.

Hint. This is hard to guess: try g(x, y) = 1/(xy).

3.4 a) The clycolytic oscillator is a model for the biochemical process glycloysis:
d
x = −x + ay + x2 y
dt
d
y = b − ay − x2 y
dt
The system depends on two parameters a > 0, b > 0. The variable x is the concen-
tration of ADP (adenosine diphosphate) and y is the concentration of F6P (fructose-
6-phosphate). The parameter space is divided into two regions. One region, where
 the fixed point (b, b/(a + b2 ) is stable, the other, where the fixed point is unstable and
where a stable limit cycle exists. Find the boundary between these two regions. When
passsing this boundary, Hopf bifurcations occur.

b) Verify that the differential equation

d
x = y − (x11 − 100x)
dt
d
y = −x .
dt
has a unique limit cycle.

The Klein bottle is an example of a two-

dimensional surface. It can not be realized with-
3.5 out selfintersection in space. Explore whether
the Poincare-Bendixon theorem holds on the
Klein bottle or not.

Hint: you can build the Klein bottle as a square

at which left and right are identified in the op-
posite orientation and top and bottom are iden-
tified with the same orientation. Start by glo-
ing the top and bottom together. This gives a
cylinder. Then, instead of glueing the cylinder
together at the end (which produces a torus),
glue them together in opposite direction.

3.6* (These are unsolved problems and therefore optional).

a) (Dulac problem) Verify that a differential equation
d
x = p(x, y)
dt
d
y = q(x, y)
dt
with polynomials p and q of degree n has only finitely many limit cycles. Find a bound
for their number H(n).

b) (Special case of Hilberts 16’th problem) Show that a Lienard system with g(x) =
x and polynomial F (x) of degree 2k + 1 has at most k limit cycles.
4. Homework set Math118, O.Knill

4.1 a) Prove the following theorem: if F (x, y, z) = (P (x, y, z), Q(x, y, z), R(x, y, z) is a
vector field in space which is divergence free div(F (x, y, z)) = 0, then the differential
d
equation dt ⃗x = F (⃗x) preserves the volume.

Remark. We have done that in two dimensions. You may be able to verify the fol-
lowing
∫ ∫ ∫ formula using Gauss theorem: if D is a region in space then d/dtvol(D) =
D div(F (x, y, z)) dxdydz.

b) Under which conditions on a, b, c does the famous ABC system

ẋ = a sin(z) + c cos(y)
ẏ = b sin(x) + a cos(z)
ż = c sin(y) + b cos(x)

preserve volume?

P.S. By the way, this is a system, where one could expect positive Lyapunov exponent
on a substantial subset of torus. But nobody knows how to estimate this.

4.2 Analyze the Lorenz system

ẋ = σ(y − x)
ẏ = rx − y − xz
ż = xy − bz .

for σ = 0. What can you say about the equilibrium points for the two-dimensional
system if x = s is fixed?

4.3 Reed section 1 and 2 in the booklet ”Chaotic evolution and strange attractors” and
write down, what Ruelle’s view of ”turbulence” and ”chaos” is. (One small paragraph
is enough).

4.4 Verify that the Lorenz system can not have quasi-periodic solutions. These are
solutions which do not close and which cover a two dimensional torus densely.

4.5 Compute the fractal dimension of the Menger sponge. This three dimensional set can
be obtained iteratively as the Cantor set by successively taking away the middle rect-
angular columns of each complete cube. You see the first two steps of the construction
in the picture.
5. Homework set Math118, O.Knill

5.1 Given a rectangle of length 1 and height b > 1. We play billiards in this table. For
which angles θ does a trajectory (which does not hit a corner) get arbitrarily close to
any point on the boundary of the table?

5.2 We have seen that for every period n with prime n, there is a periodic orbit of a bil-
liard. We have done this by by maximizing the length functional H(x1 , ..., xn ), which
is the total length of the closed trajectory. We have assumed that the integer n has
no nontrivial factor, because we did not want to have a periodic orbit of some smaller
period.

a) Can you prove that the result actually holds for any n? There is a periodic orbit of
minimal period n.

b) Prove that there are at least two periodic orbits of period 5 in the table x4 + y 4 ≤ 1.

5.3 Here is an application of the Kronecker dynamical system x → x + α. Consider the

first digits of the powers

1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, ..

Can you determine how often each digit occurs in average? Which digit does occur
more often, the digit 8 or the digit 9?

Hint. If a power 2k starts with the digit 5 then 5 · 10m ≤ 2k < 6 · 10m for some m. Take
logarithms to the base 10 of this equation and watch out for a Kronecker system. You
are allowed to use Weyls theorem without proof, which assures that for irrational α,
the frequency with which [kα] is in some interval [a, b] is equal to b − a. We will prove
that later.

5.4 a) A table is called convex, if the line segment connecting two arbitrary points in the
table is inside the table. Verify that the billiard map can not be continuous on the
annulus (R/Z) × [−1, 1], if the table is not convex.

b) Verify that the billiard in a half ellipse x2 /a2 + y 2 /b2 ≤ 1, y ≥ 0 is integrable.

5.5 The string construction allows to construct a table, with a given caustic.

a) Draw a family of tables, which have a regular triangle as a caustic.

b) Draw a family of tables, which have a regular square as caustics.

c) Is there a convex billiard table which has two different caustics, where each caustic
is a polygon?
6. Homework set Math118, O.Knill

6.1 Let X = {0, 1}Z with the distance d(x, y) = 1/(n + 1) n is the smallest index such that
xn ̸= yn . that is xk = yk for |k| ≤ n and xn ̸= yn or x−n = y−n .

a) To verify that X is a metric space, have to verify d(x, y) = d(y, x), d(x, x) = 0 and
the triangle inequality: d(x, z) ≤ d(x, y) + d(y, z).

Hint: Actually a stronger inequality holds: d(x, z) ≤ max(d(x, y), d(y, z)). If x, y agree
on some interval and y, z agree on an other say bigger interval, then x, z agree at least
on the smaller interval.

b) Verify that X is compact: every sequence x(n) in X has an accumulation point. You
have to show that there is a subsequence x(kl ) such that x(kl ) converges in X for l → ∞.

Hint. You have to construct an accumulation point using a diagonal type argument
(unlike in Cantors case, you do not need to change the diagonal).
6.2 Model an epidemic as a CA. You can use your own model. Otherwise, a suggestion
would be that people can either be healthy, sick and non-contagious or sick and conta-
gious, The disease is quite severe so that sick people close to a contagious sick person
will become sick to. After incubation, they become themselves contagious, then sick
and non-contageous, then sick and contagious and then recover again. Describe the CA
rule φ(a, b, c) which gives the state of the person depending on its own state b, and the
states a, c of two neighbors. Can you invent a numbering which includes all CA of the
type you consider, analogue to the Wolfram numbering over the alphabet {0, 1}?

Optional: simulate your CA on the computer.

6.3 a) Construct a CA with radius 1 over the alphabet {0, 1, 2, 3, 4, 5} for which the speed
c satisfies is c/R = c = 2/5.

