Celest Mech Dyn Astron manuscript No.

(will be inserted by the editor)

Phase correlations in chaotic dynamics

A Shannon entropy measure

P. M. Cincotta and C. M. Giordano

Received: date / Accepted: date

Abstract In the present work we investigate phase correlations by recourse to the

Shannon entropy. Using theoretical arguments we show that the entropy provides
an accurate measure of phase correlations in any dynamical system, in particular
when dealing with a chaotic diffusion process. We apply this approach to different
low dimensional maps in order to show that indeed the entropy is very sensitive to
the presence of correlations among the successive values of angular variables, even
when it is weak. Later on, we apply this approach to unveil strong correlations in
the time evolution of the phases involved in the Arnold’s Hamiltonian that lead to
anomalous diffusion, particularly when the perturbation parameters are compara-
tively large. The obtained results allow us to discuss the validity of several approx-
imations and assumptions usually introduced to derive a local diffusion coefficient
in multidimensional near–integrable Hamiltonian systems, in particular the so-called
reduced stochasticity approximation.
Keywords Shannon Entropy - Phase correlations - Diffusion

1 Introduction

Chaotic diffusion in multidimensional near–integrable systems is a quite relevant

dynamical process observed in different fields of astronomy, physics and chemistry.
Particularly in dynamical astronomy, diffusion plays a major role in planetary and
galactic dynamics, as recently reviewed and discussed in [26], [10], [25] and refer-
ences therein. The characterization of this instability is still an open subject, phase
correlations mainly due to stickiness, prevent in general the free diffusion that ex-
hibits for instance an ergodic system, where the variance of any phase space variable,
σ 2 , increases linearly with time. This type of behavior is known in the literature as
normal diffusion.

Grupo de Caos en Sistemas Hamiltonianos

Facultad de Ciencias Astonómicas y Geofísicas, Universidad Nacional de La Plata and
Instituto de Astrofísica de La Plata (CONICET-UNLP), Argentina
E-mail: [email protected], E-mail: [email protected]
2 P. M. Cincotta and C. M. Giordano

As shown in several works, it seems that in the limit of weak chaos1 , the diffusion
that proceeds along the chaotic layer of a single resonance (or along different layers
of the resonance web) over very large motion times might be approximated by a
normal process (see for example [22], [13], [16], [14], [21], [12] and [9]), leading then
to a slow drift of the unperturbed integrals of motion. On the other hand, in the
strongly chaotic scenario and relatively short (or physical) motion times, diffusion
could deviate significantly from the normal regime as it was presented in [10]. This
anomalous behavior of the diffusion imposes serious limitations on the derivation
of a diffusion coefficient. Indeed, within this approach to the stability problem, the
diffusion coefficient is introduced as the constant rate at which the variance evolves
with time.
Let us mention that anomalous diffusion has been observed also in many low
dimensional dynamical systems. In for instance [33],[34], [35], [36], [37] and [38] as
well as in [18], [19], [20], [30], [31], [27], a different approach to diffusion is considered
through a transport process formulation, where a generalization of the normal regime
is introduced, i.e. the variance scales as σ 2 = Dtb where the value of the exponent
b determines the regime of anomalous transport, namely super–diffusion if b > 1 or
sub–diffusion when b < 1, and the diffusion coefficient D is thus defined.
In [15] we present an alternative way to measure both, the extent and the time
rate of chaotic diffusion by means of the Shannon entropy (see [29]). A relevant
aspect of this approach is that the time evolution of the entropy is independent of
the transport process that takes place in phase space and therefore a measure of the
diffusion rate could be derived independently of the power law the variance satisfies.
In any case, as we have already mentioned, phase correlations are responsible for
anomalous diffusion. The analysis of phase correlations is a quite difficult problem
and thus restrictive assumptions are adopted in order to obtain analytical estimates
for local diffusion coefficients (see for instance [5], [8], [9]). Therefore, herein we face
with the problem of phase correlations in order to get some insight on the time
evolution of chaotic diffusion.
While the time–correlation function (see for instance [11], [28]) seems to be an
appropriate tool to quantify phase correlations, its computation is not an easy task
(as shown for example in [3]) and its numerical implementation is in fact quite
expensive. Thus, in the present effort, also by recourse to the Shannon entropy, we
provide a very efficient and simple way to clearly detect phase correlations.
In Section 2 we outline the underneath theory of the Shannon entropy in a similar
fashion as that given in [15], while in Section 3 profuse numerical experiments are
presented in several well known maps to finally investigate phase correlations in
the Arnold’s Hamiltonian [1], the very same system studied in [10] and [15]. In
this direction, we discuss the limitations of the assumptions behind the analytical
derivation of expressions for a local diffusion coefficient in Chirikov’s formulation of
Arnold Diffusion [5].

2 The Shannon entropy

As we have already discussed, it is not a simple task to provide a measure of corre-

lations among successive values of a dynamical variable, in particular at large times.
1 By weak chaos we mean the dynamical state when the unstable chaotic motion is mostly

confined to the narrow stochastic layers around resonances.

Phase correlations in chaotic dynamics 3

Thus, in the present section we propose the Shannon Entropy as an efficient tool to
measure correlations.
Theoretical background on the Shannon entropy can be found in [29] however,
different approaches to the entropy in dynamical systems are presented in, for in-
stance, [17], [2] and [32]. Applications to time series analysis can be found in [6] and
[7]. A recent approach concerning the use of the Shannon entropy to measure chaotic
diffusion and its time rate is addressed in [15].
Following for instance [2] and [23], consider the function Z defined in [0, 1] as:

−x lnx,
 x ∈ (0, 1]
Z(x) = (1)

0 x = 0.

