GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH

THE DEGREE OF FREEDOM

STEPHEN H. LIHN
arXiv:2405.04693v1 [q-fin.ST] 7 May 2024

Abstract. A Wright function based framework is proposed to combine and extend sev-
eral distribution families. The α-stable distribution is generalized by adding the degree of
freedom parameter. The PDF of this two-sided super distribution family subsumes those of
the original α-stable, Student’s t distributions, as well as the exponential power distribution
and the modified Bessel function of the second kind. Its CDF leads to a fractional exten-
sion of the Gauss hypergeometric function. The degree of freedom makes possible for valid
variance, skewness, and kurtosis, just like Student’s t. The original α-stable distribution is
viewed as having one degree of freedom, that explains why it lacks most of the moments. A
skew-Gaussian kernel is derived from the characteristic function of the α-stable law, which
maximally preserves the law in the new framework. To facilitate such framework, the sta-
ble count distribution is generalized as the fractional extension of the generalized gamma
distribution. It provides rich subordination capabilities, one of which is the fractional χ
distribution that supplies the needed ’degree of freedom’ parameter. Hence, the ”new” α-
stable distribution is a ”ratio distribution” of the skew-Gaussian kernel and the fractional χ
distribution. Mathematically, it is a new form of higher transcendental function under the
Wright function family. Last, the new univariate symmetric distribution is extended to the
multivariate elliptical distribution successfully.

1. Introduction
The aim of this work is to propose a two-sided, super distribution family, called the gen-
eralized α-stable distribution (GAS), that subsumes three major distributions, plus one tran-
scendental function:
• the α-stable distribution,
• Student’s t distribution,
• the exponential power distribution, and
• the modified Bessel function of the second kind1.

E-mail address: [email protected], [email protected] .

Date: May 12, 2024. v1.0.0512. Dedicated to my mother on the Mother’s Day of 2024. Full Version .
2020 Mathematics Subject Classification. Primary: 33C05, 33C10, 33C15, 33E20, 60E07, 60E05, 60E10, 60G53,
62E15, 62H10, 65C10.
Key words and phrases. fractionally generalized distribution, the Wright function, fractional chi distribution,
stable count distribution, generalized gamma distribution, α-stable distribution, Student’s t distribution, ex-
ponential power distribution, multivariate distribution, Feller square root process, fractional hypergeometric
function.
Disclaimer: Mr. Lihn is a quant researcher at Atom Investors LP, Austin, TX 78746, USA. This mathematical
treatise is not related to the business of Atom Investors LP. I attest there is no conflict of interest.
1It is the characteristic function of Student’s t distribution
1
2 S. H. LIHN

The literature review is as follows.

In 1908, William S. Gosset published a paper on measuring the distribution on samples
of finite size under the pseudonym ”Student”[30]. He pioneered a distribution later called
Student’s t distribution or simply the t distribution, that plays a crucial role in statistical
inferences and hypothesis testing. The distribution has a major parameter called degrees of
freedom k, hence the notation tk , with the Cauchy distribution at k = 1, and the standard
normal distribution N = N (0, 1) at k → ∞.
Student’s t distribution can be constructed as the ratio of a standard normal variable
q and
2
the square root of a χ distribution, divided by degrees of freedom. That is, tk ∼ N / χ2k /k.
The χ2 distribution was first derived by F. R. Helmert in 1872 from his work on least squares
adjustment and error propagation[12]. Later
q in 1900, Karl Pearson defined the χ distribution
as the square root of a χ2 variable: χk ∼ χ2k .
This history is summarised in our first and central guiding equation:

√
(1) tk ∼ N / χk , where χk := χk / k.

In 1925, Paul Lévy published his seminal book on the α-stable distribution[13]. The dis-
tribution has a major parameter, among others, called the stability index α ∈ (0, 2]. In this
work, we call the extension to the α dimension fractional.
The α-stable distribution family also contains the Cauchy distribution at α = 1, and the
normal distribution at α = 2. Are the two commonalities between Student’s t and α-stable
coincident?
A main goal of this paper is to show that this is not a coincidence – There is a super distribu-
tion family that subsumes both distributions seamlessly. However, its construction requires
the fractional extension of the χ distribution: χk → χα,k , and a branch of mathematical
functions called higher transcendental function.
In 1933, E. M. Wright published his work on the asymptotic theory of partitions[32, 33] that
proposed a particular type of higher transcendental function, with the notation of Wλ,δ (z)
where λ, δ are two shape parameters. It was later collected into the Bateman manuscript in
1955, but tucked under the Chapter of the Mittag-Leffler function (Chapter 18.1 of Vol 3)[2].
In recent decades, this function is found to have an essential role in the theory of fractional
calculus and the α-stable distribution, hence is called the Wright function.
In 1952, William Feller pioneered a new stable parameterization (α, θ) in his work on the
Riesz-Feller fractional derivative[8], linking the fractional calculus and the α-stable distribu-
tion together[20, 26]. It is called the Feller parameterization in this work, in which θ is the
skewness parameter in the form of a trigonometric angle. Among many parameterizations
in the α-stable distribution (sometimes very confusing), this is the primary parameterization
used here (See Definition 3.7 of [23]). Therefore, the α-stable distribution takes the notation
of Lθα .
Later in Feller’s 1971 seminal book on the probability theory[9], the series representation
for the PDF of the α-stable distribution was given. When θ = ±α, the distribution becomes
one-sided (or called extremal). Its PDF Lα (x) (x > 0) is based on one of the most important
variants of the Wright function: Lα (x) = x−1 W−α,0 (−x−α ) (See Table 3).
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 3

In a nutshell, χα,k is a Wright function extension of χk . We would properly standardize

χα,k to χα,k , then extend tk from (1) to

(2) Lα,k ∼ N / χα,k

after which the skewness θ is added, and we obtain the enlarged distribution family Lθα,k . We
make sure when k = 1, it conforms to the α-stable distribution: Lθα,1 = Lθα .
We can also look at it from a different angle. We take the α-stable distribution Lθα , add k
from tk to it, and form Lθα,k . We would get to the same place.

Since 1990’s, Mainardi et al. have been exploring extensively the relation between the
Wright function, the fractional calculus, and the α-stable distribution[18, 21, 19]. I consider
this work as a continuation of such exploration.
In 1962, E. W. Stacy proposed a super distribution family called the generalized gamma
distribution (GG). Many one-sided distributions commonly used for parametric models are
part of it, such as the exponential distribution and the exponential power distribution, the
half-normal, Weibull, Rayleigh, gamma, χ, and χ2 distributions (See Table 1).
The Wright function and GG have a deep connection. The large-x asymptotic of a PDF
based on the Wright-function becomes a GG-style PDF, apart from a constant. We can
jokingly say they are joined at the tail.
One contribution of this work is to have GG fractionally extended, called the generalized
stable count distribution (GSC), which serves as the parent of the fractional χ distribution.
Obviously, GSC is the largest one-sided distribution family in this scope. The following is a
brief review of my past effort.
In 2017, I discussed the α-stable law with Professor Mulvey at Princeton University. Later
that year, I derived the stable count distribution Nα (x) as the conjugate prior of the one-sided
stable distribution Lα (x)[16, 17, 11, 24]. Nα (x) measures the distribution of the length N
of a Lévy sum N
P
i=1 Xi where Xi ’s are i.i.d. stable random variables. It was found to be
−1
expressed elegantly in the Wright function: Nα (x) = Γ( α1 + 1) W−α,0 (−xα ).
In 2022, I began to organize the subordination structure of the GG-related distributions via
the stable count distribution, which is documented in Appendix B.1. The GG-style pattern
in the PDF of GSC, xd−1 W−α,0 (−xp ), was discovered (See Table 1).
It turned out such pattern is not a coincidence. GSC becomes GG when α → 0. GSC is
proven to be the fractional extension of GG, solidifying its theoretical position in the statistical
distribution theory. The proof is based on the nice properties of Mainardi’s auxiliary Wright
functions[18].
In 2023, the subordination study was extended to the known symmetric two-sided dis-
tributions, as shown in Appendix B.2. The main structure in the PDF of the fractional χ
distribution, χα,k (x) ∝ xk−2 W− α2 ,0 (−xα ), became clear, when the continuous Gaussian mix-
ture (2) is focused. This is elaborated in Sections 2 and 7. It combines the α and k parameters
into a single distribution. This is the cornerstone that generalizes the α-stable, Student’s t,
and exponential power distributions. All three are combined to the expanded notation Lθα,k .
These are the main contributions of this paper.
4 S. H. LIHN

The symmetric case (θ = 0) is very solid and elegant: Lα,k = L0α,k . Most of this paper aims
to explain this case in great detail. The skewness generalization is experimental. It is a much
harder task that requires future work. In a tongue-twisting manner, it goes like:
• Lθα,1 : the original α-stable distribution,
• L1,k (k > 0): Student’s t distribution,
• Lα,−1 : the exponential power distribution.
I also proposed a framework for the generation of the random variables, from the GSC
to χα,k , to Lα,k . It is based on Feller’s square-root process in 1951[7]. Hence, this new
distributional system is grounded on an elegant, well known stochastic process. Amazingly,
some of the solutions of the mean-reverting force turn out to be simple polynomials.
Lastly, the univariate Gaussian mixture is extended nicely to the multivariate distribution.
Two kinds are proposed: The first kind is a straightforward copy from the elliptical distribu-
tion. The second kind is an invention that allows each dimension to have its own shape. The
bivariate models from both kinds are applied to the VIX/SPX2 data set with success. This
is discussed in Section 4.
Figure 1 illustrates the stack of distributions involved in this work. Seven new distributions
are invented, and a handful of existing, classic distributions are subsumed into them.

The benefits from multiple degrees of freedom, aka k > 1, are two-fold. First, the original
α-stable distribution is viewed as having one degree of freedom. That explains why it lacks
most of the moments. k > 1 allows the distribution to have finite moments with increasing k:
The n-th moment should exist when k > n. This addresses, to some extent, the decades-old
issue of non-existing moments in the stable law.
Secondly, for the generalized exponential power distribution, it allows the slope of the PDF
to be continuous everywhere. This deserves some explanations.
α
In 2017, I incorporated the exponential power distribution, Eα (x) = e−|x| /2Γ α1 + 1 , into

the Hidden Markov Model (HMM) for regime identification[15]. It is especially important for
the crash regime, such that the model can capture the leptokurtic nature of the financial time
series.
A main deficiency of its high-kurtosis distribution is that, when α ≤ 1, the slope of the
PDF is discontinuous at x → 0± . A simple example is e−|x| . This appears to violate the
intuition that the low-volatility returns should distribute like normal. This deficiency is also
attributed as having one degree of freedom. It is addressed by the addition of the k dimension:
Eα → Eα,k . When k > 1, the slope of the PDF is continuous everywhere.

To showcase the new GAS distribution, it is used to fit the daily log returns of SPX from
1927 to 2020 in Figure 2. This is a very difficult data set to fit for any existing two-sided
distribution. In the pyro project[3], it is used as a failed example that this data set can not
be fitted by a single marginal distribution3.
Nonetheless, GAS produced a good fit. It fits the shape of the histogram reasonably good in
both the density and log-density scales, to the 10−3 level, which is about 6 standard deviations.
It produced the desired excess kurtosis (∼19) and standardized peak density (∼0.72), and
2VIX is the CBOE volatility index; SPX is the S&P 500 index.
3See https://pyro.ai/examples/stable.html
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 5

Generalized alpha-stable distribution Multivariate distributions – Elliptical, Adaptive

Alpha-stable distribution Generalized symmetric alpha-stable distribution

(via Inverse FCM)

Ratio
Symmetric alpha-stable distribution Generalized Feller
Square-Root Process
Generalized exponential power distribution

Student’s t distribution

Exponential power distribution

Fractional chi-mean distribution

The Wright
Legend:
Chi distribution
function Subsumed to
Generalized stable count distribution
CF Characteristic function

M-Wright functions Product or ratio distribution

Generalized gamma distribution
Stable vol distribution distribution Known distribution

Stable count distribution distribution New distribution

Figure 1. The hierarchy of the distributions involved in this work. The

Wright function plays a central role. The green boxes are the contributions
from this work.

added a small skewness. Most interestingly, it provides a set of interpretable parameters that
we are familiar with in the Student’s t and α-stable context: α = 0.813, k = 3.292 for the
shape, and θ = 0.08 for the skewness. More discussion is in Section 3.4.

2. Framework of Continuous Gaussian Mixture

The bulk of this paper is focused on the construction of a symmetric two-sided distribution
in the form of a continuous Gaussian mixture. Both the ratio and product distribution
methods are used. In the case of the symmetric α-stable distribution (SaS)[5], the exponential
power distribution comes from its characteristic function (CF). We would like to present a
unified framework and familiarize the reader with the notations, which would be otherwise
subtle and confusing.
Assume the PDF of a two-sided symmetric distribution is L(x) where x ∈ R. It has zero
mean, E(X|L) = 0. Assume the PDF of a one-sided distribution is χ(x) (x > 0) such that
6 S. H. LIHN

Figure 2. Empirical distribution of S&P 500 daily log returns. Data is stan-
dardized and centered for demonstration purpose. The return distribution is
normalized to one standard deviation with zero mean. Fit by GSaS (orange
line) and GAS (red line). The fits are reasonably good up to 6 standard devi-
ations.

Z ∞
(3) L(x) := s ds N (xs) χ(s)
0
This is nothing but the definition of a ratio distribution with a standard normal variable N .
This is the first form of the Gaussian mixture: L ∼ N /χ.
It also has the equivalent expression in terms of a product distribution by way of the inverse
distribution χ† such that L ∼ N χ† . This is the second form of the Gaussian mixture.
χ† is closer to our typical understanding of the marginal distribution of a volatility process.
For example, the Brownian motion process dXt = σt dWt , measured in a particular time
interval ∆t, we have ∆Xt ∼ L and σt ∼ χ† .
However, χ in the first form is more natural in the expression of the α-stable distribution.
So we are more inclined to use the ratio distribution. The reader should keep this subtlety in
mind.
Lemma 2.1. (Inverse distribution) The inverse distribution is defined as[10]
 
† −2 1
(4) χ (s) := s χ
s
such that
Z ∞ Z ∞
ds  x  †
(5) s ds N (xs) χ(s) = N χ (s) (x ∈ R)
0 s s
Z ∞ Z0 ∞
ds  x 
(6) s ds N (xs) χ† (s) = N χ(s)
0 0 s s
The proof is straightforward by a change of variable t = 1/s. You can move between LHS
and RHS easily.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 7

We use the notation CF{g}(t) = E(eitX |g) to represent the characteristic function transform
√ the PDF g(x). Note that N has a special property that its CF is still itself: CF{N }(t) =
of
2π N (t).

Lemma 2.2. (Characteristic

R∞ function transform of L) Let ϕ(t) be the CF of L such that
ϕ(t) := CF{L}(t) = −∞ dx exp(itx) L(x). (3) is transformed to

√ ∞
Z  
t
(7) ϕ(t) = 2π ds N χ(s) (t ∈ R)
0 s

This allows us to define a new distribution pair: Lϕ and χ†ϕ , in terms of a product distri-
bution such that
Z ∞
ds  x  †
(8) Lϕ (x) := N χϕ (s) (x ∈ R)
0 s s
s χ(s)
(9) χ†ϕ (s) :=
E(X|χ)

where E(X|χ) is the first moment of χ. Here χ†ϕ is the inverse distribution of χϕ , which can
be reverse-engineered according to (4),

s−3
 
1
(10) χϕ (s) := χ
E(X|χ) s

We are in an interesting place: We start with an one-sided distribution χ, we derive two

variants from it: χϕ and χ†ϕ . We also obtain two two-sided distributions: L and Lϕ .
We shall call χϕ the characteristic distribution of χ since it facilitates the following parallel
relation:

L ∼ N /χ
Lϕ ∼ N /χϕ

χ symbolizes the fractional χ distribution we are about to introduce. The ϕ suffix will be
replaced with the negation (sign change) of the degree of freedom.

