Period Three Implies Chaos

Author(s): Tien-Yien Li and James A. Yorke

Source: The American Mathematical Monthly , Dec., 1975, Vol. 82, No. 10 (Dec., 1975),
pp. 985-992
Published by: Taylor & Francis, Ltd. on behalf of the Mathematical Association of
America

Stable URL: https://www.jstor.org/stable/2318254

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Taylor & Francis, Ltd. and Mathematical Association of America are collaborating with JSTOR
to digitize, preserve and extend access to The American Mathematical Monthly

This content downloaded from

189.68.23.153 on Thu, 25 Apr 2024 21:48:11 +00:00
All use subject to https://about.jstor.org/terms
 PERIOD THREE IMPLIES CHAOS

TIEN-YIEN LI AND JAMES A. YORKE

1. Introduction. The way phenomena or processes evolve or change in time is often described
by differential equations or difference equations. One of the simplest mathematical situations occurs
when the phenomenon can be described by a single number as, for example, when the number of
children susceptible to some disease at the beginning of a school year can be estimated purely as a
function of the number for the previous year. That is, when the number x,,+ at the beginning of th
n + 1st year (or time period) can be written

(1. 1) x, = F(x.),

where F maps an interval J int

disease would be very simplistic and would contain only a shadow of the more complicated
phenomena. For other phenomena this model might be more accurate. This equation has been used
successfully to model the distribution of points of impact on a spinning bit for oil well drilling, as
mentioned in [8, 11], knowing this distribution is helpful in predicting uneven wear of the bit. For
another example, if a population of insects has discrete generations, the size of the n + 1st
generation will be a function of the nth. A reasonable model would then be a generalized logistic
equation

(1.2) Xn+1 = rXn [1 -xn/K].

A related model for insect populations was discusse

1.0 I T I I I

0.9

0.8

0.7

0.6

0.5

0.4

0.3
0.2

0.I

X2 X4 X6 X8 X10 X 12 X 14 X 16 X18 X20 SumTmary

FIG. 1. For K = 1, r = 3.9, with x, .5, the above graph is obtained by iterating Eq. (1.2) 19 times. At right
the 20 values are repeated in summary. No value occurs twice. While x2 = .975 and xO = .973 are close together,
the behavior is not periodic with period 8 since x8 = .222.

These models are highly simplified, yet even this apparently simple equation (1.2) may have
surprisingly complicated dynamic behavior. See Figure 1. We approach these equations with the
viewpoint that irregularities and chaotic oscillations of complicated phenomena may sometimes be
understood in terms of the simple model, even if that model is not sufficiently sophisticated to allow
accurate numerical predictions. LorenT [1-4] took this point of view in studying turbulent behavior
in a fascinating series of papers. He showed that a certain complicated fluid flow could be modelled

985

This content downloaded from

189.68.23.153 on Thu, 25 Apr 2024 21:48:11 +00:00
All use subject to https://about.jstor.org/terms
 986 TIEN-YIEN LI AND J. A. YORKE [December

by such a sequence x, F(x), F2(x), .., which retained some of the chaotic aspects of the original
flow. See Figure 2. In this paper we analyze a situation in which the sequence {Fn (x)} is non-periodic
and might be called "chaotic." Theorem 1 shows that chaotic behavior for (i.t) will result in any
situation in which a "population" of size x can grow for two or more successive generations and
then having reached an unsustainable height, a population bust follows to the level x or below.

F(Xn)= Xn+,

450 --

400

350 -

300

300 350 400 450 Xn

FIG. 2. Lorer1z [1] studied the equations for a rotating water-filled vessel which is circularly symmetric about
its vertical axis. The vessel is heated near the rim and cooled near its center. When the vessel is annular in shape
and the rotation rate high, waves develop and alter their shape irregularly. From a simplified set of equations
solved numerically, Lorenz let X, be in essence the maximum kinetic energy of successive waves. Plotting X,n
against X_, and connecting the points, the above graph is obtained.

