1
$\begingroup$

Here are some of the classical density functions for spherical distributions (on the $\mathcal{S}^{d-1}$ sphere, living in the Euclidean space $\mathbb{R}^d$):

  1. $$\mathbf{x}\mapsto \frac{(\kappa/2)^{d/2-1}}{2 \pi^{d/2} I_{d/2-1}(\kappa)} \exp(\kappa \mathbf{x}^{\top} \boldsymbol{\mu}), \qquad (\text{called the Fisher-von Mises-Langevin density}),$$

  2. $$\mathbf{x}\mapsto \frac{1}{a(\kappa,A)} \exp(\kappa \mathbf{x}^{\top} \boldsymbol{\mu} + \mathbf{x}^{\top} A \mathbf{x}), \qquad (\text{called the Fisher-Bingham density}),$$

  3. $$\mathbf{x}\mapsto \frac{\Gamma(d/2)}{2 \pi^{d/2} M(\frac{1}{2},\frac{d}{2},\kappa)} \exp(\kappa (\mathbf{x}^{\top} \boldsymbol{\mu})^2), \qquad (\text{called the Watson density}),$$

where $\kappa\geq 0$ is a concentration parameter, $\boldsymbol{\mu}\in \mathcal{S}^{d-1}$ is a location parameter, $A$ is a symmetric $d\times d$ matrix, and both $a(\kappa,A)$ and $M(\frac{1}{2},\frac{d}{2},\kappa)$ are the appropriate normalizing constants.

I've seen very few central limit theorems in the literature relating to this setting. In particular, I found absolutely nothing regarding local limit theorems. If the parameter $\kappa$ approaches some limit ($0$ or $\infty$), do any of these density functions approach a particular limit density (with a properly normalized argument)?


$\textbf{Example:}$ As the intensity parameter $\lambda$ of a Poisson$(\lambda)$ distribution tends to $\infty$, the probability mass function tends to the density of a $\text{Normal}(\lambda,\lambda)$ distribution. Is there any analogous results/conjectures in the context of spherical distributions?

$\endgroup$
8
  • $\begingroup$ I think it is worth pointing out that you also asked this question on math stackexchange. $\endgroup$ – Kernel May 14 at 7:31
  • $\begingroup$ Do you have an explicit formula/expansion for the characteristic function of such random variables? I am not familiar with this class of distributions. If do have such characterisation, the argument should be much easier, I believe. $\endgroup$ – Kernel May 14 at 7:31
  • $\begingroup$ The characteristic function of the matrix version of the Fisher-von Mises distribution is here (link.springer.com/content/pdf/10.1007%2F978-1-4612-1358-1.pdf) on page 484, but it's pretty nasty. I didn't find the characteristic function for the vector-version above. $\endgroup$ – Frédéric Ouimet May 14 at 8:11
  • $\begingroup$ Right, we actually do not need the explicit formula, but it would be convenient. In many proofs for LCLT for random walks, we just need an expansion of the characteristic function $\phi_X(\theta)$ around $\theta=0$. Because this is not a supported in a lattice, and therefore $\phi_X$ is not periodic, we will also need some estimates of the decay of $\phi_X$ at infinity as well. $\endgroup$ – Kernel May 14 at 10:06
  • $\begingroup$ Given such expansion, I do believe you could achieve a result like $\|p_{X_n}-p_{N(\mu\cdot n, \Sigma*n)}\|_\infty \le Cn^{{-d-2}/{2}}$ (where $\mu$ and $\Sigma$ respectively denote the appropriate mean vector and correlation matrix ) by formulating on the argument in Lawler-Limic book $\endgroup$ – Kernel May 14 at 10:08
0
$\begingroup$

$\textbf{Partial answer:}$ Note that $$\boldsymbol{x}^{\top} \boldsymbol{\mu} = -\frac{1}{2} (\boldsymbol{x} - \boldsymbol{\mu})^{\top}(\boldsymbol{x} - \boldsymbol{\mu}) + 1$$ since $\boldsymbol{x}^{\top}\boldsymbol{x} = \boldsymbol{\mu}^\top\boldsymbol{\mu} = 1$, and it is possible to show (using the integral representation for the modified Bessel function of the first kind) that, as $\kappa\to \infty$, $$e^{\kappa} \cdot \frac{(\kappa/2)^{d/2-1}}{2 \pi^{d/2} I_{d/2-1}(\kappa)} = \frac{1}{(2\pi \kappa^{-1})^{(d-1)/2}} \cdot \frac{1}{\sqrt{2\pi} \kappa^{1/2} e^{-\kappa} I_{d/2 - 1}(\kappa)} \approx \frac{1}{(2\pi \kappa^{-1})^{(d-1)/2}},$$ so that, as $\kappa\to \infty$, $$\frac{(\kappa/2)^{d/2-1}}{2 \pi^{d/2} I_{d/2-1}(\kappa)} \exp(\kappa \mathbf{x}^{\top} \boldsymbol{\mu}) \approx \frac{1}{(2\pi \kappa^{-1})^{(d-1)/2}} \exp\Big(-\frac{1}{2 \kappa^{-1}} (\boldsymbol{x} - \boldsymbol{\mu})^{\top}(\boldsymbol{x} - \boldsymbol{\mu})\Big).$$ This heuristic solves the question for 1.

$\textbf{Note:}$ The $d-1$ instead of $d$ comes from the restriction to the $\mathcal{S}^{d-1}$ sphere.

$\endgroup$

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.