2.1. List of Symbols. The next tables describe the notations and symbols that will be used
in this paper. Note that GSC, FCM, GAS, GSaS are new notations. Others are known.
The PDF of Simple Distribution
8 S. H. LIHN

Γ(x; s) The gamma distribution (used only as a notation)

χk (x), χ2k (x) The chi and chi-square distributions
L(x) The two-sided Laplace distribution
N (x), N (x; µ, σ 2 ) The normal distribution, N (x) ∼ N (x; 0, 1)
α
The two-sided exponential power distribution (e−|x| /2Γ α1 + 1 )


Eα (x)
Nα (ν) The stable count distribution (SC)
Vα (x) The stable vol distribution (SV)
Lα (x) The one-sided stable distribution, aka Lαα (x)
Pα (x) The symmetric α-stable distribution, aka L0α (x) (SaS)
tk (x) Student’s t distribution
Wb(x; k) Weibull distribution
IG(x; k) The inverse gamma distribution (See scipy.stats, but use a = k)
IWb(x; k) The inverse Weibull distribution (See scipy.stats, but use c = k)

The PDF of Complex Distribution

fGG (x; a, d, p) The generalized gamma distribution (GG)
GenGamma(x; s, c) The generalized gamma distribution too (c = p, sc = d)
θ
Lα (x) The α-stable distribution
Nα (x; σ, d, p) The generalized stable count distribution (GSC, new)
θ
Lα,k (x) The generalized α-stable distribution (GAS, new)
Lα,k (x) The generalized symmetric α-stable distribution (GSaS, new)
χα,k (x) The fractional χ distribution (new)
χα,k (x) The fractional χ-mean distribution (FCM, new)
Eα,k (x) The generalized exponential power distribution (GExpPow, new)

Other Symbols
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 9

α The shape parameter associated with the Lev́y stability index

C The normalization constant for Nα (x; σ, d, p)
d The degree of freedom parameter in Nα (x; σ, d, p)
θ The skewness parameter in Lθα,k (x) and gαθ (x, s)
α
e−z The one-sided stretched exponential function
Eα (z) The Mittag-Leffler function
k The degree of freedom parameter in Lθα,k (x), tk (x), χα,k (x)
E(X n |Dist) The n-th moment of the distribution ”Dist”
mn The n-th moment in a local context
p The shape parameter in Nα (x; σ, d, p), unless mentioned otherwise
Γ(z) The gamma function
Γ(s, z) The upper incomplete gamma function
γ(s, z) The lower incomplete gamma function
gαθ (x, s) The skew-Gaussian kernel
Wλ,δ " (z) # The Wright function (two-parameter)
a, b
W (z) The four-parameter Wright function
λ, µ

3. Main Results
We will go over the main results in this section, and leave the explanations and proofs for
later sections. The new symmetric α-stable distribution Lα,k is constructed in three steps.
Then its CF is used to extend the exponential power distribution. The last part is to add the
skewness θ experimentally to the new α-stable distribution Lθα,k .
3.1. Recap of the Wright Function. The mathematical foundation of this work is the
Wright function[32, 33] as shown in the center of Figure 1. Its series representation is

∞
X zn
(11) Wλ,δ (z) := (λ ≥ −1, z ∈ C)
n! Γ(λn + δ)
n=0
Four (λ, δ) pairs are used extensively in this work. The first group of the two are Fα (z) :=
W−α,0 (−z) and Mα (z) := W−α,1−α (−z) = Fα (z)/(αz), where α ∈ [0, 1]. They are used in
various definitions and proofs, such as in (14), (18), (114).
In particular, Mα (z) is called the M-Wright function or simply the Mainardi function (See
Appendix C.1)[18, 21, 19]. Conceptually, the fractional extension of a classic, exponential-
based distribution hinges on two important properties: M0 (z) = exp(−z) and M 1 (z) =
2
√1 exp(−z 2 /4), that lead to (17).
π
The second group of the two are W−α,−1 (−z) and −W−α,1−2α (−z). Their usefulness is
discovered by me for the generation of random variables, such as in (89) and (119). They
are associated with the derivatives of Fα (z) and Mα (z). In some cases, they lead to beautiful
polynomial solutions. Section 11 is dedicated to this subject.
10 S. H. LIHN

Let’s begin with the 2017 discovery of the stable count (SC) distribution Nα (x)[16]. Its
PDF was first formulated as the conjugate prior of the one-sided stable distribution Lα (x).
Soon afterwards, it was linked to the Wright function:
 
1 1 1 1
(12) Nα (x) := Lα = W−α,0 (−xα ) (0 < α ≤ 1)
Γ( α1 + 1) x x Γ( α1 + 1)
In 2020, a variant of SC was proposed, called the stable vol (SV) distribution Vα (x)[17],
that works better with a normal variable:
√ √
2π Γ( α2 + 1) 2 2π  √ α
(13) Vα (x) := N α (2x ) = W − 2 ,0 −( 2x)
α (0 < α ≤ 2)
Γ( α1 + 1) 2
Γ( α1 + 1)
As the readers can observe, Fα (z), Mα (z), Nα (x), Vα (x), Lα (x) all point to a GG-style
pattern that subsumes a large number of one-sided distributions. The pattern is as follows.
3.2. New One-Sided Distributions.
Definition 3.1 (Generalized stable count distribution (GSC)). GSC is a four-parameter one-
sided distribution family, whose PDF is defined as

 x d−1   x p 
(14) Nα (x; σ, d, p) := C W−α,0 − (x ≥ 0)
σ σ
where α ∈ [0, 1] controls the shape of the Wright function; σ is the scale parameter; p is also
the shape parameter controlling the tail behavior (p ̸= 0, dp ≥ 0); d is the degree of freedom
parameter. When α → 1, the PDF becomes a Dirac delta function: δ(x − σ) assuming σ is
finite. When d ≥ 1, all the moments of GSC exist and have closed forms. See Section 5 for
more detail.
The normalization constant C will be derived in Section 5.2 and is shown below:
 αd
|p| Γ( p )

σ Γ( d ) , for α ̸= 0, d ̸= 0.
(15) C= p
|p|
, for α ̸= 0, d = 0.

σα

The original stable count distribution (12) is simply Nα (x) = Nα (x; σ = 1, d = 1, p = α).
See Section 5.7.
It is important to note that d and p are allowed to be negative, as long as dp ≥ 0. Several
examples are found in Table 1 and Table 3. This feature is used to define the characteristic
distribution of FCM below.
Since the Wright function extends an exponential function to the fractional space, GSC
is the fractional extension of the generalized gamma (GG) distribution[29], whose PDF is
defined as:
|p|  x d−1 −(x/a)p
(16) fGG (x; a, d, p) = e .
aΓ( dp ) a
The parallel use of parameters is obvious, except that a in GG is replaced with σ in GSC to
avoid confusion with α.
GG is subsumed to GSC in two ways:
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 11

(
 σ, d = d − p, p)
N0 (x;  , at α = 0.
(17) fGG (x; σ, d, p) := σ p p
N 1 x; σ = 22/p
,d =d− 2, p = 2 , at α = 21 .
2

The first line is treated as the definition of GSC at α = 0. See Sections 5.3 and 5.4 for more
detail.
Although the first line is more obvious, it is the second line that leads to the fractional
extension of the χ distribution.
The readers are reminded that W−α,0 (−x) used here is fully supported by existing soft-
ware packages. This is due to the fact that it can be converted to Lα (x) by Lα (x) =
x−1 W−α,0 (−x−α ) from (12). And Lα (x) can be computed via scipy.stats.levy stable us-
ing 1-Parameterization with beta=1, scale= cos(απ/2)1/α for 0 < α < 1. (L1 (x) = δ(x − 1)
can’t be computed.)4 It might seem somewhat peculiar that we can use the existing imple-
mentation of Lα (x) to develop all the new distributions for proof of concept.

When working with the stable law, α in GSC may become α/2 , which produces the more
recognizable range of α ∈ [0, 2] as in the next definition.
Let χk be the χ distribution √ of k degrees of freedom, then Student’s t distribution tk can
√
be constructed by tk ∼ N /(χk / k)[28]. We prefer to standardize it by √ dividing out the k
term. So we use a new convention of a bar over χ such that χk := χk / k.
Definition 3.2 (Fractional χ-mean distribution (FCM)). FCM χα,k is the fractional extension
of χk . It is the main workhorse in this paper. It is a two-parameter sub-family of GSC,
whose PDF for k > 0 is:

(18) χα,k (x) := N α2 (x; σ = σα,k , d = k − 1, p = α) (x ≥ 0, k > 0)

|k|1/2−1/α
where σα,k = √
2
k−1
k−1 
x α
   
αΓ( 2 ) 1 k−2
= x W α2 ,0 −
Γ( k−1
α )
σα,k σα,k
and α ∈ [0, 2] matches exactly Lévy’s stability index in the stable law. See Lemma 3.3 when
k = 1.
The characteristic FCM (χϕ in Section 2) is defined in the negative k space, whose PDF is:

(19) χα,−k (x) := N α2 (x; σ = (σα,k )−1 , d = −k, p = −α) (x ≥ 0, k > 0)

This is used to subsume the exponential power distribution to GSaS below. Negative k is not
in the scope until (32).
To use a single expression for both positive and negative k’s, let h(k) = (1 + sgn(k))/2 be
the Heaviside (step) function. The PDF can be consolidated to:
(20) χα,k (x) = N α2 (x; σ = (σα,k )sgn(k) , d = k − h(k), p = sgn(k) α) (x ≥ 0, k ∈ R̸=0 )
4See Chapter 1 of [23] for more detail on different parameterizations. We wouldn’t go into the issue of stable
parameterizations.
12 S. H. LIHN

This looks cumbersome, but it can save some duplication in a few places.

Lemma 3.1. (FCM Reflection Formula) Assume k > 0, FCM has the reflection formula that
resembles (10):
 
1 1
(21) χα,−k (x) = 3 χα,k
x E(X|χα,k ) x
Γ( k−1 k
2 ) Γ( α )
where E(X|χα,k ) = σα,k (See (63))
Γ( k−1 k
α ) Γ( 2 )
It is equivalent to the following relation in terms of the moments:
E(X −n+1 |χα,k )
(22) E(X n |χα,−k ) = , k > 0.
E(X|χα,k )
In particular, when n = 1, E(X|χα,−k ) = 1/E(X|χα,k ). The first moment is reciprocal.

When k is positive, k’s meaning is exactly the same as that of Student’s t, since χ1,k =
√
χk / k, which can be proven easily from (17). On the other hand, χα,1 is called fractional
chi-1. It is used to construct the SaS distribution: Pα := L0α ∼ N /χα,1 . This important point
will be elaborated in Lemma 3.3.
The parameter space outside (α = 1 | k = 1) is our innovation - a main contribution of this
paper. See Section 7 for more detail.
Lemma 3.2. (FCM Asymptotics) As k → ±∞, FCM becomes a delta function at its asymp-
totic mean. For k ≫ 1,
(23) χα,k (x) ∼ δ(x − mα ), where mα = α−1/α .
On the other hand, χα,−k (x) ∼ δ(x − 1/mα ).
This property ensures that the two-sided distributions constructed from FCM become a
normal distribution at k → ±∞. Hence, this lemma is associated with the Central Limit
Theorem. √
Note that m1 = 1 is preserved for Student’s t; m2 = 1/ 2 is preserved for α-stable.
The speed of approaching a delta function is determined by how fast the variance decreases:
E(X 2 |χα,k ) − E(X|χα,k )2 → 0. It is faster for larger α, and can be very slow for small α below
1.

Definition 3.3. (Inverse FCM) The inverse distribution of FCM can be constructed by
”inverting σ and negating d and p” so to speak. The results are simple due to the FCM
reflection rule:
x χα,−k (x)
(24) χ†α,k (x) := = N α2 (x; σ = (σα,k )−1 , d = −(k − 1), p = −α) (k > 0)
E(X|χα,−k )
x χα,k (x)
(25) χ†α,−k (x) := = N α2 (x; σ = σα,k , d = k, p = α)
E(X|χα,k )
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 13

Note that (25) is similar to (18) except d is incremented by one. And (24) is similar to (19)
except d is incremented by one.

We are jumping ahead to emphasize that the inverse FCM allows the generalized exponen-
tial power distribution Eα,k in (30) defined as N / χ†α,−k to be mapped to the negative k space
of GSaS as N / χ†α,−k = N χα,−k = Lα,−k .
As a validation, χ†α,−1 (x) is equal to the stable vol distribution Vα (x) in (13), as expected
by its original design (explained in Section 5.7).
In Figure 3, we plot the FCM distribution that came out of the S&P 500’s GSaS fit, which
is a distribution of the inverse of volatility.

Figure 3. The FCM distribution from the S&P 500’s GSaS fit. Its scale
has been standardized because the variance of the log-return distribution was
normalized to one. This is a distribution of the inverse of volatility. The
smaller x the more volatile the market was.
All aspects of FCM are well defined up to this point. Next, we are going to take apart the
α-stable distribution.
Definition 3.4 (The skew-Gaussian kernel). The skew-Gaussian kernel gαθ (x, s) is derived
from the α-stable law.
The minus-log θ
iθπ

αof the stable CF is defined in Feller’s parameterization[9] as ψα (ζ) :=
exp sgn(ζ) 2 |ζ| , where θ is the skewness parameter confined by the so-called Feller-
Takayasu diamond : θ ≤ min{α, 2 −R α} and α ∈ (0, 2][20, 19].
1 ∞ −ixζ . We want it to be decomposed
The stable PDF is Lθα (x) := 2π θ
−∞ dζ exp(−ψα (ζ)) e
∞
according to Lθα (x) = 0 s ds gαθ (x, s) χα,1 (s). The solution of gαθ (x, s) is
R

Z ∞  
1 α x 2
(26) gαθ (x, s) = dt cos τ (st) + st e−t /2 (x ∈ R, s ≥ 0)
qπ 0 q
where q = cos(θπ/2)1/α , τ = tan(θπ/2), θ ̸= 1.
14 S. H. LIHN

Feller’s parameterization provides a clean environment to dissect the subordination struc-

ture in the stable CF. See Section 6 for more detail.
In general, gαθR(x, s) is not a distribution, since some parts of the function can be negative.
∞
But it satisfies −∞ gαθ (x, s)dx = 1/s. One exception is gα0 (x, s) = N (xs). Also note that
√
gαθ (x, 0) = 1/q 2π, independent of x.
Most of the difficulty in computing the α-stable PDF arises from integrating some kind of
highly oscillatory functions. See Section 3 of [23] for a general discussion.
In our approach, that difficulty is concentrated in gαθ (x, s) for a non-zero θ. The range of
2
the integral is confined by e−t /2 , which is nice. But when the oscillation frequency is high,
it gets harder to integrate with the ordinary quad utility in the scipy package. We switch
to the quadosc function in the mpmath package[22] with more success in accuracy, but some
performance is sacrificed.
Lemma 3.3. Fractional Chi-1. When the skewness is absent: θ = 0, we get q = 1, τ = 0,
there is no ”complicated oscillation” from the cosine function. gα0 (x, s) = N (xs) greatly
simplifies the matter. This leads to the SaS distribution as Pα := L0α ∼ N /χα,1 . That is, all
the shape information in SaS is determined by χα,1 .
The χα,1 term is called fractional chi-1, which is the fractional version of the χ distribution
of one degree of freedom from the absolute of a random variable defined as such. This is a
new concept, used as a building block for χα,k .
Here we briefly explain how to go from χα,1 to χα,k . See Section 7.3 in full detail.
χα,1 is N α2 (x; σ = √12 , d = 0, p = α), whose PDF is
 √ α
(27) χα,1 (x) = 2x−1 W− α2 ,0 −( 2x)

χα,1 has some peculiar behaviors. For α < 1, χα,1 (x) diverges as xα−1 for small x. And
lim χα,1 (x) ≈ α/(ex) (x > α approximately)
α→0
2
χ1,1 (x) = 2/π e−x /2
p
√
χ2,1 (x) = δ(x − 1/ 2)
The first line implies lim Pα (x) ≈ χα,1 (x)/2 ≈ α/(2ex), for not too small x, e.g. x > α/2.
α→0
The xk−2 Term. In the k > 0 case, the PDF χα,k (x) comes from χα,1 (x) multiplied by
the surface area of the (k − 1)-dimensional sphere, which is ∝ xk−1 . Note that χα,1 (x) ∝

α/2  
α/2 
x−1 W− α2 ,0 − 2x2 . Hence, χα,k (x) ∝ xk−2 W− α2 ,0 − 2x2 . This is the origin of the
crucial xk−2 term.
Once χα,k is defined, it is standardized to FCM such that χα,k := χα,k /σα,k . FCM is used
to construct all the subsequent two-sided distributions.
3.3. New Two-Sided Distributions. We focus on the symmetric case where θ = 0 first,
and defer the skew case θ ̸= 0 to the latter.
When θ = 0, we have gα0 (x, s) = N (xs), independent of α. The new generalized symmetric
distribution is solely based on FCM that subsumes the SaS, Student’s t, and exponential
power distributions.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 15

Definition 3.5. (Generalized symmetric α-stable distribution (GSaS)) GSaS is elegantly

constructed as a ratio distribution: Lα,k ∼ N / χα,k . Hence, its PDF is
Z ∞
(28) Lα,k (x) := s ds N (xs) χα,k (s) (x ∈ R, α ∈ (0, 2], k ∈ R̸=0 )
0
It follows naturally that Lα,1 = Pα for SaS, and L1,k = tk for Student’s t when k > 0. On the
other hand, Lα,−1 = Eα for the exponential power.
Lemma 3.4. (GSaS Version of Central Limit Theorem) When |k| ≫ 1, it is straightforward
from Lemma 3.2 that Lα,k becomes a normal distribution. This statement is similar to the
Central Limit Theorem (CLM). The variance asymptotics of Lα,k is (α2/α )sgn(k) at α → 2 or
|k| → ∞. This is consistent with both α-stable at α = 2 and Student’s t at α = 1 by design.

Its CF becomes a product distribution, according to (7), such as

√ ∞
Z  
ζ
(29) ϕα,k (ζ) := CF{Lα,k }(ζ) = 2π ds N χα,k (s) (ζ ∈ R)
0 s
For k > 0, it is an expanded form of the inverse lambda decomposition (51). It describes,
in the θ = 0 case, how the α-stable law is modified from exp(−|ζ|α ) at k = 1 to a generalized
form of k > 1.
Since Lα,k (x) with k > 0 covers Student’s t as well, mathematically speaking, ϕα,k (ζ)
forms a high transcendental function that subsumes both the stretched exponential function
and the modified Bessel function of the second kind (due to CF{tk }(ζ)). See Section 8.6 for
more detail.

