J Dyn Diff Equat (2010) 22:381–398

DOI 10.1007/s10884-010-9169-y

Lyapunov Exponents and Sensitive Dependence

Hüseyin Koçak · Kenneth J. Palmer

Received: 26 October 2009 / Published online: 13 April 2010

© Springer Science+Business Media, LLC 2010

Abstract Precise relationships between Lyapunov exponents and the instability/stability

of orbits of scalar discrete dynamical systems are investigated. It is known that positive Lyapu-
nov exponent does not, in general, imply instability, or, equivalently, sensitive dependence on
initial conditions. A notion of strong Lyapunov exponent is introduced and it is proved that for
continuously differentiable maps on an interval, a positive strong Lyapunov exponent implies
sensitive dependence on initial conditions. However, to have a strong Lyapunov exponent an
orbit must stay away from critical points. Nevertheless, it is shown for a restricted class of
maps that a positive Lyapunov exponent implies sensitive dependence even for orbits which
go arbitrarily close to critical points. Finally, it is proved that for twice differentiable maps
on an interval negative Lyapunov exponent does imply exponential stability of orbits.

Keywords Lyapunov exponent · Strong Lyapunov exponent · Sensitive dependence ·

Chaotic orbits · Chaos

Mathematics Subject Classification (2000) 58F22 · 58F13 · 65H10

1 Introduction: Definitions and Results

The notions of positive Lyapunov exponent and sensitive dependence on initial conditions
play a prominent role in chaotic dynamical systems. Indeed, it is a popular practice to use
the numerical value of a Lyapunov exponent as the quantitative measure of sensitivity or the

This paper is dedicated to Professor Jack Hale, in admiration and gratitude, on the occasion of his 80th
birthday.

H. Koçak
Departments of Computer Science and Mathematics, University of Miami, Coral Gables, FL 33124, USA

K. J. Palmer (B)
Department of Mathematics, National Taiwan University, Taipei 106, Taiwan
e-mail: [email protected]

123
382 J Dyn Diff Equat (2010) 22:381–398

lack thereof. A firm mathematical basis for this practice is often not clear. It is the purpose
of this paper to provide several precise results on the relationships of these two key concepts
in the context of interval maps.
For the sake of concreteness, we begin with several definitions concerning a mapping
f : [0, 1] → [0, 1] of the unit interval into itself and its positive orbit {xn }∞
n=0 through an
initial value x0 ∈ [0, 1], where xn+1 = f (xn ).

Definition 1 (Lyapunov exponent) The Lyapunov exponent λ(x 0 ) of a positive orbit {xn }∞
n=0
of an interval map f : [0, 1] → [0, 1] is defined as the number

1 
n
λ(x0 ) = lim ln | f  (xk )|,
n→∞ n + 1
k=0

if the limit exists.

The Lyapunov exponent is a property of the orbit {xn }∞ n=0 since itis easily seen that
1 n  (x )| exists then for m > 0 lim n 
if limn→∞ n+1 k=0 ln | f k n→∞
1
n+1 k=0 ln | f (x k+m )|
exists also and the limits are equal.

Definition 2 (Sensitive dependence) Let f : [0, 1] → [0, 1] be a mapping of the interval.

The positive orbit {xn }∞
n=0 through an initial value x 0 ∈ [0, 1] exhibits sensitive dependence
on initial conditions, if there exists ε > 0 such that given any δ > 0 there exists y with
|y − x0 | < δ and N ≥ 0 such that | f N (y) − f N (x0 )| ≥ ε.

The notion of sensitive dependence on initial conditions is nothing more than the classical
notion of Lyapunov instability, as evident from the following definition:

Definition 3 (Lyapunov stability) Let f : [0, 1] → [0, 1] be a mapping of the interval. The
positive orbit {xn }∞
n=0 through an initial value x 0 ∈ [0, 1] is said to be Lyapunov stable if
for all ε > 0 there exists δ > 0 such that if |y − x0 | < δ then | f n (y) − f n (x0 )| < ε for all
n ≥ 0.

It is clear that if the orbit starting at x0 exhibits sensitive dependence on initial conditions
then for m > 0 the orbit starting at xm exhibits sensitive dependence also. So it makes sense
to refer to the sensitive dependence of the orbit.
Before we describe our results, it is interesting to note that either a positive Lyapunov
exponent or sensitive dependence on initial conditions are used as a prerequisite for chaotic
orbits, with the tacit assumption of their equivalence. For example, Alligood et al. [2, p. 110]
offer the following definition of a chaotic orbit:

Let f be a map of the real line R, and let {x1 , x2 , · · · } be a bounded orbit of f . The
orbit is chaotic if (1) {x1 , x2 , · · · } is not asymptotically periodic; (2) the Lyapunov
exponent h(x1 ) is greater than zero.

With a somewhat different notion of sensitive dependence, Palis and Takens [17, p. 8]
propose the following definition of a chaotic orbit:

We say that the (positive) orbit {x, f (x), f 2 (x), · · · } of x is sensitive or chaotic if there
is a positive constant C > 0 such that any q ∈ ω(x) (see the definition in the previous
section) and any ε > 0 there are integers n 1 , n 2 , n > 0 such that d( f n 1 (x), q) < ε,
d( f n 2 (x), q) < ε, but d( f n 1 +n (x), f n 2 +n (x)) > C.

