CHAOTIC ATTRACTORS IN ONE-DIMENSION GENERATED BY A

SINGULAR SHILNIKOV ORBIT

Krista J. Taylor
College of Arts and Sciences
Shawnee State University
Portsmouth, OH 45662

Bo Deng
Department of Mathematics and Statistics
University of Nebraska-Lincoln
Lincoln, NE 68588

December 1999

Abstract: Chaotic attractors containing Shilnikov’s saddle-focus homoclinic or-

bits have been observed in many physical systems. Past and current researches of
this type of Shilnikov homoclinic phenomena have focused on the orbit and nearby
structures only. In this paper we will look at the role such orbits play in a type
of attractors, which arise from one-dimensional return maps at the singular lim-
its of some singularly perturbed systems. Results on symbolic dynamics, natural
measures, and Lyapunov exponents are obtained for a sequence of a one-parameter
caricature family of such attractors.
 1

1 Introduction
In the 1960’s Shilnikov studied the dynamics of a 3-dimensional system of ordinary
differential equations containing an orbit homoclinic to a saddle-focus equilibrium
point. In a series of papers he showed that such a system can produce chaos in a
neighborhood of the orbit. See Šil’nikov [20], [21], [22]. His saddle-focus homoclinic
orbit is one of the simplest structures from which chaos occurs. This type of chaos
phenomena have been observed in many physical systems, such as neural dynamics,
chaotic circuits, nonlinear laser systems, and fluid dynamics, see for examples [4],
[5], [12], [14], [15], [18], [24]. His result has also been refined and extended by others,
see e.g., [7], [10], [24].

All prior results in the literatures on Shilnikov’s orbit have been restricted to
some subsets of a small neighborhood of the orbit. However the reason that most
of the related chaotic phenomena are observable is due to the existence of a chaotic
attractor containing such an orbit. Therefore, it is equally important, if not more, to
study the attractor which the orbit spawns. The purpose of this paper is to study a
type of such attractors that arise from singularly perturbed saddle-focus homoclinic
orbits in some 3-dimensional singularly perturbed differential equations.

A brief outline of this paper is as follows. First we will follow the modeling
strategy presented in [8, 9], as well as in [18, 19], to derive our 1-dimentional models.
The singularly perturbed system in this paper has the form:

ẋ = f (x, y, z), ẏ = g(x, y, z), εż = h(x, y, z).

The attractors are constructed by limiting orbits of the singular perturbed system.
A limiting orbit is one that connects the flow of the slow subsystem to the flow of the
fast subsystem. The singular Shilnikov orbit is generated in this way so that it starts
and ends at one equilibrium point. It can be proved by Fenichel’s geometric the-
ory of singular perturbations [13] together with Bonet’s result on turning points [3]
that a Shilnikov orbit Γε persists for a small perturbation of the system at each
0 < ε ¿ 1. The singular Shilnikov orbit corresponds to the limit Γ0 = limε→0 Γε .

Studying the limiting system is considered important because whatever limiting

structure presents at ε = 0 can be considered as the origin of the perturbed, full
structure. Figure 1 illustrates the type of the 3-dimensional systems that we are
considering in this paper. For such a system, the slow subsystem is 2-dimensional
in x and y, thus any 2-dimensional Poincaré section to the flow, in particular, to the
2-dimensional slow manifold produces a 1-dimensional Poincaré return map at the
limit ε = 0. More specifically, a 1-dimensional return map is constructed by picking
an interval that crosses the sub-flow on the slow manifold transversally and map-
ping it back into itself via the flows. The 1-dimensional return map of our system
 2

is piecewise continuous as illustrated in both Figure 1 and Figure 2(a).

The full return map is difficult to analyze because of the infinite derivative at
x = A+ 0 as shown in both Figure 1 and Figure 2(a). Instead we will consider a
simplified, caricature model for the full return map. The caricature model is given
as the family of piecewise linear maps
(
2x, if 0 ≤ x ≤ 1/2
fλ (x) =
4λ | x − 3/4 |, if 1/2 < x ≤ 1,
where λ ∈ [0, 1] and A0 , B0 , and C0 are represented by 1/2, 3/4, and 1 respectively,
see Figure 2(b). The degree of difficulty between treating the full return maps and
the simplified piecewise linear models is similar to that between the logistic map
and the tent map for which the former is much harder to analyze because of its
vanishing derivative at its only critical point. Nonetheless we believe that results on
the piecewise linear models should shed some light on the more complex return maps.

Most of our results, analyses, discussions are for the piecewise linear models at
a sequence of parameter values λ = 1/2i , i = 1, 2, 3, . . .. This limitation is due to
what our method of this paper can do at this point. Nevertheless the sequence does
give a fair representation to the bifucation parameter range λ ∈ [0, 1] as we will see
throughout the paper. We will obtain results on the following aspects of the attrac-
tors: symbolic dynamics, invariant measures, and Lyapunov exponents for sensitive
dependence on initial conditions.

For symbolic dynamics, we will show that the piecewise linear maps at λ = 1/2i ,
i = 1, 2, 3, . . . are conjugate to a subshift on 3 symbols. More specifically, the
conjugate map is the itinerary map defined as

if fλj (x) ∈ IA
 A

φ(x) = (s0 s1 s2 ...) = (s) such that sj = B if fλj (x) ∈ IB

C if fλj (x) ∈ IC .


Here the intervals IA , IB , and IC represent the intervals [0,1/2], (1/2,3/4], and
(3/4,1] respectively. The sequence space Σλ is made up of sequences of the three
symbols A, B, and C that obey rules determined by the parameter value λ. Typical
sequences are made up of many A’s and a few isolated B’s and C’s separated by the
A’s. The shift map on Σλ is defined as usual: σ(s0 s1 s2 . . .) = (.s1 s2 . . .). It simply
removes the first entry and shifts all other entries to the left one place.

For the same parameter values we will show that there exists an invariant prob-
ability measure and calculate it explicitly on a Markov partition given as the set of
intervals ∪i+2
k=1 Ik with
1 1 1 1 1 1 3 3
I1 = (0, i ), I2 = ( i , i−1 ), ..., Ik = ( i+2−k , i+1−k ), ..., Ii+1 = ( , ), Ii+2 = ( , 1).
2 2 2 2 2 2 4 4
 3

The measure that we will derive is the “natural” measure, so called because it mea-
sures how frequent an orbit visits a conceivable region or the fraction of the orbit
that falls into the region. Roughly speaking, to calculate the natural measure of a
map f on a particular region, you iterate the map f , count the amount of iterates
in that region, divide by the number of iterates, and take the limit as the number
of iterates goes to infinity. We will show that the number of iterates falling in the
intervals I2 through Ii are the same, but since their lengths are different their den-
sities will be different. The number of iterates in each of these intervals will be half
the number of iterates in I1 and twice the number in Ii+1 and Ii+2 .

We will also calculate the Lyapunov numbers for orbits of the family of maps with
the same parameter values as above. Lyapunov numbers measure the separation rate
of orbits on the attractors. The Lyapunov number L(x1 ) of the orbit {x1 , x2 , ...} is
defined as
L(x1 ) = lim (| f 0 (x1 ) | · · · | f 0 (xn ) |)1/n
n→∞

if this limit exists. We will use the natural measure to calculate the Lyapunov num-
ber for λ = 1/2i , i = 1, 2, 3, . . ., all of which are greater than one, implying that
the system has a sensitive dependence on initial conditions in the sense that points
separate at an exponential rate.

We will look at numerical approximations of the Lyapunov numbers for the

piecewise linear maps and compare them to our results. They seem to suggest that
3 log 2
2 2 log 2−log λ is a good interpolation for the Lyapunov number Lfλ for all λ ∈ [0, 1]. In
other numerical simulations we will compare the piecewise linear caricature models
and the full return maps.

2 Singular Shilnikov Orbits

We begin by looking at the simplest form that contains the qualitative structure
of a Shilnikov orbit. This simplest form appears as a singularly perturbed system
which has the following form:

ẋ = f (x, y, z), ẏ = g(x, y, z), εż = h(x, y, z), (1)

where x, y, z ∈ <, 0 < ε ¿ 1. If we set ε = 0, then we get:

ẋ = f (x, y, z), ẏ = g(x, y, z), 0 = h(x, y, z), (2)

which is a system of equations restricted on the set:

M = {(x, y, z) : h(x, y, z) = 0}. (3)

 4

M is called the slow manifold, although it may not be differentiable everywhere,

and hence not a manifold in the strict sense. System (2) is called the slow subsystem
of (7) and (x, y) is called the slow variable. Rescaling the time as τ = t/ε results in:

x0 = εf (x, y, z), y 0 = εg(x, y, z), z 0 = h(x, y, z) (4)

which is a regular perturbation of equations

x0 = 0, y 0 = 0, z 0 = h(x, y, z), (5)

at ε = 0. Equations (4) are called the fast subsystem of (7) and z is called the fast
variable. The reason to call (x, y) slow and z fast is that the rate of change for z,
h/ε, is far greater than that for (x, y) for sufficiently small 0 < ε ¿ 1, and h greater
than the order of ε. Note that the slow manifold M consists entirely of equilibrium
points of the fast subsystem (5). In this paper we need only to consider a portion of
the slow manifold, and for simplicity we will denote it M. It appears as a Z-shaped
smooth surface in <3 , which we will call a Z-switch. Precise assumptions are listed
below. (See also Figure 1).

H.1. The slow manifold M is analytic and consists of three connected open com-
ponents, Sj , j = 1, 2, 3 on which Dz h is nonzero, where Dz h denotes the partial
differentiation in the fast variable z. The Sj are called normally hyperbolic and
appear as in Figure 1.

H.2. These components are separated by two 1-dimensional analytic manifolds, T1 ,

T2 , with T1 ∩ T2 = ∅. T1 separates S1 and S2 , T2 separates S3 and S2 . That is,
T1 = S1 ∩ S2 , and T2 = S3 ∩ S2 , respectively. Dz h on S1 and S3 is negative, while
Dz h on S2 is positive.

H.3. The points of Tj s are turning points in the following sense.

