J. Nonlinear Sci. Vol. 15: pp.

63–91 (2005)
DOI: 10.1007/s00332-004-0620-2

© 2005 Springer Science+Business Media, Inc.

Life Expectancy of Transient Microchaotic Behaviour

G. Csernák1 and G. Stépán1
1
Research Group on Dynamics of Machines and Vehicles, Hungarian Academy of Sciences,
Budapest University of Technology and Economics, Department of Applied Mechanics,
H-1521 Budapest, Hungary
E-mail: [email protected]
Received December 17, 2003; accepted November 30, 2004
Online publication March 29, 2005
Communicated by A. R. Champneys

Summary. The application of digital control may lead to so-called transient chaotic
behaviour. In the present paper, we analyse a simple model of a digitally controlled
mechanical system, which may create such vibrations. As a consequence of the digital
effects, i.e., the sampling and the round-off error, the behaviour of this system can be
described by a one-dimensional piecewise linear map. The lifetime of chaotic transients
is usually characterized by the so-called escape rate. In the literature, the reciprocal
of the escape rate is considered to be the expected duration of the transient chaotic
phenomenon. We claim that this approach is not always fruitful, and present a different
way of calculating the mean lifetime in the case of one-dimensional piecewise linear
maps. Our method might also be used to solve diffusion problems in one-dimensional
models of periodic arrays.

1. Introduction

There are several vibration problems where chaotic oscillations occur. These oscillations
often disappear suddenly; this phenomenon is referred to as transient or metastable
chaos [1].
Since the duration of the transient chaotic behaviour is finite, it cannot be described
as a motion on a strange attractor. There is another invariant set in the phase space of
transient chaotic systems that is responsible for the irregular vibrations, the so-called
chaotic saddle or chaotic repeller [2]. The chaotic repeller can be considered as a dense
but measure zero set of repelling (hyperbolic) points, which form a strange structure, a
“maze” for the solution curves. If a solution arrives at this maze, a very long time might
be needed for it to leave. In what follows, the leaving solutions are assumed to tend
to attracting fixed points. The observed law governing this escape is quite simple: The
probability that a solution has not yet escaped from the small neighborhood of the repeller
64 G. Csernák and G. Stépán

decays exponentially with time in an asymptotic sense. The exponent is called escape
rate κ. The reciprocal of the escape rate is usually considered as the mean lifetime of the
transient chaotic phenomenon [3], [4], [5]. However, the exponential decay of the escape
probability is only asymptotically fulfilled, and it can be shown [6] that, in the case of
maps—even in the hypothetical case of exact exponential decay—a different relation is
fulfilled between these quantities. The so-called mean kickout number Nm —the average
number of steps before escape—is given as

∞
ne−κn 1
Nm =
n=1
∞ = . (1)
n=1 e
−κn 1 − e−κ

In the case of a small escape rate and a long lifetime, 1/(1−e−κ ) ≈ 1/κ, indeed. However,
in this case, the exponential decay of probabilities is fulfilled only after a significantly
long time.
The determination of the mean lifetime of transient chaotic oscillations is important
for the design work, since it estimates the length of time needed to reach a steady state.
The machine structure has to survive these transients without damage. This is why the
life expectancy of transient chaos may become such an important design parameter, as
the damping factor or the logarithmic decrement is.
The nonlinearity caused by the quantization in digital control may lead to transient
chaotic behaviour. The amplitude of such oscillations is usually very small; this is why
they are called microchaotic vibrations [7], [8]. In this paper, we introduce and anal-
yse a mechanical system of this kind. As a consequence of the digital effects, i.e., the
sampling and the round-off error, the behaviour of this system can be described by a
one-dimensional piecewise linear map. We developed a new recursive method for the
calculation of the mean lifetime of transient chaotic phenomena in the case of such maps.
With slight modifications, our method might also be used to solve diffusion problems in
one-dimensional models of periodic arrays [9], [10].

2. Mechanical Model

The tools of polishing machines usually move slowly along the workpiece, but sometimes
the polishing tool must be stopped and its position must be kept for a while. If the machine
in question is not a hand tool, it has a support that has to make this kind of operation
possible.
The corresponding mechanical model is presented in Figure 1. It consists of a revolv-
ing cylinder sliding on the rough surface of a fixed block. The shaft of the cylinder is
attached to a support, which is modeled by the parallel connected linear spring of stiff-
ness s, viscous damper of damping coefficient k, and dry friction force C. The velocity
of the shaft is denoted by ẏ = v, while the circumferential velocity of the polishing tool
equals v0 = Rω0 . The friction force acting between the fixed block and the cylinder
depends on the difference v0 − v of these two velocities. This friction force cannot be
approximated by either a linear viscous damping force or a simple Coulomb friction
force. The commonly used friction force characteristic µ(v0 − v) = µ(vrel ) is a mixture
of these two friction characteristics (see Figure 1).
Life Expectancy of Transient Microchaos 65

0 y
s R (vrel)
!0
k
v
C
v0 v m v0 vrel = v0 v

Fig. 1. The mechanical model.

At low relative speeds vrel = v0 − v, the combined dry and viscous friction force
acting on the cylinder is locally decreasing as the relative velocity v0 − v increases. If
the shaft of the polishing tool does not move, the friction force equals mgµ(v0 ). With
certain forces arising in the support, the cylinder is in equilibrium.
To determine the properties of this possible equilibrium, one can use the linearized
equation of motion. Since the support can be characterized by linear spring and linear
dampers, only the mixed friction force must be linearized for this purpose, about ẏ = v =
0: mgµ(v0 − v) ≈ mgµ(v0 ) − mgµ (v0 ) ẏ, where the prime refers to the differentiation
with respect to the relative velocity.
Thus, the linearized equation of motion assumes the following form:

m ÿ = mgµ(v0 ) − mgµ (v0 ) ẏ − C sgn( ẏ) − k ẏ − sy, (2)

where sgn() denotes the signum function, and the latter three terms correspond to the
reaction forces acting at the support. After rearranging the terms, one obtains
 
m ÿ + k + mgµ (v0 ) ẏ + sy = mgµ(v0 ) − C sgn( ẏ). (3)

It is clear from the properties of the dry friction not only that a single point can be the
equilibrium, but that there is a stripe where the perturbed motion of the system can come
to a rest. The borders of this stripe can be given as

mgµ(v0 ) ± C
ye = . (4)
s
The equilibrium states within this stripe are stable if the coefficients of the homogeneous
part of (3) are strictly positive. This condition is fulfilled if k > −mgµ (v0 ), since m, k,
and s are positive and µ (v0 ) is negative.
The above-described support cannot provide stationary motion for the polishing tool,
nor can the position of the equilibrium be changed. A solution of this problem might be the
use of a simple control system that “imitates” the mechanical support. The corresponding
mechanical model can be seen in Figure 2.
66 G. Csernák and G. Stépán

PC

0 v
!0 R
C
DC m
v0 v
(v0 v)

Fig. 2. The mechanical model of the computer-controlled system.

In this case, the shaft of the polishing tool is driven by a DC motor, which exerts a
control force Q = −mgµ(v0 ) − Dv. With this force, the tool can be held at a certain
position without a proportional gain, i.e., without a (virtual) spring force. D denotes a
differential gain, with which the strength of the control can be tuned. Thus, the linearized,
first-order equation of motion of the controlled system is obtained as

m v̇ + mgµ (v0 )v = mgµ(v0 ) − C sgn(v) −mgµ(v0 ) − Dv . (5)

  
Q

For the sake of simplicity, we denote |µ (v0 )| = −µ (v0 ) > 0 by f . Using this notation,
we can rewrite the equation of motion as

m v̇ + (D − f mg)v = −C sgn(v). (6)

According to the Routh-Hurwitz criterion, the solution v = 0 is stable if the differential

gain D is greater than f mg. Thus, if we use an analogue control system, the appropriate
choice of D results in the asymptotic stability of the equilibrium state.
However, we would like to use digital control, where one encounters the following
deficiencies:
• The computer samples the velocity difference v at discrete time instances, v j ≡ v(t j ) ≡
v( jτ ), j ∈ {1, 2, . . .}; t j = jτ is the jth sampling instant, and τ is the sampling time.
Moreover, some time is needed to process the measured signal; thus, the force is
exerted by the motor a bit after the sampling instant. This processing delay is denoted
by t.
• The velocity measurement has a finite resolution, h.
As a first step, let us take into account the effect of temporal discretization only. In
this case, the control force is assumed to be constant while t ∈ [t j + t, t j+1 + t),
where it is given as Q = −mgµ(v0 ) − Dv(t j ). For the sake of simplicity, we assume
that the processing delay t is negligible.
Life Expectancy of Transient Microchaos 67

