1.

7 The Logistic Equation and Bifurcation 43

15. Consider Ricker’s equation

x(n + 1) = x(n) exp(r(1 − x(n))).
Find the 2-period solution when r > 2.
16. The population of a certain species is modeled by the difference equa-
tion x(n + 1) = µx(n)e−x(n) , x(n) ≥ 0, µ > 0. For what values of µ
does the equation have a 2-cycle?
17. Use Carvalho’s lemma to find the values of c for which the map
Qc (x) = x2 + c, c ∈ [−2, 0],
has a 3-cycle and then determine its stability.
18*. (Term Project). Find the values of µ where the logistic equation x(n +
1) = µx(n)[1 − x(n)] has a 3-periodic solution.
19. Use Carvalho’s lemma to find the values of µ where the logistic
equation x(n + 1) = µx(n)[1 − x(n)] has a 2-periodic solution.
20. Find the 3-periodic solutions of the equation x(n + 1) = ax(n), a $= 1.

1.7 The Logistic Equation and Bifurcation

Let us now return to the most important example in this chapter: the
logistic difference equation
x(n + 1) = µx(n)[1 − x(n)], (1.7.1)
which arises from iterating the function
Fµ (x) = µx(1 − x), x ∈ [0, 1], µ > 0. (1.7.2)

1.7.1 Equilibrium Points

To find the equilibrium points (fixed points of Fµ ) of (1.7.1) we solve the
equation
Fµ (x*) = x*.
Hence the fixed points are 0, x* = (µ − 1)/µ. Next we investigate the
stability of each equilibrium point separately.
(a) The equilibrium point 0. (See Figures 1.31, 1.32.) Since Fµ" (0) = µ, it
follows from Theorems 1.13 and 1.15 that:
(i) 0 is an asymptotically stable fixed point for 0 < µ < 1,
(ii) 0 is an unstable fixed point for µ > 1.
44 1. Dynamics of First-Order Difference Equations

x(n+1)

x(n)
x0

FIGURE 1.31. 0 < µ < 1 : 0 is an asymptotically stable fixed point.

x(n+1)

x(n)
x0 x*

FIGURE 1.32. µ > 1 : 0 is an unstable fixed point, x∗ is an asymptotically fixed

point.

The case where µ = 1 needs special attention, for we have F1" (0) = 1
and F "" (0) = −2 $= 0. By applying Theorem 1.15 we may conclude that
0 is unstable. This is certainly true if we consider negative as well as
positive initial points in the neighborhood of 0. Since negative initial
points are not in the domain of Fµ , we may discard them and consider
only positive initial points. Exercises 1.5, Problem 16 tells us that 0 is
semiasymptotically stable from the right, i.e., x∗ = 0 is asymptotically
stable in the domain [0, 1].
(b) The equilibrium point x* = (µ − 1)/µ, µ $= 1. (See Figures 1.32, 1.33.)

In order to have x* ∈ (0, 1] we require that µ > 1. Now, Fµ" ((µ−1)/µ) = 2−

µ. Thus using Theorems 1.13 and 1.16 we obtain the following conclusions:
 1.7 The Logistic Equation and Bifurcation 45

x(n+1)

x(n)
x0 x*

FIGURE 1.33. µ > 3: x∗ is an unstable fixed point.

(i) x∗ is an asymptotically stable fixed point for 1 < µ ≤ 3 (Figure 1.32).

(ii) x∗ is an unstable fixed point for µ > 3 (Figure 1.33).

1.7.2 2-Cycles
To find the 2-cycles we solve the equation Fµ2 (x) = x (or we solve x2 =
µx1 (1 − x1 ), x1 = µx2 (1 − x2 )),
µ2 x(1 − x)[1 − µx(1 − x)] − x = 0. (1.7.3)
Discarding the equilibrium points 0 and x* = µ−1 µ , one may then divide
(1.7.3) by the factor x(x − (µ − 1)/µ) to obtain the quadratic equation
µ2 x2 − µ(µ + 1)x + µ + 1 = 0.
Solving this equation produces the 2-cycle
! " #$
x(0) = (1 + µ) − (µ − 3)(µ + 1) 2µ,
! " #$
x(1) = (1 + µ) + (µ − 3)(µ + 1) 2µ. (1.7.4)

Clearly, there are no periodic points of period 2 for 0 < µ ≤ 3, and there
is a 2-cycle for µ > 3. For our reference we let µ0 = 3.