(i) The derivative Dz h(p) = 0 and the second derivative Dz2 h(p) 6= 0 at each
point p ∈ Tj s.
(ii) Dh(p) has rank 1 at each point p ∈ Tj s, where D denotes the differentiation
operator in (x, y, z).
(iii) The vector field of (7) is transverse to the slow manifold M at each point
p ∈ Tj s, except at one point, A1 on T2 which is called a branching point.
(iv) The slow flow on S3 at the branching point A1 , is tangent to T2 to the first
order.

H.4. For each point p ∈ T1 , the ω-limit set of the orbit of the fast subsystem (4)
leaving p is on S3 . Similarly, for each point p ∈ T3 , the ω-limit set of the orbit of
the fast subsystem (4) leaving p is on S1 .
 5

H.5. There exists an equilibrium point for the slow flow on S3 which is also a ω-
limit point of T1 and the flow on S3 spirals away from this equilibrium point. The
equilibrium point on S3 is a hyperbolic global source on S3 . On S1 , every point
reaches a turning points on T1 in a finite time.

From now on, we assume hypothesis H.1-H.5 are satisfied. A piecewise con-
tinuous and oriented curve in the phase space is said to be a limiting orbit of the
singular perturbed system (7) if it consists of orbits of the slow and fast subsystems
in an alternating way so that if an orbit of the slow subsystem (2) is followed by an
orbit of the fast subsystem (5), the slow orbit must terminate at a turning point on
Tj and the fast orbit must leave the same turning point. Similarly, if the fast orbit
is followed by a slow orbit, then the fast orbit must tend to a point on Sj , referred
to as a junction point, and the slow orbit must start at the same junction point.
The orientation is induced by the time variable for the slow and fast flows (see Fig-
ure 1). A limiting Shilnikov orbit starts in S3 and ends at the equilibrium point in
S3 . It can be proved by Fenichel’s geometric theory of singular perturbation [13]
together with Bonet’s result on turning points [3] that if the limiting Shilnikov or-
bit does not contain the branching point A1 as shown in Figure 1, then for each
0 < ε ¿ 1 there is always a perturbed system of (7) such that the perturbed sys-
tem has a Shilnikov’s saddle-focus homoclinic orbit. Hence a chaotic subdynamics
exists nearby by Shilnikov’s theorem (see Šil’nikov [20], [21], [22]). Since singular
perturbation provides the simplest way by which such an orbit originates, we adopt
the same approach to attractors by considering singular limiting attractors which
contain singularly perturbed Shilnikov orbits.

A standard approach to studying the flow structure of differential equations is to

study its cross section structure and the flow induced return map if such a reduction
is feasible. For a singularly perturbed system (7) satisfying hypotheses H.1.-H.5.,
such a Poincaré return map can not only be defined for the perturbed system with
0 < ε ¿ 1, but also for the unperturbed one at ε = 0. The latter case is simply a
singular limit of the former. They are one and two dimensional maps respectively.

To construct the return map we will refer to Figure 1. We will map points from
the interval labeled [0, C0 ] into itself. In this interval there are two more points that
play a significant role, labeled A0 and B0 . B0 is located on the limiting homoclinic
orbit and the flow will always end up at the fixed point corresponding to the point
0. We will look at the behavior of the return map at A0 last.

Now if you start in the interval (0, A0 ) you remain on the S3 section of the slow
manifold. Recall that the behavior of the flow on this component of the Z-switch
is a spiral away from the equilibrium point represented here by 0. To simplify our
construction, we assume the slow flow is linear. This assumption retains qualitative
 6

structure of the more general assumption H.5. So we can say the flow is given by
the system:
ẋ = ax − by, ẏ = bx + ay,
which under polar coordinates becomes

ṙ = ar, θ̇ = b.

To get the desired orientation as in Figure 1 we will require that b < 0. Solving the
equation we get r = r0 eat0 where t0 satisfies bt0 + θ0 = θ0 + 2π, the time required
to make one revolution. Therefore points starting in (0, A0 ) spiral away from the
fixed point at an expanding rate, eat0 > 1. Hence on the return map this leads to a
monotone increase in the interval. Note that if you are in the interval (0, A 0 ) close
to the point A0 the next iteration of the return map brings you close to the point C0 .

Now consider the interval (B0 , C0 ). Follow it to the interval (B1 , C1 ), which
jumps via the fast subsystem to the interval (B2 , C2 ), which traverses the S1 section
of the slow manifold to the interval (B3 , C3 ). It then drops to the S3 section to the
interval (0, C4 ). This represents one iteration of the return map. We can see that
B0 is mapped to 0, and C0 is mapped to C4 . Note that the flow on section S1 of
the slow manifold can change the location of the point C4 . We may assume a linear
increase here.

We now consider the interval (A0 , B0 ). This interval deserves a little bit of spe-
cial attention because as we follow A0 to A1 we hit a fork in the road. A1 is the
branching point referred to above. Here the flow separates with the point A 1 ac-
tually mapping to C0 , and points close to A1 mapping near A2 to A3 to A4 and
finally to A5 . The return map is sufficiently smooth everywhere in [A0 , B0 ] ex-
cept at A0 . To see the behavior of the return map at the points approaching A0 , it
is only essential to look at the S3 section of the connecting map as shown in Figure 3.

For simpler notation assume A1 = (x∗ , y1 ) = (0, 0) on the xy-plane. Note

that the xy-flows near x∗ are parabola-like curves. Assume they are given by
y = −q|x − x∗ |σ + p + h.o.t. locally near A1 , where q > 0, σ > 1, p ≥ 0, and
h.o.t. denotes terms of higher order than |x − x∗ |σ . Suppose p = 0 corresponds to
the curve tangent to y = y1 = 0 at x∗ = 0, and without loss of generality, assume
q = 1 and h.o.t. = 0 in the following analysis. Let (xa , y0 ) be a point on the curve
y = −|x − x∗ |σ as appears in Figure 3. Then define (x1 , y1 ) and (x0 , y0 ) to be points
on the parabola-like curve y = −|x − x∗ |σ + p. Let π be the map that takes (xa , x0 ]
to (x∗ , x1 ] : x0 7→ π(x0 ) = x1 . Then
( (
y0 = −|x0 − x∗ |σ + p p = y0 + |x0 − x∗ |σ
which implies
y1 = −|x1 − x∗ |σ + p p = y1 + |x1 − x∗ |σ .
 7

Recall y1 = 0 and x∗ = 0 to get p = (−x1 )σ which implies x1 = −p1/σ . Take the

derivative with respect to x0 to get:
dx1 1 dp 1
= − p1/σ−1 = − p1/σ−1 (σ(−x0 )σ−1 (−1)).
dx0 σ dx0 σ
dx1
Since p → 0 as x0 → xa , the derivative dx 0
above approaches infinity as x0 → xa
because 1/σ − 1 < 0. Thus the graph of the return map at A0 is asymptotically
tangent to a vertical line. Therefore, the one-dimensional return map at A + 0 will
appear as shown in Figure 1.

In order to have an attractor Λ, we need a closed invariant set that has a basin
of attraction B(Λ) = {x : ω(x) ⊂ Λ} with positive Lebesgue measure. For example,
if the flow on S1 were modified so that the turning points C3 and A3 were made
to be wider than the radius of the orbit below, their corresponding images on the
return map would be outside the interval [0,C0 ]. See Figure 4. In this case, intervals
of points would escape so that the remaining invariant set would have a basin of
attraction consisting of a cantor set possibly of measure zero. Thus we will assume
from now on that the widest the turning points C3 and A3 on S1 can get are directly
above the radial orbit passing through C0 . See Figure 5(c). Under this constraint,
the interval [0, C0 ] is mapped into itself by the limiting return map and there must
exist limiting attractors in the interval.

There are many ways by which the attractors can bifurcate. The number of
bifurcation parameters required is at least three: associated as to where points
A3 , B3 , C3 drop on S3 . In this paper, however, we will only consider the bifurcation
of a one-parameter family of the attractors. This restriction is described as follows.
Consider first the case that all flows on S1 tend to the homoclinic orbit before
dropping to the S3 section via the fast flow. This is represented by C2 and A2
mapping to B3 as in Figure 5(a). In this case the return map is monotonically
increasing between 0 and A0 and is the constant function 0, between A0 and C0 .
Next we change the flow on S1 continuously so that the turning points represented
by C3 and A3 fan out from B3 to the outer edges of the Z-switch until they reach
the widest points discussed above. This is represented on the return map by an
increase in the values at A0 and C0 , from 0 to C0 . See Figure 5. can be at A0
described above we case, every point from the interval [0,C0 ] maps back interval,
and hence the basin of attraction is the entire interval. By renormalizing A 0 = 1/2,
B0 = 3/4, and C0 = 1, we are dealing with a family of attractors parameterized by
two parameters; the values of A0 and C0 under the return map. In this paper, we
will restrict our analysis further to the case that both A0 and C0 have the same value
under the return map, i.e., A3 and C3 drop onto the same orbit on S3 as illustrated
in Figure 5. Let λ represent the height of the value at A0 = 1/2 and C0 = 1 and
denote the one-parameter family of maps by Fλ : [0, 1] 7→ [0, 1]. As mentioned in
 8

the introduction, to simplify the analyses we will assume the maps are also linear
on (A0 , B0 ] and consider the following family of piecewise linear models
(
2x, if 0 ≤ x ≤ 1/2
fλ (x) = (6)
4λ | x − 3/4 |, if 1/2 < x ≤ 1.

A further restriction to a sequence of the parameter values λ is to be specified in

the next section.

We end this section by emphasizing that other types of families that represent
different types of bifurcations are equally, if not more, interesting and important.
They are not treated here because we believe that they deserve separate treatments
on their own right. More discussions on this topic can be found at the end of this
paper.

3 Symbolic Dynamics
Consider the family of functions:
(
2x, if 0 ≤ x ≤ 1/2
fλ (x) = (7)
4λ | x − 3/4 |, if 1/2 < x ≤ 1

where λ ∈ [0, 1]. Note that this family of maps has two critical points, one at 1/2
and the other at 3/4. Between these points the map is linear. Let IA = [0, 1/2],
IB = (1/2, 3/4], and IC = (3/4, 1]. Let φ(x) = itinerary of x, i.e.,

if fλj (x) ∈ IA
 A

φ(x) = (s0 s1 s2 ...) = (s) such that sj =  B if fλj (x) ∈ IB (8)
C if fλj (x) ∈ IC .