Thus, we arrive at the following linearized equation of motion, which is valid between
two successive sampling instants, t ∈ [ jτ, ( j + 1)τ ):
D C
v̇(t) − f gv(t) = − v(t j ) − sgn(v(t)). (7)
m m
Now, let us introduce a nondimensional time T , which is defined as t = τ T . In the
following, the differentiation with respect to the variable T will be denoted by the prime.
Thus, v̇ ≡ v  /τ , and so
Dτ Cτ
v  (T ) − f gτ v(T ) = − v( j) − sgn(v(T )), T ∈ [ j, j + 1) (8)
m m
is the nondimensional form of the equation of motion.
Let us assume now, temporarily, that the velocity v(T ) is positive if T ∈ [ j, j + 1).
This means that the sgn() function disappears from the equation, and we arrive at a first-
order linear ordinary differential equation, with constant term on the right-hand side.
The solution of this equation can be obtained as
D C
v(T ) = vh (T ) + v p (T ) = K e f gτ T + v( j) + , (9)
f mg f mg
where vh and v p refer to the general solution of the homogeneous equation and a particular
solution of the inhomogeneous equation, respectively. K denotes a constant, which
depends on the initial condition, i.e., on the velocity at the jth sampling instant:
D C
v( j) = K e f gτ j + v( j) + . (10)
f mg f mg
Thus, K can be expressed as

D C
K = e− f gτ j 1− v( j) − , (11)
f mg f mg
and the solution of (7) is
D C f gτ (T − j) D C
v(T ) = 1 − e f g(T − j) v( j) − e + v( j) + . (12)
f mg f mg f mg f mg
Note that equations (9)–(12) are valid only for T ∈ [ j, j + 1).
Using (12), a one-dimensional mapping can be defined as follows:
D  f gτ    C
v( j + 1) = e f gτ − e − 1 v( j) − e f gτ − 1
f mg f mg
=: ãv( j) − S̃, (13)
where the sign of the last term changes in the case of negative velocity.
As is known from the theory of difference equations [11], the solution of (13) can be
given as
j−1
v( j) = v(0)ã j − S̃ ã j−i−1 . (14)
i=0
68 G. Csernák and G. Stépán

This solution is stable if |ã| < 1, i.e., if |e f gτ − f Dmg (e f gτ − 1)| < 1. After rearranging
terms, we obtain that ã < 1 if (1 − f Dmg )(e f gτ − 1) < 0. This is fulfilled if

D > f mg, (15)

which is the same stability condition as in the continuous case. However, ã > −1 must
be fulfilled, too. This condition can also be reformulated. It is fulfilled if e f gτ ( f Dmg −1) <
D
f mg
+ 1, and, consequently, if

1 D + f mg
τ< log . (16)
fg D − f mg

Thus, a sufficiently small sampling time and a properly tuned differential gain may result
in the stability of the solution (14).
Now, let us take into account the discretization in the velocity measurement, too.
Doing so means that the measured analogue data are transformed into discrete data.
Thus, the input of the control system is a multiple of the smallest number that can be
represented digitally. This number will be denoted by h, which determines quantization,
and is also equal to the round-off error.
Using this discretization, (8) turns into

Dhτ v( j) Cτ
v  (T ) − f gτ v(T ) = − Int − sgn(v(T )), T ∈ [ j, j + 1), (17)
m h m

where Int() stands for the integer-part function. The solution of this equation is similar
to (9), and the constant K can be expressed similarly to (11):

Dh  f gτ (T − j)  v( j)   C
v(T ) = e f gτ (T − j) v( j)− e −1 Int − e f gτ (T − j) −1 . (18)
f mg h f mg

Finally, instead of obtaining (13), we arrive at the following one-dimensional mapping:

Dh  f gτ  v( j)   C
v( j + 1) = e f gτ v( j) − e − 1 Int − e f gτ − 1 . (19)
f mg h f mg

By introducing a new variable x j = v( j)/h, this map can be transformed to a more

tractable form:
D  f gτ      C
x j+1 = e f gτ x j − e − 1 Int x j − e f gτ − 1 . (20)
f mg f mgh
After introducing new parameters

a = e f gτ > 1, (21)
D  f gτ 
b = e − 1 > 0, (22)
f mg
  C
S = e f gτ − 1 > 0, (23)
f mgh
Life Expectancy of Transient Microchaos 69

xj +1
2

3 2 1 x
x 1 2 3 xj

Fig. 3. The microchaos map at a = 1.5, b = 1.1, S = 0.3, and d = 3.

we obtain a modified version of the so-called microchaos map [7]:


ax j − b Int(x j ) − S if S/a < x j < d,
Fm : x j+1 = 0 if −S/a ≤ x j ≤ S/a, and (24)

ax j − b Int(x j ) + S if −d < x j < −S/a,

where the parameter d characterizes the size of the domain of definition. Note that the
original microchaos map (see [7]) is the special case of (24), with S = 0.
The boundaries S/a and −S/a refer to the fact that the mapping (24) could assume
negative values at certain positive values, and vice versa. However, if the velocity v
becomes zero while the control force is small—i.e., when the velocity at the latest
sampling instant is small—the motion is stopped and trapped by the static dry friction
force, which can be much greater than C. This is why the mapping is assumed to map
to zero if x ∈ [−S/a, S/a].
An example for such maps is presented in Figure 3. As can be seen, it is an odd,
piecewise linear, discontinuous map. This iterated map [12] may serve as a simple
cartoon model for other digitally controlled systems, too. The intersections of the graph
and the dashed line of modulus one are the fixed points of the map. Two of them—
denoted by x ∗ and −x ∗ in Figure 3—form the basin boundary of the direct attracting
domain near the origin.
It was shown in [7] that the solutions of the original microchaos map may behave
chaotically in finite parameter domains, even if −1 < a − b < 1, i.e., if inequalities
(15) and (16) are fulfilled. The amplitude of this chaotic vibration is very small; it is in
the order of the resolution of the velocity measurement h. This is where the name of the
map comes from.
Since there is an attracting domain [−S/a, S/a] in the domain of definition (−d, d)
of the map (24), the existence of transient chaos is expected in this case. In Figure 4,
the graph of the map (24) can be seen for the parameters a = 3/2, b = 6/5, S = 2/15,
70 G. Csernák and G. Stépán

1.5
xj +1

0.5

I0
0
0 S=a x 0.5 1 a S 1.5 2
xj
Fig. 4. A long trajectory at a = 1.5, b = 1.2, d = 2, and |I0 | = 0.1. The initial
point is 1.472444646985281, and the length of the trajectory is 354.

and d = 2 (thick lines), with quite a long trajectory (thin lines), which eventually
arrives at the domain of attraction of the origin. It can be seen that there is a subinterval
I0 = [a − b − S, x ∗ ) left of the fixed point x ∗ , which is directly reachable from the right
by a trajectory; see the dashed lines in the figure. If a solution arrives at I0 , the transient
chaotic behaviour is over. The size of I0 can be given as
S
|I0 | = a − (a − b). (25)
a−1
We introduce |I0 | as a new parameter instead of S, because it characterizes the system
better than the friction parameter S, and it is uniquely related to that. The position of the
fixed point expressed by this new parameter is
S a − b + I0
x∗ = = . (26)
a−1 a
Note that, despite the existence of the attracting domain at the origin, at certain
parameters the mapping can still assume negative values near the small positive integer
values. As an example, consider the case a = 2.8, b = 2.6, |I0 | = 1.04̇, d = 2. The
corresponding graph is presented in Figure 5.
Determine the condition of the occurence of such a situation. We focus our attention
on the value x = 1, since here the function (24) takes its local minimum, a − b − S =
1 + (I0 − b)/a − |I0 |.
Thus, the condition of mapping to the other half-plane is
|I0 | − b a−b
1+ − |I0 | < 0 ⇐⇒ < |I0 |. (27)
a a−1
Note that we assumed during the derivation of the map (24) that the sign of the
velocity did not change between two successive sampling instants. Thus, map (24) has
Life Expectancy of Transient Microchaos 71

xj +1
2

0 1 xj
2 1 2

Fig. 5. The mapping can map from the positive half-plane to the negative one,
and back. a = 2.8, b = 2.6, S = 0.8 (|I0 | ≈ 1.044), and d = 2.