Remark. If you would generalize your construction to any alphabet, you could proceed
to prove that the set c/R of speed/radius ratios of one-dimensional cellular automata
is dense in the interval [0, 1]. Since one can simulate any automaton with alphabet A
with an automaton with alphabet {0, 1} but with an adjustment of the radius, one can
see that also in general, the possible speed ratios c/R are dense in the interval [0, 1].
6.4 a) Verify that the CA T over the
∩ alphabet A = {0, 1, 2} defined by φ(x, y, z) =
xyz mod 3 has an attractor K = n T n (X) on which the the map is isomorphic to an
elementary CA. (An elementary CA is a CA in one dimensions over the alphabet {0, 1}).

b) (optional) The CA T over the alphabet A = {0, 1, 2} defined by φ(x, y, z) = xyz +

1 mod 3 shows ”particles”. If two particles interact, they annihilate. Can you verify
this picture?
6.5 Invent a CA in two dimensions of your choice. Either look at a simple one which you can
analyze completely. Or invent one which models some situation (life, forest fire, spread
of passion for dynamical systems etc.) It is possible to do experiments on a computer
if you wish but this is not necessary. Coming up with an interesting suggestion is also
good.
7. Homework set Math118, O.Knill

7.1 The Mandelbrot set M can be defined as the set of parameter values c for which fcn (0)
stays bounded.

a) Show that M is a subset of |c| ≤ 2.

b) Find and explain a nontrivial symmetry in the Mandelbrot set.

7.2 The filled in Julia set Kc are the set of all points z in the comple plane such that
fcn (z) stays bounded. The Julia set Jc itself is the boundary of that ”prisoner set” Kc .

a) Why are all filled in Julia sets of fc (z) = z 2 + c centrally symmetric? The picture
below shows the Douady rabbit.

b) Show that Jc is a compact set.

Hint: Find a radius r(c) such that Jc is contained in Br (0) = {|z| ≤ r}.

7.3 a) If T (z) = p(z) is a polynomial map, find the number of periodic points of period n,
where we count the periodic points with multiplicity and do not require the periodic
points to be of minimal period.

b) How many periodic points of minimal period 2 do you expect in general for a cubic
map T ?

c) Show that every cubic polynomial T can be conjugated to fa,b (z) = z 3 − 3a2 z + b by
a linear conjugation.

Hint. If c1 , c2 are the critical point of T , then conjugate with a translation S(z) = z + c
so that the new critical points are centrally symmetric a, −a. Then conjugate with
S(z) = dz so that the coefficient of z 3 becomes 1.
7.4 De Moivres formula

(cos(nz) + i sin(nz)) = z n = (cos(θ) + i sin(θ))n

shows that cos(nz) can be written as a polynomial in cos(θ). (Just look at the real part
of both sides of this identity and use that sin2 (θ) = 1 − cos2 (θ).

a) Find the Chebychev polynomials T1 (z), T2 (z), T3 (z). What is the Julia set of each of
the map Tk (z)?

b) Verify that each line segment in the complex plane through 0 which is centrally
symmetric is the Julia set of some quadratic polynomial.

7.5 We want to understand why the Julia set of the Newton method applied to f (z) = z 3 −1
must be complicated. Remember that the T (z) = z −f (z)/f ′ (z) is the Newton method.
The Julia set of T is the set of points which are not attracted to one of the fixed points
of T . It is the boundary between three regions, the attractors of the fixed points.

a) Verify that the fixed points of T are attractive. Compute the Lyapunov exponent
λ(T, x), if x is in the basin of attraction of a fixed point.

b) Show that the basins of attractions as well as the Julia set J are T invariant.

c) Verify that the Julia set is invariant under rotation by 2π/3 in the complex plane.
8. homework set Math118, O.Knill

8.1 Define a subshift X of finite type over the alphabet A = {a, b, c} by forbidding the
words aa, ab, cb.

a) Find a finite set of words, with which you can build any sequence in that subshift.

b) Draw the graph which has as vertices the list of words in a) and as directed edges
the possible transitions between these words.

c) Write down a few of words in the language of this shift.

8.2 The golden ratio subshift is the subshift of finite type over the alphabet {0, 1} for
which the single word 11 is forbidden.

a) Find a list of words, with which you can build any sequence (words of length 1 are
allowed too).

b) Find and draw the graph and the adjacency matrix A which is defined as Aij = 1 if
the word wi wj is allowed in the sequence.

c) The entropy of the subshift of finite type is the defined as log(λ), where λ is the
largest eigenvalue of A. Compute the entropy of the golden ratio subshift. Relate it to
the entropy of the full shift and the entropy of the shift for which both the words 11
and 00 are forbidden.

Remark. The entropy of a shift is a measure on how much information is in a sequence

shift. A shift with low entropy can be compressed well.
∫
8.3 A
∫ map T on the interval [0, 1] is said to preserve the measure dx if f (x) dx =
f (T (x))) dx for any continuous function f . The triple (X, T, dx) is called a measure
preserving dynamical system.

a) Show that T (x) = 2x mod 1 preserves the measure dx. In other words, show that
(X, T, dx) is a measure preserving dynamical system.

b) Verify that T (x) = x + α mod 1 preserves the measure dx so that (X, S, dx) is a
measure preserving dynamical system.

c) If A = [a, b] is an interval in [0, 1], and T is a measure-preserving system, verify that

there exist arbitrary large n such that T n (A) ∩ A has some intersection.
8.4 A function X on ([0, 1], dx) is also called a random variable. The expectation of X is
defined as ∫ 1
E[X] = f (x) dx
0
Together with a measure preserving dynamical system, we get a sequence of ran-
dom variables Xk (x) = X(T k (x)). Two random variables are called uncorrelated,
if E[XY ] = E[X]E[Y ].

a) Verify that for X(x) = sin(2πx) and the dynamical system T (x) = 2x, the random
variables X and X(T k ) are all uncorrelated.

b) Verify that for X(x) = sin(2πx) and any of the dynamical system T (x) = x + α,
some pair of random variables X and X(T k ) are correlated.

8.5 We look again at the cat map T (x, y) = (2x + y, x + y) on the torus Y which we rep-
resent as a square in which opposite sites are identified. Draw the stable and unstable
manifolds of the fixed point (0, 0) until the hit themselves. This defines a partition of Y
into three sets. Doing the symbolic dynamics gives us a map S from the torus Y to the
sequence space X over the alphabet A = {a, b, c}. This map S defines a conjugation of
the cat map T to a subshift (X, σ) of finite type defined by a set of forbidden words of
length 2.

a) Find all forbidden words of length 2 of the subshift. To do so, look at the images of the
three rectangular sets Y1 , Y2 , Y3 . Hint: You can find three rectangles T (Y1 ), T (Y2 ), T (Y3 )
covering the same space then Y1 , Y2 , Y3 . See the right picture below.

b) Find the graph which belongs to the subshift as well as the adjacency matrix A.

c) Find the entropy of the subshift (the logarithm of the maximal eigenvalue of A) and
compare it with the average Lyapunov exponent of the cat map T .
9. homework set Math118, O.Knill

