Convergence in probability iff convergence for every bounded continuous function

874 Views Asked by At

I'm trying to show the following:

$X_n \overset{p}{\to}X \iff f \circ X_n \overset{p}{\to} f \circ X$ for every continuous, bounded function $f$.

I can show ($\Rightarrow$) already using the usual almost-everywhere converging subsequences approach. The other direction is giving me difficulty.

I believe I can show it if $f$ is strictly increasing: then $f$ is a homeomorphism of $\mathbb{R}$ onto some interval $(a,b)$, thus $f^{-1}$ is continuous, so it suffices to apply the continuous function $f^{-1}$ to $f\circ X_n$. (Though if this is incorrect please do let me know.)

For the general case, however, I'm not quite sure how to proceed. My two questions are:

1) How to show $(\Leftarrow)$ when I can't rely on $f$ being strictly monotone?

2) What is the intuition behind needing $f$ to be bounded? I feel like this should be obvious but I can't seem to intuit why.

Many thanks.

Disclaimer: I'm pursuing self-study in Cinlar's 'Probability and Stochastics' in my spare time, this is problem 3.14.

1

There are 1 best solutions below

7
On BEST ANSWER

Fix $\epsilon>0$ and $r>0$. We want to show

$$\mathbb{P}(|X_n-X| \geq r) \leq \epsilon \qquad \text{for all $n$ sufficiently large.}$$

Set

$$\chi_R(x) := (-R) \vee x \wedge R = \begin{cases} R, & x \geq R, \\ x, & x \in [-R,R] \\ -R,, & x \leq -R \end{cases}.$$

Obviously, $\chi_R$ is bounded and continuous. Moreover, as $\chi_R(x)=x$ for all $|x| \leq R$, we have

$$\mathbb{P}(|X_n-X| \geq r) \leq \mathbb{P}(|\chi_R(X_n)-\chi_R(X)| \geq r) + \mathbb{P}(|X_n| \geq R) + \mathbb{P}(|X| \geq R).$$

Now

$$\begin{align*}\mathbb{P}(|X_n| \geq R) &= \mathbb{P}\left(|X_n| \geq R, |X| \leq \frac{R}{2} \right) + \mathbb{P} \left(|X|> \frac{R}{2} \right) \\ &\leq \mathbb{P} \left( |\chi_R(X_n)| \geq R, |\chi_R(X)|=|X| \leq \frac{R}{2} \right) + \mathbb{P} \left(| X|> \frac{R}{2} \right) \\ &\leq \mathbb{P} \left( |\chi_R(X_n)-\chi_R(X)| \geq \frac{R}{2} \right) + \mathbb{P} \left(| X|> \frac{R}{2} \right) \end{align*}$$

Combining both estimates, we obtain

$$\begin{align*} \mathbb{P}(|X_n-X| \geq r) &\leq \mathbb{P}(|\chi_R(X_n)-\chi_R(X)| \geq r)+ \mathbb{P}\left(|\chi_R(X_n)-\chi_R(X)| \geq \frac{R}{2} \right) \\ &\quad + \mathbb{P}(|X| \geq R) + \mathbb{P} \left(| X|> \frac{R}{2} \right) \end{align*}$$

Choose $R$ sufficiently large such that

$$\mathbb{P} \left(| X|> \frac{R}{2} \right) \leq \epsilon$$

For this (fixed!) $R$, it follows from our assumption that we can choose $N \in \mathbb{N}$ such that

$$ \mathbb{P}(|\chi_R(X_n)-\chi_R(X)| \geq r)+ \mathbb{P}\left(|\chi_R(X_n)-\chi_R(X)| \geq \frac{R}{2} \right) \leq \epsilon$$

for all $n \geq N$. Adding all up, we get

$$\mathbb{P}(|X_n-X| \geq r) \leq 3 \epsilon$$

for all $n \geq N$.

Concerning boundedness: The boundedness of $f$ is not needed. Obviously, the equivalence holds also if we drop the assumption "bounded" (however, then "$\Leftarrow$" is obvious.)