to be modified to be valid in case of sign changes, too. Consider the case when the sign
of the velocity changes from positive to negative.
Using the definition of the new variable x and parameters a, b, and S, (18) takes the
form
  b   S
x(T ) = a T − j x( j)− a T−j −1 Int (x( j))− a T − j −1 , T ∈ [ j, j+1). (28)
a −1 a −1
We exploited here that D/( f mg) = b/(a − 1) and C/( f mgh) = S/(a − 1). According
to (28), x(T0 ) = 0 is fulfilled if
S + b Int(x( j))
T0 = j + log / log(a). (29)
x( j) − ax( j) + S + b Int(x( j))
This means that v(T0 ) = 0, and (17) is reduced to
Dhτ v( j)
v  (T0 ) = − Int . (30)
m h
Thus, at T = T0 the shaft stops and the sign of its velocity changes. So, the equation of
motion (17) has to be modified for the rest of the sampling interval with opposite friction
force as follows:
Dhτ v( j) Cτ
v  (T ) − f gτ v(T ) = − Int + sgn(v(T )), T ∈ [T0 , j + 1). (31)
m h m
This equation can be solved similarly to (17) by exploiting that v(T0 ) = 0. Finally, we
arrive at
  b     S
x j+1 = 1 − a j+1−T0 Int x j − 1 − a j+1−T0
a−1 a−1
 
b Int x j − S  
= a   x j − b Int x j + S. (32)
b Int x j + S
72 G. Csernák and G. Stépán

xj +1
2

2 0 2 xj

Fig. 6. The modified map (32) at parameters a = 2.8, b = 2.6, S = 0.8 (|I0 | ≈
1.044), and d = 2.

If the sign of the velocity changed the other way, i.e., from negative to positive, only the
sign of S would change in this formula. In Figure 6, the graph of the map (32) can be
seen at parameters a = 2.8, b = 2.6, S = 0.8, and d = 2.
Now we have a map that correctly describes the solutions of (17). In the follow-
ing, we will introduce a method for determining the mean lifetime of transient chaotic
oscillations. To simplify the treatise, we restrict ourselves to the cases defined below:
• If there is only one fixed point on the positive half-plane, the structure of the repeller
is remarkably simple (see the next section), so we consider only such situations. The
condition of having single nontrivial fixed points on both half-planes is

a2 − a − b
2a − b − S < 2 ⇐⇒ |I0 | > . (33)
a−1
In this case, the chaotic repeller of the map (24) is confined in the interval Irep =
l
[brep , brep
r
) = [x ∗ , a − S) (see Figure 4). The left border is defined by the border
of the direct domain of attraction of the interval [0, S/a], while the value a − S for
the right border comes from the following consideration: a solution, starting from a
greater initial value, first tends to the left, and sooner or later it takes a value less than
1. It can be seen from definition (24), and also in Figure 4, that if a solution arrives
below 1, later it cannot take a value greater than a − S. This means that the mixing
of possible trajectories—a necessary condition of chaos—can be realized only in the
domain situated to the left of the point a − S. Using the parameter |I0 |, the interval
containing the repeller can be given as

|I0 | + a − b (b − |I0 |)(a − 1)
Irep = [brep
l
, brep
r
)= ,1 + , (34)
a a
and the size of the repeller is |Irep | = b − |I0 |.
Life Expectancy of Transient Microchaos 73

0
0 1 2 3 4 5 6 7 8
Fig. 7. a = 1.75, b = 1.4, |I0 | = 0.1, d = 11, and x0 = 9.59.

• To keep the simple form of the map (24), we do not consider cases when the sign of
the velocity changes. There is only one exception: if the solution is mapped into the
basin of the direct domain of attraction. This means that the lifetime calculation is not
influenced by the change of velocity sign. Suppose that the map (32) takes a value
greater than −x ∗ at x = 1 (see Figure 6):

b−S S
a −b+S >− . (35)
b+S a−1
This is trivially fulfilled if the sign of the velocity does not change.
• Moreover, we will restrict ourselves to the neighborhood of the repeller, so we choose
d = 2. On large scales, map (24) behaves like a linear map with slope a − b. Since
a−b < 1, a solution starting from x0 > 2 tends to [0, 2) quite quickly, so this restriction
does not have a great influence on the lifetime calculation of chaotic transients. This
situation is illustrated in Figure 7.
Function Fm is odd; thus, the behaviour of solutions in (−d, 0] is just the same as in
[0, d). Since we excluded the change of velocity sign, in the following we may consider
the case x j ∈ [0, d) only, without further loss of generality.

3. Life Expectancy Calculation by Recursive Procedure

3.1. Fundamental Concepts of the Method

Since the lifetime calculation procedure is rather complex, we introduce it by calculating

the mean kickout number Nm in the case of a certain example. We chose the parameters
a = 1.5, b = 1.3, and |I0 | = 0.25. In this case, S = 0.15. The steps of the iteration
of map (24) correspond to successive sampling instants; thus the mean lifetime can be
obtained directly from the mean kickout number as tm = τ Nm , where τ is the sampling
time.
74 G. Csernák and G. Stépán

80

70

Kickout number 60

50

40

30

20

10

0
0 0.5 1 1.5 2
x0
(a) Result of numerical simulations

A B C
2

2
I F I1
1:5
1111
0000 0
1
0000
1111 00
11 0
1
0000
1111 00
11
0
100
11
0
1 0
1
0000
1111 00
11
0
100
11
0
1 0
1
0000
1111 00
11
0
100
11
0
1 0
1
0000
1111 00
11 00
11
0
100
11
0
1 0
1
+1

1 0000
1111 00
11 00
11
0
100
11
0
1 0
1
11
00 00
11
0
100
11
0
1 0
1
xi

00
11 0000
1111
0000
1111 00
11 00
11
0
100
11
0
1 0
1
00
11 0000
1111 00
11 00
11
011
100
0
1 0
1
00 11
11
00
11
00
00
11 0000
1111 00 11
11 00
00
11 00
11
0
1
00
1100
11
0
1 0
1
0
1
00
11 00
11 0000
1111 00
11
00
11 00
11 0
1
00
1100
11
0
1 0
1
0
100
11 00
11 00
11 0000
1111 00
11 00
11 00
11 00
11 0
1
00
1100
11
0
1 0
1
0:5 00
110
100
11 00
11 00
11 0000
1111 00
11 00
11 00
11 00
11 0
1
00
1100
11
0
1 0
1
00
110
100
11 00
11 00
11 0000
1111 0
1
00
11 00
11 00
11 00
11 0
1
00
1100
11
0
1 0
1
0
10
1
00
11
0
1 0
100
11 00
11 00
11 0000
1111 00
11 0
10
1
00
11 00
11 00
11 00
11 0
1
00
1100
11
0
1 0
1
0
10
1
00
11
0
1 0
100
11 00
11 00
11 0000
1111 0
1
00
11 0
10
1
00
11 00
11 00
11 00
11 0
1
00
1100
11
0
1 0
1
0
10
1
00
11
0
1 0
100
11 00
11 00
11 0000
1111 0
1
00
11 0
10
1
00
11 00
11 00
11 00
11 0
1
00
1100
11
0
1 0
1
0
10
1
011
I0 100
0
1 0
100
11
00
11 00
11 00
11 0
1
0000 1 11
1111 00
0
1 0
10 11
1
00
11 00 11
00
11 00 00
11 011
1
00
1100
0
1
00
11
0
1 0
1
0
1
011
100
0
1 0
15
00
11 4 00
113 00
11 21111
0000 11
00
0
1 0
10
1
00
11 5
00
11 4 00
113 00
11 2 0
1
11
0000
11
0
1 504
1
0
1
00
11
0
1 0
1 00
11 00
11 0000
1111 00
11

0
10
1
00
11 00
11 00
11 0
1
00
11
 1
0 
 r
0 x0 1 1
x brep x2 2 x 3
xi

(b) Fundamental escape and fractal intervals

Fig. 8. The structure of the domain of definition at a = 1.5, b = 1.3, d = 2, and |I0 | = 0.25.

