Você está na página 1de 3

CHAOS AND NONLINEAR TIME SERIES

One of the key ideas of chaos theory is that non-stochastic systems can exhibit apparently random
("chaotic") behavior. For example, consider the non-stochastic time series {xt } defined recursively by
xt = (1+r ) xt 1 rxt 21 ,
where r is a positive constant. Equivalently, we have xt = f (xt 1), where f (x ) = (1+r ) x rx 2. This
choice of f provides a simple example of a nonlinear function. Similar results to those described here
would be found if we took f to be any polynomial of second degree or higher. If x 0 is the initial
value, then x 1 = f (x 0), x 2 = f (f (x 0)), and in general, xn = f n (x 0), where f n (x ) denotes the n th iterate
(not the n th power) of f . A realization of {xt } is completely determined by the values of x 0 and r .
Interestingly, it has been shown that the eventual qualitative behavior of {xt } depends quite sensitively on r . Some values of r produce very orderly and predictable realizations of {xt }, while other
values yield chaotic and essentially unpredictable realizations. As r is increased steadily from zero, the
behavior of {xt } makes a fascinating transition from order into chaos, which we now describe.
First note that for any r > 0, since f (0) = 0 and f (1) = 1, a starting value of x 0 = 0 will yield xt = 0
for all t , while x 0 = 1 will yield xt = 1 for all t . We say that 0 and 1 are fixed points of f .
If x 0 is small and positive, then x 1
(1+r ) x 0, and x 2
(1+r ) x 1, etc., so that, at least for a while,
the values of {xt } will move away from zero. Thus, the fixed point at zero is unstable.
Next, we study the nature of the fixed point at 1, i.e., we examine what happens to xn +1 if xn is
close to 1. If xn
1, then
xn +1 = f (xn )
f (1) + (xn 1) f (1)

1 + (xn 1) (1r ) .
If we define n = xn 1 and n is small, then the calculation above shows that n +1
n (1r ). Consequently, if 0 < r < 2, n +1 will be smaller in magnitude than n so the sequence xn , xn +1 , . . . will
approach 1, i.e., the fixed point at x = 1 is stable. We have thus shown that if 0 < r < 2 and we pick any
starting value between 0 and 1, the series {xt } will eventually tend to 1.

-2If r > 2, however, the magnitude of the deviations n will increase, so the fixed point at x = 1 is
now unstable. The typical eventual behavior for {xt } when r is slightly greater than 2 is a periodic
oscillation between two values. These values are fixed points of f 2(x ) = f (f (x )). It can be shown that
these fixed points are stable as long as r < 6 = 2.449. When r = 2.5, we get an eventual oscillation of
period 4. For somewhat larger values of r , we get oscillations of period 8, then 16, etc. Period doublings occur at ever more closely spaced values of r .
By the time we reach r = 2.570, chaos has set in. Realizations of {xt } are not periodic. There is
no way to predict the future value without actually constructing the iterates of f . Furthermore, if there
is any uncertainty whatsoever (even an infinitesimal round-off error) in xt , then the future values
become less and less predictable, even though the process is not random and f is known! This is
because any error in xt is magnified by the iteration of f , until eventually the future value is dominated
by the error. This phenomenon is called sensitive dependence on initial conditions (or the butterfly
effect), and is a hallmark of chaos.
The bifurcation diagram (attached) shows the eventual behavior of realizations of {xt } as a
function of r . The diagram shows the period doubling described above, as well as a variety of other
features. It can be seen that, as r is increased still further beyond 2.570, periodic behavior eventually
re-emerges from the chaos. For example, at r = 2.8284, a stable period 3 oscillation appears. This doubles to 6 , 12 , 24 , . . . and then dissolves back into chaos at r = 2.8495. This region, blown up in the
lower left corner of the diagram, seems remarkably similar to (but not identical to) the original diagram.

Chaos vs Nonlinear Time Series


It is intriguing that nonrandom systems can behave chaotically. Some take this to mean that the
best way to model a real time series is by means of deterministic chaos, instead of considering the
series as a stochastic process. The apparent advantage of the chaos approach is that, if the map f (.) can
be found, then the system will be completely understood. This does not mean, however, that we can
necessarily do a good job of long-range forecasting, even if the series was indeed generated by deterministic chaos with known f . The problem arises from the sensitive dependence on initial conditions. If
there is any uncertainty whatsoever about the current value of the time series, then this uncertainty will

-3-

grow exponentially with time (see previous section), so that values of xt in the distant future will
remain essentially unpredictable. Furthermore, there is almost always at least some uncertainty about
the current state. (Think of forecasting the weather. Even though we have fairly precise measurements
of the current conditions and we understand the physical laws governing the weather, it is impossible to
accurately forecast more than a few days ahead.) In fact, because of quantum mechanics, it is typically
impossible to measure things with absolute precision. Chaos magnifies these infinitesimal uncertainties
about the current state into very substantial uncertainties about the distant future.
If we would prefer to directly model a time series as a stochastic process, we can simply introduce a stochastic term into the mechanism which generated the chaos. For example, if we introduce an
additive innovation, so that xt = f (xt 1) + et , where {et } is Gaussian white noise, then we obtain Tongs
nonlinear AR (1) model. On the other hand, we might want to introduce the innovation in a multiplicative fashion. For example, if we define xt = f (xt 1)et with f (x ) = +x 2, we obtain the ARCH (1)
model. It is clear, then, that there are close connections between deterministic chaos and stochastic
nonlinear time series analysis.

Você também pode gostar