Professional Documents
Culture Documents
Mohamed S. Soliman
Nonlinear systems can exhibit multiple stable steady-state solutions at a fired set of parameter values. A
disturbance to a given attractor can cause a transition across a basin boundary resulting in a qualitative, and
often substantial quantitative, change in the long-term response of the system. In this paper we examine how
the global stability properties of steady-state solutions may be examined in the context of basins of attraction.
Such a global approach, based on geometric considerations, may usefully be combined with a linearized
stability analysis. We first examine the stability properties of point attractors and harmonic solutions, and then
we extend the analysis to higher periodic and chaotic solutions. In addition we show that in situations where
basins of attraction are highly intertwined there may be not only a loss of stability but also a loss of
predictability.
xc, g(x) x E R*
where x = xE = g(x”> is a fixed point, the associated
linearized mapping Jacobian of first order partial deriva-
tives of the perturbed system evaluated at the fixed point is
given by
If all the eigenvalues have moduli less than one then the
fixed point is asymptotically stable, whereas if one of the
eigenvalues has moduli greater than one it is unstable.
When considering periodically forced systems we shall
condense, by the Poincare mapping technique for flows,
the behavior of the three-dimensional trajectories to a
mapping of a two-dimensional surface of section onto
itself.
As an example of a linearized stability analysis, we
consider the nonlinear system whose associated potential Figure 3. (a) Undamped phase trajectories for a nonlinear
oscillator with two stable (A and C) and one unstable (B)
energy function is shown in Figure 1 (b) and whose un-
equilibrium states. (b) The damped system; white represents
forced damped equation of motion is given by
the basin of attraction of state A and the stripes represent the
basin of attraction of state C.
i+25k+f(x) =o
where f(x) is a nonlinear function given by f(x) = 0.2 -
x +x3. The system has three equilibrium states {A, B, C}
at M,n,c, Y&r,, ) = ( - 0.9, 0), (0.2, 0), (1.1, 0), respec- of the equilibrium state in relation to its basin boundary
tively. The Jacobian matrix is given by allows us to determine its global rather than local stability
properties. Although in general there are no analytical
methods to determine basins of attraction (except in simple
-25 cases) a geometric approach - the direct method of
Liapunov - may be used to describe the local, not in an
with the associated eigenvalues evaluated at each of the infinitesimal structure of the vector field in the vicinity of
equilibrium states given by hi,, = - 5 f {m. an equilibrium point.5 In general, however, since basins of
Focusing on lightly damped dissipative systems where attraction usually have complicated structures, it is often
5 > 0, we find that equilibrium states A and C are asymp- appropriate to determine basins of attraction numerically.6
totically stable whereas state B is unstable. A useful starting point in analyzing the global stability
properties of nonlinear systems is to consider the unforced
undamped system with 5 = 0. For conservative systems,
2.2 Global analysis: Domains of asymptotic stability trajectories in phase space are simply contours of constant
total energy, E, given by the sum of the potential energy,
It is evident that in real physical engineering systems,
V(x), and kinetic energy, T(y). Returning to the system
perturbations are not infinitesimal; they may assume finite
whose potential energy function is shown in Figure 1 (b),
quantities. The next question that arises concerns the de-
we obtain
gree of stability of an equilibrium state. Is a stable equilib-
rium state “weakly” stable or “strongly” stable? In other E=V+T
words although an equilibrium state may be locally stable,
how nonlinearly or globally stable is it to large perturba-
tions? To address these questions we have to examine the
global structure of the associated vector field rather than The phase trajectories at different values of E are shown
the linearized vector field in the vicinity of an equilibrium in Figure 3(a). For small amplitude motions, the closed
solution. In particular, we have to determine all initial orbits surrounding the two stable equilibrium states are
conditions that result in orbits that tend to an equilibrium almost elliptical; at larger amplitudes of oscillation they
state and hence define its neighborhood of asymptotic become less elliptical, until motions reach the central
stability - its domain or basin of attraction. In systems dividing unstable equilibrium state. Larger orbits now have
with more than one stable equilibrium state, basin bound- enough energy to cross the unstable equilibrium state and
aries separate basins of attraction. Assessing the proximity translate back and forth across both stable equilibrium
states. The critical trajectory through the unstable equilib- odic, and chaotic solutions) subjected to external determin-
rium point is called a homoclinc orbit. This is often termed istic disturbances may be examined in context of their
a seperatrix because it is the boundary between qualita- associated basins of attraction. Furthermore we see that
tively different types of behavior, namely (i) motions that since basins of attraction can be highly intertwined, a
remain bounded in one well in which orbits cross back and disturbance may not only cause a loss of stability but also
forth across one stable equilibrium point and (ii) motions a loss of long-term predictability.
