Physica A 255 (1998) 406414

Anomalous relaxation oscillations due

to dynamical traps
I.A. Lubashevsky
a
, V.V. Gaychuk
b; 1
, A.V. Demchuk
b;
a
Moscow State University, Leninsky Hills, Moscow, 117234, Russia
b
Institute for Applied Problems of Mechanics and Mathematics, National Academy of Sciences,
3 b Naukova Str., Lviv, 290603, Ukraine
Received 10 July 1997
Abstract
We consider relaxation oscillation governed by the coupled equations x =(x x
0
)
1+p
Q(x; y) ;
y =P(x; y) ; where is a small parameter (1), p and x
0
are given constants, and Q; P are
nonlinear functions. We investigate the relaxation oscillations in the system and it is shown
that their limit cycle exhibits anomalous properties. It is shown that small uctuations aect the
system behavior and its motion near the line x =x
0
can give rise to the strong chaotic behavior
of the system. Lyapunov exponents are calculated for the given system and fractal dimensions
are determined for the chaotic attractors of the system. c 1998 Elsevier Science B.V. All rights
reserved
1. Introduction
Let us begin by considering the following dynamical system:
x =(x x
0
)
1+p
Q(x; y) + (x x
0
)
1+q
f(t) ; (1)
y =P(x; y) ; (2)
where is a small parameter (1), x
0
; p; q are given constants such that p0 and
q0, Q; P are nonlinear functions. The constant 1 is a small parameter.
The nullclines of Eqs. (1) and (2) (i.e. the zero sets of equations Q(x; y) =0,
P(x; y) =0) are shown in Fig. 1 (curves A and B, respectively). The line x =x
0
is
the dynamic trap, and the nullcline y
1
(x) of the Eq. (1) is of the N-form and in-
tersects the line x =x
0
in the vicinity of the maximum of y
1
(x) attained at x
0
max
. The

Corresponding author. Fax: 380 322 596623; e-mail: [email protected].

1
E-mail: [email protected].
0378-4371/98/$19.00 Copyright c 1998 Elsevier Science B.V. All rights reserved
PII S 0378- 4371( 98) 00094- 6
I.A. Lubashevsky et al. / Physica A 255 (1998) 406414 407
Fig. 1. The general form of the nullclines and the attractor of the system.
stationary point (x
e
; y
e
) (i.e. the point of intersection of the nullclines) is assumed to
be either unstable or x
0
x
e
x
0
max
. The function f(t) is a random force such that
f(t)) =0 ; (3)
f(t)f(t

)) =(t t

) : (4)
In order to estimate the random force eect, as the rst step of our investigation we
will study the system motion when the random force is absent.
2. Relaxation oscillations
In order to analyze analytically the characteristic features of the relaxation oscillations
we will start from the case 1; =0. Under these conditions as it follows from
the system of equations (1), (2) we may assume that either dx=dt is very large in
comparison with dy=dt or the right-hand side of Eq. (1) is approximately equal to
zero. Therefore, the dynamics must combine fast motion governed by Eq. (1) for xed
y and slow motion when the time dependences y(t) and x(t) are such that the left-hand
side of Eq. (1) is about . Taking the latter into account let us analyze the right-hand
side of Eq. (1) in more detail. Let us begin by considering the case when the stationary
408 I.A. Lubashevsky et al. / Physica A 255 (1998) 406414
point (x =x
e
; y =y
e
) is unstable, i.e. the condition of instability holds, when the point
(Q=0, P =0) belongs to the decreasing part of the nullcline y =y
1
(x) which looks
like N.
According to the classical theory of relaxation oscillations [1] one could expect that
in this system the limit cycle will be of the form presented in Fig. 1 by the thin line.
Indeed, based on the system of equations (1), (2) we can show that two increasing parts
(segments I, III in Fig. 1) of the nullcline Q=0 are formally attractive for phase paths
whereas the decreasing one (segment II) is repulsive. The main characteristics of such
a limit cycle are directly determined by the specic shape of the nullcline. In particular,
the amplitude of the variable y oscillations corresponding to this limit cycle is about
(y
0
max
y
0
min
), where the points (x
0
max
; y
0
max
), (x
0
min
; y
0
min
), are the local maximum and
minimum of the curve y =y
1
(x). However, the right-hand side of Eq. (1) possesses a
certain peculiarity in the region xx
0
1, because it tends to zero as x x
0
. Therefore,
if the phase path reaches the region x x
0
1 the time scale hierarchy will be changed
and, thus, the phase path will not be able to cross the nullcline x =x
0
. For the given
system the increasing part of the nullcline y
1
(x) is located in this region. So, the
maximum y
max
of the variable y attained during oscillations cannot be determined by
the point (x
0
max
; y
0
max
). In this case the limit cycle of the relaxation oscillations involves
the conventional segments of the phase path going along the horizontal lines y =y
max
,
and the part of the nullcline from y
max
to y
0
min
as well as a phase path located in the
region x x
0
1 and joining the two horizontal segments. In order to calculate the
limit cycle parameters we need to nd the value of y
max
by solving directly the system
of (1); (2) in the region x x
0
1.
Dividing Eq. (1) by Eq. (2) we get the equation describing the phase path in the
xy-plane