To get a deeper insight into the mechanism of the transient chaotic behaviour, we per-
formed numerical simulations and determined the number of steps needed to escape—to
reach the interval I0 —from 7000 uniformly distributed points of the domain of definition
Iic ≡ [0, d) = [0, 2) of (24). The result is shown in Figure 8a.
Life Expectancy of Transient Microchaos 75

If we found the function describing the dependence of the kickout numbers on the
initial conditions, the mean kickout number could be determined as the integral of this
function divided by the size of the domain of definition.
The singularities of the distribution shown in Figure 8a seem to form a fractal. This
fractal structure can be explored by an analytical method, presented below.
If a trajectory arrives at I0 , the transient chaotic behaviour can be considered to be
over. Thus, if we want to know how many iteration steps are needed to reach I0 , it is
obvious that we must explore the structure of the domain of definition of the map by
constructing the pre-images of I0 . The first steps of this construction can be done quite
easily, by following the dashed lines shown in Figure 8b.
As can be seen in the figure, I0 has only one first pre-image, I1 = [x1l , x1r ) = [1, 1 +
|I0 |/a) = [1, 1.166̇). I1 has infinitely many pre-images on the left branch of the map,
i.e., in [x ∗ , 1) = [0.3, 1). These intervals will be referred to as fundamental escape
intervals. The sequence of the fundamental escape intervals accumulates at x ∗ ≡ x0∗ .
The left borders xil of these intervals are the pre-images of the left border of I1 , i.e., the
pre-images of x1l = 1:
1+S
ax2l − S = x1l ⇒ x2l = . (36)
a
Similarly,
xil + S
l
axi+1 − S = xil ⇒ l
xi+1 = . (37)
a
By solving this recursive equation, we obtain
i−1
S |I0 | − b b − |I0 |
xil = a 1−i + k
=1+ + . (38)
k=1
a a ai

The right borders of the pre-images of I1 are the pre-images of x1r = 1 + |I0 |/a:
|I0 | 1 + |I0 |/a + S
ax2r − S = 1 + ⇒ x2r = . (39)
a a
Similarly as above, we obtain
|I0 | − b b
xir = 1 + + i. (40)
a a
Here i refers to the number of steps necessary to reach I0 . This number will be referred
to as the kickout number. The kickout numbers of certain escape intervals are shown in
Figure 8b. The size of an interval Ii —from which the solution arrives at I0 in i steps—is
|I0 |/a i . The distance of the points xil and xir from the fixed point x ∗ is
b − |I0 | b
xil − x ∗ = and xir − x ∗ = , (41)
ai ai
respectively.
As can also be seen in in Figure 8b, there are intervening intervals between the
fundamental escape intervals (over which the shaded columns are drawn), from which
76 G. Csernák and G. Stépán

the number of steps needed to escape is not yet known. In what follows, we will refer to
these intervals as fundamental fractal intervals.
As we will see later, the structure of these intervals is self-similar. Thus, fundamental
fractal intervals contain smaller escape and fractal subintervals; this is where the name
comes from. The distance of the borders of these subintervals from the left border of the
fundamental fractal interval in question is given by (41).
The left and right borders of the fundamental fractal intervals are
|I0 | − b b
f il = 1 + + i+2 and
a a
|I0 | − b b − |I0 |
fi = 1 +
r
+ , (42)
a a i+1
respectively. The greatest fundamental fractal interval is denoted by F. The size of this
interval is
a(b − |I0 |) − b
|F| ≡ |F0 | = f 0r − f 0l = . (43)
a2
The size of the other fractal intervals can be expressed as
|Fi | = |F0 |/a i . (44)
Fi is mapped onto F ≡ F0 in i steps. This number will be referred to as the F-number
in the following. The F-number of a fractal interval is 1 less than the kickout number
of the neighboring escape interval on the right, and 2 less than the kickout number of
the escape interval on the left. For example, the two neighbors of the interval F (with
F-number zero) are the intervals I1 from the right and I2 from the left.
The sequence of fundamental escape intervals and fundamental fractal intervals forms
the higher-order interval, denoted by A in Figure 8b. This interval is situated between
the fixed point x ∗ ≡ x0∗ = 0.3 and the first pre-image of the fixed point x1∗ = x1r =
1 + |I0 |/a = 1.166̇.
One can realise that certain other pre-images of I1 —with the fundamental fractal in-
tervals between them—form other higher-order intervals, similar to A. These are denoted
by B and C. The border between B and C is the second pre-image of the fixed point,
x2∗ = 1.744̇. Every subinterval is smaller in B than the corresponding subinterval in A,
by a scaling factor a. On the other hand, the kickout numbers are greater by 1. Similarly,
the scaling factor is a 2 in the case of C, while the kickout numbers are greater by 2.
Note that since there is an infinity of fundamental intervals, B is equivalent to A \
(F ∪ I1 ). This way, B is exactly similar to A, while interval C is “mangled,” since the
corresponding third pre-image of the fixed point, x3∗ ≈ 2.1296, is a bit greater than the
right border of the domain of definition, d = 2.
Now, the left and right borders of the fundamental escape intervals in B and C can be
given easily: Their distance is (b − |I0 |)/a i and b/a i from the corresponding pre-image
of the fixed point, respectively; see (41).
The pre-images of the fixed point x ∗ ≡ x0∗ = S/(a − 1) in x ∈ [1, 2) can be obtained
similarly to the borders of fundamental escape intervals. The pre-images can be expressed
in a recursive way, for example,
S x0∗ + S + b
ax1∗ − b − S = x0∗ = ⇒ x1∗ = . (45)
a−1 a
Life Expectancy of Transient Microchaos 77

The solution of the corresponding recursive equation is

i
S S+b |I0 | 1 − a 1−i
xi∗ = + = 1 + + b . (46)
(a − 1)a i k=1
ak a a(a − 1)

Using (41) and (46), the borders of fundamental escape intervals can be expressed as
 
b − |I0 | ∗ b 1.05 1.3
x0∗ + , x 0 + = 0.3 + , 0.3 + in A,
ai ai 1.5i 1.5i
 
b − |I0 | ∗ b 1.05 1.3
x1∗ + , x 1 + = 1.166̇ + , 1.166̇ + in B, (47)
ai ai 1.5i 1.5i
 
b − |I0 | ∗ b 1.05 1.3
x2∗ + i
, x 2 + i
= 1.744̇ + i
, 1.744̇ + in C.
a a 1.5 1.5i
Note that, at different parameter values, more or less higher-order intervals occur in the
structure than in our specific example. The number N ABC of the nonmangled higher-order
intervals can be obtained by finding the greatest integer number, with which inequality
x N∗ ABC < 2 fulfills

b
N ABC = Int log / log(a) + 1 = 2. (48)
|I0 |(a − 1) − a(a − 1) + b
Thus, in the case of the example, the higher-order intervals A and B are not mangled,
while interval C is mangled, as can be seen in Figure 8b.
Note that if we considered a smaller interval as the domain of definition—e.g., the
repeller—the number of nonmangled intervals would be
   
ba
N ABC = Int log / log(a) = 1, (49)
(a + |I0 | − brep
r a)(a − 1) + b

r
where brep = a − S is the right border of the repeller (34).
The kickout number of the rightmost fundamental escape interval in C can be obtained
as follows. The solution i 0 of
b − |I0 | b − |I0 |
x N∗ ABC + ≡ x2∗ + =2 (50)
ai ai
is a power to which a must be raised in the equation in order to make the left border of the
i 0th escape interval coincide with the right border of the domain of definition. However,
variable i can assume only integer values. If one takes the integer part Int(i 0 ) of i 0 , it
0|
decreases, so x N∗ ABC + b−|I
a {i0 }
will be greater than 2. To correct this problem, the integer
(1)
part of the solution i 0 must be increased by 1 to obtain the kickout number NCr of the
rightmost fundamental escape interval:

(1) (|I0 |−b)a(a −1)

