Chapter 6

An Investigation of the Chaotic Transient

for a Boundary Crisis in the Fermi-Ulam
Model

Edson D. Leonel and Murilo F. Marques

6.1 Introduction

Dynamical systems are often tuned by control parameters [1, 2]. The variation
of them may lead to interesting properties. In conservative systems transitions
such as integrability to nonintegrability [3] are observed as the parameters are
varied, transport of particles can be investigated [4] as invariant tori are broken
due to a parameter change and many other different consequences emerge [5, 6].
In dissipative systems, no matter if described by ordinary differential equations [7]
or by mappings [8, 9], bifurcations are often observed. They can be classified into
two different types [10]: (1) local and (2) global bifurcations. The local bifurcations
[11] are identified and characterized only by the investigation of local dynamics,
mostly by the analysis of fixed points and their stability. The set of nonlinear
equations describing the system is linearized near the fixed point allowing to only
local inferences. The local knowledge of a given phenomena cannot be extended to
all the phase space due to the linear stability analysis. Birth of attractors such as fixed
points may lead to pitchfork bifurcations,1 saddle node bifurcation with one fixed
point stable and other unstable, transcritical bifurcation where there is a coexistence
of fixed points but at the bifurcation they exchange stability, period doubling
bifurcations, tangent bifurcations, Hopf bifurcations among others. In a global
bifurcation [11], the knowledge of the stability of a fixed point is not enough to

1 Depending on the type of the fixed points born at the bifurcations as well as their stability, the

pitchfork bifurcation can be classified as supercritical or subcritical.

E. D. Leonel () · M. F. Marques
Departamento de Física, UNESP - Univ Estadual Paulista, Rio Claro, SP, Brazil
e-mail: edleonel@rc.unesp.br

© Springer International Publishing AG, part of Springer Nature 2019 89

E. E. N. Macau (ed.), A Mathematical Modeling Approach from Nonlinear
Dynamics to Complex Systems, Nonlinear Systems and Complexity 22,
https://doi.org/10.1007/978-3-319-78512-7_6

edleonel@rc.unesp.br
90 E. D. Leonel and M. F. Marques

foresee a bifurcation. Generally the change in the phase space due to the occurrence
of a global bifurcation is catastrophic. It leads then to a major modification in the
form of the phase space. Such types of bifurcations are commonly called in the
literature as crisis [12, 13]. Three well-known crises [14] in the scientific community
are: (i) interior crisis; (ii) merging attractor crisis; (iii) boundary crisis. In case (i)
one can observe in the system a sudden increase in the size of the chaotic attractor
due to a collision of a periodic orbit with the chaotic attractor. For the case (ii) two or
more attractors collide simultaneously among each other and with a periodic orbit;
as a result, the attractors become merged. Finally in case (iii) there is a crossing
of an unstable with a stable manifold of a saddle fixed point leading to a sudden
destruction of the chaotic attractor. If the manifolds are produced by the same saddle
fixed point there is homoclinic crossing when a heteroclinic crossing is produced by
intersections of manifolds generated from different saddle fixed points.
In this chapter we address a specific system to describe some properties for the
chaotic transient observed in a boundary crisis. The model under consideration
is a dissipative version of the Fermi-Ulam model [15]. The system consists of a
classical particle which is confined to move between two rigid and infinitely heavy
walls. One of them is fixed while the other moves periodically in time. Collisions
of the particle with moving wall may lead to velocity changes in the particle. For
elastic collisions [16], this was a frustrated model idealized by Fermi as an attempt
to describe the acceleration of cosmic particles reaching the earth. The cosmic
particles were supposed to be affected by the moving magnetic fields from the extra-
galactic bodies therefore furnishing energy to the moving particle. The oscillating
wall was introduced with this objective, to mimic the action of the magnetic fields.
The fixed wall is an effective re-injection mechanism directing the moving particle
for a further collision with the periodically varying wall. It is known that if the
motion of the moving wall is not smooth enough, then the particle may experience
no bound for the velocity. On the other hand, for a sinusoidally motion of the
wall, the phase space of the model is composed by coexistence of periodic islands,
invariant spanning curves, and chaotic seas. The invariant spanning curves work
as barriers [17] not letting the particle to cross therefore confining the energy gain
for a bouncing particle. The existence of them as a superior limit for energy of a
chaotic orbit warrants local chaos.2 Scaling properties give an overall description
of the chaotic sea [17] proving it to be scaling invariant with respect to the control
parameter, namely, the amplitude of the oscillation wall [18]. The determinant of
the Jacobian matrix leads to measure preservation hence attractors are not observed.
The scenario changes significantly when a fractional loss of energy upon collision
is considered [19]. This is made by assuming the particle experiences an inelastic

2 By local chaos we want to say the chaos observed is confined by invariant spanning curves. In a

situation of global chaos the particle can diffuse unlimitedly in the velocity axis therefore leading
to a phenomena called as Fermi acceleration. A prototype system exhibiting such phenomena is
an analogous 1-D model called as a bouncer. The only difference from the Fermi-Ulam model is
the returning mechanism for a further collision. In the Fermi-Ulam it is provided by a fixed wall
whereas in the bouncer it is due to the gravitational field only.

