Escolar Documentos
Profissional Documentos
Cultura Documentos
of the calculations of the stable function and it presents the same asymptotic behaviour as the stable distribution function. To
obtain the uniform analytical approximation two types of approximations are developed. The first one is called “inner solution”
and it is an asymptotic expansion around x = 0. The second refers to “outer solution” and it is a series expansion for x → ∞.
Then, the uniform solution is proposed as a result of an asymptotic matching between inner and outer solutions. Finally, the
results of analytical approximation were compared to the numerical results of the stable distribution function, making this
uniform solution valid to be applied as an analytical approximation.
of them are only applicable for extreme values x → 0 or purpose, two types of approximations are developed. The
x → ∞. Mantegna [22] presented a quite similar solu- first approximation refers to the inner expansion that
tion to Elliot [21] but his algorithm is only valid when converges asymptotically to the stable distribution as
x → ∞ and 0.75 < α ≤ 1.95. Nolan [23] presented his al- x → 0. The second approximation refers to the outer
gorithms considering asymmetric distributions for large expansion that converges asymptotically as x → ∞. For
events x → ∞ focusing on tail behaviour only. Thus, the outer expansion two cases are presented, one is ob-
the stable distribution function does not have an explicit tained from the stable function in subsection IV B and
expression [24, 25] and no uniform solution of the stable the second one from trans-stable function in subsection
distribution has been proposed until now [18, 19, 23]. IV C. Finally, the uniform solution in section V is pro-
Due to the absence of an explicit expresion, numerical posed as a result of matching the inner and the outer so-
solutions were developed to evaluate the stable distri- lution. The analytical equation of uniform solution pro-
bution function by using numerical recursive quadrature posed in this paper gives an approximated solution of the
methods [26–28]. Nolan [26, 29] develops a numerical so- stable distribution function over the range −∞ < x < ∞.
lution for the estimation of stable parameters through a
maximun likelihood method for each data set of x. How-
ever, Nolan’s method converges only for α > 0.4 and II. STABLE DISTRIBUTION FUNCTION
the convergence to the stable distribution function seems
to be not accurate enough. Despite this fact, Nolan’s The stable distribution is presented in Section 34 of the
method constitutes an important method that is still be- Gnedenko-Kolmogorov book [14] as the Fourier transfor-
ing used [27]. mation:
Apart from the numerical issues in the evaluation of
the stable distribution, some authors have pointed out its
infinite variance as a drawback [30–32]. To avoid the in- Z∞
1
finite variance of the stable distribution function, several f (x; α, β, σ, µ) = ϕ(t; α, β, σ, µ)eixt dt, (2)
2π
truncations are proposed. The truncation was justified −∞
by the observed change of slope of the tails on extensive
datasets [33]. For example, by increasing the S&P 500 where ϕ(t) is defined as the characteristic function,
index data from six-year period (1984-89) to 13-year pe- α
riod (1984-96) a change of slope from α = 1.4 to α = 3 ϕ(t; α, β, σ, µ) = e(itµ−|σt| (1−iβsgn(t)Φ))
. (3)
was found. Inoue and Sakuza [34] proposed a new model
by defining the effective cutoff as the crossover point be- The four parameters involved are: the stability param-
tween the non-Gaussian Stable and the Gaussian regime. eter α ∈ h0 2], the skewness parameter β ∈ [−1 +1] , the
The truncations make the variance finite, consequently scale parameter σ ∈ (0 +∞) , and the location parameter
the distribution function of the sum of independent ran- µ ∈ (−∞ + ∞). The parameter α constitutes the char-
dom variables converges to the normal distribution due acteristic exponent of the asymptotic power-law in the
to the central limit theorem for a large N value. Nev- tails and it determines whether the mean value and the
ertheless, a time series in some stock markets indexes variance exist. The stable distribution with 0 < α ≤ 1
can exhibit infinite variance, one of these cases is the does not have a mean value and it has a define variance
variance of price fluctuations in Shanghai stock market only for α = 2 [37].
index, which increases when the time frame is enlarged
[35, 36]. The function sgn (t) represents the sign of t and the
The aim of this paper is to formulate a uniform an- function Φ is defined as:
alytical approximation for the stable distribution func-
tan πα
α 6= 1,
Φ= 2 (4)
tion based on a series expansion. To achieve this aim − π2 log |t| α = 1.
we propose several regularizations of the inner and outer
series expansions to ensure convergence. This will be an The stable distribution is the family of all attractors
important tool to get the most accurate approximation of normalized sums of independent and identically dis-
reducing numerical errors (oscillations) when the stable tributed random variables. The most well-known stable
function is evaluated. distribution functions are the Cauchy distribution with
This paper is divided in two parts. The first part α = 1 and the normal distribution function with α = 2.
introduces the stable distribution and the trans-stable Both functions have β = 0, which means they are sym-
function. They are defined by Fourier transformations in metric distributions about their mean [18].
sections II and III respectively. The trans-stable func- In this paper we will focus on symmetric distributions
tion is shown to be identical to the stable distribution (β = 0). For this case the stable distribution can be
for α < 1 and it has the same asymptotic behaviour normalized as follows:
for α > 1 for large events. The second part refers to
section IV and it deals with the closest form— analyti- x−µ
f (x; α, β = 0, σ, µ) = Re S ,α , (5)
cal approximations— of the stable distribution. For this σ
3
where the general distribution function is given by the First it will be shown that for 0 < α ≤ 1, both
following equation: stable S(x; α) and trans-stable T (x; α) functions are
identical. For 1 < α < 2 it will be demonstrated that
Z∞ both functions exhibit the same asymptotic behaviour
1 α
S(x; α) = e−t eixt dt. (6) when x → ∞.