NCr = Int log / log(a) +N ABC +1 = 4.
(a −1)(|I0 |a ABC −a 1+N ABC )+b(a N ABC −a)
N
(51)
78 G. Csernák and G. Stépán

A B C
2

F I1
1:5
00
11
111
000 0011
1100
000
111 11
00
00
110 11
011
100
1
00
11
00
00
11
000
111 0
10
1
000
111 00
11
00
110 11
011
100
1
00
11
00
00
11
000
111 11
00 0
10
1 00
11
+1

1 00
11 00
11
0
100
11
0
1 00
11
00
11
0
100
11
0
1
xi

00
11 00
1100
11
0
1
011
1 00
11
00
11 00
110 11
000
11 00
00
11 00 11
11 00 00
1100
11
0
1
011
1 00
11
00
11
00
11 00
11 111
000 00 11
11 00
00
11 00
11
00
110 11
000
11 00
00
11
0
111
00 00
11 00
11 000
111 00
1100
11 00
11 00
11 0
1
00
1100
11
0
1 00
11
0:5 00
110
100
11 00
11 00
11 000
111 00
1100
11 00
11 00
11 0
1
00
1100
11
0
1 00
11
00
110
100
11 00
11 00
11 000
111 0
1
00
1100
11 00
11
00
11 00
11 0
1
00
1100
11
0
1 00
11
0
10
1
00
11
0
1 0
100
11 00
11 00
11 000
111 00
110
10
1
00
1100
11 00
11 00
11 0
1
00
1100
11
0
1 00
11
0
10
1
00
11
0
1 0
100
11 00
11 00
11 000
111 0
1
00
110
10
1
00
1100
11 00
11 00
11 0
1
00
1100
11
0
1 00
11
0
10
1
00
11
0
1 0
100
11 00
11 00
11 000
111 0
1
00
110
10
1
00
1100
11 00
11 00
11 0
1
00
1100
11
0
1 00
11
0
10
1
00
110
100
11 00
11 000
111 0
10
10
1
00
1100
11 00
11 0
100
11
0
1 00
11
I0 1
00
1
0
1
00
11
0
1 0
15 4 11
00
11 00
00
113 11
00 2 000
111 1 00
11
0
1
00
110
10
1
00
1100
11
5 4 00
11
00
113 00
11 2 11
00
0
1
00
1100
11
0
1 00
11
5 4
011
10
1
0 100
11
0 11 00 11
00 000
111 0
10
10
1 00
11 00 11
00 0
100
11
0
1 00
11
00
1  11
00
11 00 11 00
11
 11
00 
0 x 0 1 1
x x2 2 x 3
xi

Fig. 9. Fractal intervals are mapped onto F, while F is mapped into B.

Note that if we considered the repeller as the domain of definition, the result would be
   
(1) (|I0 | − b)a(a − 1)
N Br = Int log / log(a)
(a − 1)(|I0 |a N ABC + (1 − brep
r )a 1+N ABC ) + b(a N ABC − a)

+N ABC + 1 = 5. (52)

If we search for the images of fundamental fractal intervals, we find that fundamental
fractal intervals in C are mapped onto fundamental fractal intervals in B, which are
mapped onto fundamental fractal intervals in A, and eventually onto F after some steps.
Figure 9 shows how a fundamental fractal interval in C is mapped onto F. On the other
hand, F is mapped into B, and thus these intervals have a similar structure.
According to another line of reasoning, the right borders of fundamental escape in-
tervals are pre-images of the fixed point, and these pre-images are accumulation points
of a sequence of subintervals of decreasing size. Thus, each fractal interval Fi contains
an infinite sequence of fractal and escape intervals. Since the structure of F—and con-
sequently the structure of every fractal interval—is similar to that of A or B, fractal
intervals have the property of self-similarity. This self-similarity provides the basis for
a recursive calculation [13], [14].
In Figure 10, the structure of F ≡ F0 can be seen. Since every fractal interval is
mapped onto F, these have the same structure. During the calculation of the mean
kickout number, this similarity of fractal intervals will be exploited.
However, the rightmost interval in F is a mangled fractal interval. Thus, it is not
exactly similar to F. The solution of this problem is that at a higher resolution this
fractal interval may be divided into smaller escape and fractal intervals.
Life Expectancy of Transient Microchaos 79

NF r
(1)
NF r (2)

11
00
NF r + 1
(1)
00
11
00
11
0000 11
1111 00 1111
0000
0000 11
1111 00
00
11 0000
1111
0000
1111
0000 11
00 0000
1111
001111
110000 11
00 001111
110000
001111
110000
1111 00
11 001111
110000
00
11
00 11
11 0000
1111 00
11 1 001111
0 110000
0000
1111
1011 00
00 11 0000
1111 00
11 0
1
00 1
11 00
110000
1111
0
100
11 00
00
110000
1111 00
11 0
00 1
11 00
11
00
110000
1111
00
11 0
100
11 00
11 00
11 0
00
1100
11 0
100
11 00
11
00
11
0000
1111
0000
1111 00
11
00
11 011
1
00
1100 1
00
11 0 00
11
00
11
0000
1111
0000
1111
00
1100
11 0
100
11 00
110000
1111 00
11 0
1
00
1100
11 0
1 00
110000
1111
00
1100
11 0
100
11 00
110000
1111 00
11 0
1
00
1100
11 0
1 00
110000
1111
00
1100
11 0
100
11 00
110000
1111 00
11 0
1
00
1100
11 0
1 00
110000
1111
00
1100
11 0
100
11 0000
1111 0
1
00
1100
11 0
1
f
00
11
l
0
00
11 0
1 00
11
00 1111
11 0000 00
11
f 1
l 2 2
0
1
00
1100 1111
f l 11
0
1 00
110000
00001
1111
NF r 1
0 0
(2)

NF r 2 NF r 1
(1) (1)

a) b)
Fig. 10. Structure of the greatest fractal interval, F. The kickout numbers and the F-numbers of
some subintervals are shown.

The exploration of smaller subintervals can be made in the following way. First of
all, explore the structure of F. Since F is similar to A or B, the borders of escape
intervals can be determined in a way similar to that used in (47). The left border of F is
f 0l = x0∗ + b/a 2 , so the escape intervals can be given as
 
b b − |I0 | ∗ b b 1.05 1.3
x0∗ + 2+ , x0 + 2 + i = 0.877̇ + , 0.877̇ + . (53)
a ai a a 1.5i 1.5i
(1)
The kickout number N Fr of the rightmost escape interval in F can be obtained by
solving the following equation (see (41)):

b − |I0 | |I0 | − b b b − |I0 |

f 0l + ≡1+ + 2+ = 1. (54)
ai a a ai
(1)
According to the reasoning in connection with (51), the kickout number N Fr of the
rightmost escape interval in F is

(1) b−|I0 | a 2 (b−|I0 |)

N Fr = Int log / log(a) +1 = Int log / log(a) +1 = 6;
1− f 0l a(b−|I0 |)−b
(55)
see Figure 10.
Since both the left and right borders of this rightmost escape interval are less than 1,
we can convince ourselves that this interval is not mangled, but the neighboring fractal
80 G. Csernák and G. Stépán

interval is mangled. Now,

b − |I0 |
f 0l + (1)
≈ 0.970 < 1 and (56)
a N Fr
b
f 0l + (1) ≈ 0.992 < 1. (57)
N Fr
a
Note that, at certain other values of the parameters,

b
f 0l + (1)
>1 (58)
N Fr
a
may be fulfilled. In such cases the rightmost escape interval is mangled in F.
Since the mangled fractal interval in F has a similar structure to F, it can be explored
quite simply. According to (40) and (41), the left border of this interval is

b |I0 | − b b b
f 0l2 = f 0l + (1)
=1+ + 2 + (1) ≈ 0.992; (59)
a N Fr a a a N Fr

see Figure 10.