edleonel@rc.unesp.br
6 Chaotic Transient in Boundary Crisis in a Fermi-Ulam Model 91

collision each time it interacts with the walls. The determinant of the Jacobian
matrix for the dissipative case is less than the unit. According to the Liouville’s
theorem [20], the dissipation leads to contraction of the phase space, hence yielding
the system to exhibit attractors that may be periodic [21] or chaotic [22]. Our main
goal in this chapter is then to investigate some properties for the chaotic transient
for a boundary crisis observed due to the crossing of a stable and unstable manifold
obtained from the same saddle fixed point.
We chose a specific combination of control parameters that lead the system to
exhibit two types of attractors before the crisis: (i) an attracting fixed point3 and
(ii) a chaotic attractor. The chaotic attractor was confirmed by means of a positive
Lyapunov exponent [23]. As soon as the control parameter is varied and the crossing
is observed, the chaotic attractor is destroyed. It is then replaced by a chaotic
transient [12]. Given an initial condition inside of the region of the phase space
where the attractor likely existed prior to the crisis, the dynamics of a particle will
run such region for a while until the particle finds the appropriate scape path and
heads towards the attracting fixed point being captured by it and reaching the steady
state at a sufficiently long time. Once the convergence to the steady state started
there is no way to escape from there without an external action such as application
of stochastic perturbation, a type of control with noise or other intervention on the
system. We want to investigate the behavior of the survival probability a particle
has to survive inside a region of the phase space where a likely chaotic attractor
existed prior the crisis. As we shall see, the survival probability will be described
by an exponential function whose relaxation time is given by a power law [22]
measured in a parameter given the distance of the boundary crisis. Our interest is
in the investigation of such a relaxation time, particularly the power law describing
it. We will do it using two different procedures. The first one is by using extensive
numerical simulations whereas the second considers the eigenvalues of the saddle
point at the parameters leading to the boundary crisis. Investigation of transient in a
gearbox model was also made in Refs. [24, 25].
The organization of this chapter is as follows. Section 6.2 describes the model
constructing the equations that describe the dynamics of the system. It gives also a
short presentation of the properties observed for the conservative case showing the
phase space, organization of the stability islands, illustration of the first invariant
spanning curve as well as the chaotic sea. In the presence of dissipation, the chaotic
sea is replaced, depending on the combination of control parameters and initial
conditions, by a chaotic attractor, which is also illustrated. In Sect. 6.3 we discuss
the boundary crisis observed in a dissipative Fermi-Ulam model. The stable and
unstable manifolds for a saddle point are plotted before the crisis. The inverse of
the mapping is obtained and is used to construct the stable manifolds. The basin of
attraction of either the chaotic attractor and attracting fixed point are plotted together
with the chaotic attractor. Remarkably, the stable manifolds give the border of the
two basin of attraction. A discussion of the relaxation time given the behavior of

3 Indeed it is an asymptotically stable focus.

edleonel@rc.unesp.br
92 E. D. Leonel and M. F. Marques

the survival probability for a particle to survive in the region of the phase space
where a chaotic attractor likely existed is given. Numerical simulations determine
the relaxation time and a connection with the results well settled [12] in the literature
are made. Conclusions and final remarks are presented in Sect. 6.4 while a short
Appendix at the end of the chapter gives the expressions for the eigenvalues of the
saddle point.

6.2 The Model and the Map

We describe in this section the model we consider in the chapter. Our numerical
simulations illustrating the nondissipative as well as the dissipative case will be
made here too.
The system we consider is composed of a classical particle which is confined to
move inside of two rigid and infinitely heavy walls. By choice, one of them is fixed
at x = l and the other one is periodically moving at x(t) = ε̃ cos(ωt), where ε̃ is
the amplitude of oscillation, ω is the frequency of oscillation, l is the distance of
the fixed wall from the equilibrium position of the moving wall, and t is the time.
Figure 6.1 shows a schematic description of the model.
Because there are no forces acting in the particle along the flight it moves with
a constant velocity between the collisions. The only modification in the velocity
happens at the instant of the collision. We assume the particle experiences inelastic
collisions with the walls. The restitution coefficient for collisions with the fixed
wall is denoted as α ∈ [0, 1]. The limit of α = 0 terminates the dynamics since a
single collision leads the particle to reach the state of rest therefore staying glued in
the static wall. Collisions with the moving wall are characterized with a restitution
coefficient β ∈ [0, 1]. The limit of β = 0 leads the particle to be re-injected in the
dynamics with the maximum wall’s velocity. This is a typical case of a dynamical

x(t) = ε~cos(ωt)

vn

l
Fig. 6.1 Plot of a schematic description of the model. The motion of the oscillating wall is
described by x(t) = ε̃ cos(ωt). The fixed wall is placed at x = l measured from the equilibrium
position of the moving wall. The restitution coefficients for collisions with fixed and moving walls
are, respectively, α ∈ [0, 1] and β ∈ [0, 1]

edleonel@rc.unesp.br
6 Chaotic Transient in Boundary Crisis in a Fermi-Ulam Model 93

regime called as locking [26]. The case of α = β = 1 gives the nondissipative

dynamics [27] whereas the phase space shows features typical of nonintegrable
Hamiltonian systems.
The traditional way of describing the dynamics of the particle is the use of a
mapping therefore updating the velocity of the particle after an impact with the
moving wall. We see there is a large number of control parameters and that not
all of them are relevant for the complete characterization of the model. We then
define a set of dimensionless and hence more convenient variables as V = v/(ωl),
ε = ε̃/ l and measure the time in terms of the number of oscillations that the moving
wall completes in between the collisions hence measuring phases φ = ωt. The
mapping is an application T (Vn , φn ) = (Vn+1 , φn+1 ) where V and φ denote both
the dimensionless velocity of the particle and phase of the moving wall,

Vn+1 = Vn∗ − (1 + β)ε sin(φn+1 )