In section 3 we give a well-known simple condition which guarantees that a periodic point is
stable and then in section 4 we quote a result applicable when F is like the one in Figure 2. It implies
that there is an interval J." C J such that for almost every x E J, the set of limit points of the sequence
{F (X)} is j_.
A number of questions remain unanswered. For example, is the closure of the periodic points an
interval or at least a finite union of intervals? Other questions are mentioned later.
Added in proof May has recently discovered other strong properties of these maps in his
indepei4dent study of how the behavior changes as a parameter is varied [17].

2. The main theorem. Let F: J -. For x E J, F?(x) denotes x and Fn+l(x) denotes F(Fn(x))
for n= 0,1, . We will say p is a periodic point with period n if p E J and p = Fn(p) and
p k Fk(p) for I ' k < n. We say p is periodic or is a periodic point if p is periodic for some n ' 1. We
say q is eventually periodic if for some positive integer m, p = Fm(q) is periodic. Since F need not
be one-to-one, there may be points which are eventually periodic but are not periodic. Our objective
is to understand the situations in which iterates of a point are very irregular. A special case of our
main result says that if there is a periodic point with period 3, then for each integer n = 1, 2, 3, ,
there is a periodic point with period n . Furthermore, there is an uncountable subset of points x in J
which are not even "asymptotically periodic."

This content downloaded from

189.68.23.153 on Thu, 25 Apr 2024 21:48:11 +00:00
All use subject to https://about.jstor.org/terms
 1975] PERIOD THREE IMPLIES CHAOS 987

THEOREM 1. Let J be an interval and le

which the points b = F(a), c = F2(a) and d = F3(a), satisfy

d 'a <b <c (or d _ a > b > c).

Then

TI: for every k = 1,2,... there is a periodic point in J having period k.

Furthermore,

T2: there is an uncountable set S CJ (containing no periodic points), which satisfies

the following conditions:

(A) For every p, q E S with pk- q,

(2.1) lim sup I Fn(p)-Fn(q)I > 0

and

(2.2) liminf IFn(p)-Fn(q)l = 0.

(B) For every p E S and periodic point q E J,

lim sup I Fn (p) -Fn (q) > O.

n--

REMARKS. Notice that if there is a periodic point with period 3, then th

will be ;atisfied.
An example of a function satisfying the hypotheses of the theorem is F(x) = rx[l - xIK] as in
(1.2) for r E (3.84, 4] with J = [0, K] and for r > 4, F(x) = max{O, rx[1 - xIK]} with J = [0, K]. See
[2] for a detailed description of iterates of this function for r E [0, 4). The case r = 4 is discussed in
[6, 7, 12].
While the existence of a point of period 3 implies the existence of one of period 5, the converse is
false. (See Appendix 1).
We say x E J is asymptotically periodic if -there is a periodic point p for which

(2.3) Fn(x)-Fn(p) > O as n -> o.

It follows from (B) that the set S contains no asymptotically periodic points. We remark that it is
unknown what the infimum of r is for which the equation (1.2) has points which are not
asymptotically periodic.

Proof of Theorem 1. The proof of TI introduces the main ideas for both TI and T2. We now give
the prodf of TI with necessary lemmas and relegate the tedious proof of T2 to Appendix 2.

LEMMA 0. Let G: I -> R be continuous, where I is an interval. For any compact interval II C G(I)
there is a compact interval Q CI such that G(Q) = II.

Proof. Let II = [G(p), G(q)], where p, q E I. If p < q, let r be the last point of [p, q] where
G(r) = G(p) and let s be the 1'rst point after r where G(s) = G(q). Then G([r, s]) = II. Similar
reasoning applies when p > q.

LEMMA 1. Let F: J -> J be coatinuous and let {}Inn=o be a sequence of compact intervals with
In CJ and In,+ CF(In) for all ti. Then there is a sequence of compact intervals Qn such that
Qn+1 C Qn C Io and Fn (Qn) = In Xfor n-' O. For any x E Q = n Qn we have Fn(x) E In for all n.