9.1 Given a large number n = pq, where p, q are prime numbers, consider the ”quadratic map”
T (x) = x2 + c modulo n, where c is an integer. The Pollard rho method to factor n looks at
the orbit of a point x. Since it will eventually be periodic modulo q, we have xn = xk mod q
which means that xn −xk has a common factor with n. Find the orbit structure of the dynamical
system T (x) = x2 + 1 mod n. Because we have a finite set, every point is eventually periodic.
This is the reason for the name ρ. An initial point will eventually be caught in a loop. Find all
the periods in the case n = 15.
√
Remark: Assuming the sequence xn to be random, we need to take about q iterates to find a factor
with probability 1/2. If you have q different objects and you chose k objects, the probability that two
q!
are not the same is q(q − 1)...(q − k + 1)/q k = (q−k)!q k . That these probabilities are relatively small is

called the Birthday paradox. If you have a room of 23 people,√then then the probability that two have
the same birthday is 1 − 365!/(342!3652 3) = 0.5072. Note that 365 = 19.11... is close to 23.
√
9.2 a) Find the continued fraction expansion of the golden ratio ( 5 + 1)/2 and relate the periodic
approximations pn /qn with the Fibonacci sequence.
√
b) Find the continued fraction expansion of the silver ratio 1 + 2. Can you find the rule,
which generates pn and qn in the the partial fractions pn /qn for the silver ratio?

9.3 A synodic month is defined as the period of time between two new moons. It is α =
29.530588853 days. The draconic month is the period of time of the moon to return to
the same node. It is β = 27.212220817 days. Intersections between the path of the moon and
the sun are called ascending and descending nodes.

Such an intersection is called a solar eclipse. In one

approximation, it appears in a period of a bit more
then 18 years = 6580 days which is called one Saros
cycle. This cycle and others are obtained from the
continued fraction expansion of α/β. It is said that
Thales used the Saros cycle cycle to predict the solar
eclipse of 585 B.C. The next big eclipse will happen
May 26, 2021. Explain at least two of the following
Eclipse cycles (one of them should be the soros cycle)
with the continued fraction expansion.

cycle eclipse synodic draconic

fortnight 14.77 0.5 0.543
month 29.53 1 1.085
semester 177.18 6 6.511
lunar year 354.37 12 13.022
octon 1387.94 47 51.004
tritos 3986.63 135 146.501
saros 6585.32 223 241.999
Metonic cycle 6939.69 235 255.021
inex 10571.95 358 388.500
exeligmos 19755.96 669 725.996
Hipparchos 126007.02 4267 4630.531
Babylonian 161177.95 5458 5922.999
 If you have no Mathematica installed: turn your browser to

http://sofia.fas.harvard.edu/cgi-bin/sofia
You can get continued fractions by entering something like

ContinuedFraction[Pi, 20] .
9.4 We have seen that a parabola y = p2 (x) = ax2 + by + c defines a dynamical system on the
two dimensional torus. This construction goes as follows: p1 (x) = p2 (x + 1) − p2 (x), p0 (x) =
p1 (x+1)−p1 (x) = α so that p1 (x+1) = p1 (x)+α, and p2 (x+1) = p2 (x)+p1 (x). If xn = p1 (x+n)
and yn = p2 (x + n), then (xn+1 , yn+1 ) = (xn + α, yn + xn ).

The curve f (x) = ax3 + bx2 + cx + d induces a dynamical system on the three dimensional torus.
Find this system and determine whether it preserves volume.
9.5 A widely used data encryption technique goes under the name RSA. The security of this en-
cryption is based on the empirical fact that it is hard to factor large integers n = pq. Some
of the best methods to factor integers goes back to Fermat: assume we can find a second root
y of x2 mod n, then x2 = y 2 mod n so that (x − y)(x + y) = 0 mod n and gcd(x − y, n) is
a factor of n. Finding square roots is difficult directly. The holy grail is to find numbers y
such that z = y 2 mod n is so small that one can factor them. Having enough such numbers
allows to find small squares using a sieaving technique. The Morison-Brillhard √ method starts
with constructing small integers
√ by doing the continued fraction expansion of n. Explain why
the periodic approximation n ∼ pn /qn produces numbers pn for which the square p2n is small
modulo n. How big do you expect these numbers to be?

Remark. Want to earn 20’000 Dollars? You can do so by factoring the RSA-640 which has 193
digits:

n = 3107418240490043721350750035888567930037346022842727545720161948823206
440518081504556346829671723286782437916272838033415471073108501919548529007
337724822783525742386454014691736602477652346609.

See
http://www.rsasecurity.com/rsalabs/node.asp?id=2093
Note that Mathematica has built in factorization
techniques. But typing in

FactorInteger[n]

and waiting will most likely will not earn you the
prize. Unless you are Indiana Jones ...
 10. homework set Math118, O.Knill

10.1 Consider the 3-body problem in space with interaction potential V (x) = ||x||2 /2 and where
particles have mass m. Find explicit solution formulas to this problem.

Hint. Go into a coordinate system in which the center of mass is fixed.

Remark. The natural Newton potential depends on the space. The Harmonic oscillator potential
||x||2 can be considered the natural 0 dimensional Newton potential

dimension potential force

0D Euclidean V (x) = ||x||2 F (x) = −2x
1D Euclidean V (x) = ||x|| F (x) = −x/||x||.
2D Euclidean V (x) = log ||x|| F (x) = −x/||x||2 .
3D Euclidean V (x) = 1/||x|| F (x) = −x/||x||3 .
4D Euclidean V (x) = 1/||x||2 F (x) = −2x/||x||4 .

In general one can compute the natural Newton potential by looking at the solutions of the
Poisson equation ∆V = δ0 which gives the solution V . In Euclidean space as well as tori, or
sphere, this can be solved with Fourier theory.
10.2 The antropic principle is a cheap but effective philosophical explanation for many things. It
was introduced in 1973 by the theoretical physicist Brandon Carter and has been discussed
in popularized in the bestseller of Steven Hawkings ”A short history of time”. The strong
antropic principle answers the question, why a physical law or physical fact holds by demonstrat-
ing that if the law would be violated, then human life would be impossible: no human person
(antropos) could observe it. The principle can be used for example to explain why energy con-
servation is a reasonable physical law: without it, spontaneous runaway processes could produce
an unbounded amount of energy, destroying everything near it. Use the proof of the first Kepler
law to verify that bounded planetary motion in four dimensional space is exceptional and argue
whether the antropic principle excludes a universe with four dimensional Euclidean space (five
dimensional space time).

Remark: You can assume that in d-dimensional space, the natural Newton force is −Gmi mj ⃗r/||x||d ,
where ⃗r is the vector between the two bodies. There are physical theories called Kaluza-Klein
theories which propose higher dimensional space but this is no more in the realm of classical
mechanics.
10.3 We consider in this problem set an n body problem, where particles interact only with their
neighbors. We look at the Toda system which is a famous n body problem which is integrable
and exhibits fancy solutions called solitons. The system is a discretization of the Korteweg de
Vries equation (KdV) ut = 6uux − uxxx .