Define

K 3/4 (fλ ) = φ(fλ (3/4)) = (AAA....) and K 1/2 (fλ ) = φ(fλ2 (1/2))

depending on the parameter value λ and call K 1/2 (fλ ) and K 3/4 (fλ ) the kneading
sequences of the critical points 1/2 and 3/4 respectively. From now on we will only
consider a subset of the above family fλ for

λ = 1/2i for i = 1, 2, 3, . . . .

For these specific parameter values we have

K 1/2 (fλ ) = (A...A

| {z } CA...A
| {z } C . . .),
i i
 9

where the number of A’s between C’s is equal to i. Denote this repeating subse-
quence as
| {z } C.
R(1/2) = A...A
i

Now, define Σλ to be the sequence space on the three symbols A, B, and C that
obeys the rules (1-3) determined by λ below.
(1) Every B or C in a sequence must be followed by at least i A’s.
(2) C cannot be followed by the kneading sequence K 3/4 (fλ ).
(3) B cannot be followed by the kneading sequence K 1/2 (fλ ).

Note that rules (2) and (3) correspond to sequences representing points that are
preimages of 3/4+ and 1/2+ respectively, or more precisely limx→3/4+ φ(x) and
limx→1/2+ φ(x) respectively. Also note that rule (1) implies that every sequence
in Σλ has at most one B or C in i + 1 consecutive entries. This is the case because
if sk = B or C, then fλ (xk ) = 4(1/2i ) | xk − 3/4 |< 1/2i which implies sk+1 = A
and you must apply the map at least i times to get up to 1/2.

Remark: The above rules also apply to the nonlinear family of maps Fλ with
λ = 1/2i , i = 1, 2, 3, .... We will denote the itinerary map for Fλ as φ∗ , so that

if Fλj (x) ∈ IA
 A

φ∗ (x) = (s0 s1 s2 ...) = (s) such that sj =  B if Fλj (x) ∈ IB (9)
C if Fλj (x) ∈ IC .


Now we define the topology on Σλ , λ = 1/2i . Note that for any two different
elements (s) and (t) of Σλ , there exists 0 ≤ l ≤ ∞ such that sj = tj , for 0 ≤ j ≤ l −1
and sl 6= tl . Let l ≤ m ≤ ∞ be such that the mth entry is the first entry after l − 1,
equal to B in either (s) or (t). Without loss of generality, assume sm = B. The
distance between these two elements (s) and (t) of Σλ is defined by:
m−1 ∞
X | sn − tn | X | sn+α − tn+β |
d(s, t) = K + + , (10)
n=l 2n n=m 2n
(
0, if sj = tj
| sj − tj |= (11)
1, if sj 6= tj
with α, β, and K defined as follows:
(
A and tl = C
(1) If m > l and sl = then α = β = K = 0;
C and tl = A
(2) If m = l, recall we assumed sm = B and
(a) tm = C then α = β = 1, and K = 1/2p+m−i where p is the minimum number
of consecutive A’s after sm or tm and i is the i in λ = 1/2i , or
 10

(b) tm = A and tm+1 = C then α = 1, β = 2, and K = 1/22q+m where q is the

minimum number of consecutive R(1/2)’s after sm or tm+1 . Note that R(1/2) is the
repeating subsequence in K 1/2 (fλ ). For example if


 | {z } A...)
(s) = (s0 ...sm R(1/2)R(1/2) A...A

i then q = 2;
 (t) = (t0 ...tm R(1/2)R(1/2)R(1/2) A...A
| {z } B...)


i
(
A
(c) tm = A and tm+1 = then α = β = K = 0.
B
Remark: Case (2)(a) defined above is to insure two points near the critical point
3/4 stay close under the itinerary map φ. Note that points close to 3/4 will have
more consecutive A’s after B or C in their itineraries, i.e. as x gets closer to 3/4,
φ(x) gets more and more A’s. Case (2)(b) defined above is to insure that points
near the critical point 1/2 stay close under the itinerary map φ. Note here that
points close to 1/2 will have more consecutive subsequences of R(1/2)’s after B
or AC in their itineraries. The constant K is defined so that two different points
mapped under fλ to the same value are not defined to be the same point under φ.
See Figure 6. The reason for the 2q in K = 1/22q+m is that the interval containing
points with q R(1/2)’s after B or C in their itineraries shrinks by 1/4 for every q.
The interval containing points with itineraries having p A’s after B or C shrinks by
1/2 with every p > i.

Proposition 3.1 The function d, defined above is a metric on Σλ for λ = 1/2i .

The proof of Proposition 3.1 can be found in Taylor [23]. It is lengthy and follows
a standard argument for metrics on symbol spaces by definition.

Proposition 3.2 (1) If l < m and sj = tj for j = 0, ..., l, then d(s, t) ≤ 1/2l .
(2) If l = m, without loss of generality assume sm = B.
(a) If sj = tj for j = 0, ..., m − 1, m + 1, ..., n, and if tm = C, then
(i) if n = ∞ then d(s, t) = 1/2p+m−i , where p is the minimum number of A’s
after sm or tm and i is as in λ = 1/2i .
(ii) if n < ∞ then d(s, t) ≤ 1/2p+m−i + 1/2n , where p is the minimum number
of A’s after sm or tm and i is as in λ = 1/2i . Note n ≥ p + m − i.
(b) If sj = tj for j = 0, ..., m−1, tm = A, tm+1 = C, and sj+1 = tj+2 for j = m, ..., n.
Then
(i) if n = ∞ then d(s, t) = 1/22q+m where q is the minimum number of R(1/2)’s
after sm or tm+1 .
(ii) if n < ∞ then d(s, t) ≤ 1/22q+m + 1/2n where q is the minimum number of
R(1/2)’s after sm or tm . Note n ≥ (i + 1)q + m ≥ 2q + m.
(c) If sj = tj for j = 0, ..., m − 1, tm = A, and tm+1 = A or B. Then d(s, t) ≤
1/2m+1 .
 11

Proof:
(1) If sj = tj for j = 0, ..., l, then
l m−1 ∞ ∞
X | sj − sj | X | sj − tj | X | sj+α − tj+β | X 1 1
d(s, t) = 0 + + + ≤ = l.
j=0 2j j=l+1 2 j
j=m 2 j
j=l+1 2
j 2

(2) If m = l and sm = B.

(a) If sj = tj for j = 0, ..., m − 1, m + 1, ..., n, and if tm = C, then

(i) if n = ∞, we have
m−1
1 X | sj − sj | n=∞
X | sj+1 − sj+1 | 1
d(s, t) = + + = p+m−i .
2p+m−i j=0 2 j
j=m 2 j 2

(ii) If n < ∞ then

m−1 n ∞
1 X | sj − sj | X | sj+1 − sj+1 | X | sj+1 − tj+1 |
d(s, t) = + + +
2p+m−i j=0 2j j=m 2j j=n+1 2j
∞
1 X 1
≤ +
2p+m−i j=n+1 2
j

1 1
= + .
2p+m−i 2n
(b) If sj = tj for j = 0, ..., m − 1, tm = A, tm+1 = C, and sj+1 = tj+2 for p = m, ..., n.
Then if
(i) n = ∞ we have
m−1
1 X | sj − sj | n=∞
X | sj+1 − sj+2 | 1
d(s, t) = + + = 2q+m .
22q+m j=0 2 j
j=m 2 j 2

(ii) If n < ∞ then

m−1 n ∞
1 X | sj − sj | X | sj+1 − sj+2 | X | sj+1 − tj+2 |
d(s, t) = + + +
22q+m j=0 2 j
j=m 2 j
j=n+1 2j
∞
1 X 1
≤ +
22q+m j=n+1 2
j

1 1
= + .
22q+m 2n
(c) If sj = tj for j = 0, ..., m − 1, tm = A, tm+1 = A or B, then
∞ ∞
X | sj − tj | X 1 1
d(s, t) = 0 + ≤ = .
j=m 2j j=m 2j 2m+1
 12

Next we introduce the shift map σ, on Σλ defined by: σ(s0 s1 s2 ...) = (.s1 s2 ...). It
simply removes the first entry and shifts all other entries to left one place. We now
show that the dynamics of fλ on [0, 1] and σ on Σλ are essentially the same via a
topological conjugacy.

Theorem 3.3 For λ = 1/2i where i = 1, 2, 3, ... the map fλ on [0, 1] is topologically
conjugate to the shift map on Σλ , i.e., there exists a homeomorphism φ : [0, 1] −→ Σλ
such that φ ◦ fλ = σ ◦ φ.

Proof: Let φ be the itinerary map defined in (8). First we will show that φ is one-
to-one: Let x 6= y ∈ I, we need to show that φ(x) 6= φ(y). Case (1): Suppose there
exists a first k such that (fλk (x), fλk (y)) contains one of the critical points 1/2 or 3/4.
In which case φ(x) and φ(y) differ in the kth place, and hence φ(x) 6= φ(y).

Case (2): Suppose there does not exist a k such that [fλk (x), fλk (y)] contains 1/2 or
3/4. In this case fλk+1 is continuously differentiable on the interval [fλk (x), fλk (y)] for
all k = 0, 1, 2, .... By the Mean Value Theorem there exists ck ∈ [fλk−1 (x), fλk−1 (y)]
such that

| fλk (x) − fλk (y) |=| (fλk )0 (ck ) || fλk−1 (x) − fλk−1 (y) |

for all k = 1, 2, 3, .... Hence we have

| fλk (x) − fλk (y) | = | (fλk )0 (ck ) || fλk−1 (x) − fλk−1 (y) |
= | (fλk )0 (ck )(fλk−1 )0 (ck−1 ) || fλk−2 (x) − fλk−2 (y) |
.
.
.
= | (fλk )0 (ck )(fλk−1 )0 (ck−1 )...(fλ )0 (c1 ) || x − y |

for all k = 1, 2, 3, .... Note that if there exists a k such that fλk (x) ∈ IB or IC , then
fλk+j (x) ∈ IA for j = 1, 2, ..., i where λ = 1/2i . Thus in any consecutive sequence
of length i of the above (fλn )0 (cn )’s, the sequence has at most one (fλn )0 (cn ) equal to
4λ = 1/2i−2 with the other i − 1 equal to 2. Therefore

| fλi (x) − fλi (y) |≥ 2 | x − y | . (12)

Now pick m so that 2m | x − y |> 1/2. Then

| fλmi (x) − fλmi (y) |≥ 2m | x − y |> 1/2

 13

which implies that fλmi (x) and fλmi (y) are not in the same interval, and hence
φ(x) 6= φ(y) because they differ in the mith place.

