Suppose we know that there exists sequences $a_n$ and $b_n$ such that $$\lim_{n \to \infty} F^{n}(a_n x+ b_n)=H(x)$$ with $F$ a distribution function and $H$ a continuous distribution function. Now let $x_n$ be a sequence with limit $x^{*}$. Furthermore it is known that $a_nx_n+b_n \leq a_{n+1}x_{n+1}+b_{n+1}$. I wish to prove that $\lim_{n \to \infty}F^n(a_nx_n+b_n)=H(x^{*})$

My proof: Since $F$ is a distribution function it lies between $0 $ and $1 $ and thus so does $F^n(a_nx_n+b_n)$. Since the sequence $a_nx_n+b_n$ is increasing and $F$ is monotonic we must have that $\lim_{n \to \infty}F^n(a_nx_n+b_n)$ exists. Let $m=n$, then $$\lim_{n \to \infty}F^n(a_nx_n+b_n)=\lim_{(m,n) \to \infty} F^n(a_nx_m+b_n)= \lim_{m \to \infty} \lim_{n \to \infty} F^n(a_nx_m+b_n)= \lim_{m \to \infty} H(x_m)=H(x^{*}) $$ where the last step follows by continuity of $H$.

I dont think the third equality can be justified because $m=n$ here, but it more or less illustrates how I tried to solve this problem.

  • $\begingroup$ What $F^n$ stands for? Is $n$ an exponent or it just symbolizes that there is a sequence $\{F_n\}$? $\endgroup$ Apr 1, 2018 at 21:01
  • $\begingroup$ @AlecosPapadopoulos it means that $F$ is raised to the power $n$. For context, the sequences $a_n$ and $b_n$ are used for normalizing the maximum of a random sample, which explains why the power is there. The author of the text I am looking at simply says that the result I try to prove here follows from monotonicity of $F$ and continuity of $H$, I am trying to fill in the details. $\endgroup$
    – Joogs
    Apr 1, 2018 at 21:08

1 Answer 1



$$G_n(x) \equiv \left[F(a_nx+b_n)\right]^n $$

We have per assumption that

$$G_n(x) \to H(x)$$

If we define the Identity function $g_n(x) = x_n,\;\;g_n(x) \to g(x) = x^*$, we wonder whether

$$G_n(g_n(x)) \to\;\;??\; H(g(x)) = H(x^*)$$

This is a "composition of function sequences" issue, and in general requires that the convergence of $G_n(x)$ be uniform (alongside pointwise convergence of $g_n(x)$), in order to obtain pointwise convergence of $G_n(g_n(x))$. Note that we have "compacted out of sight" the intermediate function/function-sequences compositions here, but that's ok, since ultimately $F$ is bounded.

Now, there is a "special-case" Theorem that states that if the limit distribution function is continuous, then "convergence in distribution" (which is nothing more than non-stochastic pointwise convergence, and only at the points of continuity of the limit function), implies also that the convergence of the distribution function sequence is uniform.

In our case $H(x)$ is assumed continuous, so $G_n(x_n)$ converges uniformly to it, so for the composition we have that

$$G_n(g_n(x)) \to H(g(x)) = H(x^*)$$

does indeed hold.

The theorem is stated in Lehmann, E. L. (2004). Elements of large-sample theory. Springer, p. 96 as "Theorem 2.6.1" without proof. The author refers to Parzen, E. (1960). Modern probability theory and its applications. John Wiley & Sons p. 438, where we find only a hint for a proof, specifically,

enter image description here

I just note that the necessity of $H(x)$ being continuous is rather easy to imagine: assume that at $x_d$, $H(x_d)$ is discontinuous (from the left). Then we will have $H(x_d-\epsilon) < H(x_d)$. Moreover, the "convergence in distribution" premise does not assert that $G_n(x_d) \to H(x_d)$, since $x_d$ is a discontinuity point for $H$. But then it may be the case that $H(x_d-\epsilon) < \lim G_n(x_d)< H(x_d)$, and then uniform convergence does not hold.


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.