One can visualize the particles located on a chain. The potential energy has the form
∑
V (q) = f (qi − qi−1 ) .
i

Consider a chain of particles qn = qn+N with mass mi = 1 and with potential f (q) = e−q . The
motion of these particles is given by the differential equations

d2
qn = eqn+1 −qn − eqn −qn−1 ,
dt2
 Verify that after a coordinate transformation
4a2n = eqn+1 −q , 2bn = pn ,
the equations

q̇n = pn ,
ṗn = eqn+1 −qn − eqn −qn−1

go into

ȧn = an (bn+1 − bn )
ḃn = 2(a2n − a2n−1 ) .

10.4 Given an , bn , define the matrices

b1 a 1 0 · 0 aN
 
 a 1 b2 a 2 · · 0 
 0 a2 · · · · 
 
L=
 · · · · aN −2 0 

0 · · aN −2 bN −1 aN −1
 
aN 0 · 0 aN −1 bN
0 a1 0 · 0 −aN
 
 −a1 0 a2 · · 0 
 0 −a2 · · · · 
 
B=  .
 · · · · aN −2 0 
0 · · −aN −2 0 aN −1
 
aN 0 · 0 −aN −1 0
A matrix L is called a Jacobi matrix. Verify that the Toda system is equivalent to the Lax
equations
L̇ = [B, L] = BL − LB .
10.5 We show that for a Lax equations L̇ = [B, L] with B T = −B, the eigenvalues of L are integrals
of motion.

a) Consider the differential equation Ṡ = BS with S(0) = 1 = In in the space of matrices. Show
that SS T = In for all times.
Hint: Ṡ T = S T B T = −S T B.

b) Verify the formula

L(0) = S(t)T L(t)S(t)
by verifying that d/dt(S T LS) = 0.

c) Conclude that the eigenvalues of L are preserved.

d) If the eigenvalues are preserved, then also the trace of L, the sum of the eigenvalues is
perserved. What is the physical meaning of this integral?

e) If the eigenvalues of L are preserved, then also the trace of L2 is preserved. What is the
physical meaning of this integral?
 11. homework set Math118, O.Knill

11.1 Describe the geodesic flow on the one-sheeted hyperboloid. How does a typical geodesic look
like. Find one periodic geodesic.

11.2 When you empty a paper towel roll you obtain a cylinder. On this cylinder you find a spiral
curve. Prove first that this curve a geodesic. Assume the cylinder is x2 + y 2 = 1, 0 ≤ z ≤ 1. We
play surface billiard: the geodesic curve is reflected at the boundaries z = 0, z = 1. Find the
return map T (θ, ϕ) = (θ1 , ϕ1 ), where (x, y, z) = (cos(θ), sin(θ), 0) and ϕ is the impact angle.

Optional: From what you know about billiards, can you cut away part of the cylinder to get a
chaotic surface billiard?

11.3 For a surface of revolution which is symmetric to the xy plane and for which r(z) → 0 for
|z| → ∞ we define a Poincare map for the geodesic flow. Sart with a point x on the surface in
the xy plane and a unit vector v, follow the geodesic flow along the surface until it comes back
to the surface.
This defines return map T (θ, ϕ) on the annulus (θ, ϕ) ∈ T × [0, π]. Describe this map in the case
of the sphere.
11.4 We play surface billiard in the triangular surface obtained by intersecting the unit sphere with
the first octant in space. Prove that this billiard is integrable.

Hint. Remember how you analyzed billiards in the rectangle? A similar idea applies here.

11.5 We play surface billiard in the half cone x2 + y 2 = 1 − z 2 . Prove that there are orbits which
never close.

Hint. Similar than for the flat torus or the cylinder one can find geodesics by cutting up the
surface and flattening it.

11.6 (optional) Can you prove the statement made in class that on a flat torus, the wave front Kt (x)
of a point becomes dense on the torus in the sense, given ϵ > 0, there is a s such that Kt
intersects every disc of radius ϵ for t > s.

Remark: We do not know whether this statment stays true, if you allow the torus to be bumpy.
Nor do we know whether the caustic Ct (x) becomes dense in general. (Ct (x) is the empty set
for the flat torus).
2/2/05 QUIZZ NR 0 Math118, O.Knill

ABSTRACT. This page should give you an idea, how weekly multiple choice quizzes look like.
The quizzes do not any special preparation if you follow the lectures. Several choices in the
multiple choice part are possible. The questions below address two lectures. The lecture today
as well as the lecture on Friday. Next Monday, we already have a quizz of this format.

1) How many midterms do we have in Math 118r?

a) None, we have quizzes.
b) One midterm
c) Two midterms

2) Check whatever belongs to the theory of dynamical systems:

a) A group acting on a set.
b) Predict the future of systems and explore the limitations of this predictions.
c) Compute square roots of real numbers.
d) Understand the iteration of maps.

3) Look at the map T (x) = x2 + x. Which of the following sequences form an orbit of x
through x = 1:

a) 1, 6, 12, 20, 30, ...

b)1,2,5,30,...
c) 0, 0, 0, 0, 0, ...
d)2,5,30,930,...

4) Which of the following dynamical systems have a discrete time? We replace ”map” or ”dif-
ferential equation” with ”system”.

a) Henon system
b) Van der Pool system
c) Standard system
d) Geodesic system
e) Billiard system.
f) Digits of π system.
g) Cellular automata system
5) Which of the following dynamical systems is the Lorentz system
a) ẍ + x + (x2 − 1)y = 0.
b)

ẋ = 10(y − x)
ẏ = −xz + 28x − y
8z
ż = xy −
3
   
x 10(y − x)
c) T  y  =  z + 28x − y  .
z xy − 8z
3

6) What is a semigroup?
a) A set G with an operation ∗.
b) A set G with an operation ∗ such that (x ⋆ y) ⋆ z = x ⋆ (y ⋆ z).
c) A set G with an operation ∗ such that (x ⋆ y) ⋆ z = x ⋆ (y ⋆ z) with a neutral element e
satisfying x ⋆ e = x.
d) A set G with an operation ∗ such that (x ⋆ y) ⋆ z = x ⋆ (y ⋆ z) with a neutral element e
satisfying x ⋆ e = x and such that for every x, there is a y such that x ⋆ y = e.

7) Which of the following sets are semigroups?

a) The natural numbers.
b) The set of words over a finite alphabet with the operation v ⋆ w = vw of putting the
words together.
c) The set of all subsets of a finite set with the operation A ⋆ B = A ∪ B.

8) Which of the following dynamical systems allow a numerical computation of the square root
of 7:

a) T (x, y) =
√ ((x + y)/2, 2xy/(x + y)).
b) T (x) = x − 7.
c) T (x) = x2 + 7.

9) How could dynamical systems theory helped to save lives. Check each which apply:

a) Predict wave heights from the strength of earthquakes triggering tsunamis.

b) Predict the outcome of the lotto.
c) Predict the sector in which the roulette ball falls.
d) Predict the global warming on earth.
2/7/05 QUIZZ NR 1 Math118, O.Knill

Name:

Email:

1) Look at the map T (x) = x2 + 1 on the real line. Which of the following sequences form an
orbit of x with initial condition x0 = 0:

a) 0, 1, 2, 3, 4, ...

b) 0, 1, 2, 5, 26, ...

c) 0, 0, 0, 0, 0, ...

d) 1, 1, 1, 1, ...