It is evident that Z(x) ≥ 0 and Z 00 < 0. Consider now an open domain B ∈ Rn and
let
α = {ai ; i = 1, · · · , q} (2)
be a partition of B, let us say a collection of q n-dimensional cells that completely
cover B. The elements ai are assumed to be both measurable and disjoint. We
consider n = 1 and B = S1 or the unit interval (0, 1) with opposite sides identified.
Let xi = x(ti ) be the phase variable of a given map M (or Hamiltonian flow)
such that (yn , xn ) → (yn+1 , xn+1 ), with y ∈ G ⊂ R some action variable 2 .
For any finite trajectory γ = {(yi , xi ) ∈ G × B, i = 1, . . . , N } of M , let us define
γx = {xi ∈ B, i = 1, . . . , N } ⊂ γ, then a probability density on B can be defined as
N
1 X
ρ(u, γx ), = δ(u − xi ), (3)
N
i=1

where δ denotes the delta function. It is clear that

Z
ρ(u, γx ) du = 1, (4)
B

and the measure µ(ai ) results

Z
µ(ai (γx )) = ρ(u, γx ) du. (5)
ai

Then, for the partition α the entropy for γx is defined as

q
X q
X
S(γx , α) = Z ( µ(ai (γx )) ) = − µ(ai (γx )) ln(µ(ai (γx ))). (6)
i=1 i=1

Notice that for a given partition and trajectory, the entropy is bounded. Indeed,
for the partition (2) it is 0 ≤ S(γx , α) ≤ ln q, for any γx . The minimum corresponds
to the situation in which the xi are restricted to a single element of α, say the j–th
element, which would correspond to the extreme case of full correlation of the phase
values at all times, such as for example at a first order resonance or one-period fixed
point of M (xn+1 = xn , ∀n). In such a case it is µ(aj ) = 1, µ(ai ) = 0, ∀i 6= j, leading
2 In fact, the dynamical system could involve more than one action and phase.
4 P. M. Cincotta and C. M. Giordano

to S = 0. On the other hand, the maximum value, S = ln q, would be obtained

when the q elements of the partition had an equal measure, that is µ(aj ) = 1/q,
which corresponds to the situation in which the x values are dense and uniformly
distributed in B. Therefore, in the case of a non-integrable system, the entropy seems
to be a natural measure of correlations.
Let nk be the number Pof phase values in the cell ak , thenPµ(ak ) = nk /N . From the
q q
normalization condition, k=1 µ(ak ) = 1, it follows that k=1 nk = N. Therefore,
the entropy given in (6) reduces to
q
1 X
S(γx , α) = ln N − nk ln nk . (7)
N
k=1

Take for instance γxr = {xi = θir ∈ B, i = 1, . . . , N } where the θir are random num-
bers and N  1, then the nk obey a Poisson distribution:
λn −λ N
Pλ (n) = e , λ= ,
n! q
where λ is the mean value of the distribution as well as its variance. Let us assume
λ  1 and take nk = λ + ξk with |ξk |  λ (nk ∈ R), so
q
X
ξk = 0, (8)
k=1

which clearly follows from the normalization condition. Then it is straightforward to

show that, up to second order in ξk /λ, it is
q q
X 1 X 2
nk ln nk = N ln N − N ln q + ξk , (9)
2λ
k=1 k=1

and (7) reads

q
q X 2
S(γxr , α) = ln q − ξk . (10)
2N 2
k=1

Now on introducing the information I as

1
I(γx , α) ≡ 1 − S(γx , α), (11)
ln q

and estimating ξk2 = λ we get, in this particular limit of absence of correlations, that
q
I(γxr , α) ≈ > 0, (12)
2N ln q
which results independent of γxr . Since N denotes the number of iterates of the map,
we observe that, for completely random motion, I would decrease with time as t−1 .
On the other hand, from (6) the time derivative of the entropy results
q
dS X
(γx , α) = − µ̇(ak )(1 + ln µ(ak )), (13)
dt
k=1
 Phase correlations in chaotic dynamics 5

and, for random motion, its theoretical estimate is

dS r q
(γ , α) ≈ Ṅ > 0, (14)
dt x 2N 2
which is seen to decrease as t−2 .
The above results clearly state that, for a random system, the information de-
creases as time increases, or alternatively, the entropy would reach asymptotically
its maximum value; the system approaches the full mixing as t−1 .
Now let us consider γxs = {xi = θis ∈ B, i = 1, . . . , N } with N  1 and where θis are
correlated phases, for instance θs = ωt with ω ∈ R \ Q, which represents the motion
on a torus. Being ω irrational, the motion is ergodic on S1 , thus the distribution
function of n ∈ R approaches f (nk ) ≈ δ(nk − λ), with λ = N/q. Therefore, if we
consider again nk = λ + ξk with |ξk |  λ, up to second order in ξk /λ, (10) holds.
The fluctuations should satisfy |ξk | < 1, since every nk would be the closest integer
number to λ which, in general, is a real quantity. Therefrom, setting |ξk | ≈ 1/2, the
information (12) for a single quasiperiodic trajectory on the torus reduces to

q2
I(γxs , α) ≈ , (15)
8N 2 ln q

and, in this particular case, the information would decrease as t−2 . Notice that if we
consider instead an ensemble of np nearby trajectories, the estimate |ξk | ≈ np /2 has
to be adopted (see below) and since N = np × t, (15) becomes

q2 1
I(γxs , α) ≈ . (16)
8 ln q t2
In order to check the analytical estimates given in (12) and (14), we pick up
np = 100 initial values of the phases xi ∈ (0, 1) and evolve each initial condition in a
random way taking x mod 1 to finally compute the information given by (11) and
the derivative of the entropy every ∆t = 103 . The information I for the ensemble
corresponding to a partition α defined by q = 103 cells on the unit interval, is com-
puted for N = np × t, where t = 103 , . . . , 106 . The time derivative of the entropy can

10-3 10-5
numerical numerical
theoretical theoretical
10-6

10-7
10-4
10-8
ln (dS/dt)

10-9
ln I

10-5 10-10

10-11

10-12
10-6
10-13

100000 300000 500000 700000 900000 100000 300000 500000 700000 900000
t t

Fig. 1 Left panel: evolution of I for np = 100 initial values of xi ∈ (0, 1) evolving randomly
(in red) and the expected theoretical trend given by (12) with N = np × t (in green). Right
panel: the time derivative of the entropy computed numerically (in red) and the theoretical
evolution given by (14) (in green), see text for details.
6 P. M. Cincotta and C. M. Giordano

be evaluated, either from equation (13) or by the numerical derivative corresponding

to the time evolution of (6), also every ∆t = 103 . In fact, both results completely
agree, though the latter procedure results somewhat less noisy than the former, so
from now on we will restrict ourselves to show solely the numerical derivative of
the entropy. Fig. 1 displays the results concerning the numerical simulation and the
theoretical estimates for both the information and the time derivative of S, revealing
the consistency of the proposed analytical approach.

Now, let us consider a similar example as the one above but for the 1D-map
xn+1 = xn + ω with x mod 1 and ω ∈ R. We have carried out several experiments,
taking different values of np ranging from 1 to 2000, ensembles of size 10−5 to 10−9
and varying both ∆t and q. The outcome values of I given by (11) show to be only
sensitive to the partition as (16) foresees, and to the ω value which could be either
rational or irrational. For the particular case ω = l ∈ Z we obtain I ≈ 1 (S ≈ 0)
despite of the size of the partition q > 1.

