Nonlinear Analysis 56 (2004) 643 – 677

www.elsevier.com/locate/na

Necessary conditions for a limit cycle and its

basin of attraction
Peter Giesl
Zentrum Mathematik, TU Munchen, Boltzmannstr. 3,
D-85747 Garching bei Munchen, Germany

Received 22 October 2001; received in revised form 2 January 2003; accepted 8 July 2003

Abstract
In this paper we consider a general di-erential equation of the form ẋ=f(x) with f ∈ C 1 (R n ; R n)
and n ¿ 2. Borg, Hartman, Leonov and others have studied su5cient conditions for the existence,
uniqueness and exponential stability of a periodic orbit and for a set to belong to its basin of
attraction. They used a certain contraction property of the 8ow with respect to the Euclidian or a
Riemannian metric. In this paper we also prove su5cient conditions including upper bounds for
the Floquet exponents of the periodic orbit. Moreover, we show the necessity of these conditions
using Floquet theory and a Lyapunov function.
? 2003 Elsevier Ltd. All rights reserved.

Keywords: Dynamical system; Autonomous ordinary di-erential equation; Periodic orbit; Basin of attraction

1. Introduction

We consider the following autonomous ordinary di-erential equation:

ẋ = f(x);
where f ∈ C 1 (R n ; R n ), and n ¿ 2. We denote the 8ow which maps the initial point x0
at time t = 0 to the solution at time t by St x0 . In [3] we have given the following
su5cient condition for existence, uniqueness and exponentially asymptotic stability of
a periodic orbit and, moreover, for a set K to belong to its basin of attraction. We cite
[3, Corollary 1.4].

E-mail address: [email protected] (P. Giesl).

0362-546X/$ - see front matter ? 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.na.2003.07.020
644 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

Theorem 1. Let ∅ = K ⊂ R n be a compact, connected and positively invariant set

which contains no equilibrium. Moreover assume L(p) ¡ 0 for all p ∈ K, where
L(p) := max L(p; v); (1)

v
=1;v⊥f(p)

L(p; v) := Df(p)v; v: (2)

Then there exists one and only one periodic orbit  ⊂ K, which is exponentially
asymptotically stable. Moreover its basin of attraction A() contains K.

This theorem has been proven by Borg [1] under slightly di-erent assumptions,
because he considered an open bounded set. However, his additional assumptions follow
by the smoothness of f and the compactness of K.
In this paper we deal with the question, whether the conditions of Theorem 1 are
necessary. More precisely, two questions arise, the Grst being the necessity of the
conditions for a limit cycle, the second for the basin of attraction, which is denoted
by A():

• Given an exponentially asymptotically stable periodic orbit , is there a set K, so

that the conditions of Theorem 1 are fulGlled (local necessity)?
• Given an exponentially asymptotically stable periodic orbit , and given a set K̃ ⊂
A(), is there a set K ⊃ K̃, so that the conditions of Theorem 1 are fulGlled for K
(global necessity)?

It is obvious that for a given compact and connected set K̃ ⊂ A() that contains no


equilibrium we can Gnd a set K := t¿0 St K̃ satisfying K̃ ⊂ K ⊂ A() which is
compact, connected, positively invariant and contains no equilibrium (cf. Section 4 for
details). Hence, the question is mainly reduced to whether L(p) is negative for all
p ∈ A() (global necessity) or at least for all p ∈  (local necessity).
We give an example which shows that—in general—not even the local necessity
property holds.

Example 2. We consider the system


ẋ = x(1 − x2 − y2 )(x + 12 ) − y;
(3)
ẏ = y(1 − x2 − y2 )(x + 12 ) + x

Fig. 1, left, shows the unit circle which is a periodic orbit of (3) and an adjacent orbit
(dashed lines). Moreover, the sign and the zero set of L are indicated. The sign of
the function L(p) is negative, if and only if trajectories through adjacent points move
towards each other. In our example L(p) ¡ 0 does not hold for all points p of the
periodic orbit, but still, after one period, the distance of the periodic orbit to adjacent
trajectories has decreased.
This is our motivation to prove the following Theorem 5, which is a generalization of
Theorem 1. In this theorem we do not use the Euclidian scalar product, but a Rieman-
nian metric, i.e., a point-dependent scalar product v; wp = vT M (p)w with a symmetric
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 645

2 2

_ + _
y 1 y 1

+
+ +
–1 –0.5 0.5 1 –2 –1 1 2
x _ x

–1 –1

_ _ +

–2 –2

Fig. 1. Left: The sign of L, the set L = 0 (solid lines), the periodic orbit and an adjacent trajectory (dashed
lines) of (3); right: the sign of LM , the set LM = 0 (solid lines) and the periodic orbit (dashed line) of (3)
with M (x; y) = e4y I .

and positive deGnite matrix M (p) for each p ∈ R n . We call LM the equivalent of the
function L.

Theorem 5. Let M (p) be a Riemannian metric and ∅ = K ⊂ R n be a compact, con-

nected and positively invariant set which contains no equilibrium. Moreover assume
maxp∈K LM (p)= : − ¡ 0, where
LM (p) := max LM (p; v)
vT M (p)v=1;vT M (p)f(p)=0
 
LM (p; v) := vT M (p)Df(p) + 12 M  (p) v;
Here M  (p) denotes the orbital derivative of M (p).
Then there exists one and only one periodic orbit  ⊂ K. This periodic orbit
is exponentially asymptotically stable, and the real parts of all Floquet exponents—
except the trivial one—are less than or equal to −. Moreover, the basin of attraction
A() contains K.

Other authors have generalized Borg’s su5cient conditions in this direction before;
StenstrOom [11] considers a compact set on a Riemannian manifold with a Riemannian
metric and deGnes the function L also for equilibria. He proves that in our case the set K
is isomorphic to a torus. He considers R n as a special case but only with a Riemannian
metric given by a constant matrix. Moreover, as he shows f(p + v); v ¿ 0 for v ⊥
f(p) and  ¿ 0 small by a contradiction argument, he cannot prove the exponential
stability of the periodic orbit and a bound for the Floquet exponents, as we do in
Theorem 5. Hartman and Olech [6] generalize Borgs result without using a Riemannian
metric, but they deGne an integral function along an orbit, which is not suitable for
computation. They show that the periodic orbit has n − 1 Floquet exponents with
negative real parts, but they do not give a bound. They only consider a single orbit
646 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

which has to have a positive distance from the boundary. This result can also be found
in the book [5]. They do not deal with the problem of uniqueness. Leonov gives other
su5cient conditions in a series of papers (cf. also the book [8]): In [7] he also uses a
Riemannian metric and synchronizes the times of two adjacent trajectories St p and S q
perpendicular to f(St p), however, not with respect to the Riemannian metric but with
respect to the Euclidian one. Also, he assumes f ∈ C 2 . In [10] Zhukovskij stability
and instability of solutions, which are not necessarily periodic, are studied by means
of linearization of reparameterized solutions and several su5cient conditions are given.
In [9] solutions of di-erential equations on a Riemannian manifold are considered
and su5cient conditions for their stability using Lyapunov-type functions and singular
values are presented.
There are two special cases of Riemannian metrics that will be considered in more
detail in this paper. Given a transformation of phase space  we can deGne a
Riemannian metric by M = DT D. The function LM is equal to the function L of
the transformed system. Another example is the Riemannian metric M (p) = e2W (p) I
deGned by a scalar weight function W (p). Then v; wp = v; we2W (p) holds, where
:; : denotes the Euclidian scalar product.
For the above Example 2, (3), we choose W (x; y) = 2y. Fig. 1, right, shows the
sign of LM for M (x; y) = e2W (x; y) I . For all points of the periodic orbit the function LM
is negative.
These modiGed conditions of Theorem 5 are in fact necessary in the above sense (cf.
Theorem 19). In two-dimensional systems we can even restrict ourselves to Riemannian
metrics given by a weight function (cf. Proposition 25). These results are summarized
in the following theorem.

Theorem 3. Let f ∈ C 1 (R n ; R n ), where n ¿ 2. Let  be an exponentially asymptot-

ically stable periodic orbit, let A() be its basin of attraction, and let the maximal
real part of the Floquet exponents except the trivial one be − ¡ 0.
◦
Then for all  ¿ 0 and all compact sets K with  ⊂ K ⊂ K ⊂ A() there exists
a Riemannian metric M (p), so that LM (p) 6 −  +  holds for all p ∈ K. If n = 2,
then M (p) can be chosen of the form M (p) = e2W (p) I with a C 1 -function W .

The necessity implies that Theorem 5 is an adequate tool to prove that a certain
set belongs to the basin of attraction of a unique exponentially asymptotically stable
periodic orbit. In fact, if there is such a periodic orbit, there exists a Riemannian metric
which satisGes the conditions of Theorem 5. It is still a serious problem to Gnd such
a Riemannian metric. One can use the proofs of this paper to construct such a metric
algorithmically. This will be done in a forthcoming paper.
Putting together the above theorems, we obtain the following two main results of
this paper.

Theorem 26. Let ẋ = f(x), f ∈ C 1 (R n ; R n ) with n ¿ 2, de<ne a dynamical system.

Then the following two conditions are equivalent.

(i) The system has an exponentially asymptotically stable periodic orbit.

 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 647

(ii) There is a set ∅ = K ⊂ R n and a Riemannian metric M (p), such that K is

compact, connected, positively invariant, contains no equilibrium, and LM (p) ¡ 0
holds for all p ∈ K.

Theorem 27. Let  be an exponentially asymptotically stable periodic orbit of the

system of Theorem 26. Then the following formula for its basin of attraction holds

A() = H;
H ∈A
◦
where A := {H ⊂ R n |  ⊂ H ⊂ H is compact; connected;
positively invariant; contains no equilibrium; and there is

a Riemannian metric M (p) with LM (p) ¡ 0 f or all p ∈ H }:

Let us describe how the paper is organized: In Section 2 we prove Theorem 5 which
gives a su5cient condition for the existence of a limit cycle and for a set to belong to
its basin of attraction. This condition makes use of a Riemannian metric, and we give
two examples of Riemannian metrics. In the third section we prove that the conditions
of Theorem 5 are necessary. We show both local and global necessity and deal with
the two-dimensional case separately. In the fourth section we summarize our results
and in an appendix we show the continuity of LM and a lemma on a normal form of
a matrix.

2. Sucient condition with a Riemannian metric

In the Grst subsection we state and prove Theorem 5. In the second subsection we
give examples of Riemannian metrics.
Throughout the paper we consider a dynamical system given by the autonomous
ordinary di-erential equation ẋ=f(x), where f ∈ C 1 (R n ; R n ) and n ¿ 2. A Riemannian
metric on a manifold is a point-dependent scalar product. We consider R n as the
manifold and can therefore deGne a Riemannian metric by a matrix-valued function.