123
J Dyn Diff Equat (2010) 22:381–398 383

It is popularly believed that if the Lyapunov exponent of an orbit is positive, then the orbit
exhibits sensitive dependence on initial conditions. Indeed, Palis and Takens [17, p. 146]
conjecture the following:
It seems reasonable to even expect that full sensitivity is formally implied just by the
existence of a positive Lyapunov exponent.
This optimism does not hold true for general interval maps, however, as was shown recently
by an example of Demir and Koçak [7]. They produced a piecewise linear continuous map
of the interval with an orbit which has a positive Lyapunov exponent but the orbit does not
exhibit sensitive dependence on initial conditions. Indeed, it was this example that prompted
our present quest to uncover a precise relationship between a positive Lyapunov exponent and
sensitive dependence on initial conditions. The weakest conditions we could devise under
which a positive Lyapunov exponent implies sensitive dependence on initial conditions is the
following theorem:
Theorem 1 Suppose f : [0, 1] → [0, 1] is C 1 , that the orbit {xn }∞
n=0 satisfies
inf | f  (xn )| > 0
n≥0

and that the Lyapunov exponent λ(x0 ) > 0 exists as a limit. Then the orbit {xn }∞
n=0 exhibits
sensitive dependence on initial conditions.
It is possible to replace the added condition on the derivative in the theorem above, which
means that the orbit stays away from critical points, with a reasonable strengthening of the
definition of Lyapunov exponent. To this end, we introduce the following notion of strong
Lyapunov exponent:
Definition 4 The strong Lyapunov exponent (x 0 ) of a positive orbit {xn }∞
n=0 of an interval
map f : [0, 1] → [0, 1] is defined as the number

1 
i+n−1
(x0 ) = lim ln | f  (xk )|,
n→∞ n
k=i
if the limit exists uniformly with respect to i ≥ 0.
Now, with this new notion of strong Lyapunov exponent, we can prove the following
result:
Theorem 2 Suppose f : [0, 1] → [0, 1] is C 1 . If an orbit {xn }∞
n=0 of f has a positive strong
Lyapunov exponent (x0 ) > 0, then the orbit has sensitive dependence on initial conditions.
The proofs of these theorems are presented in Sect. 2.
In general Theorem 1 could not be applied to a chaotic map as such maps usually have
critical points and most orbits would be dense and hence come arbitrarily close to critical
points. In Theorem 3 below we exhibit a class of maps with critical points for which a positive
Lyapunov exponent does imply sensitive dependence even for orbits which come arbitrarily
close to critical points. This class includes the map f (x) = 4x(1 − x).
Theorem 3 Let f : [0, 1]  → [0, 1] be a C 2 map such that f  (c) = 0 implies that f  (c)  =
0 and there exists m ≥ 0, possibly depending on c, such that f m (c) = qc is fixed and
| f  (qc )| > 1; moreover no orbit of a critical point contains another critical point. Then if
{xn }+∞n=0 is a nonconstant orbit of f with positive Lyapunov exponent λ(x 0 ) > 0, the orbit
exhibits sensitive dependence on initial conditions.
The proof of this theorem is given in Sect. 3.

123
384 J Dyn Diff Equat (2010) 22:381–398

Demir and Koçak [7] also give an example of a piecewise linear continuous interval
map with an orbit which has a negative Lyapunov exponent but which exhibits sensitive
dependence, that is, it is not Lyapunov stable. We prove the following theorem in the final
section which establishes that with sufficient differentiability negative Lyapunov exponent
does indeed imply Lyapunov stability. This theorem is a straightforward analogue of a cor-
responding theorem going back to Lyapunov for ordinary differential equations:

Theorem 4 Suppose f : [0, 1]  → [0, 1] is C 2 . If an orbit {xn }∞

n=0 of f has negative Lyapu-
nov exponent λ(x0 ) < 0, then the orbit is Lyapunov stable.

In this paper we have concerned ourselves with the Lyapunov instability or stability of
a given orbit of a scalar interval map. There are several noteworthy studies on the sensitive
dependence, or Lyapunov instability, of almost all orbits of such maps; see, for example,
Guckenheimer [8] for a topological and Abraham et al. [1] for a measure theoretic perspec-
tive.
In order to apply these theorems to particular orbits of particular maps, one would need
to rigorously verify that the Lyapunov exponent for the map is positive. The works of Day
et al. [6] and the references therein would be of assistance in such an endeavour.
In higher dimensions, in the case of invertible maps, there would be analogous results giv-
ing criteria involving Lyapunov exponents for Lyapunov stability or instability [15]. However,
for noninvertible maps, such theorems could not be applied to orbits passing arbitrarily close
to critical points and so extensions of our Theorem 3 would be needed.
The theory of Lyapunov exponents has its origins in the study of the stability of solutions
of ordinary differential equations. Therefore, it is fitting to close this introduction with a trib-
ute to some of the classical references in this area. The landmark theorems here are, roughly
speaking, that a solution x(t) of ẋ = f (t, x(t)) is stable when ẋ = f x (t, x(t))x is regular
and all Lyapunov exponents are negative and that it is unstable when ẋ = f x (t, x(t))x is
regular and at least one Lyapunov exponent is positive. The stability result was established
by Lyapunov [16]. The instability result was due to Chetaev [3]. Similar results on stabil-
ity or instability using Lyapunov functions were expounded, for example, by La Salle and
Lefschetz [12–14] and Hale [9,10].