2) Which of the following orbits are periodic cycles of the dynamical system?

d
a) (x(t), y(t)) = (sin(t), cos(t)) for the harmonic oscillator differential equation dt
x(t) =
y(t), dtd y(t) = −x(t).

b) The longest diagonal in a convex billiard table.

c) A single alive cell in the game of life.

d) The point 0 in the logistic map T (x) = 4x(1 − x).

3) Which of the following dynamical systems have a discrete time? We replace ”map” or ”dif-
ferential equation” with ”system”.

a) The game of life

b) The Lorentz system

c) The billiard system.

d
d) The harmonic oscillator system dt
x = y, dtd x = −x.
4) There is a sentence attributed to Steven Smale which appears also in the movie ”Jurassic
Park”. The statement is ”The wing of a butterfly in X can produce a tornado in Y a few
weeks later”:

a) X=Rio, Y=Texas
b) X=New York, Y = Los Angeles
c) X=Chicago, Y = New Orleans

5) We have seen the map

T (x, y) = ((x + y)/2, 2xy/(x + y))
to compute the square root of numbers. The number 2xy/(x + y) is called the
a) The geometric mean.
b) The algebraic mean.
c) The harmonic mean.

of x and y.

6) We have seen taht for a computer, iterations of the map T (x) = 4x(1 − x) or the map
S(x) = 4x − 4x2 give different results for T n (x) and S n (x) if n is large and this happened
for identical initial condition x0 . For which n, did we see different results?

a) n = 1
b) n = 10
c) n = 100
2/23/05 QUIZZ NR 3 Math118, O.Knill

Name:

1) One of the following maps is an area preserving Henon map. Which one?

a) T (x, y) = (3(1 − x2 ) − y, 2x)

b) T (x, y) = (2x − y + sin(x), x).
c) T (x, y) = (2x − y + sin(x), x/2).
d) T (x, y) = (3(1 − x2 ) − y, x)

2) For a certain parameters c, the Standard map T (x, y) = (2x + c sin(x) − y, x) mod 1 is
integrable. Which integral verifies this fact?

a) F (x, y) = x − y.
b) F (x, y) = x
c) F (x, y) = x2 + y 2 .
d) F (x, y) = 1.

3) Which
[ 2 of2 the ]following matrices is the Jacobean matrix of the transformation T (x, y) =
(x + y )/2
?
2x2 − y 2
[ ]
x y
a) DT (x, y) = .
4x −2y
[ 2 ]
x /2 y 2 /2
b) DT (x, y) = .
2x2 −y 2

4) The map T of the previous problem is area preserving.

a) True

b) False

5) A map in the plane is called an involution if T 2 = Id, that is if every point is periodic with
period 2. Which of the following statements are true?

a) In general, an involution is integrable.

b) The map T (x, y) = (−x, y + c sin(x)) is an involution.
c) All linear involutions{are area-preserving.
(−x/2, y) x > 0
d) The map T (x, y) = is an involution.
(−2x, y) x < 0
e) In general, an involution is area preserving.
6) Which of the following 6 pictures shows the eigenvalues of a Jacobean at a fixed point of a
map T in the plane which has stable and unstable manifolds:
1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

-1 -0.5 0.5 1 1.5 -1 -0.5 0.5 1 1.5 -1 -0.5 0.5 1 1.5

-0.5 -0.5 -0.5

-1 -1 -1

-1.5 -1.5 -1.5

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

-1 -0.5 0.5 1 1.5 -1 -0.5 0.5 1 1.5 -1 -0.5 0.5 1 1.5

-0.5 -0.5 -0.5

-1 -1 -1

-1.5 -1.5 -1.5

7) If the stable and unstable manifolds of a hyperbolic fixed point (x0 , y0 ) of intersect trans-
versely, then this intersection point is called

a) an equilibrium point
b) an integral
c) a homoclinic point
d) a periodic point
e) a horse shoe

8) Which of the following facts are true about the Henon attractor, obtained with parameters
a = 1.4, b = 0.3?

a) It contains the stable manifold of one of the hyperbolic fixed points.

b) It contains a horse shoe.
c) It contains infinitely many periodic points.
d) It is integrable.
2/28/05 QUIZZ NR 4 Math118, O.Knill

Name:

1) Which of the following differential equations produces an area-preserving flow?

d
a) dt
x = x + y 2 , dtd y = −y + x2
d
b) dt
x = −x + x2 , dtd y = y − y 3
d
c) dt
x = 1, dtd y = 2
d
d) dt
x = y 2 , dtd y = x2

2) Which of the following 6 pictures of eigenvalues a Jacobean DF (x0 , y0 ) at an equilibrium

point (x0 , y0 ) which is stable?
1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

-1 -0.5 0.5 1 1.5 -1 -0.5 0.5 1 1.5 -1 -0.5 0.5 1 1.5

-0.5 -0.5 -0.5

-1 -1 -1

-1.5 -1.5 -1.5

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

-1 -0.5 0.5 1 1.5 -1 -0.5 0.5 1 1.5 -1 -0.5 0.5 1 1.5

-0.5 -0.5 -0.5

-1 -1 -1

-1.5 -1.5 -1.5

3) What happens at a Hopf bifurcation?

a) A pair of eigenvalues crosses the unit circle.

b) A pair of eigenvalues crosses the imaginary axes.
c) A single eigenvalues crosses the unit circle.
d) A single eigenvalues crosses the imaginary axes.
e) An attractive equilibrium point becomes repelling.
4) Which of the following formulations is the Poincare Bendixson theorem?

a) An orbit in the plane which stays in a bounded region is either asymptotic to an equilib-
rium point or to a limit cycle.
b) An orbit in the plane which is not asymptotic to a limit cycle is attracted to an equilib-
rium point.
c) Every orbit of a differential equation in the plane is either asymptotic to a limit cycle or
to an equilibrium point.

d
5) A differential equation of the form dt
x(t) = Hy (x, y), dtd y(t) = −Hx (x, y), where H(x, y) is a
function of two variables.

a) produces an area-preserving flow.

b) is integrable.
c) has an attractive limit cycle.
d) has at least one attractive equilibrium point.

6) Which of the following differential equations is called the van der Pol oscillator ?

d2
a) dt2
x = −x.
d2
b) dt2
x + c(x2 − 1) dtd x + x = 0.
d2
c) dt2
x + c(x2 − 1) dtd x + sin(x) = 0.

7) Lienard systems have

a) exactly one repelling limit cycle.

b) exactly one attracting limit cycle.
c) exactly one repelling equilibrium point.
d) exactly one attractig equilibrium point.
3/7/05 QUIZZ NR 5 Math118, O.Knill

Name:

1) How many equilibrium points does the Lorenz system have in total for r > 1, when
σ > 0, b > 0 are fixed?

a) three.
b) two.
c) one.
d) none.

