Physica A 269 (1999) 79– 89

[Link]/locate/physa

Discrete random walk models for symmetric

Levy–Feller diusion processes
Rudolf Goren oa , Gianni De Fabritiisb , Francesco Mainardic; ∗
a Department of Mathematics and Computer Science, Free University of Berlin,
D-14195 Berlin, Germany
b CINECA, Supercomputing Center, I-40033 Casalecchio di Reno, Bologna, Italy
c Dipartimento di Fisica, Universit
a di Bologna and INFN, Sezione di Bologna, Via Irnerio 46,
I-40126 Bologna, Italy

Received 5 November 1998; received in revised form 27 November 1998

Abstract
We propose a variety of models of random walk, discrete in space and time, suitable for
simulating stable random variables of arbitrary index  (0 ¡ 62), in the symmetric case. We
show that by properly scaled transition to vanishing space and time steps our random walk models
converge to the corresponding continuous Markovian stochastic processes which we refer to as
Levy–Feller diusion processes. c 1999 Elsevier Science B.V. All rights reserved.

PACS: 02.50.-r; 02.70.-c; 05.40.+j

Keywords: Random walks; Stable probability distributions; Diusion

1. Introduction

By a Levy–Feller diusion process we mean a Markovian process governed by a

stable probability density function (pdf) evolving in time, g (x; t; ), whose spatial
Fourier transform (the characteristic function) reads
Z +∞
ĝ (; t; ) = eix g (x; t; ) d x = exp(−t|| ei(sign )=2 ) ; (1.1)
−∞

where x;  ∈ R; t ¿ 0. The two relevant parameters, ; called the index of stability, and
 (related to the asymmetry), improperly referred to as the skewness, are real numbers

∗Corresponding author. Fax: +39-051-247244.

E-mail address: mainardi@[Link] (F. Mainardi)

0378-4371/99/$ - see front matter c 1999 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 8 - 4 3 7 1 ( 9 9 ) 0 0 0 8 2 - 5
80 R. Goren o et al. / Physica A 269 (1999) 79– 89

subject to the conditions (see, e.g. [1])


; if 0 ¡ 61 ;
0 ¡ 62; ||6 (1.2)
2 − ; if 1 ¡ 62 :

By introducing the similarity variable xt −1= , we can write g (x; t; )=t −1= p (xt −1= ; ),
where p (x; ) is the stable pdf at t =1. The specic form of the characteristic function
(1.1) allows us to recognize g (x; t; ) as the Green function (fundamental solution) of
the Cauchy problem
@
u(x; t) = D [u(x; t)]; u(x; 0) = (x); x ∈ R; t ¿ 0 ; (1.3)
@t
where D is the pseudo-dierential operator with symbol
b  = −|| ei(sign )=2 :
D (1.4)

Let us recall that a generic pseudo-dierential operator A, acting with respect to the
R +∞
variable x ∈ R; is dened through its Fourier representation, namely −∞ eix A[(x)]
d x= A() ˆ
ˆ (), ˆ
where (x) denotes a suciently well-behaved function in R, and A()
is referred to as symbol of A, given as A()ˆ = (Ae−ix )e+ix .
With the names of Levy and Feller we have intended to honour both Paul Levy [4]
who rst introduced the class of stable distributions, see [2– 4] and William Feller [5]
who rst investigated the semigroups generated by a pseudo-dierential equation of
type (1.3) – (1.4). For  = 2 and  = 1 (with  = 0) we recover the standard Gaussian
and Cauchy pdf’s
 2
1 x 1 t
g2 (x; t; 0) = √ t −1=2 exp − ; g1 (x; t; 0) = : (1.5)
2  4t  x2 + t 2

