1. Let $X_1,..., X_n$ be i.i.d. $d$-dimensional standard normal random variables, and let $R_n$ be the radius of the smallest ball containing $X_1,...,X_n$. What is known about the distribution of $R_n$ as $n \to \infty$?

  2. Are there some examples for the asymptotics of $R_n$ when the independence assumption does not hold? I would appreciate any reference or suggestion here, also for a non-Gaussian case.

  • 1
    $\begingroup$ I don't think the ball was supposed to be centred at the origin. $\endgroup$ – Anthony Quas Nov 9 '16 at 17:11
  • $\begingroup$ @AnthonyQuas no, otherwise it would be the maximal distance and this I think is well known. $\endgroup$ – Sinusx Nov 9 '16 at 20:35
  • 2
    $\begingroup$ In fact, though, I suspect this is the right answer. If you look at the smallest ball centred at the origin, then this grows slowly (at rate $\sqrt{\log n}$). Hence there are many points close to the maximal radius. These should fairly closely fill out a thin spherical shell of radius $R_n$. $\endgroup$ – Anthony Quas Nov 9 '16 at 22:26

This is to confirm and give more details to Anthony Quas' comment.

Let's start with a bit of wishful thinking. Denote by $P_n$ the convex hul of the points $X_1,\dotsc, X_n$.$\newcommand{\bR}{\mathbb{R}}$ $\newcommand{\bE}{\mathbb{E}}$ Denote by $B_R(p)$ the ball or radius $R$ centered at $p\in\bR^d$.Suppose that for any $n\geq 3$ there exists radii $R_n^-<R_n^+$ such that.

  1. $B_{R_n^-}(0)\subset P_n \subset B_{R_n^+}(0)$ a.s. for large $n$.
  2. $R_n^-\sim R_n^+$ as $n\to \infty$, i.e.,

$$ \lim_{n\to\infty} \frac{R_n^+}{R_n^-}=1. $$

Then, almost surely,

$$\frac{R_n}{R_n^-}\to 1. $$

Fortunately such radii can be found and they have the growth rate suggested by Anthony Quas. This is all contained in a preprint of Fresen & Vitale arXiv: 1402.2718

Let me give a few more details. Denote by $\mu$ The standard Gaussian measure on $\bR^d$. For each $\delta>0$ we denote by $F_\delta$ the intersection of all closed half-spaces $S\subset \bR^d$ such that $\mu(S)\geq 1-\delta$.

The floating body is then the convex set $F_{1/n}$. Due to the radial symmetry of $\mu$ you can see that $F_{1/n}$ is a ball of radius $r_n$ centered at the orgin,where $r_n$ is found from the equality

$$\Psi(r_n)=\frac{1}{n}, \;\;\Psi(x):=\frac{1}{\sqrt{2\pi}}\int_{x}^\infty e^{-t^2/2} dt=\frac{1}{n}. $$

From the known estimates

$$ \Big(\frac{1}{x}-\frac{1}{x^3}\Big)\gamma(x) < \Psi(x) <\frac{1}{x}\gamma(x),\;\;\;\;\gamma(x)=\frac{1}{\sqrt{2\pi}} e^{-x^2/2}. $$

We deduce that

$$r_n=\sqrt{2\log n} +o(1), \;\;\mbox{as}\; n\to\infty. $$

In the above preprint construct (explicit) sequences of positive real numbers $\newcommand{\ve}{\varepsilon}$ $\ve_1(n),\ve_2(n)$ converging to $0$ with the following property: with probability $1$, there exists a natural number $N$ such that for all $n>N$

$$ \big(1-\ve_1(n)\;\big) F_{1/n}\subset P_n\subset \big(1+\ve_2(n)\;\big) F_{1/n}. $$

Thus proves that

$$ \frac{R_n}{\sqrt{2\log n}}\to 1\;\; \mbox{a.s.}. $$


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.