1.7.2.1 Stability of the 2-Cycle {x(0), x(1)} for µ > 3

From Theorem 1.21, this 2-cycle is asymptotically stable if
|Fµ" (x(0))Fµ" (x(1))| < 1,
or
−1 < µ2 (1 − 2x(0))(1 − 2x(1)) < 1. (1.7.5)
Substituting from (1.7.4) the values of x(0) and x(1) into (1.7.5), we obtain
√
3 < µ < 1 + 6 ≈ 3.44949.
46 1. Dynamics of First-Order Difference Equations

Conclusion This 2-cycle is attracting if 3 < µ < 3.44949 . . . .

√
Question What happens when µ = 1 + 6?
In this case,
% 2 &"
Fµ (x(0)) = Fµ" (x(0))Fµ" (x(1)) = −1. (1.7.6)
(Verify in Exercises 1.7, Problem 7.)
Hence we may use Theorem 1.16, part (i), to conclude
√ that the 2-cycle is
also attracting. For later reference, letõ1 = 1 + 6. Moreover, the 2-cycle
becomes unstable when µ > µ1 = 1 + 6.

1.7.3 22 -Cycles
To find the 4-cycles we solve Fµ4 (x) = x. The computation now becomes
unbearable, and one should resort to a computer √ to do the work. It turns
out √that there is a 22 -cycle when µ > 1 + 6, which is attracting for
1 + 6 < µ < 3.544090 . . . . This 22 -cycle becomes unstable at µ > µ2 =
3.544090 . . . .
When µ = µ2 , the 22 -cycle bifurcates into a 23 cycle. The new 23 cycle
is attracting for µ3 < µ ≤ µ4 for some number µ4 . This process of double
bifurcation continues indefinitely. Thus we have a sequence {µn }∞n=0 where
at µn there is a bifurcation from a 2n−1 -cycle to a 2n -cycle. (See Figures
1.34, 1.35.) Table 1.4 provides some astonishing patterns.
From Table 1.4 we bring forward the following observations:
(i) The sequence {µn } seems to converge to a number µ∞ = 3.57 . . . .
(ii) The quotient (µn − µn−1 )/(µn+1 − µn ) seems to tend to a number δ =
4.6692016 . . . . This number is called the Feigenbaum number after its
discoverer, the physicist Mitchell Feigenbaum [56]. In fact, Feigenbaum
made a much more remarkable discovery: The number δ is universal
and is independent of the form of the family of maps fµ . However, the
number µ∞ depends on the family of functions under consideration.

µ
1 3 1+ 6

FIGURE 1.34. Partial bifurcation diagram for {Fµ }.

 1.7 The Logistic Equation and Bifurcation 47

FIGURE 1.35. The bifurcation diagram of Fµ .

TABLE 1.4. Feigenbaum table.

µn − µn−1
n µn µn − µn−1
µn+1 − µn

0 3 — —
1 3.449499 ... 0.449499 . . . —
2 3.544090 ... 0.094591 . . . 4.752027 ...
3 3.564407 ... 0.020313 . . . 4.656673 ...
4 3.568759 ... 0.004352 . . . 4.667509 ...
5 3.569692 ... 0.00093219 . . . 4.668576 ...
6 3.569891 ... 0.00019964 . . . 4.669354 ...

Theorem 1.25 (Feigenbaum [56] (1978)). For sufficiently smooth fam-

ilies of maps (such as Fµ ) of an interval into itself, the number δ =
4.6692016 does not in general depend on the family of maps.