In physics, the rst recognition that the (symmetric) Levy distribution could be char-
acterized via a pseudo-dierential operator of type (1.3) – (1.4) was made explicitly by
West and Seshadri [6]. Recently, Goren o and Mainardi [7–9] have revised Feller’s
original arguments by interpreting (1.3) as a space-fractional diusion equation (of or-
der  and “skewness” ) and have provided a variety of related random walk models,
discrete in space and time, which by properly scaled transition to vanishing space and
time steps converge to the corresponding continuous Levy–Feller processes. In other
words the discrete probability distributions generated by the random walk models have
been proved to belong to the domain of attraction of the corresponding stable distri-
bution.
Here, limiting ourselves to the symmetric case ( = 0), we present the main features
of the random walk models by Goren o and Mainardi, and we display preliminary
results of a few numerical case studies, which can be of some interest in econophysics.
The Levy statistics in modelling uctuations of economical and nancial variables,
formerly used by Mandelbrot in the early sixties, is still adopted with success, see, e.g.
[10 –12].
 R. Goren o et al. / Physica A 269 (1999) 79– 89 81

2. Outline of the general theory

Let Y be an integer-valued random variable and let the random variables Y1 ; Y2 ; Y3 ; : : :

be iid (= independent identically distributed), all having their probability distribution
common with Y . We dene a spatial-temporal grid {(xj ; tn ) | j ∈ Z; n ∈ N0 } by xj =
xj (h) = jh; tn = tn () = n, where h ¿ 0 and  ¿ 0. Then we consider the sequence of
random variables
Sn = hY1 + hY2 + · · · + hYn ; n∈N ; (2.1)
with (for convenience) S0 =0, and interpret it as follows. A particle, sitting in x =x0 =0
at time t = t0 = 0 nds itself at a later instant t = tn in point x = Sn which is an integer
multiple of h. We recognize the pk = P(Y = k) (for k ∈ Z) as “transition probabilities”:
pk is the probability of a particle jumping from a point xj = Sn to a point xj+k = Sn+1
as time proceeds from tn to tn+1 . All pk are non-negative, and their sum equals 1.
The probability yj (tn ) of sojourn of our particle in point xj at instant tn obeys the
transition law
+∞
X
yj (tn+1 ) = pk yj−k (tn ); yj (0) = j0 ; j ∈ Z; n ∈ N0 (2.2)
k=−∞

which has the form of a discrete convolution. Hence, introducing generating functions
+∞
X +∞
X
p̃(z) = pj z j ; ỹ n (z) = yj (tn )z j ; (2.3)
j=−∞ j=−∞

we obtain
ỹ n (z) = ỹ 0 (z)[p̃(z)]n = [p̃(z)]n ; n ∈ N0 : (2.4)
The power series in (2.3) and (2.4) are absolutely and uniformly convergent on |z| = 1
and assume the value 1 at z = 1. Putting z = eih ;  ∈ R, and observing z j = eijh = ei xj ,
we recognize p̂(; h) = p̃(eih ) and ŷ(; tn ; h) = ỹ n (eih ) as characteristic functions of
the random variables hY and Sn , respectively.
Our aim is to approximate the Levy–Feller diusion process, which is governed by
the evolution equation (1.3), arbitrarily well. To this purpose we introduce a strictly
monotonic scaling relation  = (h) → 0 as h → 0. We will x t ¿ 0 and let h (and
likewise ) go to zero over such values that always n (=t= = t=(h)) is a positive
integer. Then we have the equivalences
n→∞ ⇔ h→0 ⇔ →0 ;
and h depends on , nally on n, so that h=h(n): Replacing h by h(n) in (2.1) we obtain
a sequence of random variables Xn with characteristic functions ŷ(; tn ; h) = [p̃(eih )]n
(note that now tn = t is xed). Invoking Theorem 3:6:1 of Lukacs [13], what remains
to be shown is that ŷ(; t; h) → exp(−t|| ) as h → 0, the characteristic function of the
corresponding symmetric Levy–Feller process. For this it suces that, for xed  6= 0,
t
log[ŷ(; t; h)] ≡ log[p̃(eih )] → − t|| as h → 0 : (2.5)
(h)
82 R. Goren o et al. / Physica A 269 (1999) 79– 89