2 Positive Strong Lyapunov Exponent Implies Sensitive Dependence

In this section we present the proofs of Theorem 1 and Theorem 2 that were introduced in
the previous section.
Theorem 1 Suppose f : [0, 1] → [0, 1] is C 1 , that the orbit {xn }∞
n=0 satisfies

inf | f  (xn )| = A > 0

n≥0

and that the Lyapunov exponent λ(x0 ) > 0 exists as a limit. Then the orbit {xn }∞
n=0 exhibits
sensitive dependence on initial conditions.

Proof Write
an = f  (xn ).
Let yn be an orbit of f with y0  = x0 and set
wn = yn − xn .

123
J Dyn Diff Equat (2010) 22:381–398 385

Then for n ≥ 0
wn+1 = an wn + gn (wn ),
where
gn (w) = f (xn + w) − f (xn ) − f  (xn )w
so that
|gn (w)| ≤ ω(|w|)|w|,
where ω(·) is the modulus of continuity of the derivative f  (x). Now we can write
gn (wn ) = bn wn ,
where
⎧
⎨ gn (wn )
if wn  = 0
bn = w
⎩0 n if wn = 0
Note that
|bn | ≤ ω(|wn |).
We can regard wn as the solution of the linear difference equation
wn+1 = (an + bn )wn .
Suppose that w0  = 0 and
|wn | ≤ η for all n ≥ 0,
where η > 0 satisfies
ω(η) λ(x0 )
ω(η) < A, ≤ .
A 2
Then an + bn  = 0 for all n ≥ 0 and, using the inequality ln(1 + x) ≤ x,
    
 1  n
1 
n   1  n
  ω(η)
   b
ln 1 +  ≤
k
 ln |ak + bk | − ln |ak | = 
n + 1 n+1  n + 1 ak  A
k=0 k=0 k=0

for n ≥ 0 and so it follows that

1 
n
ω(η) λ(x0 )
lim inf ln |ak + bk | ≥ λ(x0 ) − ≥ .
n→∞ n+1 A 2
k=0

It follows that there exists N > 0 such that

1 
n
ln |ak + bk | ≥ λ(x0 )/4 (1)
n+1
k=0

if n ≥ N and thus
eλ(x0 )n/4
|wn | = |an−1 + bn−1 | · · · |a N + b N ||w N | ≥ |w N | → ∞
|a N −1 + b N −1 | · · · |a0 + b0 |
as n → ∞. This contradicts |wn | ≤ η for all n ≥ 0 and completes the proof of the
theorem. 

123
386 J Dyn Diff Equat (2010) 22:381–398

What we have shown in Theorem 1 is that if η > 0 satisfies the conditions given in the
proof, then whenever 0 < |y0 − x0 | ≤ η, no matter how small, there always exists n > 0 such
that |yn − xn | > η. This is actually stronger than sensitive dependence since for sensitive
dependence we need only show that given any δ > 0 there exists y0 with |y0 − x0 | < δ such
that there exists n > 0 with |yn − xn | > η.
Theorem 1 implies that if f  (x) > 0 for all x and the Lyapunov exponent is positive, then
we get sensitive dependence since the additional condition will always hold. Note that in the
example of Demir and Koçak [7], the function is strictly increasing. So essentially the only
thing lacking in their example is smoothness.
We proceed with a lemma showing that an orbit with a positive strong Lyapunov exponent
satisfies the derivative condition in Theorem 1.

Lemma 1 Suppose f : [0, 1]  → [0, 1] is C 1 and (x0 ) > 0 is a strong Lyapunov exponent
for the orbit {xn }∞
n=0 . Then

inf | f  (xn )| > 0.

n≥0

Proof. By definition, there exists N such that

1 
i+n
(x0 )/2 < ln | f  (xk )| < 3(x0 )/2
n+1
k=i

for all i ≥ 0, provided n ≥ N . Taking n = N , this implies that for all i ≥ 0

e(N +1)(x0 )/2 < | f  (xi ) · · · f  (xi+N )| < e3(N +1)(x0 )/2 .

Then the left inequality implies that

| f  (xn )| ≥ M −N e(N +1)(x0 )/2

for all n ≥ 0, where M = sup0≤x≤1 | f (x)|. This concludes the proof of the lemma. 

Theorem 2 Suppose f : [0, 1] → [0, 1] is C 1 . If an orbit {xn }∞

n=0 of f has a positive strong
Lyapunov exponent (x0 ) > 0, then the orbit has sensitive dependence on initial conditions.

Proof This follows immediately from Lemma 1 and Theorem 1. 

3 Maps for Which Positive Lyapunov Exponent Implies Sensitive Dependence

In this section we give a proof of Theorem 3. To prove the theorem we need some preliminary
results. We begin with a lemma that uses ideas in Coppel [5] and is related to Theorem 4.5
on p. 338 in [4].

Lemma 2 Let f : I  → I be a C 2 map defined on a compact interval I such that for some p

f ( p) = p, f  ( p) = α,

where |α| < 1. Then given K > 1, there exists δ > 0 such that if |x − p| ≤ δ, then

K −1 |α|n |x − p| ≤ | f n (x) − p| ≤ K |α|n |x − p| for n ≥ 0.