The right border of this interval is 1, since this is also the right border of F. There is
a similar infinite structure between f 0l2 and 1 like that in A or in F. The distances of the
borders of the subintervals from the left border f 0l2 of the fractal interval in question are
given by (41).
Thus, the kickout number of the rightmost escape interval in this smaller subinterval
(1)
can be determined in a way similar to the determination of N Fr (see (55) and Figure 10):

(2) b − |I0 |
N Fr = Int log / log(a) + 1 = 12. (60)
1 − f 0l2
(2) (2)
It can be checked that f 0l2 + b/a N Fr > 1; and f 0l2 + (b − |I0 |)/a N Fr < 1; thus, this smaller
rightmost escape interval is mangled.
If a mangled fractal interval still occurred at 1, it could be explored, too. Only the left
(2)
border of the new interval would change to f 0l3 = f 0l2 + b/a N Fr . This procedure can be
repeated, while a mangled fractal interval occurs at the right side of F.
Note that, at certain parameters, a mangled fundamental fractal interval may occur at
the right side of interval C. In such cases, the exploration of this interval can be done
with the procedure described above.

3.2. Kickout Number Calculation

The calculation of the mean kickout number is performed as follows: One adds the
lengths of certain subintervals in Iic , weighted by the corresponding kickout numbers.
The result divided by the size of the domain of definition gives the mean kickout number.
In the case of the fundamental escape intervals, this summation can be done quite
easily. The lengths of these intervals are |I0 |/a i , with kickout number i. If there are N ABC
Life Expectancy of Transient Microchaos 81

nonmangled higher-order intervals, the contribution of the fundamental escape intervals

in them is
N ABC ∞
|I0 | a 1−N ABC (N ABC − N ABC a − 1 − a) + a(a + 1)
W1 = i = |I 0 | . (61)
k=1 i=k
ai (a − 1)3

In our example,

2 ∞
|I0 | (1 − 3 a) + a 2 (a + 1)
W1 = i = |I 0 | = 2.833̇ (62)
k=1 i=k
ai a(a − 1)3

is the weighted sum of the lengths of fundamental escape intervals in A and B.

The next step is to determine the similar weighted sum for the mangled higher-order
interval C. Note that the rightmost fundamental escape interval is mangled, so its size is
(1)
2 − (x2∗ + (b − |I0 |)/a NCr ) = 2 − (1.744̇ + 1.05/a 4 ) ≈ 0.048.
Thus, the weighted sum of the lengths of fundamental escape intervals in C is
∞
|I0 | b − |I0 | (1)
W2 = i + 2 − x N∗ ABC + (1)
NCr
(1)
ai a NCr
i=NCr +1

5a − 4
≈ |I0 | + 0.048 · 4 ≈ 0.884. (63)
a 4 (a − 1)2

Note that if a mangled fractal interval was at the right side of C, its escape subintervals
would give an additional contribution to W2 . This additional weighted sum could be
calculated during the exploration of the fine structure of the mangled fractal interval.
The weighted sum of the fundamental escape intervals is W1 + W2 . The calculation
has been quite straightforward so far, but how can we calculate similar weighted sums
for the fundamental fractal intervals?
Since the fractal intervals are self-similar, almost every escape interval in F has a
counterpart in the mangled fractal interval [ f 0l2 , 1). So do the fractal intervals.
Sort the escape and fractal intervals of F. For example, this can be done by choosing
(1)
the greatest escape interval with kickout number N Fr to be the first one, the rightmost
(2)
escape interval with kickout number N Fr to be the second, I N (1) +1 to be the third, I N (2) +1
Fr Fr
to be the fourth, and so on. This means that one choses an escape interval from Figure 10a,
then from Figure 10b, alternately.
Introduce the following notation:
• lei : the length of the ith escape interval in the previously obtained infinite sequence,
divided by the size of F, i.e., it is a relative length. Thus, the relative length of an
interval with kickout number j can be expressed as |I0 |/(a j |F|). For simplicity, lei
will denote the ith escape interval itself, too.
• Nei : the kickout number of the ith escape interval.
• l f i : the relative size of the ith fractal interval in F. According to (44), the relative
length of a fractal interval with F-number j is a − j .
• N f i : the F-number of the ith fractal interval.
82 G. Csernák and G. Stépán

le1

le1 lf 1 le1 lf 2 le1 lf 3 le1 lf 4 ...

... ... ...

...
le1 lf 1 lf 1 le1 lf 1 lf 2
... ...

S1 lf 1

S1
Fig. 11. Tree structure of relative sizes of intervals eventually mapped onto le1 .

Since every fractal interval is mapped onto F, the same happens to the fractal subin-
tervals of F, too. Thus, these intervals are stretched, and some parts of them are mapped
onto the escape subintervals of F. A part of a fractal interval in F may also be mapped
into another fractal interval, but it is eventually mapped onto an escape interval.
The overall size S1 of the intervals that are eventually mapped onto the first escape
interval le1 in F can be calculated as follows. When the first fractal interval is mapped
onto F, a part of it—whose relative size is le1l f 1 —is mapped onto the first escape
interval of relative size le1 . Similarly, a part of the second fractal interval—with relative
size le1l f 2 —is mapped onto the first escape interval. There is a subinterval in l f 1 , which
is first mapped into l f 1 and then mapped onto le1 . The relative size of this interval is
le1l f 1l f 1 . Thus, there is an infinity of intervals that are mapped onto le1 . The relative sizes
of these intervals can be put into a tree structure, as shown in Figure 11. Dots represent
that there is an infinity of branches at every node.
The self-similarity of this tree is quite apparent: the terms in the subtree with root
le1l f 1 can be obtained from the corresponding terms of the original tree by a simple
multiplication by l f 1 . Thus, if the overall sum of the terms in the tree is S1 , the overall
sum of the terms in the subtree is S1l f 1 . Similarly, the overall sum of the terms in the ith
subtree is S1l f i . This means that
∞
S1 = le1 + S1 lf i. (64)
i=1

Thus, according to Figure 10,

le1 |I N (1) |
S1 =
∞ =  
Fr

∞  . (65)
1− i=1 l f i |F| 1 − ∞
(1) a− j + (2) a− j
j=N Fr −1 j=N Fr −1

Similarly, the overall relative size of the subintervals that are eventually mapped onto
the kth escape interval is
lek
Sk =
∞ . (66)
1− i=1 l f i
Life Expectancy of Transient Microchaos 83

Ne1 le1

(Ne1 + Nf 1 ) le1 lf 1 Ne1 + Nf 2 ) le1 lf 2

( (Ne1 + Nf 3 ) le1 lf 3
...

... ...

(Ne1 + Nf 1 + Nf 1 ) le1 lf 1 lf 1 ( Ne1 + Nf 1 + Nf 2 ) le1 lf 1 lf 2 ...

...
...

Fig. 12. Tree-structure W F1 of weighted relative sizes of intervals eventually mapped onto le1 .

For the calculation of the mean kickout number, we need the weighted sum of the
lengths of the eventually escaping intervals. The weighted relative sizes of the intervals
that are mapped onto le1 can be represented in another tree structure, shown in Figure 12.
Here, for example, (Ne1 + N f 1 )le1l f 1 refers to a subinterval in l f 1 , which is mapped
onto le1 in N f 1 steps. Then Ne1 steps are needed for the solutions to escape. Denote the
overall sum of the elements of this tree by W F1 . By subtracting Ne1 -times the elements
of the tree S1 from the corresponding elements of the tree W F1 , we arrive at a more
treatable structure:
0

Nf 1 le1 lf 1 Nf 2 le1 lf 2 Nf 3 le1 lf 3

...

... ...

( Nf 1 + Nf 1 ) le1 lf 1 lf 1 ( Nf 1 + Nf 2 ) le1 lf 1 lf 2 ...

...
...

...

A
( 1+ Nf 1S1 )lf 1 A
( 1+ Nf 2S1 )lf 2 A
( 1+ Nf 3S1 )lf 3

A1
Fig. 13. Modified tree-structure A1 of relative sizes.

