5
$\begingroup$

This was previously asked on MathSE, but was not answered.

Answering a question, I realized that the nearest point for a planar Poisson point process (with constant intensity $\lambda>0$) is normally distributed.

Indeed, it is easy to see that if $R_x$ measures the distance to nearest neighbor $n_x$ of some point $x$, then $$ P(R_x>r) = P(\text{no points in a ball of radius }r\text{ around }x) = e^{-\lambda\pi r^2}, $$ so $R_x$ has the Rayleigh distribution with parameter $\sigma = 1/(\lambda\pi\sqrt{2})$. And since the direction to nearest neighbor is obviously uniformly distributed, the vector $n_x - x$ is distributed as a pair of independent centered normal variables with variance $\sigma^2$. It is not hard to see that this is also true if we speak about the nearest neighbor of a point taken from the Poisson point process.

Naturally, this property is just a coincidence, it is specific to dimension $2$, and in this sense it is similar to the conformal invariance of Brownian motion. But the latter property implies a lot of interesting facts; you can even prove theorems from complex analysis like little and big Picard, using the planar Brownian motion.

Therefore, the question:

Are there any interesting facts or properties following from the fact that the nearest neighbor distribution for a planar Poisson point process is normal?

$\endgroup$
  • $\begingroup$ This is not really an answer, but a possibly related reference. Terry Soo and co-authors have some papers where they study "Poisson thinning". That is, they construct for fixed $\mu<\lambda$ a deterministic measurable map which sends an instance of a Poisson process with rate $\lambda$ to an instance of a Poisson process with rate $\mu$. This is all about "extracting randomness" from PPs; maybe the fact that you mention is used somewhere. (NB: of course non-deterministic Poisson thinning is simple) $\endgroup$ – Anthony Quas Sep 20 '15 at 6:13
  • $\begingroup$ Normal random variables can be negative, while distances cannot... $\endgroup$ – Michael Sep 20 '15 at 17:41
  • $\begingroup$ The $e^{-x^2}$ function is associated with the Gaussian density. In your link, the $e^{-x^2}$ is associated with a CDF function. I do not think it is correct to call your distance "normally distributed," even if you overlook the non-negative issue. $\endgroup$ – Michael Sep 20 '15 at 17:49
  • $\begingroup$ @Michael, maybe I was not very careful writing this. I don't claim that the distance is normally distributed, but rather that the neighbor is. The neighbor is a two dimensional vector, whose length has Rayleigh distribution with cdf $1-e^{-ax^2}$. But the vector does have normal distribution. I copied the paragraph from my MathSE post in order to clarify. $\endgroup$ – zhoraster Sep 20 '15 at 18:01
1
$\begingroup$

One "fact" following from the Gaussian form of the nearest-neigbhor distribution is that it is identical to the spherical contact distribution, being the distribution of the radius of a sphere centered at the origin when it first makes contact with a point of the Poisson process (irrespective of whether the origin is itself a point of the Poisson process). This is a useful property if you wish to test whether a given set of points has a Poisson distribution or not. The identity of nearest-neighbor and spherical contact distributions holds in any dimension iff the process is Poisson (Slivnyak’s theorem).

$\endgroup$
  • $\begingroup$ Doesn't this exist in any dimension? $\endgroup$ – zhoraster Sep 21 '15 at 14:36
  • $\begingroup$ yes, in any dimension both distributions are the same, the Gaussian form is special for the plane $\endgroup$ – Carlo Beenakker Sep 21 '15 at 14:43

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.