Our random walk can be interpreted as a “dierence scheme” to approximate the
evolution equation (1.3), if we write (2.2) in the equivalent form, observing the scaling
relation,
" #
yj (tn+1 ) − yj (tn ) 1 X
= (p0 − 1)yj (tn ) + pk yj−k (tn ) : (2.6)
 (h)
k6=0

In fact, after division by the spatial mesh-width h, the LHS is the explicit discrete ap-
proximation to the rst-order time derivative @u(x; t)=@t, and the RHS can be considered
as a particular discrete approximation to the space pseudo-dierential term D0 [u(x; t)],
R x +h=2
provided we mean yj (tn ) = xjj−h=2 u(x; tn ) d x ≈ hu(xj ; tn ) with yj (0) = j 0 .

3. The random walk models

From the previous section we have learnt that, in order to construct discrete random
walk models which are convergent (in distribution) to the (symmetric) stable pdf’s, the
clue points are: (1) to guess a suitable generating function p̃(z), whose coecients
of its power series expansion provide the transition probabilities, (2) to determine the
corresponding scaling relation  = (h) which ensures the required convergence.
In the classical case of the Gaussian distribution (=2) the matter is easily treated if
we remember that the corresponding density is the fundamental solution of the standard
diusion equation, which is known to be well approximated via the nite-dierence
equation
yj (tn+1 ) − yj (tn ) yj+1 (tn ) − 2yj (tn ) + yj−1 (tn )
= ; yj (0) = j 0 : (3.1)
 h2
In this case, introducing the scaling parameter  = =h2 , so  = (h) = h2 , the transition
probabilities turn out to be
p0 = 1 − 2; p±1 = ; p±k = 0; k = 2; 3 : : : ; (3.2)
subject to the condition 0 ¡ 61=2. Thus the generating function is
p̃(z) = 1 + [z − 2 + z −1 ] : (3.3)
The proof of the convergence to the Gaussian is simple since one easily nds [t=(h2 )]×
log[p̃(eih ] → − t2 as h → 0. The scheme (3.2) means that for approximation of the
standard Gaussian process the corresponding random walk model exhibits only jumps
of one step to the right or one to the left or jumps of width zero. For the stable
non Gaussian processes we expect to nd a non-polynomial generating function with
innitely many transition coecients which imply the occurrence of arbitrarily large
jumps. It is common practice to refer to the corresponding random walks as to Levy
ights.
In the following, limiting ourselves to the symmetric cases ( = 0), we shall re-
sume the main features of three dierent random walk models, referred to as (RW1),
(RW2), (RW3), of which Goren o and Mainardi have proved the convergence to the
 R. Goren o et al. / Physica A 269 (1999) 79– 89 83

corresponding continuous processes. For each model we give the generating function
p̃(z) with the transition probabilities pk and the scaling relation  = (h), referring
to the original papers for details. From the analysis of the classical Gaussian case we
nd it natural to rst introduce the scaling parameter  = =h with 0 ¡ 62 but, as
we shall show later, this will not necessarily imply (h) = h , with  constant for
xed .
The model (RW1) has been introduced and discussed in [7], starting from the iden-
tication of the operator D0 in the framework of fractional calculus, see [5], and then
applying, in the authors’ original approach, the Grunwald–Letnikov discretized scheme.
For this model we need to keep distinct the two cases (a) 0 ¡  ¡ 1 and (b) 1 ¡ 62,
the case  = 1 being excluded in this treatment. The generating function is
 
  −1 
 1 − 2 cos(=2) [(1 − z) + (1 − z ) ]; 0¡¡1 ;
p̃(z) =  (3.4)