In Fig. 2 we present the representative behavior of I corresponding √ to ergodic

motion on a torus for different np values in ensembles with |ω − 3| ≤ 10−8 , x0 = 0
and N = np ×t, t = 5×103 , . . . , 5×106 , for a partition of the unit interval of q = 2×103
and ∆t = 5×103 . We also include in the figure the theoretical estimates given by (16).
Again, we notice a good agreement between the numerical and theoretical estimates
for the information in the case of ergodic motion.

The above examples, though rather simple, allow us to assert that in a non-
linear system the information, that measures the phase correlations, should decrease
at most as t−1 , the limit taking place when the system becomes completely mixed.
Strong phase correlations would provide larger values of the information and I should
scale with time as I ∼ tβ , β > −1. On the other hand, if β → −2, the phase evolves as
in an integrable system, the successive values of x though fully correlated are ergodic
on the torus and so that the approximation hsin2 xi = hcos2 xi ≈ t/2 holds, though
this result does not imply that the phase values are completely uncorrelated.

10-3
np = 50
np = 100
np = 1000
10-4 theoretical ergodic

10-5
ln I

10-6

10-7 Fig. 2 Evolution of I for√three

different np values, |ω − 3| ≤
10−8 and q = 2000 given by the
10-8
map xn+1 = xn + ω with x0 =
0, and the expected theoreti-
1x106 2x106 3x106 4x106 5x106 cal trend for ergodicity given by
t (16) which is independent of np .
Phase correlations in chaotic dynamics 7

3 Numerical experiments

In this section we describe the numerical experiments regarding three different maps
and provide evidence of how the information succeeds in yielding a measure of phase
correlation in each case.

3.1 The standard map

This well known discrete area preserving dynamical system, introduced by Chirikov
in [4], [5] and largely studied in the last four decades (see for instance [27] and
references therein), (I, θ) → (I 0 , θ0 ), is defined by
I 0 = I + K sin θ, θ0 = θ + I 0 , (17)
with I ∈ R, θ mod 2π, and K a free parameter. We numerically investigate this system
which describes the dynamics of a pendulum perturbed by an infinite set of periodic
kicks of similar amplitude and frequency Ω = 1. In fact we deal with the reduced
standard map defined in S1 × S1 as a consequence of introducing the new variables
(p, x) that comply I = 2πp, θ = 2πx, both p and x being mod 1, so that
K
p0 = p + k sin 2πx, x 0 = x + p0 , k= . (18)
2π
It is well known that for K < 1 (k < 1/2π) the motion is mostly stable, except near
the thin√chaotic layers or homoclinic tangles around the integer resonance p = 0
of size k, and the main fractional resonances p = 1/2 and p = 1/3, 2/3, whose
half-widths are k and k 3/2 respectively. For K = Kc ≈ 1 the overlap of resonances
becomes relevant and this critical value separates stable from unstable motion in a
broad sense. For K > 4 the centers of the integer resonances become unstable and
therefrom the motion is usually assumed to be almost completely uncorrelated.
We have performed several numerical experiments with values of K in the interval
0 ≤ K ≤ 100 though herein we only show a few examples corresponding to K =
0.1, 1, 5, 10, for an ensemble of np = 100 random initial conditions with p, x ≈ 10−5 . We
have iterated the map (18) up to N = 106 , introducing a partition defined by q = 103 ,
and we have computed the information (11) and the derivative of the entropy (13)
adopting a time interval of ∆ = 103 . The results are displayed in Fig. 3, where we have
also included the theoretical estimates (12) and (14) for I and dS/dt respectively,
which correspond to the case of random motion, for a partition defined by q and
N = np × t. The left panel shows the evolution of the information corresponding to
different K values. Notice that the phase values appear to be correlated for the lower
values of K; in fact, even for K values as large as 1 or 5 the successive values of the
phases seem to be correlated for large times. A similar result is obtained for K = 9
(not shown in the figure), but for K ≥ 10 the dynamics in x looks like random,
the information decaying as t−1 . The right panel of Fig. 3 corresponds to the time
derivative of the entropy. Besides the expected noisy behavior, it turns out evident
that dS/dt decreases as t−2 only for the experiment corresponding to K = 10. These
results suggest that the usual assumption for the standard map,
t
X K
h(I(t) − I(0))2 i = K hsin2 θ(l)i ≈ t,
2
l=1
 8 P. M. Cincotta and C. M. Giordano

10-4
random
K = 0.1
10-1 K=1
10 -6 K=5
K = 10
10-2
10-8

ln (dS/dt)
10-3
ln I

10-10
10-4
random
K = 0.1
10-5 K=1 10-12
K=5
K = 10
10-6
10-14
100000 300000 500000 700000 900000 100000 300000 500000 700000 900000
t t

Fig. 3 (Left panel) Evolution of I for np = 100 initial values of xi ∈ (0, 1) of the reduced
standard map (18) for different values of the parameter K and the analytical expected behavior
for random motion given by (12) displayed in red for N = np × t. (Right panel) Similar to the
plot at the left but for the time derivative of the entropy computed numerically and the
theoretical expected evolution given by (14).

is a plausible approximation only for K & 10, which is in agreement with the results
given in [27].

3.2 The whisker mapping

Let us consider another canonical map, the whisker or separatrix mapping (w, τ ) →
(w0 , τ 0 ), w ∈ R, τ mod 2π, also introduced by Chirikov in [4],[5] as
 
32
w0 = w + W sin τ, τ 0 = τ + λ ln , (19)
|w0 |

where W, λ are free parameters. This map represents the motion close to the sep-
aratrix of a non-linear resonance or pendulum subjected to a symmetric periodic
perturbation; thus |W | ∼ A2 (λ)  1 for λ  1, where A2 (λ) is the Melnikov-Arnold
integral for m = 2. This simple system models the chaotic layer around resonances.
After rescaling the a-dimensional energy w by λW and defining y = w/(λW ), the
mapping (19) reduces to