T : , (6.1)
φn+1 = [φn + ∆Tn ] mod 2π

where the expressions for Vn∗ and ∆Tn are determined by the kind of collision hap-
pened: (i) direct collisions4 or (ii) indirect collisions.5 For case (i), the expressions
are Vn∗ = −βVn and ∆Tn = φc where φc ∈ (0, 2π] is the smallest solution of
the equation G(φc ) = 0 with G(φc ) = ε cos(φn + φc ) − ε cos(φn ) − Vn φc . The
numerical procedure used to solve G(φc ) = 0 was the bisection method with an
accuracy of 10−14 . The determinant of the Jacobian matrix for this case is given by
 
2 Vn + ε sin(φn )
det(J ) = β . (6.2)
Vn+1 + ε sin(φn+1 )

For direct collisions the mapping is conservative only for the case of β = 1 where
it preserves the following measure dµ = [V + ε sin(φ)]dV dφ.
For indirect collisions the expressions assume the following forms: Vn∗ = βαVn
and ∆Tn = φr + φl + φc where the auxiliary terms are φr = (1 − ε cos(φn ))/Vn ,
corresponding the time the particle spends traveling to the right-hand side until
collides with the fixed wall and φl = (1 − ε)/(αVn ) is the time the particle spends
moving backwards until reaches the collision zone at x = ε while φc ∈ [0, 2π]
is obtained from the smallest solution of F (φc ) = 0 where F (φc ) = ε cos(φn +
φr + φl + φc ) − ε + αVn φc . Again here the same numerical method was applied
for the solution of F (φc ) = 0 with the same accuracy as before. For this case the
determinant of the Jacobian matrix is written as

4 Direct collisions are defined as the collisions the particle has with the moving wall without leaving

the collision zone, a region defined by x ∈ [−ε, +ε].

5 In an indirect collision a particle leaves the collision zone, collides with the fixed wall, and is

rebounded back for a further collision with the moving wall.

edleonel@rc.unesp.br
94 E. D. Leonel and M. F. Marques

 
2 2 Vn + ε sin(φn )
det(J ) = α β . (6.3)
Vn+1 + ε sin(φn+1 )

The measure dµ is preserved only when β = α = 1.

The term −(1 + β)ε sin(φn+1 ) appearing in first equation of mapping (6.1) was
obtained performing calculations in the referential frame of the moving wall. It is
assumed to be instantaneously at rest at the moment of the impact. In such case the
law of conservation of momentum is applied.
In what follows, we discuss two distinct regimes in separate sections. The first is
the nondissipative case where the second one is the dissipative case leading to the
boundary crisis for specific combinations of control parameters.

6.2.1 Nondissipative Case

When the two restitution coefficients assume α = β = 1, the collisions of the

particle with either walls are elastic and no fractional loss of energy is observed.
The phase space is mixed with the existence of stability islands surrounded by a
chaotic sea which is limited by a set of invariant spanning curves [6, 17]. These
curves, in particular the first one, work as barriers not letting a flux of particle
through them. Figure 6.2 illustrates well a typical phase space of the model for the
nondissipative case. The parameter used is ε = 10−2 . The first invariant spanning
curve is highlighted in the figure and is identified as fisc. It works as a barrier not
letting any flux of particle through it.

e = 10-2
0.5

0.4
Fixed
V 0.3 Points

fisc
0.2
Fixed
Points
0.1

0 Chaotic sea
0 1 2 3 4 5 6
f
Fig. 6.2 Plot of the phase space for mapping (6.1) considering ε = 10−2 . The invariant spanning
curve, represented by fisc, is shown in red (gray) color. Periodic islands are present as well as a
chaotic sea

edleonel@rc.unesp.br
6 Chaotic Transient in Boundary Crisis in a Fermi-Ulam Model 95

The periodic fixed points classified as elliptic are identified in the figure centering
the periodic islands. For the parameter ε = 10−2 there are only three elliptic fixed
points, namely Vi∗ = 1/(iπ) for i = 1, 2, 3 and φ = π. For the regime of low
velocity there is a region dominated by a chaotic sea. Some relevant information
[17] of the chaotic sea are (1) Its amplitude depends on the position of the first
invariant spanning√ curve. A first order approximation furnishes the position of the
curve as Vf isc ∼
= 2 ε/0.9716, therefore scaling with ε1/2 . Starting with a low initial
velocity, typically V0 < ε, the average velocity
√ grows to start with, its diffusion is
given by a slope 1/2 in n, i.e. V (n) ∼= ε2 n. The changeover from the growth to
a regime of constant velocity is marked by nx ∝ 1ε . A scaling law [17] describing
the invariance of the chaotic sea with respect to either the parameter ε or n is given
by z = α̃ − 2, where α̃ gives the saturation of the velocity, β̃ marks the regime of
β̃
growth and z gives the exponent of crossover. For the Fermi-Ulam model [17] the
critical exponents are β̃ = 1/2, α̃ = 1/2, and z = −1. The Lyapunov exponent [23]
was measured for the chaotic sea. Indeed the Lyapunov exponents are defined as

1
λj = lim ln | j |, j = 1, 2, (6.4)
n→∞ n

where j are the eigenvalues of the matrix M = !ni=1 Ji (V , φ) where Ji is the

Jacobian matrix of the mapping (6.1) evaluated along the orbit. Figure 6.3 shows a

2
φ0 = 1
ε = 10−2
φ0 = 2
1.5 φ0 = 3 V0 = 2.1 ε
φ0 = 4
φ0 = 5
l 1

0.5

λ = 0.766(2)