123
J Dyn Diff Equat (2010) 22:381–398 387

Proof Write xn = f n (x) = yn + p. Then for n ≥ 0

yn+1 = αyn + g(yn ),
where
g(y) = f ( p + y) − f ( p) − f  ( p)y.
Fix δ0 > 0. Then there exists a constant M such that
| f  (x)| ≤ 2M for |x − p| ≤ δ0
from which it follows that
|g(y)| ≤ M|y|2 for |y| ≤ δ0 .
We see that if |yn | ≤ δ ≤ δ0 , then
|yn+1 | ≤ (|α| + Mδ)|yn |.
Hence if
λ = |α| + Mδ < 1 and δ ≤ δ0 ,
we have
|yn | ≤ λn |y0 | for n ≥ 0,
whenever |y0 | ≤ δ.
Now set
z n = α −n yn
for n ≥ 0. Then for these n
z n+1 = z n + bn ,
where
bn = α −n−1 g(yn )
so that

n−1 
n−1
zn = z0 + bk = y0 + bk .
k=0 k=0

Now for n ≥ 0
|bn | ≤ |α|−n−1 M|yn |2 ≤ |α|−n−1 Mλ2n |y0 |2 ≤ M|α|−1 μn |y0 |2 ,
where
(|α| + Mδ)2
μ= <1
|α|
if δ sufficiently small. So for n ≥ 0

n−1
|z n | ≤ |y0 | + M|α|−1 μk |y0 |2 ≤ |y0 | + M|α|−1 (1 − μ)−1 δ|y0 |
k=0

123
388 J Dyn Diff Equat (2010) 22:381–398

and

n−1
|z n | ≥ |y0 | − M|α|−1 μk |y0 |2 ≥ |y0 | − M|α|−1 (1 − μ)−1 δ|y0 |.
k=0

It follows that for n ≥ 0

|yn | ≤ (1 + M|α|−1 (1 − μ)−1 δ)|α|n |y0 | ≤ K |α|n |y0 |,

provided that

M|α|−1 (1 − μ)−1 δ ≤ K − 1

and that

|yn | ≥ (1 − M|α|−1 (1 − μ)−1 δ)|α|n |y0 | ≥ K −1 |α|n |y0 |

provided that

M|α|−1 (1 − μ)−1 δ ≤ 1 − K −1 .

This completes the proof of the lemma. 

Corollary 1 Let f : [0, 1]  → [0, 1] be a C 2 map such that for some p

f ( p) = p, f  ( p) = α,

where |α| > 1. Then given K > 1, there exists δ > 0 such that

K |α|n |x − p| ≥ | f n (x) − p| ≥ K −1 |α|n |x − p|

for all n ≥ 0 such that | f k (x) − p| ≤ δ for k = 0, . . . , n.

Proof Since f ( p) = p and f  ( p)  = 0, the inverse function g(x) = f −1 (x) is well-

defined and strictly monotone on a compact interval I containing p as an interior point if
0 < p < 1 and otherwise as an endpoint. Note that g( p) = p and |g  ( p)| = 1/| f  ( p)| =
1/|α|<1 so that, in particular, we can assume g maps I into itself. Then, by Lemma 2, given
K > 1, there exists δ > 0 such that if |y − p| ≤ δ, then y ∈ I and

K −1 |α|−n |y − p| ≤ |g n (y) − p| ≤ K |α|−n |y − p| for n ≥ 0.

Choose x satisfying |x − p| ≤ δ and let n be the largest n ≥ 0 such that | f k (x) − p| ≤ δ for
k = 0, . . . , n. Set y = f n (x). Then

K −1 |α|−n | f n (x) − p| ≤ |x − p| ≤ K |α|−n | f n (x) − p|

so that

K |α|n |x − p| ≥ | f n (x) − p| ≥ K −1 |α|n |x − p|.

This concludes the proof of the corollary. 

The following corollary is a key element in the proof of Theorem 3. It shows that even if
two points are arbitrarily close to an expansive fixed point, then their orbits still separate by
a fixed distance, provided the ratio of their distances from the fixed point is small enough.

123
J Dyn Diff Equat (2010) 22:381–398 389

Corollary 2 f : [0, 1]  → [0, 1] be a C 2 map such that for some p

f ( p) = p, f  ( p) = α,
where |α| > 1. Then there exists δ0 > 0 and r with 0 < r < 1 such that if
|x − p| ≤ δ0 , 0 < |y − p| ≤ δ0 and |x − p| ≤ r |y − p|,
there exists N > 0 such that
| f N (y) − f N (x)| > δ0 /2.

