NSTV 2008 DCDS B

DISCRETE AND CONTINUOUS Website: http://aimSciences.
org
DYNAMICAL SYSTEMS SERIES B
Volume 10, Number 2&3, September & October 2008 pp. 621–650
PERIODIC ORBITS AND STABILITY ISLANDS IN CHAOTIC

SEAS CREATED BY SEPARATRIX CROSSINGS IN SLOW-FAST
SYSTEMS
Anatoly Neishtadt
Space Research Institute
Profsoyuznaya 84/32, Moscow 117997, Russia
and
Department of Mathematical Sciences, Loughborough University
Loughborough, LE11 3TU, UK
Carles Simó
Departament de Matemàtica Aplicada i Anàlisi, Universitat de Barcelona
Gran Via 585, 08007 Barcelona, Spain
Dmitry Treschev
V.A. Steklov Mathematical Institute
Gubkina 8, Moscow 119991, Russia
Alexei Vasiliev
Space Research Institute
Profsoyuznaya 84/32, Moscow 117997, Russia
Abstract. We consider a 2 d.o.f. natural Hamiltonian system with one degree

of freedom corresponding to fast motion and the other one corresponding to
slow motion. The Hamiltonian function is the sum of potential and kinetic
energies, the kinetic energy being a weighted sum of squared momenta. The
ratio of time derivatives of slow and fast variables is of order ε ≪ 1. At
frozen values of the slow variables there is a separatrix on the phase plane
of the fast variables and there is a region in the phase space (the domain of
separatrix crossings) where the projections of phase points onto the plane of the
fast variables repeatedly cross the separatrix in the process of evolution of the
slow variables. Under a certain symmetry condition we prove the existence of
many, of order 1/ε, stable periodic trajectories in the domain of the separatrix
crossings. Each of these trajectories is surrounded by a stability island whose
measure is estimated from below by a value of order ε. Thus, the total measure
of the stability islands is estimated from below by a value independent of ε.
We find the location of stable periodic trajectories and an asymptotic formula
for the number of these trajectories. As an example, we consider the problem
of motion of a charged particle in the parabolic model of magnetic field in the
Earth magnetotail.
2000 Mathematics Subject Classification. Primary: 34E10 ; Secondary: 37J40 .

Key words and phrases. slow-fast systems, separatrix crossings, chaotic seas, stability islands.
The work was supported in part by grants RFBR 06-01-00117, 05-01-01119, and NSh-
1312.2006.1, and NSh-691.2008.1. C.S. has been supported by grants BFM2003-09504-C02-01,
MTM2006-05849/Consolider (Spain) and CIRIT 2005 SGR-1028 (Catalonia). The cluster facility
HIDRA of the Dynamical Systems group at the Univ. de Barcelona has been widely used.
621
622 A. NEISHTADT, C. SIMÓ, D. TRESCHEV AND A. VASILIEV
1. Introduction. Under rather general assumptions, in the phase space of a math-

ematical pendulum of a slowly periodically varying length there exist many, of order
1/ε, stable periodic trajectories. Along each of them, the pendulum transits from
the oscillatory mode to the rotational one and back again [1, 2]. Here ε > 0 is a small
parameter characterizing the rate of length variation. In the Poincaré section each
of these periodic trajectories is represented by a stationary point surrounded by a
stability island whose area is estimated from below by a value of order ε. Hence, the
total area of the stability islands is estimated from below by a value independent
of ε. Numerical experiments demonstrate that these islands are immersed into a
chaotic sea and its measure can be estimated from below by a value independent of
ε. However, rigorous estimates of this measure are not known.
The transition from one mode of the pendulum’s motion to another one corre-
sponds to a crossing of the separatrix on the pendulum’s phase portrait considered
at a frozen value of its length. Similar results about stable periodic motions with
separatrix crossings are valid in the problem of a particle’s motion in a slowly pe-
riodically varying double-well potential [1, 2].
Analogous results can be obtained for 2 d.o.f. Hamiltonian systems with fast
and slow variables (with one degree of freedom corresponding to the fast variables,
and another one corresponding to the slow variables). At frozen values of the slow
variables we obtain a 1 d.o.f. Hamiltonian system for the fast variables, called
the fast system. This system involves values of the slow variables as parameters.
Assume that on the phase portrait of the fast system there exists a saddle point,
and the separatrices passing through this point form an “8”-like figure. As values
of the slow variables change, the projections of the phase points onto the plane of
the fast variables may cross the separatrix. Assume that the following symmetry
condition is valid: the areas bounded by the two separatrix loops are equal at all
possible values of the slow variables. Then, under certain additional conditions, on
every energy level there are many, of order 1/ε, stable periodic trajectories crossing
the separatrix [3]. Here ε > 0 is a small parameter characterizing the ratio of time
derivatives of slow and fast variables. Each one of these periodic trajectories is
surrounded by a stability island whose measure is estimated from below by a value
of order ε. Hence, the total area of the stability islands is estimated from below by
a value independent of ε [3].
The proofs in [1, 2] and [3] are based on the study of asymptotic formulas for the
corresponding Poincaré maps. In [1, 2], these formulas were constructed with the
use of asymptotic expressions for the jump of the adiabatic invariant at a separatrix
crossing in systems with one and a half d.o.f. [4, 5, 6] and for the variation of the
“angle” variable between separatrix crossings in such systems [7]. In [3], analogous
formulas for systems with two d.o.f. [8, 9] were used.
In the present paper we consider a natural Hamiltonian system with two degrees
of freedom corresponding to fast and slow motions; the Hamiltonian function is a
sum of potential and kinetic energies, the kinetic energy is a weighted sum of squared
momenta. This system possesses an additional symmetry allowing to locate the
periodic trajectories and to find explicitly an asymptotic formula for their number.
As main example we consider a model problem [10] of the motion of charged particles
in the Earth magnetotail. The results can be also applied to some systems that are
not natural but still possess the necessary additional symmetry. In particular, the
results are applicable to the model system introduced in [11] to describe the motion
STABILITY ISLANDS IN CHAOTIC SEAS OF SLOW-FAST SYSTEMS 623
of an asteroid near a 3:1 resonance with the Jupiter. The stability islands in this
problem were found numerically in [12].
The paper is organized as follows. In Section 2 the equations of motion are pre-
sented. These equations contain fast and slow variables. We describe the dynamics
of the fast variables at frozen values of the slow variables and the dynamics of the
slow variables, averaged over the fast variables. The dynamics is considered on a
fixed energy level. Section 3 contains the formulation of the principal analytical re-
sults of the paper, Theorems 1 and 2. The assertion of Theorem 1 is that there are
many, of order 1/ε, stable in the linear approximation periodic trajectories crossing
the separatrix. Explicit asymptotic formulae for the number of these trajectories
are given in Section 7. The assertion of Theorem 2 is that the majority of the peri-
odic trajectories whose existence is proved in Theorem 1 are stable and each such
stable trajectory is surrounded by invariant tori, which bound a domain of phase
volume estimated from below by a value of order ε.
In Section 4 we put forward a standard procedure of approximate description of
the dynamics (an improved adiabatic approximation) which describes the behavior
of the fast variables on time intervals of length of order 1/ε with accuracy tending
to 0 as ε → 0. Such an accuracy is suitable to search for periodic trajectories. In
Section 5 we introduce auxiliary variables which are needed to describe the sepa-
ratrix crossings. These variables are closely related with the variables introduced
in Section 4. One of these variables is the improved adiabatic invariant. It is a
function of the phase variables which is constant up to terms O(ε2 ) in the motion
far from the separatrix. In Section 6, we present the asymptotic formulae for the
Poincaré return map for the problem under consideration. In the adiabatic approx-
imation the behavior of the slow variables is periodic with period proportional to
1/ε. We consider a surface of section to which the phase points return over the
time close to this period. This two-dimensional surface of section is parametrized
by two variables: the value of improved adiabatic invariant and an angular variable
analogous to the phase of the fast motion for phase points near the separatrix. The
possibility to construct the return map for such a long period is based on two facts:
a) far from separatrix the improved adiabatic invariant changes are O(ε2 ), while
the passage through a narrow neighborhood of separatrices gives rise to a change
O(ε); there is a known asymptotic formula for this change, b) the change of the
phase of the fast motion with high enough accuracy can be calculated via formulae
of improved adiabatic approximation.
In Section 7, we study the problem of existence of stable in linear approxima-
tion fixed points of the return map. In Subsection 7.1, we present equations for
stable, in linear approximation, fixed points. In Subsection 7.2, it is proved that
there are many, ∼ 1/ε , such fixed points, and moreover, asymptotic formulae for
the number of such fixed points are presented. The analysis here is based on the
reduction of the problem under consideration to the following ergodic problem. On
a two-dimensional torus there is a slowly (with velocity O(ε)) moving curve. There
is also a point moving on the same torus with slowly varying velocity (the accelera-
tion is O(ε)). One should count the number of times when the point (transversely)
crosses the curve on a time interval of length const/ε. A stable, in linear approxima-
tion, fixed point of the return map corresponds to each such crossing. Asymptotic
formula for the number of crossings can be obtained easily under some generic con-
dition. This completes the proof of Theorem 1. In Subsection 7.3, some symmetry
properties of the periodic trajectories are discussed. In Section 8, Theorem 2 is

proved.
Section 9 contains an analytical and numerical investigation of an example pre-
senting a model problem of the motion of charged particles in the Earth magnetotail.
In Subsections 9.1, 9.2, we describe this model and apply to it the theoretical results
of the previous sections. In Subsection 9.3, we present results of the numerical inves-
tigation of the problem and demonstrate a good agreement of previous theoretical
results with this numerics.
2. Main equations. Consider a 2 d.o.f. Hamiltonian system with the Hamiltonian

function
1 1
H = H(p, q, y, x) = g(x)y 2 + β(x)p2 + U (q, x). (1)
2 2
−1
Here (p, q) and (y, ε x) are the pairs of canonically conjugated variables, ε > 0 is
a small parameter. All the functions are assumed to be of smoothness C ∞ (but all
results are valid for sufficiently high finite smoothness); the functions g and β are
positive. The equations of motion have the form:
∂H ∂H ∂H ∂H
ṗ = − , q̇ = , ẏ = −ε , ẋ = ε . (2)
∂q ∂p ∂x ∂y
The variables p, q are called fast, while the variables y, x are called slow.
The system with Hamiltonian H at frozen values of (y, x) is called the fast system.
It contains x as parameter. This is a natural system with one degree of freedom.
We assume that the potential U has a unique local maximum at the point q = qs (x)
(the subindex “s” stands for “saddle”) and that U is a symmetric function of q with
respect to qs (x), having two minima, for all the considered values of x.
The phase portrait of the fast system is shown in Figure 1. On it there exist
a non-degenerate saddle point C and separatrices l1 , l2 passing through C. These
separatrices divide the (p, q)-plane into domains Gi = Gi (x), i = 1, 2, 3. G1 and G2
are symmetric w.r.t. the line q = qs (x) and also w.r.t. p = 0. Introduce E by
1
E = E(p, q, y, x) = β(x)p2 + U (q, x) − U (qs (x), x). (3)
2
In the domain G3 the function E is positive, and in G1,2 (x) it is negative. At the
point C and on the separatrices E = 0.
G3
G1 G2
C
q
l1 l2
Figure 1. Phase portrait of the fast system.
Let S = S(x) denote the area of G1 (x). In the unperturbed system, one can
introduce canonical “action-angle” variables (I, ϕ) separately in each Gi . The “ac-
tion” I is a function of p, q, x given by the formula I = I(E, x), where I(e, x) is the
area bounded by a trajectory of the fast system, corresponding to E = e, divided
by 2π. The action is discontinuous on the separatrix, and hence it is convenient