2) At the parameter r = r1 = 24.74 = 470/19, at each of the two additional equilibrium points
C ± of the Lorenz system, something happens, when the parameter r increases:

a) A sub-critical Hopf bifurcation: an unstable limit cycle collides with the critical point.

b) A Hopf bifurcation: a stable equilibrium point becomes unstable and ejects a limit cycle.

c) A flip bifurcation: the equilibrium point C ± double and undergo a pitchfork bifurcation.

d) A period doubling bifurcation for cycles: periodic cycles double.

3) We define an object in the plane similar to the Shirpinski carpet by cutting away 2 squares of
length 1/4 from a square of length 1 and repeating this construction with remaining squares
of length 1/4 etc: the first three steps are shown below:

What is the dimension of this object?

a) log(14)/ log(4)
b) log(5)/ log(3)
c) log(20)/ log(5)

4) Find the box counting dimensions of the following sets:

 a) The graph of the function f (x) = sin(x) in the plane.
b) A filled triangle.
c) The set {1, 1/2, 1/3, 1/4, 1/5, 1/6, ...}
d) The Cantor set.

5) Which of the following properties does a strange attractor K of a differential equation in

space possess:

a) sensitive dependence on initial conditions.

b) the dimension must be a non-integer.
c) The set K has to be an attractor.
d) The set has to contain a sink, (this is an equilibrium point for which all eigenvalues have
negative real part).

6) Which of the following differential equations in space produces a volume preserving flow?
a)

ẋ = a(y − x)
ẏ = cx − xz − y
ż = xy − bz

b)

ẋ = −(y + z)
ẏ = x + 0.2y
ż = 0.2 + xz − cz

c)

ẋ = y
ẏ = −y − x + x3 − c cos(z)
ż = 1

d)

ẋ = a sin(z) + c cos(y)
ẏ = b sin(x) + a cos(z)
ż = c sin(y) + b cos(x)
3/14/15926... QUIZ NR 6 Math118, O.Knill

Name:

1) The billiard in an ellipse is known to be integrable. What is the integral F ?

a) The sum of the shortest distances of the trajectory line to the focal points.
b) The product of the shortest distances of the trajectory line to the focal points.
c) The impact angle θ.
d) The distance of the focal points.

2) A billiard table γ is obtained by doing the string construction to a convex set K. (For example, in the
homework for today, K was a triangle or square):
Check all which applies:

a) The billiard has a caustic.

b) The billiard map can not have glancing trajectories: trajectories for which the angle θ can become arbi-
trarily close to 0 and arbitrarily close to π. .

c) The billiard map has periodic orbits of period 17.

d) The billiard map has an invariant curve in the annulus R/Z × [−1, 1].

3) For which coordinates is the billiard map area-preserving?

a) The (s, θ) coordinates, where s is the arc length normalized that the table has length 1 and where θ is the
impact angle.

b) The (x, y) coordinates, where x is the arc length normalized so that the table has length 1 and where
y = cos(θ).

c) The (s, s′ ) coordinates, where (s, s′ ) are successive impact points of the trajectory and where s is the arc
length parameter.

4) Every strictly convex smooth Birkhoff billiard has periodic orbits, because

a) We can maximize the length functional of the polygon.

b) We can minimize the length functional of the polgyon.

c) We can maximize the area functional inside a polygon.

d) We can minimize the area functional inside a polygon.

5) Which of the following are open mathematical problems?

a) Every billiard in a triangle has a periodic orbit.

b) Every exterior billiard has the property that for (x, y) outside the table, T n (x, y) → ∞.

c) The solar system is stable in the sense that all planets remain in a bounded region near the sun for all times.

d) There exists a convex billiard for which the Lyapunov exponent is positive on a set of positive area.

e) There exists a smooth convex billiard for which there are no glancing orbits.

∂
6) Which of the following equations is called the Euler equation? We use the notation h1 (x, y) = ∂x h(x, y)
∂
and h2 (x, y) = ∂y h(x, y). Just one answer is correct.

a) h1 (xi−1 , xi ) + h2 (xi , xi+1 ) = 0. e) h1 (xi−1 , xi ) − h2 (xi , xi+1 ) = 0.

b) h2 (xi−1 , xi ) + h1 (xi , xi+1 ) = 0. f) h2 (xi−1 , xi ) − h1 (xi , xi+1 ) = 0.

c) h1 (xi−1 , xi ) + h1 (xi , xi+1 ) = 0. g) h1 (xi−1 , xi ) − h1 (xi , xi+1 ) = 0.

d) h2 (xi−1 , xi ) + h2 (xi , xi+1 ) = 0. h) h2 (xi−1 , xi ) − h2 (xi , xi+1 ) = 0.

7) Which of the following matrices is conjugated to the Jacobean DT (xi , yi ) of the billiard map. li is the length
of the trajectory before the impact with the boundary where the impact angle is θi and the curvature is κi .
[
1 0
] [ ]
1 li
a) Bi = 2κi ·
− sin(θ i ) 1 0 1
[ ] [
1 0
]
1 li
b) Bi = · 2κi
0 1 − sin(θ i)
1
[
1 0
] [ ]
1 −li
c) Bi = 2κi · .
sin(θi ) 1 0 1

8) Today is Pi-day. Somebody cuts a piece from a circular apple pie. We use the remaining part as a billiard
table. Which of the following is true:

a) The table is convex.

b) The table is not convex.

3/21 (Start of Spring!) QUIZ NR 7 Math118, O.Knill

Name:

1) Which of the following formulations is the full content of the Hedlund-Lyndon-Curtis theorem? (only one
answer is correct).

a) Any map T on X = {0, 1}Z which is continuous and commutes with the shift is of the form T (x)n =
φ(xn−1 , xn , xn+1 ).

b) A cellular automaton is a continuous map on X = {0, 1}Z .

c) A shift commuting, continuous map on X = {0, 1}Z is a cellular automaton.

d) Any continuous map on X = {0, 1} has the property that the n’th entry of T (x) depends only on finitely
many neigbors.

2) True or False?

a) There exists a cellular automaton T such that the set of periodic orbits is dense.

a) There exists a cellular automaton T such that the set of periodic orbits of period 11 are dense.

b) There exists a cellular automaton T such that {T n (x), n = 1, 2, ....} covers the entire set X = {0, 1}Z .

c) There exists a cellular automaton T, such that {T n (x), n = 1, 2, ..., } is dense in X.

3) We have T (x)n = xn+1 + xn−1 + xn mod(2). What is the image of the sequence x = (..., 1, 0, 1, 0, 1, 0, 1, ....)?

a) (...,1,1,1,1,1,1,1,1,1,1,1,...)

b) (...,0,0,0,0,0,0,0,0,0,0,0,...)

c) (...,1,0,1,0,1,0,1,0,1,0,1....)

4) Assume a cellular automaton has the property that T 3 (x) is the shift. Which of the following statements are
true?

a) T is chaotic in the sense of Devaney.

b) T 3 is chaotic in the sense of Devaney.

c) T 9 is chaotic in the sense of Devenay

5) If x = (...1, 0, 1, 0, 1, 0, 1, 0, 1, 0, ... with x0 = 1 and y = (..., 0, 0, 0, 0, 0, ...), then the distance between these
two points d(x, y) is
a) 1
b) 1/2
c) 0
d) 2
 6) True or false?

