Beruflich Dokumente
Kultur Dokumente
Tyler London
Department of Mathematics
College of Liberal Arts and Sciences
Tufts University
Senior Honors Thesis
2009
Abstract
In 1964, Sharkovsky showed that there is a linear ordering
s
of the natural numbers
such that m
s
n if every continuous interval map having an n-periodic point has an
m-periodic point. This idea is generalized to get a partial ordering of cyclic permutations
called the forcing relation by taking into consideration the pattern of the periodic orbit.
We show that forcing can be reduced to the study of certain canonical piecewise linear
maps. Combinatorial algorithms of Baldwin and Coven et. al. that decide forcing are
presented, as well as a diagram of the forcing relation on over 190 cycles of length up to
8. The global structure of forcing is described by considering ips of cycles and unimodal
cycles. It is shown that the forcing relation restricted to unimodal cycles is a total ordering,
and this relation is diagrammed on 115 unimodal cycles. The local structure of forcing
is also explored by studying doublings and extensions. A new combinatorial proof of a
theorem describing the structure of doublings in the forcing order is given. Forcing-minimal
properties are examined by considering primary and simple cycles. Finally, using the
results proven about the forcing relation, we derive Sharkovskys theorem as a consequence
of the forcing theorem.
Key words and phrases. Interval maps, periodic orbit, Markov graph, Sharkovskys theo-
rem, cycle, forcing, doubling, block structure, extension, primary cycle, simple cycle.
Contents
Acknowledgments iii
0 Introduction 1
1 Background 3
1.1 Cyclic Permutations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Continuous Maps of the Interval . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The Theorem of Sharkovsky . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 The Forcing Relation on Cycles 11
2.1 The Combinatorial Data in a Periodic Orbit . . . . . . . . . . . . . . . . . . 11
2.2 The Forcing Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 A Computational Digression 20
3.1 Description of the Algorithm of Baldwin . . . . . . . . . . . . . . . . . . . . 20
3.2 Implementation of the Algorithm in GAP . . . . . . . . . . . . . . . . . . . 21
3.3 Eciency of the Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.4 Properties of Forcing in the diagrammed data . . . . . . . . . . . . . . . . . 22
4 Global Structure of Forcing 25
4.1 The Flip Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2 The Restriction of Forcing to Unimodal Permutations . . . . . . . . . . . . 26
5 Extensions of Cycles 32
5.1 Doublings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 General Extensions of Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6 Forcing-Minimal Cycles 44
6.1 Primary Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.2 Simple Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
7 A Forcing Proof of Sharkovskys Theorem 52
8 Conclusions 53
Bibliography 54
i
List of Figures
1.1 f
for = (1 5 2 6 3 4). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 f
for = (1 4 3 2 5). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The Markov graph of = (1 4 3 2 5). . . . . . . . . . . . . . . . . . . . . . . 5
1.4 The Markov graph of q
4
(x) associated to the partition P. . . . . . . . . . . 8
1.5 The Markov graph associated to the partition J
1
, J
2
. . . . . . . . . . . . . . 9
2.1 Graphs of symmetric tent maps exhibiting dierent 5-cycles. . . . . . . . . . 12
2.2 Graphs of f
and f
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 f
for = (1 4 3 2 5). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 The signed Markov graph of = (1 4 3 2 5). . . . . . . . . . . . . . . . . . . 15
2.5 Markov graph of = (1 2 4 3) . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1 Baldwins original diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Computation of for n 8. . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1 Illustration of the complexity of on some 4-cycles. . . . . . . . . . . . . . 25
4.2 The graph of f
nN
C
n
.
Example 1.1.2. Consider the 3-cycle dened by (1) = 2, (2) = 3, and (3) = 1.
Using cycle notation, we have = (1 2 3).
Given an n-cycle , one can naturally associate to two objects: a continuous map of
the interval [1, n] and a directed graph.
3
Denition 1.1.3. Let be a cyclic permutation of length n and for each i 1, . . . , n1,
let
i
= (i +1) (i). The connect-the-dots map or canonical -linear map f
: [1, n]
[1, n] is dened by
f
(i +x) = (i) +x
i
for 0 x 1.
Note that f
1, . . . , n
= . Simply put, f
is not
monotone on the set k 1, k, k + 1 or that k is a local extremum of f
.
Example 1.1.4. Consider the 6-cycle = (1 5 2 6 3 4). The graph of f
is shown in Figure
1.1, and k = 2 is a maximum of and k = 4 is a minimum of .
1 2 3 4 5 6
1
2
3
4
5
6
Figure 1.1: f
for = (1 5 2 6 3 4).
A directed graph is a pair G = (V, E) where V is a set whose elements are called vertices
and E is a set of ordered pairs of vertices called directed edges. If G is a directed graph,
then a path of length k in G is a k-tuple (v
1
, . . . , v
k
) such that each v
i
is a vertex and
(v
i
, v
i+1
) is a directed edge for 1 i k. We will primarily be concerned with paths in
directed graphs that begin and end at the same vertex. In particular, a path (v
1
, . . . , v
k
)
in G is called a closed path of length k or a loop if (v
k
, v
1
) is also a directed edge. In graph
theory, a closed path may also be referred to as a circuit.
Denition 1.1.5. Suppose is an n-cycle. Dene J
i
= [i, i + 1] for i = 1, . . . , n 1 and
let V = J
1
, . . . , J
n1
. The Markov graph associated to is the directed graph with n1
vertices V and an arrow (directed edge) drawn from J
k
to J
l
if and only if f
(J
k
) J
l
.
Example 1.1.6. Let = (1 4 3 2 5). By looking at the graph of f
for = (1 4 3 2 5).
J1 J2
J3 J4
Figure 1.3: The Markov graph of =
(1 4 3 2 5).
Lastly, we remark that if C
n
, then we will generally consider two dierent labellings
of the elements of P = p
1
, . . . , p
n
= 1, . . . , n. The rst is the spatial labeling where
p
i
< p
j
if and only if i < j. The second is the temporal labeling where p
i
=
i1
(p
1
)
for i = 2, . . . , n. Note that the above denition of the temporal labeling is equivalent to
saying p
i
= (p
i1
) for i = 2, . . . , n and p
1
= (p
n
).
1.2 Continuous Maps of the Interval
We aim to explore the combinatorial properties of periodic points of interval maps. By
an interval map, we mean a continuous function f : I I of an interval I R into itself.
Oftentimes, the family of all continuous self-maps of an interval I will be of interest, so
A(I) shall denote this set. Given f A(I), we shall examine the iterates f
n
of f, which
are dened inductively by f
0
= id and f
n
= f f
n1
for n N. For x I and f A(I),
the orbit of x under f is dened to be the set
O(x) = f
n
(x) : n = 0, 1, 2, . . ..
A point x is said to be periodic if for some k N, f
k
(x) = x. In this case, O(x) is nite
and the set x, f(x), . . . , f
k1
(x) is called a periodic orbit. The period
1
of a periodic point
x is the least positive integer n such that f
n
(x) = x; that is, f
n
(x) = x and f
k
(x) ,= x for
1 k n 1. We may also call a periodic point of period n an n-periodic point. Lastly,
if x is a 1-periodic point, f(x) = x, then we call x a xed point. Note that an n-periodic
point x is a xed point under the map f
n
, f
2n
, f
3n
, . . .; however, a xed point of the map
f
n
need not be a periodic point of period n.
1
In other literature, what we have dened to be the period of a point is sometimes called the least or
minimal period.
5
Example 1.2.1. Let be an n-cycle. Then it is clear from the denition of that
1, 2, . . . , n is a periodic orbit of period n for f
.
If P is a nonempty subset of an interval I, then P) will denote the closed convex hull
of P; that is, the smallest closed connected subset of I containing P. For example, in
R, 2, 3) = 2, 3) = [2, 3] and (1, 5)) = [1, 5]. Let f A(I) and P = p
1
< p
n
I) = J, f(int
I) = int J
and f(
I) = J.
Proof. Let J = [a, b]. By hypothesis, f(I) J, so there exist points a
0
, b
0
I such that
f(a
0
) = a and f(b
0
) = b. First, let us assume that a
0
< b
0
. Dene
a
1
= supx : a
0
x b
0
and f(x) = a.
Observe that by continuity, f(a
1
) = a and since f(b
0
) = b > a, a
1
is strictly less than b
0
.
Next, dene
b
1
= infx : a
1
x b
0
and f(x) = b,
again noting that f(b
1
) = b. Now, f(a
1
, b
1
) = a, b, and by denition of a
1
and b
1
, no
point of (a
1
, b
1
) maps to the boundary of J. Thus, f(int(a
1
, b
1
)) = intJ = (a, b) and so
I = [a
1
, b
1
] is the desired subinterval of I. The case when b
0
< a
0
is handled analogously,
interchanging the inmum and supremum.
We shall say that an interval J minimally f-covers an interval I if J satises the
conclusion of Lemma 1.2.2.
Lemma 1.2.3. If I
k
: k = 0, 1, . . . is a (nite of innite) family of nonempty closed
subintervals of I such that f(I
k
) I
k+1
(i.e. I
k
I
k+1
), then there exists a point x I
0
such that f
k
(x) I
k
for each k.