that are unbounded, in which orbits transverse across all
three equilibria. The seperatrix may be determined from
the energy level corresponding to the unstable equilibrium
state, E,, where E, = V(xi> such that y/2 = 3. Forced linear dissipative systems
Figure 6. Concept of basins of attraction in an idealized linear oscillator. All initial conditions shaded in grey result in transients that
do not exceed the failure criterion; white represents initial conditions that generate trajectories that cross the stability boundary. Here
o = 0.85, J= 0.05, and the window of initial conditions are - 1 < x0 < 1, - 1.5 < y0 < 1.5. The Poincare mapping point is the black dot.
where converge toward it. Here too its domain of attraction is the
whole of phase space. Since the complete solution depends
(w,~-u*)F
c= upon the independent frequencies that are related to the
(to,’ - my* + (2(w)* and forcing and transient components of the response, an ideal-
ized linear analysis yields safe bounded regions which are
25wF not only a function of initial conditions and damping level
D= but also the forcing frequency and forcing amplitude.
(w,” - 63)’ + (2&w)*
Figure 6 shows at a fixed frequency value a sequence of
In order to investigate the stability characteristics of the idealized basins at different forcing amplitudes F. We
steady-state, it is perhaps more convenient to write the have also superimposed the Poincare mapping point
complementary function, a variational solution about the
particular integral, as s = x - xp,. This transformation of
variables allows us to study the stability of the flow Linear response
A=n,-C B=~(~x~+Y~-&-wD)
x+px+(x+x-“)(l+Csln(~~)=O
n=2.1,p=o.1, -2cx<2.-2<x<2
Grid=200x200:
Figure 9. Coexisting basins of attraction for parameters where the stable equilibrium state; white represents the basin of the
resonant oscillation. The central dot represents the equilibrium state, the outer dots represent the mapping points of the period-2
oscillation.
, _AL - at (xE, yE) = (1, 0). The system has the ability to escape
over this hilltop equilibrium to the attractor at infinity.
Figure Ila shows a typical response at a fixed frequency
value. For small F, the stable equilibrium is transformed
into a small stable n = 1 oscillation. We denote the path of
n = 1 cyclic attractors that originate from the equilibrium
state by S,. Similarly, under small force the unstable
equilibrium (hill-top) becomes a small unstable n = 1 os-
Figure 10. The proximity of the equilibrium state to its basin cillation, D,. The system experiences a region of resonant
boundary. Ay,,, represents the maximum impact the equilib- hysteresis between A and B, where the two stable attrac-
rium state can sustain such that there is no change in the tors, S, and 8, (r for resonant, n for nonresonant) are
long-term behavior of the system. separated by an n = 1 saddle solution, 0,. Although this
saddle point is physically unrealizable, it plays an impor-
initial convergence or divergence to the equilibrium state. tant role in determining the structure of the basins of
By numerically integrating the equation of motions to attraction of the two competing attractors. Thus by slowly
obtain estimates of the mapping Jacobian we have deter- increasing F, the nonresonant solution will move along the
mined the moduli of the mapping eigenvalues as G is lower branch of the response curve until point A, at which
increased (Figure S(b]). The equilibrium solution loses its point there is a saddle node bifurcation. The associated
stability at a subcritical bifurcation (at point S,) at G = mapping eigenvalues shown in Figure lib illustrate the
0.20. Here one of the eigenvalues exceeds the stability movement of the eigenvalues associated with the n = 1
level of moduli equal to one; infinitesimal perturbations path that originates from the stable equilibrium at (F =x
will result in a divergence away from the equilibrium state. = y = O}. Stable solutions have eigenvalues with moduli
Meanwhile at G = 0.1, a finite amplitude subharmonic less than one; unstable solutions have moduli greater than
solution (the mapping Poincare point at each forcing cycle one. By increasing F along the resonant branch, the period
is plotted) is created at a saddle node bifurcation. Over a solution loses its stability at a supercritical flip bifurcation,