1.7.4 The Bifurcation Diagram

Here the horizontal axis represents the µ values, and the vertical axis repre-
sents higher iterates Fµn (x). For a fixed x0 , the diagram shows the eventual
behavior of Fµn (x0 ). The bifurcation diagram was obtained with the aid of
48 1. Dynamics of First-Order Difference Equations

1 1
a computer' for nx0' 1=((2 , taking increments of 500 for µ ∈ [0, 4] and plotting
all points µ, Fµ 2 for 200 ≤ n ≤ 500.
Question What happens when µ > µ∞ ?
Answer From Figure 1.35 we see that for µ∞ < µ ≤ 4 we have a large
number of small windows where the attracting set is an asymptotically sta-
ble cycle. The largest window appears at approximately µ = 3.828427 . . .,
where we have an attracting 3-cycle. Indeed, there are attracting k-cycles
for all positive integers k, but their windows are so small that they may
not be noticed without sufficient zooming. As in the situation where
µ < µ∞ , these k-cycles lose stability and then double bifurcate into at-
tracting 2n k-cycles. We observe that outside these windows the picture
looks chaotic!
Remarks: Our analysis of the logistic map Fµ may be repeated for any
quadratic map Q(x) = ax2 + bx + c. Indeed, the iteration of the quadratic
map Q (with suitably chosen parameters) is equivalent to the iteration of
the logistic map Fµ . In other words, the maps Q and Fµ possess the same
type of qualitative behavior. The reader is asked, in Exercises 1.7, Problem
11, to verify that one can transform the difference equation
y(n + 1) = y 2 (n) + c (1.7.7)
to
x(n + 1) = µx(n)[1 − x(n)] (1.7.8)
by letting
µ µ µ2
y(n) = −µx(n) + , c= − . (1.7.9)
2 2 4
Note here that µ = 2 corresponds to c = 0, µ = 3 corresponds to c = −3 4 ,
and µ = 4 corresponds to c = −2. Naturally, we expect to have the same
behavior of the iteration of (1.7.7) and (1.7.8) at these corresponding values
of µ and c.
Comments: We are still plagued by numerous unanswered questions in
connection with periodic orbits (cycles) of the difference equation
x(n + 1) = f (x(n)). (1.7.10)

Question A. Do all points converge to some asymptotically stable periodic

orbit of (1.7.8)?
The answer is definitely no.
If f (x) = 1 − 2x2 in (1.7.10), then there are no asymptotically stable
(attractor) periodic orbits. Can you verify this statement? If you have some
difficulty here, it is not your fault. Obviously, we need some tools to help
us in verifying that there are no periodic attractors.
 1.7 The Logistic Equation and Bifurcation 49

Question B. If there is a periodic attractor of (1.7.10), how many points

converge to it?
Once again, we need more machinery to answer this question.
Question C. Can there be several distinct periodic attractors for (1.7.10)?
This question leads us to the Li–Yorke famous result “Period Three Implies
Chaos” [92]. To explain this and more general results requires the intro-
duction of the so-called Schwarzian derivative of f (x). We will come back
to these questions in Chapter 6.
Exercises 1.7
Unless otherwise stated, all the problems here refer to the logistic difference
equation (1.7.1).
1. Use the stair step diagram for F4k on [0, 1], k = 1, 2, 3, . . ., to demon-
strate that F4 has at least 2k periodic points of period k (including
periodic points of periods that are divisors of k).
2. Find the exact solution of x(n + 1) = 4x(n)[1 − x(n)].
3. Let x* = (µ − 1)/µ be the equilibrium point of (1.7.1). Show that:
(i) For 1 < µ ≤ 3, x∗ is an attracting fixed point.
(ii) For µ > 3, x∗ is a repelling fixed point.
1
4. Prove that limn→∞ F2n (x) = 2 if 0 < x < 1.
5. Let 1 < µ ≤ 2 and let x* = (µ − 1)/µ be the equilibrium point of
(1.7.1). Show that if x* < x < 12 , then limn→∞ Fµn (x) = x*.
√
6. Prove that the 2-cycle given by (1.7.4) is attracting if 3 < µ < 1 + 6.
7. Verify formula (1.7.6). √
Then show that the 2-cycle in (1.7.4) is
attracting when µ = 1 + 6.
8. Verify that µ2 ≈ 3.54 using a calculator or a computer.
*9. (Project). Show that the map Hµ (x) = sin µx leads to the same value
for the Feigenbaum number δ.
10. Show that if |µ − µ1 | < ε, then |F µ(x) − F µ1 (x)| < ε for all x ∈ [0, 1].
11. Show that (1.7.7) can be transformed to the logistic equation (1.7.8),
2
with c = µ2 − µ4 .