De"nition 4. We call M : R n → R n×n a Riemannian metric, if M is a C 1 -function and

M (p) is a symmetric and positive deGnite matrix for all p. Thus, vT M (p)w = v; wp
is a scalar product for each p.

2.1. Proof of Theorem 5

Now we state Theorem 5 which is a more general version of Theorem 1. This

theorem is similar to the result of [11], but it also provides precise bounds for the
Floquet exponents.

Theorem 5 (Su5ciency): Let M (p) be a Riemannian metric and ∅ = K ⊂ R n be

a compact, connected and positively invariant set which contains no equilibrium.
648 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

Moreover assume maxp∈K LM (p)= : − ¡ 0, which exists since LM is continuous

(cf. Proposition 28) and K is compact, where
LM (p) := max LM (p; v); (4)
vT M (p)v=1;vT M (p)f(p)=0
 
LM (p; v) := vT M (p)Df(p) + 12 M  (p) v; (5)
n @Mij (p)
Here M  (p) denotes the matrix with entries aij = l=1 @pl fl (p)
which is also the
d
orbital derivative of M (p), i.e., M  (p) = ( d. )M (S. p)|.=0 .
Then there exists one and only one periodic orbit  ⊂ K. This periodic orbit
is exponentially asymptotically stable, and the real parts of all Floquet exponents—
except the trivial one—are less than or equal to −. Moreover, the basin of attraction
A() contains K.

Remark 6. LM (p) is a continuous function with respect to p as we prove in

Proposition 28.

The proof of Theorem 5 proceeds in the following steps and related propositions:
(1) DeGne a time-dependent distance by means of the Riemannian metric between two
trajectories with close starting points and prove that this distance is exponentially
decreasing (Proposition 7)
(2) Show that the !-limit set of all points of a neighborhood is the same
(Proposition 8)
(3) Show that the !-limit set of all points of K is the same (Proposition 11)
(4) Show that this !-limit set is an exponentially asymptotically stable periodic orbit
(use Proposition B.1 of [3])
In Proposition 7 we deGne a distance function by measuring the distance between S. p
and ST (p + 1) with respect to the Riemannian metric M (p), where T = Tpp+1 (.) is
a synchronized time. In Proposition 7 we only consider initial values p + 1 of the
hyperplane 1 ∈ (M (p)f(p))⊥ , whereas in Proposition 8 we extend our results to all
points q of a full neighborhood of p.

Proposition 7. Let the assumptions of Theorem 5 be satis<ed.

Then for each k ∈ (0; 1) there are constants  ¿ 0 and C ¿ 1 such that for all p ∈ K
and for all 1 ∈ R n with 1T M (p)f(p) = 0 and 1 6 =2 there exists a di>eomorphism
Tpp+1 : R0+ → R0+ which satis<es 23 6 Ṫ p+1 4
p (.) 6 3 and

(STpp+1 (.) (p + 1) − S. p)T M (S. p)f(S. p) = 0

for all . ¿ 0. Tpp+1 depends continuously on 1. Moreover, we have

STpp+1 (.) (p + 1) − S. p 6 Ce−(1−k). 1 for all . ¿ 0 (6)
and !(p) = !(p + 1).

Proof. The proposition is proven in four steps. We will introduce a time-dependent

distance between two trajectories with adjacent starting points p and p + 1 where
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 649

1T M (p)f(p) = 0, and we show that this distance decreases exponentially. We deGne

the distance with respect to the Riemannian metric, but note that on the compact set K
the Riemannian and the Euclidian norms are equivalent. In the Grst step we give some
bounds, in the second step we parameterize the time of the orbit ST (p + 1) and use
this to deGne the distance A(.). In the third step we prove that this distance decreases
exponentially and in the last step we show that each point p ∈ K has a neighborhood
in p + (M (p)f(p))⊥ , the points of which have the same !-limit set as p.
(I) Since f and M  (p) are continuous functions, we can derive uniform bounds on
the compact set K. Thus,
0 ¡ fm 6 f(p) 6 fM (7)
and
M  (p) 6 MD (8)
hold for all p ∈ K. M (p) is symmetric and positive deGnite for all p ∈ K. Hence, for
the smallest eigenvalue 41 (p) ¿ 0 holds. Since the eigenvalues depend continuously
on p, there are 0 ¡ 4m 6 4M ¡ ∞ such that
4m 52 6 5T M (p) 5 6 4M 52 (9)

M (p)5 6 4M 5 (10)

n
hold for all 5 ∈ R and all p ∈ K. We set
 
k 2 4m 1
 := min ; k 6 ; (11)
3 3 MD 3
where  := −maxp∈K LM (p) ¿ 0. Also, Df is continuous and thus uniformly continu-
ous on K. Hence, there exists a 1 ¿ 0, so that (12) holds for all p ∈ K and all 5 ∈ R n
with 5 6 1
k4m
Df(p) − Df(p + 5) 6 : (12)
3(1 + )4M
Finally, there is a positive constant fD , such that
Df(q) 6 fD holds for all q ∈ K1 : (13)
We set
 
4m fm2 
 := min 1 ; (14)
fM [MD + (2 + )fD 4M ]
and

4m 
 := : (15)
4M
(II) Fix p ∈ K and 1 ∈ R n with 1T M (p)f(p) = 0 and 1 6 =2. We synchronize
the time of the trajectories through p and p + 1 by deGning Tpp+1 (.) implicitly by
Q(T; .; 1) := (ST (p + 1) − S. p)T M (S. p)f(S. p) = 0: (16)
Q(0; 0; 1) = 0 implies that Tpp+1 (0)
= 0. Since, as we show later, @T Q(0; 0; 1) = 0, the
function Tpp+1 (.) is deGned by (16) locally near . = 0 and depends continuously on 1
650 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

by the implicit function theorem. We will later show by a prolongation argument that,
in fact, Tpp+1 is deGned for all times . ¿ 0. From now on we write T = Tpp+1 . As long
as T (.) is deGned, we set
 +
 R0 → R0+ ;
A: : (17)
 . → (ST (.) (p + 1) − S. p)T M (S. p)(ST (.) (p + 1) − S. p)

A(0) = 0 implies that A(.) = 0 for all . ¿ 0. In this case set v(.) := (ST (.) (p + 1) −
S. p)=A(.). Hence, we have
ST (.) (p + 1) − S. p = A(.)v(.):
 √ √
v(.) is a vector with v(.)T M (S. p)v(.) = 1, and thus 1= 4M 6 v(.) 6 1= 4m
by (9). We have v(.)T M (S. p)f(S. p) = 0 for each . ¿ 0 by (16). We calculate the
derivative Ṫ (.)=−@. Q(T; .; 1)=@T Q(T; .; 1) by the implicit function theorem. We have,
cf. (16)
@. Q(T; .; 1) = −f(S. p)T M (S. p)f(S. p)

+ (ST (p + 1) − S. p)T M  (S. p)f(S. p)

+ (ST (p + 1) − S. p)T M (S. p)Df(S. p)f(S. p)

= − f(S. p)T M (S. p)f(S. p) + A(.)v(.)T M  (S. p)f(S. p)

+ A(.)v(.)T M (S. p)Df(S. p)f(S. p) (18)

and by the mean value theorem
@T Q(T; .; 1) = f(ST (p + 1))T M (S. p)f(S. p)

= f(S. p)T M (S. p)f(S. p)

 T
1
+ A(.) Df(S. p + 4A(.)v(.)) d4 v(.)
0

×M (S. p)f(S. p): (19)

√ √ 
We have A(0) 6 4M 1 6 4m  =2 because √ of (9), 1 6 =2 and (15). Hence, the
continuous function A satisGes A(.) 6 4m  for . small enough. We will show in
the next step that, however, this inequality holds for all . ¿ 0. √
Since K is positively invariant, S. p+4A(.)v(.) ∈ K
, supposing that A(.) 6 4m 
1
and 4 ∈ [0; 1]. We conclude  0 Df(S. p + 4A(.)v(.)) d4 6 fD using (13) and (14).
Eqs. (7)–(10) and (13) imply with (18) and (19) that
f(S. p)T M (S. p)f(S. p) +  fM (MD + fD 4M )
Ṫ (.) 6
f(S. p)T M (S. p)f(S. p) −  fM fD 4M
 fM (MD + 2fD 4M )
61+
4m fm2 −  fM fD 4M

= 1 +  6 43 by (14):
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 651

Similarly, we can conclude that Ṫ (.) ¿ 1− ¿ 23 , and in particular that @T Q(0; 0; 1) = 0
holds. This shows that T (.) is a √ strictly increasing function. The inverse map .(T )
satisGes 34 6 .̇(T ) 6 32 . If A(.) 6 4m  holds for all . ¿ 0, we can deGne T (.) by
a prolongation argument.
(III) We show that A(.) is monotonously decreasing and tends to zero exponentially.
This implies that we can deGne T (.) for all . ¿ 0. We calculate the temporal derivative
of A2 (cf. (17)) and use M (p) = M (p)T and v(.)T M (S. p)f(S. p) = 0.
 
d 2 T dT
A (.) = 2(ST (.) (p + 1) − S. p) M (S. p) f(ST (.) (p + 1)) (.) − f(S. p)
d. d.

+ (ST (.) (p + 1) − S. p)T M  (S. p)(ST (.) (p + 1) − S. p)

= 2A(.)Ṫ (.) v(.)T M (S. p)f(S. p + A(.)v(.))

+ A2 (.) v(.)T M  (S. p)v(.):

As 4A(.)v(.) 6  for 4 ∈ [0; 1] and small . we can use (12). The mean value
theorem yields with v(.)T M (S. p)f(S. p) = 0, LM (S. p) 6 −  and (12)
 
1
d 2
A (.) = 2A2 (.)Ṫ (.) v(.)T M (S. p) Df(S. p + 4A(.)v(.)) d4 v(.)
d. 0

+ A2 (.)Ṫ (.) v(.)T M  (S. p)v(.)

+ A2 (.)(1 − Ṫ (.)) v(.)T M  (S. p)v(.)





