Lets $(X_n)_{n\in\mathbb{N}}$ be an iid sequence of standard normal random variables. Define $$M_n=\max_{1\leq i\leq n} X_i.$$ Prove that $$\lim_{n\rightarrow\infty} \frac{M_n}{\sqrt{2\log n}}=1\quad\text{a.s.}$$

I used the fact that $$\left(\frac{1}{x}-\frac{1}{x^3}\right)e^{-\frac{x^2}{2}}\leq\mathbb P(X_n>x)\leq \frac{1}{x}e^{-\frac{x^2}{2}},$$

and the Borel Cantelli lemmas to prove that $$\limsup_{n\rightarrow\infty} \frac{X_n}{\sqrt{2\log n}}=1\quad\text{a.s.}$$

I used Davide Giraudo's comment to show $$\limsup_n \frac{M_n}{\sqrt{2\log n}}=1\quad \text{a.s.}$$

I have no idea how to compute the $\liminf$. Borel-Cantelli give us tools to compute the $\limsup$ of sets, I am unsure of how to argue almost sure convergence. Any help would be appreciated.


Your estimate gives that for each positive $\varepsilon$, we have $$\sum_i\mathbb P(M_{2^{i}}>(1+\varepsilon)\sqrt 2\sqrt{i+1})<\infty$$ (we use the fact that $\mathbb P(M_n>x)\leqslant n\mathbb P(X_1\gt x)$).

We thus deduce that $\limsup_iM_{2^{i}}/(\sqrt{2(i+1)})\leqslant 1+\varepsilon$ almost surely. Taking $\varepsilon:=1/k$, we get $\limsup_iM_{2^{i}}/(\sqrt{2(i+1)})\leqslant 1$ almost surely. To conclude, notice that if $2^i\leqslant n\lt 2^{i+1}$, $$\frac{M_n}{\sqrt{2\log n}}\leqslant \frac{M_{2^{i+1}}}{\sqrt{2i}}.$$

For the $\liminf$, define for a fixed positive $\varepsilon$, $$A_n:=\left\{\frac{M_n}{\sqrt{2\log n}}\lt 1-\varepsilon\right\}.$$ A use of the estimate on the tail of the normal distribution show that the series $\sum_n \mathbb P(A_n)$ is convergent, hence by the Borel-Cantelli lemma, we have $\mathbb P\left(\limsup_n A_n\right)=0$. This means that for almost every $\omega$, we can find $N=N(\omega)$ such that if $n\geqslant N(\omega)$, then $$\frac{M_n(\omega)}{\sqrt{2\log n}}\geqslant 1-\varepsilon.$$ This implies that $$\liminf_{n\to \infty}\frac{M_n(\omega)}{\sqrt{2\log n}}\geqslant 1-\varepsilon \quad \mbox{a.e.}$$ Here again, taking $\varepsilon:=1/k$, we obtain that $$\liminf_{n\to \infty}\frac{M_n(\omega)}{\sqrt{2\log n}}\geqslant 1\quad \mbox{a.e.}$$

  • $\begingroup$ So I am mainly having trouble with justifying the first sum converges. I get $$\mathbb{P}(M_{2^i}>\sqrt{(1+\epsilon)2(i+1)})=1-(1-\mathbb{P}(X_{2^i}>\sqrt{(1+\epsilon)2(i+1)}))^{2^i} \leq 1-(1-\frac{1}{\sqrt{(1+\epsilon)2(i+1)}}e^{-(1+\epsilon)(i+1)})^{2^i}$$ $\endgroup$
    – anonymous
    Oct 7 '14 at 10:15
  • $\begingroup$ I edited. Is it clearer? $\endgroup$ Oct 7 '14 at 10:18
  • $\begingroup$ yes significantly. Thank you. So that shows that $\limsup$ of $\frac{M_n}{\sqrt{2\log n}}\leq 1$, but not the $\liminf\geq1$. $\endgroup$
    – anonymous
    Oct 7 '14 at 10:27
  • $\begingroup$ It may be not the case that $\liminf\geq 1$, however the limsup is greater than $1$ because $M_n\geq X_n$. $\endgroup$ Oct 7 '14 at 11:10
  • $\begingroup$ Sorry, I missed the fact that you have a limit and not only a limsup. $\endgroup$ Oct 7 '14 at 11:22

Not an answer, but a related comment that is too long for a comment box ...

This question made me curious to compare:

  • the pdf of the sample maximum (given a sample of size $n$ drawn on a N(0,1) parent), say $f(x;n):$

$$f(x) = \frac{2^{\frac{1}{2}-n} n e^{-\frac{x^2}{2}} \left(1+\text{erf}\left(\frac{x}{\sqrt{2}}\right)\right)^{n-1}}{\sqrt{\pi }}$$ where erf(.) denotes the error function, to

  • the asymptote proposed by the question $\sqrt{2 \log n}$

... when $n$ is very large (say $n$ = 1 million).

The diagram compares:

enter image description here

The distribution does not, in fact, converge to the asymptote in any meaningful manner, even when $n$ is very large, such as $n$ = 1 million or 100 million or even a billion.

  • $\begingroup$ Are you using the right logarithm? (I.e. base $e$ and not base 10?) $\endgroup$ Oct 1 '18 at 0:13
  • 1
    $\begingroup$ Yes - $\sqrt{2 \log \left(10^6\right)}$ = 5.25 ... as shown in the diagram above (the red line). By contrast, with base 10, the same calculation would be equal to $2 \sqrt{3}$ which is approx 3.46. $\endgroup$
    – wolfies
    Oct 1 '18 at 7:17
  • $\begingroup$ Using the large-$x$ asymptotics of the error function, I found that the maximum is attended at $\sqrt {2\log n} \left( {1 + \mathcal{O}\left( {\frac{1}{{\log n}}} \right)} \right)$, showing that the convergence is very slow. $\endgroup$
    – Gary
    Nov 5 '20 at 10:47

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.