Alpha-Skew Generalized Normal Distribution

• The ASGN is a parametric family unifying Gaussian-based models by flexibly controlling tail thickness (α), skewness (θ), and moment existence (ν).
• Its formulation using a generalized Wright function and mixture structures enables closed-form expressions for moments and characteristic functions.
• ASGN offers practical modeling benefits for heavy-tailed, skewed data across fields like finance, hydrology, and signal processing.

The alpha-skew generalized normal distribution (often abbreviated as ASGN, Editor's term) is a parametric family of probability distributions that unifies and extends a diverse collection of Gaussian-based models—most notably, the classical normal, $\alpha$-stable, exponential power, Student’s $t$, and various skew-normal/generalized normal distributions—by introducing flexible control of tail behavior (via the stability index $\alpha$), skewness (via a skewness-angle parameter $\theta$), and variance/existence of moments (via a degree-of-freedom parameter $\nu$). Mathematically, the ASGN is represented through a generalized Wright function framework and incorporates advanced mixture and subordination structures, granting it remarkable adaptability for modeling data exhibiting skewness, heavy tails, and variable kurtosis (Lihn, 2024).

1. Mathematical Formulation and Core Structure

The symmetric "super-distribution" underlying the ASGN is given as a continuous Gaussian mixture: $$L_{\alpha,\nu}(x) = \int_0^{\infty} s\,ds \; N(xs)\; {}_{\alpha,\nu}(s), \qquad x \in \mathbb{R}$$ where $N(x) = (2\pi)^{-1/2} e^{-x^2/2}$ is the standard normal PDF and ${}_{\alpha,\nu}(s)$ is a fractional $\chi$-type mixing law parameterized by index $\alpha\in(0,2]$ and general degree of freedom $\nu>0$.

This construction reduces to several classical distributions as special/limiting cases:

• $\nu = 1$: recovers the classical symmetric $\alpha$-stable distribution.
• $\nu \to -1$: yields the exponential power (EP) family.
• $\alpha = 2$: gives Student’s $t$ with degree of freedom $\nu$.
• Other limits interpolate among these, depending on $\alpha$, $\nu$.

The skewed variant, the full ASGN, replaces $N(xs)$ by a skew–Gaussian kernel $g_\alpha^\theta(x,s)$ derived from the Feller parameterization:

$$g_\alpha^{\theta}(x,s) = \frac{1}{q\pi} \int_{0}^{\infty} \cos\left(\tau (s t)^\alpha + \frac{x}{q}s t \right) e^{-t^2/2} dt$$

with $q = \cos(\theta\pi/2)^{1/\alpha}$, $\tau = \tan(\theta\pi/2)$, $|\theta| \leq \min\{\alpha, 2-\alpha\}$.

The ASGN density is thus given by

$$L_{\alpha,\nu}^{\theta}(x) = \int_0^{\infty} s\,ds\; g_\alpha^{\theta}(x, s)\;{}_{\alpha,\nu}(s)$$

which has a closed-form representation in terms of the four-parameter Wright function (Lihn, 2024).

2. Parametric Roles and Distributional Properties

The ASGN possesses three principal shape parameters:

• Stability index $\alpha$ ($0<\alpha\le 2$): Governs tail thickness and generalizes the normal ($\alpha=2$) and Cauchy ($\alpha=1$) cases.
• Degree of freedom $\nu>0$: Controls existence of moments, variance, skewness, and kurtosis. For each even $n$, the moment $E[X^n]$ exists iff $\nu>n$.
• Skewness parameter $\theta$ ($|\theta|\le\min\{\alpha,2-\alpha\}$): Encapsulates asymmetric behavior via a “skewness angle.”

Key quantitative properties:

• Mean: $m_1\ne 0$ for $\theta\ne 0$, obtained by explicit formula.
• Variance: Finite when $\nu>2$.
• Skewness: Finite when $\nu>3$.
• Kurtosis: Finite when $\nu>4$.

Closed-form formulae for the first four moments and explicit existence conditions follow from the Wright-function representation and integration against the mixing law (Lihn, 2024).

Special/lower-dimensional cases recovered:

• $\theta=0$: symmetric law corresponding to the “generalized normal” of order $(\alpha,\nu)$.
• $\nu \rightarrow 1$: recovers classical $\alpha$-stable.
• $\alpha=2$: the distribution approaches the (possibly skewed) Student’s $t$; for $\nu=-1$, recovers the exponential power.

3. Analytical and Computational Tools

Characteristic Function

The characteristic function is

$$\phi_{\alpha,\nu}^\theta(t) = \int_{0}^{\infty} \exp\left(-s^{\alpha}e^{i \mathrm{sgn}(t)\theta\pi/2}|t|^{\alpha}\right){}_{\alpha,\nu}(s)ds$$

reducing to the classical $\alpha$-stable characteristic function for $\nu=1$.

Parameter Estimation

Parameter inference for ASGN can be approached via:

• Method of moments: Matching up to four empirical moments to their closed-form expressions.
• Maximum likelihood / EM: The latent subordinator $S_i$ is treated as a missing variable. The E-step integrates over its conditional posterior, and the M-step optimizes with respect to $\alpha, \nu, \theta$. Special cases admit closed-form updates (e.g., $\theta=0$, $\alpha=1$), but the general model requires numerical root-finding (Lihn, 2024).