Definition 3.6 (Generalized exponential power distribution (GEP)). According to Lemma

2.2, ϕα,k (ζ) can be treated as a distribution. In this case, its k = 1 base is the exponential
α
power distribution whose PDF is Eα (x) = e−|x| /2Γ α1 + 1 [14, 25]. We take it to the next


level:
Z ∞
1 x
(30) Eα,k (x) := ds N χα,k (s) (x ∈ R, k > 0)
E(X|χα,k ) 0 s
Z ∞
ds  x  †
(31) = N χα,−k (s)
0 s s
where the normalization constant E(X|χα,k ) is FCM’s first moment in (63). And Eα,1 = Eα .
Furthermore, according to Lemma 2.1, (31) is inverted to a ratio distribution form:
Z ∞
(32) Eα,k (x) = s ds N (xs) χα,−k (s)
0
= Lα,−k (x)
Therefore, GSaS subsumes GEP beautifully when the negative k domain is defined properly
in FCM.
The k dimension mitigates a flaw in Eα (x): when α ≤ 1, the slope of the PDF is discontinu-
d d d
ous at x = 0. That is, limx→0− dx Eα (x) ̸= limx→0+ dx Eα (x). Now as long as k > 1, dx Eα,k (x)
is continuous everywhere. See Section 9 for more detail.
16 S. H. LIHN

Since N (xs) in (28) doesn’t take either α or k. All the shape information comes from our
innovation, χα,k (s). For instance, the peak PDF Lα,k (0) is the first moment of χα,k .
Excess Kurtosis. The moments of GSaS have closed form solutions. Of particular interest
is the excess kurtosis, plotted in Figure 4 in the (k, α) coordinate. Notice that a major division
occurs along the line of k = 5 − α. In the region where 0 < k ≤ 5 − α, there are complicated
patterns caused by the infinities of the gamma function. Only small pockets of valid kurtosis
exist.
Lemma 3.5. In the region where k > 5 − α, the excess kurtosis is a continuous function with
positive values. At large k’s, the closed form of the moments can be expanded by Sterling’s
formula. The excess kurtosis (exKurt) becomes part of a linear equation:
     
1 k−3 exKurt
(33) s− = log 1 + , where s = 1/α
2 4 3
This equation shows how GSaS works under the Central Limit Theorem. GSaS ap-
proaches a normal distribution: Lα,k → N when the excess kurtosis becomes zero. This can
happen from two directions: when α → 2, or when k → ∞.

The contour plot of excess kurtosis is shown in the (k, s) coordinate in Figure 5. It is
visually amusing. Notice the singular point at s = 1/2, k = 3.
The excess kurtosis of Eα,k has a closed form too:
3 E(X|χα,k ) E(X 5 |χα,k )
(34) exKurt(Eα,k ) = −3
E(X 3 |χα,k )2
which is shown in Figure 6. The linear contour levels are clear indication that the distribution
has an orderly structure in terms of kurtosis.
Definition 3.7 (Generalized α-stable distribution (GAS)). GAS is an experimental skew
distribution based on GSaS. The GAS Lθα,k is a ratio such that Lθα,k ∼ gαθ /χα,k . Hence, its
PDF is
Z ∞
(35) Lθα,k (x) := s ds gαθ (x, s) χα,k (s) (x ∈ R)
0
However, the PDF of a distribution must be positive-definite (Bochner’s Theorem). This
narrows the range of θ as k increases.

GAS adds the k dimension to α-stable by expanding from χα,1 to χα,k in the integral. It
subsumes α-stable as Lθα ∼ Lθα,1 , and subsumes Student’s t as tk ∼ L01,k . See Section 10 for
more detail.
GAS admits valid variance, skewness, and kurtosis, as long as there are a few degrees of
freedom, just like you would expect in Student’s t. (See Lemma 3.5 about the validity of
kurtosis.) It inherits the skewness from the α-stable law - a property that very few two-sided
skew distributions have, if any.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 17

Figure 4. The contour plot of excess kurtosis in GSaS by (k, α).

Most real-world data analyses begin with the summary statistics of the first four moments.
Plus a few more, such as the peak PDF, and the 95% and 99% confidence levels. GAS can
address all of them. This is a large improvement to α-stable.
3.4. Explanation of the S&P 500 Fit. Next, we explain how we fitted the S&P 500 daily
log return data that produced Figure 2. We used GSaS in the first pass. The values of α and
k are obtained such that the GSaS agree with the empirical kurtosis (20) and standardized
peak density (0.71). In the second pass, a small skewness is added to GAS and we fine-tuned
the fit.
However, the solution for the S&P 500 fit falls in an unconventional region as shown in
Figure 7. It is in a small pocket near α = 1, k = 3. Conventionally, as we know from Student’s
t, the kurtosis only exists for k greater than 4. But in our new (α, k) space, it is possible for
k to be less than 4 with valid kurtosis, although only in small pockets. We found the solution
by tracing the contours of the peak PDF and excess kurtosis as shown in Figure 8. (The
peak PDF has a closed form solution and is more stable than the kurtosis numerically.) The
18 S. H. LIHN

Figure 5. The contour plot of excess kurtosis in GSaS by (k, s) where s =

1/α, in order to show the linear equation described in (33) for large k’s.

intersection near α = 0.81, k = 3.3 is the solution that satisfies both properties. This is the
only solution for the S&P 500 data in the low degrees of freedom region where k < 5.
GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 19

Figure 6. The contour plot of excess kurtosis (34) of the generalized expo-
nential power distribution Eα,k by (k, s) where s = 1/α. The linear contour
levels can be clearly observed. We allow the formula to be evaluated down to
k = −2 to show the singular point at s = 1/2, k = −2.
20 S. H. LIHN

Figure 7. The region near (α = 1, k = 3) that contains the S&P 500 solution.

Figure 8. The S&P 500 solution is found by tracing the contours of the
excess kurtosis (blue line) and the peak PDF (red line). The intersection is
the solution that satisfies both properties.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 21

4. Multivariate Distribution
The appearance of N (xs) makes it possible to extend GSaS to an n-dimensional multivari-
ate symmetric distribution with the concept of the elliptical distribution[31]. We propose two
viable options. And we use the daily return distributions of the S&P 500 Index (SPX) and
CBOE VIX Index (VIX) since 1990 to examine the bivariate use case.
We need to upgrade N (xs) to a multivariate normal distribution properly. To set it up, we
vectorize x to x ∈ Rn , and introduce the covariance matrix Σ ∈ Rn×n , which is invertible to
Σ−1 . Its determinant is |Σ|. There are two routes to go from here.

4.1. Multivariate of the First Kind - Elliptical. The first route is simpler and produces
very meaningful results. All dimensions share the same shape from a single FCM. So (xs)2
in N (xs) is replaced to x⊺ Σ−1 x s2 . Such elliptical distribution closely resembles that of a
multivariate normal. The ellipse structure x⊺ Σ−1 x is perserved.
We use an extended notation N (x; µ, Σ) to represent the PDF of a multivariate normal
distribution. But for the purpose of exploring fundamental properties, we don’t want to be
bothered with the location parameter. Therefore, we set its µ = 0.
Definition 4.1. (GSaS Elliptical Multivariate Distribution) The PDF of the first kind is
Z ∞  
Σ
(36) Lα,k (x; Σ) := ds χα,k (s) N x; 0, 2
0 s
Z ∞   2 
1 n s ⊺ −1
(37) = s ds χα,k (s) exp − x Σ x
(2π)n/2 |Σ|1/2 0 2
The scipy.stats.multivariate normal function is used in our prototype implementation.
It is obvious from the first line that the total density is integrated to 1.
When α = 1, k > 0, L1,k (x; Σ) becomes the multivariate t distribution. This is validated
numerically with scipy.stats.multivariate t. On the other hand, due to Lemma 3.4,
Lα,k x; α−2α Σ = N (x; 0, Σ)


lim
α→2 or k→∞
These two validations are critically important. They show that the univariate GSaS extends
naturally to multivariate through the combination of the elliptical distribution and FCM.
From (37), its peak PDF is
E(X n |χα,k )
(38) Lα,k (0; Σ) =
(2π)n/2 |Σ|1/2
Its variance (aka the second moment) is
(39) E(Xi Xj ) = E(X −2 |χα,k ) Σi,j
Its i-th dimension marginal PDF is a GSaS:
 
(i) 1 x p
(40) Lα,k (x; Σ) = Lα,k , where σi = Σi,i
σi σi
Proof. The proof of (40) is straightforward. The marginal distribution of a multivariate
normal distribution is the down-sized multivariate normal distribution with the covariance
matrix from those selected dimensions. See Section 3.3.1 of [31]. □
22 S. H. LIHN

Hence, the i-th marginal variance is E(X 2 |Lα,k ) Σi,i , the same as E(Xi2 ) in (39) since
E(X 2 |Lα,k ) = E(X −2 |χα,k ) from Lemma 8.1. They match nicely.
(i) 1
The marginal peak density is Lα,k (0; Σ) = √2πσ E(X|χα,k ).
i
The multivariate extension of GEP (aka k < 0) has not been validated since there is no
implementation in the scipy package. But we do find it useful as an alternative model in
Section 4.3.

4.2. Multivariate of the Second Kind - Adaptive. The second route is more complex,
but also more adaptive for practical use. For instance, in a multi-asset scenario, it allows each
asset to have its own shape according to its marginal distribution. The downside is that the
elliptic structure is distorted.
In the previous method, we learn the advantage of using multivariate normal to construct
our multivariate PDF. The N (x; ...) term can be integrated away easily to make sure the total
density is equal to 1. We follow the same methodology but expand on the FCM structure.
For simplicity, we use Θi to represent the i-th parameterization: (αi , ki ). And Θ is the
collection of all Θi ’s. We vectorize s to s ∈ Rn . Each si comes from a distinct χΘi in that
dimension.
The covariance matrix Σ is sandwiched by 1/s such that the covariance matrix becomes
(s† )⊺ Σ s† where s† = diag(s)−1 . It is then enveloped by the integrals of the multi-dimensional
FCMs. To put it together -
Definition 4.2. (GSaS Adaptive Multivariate Distribution) The PDF of the second kind is
YZ ∞  
(41) LΘ (x; Σ) := dsi χΘi (si ) N x; 0, (s† )⊺ Σ s† , where s† = diag(s)−1
i 0

This object is far more computationally intense than Lα,k (x; Σ) due to the multi-dimensional
integral on s.
Note that, in the case of identity Σ, LΘ (x; Σ) does factorize into the product of the marginal
one-dimensional distributions. This is a positive feature. However, this PDF retains the
circular contour only in the center region. The contour could be distorted into rectangle-
barbell shapes with round edges in the tail regions.
Its peak PDF is
Q
E(X|χΘi )
(42) LΘ (0; Σ) = i
(2π)n/2 |Σ|1/2
Its variance (aka the second moment) is
E(X −2 |χΘi ) Σi,i

, for i = j.
(43) E(Xi Xj ) =
E(X −1 |χΘi ) E(X −1 |χΘj ) Σi,j , for i ̸= j.
Its i-th dimension marginal PDF is a GSaS:
 
(i) 1 x p
(44) LΘ (x; Σ) = Lαi ,ki , where σi = Σi,i
σi σi
Such marginal distribution can adapt to each sample’s shape. This is the major improvement
over (40). The i-th marginal variance is E(X 2 |Lα,k ) Σi,i = E(Xi2 ) in (43). They match nicely.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 23

(i)
The i-th marginal peak density is LΘ (0; Σ) = √ 1 E(X|χα ,k ). Hence, the peak density
2πσi i i
ratio between marginal and multivariate is
(i)
|Σ| 1/2
Q  
i LΘ (0; Σ)
= Q
LΘ (0; Σ) i Σi,i
p
In the bivariate case, this ratio becomes 1 − ρ2 where ρ is the correlation and |ρ| ≤ 1. On
one hand, when ρ = 0, this ratio is 1. On the other hand, the closer ρ is to 1, the smaller this
ratio is. Conversely, the ratio from the sample data implies the model correlation.
In conclusion, the differences between the elliptical version and the adaptive version are
(a) how the peak density is formed; (b) the shapes of the marginal distributions; and (c) the
off-diagonal elements of the variance.

4.3. Bivariate Application - SPX and VIX. In this section, we use the bivariate dis-
tributions to model the joint daily return distributions of the S&P 500 Index (SPX) and
CBOE VIX Index (VIX) from 1/1990 to 3/2024. We use this case to demonstrate the fitting
procedure and visualize the features mentioned above. This is the pinnacle of this work.
First, the sample statistics are calculated from the full sample. The sample covariance
matrix is Σ = [[ 0.00486232, -0.00055669], [-0.00055669, 0.0001304]]. The correlation is ρ =
−0.70. The excess kurtoses are 15.8 for VIX, and 17.4 for SPX.
Second, we fit the marginal distributions of VIX and SPX with two GSaS’s. The result
is shown in Figure 9. To manage a reasonable size of the bins in the histograms, the data
is truncated at [-0.3, 0.3] for VIX, and [-0.05, 0.05]
p for SPX.p200 bins are used. Each GSaS
is rescaled to the sample’s standard deviation: Σ0,0 and Σ1,1 . The model mean is the
same as the sample mean in the plots. For demonstration purpose, the means in the bivariate
models are discarded.
Each dimension has its own shapes: α0 = 0.64, k0 = 5.50 for VIX; and α1 = 0.88, k1 = 3.20
for SPX. The parameters are chosen such that the peak density and the excess kurtosis
between the model and sample are matched reasonably well.’ Skewness is not handled.
The same search procedure shown in Figure 8 is applied to each marginal distribution.
According to Figure 4, The high sample kurtosis in conjunction with α slightly below 1
indicates k in the neighborhood of 5. This observation is true for VIX, but not for SPX.
SPX distribution has a sharper peak. The height exceeds the range that any k ≥ 4 can
afford. The solution is still located in the pocket near α ≈ 1 and k ≈ 3, as shown in Figures
7 and 8. This is a peculiar feature of SPX index, regardless much shorter history in this case.
To configure the bivariate model distributions, we designed an adjustment procedure for
the model Σ. The main obstacle is that, if we simply feed the parameters obtained above
to the bivariate models, the peak joint density will come out lower than the sample’s peak
density, which is 1186, printed in the upper right plot of Figure 10.
We attribute the shortage to insufficient correlation. The adjustment is treated as an
optimization problem: The free variable is an adjustment factor. The model variance of each
dimension is scaled down by the factor, while the absolute of the correlation is scaled up by
the factor. Construct a new model covariance matrix. The objective is to increase the model’s
peak joint density such that its percentage deviation to the sample peak density is less than
the adjustment factor.
24 S. H. LIHN

For the adaptive bivariate model, ρ is moved to -0.81. For the elliptical bivariate model,
ρ is moved to -0.79. Since the latter only takes one set of the shape parameter, we use the
simple averages from the marginals: ⟨α⟩ = (α0 + α1 )/2 = 0.76, ⟨k⟩ = (k0 + k1 )/2 = 4.35.
The model outputs are shown in the lower two contour plots in Figure 10. The results
are satisfactory. The tilts of the contours are in agreement with that of the sample (Upper
right). The adaptive contour is particularly interesting, or even a bit exaggerating. It has a
rectangle-barbell shape, with four corners sticking out, which looks more like what’s from the
data. Elliptical contour appears to be too simple and naive.

Figure 9. The The marginal distributions of VIX and SPX daily returns
from 1/1990 to 3/2024. Each is fitted with a GSaS by matching their peak
densities, standard deviations, excess kurtosis, and shapes of the distribution.
Each GSaS is shifted to match the mean of the data. The y-axes of the left
plots are in the log scale.
GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 25

Figure 10. The bivariate distribution of VIX and SPX daily returns from
1/1990 to 3/2024. Upper left: The scatter plot of the data points, from which
the sample covariance matrix and correlation are calculated. Upper right: The
contour plot from the 2D histogram of the data. 200 bins in each dimension.
This shows approximately what kind of contours should be expected. They
are more like rectangles than ellipses. Lower left: The contour plot from the
elliptical bivariate model distribution. The tilt of the ellipses agrees with the
data (about 45◦ in the chart). Lower right: The contour plot from the adaptive
bivariate model distribution. Its shape is like rectangle-barbell shape, with four
corners sticking out, which looks more like what’s in the raw data.
26 S. H. LIHN

5. GSC: Generalized Stable Count Distribution

5.1. GSC PDF and Explanation. We recap the PDF of GSC (14) from Definition 3.1 and
compare it with that of GG (16)5 side by side:
 x d−1   x p 
Nα (x; σ, d, p) := C W−α,0 −
σ σ
|p|  x d−1 −(x/a)p
fGG (x; a, d, p) := e
aΓ( dp ) a
The parallel in notations between the two PDFs should be obvious.
GSC is the fractional extension of GG. The exponential function in GG is replaced with
the Wright function: e−z → W−α,0 (−z). The additional parameter α ∈ [0, 1] controls the
shape of the Wright function. The scale parameter is changed from a to σ to avoid confusion
between a in GG and α in GSC.
The other two parameters d, p are kept the same intentionally: d is the degree of freedom
parameter; p is the second shape parameter controlling the tail behavior (p ̸= 0, dp ≥ 0). In
our use cases, p is either α or 2α. In the later case, p is exactly the stability index.
The range of α is constrained to [0, 1] due to the analytic property of W−α,0 (−z). When
α → 1, it becomes a Dirac delta function δ(z − 1). This is a well known result.
On the other hand, when α → 0, GSC becomes GG. See Section 5.3 for proof. A broad
class of distributions belong to GG, see Table 1.

