2
$\begingroup$

I am trying to compute distribution of the following random variable \begin{align*} E[(X-E[X|Y])^2|Y] \end{align*} where $X \sim \mathcal{N}(0,\sigma^2_x)$ and $Z \sim \mathcal{N}(0,\sigma^2_Z)$ where $Z$ is independent of $X$ and $Y=X+Z$.

I know that $E[(X-E[X|Y])^2|Y]$ should have Gaussian distribution but I get lost in what is conditional expectation and what is random variable.

Here is my reasoning: I know that $E[X|Y=y]= \frac{\sigma_{XY}}{\sigma_Y^2}y=\frac{\sigma^2_x}{\sigma^2_x+\sigma^2_z}y$. So, then

\begin{align*} E[(X-E[X|Y])^2|Y]&=E\left[\left(X-\frac{\sigma^2_x}{\sigma^2_x+\sigma^2_z}Y\right)^2 \Big|Y\right]=E\left[\left(Y-Z-\frac{\sigma^2_x}{\sigma^2_x+\sigma^2_z}Y\right)^2 \Big|Y\right]\\ &= \frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}E[Y^2|Y]-2\frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}E[ZY|Y]+E[Z|Y]\\ &= \frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}Y^2-2\frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}YE[Z|Y]+E[Z|Y]\\ &= \frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}Y^2-2\frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}YE[Y-X|Y]+E[Y-X|Y]\\ &= \frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}Y^2-2\frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}(Y^2-YE[X|Y])+Y-E[X|Y]\\ &= \frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}Y^2-2\frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}(Y^2-\frac{\sigma^2_x}{\sigma^2_x+\sigma^2_z}Y^2)+Y-\frac{\sigma^2_x}{\sigma^2_x+\sigma^2_z}Y\\ &= -\frac{\sigma^2_z}{\sigma^2_x+\sigma^2_z}Y^2+\frac{2\sigma^2_z\sigma^2_x}{(\sigma^2_x+\sigma^2_z)^2}Y^2+Y-\frac{\sigma^2_x}{\sigma^2_x+\sigma^2_z}Y \end{align*}

but $Y^2$ can not be Gaussian. So, I am making a mistake somewhere. Could you guys help. Thank you

$\endgroup$
1
$\begingroup$

Note that $Y \sim {\mathcal N}(0, \sigma_y^2)$ where $\sigma_y^2 = \sigma_x^2 + \sigma_z^2$ and $\text{Cov}(X,Y) = \sigma_x^2$. Thus $X = \dfrac{\sigma_x^2}{\sigma_y^2} Y + W$ where $W$ is independent of $Y$ and $W \sim {\mathcal N}(0, \sigma_x^2 \sigma_z^2/\sigma_y^2)$. You have $E[X|Y] = \dfrac{\sigma_x^2}{\sigma_y^2} Y$, so $X - E[X|Y] = W$. The rest should be easy, since $W$ is independent of $Y$.

$\endgroup$
  • $\begingroup$ Thanks. So then $E[(X-E[X|Y])^2|Y]=E[W^2|Y]=E[W^2]$, right? But $E[W^2]$ is not a random variable it's just a number, right? $\endgroup$ – Boby Jun 26 '15 at 15:39
  • 1
    $\begingroup$ Worth mentioning is that this works because (and is quite specific to the case when) (X,Y) is gaussian. $\endgroup$ – Did Jun 26 '15 at 17:25
  • $\begingroup$ You guys seem very good with conditional expectation would you care looking at this question that I asked before: math.stackexchange.com/questions/1337638/… I am also planning to start a bounty on it. $\endgroup$ – Boby Jun 26 '15 at 20:25

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.