1
y0 = y + sin τ, τ 0 = τ − λ ln |y 0 | + G, (20)
λ
where G = λ ln(32/λ|W |) is a constant. It is well known that, after the above rescal-
ing, the chaotic layer has a finite width, |y| . 1, whose central part, |y| . 1/4,
is highly chaotic, with no stability domains, while its external part reveals a di-
vided phase space with large stability islands. Therefore, the strong correlation be-
tween the successive values of the phase τ for |y| ∼ 1 should be responsible for
the finite width of the layer. Roughly, near the edges of the layer it is |y| ≈ 1
and thus from the second relation in (20) it is clear that the structure of the ex-
ternal part of the layer is dominated by G. Anyway, the study of its resonances,
yrk = (32/λW ) exp(−2kπ/λ) = exp((G − 2kπ)/λ)), k ∈ Z, by means of local stan-
dard maps clearly shows that the motion for |y| ≈ 1 or larger should be mostly
 Phase correlations in chaotic dynamics 9

10-7
random
λ = 10
10-4 10-8
λ = 102
λ = 103
10-9 λ = 104
10-5
10-10

ln (dS/dt)
ln I

10-11
10-6

10-12
random
10-7 λ = 10
λ = 102 10-13
λ = 103
λ = 104 10-14
10-8
1x106 3x106 5x106 7x106 9x106 1x106 3x106 5x106 7x106 9x106
t t

Fig. 4 (Left panel) Evolution of I for np = 100 initial values of xi ∈ (0, 1) of the whisker
mapping (20) for W = 10−6 and four different values of the parameter λ. The analytical
expected behavior for random motion given by (12) is also included and displayed in red for
N = np × t. (Right panel) Similar to the plot at the left but for the time derivative of the
entropy computed numerically and the theoretical expected evolution given by (14).

regular. Therefore it turns out fairly interesting to study phase correlations in this
particular map by means of the Shannon entropy.
Then, we have thoroughly investigated the map (20) by numerical means taking
different values of the parameters W and λ. Herein we include the results concerning
a single value of W , namely W = 10−6 , since the map is almost independent of this
scaling parameter, and four different values of λ = 10, 102 , 103 , 104 . As before, we
have considered an ensemble of np = 100 random initial conditions with (y, τ ) ≈ 10−5 .
We have iterated the map up to N = 107 , after introducing the normalized phase
x = τ /2π mod 1, having taken a similar partition of the unit interval, q = 103 , and
computed the information (11) and the derivative of the entropy (13) with a sample
time interval ∆ = 104 . The results are displayed in Fig. 4, where again we have
included the theoretical estimates (12) and (14) corresponding to random motion
for I and dS/dt respectively. In any case, the information reveals that at short
times, while the trajectories are confined to the central part of the layer, the values
of the phase τ seem to be uncorrelated, but for large times instead, when the system
has already reached the border of the layer, the correlations are quite evident, the
information I increasing significantly. Notice that the time required to become a
correlated system is seen to increase with λ, as expected. The derivative of the
entropy reveals the very same dynamical behavior.

3.3 A generalized whisker mapping

We devote this section to the numerical study of the map that describes, accordingly
to Chirikov, the diffusion across and along the chaotic layer of the resonance ω1 = 0
for the Arnold Hamiltonian (see [1], [5]), which can be recast as

1
H(I1 , I2 , θ1 , θ2 , t; ε, µ) = (I12 + I22 ) + ε(cos θ1 − 1)[1 + µ(sin θ2 + cos t)], (21)
2
with I1 , I2 ∈ R, θ1 , θ2 , t ∈ S1 , and µε  ε  1. Let us note that (21) can be re-arranged
as
10 P. M. Cincotta and C. M. Giordano

H(I1 , I2 , θ1 , θ2 , t; ε, µ) = H1 (I1 , θ1 ; ε) + H2 (I2 ) + µV (θ1 , θ2 , t; ε),

where
1 1
H1 = I12 + ε(cos θ1 − 1), H2 = I22 , V = ε(cos θ1 − 1)(sin θ2 + cos t),
2 2
and that the instantaneous change of H and H2 are given by
∂V ∂V
Ḣ = µ , Ḣ2 = I2 I˙2 = −µω2 ,
∂t ∂θ2
where ω2 = I2 6= 0 is assumed to be constant and irrational, then
µε
Ḣ = (2 sin t + sin(θ1 − t) − sin(θ1 + t)) ,
2
(22)
µε
Ḣ2 = (2ω2 cos θ2 − cos(θ2 − θ1 ) − cos(θ2 + θ1 )) .
2
√
Thus, taking θ2 (t) ≈ ω2 t + θ20 , θ1 (t) = 4 arctan (exp ( ε(t − t0 ))), i.e. the motion on
the separatrix, defined in such a way that θ1 = π for t = t0 , and neglecting the free
oscillatory terms sin t and cos θ2 3 , the variations of H and H2 over a half–period of
oscillation of θ1 result (see [5] for more details)
√   √  
µ 1 µω2 ω2
∆H = A2 √ sin t0 , ∆H2 = A2 √ cos θ20 ,
2  2 

where t0 is the value of the phase t when the pendulum crosses the surface θ1 = π
where the stable equilibrium point lies, and θ20 = θ2 (t0 ).
In Chirikov’s formulation, significant changes in H are only possible due to large
variations in H2 , since H1 is bounded to the exponentially small (wrt µ) width of
the chaotic layer. Moreover, defining DH = h(H(t) − H(0))2 i/t and D2 = h(H2 (t) −
H2 (0))2 i/t, for large times it should be DH = D2 and thus

hsin2 t0 i = v 2 hcos2 θ20 i, (23)

with  
ω2 A2 ω2
√ √ −1)π
ε sinh (π/ ε) (ω22√ (1−ω2 )π
√
v=   = ω22 √ e ε ≈ ω22 e 2 ε , (24)
A2 √1 sinh (ω2 π/ ε)
ε

where the approximation holds for ε  1. The latter expression

p for the relative√am-
plitude v is significant only in the interval |ω2 − ω2M | < 2/π where ω2M = 4 √ε/π
corresponds to its maximum value, being the latter vM ≈ (4/πe)2 ε exp √ (π/(2 ε));
and goes to zero in both limits as v ∼ ω22 , ω2 → 0 and as v ∼ exp (−ω2 / ε) , ω2 → ∞.
Then, for
√ 
ω2 > 1 it is
p  v  1. On the other hand, for ω2 in the narrow range
ω2 ≈  4/π ± 2/π , it is v  1. Therefore from (23) for ω2 away from 0 and
1, it turns out that both phases could not be simultaneously random.
The associated map represents the finite variation of the energy in H1 (pendulum
model for the resonance ω1 = 0) when the system is close to the separatrix of the
3 See the discussion below.
Phase correlations in chaotic dynamics 11