The overall sum of the nodes of the new tree A1 can be calculated in a recursive way.
One can realise that the overall sum of the elements of the subtree with root N f 1le1l f 1
is A1l f 1 + N f 1 S1l f 1 = (A1 + N f 1 S1 )l f 1 . A similar relation is fulfilled in the case of the
other subtrees, too. Thus,
∞
A1 = (A1 + N f i S1 )l f i , (67)
i=1

and so

∞
∞
S1 i=1 Nf ilf i le1 i=1 Nf ilf i
A1 =
∞ =
∞
1 − i=1 l f i (1 − i=1 l f i )2

∞
∞
|I N (1) | (1)
j=N Fr −1
ja + j=N (2) −1 ja − j
−j

= Fr
 

Fr
2 . (68)
|F| ∞ −

∞ −j
1− j=N −1 (1) a j + (2)
j=N −1
a
Fr Fr
84 G. Csernák and G. Stépán

Similarly,

∞
lek i=1 Nf ilf i
Ak =
∞ . (69)
(1 − i=1 l f i )2
Since W Fk = Ak + Nek Sk gives the weighted overall sum of the intervals that escape via
lek , the weighted overall sum of all the intervals escaping from F is
∞ ∞ ∞
∞
∞
∞
i=1 l ei
i=1 N f i l f i i=1 Nei lei
WF = W Fk = Ak + Nek Sk =
∞ +
∞ . (70)
k=1 k=1 k=1
(1 − i=1 l f i )2 1 − i=1 lf i

In the example, we obtain (see (43) and (59))

 
∞ ∞
|I0 | |I0 | b − |I0 | 1 
WF =  j |F|
+ j |F|
+ 1 − f 0l2 +
(1)
a (2)
a (2)
a N Fr |F|
j=N Fr j=N Fr +1

∞

(1)
j=N Fr −1
ja − j + ∞ (2)
j=N Fr −1
ja − j
× 
2
∞

1− j=N (1) −1
a− j + ∞ j=N (2) −1
a− j
Fr Fr

∞ |I0 |
∞ |I0 | (2) b−|I0 |

(1)
j=N Fr
j a j |F|
+ (2)
j=N Fr +1
j a j |F|
+ N Fr 1− f 0l2 + N
(2)
1
|F|
a
+ 

Fr

∞
∞
1− (1)
j=N Fr −1
a− j + (2)
j=N Fr −1
a− j
≈ 14.0272. (71)

In the first row of (71), the overall relative length of the escape intervals in F is expressed.
The first term is the overall relative length of the nonmangled escape intervals in F (see
Figure 10a). The second term is the overall relative length of the nonmangled escape
intervals in the mangled fractal interval of F (see Figure 10b). The third term is the
length of the rightmost mangled escape interval in F (Figure 10b).
In the numerator of the second row of (71), the relative lengths of the fractal intervals
of F are added, weighted by the corresponding F-numbers. Since the exploration of the
structure of F terminates when a mangled escape interval occurs at the right end of F,
the fractal intervals are not mangled, and their lengths can be expressed simply as a − j .
The denominator of this term contains terms similar to those of the numerator, without
the weights.
The numerator of the third row is similar to that of the first row, but the terms are
weighted by the kickout numbers here.
Actually, W F is the weighted sum of the relative lengths, and so it gives the mean
kickout number for the interval F.
Note that, in the example, the exploration of F terminated after the second step, i.e.,
only one subinterval of F had to be decomposed. This is why only two infinite sums occur
in each term of (71). If a mangled fractal interval occurred on the right of Figure 10b,
this interval must have been decomposed, too. Thus, the number of infinite sums would
increase, while the lower limit of the second sum in the first row would decrease by one
(2)
to N Fr .
Life Expectancy of Transient Microchaos 85

Now, the mean kickout number W F for F can be calculated at any parameter set. How
can it be used? If a fractal interval is mapped onto F in i steps, its weight is (i + W F ).
Thus,
N ABC −1 ∞
|F|
W3 = (i + W F )
j=0 i= j
ai
a−N ABC ((W F + N ABC )(1−a)−2)+W F (a −1)+2
= ((b−|I0 |)a − b) (72)
(a −1)3
≈ 10.0388

is the contribution of the fractal intervals in the nonmangled higher-order intervals A

and B, and
∞
|F|
W4 = (i + W F )
(1)
ai
i=NCr −1
(1)
b − |I0 | b 2 − a + (NCr + W F )(a − 1)
= a − ≈ 2.067 (73)
a
(1)
NCr
a
(1)
NCr +1 (a − 1)2

is the contribution of the fractal intervals in the rightmost higher-order interval C.

Here we exploited that the F-number of a fractal interval is 1 less than the kickout
number of the escape interval situated to the right of the fractal interval in question. This
is why the first sum in (72) runs from 0 to N ABC − 1 = 1.
Note that the analytical expressions in (63) and (73) are valid only if an escape
interval can be found at the right border of Iic . If not, the mangled fractal interval has to
be explored in interval C, just as the similar interval was explored on the right side of F.
In this case, additional infinite sums must be added to these expressions, corresponding
to the additional escape and fractal subintervals.
Finally, we obtain the mean kickout number as
W1 + W2 + W3 + W4
Nm = ≈ 7.91164. (74)
|Iic |
Note that if |I0 | is sufficiently great, the pre-images of I0 overlap. In such cases,
we obtain a negative value for |F|, according to the formula (43). The solution of this
problem is the rescaling of |I0 |. Since the left borders of the escape intervals can still be
expressed by (38), the new value of |I0 | is
 
|I0 | = x0l − x1l a. (75)

By taking into account that the sizes of I1 in A, I2 in B, etc, do not change, the kickout
number can be calculated in a manner similar to that in the above-introduced example.
(i)
Note that, in this case, the numbers NCr must be determined using the left borders of the
escape intervals (see (51)).
Note that if we restrict ourselves to the repeller, and

|F| = ((b − |I0 |)a − b)/a 2 < 0 ⇐⇒ (b − |I0 |)a − b < 0, (76)
86 G. Csernák and G. Stépán

the mean kickout number depends explicitly on a, only. In such cases, the higher-order
interval A is mangled, since the inequality
|I0 | (b − |I0 |)(a − 1)
x1r = 1 + >1+ = brep
r
(77)
a a
is equivalent to (b −|I0 |)a −b < 0. The overall weighted length of the escaping intervals
can be calculated simply as
∞
b − |I0 | b − |I0 |  
(1)
Wrep = − i + b r
rep − x l
(1)
−1
(N Ar − 1), (78)
(1)
a i−1 ai N Ar
i=N Ar

(1)
where N Ar is the kickout number of the rightmost nonmangled interval in A.
Since the size of the repeller is |Irep | = b − |I0 | (34), the kickout number is
∞
Wrep 1 1 a−1 1 1 (1)
Nm = = − i i+ + − (1) (N Ar − 1), (79)
|Irep | (1)
a i−1 a a a a N Ar −1
i=N Ar

(1)
where we used (38). N Ar depends only implicitly on a, b, and |I0 |.
As can be seen in (77), the right border of the repeller is greater than 1 if b > |I0 |.
(1)
This means that I1 is the mangled interval in such cases, and N Ar = 2. Thus, according
to (79), we obtain
a
Nm = . (80)
a−1
For example, if a = 1.5 and b = 1.3, the overlap of escape intervals occurs at |I0 | >
0.433̇. In this case, the mean kickout number for the repeller assumes the value Nm =
1.5/0.5 = 3.

4. Algorithm

The following algorithm is suggested for the precise calculation of the mean lifetime of
transient chaotic phenomena in the case of the map

ax j − b Int(x j ) − S if S/a < x j < d,
Fm : x j+1 = 0 if −S/a ≤ x j ≤ S/a, and

ax j − b Int(x j ) + S if −d < x j < −S/a,

which is referred to as the microchaos map (24).

• Check whether there are single nontrivial fixed points on both half-planes, i.e.,

a2 − a − b
2a − b − S < 2 ⇐⇒ |I0 | >
a−1
(see (33)). If this condition is not fulfilled, the fine structure of the repeller cannot be
explored using the proposed method.
Life Expectancy of Transient Microchaos 87

• Check whether the sign of the velocity cannot change, or the map takes a value greater
than −x ∗ at x = 1, i.e.,
b−S S
a −b+S >−
b+S a−1
(see (35)). The occurrence of the opposite case implies that an additional fractal
interval appears between F and I0 . The structure of this interval is similar to that of a
subinterval of F. Thus, our method can be extended to cover such cases as well.
• Introduce a variable corresponding to the numerator of (74). In the following, we will
refer to this variable as the numerator.
• Calculate N ABC according to (48).
• Determine |F| (43). If the result is negative, |I0 | must be rescaled (75).
• Calculate the weighted sum of the escape intervals in the nonmangled higher-order
intervals (see (61)) and add to the numerator.
(1)
• Calculate NCr (51). The weighted overall length of the escape intervals in the rightmost
higher-order interval is calculated and added to the numerator
(see (63)).
(1)
• Determine N Fr (55).