range of parameter values, 0.1 < G < 0.2, this stable peri- at C, into a stable n = 2 subharmonic solution. This is
odic solution coexists with the stable equilibrium state. followed by a series of further supercritical flip bifurca-
Figure 9 shows, in the PoincarC section, the structure of tions and ultimately a cascade to chaos; the resulting
the coexisting basins of attraction at various forcing levels. chaotic attractor loses its stability at a boundary crisis with
As can be seen as attractors evolve, the structure of their an inevitable jump to escape to the attractor at infinity
associated basins of attraction can change. Focusing on the (F = 0.10). We shall focus on the stability of higher
relative proximity of the equilibrium state to its basin periodic orbits and chaotic solutions in the next section.
boundary we see that finite disturbances may result in a In addition to the local bifurcations there are various
transition to the basin of attraction of the resonant solution. important global bifurcations that can take place in nonlin-
Figure 10 shows the magnitude of the maximum impact, ear systems that play an important role in determining the
AYlil,X7 the system can sustain without any change in the structure of coexisting basins. For further details see Soli-
long-term response. For G < 0.2, all disturbances result in man and Thompson’ who describe how the underlying
motions that return to the equilibrium state; for G > 0.2, manifold structure, determined by the stable and unstable
the creation of the resonant attractor results in impacts that manifolds of the unstable saddle solutions, dictate the
can cause a transition across the basin boundary and hence structure of coexisting basins of attraction. In that paper it
a qualitiative change in the long-term response of the
system. The sudden discontinuities in Ay,,, is here due to
relative changes in the proximity of the equilibrium state
to its basin boundary. In the next section we shall show
how global or basin bifurcations can result in discontinu-
ous changes in the distance of the attractor to its basin
boundary.
Table 1. Local and global bifurcations at (= 0.05, o = 0.83 accumulation of highly intertwined basins of attraction
Main local bifurcationsFB = 0.0488: saddle-node B, creation encircling the nonresonant solution (represented by a black
of Sr and D, dot). As before, we assess the global stability properties of
FA = 0.0800: saddle node A, destruction of S, and Dr the nonresonant solutions from a geometrical viewpoint in
Fc = 0.0915: first flip of the main period-doubling cascade the context of the coexisting basin structures. At F = 0.060,
FE = 0.0995: crisis of the main chaotic attractor although a finite disturbance may result in a transition
Main global basin bifurcations across the basin boundary, due to the smooth nature of the
F= 0.0488 (FA)
basin boundaries, there is no loss of predictability in the
Heteroclinic connections W”( D,) f- WY 4) # 0
outcome of the long-term response. On the other hand, at
and W* (D,)n W”(Dr)#O
F = 0.0623 ( FM) F = 0.075, since there is a highly fractal basin structure
Homoclinic connection W”( D,) n Ws( D,) # 0 encircling the nonresonant solution, there is not only a loss
F = 0.070 of stability but in addition a loss of predictability in
Homoclinic connection W”( 4) n W”( 4) # 0 determining the outcome to a given disturbance. Table 2
F = 0.0705 shows how such a disturbance (in this case an impact) can
Heteroclinic connection W”( D,) n W”( D,) + 0 result in changes in the long-term behaviour of the system
(here we have shown only a few selected values, although
was shown that the primary global bifurcations that cause our analysis was performed with finer increments of distur-
important changes in the structure of coexisting basins are bance). At F = 0.060, we can see for disturbances Ay <
related connections of the stable, IV’, and unstable, W”, 0.44, the system returns to the nonresonant state and hence
manifolds of the regular saddles cycles 0, and D,. For the there is no longterm change in the response; for Ay > 0.44,
purpose of our study here, we note that the most important the system deviates from the resonant state and settles onto
of these global bifurcations include (i) a homoclinic con- the resonant attractor. These results allow us to determine
nection between the stable and unstable manifolds of the the smallest disturbance required for the system to deviate