0
104 105 106 107
n
Fig. 6.3 Plot of five different Lyapunov exponent series obtained for the parameter ε = 10−2 ,
initial velocity V0 = 2.1ε and different initial phases as φ01 = 1, φ02 = 2, φ03 = 3, φ04 = 4,
φ05 = 5. An average of the asymptotic behavior gives us λ = 0.766(2). The positive finding for
the Lyapunov exponent confirms chaos in the phase space. Horizontal axis is plotted logarithmic
only for visual purposes. Due to the fluctuations the initial n = 104 were omitted from the plot

edleonel@rc.unesp.br
96 E. D. Leonel and M. F. Marques

plot of five different series generated by different initial conditions along the chaotic
sea for the parameter ε = 10−2 . The initial velocity was fixed as V0 = 2.1ε while
the initial phases chosen were φ01 = 1, φ02 = 2, φ03 = 3, φ04 = 4, φ05 = 5. One
can see from the Lyapunov exponent time series there is a relative large fluctuation at
short n when the series converge for large n. The initial 104 points of the series were
not plotted in the figure since the fluctuations obtained for short time are large for
the scale of the figure. Horizontal axis was plotted logarithmic for visual purposes
only. From the five series we obtained an average positive Lyapunov exponent as
λ = 0.766(2), therefore confirming the existence of chaos in the phase space. We
must stress here the chaos observed is nondissipative since the phase space preserves
a measure dµ. In sequence we discuss the consequences of the dissipation in the
collisions.

6.2.2 Dissipative Case

When the collisions are considered inelastic, the structure of the phase space
changes significantly [19]. The mixed properties are not observed anymore.
Depending on the combination of control parameters as well as initial conditions,
attractors—either periodic or chaotic—are observed in the phase space. Figure 6.4
shows a plot of a chaotic attractor and an attracting fixed point for the parameters
ε = 0.04, β = 1, and6 α = 0.93624. The chaotic attractor is marked by the dark
region on the bottom part of the plot and is limited from below by the velocity of the
moving wall, plotted as a continuous line (red). Attracting fixed point is highlighted
as a bullet when saddle point is plotted as a star. The Lyapunov exponent measured
for the chaotic attractor was λ = 1.7743(5) as reported in Ref. [19].
In the dissipative case, each attractor has now its own basin of attraction [10].
The basin corresponds to a set of points in the phase space whose dynamics heads
towards the attractor for sufficiently long time. It is known that a boundary crisis
may be investigated by analyzing the behavior of the basin of attraction of the
chaotic attractor. Indeed a boundary crisis is produced [14] by a collision of the
border of the basin of attraction of the chaotic attractor with the chaotic attractor
itself. Such a collision implies in a sudden destruction of the chaotic attractor and
also of its basin of attraction. The collision creates a homoclinic orbit produced by
successive crossings of the stable and unstable manifolds generated by the same
saddle fixed point and that the basin of attraction of the attracting fixed point
becomes porous and suitable to be visited by the dynamics of the chaotic attractor.
However once such visit happens the dynamics is then completely dominated by the
convergence to the steady state. The chaotic attractor is destroyed and is therefore
replaced by a chaotic transient [10].

6 This set of control parameters was chosen just before the crisis.

edleonel@rc.unesp.br
6 Chaotic Transient in Boundary Crisis in a Fermi-Ulam Model 97

0.4
e = 0.04 Period 1 fixed point
a = 0.93624
Sink Saddle
0.3 b=1
Chaotic attractor
0.2
V

0.1

0
Chaotic attractor lower limit

0 1 2 3 4 5 6
f
Fig. 6.4 Plot of a chaotic attractor and attracting fixed point for the control parameters ε = 0.04,
β = 1 and α = 0.93624. The lower limit of the chaotic attractor corresponds to the velocity of
the moving wall (red curve). Sink and saddle are identified in the figure. The positive Lyapunov
exponent measured for the chaotic attractor was λ = 1.7743(5), as obtained in Ref. [19]

One way to investigate the boundary crisis is by constructing the stable and
corresponding unstable manifolds. A stable manifold consists of points along the
eigenvector of the stable direction of the saddle point associated with the eigenvalue
smaller than the unity whose dynamics heads directly towards the saddle point. The
usual way to obtain the stable manifolds is to iterate the mapping inversely, i.e.,
instead of moving forward from the dynamics, the procedure moves backwards. For
this end a specific set of initial conditions must be placed along the eigenvectors and
the inverse mapping is updated. On the other hand, the unstable manifold consists of
points along the eigenvector of the associated eigenvalue larger than the unity whose
dynamics moves apart from the saddle point. The knowledge of the eigenvectors and
initial conditions along them allow a direct iteration of the mapping to produce such
curves.

