December – Asymptotic decay rate of a probability distribution

Let us study the tail distributions of three related distributions. The asymptotic decay rate measures the thickness of a random variable's $X$ tail distribution. It is defined by

$\displaystyle \lim_{x \rightarrow \infty} -\frac{\log\textbf{P}(X > x)}{x},$ (1)

where $\textbf{P}(\cdot)$ denotes the probability of an event. Let us calculate the decay rates for the exponential distribution, gamma distribution, and normal distribution.

For exponentially distributed $X$ with scale parameter $\theta$:

$\displaystyle \rho_{\text{Exponential}}$ $\displaystyle =\lim_{x \rightarrow \infty} -\frac{\log\textbf{P}(X > x)}{x} = \lim_{x \rightarrow \infty} -\frac{\log(e^{- x/\theta})}{x}$    
  $\displaystyle =\lim_{x \rightarrow \infty} -\frac{-x/\theta }{x } = 1/\theta.$ (2)

For gamma distributed $X$ with shape parameter $k$ and scale parameter $\theta$:

$\displaystyle \rho_{\text{Gamma}}$ $\displaystyle = \lim_{x\rightarrow \infty} -\frac{\log \mathbb{P}(X>x)}{x} = \lim_{x\rightarrow \infty} -\frac{\log (1-\gamma(k,x/\theta)/\Gamma(k))}{x}$    
  $\displaystyle =\lim_{x\rightarrow \infty} -\frac{\log (\Gamma(k,x/\theta)/\Gamma(k))}{x}$    
  $\displaystyle \overset{(a)}{=} \lim_{x\rightarrow \infty} -\frac{\log (\Gamma(k...
... = \lim_{x\rightarrow \infty} -\frac{\log\left(x^{k-1}e^{-x/\theta}\right) }{x}$    
  $\displaystyle = \lim_{x\rightarrow \infty} -\frac{\log \left(e^{-x/\theta}\right)}{x} = 1/\theta.$ (3)

In $(a)$, we used the asymptotic behavior of the gamma distribution

$\displaystyle \lim_{x \rightarrow \infty}\frac{\Gamma(s,x)}{x^{s-1}e^{-x}} = 1.

For normal distributed $X$ with mean $\mu$ and variance $\sigma^2$:

$\displaystyle \rho_{\text{Normal}}$ $\displaystyle = \lim_{x\rightarrow \infty} -\frac{\log \mathbb{P}(X>x)}{x} = -\...
...t{\frac{e}{2\pi}}e^{-2\left(\frac{x-\mu}{\sigma \sqrt{2}}\right)^2} \right)}{x}$    
  $\displaystyle \geq -\lim_{x\rightarrow \infty} \frac{\log\left( e^{-2\left(\fra...
...}{\sigma^2 }\lim_{x\rightarrow \infty} \frac{x^2-2 x \mu + \mu^2}{ x} = \infty,$ (4)

where in $(b)$, we used the inequality

erfc$\displaystyle (x) \geq \sqrt{\frac{{2 e}}{\pi}} \frac{\sqrt{\beta-1}}{\beta} e^{-\beta x^2},$

that holds for $x \geq 0$ and $\beta >1$ (we used $\beta = 2$).

The asymptotic decay rates for the exponential, gamma, and normal distribution are given in $(2)$, $(3)$, and $(4)$, respectively. We see that the decay rate $\rho$ of the normal distribution is always maximal, i.e., $\rho_{\text{Normal}} = \infty$. In contrast, the decay rates of the exponential and gamma distributions depend on the scale parameter $\theta$.

By normalizing the mean ( $\mathbb{E}[X] = 1$) by setting $\theta = 1$ and $k=1/\theta$, we can compare the decay rates of the exponential distribution and gamma distribution: then, the decay rate of the exponential distribution is $\rho_{\text{Exponential}} = 1$ and of the gamma distribution is $\rho_{\text{Gamma}} = k$. That is, keeping the mean normalized, the gamma distribution decays slower than the exponential distribution with shape parameters $k<1$, and faster than the exponential distribution with $k>1$. With $k=1$, the distributions coincide. When $k \rightarrow \infty$, the gamma distribution approaches the normal distribution and $\rho_{\text{Gamma}} \rightarrow \rho_{\text{Normal}} = \infty$.