I am trying to prove the following statement, but I do not know how to go on:

Let $F(x)$ be an arbitrary continuous distribution function. Then there are constants $A_n, B_n > 0$ such that, as $n \to \infty$, \begin{align*} \lim P(X_{N(n)} < A_n + B_n x) =: T(x) \end{align*} exists and is nondegenerate if, and only if, \begin{align*} \lim_{n \to \infty} \frac{U_F(A_n + B_nx) - n}{\sqrt n} =: g(x) \end{align*} exists and is finite on an interval, where $g(x)$ has at least two points of increase. When it exists, $T(x) = \Phi(g(x))$.

Some remarks: Let $X_1, X_2, \ldots, X_n$ be i.i.d. random variables with continuous distribution $F(x)$, then $N(n)$ is defined to be \begin{align*} N(n) := \min\{j : j > N(n-1), X_j > X_{N(n-1)}\}. \end{align*} Furthermore it is \begin{align*} U_F(x) := -\log(1 - F(x)) \end{align*} and $\Phi(x)$ is the standard normal distribution function. I know that $F(X_j)$ is uniformly distributed on $(0,1)$ and that $U_F(X_j) \sim \text{Exp}(1)$.

I also know the following fact:

Let $F(x) = 1-e^{-x}$, $x > 0$. Let $X_{N(n)}$, $n > 1$ be the records and let $Y_1 = X_{N(1)} = X_1$, $Y_j = X_{N(j)} - X_{N(j-1)}$, $j \ge 2$. Then $Y_1, Y_2, \ldots$ are i.i.d. random variables and their common distribution function is $F(x)$ itself.

Since $X_{N(n)} = \sum_{k = 1}^n Y_k$, by the Central limit theorem it is \begin{align*} \lim_{n \to \infty} P(X_{N(n)} < n + x \sqrt n) = \Phi(x). \end{align*}

Questions: 1. How do I need to define the constants $A_n$ and $B_n$?

  1. If I want to show that $T(x)$ is nondegenerate, do I need to show that $T'(x) \ge 0$ for all $x$?

  2. I really need some help to start.

Thanks in advance.


Basically you already wrote the pieces yourself: if $(X_i)$ is i.i.d. with continuous $F$, then $Z_i=U_F(X_i)$ defines a sequence $(Z_i)$ of i.i.d. standard exponential random variables.

Furthermore, $Z_{N(n)}=U_F(X_{N(n)})$ hence, for every $(x_n)$, $$[X_{N(n)}\leqslant x_n]=[Z_{N(n)}\leqslant U_F(x_n)].$$ Finally, $Z_{N(n)}$ is distributed like the sum of $n$ i.i.d. standard exponential random variables hence, by the central limit theorem, $Z_{N(n)}=n+\sqrt{n}G_n$ where $G_n$ converges in distribution to the standard normal distribution.

Can you conclude?

  • $\begingroup$ So we have as a fact that $P[X_{N(n)} \le A_n+B_nx] = P[G_n \le (U_F(A_n+B_nx) - n)/\sqrt n]$ converges to $\Phi(g(x))$ as $n \to \infty$ if $(U_F(A_n+B_nx) - n)/\sqrt n \to g(x)$. So here we assumed that $g(x)$ exists. On the other hand, if we know that $P[X_{N(n)} \le A_n+B_nx]$ does converge, then $g(x)$ must exist. $T(x)$ must be nondegenerate, because it is a limit distribution? Why must $g(x)$ be finite on an interval? I also do not see why $g(x)$ needs to have two points of increase. And do I understand the statement, if I say that $A_n$ and $B_n$ need to exist if $\endgroup$ – numerion Oct 1 '14 at 15:01
  • $\begingroup$ $(U_F(A_n+B_nx)-n)/\sqrt n$ converges to some function $g(x)$ with exactly those $A_n$ and $B_n$? Does that mean, we only need to secure that $(U_F(A_n+B_nx)-n)/\sqrt n$ converges with those constants? I am still confused by the existence of $A_n$ and $B_n$. Do they exist in general? $\endgroup$ – numerion Oct 1 '14 at 15:03

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.