Proof Without loss of generality, we can assume α > 0 since if α < 0 we can consider f 2 .
By Corollary 1 with K = 2, there exists δ > 0 such that
1 n
2α n |z − p| ≥ | f n (z) − p| ≥ α |z − p| (2)
2
for all n ≥ 0 such that | f k (z) − p| ≤ δ for k = 0, . . . , n. We can also assume δ is so small
that f  (z) > 0 for |z − p| ≤ δ so that f (z) is increasing in |z − p| ≤ δ.
Now fix δ0 such that
δ0 < δ and |z − p| ≤ δ0 ⇒ | f (z) − p| < δ.
Suppose x and y are as in the statement of the corollary where r is to be determined. If x is
on the opposite side of p to y, then since f (z) is increasing in |z − p| ≤ δ, while their orbits
are in [ p − δ, p + δ] they stay on opposite sides of p and if we let N be the first N such that
either
| f N −1 (y) − p| ≤ δ0 < | f N (y) − p| or | f N −1 (x) − p| ≤ δ0 < | f N (x) − p|
(in view of Eq. 2 there must exist such an N ), then
| f N (y) − f N (x)| > δ0 .
So we can assume x and y are on the same side of p. Note also that x is closer to p and so,
since f (z) is increasing in |z − p| ≤ δ, the orbit of x is closer to p than the orbit of y while
both orbits remain in [ p − δ, p + δ].
Now let N be the first N such that
| f N −1 (y) − p| ≤ δ0 < | f N (y) − p|.
Then, from the choice of δ0 , | f k (x) − p| ≤ | f k (y) − p| < δ for 0 ≤ k ≤ N and so, using
Eq. 2 and the hypothesis of the corollary,
| f N (y) − f N (x)| ≥ | f N (y) − p| − | f N (x) − p| > δ0 − 2α N |x − p| ≥ δ0 − 2r α N |y − p|.
But by Eq. 2 again
α N |y − p| ≤ 2| f N (y) − p| ≤ 2δ.
δ0
Hence if we r = 8δ ,
1
| f N (y) − f N (x)| > δ0 − 4r δ = δ0 .
2
Thus the corollary is proved. 

123
390 J Dyn Diff Equat (2010) 22:381–398

We can now proceed with the proof of Theorem 3.

Theorem 3 Let f : [0, 1]  → [0, 1] be a C 2 map such that f  (c) = 0 implies that f  (c)  =
0 and there exists m ≥ 0, possibly depending on c, such that f m (c) = qc is fixed and
| f  (qc )| > 1; moreover no orbit of a critical point contains another critical point. Then if
{xn }+∞n=0 is a nonconstant orbit of f with positive Lyapunov exponent λ(x 0 ) > 0, the orbit
exhibits sensitive dependence on initial conditions.

Proof We begin with some notation. The hypotheses imply there are finitely many critical
points ci , i = 1, . . . , I . Denote by m i the first m such that f m (ci ) = qci . Write qci = qi and
set

1
= min min |ci − q j |, min |ci − c j | > 0.
i, j 2 i= j
Next choose r and δ0 as in Corollary 2 applied to p = qi (we can assume they are independent
of i by taking the smallest ones).
Note next by Taylor’s theorem and since ( f m i ) (ci ) = 0,
f m i (x) − f m i (ci ) 1 1
→ Di = ( f m i ) (ci ) = f  ( f m i −1 (ci )) · · · f  ( f (ci )) f  (ci )  = 0
(x − ci )2 2 2
as x → ci . Choose η0 > 0, independent of i such that
1 | f m i (x) − qi | 3
0 < |x − ci | ≤ η0 ⇒ |Di | ≤ ≤ |Di |.
2 (x − ci )2 2
Next choose δ > 0 so that
δ < , 6 maxi |Di |δ 2 ≤ δ0 , δ ≤ η0 /2,
m δ = inf{| f  (x)| : |x − ci | ≤ δ for some i} > 0,
√
ln( δ − 1) 6M2 (1 + 3/r )
N (δ) = − max m i ≥ ,
ln M1 m δ λ(x0 )
where
M1 = max | f  (x)|, M2 = max | f  (x)|.
Then choose η1 > 0 so that
η1 ≤ η0 /2, 6 max |Di |η12 ≤ δ0
i

and η2 > 0 so that

λ(x0 )
η2 ≤ Aδ ,
6M2
where
Aδ = min{| f  (x)| : |x − ci | ≥ δ for all i}.
The basic idea of the proof is to follow the proof of Theorem 1. We want to prove that an
orbit yn starting near x0 diverges from xn . Problems arise when xn is near a critical point.
First, in Step 1 we estimate the number of points in the orbit x n which pass close to the critical
point. Now according to the hypothesis of Theorem 3, if xn is near a critical point ci , then
m i iterates later it is near the expanding fixed point qi . In Step 2 we use this together with
Corollary 2 to show that if for some n, xn is near a critical point ci and is relatively closer to

123
J Dyn Diff Equat (2010) 22:381–398 391

ci than yn , the two orbits will separate after they pass close to the fixed point qi . On the other
hand, if whenever xn is near a critical point it is never relatively closer than yn , we show in
Step 3 that the argument of Theorem 1 can be pushed through using the estimate found in
Step 1 of the number of points in the orbit xn which pass close to the critical point. The proof
is concluded in Step 4. Now we turn to the first step in the proof.
Step 1: Let n 1 < n 2 < · · · be the sequence of those n ≥ 0 such that |xn − ci | < δ for some
i (note that because λ(x0 ) is finite, xn  = ci for all n and i). We define the set

In = {k : 0 ≤ k ≤ n : there exists j with k = n j }

and we denote by Jn the cardinality of In . Our aim in this step is to obtain an upper bound
for Jn .
Let n be one of the n j . Since f m i (ci ) = qi , we have for k ≥ 0

|xn+m i +k − qi | = | f m i +k (xn ) − f m i +k (ci )| ≤ M1m i +k δ.