  
2 T 1 
6 2A (.)Ṫ (.) v(.) M (S. p)Df(S. p) + M (S. p) v(.)

 2

  
=LM (S. p)

 


1 
+ v(.)T M (S. p) [Df(S. p + 4A(.)v(.)) − Df(S. p)] d4v(.)
0 




+ A2 (.)|1 − Ṫ (.)| · |v(.)T M  (S. p) v(.)|

A2 (.)Ṫ (.)4M k4m MD

6 −2A2 (.)Ṫ (.) + 2 + A2 (.)|1 − Ṫ (.)|
4m 3(1 + )4M 4m
 
2 MD
6 −2 + 2 + k +  A2 (.)
3 4m

6 −2(1 − k)A2 (.) by (11):

652 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

Thus we have
 
A(.) 6 A(0) e−(1−k). 6 4m e−(1−k). : (20)
2
√
This proves in particular A(.) 6 A(0) 6 4m  =2 for all . ¿ 0 and thus that both T (.)
and A(.) are deGned for all . ¿ 0 by a prolongation argument. Eq. (20) also shows

4m STpp+1 (.) (p + 1) − S. p 6 A(.)

6 A(0)e−(1−k).

6 4M 1e−(1−k). :

Hence, (6) follows with C := 4M =4m ¿ 1.
(IV) Now we show that all points p + 1 with 1 as above have the same !-limit set
as p itself. Assume w ∈ !(p). Then we have a strictly increasing sequence .n → ∞
satisfying w − S.n p → 0 as n → ∞. Because of (6) and Ṫ ¿ 23 the sequence T (.n )
satisGes T (.n ) → ∞ and ST (.n ) (p + 1) − S.n p 6 Ce−(1−k).n → 0 as n → ∞. This
proves ST (.n ) (p + 1) → w and w ∈ !(p + 1). The inclusion !(p + 1) ⊂ !(p) follows
similarly. Thus we have shown Proposition 7.

Proposition 7 proves that all points of a neighborhood of p in the hyperplane p +

(M (p)f(p))⊥ have the same !-limit set. In Proposition 8 we show this property for
all points of a full neighborhood of p.
Here we cannot follow the proof of [3, Proposition 2.2]. Instead, we use the con-
struction of [3, Appendix B], where we deGned a new coordinate system around p. In
[3, Proposition 2.2], we proved the existence of a t0 such that q − St0 p; f(St0 p) = 0.
Now we will prove the existence of a t0 such that (St0 q − p)T M (p)f(p) = 0.

Proposition 8. Let the assumptions of Theorem 5 be satis<ed.

Then there is a constant ∗ ¿ 0 such that !(p) = !(q) holds for all p ∈ K and
p − q 6 ∗ .

Proof. This proposition is proven in three steps. For a given point q in the neigh-
borhood of p we Gnd a point St0 q= : q , such that we can write q = p + 1 with
1T M (p)f(p) = 0. Then the statement follows by Proposition 7.
(I) New coordinates. Choose a point p ∈ K and Gx k ∈ (0; 1). We will deGne a
new coordinate system centered in p. Set F(p) := M (p)f(p)=M (p)f(p), then we
have F(p) = 1. We deGne y as the amount in F(p)-direction and x as the vectorial
amount perpendicular to F(p). The hyperplane H := p + F(p)⊥ consists precisely of
the points q, where y(q) = 0. Set 8 := F(p); f(p). We have 8M := fM ¿ 8 ¿ 8m :=
4m fm =4M ¿ 0 for all p ∈ K using the bounds of I. of the proof of Proposition 7 with
the above k. We deGne for arbitrary q ∈ R n
y(q) := q − p; F(p) ∈ R ; (21)

x(q) := q − p − y(q)F(p) ∈ F(p)⊥ : (22)

 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 653

Then we can express q = p + y(q) F(p) + x(q) and q − p2 = |y(q)|2 + x(q)2 . We
also write the vectors f(q) in the new coordinates. We set
4(q) := f(q) − f(p); F(p) ∈ R ; (23)

w(q) := f(q) − f(p) − 4(q)F(p) ∈ F(p)⊥ : (24)

Thus we can write
f(q) = f(p) + 4(q) F(p) + w(q): (25)
As f is uniformly continuous in p ∈ K, there is a 0 ¡ 2 6  such that for all p ∈ K
and all q ∈ B2 (p)
8 
m
f(q) − f(p) 6 min ; fM holds: (26)
2
Hence, the following inequalities hold for all q ∈ B2 (p) by multiplying (25) by F(q)
and w(q), respectively.
8m
|4(q)| 6 f(q) − f(p) 6 ; (27)
2
w(q) 6 f(q) − f(p) 6 fM : (28)
Inside the ball B2 (p) orbits can only move within a cone. This is shown in the next
lemma.

Lemma 9. Let St q ∈ B2 (p) for all t ∈ [0; ] ˜ (t ∈ [ − ;

˜ 0]) with a positive constant .
˜
Then for all 1 ; 2 with 0 6 1 6 2 6 ˜ (−˜ 6 1 6 2 6 0) the following inequal-
ities hold:
1 d 1
8m 6 y(St q) 6 8m + 8M ; (29)
2 dt 2  
1 1
8m (2 − 1 ) 6 y(S2 q) − y(S1 q) 6 8m + 8M (2 − 1 ); (30)
2 2
x(S2 q) − x(S1 q) 6 k0 (y(S2 q) − y(S1 q)); (31)
where k0 := 4 f8mM .

Proof. We only prove the statement for positive times—for negative times it is shown
in the same way. Writing St q = p + y(St q) F(p) + x(St q) we conclude
d
f(St q) = St q
dt
d d
= y(St q) · F(p) + x(St q): (32)
dt dt
As x(St q) ⊥ F(p) for all t ∈ [0; ˜ ], we have ( dtd )x(St q) ⊥ F(p) for all t ∈ [0; ˜ ], too.
Multiplying (32) by F(p) yields because of (25)
d
8 + 4(St q) = y(St q):
dt

Now we can conclude (29) by (27). Since 12 ( dtd )y(St q) dt = y(S2 q) − y(S1 q), (30)
follows immediately.
654 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

Multiplying (32) by ( dtd )x(St q) ⊥ F(p) we get

 2 !
d 
 x (St q) = f(St q); d x(St q)
 dt  dt
!
d
= f(p) + w(St q); x(St q) by (25); i:e:;
dt
 
d 
 x(St q) 6 f(p) + w(St q): (33)
 dt 

Hence,
 
 2
d 
x(S2 q) − x(S1 q) = 
 x(St q) dt 

1 dt
 2  
d 
6  
 dt x(St q) dt
1
 2
6 (f(p) + w(St q)) dt by (33)
1

6 2(2 − 1 )fM by (28)

6 k0 (y(S2 q) − y(S1 q)) by (30):

This proves (31) and the lemma.

(II) Intersection with the hyperplane H . Orbits through points in a small ball around
p intersect the hyperplane H := p + F(p)⊥ . This enables us to deGne the time t0 (q)
such that St0 (q) q ∈ H .

Lemma 10. Set 0 := 2fM + 12 8m and 3 := 12 (20 =82m )+1 .

Then for all q ∈ B3 (p) there is a t0 (q) ∈ R such that

(St0 (q) q − p)T M (p)f(p) = 0; (34)

St0 (q) q − p 6 (k0 + 1)q − p: (35)

Proof. We have to show that for q ∈ B3 (p) the continuous function y(S q) vanishes
for some  = t0 . We prove this using Lemma 9 and show afterwards that t0 is close
enough to zero so that S q remains in B2 (p) for all  between 0 and t0 .
We only consider the case y(q) 6 0—the case y(q) ¿ 0 corresponds with the parts
in brackets of Lemma 9 and is shown similarly. As long as S q ∈ B2 (p), (30) implies
that y(S q) ¿ y(q) + (=2)8m . For ˜ = −(2=8m )y(q) ¿ 0 we have y(S˜q) ¿ 0. By the
intermediate value theorem there exists a t0 ∈ [0; ˜ ] satisfying y(St0 q) = 0. As y(S q)
is strictly monotonously increasing with respect to  by Lemma 9, t0 is unique and we
can deGne t0 (q) by y(St0 (q) q) = 0, i.e. (34) is shown.
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 655

Now we check that S q remains in B2 (p) for all  ∈ [0; ˜ ]. Assuming the opposite,
˜ with S0 q − p = 2 and S q − p ¡ 2 for all  ∈ [0; 0 ). By
there is a 0 ∈ [0; ]
(25), (27) and (28) we have f(q) 6 2fM + 12 8m = 0 for all q ∈ B2 (p). This yields
 0 
 
2 = S0 q − p 6  f(S  q) d  + q − p

0
 
20 2
6 ||
˜ 0 + 3 6 3 +1 = ;
8m 2
which is a contradiction. Now we prove (35). We have
St0 (q) q − p = x(St0 (q) q)

6 x(St0 (q) q) − x(q) + x(q)

6 k0 |y(q)| + x(q) by (31)

6 (k0 + 1)q − p:

This completes the proof of Lemma 10.

(III) Set ∗ := min(3 ; =(2(k0 + 1))), having deGned  in Proposition 7. Choose a

q with p − q 6 ∗ . By Lemma 4 there exists a t0 = t0 (q) such that St0 q = p + 1,
where St0 q − p = 1 ⊥ M (p)f(p) and 1 6 (k0 + 1)∗ 6 2 . By Proposition 7 we have
!(St0 q) = !(p) and thus !(q) = !(p).

Using the fact that K is connected we can now prove Proposition 11 which shows
that all points of K have the same !-limit set.

Proposition 11. Let the assumptions of Theorem 5 be satis<ed. Then ∅ = !(p) =

!(q)= :  ⊂ K for all p; q ∈ K.

Proof. Let p0 ∈ K. Since for all . ¿ 0 we have S. p0 ⊂ K, which is a compact set,

∅ = !(p0 )= :  ⊂ K.
Now consider an arbitrary point p ∈ K. By Proposition 8 we have !(p) = !(q)
for all q in a neighborhood of p. Hence, V1 := {p ∈ K|!(p) = !(p0 )} and V2 :=
{p ∈ K|!(p) = !(p0 )} are open sets. Since K = V1 ∪˙ V2 , p0 ∈ V1 and K is connected,
V2 must be empty and V1 = K.

To complete the proof of Theorem 5, we must show that  is an exponentially

asymptotically stable periodic orbit and that − provides an upper bound for the Floquet
exponents.