6.3 Boundary Crisis Characterization

In this section we investigate the boundary crisis properly. To start with we obtain
the fixed points that include also the saddle point from where the manifolds
responsible for generating a homoclinic orbit are emanating. The fixed points [28]
are obtained from the condition Vn+1 = Vn = V ∗ and φn+1 = φn = φ+2mπ where
m = 1, 2, 3, . . . is any integer number corresponding to the number of oscillations

edleonel@rc.unesp.br
98 E. D. Leonel and M. F. Marques

that the moving wall completed between collisions. Matching these two conditions
in the mapping (6.1) we obtain the coordinates of the fixed points are
 
1+β
V = ε sin(φ), (6.5)
βα − 1
  √ 
ε+γ̃ ε 2 +γ̃ 2 −1
arccos ,


ε 2 +γ̃ 2


φ± =  √  ,
 ε−γ̃ ε 2 +γ̃ 2 −1
 2π − arccos


ε 2 +γ̃ 2

where the auxiliary term γ̃ assumes the following expression

 
2εαmπ 1+β
γ̃ = , (6.6)
α+1 βα − 1

for any m = 1, 2, 3, . . .. The classification of the fixed points are then Psink and
Psaddle with
   
ε + γ̃ ε2 + γ̃ 2 − 1

1+β
Psink = ε sin(φ), φ = arccos (6.7)
βα − 1 ε2 + γ̃ 2
   
ε−γ̃ ε2 +γ̃ 2 −1

1+β
Psaddle = ε sin(φ), φ=2π−arccos . (6.8)
βα−1 ε2 +γ̃ 2

The construction of the unstable manifolds is simple. They are generated by the
evolution of the dynamics produced by the mapping (6.1) with an appropriate set
of initial conditions. Indeed the initial conditions must lie along the eigenvectors
of the corresponding unstable eigenvalue. There are two branches of the unstable
manifolds of the saddle point, as we see from Fig. 6.5. One of them, in black, leaves
the saddle point and converges towards the attracting fixed point spiraling around it
counterclockwise reducing gradually its amplitude of motion as the time goes one.
For long enough time such branch converges to the steady state and gets trapped
there forever. The second branch of the unstable manifold, in red color, moves
away from saddle point in the downwards direction converging then to the chaotic
attractor. The construction of the stable manifolds requests an explicit form for the
inverse of mapping (6.1). Therefore instead of iterate T (Vn , φn ) = (Vn+1 , φn+1 ) we
must obtain the inverse of T such that given a set of pair (V , φ) at instant n + 1 we
obtain the new pair at instant n. The expression for T −1 is given by

 Vn = Vn+1 +(1+β)ε sin(φn+1 )
−1 α
T : , (6.9)
 h(φ ) = φ + 1−ε cos(φn ) 1−ε cos(φn+1 )
n n Vn + αVn − φn+1

edleonel@rc.unesp.br
6 Chaotic Transient in Boundary Crisis in a Fermi-Ulam Model 99

0.6

S
V 0.3

0
0 1 2 3 4 5 6
f
Fig. 6.5 Plot of the manifolds for the saddle point for the control parameters ε = 0.04, β = 1,
and α = 0.93624. Black denotes one branch of the unstable manifold leaving the saddle point
and converging towards the attracting fixed point. Red represents the other branch of the unstable
manifold converging to the chaotic attractor itself. The stable branches, plotted as green and blue
draw the corresponding border of influence for the basin of attraction of the attracting fixed point
and the chaotic attractor

where φn is obtained by the numerical solution of h(φn ) = 0. The numerical

procedure used to solve h(φn ) = 0 was the well-known Newton’s method with
an accuracy of 10−14. Because the stable manifolds always grow in velocity and
since the velocity never gets smaller than ε, the condition of direct collisions7 does
not need to be considered in the stable manifold construction therefore leading us to
take into account only the case of indirect collisions.
Let us now discuss the properties observed for the basin of attraction of the
chaotic attractor at a boundary crisis. As soon as the control parameter α is varied,
say increased, the distance between the attractor and its basin of attraction decreases
until a limit they touch each other at a parameter αc , leading to the occurrence of
a crisis. The encounter of the chaotic attractor with its basin of attraction happens
at the same instant that an unstable orbit touches the chaotic attractor producing a
homoclinic tangency. After the event the chaotic attractor no longer exists anymore
but is rather replaced by a chaotic transient. If an initial condition is given in the
region of the phase space where the chaotic attractor likely existed prior to the
crisis, the dynamics will follow it for a while, as a type of ghost version of the
chaotic attractor, until it finds the appropriate route of escape and makes an incursion
towards the direction of the fixed point therefore getting trapped there for further
time. The region where the chaotic attractor existed before the crisis is now part of

7 Such condition is given by solution of G(φc ) in mapping (6.1).

edleonel@rc.unesp.br
100 E. D. Leonel and M. F. Marques

Fig. 6.6 Plot of the basin of

attraction of the chaotic 0.6
attractor (black) and the sink
(gray). The boundary of the
two basin of attraction is
delimited by the stable
manifolds, as painted in 0.4
yellow. Attracting fixed point
is plotted as a bullet while V
saddle is plotted as a star.
Chaotic attractor is shown in 0.2
red while the branch of the
unstable manifold converging
to the fixed point is shown in
green. The control parameters
used were β = 1, 0
α = 0.93624, and ε = 0.04
0 2 4 6
f

the basin of attraction of the attracting fixed point. Figure 6.6 shows a plot of the
basin of attraction of the chaotic attractor as well as the basin of attraction of the
attracting fixed point. Black identifies the basin of attraction of the chaotic attractor
whereas gray gives the basin of attraction for the attracting fixed point. The chaotic
attractor is painted as red in the figure. The attracting fixed point (bullet) and saddle
(star) points are also shown in the figure. The control parameters used for construct
Fig. 6.6 were ε = 0.04, β = 1, and α = 0.93624, immediately before the boundary
crisis. We notice the border separating the two basin of attractions are constructed
by the stable manifolds, as shown in yellow. Figure 6.7 shows a plot of V vs. φ
for the parameters ε = 0.04, β = 1, and α = 0.9375, immediately after the crisis,
exhibiting a chaotic like dynamics in black for short time. We see the particle moves
along the previous region where the chaotic attractor likely existed until escape the
region and heads towards the attracting fixed point spiraling around it. There is no
return once the dynamics dives into the attracting fixed point. It spirals around the
fixed point reducing gradually the amplitude of oscillation until eventually reaches
the fixed point at a sufficiently long time.
The crossing of the unstable and stable manifolds is a clear evidence that the
boundary crisis has happened. Figure 6.8 shows the stable and unstable manifolds
for the saddle point constructed after the boundary crisis. The parameters used to
construct the figure were ε = 0.04, β = 1, and α = 0.9375.
The time measured in the number of collisions a particle stays wondering in
a pseudo-chaotic motion after the crisis depends on the initial condition. The
probability an initial condition is given in the chaotic region prior to the crisis
to survive such region without escaping it and being trapped permanently by the
attracting fixed point is described by an expression of the type
n
P (n) ∼
= e− τ , (6.10)