It follows that for all j

|xn+m i +k − c j | ≥ |qi − c j | − |xn+m i +k − qi | ≥ − M1m i +k δ ≥ δ

provided that k ≤ N (δ). Hence for all j

n j+1 − n j ≥ N (δ).

Then

Jn
n ≥ n Jn ≥ n Jn − n 1 = (n j − n j−1 ) ≥ (Jn − 1)N (δ).
j=2

so that
Jn 1 1
≤ + . (3)
n+1 N (δ) n+1

Step 2: Suppose that n = n j so that there exists i such that 0 < |xn − ci | < δ. Let {yk }∞
k=0
be another orbit such that for n = n j we have |yn − xn | ≤ η1 and

r
|xn − ci | ≤ |yn − ci |. (4)
3
Then we show that the x and y orbits separate by the distance δ0 /2.
Since 0 < |xn − ci | < δ ≤ η0 /2
1 3
|Di ||xn − ci |2 ≤ |xn+m i − qi | ≤ |Di ||xn − ci |2
2 2
and similarly since 0 < |yn − ci | < δ + η1 ≤ η0
1 3
|Di ||yn − ci |2 ≤ |yn+m i − qi | ≤ |Di ||yn − ci |2 .
2 2
It follows from the last two relations and Eq. 4 that
|xn+m i − qi | |xn − ci |2
≤3 ≤ r.
|yn+m i − qi | |yn − ci |2

123
392 J Dyn Diff Equat (2010) 22:381–398

Hence we have
3
|xn+m i − qi | ≤ r |yn+m i − qi | < |yn+m i − qi | ≤ |Di |(δ + η1 )2 ≤ δ0 .
2
Then by Corollary 2 with p = qi applied to xn+m i and yn+m i , there exists N > 0 such that
1
| f N (yn+m i ) − f N (xn+m i )| > δ0 .
2

It follows that if for some i and j, 0 < |xn j − ci | < δ, |xn j − ci | ≤ 3 |yn j
r
− ci | and
|yn j − xn j | ≤ η1 the orbits will separate by at least a distance of δ0 /2.

Step 3: In this step we suppose that the situation described in Step 2 does not arise and find
with the help of the estimate in Step 1 that the proof of Theorem 1 can be pushed through
when y0 ∈ / S, where

S = {y ∈ [0, 1] : f n (y) = xn for some n > 0}.

Note if y0 ∈ S, then the orbit of y0 will land on the orbit of x0 at some time. However the
set S is at most countable. This follows from the fact that the preimage under f of any point
must be finite. For if it were infinite, then there would be an accumulation point which must
be a critical point. However near a critical point the function is strictly increasing on one side
and strictly decreasing on the other side and so there are at most two points in any preimage.
So we assume that y0  = x0 , y0 ∈ / S and that for all j,

r
|xn j − ci | > |yn − ci |, (5)
3 j
where i is the unique i such that 0 < |xn j − ci | < δ. Then, proceeding as in the proof of
Theorem 1, we suppose that

|yn − xn | ≤ η2 for all n ≥ 0

and get a contradiction.

As in the proof of Theorem 1, set

wn = yn − xn .

wn satisfies

wn+1 = [an + bn ]wn ,

where

| f  (xn ) − f  (ci )| ≥ m δ |xn − ci | if |xn − ci | < δ
|an | = | f  (xn )| =
≥ Aδ if |xn − ci | ≥ δ
and
1
|bn | ≤ M2 |wn |.
2
Then if n = n j for some j so that 0 < |xn − ci | < δ for some i,
  √
 bn 
  ≤ M2 |yn − xn | ≤ M2 (|yn − ci | + |xn − ci |) ≤ M2 (1 + 3/r ) . (6)
 a  2m |x − c | 2m δ |xn − ci | 2m δ
n δ n i

123
J Dyn Diff Equat (2010) 22:381–398 393

On the other hand, if n  = n j for all j,

 
 bn 
  ≤ M2 |yn − xn | ≤ M2 η2 . (7)
a  2| f  (xn )| 2 Aδ
n

Next, noting that an + bn  = 0 for all n ≥ 0 since y0 ∈

/ S,
 
1  1  1   bk 
n n n
ln |ak + bk | = ln |ak | + ln 1 + 
n+1 n+1 n+1 ak
k=0 

k=0 k=0
 1  

n
 bk  
> λ(x0 )/2 −  ln 1 +  (8)
n + 1 ak 
k=0

if n is sufficiently large. Then, using Eqs. 3, 6, 7 and the inequality ln(1 + x) ≤ x for x ≥ 0,
and recalling the definitions of In and Jn from Step 1,
  
 1  n
 bk 
 
 ln 1 + 
n + 1 ak 
k=0
 ⎡ ⎤
       
 1 b   b  
=  ⎣ ln 1 +  + ln 1 + ⎦
k k
 n + 1 k ∈I / n
ak
k∈In
ak 
   √ 
n + 1 − Jn M 2 η2 Jn M2 (1 + 3/r )
≤ ln 1 + + ln 1 +
n+1 2 Aδ n+1 2m δ
  √
M 2 η2 1 1 M2 (1 + 3/r )
≤ + +
2 Aδ N (δ) n+1 2m δ
≤ λ(x0 )/12 + λ(x0 )/12 + λ(x0 )/12
= λ(x0 )/4, (9)

provided
√
6M2 (1 + 3/r )
n≥ − 1.
m δ λ(x0 )
From Eqs. 8 and 9, we conclude that

1 
n
ln |ak + bk | > λ(x0 )/4
n+1
k=0

if n is sufficiently large. Reasoning as after Eq. 1 in the proof of Theorem 1, we obtain a

contradiction. It follows that if y0  = x0 , y0 ∈
/ S and Eq. 5 holds for all j, then for some
n ≥ 0, we must have |yn − xn | > η2 .
Step 4: From Steps 2 and 3 we conclude that for any orbit yn with y0  = x0 and y0 ∈
/ S, there
exists n ≥ 0 such that

|yn − xn | > ε = min(δ0 /2, η1 , η2 ).