Proof. By Lemma 1.2.2, since f(I
0
) I
1
, there is a closed subinterval J
1
of I
0
that
minimally f-covers I
1
. Proceeding by induction, suppose that we have dened subintervals
J
1
J
2
J
n
in I
0
such that f
k
(J
k
) = I
k
and J
k
minimally f
k
-covers I
k
for every
1 k n. Then,
f
n+1
(J
n
) = f(f
n
(J
n
)) = f(I
n
) I
n+1
.
Appealing to Lemma 1.2.2, there exists a closed subinterval J
n+1
J
n
that minimally
f
n+1
-covers I
n+1
, thus f
n+1
(J
n+1
) = I
n+1
. We now have a nested sequence J
n
nN
of
6
non-empty closed intervals, so the intersection
n=1
J
n
is nonempty, and for any x in this
intersection, we have f
k
(x) I
k
.
Lemma 1.2.4. If J I is a closed subinterval such that f(J) J, then f has a xed
point in J.
Proof. Let a < b be the endpoints of J. Since f(J) J, there exist points , J such
that f() a and f() b. If we let g(x) = f(x) x, then
g() = f() f() a 0
and
g() = f() f() b 0.
As g is continuous, by the Intermediate Value Theorem, there exists a point x J such
that g(x) = 0; that is, x is a xed point of f.
Before we discuss the implications of the above lemmas, let us rst associate to each
interval map a directed graph which has the property that f-covering relations are encoded
as paths.
Given an interval I, a partition is a (nite or innite) collection of closed subintervals
J
k
of I, with pairwise disjoint interiors such that
k
J
k
= I. Given a partition J
k
of
I, an interval map f A(I) and a point x I, the itinerary of x of length n is the word
itin
n
(x) = (J
i1
, . . . , J
in
) in the letters J
1
, J
2
, . . ., such that f
k
(x) I
i
k
for 1 k n.
Itineraries give us information about where a point has been and where it might go,
without precisely telling us where the point is. Itineraries are very useful tools in symbolic
dynamics, and we will revisit their use in later chapters when unimodal permutations are
studied.
Denition 1.2.5. The Markov graph of f A(I) associated to the partition J
k
is
the directed graph with vertices J
k
, and a directed edge from J
l
to J
m
if and only if J
i
f-covers J
m
.
Example 1.2.6. The family of interval maps q
a
(x) = ax(1 x) : a R is referred
to as the quadratic or logistic family. Consider the partition P = J
1
= [0, 1/4], J
2
=
[1/4, 3/4], J
3
= [3/4, 1] of the interval [0, 1]. For a = 4, the quadratic map q
4
(x) =
4x(1x) maps [0, 1/4] to [0, 3/4], [1/4, 3/4] to [3/4, 1] and [3/4, 1] to [0, 3/4]. The Markov
graph of f associated to the partition P is shown in Figure 1.4.
Example 1.2.7. If we consider the partition of the interval [1, 5] given by P =
[1, 2], [2, 3], [3, 4], [4, 5], then for = (1 4 3 2 5), the Markov graph of f
associated to
the partition P is the same as in Example 1.1.6. See Figure 1.3. This illustrates that for
a connect-the-dots map, Denitions 1.1.5 and 1.2.5 coincide.
Given Denition 1.2.5, Lemma 1.2.3 reveals that every path in the Markov graph of
a partition corresponds to a point whose itinerary is exactly the sequence of vertices in
this path. Suppose, in particular, that the path is a closed path, i.e. a path of type
J
i0
J
i2
J
in1
J
i0
. Then J
i0
f
n
-covers itself, so Lemma 1.2.4 guarantees
the existence of a xed point x J
i0
of f
n
such that f
k
(x) J
i
k
and f
n
(x) = x. However,
if this loop is repetitive in the sense that it is formed by going around smaller loop several
times, then the period of the xed point may be strictly smaller than n. The above
discussion gives us the following results:
7
0
1
4
3
4
1
3
4
1
q4
J
3
J
1
J
2
Figure 1.4: The Markov graph of q
4
(x) associated to the partition P.
Corollary 1.2.8. If J
i0
J
i1
J
in1
J
n
= J
i0
is a loop in the Markov
graph of f of a partition, then there exists a xed point x J
i0
of f
n
with f
k
(x) J
i
k
for
0 k n 1.
Lemma 1.2.9. If J
i0
J
i1
J
in1
J
in
= J
i0
is a loop in the Markov graph
of f of a partition that is not formed by going p times around a shorter loop of length m
where mp = n, then the periodic point x of Corollary 1.2.8 has period n.
Any loop that cannot be formed by going multiple times around a shorter loop as in
the statement of Lemma 1.2.9 will be called a simple or nonrepetitive loop.
1.3 The Theorem of Sharkovsky
In the previous section, we showed that given an interval map and a partition of the
interval, by analyzing the Markov graph of a partition we can gain useful information
about the existence of periodic orbits. Suppose now we know that x
1
< < x
n
is a
periodic orbit of the interval map f A(I). Then, the set P = [x
1
, x
2
], , [x
n1
, x
n
]
is a partition of the interval [x
1
, x
n
]. By Lemma 1.2.3, closed paths in the Markov graph
of this partition correspond to xed points of some iterate of f in [x
1
, x
n
]. A very natural
question then arises: does the existence of a periodic orbit of period n guarantee the
existence of an orbit of some other period? This question was independently answered
by the Ukrainian mathematician Alexander Sharkovsky in two papers between 1964 and
1965 [Sha64, Sha65] and by Tien-Yien Li and James A. Yorke in 1975 [LY75].
To address this question, rst consider the following ordering
s
on the natural num-
bers dened by:
1
s
2
s
2
2
s
2
3
s
s
2
2
5
s
2
2
3
s
s
2
k
5
s
2
k
3
s
s
2 5
s
2 3
s
s
7
s
5
s
3.
This ordering is eponymously named the Sharkovsky ordering and encodes information
about which periodic orbits force other periodic orbits to exist. This is stated explicitly
in the following theorem:
8
Theorem 1.3.1 (Sharkovsky). Let m, n N. Then m
s
n if and only if every continuous
f : I I having a periodic point of period n has a periodic point of period m.
The proof of this theorem will be given at the end of the paper once the notions
of forcing have been built up suciently. Nevertheless, there are several nice proofs of
Theorem 1.3.1 that do not use forcing. The standard proof can be found in [BGMY80],
whereas slightly dierent and newer proofs can be found in [ALM93] and [BH]. For the
earlier versions, the reader is directed to [
a[0,2]
is given by T
a
(x) = ax for
x [0,
1
2
] and T
a
(x) = a(1x) for x [
1
2
, 1] [BB04]. For two values of a, a = a
1
1.72208
and a = a
2
1.51288, x
0
=
1
2
is a periodic point of period 5 for the tent maps T
a1
and T
a2
.
Denote the period 5 orbit for each map by x
0
, x
1
, x
2
, x
3
, x
4
with x
i
x
i+1
. Relabeling the
points in a linearly increasing manner, the permutations exhibited are obtained. As can
be seen in Figure 2.1 [BB04, Figure 3.3], T
a1
exhibits the permutation (1 2 4 3 5) whereas
T
a2
exhibits the permutation (1 3 4 2 5).
Denition 2.1.1 points towards a generalization of the Sharkovsky ordering, one that
considers orbit types. Example 2.1.2 shows that we may certainly have two distinct pe-
riodic orbits of period 5. In fact, a simple counting argument shows us that there are
11
x2 x3 x0 x4 x1
1
2
1 2 3 4 5
(1 2 4 3 5)
Ta
1
x2 x0 x3 x4 x1
1
2
1 2 3 4 5
(1 3 4 2 5)
Ta
2
Figure 2.1: Graphs of symmetric tent maps exhibiting dierent 5-cycles.
(n 1)! possible distinct orbit types for a periodic point of period n. In our more general
setting, what does the existence of a periodic orbit of type imply? This motivates the
following denition.
Denition 2.1.3. Let and be cycles. We say forces , denoted , if every
continuous map f : I I that has an orbit of type also has an orbit of type .
Equivalently, if f : Perm(f) f : Perm(f).
A priori, Denition 2.1.3 merely denes a relation, that is, a subset of C C. The
usefulness of this denition thus depends on how we can characterize this relation. First,
recall the denition of a partial order:
Denition 2.1.4. A partial order is a relation on a set X that is reexive, transitive and
antisymmetric. A set equipped with a partial order is called a partially ordered set.
As it turns out, forcing is a partial order on the set of cycles C but not a linear ordering.
Theorem 2.1.5 ([Bal87, Theorem 1.4]). Forcing is a partial order.
Proof. We need to show that is a reexive, transitive and antisymmetric relation. Let
, and be cycles. Reexivity is trivial: if a map exhibits , then it exhibits . Hence,
. For transitivity, suppose and . If a map exhibits , then it exhibits , and
since it exhibits , it exhibits . Thus, we have and so is a transitive relation.
Antisymmetry, on the other hand, requires a bit more work. Let n, m N and
C
n
, C
m
. Suppose that and . We aim to show that this implies = .