hilltop saddle [WYD,,) n W”(D,) # 01 which results in a away from the nonresonant state. Alternatively, since in
fractal basin boundary between the basins of attraction of engineering systems, small fluctuations around the nonres-
the bounded solutions and the attractor at infinity, and (ii) onant state (i.e., large but bounded oscillations) may still
the heteroclinic connection between the stable manifold of be considered as a form of stability (in the sense of
the hilltop saddle and unstable manifold of the resonant Liapunov), one can consider unbounded behavior as the
saddle [Ws(D,) n W”(Q) # 01. A summary of the main criterion for loss of stability, and hence Ay,,, = 0.84
local and global bifurcations are summarised below in would be our measure of local stability. For the case of
Table 1. F = 0.075, slightly different disturbances result in qualita-
Figure 12 shows the attractor-basin phase portraits at two tively different types of long-term behavior. Since the
different parameter values. At F = 0.06 no major global basins of attraction are fractal, and hence infinitely inter-
bifurcations have occured, whereas at F = 0.075 homo- twined, this implies that if we do not know the magnitude
clinic and heteroclinic bifurcations, resulting in significant of the disturbance with certainity we cannot predict the
qualitative changes in the basin structures, have taken final outcome. These results further highlight the added
place. We see at this higher forcing level there is a fractal complexity in the stability analysis of nonlinear systems.’
FHACTAL UASIN DOUNDARIES: A LOSS OI; STAIIILITY AND LOSS OF PREDICTADILITY IN NONLINEAR SYSTEMS
I I
F=O.60 F=0.075
r+s*+x-x’--Fslnwl
U)=o.a3,p-0.1, -0.8<xc1.2.-1~w~l
Grid=250x250
Figure 12. Coexisting basins of attraction. Grey represents the basin of attraction of the nonresonant solution; black represents the
basin of the resonant solution, and white represents the basin of the attractor at infinity. The black dots represent the mapping point
of the nonresonant solution; the white dot represents the resonant solution. Window of initial conditions are 0.8 < x,, < 1.2, 1 < y0 < 1.
Here o = 0.83, t= 0.05. At F= 0.06 basin boundaries are smooth: at F = 0.075 basin boundaries are fractal.
Table 2 A loss of stability and predictability may be asymptotically stable it lies very close to its basin
f= 0.060 F= 0.075 boundary; small disturbances can cause a transition across
its basin boundary.‘0
4 Amplitude Type Amplitude Type
0.00000 0.21600 Small 0.30943 Small
0.04000 0.21600 Small 0.30943 Small
0.08000 0.21600 Small 0.30943 Small
6. Stability of higher order periodic and chaotic
0.12000 0.21600 Small 0.30943 Small
solutions
0.16000 0.21600 Small 0.30943 Small
0.20000 0.21600 Small cc Escape So far we have considered the bifurcational behavior and
0.24000 0.21600 Small m Escape
global stability characteristics of point attractors and peri-
0.28000 0.21600 Small 0.76596 Large
odic solutions of order n = 1. We can extend our analysis
0.32000 0.21600 Small Escape
0.36000 0.21600 Small 0.;6596 Large
to periodic orbits of order n = k, which in terms of the
0.40000 0.21600 Small Escape assocaited mapping are given by
0.44000 0.21600 Small 0.76596 Large
xi+l = F( xi)
0.48000 0.73174 Large 0.76596 Large
0.52000 0.73174 Large 0.30943 Small =F(F(x,))=F’(q)
xi+2
0.56000 0.73174 Large Escape
0.60000 0.73174 Large 0.76596 Large
0.64000 0.73174 Large Escape
0.68000 0.73174 Large 0.76596 Large Xi+k = Fk( xI)
0.72000 0.73174 Large 0.76596 Large
0.76000 0.73174 Large 0.76596 Large For the sake of simplicity we shall here consider two-di-
0.80000 0.21600 Small 00 Escape mensional maps (which are implicitly obtained for flows
0.84000 m Escape m Escape by Poincare sampling) of the form
0.88000 m Escape m Escape
0.92000 m Escape m Escape X n+l =(3X,, Y,)
0.96000 m Escape m Escape
1 .ooooo m Escape CC Escape Yn+l =fJ(Xn, Y,)
Ay represents the magnitude of the impact given to the fixed where steady-state periodic solutions of order n are given
point. The subsequent columns represent the amplitude of the
by
resulting long-term response and the type of response. An
amplitude of 00 represents the attractor at infinity. Bold type- xE = G”( xE, y”)
scripts represent the maximum impact the fixed point can
sustain such that there is no change in the long-term response yE = H”( xE, y”)
of the system.