resonance ω1 = 0 and it can be computed as the difference ∆H1 = ∆H − ∆H2 after

a half period of oscillation or a period of rotation of the pendulum and it has the
form (y, τ ) → (y 0 , τ 0 ), y ∈ R, τ mod 2π, where y measures the energy changes in H1
relative to the separatrix energy (H1 = 0), τ represents the successive values of t0
and it reads
1
y 0 = y + sin τ − v cos(ω2 (τ + β)), τ 0 = τ − √ ln |y 0 | + η, (25)
ε

where η is a constant similar to G introduced in the whisker mapping, that depends

on both µ and ε but not on ω2 , ω2 β being the initial value of θ20 .
Chirikov proposed that the map (25) not only describes the thin chaotic layer
around the main or guiding resonance √ ω1 = 0 but the diffusion along the latter as
well. For ε  1 and ω2 > 1 (0 < ω2 ∼ 4 ε/π) since v  1 (v  1), in the right hand side
of the first equation of (25) a clearly dominant term is present. Thus at this order,
the map reduces to the whisker mapping, being the largest term (layer resonances
in Chirikov’s terminology) responsible for the properties of the chaotic layer, such as
its width and resonances’ structure. Therefore, for large times, the successive values
of the phase involved in such √ a dominant term would be strongly correlated (τ for
ω2  1 and θ2 for ω2 ∼ 4 ε/π). Note that for v  1 (ω2  1) the map becomes
independent of ω2 and thus, while τ values should present correlations, phase values
θ20 = ω2 τ could evolve, in principle, nearly in a random or ergodic fashion when ω2
is irrational and large.
Since when the system is close to the separatrix of the resonance ω1 = 0, the
variation in H1 is given by the largest term in the map (y), the successive values
of the smaller term in the first of (25) are assumed to be nearly random, and this
perturbation term (or driving resonances accordingly Chirikov) is responsible for a
nearly normal diffusion process along the resonance, i.e. large variations in√I2 such
that h(H(t) − H(0))2 i ∝ t as in the standard map. In other words, for ω2 ∼ 4 ε/π < 1
it is assumed that hsin2 τ i = R/2, while for ω2 > 1, hcos2 θ20 i = R/2, where R < 1 is
the so–called reduction factor that takes into account some correlations among the
successive phase values when the system is moving within the external part of the
chaotic layer (see [5], [8], [9]).
Let us remark that the derivation of the map (25) rests, among other simplifica-
tions, on the following assumptions: (i) ε  1 such that no overlapping with other
first order resonance takes place (as for instance among the resonances ω1 = ±1
and ω1 = 0); (ii) εµ  1 such that we could consider ω2 nearly constant and (iii)
√
|I2 |  2 εµ, the width of the resonance ω2 = 0, such that cos θ2 can be regarded as
a purely oscillatory term.
Therefore, the expected behavior of both phases depending on ω2 encourages the
investigation of correlations in this map by means of the Shannon entropy.
We present the results of only one of the large set of experiments
√ performed for
ε = 0.05 and µ = 10−3 ε and eight different values of ω20 = κ 3, κ ∈ K where

K = {κ : κ = 0.06, 0.16, 0.25, 0.62, 1.75, 2.25, 11.0, 19.0},

such that 0.10 . ω20 . 33 and for κ = 0.62, ω2 ≈ 1. For these values of ε and κ ∈ K,
the relative amplitude v (24) ranges from 2 × 10−95 for the largest κ value up to
12 for κ = 0.16, which it is in agreement with the analytical estimates given above.
The resonances’ width is about 0.447 for the resonance ω1 = 0 while it equals 0.003
12 P. M. Cincotta and C. M. Giordano

for the remainder ones, so that no overlap would be expected between the main
resonance and ω1 = ±1. For the smallest value of κ ∈ K, ω20 ≈ 0.104 and it would be
plausible to assume that it is far from the resonance ω2 = 0 since its width is quite
small. Anyway, we numerically integrated the Hamiltonian (21) for the same set of
parameters and np = 100 random initial 0 0
0 √ values of ω2 and I1 on the chaotic layer of
the resonance ω1 = 0 such that I1 = 2 ε, and we measured the variation of ω2 ≈ I2
to find that the largest value of |ω2 − ω20 | ≤ 0.01, which corresponds to κ = 0.06 after
a total motion time t = 106 . It outcomes then that the approximations assumed to
derive the map (25) are valid for the adopted values of the parameters (see below).
In Fig. 5 we present a three dimensional visualization of the diffusion for the smaller
four values of κ ∈ K for the section |θ2 | ≤ 10−5 where we observe that at least up to
t = 106 the diffusion is almost irrelevant.

I1

0.4
0.2
0
-0.2
-0.4
1 6
0.8 5 Fig. 5 Visualization of the dif-
4 fusion along the chaotic layer of
0.6 3 the resonance ω1 = 0 for four
0.4 2
I2 1 θ1
values of κ ∈ K, 0.06 in red, 0.16
0.2
0 in green, 0.25 in blue and 0.62 in
magenta.

Again, we take for the map an ensemble of np = 1000 random initial conditions
with y ≤ 10−5 , |ω2 − ω20 | ≤ 10−5 , τ = 10−5 , β = 0 and iterate the map (25) up to
N = 106 . We normalize both phases τ and θ2 to the unit interval and compute the
information (11) and the time derivative (13) after time intervals of length ∆ = 103
with q = 103 . The results for the information are displayed in Fig. 6, where I as a
function of time is presented for both phases, τ, θ2 , mod(1), for the eight different
values of ω20 . The left panel in Fig 6 shows the results of the information concerning
the phase τ while the right panel does so for the phase θ2 . We observe that Chirikov’s
assumption might apply for ω2 < 1, where the phase τ behaves almost randomly
though some correlations are clearly present while θ2 shows strongly correlated. On
the other hand for ω2 > 1, while it is evident that the successive values of τ are not
sensitive to ω2 and highly correlated, only for large values of ω2 the phase θ2 seems to
be uncorrelated, when the amplitude v is negligible. The results obtained from several
experiments with different values of ε, µ, κ and N are quite similar when considering
µε    1 and the mentioned restrictions to the location of the ensemble. In sum,
it seems that the conjecture that τ or θ2 are partially random, in the sense that
either hsin2 t0 i or hcos2 θ20 i (depending on ω2 ’s value) grows in mean linearly
√ with
time, could be a fairly good approximation provided that either ω2 ∼ 4 ε/π < 1 or
ω2  1. Nevertheless, some departures from the linear trend should be expected due
to the presence of correlations.
 Phase correlations in chaotic dynamics 13