To show that φ is onto, we first introduce the following notation. Let J =

[a, b] ⊂ I = [0, 1] be a closed interval 0 < a < b < 1. Let

fλ−n (J) = {x ∈ I | fλn (x) ∈ J}.

In particular, fλ−1 (J) denotes the preimage of J. Note that the preimage of J is:
(a) 3 subintervals if and only if a < 1/2i , one contained in each of the subintervals
IA , IB , and IC . Or (b) 1 closed subinterval contained in IA . See Figure 7.

To find x ∈ [0, 1] for a given (s) = (s0 s1 s2 ...) ∈ Σλ such that φ(x) = (s), we
define

Is0 s1 s2 ...sn = {x ∈ I | x ∈ Is0 , fλ (x) ∈ Is1 , ..., fλn (x) ∈ Isn }

= Is0 ∩ fλ−1 (Is1 ) ∩ ... ∩ fλ−n (Isn ).

Note that sj ∈ {A, B, C} for each j, so Isj is really one of IA , IB , or IC .

For any pair of intervals K and L, we have

fλ−1 (K ∩ L) = fλ−1 (K) ∩ fλ−1 (L).

Thus we may also write

−(n−1)
Is0 s1 s2 ...sn = Is0 ∩ fλ−1 (Is1 ∩ ... ∩ fλ (ISn ))
= Is0 ∩ fλ−1 (Is1 ...sn ).

We claim that the I s0 s1 s2 ...sn , the closure of Is0 s1 s2 ...sn , are nonempty closed in-
tervals that are nested. Clearly the closure of any set is closed, so it is left to show
that I s0 s1 s2 ...sn are nested and nonempty.

To show I s0 s1 s2 ...sn are nonempty we will use an induction argument. Clearly

I sn ∈ {I A , I B , I C } is nonempty. Assume I s1 ...sn is nonempty then by previous obser-
vation fλ−1 (I s1 ...sn ) is (a), the union of three disjoint subintervals or (b), one closed
subinterval contained in IA . If we look at case (a), since only one of the three subin-
tervals are contained in each of the subintervals IA , IB , and IC , it follows that the
closure of Is0 ∩ fλ−1 (Is1 ...sn ) ⊂ I s0 is a nonempty closed interval. If the preimage of
I s1 ...sn is the one closed interval contained in IA , then I s1 ...sn is contained in IB ∪ IC .
But Is1 ...sn = Is1 ∩ fλ−1 Is2 ...sn ⊂ Is1 , and hence Is1 ...sn is actually contained in either
IB or IC . Thus s1 = B or C, in which case s0 must be equal to A, since (s) ∈ Σλ .
This being the case, implies that the closure of Is0 ∩ fλ−1 (Is1 ...sn ) ⊂ IA is a nonempty
 14

closed interval. Note IA is a closed interval.

These intervals are nested because

I s0 s1 s2 ...sn = Is0 s1 s2 ...sn−1 ∩ fλ−n (Isn ) ⊆ I s0 s1 s2 ...sn−1 ∩ fλ−n (Isn ) ⊂ I s0 s1 s2 ...sn−1 .

Therefore we conclude that \

I s0 s1 s2 ...sn
n≥0

is nonempty since a nested intersection of compact intervals is always nonempty.

Claim: ∩n≥0 I s0 s1 s2 ...sn = ∩n≥0 Is0 s1 s2 ...sn .

proof of claim: Suppose on the contrary that there exists x ∈ ∩n≥0 I s0 s1 s2 ...sn −
∩n≥0 Is0 s1 s2 ...sn . Then x is either equal to or the preimage of one of the criti-
cal points 1/2 or 3/4. This is because the only points not included in the clo-
sure of IA , IB , and IC are 1/2 or 3/4, and the only points not included in the
preimages of IA , IB , and IC are preimages of 1/2 or 3/4. In which case there
exists m ≥ 0 such that f m (x) = 1/2 or f m (x) = 3/4. If f m (x) = 1/2 then
φ(x) = (s0 ...sm−1 ACK 1/2 (fλ )), and ∩n≥0 I s0 s1 s2 ...sn = I s0 ...sm−1 ACK 1/2 (fλ ) . Now if
x∈ / ∩n≥0 Is0 s1 s2 ...sn then we have I s0 ...sm−1 ACK 1/2 (fλ ) 6= Is0 ...sm−1 ACK 1/2 (fλ ) which im-
plies φ(x) 6= (s0 ...sm−1 ACK 1/2 (fλ )) contrary to f m (x) = 1/2. Similarly if f m (x) =
3/4, then ∩n≥0 I s0 s1 s2 ...sn = Is0 ...sm−1 BK 3/4 (fλ ) and if x ∈/ ∩n≥0 Is0 s1 s2 ...sn we get that
3/4
φ(x) 6= (s0 ...sm−1 BK (fλ )) which is a contradiction. Therefore ∩n≥0 I s0 s1 s2 ...sn =
∩n≥0 Is0 s1 s2 ...sn and we have proved the claim.

Suppose that ∩n≥0 Is0 s1 s2 ...sn contains more than one point. Let xs and ys be
two of these points contained in ∩n≥0 I s0 s1 s2 ...sn . By the definition of Is0 s1 s2 ...sn ,
the itineraries of xs and ys are the same, i.e. φ(xs ) = φ(ys ). But φ is one-
to-one, hence xs = ys . Say ∩n≥0 I s0 s1 s2 ...sn = {xs }. In which case xs ∈ Is0 ,
fλ (xs ) ∈ Is1 ,...fλn (xs ) ∈ Isn ... and so forth. Hence (s0 s1 s2 ...) = φ(xs ), and φ is
onto.

To prove continuity of φ, let x ∈ [0, 1] and suppose that φ(x) = (s0 s1 s2 ...). The
idea is that since the intervals defined by Is0 ,...,sq get smaller as q → ∞, we need to
pick a large enough q so that x, y ∈ Is0 ,...,sq implies d(φ(x), φ(y)) < ε. In which case
it is necessary to know the length of the interval Is0 ,...,sq , so we can pick the proper δ.

Let ε > 0. Pick n so that 1/2n < ε. Choose r > n, and suppose that
x ∈ Is◦0 s1 s2 ...sr , the interior of Is0 s1 s2 ...sr . Pick δ equal to the minimum distance
between x and the endpoints of Is0 s1 s2 ...sr . If | x − y |< δ, both x and y are in
Is0 s1 s2 ...sr and hence have the same first r + 1 entries in their itineraries. In which
 15

case d(φ(x), φ(y)) ≤ 1/2r < 1/2n < ε.

If x 6∈ Is◦0 s1 s2 ...sr , then there exists a first 0 ≤ m ≤ r such that x 6∈ Is◦0 s1 s2 ...sm .
Hence x is an endpoint of Is0 ...sm and so a preimage of 0, 1/2, 3/4, or 1, i.e.
fλm (x) ∈ {0, 1/2, 3/4, 1}. If f m (x) = 0 or 1, then x is either equal to or a preimage
of 0 or 1. If it is a preimage of either 0 or 1, then it is also a preimage or equal to
3/4 or 1/2 respectively. If x is a preimage or equal to 3/4 or 1/2 the argument of
continuity follows that as if f m (x) = 3/4 or f m (x) = 1/2, respectively.

First consider if x = 0, then x ∈ IAAA... and the length of Is0 ....sr = 1/2r+1 , where
sj = A for j = 0, ..., r. So pick δ = 1/2r+1 . If | x − y |< δ, both x and y are in
Is0 ....sr and hence have the same first r + 1 entries in their itineraries. In which case
d(φ(x), φ(y)) ≤ 1/2r < 1/2n < ε.

Second consider if x = 1 then x ∈ ICR(1/2)R(1/2)... and the length of ICs1 ....sr =

1/22ξ , where ξ is the number of consecutive R(1/2)’s in the sequence s1 ...sr . Recall
that R(1/2) represents the repeating subsequence in the kneading sequence K 1/2 (fλ ).
Pick δ = 1/22ξ . If | x − y |< δ, both x and y are in Is0 ....sr and hence have the same
first r + 1 entries in their itineraries. In which case d(φ(x), φ(y)) ≤ 1/2 r < 1/2n < ε.

Third consider if f m (x) = 3/4 then x ∈ Is0 ...sm−1 BAAA... . The length of It0 ....tu =
1/2u−i+2 , where t0 = B or C and tj = A for j = 1, ..., u, and the length of
Is0 ...sm−1 ≥ 1/2m−1 . So the length of Is0 ...sm−1 sm ...sm+u ≥ 1/2u−i+2 × 1/2m−1 , where
sm = B or C and sj = A for j = m+1, ..., m+u. Choose u so that u > n−m+i+1,
and pick δ = 1/2u−i+2 × 1/2m−1 . Then if | x − y |< δ, y ∈ Is0 ....sm−1 sm ...sm+u , and by
Proposition 3.2, d(φ(x), φ(y)) ≤ 1/2u+m−i + 1/2u+m−i = 1/2u+m−i−1 < 1/2n < ε.