Proof of Theorem 5. Choose a point p0 ∈ . By Proposition 11 we have p0 ∈  =

!(p0 ). We apply Proposition B.1 of [3] with C of Proposition 7 and g(.) := M (S. p)f
(S. p)=M (S. p)f(S. p), so that g(.); f(S. p) ¿ 4m fm =4M ¿ 0. Hence,  is an ex-
ponentially asymptotically stable periodic orbit and by Proposition 11 !(q) =  for all
q ∈ K. Since  is asymptotically stable, q ∈ A() follows for all q ∈ K.
656 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

Finally, we must show that − is an upper bound for the real parts of all but
the trivial Floquet exponent. By [5, Lemma 10.2] we can also calculate the Floquet
exponents by regarding a PoincarTe map. Hartman proves this only for the PoincarTe
section perpendicular to f, but with the remark before [5, Lemma 10.2] one sees that
the same is true for all transversal sections.
Fix a point p ∈  and choose the PoincarTe map < which corresponds to the hyper-
plane H = {p + v|vT M (p)f(p) = 0} such that <v := STpp+v (T ) (p + v) − ST p with the
notations of Proposition 7, where T is the period of  and v is small enough. Assume
there was a real Floquet exponent −0 with −0 ¿ −  besides the Floquet exponent
0. We will deal with the case of a complex Floquet exponent later. Then, for k ¿ 0
small enough there is an l ¿ 0 such that
√
e−(1−k)T = 1 − l e−0 T :

By the above argumentation e−0 T is an eigenvalue of the linearized PoincarTe map

D<. We assume v0 to be a corresponding eigenvector such that D< v0 = e−0 T v0 and
v0  = 1. Then for any  ¿ 0

<(v0 ) = e−0 T v0 + k() holds;

where k() is a vector with lim→0 k()= = 0 by the Taylor approximation. For all
vectors v = v0 with  6 =2 we have
√ 
1 − l e−0 T vT M (p)v

= e−(1−k)T vT M (p)v

¿ (<v)T M (p)<v by (20) of Proposition 7

= e−20 T vT M (p)v + 2e−0 T vT M (p)k() + k()T M (p)k()

¿ e−20 T vT M (p)v + 2e−0 T vT M (p)k():

Dividing both sides by e−0 T vT M (p)v we get

√ vT M (p)k()
1 − l ¿ 1 + 2e0 T T ;
v M (p)v

lvT M (p)v 6 − 2e0 T vT M (p)k(): (36)

By (9) we have with (36)

4m lv2 6 lvT M (p)v 6 2e0 T |vT M (p)k()| 6 2e0 T 4M k() v:

Hence,
4m k() k()
l 6 2e0 T = 2e0 T :
4M v 
For  → 0 the right-hand side tends to zero and thus we have found a contradiction.
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 657

Now consider the case where −0 ¿ −  and −0 + i=0 is a Floquet exponent. Let
v0 be a complex eigenvector, i.e., D< v0 = e(−0 +i=0 )T v. We assume v0  = 1. Then vU0
is an eigenvector with corresponding eigenvalue e(−0 −i=0 )T and we have for the real
vectors v10 := (v + v)=2
U and v20 := (v − v)=(2i)
U
D<v10 = e−0 T (cos (=0 T )v10 − sin (=0 T )v20 );

D<v20 = e−0 T (sin (=0 T )v10 + cos (=0 T )v20 ):

Then we have for v1 = v10 and v2 = v20
<v1 = e−0 T (cos (=0 T )v1 − sin (=0 T )v2 ) + k1 ()
with lim→0 k1 ()= = 0. An analogous statement holds for v2 . As in the real case, we
have with R1 (v1 ; v2 ) := 2e0 T (cos (=0 T )v1 − sin (=0 T )v2 )T M (p)k1 () and R2 (v1 ; v2 ) :=
2e0 T (sin (=0 T )v1 + cos (=0 T )v2 )T M (p)k2 ()
√
1 − l e−0 T v1T M (p)v1

¿ [e−20 T (cos (=0 T )v1 − sin (=0 T )v2 )T M (p)(cos (=0 T )v1 − sin (=0 T )v2 )
+ e−20 T R1 (v1 ; v2 )]1=2

(1 − l)v1T M (p)v1 ¿ cos2 (=0 T )v1T M (p)v1 + sin2 (=0 T )v2T M (p)v2

− 2 cos (=0 T ) sin (=0 T )v1T M (p)v2 + R1 (v1 ; v2 )

−lv1T M (p)v1 ¿ sin2 (=0 T )(−v1T M (p)v1 + v2T M (p)v2 )

− 2 cos (=0 T ) sin (=0 T )v1T M (p)v2 + R1 (v1 ; v2 ):
Analogously we get
−lv2T M (p)v2 ¿ sin2 (=0 T )(v1T M (p)v1 − v2T M (p)v2 )
+ 2 cos (=0 T ) sin (=0 T )v1T M (p)v2 + R2 (v1 ; v2 )
and hence, by adding both inequalities,
4m l(v1 2 + v2 2 ) 6 l[v1T M (p)v1 + v2T M (p)v2 ]
6 |R1 (v1 ; v2 )| + |R2 (v1 ; v2 )|:
Thus, we have

4m 2e0 T k1 ()
l6 0 2 |cos (=0 T )v10 − sin (=0 T )v20 |
4M v1  + v20 2 

k2 ()
+ |sin(=0 T )v10 + cos (=0 T )v20 | :

658 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

The right-hand part tends to zero as  → 0 and thus we have a contradiction. This
shows that the largest real part of all Floquet exponents except the trivial one is less
than or equal to −.

2.2. Examples of Riemannian metrics

We discuss two examples of Riemannian metrics. The Grst one is a di-eomorphism

of the phase space. The second is given by a scalar-valued function, a so-called weight
function.

2.2.1. Transformations of the phase space

Let  : K → K  be a C 2 -di-eomorphism of the phase space with K; K  ⊂ R n . Then
M (p) := D(p)T D(p) deGnes a Riemannian metric. Thus we have the following
corollary of Theorem 5.

Corollary 12 (Su5ciency, transformation). Let ∅ = K ⊂ R n be a compact, connected

and positively invariant set which contains no equilibrium. Let  : K → K  ⊂ R n be
a C 2 -di>eomorphism such that L (p) 6 −  ¡ 0 holds for all p ∈ K, where
L (p) := max L (p; w)
wT D(p)−1 D(p)−T w=1;w⊥f(p)

L (p; w) := wT Df(p)D(p)−1 D(p)−T w

n
"
+ (D(p)−T w)i f(p)T Hi (p)D(p)−1 D(p)−T w
i=1
 2

@ i
with Hi (p) := (p): (37)
@pk @pj kj

We use the notation D(p)−T := [D(p)−1 ]T .

De<ning M (p) := D(p)T D(p) we have the formula
LM (p) = L (p): (38)
Then there exists one and only one periodic orbit  ⊂ K. This periodic orbit is
exponentially asymptotically stable, and the real parts of all Floquet exponents—
except the trivial one—are less than or equal to −. Moreover, the basin of attraction
A() contains K.

Proof. Set M (p) := D(p)T D(p). Since  is a C 2 -di-eomorphism, D has full

rank. Furthermore, M (p) is a symmetric positive deGnite matrix for each p, and M (p)
is C 1 with respect to p. We will show that LM (p) = L (p) holds (cf. (38)). Then the
corollary follows by Theorem 5.
Set w := M (p)v. Then vT M (p)v = 1 ⇐ wT M (p)−1 w = 1 and vT M (p)f(p) = 0 ⇐
T
w f(p) = 0. Now we show LM (p; v) = L (p; w).
 
vT M (p)Df(p) + 12 M  (p) v = wT Df(p)M (p)−1 w + 12 vT M  (p)v:
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 659

Since
n
" @i @i
vT M (p)v = vj (p) (p)vl ;
@pj @pl
i; j;l=1

the second term is

n  2 
1 T  1 " @ i @i @ 2 i @i
v M (p)v = vj (p) (p) + (p) (p) vl fk (p)
2 2 @pk @pj @pl @pk @pl @pj
i; j; k;l=1

n
" @i @2  i
= (p)vl fk (p) (p)vj
@pl @pk @pj
i; j; k;l=1

n
"
= (D(p)v)i f(p)T Hi (p)v:
i=1

With v = M (p)−1 w = D(p)−1 D(p)−T w the corollary is shown.

The function L can be interpreted as the corresponding function L of the trans-

formed system with the Euclidean metric.

Lemma 13. Let  : K → K  ⊂ R n be a C 2 -di>eomorphism. Denote the solution of the

transformed system by y(t) = (x(t)), where x(t) is a solution of the original system.
For the transformed system

ẏ = D(−1 (y))f(−1 (y)) = : g(y) holds:

The conditions of Corollary 12 hold with K and L , if and only if the conditions
of Theorem 1 hold for the transformed system ẏ = g(y) with K  and L (y) :=
max
u
=1; u⊥g(y) uT Dg(y)u. Moreover, L (x) = L ((x)).

Proof. Since  is a di-eomorphism, K is compact, connected and positively invariant,

if and only if K  is. g(y) = 0, if and only if f(−1 (y)) = 0, since D(−1 (y)) has
full rank.
Now we calculate L (y; u). We have with x = −1 (y)

L (y; u) = uT Dg(y)u
n
"
= uT D(x)Df(x)D(x)−1 u + ui f(x)T Hi (x)D(x)−1 u:
i=1

The map w → u := D(x)−T w is a bijection from f(x)⊥ to g(y)⊥ . Thus

L (y) = max L (y; u)

u
=1;u⊥g(y)
660 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

= max L (y; D(x)−T w)

wT D(x)−1 D(x)−T w=1;w⊥f(x)

= max L (x; w):

wT D(x)−1 D(x)−T w=1;w⊥f(x)

Hence, L (y) = L (x).

Both local and global necessity can be obtained through Riemannian metrics given
by transformations, if we assume f ∈ C 2 (R n ; R n ) as we will show in the third section
in Proposition 22.

2.2.2. Weight function

Let W : K → R be a C 1 -function. Then M (p) := e2W (p) I deGnes a Riemannian
metric and Corollary 14 follows by Theorem 5.

Corollary 14 (Su5ciency, weight function). Let ∅ = K ⊂ R n be a compact, con-

nected and positively invariant set which contains no equilibrium. Let W : K → R be
a C 1 -function. Moreover, assume that
max [L(p) + W  (p)] = : − ¡ 0;
p∈K

where W  (p) := ∇W (p); f(p) denotes the orbital derivative. For the de<nition
of L cf. (1).
De<ning M (p) := e2W (p) I we have the formula
LM (p) = L(p) + W  (p): (39)
Then there exists one and only one periodic orbit  ⊂ K. This periodic orbit is
exponentially asymptotically stable, and the real parts of all Floquet exponents—
except the trivial one—are less than or equal to −. Moreover, the basin of attraction
A() contains K.

Proof. DeGne M (p) := e2W (p) I . We have with (5) and (4)
LM (p; v) = e2W (p) vT Df(p)v + 12 2∇W (p); f(p)e2W (p) v2

and LM (p) = max LM (p; v):

v
=e−W (p) ;v⊥f(p)

Setting w := eW (p) v, we have LM (p) = L(p) + W  (p). Hence, the assumptions of

Theorem 5 are satisGed and this shows the corollary.