10-2
random κ = 0.25 κ = 2.25 random κ = 0.25 κ = 2.25
τ κ = 0.06 κ = 0.62 κ = 11.0 θ2 κ = 0.06 κ = 0.62 κ = 11.0
κ = 0.16 κ = 1.75 κ = 19.0 κ = 0.16 κ = 1.75 κ = 19.0
-3
10
10-1

10-4 10-2

10-3
ln I

ln I
10-5
10-4

10-5
10-6
10-6

10-7 10-7
100000 300000 500000 700000 900000 100000 300000 500000 700000 900000
t t

Fig. 6 (Left panel) Evolution of I for np = 1000 √ initial values of τ ∈ (0, 1) for the map (25)
for ε = 0.05, µ = 10−3 ε and eight values of ω20 = κ 3. The analytically expected behavior for
random motion given by (12) is also included for N = np × t. (Right panel) Similar to the plot
at the left but for the phase θ2 ∈ (0, 1).

Let us now consider the ensemble variance,

np 2
1 X  (k) (k)
h∆H2 (tj )2 i = H2 (tj ) − H2 (0) , (26)
np
k=1

where tj = t0 + jδt, j ∈ Z+ , δt  1 (see below) being the integration time step of

the Hamiltonian (21). We study the temporal evolution of h∆H2 (tj )2 i over the time
span t = 106 . Fig. 7 presents the evolution of H2 ’s variance for the four smaller values
of κ ∈ K where we observe that, within the regarded time interval, the variances
increase in mean linearly as expected. Moreover, though the diffusion domains do
not overlap, the general properties of the chaotic dynamics provided by h∆H22 i seems
to be quite similar since the variances take nearly the same values. However some
departures from the linear trend are noticeable, in particular for κ = 0.62, which
are due precisely to phase correlations. The variances for the remainder values of
κ, which are not regarded for the figure, require a much longer time span for any
increasing behavior to show up.

1.4x10-9 κ = 0.06 1.8x10-10 κ = 0.62

κ = 0.16
-9 κ = 0.25 1.6x10-10
1.2x10
1.4x10-10
1x10-9
1.2x10-10
< ∆H22 >

< ∆H22 >

8x10-10 1x10-10

6x10 -10 8x10-11

6x10-11
4x10-10
4x10-11
2x10-10
2x10-11

0 0
100000 300000 500000 700000 900000 100000 300000 500000 700000 900000
t t

Fig. 7 Evolution of h∆H22 i given in (26) for κ = 0.06, 0.16, 0.25, 0.62 in different panels due to
their different scale in the variance. The adopted time-step in this computation is δt = 5 × 10−4
14 P. M. Cincotta and C. M. Giordano

Let us now consider larger values of the parameters, namely, ε = 0.25 and µ = 0.1ε,
the ones adopted in [10]. In this case the approximation for the relative amplitude
v given in (24) presents a maximum v ≈ 1.27 at κ ≈ 0.36. Thus we consider the
following values of κ ∈ K 0 :

K 0 = {κ : κ = 0.25, 0.57, 0.75, 1.00, 1.50, 2.00, 9.00},

such that v ranges from 3 × 10−18 for the larger value of κ up to 1.22 for the smallest
one. For the remainder values of κ = 0.57, 0.75, 1.00, 1.50, 2 the relative amplitude
amounts 1.02, 0.66, 0.31, 4.5 × 10−2 and 5 × 10−3 respectively. We observe that for all
values of κ, v ∼ 1 or smaller and thus τ should exhibit strong correlations always
while it is expected that for the largest values of κ, θ2 behaves nearly as random,
accordingly to the discussion given from the map (25). However, as it is shown in [10],
since both ε and µ are comparatively large, resonance interaction is strong leading to
large variations of I2 , −2 . I2 . 2 for all values of κ (except perhaps for the largest
one, see below) and thus the approximations introduced in the derivation of the map
seem to be no longer valid, i.e., ω2 could not be regarded as constant and moreover,
the motion may be trapped in the resonance ω2 = 0. Additionally, the resonances
ω1 = 0 and ω1 = ±1 are close to be in overlap and the amplitude of the resonances
ω1 ± ω2 = 0 are not negligible, thus around ω2 = 1 the motion could not be confined
to the chaotic layer of the main resonance, as it was discussed in [10].

I1

1
0.5 6 Fig. 8 Visualization of the dif-
0 5 fusion along the chaotic layer of
-0.5 4 the resonance ω1 = 0 for ε =
3
-1 0.25 and µ = 0.1ε and six values
2
3
θ1
of κ ∈ K 0 , 0.25 in red, 0.57 in
2 1
1 0 green, 0.75 in blue, 1.00 in ma-
-1 0
genta, 1.50 in cyan and 2.00 in
I2
yellow.

Fig. 8 presents the visualization of the diffusion for the adopted values of ε and
µ and several initial locations of the ensembles given by κ ∈ K 0 . We observe that for
0.25 ≤ κ ≤ 1 the diffusion domains overlap and the action has a significant variation;
−2 . I2 . 2, while for κ > 1, the diffusion in I2 is much more confined, similar to
what is observed in Fig. 5. Notice that near I2 = 1 diffusion spreads out of the chaotic
layer of the resonance ω1 = 0 where the motion is under the influence of at least three
resonances, ω1 = 0, ω1 ± ω2 = 0 and ω1 = ±1. Thus, the numerical results confirm
that for ω20 . 2 the map (25) does no longer describe the diffusion along the chaotic
layer of the main resonance.
Therefore for these large values of the perturbation parameters we numerically
investigate the Hamiltonian (21) instead of the map (25). To this aim, we take an
ensemble of np = 100 random initial conditions on the separatrix of the resonance
 Phase correlations in chaotic dynamics 15

10-2 10-2

10-3 10-3
ln I

ln I
10-4 10-4

10-5 τ 10-5 θ2
random κ = 0.57 κ = 1.50 random κ = 0.57 κ = 1.50
ergodic κ = 0.75 κ = 2.00 ergodic κ = 0.75 κ = 2.00
-6
κ = 0.25 κ = 1.00 κ = 9.00 -6
κ = 0.25 κ = 1.00 κ = 9.00
10 10
1000 10000 100000 1000 10000 100000
ln t ln t

Fig. 9 (Left panel) Evolution of I for the phase τ ∈ (0, 1) corresponding to the values of t
when the the pendulum crosses the surface θ1 = π for np = 100 random initial values on the
separatrix of the√resonance ω1 = 0 for the Hamiltonian (21) with ε = 0.25, µ = 0.1ε and seven
values of ω20 = κ 3. The analytical expected behavior for random and ergodic motion given
by (27) are also included for N = βnp × t/δt, with β ≈ 3 × 10−4 . (Right panel) Similar to the
plot at the left but for the phase θ2 ∈ (0, 1).