If f m (x) = 1/2 then x ∈ Is0 ...sm−1 ACR(1/2)R(1/2)... and the length of IACt2 ....tr+1 =
1/22ρ+3 , where ρ is the number of consecutive R(1/2)’s in the sequence t2 ...tr+1 start-
ing at t2 . Note that the length of IBz1 ....zr = 1/22ρ+2 , where zj = tj+1 for j = 1, ..., r.
The length of Is0 ...sm−1 ≥ 1/2m−1 , so the length of Is0 ...sm−1 sm sm+1 ...sm+r+β ≥ 1/22ρ+3 ×
1/2m−1 , where sm = A and sm+1 = C or sm = B, and ρ is the number of consecutive
R(1/2)’s in the sequence sm+α ...sm+r+β starting at sm+α , where α and β come from
the definition of the distance between two elements of Σλ . If sm = A then α = 2
and β = 1, else α = 1 and β = 0. Choose r so that 1/22ρ+m + 1/2m+r+β < ε and
pick δ = 1/22ρ+3 × 1/2m−1 . If | x − y |< δ, then y is in Is0 ...sm−1 sm sm+1 ...sm+r+β . By
Proposition 3.2, d(φ(x), φ(y)) ≤ 1/22ρ+m + 1/2m+r+β < ε.

Note that φ−1 is continuous since φ is a bijective continuous mapping from a

compact space into a Hausdorff space.
T
We claim that φ ◦ fλ = σ ◦ φ. First note that since φ is one-to-one, n≥0 Is0 s1 s2 ...sn
 16

consists of a unique point. Let x ∈ [0, 1] have the itinerary (s0 s1 s2 ...) = φ(x), then
by definition x ∈ Is0 , fλ (x) ∈ Is1 , fλ2 (x) ∈ Is2 ,..., and so forth. Now let fλ (x) have
itinerary (t0 t1 t2 ...) = φ(fλ (x)), then by definition fλ (x) ∈ It0 , fλ2 (x) ∈ It1 , fλ3 (x) ∈
It2 ,..., and so forth. By uniqueness t0 = s1 , t1 = s2 , .... Thus, φ(fλ (x)) = σ(φ(x))
as claimed. ♦

Remark: From Theorem 3.3 it follows that for the parameter values λ = 1/2i ,
i = 1, 2, 3, ... the dynamics of the map fλ are chaotic, in the sense they satisfy the
three properties of the following result.

Corollary 3.4 For the parameter values λ = 1/2i , i = 1, 2, 3, ..., the piecewise
linear family of maps fλ satisfies the following three properties;
(1) Periodic points are dense.
(2) There exists a dense orbit.
(3) Orbits are sensitive to initial conditions.

The result of the above corollary follows from the fact that fλ is conjugate to
the shift map σ on the sequence space Σλ . Readers are referred to Devaney [11] or
Alligood et al. [2] for a standard argument for σ to satisfy these properties.

Remark: For λ = 1/2i where i = 1, 2, 3, ... we believe the same result applies to the
map Fλ . We cannot rigorously prove this claim because of the nonlinearity of Fλ on
the interval IB . To be more specific, recall that the itinerary map φ is one-to-one
because we can show that points separate after time under the iteration of the map
fλ . We are unable to do the same for Fλ at this point using similar argument. The
problem lies in that even though there is a minimum derivative for the map Fλ on its
nonlinear part, we cannot guarantee that a typical orbit will spend enough time in
the interval IA with slope equal to 2 to compensate. Instead one might have to try
an alternative strategy by which one argues that if points were not to separate off
over the nonlinear part then they would be attracted to 3/4− , where the minimum
derivative is. But this corresponds to the homoclinic orbit and would indicate that
it is stable which should be a contradiction. Then the same argument as above for
fλ that uses the injectivity to show the surjectivity would apply to φ∗ as well. Also
note that for the linear map we were able to explicitly find the length of the needed
interval Is0 ,...,sq . This length is not so easily derived for the Fλ case. For this reason,
we were unable to directly show the continuity of φ∗ for now.

We end this section by pointing out a connection between our result and Shilnikov’s
result on saddle-focus homoclinic orbits. His symbolic dynamics in the context of
singularly perturbed systems (2.1) for 0 < ε ¿ 1 would correspond to a subshift
of σ restricted on an invariant subspace of Σλ for which the number of consecu-
tive A’s, as in rule (1) is limited to a proper subset of the natural numbers. In
fact, his symbolic dynamics would require that the numbers of consecutive A’s be
 17

sufficiently large and that any two adjacent numbers m, n, with n following m, sat-
isfy the relation n ≤ O(1/ε)m. Note that his result can be formally extended to
the singular limit ε = 0, and our result precisely justify that. In other words, our
result can be considered as the origin of his in terms of singularly perturbed systems.

4 Natural Measure
The term “measure”, to which this section is devoted, refers to a way of specifying
how much of the attractor is in each conceivable region. We would like to guarantee
that an orbit chosen at random spends the same portion of it’s iterates in a given
region as any other such orbit would. To this end we will introduce what we will
call the natural measure and use the definition in Alligood et al. [2]. See also the
average time spent by a point in a set A, as in Mané [17], and Bowen-Ruelle-Sinai
measure as in de Melo [6].

The idea of the natural measure is to iterate a map f , count the amount of
iterates in a particular region, divide by the number of iterates, and take the limit
as the number of iterates goes to infinity. More precisely, this defines the fraction
of iterates of the orbit {f n (x0 )} lying in a set S by
#{f j (x0 ) ∈ S; 0 ≤ j < n}
Ψf (x0 , S) = n→∞
lim .
n
Let Nr (S) denote the set of points that are within r distance to the set S. The
natural measure generated by the map f , is defined by

µf (S) = lim Ψf (x0 , Nr (S))

r→0

for each closed set S, and Lebesgue almost all x0 ∈ [0, 1]. Where the map f is
clear we will omit the subscript f . Note that if it exists, the natural measure is a
f -invariant probability measure, meaning µ(A) = µ(f −1 (A)) and µ([0, 1]) = 1.

The main result of this section is to prove the following theorem.

Theorem 4.1 The natural measure exists for the map

(
2x, if 0 ≤ x ≤ 1/2
fλ (x) =
4λ | x − 3/4 |, if 1/2 < x ≤ 1

on [0, 1] with λ = 1/2i , i = 1, 2, 3, .... Moreover, let ∪i+2

k=1 Ik be a partition for [0, 1]
with
1 1 1 1 1 1 3 3
I1 = (0, ), I 2 = ( , ), ..., I k = ( , ), ..., I i+1 = ( , ), I i+2 = ( , 1).
2i 2i 2i−1 2i+2−k 2i+1−k 2 4 4
 18

Then the natural measures of these intervals are given as

2 1 1
µ(I1 ) = , µ(I2 ) = µ(I3 ) = · · · = µ(Ii ) = , and µ(Ii+1 ) = µ(Ii+2 ) = .
2+i 2+i 2(2 + i)

Remark: Note that the natural measure for the i − 1 intervals after I1 are the same
but their densities are different. Each interval is half the length of the next and so
its density is doubled.

The proof of this theorem consists of a few lemmas. The strategy is to show
the existence of the natural measure first and then to calculate the measures on the
partitioning intervals.

To show the existence of the natural measure we need to use some fundamental
results from ergodic theory for one-dimensional maps which are Markov maps. By
definition, a piecewise C 1 map f : N → N is called Markov if there exists a finite
or countable family Ij of disjoint open intervals in N such that
(a) N − ∪j Ij has Lebesgue measure zero and there exists C > 0 and γ > 0 such
that for each n = 1, 2, 3, ... and each interval I such that f k (I) is contained in one
of the intervals Ij for each k = 0, 1, ..., n one has
¯ ¯
¯ Df n (x) ¯
¯ ≤ C· | f n (x) − f n (y) |γ for all x, y ∈ I.
¯ ¯
¯ − 1
¯ Df n (y) ¯

(b) if f (Ik ) ∩ Ij 6= ∅ then f (Ik ) ⊃ Ij ;

(c) there exists r > 0 such that | f (Ij ) |≥ r for each j.

Lemma 4.2 Each of our maps fλ with parameter values λ = 1/2i , i = 1, 2, 3, ... is
Markov with the same partition as in the statement of the theorem, ∪i+2
k=1 Ik .

Proof: By definition, we need to verify the definition properties as below.

(a) [0,1]-∪i+1
j=0 Ij is a finite discrete set and has Lebesgue measure zero. Because the
map fλ is linear on each subinterval it is clear that for each n = 1, 2, 3, ... and each
interval I such that fλk (I) is contained in one of the intervals Ij for each k = 0, 1, ..., n
one has ¯ ¯
¯ Df n (x) ¯
λ
¯ = 0 ≤ C· | fλn (x) − fλn (y) |γ for all x, y ∈ I.
¯ ¯
¯
¯ Df n (y)
− 1 ¯
λ

(b) It is straightforward to verify that fλ (Ij ) is exactly equal to the union of some
other subintervals in the partition. It is also clear that if f (Ik ) ∩ Ij 6= ∅ then
f (Ik ) ⊃ Ij .
 19

(c) | f (Ij ) |≥ 1/2i for each j. ♦

1
Lemma 4.3 The natural measure exists for our family of maps fλ with λ = 2i
,i =
1, 2, 3, ....

Proof: Since our map is Markov, by Theorem 2.2 from Melo [6] there exists an
invariant probability measure µ which is absolutely continuous with respect to the
Lebesgue measure. Also, by the proof of that theorem, the map is ergodic with re-
spect to the invariant probability measure. Hence by the Birkhoff ’s Ergodic theorem
which can be found in any ergodic theory book, see Mané [17], the measure of a
Borel measurable set is given by the fraction of iterates as defined above. That is
the invariant probability measure is indeed the natural measure. ♦

Next, we want to calculate the natural measures of the partitioning intervals.

To do this we will use some results from the theory of probability on Markov maps,
closely following the treatment by Kemeny and Snell [16]. First, we introduce some
concepts . The first of which is a transition matrix. By definition 2.1.4 of Kemeny
and Snell [16] the transition matrix of a Markov map f : N → N is the matrix
P with entries pkj , where the pkj ’s represent the probability f (x) ∈ Ij given that
x ∈ Ik . Where the Ij form a partition that satisfy property (b) from the definition
of a Markov map and N − ∪Ij has Lebesgue measure zero. For our family of maps
fλ , this probability can be calculated explicitly. For example, fλ (x) ∈ I1 given that
1
x ∈ I1 if and only if x ∈ (0, 2i+1 ), the left half of the interval I1 . So p11 = 21 . The
other entries are calculated similarly. Thus, the transition matrix for our family of
maps with λ = 1/2i is an i + 2 by i + 2 matrix with the form:
 1 1

0 0 ···
2 2
0 0 0
 
 0 0 1 0 ··· 0 0 0 
 
 0 0 0 1 ··· 0 0 0 
 
 . ... .. 
 .. . 
P =



 0 0 0 0 ··· 1 0 0 
 
 0 0 0 0 ···
 0 21 12 

 1 0 0 0 ···

0 0 0 

1 0 0 0 ··· 0 0 0
except for the case i = 1 which has the form
 
1/2 1/4 1/4
P =
 1 0 0 .
1 0 0
Next, we introduce the concept of regular. By definition, a transition matrix
is regular if for some N , P N has no zero entries. We now demonstrate that our
transition matrices are indeed regular.
 20

Lemma 4.4 The transition matrices P defined above for our maps fλ and Fλ with
λ = 21i , i = 1, 2, 3, ... are regular.