This content downloaded from

189.68.23.153 on Thu, 25 Apr 2024 21:48:11 +00:00
All use subject to https://about.jstor.org/terms
 988 TIEN-YIEN LI AND J. A. YORKE [December

Proof Define Qo = Io. Then F0(Qo) = IO. If Qn-I

In CF(In1) = Fn(Qn,). By Lemma 0 applied to G = Fn on Qn-I there is a compact interval
Qn C Qn - such that Fn(Qn) = In. This completes the induction.
The technique of studying how certain sequences of sets are mapped into or onto each other is
often used in studying dynamical systems. For instance, Smale uses this method in his famous
"horseshoe example" in which he shows how a homeomorphism on the plane can have infinitely
many periodic poinrts [13].

LEMMA 2. Let G: J -> R be continuous. Let I CJ be a compact interval. Assume I C G(I). Then
there is a point p E I such that G(p) = p.

Proof Let I= [go,30 f. Choose ai(i=O,1) in I such that G(ai)= i. It follows ao-G(cao)>O
and a, - G(aj) ?0 and so continuity implies G(f3) - 3 must be 0 for some f in I.
Assume d ' a < b < c as in the theorem. The proof for the case d ' a > b > c is similar and so
is omitted. Write K = [a, b] and L = [b, c].

Proof of T1: Let k be a positive integer. For k > 1 let {In} be the sequence of intervals In = L for
n = 0, , k - 2 and Jk-, = K, and define In to be periodic inductively, In+k = In for n = 0, 1, 2, . If
k = 1, let In = L for all n.
Let Qn be the sets in the proof of Lemma 1. Then notice that Qk C Qo and Fk (Qk) = Qo and so by
Lemma 2, G = Fk has a fixed point Pk in Qk. It is clear that pk cannot have period less than k for F
otherwise we would need to have Fk-l(Pk) = b, contrary to Fk+l(pk) E L. The point Pk is a periodic
point of period k for F.

3. Behavior near a periodic point. For some functions F, the asymptotic behavior of iterates of a
point can be understood simply by studying the periodic points. For

(3.1) F(x) = ax(I - x)

a detailed discussion of the points of period 1 and 2 may be found in [1] for a E [0, 4] and we now
summarize some of those results. For a E [0, 4], F: [0, 1]-* [0, 1].
For a E [0, 1], x = 0 is the only point of period 1; in fact, for x E [0, 1], the sequence Fn (x) -> as
n -* oo.
For a E (1, 3], there are two points of period 1, namely 0 and 1 - a-', and for x E (0, 1),
Fn (X) ___>I- a -as n->.oo.
For a > 3 there are also two points of period 2 which we may call p and q and of course F(p) = q
and F(q) = p. For a E (3, 1 + 6/-: 3.449) and x E (0, 1), F2n(x) converges to either p or q while
F2n+1(x) converges to the other, except for those x for which there is an n for which Fn(x) equals
the point 1 - a-l of period 1. There are only a countable number of such points so that the behavior
of {Fn (x)} can be understood by studying the periodic points.
For a > 1 + \/6, there are 4 poiiits of period 4 and for a slightly greater than 1 + \/6, F4n (x) tends
to one of these 4 unless for some n, Fn(x) equals one of the points of period 1 or 2. Therefore we
may stimmarize this situation by saying that each point in [0, 1] is asymptotically periodic.
For those values of a for which each point is asymptotically periodic, it is sufficient to study only
the periodic points and their "stability properties." For any function F a point y E J with period k is
said to be asymptotically stable if for some interval I = (y - 8, y + 8) we have

IFk(X)_y I< IX_yI for all xE L

If F is differentiable at the points y, F(y), .., Fk-l(y), there is a simple condition that will guarantee
this behavior, namely

d Fk(X) <
dx

This content downloaded from

189.68.23.153 on Thu, 25 Apr 2024 21:48:11 +00:00
All use subject to https://about.jstor.org/terms
 19751 PERIOD THREE IMPLIES CHAOS 989