edleonel@rc.unesp.br
6 Chaotic Transient in Boundary Crisis in a Fermi-Ulam Model 101

Chaotic attractor
Attracting sink
0.4

0.3

V
0.2

0.1

0 1 2 3 4 5 6
f
Fig. 6.7 Plot of V vs. φ for the parameters β = 1, ε = 0.04, and α = 0.9375, immediately
after the boundary crisis. Black points indicate the chaotic like dynamics (transiently) while the
red small bullets show the convergence to the attracting fixed point. The dotted line is used only as
a guide to the eye

0.6

0.4

0.2

0
0 1 2 3 4 5 6
f
Fig. 6.8 Plot of the manifolds for the saddle point for the control parameters ε = 0.04, β = 1,
and α = 0.9375. Black denotes one branch of the unstable manifold leaving the saddle point and
converging to the attracting point. Red represents the other branch of the unstable manifold passing
near the region of the phase space where the chaotic attractor likely existed before the crisis. The
stable branches, plotted as green and blue draw are also evident. The box highlights the many
crossings of the manifolds confirming the boundary crisis

edleonel@rc.unesp.br
102 E. D. Leonel and M. F. Marques

Numerical data
Power law fit
105

t
104
d = - 2.01(2)

103
10-3 10-2
m
Fig. 6.9 Plot of τ vs. µ. A power law fitting furnishes δ = −2.01(2). We considered an ensemble
of 5×103 different initial conditions along the region in the phase space where the chaotic attractor
existed before the crisis. The parameters considered were β = 1 and ε = 0.04. The parameter α is
varied as the horizontal axis with αc = 0.93624. Dotted line is used only as a guide to the eye

where n is the number of collisions of the particle with the moving wall and τ is a
relaxation time given by

τ ∝ µδ , (6.11)

with µ = α − αc , for α > αc where αc denotes the critical control parameter where
the boundary crisis happens and δ is an exponent that must be obtained.
The exponent δ can be obtained from different ways. A first description is
by using numerical simulations. Since that the time measured in the number of
collisions a particle leaves to escape the region where a chaotic attractor existed
prior to the crisis depends on the initial condition, it turns out to be convenient to
run an average over an ensemble of different initial conditions along the previous
chaotic attractor. Figure 6.9 shows a plot of the average relaxation time τ against µ.
A power law fitting to the data gives δ = −2.01(2). The parameters considered
were ε = 0.04, β = 1, and αc = 0.93624. The resulting exponent obtained in
Fig. 6.9 was also obtained in other crisis for different parameters in the model. For
ε = 0.033, αc = 0.96375 for m = 1 we found δ = −2.11(9) while for m = 2,
αc = 0.90232 and ε = 0.02 we obtained δ = −1.98(3). The numerical fittings give
the two exponents close to −2 in well agreement with the results obtained in [22].
A second procedure we consider involves a description proposed by Grebogi et
al. in Ref. [12]. It considers the eigenvalues of the saddle point at the parameter
where the boundary crisis happens. In their investigation they also assumed the

edleonel@rc.unesp.br
6 Chaotic Transient in Boundary Crisis in a Fermi-Ulam Model 103

survival probability is described by an exponential function P (t) ∝ e−t /τ with t

denoting the time and where the average τ has an expression of the type τ ∝ µ−γ .
In Grebogi’s formalism (see Refs. [12, 13]), the exponent γ is obtained from the
equation

ln |β2 |
γ = , (6.12)
ln |β1 β2 |2

where β1 and β2 are the corresponding expanding and contracting eigenvalues

evaluated at the saddle point. Using their notation and considering the eigenvalues
at the saddle point, we obtain an exponent γ = 0.4991 . . ., which is significantly
close to 1/2, as discussed in their paper [12]. This result however is not matching
with our numerical simulations as obtained in Figs. 6.9 and 6.10. We now discuss
this apparent discrepancy among the results.
When comparing the two equations describing the relaxation time, those heading
the exponential, we see an ingenious difference on a minus signal that is heading the
parameter µ. Because of such signal and considering Grebogi’s approach [12], the
expression for δ used in our simulations must be written as

ln |β2 |
−δ = ,
ln |β1 β2 |2
ln |β2 |
δ=− ,
ln |β1 β2 |2
ln |β1 β2 |2
= . (6.13)
ln |β2 |