Proof: By definition, we need to show that there exists an N such that P N has
no zero entries. First note that since all the entries are positive or zero, there is
no negative entry for any of the iterates P N . Second, if we have the first row and
column filled with nonzero entries for one of the iterates, then this matrix squared
will contain no zero entries. That is, once we find an M such that P M has positive
entries in the entire first row and column, then P 2M has no zero entries. We now
show that there is such an M . Because of the structure of our transition matrix P ,
after each multiplication we add another nonzero entry to the first row and the first
column, maintaining the sign of all the original nonzero entries. The first row fills
in from left to right and the first column fills in from bottom to top. So after i − 1
multiplications of P to itself we will have nonzero entries in the entire first row and
first column. Therefore the 2ith power of P will have no zero entries. Hence, the
transition matrix P is regular. ♦

As an example of a particular case of the proof, the case with i = 3 and P, ..., P 2i
are explicitly calculated as follows.
 1 1
  1 1 1

2 2
0 0 0 4 4 2
0 0
0 0 12 12 
   
 0 0 1 0 0   0
 1

1  2
 
 0 0 0
P = , P =  1 0 0 0 0 
2 2   
 1 1
 1 0 0 0 0  0 0 0
  
 2 2 
1 1
1 0 0 0 0 2 2
0 0 0
 1 1 1 1 1
  9 1 1 1 1

8 8 4 4 4 16 16 8 8 8
   1 1 
 1 0 0 0 0   2 2
0 0 0 
3
 1 1
 4
 1 1 1

P =
 2 2
0 0 0  ,P =

 4 4 2
0 0 

 1 1 1 1 1 1 1 1
 4 4 2
0 0 


 8 8 4 4 4


1 1 1 1 1 1 1 1
4 4 2
0 0 8 8 4 4 4
 17 9 1 1 1
  25 17 9 1 1

32 32 16 16 16 64 64 32 32 32
 1 1 1   1 1 1 1 1 

 4 4 2
0 0 


 8 8 4 4 4


5 1 1 1 1 1  , P6 =  9 1 1 1 1
P =
 8 8 4 4 4   16 16 8 8 8
.

 9 1 1 1 1   17 9 1 1 1 
 16 16 8 8 8   32 32 16 16 16 
9 1 1 1 1 17 9 1 1 1
16 16 8 8 8 32 32 16 16 16
Now our calculation will be based on the following lemma taken from Theorems
4.1.2, 4.1.4, 4.1.6, and 4.2.1 of Kemeny and Snell [16]. A proof of the lemma is
omitted and is referred to the cited reference. A vector is said to be a probability
vector if all of its entries are positive and sum to 1.

Lemma 4.5 If P is a regular transition matrix then there exists a unique probability
vector α = {a1 , a2 , ..., an }T such that P T α = α, where P T is the transpose of the
 21

transition matrix P of the map f and f = fλ . Moreover, for almost every x0 ∈

[0,1], we have Ψf (x0 , Ij ) = aj , where Ψf (x0 , Ij ) is the fraction of the f -orbit lying
in the interval Ij .

Remark: Since the natural measure exists for the family of maps fλ , λ = 1/2i ,
i = 1, 2, 3, ... we have that Ψfλ (x0 , Ij ) in the above lemma actually represents the
natural measure for the interval Ij . In fact all the end points of Ij and their preim-
ages are eventually periodic and they are countable. Thus the natural measures of
these end points are zero because almost all orbits will not land on them. Therefore
µf (Ij ) = limr→0 Ψf (x0 , Nr (Ij )) = Ψf (x0 , Ij ) as claimed.

Thus to calculate the natural measures of Ik we need to solve for the probability
vector α. Since we know Ψfλ (x0 , Ij ) represents the natural measure we will rewrite
the previous equation as µ(Ij ) = aj , where µ is the natural measure. To solve
for α we need to find the probability eigenvector corresponding to the eigenvalue
1 of the transpose of the transition matrix. Solving the equation P T α = α, we
i−1
z }| {
get α = {1, 1/2, 1/2, ..., 1/2, 1/4, 1/4} before normalizing it as a probability vector.
2 1 1 1 1 1
After normalization we get α = { 2+i , 2+i , 2+i , ..., 2+i , 2(2+i) , 2(2+i) }, which gives rise
to what are stated in Theorem 4.1.

Now the theorem is proved by Lemma 4.3 and the calculations given above.

We also have a more general result about the existence of natural measures on
the piecewise linear family of maps.

Theorem 4.6 The natural measure exists for the maps fλ with parameter values λ
such that the orbit of λ under the map fλ is eventually periodic.

Proof: From Theorem 2.2 of de Melo [6] it is only necessary to show that the maps
fλ satisfying the above condition are Markov. Since the orbit of λ is eventually
periodic say of period n, there exists m such that fλm+n (λ) = fλm (λ). So we can
write out the forward iterates of λ along with 1/2 and 3/4 and order them so that

0 < j0 (λ) < j1 (λ) < · · · < jm (λ) < · · · < jn+m+h (λ),

where jk (λ) is one of the iterates of λ or equal to 1/2 or 3/4. Also h = 0, 1 or

2, depending on whether 1/2 and/or 3/4 are iterates of λ. Let I1 = (0, j0 (λ));
n+m+h
I2 = (j0 (λ), j1 (λ)); · · ·; Im+n = (jn+m+h−2 (λ), jn+m+h−1 (λ)). Then {Ik }k=1 is a
n+m+h
family of disjoint open intervals in [0,1] such that [0, 1] − ∪k=1 Ik has Lebesgue
measure zero. Also since fλ | Ik is linear for all k = 1, 2, ..., n + m + h, we have
that Dfλ (x) = Dfλ (y) for all x, y ∈ Ik and all k = 1, 2, ..., n + m + h. By the
chain rule Dfλn (x) = Πn−1 j=0 Dfλ (x) if x ∈ Ik for all n. Therefore for each n = 1, 2, ...
 22

and each interval I such that fλj (I) is contained in one of the intervals Ik for each
j = 0, 1, ..., n one has
Dfλn (x)
| − 1 |= 0 ≤ C· | f n (x) − f n (y) |γ for all x, y ∈ I.
Dfλn (y)
Thus property (a) of the definition of Markov is confirmed. Now let’s check property
(b). If fλ (Ik ) ∩ Ij 6= ∅ then in fact, fλ (Ik ) equals a union of the Ij , j ∈ {1, 2, ..., n +
m+2}. See Figure 8. Now because our partition ∪Ij is finite, property (c) is trivially
true. Thus by Theorem 2.2 we have a natural measure for fλ such that the orbit of
λ is eventually periodic. ♦

5 Lyapunov Numbers
An important measure of chaotic dynamics is sensitive dependence on initial condi-
tions, that is to say that the orbits of points close to each other will separate over
time. One way to measure the rate of this separation is through the calculation of
Lyapunov numbers. The Lyapunov number L(x1 ) of the orbit {x1 , x2 , ...} is defined
as
lim (| f 0 (x1 ) | · · · | f 0 (xn ) |)1/n
L(x1 ) = n→∞
if this limit exists. Note that orbits separate at an average rate roughly equal to L
if L > 1. The main result of this section is the following theorem.

Theorem 5.1 The Lyapunov number for the fλ -orbit of almost every x ∈ [0, 1]
where λ = 1/2i , i = 1, 2, 3, ... is given by
3
Lfλ (x) = 2 i+2 .
Proof: We will explicitly calculate Lfλ using the calculated natural measures on the
intervals [0,1/2] and (1/2, 1]. The slope of all points in [0,1/2] is equal to 2, while
1
the absolute value of the slope of all points in (1/2, 1] is 2i−2 . The natural measure
1
of (1/2, 1] is given by µ(Ii+1 ) + µ(Ii+2 ) = 2+i . Thus the natural measure for [0,1/2]
1
is equal to 1 − 2+i = 2+i−1
2+i
. Therefore
n
Y 1 1 1 3
Lfλ (x) = n→∞
lim ( lim (2(1− 2+i )n · (
| fλ0 (xk ) |)1/n = n→∞ i−2
)( 2+i )n )1/n = 2 i+2 .
k=1 2
♦

Remark: Note that this number is given in terms of the variable i which is related
to the parameter value via λ = 1/2i . Solve for i in λ from λ = 1/2i and write the
Lyapunov number in terms of λ. This results in
3 log 2
lλ = 2( 2 log 2−log λ ) .
 23

As to be shown in the following section, this function lλ seems to give a good

interpolation to the Lyapunov number Lfλ for all λ > 0.

6 Numerical Simulations
One of the ways to view the change of a one-dimensional attractor with the change
in a parameter is through the use of a bifurcation diagram, which is generated by
plotting the tail end of randomly chosen orbits versus the parameter. The bifurcation
diagrams Figures 9, 10 for a qualitative return map and the corresponding piecewise
linear caricature map used the last 200 iterates of 300 at 5000 parameter values
between 0 and 1 evenly incremented. The orbits were chosen at random by randomly
selecting their initial points. For simulation purposes we used the following map for
the nonlinear return map

 2x,

 if 0 ≤ x ≤ 1/2
8λ(x−3/4)2
Fλ (x) =  + 4λ | x − 3/4 |, if 1/2 < x ≤ 3/4
(x−1/2)1/2 +1

 4λ | x − 3/4 |, if 3/4 < x ≤ 1.