We will show later that in two-dimensional systems local and global necessity can
be obtained by Riemannian metrics given by a weight function. In higher dimen-
sions, however, local necessity cannot be obtained with a Riemannian metric given
by a weight function. The reason is that for two-dimensional systems there is only a
one-dimensional family of vectors v ⊥ f(p), which is not the case in higher dimen-
sions. Let us consider a three-dimensional example to clarify this.
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 661

1
–0.09
t
1 2 3 4 5 6
0 –0.095

–0.1
–1

–0.105

–2
–0.11

–3 0 1 2 3 4 5 6
t

Fig. 2. Left: The functions L((cos t; sin t; 0); v1 (t)) and L((cos t; sin t; 0); v2 (t)) (dashed line) for the system
(40)—for each t at least one of them is positive; right: The function (cos t + 0:55)[(cos t − 0:7)2 +
0:1] − sin t Wx (cos t; sin t) + cos t Wy (cos t; sin t), which is negative for all t ∈ [0; 2<].

Example 15. Consider the equations

 # $

 ẋ = x(1 − x2 − y2 ) x + 12 − y;
 # $
ẏ = y(1 − x2 − y2 ) x + 12 + x;



z˙ = z(x + 0:55)[(x − 0:7)2 + 0:1]: (40)
This is an extended version of Example 2, cf. (3). A periodic orbit is given by  =
{(x; y; 0)|x2 +y2 =1}. More precisely, the periodic solution is given by (cos t; sin t; 0), the
minimal period is 2< and f(cos t; sin t; 0) = (−sin t; cos t; 0). Hence, the vectors v1 (t) =
(cos t; sin t; 0) and v2 (t) = (0; 0; 1) satisfy vi (t) = 1 and vi (t) ⊥ f(cos t; sin t; 0). Fig. 2,
left, shows the values of the functions L((cos t; sin t; 0); v1 (t)) and L((cos t; sin t; 0); v2 (t)).
Since for all t ∈ [0; 2<] at least one of them is positive, we have L(p) ¿ 0 for all
points p ∈ . Now assume there was a weight function W :  → R such that L(p) +
W  (p) ¡ 0 for all p ∈ . Let p0 ∈ . Then in particular S2< p0 = p0 . We have 0 =
 2<
W (S2< p0 ) − W (p0 ) = 0 W  (St p0 ) dt. Hence, by assumption
 2<  2<
0¿ [L(St p0 ) + W  (St p0 )] dt = L(St p0 ) dt ¿ 0;
0 0

which is a contradiction. Thus, there is no weight function.

But the periodic orbit is exponentially asymptotically stable. To show this, we use
Theorem 5 for K = . We deGne a Riemannian metric by
 4y 
e 0 0
 
M (x; y; z) = 
0 e4y 0 ;

2W (x; y)
0 0 e
662 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

where W (x; y) := − 13 x2 y− 12 y+ 17 40 xy. We calculate the matrix M̃ (x; y; 0)=M (x; y; 0)Df
1
(x; y; 0) + 2 M (x; y; 0) for points with x2 + y2 = 1, i.e. (x; y; 0) ∈ ,


         
4y 2 1 4y 1
 e −2x x + + 2x e −2xy x + − 1 0 
 2 2 
         
M̃ (x; y; 0) =  e4y −2xy x + 1 4y 2 1 

 +1 e −2y x + + 2x 0 
 2 2 
0 0 c(x; y)

with c(x; y) := [(x + 0:55)[(x − 0:7)2 + 0:1] − yWx (x; y) + xWy (x; y)]e2W (x; y) . We
parameterize the periodic orbit by x = cos t and y = sin t. The vectors v ⊥ M (cos t;
sin
√ t; 0)f(cos t; sin t; 0) with vT M (cos t; sin t; 0)v=1 are given by v(t)=±(ke−2 sin t v1 (t)+
1 − k 2 e−W (cos t; sin t) v2 (t)) with k ∈ [−1; 1]. It su5ces to show that LiM (cos t; sin t; 0) :=
LM ((cos t; sin t; 0); vi (t)) ¡ 0 for i = 1; 2 and for t ∈ [0; 2<], because the structure of M̃
yields

LM ((cos t; sin t; 0); v(t))

= v(t)T M̃ (cos t; sin t; 0)v(t)

= k 2 e−4 sin t L1M (cos t; sin t; 0) + (1 − k 2 )e−2W (cos t; sin t) L2M (cos t; sin t; 0)

¡ 0;

if L1M (cos t; sin t; 0) ¡ 0 and L2M (cos t; sin t; 0) ¡ 0. We have

L1M (cos t; sin t; 0) = −e4 sin t ¡ 0

and L2M (cos t; sin t; 0) = c(cos t; sin t; 0)

= e2W (cos t; sin t) [(cos t + 0:55)[(cos t − 0:7)2 + 0:1]

− sin t Wx (cos t; sin t) + cos t Wy (cos t; sin t)]

¡ 0 (cf : Fig: 2; right):

3. Necessity

This section deals with the question whether the assumptions of Theorem 5 are
necessary. In the Grst subsection we show the existence of a Riemannian metric M ,
so that LM is strictly negative in a neighborhood of  (local necessity). In the second
subsection we show that for each given compact set K ⊂ A() there is a Riemannian
metric M , so that LM is strictly negative in K (global necessity). In the third subsection
we consider the two-dimensional case where we can obtain the same results with a
Riemannian metric deGned by a weight function.
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 663

3.1. Local necessity

We assume  to be an exponentially asymptotically stable periodic orbit. In order to

prove the necessity of the assumptions of Theorem 5 we construct a Riemannian metric
on a given exponentially asymptotically stable periodic orbit using Floquet theory.
Since f(S. p) is a solution of the Grst variation equation, 0 is a Floquet exponent, let
us say 81 . The exponential asymptotic stability implies that all other Floquet exponents
have negative real parts (cf. [5]). Denote the maximal real part of all Floquet exponents
82 ; : : : ; 8n by − ¡ 0.

Theorem 16 (Local necessity). Let f ∈ C 1 (R n ; R n ) with n ¿ 2. Let  be an expo-

nentially asymptotically stable periodic orbit with (minimal) period T , and let the
maximal real part of the Floquet exponents except the trivial one be − ¡ 0.
Then for all  ¿ 0 there exists a Riemannian metric M (p), so that LM (p) 6−+
holds for all p ∈ . M (p) is de<ned in an open neighborhood W of .

Proof. Fix a point p ∈ . We consider the Grst variation equation along the periodic
orbit, namely

ẏ = Df(S. p)y: (41)

By Floquet theory, every fundamental matrix W (.) of the linear system (41) can be
expressed by W (.) = P(.)eB. , where P(.) is a .-periodic matrix with period T and B
is a constant (n × n)-matrix. P(.) is C 1 with respect to .. The real parts of the Floquet
exponents, which are the eigenvalues of B, are uniquely deGned.
We Gx  ¿ 0. In Lemma 29 we transform the matrix B into a special normal form
T −1 BT =A where T depends on  such that the function uT Au assumes negative values
for all 0 = u ⊥ e1 . This is needed in the sequel to ensure that also LM assumes only
negative values. The lemma is stated and proven in the appendix.
We use the following transformation (cf. Lemma 29)  : x → T −1 P(.)−1 x which is
linear with respect to x. We deGne the Riemannian metric for all points of the periodic
orbit by

M (S. p) = P(.)−T T −T T −1 P(.)−1 : (42)

We check that LM (S. p) is strictly negative for all . ∈ [0; T ], i.e. for all points of the
periodic orbit. #d$
Let us calculate LM . Mind that M  (S. p) = d. M (S. p). Since M = M T we have
d
M  (S. p) = 2P(.)−T T −T T −1 [P(.)−1 ]:
d.
#d$ #d$
By using d. (P(.)−1 P(.)) = 0 we get d. [P(.)−1 ] = −P(.)−1 Ṗ(.)P(.)−1 . Also,
since P(.)eB. is a solution of (41) we have Ṗ(.) = Df(S. p)P(.) − P(.)B. Altogether,
we get

M  (S. p) = −2M (S. p)Df(S. p) + 2P(.)−T T −T T −1 BP(.)−1 :

664 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

Hence, we have
 
LM (S. p; v) = vT M (S. p)Df(S. p) + 12 M  (S. p) v

= vT P(.)−T T −T T −1 BP(.)−1 v:
After deGning w := T −1 P(.)−1 v we have wT T −1 P(.)−1 f(S. p) = 0 and w = 1, if
and only if vT M (S. p)f(S. p) = 0 and vT M (S. p)v = 1. Thus,
LM (S. p; v) = wT T −1 BTw = wT Aw
and hence,
LM (S. p) = max LM (S. p; v)
vT M (S. p)v=1;vT M (S. p)f(S. p)=0

= max wT Aw:

w
=1;wT T −1 P(.)−1 f(S. p)=0

If we can show that w ⊥ e1 holds for all w with wT T −1 P(.)−1 f(S. p)=0, then Lemma
29 yields LM (S. p) 6 −  + .
There is a 0 = c ∈ R n such that the solution f(S. p) of (41) is given by f(S. p) =
P(.)eB. c. Since P(.)−1 f(S. p) is .-periodic with period T , we have eBT c = c. Since
0 is the only eigenvalue of B with nonnegative real part, we have Bc = 0. Hence,
P(.)−1 f(S. p) = eB. c = c is constant and an eigenvector of B with corresponding
eigenvalue 0. Thus, T −1 BT [T −1 P(.)−1 f(S. p)] = A[T −1 P(.)−1 f(S. p)] = 0 holds for
all . ∈ [0; T ]. Thus, Lemma 29 implies that T −1 P(.)−1 f(S. p) = 4e1 with 4 = 0, and
hence w ⊥ e1 .
Now we prolongate M in a C 1 -way to a neighborhood of . For instance, we can
deGne a C 1 -map .(x) implicitly by
x − S. p; f(S. p) = 0
in a neighborhood of  and can set
M (x) := M (S.(x) p);
i.e., M is constant in the hyperplane perpendicular to f(S. p) in S. p. This is clearly
a C 1 -prolongation.

Remark 17. A similar theorem to Theorem 16 holds which guarantees the existence
of a Riemannian metric given by a di-eomorphism. The di-eomorphism is deGned by
x → T −1 P(.)−1 x for all points x of the hyperplane S. p + f(S. p)⊥ . Here we have to
assume that f ∈ C 2 in order to obtain a C 2 -di-eomorphism and a C 1 -metric (cf. [2]).
The proof follows from the fact that the Riemannian metric deGned through the
di-eomorphism is equal to the one deGned in Theorem 16 (cf. (42)) for all points
of the periodic orbit. Note, however, that the Riemannian metrics are di-erent for all
other points in general.