5.2. Determination of C. We derive the value of the normalization constant C in (15) as

following.

Proof. The normalization constant C in (14) is obtained from the requirement that the integral
of the PDF must be 1:
Z ∞ d
Cσ Γ( p )
Nα (x; σ, d, p) dx = =1
0 |p| Γ( dp α)

where the integral is carried out by the moment formula of the Wright function (109).
Hence, C is determined as
d
|p| Γ( p α)
C= (α ̸= 0, d ̸= 0)
σ Γ( dp )

We typically constrain dp ≥ 0 and p is typically positive. But it becomes negative in the

inverse distribution and/or characteristic distribution types. So we need |p| to ensure C is
positive.
For the case of α ̸= 0 and d → 0, due to (123), we have
|p|
C= (α ̸= 0, d = 0)
σα
These two cases are combined to form (15). □
5The CDF of a GG is a regularized incomplete gamma function Γ(·, ·). Hence, it can also be parameterised
x d x c



such as GenGamma( a ; s = p , c = p) := Γ s, a
.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 27

5.3. Mapping between GSC and GG at alpha = 0. All of the gamma-based, exponential-
based distributions can be expressed by GSC at α = 0. We prove the first line of (17) as
following.

Proof. For certain proofs, it is more convenient to express the PDF of GSC in terms of the
M-Wright function Mα (z) in (114):

 x d+p−1  x p 
(45) Nα (x; σ, d, p) = αC Mα
σ σ

We apply the property M0 (z) = exp(−z) mentioned below (116) to Nα (x; σ, d, p) from (45),
and obtain:

 x d+p−1   x p 
lim Nα (x; σ, d, p) = lim αC exp −
α→0 α→0
 σ  σ
d p/σ x d+p−1
    x p 
= lim αΓ α exp −
α→0 p Γ( dp ) σ σ
p/σ  x d+p−1   x p 
= exp −
Γ( d+p ) σ
p
σ
= fGG (x; σ, d + p, p)

Hence, we reach the mapping of GG to GSC at α = 0 described in the first line of (17):

(46) fGG (x; σ, d, p) = N0 (x; σ, d − p, p)

The GSC mapping of classic distributions at α = 0 is illustrated in Table 1.

28 S. H. LIHN

GG: fGG (x; σ, d, p) GSC: N0 (x; σ, d, p)

Distribution (PDF) Classic Equiv. σ d p α σ d p
(1)
Stretched Exp: Eα (x) 1 1 α 0 1 1−α α
√ √
Half-Normal: 2 N (x) 2 1 2 0 2 -1 2
Weibull: Wb(x; k) 1 k k 0 1 0 k
Exponential: 2 L(x) Wb(x; k = 1) 1 1 1 0 1 0 1
√ √
Rayleigh(x) Wb( √x2 ; k = 2) 2 2 2 0 2 0 2
Gamma: Γ( σx ; s) σ s 1 0 σ s−1 1
2 √ √
χk (x) Γ( x2 ; k2 ) 2 k 2 0 2 k−2 2
χ2k (x) Γ( x2 ; k2 ) 2 k
2
1 0 2 k
2
−1 1
GenGamma( σx ; s, c) Γ(( σx )c ; s) σ sc c 0 σ c(s − 1) c
or fGG (x; σ, d, p) σ d p 0 σ d−p p
IG(x; k) 1 −k −1 0 1 −k + 1 −1
IWb(x; k) 1 −k −k 0 1 0 −k

(1)
Table 1. Mapping classic distributions to N0 (x; σ, d, p). Eα (x) =
1
exp(−xα )/Γ


α + 1 is a one-sided exponential power distribution.

5.4. Mapping between GSC and GG at alpha = 1/2. We prove the second line of (17)
as following.
2
Proof. From the property mentioned below (116) that M 1 (z) = z2 W− 1 ,0 (−z) = √1 e−z /4 , the
π
√ 2 √ 2
exponential function can be re-interpreted as e−z = π M 1 (2 z).

p 2
By a change of variable z = xa , GG is mapped to GSC at α = 1/2 as
√    
π p/a  x d−1 x p/2
fGG (x; a, d, p) = d
M1 2
Γ( p ) a 2 a
√    
π p/a  x d−1  x −p/2 x p/2
= W − 1
,0 −2
Γ( dp ) a a 2 a
√
 !p/2 
π p/a x
  d−p/2−1 2/p
2 x
= d
W− 1 ,0 − 
Γ( p ) a 2 a
 a p p
= N 1 x; σ = 2/p , d = d − , p =
2 2 2 2
Hence, we reach the second line of (17). □
If a classic distribution can be mapped to N0 (x; ...), it can also be mapped to N 1 (x; ...).
√ 2
That is how we map the χ distribution from χk (s) = N0 (s; σ = 2, d = k, p = 2) to its
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 29

fractional counterpart χk (s) = N 1 (s; σ = √12 , d = k − 1, p = 1) in (55) that eventually leads

2
to (18).
The α = 1/2 mapping is illustrated in Table 2.

GSC: Nα (x; σ, d, p)
Distribution (PDF) Classic Equiv. α σ d p
1 −2/k k k
Weibull: Wb(x; k) 2 2 2 2

Gamma: Γ( σx ; s) 1
2 σ/4 s− 1
2
1
2
2
χk (x) Γ( x2 ; k2 ) 1
2
√1
2
k−1 1
χ2k (x) Γ( x2 ; k2 ) 1
2
1
2 (k − 1)/2 1
2

GenGamma( σx ; s, c) Γ(( σx )c ; s) 1
2 2−2/c σ (s − 21 )c c/2
p p
or fGG (x; σ, d, p) 1
2 2−2/p σ d− 2 2
1 1
IG(x; k) 2 4 2 −k − 12
1
IWb(x; k) 2 22/k −k/2 −k/2

Table 2. Mapping of some classic distributions to N 1 (x; σ, d, p)

2

5.5. GSC Hankel Integral. The Hankel integral establishes another connection between
GSC R and GG. We begin with the Hankel integral of Fα (z) in (115) where Fα (z) = W−α,0 (−z) =
1 α
2πi H dt exp (t − zt ). Hence, for αd/p ̸= 0, the Hankel integral of GSC is

Z
1  x d−1   x p 
Nα (x; σ, d, p) = dt C exp t − tα
2πi H σ σ
!p !
1
Z  x d−1 x tα/p
= dt et C exp −
2πi H σ σ
  Z  t 
αd 1 e  σ 
(47) = Γ dt fGG x; a = , d = d, p = p
p 2πi H tαd/p tα/p
This shows that GSC can be obtained from the Hankel integral over its GG counterpart.

5.6. GSC Moments. GSC’s moment formula plays a foundational role since FCM and
GSaS’s moment formulas are derived from it.

Lemma 5.1. The n-th moment of GSC is

Γ( dp α) Γ( n+d
p )
(48) E(X n |Nα (x; σ, d, p)) = σ n , for α ̸= 0, d ̸= 0.
Γ( dp ) Γ( n+d
p α)
30 S. H. LIHN

Proof. The integral of the n-th moment is

Z ∞  x d−1   x p 
n
E(X |Nα (x; σ, d, p)) = dx xn C W−α,0 −
0 σ σ
Z ∞
C n+d
−1
 x p
= σ n+1 dz z p W−α,0 (−z) , where z =
p 0 σ
 
This is basically the n+dp -th moment of the Wright function in (109). We arrive at
n+d
n n+1 C Γ( p )
E(X |Nα (x; σ, d, p)) = σ
p Γ( n+d
p α)
Γ( dp α) Γ( n+d
p )
= σn , for α ̸= 0, d ̸= 0.
Γ( dp ) Γ( n+d
p α)
□
Although the formula exists, whether a specific moment exists can be complicated by the
choice of n relative to the shape parameters α, d, p.
We caution a common scenario in this paper that such formula needs special handling due
to Γ(z) ≈ 1/z when z → 0. For instance, in this case, d = 0 or n = −d. This is documented
in (123). There are numerous situations like these, we may not mention them repeatedly in
subsequent encounters.
As a validation, when α → 0, it becomes the moments of fGG (x; σ, d + p, p), as expected
from Section 5.3:
n n
Γ( n+d+p
p )
E(X |Nα (x; σ, d, p)) = σ d+p
Γ( p )
When α → 1, all gamma terms cancel out in E(X n ), leaving it to E(X n ) = σ n . Therefore,
the mean is σ, and the variance is 0, which indicates it is a Dirac delta function δ(x − σ).
5.7. Stable Count and Stable Vol Distributions. GSC traces its root to the discovery
of the stable count distribution in 2017[16] and the stable vol distribution in 2020[17]. It is
important to reiterate the following relations.
The PDF of the original stable count distribution (SC) is defined as
 
1 1 1 1
(49) Nα (ν) := 1

Lα = 1

W−α,0 (−ν α )
Γ α +1 ν ν Γ α +1
where Lα (x) is the PDF of the one-sided stable distribution. Lα (x) = x−1 W−α,0 (−x−α ). SC
is subsumed to GSC as Nα (ν; σ = 1, d = 1, p = α).
The stable vol distribution (SV) is a variant of SC where
√ √
2πΓ α2 + 1  √ α


2 2π
(50) Vα (s) := N α (2s ) = W − α
,0 −( 2s) .
Γ α1 + 1 Γ α1 + 1



2 2

√
It is subsumed to GSC as N α2 (s; σ = 1/ 2, d = 1, p = α).
The original idea of having two variants, SC and SV, came from the so-called ”Lambda
α
decomposition”: transforming e−|z| as a product distribution with the Laplace distribution
L(x) and the normal distribution N (x):
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 31

 Z ∞  
dν |z|
L Nα (ν) , for 0 < α ≤ 1.


1
Z0 ∞ ν  ν 
α

(51)
e−|z| =
2Γ 1
+1 ds |z|
α N Vα (s) , for 0 < α ≤ 2.


s s

0

It turns out the second line in (51) is very useful. First, the normal distribution represents
the Brownian noise. This is the early form of the Gaussian mixture presented in Section 2.
Secondly, in the fractional χ notation, Vα (x) = χα,2 (x), which has two degrees of freedom.
This laid the groundwork for FCM.
Thirdly, it is in fact decomposing the CF of the α-stable law into a Gaussian mixture. It
is used in the proof of Section 6 that follows.
In closing, several known fractional distributions are expressed in GSC in Table 3.

GSC: Nα (x; σ, d, p)
Distribution (PDF) Wright Equiv. α σ d p
One-sided stable: Lα (x) x−1 W−α,0 (−x−α ) α 1 0 −α
SC: Nα (ν) α 1 1 α
α √1
SV: Vα (s) 2 2
1 α
1
M-Wright: Mα (z) αz W−α,0 (−z) α 1 0 1
M-Wright: Γ(α)Fα (z) Γ(α)W−α,0 (−z) α 1 1 1

Table 3. GSC mapping of several known fractional distributions in the literature.

6. The Skew-Gaussian Kernel and Fractional Chi-1

6.1. Derivation of the Skew-Gaussian Kernel. We prove how the kernel gαθ (x, s) in (26)
comes into existence as following.

Proof. The minus-log of α-stable CF in Feller’s parameterization is defined as (Section XVII.6,

p.581-583 of [9]):

 
iθπ
ψαθ (ζ) = exp sgn(ζ) |ζ|α
2
θπ
= (cos(Θ) + i sin(Θ)) |ζ|α , where Θ = sgn(ζ)
2
and 0 < α ≤ 2, |θ| ≤ min{α, 2 − α}.

Put it into the integral of the α-stable PDF (See (F.35) of [18]), we get
32 S. H. LIHN

Z ∞
1
Lθα (x) = dζ exp(−ψαθ (ζ)) e−ixζ
2π −∞
Z ∞
1 α
= dζ e− cos(Θ)|ζ| e−i(B+xζ) , where B = sin(Θ)|ζ|α
2π −∞
The imaginary part of the integral cancels out due to the sign of ζ. The real part of the
integral can be simplified to the integral on the positive axis (ζ ≥ 0). We can remove sgn(ζ)
from Θ. Hence,
Z ∞
1 α θπ
Lθα (x) = dζ e− cos(Θ)ζ cos(B + xζ), where Θ = , B = sin(Θ)ζ α
π 0 2
α
Next, we apply the Lambda decomposition (51) for the stable-vol Vα (ζ) on the e− cos(Θ)ζ
α α
term. Let u = cos(Θ)1/α ζ, then e− cos(Θ)ζ = e−u . We have
!
∞ ∞ 2Γ( α1 + 1)
Z Z  
1 −(u/s)2 /2 ds
Lθα (x) = dζ cos(B + xζ) √ e Vα (s)
0 0 2π s π
!
Z ∞  Z ∞  1
1 2 2Γ( + 1)
(52) = s ds dζ cos(B + xζ)e−(u/s) /2 √α s−1 Vα (s)
0 πs 0 2π
The terms inside the first big parenthesis of (52) is the skew-Gaussian kernel in its raw
format:
Z ∞
1 2 /2
gαθ (x, s) = dζ cos(B + xζ)e−(u/s) , where u = cos(Θ)1/α ζ
πs 0

Make one more change of variable: t = u/s = cos(Θ)1/α ζ/s, and B = tan(Θ)(st)α , we arrive
at the final form presented in (26) as
Z ∞  
1 x α 2
gαθ (x, s) = dt cos τ (st) + st e−t /2 (s ≥ 0)
qπ 0 q
where q = cos(θπ/2)1/α , τ = tan(θπ/2)
□
R∞
Next, we prove that Lθα (x) is indeed 0 s ds gαθ (x, s) χα,1 (s).
Proof. The terms inside the second big parenthesis of (52) is simplified by (50),
!
2Γ( α1 + 1) −1  √ α
(53) √ s Vα (s) = 2s−1 W− α2 ,0 −( 2s)
2π
1
= N α2 (s; σ = √ , d = 0, p = α)
2
= χα,1 (s)
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 33

So according to (53) and (56), the α-stable PDF (52) becomes a ratio distribution of the
kernel and an FCM of one degree of freedom:
Z ∞
(54) Lθα (x) = s ds gαθ (x, s) χα,1 (s)
0
□
6.2. Some Properties of the Skew Kernel. Some properties of gαθ (x, s) are analyzed in
this section. This will help us better understand its behavior intuitively, as well as provide a
better implementation. R∞ 2
First, the integral inside gαθ (x, s) is bounded by the Gaussian integral 0 dt e−t /2 , where
2
e−t /2 is the upper bound to the absolute of the integrand since | cos(z)| ≤ 1. Therefore, the
contribution of the integrand decreases rapidly as t gets larger. This is the main reason such
representation is chosen. Implementation-wise, it is better to cut off the integral at a certain
large t (e.g. 100).
Indeed, when θ = 0, the Gaussian integral becomes exact, that is, gα0 (x, s) = N (xs); and
0
Lα (x) becomes symmetric.
Proof. When θ = 0, q = 1 and τ = 0. Therefore,
1 ∞
Z
2
0
gα (x, s) = dt cos (xst) e−t /2
π 0
2
From (128), this is √1 e−(xs) /2 = N (xs). □
2π

When θ ̸= 0, some parts of gαθ (x, s) can 

be negative.
  Hence, it is not a distribution per se.
θ 1 x
When α = 1, we have g1 (x, s) = q N τ + q s , and Lθ1 (x) becomes symmetric over
x0 = −qτ = − sin(θπ/2). That is, the effect of θ is no longer skewness, but a shift in location.
x
Proof. When α = 1, the argument of cos() in (26) becomes τ (st) + q st = Y t where Y =
 
τ + xq s. Therefore, from (128),
Z ∞
1 2
g1θ (x, s) = dt cos (Y t) e−t /2
qπ 0
  
1 1 x
= N (Y ) = N τ+ s
q q q
□
√
At s = 0, gαθ (x, 0) is simply a constant 1/ 2πq. This is straightforward
√ since gαθ (x, 0) =
1
R∞ −t2 /2 θ
qπ 0 dt e . Furthermore, for a given (α, θ) pair, |gα (x, s)| ≤ 1/ 2πq for all x’s and s’s.
Next we prove:
R∞
Lemma 6.1. For a given s > 0, −∞ gαθ (x, s) dx = 1/s.
Proof. We integrate the α-stable PDF (54) by the full range of x, and it should be one:
Z ∞ Z ∞ Z ∞
Lθα (x) dx = dx s ds gαθ (x, s) χα,1 (s) = 1
−∞ −∞ 0
34 S. H. LIHN

Rearranging the double integrals, it becomes

Z ∞ Z ∞ 
θ
s ds gα (x, s) dx χα,1 (s) = 1
0 −∞
R∞
Since χα,1 (s) is a PDF, we have 0 χα,1 (s) ds = 1. This forces the integral inside the big
R∞
parenthesis to be 1/s, which is −∞ gαθ (x, s) dx = 1/s.
□

7. FCM: Fractional Chi-Mean Distribution

FCM is the core of the innovation in this paper. We elaborate the thought process that
led to FCM in this section. There is a theoretical side of thinking. There is also the aspect
of design, mainly on how to choose the scale of standardization such that it is consistent at
both k = 1 and k → ∞ limits.