Following [Bal87] and [ALM93, p.53], let f be a polynomial of degree at least 2 which
passes through all of the points (i, (i)) for 1 i n, so that f has an orbit of type
(and therefore, by hypothesis, an orbit of type ). Now, as f is a polynomial, the equation
f
n
(x) = x has only nitely many solutions, so f can have only nitely many orbits of type
. Let P = x
1
, . . . , x
n
, x
1
< < x
n
, be one of these orbits having the property that
12
max P min P = x
n
x
1
is the least possible. Dene a continuous function g by
g(x) =
f(x
1
) : x x
1
f(x) : x [x
1
, x
n
]
f(x
n
) : x x
n
By construction, g has an orbit of type , namely P, and therefore, by hypothesis, has an
orbit of type , say Q = y
1
, . . . , y
m
, y
1
< < y
m
. Note that since g(P) = P and for
all x / [x
1
, x
n
], g(x) P, we have Q [x
1
, x
n
]. By the same construction as above, there
is a continuous map h given by
h(x) =
f(y
1
) : x y
1
f(x) : x [y
1
, y
m
]
f(y
m
) : x y
m
Since Q is an orbit of type for h, h must also have an orbit of type , say R = x
1
, . . . , x
n
.
But, R [y
1
, y
m
], so R is also an orbit of f of type . By the minimality of P, we have
R = P, so P = Q and thus = .
Theorem 2.1.6 ([Bal87, Theorem 1.5]). Forcing is not a linear ordering.
Proof. To show that forcing is not a linear ordering it suces to show that there exist
two cycles and such that , and ,. By denition of the forcing relation, it is
sucient to show that there exist two continuous maps, one exhibiting but not , and
the other exhibiting but not . To that end, let = (1 2 3) and = (1 3 2) and recall the
denition of the connect-the-dot maps f
and f
(x
1
) = x
3
> x
1
= f
(x
2
), we must have that x
1
J
2
. Similarly, as x
2
< x
3
and f
(x
2
) = x
1
< x
2
= f
(x
3
), x
2
must be in J
1
. But, as x
2
,= 2, this contradicts our
assumption that x
1
< x
2
. Thus, f
(x
2
) < f
(x
3
) < f
(x
1
).
On the other hand, suppose that f
(y
1
) = y
2
< y
3
= f
(y
2
) so y
1
J
2
. However, y
2
< y
3
and f(y
2
) = y
3
> y
1
= f(y
3
), so
y
2
J
1
. This again contradicts the assumption that y
1
< y
2
. Therefore, f
is a continuous
map that exhibits but not , so ,. This concludes the proof of showing that forcing
is not a linear ordering.
2.2 The Forcing Theorem
Thus far, we have shown that the set C of cycles is partially ordered by the forcing
relation , and moreover, this relation is not a linear ordering. Nevertheless, a criterion
for deciding when a cycle forces another cycle has yet to be given. The main goal
of this section, therefore, is to state and prove the Forcing Theorem (Theorem 2.2.1)
which gives a necessary and sucient condition for a cycle to force another. First proven
13
1 2 3
1
2
3
1 2 3
1
2
3
f
and f
.
in [Bal87, Main Theorem 2.4], this theorem and its proof plant the seeds for Chapter 3
where an explicit algorithm is presented that decides the forcing relation.
Theorem 2.2.1 (Forcing Theorem). If C
n
and C
m
, then if and only if
either = or there is a loop of length m in the Markov graph of with orbit type .
To begin with, it is clear that given some cyclic permutation , the family of interval
maps that exhibits can be extremely large. An interesting question is then whether or
not there exists some canonical interval map that exhibits only the cyclic permutations
forced by . If this were to be the case, then we could reduce forcing to the study of
properties of these maps alone. In particular, our approach will be to show that such
canonical maps exist indeed, they are the connect-the-dots maps from Chapter 1 and
that by studying the combinatorial properties of the Markov graphs of these maps, the
desired criterion for forcing can be derived.
Let C
n
and consider the interval map f
+1 if f
increasing on I
1 if f
decreasing on I
Example 2.2.2. The graph of f
for = (1 4 3 2 5).
J
1
,+ J
2
,
J
3
,+ J
4
,
Figure 2.4: The signed Markov graph of =
(1 4 3 2 5).
Denition 2.2.3. For a loop (J
i1
, . . . , J
in
), we dene the shift operation by (J
i1
, . . . , J
in
) =
(J
i2
, J
i3
, . . . , J
in
, J
i1
). Given two loops = (J
i1
, . . . , J
in
) and = (J
k1
, . . . , J
kn
) of the
same length, we dene <
if and only if
M1
j=1
z(J
ij
)
i
M
<
M1
j=1
z(J
ij
)
k
M
,
where M n is the least value of j for which i
j
,= k
j
. This is a linear ordering.
When we are dealing with the Markov graphs of cycles (resp. connect-the-dots maps)
it will be notationally easier sometimes to label the vertices as integers. Thus, instead
of labeling the n 1 vertices of the Markov graph of an n-cycle by J
1
, J
2
, . . . , J
n1
, we
will simply label them 1, 2, . . . , n 1, where the integer i corresponds to the subinterval
[i, i +1]. Then, if = (a
1
, a
2
, . . . , a
n
) and = (b
1
, b
2
, . . . , b
n
) are loops of the same length,
we have <
if and only if
M1
j=1
z(a
j
)
a
M
<
M1
j=1
z(a
j
)
b
M
.
where M n is the least value of j for which a
j
,= b
j
.
The most important aspect of this ordering is the following: suppose f A(I) is an
interval map, J
1
, . . . , J
N
is a partition of I (labeled so that min J
i
< min J
i+1
) and we
have loops = (J
i1
, . . . , J
in
) and = (J
k1
, . . . , J
kn
) in the Markov graph of f associated
to this partition. These loops can be thought of as signed itineraries of periodic points
in I, say x corresponding to and y corresponding to . We will see in the proof of the
forcing theorem that the order of loops tells us the order of the points to which the loops
correspond:
Proposition 2.2.4. <
implies x < y.
15
Example 2.2.5. In the Markov graph of = (1 2 4 3) (Figure 2.5), one can nd the
nonrepetitive loops (1, 2, 2), (2, 2, 1) and (2, 1, 2). Using the order from Denition 2.2.3,
observe that (1, 2, 2) <
(2, 2, 1) <
(2, 1, 2).
1, + 2,
3, +
Figure 2.5: Markov graph of = (1 2 4 3)
Denition 2.2.6. Let be a nonrepetitive loop of length mand let W = , (),
2
(), . . . ,
m1
().
Order W by <
(2, 2, 1) <
we must have 1 x
1
and x
m
n.
Assume ,= , so that none of x
1
, . . . , x
m
are integers. Dene the path = (a
1
, . . . , a
m
)
in the Markov graph of by
a
j
= i f
j
(x
1
) [i, i + 1],
16
and observe that is actually a loop in the Markov graph of . We rst want to show
that is nonrepetitive, so suppose that is a loop in the Markov graph of of length
k such that k divides m and a
j1
= b
j2
if and only if j
1
j
2
(mod k). Dene intervals
I
j
, 0 j m by backwards induction from m. Let I
m
= [a
m
, a
m
+ 1], and if I
j
has
been dened for j 1, let I
j1
= f
1
(I
j
) [a
j1
, a
j1
+ 1] (where we let a
0
= a
m
for convenience). Now a
jk
= a
j
, so we can show by induction that I
jk
I
j
whenever
k j m and that f
k
: I
0
I
k
has a xed point y, and clearly f
m
(y) = y also, since k divides
m. However, x
1
I
0
(since by denition f
j
(x
1
) [a
j
, a
j+1
] for all j) and f
m
(x
1
) = x
1
,
but x
1
,= y (since x
1
has least period m and y has least period k < m). Thus f
m
: I
0
I
m
is one-to-one, onto, linear, and has two xed points, so I
0
= I
m
and f
m
is the identity
function. Furthermore, since [f
(x)[ 1 whenever f
: I
j
I
j+k
is a linear map of slope either 1 or 1.
Thus if x
1
= a
m
+
1
2
, then f
k
(x) = a
m
+
1
2
contradicting x
1
has least period m > k, and
if x
1
is some element of [a
m
, a
m
+ 1] other than a
m
+
1
2
, then it is easy to see that X has
type , contradicting ,= . Thus must be nonrepetitive.
Next, we must show that the orbit type of equals . Let (k) =
k
() with (k) =
(a
1
(k), a
2
(k), . . . , a
m
(k)) and note that a
j
(k) = i if and only if f
j+k
(x
1
) [i, i + 1].
Suppose (k
1
) <
(k
2
). Let j be least such that a
j
(k
1
) ,= a
j
(k
2
). Then if we look
at f
k1
(x
1
) and f
k2
(x
1
) and follow their progress as f
(f
k1
(x
1
)) [a
j
(k
1
), a
j
(k
1
) + 1] and f
j
(f
k2
(x
1
)) [a
j
(k
2
), a
j
(k
2
) + 1] we see that
f
k1
(x
1
) < f
k2
(x
1
).
Thus we have shown that
f
k1
(x
1
) < f
k2
(x
1
) (k
1
) <
(k
2
),
from which it follows that the orbit type of is .