Note that since S is at most countable, given any δ > 0 there exists y0 ∈
/ S such that
0 < |y0 − x0 | < δ. The sensitive dependence follows and the proof of the theorem is
complete. 

123
394 J Dyn Diff Equat (2010) 22:381–398

4 Negative Lyapunov Exponent Implies No Sensitive Dependence

In this section we give a proof of Theorem 4. This result was suggested by Lyapunov’s
stability theorem for ODEs [15].

Theorem 4 Suppose f : [0, 1]  → [0, 1] is C 2 . If an orbit {xn }∞

n=0 of f has negative Lyapu-
nov exponent λ(x0 ) < 0, then the orbit is Lyapunov stable.

Proof Given ε > 0, there exists an integer N such that

1 
n
λ(x0 ) − ε < ln |ak | < λ(x0 ) + ε
n+1
k=0

if n > N , where

an = f  (xn ).

Then if N < k < n

|an−1 · · · a0 |
|an−1 · · · ak | = < en(λ(x0 )+ε)−k(λ(x0 )−ε) = e(n−k)(λ(x0 )+ε)+2kε .
|ak−1 · · · a0 |
From this it is not hard to show that there exists a constant C ≥ 1 depending on ε such that
if 0 ≤ k < n, then

|an−1 · · · ak | ≤ Ce(n−k)(λ(x0 )+ε)+2kε . (10)

Note this also holds for k = n if we interpret empty products as 1. From now on we write

λ(x0 ) + ε = λ

and we fix ε > 0 satisfying

λ(x0 ) + 3ε = λ + 2ε < 0.

Now let yn be another orbit of f and set

yn = xn + wn .

Then for n ≥ 0

wn+1 = an wn + gn (wn ),

where

gn (w) = f (xn + w) − f (xn ) − f  (xn )w.

Note that if M = max | f  (x)|, then

1
|gn (w)| ≤ M|w|2 . (11)
2
Now for n ≥ 0 we can write, interpreting empty sums as zero,

n
wn = an−1 · · · a0 w0 + an−1 · · · ak gk−1 (wk−1 )
k=1

123
J Dyn Diff Equat (2010) 22:381–398 395

so that for n ≥ 0, using Eqs. 10 and 11,

1  n
|wn | ≤ Ceλn |w0 | + MC eλ(n−k)+2kε |wk−1 |2 . (12)
2
k=1

Now suppose for a given η > 0 that

|wn | < ηeλn for 0 ≤ n < T. (13)

We will show provided |w0 | is sufficiently small depending on η but not on T that Eq. 13
holds for n = T also and hence for all n ≥ 0. From this the theorem will follow immediately.
To this end, observe from Eq. 12 for 0 ≤ n ≤ T (note the ≤ T here) that,

1  n
|wn | ≤ Ceλn |w0 | + MC eλ(n−k)+2kε ηeλ(k−1) |wk−1 |.
2
k=1

Next set

z n = |wn |e−λn .

Then for 0 ≤ n ≤ T

1  n
z n ≤ C|w0 | + MCη e2kε+λ(k−2) z k−1 .
2
k=1

That is, for 0 ≤ n ≤ T ,


n
zn ≤ B + μk z k−1 , (14)
k=1

where
1
B = C|w0 |, μn = MCηe2nε+λ(n−2) .
2
Now we need a discrete Gronwall Lemma, which was suggested by Lemma B.3.1 on page
347 of Winkler [18]. First we prove another lemma.

Lemma 3 If the μn are nonnegative numbers, then for n ≥ 1

k−1   

n  
n
1+ μk exp μ
≤ exp μ
,
k=1
=1
=1

empty sums being interpreted as zero.

Proof For n = 1 we get

1 + μ1 ≤ exp(μ1 ),

123
396 J Dyn Diff Equat (2010) 22:381–398

which is true. Assuming the inequality holds for n, we prove it for n + 1. We get
n+1   n 
 
exp μ
= exp μ
exp(μn+1 )

=1
=1
   
n 
n
≥ exp μ
+ μn+1 exp μ

=1
=1
since exp(x) ≥ 1 + x
k−1   

n  
n
≥ 1+ μk exp μ
+ μn+1 exp μ

k=1
=1
=1
using the induction hypothesis
k−1 

n+1 
= 1+ μk exp μ
.
k=1
=1
This completes the proof of the lemma by induction. 

