7
$\begingroup$

The von Mises distribution is entirely defined on the circle with a density given by $$f(x) = (2\,\pi\, I_0(\kappa))^{-1} \exp(\kappa \cos(x-\mu))\ ,$$ where $x$ is in an arbitrary real interval of length $2\pi$, $I_0$ is the Bessel-I function of order 0, $\mu$ is a location parameter, and $\kappa>0$ is a concentration parameter.

My question is: What transformations of a von Mises random variable preserve the distribution family?

A slightly simplified version of this question (considering an important special case) can be restated as follows: Is there a bijective function (defined on the circle) $T$ and a $\kappa>0$ solving $$(I_0(1))^{-1}\, \left|T'(x)\right|\, \exp(\cos(T(x))) = (I_0(\kappa))^{-1} \exp(\kappa \cdot \cos(x))\ ,$$

The answer is easy for $\kappa=1$, because $T$ might be chosen as the identity, a simple shift, or a suitable reflection, Thus, I am interested if this can be solved when $\kappa\neq 1$. Unfortunately, I am not aware whether there are any results available on this issue.

This is a crosspost from Mathoverflow.

$\endgroup$

1 Answer 1

6
$\begingroup$

Obviously $\mu$ is a location parameter, meaning that translations of the variable preserve the family.

Focus now on the shape parameter $\kappa$. Consider any family $\Omega=\{F_\theta|\theta\in\Theta\}$ of continuous distributions. By virtue of this continuity, whenever $X\sim F_\theta$ and $0\le q\le 1$,

$$\Pr(F_\theta(X)\le q) = q.$$

The transformation

$$G_{\theta^\prime,\theta}(X) = F_{\theta^\prime}^{-1}(F_\theta(X))$$

maps any such random variable into $Y = G_{\theta^\prime,\theta}(X)$ and

$$\Pr(Y \le y) = \Pr(F_{\theta^\prime}^{-1}(F_\theta(X)) \le y) = \Pr(F_\theta(X) \le F_{\theta^\prime}(y))=F_{\theta^\prime}(y)$$

shows that $Y \sim F_{\theta^\prime}.$

The question, then, is whether the family $\{G_{\theta^\prime,\theta}| \theta\in\Theta, \theta^\prime\in\Theta\}$ is closed under composition. Suspecting that it should not be for the shape family of the Von Mises distribution (with $\mu=0$ and $\theta=\kappa$), I numerically searched for a solution $(\alpha,\beta)$ to the equation

$$ G_{\alpha,\beta} = G_{2,1} \circ G_{1/2,1}$$

by minimizing the $L^2$ norm between the two sides. The difference between the best solution (with $\alpha=2.96234\ldots$ and $\beta = 2.48773\ldots$) and the right hand side is small but so clear that I doubt there was an error in the calculation.

Figure

Consequently the answer to the question--as understood in the sense described here--appears to be that only the translations (modulo $2\pi$) and, of course, the reflections $x\to a-x \mod 2\pi$ preserve the entire family of Von Mises distributions.

$\endgroup$

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.