Now we need to show the other direction: if there is a loop in the Markov graph of
of loop type , then any interval map with an orbit of type has an orbit of type . First,
we need to prove the following lemma:
Lemma 2.2.9 ([Bal87, Lemma 3.2]). Suppose that f : R R is continuous, a < b,
f(a) < f(b), and for some N N, (a
i
, b
i
) : 1 i N is a collection of intervals having
the properties that
(i) f(a) f(a
i
) < f(b
i
) f(b);
(ii) a a
i
< b
i
b;
(iii) (a
i
, b
i
) minimally f-covers itself;
(iv) If i < j, then either (a
j
, b
j
) (a
i
, b
i
) or (a
i
, b
i
) (a
j
, b
j
) = ;
(v) f is strictly increasing on the set a
i
: i < N b
i
: i < N.
If c < d, c, d [f(a), f(b)] are such that for all i either (c, d) (f(a
i
), f(b
i
)) or
(c, d) (f(a
i
), f(b
i
)) = , then there exist points a
N+1
and b
N+1
such that f(a
N+1
) =
c, f(b
N+1
) = d and (a
i
, b
i
) : 1 i N + 1 satises (i)-(v) with N replaced by N + 1.
17
Proof.
Case 1. (c, d) (f(a
i
), f(b
i
)) for some i.
Assume that i is the largest possible, so that (f(a
i
), f(b
i
)) (f(a
j
), f(b
j
)) or
(f(a
i
), f(b
i
)) (f(a
j
), f(b
j
)) = for all i ,= j (by condition (iv) and (v)). Let a
N+1
be the greatest element of [a
i
, b
i
] such that f(a
N+a
) = c and then let b
N+1
be the least
element of [a
N+1
, b
i
] such that f(b
N+1
) = d.
Case 2. (c, d) (f(a
i
), f(b
i
)) = for all i.
Now let a
be the least
element of b a
i
: f(a
i
) d. Let a
N+1
be the greatest element of [a
, b
] such that
f(a
N+1
) = c and let b
N+1
be the least element of [a
N+1
, b
, and we need only consider this function so that we can construct the Markov graph
of . However, the function f
(2, 2, 1) <
(2, 1, 2), and from Example 2.2.7, the orbit type of this
ordered list is (1 2 3). Therefore, we conclude that (1 2 3).
20
L
2
: Next, using the same strategy as above, we can order the list of loops that are shift
equivalent to L
2
by (2, 3, 2) <
(2, 2, 3) <
. Finally, for each of these ordered lists of walks closed under the shift operation, the
function baldwinForcedPermutation computes its loop type. The list of all loop types is
returned; that is, the list of all m-cycles forced by .
3.3 Eciency of the Algorithm
Theorem 2.2.1 states that a necessary and sucient condition for a cycle to force a
cycle is that there is a loop of length [[ in the Markov graph of that corresponds to
21
an orbit of type . Recall from Denition 2.1.3 that forcing is dened as a property of
all continuous maps exhibiting a certain cyclic permutation. In fact, there is an inherent
problem with using the connect-the-dots maps to determine forcing. The problem with
using f
is in
one-to-one correspondence with a subset of the closed paths of length [[ in the Markov
graph of , there is no ecient way known of determining whether has a representative
in f
. Baldwin solved this problem by ignoring it. In his algorithm, one examines all
closed paths of length [[ in the graph to see if any of them correspond to a representative
of . The number of such paths is exponential in [[ [BCMM92].
However, it is possible to speed up the algorithm considerably. In [BC92a], Bernhardt
and Coven present an eective polynomial-time algorithm for deciding forcing. This paper,
which can be seen as an extension of [BCMM92], considers horseshoe maps instead of
connect-the-dots maps.
Denition 3.3.1. Let be an n-cycle and let M be the number of nondegenerate intervals
I maximal with respect to f
(0) = 0
or 1 according to whether (1) < (2) or (1) > (2), and H maps each subinterval
[(i 1)/M, i/M] linearly onto [0, 1] for i = 1, . . . , M 1.
The essential idea to check whether forces is to rst nd a certain canonical repre-
sentative, say Q, of in H
has a representative P of
such that
min
qQ
q
i
M
min
pP
p
i
M
, i = 1, . . . , M 1.
This algorithm is similar to an earlier one given by Jungreis [Jun91] which used
itineraries of points.
3.4 Properties of Forcing in the diagrammed data
To begin with, Baldwins original diagram [Bal87, Figure 4], showing the forcing relation
on all 36 cycle of length up to 5, is shown in Figure 3.1. In Figure 3.2, the forcing relation
on almost 200 cycles of order up to 8 has been diagrammed. This diagram contains the 36
cycles from Baldwins original diagram of forcing on all cycles of length up to 5. We adopt
the convention that if and only if is above and there is a line segment between
and . There are several aspects of this diagram which warrant being pointed out, as
they will be described at length in later chapters. The rst clear property of Figure 3.2
is that it is symmetric about a vertical axis extending from (1 2). This symmetry could
simply be an artifact of the way in which the forcing relations were diagrammed; however,
we will show in Chapter 4.1 that this is not the case. In fact, in Proposition 4.1.4, we will
prove that for every cycle there exists a cycle
with the property that forces if and
only if
forces .
Though more subtle than the symmetry in Figure 3.2, one characteristic is that there
is a long vein of cycles on which the forcing relation is a linear ordering; that is, for any
two cycles in this vein, forces or forces . We will outline in Chapter 4.2 that the
forcing relation restricted to unimodal cycles is a linear ordering.
Lastly, of the cycles that have been diagrammed, if one looks at the lengths of the
cycles immediately above a given cycle (Denition 5.1.5), then the length turns out
to be 2[[. In Chapter 5.1, we will prove that immediately above any cycle lies all of
22
Figure 3.1: Baldwins original diagram.
its doubles. Consequently, we note that the forcing order on cycles is not locally nite:
between any two odd length cycles lie innitely many even length cycles.
23
1324
1426375
14253
16374285
1627435
1635274
1634275
17435286
17436285
153264
132
18527436
1742635
15324
1743526
18453627
163425
18452736
1423
16284537
15384627
142635
15384726
16273845
15273846
15483726
143625 152634
15482637 17362845
1437526
17284536
1527436
13254
15274386
16274385
1526374
1362745
15463827
15372846
15462837
12
123
14725638
1462537
13425
1453627
15463728
15472638
136245
14825637
13825647
1372546
1374625
1473625
14837256
13524
14836257
1452736
12534
13847256
1453726
1472536
indicates new data
indicates data from Baldwin's digram
indicates unimodalpermutation
143526
1243
135264 136425
134265 126435
1274536
1374526
1435276
1436275
13856247
13854627
12854637
14537286
14536287
14375286 13728546
14628537
14538627
1375246 1364275
13862547
13852746
13745286
13746285 14862537
14625387
14725386
14853726
14738526
14527386
14627385
12853746
13854726
15342 13245
14523 14352
1342
13452
15432 12345
14532
12543
12453 13542
12354
1432 1234
12435
15243
15423
164253
1752634
18635274
18627435
13647258
1362547
13725648
135246
163254
1734265
1735264
1726435
18435276
18364275
18436275
18362745
18274536
162534
14235
16357248
1536247
1546237
15647238
15637248
145236
1624537
17254638
152436
16372548
14325
153426
1543627
1635427
1643527
1726354
16453728
17436528
17364528
17453628
18274365
18273645
15234
18256347
1724536
18254637
162435
1734526
1725346
1625347
18263547
17256348
17263548
18354627
124536
12573648
1254637
12564738
1356247
13657248 18634275
1763425
18743526
165324
18742635
1753264
12467358
1235647
123546
12473658
1246357
13572468
12365748
18642753
1764253
18752634
165423
18763425
1765324
18753264
Figure 3.2: Computation of for n 8.
24
Chapter 4
Global Structure of Forcing
In application, one major dierence that the ordering has when compared to the
Sharkovsky ordering, is that forcing is only a partial ordering (Theorem 2.1.6). Whereas
any pair of elements in N are mutually comparable under the Sharkovsky ordering (that
is,
s
is a linear ordering), this need not hold for any pair of elements in C under . For
example, consider the 4-cycles (1 2 3 4) and (1 4 3 2). It can be shown (see Chapter 3) that
the only 4-cycle that (1 2 3 4) forces is (1 3 2 4), and the only 4-cycle that (1 4 3 2) forces
is (1 4 2 3). Thus, (1 2 3 4),(1 4 3 2) and (1 4 3 2),(1 2 3 4). However, the 4-cycle (1 2 4 3)
forces both (1 4 2 3) and (1 3 2 4), yet is incomparable with both (1 2 3 4) and (1 4 3 2)!
These relationships are illustrated in Figure 4.1.
(1 2 3 4)
(1 2 4 3)
(1 4 3 2)
(1 3 2 4) (1 4 2 3)
Figure 4.1: Illustration of the complexity of on some 4-cycles.
This one example begins to illustrate the complex global structure of (C, ). However,
there is some symmetry and nice structure to (C, ) when we either consider certain
pairs of cycles or restrict ourselves to certain types of cycles.
