8
$\begingroup$

$(\Omega, \mathcal{F}, P)$ is a probability space. $X$ is a r.v. defined on it, and $\mathcal{G}_1, \mathcal{G}_2$ are two $\sigma$-algebra, can we claim the following: $$ \mathbb{E}\{\mathbb{E}[X|\mathcal{G}_1]|\mathcal{G}_2\}=\mathbb{E}\{\mathbb{E}[X|\mathcal{G}_2]|\mathcal{G}_1\}=\mathbb{E}[X|\mathcal{G}_1\cap\mathcal{G}_2]. $$ If this is not true, why the following is true: $X_1,X_2,\ldots,X_n$ are independent r.v. on the same probability space, $Z=f(X_1,X_2,\ldots,X_n)$ is another r.v. And we have: $$ \mathbb{E}\{\mathbb{E}[Z|X_1,\ldots,X_{i-1},X_{i+1},\ldots,X_n]|X_1,\ldots,X_i\}=\mathbb{E}[Z|X_1,\ldots,X_{i-1}]. $$

$\endgroup$

1 Answer 1

5
$\begingroup$

It's not true. For a counterexample, take $\Omega = \{a,b,c\}$ to be a sample space with 3 points, $\mathcal{F} = 2^{\Omega}$, and $P(A) = \frac{1}{3} |A|$ to be the uniform probability measure assigning probability 1/3 to each outcome. Let's represent a random variable $X : \Omega \to \mathbb{R}$ as the ordered triple $(X(a), X(b), X(c))$.

Set $\mathcal{G}_1 = \{\Omega, \emptyset, \{a\}, \{b,c\}\}$, and $\mathcal{G}_2 = \{\Omega, \emptyset, \{b\}, \{a,c\}\}$. Let $X$ be the random variable $(1,2,3)$. Then one can directly compute $$\begin{align*} E[X \mid \mathcal{G}_1] &= (1, 2.5, 2.5) \\ E[X \mid \mathcal{G}_2] &= (2,2,2) \\ E[E[X \mid \mathcal{G}_1] \mid \mathcal{G}_2] &= (1.75,2.5,1.75) \\ E[E[X \mid \mathcal{G}_2] \mid \mathcal{G}_1] &= (2,2,2). \end{align*}$$

To prove the second claim, I assert the following: suppose $U,V,W$ are mutually independent random variables or vectors, and $Y$ is $\sigma(U,V)$-measurable and integrable. Then $E[Y \mid V,W] = E[Y \mid V]$.

Clearly $E[Y \mid V]$ is $\sigma(V,W)$ measurable, so it suffices to show $E[Y 1_A] = E[E[Z \mid V] 1_A]$ for any $A \in \sigma(V,W)$. Suppose first that $A$ is of the form $A = B \cap C$ for $B \in \sigma(V)$, $C \in \sigma(W)$. Then $$\begin{align*} E[E[Y \mid V] 1_A] &= E[E[Y\mid V] 1_B 1_C] \\ &= E[E[Y 1_B \mid V] 1_C] && \text{since $B \in \sigma(V)$} \\ &= E[E[Y 1_B \mid V]] E[1_C] && \text{since $1_C \in \sigma(W)$ is independent of $\sigma(V)$} \\ &= E[Y 1_B] E[1_C] \end{align*}$$ while $E[Y 1_A] = E[(Y 1_B) 1_C] = E[Y 1_B] E[1_C]$ since $Y 1_B \in \sigma(U,V)$ is independent of $\sigma(W)$. The general case follows by a monotone class or $\pi$-$\lambda$ argument (the collection of all $B \cap C$ is a $\pi$-system that generates $\sigma(V,W)$).

Now apply this taking $U = (X_{i+1}, \dots, X_{n})$, $V = (X_{1}, \dots, X_{i-1})$ and $W = X_i$, and $Y = E[Z \mid U,V]$.

$\endgroup$
4
  • $\begingroup$ Thank you so much for your counterexample. But how can we easily see that the second assertion is true? $\endgroup$
    – edulearner
    Oct 15, 2015 at 19:02
  • $\begingroup$ @edulearner: See edit. $\endgroup$ Oct 15, 2015 at 19:03
  • $\begingroup$ I see, the proof is nice. I think that is not a trivial thing, but the paper which I looked at just states this without any remark as if it is evident. $\endgroup$
    – edulearner
    Oct 15, 2015 at 19:33
  • $\begingroup$ It's not completely trivial, no, but I think facts of this kind are pretty well-known among probabilists - it's the sort of thing that one would typically have done as an exercise in a graduate course. $\endgroup$ Oct 15, 2015 at 19:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.