dx
dy
=
(x x
0
)
1+p
Q(x; y)
P(x; y)
: (5)
To analyze the asymptotic behavior of the phase path in the region (x x
0
)1 we
can conne our consideration to the following approximations. First, we may set x =x
0
in the smooth functions Q(x; y); P(x; y) and assume that the function P(x
0
; y) meets
the condition P(x
0
; y)C
b
where C
b
is a certain positive constant. The latter is due
to the fact that in the region x x
0
1 the rate y must be positive and the null-
cline y =y
2
(x) of Eq. (2) goes suciently far from this region. Second, let y

be the
y-coordinate of the point at which the line x =x
0
crosses the nullcline y =y
1
(x) of
Eq. (1), that is y

=y
1
(x
0
) or what is the same Q(x
0
; y

) =0. Therefore, we as-

sume that the function Q(x
0
; y)0 for yy

, Q(x
0
; y)0 for yy

and, in ad-
dition, [Q(x
0
; y)[C
q
[y y

[1 + [y y

[
1
, where C
q
is also a certain positive
constant. The former inequalities result from the fact that in the region x x
0
1
the system motion is directed from left to right for yy

and in the opposite di-

rection for yy

. The latter inequality is due to the smoothness of the function

Q(x
0
; y).
I.A. Lubashevsky et al. / Physica A 255 (1998) 406414 409
In this way we can represent the solution of Eq. (5) in the following form:

1
(x
min
x
0
)
p

1
(x x
0
)
p

=
y

Q(x
0
; y

)
P(x
0
; y

)
dy

; (6)
where x
min
is the minimal value of x attained during the motion along the phase path.
Under the adopted assumptions the right-hand side of expression (6) formally diverges
as y . So, there should be two values y
min
y

and y
max
y

for which
y
max

Q(x
0
; y

)
P(x
0
; y

)
dy

=
y

y
min
Q(x
0
; y

)
P(x
0
; y

)
dy

=

p(x
min
x
0
)
p
: (7)
The curve x =x(y) specied by expression (6) in the region y
min
yy
max
for-
mally tends to innity, x , as y y
min
or y y
max
. This behavior of the curve
x =x(y) actually describes the crossover of the limit cycle between its horizontal seg-
ments y y
min
, y y
max
and the segment going near the vertical line x x
0
. Therefore,
formula (7) can be regarded as the expression relating the values y
max
and x
min
to the
value y
min
which is directly determined by the form of the nullcline y =y
1
(x). In
particular, for P(x
0
; y) =constant and Q(x
0
; y) =constant (y y

) from Eq. (7) we

get (y
max
y

) =(y

y
min
).
We note that the constructed limit cycle and that predicted by the classical theory of
relaxation oscillations [1] dier signicantly in the shape. In particular, the maximum
y
max
of the variable y attained during the oscillations exceeds the value y
0
max
. In
addition, the limit cycle does not come close to the stationary point (y
0
max
; x
0
max
), so
the system motion along the limit cycle is practically independent of the behavior of
the nullcline y
1
(x) near the maximum. Thus, the relaxation oscillations can occur even
when the steady solution is stable with respect to innitely small perturbations. Besides,
oscillations can be regarded as spikewise.
3. Chaotic oscillations. Qualitative description
In order to understand the eect of the random forces let us simplify the system under
consideration. Roughly speaking, in the region Q

(see Fig. 1) the system motion of

duration about unity is described by the equations (where x (x x
0
))
x x
1+p
; y 1 :
So,
x (=t)
1=p
and during the motion in the region Q

the system reaches the points where x

tr

1=p
.
In the region Q
+
the system dynamics is approximately governed by the equations
x x
1+p
; y 1 :
410 I.A. Lubashevsky et al. / Physica A 255 (1998) 406414
Thus,
x(t) (1=x
p
tr
t=)
1=p
and it takes about the time t 1 for the system to leave the region Q
+
. This is actually
the essence of the dynamic trapping.
Let us now consider the system motion under the random force. In the region Q

the motion is governed by the equation

x x
1+p
+ x
1+q
f(t) : (8)
Therefore, converting to the variable
z =
1
x
q
we can rewrite Eq. (8) in the form
z =z
(qp)=q
f(t) : (9)
By virtue of the rst term on the right-hand side of Eq. (9) during the time of order
unity the system can penetrate into the region Q