1− [z −1 (1 − z) + z(1 − z −1 ) ]; 1 ¡ 62 :
2 cos(=2)
For both cases the scaling relation is conrmed to be  = (h) = h , but the parameter
 is subject to dierent restrictions to ensure that 06p0 ¡ 1.
In the case (a) we have 0 ¡ 6 cos(=2) and
 

 p0 = 1 − cos(=2) ;
  (3.5a)

 p±k = (−1)k+1 for k = 1; 2; 3; : : : :
2 cos(=2) k
In the case (b) we have 0 ¡ 6|cos(=2)|= and
   h   i

 p = 1 − ; p = + 1 ;
 0 |cos(=2)|
±1
2|cos(=2)| 2
  (3.5b)

  
 p±k = (−1)k+1 for k = 2; 3; 4; : : : :
2|cos(=2)| k + 1
We note that, whereas the classical Gaussian random walk (3.2) is promptly recovered
from (3.5b) for  = 2, the random walk for the Cauchy process ( = 1) cannot be
obtained, neither directly nor by a passage to the limit  → 1. Indeed, in both the
limits  → 1− and  → 1+ the permissible range of the scaling factor  is vanishing.
In numerical practice the consequence will be that if  is near 1 the convergence is
slow: for good approximation we will need a very small step-time  with respect to
the step-length h.
Apart from (RW1) we shall show how the two other random walk models, (RW2)
and (RW3), that we are going to brie y discuss, exhibit a smooth scaling law in the
range 0 ¡  ¡ 2, so the case  = 1 is no longer singular. However, “there is no free
lunch”. We have to pay for the good behaviour at  = 1 with bad behaviour at  = 2
in a sense to be seen later.
The model (RW2) is obtained by imposing for the transition probabilities an expres-
sion in terms of binomial coecients, as suggested from (3.5b), which is required to
84 R. Goren o et al. / Physica A 269 (1999) 79– 89

be regular in the limit as  → 1, namely

 
 
p0 = 1 − 2; p±k = (−1)k+1 for k = 1; 2; 3; : : : ; (3.6)
−1 k +1
where , subject to the condition 0 ¡ 61=2, is to be determined. The model has been
extensively discussed in [9] whereas the particular case  = 1 (related to the Cauchy
process) has been formerly presented in [8]. To guarantee the convergence for all
 (0 ¡ 62) to the corresponding continuous process, the asymptotics as h → 0 still
requires a scaling relation of the kind  = h (with  constant), namely

 2 cos(=2) 

 h if 0 ¡  ¡ 2;  6= 1 ;
 1−
 = (h) = h if  = 1 ; (3.7)



 2
h if  = 2 :
We recognize that for 0 ¡  ¡ 2 the model (RW2) exhibits a smooth scaling law but
a discontinuity is present at  = 2 as can be seen by taking the limit as  → 2− . This
allows us to recover for  = 2 the Gaussian model (3.2) – (3.3), but it shows that
in numerical practice when  is near 2 the convergence is expected to be slow. For
0 ¡  ¡ 2 we obtain the generating functions and the transition probabilities as follow.
Putting (z) = [(1 − z)−1 − 1] for  6= 1, and introducing the scaling parameter , we
have
 

1− [(1 − z −1 )(z) + (1 − z)(z −1 )];  6= 1 ;
2 cos(=2)
p̃(z) = (3.8)
 1 −  [(1 − z −1 )log(1 − z) + (1 − z)log(1 − z −1 )];  = 1 :


For 0 ¡  ¡ 2;  6= 1, we have 0 ¡ 6cos(=2)=(1 − ) and
 (1 − )

 p0 = 1 − cos(=2) ;

  (3.9)

  
 p±k = (−1)k for k = 1; 2; 3; : : : ;
2 cos(=2) k + 1
for  = 1 we have 0 ¡ 6=2 and
2  1
p0 = 1 − ; p±k = for k = 1; 2; 3; 4; : : : : (3.10)
  k(k + 1)
In both models (RW1), (RW2), the generating function is expressed in terms of ele-
mentary functions and the transition coecients, for 0 ¡  ¡ 2; exhibit an asymptotic
behaviour as |k| → ∞ consistent with that of the power-law tails of the stable densities
as |x| → ∞; see, e.g. [1]. Indeed we obtain
sin(=2) −(+1)
pk ∼  ( + 1) |k| as |k| → ∞; 0 ¡  ¡ 2 ; (3.11)


sin(=2) −(+1)

p (x; 0) ∼ ( + 1) |x| as |x| → ∞; 0 ¡  ¡ 2 : (3.12)