ω1 = 0. We integrate the equations of motion up to t = 5 × 105 with a small time

step δt = 5 × 10−4 and take the values of θ2 and t ≡ τ each time |θ1 − π| ≤ 0.001,
avoiding two consecutive crossings of the surface θ1 = π by a single trajectory4 .
After normalizing θ2 and τ to the unit interval, on adopting a sample time interval
∆t = 500 √and setting q = 103 , we compute the information (11) for the seven values
of ω2 = κ 3, with κ ∈ K 0 .
0

For the theoretical estimates, the total number of points is now N = np ×t/δt, but
since we are considering only those satisfying |θ1 − π| ≤ 0.001, we set N = βnp × t/δt,
with β  1, and therefore (12) and (16) become

qδt 1 q 2 δt2 1
I(γxr , α) ≈ , I(γxs , α) ≈ , (27)
2βnp ln q t 8β 2 ln q t2

respectively, where β ≈ 3 × 10−4 obtained numerically. √

Fig. 9 shows the results for I for all the ensembles defined by ω20 = 3κ, with
κ ∈ K 0 . It becomes evident that in any case, the phase τ is always correlated, though
up to t ≈ 104 I decreases with a power law near to t−1 , at larger times decreases in a
much slower way and for large κ it reaches a nearly constant value at relatively short
times. On the other hand, while θ2 decreases with time for all ensembles, it could be
assumed nearly random only for large values of ω2 , precisely when ω2 ≈ const is a
rough approximation. We observe that up to t ≈ 2×104 it is I ∼ tp with −2 < p < −1,
but for larger times only those ensembles located at large ω20 , I ∼ t−1 . Note that for
κ = 9.00, I behaves initially in an ergodic rather than random fashion, as expected
since θ2 ≈ ω2 t.
In [10] as well as in [15] it was shown that for ε = 0.25 and µ = 0.1ε, the evo-
lution of the ensemble variance (26) for |ω20 | < 2 reveals that the diffusion is clearly
anomalous, in particular a sub-diffusive process. This fact motivated our introducing
an alternative way of measuring both the extent and the rate of the diffusion in [15],
namely by means of the Shannon entropy.

4 This could occur due to the smallness of the time step.

16 P. M. Cincotta and C. M. Giordano

Regarding the diffusion observed in I2 , we can say that for ω2 . 1, the diffusion
is certainly driven by all first order resonances and it has no sense to speak about
layer or driving resonances, all of them contributing to the motion along and across
the chaotic layer of the guiding resonance. On the other hand, for ω2  1 Chirikov’s
description could apply, resonance interaction is weak so the assumptions beneath the
map (25) become plausible, the amplitude being v  1 and regarding the evolution
of θ2 for large ω2 we could take hcos2 θ2 i ≈ Rt/2 with R . 1.
Let us mention that in case of a general multidimensional Hamiltonian, Chirikov
derived a similar map than (25) but involving several phases and the discussion
about the behavior of the different phases is quite similar to the one given above
(see [5], [8], [9]).

4 Discussion

Herein we have shown that the Shannon entropy turns out to be a rather simple and
effective way to measure phase correlations. Even though we apply this technique
to area–preserving maps or near–integrable Hamiltonian systems, its formulation
–given in Section 2– reveals that it can be used in a wide range of problems be-
yond dynamical systems. The numerical experiments included therein unfold that
the derived analytical estimates for absence of correlations or ergodicity are quite
accurate.
The numerical simulations presented in this work concern in general chaotic
motion, thus a given phase variable cover completely the unit interval. Therefore, the
key point in the present approach is the probability of occupation of each element of
the partition, µ(ak ) = nk /N . If the nk are randomly distributed, thus the information
I decreases as the inverse of the integer time or total number of phase values. On
the other hand, if the nk have a uniform distribution, then the information decreases
faster, following an inverse square law. Any other dependence of I with N reveals
the existence of correlations among the phase values. Therefore a reference level for
the information could be adopted for a given distribution of N values of the phases,
in order to discriminate between strong and weak correlations in the sample.
In dynamical systems like those studied in Section 3, it would be much more
interesting to study how the information evolves with time. In such a way it is
possible to reveal and understand the underlying dynamics, as for instance in Figs. 3
and 4.
More interesting is the use of this novel tool to revisit Chirikov’s approach to
diffusion along the chaotic layer of a single resonance or Arnold diffusion, in a broad
sense of this term. We succeed in revealing strong and weak correlations that prevent
the free diffusion and therefore the so–called reduced stochasticity approximation,
in the sense that hcos2 θ2 i ≈ Rt/2 or hsin2 τ i ≈ Rt/2 with R < 1 is not applicable
and thus the obtained estimates for the diffusion coefficient are not correct (see [5]–
Section 7, [10] for more details). Particularly, the map (25) could be a reliable model
to investigate diffusion along the chaotic layer of resonance ω1 = 0 provided that
µε  ε  1 and ω2 is far away from 0 and 1 as Fig. 6 shows. On the other hand,
if the parameters are not too small, the map would provide a fair approximation
to the diffusion that actually takes place only if ω2  1. In other words, Chirikov’s
formulation seems to be adequate when the diffusion extent is quite restricted and
thus it makes sense to derive a local diffusion coefficient.
Phase correlations in chaotic dynamics 17

Acknowledgements This work was supported by grants from Consejo Nacional de Investiga-
ciones Científicas y Técnicas de la República Argentina (CONICET), the Universidad Nacional
de La Plata and Instituto de Astrofísica de La Plata. We acknowledge two anonymous reviewers
for their valuable comments and suggestions that allow us to improve this manuscript.

Conflict of interest

The authors declare that they have no conflict of interest.