to introduce a modified action Iˆ = I(E, ˆ x) coinciding with I at the points of G1,2
and equal to I/2 at the points of G3 . On the separatrices and at the point C put
Iˆ = S(x)/(2π). The function Iˆ is continuous.
For an approximate description of the motion in system (2) one can use the
adiabatic approximation (see, e.g. [13]). In this approximation Iˆ = const along
a phase trajectory, and the variation of the slow variables y, x is defined by a
Hamiltonian system with Hamiltonian εH0 (I, y, x), where H0 is the function H
expressed in terms of I, y, x. Thus, H0 = 12 g(x)y 2 + V (I, x), where V (I, x) is the
effective potential given by the function 12 β(x)p2 +U (q, x) expressed in terms of I, x.
The system with the Hamiltonian H0 is called the slow system. Let Ĥ0 = Ĥ0 (I, ˆ y, x)
denote the Hamiltonian H0 expressed in terms of I, y, x.ˆ
We shall consider trajectories of the slow system on an energy level H0 = h0 .
Along a trajectory Iˆ = I(h ˆ 0 − 1 g(x)y 2 − U (qs (x), x), x) = const. Assume that on
2
the plane of the slow variables there exists a domain (an annulus) D filled up by
closed trajectories of the slow system with H0 = h0 . The values of Iˆ corresponding
to these trajectories fill up a segment Ξ.
Consider on the plane of variables (y, x) the uncertainty curve [11] Γ = {y, x :
1 2
2 g(x)y + U (qs (x), x) = h0 }. We assume that every trajectory of the slow system
that belongs to D crosses the uncertainty curve at exactly two points. Assume also
that at these points Θ(y, x) ≡ gy ∂S/∂x 6= 0. Note that the points of intersection
of a trajectory of the slow system with the uncertainty curve are symmetric with
respect to the axis Ox. The uncertainty curve divides D into two subdomains D3
and D1,2 . Consider a point (y, x) ∈ D3 . The corresponding energy level line H = h0
on the plane of the fast variables (p, q) belongs to the domain G3 (x) and gives a
trajectory of the fast system. For a point (y, x) ∈ D1,2 the corresponding energy
level consists of two trajectories of the fast system, belonging to the domains G1
and G2 respectively. For a point (y, x) ∈ Γ the energy level is the union of the
separatrices l1 (x), l2 (x) and the point C(x).
Remark. The value Θ(y, x) is, up to the multiplier ε, the rate of change of the
area surrounded by the separatrix loop on the plane of fast variables as the slow
variables are evolving according to the slow system. This value plays an impor-
tant role in the subsequent analysis. In the adiabatic approximation the “action”
I remains constant. However, the area surrounded by the separatrix loop is chang-
ing. At the points of the uncertainty curve, where Θ(y, x) > 0 (correspondingly,
Θ(y, x) < 0), the phase points of the slow system cross the uncertainty curve from
D3 to D1,2 (correspondingly, from D1,2 to D3 ). Accordingly, the projections of the
phase points of the exact system onto the plane of the fast variables leave G3 and
enter G1 or G2 at Θ(y, x) > 0, and leave G1 , G2 and enter G3 at Θ(y, x) < 0. There
is a useful formula
∂S ∂E
I
=− dt. (4)
∂x l1 ∂x
(The separatrix l1 is parametrized by time t of motion in the fast system along it,
−∞ < t < +∞. So, on the separatrix ∂E/∂x is a function of t, and the integral
in (4) is taken from t = −∞ to t = +∞.) Formula (4) shows that −εΘ(y∗ , x∗ )
gives an approximate value of the change of the function E along a segment of a
trajectory of the exact system, such that the projection of this segment onto the
plane of the fast variables is close to the separatrix l1 for x ≈ x∗ , y ≈ y∗ .
3. Formulation of results. We look for stable periodic trajectories of system (2),

belonging to one of the following three families F1, F2, F(1,2).
F1. The projection of the phase trajectory onto the plane of the slow variables
is a closed curve. Along this curve the value of Iˆ oscillates with an amplitude O(ε)
around a quantity Iˆ0 ∈ Ξ. The period of the motion along this trajectory differs by
O(1) from the period of the motion along the trajectory Iˆ = Iˆ0 = const of the slow
system. When the projection of the phase point onto the plane of the slow variables
is in the domain D1,2 at a distance larger than c0 ε from the uncertainty curve, the
projection of the phase point onto the plane of the fast variables lies in G1 . Here
c0 is a large enough positive constant.
F2. The same as F1 with G1 replaced by G2 .
F(1,2). The same as F1, but the period of motion is O(1)-close to the double
of the period of motion along the corresponding trajectory of the slow system, and
when the projection of the phase point onto the plane of the slow variables is in the
domain D1,2 at a distance larger than c0 ε from the uncertainty curve, the projection
of the phase point onto the plane of the fast variables is in turn in the domains G1
and G2 .
Theorem 1. Under additional generic conditions, on the energy level H = h0
each of the families F1, F2, and F(1,2) consists of Kε (1 + o(1)) linearly stable peri-
odic trajectories. Among the trajectories of both F1 and F2 there are 1ε Ksm (1+o(1))
trajectories passing through points where y = 0, p = 0 (hence, they are symmetric).
Among the trajectories of F(1,2) there are 1ε Ksm (1 + o(1)) trajectories passing
through points where y = 0, q = qs (x) + o(1) (hence, they are nearly symmetric; if
qs ≡ const, these trajectories pass through the points with y = 0, q = qs and are
symmetric). Here K, Ksm are positive constants; formulas for them are given in
Section 7. The required generic conditions are given in Subsection 7.2 (conditions
H1).
This theorem differs from the corresponding result of [3] in the following aspects:
here we provide an explicit asymptotic formula for the number of periodic trajecto-
ries, we include trajectories of F(1,2) into our consideration, and we describe the
locations of the periodic trajectories (defined by the location of a point on the fast
plane at y = 0).
The phase space of system (2) is four-dimensional, and the energy level H = h0
is three-dimensional. Therefore, if there is a two-dimensional invariant torus of the
system, it divides the energy level into two invariant domains.
Fix an arbitrary positive constant c00 .
Theorem 2.Under additional generic conditions, among the trajectories of each
K−c−1
family, specified in Theorem 1, there are not less than ε
00
(and, for symmetric
K −c−1
and nearly symmetric trajectories, not less than smε 00 ) stable trajectories. Each
of these trajectories is surrounded by an invariant torus, which bounds a domain of
(3-dimensional) phase volume larger than C1−1 ε. The variation of Iˆ in this domain
is smaller than C2 ε. The required generic conditions are given in Subsection 8
(conditions H2).
Corollary. The total measure of these domains is larger than C3−1 .
Here Ci are positive constants, i.e. values independent of ε.
This theorem differs from the corresponding result of [3] by the explicit estimate
of the number of the stable trajectories.
For any periodic orbit of Theorem 2 there is a Cantor set of tori surrounding
it, whose existence is provided by KAM theory. Each one of these tori bounds
a solid torus which contains the periodic orbit. The closure of the union of all
these solid tori is known as the stability domain associated to the periodic orbit.
The intersection of this domain with a surface of a Poincaré section is known as a
stability island.
4. Adiabatic and improved adiabatic approximations. In the fast system,

the action-angle variables I, ϕ mod 2π are introduced separately in each Gi by
a canonical transformation of variables. The corresponding generating function
W (I, q, x) contains x as a parameter [for brevity, we omit the i subscripts]. We take
this function in the form
Z q
W (I, q, x) = P(I, q ′ , x)dq ′ , (5)
q0 (I,x)
where P is the value of the p-variable along the trajectory with the prescribed value
of the action I. In G1 , G2 , q0 (I, x) is the value of q at one of the two points where
this trajectory crosses the axis p = 0; in G1 we take the right one of these points, and
in G2 we take the left one. In G3 we take q0 (I, x) = qs (x) and assume that at this
point P > 0. In the new variables the Hamiltonian has the form H = H0 (I, y, x).
Now make a canonical transformation of variables (p, q, y, x) 7→ (I, ¯ ϕ̄, ȳ, x̄) with
−1 ¯
generating function ȳε x + W (I, q, x). The canonically conjugated pairs of vari-
¯ ϕ̄) and (ȳ, ε−1 x̄). The formulas for the transformation of variables
ables are (I,
are:
¯ p = ∂W/∂q, x̄ = x, y = ȳ + ε∂W/∂x.
ϕ̄ = ∂W/∂ I, (6)
In the new variables, the Hamiltonian H has the form
¯ ȳ, x̄) + εH1 (I,

H = H0 (I, ¯ ϕ̄, ȳ, x̄) + ε2 H2 (I,
¯ ϕ̄, ȳ, x̄, ε), (7)
where
∂W
H1 = g ȳ . (8)
∂x
In the adiabatic approximation, the dynamics is described by the Hamiltonian H0 .
In this approximation I¯ = const along a phase trajectory.
One can also construct a canonical, close to the identity, transformation of vari-
¯ ϕ̄, ȳ, x̄) 7→ (J, ψ, ŷ, x̂) in order to make the terms of order ε in the Hamil-
ables (I,
tonian independent of the phase (see, for example, [9]). In the new variables, the
Hamiltonian takes the form:
H = H0 (J, ŷ, x̂) + ε2 H2 (J, ψ, ŷ, x̂, ε). (9)

The term of order ε is absent because it is equal to the average of H1 with respect
to ϕ̄, and this average is zero due to the symmetry with respect to the axis p = 0.
In the improved adiabatic approximation, the dynamics is described by the
Hamiltonian H0 (J, Y, X). In this approximation J is an integral of motion. With
an accuracy O(ε2 ), the following formula for J is valid (see [8]):