 R. Goren o et al. / Physica A 269 (1999) 79– 89 85

The asymptotic behaviour of the stable densities is the starting point for the random
walk model (RW3) in that, following Gillis and Weiss [14], we require that all pk
for k 6= 0 are proportional to |k|−(+1) : However, the simplicity of the starting point
leads to diculties for treating this model since the generating function is no longer
elementary, a fact that makes the convergence proof a really hard aair. We need to
recall the following special functions
∞
X ∞
X zk
() = k − ;  ¿ 1; (z; ) = ; |z| ¡ 1;  ∈ R ; (3.13)
k
k=1 k=1

respectively, known as the Riemann zeta function and the polylogarithmic function.
This model has been extensively discussed in [9] (see also [14]), where the following
generating function is derived:
p̃(z) = 1 − 2( + 1) + [(z;  + 1) + (z −1 ;  + 1)] : (3.14)
Here  =  + 1 ¿ 1 so (z; ) is by its power series also dened on the periphery
|z| = 1 of the unit circle. Then we get a pure power-law random walk with
p0 = 1 − 2( + 1); pk = |k|−(+1) for k 6= 0 ; (3.15)
where ; subject to the condition 0 ¡ 61=[2( + 1)]; is to be determined. To ensure
the convergence to the corresponding continuous process the following scaling relation
must hold:

 
h if 0 ¡  ¡ 2 ;
 = (h) = ( + 1)sin(=2) (3.16)
 2
h |log h| if  = 2 :
It is interesting to note that in the case  = 2 the classical random walk model (3.2) –
(3.3) is no longer recovered, since now arbitrarily large jumps occur (with a probability
decaying as k −3 ) as in the Levy ights. Nevertheless, through the scaling relation
(3.16), this random walk converges to the continuous Gaussian process. Thus this
model gives us the opportunity to verify that a random walk with innite variance
steps may (slowly) converge to the Gaussian, in agreement with a general theorem on
the domain of attraction of the normal law, see, e.g. [15].

4. Numerical results

In general the random walk models are not only valuable from the conceptual point
of view for visualizing what the diusion means but also for numerical calculations,
either as Monte Carlo simulation of particle paths in a diusion process or as discrete
imitation of the process in form of redistribution (from one time level to the next) of
clumps of an extensive quantity (across the spatial grid points). Our models can be
used in at least three dierent ways: (a) as nite dierence schemes for approximate
calculation of symmetric stable densities; (b) for producing sample paths of individual
particles performing the random walk; (c) for producing histograms of the approximate
86 R. Goren o et al. / Physica A 269 (1999) 79– 89

realization of the densities g by simulating many individual paths with the same
number of time steps and making statistics of the nal positions of the particles.
For numerical simulations of stable random variables dierent algorithms have been
provided by a number of specialists, including Chambers et al. [16], Bartels [17],
Mantegna [18], Janicki and Weron [19], Samorodnitsky and Taqqu [20]. Our present
approach for treating Levy statistics has been carried out independent of all the above
references but uniquely based on the random walk models presented here, so, as far
as we know, our results would be in the great part original.
Having preliminarily checked a sucient level of accuracy for our nite dierence
schemes with the existing tables of stable densities, here we present some results on the
simulation of the sample paths and histograms corresponding to some typical values of
the index of stability, namely  = 1; 1:5; 2, in Figs. 1, 2, 3, respectively. In practice, in
our numerical studies truncation is required in two ways. It is impossible to simulate
all innitely many discrete probabilities, so the size of possible jumps must be limited
to a maximal possible jump length. The other truncation is required if a priori one
wants a denite region of space to be considered in which the walk takes place. Then,
particles leaving this space have been ignored. Our simulations, based on one million
of realizations, have been carried out in the interval |x|64. All the histograms refer
to stable densities at t = 1 for |x|63; the space interval being reduced to avoid the
border eects. The sample paths are plotted against the time steps, up to 1200 for
 = 1 and up to 2000 for  = 1:5; 2; so they refer to dierent nal times, namely
t = 1:87; 1:33; 0:8; respectively. The transition probabilities have been chosen from our
random walk models as follows:  = 1 from (RW2), see (3.10), with scaling parameter
 = =4;  = 1:5; from (RW1), see (3.5b), with  = (2=3)cos(3=4); for  = 2; from the
standard model, see (3.2), with  = 1=4: The cases  = 1 (Cauchy process) and  = 2
(normal process) have been considered for a possible comparison with the standard

