3

I believe it is easier if I print the proof below:

enter image description here

Several other papers uses the same argument to bound $\lvert R_n(x)-E R_n(x) \rvert$ almost surely, so I believe it is correct. It should be clear that the aim of this result is to establish the (uniform) almost sure convergence of $R_n(x)$ to its expectation (but here, drop the uniformity part, for the sake of simplicity). It is not clear to me what "$O(a_n)$ almost surely" exactly means, but since the aim is show the rate that the almost surely convergence occurs, I believe that a suitable definition could be \begin{align} \lvert R_n(x)-E R_n(x) \rvert=O(a_n) \text{ almost surely}\\ \text{ iff } P(\lim\sup_n\{\lvert R_n(x)-E R_n(x) \rvert\leq Ca_n\})=1,\text{ for some }C>0 \end{align} this means that $\lim\sup_n\{\lvert R_n(x)-E R_n(x) \rvert\leq Ca_n\}$ holds $P$-almost everywhere. Please, correct me if you have further information.

As far as I understand, he is trying to give a bound from $\lvert R_n(x)-E R_n(x) \rvert\leq\lvert R_n(x) \rvert +\lvert E R_n(x) \rvert$.

As shown above, $\lvert E R_n(x) \rvert=O(a_n)$, or equivalentely, $\lvert E R_n(x) \rvert\leq C_1 a_n$ for some $C_1>0$ and all large $n$.

Further, he is saying that $\lvert R_n(x) \rvert=0$ almost everywhere (or, with probability one) when $n$ is large enough. This step is not clear to me, and I hope you can help me.

My attempt (Updated)

I'm on this problem for some days, but not suceeded. I'll try something different now.

I want to show that \begin{align} P(\lim\sup_n \{\lvert R_n-ER_n\rvert\leq C a_n\})=1 \end{align} for some $C>0$ with $a_n\to 0$. Note that \begin{align} &w\in\bigg\{\frac{1}{n}\sum_{i=1}^n Y_i1(\vert Y_i\rvert>\tau_n)>0\bigg\}\\ \implies &\exists i\in\{1,\dotsc,n\}:w\in\{\vert Y_i\rvert>\tau_n\}. \end{align} Using this result, the triangle inequality and the fact $\lvert E R_n\rvert=O(a_n)$, for any $n:n>n^*$, \begin{align} &P( \lvert R_n-ER_n\rvert>C_1 a_n) \\ \leq &P( \lvert R_n\rvert+C_1a_n>C_1 a_n)\\ =&P( \lvert R_n\rvert>0)\leq P(\lvert Y_i\rvert>\tau_n \text{ for some }i\in\{1,\dotsc,n\}) \end{align} If $Y_i$ were strictly stationary (he assumed $Y_i$ i.i.d.), then we could put $ P(\lvert Y_i\rvert>\tau_n \text{ for some }i\in\{1,\dotsc,n\}) = P(\lvert Y_n\rvert>\tau_n) $ and so \begin{align} \sum_{i=1}^\infty P( \lvert R_n-ER_n\rvert>C_1 a_n)&\leq n^*+\sum_{i=n^*}^\infty P( \lvert R_n-ER_n\rvert>C_1 a_n)\\ &\leq n^* +\sum_{i=n^*}^\infty P( \lvert Y_n\rvert>\tau_n)<\infty \end{align} The application of Borel-Cantelli's Lemma gives the result. Thanks in advance!

  • Are the curly brackets in all places that you use them, placed correctly? It appears that you put inside the curly brackets an inequality, which would then represent an event, but then you take limsup / liminf of this event? – Alecos Papadopoulos May 13 '20 at 10:05
  • @AlecosPapadopoulos I updated the "My attempt" section. This is how I view the above question today, which I believe is correct. About your question, what is inside the curly brackets is a set (the inverse image of a random variable or measurable function) which depends on $n$. Then I want to take the measure of the limsup of this set. – Celine Harumi May 13 '20 at 17:57

0 Answers0