J = J(p, q, y, x) = I + εu(p, q, y, x), (10)
T
1 T ∂E
Z
u= gy −t dt. (11)
2π 0 2 ∂x
The integral here is calculated along a phase trajectory of the fast system passing
through the point (p, q); t is the time of motion along this trajectory starting from
this point, T is the period of the motion. The function J is the improved adiabatic
invariant. In the complete system far from separatrices its value along a phase
trajectory is constant with an accuracy O(ε2 ) on time intervals of order ε−1 . In
what follows, it is convenient to use the function J(p,ˆ q, y, x) equal to J in the
domains G1 , G2 and equal to J/2 in G3 .
5. Description of separatrix crossing. On the phase plane Psl of the slow vari-
ables (y, x), the separatrix is represented by a curve Γ(h0 ) (see Figure 2). The
annulus D filled up with trajectories of the slow system with Iˆ ∈ Ξ is divided by
the uncertainty curve into two domains D3 and D1,2 . The ring D intersects the
axis y = 0 along two segments S3 and S1,2 , which belong to D3 and D1,2 respec-
tively. For (y, x) ∈ S3 the points of the plane of the fast variables on the energy
level H = h0 belong to G3 , and for (y, x) ∈ S1,2 the points of the plane of the fast
variables on this energy level belong to G1 or G2 . Fix an interval Ξ0 ∈ Ξ, such that
the endpoints of Ξ0 and Ξ are different.
Γ(h0)
S3 S1,2
x
Figure 2. Schematic picture of the motion on the slow plane.
Consider the motion in the exact system (2) with initial condition M(0) =
(p(0) , q (0) , y (0) , x(0) ) at t = 0. We assume that (ŷ (0) , x̂(0) ) ∈ S3 . Hence, (p(0) , q (0) ) ∈
G3 (x(0) ). Let Iˆ = Iˆ(0) , Jˆ = Jˆ(0) , ψ = ψ (0) at the point M(0) ; Iˆ(0) ∈ Ξ0 . Consider
also the motion in the slow system with Hamiltonian Ĥ0 (Iˆ(0) , y, x) and initial (at
t = 0) condition y = 0, x = x̃(0) , where x̃(0) is determined from Ĥ0 (Iˆ(0) , 0, x̃(0) ) =
h0 . The trajectory of this motion is B(Iˆ(0) ) = {y, x : Ĥ0 (Iˆ(0) , y, x) = h0 }. This
trajectory crosses Γ(h0 ) at points (x∗ , y∗ ) and (x∗ , −y∗ ) at slow time moments τ∗
and τ∗∗ (slow time is τ = εt). We will call τ∗ and τ∗∗ the slow time moments of the
separatrix crossing in the adiabatic approximation.
Assume, for definiteness, that y∗ > 0, Θ∗ ≡ Θ(y∗ , x∗ ) > 0. Thus, a phase
point of the slow system passes from D3 to D1,2 at the point (x∗ , y∗ ), and at the
point (x∗ , −y∗ ) it passes from D1,2 to D3 . In the exact system, in the process of
evolution the projection of the phase point onto the plane of the slow variables
(ŷ, x̂) approaches the curve Γ(h0 ), and accordingly, its projection onto the plane
(p, q) approaches the separatrix. We are interested in the dynamics of phase points
that are being captured into Gν , ν = 1, 2 after the separatrix crossing. For such
a phase point, the first crossing of Cq-axis in Gν (see Figure 1) occurs near C at
time τ = τ∗ + O(εlnε). Assuming that at the point of this crossing E = h(0) , we
introduce η (0) = 1 − |h(0) /εΘ∗ | .
After the separatrix crossing the projection of the phase point onto the plane of
the slow variables (ŷ, x̂) moves towards S1,2 . When it crosses S1,2 , the projection of
the phase point onto the plane of fast variables is deep inside the region Gν . Denote
the value of Jˆ at this time moment as Jˆ(1) .
Then the phase point starts again approaching the separatrix. At τ = τ∗∗ +
O(εlnε) the projection of the phase point onto the plane of fast variables crosses
Cq-axis in Gν near the point C for the last time before entering G3 . Assuming that
at the point of this crossing E = h(1) , we introduce η (1) = |h(1) /εΘ∗ | .
After crossing Γ(h0 ), the projection of the phase point onto the plane (ŷ, x̂)
crosses again the segment S3 . Let Jˆ(2) , ψ (2) denote the values of J,ˆ ψ at this time
moment. Then the projection of the phase point onto the plane of fast variables
approaches the separatrix again, crosses it and gets captured into the domain Gl , l =
1 or l = 2. Let E be equal to h(2) at the first crossing of the Cq-axis in Gl near
C. This crossing occurs at time τ = τ∗ + T0 + O(εlnε), where T0 is the slow time
period of motion along the trajectory Iˆ = Iˆ(0) in the adiabatic approximation. We
introduce η (2) = 1 − |h(2) /εΘ∗ | .
6. The return map. In the energy level H = h0 the segment S3 is represented by a

piece of a two-dimensional surface {p, q, y, x : H(p, q, y, x) = h0 , (ŷ, x̂) ∈ S3 }. This
piece can be parametrized by variables J, ˆ ψ. The corresponding Poincaré return
map M : (Jˆ(0) , ψ (0) ) → (Jˆ(2) , ψ (2) ) produced by trajectories that pass through Gν
is symplectic. Its stable stationary points correspond to stable periodic orbits of
the original problem, of period approximately equal to T0 /ε. It is convenient to
study these stationary points using a different set of variables, namely to consider
the map M̂ : (Jˆ(0) , η (0) , ν) → (Jˆ(2) , η (2) , l).
The map M̂ is the composition of two maps: M̂ = M (2) ◦ M (1) ,
M (1) : (Jˆ(0) , η (0) , ν) → (Jˆ(1) , η (1) , ν) , M (2) : (Jˆ(1) , η (1) , ν) → (Jˆ(2) , η (2) , l) .
The results of [8, 9] give the following formulas for M (k) . Suppose that
η (0) , η (1) , η (2) ∈ [c−1 −1

1 , 1 − c1 ] . (12)
Then, as we shall prove immediately, we obtain the maps
Jˆ(1) = Jˆ(0) − (2π)−1 εa∗ Θ∗ ln(2 sin πη (0) ) + O(ε3/2 lnε) , (13)
η (1) = {η (0) + ε−1 Φ1 (Jˆ(1) ) + O(ε1/3 ln−1/3 ε)} , (14)
Jˆ(2) = Jˆ(1) + (2π)−1 εa∗ Θ∗ ln(2 sin πη (1) ) + O(ε3/2 lnε) , (15)
η (2)
= {η (1) + ε−1 Φ2 (Jˆ(2) ) + O(ε1/3 ln−1/3 ε)} , (16)
l = ν, if 0 < {(η (1) + ε−1 Φ2 (Jˆ(2) ))/2} < 1/2, (17)

l 6= ν, if 1/2 < {(η (1) + ε−1 Φ2 (Jˆ(2) ))/2} < 1. (18)
Here√{ · } denotes the fractional part, a∗ = a(x∗ ). The value of a is given by

a = 1/ −d, where d is the Hessian of E at the point C. In (14), (16) one has
introduced
Z τ+(1)
ˆ = 1
Φ1 (J)
(1) ˆ
ω (J, Y1 (τ ), X1 (τ )) dτ , (19)
2π τ−(1) 0
Z (3)
ˆ 1 τ− (3) ˆ
Φ2 (J) = ω (2J, Y3 (τ ), X3 (τ )) dτ. (20)
π τ+(3) 0
(j)
Here ω0 = ∂H0 /∂J, with H0 calculated in the region Gj , and (Yj , Xj ) is a solution
of the Hamiltonian system with Hamiltonian H0 (J, y, x) on the energy level H0 =
(j)
h0 . The values τ± are the slow time moments when the phase point corresponding
to this solution arrives at the separatrix. At these moments S(Y, X) = 2π J. ˆ (One
can take any of such solutions, they differ by a time shift which does not change the
value of the integrals in (19), (20)). Conditions (12) are understood inductively:
if η (0) ∈ [c−1 −1
1 , 1 − c1 ], and η
(1)
, found according to (13), (14) lies in the segment
−1 −1
[c1 , 1 − c1 ], then (14) is valid, and analogously for η (2) and expressions (15), (16).
All estimates of the form O(·) are uniform on the domain of phase variables under
consideration.
Equations (13), (15) follow directly from the formula for the jump of the adiabatic
invariant at a separatrix [8] if the assumed symmetry properties hold. Equation (14)
follows directly from the formula for the phase change between separatrix crossings
[9].
Equations (16)-(18) follow from results of [9]. Indeed, let h+ be the value of E
when the phase point crosses the axis Cp near C for the first time after exit from
Gν . At this time moment the phase point is in G3 on the positive or the negative
part of the axis Cp. Let h− be the value of E when the phase point crosses the
same part of Cp for the last time before exit from G3 . Introduce the notations:
h+ h−
ζ̄ = 2ε|Θ ∗|
, ζ = 2ε|Θ ∗|
. According to [9],
1
ζ̄ + ζ = Φ2 (Jˆ(1) ) + O(ε1/3 ln−1/3 ε) mod 1, (21)
2ε
provided ζ̄, ζ ∈ [c−1 −1
2 , 1 − c2 ]. The values h
(1)
(see Section 5) and h+ are related as
h+ = h + εΘ∗ + O(ε ) [8]. Therefore, ζ̄ = 12 (1 − η (1) ) + O(ε1/2 ). The values h(2)
(1) 3/2
(see Section 5) and h− are related as h− = h(2) + 2εΘ∗ + O(ε3/2 ) if l = ν, and as

h− = h(2) + εΘ∗ + O(ε3/2 ) if l 6= ν. We have l = ν provided 1/2 < ζ < 1. Consider
this case. We obtain that ζ = 21 (1 + η (2) ) + O(ε1/2 ). From (21) we find
1 1 1
(1 − η (1) ) + (1 + η (2) ) = Φ2 (Jˆ(2) ) + O(ε1/3 ln−1/3 ε) mod 1.
2 2 2ε
Therefore,
1
η (2) = η (1) + Φ2 (Jˆ(2) ) + O(ε1/3 ln−1/3 ε) mod 2. (22)
ε
Then condition 1/2 < ζ < 1 implies equation (17).
Now consider the case l 6= ν. In this case 0 < ζ < 1/2. We get that ζ and η (2)
are related as ζ = 12 η (2) + O(ε1/2 ). From (21) we find
1 1 1
(1 − η (1) ) + η (2) = Φ2 (Jˆ(2) ) + O(ε1/3 ln−1/3 ε) mod 1.
2 2 2ε
Therefore,
1
η (2) = η (1) + Φ2 (Jˆ(2) ) − 1 + O(ε1/3 ln−1/3 ε) mod 2. (23)
ε
Combining (22) and (23) we obtain equation (16). Condition 0 < ζ < 1/2 implies
(18).
Terms O(·) in (13), 14) (correspondingly, in (15), (16)) are functions of Jˆ(0) , η̂ (0)
(correspondingly, Jˆ(1) , η̂ (1) ).
Proposition. Terms O(·) in (13), (14) (correspondingly, in (15), (16)) can be
differentiated an arbitrary number of times with respect to Jˆ(0) , η̂ (0) (correspond-
ingly, Jˆ(1) , η̂ (1) ) without changing form of arguments in symbols O(·).
We omit the proof of this proposition. The proof consists of repetition of the
estimates in [8, 9] and additional estimates for derivatives.
7. Linearly stable periodic trajectories.
7.1. Equations for stable fixed points. Linearly stable periodic trajectories
of the original problem belonging to the families F1, F2 correspond to linearly
stable fixed points of the map M̂ . These points are defined by the following set of
conditions:
Jˆ(2) = Jˆ(0) , η (2) = η (0) , l = ν, (24)

′
| tr M̂ | < 2, (25)
where M̂ ′ is the linearization of M̂ at the fixed point, ν equals 1 or 2. To find

solutions of the family F(1, 2) one should replace the last relation in (24) with
l 6= ν.
Consider, for definiteness, solutions of the family F1. Introduce the following
notation:
Jˆ(s) = εξ (s) , s = 0, 1, 2; (26)

dΦk (εξ)
ε−1 = γk (εξ) , k = 1, 2 , a∗ Θ∗ = 2πα. (27)
dξ
Suppose that (ξ (0) , η (0) ) = (ξ, η) correspond to a fixed point of the map M̂ . We
rewrite equations (24) using formulas (12)-(16), notations (26), and neglecting terms
O(ε1/3 ln−1/3 ε):
sin πη = sin πη (1) , (28)

(1) −1
η − η + γ1 α ln(2 sin πη) = ε Φ1 (εξ) mod 1, (29)
(1) −1
η−η =ε Φ2 (εξ) mod 2, (30)
η, η (1)
∈ (c−1
1 ,1 − c−1
1 ). (31)
Note that equation (30) implies l = ν (cf. (22)).