12. (a) Find the equilibrium points y1* , y2* of (1.7.7).

(b) Find the values of c where y1* is attracting or unstable.

(c) Find the values of c where y2* is attracting or unstable.

H WithApplications to
~hyaib,
@ Chsmtstry,
10.2 Logistic Map: Numerics
In a fascinating and influential review article, Robert May (1976) emphasized that
even simple nonlinear maps could have very complicated dynamics. The article
ends memorably with "an evangelical plea for the introduction of these difference
equations into elementary mathematics courses, so that students' intuition may be
enriched by seeing the wild things that simple nonlinear equations can do."
May illustrated his point with the logistic map

a discrete-time analog of the logistic equation for population growth (Section 2.3).
Here x,, 2 0 is a dimensionless measure of the population in the nth generation
and r 2 0 is the intrinsic growth rate. As shown in Figure 10.2.1, the graph of ( I ) is
a parabola with a maximum value of 1-14 at x = i.We restrict the control parame-
ter r to the range 0 5 r <_ 4 so that (1) maps the interval 0 Ix <: 1 into itself. (The
behavior is much less interesting for other values of x and r-see Exercise
10.2. I.)

Figure 10.2.1

Period-Doubling
Suppose we fix r, choose some initial population I,,, and then use ( I ) to gener-
ate the subsequent x,, . What happens?
For small growth rate r < 1 , the population always goes extinct: x , + 0 as
n + . This gloomy result can be proven by cobwebbing (Exercise 10.2.2).
03

For 1 < r < 3 the population grows and eventually reaches a nonzero steady
state (Figure 10.2.2). The results are plotted here as a time series of x,, vs. n.
To make the sequence clearer, we have connected the discrete points (n,x,,) by
line segments, but remember that only the corners of the jagged curves are
meaningful.

10.2 L O G I S T I C M A P : N U M E R I C S 353
 Figure 10.2.2

For larger r , say r = 3.3, the population builds up again but now oscillates
about the former steady state, alternating between a large population in one gener-
ation and a smaller population in the next (Figure 10.2.3). This type of oscillation,
in which x,,repeats every two iterations, is called aperiod-2 cycle.

Figure 10.2.3

At still larger r, say r = 3.5, the population approaches a cycle that now repeats
every four generations; the previous cycle has doubled its period to period-4 (Fig-
ure 10.2.4).

Figure 10.2.4

354 ONE-DIMENSIONAL M A P S
 Further period-doublings to cycles of period 8, 16, 32, . . . , occur as r in-
creases. Specifically, let r, denote the value of r where a 2"-cycle first appears.
Then computer experiments reveal that

(period 2 is born)
4
8
16
32

Note that the successive bifurcations come faster and faster. Ultimately the r, con-
verge to a limiting value r_ . The convergence is essentially geometric: in the limit
of large n , the distance between successive transitions shrinks by a constant factor

j We'll have a lot more to say about this number in Section 10.6.
I
1
Chaos and Periodic Windows
I

I According to Gleick (1987, p. 69), May wrote the logistic map on a corridor
blackboard as a problem for his graduate students and asked, "What the Christ

i happens for r > r,?'The answer turns out to be complicated: For many values of
r , the sequence { x , , ) never settles down to a fined point or a periodic orbit-
instead the long-term behavior is aperiodic, as in Figure 10.2.5. This is a discrete-
time version of the chaos we encountered earlier in our study of the Lorenz
( equations (Chapter 9).

Figure 10.2.5

The corresponding cobweb diagram is impressively complex (Figure 10.2.6).