This result leads us therefore to a correct estimation of the exponent with δ =

2.003 . . . obtained from the corresponding expanding and contracting eigenvalues
of the saddle point. This result is in well agreement with the result obtained from
numerical simulations and with the existing literature [22] and explains the apparent
discrepancy of the results which indeed does not exist. In the Appendix of this
chapter we present the relevant expressions obtained from the Jacobian matrix that
lead to the eigenvalues of the saddle point.
As we have discussed along the chapter, the only dissipative parameter varied
was α while the restitution coefficient β remained constant all the discussion. One
may think what happens if the parameter β is allowed to change to. Does the
boundary crisis move? Is there any significant change in the dynamics, i.e. in the
phenomena itself? Indeed when dissipation is considered in both the walls, there is
a combination between then to produce a type of effective restitution coefficient.
As discussed in Ref. [22], the combination of the two control parameters α and β
may lead to a family of boundary crisis to be observed in the parameter space. It
was shown in Ref. [22] that when the parameters are varied in such a way obeying
the law β = A + αB where the authors of Ref. [22] found A = 1.927(1) and
B = −0.989(1) for the parameter ε = 0.04. They also considered other crisis

edleonel@rc.unesp.br
104 E. D. Leonel and M. F. Marques

(a)

Numerical data
Power law fit
105

t
104
d = - 2.11(9)

103
10-3 10-2
m
(b)

Numerical data
Power law fit
105

t
104
d = - 1.98(3)

103
10-3 10-2
m
Fig. 6.10 Plot of average transient τ vs. µ for two other boundary crisis observed in the Fermi-
Ulam model. The parameters used were β = 1 and: (a) ε = 0.033 and αc = 0.96375 considering
m = 1 and (b) for m = 2, αc = 0.90232 and ε = 0.02. The numerical fittings furnish the exponent:
(a) δ = −2.11(9) and (b) δ = −1.98(3). Each exponent was obtained considering an ensemble
of 5 × 104 different initial conditions from a region in the phase space where the chaotic attractor
likely existed before the crisis. The two exponents obtained were close to −2 in well agreement
with the results reported in Ref. [22]. Dotted lines are used only as guide to the eye

edleonel@rc.unesp.br
6 Chaotic Transient in Boundary Crisis in a Fermi-Ulam Model 105

happening around ε = 0.033, therefore finding A = 1.956(1) and B = −0.992(1).

Therefore the variation of the parameter β retains the crisis and allows to find a
family of control parameters where it is observed.

6.4 Conclusions

In this chapter we studied the chaotic transient for a boundary crisis observed
in a dissipative version of a Fermi-Ulam model. The inelastic collisions with the
walls led to measure contraction in the phase space hence producing attractors. We
constructed the manifolds from a saddle point and show both the stable and unstable
branches of the manifold tangency each other at the boundary crisis. After the crisis
there are infinitely many crossings of these manifolds. Consequently the basin of
attraction of the fixed point becomes porous and suitable to be visited by a likely
chaotic motion. The attractor that existed prior to the boundary crisis is destroyed
therefore being replaced by a chaotic transient. The survival probability a moving
particle has to survive the region of the phase space where the chaotic attractor likely
existed before the crisis is decaying exponentially. The relaxation time is described
by a power law and that the numerical value of the exponent can be obtained from
the eigenvalues of the saddle point at the critical parameter where crisis happens.
For a homoclinic tangency [12] the expression for the exponent is γ = lnln|β|ββ2 ||2
1 2
leading to a numerical value of γ ∼ = 12 . Our definition however leads to the inverse
of this exponent, hence our numerical findings gave δ = γ1 = ∼ 2 and seem to be
universal for this type of global bifurcation.

Acknowledgements EDL acknowledges the support from CNPq (303707/2015-1), FAPESP

(2017/14414-2), and FUNDUNESP. MFM thanks to CNPq and CAPES for support.

Appendix

We present in this Appendix the expressions used to obtain the eigenvalues of the
saddle point at the parameters leading to the boundary crisis. Since there is no fixed
point for solving G(φc ), i.e., the direct collisions, we must consider only the case of
the Jacobian matrix for indirect collisions. The Jacobian matrix is written as
 ∂Vn+1 ∂Vn+1 
∂Vn ∂φn
J = ∂φn+1 ∂φn+1
,
∂Vn ∂φn

where the coefficients of J must be evaluated at the saddle point. The terms in J are
written as

edleonel@rc.unesp.br
106 E. D. Leonel and M. F. Marques

∂Vn+1 ∂φn+1
j11 = = αβ − (1 + β)ε cos(φ) , (6.14)
∂Vn ∂Vn
∂Vn+1 ∂φn+1
j12 = = −(1 + β)ε cos(φ) , (6.15)
∂φn ∂φn
∂φn+1 ∂φr ∂φl ∂φc
j21 = = + + , (6.16)
∂Vn ∂Vn ∂Vn ∂Vn
∂φn+1 ∂φr ∂φl ∂φc
j22 = = 1+ + + , (6.17)
∂φn ∂φn ∂φn ∂φn

where the auxiliary derivatives are given by

∂φr 1 − ε cos(φ)
=− , (6.18)
∂Vn V2
∂φl (1 − ε)
=− , (6.19)
∂Vn αV 2
∂φr ε sin(φ)
= , (6.20)
∂φn V
∂φl
= 0. (6.21)
∂φn

∂φc ∂φc
The two terms ∂Vn and ∂φn are determined by implicit derivatives and are given by
 
∂φr ∂φl
∂φc ε sin(φ) ∂V n
+ ∂V n
− αφc
= , (6.22)
∂Vn αV − ε sin(φ)
 
∂φr ∂φl
∂φc ε sin(φ) 1 + ∂φn + ∂φn
= , (6.23)
∂φn αV − ε sin(φ)

where
ε − ε cos(φ)
φc = . (6.24)
αV
The eigenvalues are then given by the following expressions