The stability condition | tr M̂ ′ | < 2 can be written using (13)-(16) (we again
neglect terms O(ε1/3 ln−1/3 ε)):
− 4 < Q < 0, (32)
Q = (u2 − u1 )(γ2 + γ1 ) − u2 u1 γ2 γ1 , (33)
u1 = απ cot πη, u2 = απ cot πη (1) .
The exact set of conditions for a stable periodic solution differs from (28)-(31)
by terms O(ε1/3 ln−1/3 ε) that can be differentiated in εξ , η , η (1) without changing
their smallness order. Below, we find non-degenerate solutions of system (28)-
(31). According to the implicit function theorem, for small enough ε, each of these
solutions corresponds to a solution of the exact system.
7.2. Existence of stable fixed points. We assume that the following generality
conditions are valid.
H1. For almost all I ∈ Ξ0
d
γ1 (I) · γ2 (I) 6= 0 , (γ1 (I)/γ2 (I)) 6= 0 . (34)
dI
Equation (28) implies that either η (1) = η or η (1) = 1 − η. First, consider the
case η (1) = η. System (29), (30) takes the form:
γ1 αln(2 sin πη) = ε−1 Φ1 (εξ) mod 1, (35)
−1
0=ε Φ2 (εξ) mod 2. (36)
From (33) we have:

− 4 < −u2 γ2 γ1 < 0, u = απ cot πη. (37)
If γ2 γ1 < 0, inequality (37) cannot be satisfied.
If γ2 γ1 > 0, inequality (37) gives
1 2
|η − 1/2| < arctan √ . (38)
π απ γ1 γ2
This condition implies (31), provided the constant c1 is large enough.
Now consider the second case, η (1) = 1 − η. System (29), (30) takes the form:
− 2η + γ1 αln(2 sin πη) = ε−1 Φ1 (εξ) mod 1, (39)
−1
2η − 1 = ε Φ2 (εξ) mod 2. (40)
From (33) we have:
− 4 < −2u(γ2 + γ1 ) + u2 γ2 γ1 < 0, u = απ cot πη. (41)
This inequality has solutions for any γ1 , γ2 . Consider, for example, the case γ1 <
0, γ2 > 0, |γ1 | > γ2 . In this case, (41) is satisfied if

2 1 1 2
− < u < 0 or 2 + <u< , (42)
|γ1 | γ1 γ2 γ2
and the corresponding intervals of η are:
1 1 2
0<η−
< arctan ,
2 π απ|γ1 |

1 2 1 1 2 1 1
− arctan < η − < − arctan + . (43)
π απγ2 2 π απ γ1 γ2
These inequalities also imply (31), provided the constant c1 is large enough.
System (28)-(30), equivalent to (35), (36) unified with (39), (40), can be in-
terpreted geometrically as follows. The right hand side of equations (29), (30)
defines a point ε−1 Φ(εξ) = (ε−1 Φ1 (εξ) mod 1, ε−1 Φ2 (εξ) mod 2) on the torus T2 =
{(s1 mod 1, s2 mod 2)}. As ξ varies, this point is moving fast, with its velocity vector
γ(εξ) = (γ1 (εξ) , γ2 (εξ)) varying slowly. L.h.s. of equations (29), (30), and (28),
together with stability condition (32) define a set Λ(εξ) ⊂ T2 parametrized by η
and an index distinguishing systems (35), (36) and (39), (40). It was shown above,
that this set is the union of several (two or three) segments of curves. For brevity we
call the set Λ(εξ) a curve. As ξ varies, Λ(εξ) slowly moves on the torus. Solutions
of (28)-(30), (32) correspond to values of “time” ξ when the point ε−1 Φ(εξ) crosses,
in its motion on the torus, the curve Λ(εξ).
It follows from assumption H1 that at almost all I ∈ Ξ the winding on the torus
with frequency vector γ(I) is ergodic and transversal to Λ(I) at almost all points
of Λ(I). Hence, it is possible to calculate, in the main approximation, the number
of the points of intersection (cf. [14]). This number is (K + o(1))/ε, where
!
1
Z Z
K= |(γ(I), n(l, I))| dl dI. (44)
2 Ξ0 Λ(I)
Here l is the natural parameter on Λ(I), and n(l, I) is a unit normal vector on Λ(I).
The factor 1/2 in this formula is due to the fact that the area of the surface of the
torus equals 2. The value of the inner integral in (44) equals the product of the
length of vector γ(I) and the full length of the projection of the curve Λ(I) onto
the normal to vector γ(I).
The curve Λ(I) is the union of two curves Λnsm (I) and Λsm constructed according
to systems (35)-(37) and (39)-(41) respectively. Hence the value K can also be
represented as K = Knsm + Ksm , where Knsm and Ksm are given by (44) with Λ
replaced by Λnsm (I) or Λsm . The subscripts “sm” and “nsm” are abbreviations for
“symmetric” and “non-symmetric”. In the following section we show that indeed,
system (39)-(41) gives symmetric or nearly symmetric periodic trajectories.
From the results obtained above we find that if γ2 γ1 < 0, Knsm = 0. If γ2 γ1 > 0,
Λnsm is a doubled segment of the s1 -axis on the torus: as η covers the interval (38),
this segment is covered twice. Hence, to calculate the length of Λnsm , one should
multiply by four the magnitude of the difference between the values of the function
γ1 αln(2 sin πη) at the middle point and at the end point of the interval (38). Thus,
this length equals

2
4|γ1 |α ln2 − ln 2 cos arctan √ =
απ γ1 γ2

4
2|γ1 |αln 1 + 2 2 .
α π γ1 γ2
From (44) we obtain:

4
Z
Knsm = αγ1 γ2 ln 1 + dI. (45)
Ξ0 α2 π 2 γ1 γ2
Now we shall obtain a formula for Ksm under the assumptions γ1 < 0, γ2 > 0, |γ1 | >
γ2 ; the results are analogous for the other cases. The projection of a point of the
curve Λsm onto the normal to the vector γ multiplied by the length of γ is
f (η) = γ2 (−2η + γ1 αln(2 sin πη)) − γ1 (2η − 1). (46)

The derivative of this function is zero only at the right end of the second interval
in (43). Therefore, on both intervals (43) the function f is monotonous, increasing
on the first interval and decreasing on the second one. Hence, the product χ(I)
of the length of γ and the length of the projection of Λsm onto the normal to the
vector γ equals the sum of the magnitudes of the differences between values of f at
the endpoints of the intervals (43):

2 2 1 4
χ(I) = − (γ1 + γ2 ) arctan − γ1 γ2 ln 1 + 2 2 2
π απ|γ1 | 2 α π γ1

2 2 2 1 1
+ (γ1 + γ2 ) arctan − arctan +
π απγ2 απ γ1 γ2
2 ! !
1 4 1 1 4
+ γ1 γ2 α ln 1 + 2 2 + − ln 1 + 2 2 2 .
2 α π γ1 γ2 α π γ2
For Ksm we have
1
Z
Ksm = χ(I) dI. (47)
2 Ξ0
The obtained formulas give the number of periodic trajectories of the family F1 in
the main approximation. The number of periodic trajectories of F2 or F(1,2) is
the same in the main approximation. For trajectories of F2 the argument is the
same. For trajectories of F(1,2) equation (30) should be replaced by
η − η (1) = ε−1 Φ2 (εξ) + 1 mod 2 (48)

and equations (36), (40) should be changed accordingly. The rest of the argument
is as before.
If we omit the stability condition | tr M̂ ′ | < 2 in the definition of the curves
Λ, Λnsm , Λsm , the obtained formulas for K, Knsm , Ksm give an asymptotic estimate
of the total number of periodic trajectories and the number of periodic trajectories
with η (1) = η or η (1) = 1 − η satisfying the condition η, η (1) , η (2) ∈ [c−1 −1
1 , 1 − c1 ].
Hence the number of unstable periodic trajectories is bounded from below at least
by a quantity of order ε−1 .
In the limit as ε → 0 one can define in a natural way the distribution density of
the number of periodic trajectories at a fixed value of I as a function of η. From
formulas (44), (35), (36) we find that for trajectories with η (1) = η this density
equals
const(I)| cot πη|.
From formulas (44), (39), (40) we find that for trajectories with η (1) = 1 − η this
density is
2(γ1 + γ2 )
const(I) − cot πη . (49)
απγ1 γ2
7.3. Symmetry properties of the trajectories. Periodic trajectories of the fam-
ilies F1 and F2 that correspond to the set Λsm are symmetric. Indeed, for these
trajectories η (1) = 1 − η and, according to (40),
2η − 1 = ε−1 Φ2 (Jˆ(0) ) + O(ε1/3 ln−1/3 ε) mod 2.
For the value ζ introduced at the end of Section 6 we have

