Weak convergence when different parts of the the support of the distribution "converge at different rates"

62 Views Asked by At

Let $F$ be the CDF of a distrubtion, supported on $[-1,1]$. It has a density $f$ (See footnote 1). Consider the sequence of finitely supported distributions defined as follows.

Take $t\in [-1,1]$, $n \in \mathbb{N}$. Define:

$$\widetilde{f}_n(t)=\begin{cases}f(t),\;\text{if}\;|t|=\frac{1}{n}+\frac{k}{2^n}\;\text{for some}\;k \in \mathbb{N} \cup \{0\},\\ 0,\;\text{otherwise}\;\end{cases},\\ \\ f_n(t)=\frac{\widetilde{f}_n(t)}{\sum\limits_{\widetilde{f}_n(t')>0}\widetilde{f}_n(t')}$$

My question is: Does the probability distribution with probability of $t=f_n(t)$, converge (weakly) to $F$?

As you can see, in the sequence $f_n$, the distance between points in the positive or negative part of the support of $f_n$ $(1/2^n)$, reduces at a rate faster than the distance between the positive and negative parts of the support $(2/n)$.

I think the answer is yes, using the standard definition of distributional convergence as pointwise convergence of the CDF (in this case, because all points are points of continuity). But I still wanted to double check what exactly, if anything, we need to be careful about, regarding the differing rates of convergence in different parts of the support.

(Edit) Footnote 1: We can assume $f(t)>0$ for all $t \in [-1,1]$. I did not edit the question as it was, because it contained a mistake as pointed out by Snoop in the first answer. However I would like to allow ourselves to make the aforementioned simplifying assumption so we can focus on the role played by the differing rates of convergence.

1

There are 1 best solutions below

5
On BEST ANSWER

The answer seems to be no for me. Let $f(t)=g(t)\mathbb{1}_{(\mathbb{R}/\mathbb{Q})\cap [-1,1]}(t)$ for some Lebesgue density $g$ on $[-1,1]$ with cdf $G(t)=\int_{[-1,t]}g(s)\lambda(ds)$. Then $F(t)=\int_{[-1,t]}f(s)\lambda(ds)=\int_{[-1,t]}g(s)\lambda(ds)=G(t)$. However, $f_n(t)=0/0$ for all $n$ and $t \in [-1,1]$.