(z1) by a distance
z
reg

q=p
; (10)
which corresponds to x
tr
. By virtue of the second term the system can reach the points
at which
z
rand
(t =1)

0
f(t) dt : (11)
Taking into account Eqs. (3), (4) from Eq. (11) we get
z
rand

since

0
f(t) dt

=1 :
Comparing Eqs. (10) and (11) we nd that the random force aects the system motion
in the region Q

substantially (i.e. z
rand
z
reg
) if
pq;
(pq)=p
: (12)
Under such conditions during the motion in the region Q

the system can come so

close to the trap that solely the random force can pull the system out of the neighbor-
hood of the trap during the motion in the region Q
+
. This process is random and it
I.A. Lubashevsky et al. / Physica A 255 (1998) 406414 411
Fig. 2. The attractor in the phase space.
is possible to consider that provided inequalities (12) are fullled the system behavior
will be chaotic.
As an example the properties of the given model have been analyzed numerically
for the following sources:
Q(x; y) =x x
3
+ y; P(x; y) =y x : (13)
We have considered the solutions of Eqs. (1), (2) under dierent parameters of
the model. Figs. 2 and 3 show the attractor of the system under consideration and
corresponding oscillations (we have set =10
4
, x
0
=0:9, =100). The chaotic
oscillations have been presented in Fig. 4 ( =10
5
, x
0
=0:9, =100) and Fig. 5
(x
0
=0:9, =100). The amplitudes of the noise varied from zero up to the value
equal to 0:2.
4. Lyapunov exponents
At the conditions when the system (1)(2) demonstrates the chaotic behavior we
are interested in calculating the Lyapunov exponents for the given system. These
exponents describe the temporal evolution of small perturbations of the initial con-
ditions and enable us to calculate the fractal dimension of the chaotic attractor
(Fig. 5).
Let us rewrite the system (1)(2) in the form:

(t) =F( (t)) ; (14)

412 I.A. Lubashevsky et al. / Physica A 255 (1998) 406414
Fig. 3. The periodic oscillations.
Fig. 4. The oscillating chaotic trajectories.
where (t) =x(t); y(t) and F =

Q; P are the vector functions. Functions

Q and P
determine the right-hand sides of Eqs. (1) and (2), respectively.
Since we are interested in the evolution of a small perturbation (t) =(x(t); y(t))
of a trajectory (t), we linearize the equations of motion obtaining

(t) =
@F
@

= (t)
(t) (15)
for the temporal evolution of the vectors (t) in the tangent bundle of the system.
D=@F=@ is the 2 2 Jacobian matrix of the system (14). This matrix D depends on
the instantaneous state vector (t). As it was shown by Oseledec [2] there exists a set
I.A. Lubashevsky et al. / Physica A 255 (1998) 406414 413
Fig. 5. Chaotic attractors: (a) =10
4
; (b) =10
5
.
of initial oset vectors
1
(0),
2
(0) such that the numbers

i
= lim
T
1
T
ln
[
i
(T)[
[
i
(0)[
(16)
exist and are independent of the initial conditions. The set
1
;
2
is called the
Lyapunov spectrum of the system.
A dynamical system with chaotic behavior has at least one positive Lyapunov expo-
nent. The Lyapunov spectrum is usually ordered from the largest to the smallest values
and in our case we will have the order
1

2
.
414 I.A. Lubashevsky et al. / Physica A 255 (1998) 406414
Table 1
x
0
=0:9 =10
3
=0:2
1
=2:20 d
KY
=1:000043
=100
2
=4895:0
=10
4
=0:1
1
=1:92 d
KY
=1:000043

2
=44503:0
=10
4
=0:2
1
=4:64 d
KY
=1:000103

2
=44922:0
=10
5
=0:2
1
=170:44 d
KY
=1:000305

2
=55806:5
Lyapunov exponents are used to determine the fractal dimension of the chaotic at-
tractor by means of a relation rst suggested by Kaplan and Yorke [3]:
d
KY
=k +

k
i =1

i
[
i+1
[
; (17)
where the index k is dened from the conditions
1
+ +
k
0,
1
+ +
k
+
k+1
0.
We calculated the Lyapunov exponents for the system (1)(2), where functions
Q(x; y) and P(x; y) are determined by the equalities (13), at the values of noise ampli-
tudes =0:1 and =0:2 for the T =700 time units. The results of these calculations
are presented in Table 1.
References
[1] A.A. Andronov, A.A. Vitt, S.E. Haikin, Theory of the oscillations. Nauka, Moscow, 1981 (in Russian).
[2] V.I. Oseledec, Trans. Moskow Math. Soc. 19 (1968) 197.
[3] P. Fredericson, J. Kaplan, E. Yorke, J. Yorke, The Lyapunov dimension of strange attractors, J. Di.
Eqs. 49 (1983).