π
0
This demonstration is based on the evaluation of the
The real part of this function corresponds to the normal- complex trans-stable integral Eq. (8) using polar repre-
ized stable distribution, sentation for α ≤ 1 and rectangular representation for
α > 1 on the complex integrand. The demonstrations
s(x; α) = Re(S(x; α)). are presented in the following subsections.
Consequently,
Im(z)
Z∞
1 α i
s(x; α) = e−t cos(tx)dt. (7)
π
0
C
III. TRANS-STABLE FUNCTION i 3
α
I (x, z; α) = e−z eixz . (9)
I
I (x, z; α) dz = 0. (12)
The relation of this function to the stable S(x; α) and
the trans-stable T (x; α) functions is obtained by choosing Let us take the contour in Figure 1 that can be divided
a particular path C in the complex plane. Then, the into four straight paths so that:
stable distribution and trans-stable function are given by
Eq. (10) and (11) respectively: 4 Z
X
Z∞ I (x, z; α) dz = 0. (13)
1 −t
α
ixt 1 Ci
S(x; α) = G[0,∞) (x; α) = e e dt, (10)
π Now, we will take the limit when τ → ∞ in Figure 1.
0
Using Eq. (10,11,13) the following equation is obtained:
Zi∞ 3 Z
1 α
T (x; α) = G[0,i∞) (x; α) = e−t eixt dt. (11)
X
π S(x; α)−T (x; α) = −limτ →∞ I (x, z; α) dz. (14)
0 i=2 Ci
4
To evaluate the right hand side in Eq. (14) it is con- Stable distribution Trans-stable function
venient to use the polar representation of the complex S(x; α) T (x; α)
Fourier R∞ −tα ixt i∞
number z = reiθ and express Eq. (9) in polar coordinates: 1
e e dt 1
R α
e−t eixt dt
transform π π
0 0
Laplace −i∞
−(it)α −xt
R∞ α
1 1
e−(it) e−xt idt
R
e e idt π
I (x, z; α) = eg(x,r,θ;α)+ih(x,r,θ;α) , (15) transform π
0 0
h(x, r, θ; α) = −rα sin (θα) + rx cos θ. (17) 100 =0.90 100 =0.90
=0.70 =0.70
S(x; )-Fourier
T(x; )-Laplace
It will be adopted the nomenclature of signal theory, =0.50 =0.50
where the polar notation separates the effects of instan- =0.40 =0.40
taneous amplitude |I| = eg and its instantaneous phase
h of a complex function [38]. Consequently, g(x, r, θ; α)
represents the attenuation factor and h(x, r, θ; α) repre-
sents the oscillation factor. 10-2 10-2
Now let us notice that limr→∞ g(x, r, θ; α) = −∞ for
0 < α ≤ 1 at any value of x. This statement is based on 0 10 0 10
the fact that cos(θα) ≥ 0 in the first quadrant for α ≤ 1. x
Consequently, limr→∞ I (x, z; α) = 0 so that the integral
x
of the right side of the Eq. (14) vanishes at τ → ∞,
100 =0.90 100 =0.90
therefore: =0.70 =0.70
S(x; )-Fourier
T(x; )-Laplace
=0.50 =0.50
S(x; α) = T (x; α) if 0 < α ≤ 1. (18) =0.40 =0.40
So, the Eq. (18) will allow to use the trans-stable function
T (x; α) instead of the stable distribution function S(x; α)
for 0 < α ≤ 1 in the numerical integration. This is with
10-2 10-2
the aim to remove numerical oscillation, specifically in
the tails. It is noticeable that the integration of the trans-
stable function T (x; α) in Eq. (11)is performed over the 100 101 100 101
imaginary axis. Applying the following change of variable x x
t → −it (formally done by defining u = −it so that du =
−idt and later replacing the dummy variable u by t inside Figure 2. Comparison of numerical integration 0 < α < 1
the integral), the trans-stable function is converted into between Fourier and Laplace transform of the stable S(x; α)
and the trans-stable T (x; α) functions using recursive adap-
a Laplace transformation. Consequently, the integration
tive Simpson quadrature method [39]. The absolute error
is performed over the real axis. The Fourier and Laplace tolerance of the method is ξ = 3.5 × 10−8 . Top plots are
representations for T (x; α) are shown in Eq. (19), shown in semi-logarithmic scale and in logarithmic scale.
Zi∞ Z∞
1 −tα ixt 1 α
T (x; α) = e e dt = e−(it) e−xt idt. (19) 2. For 1 < α < 2
π π
0 0
Figure 2 compares the Fourier representation of the Here we will shown that for 1 < α < 2 the stable
stable distribution function S(x; α) and the Laplace rep- and trans-stable functions have the same asymptotic be-
resentation of trans-stable function T (x; α). The inte- haviour on large events if the integrals are appropriately
gration is performed using a recursive adaptive Simpson truncated.
quadrature method [39]. It is evident that the Laplace Let us recall Eq. (16) for the attenuation factor,
representation removes the oscillations of the Fourier rep- g(x, r, θ; α) = −rα cos (θα) − rx sin θ.
resentation of the stable distribution for α < 1.
It is important to add that the stable distribution func- In the previous section, it was shown that cos(θα) is al-
tion and trans-stable function hold the same value for ways positive in the first quadrant of the complex plane
their Fourier and Laplace transform representations. The if 0 < α ≤ 1. Otherwise, if α > 1, then cos(θα) < 0 when
difference between each transform representation is the θ = π/2. Consequently, limr→∞ I(x, r, θ; α) = ∞ in this
axis in which each function is integrated. The expressions range, so that the right hand side of Eq. (14) can not be
are shown in Table I. neglected. Therefore S(x) 6= T (x) if α > 1.