1
ζ = (1 + η) + O(ε1/2 ), (50)
2
therefore,
4ζ − 1 = ε−1 Φ2 (Jˆ(0) ) + O(ε1/3 ln−1/3 ε) mod 2. (51)
(0)
Let ϕ0 denote the value of the “angle” variable ϕ at the initial point M of the
considered trajectory (for definiteness, let the trajectory belong to F1). According
to [9]
1 1
ζ= ϕ0 + Φ2 (Jˆ(0) ) + O(ε1/3 ln−1/3 ε) mod 1. (52)
2π 4ε
Expressing ζ from (51) and substituting in (52) we find:
ϕ0 1
= + O(ε1/3 ln−1/3 ε) mod 1/2.
2π 4
Hence, possible values of ϕ0 are π/2 + O(ε1/3 ln−1/3 ε) and 3π/2 + O(ε1/3 ln−1/3 ε).
Therefore at t = 0, when y = 0, the phase point on the plane of the fast vari-
ables is located nearby the axis p = 0. In fact, at t = 0 we have p = 0. Indeed,
it follows from the symmetry of the problem that together with the periodic solution
(p(t), q(t), y(t), x(t)) also exists the periodic solution (−p(−t), q(−t), −y(−t), x(−t)).
At t = 0, on these two solutions the values of y, x, J coincide, and the values of ϕ
are close. If these two periodic solutions were different, system (39)-(41) would
have two solutions with εξ = Jˆ(0) and close values of η. This contradicts the non-
degeneracy of system (39)-(41). Hence, initially p = 0 and the considered periodic
trajectory is symmetric:
(p(t), q(t), y(t), x(t)) ≡ (−p(−t), q(−t), −y(−t), x(−t)).
Periodic trajectories of the family F(1,2) that correspond to the set Λsm are
nearly symmetric. Using (23) and η + η (1) = 1, we obtain for such trajectories
2η = ε−1 Φ2 (Jˆ(0) ) + O(ε1/3 ln−1/3 ε) mod 2. (53)
From (50) we find
4ζ = ε−1 Φ2 (Jˆ(0) ) + O(ε1/3 ln−1/3 ε) mod 2. (54)
Using (52) we obtain
ϕ0
= O(ε1/3 ln−1/3 ε) mod 1/2.
2π
Therefore, possible values of ϕ0 are O(ε1/3 ln−1/3 ε) and π + O(ε1/3 ln−1/3 ε). Hence,
at t = 0, when y = 0, the phase point on the plane of the fast variables has a value of
q close to qs (x(0) ). If qs ≡ const, the reflection w.r.t. the plane q = qs , y = 0, t = 0
is a canonical transformation. In this case, it can be shown as above that at t = 0
we have q = qs , and the periodic trajectory is symmetric with respect to the plane
q = qs .
8. Stability islands. Let Jˆ(0) = Ir = εξr , η (0) = ηr be one of the stable, in the
linear approximation, fixed points of the map M̂ found in the previous section.
(1) (1)
Let ηr be the corresponding value of the variable η (1) (note that either ηr =
1/3 −1/3 (1) 1/3 −1/3
ηr + O(ε ln ε) or ηr = 1 − ηr + O(ε ln ε)). According to Kolmogorov-
Arnold-Moser (KAM) theory, under certain conditions (absence of resonances up
to the 4th order, a non-zero coefficient in the normal form) this fixed point is stable
and is surrounded with a Cantor family of invariant curves (see, e.g. [13]). KAM-
theory allows also to show that under certain conditions the majority of such fixed
points are stable simultaneously, and to give an estimate from below of a size of
a stability island around a fixed point which is uniform for the set of these fixed
points. To show this we will proceed as follows.
Put ξ˜ = 1ε (Jˆ(0) − Ir ), ξ˜(1) = 1ε (Jˆ(1) − Ir ), ξ˜(2) = 1ε (Jˆ(2) − Ir ), η = η (0) . Substi-
tuting these expressions into (13) - (16)), assuming |ξ| ˜ < 1 and expanding the r.h.s.
in ε we get
ξ˜(1) = ξ˜ − αln(2 sin πη) + O(µ) , (55)
η (1)
= {η + ε−1 Φ1 (Ir ) + γ1 (Ir )ξ˜(1) + O(µ)} , (56)
ξ˜(2) = ξ̃ (1) + αln(2 sin πη (1) ) + O(µ) , (57)

η (2)
= {η (1) + ε−1 Φ2 (Ir ) + γ2 (Ir )ξ˜(2) + O(µ)} . (58)
Here µ = ε1/3 ln−1/3 . The variable α was introduced in (27); in the expression
above α is considered as a function of Ir , α = α(Ir ).
In the new variables the map M̂ is again a composition of two symplectic maps.
It has a fixed point ξ˜ = 0, η = ηr . Because of this
ε−1 Φ1 (Ir ) = ηr(1) − ηr + γ1 (Ir )α(Ir )ln(2 sin πηr ) + O(µ) mod 1, (59)
−1
ε Φ2 (Ir ) = ηr − ηr(1) + O(µ) mod 1. (60)
(1)
Denote the principal terms in the r.h.s. of (59) and (60) as and g1 (ηr , ηr , Ir )
(1)
g2 (ηr , ηr , Ir ) respectively. Thus, maps (55), (56) and (57), (58) take the form
ξ˜(1) = ξ̃ − α(Ir )ln(2 sin πη) + O(µ) , (61)
η (1)
= {η + g1 (ηr , η (1) , Ir ) + γ1 (Ir )ξ˜(1) + O(µ)} ,
r (62)
ξ˜(2) = ξ˜(1) + α(Ir )ln(2 sin πη (1) ) + O(µ) , (63)

η (2)
= {η (1) + g2 (ηr , η (1) , Ir ) + γ2 (Ir )ξ˜(2) + O(µ)} .
r (64)
Consider a truncated map: discard terms O(µ) in (61) - (64). Replace in this new
map symbols Ir , ηr with I, ¯ η̄ and symbol ηr(1) either with η̄ or with 1 − η̄. Denote
nsm (1) sm (1)
the obtained map as M̂trunc for ηr replaced with η̄ and as M̂trunc for ηr replaced
with 1 − η̄. While the original map was defined for a discrete set of parameters
Ir , ηr , the new map is well defined for all I¯ ∈ Ξ0 , η̄ ∈ [c−1 −1
1 , 1 − c1 ] and has a fixed
˜
point ξ = 0, η = η̄.
Consider the rectangle K = {I, ¯ η̄ : I¯ ∈ Ξ0 , η̄ ∈ [c−1 , 1 − c−1 ]}. Let V a be the
1 1
part of K such that for (I,¯ η̄) ∈ V a the eigenvalues of the fixed point ξ˜ = 0, η = η̄
lie on the unit circle; here a is one of the symbols nsm, sm. On the boundary of
V a both eigenvalues equal either +1 or -1. Let λ(a) (I, ¯ η̄) be the eigenvalue which
has a non-negative imaginary part. The conditions of absence of resonances up to
4th order have the form
(λa )k 6= 1, k = 1, 2, 3, 4. (65)
If these conditions are satisfied, then in a neighborhood of the fixed point one
a
can transform the map M̂trunc to the Birkhoff normal form up to terms of the 4th
order (see, e.g. [15]). In the coordinates ρ, χ mod 2π which are close to symplectic
polar coordinates [15] this form is
ρ 7→ ρ + . . . , (66)
χ 7 → ¯ η̄) + Da (I,
a
χ + ω (I, ¯ η̄)ρ + . . . (67)
a a a
Here ω = arg λ , D is a smooth function in the domain where the conditions
(65) are satisfied. Lower dots denote terms of higher order in ρ. Conditions (65)
and condition Da 6= 0 can be written as one condition of the form F a (I, ¯ η̄) 6= 0,
a a
where F is a smooth function in the domain V .
We assume that the following condition is satisfied.
H2. For all I¯ ∈ Ξ0 we have F a (I,¯ η̄) 6= 0 for all η̄ such that (I,
¯ η̄) ∈ V a except,
¯
maybe, at a finite number of points whose location depends on I piecewise smoothly;
a = nsm, sm.
Now we will proceed with the case a = sm, the case a = nsm is completely
analogous. The set Λsm (I) introduced in Section 7.2 is parametrized by values
η from a set Lsm (I) ⊂ R1 ((41) implies that Lsm (I) consists of two intervals).
Fix an arbitrary constant c2 > 0. Denote as L̃sm (I) the part of Lsm (I) where
|F sm (I, η)| > c−1
2 . Condition H2 implies that the set L̃sm (I) consists of several
intervals, and the number of these intervals is bounded uniformly for I ∈ Ξ0 . The
length of Lsm (I) \ L̃sm (I) can be made arbitrarily small by taking a sufficiently
¯ the map M̂trunc
large value of c2 . For η̄ ∈ L̃sm (I), sm
in the neighborhood of fixed
˜
point ξ = 0, η = η̄ can be written in the form
ρ 7→ ρ + O(ρ5/2 ), (68)
χ 7→ ¯ η̄) + Da (I,
χ + ω a (I, ¯ η̄)ρ + O(ρ3/2 ), (69)
¯
where the estimate O(·) is uniform over the set of values I, η̄ under consideration.
sm
Any symplectic map of the form M̂trunc + O(µ) such that: a) ξ̃ = 0, η = η̄ is its
¯ η̄ under consider-
fixed point; b) the estimate O(µ) holds uniformly on the set of I,
ation and can be differentiated an arbitrary number of times without changing the
order of smallness, can be written in the form
ρ 7→ ρ + O(ρ5/2 ) + O(ρµ), (70)
χ ¯ η̄) + Da (I,
7→ χ + ω a (I, ¯ η̄)ρ + O(ρ3/2 ) + O(µ). (71)
This map for small enough µ satisfies KAM stability conditions: there are no reso-
nances up to the 4th order, the coefficient in the normal form that ensures the twist
property is different from 0. So, the fixed point is stable. Obviously, the required
choice of µ can be made the same for all pairs (I, ¯ η̄) under consideration.
The original map M̂ in the neighborhood of the fixed point Ir , ηr such that
(1)
ηr = 1 − ηr + O(µ) (see notations in the beginning of this Section), provided
ηr ∈ L̃sm (I), can be written in the form (70), (71) with I¯ = Ir , η̄ = ηr . So, this
fixed point is stable for small enough ε, and the required choice of ε can be made
the same for all pairs (Ir , ηr ) under consideration.
Now introduce ρ̄ = ρ/κ, where κ is a small constant to be defined later. Assuming
that ρ̄ = O(1) we will write map M̂ in the form
ρ̄ 7→ ρ̄ + κ3/2 O(1) + O(µ), (72)
χ 7→ χ + ω a (I¯r , η̄r ) + κDa (I¯r , η̄r )ρ̄ + κ3/2 O(1) + O(µ). (73)
In the annulus 1 ≤ ρ̄ ≤ 2 this map satisfies the conditions of the Moser theorem [16]
for the case of a small twist, provided κ is taken small enough and after that ε is
taken small enough. So, in this annulus there exists an invariant curve of the map.
Obviously, the choice of κ, ε to meet assumptions of Moser theorem can be made
the same for all pairs (Ir , ηr ) under consideration. So, there is an invariant curve
at a distance larger than c−1 3 from the point ξ˜ = 0, η = ηr . We call the domain
surrounded by this curve a stability island of the Poincaré map. The area of this
island is larger than c−1
4 . Returning from ξ̃, η to the original variables I, η, we find
that the area of the stability island is estimated from below as c−14 ε.
The boundary of the stability island (an invariant curve) is a section of the
invariant torus on the energy level of the original system. The volume of the domain
surrounded by this torus is larger than C1−1 ε. The variation of Iˆ inside this domain
is smaller than C2 ε. This completes the proof of Theorem 2.
9. Example: stable periodic trajectories and stability islands in the prob-

lem of the motion of charged particles in the parabolic model of magnetic
field.
9.1. The model. The parabolic model (see, e.g. [10]) is one of the simplest mod-
els of the magnetic field in the Earth magnetotail. Introduce a right Cartesian
coordinate system Oxξq with the origin located in the Earth’s equatorial plane,
the Earth being located between the origin and the Sun; the Ox-axis lies in the
equatorial plane and is directed towards the Earth, and the Oq-axis is orthogonal
to the equatorial plane. In this coordinate system the magnetic field vector has the
form B = (B0 q/L, 0, Bn). Here B0 , Bn , L are positive constants. The field lines of
this field are parabolas Bn x − B0 q 2 /(2L) = const (see Figure 3).
-2
-4
0 2 4 6 8
Figure 3. Fieldlines of the magnetic field in the parabolic model.
The vector potential of the magnetic field has the form