If d(x, y) = 1/10, then d(σ(x), σ(y)) = 1/10, where σ is the shift.

7) A lattice gas cellular automaton

a) conserves the total momentum of the particles

b) is used to simulate fluids

c) is used to simulate sand dynamics.

d) conserves the total angular momentum of the particles.

e) conserves the total energy of the particles.

8) What is a ”glider” in the game of life (X, T )?

a) A configuration x which satisfies T n (x) = σ m (x) for n, m > 0.

b) A configuration with finitely many living cells which satisfies T n (x) = σ m (x) for n, m > 0.

c) A configuration which satisfies T n (x) = x for n > 0.

d) A fixed point of T .
9) Who is believed to have first come up with the notion of cellular automata?

a) Hedlund at Harvard

b) Wolfram at Caltech

b) Ulam and von Neuman at Los Alamos

10) If you allow the alphabet of a cellular automaton to become a continuum, then the corresponding dynamical
system is called a

a) partial differential equation.

b) coupled map lattice.

c) map on an infinite dimensional space

d) an infinite system of coupled ordinary differential equations.

11) (5 points if correct) A one dimensional automaton maps x to y, where

x= ... 1 1 1 0 1 1 0 0 1 1 0 0 0 0 1 0 ...
y= ... 1 0 0 1 1 0 0 0 1 0 0 1 1 0 0 1 ...
What is the Wolfram number of this cellular automaton?
4/4 QUIZ NR 8 Math118, O.Knill

Name:

1) With fc (z) = z 2 + c, which of the following statements are true?

a) The Mandelbrot set is the set of complex numbers such that fcn (c) → ∞.

b) The Mandelbrot set is the set of complex numbers such that fcn (c) ≤ 2 for all n. (You
have shown in the homework that if fcn (c) > 2 then c is not in the Mandelbrot set).

c) The Mandelbrot set is the set of complex numbers such that fcn (c) stays bounded.

2) True or False?

a) Every quadratic polynomial is conjugated to the polynomial fc (z) = z 2 + c.

b) Every cubic polynomial is conjugated to a polynomial gc (z) = z 3 + c.

c) Every cubic polynomial is conjugated to a polynomial ga,b (z) = z 3 − 3a2 z + b.

3) True or False?

a) The union of the Julia set and the Fatou set is the entire complex plane.

b) The Fatou set is the complement of the Mandelbrot set.

4) Who was historically first to have made pictures of the Mandelbrot set?

a) John Hubbard.

b) Douady and Hubbard.

c) Benoit Mandelbrot.

d) Brooks and Matelski.

 5) A fixed point of a quadratic map f (z) is defined to be stable, if (only one answer applies):
a) f ′ (z) < 1.

b) |f ′ (z)| ≤ 1

c) |f ′ (z)| = 0.

d) |f ′ (z)| < 1.

6) True or False?
a) The Ulam map f (z) = 4z(1 − z) is in the complex plane conjugated to f−2 (z) = z 2 − 2.

b) The Julia set of the polynomial f0 (z) = z 2 is the circle with radius 1.

c) The filled in Julia set of the polynomial f (z) = 4z 2 is the disc of radius 1/2.

7) True or False? The Ulam map f (z) = 4z(1 − z) restricted to its Julia set is conjugated to
x 7→ 2x mod 1.

8) Which of the following dynamical systems is called the Newton iteration to find the root
f (z) = 0:

a) T (z) = 1 − f (z)/f ′ (z)

b) T (z) = z − f ′ (z)/f (z)

c) T (z) = 1 − f ′ (z)/f (z)

d) T (z) = z − f (z)/f ′ (z)

9) In order to find a fixed point of a map S, we can try to apply the Newton method to one of
the following:

a) T (z) = S(z) − z

b) T (z) = S ′ (z) − z

c) T (z) = z − S(z)/S ′ (z).

10) True or False? The Mandelbrot set is a fractal because its dimension has shown to be smaller
then 2 and bigger than 1.
4/11 QUIZ NR 9 Math118, O.Knill

Name:

1) Which of the following properties apply to the Baker transformation T on the square
[0, 1) × [0, 1).

a) The map is continuous.

b) There is a conjugation of the map to a subshift S(Y ) ⊂ {0, 1}Z
c) There is a conjugation of the map to the shift S(Y ) ⊂ {0, 1}N
d) The map is area-preserving.
e) The map has many periodic points.
f) The map has no periodic points.
g) The map is invertible.

2) True or False: If you take a subshift X of finite type, and a cellular automaton φ, then φ(X)
is a subshift of finite type.

3) True or False: If you take a sophic subshift and a cellular automaton φ, then φ(X) is a
sophic subshift.

4) Which of the following inclusions are true? (I had this once wrong on the blackboard and
Orr had corrected it):

a) subshifts ⊃ subshifts of finite type ⊃ sophic subshifts.

b) subshifts ⊃ sophic subshifts ⊃ subshifts of finite type.

5) True or False: the language of a subshift of finite type is the set of forbidden words.

6) What can you say about the subshift X of finite type over the alphabet {a, b, c} defined by
the forbidden words {aa, bb, cc, ac, ba, cb}?

a) X does not contain any point.

b) X contains only finitely many points.
c) X contains infinitely many points.

7) Which of the following subshifts is the shift over the alphabet {a, b} for which all words
bab, baaab, baaaaab, baaaaaaab, baaaaaaaaab, ... etc. are forbidden?

a) The Fibonacci shift

b) The even shift
c) The golden mean shift
d) The full shift.
 8) When doing symbolic dynamics for the Arnold cat map T (x, y) = (2x + y, x + y) mod 1,
one uses a subshift of finite type over an alphabet with a minimal amount of letters. This
alphabet has

a) 2 elements.
b) 3 elements.
c) 5 elements.
d) 6 elements.

9) Two random variables Y and Z taking finitely many values are called uncorrelated if and
only if

a) P [Y = a, Z = b] = P [Y = a]P [Z = b] for all possible numbers a, b.

b) E[Y Z] = E[Y ]E[Z].

10) Assume, a sequence of independent identically distributed random variables Y1 , Y2 , Y3 , ... de-
scribes drawing a card from an infinite deck containing 52 types of cards. It is assumed
that each card appears with the same probability 1/52 and that a card can appear multiple
times. How do you model these random variables?

a) Yk (y) = yk , where y ∈ {1, ..., 52}N .

b) Yk (y) = k, where y ∈ {1, ...., 52}N .

c) Yk (y) = y, where y ∈ {1, ..., 52}N .

4/18 QUIZ NR 10 Math118, O.Knill

Name:

√
1) What is special about the golden ratio θ = ( 5 − 1)/2? Check everything which applies:

a) It has the smallest possible continued fraction expansion θ = [a0 ; a1 , a2 , a3 , ....].

b) The partial fractions pn /qn have the property that qn and pn grow like Fibonnacci num-
bers.
c) θ is a solution of x = 2/(2 + x).