The only difference between this Fλ and the piecewise linear caricature map fλ is
+
that it is nonlinear in (1/2, 3/4] and has an infinite derivative at 12 asymptotically.

In Figures 9, 10 we observe an apparent random distribution of points over any

specific parameter value. Looking over all parameter values at once there appears to
be a linear pattern to the density of points. For example at the parameter value 1/2,
there appears to be twice as many points in the interval [0, 1/2], as in the interval
(1/2, 1]. At the parameter value 1/4, there appears to be twice as many points in
the interval [0, 1/4], as in the interval (1/4, 1/2], which appears to have twice as
many points as the interval (1/2, 1]. Note that this appears to be the case for both
the piecewise linear and piecewise nonlinear cases. This observation for parameter
values λ = 1/2i , i = 1, 2, 3, ... is confirmed in Theorem 4.1 for fλ .

Next, we point out the difference in appearance of the two bifurcation diagrams.
Notice that after the parameter value λ = 1/2, there is an area of higher density
that does not appear in the nonlinear return map. To explain this we will look at
Figure 11. The derivative at the fixed point of the piecewise linear map is greater
than one and repelling but does not repel as fast as the fixed point of the piecewise
nonlinear return map. Thus, orbits close to the fixed point take a few more itera-
tions to escape on the piecewise linear map than on the piecewise nonlinear return
map.

Next we numerically approximate the fraction of iterates in a neighborhood of

zero over the λ-parameter values 0 through 1. This is done for both the piecewise
 24

nonlinear return maps and the corresponding piecewise linear maps. Note that zero
on the return map represents the equilibrium point of the homoclinic orbit. Thus
this simulation will show the change in orbit distribution near the equilibrium point
or equivalently the homoclinic orbit. Figures 12, 13 are generated by taking 10,000
iterations of a random orbit, counting how many land in the interval [0,.01) and
then dividing by 10,000. Notice that at λ = 0 the fraction of iterates of a random
orbit equals one, due to the fact that every orbit eventually lands on x = 0. Note
also that the differences between the piecewise linear and the piecewise nonlinear
maps seems negligible.

In Figure 14 we compare the Lyapunov numbers we calculated in Section 5 with

ones we approximated using 10,000 iterations. The approximated Lyapunov num-
bers for the piecewise linear maps are calculated for 200 parameter values 0 ≤ λ ≤ 1,
evenly incremented. The filled dots represent the approximate Lyapunov num-
bers while the circles correspond to the Lyapunov numbers at the parameter values
3 log 2
λ = 1/2i , i = 1, 2, ..., 10. The interpolating function lλ = 2 2 log 2−log λ with 0 ≤ λ ≤ 1
for the Lyapunov numbers is represented by the curve on Figure 14. Note also that
the Lyapunov number is greater than one for all parameter values greater than zero.
At λ = 0 all orbits are attracted to the fixed point.

We now turn our discussion to other possible bifurcations of the singular at-
tractors derived in Section 2. For example we can look at what happens when the
flow on the top section of the Z-switch moves from left to right. Figure 15 shows
how this phenomena looks as well as how the corresponding return maps behave.
Changing the flow in this way corresponds to changing the homoclinc orbit in such
a way so that on the return map the point that gets mapped to 0 is the bifurcation
parameter value, ranging from 1/2 to 1. This is a very interesting bifurcation, for if
we look at the corresponding piecewise-linear maps, it then changes from the baker
map to the tent map. These two maps are completely understood individually but
have not been studied as a bifurcation problem from each other in any plausible
system of differential equations. Bifurcation diagrams Figures 16,17 are constructed
by plotting the last 200 iterates of 300 at 5000 parameter values between 1/2 and
1 evenly incremented. One can analyze part of this bifurcation problem using the
similar techniques developed in this paper.

Another bifurcation problem one can look at corresponds to changing where the
flow from the top section of the Z-switch lands. For example consider changing
where the flow drops so that it over shoots the equilibrium point, see Figure 18.
In this case we do not have the homoclinic orbit and so on the return map, no
points map to zero. Note also that at 3/4 the slope of the return map is zero. The
bifurcation parameter represents the value of the return map at 3/4. Figure 19
shows the bifurcation diagram constructed by plotting the last 200 iterates of 300 at
 25

5000 parameter values between 0 and 1/2 evenly incremented. For this bifurcation
diagram we assumed that the difference in values of the map between x = 1/2 and
x = 3/4 is kept constant at 1/2. We also assumed the same for the values of the
map between x = 1 and x = 3/4.

We end this paper by pointing out that there are many other questions one might
ask about the singular attractors. For examples, does the natural measure exist for
all λ for the piecewise linear models? Can results for the piecewise linear models at
the sequence λ = 1/2i be extended to the piecewise nonlinear return maps? Do the
attractors persist for the perturbed systems at 0 < ε ¿ 1 and in what sense? This
paper only raises more questions than it can answer.

References
[1] Arnéodo, A., P. Coullet, and C. Tresser, Oscillations with chaotic behavior: an
illustration of a theorem by Šilnikov, J. Statist. Phys. 27(1982), pp.171–182.

[2] Alligood, K.T., T. Sauer, and J.A. Yorke, Chaos–An Introduction to Dynamical
Systems, Springer-Verlag New York, 1996.

[3] Bonet, C., Singular perturbation of relaxed periodic orbits, J.D.E., 66(1987),
pp.301–339.

[4] Chua, L.O., The genesis of Chua’s circuit, AEÜ, 46(1992), pp.250–257.

[5] Coullet, P., C. Tresser, and A. Arnéodo, Transition to stochasticity for a class
of forced oscillators., Phys. Lett. A 72(1979), pp.268–270.

[6] de Melo, W., and S. van Strien, One-Dimensional Dynamics, Springer-Verlag

Berlin Heidelberg, 1993.

[7] Deng, B., On Šil’nikov’s homoclinic-saddle-focus theorem, J.D.E., 102(1993),

pp.305–329.

[8] Deng, B., Folding at the genesis of chaos, Proceedings of the First World
Congress of Nonliniear Analysts’ 92., ed. V. Lakshmikantham, Walter de
Gruyter, Berlin/New York,(1996) pp.3765–3777.

[9] Deng, B., Construction homoclinic orbits and chaotic attractors., Int. J. Bifur-
cation and Chaos, 4(1994), pp.823–841.

[10] Deng, B., and Sakamoto, K., The subshift dynamics for the attractor near an
orbit homoclinic to a Hopf equilibrium. J. D. E. 119(1995), pp.1–23.
 26

[11] Devaney, R.L., An Introduction to Chaotic Dynamical Systems, Benjamin Cum-

mings Publishing Company, Inc., 1986.

[12] Evans, J.W., N. Fenichel, and J.A. Feroe, Double impulse solutions in nerve
axon equations, SIAM J. Appl. Math. 42(1982), pp.219–234.

[13] Fenichel, N., Geometric singular perturbation theory for ordinary differential
equations., J. D. E., 73(1979), pp.513–527.

[14] Guckenheimer, J., and P. Holmes, Nonlinear Oscillations, Dynamical Systems,

and Bifurcations of Vector Fields., Springer-Verlag, New York, 1983.

[15] Hastings, S.P., Single and multiple impulse waves for the FitzHugh-Nagumo
equations, SIAM J. Appl. Math. 42(1982), pp.247–260.

[16] Kemeny, J.G., J.L. Snell, Finite Markov Chains., Springer-Verlag New York,
1976.

[17] Mãné, R., Ergodic Theory and Differentiable Dynamics, Springer-Verlag Berlin
Heidelberg, 1987.

[18] Rössler, O.E., Chaotic behavior in simple reaction systems, Z. Naturforsch,

A31(1976), pp.259–264.

[19] Rössler, O.E., Continuous chaos—four prototype equations, in Bifurcation The-

ory and Applications in Scientific Disciplines, eds. Gurd, O. and Rössler, O. E.,
Annals of The New York Academy of Science, 316(1979), pp.376–392.

[20] Šil’nikov, L.P., A case of the existence of a denumerable set of periodic motions,
Sov. Math. Dokl., 6(1965), pp.163–166.

[21] Šil’nikov, L.P., The existence of a countable set of periodic motions in the neigh-
borhood of a homoclinic curve, Sov. Math. Dokl., 8(1967), pp.102–106.

[22] Šil’nikov, L.P., A contribution to the problem of the structure of an extended

neighborhood of a rough equilibrium state of a saddle-focus type, Math. USSR-
Sb, 10(1970), pp.91–102.

[23] Taylor, K., Chaotic Attractors in One-Dimension Generated by a Singular

Šil’nikov Orbit, Dissertation, University of Nebraska–Lincoln, 1999.

[24] Tresser, C., About some theorems by L. P. Šilnikov, Ann. Inst. H. Poincaré,
40(1984), pp.440–461.
 26

Due to a TeX problem ignor this page and the text pages 41, 40.

7 Introduction
In the 1960’s Shilnikov studied the dynamics of a 3-dimensional system of ordinary
differential equations containing an orbit homoclinic to a saddle-focus equilibrium
point. In a series of papers he showed that such a system can produce chaos in a
neighborhood of the orbit. See Šil’nikov [20], [21], [22]. His saddle-focus homoclinic
orbit is one of the simplest structures from which chaos occurs. This type of chaos
phenomena have been observed in many physical systems, such as neural dynamics,
chaotic circuits, nonlinear laser systems, and fluid dynamics, see for examples [4],
[5], [12], [14], [15], [18], [24]. His result has also been refined and extended by others,
see e.g., [7], [10], [24].

All prior results in the literatures on Shilnikov’s orbit have been restricted to
some subsets of a small neighborhood of the orbit. However the reason that most
of the related chaotic phenomena are observable is due to the existence of a chaotic
attractor containing such an orbit. Therefore, it is equally important, if not more, to
study the attractor which the orbit spawns. The purpose of this paper is to study a
type of such attractors that arise from singularly perturbed saddle-focus homoclinic
orbits in some 3-dimensional singularly perturbed differential equations.

A brief outline of this paper is as follows. First we will follow the modeling
strategy presented in [8, 9], as well as in [18, 19], to derive our 1-dimentional models.
The singularly perturbed system in this paper has the form:

ẋ = f (x, y, z), ẏ = g(x, y, z), εż = h(x, y, z).