5
3 Z
πα π
X h̄(x, t; α) = −tα sin + . (28)
S(x; α, τ ) − T (x; α, τ ) = − I (x, z; α) dz, (20) 2 2
i=2 Si
The instantaneous amplitude |I| ¯ = eḡ will be determined
where S(x; α, τ ) and T (x; α, τ ) are the truncated inte- by the attenuation factor in Eq. (27). For that reason,
grals in Eqs. (10) and (11) respectively: an analysis of ḡ(x, t; α) was made in Figure 3. The curve
ḡ(x, t; α) = 0 divides two regions, one with exponential
growth (ḡ > 0) and the other with exponential decay
Zτ
1 α (ḡ < 0).
S(x; α, τ ) = e−t eixt dt, (21) In Figure 3, two sub-regions can be recognized in ḡ <
π
0 0. The first one, “Zone A” which contains negative ḡ
values with downward trend ∂g/∂t < 0 that is faster as
x → ∞. The second sub-region is “Zone B”, it contains
Ziτ
1 α smaller negative ḡ values that follow an upward trend
T (x; α, τ ) = e−t eixt dt. (22) and ∂g/∂t > 0 displaying an increase behaviour when
π
0 x → 0. Considering these sub-regions, the truncation τ
in Eq. (24) will depend on x value as follows:
Now, the right hand of Eq. (20) can be evaluated in
the limit where x → ∞. First, notice that in the contour • For x → 0, The integration must avoid zone B. The
of integration in Figure 1√
the magnitude of r is bounded values of ḡ(x, t; α) in this zone lead to an exponen-
by the condition 0 < r < 2 τ and sin(θ) > 0 in the first tial growth due to an upward trend ∂g/∂t > 0,
quadrant, thus: consequently |I| ¯ 9 0.
0.05
0.4
1e
0.01
0.0
=x/ if x<xc
-0
1 -wavelength
5
01
0 0
4 1
=tc if x>xc
1e- 1e-0
06 7
(tc,x c) 1e- -1 -1
3.5 0.0 05 0 3 t 6 9
00 1e-07 x=2.9 =2.07
0.0
0. 1 1e-0 1
00
02
6
0.75
1 -truncation 1
0.0
5
1
0.05
0.4
1e-0 1 1
1
5 1e-06
3 0.00
01
-wavelength
1e-05 0 0
0.0 0.00 0.0001
025 1
-1 -1
2.5 0.0
1 0.001
0.0001
0.0 0.0025 0 3 t 6 9
x
5 0.001
0.0025 x=1.6 1
=1.14
0.01 0.0025 2 2
1 0.75
2 -truncation
0.4
0.01 0.01 1
0.05 1 1
-wavelength
0.05 0.05 0.05
0 0
1.5 0.4 0.75 -1 -1
0.4 1 2
0.4 0.4 0.75
0.4 0.75 1 2 4 -2 -2
0.75 1 2 0 3 t 6 9
1 0.7 4 10 x=0.5 1=0.36
5 0.75 1 40
1000 1000
1
2 4 10
1 40 200 500
2 4 10
0.5 40 200 1000 0 0
1 500
-truncation
1
10 0
4 40 200 100 -wavelength
0 -1000 -1000
0 1 2 3 4 5 6 7 8 9 0 3 t 6 9
t
ḡ
Figure 4. Contour plot of e that represents the instantaneous amplitude |I| ¯ for α = 1.4 and tolerance = 10−3 in Eq. (26)
and (27). The red line x was drawn by considering ḡ = 0 which represents the limit between the positive and negative values
of the attenuation factor ḡ. The truncation τ1 is applied following Eq. (30). Downward trend before τ1 and upward trend
¯ can be observed in the right part of the figure for x = 0.5 and x = 1.6. For x = 2.9 and x = 3.8 the truncation
after τ1 of |I|
is made after reaching a small value on the modulation I. ¯
the boundary between zones A and B for all range of α αtc 2 + tc α cos (πα/2) + ln() = 0,
(31)
values. xc = αtc .
The cut off τ1 obeys a linear equation and is obtained
from the following equations:
T(x; )-Laplace
-1 =1.60 =1.60
10 and the domain of integration is a closed interval [40, 41].
=1.90 =1.90
10-2 10-2
A. Inner Expansion
-3
10
The inner expansion is obtained making a substitution
-4 -4 of eixt by its Taylor series expansion given by Eq. (32)
10 10
0 10 0 10 in the integrand of the stable distribution I. After this
x x substitution, the integrals in Eq. (10) and (21) can be
analytically solved. The difference between these two
100 100 equations are the truncation on the interval of integra-
x c( ) tion.
S(x; )-Fourier
T(x; )-Laplace
n ln()
X zk τ2 (x, ) = − . (35)
ez ∼ as n → ∞. (32) x
k!
k=0
This truncation results from the equation eixτ2 = ,
Two different asymptotic expansions will be performed. where represents the tolerance that needs to be small to
The first one corresponds to the ‘inner expansion’. To get ensure a cut-off when negligible quantities of |I| and |In |
this solution the stable distribution function is evaluated are obtained. Consequently, the area under the curve of
by expanding eixt of Eq. (10) and (21) around x = 0. both functions are similar.