3.2. Global necessity

In this section we deal with the following problem: given an exponentially asymp-
totically stable periodic orbit  and a set K ⊂ A(), is there a Riemannian metric M ,
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 665

◦
such that LM (p) is strictly negative for all p ∈ K? In fact, we have to assume  ⊂ K
and K to be compact.
In Theorem 19 we show the existence of a Riemannian metric using the local The-
orem 16 and a weight function given by a Lyapunov function. In Proposition 22 we
give a second way to obtain a Riemannian metric. This time we show the existence
of a transformation using the 8ow. As in Remark 17, here we have to assume f ∈ C 2 .
For the proof of Theorem 19 we will use a Lyapunov function. Therefore we cite
the following theorem.

Theorem 18 (Lyapunov function). Let f ∈ C 1 (R n ; R n ), where n ¿ 2. Let  be an ex-

ponentially asymptotically stable periodic orbit, and A() its basin of attraction. Let
K ⊂ A() be a compact set.
Then there exists a C 1 -Lyapunov function V : K → R0+ , i.e., V (q) = 0 ⇔ q ∈  and
for the orbital derivative V  (q) := ∇V (q); f(q)

V  (q) 6 0 for all q ∈ K

and V  (q) = 0 ⇔ q ∈  hold:

Proof. The proof is similar to the one of [12, Theorem 8], where the analogous state-
ment is proven for an equilibrium. Giesl [2, Satz A.1] extends the result to periodic
orbits, but only constructs an orbitally di-erentiable function V .

The theory of Lyapunov functions also gives necessary and su5cient conditions for
the basin of attraction, but to use these su5cient conditions one has to determine the
periodic orbit explicitly, which is not the case in our approach.
Now we prove the global necessity. First we prolongate the local Riemannian metric
of Theorem 16 and then we multiply it by a weight function given by a Lyapunov
function.

Theorem 19 (Global necessity). Let f ∈ C 1 (R n ; R n ) with n ¿ 2. Let  be an expo-

nentially asymptotically stable periodic orbit, let A() be its basin of attraction, and
let the maximal real part of the Floquet exponents except the trivial one be − ¡ 0.
◦
Then for all  ¿ 0 and all compact sets K with  ⊂ K ⊂ K ⊂ A() there exists a
Riemannian metric M (p), so that LM (p) 6 −  +  holds for all p ∈ K.

Proof. Fix a set K and  ¿ 0. Denote the Riemannian metric of Theorem 16 for 12 
by M ∗ (p), which is deGned in an open neighborhood W ⊂ K of  and satisGes
LM ∗ (p) 6 −  + 12  for all p ∈ .
We deGne a prolongation M̃ (p) for all points of K. Choose a partition of unity
corresponding to W and K
\ , where K
= {x ∈ R n | dist(x; K) ¡  } and  ¿ 0, i.e.
there are C 1 -functions g1 (p), g2 (p) such that gi (p) ¿ 0 for i=1; 2 and g1 (p)+g2 (p)=1
for all p ∈ K. Furthermore, supp g1 ⊂ W and supp g2 ⊂ K
\  hold. Hence, for p ∈ 
we have g1 (p) = 1 and g2 (p) = 0.
666 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

DeGne
M̃ (p) := g1 (p)M ∗ (p) + g2 (p)I;
where M ∗ (p) = 0 outside W . This is a C 1 -function with respect to p and clearly M̃ (p)
is a symmetric matrix. We have to show that it is positive deGnite for all p ∈ K. Indeed,
we have vT M̃ (p)v = g1 (p)vT M ∗ (p)v + g2 (p)v2 ¿ 0 for all v = 0 since M ∗ (p) is
positive deGnite.
Thus, M̃ (p) is a Riemannian metric for all p ∈ K. Note that M̃ (p) = M ∗ (p) for
p ∈  and hence, also LM ∗ (p) = LM̃ (p) 6 −  + =2 holds for p ∈ , because the
orbital derivative is tangential to . Since LM̃ is continuous and  is compact there is
a neighborhood  such that LM̃ (p) 6 −  +  for all p ∈  . Set = := maxp∈K LM̃ (p).
If = 6 −  + , then choose M̃ (p) as the Riemannian metric M (p).
Now let us assume = +  −  ¿ 0. By Theorem 18 there is a Lyapunov function
V : K → R0+ so that maxp∈K∩(R n \ ) V  (p) = : − ¡ 0 holds. Set c := (=+−)= ¿ 0
and W (p) := cV (p). DeGne the Riemannian metric by
M (p) := e2W (p) M̃ (p):
We show that LM (p) = LM̃ (p) + W  (p). In the proof we set w := eW (p) v.
LM (p)
+
= max vT e2W (p) M̃ (p)Df(p) + W  (p)e2W (p) M̃ (p)
vT M (p)v=1;vT M (p)f(p)=0


,
+ 1
2 e2W (p) M̃ (p) v
+ 
,
= max e2W (p) vT M̃ (p)Df(p) + 12 M̃ (p) v + W  (p)
vT M̃ (p)v=e−2W (p) ;vT M̃ (p)f(p)=0
+ 
,
= max wT M̃ (p)Df(p) + 12 M̃ (p) w + W  (p)
wT M̃ (p)w=1;wT M̃ (p)f(p)=0

= LM̃ (p) + W  (p):

Now we calculate LM (p) and distinguish between the cases p ∈  and p ∈ K ∩ (R n \
 ). In the Grst case we have LM̃ (p) 6 −  +  and W  (p) = cV  (p) 6 0. In the
second case we have LM̃ (p) 6 = and W  (p) 6 − c = −(= +  − ). Thus in both
cases LM (p) = LM̃ (p) + W  (p) 6 −  + .

If we assume f ∈ C 2 , then we can deGne a C 2 -di-eomorphism  in a neighborhood

 such that for the corresponding Riemannian metric M̃ (p) = D(p)T D(p) the
inequality LM̃ (p) 6 −  +  holds for all p ∈  by Remark 17. In Proposition 22
we show that—more generally—for every compact subset of the basin of attraction of
 there is a di-eomorphism @, so that LM is strictly negative for all points of this
compact subset.
We need two lemmas for the proof. First we state and prove Lemma 20, which
shows that an asymptotically stable periodic orbit attracts the compact subsets of its
basin of attraction uniformly. This result is found in [4, Lemma 3.3.1].
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 667

Lemma 20. Let  be a compact, invariant and asymptotically stable set, and let
K ⊂ A() be a compact set. Then  attracts K uniformly.

Proof. Assuming the opposite, there is an  ¿ 0 and there are sequences tn → ∞ and
pn ∈ K, such that Stn pn ∈  for all n ∈ N , where  := {x ∈ R n |dist(x; ) ¡ }.  is
stable, so for the above  there is a  ¿ 0 satisfying p ∈  ⇒ St p ∈  for all t ¿ 0.
The sequence pn remains in the compact set K and therefore has a subsequence,
which we again denote by pn , with limit p ∈ K. K ⊂ A() implies that there is a
T0 ¿ 0 with ST0 p ∈ =2 . Since ST0 is continuous, there is a  ¿ 0, so that q − p
6  implies that ST0 q − ST0 p 6 =2. Let N be so large, that both tN ¿ T0 and
pN − p 6  hold. This yields ST0 pN ∈  and hence StN pN ∈  by the stability of
. This is a contradiction to the assumption.

In the following Lemma 21 we use the 8ow ST0 as a transformation and derive a
formula for the corresponding function L.

Lemma 21. Let f ∈ C 2 (R n ; R n ), where n ¿ 2. Let M (p) be a Riemannian metric.

Set M ∗ (p) := DST0 (p)T M (ST0 p)DST0 (p), where ST0 denotes the Bow.
Then LM ∗ (p) = LM (ST0 p).

Proof. We calculate (d=dt)D(ST0 +t p) twice. On the one hand we have with the chain
rule D(ST0 +t p) = DSt (ST0 p)DST0 (p). The matrix DSt (q) is a solution of the Grst vari-
ation equation Y˙ = Df(St q)Y with initial condition DSt (q)|t=0 = I . Thus, we have
d
D(ST0 +t p) = Df(ST0 +t p)DSt (ST0 p)DST0 (p) (43)
dt
and at t = 0;
d
D(ST0 +t p)|t=0 = Df(ST0 p)DST0 (p): (44)
dt
On the other hand we have with the chain rule D(ST0 +t p) = DST0 (St p)DSt (p). We
again use the fact that DSt (p) solves the Grst variation equation.
d d
D(ST0 +t p) = [DST0 (St p)DSt (p)]
dt dt
d
= [DST0 (St p)] · DSt (p) + DST0 (St p)Df(St p)DSt (p):
dt
Hence, we have at t = 0 with (44)
d
[(DST0 )(St p)]|t=0 = Df(ST0 p)DST0 (p) − DST0 (p)Df(p): (45)
dt
We also have
d
f(ST0 +t p) = ST0 +t p = DST0 (St p)f(St p):
dt
At t = 0 this yields
DST0 (p)f(p) = f(ST0 p): (46)
668 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

Set DST0 (p)u = v. Then uT DST0 (p)T M (ST0 p)DST0 (p)f(p) = 0, if and only if
vT M (ST0 p)f(ST0 p) = 0 by (46). uT DST0 (p)T M (ST0 p)DST0 (p)u = 1 holds, if and only if
vT M (ST0 p)v = 1. With (45) and M = M T we Gnd
LM ∗ (p; u) = uT [DST0 (p)T M (ST0 p)DST0 (p)Df(p)]u
 -
1 T d T-
-
+ u [(DST0 )(St p)] - M (ST0 p)DST0 (p)
2 dt t=0
-
d -
+ DST0 (p)T (M (ST0 (St p)))-- DST0 (p)
dt t=0
- 
d -
T
+ DST0 (p) M (ST0 p) [(DST0 )(St p)]-- u
dt t=0

= uT [DST0 (p)T M (ST0 p)DST0 (p)Df(p)]u

+ 12 vT [Df(ST0 p)T M (ST0 (p)) + M  (ST0 p)

+ M (ST0 p)Df(ST0 p)]v

− 12 uT [Df(p)T DST0 (p)T M (ST0 (p))DST0 (p)

+ DST0 (p)T M (ST0 p)DST0 (p)Df(p)]u

 
= vT M (ST0 p)Df(ST0 p) + 12 M  (ST0 p) v
= LM (ST0 p; v):
The maximum with respect to u and v, respectively, completes the proof of the
lemma.