(j11 + j22 ) ± (j11 + j22 )2 − 4(j11 j22 − j12 j21 )
λ1,2 = . (6.25)
2
To apply on the Grebogi’s expressions [12] we have to normalize the eigenvalues.
For the control parameters ε = 0.04, β = 1, and α = 0.93264 we obtain that the
expanding and contracting eigenvalues are β1 = 3.0739934591561680 and β2 =
0.28514874278495017, respectively.

edleonel@rc.unesp.br
6 Chaotic Transient in Boundary Crisis in a Fermi-Ulam Model 107

References

1. Lichtenberg, A. J., & Lieberman, M. A. (1992). Regular and chaotic dynamics. Applied
mathematical sciences (Vol. 38). New York: Springer.
2. Hilborn, R. C. (1986). Chaos and nonlinear dynamics. New York: Oxford University Press.
3. Tabor, M. (1989). Chaos and integrability in nonlinear dynamics: An introduction. New York:
Wiley.
4. Livorati, A. L. P., Kroetz, T., Dettmann, C. P., Caldas, I. L., & Leonel, E. D. (2012). Stickiness
in a bouncer model: A slowing mechanism for Fermi acceleration. Physical Review E, 86,
036203.
5. Leonel, E. D. (2007). Corrugated waveguide under scaling investigation. Physical Review
Letters, 98, 114102.
6. Leonel, E. D., & McClintock, P. V. E. (2005). A hybrid Fermi Ulam-bouncer model. Journal
of Physics A: Mathematical and General, 38, 823.
7. Leonel, E. D. (2016). Defining universality classes for three different local bifurcations.
Communications in Nonlinear Science and Numerical Simulation, 39, 520.
8. Teixeira, R. M. N., Rando, D. S., Geraldo, F. C., Costa Filho, R. N., de Oliveira, J. A., & Leonel,
E. D. (2015). Convergence towards asymptotic state in 1-D mappings: A scaling investigation.
Physics Letters A, 379, 1246.
9. Leonel, E. D., Teixeira, R. M. N., Rando, D. S., Costa Filho, R. N., & Oliveira, J. A. (2015).
Addendum to: Convergence towards asymptotic state in 1-D mappings: A scaling investigation
[Phys. Lett. A 379 (2015) 1246]. Physics Letters A, 379, 1796.
10. Ott, E. (2002). Chaos in dynamical systems. Cambridge: Cambridge University Press.
11. Strogatz, S. H. (2015). Nonlinear dynamics and chaos: With applications to physics, biology,
chemistry and engineering. Bolder: Westview Press.
12. Grebogi, C., Ott, E., & Yorke, J. A. (1986). Critical exponent of chaotic transients in nonlinear
dynamical systems. Physical Review Letters, 57, 1284.
13. Grebogi, C., Ott, E., Romeiras, F., & Yorke, J. A. (1987). Critical exponents for crisis-induced
intermittency. Physical Review A, 36, 5365.
14. Grebogi, C., Ott, E., & Yorke, J. A. (1982). Chaotic attractors in crisis. Physical Review Letters,
48, 1507.
15. Fermi, E. (1949). On the origin of the cosmic radiation. Physical Review, 75, 1169.
16. Lichtenberg, A. J., Lieberman, M. A., & Cohen, R. H. (1980). Fermi acceleration revisited.
Physica D, 1, 291.
17. Leonel, E. D., McClintock, P. V. E., & Silva, J. K. L. (2004). Fermi-Ulam accelerator model
under scaling analysis. Physical Review Letters, 93, 014101.
18. Silva, J. K. L., Ladeira, D. G., Leonel, E. D., McClintock, P. V. E., & Kamphorst, S. O. (2006).
Scaling properties of the Fermi-Ulam accelerator model. Brazilian Journal of Physics, 36, 700.
19. Leonel, E. D., & McClintock, P. V. E. (2005). A crisis in the dissipative Fermi accelerator
model. Journal of Physics. A, Mathematical and General, 38, L425.
20. Sussman, G. J., Wisdom, J., & Mayer, M. E. (2001). Structure and interpretation of classical
mechanics. Cambridge: MIT Press.
21. Tavares, D. F., & Leonel, E. D. (2008). A simplied Fermi accelerator model under quadratic
frictional force. Brazilian Journal of Physics, 38, 58.
22. Leonel, E. D., & Carvalho, R. E. (2007). A family of crisis in a dissipative Fermi accelerator
model. Physics Letters A, 364, 475.
23. Eckmann, J.-P., & Ruelle, D. (1985). Ergodic theory of chaos and strange attractors. Reviews
of Modern Physics, 57, 617.
24. Souza, S. L. T., & Caldas, I. L. (2001). Basins of attraction and transient chaos in a gear-rattling
model. Journal of Vibration and Control, 7, 849.
25. Souza, S. L. T., Caldas, I. L., Viana, R. L., & Balthazar, J. M. (2004). Sudden changes in
chaotic attractors and transient basins in a model for rattling in gearboxes. Chaos, Solitons,
and Fractals, 21, 763.

edleonel@rc.unesp.br
108 E. D. Leonel and M. F. Marques

26. Luck, J. M., & Mehta, A. (1993). Bouncing ball with a finite restitution: Chattering, locking,
and chaos. Physical Review E, 48, 3988.
27. Leonel, E. D., Silva, J. K. L., & Kamphorst, S. O. (2004). On the dynamical properties of a
Fermi accelerator model. Physica A, 331, 435.
28. Galor, O. (2007). Discrete dynamical systems. Heidelberg: Springer.

edleonel@rc.unesp.br