The second one refers to the ‘outer expansion’, which The convergence of In to I demands a large value of
is the asymptotic series expansion for x → ∞. When order n in Eq. (32), as it can be observed in Fig (6). This
α
x >> 1, the oscillations of the integrands in Eq. (10) occurs because of slow decay of e−t value for α < 1. This
and (21) are large. Consequently, there are important is the reason to evaluate the integral in the closed interval
cancellations due to factor eixt in the integral. Thus, [0,τ2 ], where the original integrand I and its Taylor series
we focus our integration in the region with the major approximation In are similar.
contribution in the integral, that is around t = 0. In The integrals in Eq. (33) can be solved without diffi-
α
consequence, the amplitude of the integral et is replaced culty. Then, the inner expansion si is given by the real
by its Taylor expansion around t = 0. To guarantee part of this solution,
the convergence of the series expansion, the improper
integrals are truncated. The truncation occurs because of
the sufficient conditions for Riemann integral existence. si (x; α, ) = Re(Si (x; α, )).
8
I- In
I- In
0 100 0 100
I
I
-1 -1
-2 10-20 -2 10-20
0 1 2 t 3 4 5 0 1 2 t 3 4 5 0 1 2
t 3 4 5 0 1 2 t 3 4 5
x=4.5 n=30 =0.75 x=4.5 n=30 =1.8
2 1020 2 1020
I I - In I I- In
1 In 1 In
2 2
I- In
I- In
0 100 0 100
I
I
-1 -1
-2 10-20 -2 10-20
0 1 2
t 3 4 5 0 1 2 t 3 4 5 0 1 2 t 3 4 5 0 1 2 t 3 4 5
x=4.5 n=50 =0.75 x=4.5 n=50 =1.8
2 1020 2 1020
I I - In I I- In
1 In 1 In
2 2
I- In
I- In
0 100 0 100
I
-1 I -1
-2 10-20 -2 10-20
0 1 2 t 3 4 5 0 1 2 3 4 5 0 1 2 t 3 4 5 0 1 2 3 4 5
t t
n
α α (ixt)k
Figure 6. Comparison of I = et eixt and In = et of Eq. (33) and (38), where In is obtained by replacing eixt in
P
k!
k=0
I by its Taylor expansion. The plots are for α = 0.75 (left) and α = 1.8 (right). For α ≤ 1, the truncation is required to en-
sure a cut-off when negligible quantities of |I| and |In | are obtained. The truncated error of the Taylor expansion is measured
by using the absolute value of the difference |I − In |. For α > 1, since the convergence is fast the truncation is unnecessary.
For these particular examples, the integrand I is evaluated at x = 4.5 for three cases of n = 20, 30, 50 with = 10−9 .
∞
1 X xk
k+1 α πk Here it is derived the inner expansion si for α > 1 from
si (x; α, ) = γ, τ2 (x; ) cos ,
πα k!
α 2 the non-truncated form of stable distribution function.
k=0
(36) This derivation is made by substituting eixt in Eq. (10)
where γ represents the incomplete gamma function [42], by its Taylor expansion in Eq. (32), then:
Z b Z∞ Z∞
γ(z, b) = xz−1 e−x dx. (37) 1 α 1
Si (x; α) = e−t eixt dt ∼ In dt as n → ∞.
0 π π
0 0
Due to a computation of the incomplete gamma func- (38)
tion γ, Eq. (36) was modified for numerical analysis in For α > 1, the convergence of integrand I and the in-
Matlab2 [43]. tegrand after the substitution In occurs faster than for
α < 1. This feature is observed in Figure (6), where
an acceptable convergence between I and In is obtained
with a small n value. Consequently, the integral is eval-
uated without truncation or taking the limit → 0 in
2 Note: Matlab defines the incomplete gamma function as γ ∗ Eq. (33).
Z b
Then, it is only considered the real part of the solution
1
∗
γ (b, z) = x z−1 −x
e dx of Eq. (38),
Γ(z) 0
3
sni (x; ,n)
I- Gn
I- Gn
0 100 0 100
I
I
-0.5 10-20 -0.5 10-20
-1 -1
10-40 10-40
0 20 40 60 0 20 40 60 0 2 4 6 0 2 4 6
t t t t
x=4.5 n=30 =0.75 x=4.5 n=30 =1.8
1 1040 1 1040
I I- Gn I I- Gn
0.5 Gn 1020 0.5 Gn 1020
3 3
I- Gn
I- Gn
0 100 0 100
I
I
-0.5 10-20 -0.5 10-20
-1 -1
10-40 10-40
0 20 40 60 0 20 40 60 0 2 4 6 0 2 4 6
t t t t
x=4.5 n=60 =0.75 x=4.5 n=60 =1.8
1 1040 1 1040
I I- Gn I I- Gn
0.5 Gn 1020 0.5 Gn 1020
3 3
I- Gn
I- Gn
0 100 0 100
I
I
-0.5 10-20 -0.5 10-20
-1 -1
10-40 10-40
0 20 40 60 0 20 40 60 0 2 4 6 0 2 4 6
t t t
t
α α k
Figure 9. Comparison of I = et eixt and Gn = (−tk! ) eixt dt of Eq. (40), where Gn is the Taylor expansion of I around
t = 0. The plots correspond to α = 0.75 (left) and α = 1.8 (right). Truncation of the integral is required for both cases. The
reason of that is to reach an accurate approximation between the original integrand I and the one after the series expansion
Gn . The error is measured by the absolute value of the difference |I − Gn |. For these particular examples, the integrand I is
evaluated at x = 4.5 for three cases of n = 20, 30, 50 with = 10−9 .
Consequenlty,
−ixτ ∞ k
n k kα+1 Z 3 () 1 X (−1) cos (παk)
1 X (−1) −1 so (x; α, ) = − ...
So (x; α, ) ∼ ukα e−u du. π k! kα + 1
π k! ix k=1
k=0 0 kα+1
(43) ...(−τ3 ()) 1 F1 (kα + 1, kα + 2, ixτ3 ()).