7.1. Derivation of Fractional Chi. First of all, after many experiments (elaborated here
and also the subordination results in Appendix B), I’ve come to realization that we need to
use a ratio distribution (108), instead of a product distribution (107), to describe the inner
structures of both Student’s t and α-stable. This is a paradigm change.
Secondly, after many analyses, I’ve settled down on using a normal distribution, instead of
other alternatives, such as Laplace or Cauchy, as the unit distribution in the middle. This
has been formalized into the continuous Gaussian mixture in Section 2.
The advantage of a normal distribution is obvious from the perspective of stochastic cal-
culus. Whatever distributional equation we come up R ∞with can be translated to the random
variable generation directly. Symbolically, L(x) = 0 sdsN (sx)S(s) in the former becomes
dLt = dWt /St in the later, where dWt is the Brownian noise. Section 11 is dedicated to the
subject of random variable generation, that is, how St can be generated.
Thirdly, for the expression of the normal distribution, we’ve evaluated two options: (a) use
the notion of variance (which we like to use ν for it in (51)); or (b) use the notion of volatility
(that is, squared root of variance, which we like to use s for it in (51)). After many thoughts,
we didn’t chose the variance route.
We wouldn’t say the last point is totally set in stone. Searches on ”chi-squared distribution”
vs ”chi distribution” showed that people like to talk about sum of square of variables, instead
of squared root of that sum. But we find that our formula is better presented by the later,
instead of the former.
During the following discussions, we assume k > 0. k has the ordinary physical meaning
as the degrees of freedom.
Now let’s begin with an expression of Student’s
q t distribution tk as the ratio of a normal
and the root-mean of a chi-squared: tk ∼ N / χ2k /k as described in [28] and [31].
We prefer to combine
q the root-mean-square together. Therefore, we define the χ-mean
√
2
distribution as χk = χk /k = χk / k, which turns out to be a GSC:
 
1
(55) χk (s) := N 1 s; σ = √ , d = k − 1, p = 1
2 2k
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 35

Next, we consider the more complicated matter of α-stable. In Section 6, after a lengthy
derivation, we arrive at (53), where we find ourselves dealing with a similar GSC:
1
(56) χα,1 (s) := N α2 (s; σ = √ , d = 0, p = α)
2
Here χα,1 has a physical interpretation: It is the α-stable’s χ distribution with one degree
of freedom. This interpretation will be used in (58) below.
By comparing the GSC parameters between (55) and (56), it becomes obvious that it is
possible to merge χk (s) and χα,1 (s) carefully and come up with a super distribution family
nicknamed FCM. The result is presented in (61). But it takes quite some reasonings to get
there...

7.2. Fractional Chi-1. In Lemma 3.3, we mentioned that χα,1 is the fractional version of
classic χ1 and it is the building block of SaS since Pα = L0α ∼ N /χα,1 , or in a distributional
form:
Z ∞
Pα (x) := L0α (x) = s ds N (xs) χα,1 (s)
0

Compare it to tk ∼ N /χk . We can re-interpret Pα as a sort of t-statistics in the fractional

space. And χα,1 is the fractional chi of one degree of freedom 6.
The PDF of χα,1 has some peculiar behaviors (listed under Lemma 3.3) that span from a
Dirac delta function, to a half-normal distribution, to 1/x type. This affects the behavior of
SaS directly. We prove them here.
First, we show χ2,1 (x) is a Dirac delta function. From FCM’s moment formula (63),
E(X n |χ2,1 ) = 2−n/2 . Its first moment is √12 . Its second moment is 21 . Its variance is zero. So
√

it is a Dirac delta function at √12 , aka χ2,1 (x) = δ x − 1/ 2 .
√
Secondly, we show χ1,1 (x) is a half-normal distribution. From (61), its PDF is 2x−1 W− 1 ,0 (− 2x).
2
2
From Section 5.4, it becomes 2/π e−x /2 .
p

Thirdly, we prove that χα,1 (x) ≈ α/(ex), and subsequently, Pα (x) ≈ χα,1 (x)/2 ≈ α/(2ex),
when α is very small but x is not too small.

Proof. We use the result of GSC at α → 0 in Section 5.3. Combine (56) with (46), we obtain
1
χα,1 (x) → fGG (x; σ = √ , d = α, p = α)
2
α √ α  √ α
= ( 2x) exp −( 2x)
x
√ α
The condition that α is very small but x is not too small means that ( 2x) ≈ 1. That is,
α| log(x)| → 0. Under such condition, χα,1 (x) ≈ α/(ex).

6Wikipedia has a paragraph for this subject: https://en.wikipedia.org/wiki/Proofs_related_to_

chi-squared_distribution#Derivation_of_the_pdf_for_one_degree_of_freedom
36 S. H. LIHN

For Pα (x) at α → 0, we have

Z ∞
Pα (x) = sds N (sx) χ̄α,1 (s)
0
Z ∞
= (α/e) ds N (sx)
0
= α/(2ex)
□
This example illustrates it is increasingly difficult to calculate Pα (x) through integrating
s χα,1 (s) numerically when α is small (e.g. α < 0.3) and x → 0.

7.3. Physical Interpretation of Fractional Chi. To gain physical understanding on χα,k ,

let’s repeat how the χ2 distribution is derived classically, as described in [27], but adapt it to
the fractional interpretation needed for our purpose.
A fractionally generalized chi-distributed random variable R with k degrees of freedom is
defined as the squared root of the sum of k squared random variables of the Sα distribution,
which shall be clarified later that Sα is χα,1 . So we have
v
u k
uX
X1 , ..., Xk ∼ Sα =⇒ R := t Xi2
i=1

Let x1 , ..., xk be the values of X1 , ..., Xk and consider (x1 , ..., xk ) to be a point in the k-
dimensional space. Define the distance to that point:
v
u k
uX
r=t x2 i
i=1

Let fR (r) and FR (r) be the PDF and CDF of R. Their relation is
FR (r + dr) − FR (r) = fR (r)dr
Z Y k
= Sα (xi )dxi
V i=1

where Sα (xi ) is the PDF of the Sα distribution, and V is the volume on the (k−1)-dimensional
surface in the k-space at the radius r.
Assume we can refactor the above integral such that
k
Z  Y !
fR (r)dr = dx1 ...dxk Sα (xi )
V i=1
where we deal with the two big parentheses separately.
The first part is an integral of the surface area of the (k − 1)-dimensional sphere with radius
r, which is
π k/2
Z 
(57) A(r)dr = dx1 ...dxk = 2 rk−1 dr
V Γ(k/2)
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 37

For the second part, we argue it should just be χα,1 in (56):

k
!
Y 2  √ α
(58) Sα (xi ) ∼ χα,1 (r) = W α2 ,0 −( 2r)
r
i=1
Combining (57) and (58), we arrive at the PDF for R (apart from a normalization constant):
 √ α
fR (r) ∝ rk−2 W α2 ,0 −( 2r)
1
= N α2 (r; σ = √ , d = k − 1, p = α)
2
Therefore, fR (r) is the fractional-chi χα,k (r) that we seek.
Note that we glossed over the constant term. We picked up the rk−1 factor from A(r)
but ignored the rest of it. This ”surface area of the (k − 1)-dimensional sphere” is more
complicated than a simple Euclidean sphere. This topic is left for future research.
7.4. From Fractional-Chi to Fractional Chi-Mean. Now we have χα,k (x) = N α2 (x; σ =
√1 , d = k − 1, p = α). We need to ”standardize” its mean, so to speak. Our cue is that
2 √ √
χk = χk / k works for Student’s t. A simple multiplier of 1/ k for σ suffices. The question
is: What works for χα,k in general?
We examine the mean of χα,k . From (48), its mean is
1 Γ( k−1 k
2 ) Γ( α )
E(X|χα,k ) = √
2 Γ( k−1 k
α ) Γ( 2 )
q
At k = 1, E(X|χα,1 ) = π2 Γ α1 + 1 , which we want to preserve for α-stable.

Its large k asymptotic is

(59) E(X|χα,k ) ∼ k 1/α−1/2 α−1/α , k ≫ 1;
whose k dependency we would like to eliminate so that the mean doesn’t go to 0 or infinity.
So our requirement for FCM is to modify the σ parameter in χα,k , according to the k
component of χα,k ’s asymptotic mean, such that FCM has an asymptotic mean independent
of k.
The solution is the following. Define
k 1/2−1/α
(60) σα,k := √
2
√ √
which has the desired properties: σ1,k = 1/ 2k, and σα,1 = 1/ 2, in order to preserve both
Student’s t and α-stable distributions.
Hence, FCM is constructed as
(61) χα,k (x) := N α2 (x; σ = σα,k , d = k − 1, p = α)
1  
= N α2 (x; σ = √ k 1/2−1/α , d = k − 1, p = α)
2
√ !k−1 √ !α !
αΓ( k−1 2 ) 2 k−2 2x
= x W α2 ,0 − 1/2−1/α
Γ( k−1α )
k 1/2−1/α k
38 S. H. LIHN

To verify, its large k asymptotic mean doesn’t have k dependency: (See Section 7.8)
lim E(X|χα,k ) = α−1/α
k→∞
This meets our design goal. E(X|χα,1 ) = E(X|χα,1 ) and limk→∞ E(X|χ1,k ) = 1 as expected.
√
And it is nice to have E(X|χ2,k ) = 1/ 2 for every k.
Interpretation: The α-stable law states that
X1 + X2 + ... + Xk ∼ k 1/α X
Therefore, the k 1/2−1/α term can be viewed as the relative scale (ratio) between the normal

scaling factor k 1/2 and α-stable scaling factor k 1/α . That is what we added to σα,k as the way
to ”demean” the k dependency.
7.5. FCM Hankel Integral. From (47), make the following substitutions: α/p = 1/2, d =
k − 1, αd/p = (k − 1)/2, p = α, and σ = σα,k . The Hankel integral of FCM for k > 0 is

et
     
k−1
Z
1 σα,k
(62) χα,k (x) = Γ dt fGG x; a = √ , d = k − 1, p = α
2 2πi H t(k−1)/2 t
This expression requires k ̸= 1.
7.6. FCM Moments. FCM’s moment is a special case of GSC’s moment, which is carried
out in Section 5.6.
For k > 0, based on (18) and (48), substitute α/p = 1/2, d = k − 1, p = α and σ = σα,k ,
and we obtain FCM’s moments as
!n
k 1/2−1/α Γ( k−1
2 ) Γ(
n+k−1
)
(63) E(X n |χα,k ) = √ k−1
α
n+k−1
, for α ̸= 0, k > 0, k ̸= 1.
2 Γ( α ) Γ( 2 )
When k = 1 or k + n = 1, it requires typical handling of gamma functions at zero (123).
For k < 0, based on (19) and (48), substitute α/p = −1/2, d = k, p = −α and σ = 1/σα,k ,
and we obtain FCM’s moments as
1/2−1/α −n
Γ( |k| ) Γ( |k|−n
!
n |k|
√ 2 α )
(64) E(X |χα,k ) = |k| |k|−n
, for α ̸= 0, k < 0.
2 Γ( α ) Γ( 2 )
When n = |k|, it requires typical handling of gamma functions at zero (123). Higher moments
may not exist when n > |k|.
As a validation, when α = 1, (63) is reduced
√ to the moment formula of a χ distribution
with k degrees of freedom and scale of 1/ k due to (122).
FCM’s first moment is directly related to GSaS peak PDF in (66). Numeric calculation of
FCM’s moments is important in that it directly impacts the accuracy of GAS PDF calculation,
especially at small α and when x is near zero.
7.7. FCM Reflection Formula. We prove the reflection formula shown in (21) and (22).
E(X −n+1 |χα,k )
E(X n |χα,−k ) = , k > 0.
E(X|χα,k )
The approach is to show that the moments evaluated from the RHS via (63) is the same as
those of the LHS from (64) with k replaced with −k.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 39

Proof. Assume k > 0, the n-th moment on the LHS is

!−n
k 1/2−1/α Γ( k2 ) Γ( k−n
α )
E(X n |LHS) = E(X n |χα,−k ) = √
2 Γ( αk ) Γ( k−n
2 )
The n-th moment on the RHS is
∞
xn−3
Z  
n 1
E(X |RHS) = χ dx
0 E(X|χα,k ) α,k x
Make a change of variable t = 1/x, we get
∞
t−n+1
Z
n
E(X |RHS) = χ (t) dt
0 E(X|χα,k ) α,k
E(X −n+1 |χα,k )
=
E(X|χα,k )
Hence, (22) is true if E(X n |RHS) = E(X n |LHS).
Plug (63) to both the numerator and denominator,
 
1/2−1/α −n+1
! ! !
−n+k
k−1
|k|1/2−1/α Γ( k−1 k

n |k| Γ( 2 ) Γ( α ) 2 ) Γ( α )
E(X |RHS) =  √ −n+k
 √
2 Γ( k−1
α ) Γ( 2 ) 2 Γ( k−1 k
α ) Γ( 2 )
!−n
|k|1/2−1/α Γ( k2 ) Γ( k−n
α )
= √ k k−n
2 Γ( α ) Γ( 2 )
It is clear that E(X n |RHS) and E(X n |LHS) are the same.
□
7.8. FCM at Infinite Degrees of Freedom. First, we establish the following lemma:
Lemma 7.1. We use (124) to simplify (63) and obtain the asymptotic value of the n-th
moment for k ≫ 1:
!n  
|k|1/2−1/α k −n/2 k n/α
 
(65) n
E(X |χα,k ) ∼ √ = α−n/α
2 2 α
It follows from the reflection rule that E(X n |χα,−k ) ∼ αn/α . Notice that the asymptotic value
has no dependency on k.
The proof for Lemma 3.2 becomes very easy, which states that FCM becomes a delta
function at x = mα when k ≫ 1, where mα is its first moment. We prove it by showing mα
is α−1/α and its variance is zero.
Proof. Let mα = α−1/α , we have E(X n |χα,k ) ∼ (mα )n . Hence, the first moment is obviously
E(X|χα,k ) ∼ α−1/α
The second moment is E(X 2 |χα,k ) ∼ m2α . Therefore, its variance is E(X 2 |χα,k ) − m2α = 0.
FCM becomes a delta function as k → ∞.
On the negative side, the reflection rule ensures the same proof applies such that E(X 2 |χα,−k )−
E(X|χα,−k )2 = 0. FCM becomes a delta function as −k → −∞.
□
40 S. H. LIHN

8. GSaS: Generalized Symmetric alpha-Stable Distribution

The PDF of a GSaS takes the form of an elegant ratio distribution:
Z ∞
Lα,k (x) := s ds N (xs) χα,k (s)
0
This two-sided distribution is one of the most important outcomes of this work. We prove
some closed form results in this section.

8.1. GSaS Peak PDF. The peak PDF of a GSaS at x = 0 comes from the first moment of
FCM in Section 7.6:
Z ∞
1 1
Lα,k (0) = √ s ds χα,k (s) = √ E(X|χα,k )
2π 0 2π
k−1 k

k 1/2−1/α Γ( 2 ) Γ( α )
√ , k > 0.


 2 π Γ( k−1 k

α ) Γ( 2 )


(66) =
 1 Γ( |k|
2 ) Γ( |k|−1
α )


 √  
|k| |k|−1
, k < 0.
 π |k|
 1/2−1/α Γ( α ) Γ( 2 )

When k = ±1, it requires the typical special handling of the gamma function at zero.
The pstandardized peak PDF is the peak PDF Lα,k (0) multiplied by the standard devi-
ation, E(X 2 |Lα,k ), if it exists. We argue that the standardized peak PDF is an important
statistic since it is invariant to the scale of the sample data. In finance, this means that the
standardized peak PDF measures the similarity of two data sets taken from two periods of
different volatility environments, or from two different assets altogether.
Aligning the peak PDF was often impossible in the past due to the limited one dimensional
parameter space in either α or k. GSaS overcomes it with a two-dimensional parameter space.
This opens up much greater flexibility, e.g., to match the standard deviation, the peak density,
and the kurtosis simultaneously.
The first validation of Lα,k (0) is that it becomes the peak PDF of the original α-stable
distribution when k → 1:
1
Lα,1 (0) = Pα (0) = Γ (1 + 1/α)
π
The second validation is that it becomes the peak PDF of the exponential power distribution
when k → −1:
1
Lα,−1 (0) = Eα (0) =
2Γ (1 + 1/α)

8.2. GSaS Moments. GSaS has closed form solution for all the moments. This is a great
improvement over the original α-stable distribution.
Lemma 8.1. The n-th moment of GSaS is
2n/2
 
n+1
(67) n
E(X |Lα,k ) = √ Γ E(X −n |χα,k ), k ̸= 0
π 2
when n is even. The odd moments are zero.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 41

Its closed form is

 !
n
  k−1 k−n−1
 2 n + 1 Γ( 2 ) Γ( α )
√
Γ , k > 0.


 π k 1/2−1/α n Γ( k−1 k−n−1

 2 α ) Γ( 2 )
n
(68) E(X |Lα,k ) =
1/2−1/α n |k| |k|+n
  !