Fig. 1. Sample path (left) and histogram (right) for  = 1 (Cauchy).

 R. Goren o et al. / Physica A 269 (1999) 79– 89 87

Fig. 2. Sample path (left) and histogram (right) for  = 1:5.

Fig. 3. Sample path (left) and histogram (right) for  = 2 (Gauss).

and accurate algorithms existing in the literature, whereas  = 1:5 has been chosen
in view of possible applications in econophysics where usually the index of stability
ranges from 1.4 to 1.7, see, e.g. [10,11].

5. Conclusions

For the simulation of Markovian processes characterized by symmetric Levy proba-

bility densities evolving in time we have presented three dierent random walk models,
discrete in space and time, by giving their respective transition probabilities. We have
88 R. Goren o et al. / Physica A 269 (1999) 79– 89

indicated how, by use of generating functions, convergence for properly scaled tran-
sition to vanishing steps of space and time can be analysed, and we have hinted at
peculiarities occurring at the particular stability indices  = 1 (the Cauchy process) and
 = 2 (the Gauss process). We have displayed preliminary results of a few numerical
case studies concerning sample paths and histograms to check the eciency of our
algorithms. From the sample paths one can recognize the “wild” character of the Levy
ights with respect to the “tame” character of the Brownian motion.
We expect that our arguments can be relevant in dierent elds of physics includ-
ing the emerging one of econophysics, where stable distributions are becoming more
common. In statistical physics the stable distributions play a key role in the (wonder-
ful) world of random walks constructed by the late Montroll and continued through
his school, see, e.g. [21–23]. Also Tsallis and his associates have recognized the key
role of stable distributions with respect to a generalized theory of thermostatics, see,
e.g. [24], and references therein. Here, we have (only brie y) pointed out the rela-
tion between Levy statistics and space-fractional diusion equations. However, stable
distributions turn out to be related also with time-fractional diusion equations, see,
e.g. [25]. Furthermore, this topic is relevant for fractal phenomena, where dierential
equations of fractional order are usually adopted to describe their evolution, see, e.g.
[26 –28], and reference herein.

Acknowledgements

We gratefully acknowledge fruitful discussions with Rosario Mantegna. This work

was supported in part by the Italian CNR and INFN, and by the Research Commission
of Free University of Berlin.

References

[1] W. Feller, An Introduction to Probability Theory and its Applications, vol. 2, Wiley, New York, 1971.
[2] P. Levy, Theorie des erreurs. La lois de Gauss et les lois exceptionnelles, Bull. Soc. Math. France 52
(1924) 49–84.
[3] P. Levy, Calcul des probabilites, Gauthier-Villars, Paris, 1925.
[4] P. Levy, Theorie de l’Addition des Variables Aleatoires, Gauthier-Villars, Paris, 1937.
[5] W. Feller, On a generalization of Marcel Riesz’ potentials and the semi-groups generated by them,
Meddelanden Lunds Universitets Matematiska Seminarium (Comm. Sem. Mathem. Universite de Lund)
Tome suppl. dedie a M. Riesz, Lund (1952) 73–81.
[6] B.J. West, V. Seshadri, Linear systems with Levy uctuations, Physica A 113 (1982) 203–216.
[7] R. Goren o, F. Mainardi, Random walk models for space-fractional diusion processes, Fractional
Calculus Appl. Anal. 1 (1998) 167–191.
[8] R. Goren o, F. Mainardi, Approximation of Levy-Feller diusion by random walk, J. Anal. Appl. 18
(1999), in press.
[9] R. Goren o, F. Mainardi, A collection of random walks for Levy-Feller diusion processes, in
preparation.
[10] B.B. Mandelbrot, Fractals and Scaling in Finance, Springer, New York, 1997.
 R. Goren o et al. / Physica A 269 (1999) 79– 89 89