(45)
To solve the integral, the incomplete gamma function Figure 10 shows the calculation of Eq. 45 for α = 1.8.
of imaginary argument γ(v, iz) is used [42, 44]. The fol- In this figure is evident that the outer expansion so con-
lowing solution is presented by Barak as a special case of verges slowly. This occurs due to computation of the con-
confluent hypergeometric function [44], fluent hypergeometric function 1 F1 which demands con-
siderable computational time. The series that define the
Z iz function 1 F1 do not have a trivial structure, this creates
γ(v, iz) = tv−1 e−t dt numerical issues which makes the calculation computa-
0 tional inefficient [45]. The approximation in Figure 10
(44)
shows how the convergence demands a large value of or-
= (iz)v v −1 1 F1 (v, 1 + v, −iz), der n to obtain an accurate approximation at the tail.
The convergence resembles waves that slowly start to de-
where 1 F1 (v, 1 + v, −iz) represents the Confluent Hyper- crease from the tails to the peak of stable distribution.
geometric function. Then, comparing Eq. (43) and (44), The series until n = 30 does not show an acceptable
we obtain the following relation betwen the variables, approximation. Only an approximation on tails are ob-
v = kα + 1, z = −xτ3 and t = u. tained after n = 40. For x → 0, the series of so converges
Finally, the real part of the solution is: to a specific value different of the stable distribution func-
tion. The convergence to the stable distribution function
so (x; α, ) = Re(So (x; α, )). is observed only for x → ∞.
11
0.1 n=100 but still remains on the real axis. The integral above can
n=40
be solved using the incomplete gamma function defined
0 in Eq. (37). Then, the real part of the result is obtained,
-0.1
to (x; α, ) = Re(To (x; α, )).
(b)
-0.2
0 1 2 4 3
5 6 7 Consequently,
x
Figure 10. Outer expansion so for α = 1.80 with tolerance to (x; α, ) =
= 10−6 in Eq. (45). The subfigure (a) and (b) correspond ∞ k kα+1
to n = 10, 30, 40, 100 respectively. 1 X (−1) 1 παk
− sin γ (kα + 1, xτ1 (x, )).
π k! x 2
k=1
(49)
C. Outer expansion by Trans-Stable distribution
The determination of τ1 for α ≤ 1 and α > 1 is presented
in the following subsections.
Because of the slow convergence of the outer expansion
so and its wave-like behavior, an alternative approxima-
tion is obtained using the trans-stable function T (x; α). 1. For 0 < α ≤ 1
As it was previously explained in section III, the solu-
tions of trans-stable T (x; α) and stable S(x; α) functions For α ≤ 1, the cut-off τ1 in Eq. (49) is given by the
are identical for 0 < α ≤ 1 and similar for 1 < α < 2 af- following equation:
ter the xc value. Consequently, the improper integral
in Eq. (19) is used to calculate the series expansions for
1/α
0 < α ≤ 1 and the truncated trans-stable integral in its τ1 () = [− ln()] f or α ≤ 1. (50)
Laplace representation in Eq. (24) for 1 < α < 2. α
This outer expansion to is given by the analytical solu- This truncation is obtained from e−τ1 = , where the
tion of the trans-stable function after applying the Taylor tolerance represents a negligible instantaneous ampli-
α
series of e−(it) around t = 0 in the trans-stable integrand tude for the integrands in Eq. (46).
I¯ using Eq. (32) and (19). Then, the following equation
is shown:
2. For 1 < α < 2
τ1Z(x,) τ1Z(x,) The truncation τ1 in Eq. (49) for 1 < α < 2 was already
1 −(it)α −xt 1
To (x; α, ) = e e idt ∼ Kn dt, obtained in subsection III-2 and defined by Eq. (30) as:
π π
0 0
t () if x > xc ,
(46) τ1 (x, ) = c f or α > 1,
x/α if x < xc ,
where Kn is given by:
where tc and xc were defined by Eq.(31). As indicated
n k in subsection III-2, the value of τ1 is used to minimize
X (−(it)α ) the truncation error and at the same time to make the
Kn (x) = e−xt i. (47)
k! domain of integration as small as possible.
k=0
12
The outer expansion by the trans-stable function con- 100 (a) Stable Function s(x; )
7
verges to the original trans-stable function. Examples 5 Trans-stable Function t(x; )
3 Outer Expansion t no(x; ,n)
are shown in Figure 11 for α ≤ 1 and Figure 12 for
tno(x; ,n)
α > 1. Note that in both cases the truncation τ1 allows 1
tno(x; , ,n)
7
-2 3
10 (xc)
10-1
2 Figure 12. Outer expansion of the trans-stable for α = 1.80
as a result of applying Taylor expansion of the integrand
25 around t = 0 in Eq. (30) and (49). The subfigure (a) shows
32
50
that the non-truncated integral does not converge to the
(a) trans-stable function. The subfigure (b) corresponds to the
10-2 truncated integral with tolerance = 10−6 . The subfigure (b)
0 1 2 3 4 5 6 7 displays a fast convergence as a result of the truncation of
x the integral.