4.1 The Flip Symmetry
One distinctive characteristic of Figure 3.2 is that it is symmetric about a vertical axis
extending from (1 2). The aim of this section is to show that to every cycle we can asso-
25
ciate a cycle
with the property that if and only if
. Therefore, by diagramming
and
across from each other, we obtain a sort of global symmetry in (C, ).
Denition 4.1.1. Given C
n
, the ip of , denoted
, is dened to be the n-cycle such
that
(k) = n + 1 (n + 1 k).
Example 4.1.2. Consider the 4-cycle = (1 3 2 4). The ip of is dened by
(1) = 4 + 1 (4 + 1 1) = 4
(2) = 4 + 1 (4 + 1 2) = 3
(3) = 4 + 1 (4 + 1 3) = 1
(4) = 4 + 1 (4 + 1 4) = 2
Thus,
= (1 4 2 3).
Example 4.1.3. As Figure 3.2 would suggest, the ip of a cycle may equal . Indeed
for the cycles (1 2), (1 2 4 3) and (1 4 7 3 8 5 2 6) this is precisely the case.
We have chosen to call
the ip for the following reason: the graph of the connect-
the-dots map f
is a 180
. A consequence of this
observation is that the Markov graph of
is obtained from the Markov graph of by
relabeling the vertex i by n i.
Proposition 4.1.4 ([Bal87, Proposition 4.2], [ALM93, Proposition 2.4.6]). Let , C.
Then, if and only if
.
Proof. Let C
n
and suppose that for some cycle . Let h : [1, n] [1, n] be the
orientation-reversing homeomorphism h(x) = x +n + 1. Then h f
h
1
= f
. This
can be seen by the fact that for any k 1, . . . , n,
(h f
h
1
)(k) = h(f
(h
1
(k)))
= h(f
(n + 1 k)) def. of h
= h((n + 1 k)) def. of f
= h(n + 1
(k)) def. of
= (n + 1
(k)) +n + 1
= f
(k).
Since f
is linear on each interval of the form (k, k +1) for k 1, . . . , n1, we conclude
that h f
h
1
is also linear on each such interval and connects the dots (k,
(k)).
Thus, this map is indeed f
h
1
)(X) is an orbit of type for f
. Therefore,
by Theorem 2.2.1,
. By using the same h, but starting with f
is shown in
Figure 4.2.
1 2 3 4 5
1
2
3
4
5
f
and I(
1
2
) = RRR = R
.
For x .80901542, we have
I(x) = RLRLRL = (RL)
,
and lastly, for x .382683432, we have
I(x) = RRC.
We shall call a sequence I of symbols L, C, R admissible if either I is an innite sequence
of Ls and Rs or if I is a nite (or empty) string of Ls and Rs followed by C. We note
that (trivially) every itinerary is admissible. Just like in Chapter 2, we can use the shift
operation on itineraries of unimodal maps. Next, let us introduce an ordering <
u
on the
set of all admissible sequences.
Denition 4.2.5. We dene L < C < R. Suppose now that A ,= B are two admissible
sequences. Let i be the rst index for which A
i
,= B
i
. We say that A <
u
B if either
O1. There are an even number of Rs in A
0
A
1
A
i1
= B
0
B
1
B
i1
and A
i
< B
i
; or
O2. There are an odd number of Rs in A
0
A
1
A
i1
and A
i
> B
i
.
If none of these hold then we say B <
u
A. This ordering is called the unimodal order or
the parity-lexicographic order.
On the surface, this denition looks entirely new. However, suppose f is a unimodal
map and consider the partition J
1
= [1, 0], J
2
= [0, 1]. Note that the sign of J
1
is +1
and the sign of J
2
is 1. Rewrite A and B in terms of J
1
and J
2
; that is, replace each
L (resp. R) with a J
1
(resp. J
2
). Then, in Denition 4.2.5, condition O1. amounts to
saying that the product z(A
0
)z(A
1
) z(A
i1
) is positive and A
i
< B
i
, and condition
O2. amounts to saying that the product z(A
0
)z(A
1
) z(A
i1
) is negative and A
i
> B
i
.
Thus, looking back at Denition 2.2.3, it becomes clear that <
u
is equivalent to the order
<
, we can use the unimodal order to show that the order of itineraries
of two points gives us the order of the two points.
Lemma 4.2.7 ([CE80, Lemma II.1.2]). If f is unimodal and I(x) <
u
I(y) then x < y.
28
Proof. We prove this by induction on the smallest index i for which I
i
(x) ,= I
i
(y). The
statement is true when i = 0. Suppose that i > 0 and that we have already shown the result
for i 1. When I
0
(x) = I
0
(y) = L, then if I(x) <
u
I(y), we nd that I(f(x)) <
u
I(f(y)).
But, I(f(x)) = I
1
(x)I
2
(x) , and hence I
i1
(f(x)) ,= I
i1
(f(y)). By the induction
hypothesis f(x) < f(y). Since I
0
(x) = I
0
(y) = L, we have x, y [1, 0) and since
f
[1, 0]
is strictly increasing, f(x) < f(y) implies x < y.
When I
0
(x) = I
0
(y) = R, then if I(x) < I(y) we nd I(f(x)) > I(f(y)) since I(f(x))
has one R less than I(x) before i. Hence f(x) > f(y). Since I
0
(x) = I
0
(y) = R, we have
x, y (0, 1] and since f
[0, 1]
is strictly decreasing, f(x) > f(y) implies x < y.
Lemma 4.2.8 ([CE80, Lemma II.1.3]). If f is unimodal and x < y then I(x)
u
I(y).
Proof. Since <
u
is a linear ordering, if x < y and I(x) > I(y), then Lemma 4.2.7 implies
a contradiction.
Denition 4.2.9 ([CE80, p.87]). Let f be a unimodal function and let A be an admissible
sequence. Suppose that for all k 0,
k
(A) < I(1) if I(1) is innite,
k
(A) < (DL)
k
(A) < (DR)
or f = some conjugate to f
d|n
d odd
(d)2
(n/d)1
,
where is the Mobius function.
Aside from the results presented in this chapter, there are very few other results avail-
able which convey a good conception of the global structure of forcing in the same way
that Sharkovskys theorem does for the relation
s
[Hal90]. However, in the next chapter
we will show that the local structure of forcing is completely describable.
31
Chapter 5
Extensions of Cycles
Whereas in the preceding chapter we analyzed the global features of (C, ), this chapter
takes a look at the local structure of forcing by considering certain cycles called extensions.
In Section 5.1, we will study doublings of cycles. It will be shown that doublings completely
describe the local structure of forcing: immediately above any n-cycle lies all of its 2
n1
doubles. Doublings are in fact only one particular type of an extension, and in Section 5.2,
we consider more general extensions of cycles such as block structures and -extensions.
These concepts will allow us to further characterize sucient conditions for one cycle to
force another and give us a better idea of the local structure of forcing.
5.1 Doublings
The basic idea of a doubling of a cycle C
n
is that we thicken each point 1, . . . , n
into a block containing two points, and then these n blocks are permuted according to .
Denition 5.1.1. If n is even, then a cyclic permutation of length n is a double (or
doubling)
1
if it cyclically permutes the n/2 sets 1, 2, . . . , n1, n. In this case, is said
to be a double of the permutation dened by (i) = j if (2i 1, 2i) = 2j 1, 2j.
Example 5.1.2. Let be the 6-cycle (1 5 3 2 6 4). Then cyclically permutes the pairs
1, 2, 3, 4, 5, 6 and is a double of the permutation (1 3 2). This is illustrated in Figure
5.1. Note that the boxes are permuted according to (1 3 2 ).
Let us rst show how one may go about constructing doublings of a given n-cycle
. First, write out the points 1, . . . , 2n on a horizontal line and draw a block around
each successive pair of points 2i 1, 2i. Label the n blocks by P
1
, P
2
, . . . , P
n
so that
P
i
= 2i 1, 2i. Starting in block P
1
with the point 1, draw an arrow to a point
in P
(1)
which we will call (1). From (1) draw an arrow to a point in P
2
(1)
and
call this point
2
(1). Continue this process n 2 times so that we have the n points
0
(1) = 1, (1),
2
(1), . . . ,
n1
(1) with
i
(1) P
i
(1)
. Since, P
n
(1)
= P
1
, we have no
choice but to draw an arrow from
n1
(1) to 2. Now we begin the process over again, only
this time we rst draw an arrow from 2 to the point in P
n+1
(1)
= P
(1)
which previously
did not have an arrow drawn to it and label this point
n+1
(1). After n 1 more steps,
1
We follow the language of [BCMM92, BC92a, MN91] by referring to 2-extensions as doubles or dou-
blings, as this seems to be the standard now. However, in [Ber87], Bernhardt refers to a doubling by
saying that a cycle is splittable, and in [Bal87], Baldwin says that a cycle is divisible by 2 if it is a double.
32
1 2 3 4 5 6
Figure 5.1: The 6-cycle (1 5 3 2 6 4) as a doubling of (1 3 2).
we are again back in the block P
n1
(1)
. To nish the construction, we draw an arrow
from the point
2n1
(1) back to 1. If we follow the arrow path that leaves the point
1, we observe that this path hits every number before hitting 1 again; consequently, the
permutation on the 2n points is in fact a cyclic permutation. Moreover, that is a
doubling of is clear from the above construction and Denition 5.1.1.