[11] R.N. Mantegna, H.E. Stanley, Physics investigation of nancial markets, in: F. Mallamace,
H.E. Stanley (Eds.), Proceedings International School of Physics “E. Fermi”, Course CXXXIV, IO
Press, Amsterdam, 1997, pp. 473– 489.
[12] J.-P. Bouchaud, M. Potters, Theorie des Risques Financiers, Comm. Energie Atomique, Alea Saclay,
Paris, 1997.
[13] E. Lukacs, Characteristic Functions, Grin, London, 1960.
[14] J.E. Gillis, G.H. Weiss, Expected number of distinct sites visited by a random walk with an innite
variance, J. Math. Phys. 11 (1970) 1307–1312.
[15] B.V. Gnedenko, A.N. Kolmogorov, Limit Distributions for Sums of Independent Random Variables,
Addison-Wesley, Cambridge, MA, 1954.
[16] J.M. Chambers, C.L. Mallows, B.W. Stuck, A method for simulating stable random variables, J. Amer.
Statist. Assoc. 71 (1976) 340 –344.
[17] R. Bartels, Generating non-normal stable variates using limit theorems properties, J. Statist. Comp.
Simulation 7 (1978) 199–212.
[18] R.N. Mantegna, Fast, accurate algorithm for numerical simulation of Levy stable stochastic processes,
Phys. Rev. E 49 (1994) 4677– 4683.
[19] A. Janicki, A. Weron, Simulation and Chaotic Behavior of -Stable Stochastic Processes, Marcel Dekker,
New York, 1994.
[20] G. Samorodnitsky, M.S. Taqqu, Stable non-Gaussian Random Processes, Chapman&Hall, New York,
1994.
[21] E.W. Montroll, B.J. West, On an enriched collection of stochastic processes, in: E.W. Montroll,
J. Leibowitz (Eds.), Fluctuation Phenomena, North-Holland, Amsterdam, 1979, pp. 61–175.
[22] E.W. Montroll, M.F. Shlesinger, On the wonderful world of random walks, in: J. Leibowitz, E.W.
Montroll (Eds.), Nonequilibrium Phenomena II: from Stochastics to Hydrodynamics, North-Holland,
Amsterdam, 1984, pp. 1–121.
[23] J. Klafter, M.F. Shlesinger, G. Zumofen, Beyond Brownian motion, Phys. Today 49 (2) (1996) 33–39.
[24] C. Tsallis, Levy Distributions, Phys. World 10 (7) (1997) 42– 45.
[25] F. Mainardi, P. Paradisi, R. Goren o, Probability distributions generated by fractional diusion equations,
in: J. Kertesz, I. Kondor (Eds.), Econophysics: an Emerging Science, Kluwer, Academic Publishers,
Dordrecht, 1999, in press, pp. 39.
[26] A. Carpinteri, F. Mainardi (Eds.), Fractals and Fractional Calculus in Continuum Mechanics, Springer,
Wien, 1997.
[27] G.M. Zaslavsky, Physics of Chaos in Hamiltonian Systems, Imperial College Press and World Scientic,
London and Singapore, 1998.
[28] R. Hilfer (Ed.), Applications of Fractional Calculus in Physics, World Scientic, Singapore, 1999.