By the chain rule

d Fk(y) d F(Fk-l(Y)) d Fk-l(Y)

dx dx dx

dX k-l(y)) X dF(
(3.2) -~ F(F -x - FF(y)
k-i d
=H XF(y.),

where y,, is the nth iterate, F"n(y). Therefore y is asymptotically stable if

k-i1 d
F d- F(yi) < 1, where y= F=(y).
This condition of course guarantees nothing about the limiting behavior of points which do not start
"near" the periodic point or one of its iterates. The function in Figure 2 which was studied by
Lorenz has the opposite behavior, namely, where the derivative exists we have

d F(x) > 1.
dx

For such a function every periodic point is "unstable" since for x near a periodic point y of period k,
the kth iterate Fk(X) is further from y than x is. To see this, approximate Fk(x) by

I d d+k
Fk(y) y[
d-Fk(y)[Y - X] y +-

Thus for x near y, J Fk(X) - y Iis approxi

greater than 1. Therefore Fn (x) is further from y than x is.
We do not know when values of a begin to occur for which F in (3.1) has points which are not
asymptotically periodic. For a = 3.627, F has a periodic point (which is asymptotically stable) of
period 6 (approx. x = .498). This x is therefore a point of period 3 for F2 and so Theorem 1 may be
applied to F2. Since F2 has points which are not asymptotically periodic, the same is true of F.
In order to contrast the situations in this section with other possible situations discussed in the
next section, we define the limit set of a point x. The point y is a limit point of a sequence {xn} CJ if
there is a subsequence x ni converging to y. The limit set L (x) is defined to be the set of limit points of
{Fn(x)}. If x is asymptotically periodic, then L(x) is the set {y, F(y), , Fk-i(y)} for some periodic
point y of period k.

4. Statistical properties of {FF(x)}. Theorem 1 establishes the irregularity of the behavior of

iterates of points. What is also needed is a description of the regular behavior of the sequence
{Fn (x)} when F is piecewise continuously differentiable (as is Lorenz's function in Figure 2) and

dF [ dF.
(4.1) inf > I x
where
E=J J1dx
=ix: dx
-exists
j

One approach to describing the asymptotic behavior for such functions is to describe L(x), if
possible. A second approach, which turns out to be related, is to examine the average behavior of
{Fn(x)}. The fraction of the iterates {x,. , FN-l(x)} of x that are in [a, a2] will be denoted by
+(x, N, [al, a2]). The limiting fraction will be denoted

+(x, [a1, a2]) = lirn (x, N, [al, a2])

Nwtt

when the limit exists. The subject of ergodic theory, which studies transformation on general spaces,

This content downloaded from

189.68.23.153 on Thu, 25 Apr 2024 21:48:11 +00:00
All use subject to https://about.jstor.org/terms
990 TIEN-YIEN LI AND J. A. YORKE [December

motivates the following definition. We say g is the dens

ra2

(x, [al, a2]) = g(x)dx forall al,a2EJ; al<a2.

The techniques for the study of densities use non-elementary techniques of measure
functional analysis, so that we shall only summarize the results. But their value lies in t
for certain F almost all x E J have the same density. Until recently the existence of su
had not been proved, except for the simplest of functions F. The following result has r
proved:

THEOREm 2. [5]. Let F: J--->J satisfy the following conditions:

1) F is continuous.
2) Except at one point t E J, F is twice continuously differentiable.
3) F satisfies (4.1).
Then there exists a function g: J -> [0, xc), such that for almost all x E J, g is the density of x. Als
almost all x E J, L(x) = {y: g(y) > O} which is an interval. Moreover, the set J. {y: g(y) > O} is a
interval, and L(x) = J. for almost all x.

The proof makes use of results in [8]. The problem of computationally finding the density
solved in [9].
A detailed discussion of (3.1) is given in [16], describing how L(x) varies as the parameter a in
(3.1) varies between 3.0 and 4.0.
A major question left unsolved is whether (for some nice class of functions F) the existence of a
stable periodic point implies that almost every point is asymptotically periodic.

Appendix 1: Period 5 does not imply period 3. In this Appendix we give an example which has a
fixed point of period 5 but no fixed point of period 3.
Let F: [1, 5]- [1, 5], be defined such that F(1) = 3, F(2) = 5, F(3) = 4, F(4) = 2, F(S) = 1 and on
each interval [n, n + 1], 1 n -?4, assume F is linear. Then

F3([1, 2])= F2([3, 5])= F([1, 4])= [2, 5].