The above construction of a doubling leads one to the following question: how many
doublings are there of a given n-cycle? Observe that for the rst n 1 steps in the
construction of above, we had precisely two choices. However, at the n
th
step, an arrow
had to be drawn back to the point 2 and from then on, all of the choices were determined.
This argument proves the following result:
Proposition 5.1.3. For any n-cycle , there are 2
n1
doublings of .
One of the most important uses of doublings is in classifying immediate successors and
predecessors in (C, ). Intuitively, an immediate successor to a cycle is a cycle such
that and there is nothing else in between and . For example, if we consider the
integers partially ordered by the less-than-or-equal-to relation , then 4 is an immediate
successor to 3, since 3 4 and x Z : 3 x 4 = . The following denitions make
precise what we mean by immediate successors and predecessors.
Denition 5.1.4. If is a cycle, let For() denote the set of cycles forced by , i.e.
For() = C : .
Denition 5.1.5. Let , C. Then is said to be an immediate successor to (or
is an immediate predecessor of ) if ,= and For() = For() .
A priori, given an arbitrary partially ordered set, one need not have an immediate
predecessor for every element. For example, if we consider the reals instead of the integers
partially ordered by , then there is no immediate predecessor to 3. Nevertheless, with
regards to the forcing order, Bernhardt proved that the following are equivalent [Ber87,
Theorem 1.12]:
B1. has an immediate predecessor;
B2. is a doubling;
B3. there is only one nonrepetitive loop from J
1
to itself in the Markov graph of ;
B4. there does not exist a loop in the Markov graph of whose orbit type is .
33
We will prove a result similar to this, which classies the most important properties of
doublings.
Theorem 5.1.6 (Doubling Theorem). Let C
n
, P = p
1
< p
2
< < p
n
=
1, 2, . . . , n and consider the connect-the-dots map f
that exhibits .
2. There is a loop in the Markov graph of with loop type .
3. is not a doubling.
Lemma 5.1.7 ([Ber87, Lemma 2.4]). Suppose that C
2n
is a doubling of C
n
. There
is precisely one loop in the Markov graph of beginning (and ending) at J
1
. Moreover,
this loop has orbit type .
Proof. The Markov graph of has 2n 1 vertices J
1
, . . . , J
2n1
. Since is a doubling,
cyclically permutes the n pairs P
k
= 2k 1, 2k, k = 1, . . . , n. In particular, since is
a doubling of , we have (P
k
) = P
(k)
. This implies that for each odd index 2
k
(1) 1,
the vertex J
2
k
(1)1
has precisely one edge leaving it, and this edge goes to the vertex
J
2
k+1
(1)1
. Therefore, the only loop beginning at J
1
is the loop
J
1
J
2(1)1
J
2
n1
(1)1
J
1
.
This loop is of length n, has no arrows leaving the loop, and corresponds to an orbit of
type for the map f
.
As a result of Lemma 5.1.7, we have the following:
Corollary 5.1.8. If is a doubling of , then forces .
Lemma 5.1.9 ([Bal87, Lemma 4.7]). Suppose that C
2n
is a doubling of C
n
. The
subgraph of the Markov graph of consisting of all even-indexed vertices and all arrows
from a vertex of even index to another vertex of even index is graph isomorphic to the
Markov graph of .
Proof. Let k, i 1, . . . , n. By denition, a directed edge exist from J
2k
to J
2i
when
f
([2k, 2k + 1]) [2i, 2i + 1]. There are two possibilities for this to happen, either f
is
increasing on [2k, 2k +1] or f
(q
i
) = q
(i)
. Trivially, P exhibits
f
(x) f
(y)]) .
is a symmetric function, and locally, tells us whether f
is increasing or decreasing.
In particular, suppose that p is a critical point and x < p < y are two nearby points.
Then, if (p, x) = 1 and (p, y) = 1, by denition of , we see that p must be a relative
maximum (so long as x and y are close enough to p that there are no other extrema in
[x,y]). On the other hand, suppose that x
1
< x
2
< x
3
< < x
n
are points such that
(x
1
, x
2
) = (x
2
, x
3
) = (x
j
, x
j+1
),
but (x
j+1
, x
j+2
) is not equal to (x
1
, x
2
) for some 1 < j < n2. We see that essentially
detects the existence of critical points, in particular in the above example, there must be
a critical point in [x
j
, x
j+1
].
Now, since p
c
and q
c
are critical points of the same type and q
i
q
c
, p
c
), we must
have (q
i
, q
c
) = (q
i
, p
c
). But, only changes sign in the presence of a critical point,
35
and since all critical points of f
(q
c
) < f
(q
i
) < f
(p
c
). But, q
c
is a relative maximum of Q, so we should actually
have f
(q
c
) > f
(q
i
). Since all critical points of f
that lies in [q
c
, p
c
]. From what we have, we cannot decide
whether q
c
< p
c
< q
i
or q
i
< p
c
< p
c
; however, this is not important. By Lemma
5.1.10, we cannot have q
c
[q
c
, p
c
]. But, since p
c
< p
c
, the spatial labeling on Q
requires that q
c
< q
c
(Figure 5.3). The argument begins again: q
c
, a point from Q,
q
c
q
i
p
c
q
c
q
i
p
c
q
c
p
c
p
c
(q
c
) > f
(q
i
) > f
(p
c
).
This does not aect the reasoning for the existence of a critical point to the left of
[q
c
, p
c
].
3. This leaves the case when p
c
< q
i
< q
c
, which is handled similarly as the above
cases.
36
Lemma 5.1.12. For each i, q
i
, p
i
) is mapped monotonically under f.
Proof. Suppose that for some i 1, . . . , n, q
i
, p
i
) is not mapped monotonically under
f
for each i.
Now, we are ready to prove the Doubling Theorem:
Proof of Theorem 5.1.6.
2 3. Suppose that C
2n
is a doubling of C
n
and let = (J
i1
, . . . , J
i2n1
, J
i1
) be a
loop in the Markov graph of of length 2n. If J
1
is a vertex in , then by Lemma
5.1.7, modulo the shift operation, we have
= (J
1
, J
2(1)1
, , J
2
n1
(1)1
, J
1
, J
2(1)1
, , J
2
n1
(1)1
, J
1
)
and, moreover, does not correspond to an orbit of length 2n. Therefore, J
1
is not
a vertex in . Since no arrow in the Markov graph of goes from an odd-index
vertex to an even-index vertex, we must have consisting of only vertices with even
indices. By Lemma 5.1.9, the subgraph consisting of only even-index vertices is
graph isomorphic to the Markov graph of . Therefore, if were to correspond to
an orbit type , then there would be a loop in the Markov graph of of orbit type
, implying that . But, by Corollary 5.1.8, is a doubling of , so , and
therefore by the antisymmetry of forcing (Theorem 2.1.5), we must have = .
Contradiction. Thus, if is a doubling of , then there is no loop in the Markov
graph of having orbit type .
3 1. By contrapositive. Suppose that Q ,= P is another orbit of f
that exhibits . By
Lemma 5.1.12, f
is monotone on q
i
, p
i
) for each i, so, in fact, any iterate of f
is
monotone on q
i
, p
i
). Therefore, since P and Q are orbits of n-periodic points, f
n
takes 1 to 1, q
1
to q
1
and is monotone on [1, q
1
]. This shows that f
n
is the identity
on [1, q
1
], and so has derivative equal to 1 on the entire interval [1, q
1
]. We want
to show that, in fact, f
n
(all of which
are in P) for such an iterate. Now, as f
k
(P) = P for all k 0, once an iterate of x
lands in P, it stays in P. Since k < n, this implies f
n
be the
iterate that takes the point 2 to 1. The slope 1 condition tells us that f
k
(1)
must be equal to 2, and then we are done: since is a cycle, if the same iterate takes
1 to 2 and 2 to 1, then is a doubling.
1 2. Suppose that P is the only orbit that exhibits . Dene a loop in the Markov
graph of in the following way: choose small enough so that the length of the
interval f
k
p
= J
i
: 1 i n 1, k i.
Proof of (1). The invariance of
p
follows directly from the denition of a block
structure since the block
P
i
= J
ki+1
, J
ki+2
, . . . , J
ki+k1
: 1 i m
moves to
P
(i)
= J
k(i)+1
, J
k(i)+2
, . . . , J
k(i)+k1
: 1 i m,
and as is an m-cycle, no arrow leaves
p
and goes to a vertex with an index divisible by
k.
Proof of (2). The map taking the index j of a vertex in
g
to j/k shows that
g
is
isomorphic to the Markov graph of .
As an immediate consequence of Theorem 5.2.3, we get the following result which will
be important in the forcing proof of Sharkovskys theorem:
Corollary 5.2.4. If C
n
has a block structure over C
m
, then .
In Theorem 2.1.5 we proved that forcing is a partial order, hence a transitive relation.
Therefore, as a result of Corollary 5.2.4, we also have
Corollary 5.2.5. Let , , C. If has a block structure over and , then .
The next type of extension that we will consider is a particular type of block structure.
Denition 5.2.6. Let be an n-cycle and be an m-cycle for some n, m N. A nm-cycle
is called an extension of by or a -extension of if
1. has a block structure over .