References

1. Arnold, V. I., On the nonstability of dynamical systems with many degrees of freedom,
Soviet Mathematical Doklady 5, 581-585 (1964)
2. Arnol’d, V., Avez, A. Ergodic Problems of Classical Mechanics, 2nd. Ed. (Addison-Wesley,
New York), (1989)
3. Berne, B.J.,Harp, G.D. On the Calculation of Time Correlation Functions in I. Prigogine,
Stuart A. Rice eds., Advance in Chemical Physics, Vol XVII (1970)
4. Chirikov, B. V., Institute of Nuclear Physics, Novosibirsk (in Russian), Preprint 267, (1969),
Engl. Transl., CERN Trans. 71 - 40, Geneva, October (1971)
5. Chirikov, B. V., A Universal instability of many-dimensional oscillator systems, Physics
Reports, 52, 263 (1979)
6. Cincotta, P. M., Helmi, A., Méndez, M., Núñez, J. A., Vucetich, H., Astronomical time-
series analysis -II. A search for periodicity using the Shannon entropy, Monthly Notices of
the Royal Astronomical Society, 302, 582 (1999)
7. Cincotta, P. M., Astronomical time-series analysis -III. The role of the observational errors
in the minimum entropy method, Monthly Notices of the Royal Astronomical Society, 307,
941 (1999)
8. Cincotta, P. M., Arnold diffusion: an overview through dynamical astronomy, New Astron-
omy Reviews, 46, 13-39 (2002)
9. Cincotta, P. M., Efthymiopoulos, C., Giordano, C. M., Mestre, M. F., Chirikov and
Nekhoroshev diffusion estimates: bridging the two sides of the river, Physica D, 266, 49
(2014)
10. Cincotta, P. M., Giordano, C. M., Martí, J. G., Beaugé, C., On the chaotic diffusion in
multidimensional Hamiltonian systems, Celestial Mechanics and Dynamical Astronomy, 130
article id. #7 (2018)
11. Chandler, D., Introduction to Modern Statistical Mechanics, (New York: Oxford University
Press), (1987)
12. Efthymiopoulos, C., Harsoula, M., The speed of Arnold diffusion, Physica D, 251, 19
(2013)
13. Froeschlé, C., Guzzo, M., Lega, E., Local and global diffusion along resonant lines in
discrete quasi-integrable dynamical systems Celestial Mechanics and Dynamical Astronomy,
92, 243 (2005)
14. Froeschlé, C., Lega, E., Guzzo, M., Analysis of the chaotic behavior of orbits diffusing
along the Arnold web, Celestial Mechanics and Dynamical Astronomy, 95, 141 (2006)
15. Giordano, C.M, Cincotta, P.M., The Shannon entropy as a measure of diffusion in multi-
dimensional dynamical systems, CMDA 130, article id. N◦ 35 (2018)
16. Guzzo, M., Lega, E., Froechlé, C., First numerical evidence of global Arnold diffusion in
quasi–integrable systems Discrete and Ccontinuous Dynamical Systems B 5, 687 (2005)
17. Katz, A., Principles of Statistical Mechanics, The Information Theory Approach, (San
Francisco: W.H. Freeman & Co.), (1967)
18. Klafter, J., Blumen, A., Zumofen, G., Shlesinger, M., Lévy walk approach to anomalous
diffusion, Physica A, 168, 637 (1990)
19. Klafter, J., Zumofen, G., Shlesinger, M., Lévy walks in dynamical systems, Physica A,
200, 222 (1993)
20. Korabel, N., Klages, R., Microscopic chaos and transport in many-particle systems, Phys-
ica D, 187, 66 (2004).
18 P. M. Cincotta and C. M. Giordano

21. Lega, E., Froeschlé, C., Guzzo, Diffusion in Hamiltonian quasi-integrable systems, Lecture
Notes in Physics, 729, 29 (2008)
22. Lega, E., Guzzo, M., Froeschlé, C., Detection of Arnold diffusion in Hamiltonian systems,
Physica D, 182, 179 (2003)
23. Lesne, A., Shannon entropy: a rigorous notion at the crossroads between probability, in-
formation theory, dynamical systems and statistical physics, Mathematical Structures in
Computer Science, 24, e240311, doi:10.1017/S0960129512000783 (2014)
24. Maffione, N. P., Gómez F. A., Cincotta, P. M.,Giordano, C. M., Cooper, A. P., O’ Shea,
B.W., On the relevance of chaos for halo stars in the solar neighbourhood, Monthly Notices
of the Royal Astronomical Society, 453, 2830 (2015)
25. Maffione, N. P., Gómez F. A., Cincotta, P. M.,Giordano, C. M., Grand, R., Marinacci, F.,
Pakmor, R., Simpson, C., Springel, V., Frenk, C., On the relevance of chaos for halo stars in
the solar neighbourhood II, Monthly Notices of the Royal Astronomical Society, 478, 4052
(2018)
26. Martí, J. G.; Cincotta, P. M.; Beaugé, C. Chaotic diffusion in the Gliese-876 planetary
system, Monthly Notices of the Royal Astronomical Society, 460, 1094 (2016)
27. Miguel, N, Simó, C, Vieiro, A., On the effect of islands in the diffusive properties of the
standard map, for large parameter values, Foundations of Computational Mathematics, 15,
89 (2014)
28. Reichl, L.E., A Modern Course in Statistical Physics, (New York: Wiley-Interscience),
1998.
29. Shannon, C., Weaver, W., The Mathematical Theory of Communication, (Urbana: Illinois
U.P.), (1949)
30. Schwarzl, M., Godec, A., Metzler, R., Quantifying non-ergodicity of anomalous diffusion
with higher order moments, Scientific Reports, 7, doi:10.1038/s41598-017-03712-x (2017)
31. Venegeroles, R. Calculation of superdiffusion for the Chirikov-Taylor model, Physical Re-
view Letters, 101, 54102 (2008)
32. Wehrl, A., General properties of entropy, Review of Modern Physics, 50, 2, 221 (1978)
33. Zaslavsky, G. M., Fractional kinetic equation for Hamiltonian chaos, Physica D, 76, 110
(1994)
34. Zaslavsky, G. M., Abdullaev, S.S., Scaling properties and anomalous transport of particles
inside the stochastic layer, Physical Review E, 51, 3901, 1995
35. Zaslavsky, G. M., Edelman, M., Niyazow, B. A., Self-similarity, renormalization, and phase
space nonuniformity of Hamiltonian chaotic dynamics, Chaos, 7, 159, (1997)
36. Zaslavsky, G. M., Niyazow, B. A., Fractional kinetics and accelerator modes, Physics
Reports, 283, 73 (1997)
37. Zaslavsky, G. M., Edelman, M., Hierarchical structures in the phase space and fractional
kinetics: I. Classical systems, Chaos, 10, 135 (2000)
38. Zaslavsky, G. M. Chaos, fractional kinetics, and anomalous transport Physics Reports,
371, 461 (2002)