Hence, F3 has no fixed points in [1, 2]. Similarly, F3([2, 3]) = [3, 5] and F3([4, 5]) = [1, 4], so ne
of these intervals contains a fixed point of F3. On the other hand,

F3([3, 4]) = F2([2, 4]) = F([2, 5]) = [1, 5] D [3, 4].

Hence, F3 must have a fixed point in [3, 4]. We shall now demonstrate that the fixed point of
unique and is also a fixed point of F.
Let p E [3, 4] be a fixed point of F3. Then F(p) E [2, 4]. If F(p) E [2, 3], then F3(p ) would be
[1, 2] which is impossible since then p could not be a fixed point. Hence F(p) E [3, 4] and
F2(p) e [2, 4]. If F2(p) E [2, 3] we would have F3(p) E [4, 5], an impossibility. Hence p, F(p), F2(p )
are all in [3, 4]. On the interval [3, 4], F is defined linearly and so F(x) = 10- 2x. It has a fixed point
10/3 and it is easy to see that F3 has a unique fixed point, which must be 10/3. Hence there is no point
of period 3.

Appendix 2. Proof of T2 of Theorem 1. Let Eft be the set of sequences M = {Mn}jn= of intervals
with

(A.1) Mn = K or Mn CL, and F(Mn) D Mn+

if Mn = K then

(A.2) n is the square of an integer and Mn+1, Mn+2 C L,

This content downloaded from

189.68.23.153 on Thu, 25 Apr 2024 21:48:11 +00:00
All use subject to https://about.jstor.org/terms
 1975] PERIOD THREE IMPLIES CHAOS 991

where K = [a, b] and L = [b, c]. Of course if n

not, so the last requirement in (A.2) is redund
in {1, , n} for which Mi = K. For each r E
such that

(A.3) im P(Mr, n 2)/n = r.

n-

Let MO {Mr: r E (3/4, 1)} CM. Then /f0 is uncountable s

Mr E 0, by Lemma 1, there exists a point Xr with F (Xr)
Then S is also uncountable. For x E S, let P(x, n) denote
F'(x) E K. We can never have Fk(Xr) = b, because then Xr
to (A.2). Consequently P(Xr, n) = P(Mr, n) for all n, and so

p(Xr) = lim P(Xr, n 2) = r

n-x

for all r. We claim that

(A.4) for p, q E S, with pi q, there exist infinitely m

Fn(q) E L or vice versa.

We may assume p(p) > p(q). Then P(p, n) - P(q, n) -> oo, and so there must be infintely m
such that Fn(p)CK and Fn(q)EL.
Since F2(b) = d -' a and F2 is continuous, there exists 6 > 0 such that F2(x) < (b + d
x E [b - 6, b] CK. If p E S and Fn(p) E K, then (A.2) implies F n+(p) E L and Fn
Therefore Fn(p)<b-6. If Fn(q)EL, then Fn(q)>b so

IF n(p) -Fn(q) I > 5.

By claim (A.4), for any p, q E S, p k q, it follows

limsup
n--
IFn(p)-Fn(q)l - 6 > 0.

Hence (2.1) is proved. This technique may be similarly used to prove (B) is

Proof of 2.2. Since F(b) = c, F(c) = d ' a, we may choose intervals [b n, c

such that
(a) [b c]= [b?,co] D[b',c']D D [b n, Cn
(b) F(x) E (bn cn) for all x E (bn+' Cn1l)
(c) F (b n+1) = c n F(c Cn+l) = b n.
Let A= nn=o[b n, cn ] b* inf A and c sup A, then F(b*) c * and F(c*) b *, because of
(c).
In order to prove (2.2) we must be more specific in our choice of the sequences Mr. In addition to
our previous requirements on M E M, we will assume that if Mk = K for both k = n2 and (n + 1)2
then Mk [b2n-2j, ,b*] for k = n2+ (2j - 1), Mk = [c* C2n-2j] for k = n2+ 2j where j = I .
For the remaining k's which are not squares of integers, we assume Mk = L.
It is easy to check that these requirements are consistent with (A.1) and (A.2), and that we can
still choose Mr so as to satisfy (A.3). From the fact that p(x) may be thought of as the limit of the
fraction of n's for which Fn2(x) E K, it follows that for any r*, r E (3/4, 1) there exist infinitely many
n such that Mr = Mr* = K for both k = n2 and (n + 1)2. To show (2.2), let Xr E S and Xr* E S. Since
bn b*, cn n->* as n -*oo, for any ? > 0 there exists N with I bn - VI < 8/2, 1cC - c_j < 8/2 for
all n > N. Then, for any n with n > N and Mr = Mr = K for both k = n2 and (n + 1)2, we have