2. There is some block P
i
such that P
i
is an orbit of type under f
m
.
If in addition, satises
3. f
in the blocks of P
2
and P
3
was chosen so
that satises Condition 1 of Denition 5.2.6, i.e., has a block structure over . The
graph of f
is shown in Figure 5.6. Observe that if we were to collapse each of the blocks
to a point, then this graph is the graph of f
(P
i
) is either or
for
i = 1, 2, . . . , m.
Proof. Let j
0
denote the index of the block guaranteed by Denition 5.2.6 which is an
orbit of type under f
n
. For j ,= j
0
, the map
f
j
:= f
n
P
j
: P
j
P
(j)
41
is a homeomorphism. Therefore, the map
F = f
1
(j)
f
2
(j)
f
j0
f
(j)
f
j
: P
j
P
j
can be written in the form h
1
F
j0
h
2
for homeomorphisms h
1
and h
2
. Thus, since P
j0
is an orbit of type under f
n
, P
j
is an orbit of type or
depending on whether h
1
and
h
2
are orientation preserving or reversing.
Now, given , a simple -extension of some cycle , we want to know which cycles are
forced by .
Proposition 5.2.9 ([ALM93, Proposition 2.10.6]
2
). Let be a simple -extension of
and suppose that . Then either or is a
forced by
or
. If ,= then in the last case both
,= and
,=
.
Proof. Assume that C
n
, C
m
and let P = 1, 2, . . . , n. Using the denition of a
-extension of , construct the m-blocks P
1
, . . . , P
m
. Since , there is an orbit R of f
m
i=1
P
i
) is f
-invariant. Hence
either R M = or R M.
Assume rst that R M = . We collapse each block P
i
) into a point and we
obtain from [1, n] an interval which can be stretched back to [1, n]. More precisely, there
exists a non-decreasing continuous map : [1, n] [1, n] such that is constant on
P
i
) and increasing on each component of [1, n]M. Then, the map
f dened by
f(x) =
f
(
1
(x)) is a continuous map of [1, n] and satises f
=
f . The set (M)
is an orbit of
f and by the denition of an extension, it exhibits the cycle . Moreover,
f
is (M)-monotone. Since is non-decreasing, (R) is an orbit of
f exhibiting the cycle
. Therefore, by the Theorem 2.2.1, .
Now, assume that R M. Denote R
i
= RP
i
). If f
(P
i
) = P
j
, then f
(P
i
)) = P
j
)
and hence R has a block structure over and f
-extension of for
. However, R
1
is an orbit of the P
1
-monotone map f
m
P
1
)
.
Therefore, by the forcing theorem, we conclude that the rst-return pattern of the block
P
1
under f
m
forces
is either or
, so we conclude that
or
.
Suppose that =
) and
(R
1
, f
m
(p
j
) = p
1(j)
and f
(r
j
) = r
2(k)
for cyclic permutations
1
and
2
. If f
is monotone on P
i
), then
1
(j) =
2
(j) for j = (i 1)s + 1, (i 1)s +
2, . . . , (i 1)s + s where s = n/m. However, if f
1
(k) and
2
(k) for k 1, 2, . . . , n(i 1)s + 1, (i 1)s + 2, . . . , (i 1)s + s (where
they are equal) and by the cycles (P
1
, f
m
) and (R
1
, f
m
1
=
2
, which implies = .
Lemma 5.2.10 ([ALM93, Lemma 2.11.4]). Let be a doubling of and suppose that
. Then either = or .
Proof. Assume that ,= . Since the unique 2-cycle (1 2) only forces itself and trivial cycle
(1), by Proposition 5.2.9, either or is a (1)-extension of . However, a (1)-extension
of is itself, so in every case .
2
Here dening and
to be equivalent would simplify the statement of this theorem (see Chapter 8).
42
Lemma 5.2.10 proves the earlier assertion that given a cycle , every doubling of is
an immediate successor to in the forcing order. The following is a sort of partial converse
to Proposition 5.2.9:
Proposition 5.2.11 ([ALM93, Lemma 2.10.8]). Let ,
or
-extension
of .
Proof. Assume that C
n
and C
m
. Since is a simple -extension of , has a
block structure P
1
, . . . , P
m
over . Let us give these blocks the temporal labeling so that
f
(P
i
) = P
i+1
(where addition of indices is modulo m). Proposition 5.2.8 tells us that
the block P
1
has the rst-return pattern of either or
; that is, P
1
is an orbit of type
or
for the map f
m
or
(Q
1
) and let Q =
m
i=1
Q
i
. Then we have
Q
i
) P
i
) for each i. Moreover the orbit Q exhibits a cycle that necessarily has a block
structure over . Since is a simple extension, the map f
is monotone on at east m1
intervals P
i
). Therefore, the cycle that Q exhibits is in fact a simple
-extension of .
By Theorem 2.2.1, forces this cycle. The above argument needs only slight changes if
we assume that
.
43
Chapter 6
Forcing-Minimal Cycles
This chapter aims to study minimality properties of forcing. The results presented here
provide the nal tool necessary to complete the forcing proof of Sharkovskys theorem
(Chapter 7). Looking ahead, to prove Sharkovskys theorem, we will begin with some
continuous map that has a periodic point of period n. This periodic point exhibits some
n-cycle, say , and it will be helpful to know if there exists some elementary n-cycle that
forces. If this were to be the case, then we could assume without loss of generality from
the beginning that the periodic point exhibits such an elementary cycle. We will show
that there are two notions of elementary cycles that of primary cycles (Denition 6.1.1)
and simple cycles (Denition 6.2.6) and will show that these notions actually coincide,
even though on the surface they are entirely dierent concepts. The general exposition of
this chapter very closely follows [ALM93, Chapter 2.9-2.11].
6.1 Primary Cycles
Looking at Figure 3.2, we see that among all 4-cycles, (1 3 2 4) and (1 4 2 3) are the only ones
that do not force any other 4-cycles. Considering 5-cycles now, we see that (1 3 2 4 5) and
(1 5 3 2 4) are minimal as well in this respect. This section aims to characterize properties
of such forcing-minimal cycles.
Denition 6.1.1. An n-cycle will be called primary
1
if it does not force any other n-cycle.
Equivale
Observe that for n = 1, the trivial cycle (1) is primary; for n = 2, (1 2) is the unique
primary cycle; and, for n = 3, (1 2 3) and (1 3 2) are the only 3-cycles and neither forces
the other, so they are primary. In the introduction, it was observed that there seem to be
two primary 4-cycles as well as two primary 5-cycles. In particular, note that (1 3 2 4) is
the ip of (1 4 2 3), and (1 3 2 4 5) is the ip of (1 5 3 2 4). The following proposition tells
us that this is always the case.
Proposition 6.1.2 ([ALM93, Proposition 2.9.1]). If is primary then
is primary.
Proof. Suppose that C
n
is primary. If
is not primary, then there exists some
C
n
, ,=
such that
. By Proposition 4.1.4, this implies that , and since
1
The notion of primary cycles, which we have adapted from [ALM93], is referred to as -minimal
in [Bal87].
44
is primary, = . But, since equal cycles have equal ips, we must have that
= .
Contradiction.
Thus, if we know that an n-cycle is primary, we know that there is at least one other
n-cycle that is primary. The next theorem will justify our interest in primary cycles by
showing that every n-cycle forces a primary cycle.
Proposition 6.1.3 ([ALM93, Theorem 2.9.4]). If C
n
, then there exists a primary
cycle C
n
such that .
Proof. Let C
n
. Theorem 2.1.5 tells us that is a partial order, and therefore,
restricted to the set of all n-cycles is also a partial order. This set is nite for any n N,
so there must exist some C
n
such that and forcing-minimal; that is, is
primary.
Before moving onto the concept of a simple cycle, we note that there is an interesting
corollary of Proposition 5.2.9 that relates the primarity of a cycle to extensions.
Corollary 6.1.4 ([ALM93, Corollary 2.9.5]). Let be a -extension of , say C
n
. If
is primary and does not force any n-cycles, then is primary.
6.2 Simple Cycles
Some of the earliest types of cycles on which the forcing relation was analyzed are the
so-called simple cycles. In [Ber84], Bernhardt described the structure of forcing on simple
permutations of a power two. If n is even, the essential idea of a simple cycle is that we have
a block structure consisting of two blocks P
1
= 1, . . . , n/2 and P
2
= n/2 + 1, . . . , n,
and P
1
is sent to P
2
and vice-versa. Thus, if we consider n = 2k points equally spaced on
a horizontal line, then one can draw a vertical line between the rst k points and the last
k points, and all points left (resp. right) of this line move to the right (resp. left) of the
line under an even simple cycle.
Simple cycles of an odd length are known as
Stefan cycles after the Czech mathemati-
cian Peter
Stefan who explicitly used them in his proof of Sharkovskys theorem [
Ste77].
Denition 6.2.1. Let q = 2k 1 for some k 2. A q-cycle is called a
Stefan cycle if
when expressed in cycle notation it is equal to either
(k (k 1) (k + 1) (k 2) (k + 2) (k j) (k +j) (2k 1))
or
(k (k + 1) (k 1) (k + 2) (k +j) (k j) (2k 1) 1) .
