1
$\begingroup$

This relate to that paper:

http://www.stat.purdue.edu/docs/research/tech-reports/1982/tr82-17.pdf

Let $U_1,...,Un$ be iid uniform on (0,1). Set $L_n=\max_{i\leq n} U_i$.

Also $S(n)= \inf\{i\leq n| U_i = L_n \}$ the time were the highest value is attained and

$Z(n)= \inf\{i\leq n| U_i = L_{S(n)-1} \} \vee 0$ the time were the highest value before the very highest is attained (not necessary the second highest globally!)

Proof that there are $V$ ,$V' \sim Exp(1)$ and $W$, $W' \sim U(0,1)$ all mutually independent, such that:

$(n(1-L_n),(S(n)-1)(1-\frac{L_{S(n)-1}}{L_n}),\frac{S(n)}{n},\frac{Z(n)}{S(n)-1})\overset{d}{\rightarrow}(V,V',W,W').$

A proof sketch will be sufficient.

I also posted the question here:https://math.stackexchange.com/questions/1844238/limit-of-a-probability-vector

$\endgroup$

1 Answer 1

1
+100
$\begingroup$

For $1 \leq s' < s \leq n$ and $t,t' \in \mathbb{R}$, one has $$ n(1-L_n) > t, (S(n)-1)(1-\frac{L_{S(n)-1}}{L_n}) > t',S(n)=s, Z(n) = s' $$ precisely when $$ U_1,\dots,U_{s'-1} < U_{s'},\\ U_{s'+1},\dots,U_{s-1} \leq U_{s'}, \\ U_{s+1},\dots,U_{n} \leq U_{s},\\ U_{s'} < \left(1 - \frac{t'_+}{s-1} \right)_+ U_s, \\ U_{s} < \left(1 - \frac{t_+}{n} \right)_+ , $$ where $x_+ = \max(x,0)$. Setting $\alpha =\left(1 - \frac{t_+}{n} \right)_+ $ and $\alpha' = \left(1 - \frac{t'_+}{s-1} \right)_+$, the probability of these events is $$ \mathbb{E}[U_{s'}^{s-2} \mathbb{1}_{U_{s'} \leq \alpha' U_s} U_{s}^{n-s} \mathbb{1}_{U_{s} \leq \alpha} ] = \mathbb{E}[\frac{(\alpha' U_s)^{s-1}}{s-1}U_{s}^{n-s} \mathbb{1}_{U_{s} \leq \alpha}] \\ = \frac{\alpha'^{s-1} \alpha^n}{n(s-1)} = \frac{1}{n(s-1)}\left(1 - \frac{t'_+}{s-1} \right)_+^{s-1} \left(1 - \frac{t_+}{n} \right)_+^n $$ The conclusion follows from this computation and from the following fact : for any $w,w' \in ]0,1]$, one has $$ \sum_{\substack{2\leq s \leq wn \\ 1 \leq s' \leq w'(s-1)}} \frac{1}{n(s-1)}\left(1 - \frac{t'_+}{s-1} \right)_+^{s-1} \left(1 - \frac{t_+}{n} \right)_+^n \longrightarrow w w' e^{-t'_+ - t_+} , $$ as $n \rightarrow + \infty$.

$\endgroup$
3
  • $\begingroup$ I suppose we also might restrain ourself to $t,t'>0$? $\endgroup$
    – Thomas E
    Jun 30, 2016 at 10:58
  • $\begingroup$ Everything I wrote also holds for negative values of $t$ (except perhaps some strict inequalities between $U_s$'s becoming non-strict). $\endgroup$
    – js21
    Jun 30, 2016 at 11:01
  • $\begingroup$ I give a bounty, because this really was great help. $\endgroup$
    – Thomas E
    Jul 2, 2016 at 10:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.