Suppose we have a bivariate normal variable $\mathbf{x}= (x_1, x_2)$ with mean $\mu_1$ and $\mu_2$ and variances $\sigma_1^2$ and $\sigma_2^2$ and correlation $\rho$. I need to obtain the pdf of the transformation $\Vert\mathbf{x}\Vert^{-1} \mathbf{x}$: the coordinates of a variable in the unit circle.

In directional statistic this is a well known distribution called the Projected normal, or offset normal. I have already the pdf but i want to derive it by myself. I'm able to do a variable transformation when there is a 1 to 1 relation but this is not the case.

Can someone put me in the right way?

Just for completion: A pdf where it is explained the projected normal distribution.


1 Answer 1


There is a trick for calculating the pdf of 2-d projected normal, which can be found on pp. 52 of Mardia (1972).

After change of variable $x_1 = r\, \cos\theta$ and $x_2 = r\,\sin\theta$, we have $f(\theta)= \int_{0}^{\infty}\frac{1}{\sqrt{2\pi}\sigma_1\sigma_2\sqrt{1-\rho^2} }r\,\mbox{exp}\{-\frac{\sigma_2^2(r\cos\theta-\mu_1)^2-2\rho(r\cos\theta-\mu_1)(r\sin\theta-\mu_2)+\sigma_1^2(r\sin\theta-\mu_2)^2}{2\sigma_1^2\sigma_2^2(1-\rho^2)}\}\mbox{d}r$,

Use the following result, $d^2\int_{0}^{\infty}r\,\mbox{exp}\{-\frac{1}{2}d^2(r^2-2br)\}\mbox{d}r = 1+ (2\pi)^{\frac{1}{2}}bd\,\mbox{e}^{\frac{1}{2}b^2d^2}\Phi(bd),$

where $\Phi(x)$ is the cdf of $\mbox{N}(0,1)$.

With $d^2 = \frac{\sigma_2^2\cos^2\theta-\rho\sigma_1\sigma_2\sin2\theta+\sigma_1^2\sin^2\theta}{(1-\rho^2)\sigma_1^2\sigma_2^2}$ and $b = \frac{\mu_1\sigma_2(\sigma_2\cos\theta-\rho\sigma_1\sin\theta)+\mu_2\sigma_1(\sigma_1\sin\theta-\rho\sigma_2\cos\theta)}{\sigma_2^2\cos^2\theta-\rho\sigma_1\sigma_2\sin2\theta+\sigma_1^2\sin^2\theta}$, you are going to obtain the pdf listed in the above paper.

  • $\begingroup$ Could you explicitly connect $d$ and $b$ in this answer to the $\mu_i$, $\sigma_i$, and $\rho$ of the question? $\endgroup$
    – whuber
    Dec 1, 2015 at 22:01
  • 1
    $\begingroup$ Expressions of $d$ and $b$ with respect to $\mu_i$, $\sigma_i$ and $\rho$ added in the answer. $\endgroup$
    – polomao
    Dec 2, 2015 at 0:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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