I doing Bayesian modelling these days. I found that cauchy distribution can be written as a scale mixture of normal based on following source. Link

So I started to derive this. Somehow, I am not getting what I want. I am trying to prove the third implementation in the above source.

It says :

$Cauchy(x|0,1)=\int N(x|0,\tau^{-1/2})InvGamma(\tau|1/2,1/2)d\tau$

Or in other words $X=x_a*\sqrt{x_b}$ follows a cauchy(0,1) if $x_a \sim N(0,1)$ and $x_b \sim invGamma(1/2,1/2)$

So here is my working :

$N(x|0,\tau^{-1/2}) \propto \tau \times e^{x^2\tau/2}$ and $IG(\tau|1/2,1/2) \propto \tau^{1/2-1} \times e^{-2\tau}$.

So the integral,

$\int N(x|0,\tau^{-1/2})$ $IG(\tau|1/2,1/2)d\tau \propto \int\tau^{1/2} \times e^{-2\tau(x^2+1)} d\tau$

$\propto 1/(x^2+1)^{3/2}$. (by creating the inverse gamma integral inside)

Ideally, I should get 1 as the exponent of the $x^2$+1 . What am I missing here ?

Can anyone help me to figure this out ?

Thank you.

  • 1
    $\begingroup$ Shouldn't the leading term in your Normal be $t^{1/2}$ (reciprocal sd rather than reciprocal variance)? $\endgroup$ – Thomas Lumley Oct 4 '20 at 1:25
  • $\begingroup$ thank you. I missed that. $\endgroup$ – student_R123 Oct 4 '20 at 1:36

If you just want to prove that it's a scale mixture of Normals, it's easier to work directly. The Cauchy distribution is the $t_1$ distribution.

The $t_1$ distribution is a Normal divided by the square root of an independent chi-squared. So, let $Z\sim N(0,1)$ and $S^2\sim \sigma^2\chi^2_1$. The distribution of $Z/S$ is Cauchy.

But that's equivalently an $N(0,T^2)$ conditional on $T=1/S$, so it's a scale mixture of Normals.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.