Calculate the overall lengths l ei of the escape intervals and
the overall lengths
l f i of the fractal

intervals in F, and the weighted overall lengths
of these intervals: Nei lei , and N f i l f i , respectively. If a mangled fractal interval
can be found at the right border of F, its structure must be explored by calculating the
(i)
numbers N Fr . The variables corresponding to the overall lengths and the weighted
overall lengths are increased at every step.
• Calculate W F (70).
• Calculate the weighted overall length of the nonmangled fractal intervals in the non-
mangled higher-order intervals (see (72)) and add to the numerator. The weight is
W F + i, where i is the corresponding F-number.
• Calculate the weighted overall length of the nonmangled fractal intervals in the man-
gled higher-order interval (see (73)) and add to the numerator.
• If a mangled fractal interval was at the right border of Iic , its structure could be
(i)
explored by calculating the numbers NCr , until a mangled escape interval is found.
The weighted overall lengths of the newly explored escape intervals are added to
the numerator at every step. The weighted overall lengths of the fractal intervals are
calculated. The weight is W F + i. The sum is added to the numerator.
• Calculate the mean kickout number: Nm = numerator/|Iic |.
• Calculate the mean lifetime: tm = τ Nm .

5. Results

By applying the algorithm described above, a C++ computer code was compiled (http:\\
www.mm.bme.hu\∼csernak). This way, we determined the kickout number at sev-
eral parameter values, three ranges faster than by the statistical analysis of simulation
results.
In Figure 14, the results of the analytical calculation can be seen at a = 1.5, b = 1.3,
and d = 2, together with the results of numerical simulations. Note that computer
simulations do not always provide reliable results, due to the quick propagation of the
88 G. Csernák and G. Stépán

45
Analytical method
Simulation
40
Deepness

35

30

25
Nm

20

15

10

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
jI0 j

Fig. 14. Mean kickout numbers at parameters a = 1.5, b = 1.3, d = 2, and Iic = [0, 2).

roundoff errors [15]. However, as the figure shows, the two graphs almost coincide. Thus,
the analytical and the numerical results support each other.
It can be seen in (61) and (63) that these expressions depend linearly on |I0 |. Moreover,
we realized after substituting (71) into (72), and then into (73), that these are also linear
(i) (i)
in |I0 |. Note that the numbers N ABC , N Fr , or NCr also depend implicitly on |I0 |, but
these numbers can change only in a discrete way. Introduce the notion of deepness as the
(i) ( j)
average of the greatest possible index i in N Fr and the greatest possible j in NCr . Using
this notion, we can conclude that every term in the numerator of Nm depends linearly
on |I0 | if the deepness of the method does not change. This linear behaviour can be seen
in the figure in |I0 | ∈ (0.2622, 0.3018) or in |I0 | ∈ (0.4234, 0.9). Thus, the graph of
the function Nm (|I0 |) is piecewise linear. Moreover, since the change of the deepness
corresponds to the appearance of a new mangled fractal interval, this graph has a fractal
structure similar to the “devil’s staircase” [16]. In our case, the function is piecewise
linear instead of being piecewise constant. To illustrate the fractal nature of the graph,
we determined the deepness, too, and it is also presented in the figure.
Using the method described above, we were also able to examine the dependence of
the mean kickout number on the parameter a at |I0 | = 0.17, b = 1.2, and d = 2. The
corresponding results are shown in Figure 15.
By a slight modification of the method, the kickout number can also be calculated in
the case of Iic = Irep . The corresponding results are shown in Figures 16 and 17.
It can be seen in Figure 16 that, if |I0 | > 0.433̇, the mean kickout number remains
constant, according to (80).
Life Expectancy of Transient Microchaos 89

10
Analytical
Simulationmethod
Deepness
9

7
Nm

3
1.1 1.2 1.3 1.4
a 1.5 1.6 1.7 1.8 1.9

Fig. 15. Mean kickout numbers at parameters |I0 | = 0.25, b = 1.3, and d = 2.

45
Analytical method
Simulation
40
Deepness

35

30

25
Nm

20

15

10

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
jI0 j

Fig. 16. Mean kickout numbers at parameters a = 1.5, b = 1.3, and Iic = Irep .
90 G. Csernák and G. Stépán

11
Analytical
Simulationmethod
Deepness
10

8
Nm

3
1.1 1.2 1.3 1.4
a 1.5 1.6 1.7 1.8

Fig. 17. Mean kickout numbers at parameters |I0 | = 0.25, b = 1.3, and Iic = Irep .

6. Conclusions

A very efficient, fast, and also very precise algorithm was constructed to calculate the
mean kickout numbers for transient microchaotic systems. With the help of this algorithm
and the corresponding computer code, we explored the fine and often fractal structure
of the parameter dependence of the mean kickout numbers. Although the exploration of
this fine structure seems to be less relevant in applications, we note that at low kickout
numbers (say, below 10) the 10–20% deviation between estimated and exact values may
lead to the same difference in life expectancy of digitally controlled machines where the
occurrence of microchaos is typical.

Acknowledgments

This research was supported by the Hungarian National Science Foundation under grant
no. OTKA T030762/03, and by the János Bolyai Research Scholarship of the Hungarian
Academy of Sciences.

References

[1] J. A. Yorke, E. D. Yorke, Metastable Chaos: The Transition to Sustained Chaotic Behavior
in the Lorenz Model, J. Stat. Phys., 21 (3), 1979.
[2] T. Tél, Transient Chaos, published in Directions in Chaos, vol. 3, edited by Hao Bai-lin,
World Scientific Publishing Company, Singapore, pp. 149–211, 1990.
Life Expectancy of Transient Microchaos 91

[3] G. Pianigiani, J. A. Yorke, Expanding Maps on Sets which Are Almost Invariant: Decay and
Chaos, Trans. Am. Math. Soc., 252, pp. 351–366, 1979.
[4] A. Lasota, J. A. Yorke, The Law of Exponential Decay for Expanding Mappings, Rend. Sem.
Math. Univ. Padova, 64, pp. 141–157, 1981.
[5] H. van den Bedem, N. Chernov, Expanding Maps of an Interval with Holes, Ergodic Theory
Dynam. Syst., 22 (3), pp. 637–654, 2002.
[6] G. Csernák, G. Stépán, Fractal Dimension as Measure of Control Time, Periodica Polytech-
nica Ser. Mech. Eng., 48 (1), pp. 17–25, 2004.
[7] G. Haller, G. Stépán, Micro-Chaos in Digital Control, J. Nonlinear Sci., 6, pp. 415–448,
1996.
[8] E. Enikov, G. Stépán, Micro-Chaotic Motion of Digitally Controlled Machines, J. Vibration
and Control, 4, pp. 427–443, 1998.
[9] R. Klages, J. R. Dorfman, Simple Maps with Fractal Diffusion Coefficients, Phys. Rev. Lett.
74, pp. 387–390, 1995.
[10] P. Cvitanović et al. Classical and Quantum Chaos, www.nbi.dk/ChaosBook, Niels Bohr
Institute, Copenhagen, 2001.
[11] V. Lakshmikantham, D. Trigiante, Theory of Difference Equations—Numerical Methods and
Applications, Academic Press, San Diego, 1988.
[12] P. Collet , J.-P. Eckmann, Iterated Maps on the Interval as Dynamical Systems, Birkhäuser,
Boston, 1980.
[13] G. Csernák, G. Stépán, Life Expectancy Calculations of Transient Chaotic Behaviour Using
Approximate 1D Maps, Meccanica 35 (6), pp. 547–562, 2000.
[14] G. Csernák, G. Stépán, Escape Rate of Transient Chaotic Phenomena in Digitally Con-
trolled Systems, Proceedings of the Fifth World Congress on Computational Mechan-
ics, online publication, Vienna University of Technology, Austria, ISBN 3-9501554-0-6,
http://wccm.tuwien.ac.at, Vienna, 2002.
[15] E. Ott, Chaos in Dynamical Systems, Cambridge University Press, Cambridge, 1993.
[16] B. Mandelbrot, Fractals: Form, Chance, and Dimension, Freeman, San Francisco, 1977.