 (|k| ) n + 1 Γ( 2 ) Γ( α )

 √ Γ |k| |k|+n
, k < 0.
π 2

 Γ( ) Γ( α 2 )
When k = ±1, n + 1, −n, it requires the typical special handling of the gamma function at
zero.
This formula is continuous and well-behaves when k is sufficiently large, but it gets very
complicated when k is small. This is a complexity inherent from the α-stable distribution.
Proof. The integral of the n-th moment is
Z ∞ Z ∞
n n
E(X |Lα,k ) = dx x s ds N (xs) χα,k (s)
−∞ 0
Z ∞  Z ∞ 
n
= s ds dx x N (xs) χα,k (s)
0 −∞

Inside the big parenthesis is the n-th moment of a normal distribution (see (130)), which
n/2
is 2√π s−n−1 Γ n+1


2 . Therefore, we get
Z ∞
2n/2

n+1
n
E(X |Lα,k ) = √ Γ s−n ds χα,k (s)
π 2 0

The remaining integral is the ”−n”-th moment of FCM. Hence, we get (67).
To obtain the first line of (68), copy (63) into (67), and replace every n with −n.
Similarly, copy (64) into (67), replace every n with −n, and we reach the second line of
(68).
□
GSaS Variance: The second moment m2 , aka the variance, is
m2 = E(X 2 |Lα,k ) = E(X −2 |χα,k )
In general, m2 ̸= 1 for finite k.
At k ≫ 1, according to (65), the variance of the normal-like distribution is
lim m2 = α2/α
k→∞
lim m2 = α−2/α
k→−∞

Both are independent of k. This is the intended result from Section 7.8.

8.3. GSaS Kurtosis. Let m4 = E(X 4 |Lα,k ), it is straightforward to get the kurtosis as
m4 E(X −4 |χα,k )
(69) = 3
m22 E(X −2 |χα,k )2
The excess kurtosis is exKurt = m4 /m22 − 3.
42 S. H. LIHN

For k > 0, the closed form is

k−1

k−5

! k−1

k−3

!2
m4 Γ 2 Γ α Γ 2 Γ α
=3
m22 k−1 k−5 k−1 k−3





Γ αΓ 2 Γ Γ α 2
 
k − 5 Γ(s(k − 5))Γ(s(k − 1))
(70) =3 , where s = 1/α.
k−3 Γ(s(k − 3))2
We prove the linear relation of excess kurtosis in (33) for k ≫ 1 as follows.

Proof. In Figure 5, notice the singular point at s = 1/2, k = 3. Hence we make a change of
variable by k3 := k − 3:
 
m4 2 Γ(s(k3 − 2))Γ(s(k3 + 2))
= 3 1 −
m22 k3 Γ(sk3 )2
Next, the gamma functions are expanded via the Sterling’s formula (125) for large k3 . It
is simplified to

2 sk3 −2s+1/2 2 sk3 +2s−1/2

   
m4
=3 1− 1+
m22 k3 k3
Take the log on both sides, and retain the Taylor series of log(1 + x) up to k3−2 . We have
   
m4 4 1
log = s−
3 m22 k3 2
Since m4 /m22 = 3 + exKurt where exKurt is the excess kurtosis, we arrive at the proof that
     
1 k−3 exKurt
s− = log 1 +
2 4 3
□

As k → ∞, the excess kurtosis approaches zero. GSaS becomes a normal distribution

asymptotically, N (0, m2 ). Only when α = 1, limk→±∞ m2 = 1. It becomes a N (0, 1).
Rx 1
8.4. GSaS CDF. The CDF formula of GSaS is straightforward since −∞ N (z) dz = 2 +
 
1 √x .
2 erf 2

Lemma 8.2. The CDF of a GSaS is

1 1 ∞
Z x Z  
xs
(71) Φ[Lα,k ](x) := Lα,k (z) dz = + ds erf √ χα,k (s)
−∞ 2 2 0 2
Proof. We begin with the double integral via the PDF, such that
Z x
Φ[Lα,k ](x) = Lα,k (z) dz
−∞
Z x Z ∞
= dz s ds N (zs) χα,k (s)
−∞ 0
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 43

The order of the two integrals is switched, and we make a change of variable u = zs:
Z ∞ Z x 
Φ[Lα,k ](x) = s ds dzN (zs) χα,k (s)
0 −∞
Z ∞ Z xs 
= ds duN (u) χα,k (s)
0 −∞
The inner integral is replaced with the error function:
Z ∞   
1 1 xs
Φ[Lα,k ](x) = ds + erf √ χα,k (s)
0 2 2 2
1 1 ∞
Z  
xs
= + ds erf √ χα,k (s)
2 2 0 2
We arrive at the desired formula for the CDF. □

8.5. Fractional Hypergeometric Function. Due to the connection between the CDF
of Student’s t and the Gauss hypergeometric function 2 F1 (a, b; c; z) (defined in 15.2.1 of
DLMF[6]), we propose a framework to extend 2 F1 fractionally.
Definition 8.1 (Fractional hypergeometric function (FHF)). Let M (b, c; z) be the confluent
hypergeometric function, or called the Kummer function, defined in 13.2.2 of DLMF. (Note
we avoid using a because it is confusing to α.)
The fractional hypergeometric function is defined in a ratio-distribution style as
r Z
∞
xks2
 
k
(72) Mα,k (b, c; x) := s ds M b, c; χα,k (s)
2π 0 2
Note that x usually takes a negative value. This is due to the asymptotic behavior:
M (b, c; z) ∼ ez z b−c /Γ(b) as z → ∞ (13.2.23 of DLMF).

Lemma 8.3. The CDF of a GSaS can be expressed by the fractional hypergeometric function
as
1 3 x2
 
1 x
(73) Φ[Lα,k ](x) = + √ Mα,k , ;−
2 k 2 2 k
Proof. The error function can be expressed by the Kummer function as (13.6.7 of DLMF)
 
2x 1 3 2
erf(x) = √ M , ; −x
π 2 2
Substitute it to (71), we obtain (73). □
Lemma 8.4. The fractional hypergeometric function subsumes the Gauss hypergeometric
function at α = 1:
   
k 1 k+1
(74) M1,k (b, c; x) B , = 2 F1 b, ; c; x
2 2 2
where B(a, b) is the beta function (8.17.3 of DLMF). Notice that k plays the role of the second
parameter in 2 F1 .
44 S. H. LIHN

kk/2 2
Proof. Since χ1,k (s) = 2k/2−1 Γ(k/2)
sk−1 e−ks /2 , (74) becomes
∞   (k+1)/2
xks2
Z     
2 1 k+1 2 k+1
ds M b, c; sk e−ks /2 = Γ F
2 1 b, , c; x
0 2 2 2 k 2
We make two changes of variable selectively, a = (k + 1)/2 and z = k/2, such that
Z ∞
2 1
ds M b, c; xzs2 s2a−1 e−zs = Γ(a) z −a 2 F1 (b, a, c; x)


0 2
Lastly, let t = s2 , and x = q/z, we get
Z ∞
dt M (b, c; qt) ta−1 e−zt = Γ(a) z −a 2 F1 (b, a, c; q/z)
0
which is the Laplace transform of the Kummer function mentioned in 13.10.3 of DLMF.
Therefore, this Lemma is essentially a disguised Laplace transform.
□

It is then straightforward to obtain the known expression of the CDF of Student’s t as7
1 k + 1 3 x2
 
1 x 1
(75) Φ[L1,k ](x) = + √ F
k 1 2 1 2
, ; ; − (x2 < k)
2 k B( 2 , 2 ) 2 2 k
Next, the CDF of Student’s t can also be expressed in the regularized incomplete beta
function Ip (a, b), where 0 ≤ p ≤ 1 (See 8.17.2 in DLMF) such that
   
1 k 1 1 1 1 k
(76) Φ[L1,k ](x) = 1 − Ip , = + Iq , (x ≥ 0)
2 2 2 2 2 2 2
k
where p = 2 , q = 1 − p.
x +k
This is an application of the following equation (by combining 8.17.7 and 15.8.1 in DLMF):
 a  
1 q q
Iq (a, b) = F
2 1 a, a + b, a + 1; −
aB(a, b) p p
where we assign a = 1/2, b = k/2, and q/p = x2 /k.

8.6. GSaS Characteristic Function. We explore the CF of GSaS that was mentioned in
(29):
√ Z ∞
 
ζ
ϕα,k (ζ) = 2π ds N χα,k (s) (k > 0)
0 s
It forms a high transcendental function, fractionally extended such that it subsumes both the
α
stretched exponential
√ function e−ζ in (77), and the modified Bessel function of the second
kind 8 K k (ζ k) in (78) (See scipy.special.kn and (11.117) in [1]).
2

7See also Wikipedia at https://en.wikipedia.org/wiki/Student%27s_t-distribution#Cumulative_

distribution_function
8See also Wikipedia at https://en.wikipedia.org/wiki/Bessel_function#Modified_Bessel_functions.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 45

The first part is straightforward. Set k = 1, the LHS by definition of the α-stable distribu-
tion is
√ Z ∞
 
−ζ α ζ
(77) ϕα,1 (ζ) = e = 2π ds N χα,1 (s) (ζ ≥ 0)
0 s
Z ∞ √
 
ds ζ 
= N 2πs χα,1 (s)
0 s s
The second line
√ is essentially the same as the second line of (51). Note the term in the
parenthesis: 2πs χα,1 (s) ∝ Vα (s). And Vα (s) = χα,2 (s). We can see that, dimensionally
speaking, the sk term is correct.
The second part is based on the CF of Student’s t, ϕ1,k (ζ), such that
tk/2 √ Z ∞ √
 
ζ
(78) ϕ1,k (ζ) = k
K k (t) = 2π ds N χ1,k (s), where t = ζ k, ζ ≥ 0
2k/2−1 Γ( 2 ) 2 0 s
The complex normalization constant between ϕ1,k (ζ) and K k (t) is to ensure the general
2
requirement of a CF such that ϕα,k (0) = 1. The tk/2 term is due to the divergent asymptotic
of K k (t) at t = 0 (See (11.124) in [1]):
2

k
lim Kν (t) = 2ν−1 Γ(ν) t−ν , ν= >0
t→0 2
8.7. GSaS Series Representation of PDF and Tail Behavior. The PDF of GSaS (k >
0) has two known series representations. One in term of x and the other in terms of 1/x.
Both have issues in terms of radius of convergence. More future work is needed in this area.
Here we present the 1/x result to investigate the tail behavior. This is a working example
of the four-parameter Wright function (112). (The small x series is presented in Section A.1.)
Lemma 8.5. The series representation of GSaS PDF in terms of 1/x is
−α !
x −k+1
     α k
 
Sα,k 1 , x
(79) Lα,k (x) = √ W 2α 2 −
2 π x σ eα,k −2, 0 σ
eα,k
∞
"  −α #n
x −k X Γ αn k
  

Sα,k 2 + 2 x
= √ −
2 πσ eα,k σ
eα,k Γ(n + 1) Γ(− αn 2 ) σ
eα,k
n=1
√
αΓ( k−1
2 ) 2   1/2−1/α 
(80) where Sα,k := and σ α,k := = 2 k
Γ( k−1
e
α )
σα,k
This series works better for smaller α where the Pareto tails are more prominent.
Proof. From (61), we get
k−1
∞
s α
Z     
s
Lα,k (x) = Sα,k ds N (xs) W α
,0 −
0 σα,k 2 σα,k
Z ∞
= Sα,k σα,k dz N (xz σα,k ) z k−1 W α2 ,0 (−z α )
0
where z = s/σα,k .
46 S. H. LIHN

Substitute the Wright function with its series representation,

√  ∞
1 −(zx/eσα,k )2 X (−1)n z αn+k−1
Z ∞ 
2Sα,k
Lα,k (x) = dz √ e
σ
eα,k 0 2π n! Γ(− αn
2 )
n=1
Integrate term by term, each integral is just the moment formula of a normal distribution.
∞
(−1)n eα,k αn+k
  
Sα,k X 1 αn + k σ
Lα,k (x) = √ Γ
πe
σα,k 2 Γ(n + 1) Γ(− αn2 ) 2 x
n=1
Re-arrange the term inside the summation, we arrive at the second line of (79). Then replace
the summation with the four-parameter Wright function (112). □
Take the first term of the summation, we observe that the tail behavior of GSaS is propor-
tional to x−(α+k) . This illustrates the effect of k in which larger k makes the tail lighter.
To validate the Pareto tail behavior, when k → 1, it becomes αcα x−(1+α) , where cα =
Γ(α) sin( απ
2 )/π, consistent with (3.43) of Nolan (2020)[23].

9. GEP: Generalized Exponential Power Distribution

α
The exponential power distribution Eα (x) = e−|x| /2Γ α1 + 1 is mentioned in Definition

3.6. It has been used to model asset returns in finance, as shown in the LHS of (51). Its
usefulness stems from its simplicity. Unlike its α-stable counterpart, the existence of moments
is never an issue.
GEP Eα,k enriches Eα with the degree of freedom parameter k. We have shown that Eα,k =
Lα,−k by the clever design of FCM. Therefore its properties can be obtained indirectly from
those of GSaS. We still carry out some proofs here based on its product distribution definition
(30), for completeness sake as well as for validation.
Its peak PDF is simply Eα,k (0) = √12π /E(X|χα,k ). It can be standardized by multiplying
with its standard deviation, which is derived below.
All the moments of Eα,k exist. The reader can take a closer look at Figure 6. All the kurtosis
contours are continuous on the half plane of k > 0. Its moment formula is straightforward:
Lemma 9.1. The n-th moment of Eα,k is
E(X n |N ) E(X n+1 |χα,k )
(81) E(X n |Eα,k ) =
E(X|χα,k )
where n is even. All the odd moments are zero.
Proof. The integral of the n-th moment from (30) is
Z ∞ Z ∞
n 1 n
x
E(X |Eα,k ) = x dx ds N χα,k (s)
E(X|χα,k ) −∞ 0 s
With a change of variable z = x/s, we can separate the two integrals:
Z ∞ Z ∞
n 1 n
E(X |Eα,k ) = z N (z) dz sn+1 χα,k (s) ds
E(X|χα,k ) −∞ 0
E(X n |N ) E(X n+1 |χα,k )
=
E(X|χα,k )
□
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 47

The excess kurtosis of Eα,k in (34) is derived from E(X 4 |Eα,k )/E(X 2 |Eα,k )2 .

The CDF of Eα,k is derived below by following the same logic in Section 8.4.
Lemma 9.2. The CDF of Eα,k is
Z x Z ∞  
1 1 x
(82) Φ[Eα,k ](x) := Eα,k (z) dz = + s ds erf √ χα,k (s)
−∞ 2 2E(X|χα,k ) 0 2s
Proof. We begin with the double integral via the PDF, such that
Z x Z ∞
1 z 
Φ[Eα,k ](x) = dz ds N χα,k (s)
E(X|χα,k ) −∞ 0 s
The order of the two integrals are switched, and we make a change of variable u = z/s:
Z ∞ Z x  z 
1
Φ[Eα,k ](x) = ds dz N χα,k (s)
E(X|χα,k ) 0 −∞ s
Z ∞ Z x/s !
1
= s ds du N (u) χα,k (s)
E(X|χα,k ) 0 −∞

The inner integral is replaced with the error function:

Z ∞   
1 1 1 x
Φ[Eα,k ](x) = s ds + erf √ χα,k (s)
E(X|χα,k ) 0 2 2 2s
Z ∞  
1 1 x
= + s ds erf √ χα,k (s)
2 2E(X|χα,k ) 0 2s
We arrive at the desired formula for the CDF. □

10. GAS: Generalized Alpha-Stable Distribution

The invention of GAS is experimental. Since GSaS is very successful, it is a natural next
step to add skewness to it.
We take a simple approach – The separation of concerns between the skew Gaussian kernel
gαθ (x, s) and fractional chi-1 in Lθα is extended along the FCM side, as shown below from the
first line to the second line:
Z ∞
θ θ
Lα (x) = Lα,1 (x) = s ds gαθ (x, s) χα,1 (s)
Z0 ∞
θ
(83) =⇒ Lα,k (x) := s ds gαθ (x, s) χα,k (s)
0

The skewness of GAS comes from gαθ (x, s)

when θ ̸= 0. When θ = 0, gαθ (x, s) = N (xs) and
GAS becomes GSaS.
Analyzing the behavior of skewness in depth when |k| > 1 is a complicated topic. θ is
confined by Lθα,k (x) > 0 for all x ∈ R. When |θ| is small, the skew PDF behaves well.
But the valid range of θ appears to be much smaller than the Feller-Takayasu diamond :
θ ≤ min{α, 2 − α} [20, 19]. The larger k is, the smaller range of θ. This issue is left for future
research.
48 S. H. LIHN

Lemma 10.1. (GAS Symmetry Relation) The following symmetry relation holds, which is
an extension from α-stable (See Section 4 in [20]):
Lθα,k (−x) = L−θ
α,k (x)

Therefore, one only needs to consider the case of x > 0.

Proof. Both α, θ only show up in the skew kernel, not in FCM. We just need to prove the
relation: gαθ (−x, s) = gα−θ (x, s).
This is easy to show since the cosine function is symmetric and the tangent
 function is 
asymmetric. In (26), when x and τ change sign simultaneously, the term cos τ (st)α + xq st
remains the same.
□
Two subjects are unresolved for GAS as of this writing. First, the analytic form for the
moments of GAS is wanting. Second, the k extension of a well-known reciprocal relation
between α and 1/α is unknown yet. When k = 1, Feller first showed
1 ∗
Lθ1/α (x−α ) = Lθα (x), θ∗ = α(θ + 1) − 1,
xα+1
where 1/2 ≤ α < 1 and x > 0 (See Section 4 in [20] and Lemma 2 in p. 583 of [9]). What is
the relation when k > 1?

