118 Dynamicalsystems
118 Dynamicalsystems
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.
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.
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.
– 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.
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.
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.
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:
• 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 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.
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.
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.
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 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.
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
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) .
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
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):
EXAMPLES:
1) x = 4/5 is a periodic point of period 2.
x0 = 4/5, x1 = 2/5, x2 = 4/5 etc
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.
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.
xn = T n (x) = cn x0 .
2000
We see that for c > 1, the population grows exponentially for c < 1, 1000
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
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.
(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.
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].
a) What is the Lyapunov exponent of the orbit of the map f wich an initial
condition x0 = 1/2?
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.
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 :
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
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.
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.
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).
−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.
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
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.
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.
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
is an integral. It is not easy to verify that. Don’t ask how F was found!
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.
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.
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:
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.
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
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 .
It is an equation for the unknown map h ∈ C(X, X). We first consider the linearized problem
∑
∞
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
h = h+ + h− = (U − A+ )−1 f+ + (U − A− )−1 f− .
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
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
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
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
PROOF.
∑
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
(iv) There is only one fixed point. Assume, there were two fixed points x̃, ỹ of ϕ. Then
PROOF.
(i)
Consider for every τ > 0 and r > 0 the complete metric space
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}.
If τ ≤ 1 is small enough, then M · τ < (1 − λ)r. Using the triangle inequality, we obtain
(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 .
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.
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
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.
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).
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
ẋ = ∂y H(x, y)
ẏ = −∂x H(x, y)
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.
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
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
-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
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:
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.
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
′
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
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
depending on the intersection point (x1 , F (x1 )) with the null cline.
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.
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 .
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.
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.
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.
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.
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).
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.
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:
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.
• 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
ẋ = −(y + z)
ẏ = x + 0.2y
ż = 0.2 + xz − cz
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
ẋ = a sin(z) + c cos(y)
ẏ = b sin(x) + a cos(z)
ż = c sin(y) + b cos(x)
ẋ = y
ẏ = cos(x) + g sin(z)
ż = 1 .
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.
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
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 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.
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
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.
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
Therefore
log(Mk (A)) log(N2−k (A))
lim = lim .
k→∞ log(2k ) k→∞ log(2k )
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?
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.
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
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.
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.
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.
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.
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.
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.
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
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
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.
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 )
THE EULER EQUATIONS. The billiard map can be described by the equation
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.
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.
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
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.
T (x, y) = (x + 2arccos(y)/(2π), y)
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.
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.
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.
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.
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
-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
-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
-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.
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.
• 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.
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).
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
(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)
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.
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.
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.
AN UNSOLVED PROBLEM. Is the ellipse the only smooth convex table for
which exterior billiard is integrable?
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
This theorem can be used for example to prove the existence of solutions to differential equations.
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.
(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 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.
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.
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.
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
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.
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 .
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.
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:
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
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.
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.
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.
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.
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 .
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.
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
(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
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|.
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”.
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?
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
1950 Idealized models of biological systems were studied using CA. Ulam and von Neuman called this ”nearest neighbor-connected cellular spaces”. Source: From
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
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:
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.
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.
• 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.
(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.
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.
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.
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:
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.)
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
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
fc : z 7→ z 2 + c
PARAMETRIZING ALL QUADRATIC MAPS. The quadratic family fx is not as special as one might think:
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 .
√
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
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
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.
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.
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
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:
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.
Φ : c 7→ Gc (c)
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.
(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))
(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.
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 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.
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.
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.
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.
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
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
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.
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 .
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
Two random variables Y, Z are called uncorrelated if E[Y Z] = E[Y ]E[Z]. Two independent random variables
are automatically uncorrelated.
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
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:
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.
√
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
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 .
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
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
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
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
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.
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.
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).
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: The full shift as well as subshifts of finite type are not minimal. They have many periodic orbits.
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
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.
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.
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).
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
.
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.
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 .
ABSTRACT. Finding lattice points close to curves leads to problems in dynamical systems theory.
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).
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).
If we look at the lines y = const, then these lines are tossed around
in a regular way by the dynamics.
