# A Gaussian scale mixture representation of the logistic distribution

Can the logistic distribution with density function $$f(x) = \frac{e^{-x}}{\left(1 + e^{-x}\right)^2}$$ be represented as a Gaussian scale mixture? In other words, if \begin{align*} X|V &\sim N(0, V) \\ V &\sim g, \end{align*} is there a choice for $$V$$ which leads to $$X$$ having a marginal logistic distribution?

This paper by Leonard A. Stefanski (1991) illustrates that the answer to this question is yes, although the mixing density $$g$$ is difficult to work with in practice. Mixing over a Gamma density with a particular choice of parameters can also be shown to provide an extremely good approximation to the logistic distribution.

## Exact Result

Stefanski (1991) demonstrates that the logistic distribution can be represented as a Gaussian scale mixture by setting $$\sqrt V \sim q$$, where $$q(x) = \frac{d}{dx}L(x/2)$$ and $$L$$ is the Kolmogorov-Smirnov cumulative distribution function. In the present notation, the density corresponding to $$g$$ becomes $$g(v) = \sum_{n=1}^\infty (-1)^{n+1}n^2\exp\left(-\frac{n^2v}{2}\right)$$

This distribution can be difficult to work with in practice. This infinite series must be truncated after a finite number terms, and may be expensive to compute when $$v$$ is small (in fact, $$g(v)$$ will be negative for small $$v$$, unless you remember to choose an odd number for the truncation parameter). Working with $$\log g(v)$$ offers no improvement. Moreover, it is not easy to generate random variates from $$g(v)$$ and full conditionals will be intractable when using Gibbs sampling.

## Approximation via Gamma

For the reasons described in the previous paragraph, it would be convenient to have an alternative choice for $$g$$ which leads to a marginal logistic distribution for $$X$$. The gamma distribution is easy to compute, easy to simulate from and often leads to tractable distributions during Gibbs sampling. Thus we consider the scale mixture representation. \begin{align*} X|V &\sim N(0, V) \\ V &\sim \text{Gamma}(\alpha, \beta), \end{align*} Note the following. \begin{align*} \mu = E(X) &= 0 \\[1.5ex] \sigma^2 = Var(X) = E(Var(X|V)) + Var(E(X|V)) = E(V) &= \frac{\alpha}{\beta} \\[1.5ex] \kappa = E\left[\left(\frac{X-0}{\sqrt{\alpha/\beta}}\right)^4\right] = \frac{\beta^2}{\alpha^2} E\left[V^2\left(\frac{X}{\sqrt V}\right)^4\right] = \frac{3\beta^2}{\alpha^2}E(V^2) &= 3\left(\frac{1}{\alpha}+1\right) \end{align*} Using a simple moment-matching approach, we set $$3\left(\frac{1}{\alpha}+1\right) = \frac{21}{6} \quad\quad\quad \frac{\alpha}{\beta} = \frac{\pi^2}{3},$$ where $$21/6$$ and $$\pi^2/3$$ are the kurtosis and variance, respectively, of the logistic distribution. Solving for $$\alpha$$ and $$\beta$$ gives $$\alpha = 2.5 \quad\quad \beta = \frac{7.5}{\pi^2} = 0.75991\ldots.$$

## Marginal Distribution of $$X$$

The joint distribution of $$X$$ and $$V$$ can be written as $$f(x, v) = cv^{\left(\alpha-\frac{1}{2}\right) - 1}\exp\left(-\frac{1}{2}\left(2\beta v + x^2\frac{1}{v}\right)\right), v > 0, x \in \mathbb R.$$ By noticing that this function is proportional (in $$v$$) to a generalized inverse Gaussian distribution, we can derive the marginal distribution of $$X$$ as having density $$f(x) = \frac{\beta^{\frac{\alpha}{2} + \frac{1}{4}}}{2^{\frac{\alpha}{2} - \frac{3}{4}}\sqrt{\pi}\Gamma(\alpha)}|x|^{\alpha-1/2}K_{\alpha-1/2}\left(\sqrt{2\beta}|x|\right),$$ where $$K_p(x)$$ is the modified Bessel function of the second kind. In the special case where $$\alpha=2.5$$ and $$\beta = 7.5/\pi^2$$, this density reduces to

$$f(x) = \frac{\sqrt{3375}}{3\pi^4}x^2K_2\left(\frac{\sqrt{15}}{\pi}|x|\right)$$

Note that this is a special case of the Bessel function distribution.

## Numerical Simulations

#Simulate data
set.seed(1234)
N <- 1e5
V <- rgamma(N, 2.5, 7.5/pi^2)
X <- rnorm(N, 0, sqrt(V))

#Plot samples
hist(X, freq=F, breaks=100, xlim=c(-8, 10), main="")

# Plot logistic density