Equivalence with norm of conditional expectation implies independence

57 Views Asked by At

I have the following problem. Upon staring at it, I think I've developed an intuition of how it should go but am unsure of the exact steps to follow.

Problem: Let $X\in L_2(\Omega, F,\mathbb{P})$ and $Y\in L_2(\Omega, G, \mathbb{P})$ where $G$ is a $\sigma$-field $\subseteq F$. Suppose that $Y=\mathbb{E}(X|G)$ and that $\lVert X\rVert _2 = \lVert Y \rVert_2$. Prove that in fact $X=Y$ almost surely.

Intuition: Through conditioning, $Y$ acts as a projection of $X$ onto the $G$ subspace. In laymen's terms, such a projection "reduces" or "coarsens" the original random variable, and hence we should expect its $L_2$ norm to decrease. However, that the $L_2$ norm remains the same shows that conditioning on $G$ has no effect and therefore $X=Y$ almost surely. Also, the $L_2$ norm amplifies the "coarsening" more than the $L_1$ norm, so it is possible that two random variables are equivalent in the latter but not in the former.

P.S. Apologies in advance if the title is wrong, as it is currently how I understand the problem. I will probably edit it to something more suitable later. Let me know if there is any better way to clarify what I've written above and I'll be glad to accomodate.

1

There are 1 best solutions below

0
On BEST ANSWER

I'll give you two ways of seeing this, one from basic definitions and a faster way. $\|X\|_2 = \|Y\|_2$ is equivalent to $E[X^2] = E[Y^2]$. But one of the defining properties of the conditional expectation (well, I suppose that this depends on your prefered text) is that for any $Z \in L_2(\Omega, G, P)$, $Y = E[X \mid G]$ is the (up to null) unique random variable $L_2(\Omega, G , P)$ such that $$ E[XZ] = E[YZ]. $$ In particular, $$ E[XY] = E[Y^2]. $$ On the other hand, $Y = E[X \mid G]$ also minimizes $$ E[(X - Y)^2] = E[X^2 + Y^2 - 2XY] $$ so if $E[X^2] = E[E[X \mid G]^2]$, then $E[(X - Y)^2] = 0$ is possible, and thus it must be that $X = Y$ almost surely.

Alternatively, $$E[Y^2] = E[E[X \mid G]^2] \leq E[E[X^2 \mid G]] = E[X^2]$$ where we use Jensen and then the tower property; conditional Jensen's is an equality in this case only if $X$ is already $G$-measurable or degenerate, in which case $Y = X$ almost surely.