10.2 L O G I S T I C M A P : N U M E R I C S 355
 xn
Figure 10.2.6

You might guess that the system would become more and more chaotic as r
increases, but in fact the dynamics are more subtle than that. To see the long-
term behavior for all values of r at once, we plot the orbit diagram, a magnifi-
cent picture that has become an icon of nonlinear dynamics (Figure 10.2.7).
Figure 10.2.7 plots the system's attractor as a function of r. T o generate the
orbit diagram for yourself, you'll need to write a computer program with
two "loops." First, choose a value of r. Then generate an orbit starting
from some random initial condition x , . Iterate for 300 cycles or so, to allow
the system to settle down to its eventual behavior. Once the transients have
decayed, plot many points, say x,,, , . . . , x,,, above that r. Then move to
an adjacent value of r and repeat, eventually sweeping across the whole pic-
ture.
Figure 10.2.7 shows the most interesting part of the diagram, in the region
3.4 I r 5 4 . At r = 3.4, the attractor is a period-2 cycle, as indicated by the two
branches. As r increases, both branches split simultaneously, yielding a period-4
cycle. This splitting is the period-doubling bifurcation mentioned earlier. A cas-
cade of further period-doublings occurs as r increases, yielding period-8, period-
16, and so on, until at r = r_ = 3.57, the map becomes chaotic and the attractor
changes from a finite to an infinite set of points.
For r > r_ the orbit diagram reveals an unexpected mixture of order and chaos,
with periodic windows interspersed between chaotic clouds of dots. The large win-
dow beginning near r = 3.83 contains a stable period-3 cycle. A blow-up of part of
the period-3 window is shown in the lower panel of Figure 10.2.7. Fantastically, a
copy of the orbit diagram reappears in miniature!

356 ONE-DIMENSIONAL MAPS I

I Figure 10.2.7 Campbell ( 1 979), p 35, courtesy of Roger Eckhardt

1 10.3 Logistic Map: A n a l y s i s

The numerical results of the last section raise many tantalizing questions. Let's try
to answer a few of the more straightforward ones.

i EXAMPLE 10.3.1 :
Consider the logistic map x,,,, = rx,,(l- x,,)for 0 < x , I I and 0 I r 5 4 . Find
all the fixed points and determine their stability.
Solution: The fixed points satisfy .r2:= f (x*) = rx *(I - x*) . Hence .r* = 0 or
1 = r(l - x*), i.e., x* = 1 - j.The origin is a fixed point for all r, whereas
x* = 1 - $ is in the range of allowable x only if r 2 1 .
Stability depends on the multiplier f f ( x * )= r - 2r.r '*. Since f '(0) = r, the ori-
gin is stable for r < 1 and unstable for r > I. At the other fixed point,

10.3 L O G I S T I C M A P : A N A L Y S I S 357
f '(x*) = r - 2 r ( l - 5 ) = 2 - r. Hence x* = 1 - $ is stable for -I < (2 - r) < 1, i.e.,
for I < r < 3 . It isunstable for r > 3 .
The results of Example 10.3.1 are clarified by a graphical analysis (Figure
10.3.1). For r < 1 the parabola lies below the diagonal, and the origin is the only
fixed point. As r increases, the parabola gets taller, becoming tangent to the diag-
onal at r = 1 . For r > 1 the parabola intersects the diagonal in a second fixed point
x* = 1 - $, while the origin loses stability. Thus we see that x * bifurcates from the
origin in a transcritical bifurcation at r = 1 (borrowing a term used earlier for dif-
ferential equations).

Figure 10.3.1

Figure 10.3. I also suggests how x * itself loses stability. As r increases beyond
1, the slope at x * gets increasingly steep. Example 10.3.1 shows that the critical
slope f '(x*) = -1 is attained when r = 3. The resulting bifurcation is called aflip
bifurcation.
Flip bifurcations are often associated with period-doubling. In the logistic map,
the flip bifurcation at r = 3 does indeed spawn a 2-cycle, as shown in the next ex-
ample.