10.1. GAS PDF at Small Alpha. When α is small and θ ̸= 0, the GAS integral (83) can
be difficult to converge. Its behavior can be better understood by splitting the skew kernel in
the integral into two parts: (a) The Gaussian kernel for small s; (b) The tail kernel for large
s. The integral over the former is much easier to converge than the later.
Let gαθ (x, s) = n(x, s, q) + hθα (x, s), where
Z ∞  
1 x 2
n(x, s, q) = dt cos st e−t /2 (s ≥ 0)
qπ 0 q
and
Z ∞     
1 α x x 2
(84) hθα (x, s) = dt cos τ (st) + st − cos st e−t /2 (s ≥ 0)
qπ 0 q q
Note that n(x, s, q) can be simplified by the Gaussian integral to
   
1 0 x 1 xs
n(x, s, q) = gα ,s = N
q q q q
Hence, (83) can be rewritten as
1 ∞
Z   Z ∞
θ xs
(85) Lα,k (x) = s ds N χα,k (s) + s ds hθα (x, s) χα,k (s)
q 0 q 0

The first integral converges quickly over several multiples of q/x (assuming x ̸= 0), and
most weight of Lθα,k (x) can be obtained from it.
However, the second integral can be hard to converge. It gets its contribution when τ ̸= 0
and s is away from zero. The tails of both hθα (x, s) and χα,k (s) can decay very slowly when
α is small. The larger |τ | is, the slower the decay.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 49

Empirically, at α = 0.1, the integral over s may have to extend to 103 in order to
achieve 10−2 relative tolerance. And calculating hθα (x, s) (and/or gαθ (x, s)) at large s is time-
consuming.

11. Generation of Random Variables for GSC, FCM, and GSaS

In this section, we address the issue of generating random numbers for a GSC. An es-
tablished stochastic framework has been written on the Wikipedia page9 of the stable count
distribution[11]. That framework is formalized and enlarged to GSC here. If one can generate
random numbers {St } for GSC, then it is straightforward to simulate random numbers {Xt }
for GSaS by a ratio distribution: Xt = N /St , where N is a standard normal variable.
Given a user-specific volatility σu describing how fast St should change, and a scale pa-
rameter θu (default to 1), we assume the random variable St should evolve by the following
generalized Feller square-root process[7]:
 
St p
(86) dSt = σu2 µ dt + σu St dWt
θu
As t → ∞, {St } will distribute like the equilibrium distribution that µ(x) is designated for.
11.1. Generation of Random Variables for GSC.
Lemma 11.1. The µ(x) solution is derived from the Fokker-Planck equation with Nα (x; σ, d, p)
as the equilibrium distribution, which yields (assume θu = 1)
d
(x dx + 1) Nα (x; σ, d, p)
(87) µ(x) =
2 Nα (x; σ, d, p)
x d 1
= [log Nα (x; σ, d, p)] +
2 dx 2
The second line is elegant, but only useful in limited cases.
Proof. As t → ∞, the stationary Fokker-Planck equation of (86) is
d  2  1 d2  2 
σu µ (x) Nα (x; σ, d, p) = σ u x N α (x; σ, d, p)
dx 2 dx2
2
R∞
First, σu cancels out from both sides and is irrelevant to the solution. Second, apply x dx
to both sides. And we know the constant at x = ∞ should be zero, because Nα (x; σ, d, p) is
the PDF of a distribution. Hence,
1 d
µ (x) Nα (x; σ, d, p) = [x Nα (x; σ, d, p)]
2 dx
Move Nα (x; σ, d, p) from LHS to RHS, we obtain (87). The second line is from the simple
d 1 d
application of dx (log f (x)) = f (x) dx f (x).
□
Note that this model subsumes the renown Cox–Ingersoll–Ross (CIR) model[4] because
CIR’s equilibrium distribution is a gamma distribution which is subsumed by GSC at α =
0, p = 1.
9See https://en.wikipedia.org/wiki/Stable_count_distribution#Generation_of_Random_Variables
50 S. H. LIHN

To simplify the symbology going forward, we define

W−α,−1 (−z)
(88) Qα (z) := −
W−α,0 (−z)
Lemma 11.2. The mean-reverting solution of a GSC in terms of the Wright functions is

p

p W−α,−1 − σx

d p
(89) µ(x) = −

+ − α ∈ (0, 1)
2α W−α,0 − σx p 2 2α
p  x p   d p

= Qα + −
2α σ 2 2α

p
Proof. From (87), we make a change of variable, z = σx , and d′ = (d − 1)/p + 1. Then
Nα (x; σ, d, p) = Nα (z; σ = 1, d′ , p = 1)
d d
The change of variable also yields pz dz = x dx , which simplifies µ(x) to
d
(pz dz + 1) Nα (z; σ = 1, d′ , p = 1)
(90) µ(x) =
2 Nα (z; σ = 1, d′ , p = 1)
The critical algebra is
d
pz Nα (z; σ = 1, d′ , p = 1)
dz  
′ d′ −1 d′ −1 d
= Cp(d − 1)z W−α,0 (−z) + Cp(z ) z W−α,0 (−z)
dz
 
d′ −1 ′ p
= Cz p(d − 1) W−α,0 (−z) + (W−α,−1 (−z) + W−α,0 (−z))
−α
p ′
 p  ′
= − Cz d −1 W−α,−1 (−z) + pd′ − − p Cz d −1 W−α,0 (−z)
α α
We use the following recurrence relation:
d
(91) −αz
W−α,0 (−z) = W−α,−1 (−z) + W−α,0 (−z)
dz
to move from the second line to the third line above. (To derive it, set λ = −α, µ = 0 in (110)
and (111)).
The µ(x) solution becomes
  1
1 −p W−α,−1 (−z)  ′ p
µ(x) = + pd − − p +
2 α W−α,0 (−z) α 2
p
p W−α,−1 − σx


 
d p
=− x

p
+ −
2α W−α,0 − σ 2 2α
which is the intended solution.
□

The next question is how to compute Qα (z). In addition to going through the series
representation of W−α,−1 (−z), the following lemma addresses such issue.
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 51

Lemma 11.3. There are two methods to compute Qα (z) depending on the knowledge of
Fα (z) or Mα (z). They are
d
W−α,−1 (−z) αz dz Fα (z)
(92) Qα (z) = − =1+
W−α,0 (−z) Fα (z)
d
αz dz Mα (z)
(93) = (α + 1) +
Mα (z)
The second line (93) allows for high-precision implementation through Mα (z) and its first
derivative (119), both of which have nice analytic properties.
Proof. Note that (91) is the same as
 
d
W−α,−1 (−z) = − αz + 1 W−α,0 (−z)
dz
Divide both sides by Fα (z) = W−α,0 (−z), we arrive at (92).
To move from (92) to (93), simply replace Fα (z) by αz Mα (z), and expand the derivative
into it.
□
To validate µ(x), first we note that,
√ by setting σ = 1, d = 1, p = α, it is reduced to that of
a SC. Secondly, by setting σ = 1/ 2, d = 1, p/α = 2 and α replaced with α/2, it is reduced
to that of a SV. Both are known solutions on the Wikipedia page[11].
Alternatively, when the closed form of Nα (z; σ = 1, d′ , p = 1) in (90) is known, it is better
to use the log approach to obtain µ(x):

pz d   1
(94) µ(x) = log Nα (z; σ = 1, d′ , p = 1) +
2 dz 2
′ 2 /4
For instance, when α = 1/2, Nα (z; σ = 1, d′ , p = 1) = 2√
C
π
z d e−z . We have the closed-
form solution:
p  x 2p d + p
(95) µα= 1 (x) = − +
2 4 σ 2
which is reduced to the known results for SC and SV:
 1
(96) µα= 1 (x) = 8 (6 −2 x), for SC;
x
2 1− 2 , for SV.
Both agree with what have been derived on the Wikipedia page[11].
11.2. Generation of Random Variables for FCM. Obviously, what really matters for
GSaS is the solution of FCM, denoted as µα,k (x). Note that from this point on, α ∈ (0, 2).
To further simplify the symbology for FCM, define
Qα(χ) (z) = Q α2 (z α )
Assume k > 0, we set σ = σα,k , d = k − 1, p/α = 2 and α replaced with α/2 in (89). We get
   
(χ) x k−3
(97) µα,k (x) = Qα +
σα,k 2
52 S. H. LIHN

For validation, µ1,k (x) = k(1 − x2 )/2 can be used to simulate Student’s t. And µα,1 (x)
provides a method to simulate an SaS Lα,1 (x):
√ 
µα,1 (x) = Qα(χ) 2x − 1

Fig. 11 shows a simulation of random variables based on the (α, k) parameters obtained
from the fit of the S&P 500 daily log returns. The rest of parameters are in the caption of
the figure. First, as outlined above, µα,k (s) is calculated analytically, as shown in the right
chart. Second, it enables the GSC simulation {St } as shown in the left chart. Third, GSaS
{Xt } is simulated via Xt = N /St , where N is drawn from a standard normal variable.
Simulation is performed daily. Sampling duration is 200,000 years. The red areas are the
histograms from simulated data. The blue lines are from theoretical formulas. They match
nicely.

Figure 11. Simulation of random variables based on the (α, k) parameters

obtained from the fit of the S&P 500 daily log returns. The red areas are the
histograms from simulated data. The blue lines are from theoretical formulas.
The settings of the simulation are α = 0.813, k = 3.292, dt = 1/365, σu =
0.85. Sampling duration is 200,000 years. The simulation takes 11 minutes in
python. µα,k (s) is discretized to 0.01 and cached to increase performance.

11.3. Generation of Random Variables for Inverse FCM. The simulation of GEP is
complicated by two options: ratio vs product. If we go with the ratio distribution, we have
to deal evaluating Fα (z) at very large z, which is technically more difficult.
To simulate Eα,k , we set σ = 1/σα,k , d = −k, p/α = −2 and α replaced with α/2 in (89),
we get
   k

(χ) −1
(98) µα,−k (x) = −Qα (x σα,k ) + 1− (k > 0)
2

The x−1 term is the added complication.

 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 53

To avoid such complication, it is more straightforward to take the product distribution

route, that is, Xt = St N . And St is generated from an inverse FCM:
   
† (χ) x k
(99) µα,−k (x) = Qα + −1 (k > 0)
σα,k 2

And µ†α,−1 (x) provides a method to simulate an exponential power distribution Eα :

√  1
µ†α,−1 (x) = Q(χ)
α 2x −
2
As is expected, this is identical to µ(x) of SV.
The simplest validation is µ†1,−1 (x) = 1 − x2 which can simulate an exponential distribution
2

E1 . On the other hand, to simulate E1 by a ratio distribution, the polynomial solution is

µ1,−1 (x) = 1/(2x2 ) − 1. The divergence at x = 0 is a very different behavior.
In summary, µα,k (x) is an amazing function that generates not only χα,k but also Lα,k . In
some cases, they are just simple polynomials. This is very impressive.

12. Acknowledgement
I would like to thank Professor John Mulvey at Princeton University for his guidance and
unwavering support. It took many years and several stages of advancements to get to this
point since we first discussed the heavy-tail distribution in 2017. Thank Professor Francesco
Mainardi at the University of Bologna, although I have not met him personally, for his decades-
long work on the Wright function. Thank Professor Petter Kolm at New York University for
introducing the concept of multivariate elliptical distribution.
My wonderful son, Zachary A. Lihn10 has provided valuable technical support for writing
this paper. I also thank him for several stimulating discussions.

Appendix A. GSC Moment Generation Function and Implementation

A.1. From GSC MGF to GSaS series representation. Since we have the analytic so-
lution of GSC moments E(X n ) from (48), the moment generating function (MGF) can be
defined from the sum of its moments:
∞

X E(X n )
MGSC (t) := E etX =
n!
n=0
∞
Γ( dp α) X n
Γ( n+d
p )
= (σt)
Γ( dp ) n=0
Γ(n + 1)Γ( n+d
p α)
Γ( dp α)
" #
1 d
p, p
= W α αd (σt)
Γ( dp ) p, p

Here we utilize the four-parameter Wright function defined in (112). The characteristic func-
tion of GSC is simply ϕGSC (z) = MGSC (iz).

10Zachary A. Lihn. Department of Mathematics, Columbia University, New York, NY 10027, USA.
[email protected]
54 S. H. LIHN

We would like to present an extraordinary approach to derive the small x series represen-
tation of GSaS PDF using the MGF of GSC. The MGF formula above can be transformed to
a ratio distribution with a normal variable, which resembles how GSaS is constructed in (28).
We show that GSC has an inherent closeness that can take a Laplace transform-like integral
to a Gaussian mixture.
Lemma A.1. The series representation of GSaS PDF in terms of x is
" #  !
Γ( k−1 2
, αk x 2

1 2 ) α
(100) Lα,k (x) = √ W −
eα,k π Γ( k−1
σ α )
1, k2 σ
eα,k

where W [...](x) is the four-parameter Wright function (112).

Proof. Equate the series form of MGF with its integral form,
Z ∞
MGSC (t) = dx etx Nα (x; σ, d, p)
0
Γ( dp α)
" #
1 d
p, p
= W α αd (σt)
Γ( dp ) p, p

Replace t with −σt2 /2 and x = s2 /σ, we get the Gaussian mixture:

 Z ∞
σt2
  
−(ts)2 /2 2 2
MGSC − = s ds e Nα (s /σ; σ, d, p)
2 0 σ
Γ( dp α)
" #
1 d
(σt)2

p, p
= W α αd −
Γ( dp ) p, p 2

Note that GSC is close under such transform:

" αd
# 2d−1

|p| Γ( ) Γ( )
Nα (s2 /σ; σ, d, p) =
p  σ 2p  Nα (s; σ, 2d − 1, 2p)
σ Γ( dp ) |2p| Γ( α(2d−1) )
2p
 
αd d−1/2
1 Γ( p ) Γ( p ) 
= Nα (s; σ, 2d − 1, 2p)
2 Γ( dp ) Γ( α(d−1/2) )
p

Now we can replace t with x, and define the new PDF as:
Z ∞
f (x) = s ds N (xs) Nα (s; σ, 2d − 1, 2p)
0
α(d−1/2) "
1 d
#
σ Γ( ) (σx)2

p p, p
=√ W α αd −
2π Γ( d−1/2
p ) p, p 2

We are in a good place to derive the series representation of GSaS in terms of x, which is
skipped in Section 8.7. But it takes some tongue-twisting substitutions to accomplish.
To create χα,k (k > 0) in the integrand of f (x) above, both α and p become α/2, thus α/p =
√
1. And d = k/2, σ = σα,k = 2/e σα,k from (80). With these substitutions, Nα (s; σ, 2d − 1, 2p)
becomes χα,k (s); and f (x) is transformed to Lα,k (x), which is (100). □
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 55

To validate, note that W [...](0) = Γ( αk )/Γ( k2 ). It is straightforward to see that Lα,k (0) in
(100) is the same as Lα,k (0) in (66).

A.2. PDF Implementation of GSC and Large-x Asymptotic. The PDF of GSC could
be converted to the PDF of SC, which is in turn implemented by the scipy.stats.levy stable
package. The direct computation on the Wright function can be safely bypassed. We have
 x d−1  x p/α
Nα (x; σ, d, p) = C W−α,0 (−ν α ) , where ν =
σ      σ
1 x d−1 x p/α
=CΓ +1 Nα
α σ σ
which is due to (49).
A naive numeric implementation has its limitation, especially it can lose numerical precision
quickly when d is large (e.g. d > 60). Since we often want to check the behavior of certain
scenarios involving d → ∞, it is highly desirable the implementation can handle such scenario.
For the asymptotic at large x, especially when d is also large, Nα (...) could be replaced
with its asymptotic from Section 4.4.2 of [17]:
α α
Nα (ν) ≈ B(α) ν 2(1−α) e−A(α) ν
1−α
, ν≫1
1
α B ′ (α) α 2(1−α)
where A(α) = (1 − α) α 1−α , B(α) = 1

, B ′ (α) = p
Γ α +1 2π(1 − α)
However, this formula doesn’t work for small α because of another competing limit per Section
5.3. This is embedded in the fact that A(α) and B(α) are undefined at α = 0.
Therefore, as long as α is not too small, GSC at large x becomes a GG-variant:
   α
1 x d−1 2(1−α)
α
 x p/α
e−A(α) ν
1−α
Nα (x; σ, d, p) ≈ B(α)C Γ +1 ν , ν=
α σ σ
 x d+ p −1  x p 
= B ′ (α)C
2(1−α) 1−α
(101) exp −A(α)
σ σ

A.3. Large-x Asymptotic of FCM. Large-x Asymptotic of FCM follows naturally from
(18) and (101). For an FCM,

α Γ( k−12 ) |k|1/2−1/α
C= , where σα,k := √
σα,k Γ( k−1
α ) 2
Therefore, at large x, an FCM of k > 0 becomes
α 2α !
α  k+ −2  α   x  2−α
x 2−α
(102) χα,k (x) ≈ B ′ C exp −A
2 σα,k 2 σα,k

This formula provides insight to the tail behavior of the FCM.