Proposition 22 (Global necessity, transformation). Let f ∈ C 2 (R n ; R n ) with n ¿ 2. Let

 be an exponentially asymptotically stable periodic orbit, let A() be its basin of
attraction, and let the maximal real part of the Floquet exponents except the trivial
one be − ¡ 0.
◦
Then for all  ¿ 0 and all compact sets  ⊂ K ⊂ K ⊂ A() there exists
a C 2 -di>eomorphism @ : K → K  , so that L@ (p) 6 −  +  holds for all p ∈ K
(cf. Corollary 12 for the de<nition of L@ ).

Proof. Remark 17 shows that there is a di-eomorphism  such that M̃ (p) := D(p)T
D(p) fulGlls LM̃ (p) 6 −  + =2 for all p ∈ . There is an 0 ¿ 0 such that LM̃ (p) 6
−  +  holds for all p ∈ 0 .  attracts the compact set K uniformly by Lemma 20.
Hence, for the above 0 there is a T0 ¿ 0 with dist(ST0 K; ) ¡ 0 , i.e., ST0 K ⊂ 0 .
f ∈ C 2 (R n ; R n ) implies that the map ST0 : q → ST0 q is C 2 and so is S−T0 . This shows
that ST0 is a C 2 -di-eomorphism. We set @ :=  ◦ ST0 and M (p) := D@(p)T D@(p) =
DST0 (p)T D(ST0 p)T D(ST0 p)DST0 (p). By (38) of Corollary 12 we have L@ (p) =
LM (p) and by Lemma 21 we have LM (p) = LM̃ (ST0 p). Since ST0 K ⊂ 0 , we have
thus L@ (p) 6 −  +  for all p ∈ K.
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 669

3.3. The two-dimensional case

In the two-dimensional case we can solve both the local and the global necessity with
Riemannian metrics given by weight functions, even for f ∈ C 1 (R 2 ; R 2 ). Lemma 23
deals with the local necessity and Proposition 25 uses Theorem 18 to construct a weight
function in order to prove global necessity.

Lemma 23 (Local necessity, dimension two). Let f ∈ C 1 (R 2 ; R 2 ) and  be an

exponentially asymptotically stable periodic orbit. Let 0 and − ¡ 0 be its Floquet
exponents.
Then there is a C 1 -weight function W (p) on the periodic orbit (which can be pro-
longated arbitrarily outside in a C 1 -way) such that LM (p) = − holds for all p ∈ 
with the Riemannian metric M (p) = e2W (p) I .

Proof. Let  = {S. p | . ∈ [0; T ]} be the exponentially asymptotically stable periodic

orbit. Introduce curvilinear coordinates (.; n) in a small neighborhood of  (cf. [13,
Section 2]), where . is the time on the periodic orbit and n is the length of a normal
to . The old coordinates (x; y) and the new coordinates (.; n) satisfy
f2 (S. p)
x = (S. p)1 − n ;
f(S. p)
f1 (S. p)
y = (S. p)2 + n :
f(S. p)
Then
dn f1 (S. p) d f1 (S. p)
f2 (S. p) + +n
dy d. f(S. p) d. f(S. p) f2 (x; y)
= = :
dx dn f2 (S. p) d f2 (S. p) f1 (x; y)
f1 (S. p) − d. − n d.
f(S. p) f(S. p)
Thus, we have the following equation for dn=d..
dn f(S. p)
(.; n) =
d. f1 (S. p)f1 (x; y) + f2 (S. p)f2 (x; y)

· −f2 (S. p)f1 (x; y) + f1 (S. p)f2 (x; y)

     .
d f2 (S. p) d f1 (S. p)
−n f2 (x; y) + f1 (x; y)
d. f(S. p) d. f(S. p)

= : F(.; n):
Then we have the following equations for the evolution of . and n

.̇ = 1;
ṅ = F(.; n)
670 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

The linearization with respect to n reads


.̇ = 1;
(47)
ṅ = Fn (.; 0) n
and we have—note that we set n = 0—
Fn (.; 0)
/  
1 f2 (S. p) f1 (S. p)
= −f2 (S. p) − f1x (S. p) + f1y (S. p)
f(S. p) f(S. p) f(S. p)
 
f2 (S. p) f1 (S. p)
+ f1 (S. p) − f2x (S. p) + f2y (S. p)
f(S. p) f(S. p)
     0
d f2 (S. p) d f1 (S. p)
− f2 (S. p) + f1 (S. p)
d. f(S. p) d. f(S. p)
=L(S. p);
because the last term is
!
1 d f(S. p) f(S. p)
f(S. p) ; =0
2 d. f(S. p) f(S. p)
and
.
L(S. p) = L(S. p; v) with v = (f2 (S. p); −f1 (S. p))T )=f(S. p). Thus, n(.) =
e 0  p) d is a solution of (47). A fundamental matrix is given by
L(S
  1  2
. 0 0 0
. exp .
0 e.+ 0 L(S p) d 0 −
T
with the Floquet exponent − := 1=T 0 L(S p)d.
.
DeGne W (S. p) := −. − 0 L(S p) d on the periodic orbit and prolongate it
smoothly. Then W  (S. p) = − − L(S. p) and thus, by (39) of Corollary 14, LM (S. p) =
L(S. p) + W  (S. p) = − for all points of the periodic orbit.

Example 24. Let us give an example where we can calculate the weight function W (p)
analytically and use Theorem 5 to determine a part of the basin of attraction of a given
limit cycle. Consider

ẋ = −(x2 + y2 − 1)(x + 4y) − y;
(48)
ẏ = −(x2 + y2 − 1)y + x:
The system has the periodic orbit (x(t); y(t)) = (cos t; sin t). We calculate the
function L.
1
L(x; y) = [ − f2 (x; y)2 (3x2 + y2 + 8xy − 1)
f1 (x; y)2 + f2 (x; y)2

− f1 (x; y)2 (x2 + 3y2 − 1) + f1 (x; y)f2 (x; y)(4x2 + 12y2 + 4xy − 4)]:
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 671

− +
1

+ +
–2 –1 0 1 2
+ +

–1

+ −

–2

Fig. 3. A subset of the basin of attraction (thin lines), the sign of LM and the set LM = 0 (thick lines)
for (48).

We have L(cos t; sin t) = −2 − 8 cos t sin t. Hence, there are points of the periodic orbit,
where L is positive. We calculate the weight function W according to Lemma 23:
 t
W (cos t; sin t) := −2t + (2 + 8 cos  sin ) d = 4 sin2 t:
0

Now we prolongate W such that W is constant along the angles by setting W (x; y) :=
4y2 =(x2 + y2 ) for (x; y) = (0; 0). Then
 
8xy −y
∇W (x; y) = 2
(x + y2 )2 x
and
8xy
LM (x; y) = L(x; y) + (−yf1 (x; y) + xf2 (x; y)); (49)
(x2 + y2 )2
where M (p) = e2W (p) I . We deGne a tube around the periodic orbit by
 
cos t
r± := (1 ± a(t))
sin t
2
and a(t) := e−4 sin t−–t , where  := 0:2 and – := 0:05. We can show that the tube is
positively invariant and lies in the region where LM is negative (cf. Fig. 3). Hence,
by Theorem 5, it belongs to the basin of attraction of the periodic orbit.

In the following, we will prove the global necessity for two-dimensional systems,
which can be achieved by weight functions, too.
672 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

Proposition 25 (Global necessity, dimension two). Let f ∈ C 1 (R 2 ; R 2 ). Let  be an

exponentially asymptotically stable periodic orbit with Floquet exponents 0 and −¡0,
◦
and denote its basin of attraction by A(). Let K be a compact set with  ⊂ K ⊂
K ⊂ A().
Then for all  ¿ 0 there exists a C 1 -weight function W : K → R , such that the
Riemannian metric M (p) := e2W (p) I satis<es LM (p) 6 −  +  for all p ∈ K.

Proof. The proof is the same as the proof of Theorem 19. First we prolongate the
weight function W1 (p) of Lemma 23 to K in a C 1 -way and we deGne the Riemannian
metric M̃ (p) = e2W1 (p) I . Then we deGne a second weight function W2 (p) through a
Lyapunov function as in Theorem 19 and set M (p) := e2W2 (p) M̃ (p) = e2W (p) I with
W (p) = W1 (p) + W2 (p).

4. Main results

We summarize our results in two theorems. Theorem 26 gives a su5cient and neces-
sary condition for the existence of an exponentially asymptotically stable periodic orbit.
Theorem 27 provides a characterization of the basin of attraction of such a periodic
orbit.

Theorem 26. Let ẋ = f(x), f ∈ C 1 (R n ; R n ) with n ¿ 2, de<ne a dynamical system.

Then the following two conditions are equivalent.

(i) The system has an exponentially asymptotically stable periodic orbit.

(ii) There is a set ∅ = K ⊂ R n and a Riemannian metric M (p), such that K is
compact, connected, positively invariant, contains no equilibrium, and LM (p) ¡ 0
holds for all p ∈ K.

Proof. (ii) ⇒ (i) follows by Theorem 5. For (i) ⇒ (ii) let − ¡ 0 be the largest real
part of all Floquet exponents except the trivial one. With  := =2 deGne a Riemannian
metric M (p) according to Theorem 16 and prolongate it to a neighborhood of . There
is a  such that for all q ∈ M := {q ∈ R n | minp∈ [(q − p)T M (p)(q − p)]1=2 ¡ } the
inequality LM (q) ¡ 0 holds. Since  is stable there is a  ¿ 0 such that q ∈ M
⇒
St q ∈ M for all t ¿ 0. Fix p ∈  and set H := {q ∈ M
|(q − p) ⊥ M (p)f(p)} and

K := {STpq (.) q|q ∈ H; . ¿ 0} = {STpq (.) q|q ∈ H; . ∈ [0; T ]}

by (2.3) of Proposition 7. In particular, K is positively invariant. We have K ⊂ M ,

and hence K is compact and LM (p) ¡ 0 holds for all points of K. K also fulGlls the
other properties and thus the theorem is proven. Moreover, for this set K the inclusion
◦
 ⊂ K holds.
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 673

Theorem 27. Let  be an exponentially asymptotically stable periodic orbit of the

system of Theorem 26. Then the following formula for its basin of attraction holds

A() = H;
H ∈A
◦
where A := {H ⊂ R n | ⊂ H ⊂ H is compact; connected;

positively invariant; contains no equilibrium; and there is

a Riemannian metric M (p) with LM (p) ¡ 0 for all p ∈ H }:

Proof. The inclusion ‘⊃’ follows by Theorem 5.