A = (0, Bn x − B0 q 2 /(2L), 0). The motion of a charged particle in such a field
is described by the Hamiltonian system with the following Hamiltonian function:
" 2 #
q2

1 2 e 2
H= y + η− Bn x − B0 +p . (74)
2m c 2L
Here m and e are the mass and the charge of the particle, c is the speed of light. The
pairs of canonically conjugated variables are (y, x), (η, ξ), (p, q). The Hamiltonian
does not depend on ξ, and hence η ≡ const along the motion. Without restricting
the generality we put η ≡ 0.
Introduce new pairs of canonically conjugated coordinates (y1 , ε−1 x1 ), (p1 , q1 ),
new time t1 and a new Hamiltonian H1 according to the formulas:
x1 = εx(ρL)−1/2 , y1 = y(mv)−1 ,
−1/2
q1 = q(ρL) , p1 = p(mv)−1 ,
t1 = tv(ρL)−1/2 , H1 = H(mv 2 )−1 .
Here v is a p
typical velocity of the particle, ρ = cmv/(eB0 ) is a typical Larmor
radius, ε = Bn
B0 L/ρ; v, ρ = const. In the new variables the Hamiltonian takes the
form (we omit the subscript “1”):
" 2 #
1 2 2 1 2
H= y +p + x− q . (75)
2 2
Following [10], we consider the case ε ≪ 1 (it is valid for ions). In this case, p, q are
the fast variables, and y, x are the slow ones. The phase portrait of the fast system
is represented in Figures 4a and 4b for x > 0 and x < 0 respectively.
p (a) p (b)
G3
G1 G2
q q
Figure 4. Typical phase portraits of the fast system for a) x > 0

and b) x < 0.
Consider the motion on the energy level H = 1/2. On the configuration plane
(x, q) the motion is possible between two parabolas (see Figure 3). The two field
lines are also the zero velocity curve (zvc). On the plane of the slow variables
the motion is possible in the domain obtained by the union of the half-disk {x, y :
x2 + y 2 ≤ 1, x ≤ 0} and the half-strip {x, y : |y| ≤ 1, x > 0}. The uncertainty curve
Γ is the semicircle {x, y : x2 + y 2 = 1, x ≥ 0}. The phase trajectories on the fast
plane that lie inside the domains G1 and G2 correspond to the points of the plane
of the slow variables with x2 + y 2 > 1 (see Figure 4a). The phase trajectories that
lie in the domain G3 correspond to the points on the plane of the slow variables
with x2 + y 2 < 1, x > 0.
The phase portrait of the slow system at H = 1/2 is presented in Figure 5 (see
[10]). The boundary of the domain of possible motion on the portrait corresponds
to the motion with zero value of the “action”. For points of the plane of the slow
variables with y = ±1, x > 0, the corresponding phase point on the fast plane in
the adiabatic approximation always coincides with one of the two stable fixed points
of the fast system. For points of the semicircle {x, y : x2 + y 2 = 1, x < 0}, the
corresponding phase point on the fast plane is always at the only stable fixed point
of the fast system. The stable fixed point on the phase portrait of the slow system
has the coordinates y = 0, x = xe ≈ 0.65.
1.5
y
1
Γ
0.5
xe
−0.5
−1
x
−1.5
−1 0 1 2 3 4
Figure 5. Phase plane of the slow variables.
In [10], the chaotic dynamics in the domain filled with trajectories of the slow
system that cross the uncertainty curve was found and studied.
9.2. Formulas for the dynamics in the slow system and separatrix cross-
ings. The following formulas are valid for the values S, Θ∗ , a∗ , α introduced in Sec-
tions 2, 5, and 6 [10]:
8 3/2 1/2
S= x , Θ∗ = 4y∗ x∗ , (76)
3
−1/2 2
a∗ = x∗ , α= π y∗ . (77)
The modified “action” Iˆ in the domain Gi is [10]
4
Iˆ = (2E)1/4 fi (k), (78)
3π
where
f3 (k) = (1 − k 2 )K(k) + (2k 2 − 1)E(k),
f1,2 (k) = 2(1 − k 2 )kK(1/k) + (2k 2 − 1)kE(1/k),
s
1 x 1 2
p + (x − q 2 /2)2 ;

k= 1+ √ , E=
2 2E 2
K( · ), E( · ) are the complete elliptic integrals of first and second kinds, respectively.
It turns out that in the considered specific problem we have: γ1 < 0, γ2 >
0, |γ1 | ≫ 1, γ2 ≪ 1 (for the definition of γi , see (26)). Therefore, according to
Section 7.2, there are no non-symmetric stable periodic trajectories. Symmetric
stable periodic trajectories correspond to the two intervals (43) of the variable η.
The corresponding intervals of the variable u = απ cot πη are of equal small length
2/|γ1 | (see (42)) and in the variable η they are ≈ 2/(απ 2 |γ1 |) and ≈ αγ22 /(2|γ1 |),
respectively. Hence, the number of stable periodic trajectories is much smaller than
the number of unstable periodic trajectories. The minimum of f in (46) is attained
at the right endpoint of the second of these two intervals. Therefore, the variation
of f on this interval is much smaller than its variation on the first one. Hence, the
second interval corresponds to a much smaller number of stable periodic trajectories
than the first one. Taking into account the relations between γ1 and γ2 , we find for
Ksm (see (45)) in the main approximation:
1 4
Z
Ksm = ˆ
dI. (79)
2 Ξ0 απ 2
The quantities Iˆ and x∗ are related as 2π Iˆ = S(x∗ ). Using (76), (77), we obtain
1 8 3/2 4
Iˆ = x∗ = (1 − y∗2 )3/4 .
2π 3 3π
Hence,
dIˆ 2
Z Z
Ksm = = 2 (1 − y∗2 )−1/4 dy∗ . (80)
Ξ0 πy∗ π Ξ0
Let v = y∗2 . Then
v2
1
Z
Ksm = (1 − v)−1/4 v −1/2 dv, (81)
π2 v1
where v1 and v2 are the values of v = y∗2 at the endpoints of Ξ0 . In particular, if
we take for Ξ0 the complete interval of values of Iˆ corresponding to the separatrix
crossings, we have v1 = 0, v2 = 1. Accordingly, for the value Kc of the function
Ksm we find:
Z 1
1 1 1 Γ(1/2)Γ(3/4)
Kc = 2 (1 − v)−1/4 v −1/2 dv = 2 B(1/2, 3/4) = 2 . (82)
π 0 π π Γ(5/4)
Here, as usual, B( · ) and Γ( · ) denote the beta and gamma functions respectively.
Using the reflection formula for the gamma function [17], we find
r
1 √ 1 π 2 1
Kc = 2 π = ≈ 0.21 . (83)
π Γ(5/4)Γ(1/4) sin(π/4) π 4(Γ(5/4))2
In the next section we consider solutions with initial conditions p = y = 0, q = q0 .
On the energy level H = 1/2, for the corresponding initial value x0 of x we have
the relation 1/2(x0 − q02 /2)2 = 1/2 and, hence x0 = q02 /2 ± 1. Take the “minus”
sign in the latter expression. Calculate for such a solution the value v = y∗2 in the
adiabatic approximation. We have
4/3
2 2 3π ˆ
v = y∗ = 1 − x∗ = 1 − I ,
4
s
1 2 1 1 x0 |q0 |
E=H− y = , k= 1+ √ = ,
2 2 2 2E 2
4 4
Iˆ = (2E)1/4 f3 (k) = f3 (|q0 |/2).
3π 3π
Hence,
v = 1 − (f3 (|q0 |/2))4/3 . (84)
9.3. Results of the numeric studies. For the exact system the 2D surface y = 0
will be used as a surface of the Poincaré section, despite it is not globally transversal
to the phase flow. Figure 6 displays a representation of a part of this surface. The
horizontal variables are q and p and x is the vertical one. We see that, beyond the
fact that it is not a global Poincaré section, to every point (q, p) on the strip |p| < 1
correspond two possible values of x.
0
1
0.5
-1-2
0
-1
0 -0.5
1
2 -1
Figure 6. The surface y = 0 used as the surface of the Poincaré section.
These effects are illustrated in Figure 7 for ε = 0.04. The same value of ε has
been used up to Figure 13. For smaller ε many pictures look quite wild. The initial
point ε−1 x = 10, q = 0, y = 0, p > 0 seems to be on an invariant torus. The section
by the surface y = 0, disregarding the value of x and the sign of ẏ is shown to the
left. The second plot contains only the points for which ẏ > 0. Third one contains
points with ẏ either positive or negative, but only for the values of x < xe (see
Section 9.1). The rightmost plot displays only points with ẏ > 0 and x < xe .
1 1 1 1
0.5 0.5 0.5 0.5
0 0 0 0
-0.5 -0.5 -0.5 -0.5
-1 -1 -1 -1
-2 -1 0 1 2 -2 -1 0 1 2 -2 -1 0 1 2 -2 -1 0 1 2
Figure 7. Various versions of the Poincaré section. See the text.
Using a 3D representation, Figure 8 shows a Poincaré section, disregarding the

value of x and the sign of ẏ at y = 0. It has been obtained using initial conditions
of the form q = y = 0 and ε−1 x = 3, 4, . . . , 15 and computing 104 iterates and then
points with the same initial conditions for q and y but with ε−1 x < 3 and step 0.05.
We note that most of these points are in chaotic regions and their orbits approach
large values of ε−1 x. Typically, the computation is stopped when ε−1 x > 104 (or
when the number of integration steps using a Taylor method at order 28 exceeds
106 steps to compute a Poincaré iterate) to obtain this figure. In the figure there are
regions where the density of points is smaller. They correspond to a neighborhood
of the lines where y = 0 is no longer transversal to the flow.
Close to x = xe , y = 0 the system has a periodic orbit. It oscillates mildly around
the plane y = 0 and intersects it four times. This can be clearly seen in Figure 8,
with the corresponding narrow island around it, in the zone of regular motion.
Other tiny islands appear inside the chaotic region and the study of them is the
main goal of this paper. Some of them can be also seen in Figure 8. However, for a
better understanding, we shall consider only the (q, p) projections of the Poincaré
section, using the constrains x < xe , ẏ > 0. This can be seen in Figure 9. The outer
part of Figure 9 corresponds to trajectories without separatrix crossings. This part
1
0.5
-1-2
0
-1
0 -0.5
1
2 -1
Figure 8. The Poincaré section of phase trajectories with initial

conditions q = y = 0, while for ε−1 x we use the values 3, 4, . . . , 15,
and also ε−1 x < 3 with step 0.05. Axes as in Figure 6.
1
0.5
-0.5
-1
-2 -1 0 1 2
Figure 9. The Poincaré section of Figure 8 projected onto the

