Gamma-order Generalized Normal distribution

Gamma-Ordered Generalized Normal Distribution

The multivariate Normal distribution, [1], is extended due to the Logarithmic Sobolev Inequalities (LSI), [2], and can act as a family of distributions based on a “shape” parameter. This shape parameter $\gamma \in \mathbb {R} -[0,1]$ creates along with parameters of position, $\mu \in \mathbb {R} ^{p}$, and dispersion, $\Sigma \in \mathbb {R} ^{p\times p}$, the $N_{\gamma }(\mu ,\Sigma )$ family of distributions with probability density function, [3]

(1) $\varphi _{\gamma }(x)=C\exp \left\{-{\frac {\gamma -1}{\gamma }}\left[Q(x)\right]^{\frac {\gamma }{2(\gamma -1)}}\right\},$

with

(2) $C=C_{p}(\mu ,\Sigma ;\gamma )={\frac {1}{\pi ^{p/2}|\Sigma |^{1/2}}}\cdot {\frac {\Gamma \left({\frac {p}{2}}+1\right)}{\Gamma \left({\frac {p(\gamma -1)}{\gamma }}+1\right)}}\left({\frac {\gamma -1}{\gamma }}\right)^{p(\gamma -1)/\gamma },\quad \gamma _{0}={\frac {\gamma -1}{\gamma }},\quad \gamma _{0}\gamma _{1}=1,$

(3) $Q(x)=\langle x-\mu ,\Sigma ^{-1}(x-\mu )\rangle ,\quad \mu \in \mathbb {R} ^{p},\quad \Sigma \in \mathbb {R} ^{p\times p},$

For $p=2$ a typical plot is Figure 1 Consider the $N_{\gamma }(\mu ,\sigma ^{2}),\ p=1$, see [6], with position (mean) $\mu$, positive scale parameter $\sigma$, extra shape parameter $\gamma \in \mathbb {R} -[0,1]$ and pdf $\varphi _{\gamma }(x;\mu ,\sigma ^{2})$ coming from (1)–(3) and given by, see Figure 2,

(4) $\varphi _{\gamma }(x;\mu ,\sigma ^{2})={\frac {\lambda _{\gamma }}{\sigma {\sqrt {\pi }}}}\exp \left\{-\gamma _{0}\left({\frac {|x-\mu |}{\sigma }}\right)^{\gamma _{1}}\right\},$

Figure 1: 3D plots of φ_γ(x; 0, I) for γ = 2 (left) and γ = 3 (right), with p = 2

with

(5) $\lambda _{\gamma }={\frac {\Gamma \left({\frac {1}{2}}+1\right)}{\Gamma \left({\frac {\gamma -1}{\gamma }}+1\right)}}\left({\frac {\gamma -1}{\gamma }}\right)^{\frac {\gamma -1}{\gamma }}={\frac {\Gamma \left({\frac {1}{2}}+1\right)}{\Gamma (\gamma _{0}+1)}}\cdot \gamma _{0}^{\gamma _{0}}.$

Let $Y\sim N_{\gamma }(\mu ,\sigma ^{2})$ then

(6) $\beta _{n}=\mathbb {E} (Y^{n})=\sum _{{\text{even }}r=0}^{n}{\binom {n}{r}}\mu ^{r}\sigma ^{n-r}\cdot \gamma _{0}^{-n}\cdot \gamma _{0}^{n-r}\cdot {\frac {\Gamma ((n-r+1)\gamma _{0})}{\Gamma (\gamma _{0})^{n-r}}}.$

When $\gamma =2$, then $\beta _{1}=\mu ,\quad \beta _{2}=\mu ^{2}+\sigma ^{2},\quad \beta _{3}=\mu ^{3}+3\mu \sigma ^{2},\quad \beta _{4}=\mu ^{4}+6\mu ^{2}\sigma ^{2}+3\sigma ^{4}.$

Moreover, [6],

(7) $\operatorname {Var} (Y)=\gamma _{0}^{-2}\cdot {\frac {\Gamma (3\gamma _{0})}{\Gamma (\gamma _{0})}}\cdot \sigma ^{2}$

and

(8) $\operatorname {Kurt} (Y)={\frac {\Gamma (\gamma _{0})\Gamma (5\gamma _{0})}{\Gamma ^{2}(3\gamma _{0})}}-3.$

The Laplace transform can be obtained, [4],

Figure 2: The pdf of the standardized φ₍γ₎(x) for γ = 2 (Normal), γ = −0.1 (near to Dirac), γ = 1.05 (near to Uniform) and γ = 30 (near to Laplace), with p = 1.