This content downloaded from

189.68.23.153 on Thu, 25 Apr 2024 21:48:11 +00:00
All use subject to https://about.jstor.org/terms
 992 TIEN-YIEN LI AND J. A. YORKE

with k = n2+ 1 and F 2+'(Xr) and Fn2+'(xr*

F n2+1(Xr)I <?. Since there are infinitely ma
=0. 0

REMARK. The theorem can be generalized by assuming that F: J -> R without assuming that
F(J) C J and we leave this proof to the reader. Of course F(J) n J would be nonempty since it would
contain the points a, b, and c, assuming that b, c, and d, are defined.

Research partially supported by National Science Foundation grant GP-31386X.

References

1. E. N. Lorenz, The problem of deducing the climate from the governing equations, Tellus, 16 (1964) 1-11.
2. , Deterministic nonperiodic flows, J. Atmospheric Sci., 20 (1963) 130-141.
3. , The mechanics of vacillation, J. Atmospheric Sci., 20 (1963) 448-464.
4. , The predictability of hydrodynamic flow, Trans. N.Y. Acad. Sci., Ser. II, 25 (1963) 409-432.
5. T. Y. Li and J. A. Yorke, Ergodic transformations from an interval into itself, (submitted for publication).
6. P. R. Stein and S. M. Ulam, Nonlinear transformation studies on electronic computers, Los Alamos Sci.
Lab., Los Alamos, New Mexico, 1963.
7. S. M. Ulam, A Collection of Mathematical Problems, Interscience, New York, 1960, p. 150.
8. A. Lasota and J. A. Yorke, On the existence of invariant measures for piecewise monotonic transforma-
tions, Trans. Amer. Math. Soc., 186 (1973) 481-488.
9. T. Y. Li, Finite approximation for the Frobenius-Perron operator-A Solution to Ulam's conjecture,
(submitted for publication).
10. Syunro Utida, Population fluctuation, an experimental and theoretical approach, Cold Spring Harbor
Symposia on Quantitative Biology, 22 (1957) 139-151.
11. A. Lasota and P. Rusek, Problems of the stability of the motion in the process of rotary drilling with
cogged bits, Archium Gornictua, 15 (1970) 205-216 (Polish with Russian and German Summary).
12. N. Metropolis, M. L. Stein and P. R. Stein, On infinite limit sets for transformations on the unit interval, J.
Combinatorial Theory Ser. A 15, (1973) 25-44.
13. S. Smale, Differentiable dynamical systems, Bull. A.M.S., 73 (1967) 747-817 (see ?1.5).
14. G. Oster and Y. Takahashi, Models for age specific interactions in a periodic environment, Ecology, in
press.
15. D. Auslander, G. Oster and C. Huffaker, Dynamics of interacting populations, a preprint.
16. T. Y. Li and James A. Yorke, The "simplest" dynamics system, (to appear).
17. R. M. May, Biological populations obeying difference equations, stable cycles, and chaos, J. Theor. Biol.
(to appear).

DEPARTMENT OF MATHEMATICS, .UNIVERSITY OF UTAH, SALT LAKE CITY, UT 84112.

INSTITUTE FOR FLUID DYNAMICS AND APPLIED MATHEMATICS, UNIVERSITY OF MARYLAND, COLLEGE PARK, MD
20742.

This content downloaded from

189.68.23.153 on Thu, 25 Apr 2024 21:48:11 +00:00
All use subject to https://about.jstor.org/terms