(q, p)-plane.
is mainly filled with segments of invariant curves as it should be by the Arnold

perpetual adiabatic invariance theorem. The inner part of Figure 9 is a chaotic sea
[10].
The two central holes inside the chaotic region correspond to points coming from
large values of ε−1 x (either in q > 0 or in q < 0). The objects inside the chaotic sea
which look as segments of fat curves are actually islands with a fixed elliptic point
at its center.
A part of Figure 9 is shown in Figure 10, left. Four islands are clearly seen in that
plot. A magnification of the rightmost one is shown in Figure 10, right. Because
the island is narrow and close to a parabola with horizontal axis, instead of (q, p)
we display (q + 0.395p2 , p).
For a clear identification of the different zones in the configuration plane (ε−1 x, q)
we show, in Figure 11, the results presented in Figure 9 in these new variables. The
continuous line corresponds to the left boundary of the zero velocity curve (zvc)
1
0.3
0.2
0.5
0.1
0 0
-0.1
-0.5
-0.2
-0.3
-1
1 1.1 1.2 1.3 1.4 1.5 1.404 1.405 1.406 1.407
Figure 10. A zoom of the previous picture (left) and magnifica-

tion of the rightmost island (right). For the last plot the variables
are (q + 0.395p2, p).
p = y = 0, while the discontinuous line shows the location of ẏ = 0. One can check
once more the sharp transition between regular and chaotic motion. The islands
seen in this picture are hardly visible in Figure 9: they are on the top close to p = 1.
Note that there are islands of period 2, without points on q = 0. The two zones
without points of the chaotic sea which appear close to ε−1 x = −25 correspond to
the two holes near the center of Figure 9. They will be filled only if orbits reaching
extremely large values of ε−1 x (and, hence, extremely large values of the return
time) are computed.
-1
-20 -10 0 10
Figure 11. Results of Figure 9 projected onto the configuration

plane (ε−1 x, q).
As an illustration, Figure 12, top left, shows the first (going from left to right)
stable periodic orbit of Figure 10 left, and on the right part the third one. On the
bottom we display, for the first of these orbits, the projections on (ε−1 x, y) (left)
and on (q, p) (right). On the (ε−1 x, y) plane, it can be seen that the orbit is close to
the one theoretically predicted (except for some minor oscillations). On the (q, p)
plane the transition of the fast system across the separatrix is clearly seen.
3 3
2 2
1 1
0 0
-1 -1
-2 -2
-3 -3
-20 0 20 40 60 -20 0 20 40 60
1 1
0.5 0.5
0 0
-0.5 -0.5
-1 -1
-20 0 20 40 60 -1 0 1 2 3
Figure 12. Various projections of stable periodic orbits. See the text.
We have computed simple periodic orbits (i.e., fixed points of the Poincaré map)
with initial conditions on the zvc: y = p = 0. They appear on Figures 9 and 10 on
the horizontal axis. Due to the symmetry it is enough to start with q > 0. We note
that for y = q = p = 0 there is a periodic orbit fully contained in the (x, y) plane.
In a similar way one can compute periodic orbits starting on the q = 0 axis.
Again, due to the symmetry, it is enough to use p > 0. To grasp the wild character
of the flow Figure 13 shows, on the left, the first image under the Poincaré map of
the line q = 0 with values of p ∈ (−1, 1) and step size 10−5 . On the right we display,
in log10 scale, the maximum value of ε−1 |x| reached (lower curve) and the return
time (upper curve) as a function of p. The plot would be similar taking points of
the zvc (p = 0 in this section).
To
√ compute p.o. on the zvc we have restricted our interest to the domain q ∈
[1, 2]. (Smaller values of q correspond to extremely large return times.) This is
enough to illustrate the theory. This range of q is scanned using a control on the
distance between consecutive points, in order to follow carefully all the foldings of
the image. Steps in q as small as 10−12 are accepted. The results of the computation
are shown in Table 1. As predicted by the theory, both the total number of p.o.
and the number of stable ones (s.p.o.) behave as ε−1 .
The p.o. in p = 0 appear also (if we display the period or the maximal value
of x along the orbit) in bumps similar to the ones displayed in Figure 13, right.
Decreasing ε the number of bumps increases like ε−1 . On each bump the behavior
of the system is similar.
1
5
0.5 4
0 3
-0.5 2
-1 1
-1.5 -1 -0.5 0 0.5 1 1.5 0 0.2 0.4 0.6 0.8 1
Figure 13. The first image under the Poincaré map of the line
q = 0 with values of p ∈ (−1, 1) and step size 10−5 (left). The
maximum value of ε−1 |x| reached (lower curve) and the return
time (upper curve) as a function of p (right). In the vertical axes
of the right plot (both for the maximum and for the return time)
the log10 scale has been used.
ε periodic orbits stable periodic orbits Ksm /ε

0.00025 24235 370 396
0.001 6351 94 99
0.0025 2662 38 39.6
0.01 632 11 9.9
0.04 144 4 2.5
Table 1. Number of periodic orbits.
4
2
0 0
1 1.1 1.2 1.3 1.4 0 0.2 0.4 0.6 0.8 1
Figure 14. Left: the cumulative number of p.o. found from q = 1

on, multiplied by the corresponding value of ε as a function of q.
Right: the distribution of p.o. inside one bump and the theoretical
curve.
A representation of the location of the p.o. is displayed on Figure 14, left. On

it the cumulative number of p.o. which are found from q = 1 on, multiplied by
the corresponding value of ε, is displayed as a function of q. The curves are shown
for ε = 0.01, 0.0025, 0.001 and 0.00025. The smaller is the value of ε, the smaller
is the effect of the bumps. Then, on the right part of Figure 14 we have collected
the distribution of p.o. inside a bump as follows. If ε is small the bumps are also
small. The extrema of the bump can be taken as the places where the return time
has a local minima (or, equivalently, the maximal x is minimum or, also, the trace
of the nearby p.o. is also a minimum). If q1 and q2 are the extrema of some bump
(q1 < q2 ), and q ∗ denotes the position of a p.o., it can be referred to the bump
by means of the parameter q̃ = (q ∗ − q1 )/(q2 − q1 ). Collecting all the values of q̃
for the different p.o. we can get the desired distribution. This is plotted to √ the
right in Figure 14. Despite a total of 359 complete bumps is found for q ∈ [1, 2]
for ε = 0.00025 (this reduces to 89 for ε = 0.001), only the bumps for q ∈ [1, 1.1]
have been used. The reason is that for that range of q the computed value of q̃ is
close enough to the theoretical parameter η introduced in Section 5. In this range
there are 58 full bumps which contain 10373 p.o. We selected 200 intervals of equal
length in the q̃ variable. The scale has been normalized to have total area equal to
1. In the same plot, and in discontinuous lines, the theoretical density of the p.o.
as a function of η is also shown. It is given by an expression of the form (49)
ρ(η) = C |c − 1/tan(πη)| ,
where c = 2(γ1 + γ2 )/(πγ1 γ2 ) (in notations of Section 7) and C is a normalization

constant. The approximate values c = 1.8, C = 0.3 have been used. Due to the
relatively small number of p.o. the experimental density shows some irregularities,
but the global agreement is remarkable. We recall that close to the endpoints
of the interval the theoretical approximation fails. The density becomes zero at
1 −1
π tan (1/c) ≈ 0.16, in good agreement with numerical results.
Concerning the s.p.o., in each bump one should find two narrow intervals in η
(43) which can contain such orbits. The largest one is close to η = 1/2. However,
as the intervals are narrow and the transformation from q̃ to η depends also on the
location of the bump, it is not suitable to look at results considering all the bumps
together. As before, it is enough to use only a part of the bumps (for instance,
with q0 ∈ [1, 1.1]). It is checked that the stable p.o. are located in the interval
q̃ ∈ [0.5, 0.515]. A very narrow interval of s.p.o. appears also near q̃ = 0.2, but it
contains less than 2% of the total number of s.p.o. This is in agreement with the
theoretical prediction in Section 9.2.
We compare the total number of s.p.o. √ to the theoretical predictions as follows.
Orbits starting on the zvc for q0 ∈ [1, 2] cross the separatrix (x2 + y 2 = 1) at
values of y ∈ [y1 , y2 ]. Let vi = yi2 . According to (84), we have in the adiabatic
approximation
4/3
√ √

1
v1 = 1 − (f3 (1/ 2))4/3 = 1 − K(1/ 2) ≈ 0.096 ,
2
4/3
3 1
v2 = 1 − (f3 (1/2))4/3 =1− K(1/2) − E(1/2) ≈ 0.570 .
4 2
Calculating the integral in (47) with these values of v1,2 we obtain Ksm ≈ 0.099.
The quantity Ksm /ε is, in the main approximation, the √ number of s.p.o.
√ of the
family F1 (or F2) that start from the zvc with q0 ∈ [1, 2] and q0 ∈ [− 2, −1]. It
is also equal in the main approximation
√ to the number of s.p.o. of the two families
that start from the zvc with q0 ∈ [1, 2]. For Ksm /ε we find values which are in
good agreement with the numerical results (see Table 1).
Appendix. The planar periodic orbit. Here we discuss the special role played
by the periodic orbit sitting on (q, p) = (0, 0). As it is contained in the (x, y)
plane, we shall call it the planar p.o. If y(0) = 0, it is given by x = cos(εt),
and the normal variational equation is given (as a second order equation) by ξ¨ =
Rcos(εt)ξ.
πp
This is a particular case of Mathieu equation, with zero average. Let M =
0 sin(τ )dτ ≈ 2.39628. For small ε the monodromy matrix can be approximately
written as the product of rotation by the angle M/ε and an hyperbolic matrix
of dominant eigenvalue exp(M/ε). Figure 15, left, displays the variation of the
trace tr of the monodromy matrix as a function of 1/ε. In the horizontal axis we
display 1/ε, while in the vertical one sinh−1 (tr) is shown. It is straightforward to
show that, despite infinitely many intervals of stability occur they are extremely
narrow. More precisely: stability domains appear near values of 1/ε of the form
M −1 (k + 1/2)π, k ∈ Z, k > 0, while the width of these intervals is, up to a constant
factor, of the form exp(−M/ε).
0.2
200
0.1
100
0 0
-100
-0.1
-200
-0.2
0 20 40 60 80 100 -0.001 -0.0005 0 0.0005 0.001
Figure 15. Left: variation of the trace of the monodromy matrix