As a validation, the asymptotic formula becomes exact at α = 1, where χα,k becomes
√
χk / k.
56 S. H. LIHN

Appendix B. Subordination Results

Lemma B.1. Assume f (x), g(x), N(x) are the PDFs from three distributions. f (x), g(x) are
two-sided; and N(x) is one-sided. They form the following pair of transform:
Z ∞
(103) f (x) = s ds g(xs) N(s) (x ∈ R)
0
Z ∞   
ζ
(104) CF{f }(ζ) = ds CF{g} N(s) (ζ ∈ R)
0 s
Z ∞  
−3 1
(105) = s ds [CF{g} (ζs)] Nϕ (s), where Nϕ (s) := s N
0 s
This is an extended form of Lemma 2.2.
R∞
Proof. For a given distribution f (x), its CF transform is CF{f }(ζ) := −∞ dx exp(iζx) f (x).
R∞
And for g(x), its CF transform is CF{g}(ζ) := −∞ dx exp(iζx) g(x). Apply them to (103),
we get
Z ∞ Z ∞ 
(106) CF{f }(ζ) = s ds dx exp(iζx) g(xs) N(s)
0 −∞
Let u = xs, the integral inside the parenthesis becomes
Z ∞ Z ∞  
−1 ζu
dx exp (iζx) g(xs) = s du exp i g(u)
−∞ −∞ s
 
ζ
= s−1 CF{g}
s
Put it back to (106), we arrive at (104).
To move from the product distribution form of (104) to the ratio distribution form of (105),
let t = 1/s, then ds/s = −dt/t. Furthermore, s ds = −t−3 dt. It is straightforward to obtain
(105).
□
The continuous Gaussian mixture in Section
√ 2 is a special case where g(x) = N (x). The
CF of N (x) is still itself: CF{N }(ζ) = 2π N (ζ). Another case we encountered is the
Laplace-Cauchy pair: CF{L}(ζ) = π Cauchy(ζ).
If g(x) and CF{g}(ζ) are well understood functions, this lemma captures many relations
between f (x) and N(x), for instance, when N(x) is a high transcendental function such as
GSC.
When N(x) is a GSC, it is easy to see that Nϕ (s) is also a GSC. Therefore, GSC is closed
under the CF transform: N(x) → Nϕ (s).
We studied many of the product and ratio distributions and tabulated the results as fol-
lowing.
B.1. Product Distribution Results. Table 4 below illustrates how a classic distribution
on the left can be decomposed to a GSC on the right via the product distribution relation:
Z ∞
dt x
(107) Left(x) := Unit( ) Right(t)
0 t t
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 57

The discovery of Table 4 directly paved the way to generalize the stable count distribution.

Right GSC Equiv.: Nα (t; σ, d, p)

Left Dist. (PDF) Unit Dist. Right Dist. α σ d p
Exp Pow: Eα (z) Laplace: L( |z|
t
) SC: Nα (t) α 1 1 α
Exp Pow: Eα (z) Normal: N (z/t) SV: Vα (t) α/2 √1 1 α
2

Exp Pow: Eα (z) Ek (z/t), any k ∈ Z GSC α/k 1 1 α

Weibull: Wb(x; k) L( xt ) aka Wb( xt ); 1) GSC: t−1 Nk (t) k 1 0 k
−1
Weibull: Wb(x; k) Rayleigh( xt ) χk,1 : t Vk (t) k/2 √1
2
0 k
aka Wb( xt ); 2)
Gamma: Γ(x; s) Wb( xt ; s) GSC 1/s 1 s 1
Poisson: f (k; λ) Wb( λt ; k + 1) GSC 1/(k + 1) 1 k+1 1
= Γ(λ; s = ⌊k + 1⌋)
χ2k : Γ( x2 ; k2 ) Wb( xt ; k2 ) GSC 2/k 2 k/2 1
2 √
χk : Γ( x2 ; k
2
) Wb( xt ; k) GSC 2/k 2 k 2
GenGamma( σx ; s, c) Wb( xt ; sc) GSC 1/s σ sc c

Table 4. Subordination by a product distribution to GSC.

B.2. Ratio Distribution Results. Table 5 below illustrates how a Student’s t and/or SaS
distribution on the left can be decomposed to a GSC on the right via a ratio distribution:

Z ∞
(108) Left(x) := s ds Unit(sx) Right(s)
0

We’ve also included the new GSaS from this work. This table shows that we should be
flexible in choosing either the path of product or ratio distributions. As Lemma B.1 shows,
they are really the flip side of the same coin.
58 S. H. LIHN

Right GSC: Nα (s; σ, d, p)

Left Dist. (PDF) Unit Dist. Right Dist. α σ d p
IG(x; k) IWb(sx; k) GSC 1/k 1 k k
IWb(x; α) IWb(sx; k), GSC α/k 1 0 α
any k ∈ Z
√
Student’s t: tk (x) Normal: N (sx) χ1,k : sk−2 V1 ( ks) 1/2 √1
2k
k−1 1
−1
SaS: Pα (x) Cauchy(sx) GSC: s Nα (s) α 1 0 α
−1 √1
SaS: Pα (x) Normal: N (sx) χα,1 : s Vα (s) α/2 2
0 α
k−2
√
GSaS: Lα,k (x) Normal: N (sx) χα,k : s Vα ( ks) α/2 √1 k−1 α
2k
−1 +
ML: Eα (−x) Expon: exp(−sx) s Mα (s) α 1 −1 1
2 2
ML: E α2 (− x2 ) Normal: N (sx) GSC: M α2 (s ) α/2 1 -1 2

Table 5. Subordination by a ratio distribution to GSC. Note on the Mittag-

Leffler (ML) relations: The first relation is a Laplace transform, neither side
is normalized properly. The LHS of the second relation needs a normalization
constant of Γ(1−α/4)
√
2π
.

Appendix C. Useful Results on the Wright Function

This section collects useful results of the Wright function that are used in this paper. It is
expanded from Section 3.1.
The moments of the Wright function are (See (1.4.28) of [21])
Z ∞
Γ(d)
(109) E(X d−1 ) = xd−1 W−λ,δ (−x)dx =
0 Γ(dλ + δ)
The recurrence relations of the Wright function are (Chapter 18, Vol 3 of [2])
(110) λz Wλ,λ+µ (z) = Wλ,µ−1 (z) + (1 − µ)Wλ,µ (z)
d
(111) Wλ,µ (z) = Wλ,λ+µ (z)
dz
The four-parameter Wright function is defined as

∞
zn
 
a, b X Γ(an + b)
(112) W (z) :=
λ, µ Γ(n + 1) Γ(λn + µ)
n=0
As far as I know, this function is used seriously for the first time in this work.
C.1. The M-Wright Functions. Mainardi introduced two auxiliary functions of the Wright
type (See F.2 of [18]):
(113) Fα (z) := W−α,0 (−z) (z > 0)
1
(114) Mα (z) := W−α,1−α (−z) = Fα (z) (z > 0)
αz
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 59

The relation between Mα (z) and Fα (z) in (114) is an application of (110) by setting λ =
−α, µ = 1.
Fα (z) has the following Hankel integral representation:
Z
1
(115) Fα (z) = W−α,0 (−z) = dt exp (t − ztα )
2πi H
Fα (z) is used to define GSC. But its series representation isn’t very useful computationally.
It requires a lot more terms to converge to a prescribed precision.
On the other hand, Mα (z) has a more computationally friendly series representation, espe-
cially for small α’s:
∞ ∞
X (−z)n 1 X (−z)n−1
(116) Mα (z) = = Γ(αn) sin(αnπ) (0 < α < 1)
n! Γ(−αn + (1 − α)) π (n − 1)!
n=0 n=1

Mα (z) also has very nice analytic properties at α = 0, 1/2, where M0 (z) = exp(−z) and
M 1 (z) = √1π exp(−z 2 /4).
2
Mα (z) can be computed to high accuracy when properly implemented with arbitrary-
precision floating-point library, such as the mpmath package. In this regard, it is much more
”useful” than Fα (z). Lα (x) (therefore, Nα (x)) can be computed via Lα (x) = αx−α−1 Mα (x−α )
without relying on the scipy.stats.levy stable package.
This is particularly important in working with large degrees of freedom and extreme values
of α, mainly at 0 and 1. Typical 64-bit floating-point is quickly overflowed.
Mα (z) has the asymptotic representation in GG-style: (See F.20 of [18])
x p
(117) Mα = A xd−1 e−B x
α p
where p = 1/(1 − α), d = p/2, A = p/(2π), B = 1/(αp).

This formula is important in guiding the series sum to high precision.R

∞
Treating Mα (x) as the PDF of a one-sided distribution such that 0 Mα (x)dx = 1, then
its n-th moment is (See Section 4 of [19])
Z ∞
n Γ(n + 1)
(118) E(X |Mα ) = xn Mα (x)dx =
0 Γ(nα + 1)
Differentiating Mα (z), and from (116), we get
∞
d −1 X (−z)n−2
(119) Mα (z) = −W−α,1−2α (−z) = Γ(αn) sin(αnπ)
dz π (n − 2)!
n=2

d
Note that dz Mα (0) = − π1 Γ(2α) sin(2απ). This also indicates that
 
d d
(120) Fα (z) = α 1 + z Mα (z)
dz dz
which can be implemented from Mα (z) through (116) and (119). These results are important
for the implementation of Qα (z) in (93).
60 S. H. LIHN

Appendix D. List of Useful Formula

D.1. Gamma
√ Function. Gamma function is used extensively in this paper. First, note that
Γ( 12 ) = π. Its reflection formulas are:
π
(121) Γ(z)Γ(1 − z) =
sin(πz)
1 √
(122) Γ(z)Γ(z + ) = 21−2z π Γ(2z)
2
Gamma function Asymptotic: At x → 0, gamma function becomes
1
(123) lim Γ(x) ∼
x→0 x
Γ(ax) b
lim = (ab ̸= 0)
x→0 Γ(bx) a
For a very large x, assume a, b are finite,
Γ(x + a)
(124) lim ∼ xa−b
x→∞ Γ(x + b)

Sterling’s formula is used to expand the kurtosis formula for a large k, which is:
√  x x
(125) lim Γ(x + 1) ∼ 2πx ,
x→∞
√ e
(126) or lim Γ(x) ∼ 2π xx−1/2 e−x .
x→∞

D.2. Transformation. Laplace transform of cosine is11

Z ∞
ν −1 ν
(127) dt cos(xt)e−t/ν = 2 −2
=
0 x + ν (νx)2 + 1
Gaussian transform of cosine is12
Z ∞ r
−t2 /2 π −x2 /2
(128) dt cos(xt) e = e
0 2
Z ∞ r
2 2 π −(sx)2 /2
Hence dt cos(xt) e−t /2s = se
0 2
D.3. Half-Normal Distribution. The moments of the half-normal distribution (HN)13 are
used several times. Its PDF is defined as
r
2 1 −x2 /(2σ2 )
(129) pHN (x; σ) := e , x>0
πσ
√
which is a special case of GG with d = 1, p = 2, a = 2σ. Its moments are
2n/2
 
n+1
(130) EHN (T n ) = σ n √ Γ
π 2
which are the same as those of a normal distribution.
11See https://proofwiki.org/wiki/Laplace_Transform_of_Cosine
12See https://www.wolframalpha.com/input?i=integrate+cos%28a+x%29+e%5E%28-x%5E2%2F2%29+dx+
from+0+to+infty
13See https://en.wikipedia.org/wiki/Half-normal_distribution
 GENERALIZATION OF THE ALPHA-STABLE DISTRIBUTION WITH THE DEGREE OF FREEDOM 61

References
[1] George Arfken. Mathematical Methods for Physicists. Academic Press, Inc., third edition, 1985.
[2] H. Bateman and A. Erdélyi. Higher Transcendental Functions, volume 3 of California Institute of Tech-
nology Bateman Manuscript Project. McGraw-Hill Book Company., 1955.
[3] Eli Bingham, Jonathan P. Chen, Martin Jankowiak, Fritz Obermeyer, Neeraj Pradhan, Theofanis Karalet-
sos, Rohit Singh, Paul Szerlip, Paul Horsfall, and Noah D. Goodman. Pyro: Deep Universal Probabilistic
Programming. Journal of Machine Learning Research, 2018. See https://arxiv.org/abs/1810.09538.
[4] John C. Cox, Jonathan E. Ingersoll, and Stephen A. Ross. A theory of the term structure of interest rates.
Econometrica, 53(2):385–407, 1985.
[5] J.C. Crisanto-Neto, M.G.E. da Luz, E.P. Raposo, and G.M. Viswanathan. An efficient series approximation
for the lévy α-stable symmetric distribution. Physics Letters A, 382(35):2408–2413, 2018.
[6] NIST Digital Library of Mathematical Functions. https://dlmf.nist.gov/, Release 1.2.0 of 2024-03-15.
F. W. J. Olver, A. B. Olde Daalhuis, D. W. Lozier, B. I. Schneider, R. F. Boisvert, C. W. Clark, B. R.
Miller, B. V. Saunders, H. S. Cohl, and M. A. McClain, eds.
[7] William Feller. Two singular diffusion problems. Annals of Mathematics, 54(1):173–182, 1951.
[8] William Feller. On a generalization of marcel riesz’ potentials and the semi-groups generated by them.
Meddelanden Lunds Universitets Matematiska Seminarium, pages 73–81, 1952.
[9] William Feller. An Introduction to Probability Theory and its Applications (2nd Edition), volume 2. John
Wiley & Sons, Inc., 1971.
[10] Wikimedia Foundation. Inverse distribution, 2023. Retrieved December 1, 2023, from https://en.
wikipedia.org/wiki/Inverse_distribution.
[11] Wikimedia Foundation. Stable count distribution, 2023. Retrieved December 1, 2023, from https://en.
wikipedia.org/wiki/Stable_count_distribution.
[12] F. R. Helmert. Die Ausgleichsrechnung nach der Methode der kleinsten Quadrate (Adjustment Calculus
by the Method of Least Squares). Teubner (Leipzig), 1872. See https://eudml.org/doc/203764.
[13] Paul Lévy. Calcul des Probabilités. Gauthier-Villars, 1925.
[14] GNU Scientific Library. Exponential power distribution, 2021. Available at https://www.gnu.
org/software/gsl/doc/html/randist.html#the-exponential-power-distribution. Accessed: 09-
December-2023.
[15] Stephen H.T. Lihn. Hidden Markov model for financial time series and its application to S&P 500 index.
SSRN Working Paper, 6 2017. Available at SSRN: https://ssrn.com/abstract=2979516.
[16] Stephen H.T. Lihn. A theory of asset return and volatility under stable law and stable lambda distribution.
SSRN Working Paper, 11 2017. Available at SSRN: https://ssrn.com/abstract=3046732.
[17] Stephen H.T. Lihn. Stable count distribution for the volatility indices and space-time generalized stable
characteristic function. SSRN Working Paper, 7 2020. Available at SSRN: https://ssrn.com/abstract=
3659383.
[18] F. Mainardi. Fractional Calculus and Waves in Linear Viscoelasticity: An Introduction to Mathemati-
cal Models. Imperial College Press, 2010. See Appendix F at https://appliedmath.brown.edu/sites/
default/files/fractional/36%20TheWrightFunctions.pdf.
[19] Francesco Mainardi and Armando Consiglio. The wright functions of the second kind in mathematical
physics. Mathematics, 8(6), 2020.
[20] Francesco Mainardi, Yuri Luchko, and Gianni Pagnini. The fundamental solution of the space-time frac-
tional diffusion equation. Frac. Calc. Appl. Anal., 4, 03 2007. See https://arxiv.org/abs/cond-mat/
0702419.
[21] Arak Mathai and Hans Haubold. Fractional and Multivariable Calculus: Model Building and Optimization
Problems. Springer Cham, 12 2017.
[22] The mpmath development team. mpmath: a Python library for arbitrary-precision floating-point arithmetic
(version 1.3.0), 2023. http://mpmath.org/.
[23] John Nolan. Univariate Stable Distributions, Models for Heavy Tailed Data. Springer Cham, 01 2020.
[24] K. A. Penson and K. Górska. Exact and explicit probability densities for one-sided Lévy stable distribu-
tions. Phys. Rev. Lett., 105:210604, Nov 2010.
[25] Wolfram Research. Exponential power distribution, 2016. Available at https://reference.wolfram.com/
language/ref/ExponentialPowerDistribution.html. Accessed: 09-December-2023.
62 S. H. LIHN

[26] Ram K. Saxena, Arak M. Mathai, and Hans J. Haubold. Space-time fractional reaction-diffusion equations
associated with a generalized riemann–liouville fractional derivative. Axioms, 3(3):320–334, 2014.
[27] Joram Soch. Proof: Probability density function of the chi-squared distribution. The Book of Statistical
Proofs. github.io, 2019. Available at https://statproofbook.github.io/P/chi2-pdf.html.
[28] Joram Soch. Proof: Probability density function of the t-distribution. The Book of Statistical Proofs.
github.io, 2019. Available at https://statproofbook.github.io/P/t-pdf.
[29] E. W. Stacy. A generalization of the gamma distribution. The Annals of Mathematical Statistics, 33(3):1187
– 1192, 1962. See also https://en.wikipedia.org/wiki/Generalized_gamma_distribution.
[30] Student. The probable error of a mean. Biometrika, 6:1–25, 1908. See also https://en.wikipedia.org/
wiki/Student%27s_t-distribution.
[31] Y.L. Tong. The Multivariate Normal Distribution. Research in Criminology. Springer New York, 1990.
[32] E. M. Wright. On the coefficients of power series having exponential singularities. Journal London Math.
Soc., s1-8, Issue 1 (1933):71–79, 1933.
[33] E. M. Wright. The asymptotic expansion of the generalized bessel function. Proc. London Math. Soc.,
s2-38, Issue 1 (1935):257–270, 1935.