EXAMPLE 10.3.2:
Show that the logistic map has a 2-cycle for all r > 3.
Solution: A 2-cycle exists if and only if there are two points p and q such that
f (p) = q and f (q) = p. Equivalently, such a p must satisfy f (f (p)) = p, where
f ( x ) = rx(l -x). Hence p is a fixed point of the second-iterate map

358 ONE-DIMENSIONAL MAPS

f '(x) = f (f (x)) . Since f (x) is a quadratic polynomial, f '(x) is a quartic polyno-
mial. Its graph for r > 3 is shown in Figure 10.3.2.

Figure 10.3.2

To find p and q, we need to solve for the points where the graph intersects the
diagonal, i.e., we need to solve the fourth-degree equation f '(x) = x. That sounds
hard until you realize that the fixed points x* = 0 and x* = 1- f are trivial solu-
tions of this equation. (They satisfy f (x*) = x * , so f 2 (x*) = x * automatically.)
After factoring out the fixed points, the problem reduces to solving a quadratic
equation.
We outline the algebra involved in the rest of the solution. Expansion of the
equation f 2 ( ~-)x = 0 gives r2x(l - x)[l - rx(1- x)] - x = 0 . After factoring out
x and x - (I - +) by long division, and solving the resulting quadratic equation, we
obtain a pair of roots

which are real for r > 3. Thus a 2-cycle exists for all r > 3, as claimed. At r = 3, the
roots coincide and equal x* = 1- $ = 3, which shows that the 2-cycle bifurcates
continuously from x * . For r < 3 the roots are complex, which means that a 2-cy-
cle doesn't exist.

A cobweb diagram reveals how flip bifurcations can give rise to period-
doubling. Consider any map f , and look at the local picture near a fixed point
where f '(x*) = -1 (Figure 10.3.3).

10.3 LOGISTIC MAP: ANALYSIS 359

 slope = -1 ,

xn
Figure 10.3.3

If the graph of f is concave down near x *, the cobweb tends to produce a small,
stable 2-cycle close to the fixed point. But like pitchfork bifurcations, flip bifurca-
tions can also be subcritical, in which case the 2-cycle exists below the bifurcation
and is unstable-see Exercise 10.3.1 1.
The next example shows how to determine the stability of a 2-cycle.

EXAMPLE 10.3.3:
Show that the 2-cycle of Example 10.3.2 is stable for 3 < r < 1+ & = 3.449.
(This explains the values of r, and r2 found numerically in Section 10.2.)
Solution: Our analysis follows a strategy that is worth remembering: To ana-
lyze the stability of a cycle, reduce the problem to a question about the stability of
afixedpoint, as follows. Both p and q are solutions of f 2(x) = x , as pointed out
in Example 10.3.2; hence p and q are fixed points of the second-iterate map
f 2 ( ~ )The
. original 2-cycle is stable precisely if p and q are stable fixed points
for f 2 .
Now we're on familiar ground. To determine whether p is a stable fixed point
of f 2, we compute the multiplier

(Note that the same A is obtained at x = q, by the symmetry of the final term
above. Hence, when the p and q branches bifurcate, they must do so simultane-
ously. We noticed such a simultaneous splitting in our numerical observations of
Section 10.2.)

360 ONE-DIMENSIONAL MAPS

 After carrying out the differentiations and substituting for p and q, we obtain

Therefore the 2-cycle is linearly stable for 4 + 2 r - r' I < 1, i.e., for 3 < r < 1 + &.
rn

Figure 10.3.4 shows a partial bifurcation diagram for the logistic map, based
on our results so far. Bifurcation diagrams are different from orbit diagrams in that
unstable objects are shown as well; orbit diagrams show only the attractors.

Figure 10.3.4

Our analytical methods are becoming unwieldy. A few more exact results can
be obtained (see the exercises), but such results are hard to come by. To elucidate
the behavior in the interesting region where r > r _ , we are going to rely mainly on
graphical and numerical arguments.