Equivalently, for q odd, q 3, we say that C
q
is a
Stefan cycle if there exists a
temporal labeling of P = p
1
, . . . , p
n
= 1, . . . , n such that either
p
n
< p
n2
< < p
5
< p
3
< p
1
< p
2
< p
4
< < p
n3
< p
n1
or
p
n1
< p
n3
< < p
4
< p
2
< p
1
< p
3
< p
5
< < p
n2
< p
n
The denition of a
Stefan cycle given above is far more complicated than the actual
concept of a
Stefan cycle. Essentially, for q = 2k 1, one starts at the point k, moves
left to the point k 1, then right to the point k + 1, then left to the point k 2, and so
45
on continuing in this spiraling out fashion until one reaches 2k 1, at which point the
cycle starts over at k. This construction can also be done by starting out moving from k
to k + 1, hence the two dierent forms of a
Stefan cycle given in Denition 6.2.1.
Example 6.2.2. Let us illustrate the characteristic spiraling out property of
Stefan
cycles described above. Let q = 5 = 2 3 1. According to Denition 6.2.1, there are two
.
Next, consider the case when m < n. We need to show that f
.
Proposition 6.2.5 ([ALM93, Proposition 2.11.3]). Each Stefan cycle is primary.
Proof. Let C
n
be a
Stefan cycle. Lemma 6.2.3 describes the Markov graph of . In
particular, this lemma shows us that there are precisely two loops of length n in this graph:
1
= J
1
J
2
J
n1
J
1
J
1
and
2
= J
1
J
1
J
1
. Now,
2
is not
a simple loop so it does not correspond to an orbit of length n for f
Stefan extension.
Observe that for q 3 and k = 0 (as above), there are precisely two simple cycles,
namely the
Stefan q-cycles. However, for k 1 there can be very many simple cycles.
We are using a notion of a simple cycle that is somewhat stronger than the notion found
in [Ber84] and [Bal87]. In these papers, a simple cycle need only have a 2-block structure
and no
Stefan mixing of the blocks is required. An interesting result of Bernhardt [Ber84,
Lemma 2.2] says that there 2
2
n
(n+1)
simple permutations of period 2
n
.
Example 6.2.7. Let us begin with some examples of simple 6-cycles. Since 6 = 3 2
1
,
Denition 6.2.6 tells us to extend the trivial cycle (1) by the two cycle (1 2), then extend
this by a
Stefan 3-cycle. Recall that all 3-cycles are
Stefan cycles, so let us extend (1 2)
by (1 3 2). This gives us either the 6-cycle = (1 6 3 4 2 5) or = (1 6 2 5 3 4). Now, by
analyzing Figure 3.2 in which all 6-cycles are diagrammed one sees that both and
are forcing-minimal, i.e., they forces no other 6-cycles. Thus, and are primary and so
2
The denition that Baldwin uses for simple [Bal87, Denition 4.8] refers only to a 2-extension. Thus,
for Baldwin, there exist simple cycles that are not primary.
47
by Proposition 6.1.2,
= (1 4 3 5 2 6) and = (1 5 2 4 3 6) should also be primary. Looking
again at Figure 3.2, we see that indeed
and are primary. Lastly, we note that had we
chosen to extend (1 2) by (1 2 3) instead of (1 3 2), we would have have produced
and .
Consequently, the above discussion gives all of the primary and simple 6-cycles. Figure
6.3 illustrates the fact that all of the simple 6-cycles have the characteristic two block
and spiraling
Stefan structure.
1 2 3 4 5 6 1 2 3 4 5 6
1 2 3 4 5 6 1 2 3 4 5 6
= (1 6 3 4 2 5) = (1 6 2 5 3 4)
= (1 4 3 5 2 6) = (1 5 2 4 3 6)
Figure 6.3: The primary and simple 6-cycles.
Example 6.2.7 points towards the most important theorem of this section, namely
Theorem 6.2.10, which says a cycle is primary if and only if it is simple. This will be
one of the keys to proving Sharkovskys theorem via forcing. We will rst work towards
showing that simple primary. To do so, wee need a couple of lemmas. The rst lemma,
Lemma 6.2.8, essentially tells us that the simple cycles of length 2
k
form a tree in the
forcing relation. The second lemma, Lemma 6.2.9, shows that any primary extension of a
simple cycle of length a power of 2 is itself primary.
Lemma 6.2.8 ([ALM93, Lemma 2.11.5]). Let be a simple cycle of length 2
k
for some
k 0. If forces , then is a 2
j
-cycle for some j k such that C
2
j .
Proof. This result follows almost immediately from the section on doublings (Section 5.1).
We prove this by induction on k. For k = 0, the result is trivial. Now, assume that the
result holds for simple cycles of length 2
k1
. If is a simple 2
k
-cycle, then is a doubling
of a simple cycle of length 2
k1
. Therefore, if forces , then by Lemma 5.2.10 and
the inductive hypothesis, we must have either is a 2
k
-cycle or is a 2
j
-cycle for some
j k 1.
Lemma 6.2.9 ([ALM93, Lemma 2.11.6]). Let be a simple 2
k
-cycle for some k 0. If
is a nontrivial extension of , then is primary.
Proof. The restriction that is not an extension guarantees that the length of is strictly
greater than the length of . Now, from the previous lemma, Lemma 6.2.8, it follows that
48
does not force any cycle of length larger than 2
k
. By Corollary 6.1.4, we conclude that
is indeed primary.
Theorem 6.2.10 ([ALM93, Theorem 2.11.7]). A cycle is primary if and only if it is
simple.
Proof that simple implies primary. Let be a simple cycle. First, if is of odd length,
say q 3, then is a
Stefan cycle. By Proposition 6.2.5, is primary. Next, suppose
that is of length 2
k
for k 0. As remarked after the denition of primarity, if k = 0
or k = 1 then must be primary. If, however, k > 1, then since (1 2) is primary, taking
the contrapositive of Lemma 6.2.9, we see that is primary. Lastly, suppose that is of
length q 2
k
with q 3 odd. Then, combining both Lemma 6.2.9 and Proposition 6.2.5,
we get that is primary.
To prove the reverse direction that a primary cycle is simple requires a considerable
amount of work that would amount to reproducing [BC86] or several sections of [ALM93].
Therefore, we will try to give a detailed sketch of how one shows that primary implies
simple.
Let us rst show how Alseda, LLibre and Misiurewicz proved the result in [ALM93].
First, they show that depending on the length of a cycle, a primary cycle is either a
Stefan
cycle or has a block structure over (1 2).
Proposition 6.2.11 ([ALM93, Proposition 2.11.8]). Let be a nontrivial primary cycle.
Then the following statements hold.
(a) If [[ is even, then has a block structure over (1 2).
(b) If [[ is odd, then is a
Stefan cycle.
The proof of this proposition requires the result that given any n-cycle , there exists
an interval map f such that f has an n-periodic point x, f is O(x)-monotone and O(x)
is the unique orbit of f that exhibits [ALM93, Proposition 2.5.4]. This can be proven
by using the pouring water construction [ALM93, p. 54]. Observe that this result is
almost guaranteed by the doubling theorem (Theorem 5.1.6) which says that so long as
is not a doubling, P = 1, 2, . . . , n is the unique invariant set of f
that exhibits .
Now, as an immediate consequence of Proposition 6.2.11 we have that for nontrivial
odd-lengthed cycles, primary implies simple. This is because given an odd-length primary
cycle , Proposition 6.2.11 implies is a
Stefan cycle, and by the denition of a simple cycle,
a
Stefan cycle is simple. What remains to be shown is that any even-lengthed primary
cycle (which by Proposition 6.2.11 has a block structure over (1 2)) is in fact a simple cycle.
To do so, requires very careful analysis of these block structures. Ultimately, one shows
that for a primary even-lengthed cycle , the rst-return patterns of the blocks are
Stefan
cycles of the same length and so is a
Stefan extension of some cycle [ALM93, Lemmas
2.11.9 - 2.11.16]. Moreover, they show that if is a primary n 2
k
-cycle, then has a block
structure over a simple 2
k
-cycle [ALM93, Proposition 2.11.17].
Block and Coppel [BC86] and Baldwin [Bal87, Theorem 4.9] (which heavily uses Block
and Coppels paper) proved the result that primary implies simple in a slightly dierent
way. Block and Coppel prove the following powerful theorem which essentially completes
the proof:
Theorem 6.2.12 ([BC86, Theorem 1]). If f has an orbit of period n, then it has a
(strongly) simple orbit of period n.
49
Now, suppose that is a primary cycle. By Theorem 6.2.12, f
h = h f
Ste77] P.
Stefan, A theorem of
Sarkovskii on the existence of periodic orbits of con-
tinuous endomorphisms of the real line, Comm. Math. Phys. 54 (1977), no. 3,
237248. MR MR0445556 (56 #3894)
[WR87] A. Weiss and T. D. Rogers, The number of orientation reversing cycles in
the quadratic map, Oscillations, bifurcation and chaos (Toronto, Ont., 1986),
CMS Conf. Proc., vol. 8, Amer. Math. Soc., Providence, RI, 1987, pp. 703
711. MR MR909946 (89b:58174)
57