We now prove the inclusion ‘⊂’. Let a ∈ A(). There is a T0 ¿ 0, so that ST0 a ∈ K,
where K is the set of Theorem 26. Hence, a ∈ S−T0 K= : H , and H is compact, con-
nected, positively invariant, and does not contain any equilibrium. Furthermore  ⊂
◦
H ⊂ H ⊂ A(). By Theorem 19 with  := =2 there exists a Riemannian metric
M (p) such that LM (p) ¡ 0 holds for all p ∈ H . Hence, H ⊂ A and the theorem is
proven.

Acknowledgements

I thank the referee for bringing Refs. [8–10] to my attention.

Appendix A. Continuity of L

In Proposition A.1 we will show that the function LM (p) depends continuously
on p.

Proposition A.1. Assume f ∈ C 1 (R n ; R n ) and M ∈ C 1 (R n ; R n×n ) such that M (p) is a

symmetric and positive de<nite matrix for all p ∈ R n . Let D := {x ∈ R n |f(x) = 0}.
Then

D → R

 
LM : 1 

 p →
 max T T
v M (p)v=1; v M (p)f(p)=0 v T
M (p)Df(p) + M (p) v
2
is a continuous function.

Proof. We assume in contradiction that there is a p ∈ D, an  ¿ 0 and a sequence 5n

with p + 5n ∈ D and 5n → 0, so that |LM (p + 5n ) − LM (p)| ¿  for all n ∈ N . So there
is either a subsequence satisfying LM (p + 5n ) ¿ LM (p) +  or a subsequence satisfying
LM (p + 5n ) 6 LM (p) − . These cases will be dealt with separately.
Case 1: Let 5n be a subsequence with LM (p + 5n ) ¿ LM (p) + . For LM (p + 5n )
there exists by deGnition a wn with wn ⊥ M (p + 5n )f(p + 5n ), wnT M (p + 5n )wn = 1 and
LM (p+5n ; wn )=LM (p+5n ) ¿ LM (p)+. The vectors wn lie in a compact set and hence
674 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

there is a subsequence, which we still denote by wn , such that wn → w. The vector w

satisGes by continuity LM (p; w) ¿ LM (p) + , wT M (p)w = 1 and wT M (p)f(p) = 0.
Hence,
LM (p) ¿ LM (p; w) ¿ LM (p) + 
which is a contradiction.
Case 2: Now let 5n be a subsequence satisfying LM (p + 5n ) 6 LM (p) −  and v be
a vector with vT M (p)v = 1 and v ⊥ M (p)f(p) such that LM (p; v) = LM (p). Set
vT M (p + 5n )f(p + 5n )
w̃n := v − f(p) = 0
f(p)T M (p + 5n )f(p + 5n )
for n ¿ N where N is chosen so large that f(p)T M (p + 5n/ )f(p + 5n ) = 00 for all
n ¿ N . We have w̃n M (p + 5n )f(p + 5n ) = 0. Now set wn := √ T 1
T
w̃n for
w̃n M (p+5n )w̃n )
n ¿ N . We have wnT M (p
+ 5n )wn = 1 and wnT M (p
+ 5n )f(p + 5n ) = 0. The vectors
wn form a convergent sequence with limit wn → v. Thus, there is a Ñ ¿ N such
that LM (p + 5n ; wn ) ¿ LM (p; v) − =2 for all n ¿ Ñ . For such an n ¿ Ñ we have the
following contradiction

LM (p) = LM (p; v) 6 LM (p + 5n ; wn ) +
2
 
6 LM (p + 5n ) + 6 LM (p) − :
2 2
This proves the continuity of LM (p).

Appendix B. A special normal form

In this section we prove Lemma B.1, which has been used in the proof of Theorem
16. A special normal form A of a matrix B is given to ensure that uT Au assumes
negative values for u ⊥ e1 and u = 0.

Lemma B.1. Let B ∈ R n×n . Let 81 =0 be an eigenvalue of B with algebraic multiplicity

one and let Re(8) 6 −  ¡ 0 hold for all other eigenvalues 8 of B.
Then for all  ¿ 0 there exists an invertible matrix T , such that A := T−1 BT
satis<es
uT Au 6 (− + )u2 for all u ⊥ e1 :
If A5 = 0, then 5 is of the form 5 = 4e1 with 4 ∈ R .

Proof. Let 81 =0, 82 ; : : : ; 8r ∈ R and 8r+1 ; 8r+1 ; : : : 8r+c ; 8r+c ∈ C \R denote the eigenval-
ues of B. Set 8r+j = : 4j +i=j , where 4j ; =j ∈ R for 1 6 j 6 c. Let m1 =1;m2 ; : : : ; mr+c ¿1
r c
be the algebraic multiplicities of the eigenvalues, so that i=1 mi + j=1 2mr+j = n.
By the Jordan normal form theorem there is a complex matrix S, so that S −1 BS is
the Jordan normal form of B. Let wj = uj + ivj be a basis of the eigenspaces of the
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 675

complex eigenvalues. Writing uj ; vj instead of wj ; wj in the columns of the matrix S,

we get a real-valued matrix S̃, which satisGes

 
J1
 
 .. 
 . 
 
 
 Jr 
−1  
S̃ BS̃ =  ;
 Kr+1 
 
 
 .. 
 . 
 
Kr+c

where Ji are the Jordan matrices and

 
4j =j 1
 
 −=j 4j 0 1 
 
 
 4j =j 
 
 
 .. 
Kr+j := 
 −=j 4j . 1 :

 
 .. 
 . 0 1
 
 
 4j =j 
 
−=j 4j

DeGne

 
S1
 
 .. 
 . 
 
 
 Sr 
 
S :=  ;
 Sr+1 
 
 
 .. 
 . 
 
Sr+c

where Sj := diag(1; ; 2 ; : : : ; mj −1 ) for j 6 r and Sj := diag(1; 1; ; ; : : : ; mj −1 ; mj −1)

for j ¿ r. Set T := S̃S .
676 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677

Then
 
M1 0
 
 .. 
 . 0 
 
 
 0 Mr 
 
T−1 BT = A :=   with
 Z1 0 
 
 
 .. 
 0 . 
 
0 Zc
 
8j  0
 
 .. .. 
 . . 
 
Mj :=  
 .. 
 .  
 
8j

and
 
4j =j 
 
 −=j 4j 0  
 
 
 4j =j 
 
 
 .. 
Zj := 
 −=j 4j .  :

 
 .. 
 . 0  
 
 
 4j =j 
 
−=j 4j

Since 81 = 0 is a single eigenvalue, we have M1 = 0. If now A5 = 0, then the form of

A implies that 5 = 4e1 .
Now we prove the inequality. First we consider a vector of a real eigenspace with
eigenvalue = 0. Let 2 6 j 6 r and v ∈ R mj . Then

v T Mj v
 
= 8j v12 + v22 + · · · + vm
2
j
+ (v1 v2 + v2 v3 + · · · + vmj −1 vmj )
    
6 8j v12 + v22 + · · · + vm
2
+ v 2
1 + v 2
2 + v 2
2 + v 2
3 + · · · + v 2
m −1 + v 2
m
j
2 j j

# $ 2 
6 8j +  v1 + v22 + · · · + vm 2
j
:
 P. Giesl / Nonlinear Analysis 56 (2004) 643 – 677 677

Now let 1 6 j 6 c and v ∈ R 2mr+j .

 
 
vT Zj v = 4j v12 + v22 + · · · + v2m
2
+ =j v1 v2 − v1 v2 + : : :
r+j   
=0

+ (v1 v3 + v2 v4 + v3 v5 + · · · + v2mr+j −2 v2mr+j )

 
6 4j v12 + v22 + · · · + v2m
2
r+j
 

 2


+ v1 + v32 + v22 + v42 + v32 + v52 + : : : + v2m
2
−2 + v 2
2m 
2  
r+j

r+j

2
=v12 +v22 +v2m 2
+v2m +2(v32 +v42 +:::+v2m
2 )
r+j −1 r+j r+j −2

# $ 
6 4j +  2
v12 + v22 + · · · + v2m r+j
:
So we get by deGnition of A for all u ⊥ e1
   
# $
u Au 6 max max 8j ; max 4i +  u22 + · · · + un2
T
26j6r 16i6c

6 (− + )u2 ;
which completes the proof of the lemma.

References

[1] G. Borg, A condition for the existence of orbitally stable solutions of dynamical systems, Kungl. Tekn.
HOogsk. Handl. 153 (1960).
[2] P. Giesl, Eine Charakterisierung der Einzugsbereiche von Gleichgewichtspunkten und periodischen
Orbits dynamischer Systeme, Ph.D. Thesis, Technical University of Munich, MOunchen, 2000
(in German).
[3] P. Giesl, Unbounded Basins of Attraction of Limit Cycles, Acta Math. Univ. Comenianae 72 (1) (2003)
81–110.
[4] J. Hale, Ordinary Di-erential Equations, Wiley-Interscience, New York, 1969.
[5] P. Hartman, Ordinary Di-erential Equations, Wiley-Interscience, New York, 1964.
[6] P. Hartman, C. Olech, On global asymptotic stability of solutions of di-erential equations, Trans. Amer.
Math. Soc. 104 (1962) 154–178.
[7] A.Yu. Kravchuk, G.A. Leonov, D.V. Ponomarenko, A criterion for the strong orbital stability of the
trajectories of dynamical systems I, Di-erents. Uravn. 28 (9) (1992) 1507–1520.
[8] G.A. Leonov, I.M. Burkin, A.I. Shepelyavyi, Frequency Methods in Oscillation Theory, in: Series in
Mathematics and Its Applications, Vol. 357, Kluwer, Dordrecht/Boston/London, 1996.
[9] G.A. Leonov, A. Noack, V. Reitmann, Asymptotic orbital stability conditions for 8ows by estimates of
singular values of the linearization, Nonlinear Anal. 44 (2001) 1057–1085.
[10] G.A. Leonov, D.V. Ponomarenko, V.B. Smirnova, Local instability and localization of attractors, from
stochastic generator to Chua’s systems, Acta Appl. Math. 40 (1995) 179–243.
[11] B. StenstrOom, Dynamical systems with a certain local contraction property, Math. Scand. 11 (1962)
151–155.
[12] M. Urabe, Moving orthonormal system along a closed path of an autonomous system, J. Sci. Hiroshima
Univ. Ser. A 21 (1958) 177–192.
[13] Y. Ye, et al., Theory of Limit Cycles, Translations of Mathematical Monographs, AMS 66, American
Mathematical Society, Providence, RI, 1986.