5
$\begingroup$

I'm trying to solve an exercise form Rick Durrett's book on probability, following a section on almost sure convergence of martingales:

Let $Z_1,Z_2,...$ be i.i.d standard normal random variables, let $\theta$ be an independent random variable with finite mean, and let $Y_i:=Z_i+\theta$. $\;$ Show that $\mathbb{E}[\theta \vert Y_1,...,Y_n] \overset{n\rightarrow \infty}{\rightarrow}\theta$ almost surely.

I know that $\mathbb{E}\Big[ \frac{1}{n} \sum_{i=1}^n Y_i \Big \vert Y_1,...,Y_n \Big]= \frac{1}{n} \sum_{i=1}^nY_i$, and that by the strong law of large numbers we know that $\frac{1}{n} \sum_{i=1}^nY_i \overset{n\rightarrow \infty}{\rightarrow}\theta$ almost surely. I'm unsure as to how best to proceed from here, or in fact how to use the independence of $\theta$. $\;$ I would appreciate any help or hints.

$\endgroup$
1
  • $\begingroup$ Isn't it a direct consequence of theorem 5.5.7: suppose $\mathcal{F}_n\uparrow\mathcal{F}_\infty$, i.e., $\mathcal{F}_n$ is an increasing sequence of $\sigma$-algebras and $\mathcal{F}_\infty=\sigma(\cup_n\mathcal{F}_n)$ then $E(X|\mathcal{F}_n)\to E(X|\mathcal{F}_\infty)$ almost surely? $\endgroup$ Feb 1 '18 at 20:17
2
$\begingroup$

$\newcommand{\F}{\mathcal{F}}\newcommand{\P}{\mathbb P}\newcommand{\E}{\mathbb E}$Edit. First I was not very careful enough with the measurability of the almost surely limit.

Define: \begin{align} \F_n=\sigma(Y_1,...,Y_n) \end{align} As noted in the comments we have: \begin{align} \E[\theta|\F_n]\to \E[\theta|\F_\infty] \ \ \ \ \ \text{ a. s.} \end{align} But unfortunately that is not enough. We don't know whether $\E[\theta|\F_\infty]=\theta$ holds. That is what we will show. Before showing that define $$X_n:= \frac{1}{n}\sum_{k=1}^n Y_k$$ Clearly $X_n$ is $\F_\infty$-measurable.

Conditioning on $\theta$ we know that $Y_i\sim \mathcal N(\theta,1)$ i.i.d. So: \begin{align} \E[X_n|\theta]=\theta \ \ \ \ \ \text{ and } \ \ \ \ \ \E[(X_n-\theta)^2|\theta]=\frac{1}{n} \end{align} Taking the expectation of both sides of the latter: \begin{align} \E[(X_n-\theta)^2]=\frac{1}{n} \end{align} That implies $X_n\to \theta $ in $L^2$. One knows that there is a subsequence $X_{n_k}$ that converges to $\theta$ a. s.. There is $\hat\theta$ which is $\F_\infty $-measurable and $\theta=\hat \theta $ a. s. (why?). Hence $$\E[\theta|\F_\infty] = \E[\hat\theta |\F_\infty] =\hat\theta$$ Conclusion: \begin{align} \E[\theta|Y_1,...,Y_n]\to \theta \ \ \ \ \ \text{ a. s. } \end{align}

$\endgroup$
3
  • $\begingroup$ This may be something trivial which I am not seeing, but why is $\mathbb{E}[Y_i\vert \theta] \sim \mathcal{N}(0,1)$ and i.i.d? $\endgroup$ Feb 2 '18 at 6:33
  • 1
    $\begingroup$ @Keen-ameteur I did not claim $E[Y_i|\theta] $ are iid normally distributed, but $Y_i|\theta =t\sim N(t, 1)$. Okay, $\theta$ is a random variable. If I would now it's value, say $\theta=t$ then $Y_i=Z_i+\theta=Z_i+t$, right? Since $Z_i$ are iid standard normally distributed, so are $Y_i$ iid normally distributed with mean $t$. Actually I should say $Y_i|\theta=t\sim N(t, 1)$ iid. Is it clear? $\endgroup$
    – Shashi
    Feb 2 '18 at 10:01
  • $\begingroup$ @Keen-ameteur $E[Y_i|\theta] =\theta$. So it's not normally distributed as long as $\theta$ is not normally distributed. The statement I have made is quite different as you could see in my previous comment. $\endgroup$
    – Shashi
    Feb 2 '18 at 10:08

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.