10.4 Periodic Windows

One of the most intriguing features of the orbit diagram (Figure 10.2.7) is the oc-
currence of periodic windows for r > r_. The period-3 window that occurs near
3.8284. . . < r 5 3.8415. . . is the most conspicuous. Suddenly, against a backdrop
of chaos, a stable 3-cycle appears out of the blue. Our first goal in this section is to
understand how this 3-cycle is created. (The same mechanism accounts for the cre-
ation of all the other windows, so it suffices to consider this simplest case.)
First, some notation. Let f (x) = rx(l - x) so that the logistic map is x,,,, = f(x,,).
or more sinlply, x,,,? = f '(x,,) . Similarly, x,),, = f '(x,,) .
Then x,,,, = f ( f (s,,))

10.4 PERIODIC WINDOWS 36 1

 The third-iterate map f 3(x) is the key to understanding the birth of the period-
3 cycle. Any point p in a period-3 cycle repeats every three iterates, by definition,
so such points satisfy p = f 3(p) and are therefore fixed points of the third-iterate
map. Unfortunately, since f (x) is an eighth-degree polynomial, we cannot solve
for the fixed points explicitly. But a graph provides sufficient insight. Figure
10.4.1 plots f 3(x) for r = 3.835.

Figure 10.4.1

Intersections between the graph and the diagonal line correspond to solutions
of f3(x) = X. There are eight solutions, six of interest to us and marked with
dots, and two imposters that are not genuine period-3; they are actually fixed
points, or period-1 points for which f (x*) = x *. The black dots in Figure 10.4.1
correspond to a stable period-3 cycle; note that the slope of f 3(x) is shallow at
these points, consistent with the stability of the cycle. In contrast, the slope
exceeds 1 at the cycle marked by the open dots; this 3-cycle is therefore unsta-
ble.
Now suppose we decrease r toward the chaotic regime. Then the graph in Fig-
ure 10.4.1 changes shape-the hills move down and the valleys rise up. The curve
therefore pulls away from the diagonal. Figure 10.4.2 shows that when r = 3.8 , the
six marked intersections have vanished. Hence, for some intermediate value be-
tween r = 3.8 and r = 3.835, the graph of f 3(x) must have become tangent to the
diagonal. At this critical value of r, the stable and unstable period-3 cycles coa-
lesce and annihilate in a tangent biJurcation. This transition defines the beginning
of the periodic window.

362 ONE-DIMENSIONAL MAPS

 Figure 10.4.2

One can show analytically that the value of r at the tangent bifurcation is
1 + & = 3.8284. . . (Myrberg 1958). This beautiful result is often mentioned in
1 textbooks and articles-but always without proof. Given the resemblance of this
I result to the 1 + & encountered in Example 10.3.3, I'd always assumed it should
be comparably easy to derive, and once assigned it as a routine homework prob-
lem. Oops! It turns out to be a bear. See Exercise 10.4.10 for hints, and Saha and
Strogatz (1994) for Partha Saha's solution, the most elementary one my class

I could find. Maybe you can do better; if so, let me know!

For r just below the period-3 window, the system exhibits an interesting kind
of chaos. Figure 10.4.3 shows a typical orbit for r = 3.8282.
nearly
period-3 chaos

0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
0 50 100 150
n
Figure 10.4.3

Part of the orbit looks like a stable 3-cycle, as indicated by the black dots. But this
is spooky since the 3-cycle no longer exists! We're seeing theghost of the 3-cycle.

10.4 PERIODIC WINDOWS 363

We should not be surprised to see ghosts-they always occur near saddle-node bi-
furcations (Sections 4.3 and 8.1) and indeed, a tangent bifurcation is just a saddle-
node bifurcation by another name. But the new wrinkle is that the orbit returns to
the ghostly 3-cycle repeatedly, with intermittent bouts of chaos between visits. Ac-
cordingly, this phenomenon is known as intermittency (Pomeau and Manneville
1980).
Figure 10.4.4 shows the geometry underlying intermittency.