The attractors are constructed by limiting orbits of the singular perturbed system.
A limiting orbit is one that connects the flow of the slow subsystem to the flow of the
fast subsystem. The singular Shilnikov orbit is generated in this way so that it starts
and ends at one equilibrium point. It can be proved by Fenichel’s geometric the-
ory of singular perturbations [13] together with Bonet’s result on turning points [3]
that a Shilnikov orbit Γε persists for a small perturbation of the system at each
0 < ε ¿ 1. The singular Shilnikov orbit corresponds to the limit Γ0 = limε→0 Γε .

Studying the limiting system is considered important because whatever limiting

structure presents at ε = 0 can be considered as the origin of the perturbed, full
structure. Figure 1 illustrates the type of the 3-dimensional systems that we are
considering in this paper. For such a system, the slow subsystem is 2-dimensional
in x and y, thus any 2-dimensional Poincaré section to the flow, in particular, to the
2-dimensional slow manifold produces a 1-dimensional Poincaré return map at the
 27
A2
B2 T1
S1 C2 A3
B3
C3

S2
0 A0 B 0 C 0
A1
A4
B1
0 S3
C1 A0
C4
T2 B0

C0
A5

Figure 1: The Z-switch and corresponding return map. The interval [0,C0 ] is marked
on the S3 section of the slow manifold by a dotted line.
 28

λ λ

(a) (b)

Figure 2: The return map and its piecewise linear caricature.

 29

x1 x*
x
x0 xa

S3
y

Figure 3: The homoclinic orbit on the S3 section of the Z-switch. Point A0 is labeled
xa , and x∗ corresponds to A1 , the branching point.
 30
A2
B2 T1
S1 C2
B3 A3

C3
S2
0 A0 B 0 C0
A1

B1
0 S3
C1 A0

T2

C0
B0

Figure 4: Points starting in the shaded region will escape to infinity along with all
of their preimages.
3
 31

A2
B T1
S1
2
C2
B 3

S2
(a) 0 A0 B 0 C0
A1
B 1

0 S3
C1 A0

T2 B 0

C0

A2
B T1
S1
2
C2 A3
B 3

C3

S2 0 A0 B 0 C0
(b)
A1
A 4
B 1

0 S3
C1 A0

T2 B 0

C0
A 5 = C4

A2
B T1
S1
2
C2 A3
B 3

C3
S2
0 A0 B 0 C0
(c) A1

B 1

0 S3
C1
A0
T2 B 0

C0

Figure 5: Changing the flow on the S1 section of the Z-switch can be represented
on the return map by changes in the height of the images at A0 and C0 .
 32

λ = 1/2

f λ (x2 ) = f λ (x1 )

x1 x2

Figure 6: The two points x1 and x2 map to the same value and so their itineraries
are φ(x1 ) = .BAs2 s3 ... and φ(x2 ) = .CAs2 s3 ... respectively, the only differences in
the sequences being the first entry. The distance between the sequences φ(x1 ) and
φ(x1 ), d(φ(x1 ), φ(x2 )) = K.
 33
( a) ( b)

f λ−1 ( J ) f λ−1 ( J )

Figure 7: In (a), if a < λ we see that the preimage of J consists of three open
intervals, one contained in each of the intervals IA , IB , or IC . In (b), if a > λ we
see that the preimage of J is one open interval contained in IA .
 34

13
λ =
16

13 13 1 3 13
64 32 2 4 16

Figure 8: Here λ = 13/16 and fλ3 (λ) = λ and so n = 3. Since 1/2 and 3/4 are not
an iterate of λ, h = 2, and m = 0.
 35

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 9: Bifurcation diagram of the nonlinear return map.

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 10: Bifurcation diagram of the piecewise linear return map.

 36
(a) (b)

Figure 11: Figure (a) shows how points near the fixed point on the piecewise linear
return map take a few more iterations before escaping its neighborhood than the
nonlinear return map does as in (b).
 37

0.8

0.6
Measure

0.4

0.2

0 0.2 0.4 0.6 0.8 1

a

Figure 12: The fraction of iterates of the nonlinear return map Fλ out of 10,000 that
land in the interval [0,.01). There are 301 points plotted taken over the parameter
values 0 through 1 at even increments.

0.8

0.6
Measure

0.4

0.2

0 0.2 0.4 0.6 0.8 1

a

Figure 13: The fraction of iterates of the linear return map fλ out of 10,000 that
land in the interval [0,.01). There are 301 points plotted taken over the parameter
values 0 through 1 at even increments.
 38

3.5

2.5
Lyapunov Number

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1
lambda

Figure 14: Approximate Lyapunov numbers taken over 10,000 iterates compared to
the Lyapunov numbers calculated in Section 5.
 39
A2
B2 T1
S1 C2
A3
B3
C3

S2 0 A0 B 0 C 0

A1
B1
0 S3
C1

T2 B0

C0 =C4
A 0= B 5

A2
B2 T1
S1 C2 A3
B3
C3

S2 0 A0 B 0 C 0

A1
B1 A4
0 S3
C1 A0

T2 B0

C0
A 5= C 4

Figure 15: The change of the flow on the S1 section of the Z-switch and the corre-
sponding return maps.
 40

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

Figure 16: Bifurcation diagram of the nonlinear return map.

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

Figure 17: Bifurcation diagram of the nonlinear return map.

 41

limit ε = 0. More specifically, a 1-dimensional return map is constructed by picking

an interval that crosses the sub-flow on the slow manifold transversally and map-
ping it back into itself via the flows. The 1-dimensional return map of our system
is piecewise continuous as illustrated in both Figure 1 and Figure 2(a).

The full return map is difficult to analyze because of the infinite derivative at
x = A+ 0 as shown in both Figure 1 and Figure 2(a). Instead we will consider a
simplified, caricature model for the full return map. The caricature model is given
as the family of piecewise linear maps
(
2x, if 0 ≤ x ≤ 1/2
fλ (x) =
4λ | x − 3/4 |, if 1/2 < x ≤ 1,

where λ ∈ [0, 1] and A0 , B0 , and C0 are represented by 1/2, 3/4, and 1 respectively,
see Figure 2(b). The degree of difficulty between treating the full return maps and
the simplified piecewise linear models is similar to that between the logistic map
and the tent map for which the former is much harder to analyze because of its
vanishing derivative at its only critical point. Nonetheless we believe that results on
the piecewise linear models should shed some light on the more complex return maps.

Most of our results, analyses, discussions are for the piecewise linear models at
a sequence of parameter values λ = 1/2i , i = 1, 2, 3, . . .. This limitation is due to
what our method of this paper can do at this point. Nevertheless the sequence does
give a fair representation to the bifucation parameter range λ ∈ [0, 1] as we will see
throughout the paper. We will obtain results on the following aspects of the attrac-
tors: symbolic dynamics, invariant measures, and Lyapunov exponents for sensitive
dependence on initial conditions.

For symbolic dynamics, we will show that the piecewise linear maps at λ = 1/2i ,
i = 1, 2, 3, . . . are conjugate to a subshift on 3 symbols. More specifically, the
conjugate map is the itinerary map defined as

 A
 if fλj (x) ∈ IA
φ(x) = (s0 s1 s2 ...) = (s) such that sj = B if fλj (x) ∈ IB

C if fλj (x) ∈ IC .


Here the intervals IA , IB , and IC represent the intervals [0,1/2], (1/2,3/4], and
(3/4,1] respectively. The sequence space Σλ is made up of sequences of the three
symbols A, B, and C that obey rules determined by the parameter value λ. Typical
sequences are made up of many A’s and a few isolated B’s and C’s separated by the
A’s. The shift map on Σλ is defined as usual: σ(s0 s1 s2 . . .) = (.s1 s2 . . .). It simply
removes the first entry and shifts all other entries to the left one place.
 40

For the same parameter values we will show that there exists an invariant prob-
ability measure and calculate it explicitly on a Markov partition given as the set of
intervals ∪i+2
k=1 Ik with

1 1 1 1 1 1 3 3
I1 = (0, i
), I2 = ( i , i−1 ), ..., Ik = ( i+2−k , i+1−k ), ..., Ii+1 = ( , ), Ii+2 = ( , 1).
2 2 2 2 2 2 4 4
The measure that we will derive is the “natural” measure, so called because it mea-
sures how frequent an orbit visits a conceivable region or the fraction of the orbit
that falls into the region. Roughly speaking, to calculate the natural measure of a
map f on a particular region, you iterate the map f , count the amount of iterates
in that region, divide by the number of iterates, and take the limit as the number
of iterates goes to infinity. We will show that the number of iterates falling in the
intervals I2 through Ii are the same, but since their lengths are different their den-
sities will be different. The number of iterates in each of these intervals will be half
the number of iterates in I1 and twice the number in Ii+1 and Ii+2 .

We will also calculate the Lyapunov numbers for orbits of the family of maps with
the same parameter values as above. Lyapunov numbers measure the separation rate
of orbits on the attractors. The Lyapunov number L(x1 ) of the orbit {x1 , x2 , ...} is
defined as
lim (| f 0 (x1 ) | · · · | f 0 (xn ) |)1/n
L(x1 ) = n→∞
if this limit exists. We will use the natural measure to calculate the Lyapunov num-
ber for λ = 1/2i , i = 1, 2, 3, . . ., all of which are greater than one, implying that
the system has a sensitive dependence on initial conditions in the sense that points
separate at an exponential rate.

We will look at numerical approximations of the Lyapunov numbers for the

piecewise linear maps and compare them to our results. They seem to suggest that
3 log 2
2 2 log 2−log λ is a good interpolation for the Lyapunov number Lfλ for all λ ∈ [0, 1]. In
other numerical simulations we will compare the piecewise linear caricature models
and the full return maps.
 41

S1 T1

0 A0 B 0 C 0

S2

0 S3
A0
T2 B0

C0

Figure 18: The change in the Z-switch so that the flow on the S1 section drops onto
the S3 section past the equilibrium point and the corresponding return map.
 42

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Figure 19: Bifurcation diagram of the nonlinear return map as the value at 3/4
changes from 0 to 1/2.