By adopting the appropriate iterate of the map, the eigen-
values of the associated linearized Jacobian of partial
The above local stability approach was carried out for derivatives allows us to determine the local stability char-
different values of F over a range of values where the acteristics of a given periodic solution. Often it is possible
nonresonant state is asymptotically stable. As seen in to determine analytically the stability characteristics of low
Figure 13, there is a general downward trend in Ay,,, as order periodic solutions, but this becomes increasingly
F is increased. These results quantify the relative stability difficult for higher periodic solutions and hence it is
characteristics of the resonant and nonresonant states; as appropriate to resort to numerical simulations. As for the
can be seen, for small values of F the nonresonant state is case of period-l solutions, we can assess the global stabil-
more robust to disturbances. Although the resonant state ity characteristics of higher order solutions by examining
the structure of their basins of attraction and in particular
their proximity to the basin boundary. As an example we
consider the two-dimensional Henon map:
X ,,+l =A-xi-Jyn
Yn+1= xl8
where J is the net contraction ratio for any finite area in
the x-y plane under the action of the map. In the following
we fix J = 0.3. The period-l solutions are given by:
-(J+l)+/m, x:
{x:7YlE} = 2
i
Figure 13. The relative robustness of the resonant and non- -(J+1)-\l(J+1)2+4A,x;
resonant solutions to an impact. Points A, B, and C represent {XZEY
Y2E) =
2
the main bifurcational parameters of the period-l solution. i
-2x -J
J=
1 0
in higher periodic orbits and ultimately chaos. The result- Figure 14. Bifurcationdiagram for the Henon map.
ing chaotic attractor will ultimately lose its stability at a
boundary crisis, resulting in motions that tend to the
attractor at infinity. Table 2 shows selected eigenvalues of 300 X 300 grid of initial conditions of {x,, yO) in which
various periodic solutions. Unstable solutions have eigen- the map was iterated until the system settled down onto a
values with moduli greater than one. It becomes increas- steady-state solution. Initial conditions that constituted the
ingly difficult to obtain eigenvalues for higher periodic basin of attraction of infinity are shaded in black. White
solutions. Figure I4 shows the full bifurcation diagram. represents initial conditions that generate bounded orbits;
Figure 15 shows a sequence of basins of attraction at in general, these are attractors that originate from the
various values of A. These were constructed using a fundamental period-l solution; although other bounded
11 12 11 12 1, 12
RE IM RE IM RE IM RE IM RE IM RE IM
7. Conclusions
Y”., -*n
In this paper we have outlined how a geometrical analysis
J=O.3.-2<x,~2.-Z~y,<12 may help us understand the dynamics of nonlinear systems
Figure 15. Coexisting basins of attraction for the Henon map. subjected to local but finite disturbances. In particular we
White represents the basin of attraction of the bounded attrac- have demonstrated that various stable solutions may coex-
tor; black represents the basin of the attractor at infinity. The ist at a fixed set of parameters; an infinitesimal stability
dots represent the mapping points of the respective bounded analysis may often be inadequate and an approach that
attractors. considers basins of attraction essential in understanding the