V. UNIFORM SOLUTION
100 Stable Function s(x; ) The uniform solution is presented as the combination
27
Trans-stable Function t(x; )
50 of the inner solution and the outer solution to construct
27 Outer Expansion t no(x; , ,n)
32
8
an approximation valid for all x ∈ [−∞, ∞]. To construct
25
tno(x; , ,n)
Normalized stable R∞ α
s(x; α) = 1
π
e−t cos(tx)dt
0
distribution (s)
Normalized trans-stable R∞ α
cos( πα
e−t 2 )
−xt
t(x; α) = 1
π
sin(tα sin( πα
2
))dt
0
distribution (t)
n
1 X xk
k+1 πk
sn
i (x; α, ) = γ , τ2 α cos n
πα k! α 2 1 X xk
Inner expansion (sn k+1 πk
i ) k=0 sn
i (x; α) = Γ cos
πα k! α 2
k=0
− ln()
τ2 =
x
n
1 X (−1)k cos (παk)
Outer expansion (sn
o) sn
o (x; α, ) = − (−τ3 )kα+1 1 F1 (kα + 1, kα + 2, ixτ3 )
π k! kα + 1
k=1
τ3 = [− ln()]1/α
n kα+1
1 X (−1)k 1
παk
Outer solution 3
(tn
o) tn
o (x; α, ) =− γ (kα + 1, xτ1 ) sin
π k! x 2
k=1
τ1 = (− ln())1/α tc if x > xc
τ1 =
x/α if x < xc
Z ∞ Z b
Complete and incomplete
Γ(z) = xz−1 e−x dx γ(z, b) = xz−1 e−x dx
0 0
gamma functions (Γ) & (γ)
contains a summary of inner and outer expansions pre- For the inner expansion si , the solution for α ≤ 1
viously obtained. In sub-section B the steps taken to corresponds to a truncated stable solution which al-
obtain su are explained. lows a faster convergence. For α > 1 the series is
obtained from the non-truncated stable solution. The
only difference between them is the use of the incom-
A. Summary of inner and outer expansions plete gamma function γ in the solution for α ≤ 1, where
Γ(z) = limb→∞ γ(z, b). Consequently, for both cases the
truncated series can provide a good approximation. How-
Table II contains the normalized stable and trans-
ever, in the case of α ≤ 1 we must take the limit as:
stable distribution and the summary of previous results
obtained from stable and trans-stable functions by ap-
plying Taylor expansions. The series refers to one inner s(x; α) = lim lim sni (x; α, ) f or x < ∞.
expansion si and two outer expansions so and to . →0 n→∞
t(x; α) = lim lim tno (x; α, ) f or x > 0. 2. Defining the inner and the outer solutions yin
→0 n→∞ and yout
Exchanging the order of the limits will affect the con-
vergence. The outer expansion that will be used is to , To obtain the uniform solution su , the asymptotic
because it displays a faster convergence and it does not matching method based on boundary layer theorem [46]
exhibit wavelike behavior. is applied. Consequenlty, the inner solution yin and the
outer solution yout must have a matching asymptotic be-
haviour. More precisely, the limit of the outer solution
B. Steps to obtain the uniform solution yout when x → 0 should correspond to the limit of the
inner solution yin when x → ∞. To obtain yin and yout
To obtain the uniform solution su the condition in solutions, the series expansions si and to are multiplied
Eq. (52) needs to be satisfied. The inner expansion si by an appropiate coefficients to meet the requirements of
and the outer expansion to have to be multiplied with matching asymptotic expansions, so the yin and yout are
an appropriate coefficient A(x) to obtain the asymptotic defined as follows:
solutions with a common matching value ym . These oper-
ations will allow us to obtain yout and yin . Consequently, yin (x; α, , µ) = (si (x) − ym ) (1 − A(x; µ)) + ym , (55)
Eq. (51) will be applied to obtain the closed-form solution
of the stable distribution function.
Below the steps are explained to obtain the location of yout (x; α, , µ) = (to (x) − ym )A(x; µ) + ym , (56)
the matching between the inner and the outer solutions
(xm , ym ), the coefficient A(x), and the uniform solution where the overlapping factor A(x) is defined as:
su .
1 x − xm
A(x; µ) = 1 + tanh . (57)
2 µ
1. Finding inner and outer limit (xm , ym )
The A(x; µ) is used to smooth si and to and provides
them with a symmetric overlap section around xm and
Considering si and to as good approximations to the gives yin and yout an asymptotic behaviour. The variable
stable distribution function, we must require that the µ determines the width of the overlap between yin and
inner and the outer expansions will be close enough be- yout .
fore matching them [48]. Consequently, the point where It is easy to see that Eq. (55) and Eq. (56) satisfy
the matching between si and to takes place is (xm , ym ) Eq. (52), where the limits of yout and yin converge to a
and it represents the location where the minimal vertical constant value ym .
distance between the inner si and the outer solution to
occurs.
The distance function between si and s is defined as δi 3. Defining the Uniform Solution su
and the distance function between to and s is δo . Con-
sequently, (xm , ym ) is the point where the Pythagorean
The inner solution yin Eq.(55) and the outer solution
addition of these distances is minimal Eq. (53):
yout in Eq.(56) were defined to fulfill the requirements
for matching asymptotic expansions. Then, Eq.(51) is
δi 2 (x; α, ) = (s(x; α) − si (x; α, )) ,
2 applied to obtain the uniform solution su ,
tno o (x; α, ) sni i (x; α, )
snui ,no (x; α, , µ) = + ....
2 2 2no
δo 2 (x; α, ) = (s(x; α) − to (x; α, )) ,
to (x; α, ) sni i (x; α, )
x − xm
....+ tanh − .