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
√ 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.
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
-4
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.
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.
∫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.
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.
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 .
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 x1 + ... + 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):
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)).
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.
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).
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
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.
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.
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.
DECIMAL EXPANSION. Let T (x) = 10x and f (x) = [10x] where [r] is the 0.8
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
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
√ √
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
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.
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
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:
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
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.
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 OF XIA.
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:
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 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.
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.
d2 z z
=− 2 ,
dt2 (z + r(t)2 )3/2
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.
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.
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
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
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:
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.
COLLISIONS.
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.
(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 ).
/ 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 .
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.
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.
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 ,
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
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.
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:
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.
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.
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.
using det(⃗a, ⃗b +⃗a) = det(⃗a, ⃗b). The caustic of the curve γ is called
the evolute of the curve.
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.
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.
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.
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.
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
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)
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
(az + b) 5
T (z) =
(cz + d)
4
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.
g11 = 4π 2 (a + b cos(2πv))2 = 4π 2 r2
g22 = 4π 2 b2
g12 = g21 = 0
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.
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.
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.
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.
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.
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).
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. 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.
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. 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.
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.
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.
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?
COHEN MAP. Does the hyperbolic point of period 14 found by Hubbard in the map
√
T (x, y) = ( x2 + 1 − y, x)
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:
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.
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.
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.
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.
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.
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).
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.
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 .
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
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
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.
∂ 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
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
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
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 |x − y|R
g(x, y) = − log( )
2π |y||x − y|
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).
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 .
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.
ABSTRACT. When studing linear maps or curves on finite tori Y = Zm1 × · · · × Zmn , we are in the ream of
elementary number theory.
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.
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.
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.
a) What is the Lyapunov exponent of the orbit of the map f wich an initial
condition x0 = 1/2?
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.
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.
b) Find the bifurcation points, which are parameter values, where the
stability of one of these fixed points changes.
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.
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].
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.
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
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.
3.4 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) (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.
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.
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.
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.
5.5 The string construction allows to construct a table, with a given caustic.
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}?
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}).
7.1 The Mandelbrot set M can be defined as the set of parameter values c for which fcn (0)
stays bounded.
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.
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
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.
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.
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.
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.
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.
Remark. The natural Newton potential depends on the space. The Harmonic oscillator potential
||x||2 can be considered the natural 0 dimensional Newton potential
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 ) .
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.
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.
3) Look at the map T (x) = x2 + x. Which of the following sequences form an orbit of x
through x = 1:
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.
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:
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).
3) Which of the following dynamical systems have a discrete time? We replace ”map” or ”dif-
ferential equation” with ”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
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?
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
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?
1 1 1
-1 -1 -1
1 1 1
-1 -1 -1
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?
Name:
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
1 1 1
-1 -1 -1
1 1 1
-1 -1 -1
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.
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.
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.
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:
a) log(14)/ log(4)
b) log(5)/ log(3)
c) log(20)/ log(5)
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:
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:
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. .
d) The billiard map has an invariant curve in the annulus R/Z × [−1, 1].
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
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.
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:
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 ).
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 .
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?
b) A configuration with finitely many living cells which satisfies T n (x) = σ m (x) for n, m > 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
10) If you allow the alphabet of a cellular automaton to become a continuum, then the corresponding dynamical
system is called a
Name:
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?
3) True or False?
a) The union of the Julia set and the Fatou set is the entire complex plane.
4) Who was historically first to have made pictures of the Mandelbrot set?
a) John Hubbard.
c) Benoit Mandelbrot.
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:
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
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).
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):
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}?
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) 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
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?
Name:
√
1) What is special about the golden ratio θ = ( 5 − 1)/2? Check everything which applies:
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++....
3) Which of the following statements is called Dirichlets theorem for an irrational number α.
|α − p/q| ≤ 1/q 2 .
|α − p/q| ≤ 1/q .
|α − p/q| ≤ 1/q 2 .
|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
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.
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
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:
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).
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 .
a) The sphere
b) The flat torus
c) The one sheeted hyperboloid
d) The torus embedded in space (the doughnut)