Lemma 4 (Discrete Gronwall). Let z n be nonnegative numbers such that for 0 ≤ n ≤ T ,


n
zn ≤ B + μk z k−1 ,
k=1
where B and μn are nonnegative numbers. Then for 0 ≤ n ≤ T ,
 n 

z n ≤ B exp μk .
k=1
Proof. For n = 0 it is clearly true. Assuming it is true for all nonnegative integers ≤ n − 1,
we prove it for n. We have,

n
zn ≤ B + μk z k−1
k=1
k−1 
n 
≤ B+ μk B exp μ

k=1
=1
by the induction hypothesis
 k−1 

n 
= B 1+ μk exp μ

 k=1 
=1
 n
≤ B exp μ
using Lemma 3.

=1
Thus the lemma is proved. 

Applying Lemma 4 to Eq. 14, where


n 
n
1 1
μk = MCηe2kε+λ(k−2) < MCηe−2λ eλ+2ε (1 − eλ+2ε )−1 = Dη,
2 2
k=1 k=1
we get for 0 ≤ n ≤ T
z n ≤ C|w0 |e Dη

123
J Dyn Diff Equat (2010) 22:381–398 397

and hence
|wn | ≤ C|w0 |e Dη eλn . (15)
In particular,
|wT | ≤ C|w0 |e Dη eλT < ηeλT ,
provided that
|w0 | < C −1 e−Dη η. (16)
Now if |w0 | satisfies Eq. 16 then certainly Eq. 13 holds for some T > 0 and so by repeating
the above argument, we conclude that Eq. 13 holds for all T > 0 so that
|wn | < ηeλn for n ≥ 0.
Moreover, it follows also from the above argument that then Eq. 15 holds for n ≥ 0. So if
we fix η > 0 and assume
|y0 − x0 | < C −1 e−Dη η,
we conclude that for all n ≥ 0
|yn − xn | ≤ C|y0 − x0 |e Dη eλn (17)
from which the Lyapunov stability clearly follows. So the theorem is proved. 

Remark Note if the Lyapunov exponent exists in the uniform sense as in Definition 4 and if
it is negative, it follows by a slight modification of the above proof that the orbit {x n }∞
n=0 is
uniformly asymptotically stable and hence by a well-known result (confer [11]) its ω-limit set
consists of a single asymptotically stable periodic orbit. When the exponent is not uniform,
the orbit is asymptotically stable (in fact exponentially stable as in Eq. 17) but perhaps not
uniformly so and its ω-limit set may be more complicated.

Acknowledgements H. K. is supported in part by the National Science Foundation grants CMG0417425

and CMG0825547, and K.P. by NSC (Taiwan) 97-2115-M-002-011-MY2.

References

1. Abraham, C., Biau, J., Cadre, B.: On Lyapunov exponents and sensitivity. J. Math. Anal. Appl. 290,
395–404 (2004)
2. Alligood, K., Sauer, T., Yorke, J.: Chaos: An Introduction to Dynamical Systems. Springer, New
York (1997)
3. Chetaev, N.G.: Stability of Motion. GITTL, Moscow-Leningrad (1946) (in Russian)
4. Coddington, E.A., Levinson, N.: Theory of Ordinary Differential Equations. McGraw-Hill, New
York (1955)
5. Coppel, W.A.: Stability and Asymptotic Behavior of Differential Equations. D.C. Heath and Company,
Boston, MA (1965)
6. Day, S., Kokubu, H., Luzzatto, S., Mischaikow, C., Oka, H., Pilarczyk, P.: Quantitative hyperbolicity
estimates in one-dimensional dynamics. Nonlinearity 21, 1967–1987 (2008)
7. Demir, B., Koçak, S.: A note on positive Lyapunov exponent and sensitive dependence on initial
conditions. Chaos Solitons Fractals 12, 2119–2121 (2001)
8. Guckenheimer, J.: Sensitive dependence to initial conditions for one dimensional maps. Commun. Math.
Phys. 70, 133–160 (1979)
9. Hale, J.: Ordinary Differential Equations. Wiley-Interscience, New York (1969)

123
398 J Dyn Diff Equat (2010) 22:381–398

10. Hale, J.: Studies in Ordinary Differential Equations. MAA Studies in Mathematics, Vol. 14. Mathematical
Association of America, Providence, RI (1977)
11. La Salle, J.P.: A Study of Synchronous Asymptotic Stability. Ann. Math. (Second Series) 65,
571–581 (1957)
12. La Salle, J.P.: The Stability of Dynamical Systems. S.I.A.M., Philadelphia, PA (1976)
13. La Salle, J.P., Lefschetz, S.: Stability by Liapunov’s Direct Method, With Applications. Academic
Press, New York, NY (1961)
14. Lefschetz, S.: Differential Equations: Geometric Theory, 2nd edn. Wiley-Interscience, New York,
NY (1963)
15. Leonov, G.A., Kuznetsov, N.V.: Time-varying linearizations and the Perron effects. Int. J. Bifurcat
Chaos 17, 1079–1107 (2007)
16. Lyapunov, A.M.: The General Problem of the Stability of Motion. Taylor and Francis, London (English
translation of the original book published by the Mathematical Society of Kharkov in 1892) (1992)
17. Palis, J., Takens, F.: Hyperbolicity and Sensitive Chaotic Dynamics at Homoclinic Bifurcations.
Cambridge University Press, Cambridge (1993)
18. Winkler, G.: Image Analysis, Random Fields and Markov Chain Monte Carlo Methods: A Mathematical
Introduction. Applications of Mathematics, 2nd edn, vol. 27. Springer, Berlin (2003)

123