as a function of 1/ε for the planar periodic orbit. Right: a stability
island for ε = 0.5155 inside a “relatively large” domain of stability.
In each one of these stability intervals there is an invariant island around (0, 0) in
the (q, p) plane. As an example, Figure 15 right shows the island for ε = 0.5155, in
the stability domain with the left endpoint at ε = 0.5136759883096067 and width
≈ 2.25 × 10−3 . Even for that large value of ε the width of the island in q is small
(less than 0.002). Furthermore there exists a period 4 p.o. (displayed as black big
points) which is hyperbolic. The presence of nearby hyperbolic periodic orbits has
been checked for other values of ε in the stability domains. Hence, when the planar
p.o. becomes stable (for these narrow intervals) there appear other unstable p.o.
which play a similar role. Using an index argument in a suitable ball containing
the island inside, one can prove that the global dynamics outside this ball is only
affected for very narrow intervals of initial conditions. From now on we shall not
consider these stability intervals.
Dependence of the return time on the initial value of q contains certain charac-
teristic structures that we call bumps, see Figure 13, right. Figure 16 shows, in the
upper part, a sample of results, and several details in the lower part. Concretely, on
the left upper plot
√ we present the results for trajectories with initial conditions at
p = 0 and q ∈ [0, 2] for ε = 0.3, 0.2, 0.1, 0.04. We plot the logarithm of the return
time as a function of the initial value of q. The smaller ε is, the higher are the
bumps and the smaller is the distance between endpoints of a bump, as it was said
before. For instance, for ε = 0.04 the maximal return√ time from the zvc exceeds
107 . In the central upper plot we show, for q ∈ [1, 2] the logarithms of the return
time multiplied by ε for ε = 0.04, 0.01, 0.0025. The curves on these plots tend to a
limit curve when ε → 0. On the right upper part we show the end of the big bump
close to q = 1.15 for ε = 0.04, displaying the return time as a function of q on the
narrow window [1.15027, 1.1502744]. Five ends of small bumps can be seen to the
left of the endpoint of the big bump, one of them very close to this endpoint. All
this detailed structure is not visible in the previous plot (the middle upper one).
3 170
15
160
2.5
10
150
5 2
140
0 0.5 1 1 1.2 1.4 1.15027 1.150274
170
148
150
146
165
145
144
Figure 16. The bumps and their structure. See the text for a
detailed explanation.
In the lower part of the figure, and from left to right we show a detail of the end
of the leftmost small bump of the previous plot (window [1.150270992 ± 2 × 10−9 ]),
a view of the end of the rightmost small bump together with the end of the big
bump (window [1.150274085 ± 5 × 10−9]) and a great enlargement of the end of the
big bump (window [1.150274092745 ± 25 × 10−12]. The vertical jumps in the return
time, clearly visible in last plot, are due to the fact that px = 0 is not everywhere
transversal to the flow, as it was discussed (see Figure 7 right, for instance). To
the right of the endpoint of the big bump the pattern of small bumps is almost
symmetric to the one displayed here.
Now we pass to the reasons for the creation of big and small bumps in geometric
terms. Taking, for instance, an initial point on the zvc, it can approach the separa-
trix x2 + y 2 = 1 moving close to the planar p.o. Then it can move back to smaller
values of x, or it can enter one of the branches of the region x > 1, either the upper
one or the lower one. A critical orbit would tend to the planar one along its stable
manifold. Hence, looking at the maximal value of x along the orbit, it would never
exceed 1. Furthermore, moving on the stable manifold of the planar p.o. it would
reach a point with y = 0, ẏ > 0 close to x = −1.
Hence, the endpoints of the big bumps can be seen as the place where there is
a minimum of the maximum value of x along the orbit (as a function of the initial
value of q in the zvc) or as the place where orbits switch from entering the upper
branch to the lower one or vice versa (see below for an additional explanation). Or,
in other words, to switching between entering one or the other of the lobes at the
crossing of the separatrix.
Small bumps are related to the crossing of the separatrix when we go outside, to
the circulation region, that is, x decreases and enters the domain x < 1. An orbit
entering from one of the branches to the region x < 1 can also approach the planar
p.o. along the stable manifold. This occurs depending on the phase it has when it
reaches the maximal value of x in one of the branches. As it changes quickly for
ε small, there are many small bumps inside one big bump. Again, entering x < 1
we can pass close to the planar p.o. and leave its vicinity following one of the two
branches of their unstable manifold.
REFERENCES
[1] A. I. Neishtadt, V. V. Sidorenko and D. V. Treschev, Stable periodic motions in the problem
of passage through a separatrix, Chaos, 7 (1997), 2–11.
[2] A. I. Neishtadt, V. V. Sidorenko and D. V. Treschev, On stability islands in the domain
of separatrix crossings, Nonlinear Mechanics, ed. V. M. Matrosov, V. V. Rumyantsev, et al
(Moscow: Fizmatlit), (2001), 192–203 (in Russian).
[3] A. I. Neishtadt, C. Simó, D. V. Treschev and A. A. Vasiliev, Stability islands in domains
of separatrix crossings in slow-fast Hamiltonian systems, Proceedings of the V. A. Steklov
Mathematical Institute, 259, (2007), 236-247.
[4] A. V. Timofeev, On the constancy of an adiabatic invariant when the nature of the motion
changes, Sov. Phys. JETP, 48 (1978), 656–659.
[5] J. R. Cary, D. F. Escande and J. Tennyson, Adiabatic invariant change due to separatrix
crossing, Phys. Rev., A34 (1986), 4256–4275.
[6] A. I. Neishtadt, Change of an adiabatic invariant at a separatrix, Sov.J.Plasma Phys., 12
(1986), 568–573
[7] J. R. Cary and R. T. Skodje, Phase change between separatrix crossings, Physica D, 36
(1989), 287–316.
[8] A. I. Neishtadt, On the change in the adiabatic invariant on crossing a separatrix in systems
with two degrees of freedom, PMM USSR, 51 (1987), 586–592.
[9] A. I. Neishtadt and A. A. Vasiliev, Phase change between separatrix crossings in slow-fast
Hamiltonian systems, Nonlinearity, 18 (2005), 1393–1406.
[10] J. Büchner and L. M. Zelenyi, Regular and chaotic charged particle motion in magnetotail
like field reversals, 1, Basic theory of trapped motion, Journal of Geophysical Research, 94
(1989), 11821–11842.
[11] J. Wisdom, A perturbative treatment of motion near the 3/1 commensurability, Icarus, 63
(1985), 272–289.
[12] A. I. Neishtadt and V. V. Sidorenko, Wisdom system: dynamics in the adiabatic approxima-
tion, Celestial Mechanics and Dynamical Astronomy, 90 (2004), 307–330.
[13] V. I. Arnold, V. V. Kozlov and A. I. Neishtadt, “Mathematical Aspects of Classical And
Celestial Mechanics,” (Encyclopedia of Mathematical Sciences 3), (Berlin: Springer), 2006.
[14] F. A. Tal and E. Vanden-Eijnden, Transition state theory and dynamical corrections in ergodic
systems, Nonlinearity, 19 (2006), 501–509.
[15] V. I. Arnold, “Mathematical Methods of Classical Mechanics,” (New York: Springer), 1978.
[16] J. Moser, On invariant curves of area-preserving mappings of an annulus, Nachr. Akad. Wiss.
Göttingen, Math.-Phys. Kl. II, (1962), 1–20.
[17] M. Abramowitz and I. A. Stegun, “Handbook of Mathematical Functions, With Formulas,
Graphs, and Mathematical Tables,” (Dover Publications), 1965.
Received November 2006; revised June 2007.

E-mail address: [email protected]

NSTV 2008 DCDS B

Uploaded by

Copyright:

Available Formats

NSTV 2008 DCDS B

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

NSTV 2008 DCDS B

Uploaded by

Copyright:

Available Formats

DISCRETE AND CONTINUOUS Website: http://aimSciences.

PERIODIC ORBITS AND STABILITY ISLANDS IN CHAOTIC

Abstract. We consider a 2 d.o.f. natural Hamiltonian system with one degree

2000 Mathematics Subject Classification. Primary: 34E10 ; Secondary: 37J40 .

1. Introduction. Under rather general assumptions, in the phase space of a math-

properties of the periodic trajectories are discussed. In Section 8, Theorem 2 is

2. Main equations. Consider a 2 d.o.f. Hamiltonian system with the Hamiltonian

Figure 1. Phase portrait of the fast system.

by 2π. The action is discontinuous on the separatrix, and hence it is convenient

3. Formulation of results. We look for stable periodic trajectories of system (2),

4. Adiabatic and improved adiabatic approximations. In the fast system,

¯ ȳ, x̄) + εH1 (I,

H = H0 (J, ŷ, x̂) + ε2 H2 (J, ψ, ŷ, x̂, ε). (9)

an accuracy O(ε2 ), the following formula for J is valid (see [8]):

Figure 2. Schematic picture of the motion on the slow plane.

6. The return map. In the energy level H = h0 the segment S3 is represented by a

η (0) , η (1) , η (2) ∈ [c−1 −1

l = ν, if 0 < {(η (1) + ε−1 Φ2 (Jˆ(2) ))/2} < 1/2, (17)

Here√{ · } denotes the fractional part, a∗ = a(x∗ ). The value of a is given by

(see Section 5) and h− are related as h− = h(2) + 2εΘ∗ + O(ε3/2 ) if l = ν, and as

7. Linearly stable periodic trajectories.

Jˆ(2) = Jˆ(0) , η (2) = η (0) , l = ν, (24)

where M̂ ′ is the linearization of M̂ at the fixed point, ν equals 1 or 2. To find

Jˆ(s) = εξ (s) , s = 0, 1, 2; (26)

sin πη = sin πη (1) , (28)

Note that equation (30) implies l = ν (cf. (22)).

From (33) we have:

f (η) = γ2 (−2η + γ1 αln(2 sin πη)) − γ1 (2η − 1). (46)

η − η (1) = ε−1 Φ2 (εξ) + 1 mod 2 (48)

For the value ζ introduced at the end of Section 6 we have

ξ˜(2) = ξ̃ (1) + αln(2 sin πη (1) ) + O(µ) , (57)

ξ˜(2) = ξ˜(1) + α(Ir )ln(2 sin πη (1) ) + O(µ) , (63)

9. Example: stable periodic trajectories and stability islands in the prob-

Figure 3. Fieldlines of the magnetic field in the parabolic model.

The vector potential of the magnetic field has the form

Figure 4. Typical phase portraits of the fast system for a) x > 0

Figure 5. Phase plane of the slow variables.

Figure 6. The surface y = 0 used as the surface of the Poincaré section.

0.5 0.5 0.5 0.5

-0.5 -0.5 -0.5 -0.5

Figure 7. Various versions of the Poincaré section. See the text.

Using a 3D representation, Figure 8 shows a Poincaré section, disregarding the

Figure 8. The Poincaré section of phase trajectories with initial

Figure 9. The Poincaré section of Figure 8 projected onto the

is mainly filled with segments of invariant curves as it should be by the Arnold

Figure 10. A zoom of the previous picture (left) and magnifica-

Figure 11. Results of Figure 9 projected onto the configuration

ε periodic orbits stable periodic orbits Ksm /ε

Figure 14. Left: the cumulative number of p.o. found from q = 1

A representation of the location of the p.o. is displayed on Figure 14, left. On

where c = 2(γ1 + γ2 )/(πγ1 γ2 ) (in notations of Section 7) and C is a normalization

Figure 15. Left: variation of the trace of the monodromy matrix

Received November 2006; revised June 2007.

You might also like