µ 2 2
δ 2 (x; α, ) = δo 2 (x; α, ) + δi 2 (x; α, ), (58)
15
4. Find the best su by choosing the most appropiate µ 100 Stable Function s
Inner Solution y in
(xm,ym) Outer Solution y out
The width of the overlap between yin and yout can be 10-1 Uniform Solution s u
s(x; )
optimized to obtain the closest solution su of the stable
distribution function. The most appropriate value of µ
needs to be obtained for each particular value of α. For 10-2
that, the least square method will be applied between the
original s(x; α) and the new closest solution su (x; α, , µ).
Applying Eq. (7) and (58) the following equation is ob- 10-3
tained: 0 1 2 3 4 5 6 7
x
Figure 14. Uniform solution su for α = 1.80 as a result of
N joining the inner solution yin with the outer solution yout .
The tolerance = 10−6 , µ = 0.4 and n = 8.
X 2
L(µ) = (su (xi ; α, , µ) − s(xi ; α)) , (59)
i=1
where the N value represents the length of the sample VI. CONCLUSIONS
used to minimize L.
In this paper we presented a uniform solution of the
The similarity between the exact solution of s(x; α) stable distribution. This solution converges to the sta-
and the uniform solution su (x; α, , µ) is observed in Fig- ble distribution function in the full range of x values
ure 13 and 14 for α = 0.75 and α = 1.80 respectively. −∞ < x < ∞. This condition makes our uniform so-
For α < 1, a good approximation between s(x; α) and lution more robust than previous analytical expressions
su (x; α, , µ) is obtained in the tails after mixing two dif- that were only applicable for extreme values x → 0 or
ferent orders. The order for the inner solution is ni = 6, x → ∞. Also, our uniform solution removes the neg-
which makes the solution concave upward. The order for ative values obtained in previous numerical solutions of
the outer solution is no = 17, which makes the solution the stable distribution function for all α values, which
concave downward. This combination of orders will en- makes this solution more reliable because a probability
sure a good matching asymptotic behaviour. For α > 1, density function must be always positive.
the uniform solution works well, and a good uniform so- The uniform solution is the result of an asymptotic
lution is obtained quickly with a lower order n = 6. matching between the inner and outer expansions. The
Lower orders can be used for both cases, where the inner expansion results from the Taylor series expansion
most important aspect to consider is the different con- of the characteristic function of the stable distribution
cavity between yin and yout for the matching asymptotic around x = 0. The outer expansion is obtained from
behaviour. The concavity of the inner and outer solution the Taylor expansion of the integrand of the trans-stable
is defined by the trigonometric element in each solution. function around t = 0. The convergence of these expan-
sions is guaranteed if the integrands are truncated, and
the speed of convergence depends on how is the trunca-
tion implemented.
100
(xm,ym)
Stable Function s
Inner Solution y in
For α ≤ 1, the uniform solution provides a good ap-
Outer Solution y out
proximation for all the range of x values. Also, the nu-
Uniform Solution s u merical integration of the trans-stable function consti-
tutes a second option which allows us to obtain a robust
s(x; )
[1] Bruce D Malamud, Gleb Morein, and Donald L Turcotte. ior. Science, 281(5384):1840–1842, 1998.
Forest fires: an example of self-organized critical behav-
16
[2] Bruce D Malamud. Tails of natural hazards. Physics [22] Rosario Nunzio Mantegna. Fast, accurate algorithm for
World, 17(8):25, 2004. numerical simulation of levy stable stochastic processes.
[3] Luciano Telesca, Gerardo Colangelo, Vincenzo Lapenna, Physical Review E, 49(5):4677, 1994.
and Maria Macchiato. On the scaling behavior of rain [23] John Nolan. Stable distributions: models for heavy-tailed
event sequence recorded in basilicata region (southern data. Birkhauser New York, 2003.
italy). Journal of Hydrology, 296(1):234–240, 2004. [24] William Feller. An introduction to probability theory and
[4] James A Roberts, Tjeerd W Boonstra, and Michael its applications, volume 2. Wiley, New York;Sydney;,
Breakspear. The heavy tail of the human brain. Cur- 1950.
rent Opinion in Neurobiology, 31:164–172, 2015. [25] Neil F. Johnson, Paul Jefferies, and Pak M. Hui. Finan-
[5] L Michiels Van Kessenich, Lucilla de Arcangelis, and cial market complexity. Oxford University Press, Oxford,
HJ Herrmann. Synaptic plasticity and neuronal refrac- 2003. ISBN 0191712108;9780191712104;.
tory time cause scaling behaviour of neuronal avalanches. [26] John P Nolan. Maximum likelihood estimation and diag-
Scientific Reports, 6:32071, 2016. nostics for stable distributions. Levy Processes: Theory
[6] Sidney I Resnick et al. Heavy tail modeling and teletraffic and Applications, pages 379–400, 2001.
data: special invited paper. The Annals of Statistics, 25 [27] John P Nolan. Multivariate elliptically contoured sta-
(5):1805–1869, 1997. ble distributions: theory and estimation. Computational
[7] Donald L Turcotte. Self-organized criticality. Reports on Statistics, 28(5):2067–2089, 2013.
Progress in Physics, 62(10):1377, 1999. [28] Yingjie Liang and Wen Chen. A survey on computing
[8] Thomas Maillart and Didier Sornette. Heavy-tailed dis- levy stable distributions and a new matlab toolbox. Sig-
tribution of cyber-risks. The European Physical Journal nal Processing, 93(1):242–251, 2013.