Multivariate Extension

A multivariate elliptical generalization is formulated by substituting the scalar normal kernel with the multivariate normal: $$L^\theta_{\alpha,\nu}(\mathbf{x};\Sigma) = \int_0^\infty |s|^{n} ds\,\mathcal N_n(\mathbf{x};0,\Sigma/s^2){}_{\alpha,\nu}(s)$$ with characteristic function reflecting the elliptical symmetry (Lihn, 2024).

4. Relation to Other Generalized and Skew-Normal Families

The ASGN synthesizes and strictly generalizes a host of earlier normal-based distributions:

• Alpha-Skew Normal (ASN): The ASN, with standardized PDF $$f(x|\alpha) = [(1-\alpha x)^2 +1]/(2+\alpha^2)\cdot \varphi(x)$$, introduces skewness via a parameter $\alpha \in \mathbb{R}$ and can exhibit bimodality and fat tails for large $|\alpha|$; for $\alpha=0$ it reduces to the normal (Nascimento et al., 2021). The ASGN contains ASN as a special symmetric, variance-finite subfamily for suitable $(\alpha,\nu,\theta)$.
• Generalized Alpha-Beta-Skew-Normal (GABSN): The GABSN further generalizes ASN and includes a cubic shape parameter $B$ and classical Azzalini-type skew parameter $\lambda$, with closed-form moments and maximum-likelihood estimation via five nonlinear equations (Shah et al., 2019). The ASGN, when specified for certain limiting parameters and with appropriate reinterpretation of skew and shape indices, includes these cases as submodels, offering a broader functional class with more flexible control over modality and tail geometry.

All these distributions are part of a hierarchy where polynomial and/or kernel mixture-type modifications of the normal law allow for increased skewness, multi-modality, and flexible tail scaling, but ASGN uniquely subsumes $\alpha$-stable, $t$-distribution, and exponential power laws directly via its advanced mixture and subordination structure.

5. Special Functions and Theoretical Framework

The ASGN's tractability relies on higher transcendental functions, specifically the Wright function $W[a,b](z)$, which enables explicit forms for the PDF, CDF, and mixing laws:

• The fractional $\chi$-mixing law, ${}_{\alpha,\nu}(s)$, is expressed as a Wright function, accommodating the desired moment structure.
• The ASGN CDF is a fractional extension of the Gauss hypergeometric function.
• The distribution is constructed as a "ratio distribution" of the skew-Gaussian kernel to a fractional $\chi$ distribution, providing a probabilistic interpretation involving subordination and stable counts (Lihn, 2024).

This theoretical apparatus allows interpolation and analytic continuation between various standard families, notably bridging the gap between $\alpha$-stable distributions (which typically lack finite moments except for Gaussian) and robust heavy-tailed symmetric or skewed models admitting higher moments.

6. Practical Implications and Applications

The ASGN is suitable for modeling data that demand heavy tails, skewness, and more nuanced moment control than offered by the classical normal, stable, or $t$ distributions. The existence of moments is governed entirely by the degree-of-freedom $\nu$, similar to Student’s $t$ but generalized to incorporate stable-like and exponential power behaviors.

Inference methods such as the Anderson–Darling estimator (ADE) are especially effective for capturing tail behavior and fitting to heavy-tailed, multimodal, or skewed data (Nascimento et al., 2021); ASGN's mixture structure provides further flexibility for risk quantile estimation and robust modeling in disciplines such as hydrology, finance, and signal processing.

Multivariate extension supports construction of elliptical models with prescribed covariance structure and heavy-tailed marginal distributions, generalizing classical Gaussian graphical models to the skewed and heavy-tailed domain (Lihn, 2024).

7. Model Selection and Empirical Performance

Empirical studies comparing ASGN-type distributions (and their nested subfamilies) with competing models use criteria such as AIC, BIC, and likelihood-ratio tests. In practical applications, extended families such as GABSN or ASGN often achieve lower AIC/BIC and greater log-likelihood than normal, skew-normal, or simpler alpha-skew-normal families, especially when the data exhibit pronounced skewness, multimodality, or heavy tails (Shah et al., 2019). This suggests the importance of considering broad, flexible normal-mixing laws, such as the ASGN, when classical models fail to adequately capture empirical distributions.

Extensive simulation and real-data illustrations confirm that standard inference routines (MLE, EM, ADE) coupled to ASGN and its relatives offer robust, flexible, and theoretically grounded solutions for real-world data that deviate significantly from normality (Nascimento et al., 2021, Lihn, 2024).

---

References:

• Lihn, "Generalization of the Alpha-Stable Distribution with the Degree of Freedom," (Lihn, 2024)
• Shah et al., "The Generalized-Alpha-Beta-Skew-Normal Distribution: Properties and Applications," (Shah et al., 2019)
• Silva et al., "Generalizing the normality: a novel towards different estimation methods for skewed information," (Nascimento et al., 2021)