Figure 10.4.4

In Figure 10.4.4a, notice the three narrow channels between the diagonal and the
graph of f 3(x). These channels were formed in the aftermath of the tangent bifur-
cation, as the hills and valleys of f (x) pulled away from the diagonal. Now focus
on the channel in the small box of Figure 10.4.4a, enlarged in Figure 10.4.4b. The
orbit takes many iterations to squeeze through the channel. Hence f 3(x,) = x, dur-
ing the passage, and so the orbit looks like a 3-cycle; this explains why we see a
ghost.
Eventually, the orbit escapes from the channel. Then it bounces around chaoti-
cally until fate sends it back into a channel at some unpredictable later time and
place.
Intermittency is not just a curiosity of the logistic map. It arises commonly in
systems where the transition from periodic to chaotic behavior takes place by a
saddle-node bifurcation of cycles. For instance, Exercise 10.4.8 shows that inter-
mittency can occur in the Lorenz equations. (In fact, it was discovered there; see
Pomeau and Manneville 1980).
In experimental systems, intermittency appears as nearly periodic motion inter-
rupted by occasional irregular bursts. The time between bursts is statistically dis-
tributed, much like a random variable, even though the system is completely
deterministic. As the control parameter is moved farther away from the periodic
window, the bursts become more frequent until the system is fully chaotic. This
progression is known as the intermittency route to chaos.

364 ONE-DIMENSIONAL M A P S
 Figure 10.4.5 shows an experimental example of the intermittency route to
chaos in a laser.

Time ( p s )

Figure 10.4.5 Harrison and Biswas ( 1 986),p. 396

The intensity of the emitted laser light is plotted as a function of time. In the lowest
panel of Figure 10.4.5, the laser is pulsing periodically. A bifurcation to intermit-
tency occurs as the system's control parameter (the tilt of the mirror in the laser
cavity) is varied. Moving from bottom to top of Figure 10.4.5, we see that the
chaotic bursts occur increasingly often.
For a nice review of intermittency in fluids and chemical reactions, see Berg6 et
al. (1984). Those authors also review two other types of intermittency (the kind
considered here is Type I intermittency) and give a much more detailed treatment
of intermittency in general.

Period-Doubling in the Window

We commented at the end of Section 10.2 that a copy of the orbit diagram ap-
pears in miniature in the period-3 window. The explanation has to do with hills and
valleys again. Just after the stable 3-cycle is created in the tangent bifurcation, the
slope at the black dots in Figure 10.4.1 is close to +I. As we increase r, the hills
rise and the valleys sink. The slope of f 3 ( x ) at the black dots decreases steadily
from +1 and eventually reaches -1. When this occurs, a flip bifurcation causes

10.4 PERIODIC WINDOWS 365

each of the black dots to split in two; the 3-cycle doubles its period and becomes a
6-cycle. The same mechanism operates here as in the original period-doubling cas-
cade, but now produces orbits of period 3 . 2 " . A similar period-doubling cascade
can be found in all of the periodic windows.

10.5 Liapunov Exponent

We have seen that the logistic map can exhibit .aperiodic orbits for certain parame-
ter values, but how do we know that this is really chaos? To be called "chaotic," a
system should also show sensitive dependence on initial conditions, in the sense
that neighboring orbits separate exponentially fast, on average. In Section 9.3 we
quantified sensitive dependence by defining the Liapunov exponent for a chaotic
differental equation. Now we extend the definition to one-dimensional maps.
Here's the intuition. Given some initial condition x,,, consider a nearby point
xo + 6,,, where the initial separation 6,) is extremely small. Let 6,, be the separation
1
after n iterates. If 16,, = 16,, le"', then 1 is called the Liapunov exponent. A positive
Liapunov exponent is a signature of chaos.
A more precise and computationally useful formula for can be derived. By
taking logarithms and noting that 6,, = f"(x,, + 6,)) - f"(x,,), we obtain

where we've taken the limit 6, -+ 0 in the last step. The term inside the logarithm
can be expanded by the chain rule:

(We've already seen this formula in Example 9.4.1, where it was derived by heuristic
reasoning about multipliers, and in Example 10.3.3, for the special case n = 2.) Hence

366 ONE-DIMENSIONAL MAPS