B-Condensed Matter and Complex Systems, 75(3):357– [29] John P Nolan. Modeling financial data with stable dis-
364, 2010. tributions. Handbook of Heavy Tailed Distributions in
[9] Bruce D Malamud, Donald L Turcotte, Fausto Guzzetti, Finance, Handbooks in Finance: Book, 1:105–130, 2003.
and Paola Reichenbach. Landslide inventories and their [30] Rosario N Mantegna and H Eugene Stanley. Stochastic
statistical properties. Earth Surface Processes and Land- process with ultraslow convergence to a gaussian: the
forms, 29(6):687–711, 2004. truncated levy flight. Physical Review Letters, 73(22):
[10] Paul L Anderson and Mark M Meerschaert. Modeling 2946, 1994.
river flows with heavy tails. Water Resources Research, [31] Ismo Koponen. Analytic approach to the problem of con-
34(9):2271–2280, 1998. vergence of truncated levy flights towards the gaussian
[11] Szymon Borak, Wolfgang Hardle, and Rafal Weron. Sta- stochastic process. Physical Review E, 52(1):1197, 1995.
ble distributions. In Statistical Tools for Finance and [32] Giovani L Vasconcelos. A guided walk down wall street:
Insurance, pages 21–44. Springer, 2005. an introduction to econophysics. Brazilian Journal of
[12] Ashadun Nobi, Seong Eun Maeng, Gyeong Gyun Ha, and Physics, 34(3B):1039–1065, 2004.
Jae Woo Lee. Effects of global financial crisis on network [33] Ofer Biham, Zhi-Feng Huang, Ofer Malcai, and Sorin
structure in a local stock market. Physica A: Statistical Solomon. Long-time fluctuations in a dynamical model
Mechanics and its Applications, 407:135–143, 2014. of stock market indices. Physical Review E, 64(2):026101,
[13] Didier Sornette and Daniel Zajdenweber. Economic re- 2001.
turns of research: the pareto law and its implications. [34] Jun-ichi Inoue and Naoya Sazuka. Crossover between
The European Physical Journal B-Condensed Matter and levy and gaussian regimes in first-passage processes.
Complex Systems, 8(4):653–664, 1999. Physical Review E, 76(2):021111, 2007.
[14] Boris V. Gnedenko and A. N. Kolmogorov. Limit dis- [35] Bwo-Nung Huang. Do asian stock market prices follow
tributions for sums of independent random variables. random walks? evidence from the variance ratio test.
Addison-Wesley Pub. Co, Cambridge, Mass, 1954. Applied Financial Economics, 5(4):251–256, 1995.
[15] Leo Breiman. Probability. Addison-Wesley Pub. Co, [36] Cheng F Lee, Gong-meng Chen, and Oliver M Rui. Stock
Reading, Mass, 1968. returns and volatility on china’s stock markets. Journal
[16] Benoit Mandelbrot. The variation of certain speculative of Financial Research, 24(4):523–543, 2001.
prices. The Journal of Business, 36(4):394–419, 1963. [37] Barry C. Arnold and Dean L. Isaacson. On normal char-
[17] Vladimir V. Uchaikin and Vladimir M. Zolotarev. Chance acterizations by the distribution of linear forms, assum-
and Stability: Stable Distributions and their Appli- ing finite variance. Stochastic Processes and their Ap-
cations. De Gruyter, Berlin ;Boston, 2011. ISBN plications, 7(2):227 – 230, 1978. ISSN 0304-4149. doi:
311093597X;9783110935974;. http://dx.doi.org/10.1016/0304-4149(78)90019-4.
[18] V. M. Zolotarev. One-dimensional stable distributions, [38] Michael Feldman. Hilbert transform in vibration anal-
volume 65. American Mathematical Society, Provin- ysis. Mechanical Systems and Signal Processing, 25(3):
dence, R.I, 1986. ISBN 9780821845196;0821845195;. 735–802, 2011.
[19] William Feller. An introduction to probability theory and [39] Walter Gander and Walter Gautschi. Adaptive quadra-
its applications, volume II. Wiley, New York;Sydney;, ture—revisited. BIT Numerical Mathematics, 40(1):84–
1950. 101, 2000.
[20] Carmen Fernandez and Mark FJ Steel. On bayesian mod- [40] Kaoru Yoneda. Riemann—lebesgue theorem. Analysis
eling of fat tails and skewness. Journal of the American Mathematica, 7(4):297–302, 1981.
Statistical Association, 93(441):359–371, 1998. [41] Robert G Bartle. Return to the riemann integral. The
[21] Elliott W Montroll and John T Bendler. On levy (or American Mathematical Monthly, 103(8):625–632, 1996.
stable) distributions and the williams-watts model of di- [42] Milton Abramowitz and Irene A. Stegun. Handbook of
electric relaxation. Journal of Statistical Physics, 34(1): mathematical functions with formulas, graphs, and math-
129–162, 1984. ematical tables. Dover Publications, New York, 1965.
17
[43] Armido R DiDonato and Alfred H Morris Jr. Computa- [46] Carl M. Bender and Steven A. Orszag. Advanced math-
tion of the incomplete gamma function ratios and their ematical methods for scientists and engineers. Springer,
inverse. ACM Transactions on Mathematical Software New York, 1999. ISBN 9780387989310;0387989315;.
(TOMS), 12(4):377–393, 1986. [47] Hans-Gorg Roos, Martin Stynes, and Lutz Tobiska. Ro-
[44] Richard Barakat. Evaluation of the incomplete gamma bust numerical methods for singularly perturbed differ-
function of imaginary argument by chebyshev polynomi- ential equations: convection-diffusion-reaction and flow
als. Mathematics of Computation, pages 7–11, 1961. problems, volume 24. Springer Science & Business Me-
[45] John William Pearson. Computation of hypergeometric dia, 2008.
functions. PhD thesis, University of Oxford, 2009. [48] Mark H Holmes. Introduction to perturbation methods,
volume 20. Springer Science Business Media, 2012.