2) Which of the the following numbers has the continued fraction expansion
1
x = [0; 2, 1, 2, 1, 2, 1, ...] = 1
2+ 1
1+ 2++....

a) x is a solution of x = 1/(2 + 1/(1 + x)).

b) x is a solution of x = 1/(2 + 1/x)).
c) x is a solution of x = 1/(1 + 2/x)).
d) x is a solution of x = 1/(1 + 1/2 + x)).

3) Which of the following statements is called Dirichlets theorem for an irrational number α.

a) For every n, there exists q ≤ n such that

|α − p/q| ≤ 1/q 2 .

b) For every n, there exists q ≤ n such that

|α − p/q| ≤ 1/q .

c) For every q, there exists p such that

|α − p/q| ≤ 1/q 2 .

d) For every q, there exists p such that

|α − p/q| ≤ 1/q .

4) Which of the following three formulas give the correct recursion for the partial fractions
pn /qn :
a) pn+1 = an pn + pn−1 , qn+1 = an qn + qn−1 .
b) pn+1 = pn + an pn−1 , qn+1 = qn + an qn−1 .
c) pn+1 = an pn + an−1 qn−1 , qn+1 = an pn + an−1 qn−1
 5) The dynamical logarithm problem is the problem

a) to find the time to reach from a point x to a neighborhood of a point y.

b) to find the point y which is reached after time t when starting from x.
c) to find the initial point x, when reaching the point y after time t.

6) Which dynamical system is involved when making a decimal expansion of a real number.

a) T (x) = 10x.
b) T (x) = 10x mod 1.
c) T (x) = 10/x mod 1.

7) The quadratic map T (x) = x2 + c is also useful in number theory. Where?

a) to compute the eclipse times in calendars.

b) to factor large integers.
c) to understand why our tonal system has 12 scales between an octave and 19 scales for a
perfect fifth .

8) Who came up with the idea to factor integers n by finding two numbers x and y satisfying
x2 = y 2 mod n and then having a common nontrivial factor of x − y with n?
a) Fermat.
b) Tchebychev.
c) Minkovsky.

9) When finding lattice points close to graphs of quadratic polynomials, we were led to a dy-
namical system on the two dimensional torus. This system is

a) (x, y) → (x + 2a, x + y) mod 1.

b) (x, y) → (2x + y, x + y) mod 1.
c) (x, y) → (x + y, x − y) mod 1.
d) (x, y) → (ax2 + bx + c, x) mod 1.

10) Check for whatever continued fractions are useful:

a) To compute eclipse cycles.
b) To justify why we use a 12 scale system in music.
c) To find lattice points close to lines in the plane.
d) To factor integers.
4/25 QUIZ NR 11 Math118, O.Knill

Name:

1) True or False: there are always equilibrium solutions to the Newtonian n-body problem.

2) What is the minimal number of bodies for which one can prove that an escape to infinity in
finite time is possible?
a) 2
b) 3
c) 4
d) 5

3) Which of the following problems are considered ”restricted three body problems”:
a) The Earth-Moon-Sun system.
b) The Sitnikov problem.
c) A planet moving in the influence of a uniformly rotating binary star system.
d) The Kepler problem.

4) Who was the first to find non-collision singularities of the Newtonian n-body problem?
a) Joseph Gerver
b) Jeff Xia
c) Jürgen Moser
d) Henri Poincare

5) If for all t, we have a skew-symmetric matrix B(t) and S(t) satisfies the matrix differential
equation Ṡ = BS with the initial condition S(0) = I, then S is

a) orthogonal
b) skew symmetric
c) symmetric
d) the identity matrix

6) Which of the following are ingredients of the proof of chaotic orbits in the Sitnikov problem
a) A horse shoe construction.
b) Stable and instable manifolds
c) The Jacobi integral.
d) The Poincare return map.
e) The Poincare recurrence theorem.
f) Continued fraction expansion.
 7) Which of the following statements is called the third Kepler law:

a) The radius vector covers equal area in equal time.

b) Each of the bodies moves on an ellipse.
c) T 2 /a3 is constant.

8) The solar system is a dynamical system which shows very weak type of chaos. If one knows
the position of the earth with accuracy 1km, how long does one have to wait until the uncer-
tainty of the orbit has grown to about 1 astronomical unit (the mean distance of the earth
to the sun)?

a) 10’000 years
b) 1 Million years
c) 100 Million years
d) 10 Billion years

9) Which periodic three body solution has been observed in our solar system?

a) Euler motion.
b) Lagrange motion.
c) Hills solutions.
d) Moore choreographies.

10) How many integrals can one find for a general n-body problem:

a) 1
b) 10
c) 3n
d) 6n

11) How many integrals (conserved quantities) did you find for the n-body Toda system?

b) 2
b) n
c) 2n

12) Which of the following forces occur in a rotating coordinate system and depend on the an-
gular speed of the rotation?

a) Centrifugal force
b) Coriolis force
5/2 QUIZ NR 12 Math118, O.Knill

Name:

1) True or False?
A geodesic on a surface connecting two points P and Q is always the shortest path between
P and Q.

2) True or False?
For any surface the following is true: between two points, there is exactly one geodesic which
connects them.

3) On a surface of revolution, the geodesic flow is integrable. Which of the following integrals
are preserved along the orbit? (Several can apply).

a) the Clairot integral

b) the energy integral
c) the momentum integral
d) the angular momentum integral

4) On the hyperbolic plane, a geodesic is

a) either a circle or a vertical line.

b) a line
c) a circle

5) True or False? On any surface of revolution, the longitudinal lines (the intersection with a
plane through the axes of rotation) is a geodesics.

6) True or False? On a surface of revolution symmetric with respect to the z-axes, each inter-
section of the surface with z = const is a geodesic.

7) Snells law tells us something about the slope of the shortest connections between two
points. Assume to the left of the plane, moving is twice as hard as moving to the right and
the angle to the left is α and the angle to the right is β, then

a) sin(α) = 2 sin(β)
b) 2 sin(α) = sin(β)
c) sin(α) = sin(β)
d) cos(α) = 2 cos(β)
e) 2 cos(α) = cos(β)
f) cos(α) = cos(β)
 8) True or False? If we have a surface of revolution and we look at the Poincaré map at an
”equator” which is the intersection of the surface with a plane perpendicular to the axes of
rotation, then a geodesic which starts with an angle 0 < φ < π to the equator, then the
geodesic curve returns to the equator at some later time.

9) Which of the following equations are called the Euler-Jacobi equations for the action
∫b
functional I(γ) = a F (t, x, ẋ) dt?

a) Fp = dtd Fx .
b) Fx = dtd Fp .
c) Fx = − dtd Fp .
d) Fp = − dtd Fx .

Hint: If F (t, x, ẋ) = ẋ2 /2 − V (x) = T − V is the kinetic minus potential energy, then the
Euler-Jacobi equations are the Newton equations ẍ = −Vx .

10) On which of the following surfaces is every geodesic periodic?

a) The sphere
b) The flat torus
c) The one sheeted hyperboloid
d) The torus embedded in space (the doughnut)