1

What is the expected minimum euclidean distance between $N$ points uniformly and independently chosen on the shell(sphere) of 3-D ball of radius $R$? Note that the expected minimum distance might be difficult to compute, so a good lower bound is also fine.

My approach I take the approach similar to this question.

Suppose we put $N$ circles of radius $r$ uniformly on the surface of the ball.
Therefore we could compute the expected minimum distance by the following expression \begin{align} E[D]=\int_0^\infty P(D>t) dt. \end{align}

So, it remains to compute $P(D>t)$. Actually, since we are after a lower bound on $E[D]$ it is enought to give a lower bound on $P(D>t)$.

Let $S_i$ be the event that the pair of circles does not intersect which should be given by \begin{align} P(S_i) = \frac{ Surf(R)-Surf(r)}{Surf(R)}= 1-\frac{ \pi r^2}{ 4\pi R^2}=1-\frac{ r^2}{ 4 R^2} \end{align}

Now the probability that euclidean minimum distance $D$ between $N$ balls is bigger than $t$ should be similar to \begin{align} P[ D \ge t] \stackrel{?}{=} P[ \cap S_i] \end{align} However, I think we somehow need to converte from the geodesic distance to euclidean distance but I am not very sure how to do that?

Community
  • 1
Boby
  • 5,985
  • 1
    That “shell of a ball” would be a sphere. Computing the distribution for the distance between a pair of points should be relatively easy, but several such distances would no longer be independent, thus complicating matters. It might be worthwhile to randomly pick some points in a Monte-Carlo fashion, so that you have some numbers against which you can compare proposed formulas. – MvG Nov 17 '16 at 14:21
  • @MvG Yes, the shell of the ball is a sphere. You are correct in saying that the distances are no longer independent. However, since we are after a lower ground we can do the following $P[ \cap S_i ]=1- P[ \cup S_i^c ]\ge 1- \sum P[S_i^c]$. My concern, if there are any issues with going from geodesic distance to euclidean? – Boby Nov 17 '16 at 14:25
  • Your main question (first paragraph) doesn't mention that you're looking for a bound on the expected value; it sounds like you'd want an exact formula. Some readers might loose interest there even though they would be able to help find a bound. – MvG Nov 17 '16 at 14:29
  • @MvG thanks. I added a correction. – Boby Nov 17 '16 at 14:36
  • It's not clear whether you are interested in a bound or an asymptotic for large $N$ – leonbloy Nov 17 '16 at 15:10
  • @leonbloy I am interested in the bound for fixed $N$. However, an asymptotic result when $N$ is large would also be interesting and acceptable. – Boby Nov 17 '16 at 15:18
  • @leonbloy Do you think the approach you used before can be applied in this case? – Boby Nov 17 '16 at 20:53

1 Answers1

0

The asymptotics is a straightforward (or boring) generalization of my previous answer.

Sketch: Take a point over the unit sphere ($R=1$). Then the "excluded area" $a(D)$ (points that lie at a -euclidean- distance less than $D$ from the point) is $$a(D)=\pi D^2$$

The probability that a "link is free" (distance between points pair $i$ isgreater than $D$ is

$$p(S_i) =1-\frac{1}{4}D^2 \hspace{1cm} 0\le D \le 2 \tag{1}$$

Assuming independence as asympotical approximation:

$$p(\cap S_i) \approx (1-\frac{1}{4}D^2 )^M, \hspace{1cm} M=\frac{N(N-1)}{2} \tag{2}$$

Then, letting $t=$ minimum distance between points

$$E(t) = \int P(t >D) \, dD = \int p(\cap S_i)\, dD \approx \int_{0}^2 (1-\frac{1}{4}D^2 )^M \, dD \tag{3}$$

For large $M$ this tends to

$$E(t) \approx \frac{\sqrt{\pi}}{\sqrt{M}}\approx \frac{\sqrt{2 \pi}}{N} \tag{4}$$

For general ball radius $R$:

$$E(t) \approx R \frac{\sqrt{\pi}}{\sqrt{M}}\approx R \frac{\sqrt{2 \pi}}{N} $$

Community
  • 1
leonbloy
  • 63,430