1
$\begingroup$

I am given a set of i.i.d RVs $\{ X_1,X_2 \dots X_n\}$. Let $\mathbf{Z} = X_1+X_2 + \dots + X_n$ and define $\mathbf{E}_i(\mathbf{Z}) = \mathbf{E}(\mathbf{Z}|X_1,X_2\dots X_i)$ and $\Delta_i = \mathbf{E}_i-\mathbf{E}_{i-1} ~~(with~~ \mathbf{E}_0 = \mathbf{E})$.

In a book I follow, it says if $j > i$, then $\mathbf{E}_i(\Delta_j) = 0$. I tried to work it out, and at the end it boils down to calculating

$\mathbf{E}(\mathbf{E}( (\mathbf{Z}|X_1,\dots X_j))|X_1,\dots X_i)$. What is the definition of the conditional expectation of a conditioned variable, and how can I calculate it ?

$\endgroup$
  • 1
    $\begingroup$ Does the book mention something called tower rule or law of iterated expectation? $\endgroup$ – E-A Dec 30 '17 at 10:11
  • 1
    $\begingroup$ The object $\mathbf{E}( (\mathbf{Z}|X_1,\dots X_j)|X_1,\dots X_i)$ does not exist because $(\mathbf{Z}|X_1,\dots X_j)$ is not a random variable (actually, $(\mathbf{Z}|X_1,\dots X_j)$ does not exist at all). The object $\mathbf{E}( \mathbf{E}(\mathbf{Z}|X_1,\dots X_j)|X_1,\dots X_i)$ does exist. $\endgroup$ – Did Dec 30 '17 at 10:15
  • $\begingroup$ @Did, Yes, one $\mathbf{E}$ was missing in my equation. $\endgroup$ – Shew Dec 30 '17 at 12:46
2
$\begingroup$

Since your answer somewhat immediately follows from tower rule, I will try to give some insight into conditional expectation and how you can understand the tower rule.

Let me try to answer this in two levels:


Conditional expectation of $X$ given $Y$, also more commonly written as $\mathbb{E}[X \mid Y]$, is what you would guess $X$ to be if you know $Y$.

For instance, let $X = Y + Z$, where $Y$ and $Z$ are independent. Then, $\mathbb{E}[X \mid Y] = \mathbb{E}[Y + Z \mid Y]$. Now, let's think about this; what is the best guess for $Y + Z$ given $Y$? Well, it is $Y + \mathbb{E}[Z]$; i.e. it is what I know for $Y$ and then what I would expect Z to be.

Now, we can try to understand the tower rule in this simple perception: $ \mathbb{E}[\mathbb{E}[Z \mid X_1, X_2]\mid X_1]$ = $ \mathbb{E}[Z\mid X_1]$, since for the outer step of the iterated expectation on the left hand side, we give up our information about $X_2$ and simply stick with our knowledge of $X_1$. Similarly, $\mathbb{E}[\mathbb{E}[Z \mid X_1]\mid X_1, X_2]$ = $ \mathbb{E}[Z\mid X_1]$, since when we first evaluate the inner step of the iterated expectation of the left hand side, we give up our information about $X_2$.

Thus, we get the tower rule (or law of iterated expectation):

$\mathbb{E}[\mathbb{E}[Z \mid X_1]\mid X_1, X_2]$ = $ \mathbb{E}[\mathbb{E}[Z \mid X_1, X_2]\mid X_1]$ = $ \mathbb{E}[Z\mid X_1]$.


More formally, $\mathbb{E}[X \mid Y]$ is actually a shorthand for $\mathbb{E}[X \mid \sigma(Y)]$.

To understand this, we will need to understand what is meant by $\sigma(Y)$. $\sigma(Y)$ contains sets of outcomes (elements of your event space $\Omega$) that, if I know some "nice" set $Y$ belongs to, I know that all $w$ that could have mapped to that "nice" set is in $\sigma(Y)$. (for instance, $\sigma(Y)$ would contain the following sets: $A_1 = \{w \mid Y(w) = 5, w \in \Omega \}$, $A_2 = \{w \mid Y(w) > 10, w \in \Omega \}$, since if I tell you Y = 5, then you can safely say that the outcome that produced it is in $A_1$.)

(If you know what is meant by a Borel sigma algebra $\mathcal{B}$, $\sigma(Y)$ is the collection of sets $A \subset \Omega$, s.t. $A = Y^{-1} (B)$ for some $B \in \mathcal{B}$.)

Now, since the elements of $\sigma(Y)$ are subsets of $\Omega$, that means we can integrate over them using the probability measure that measures the sizes of sets in $\Omega$. $\int_{A_1} Y d\mu$ is a perfectly valid expression to write (if $A_1 \subset \Omega$). For instance, if $Y$ was an indicator for some set $A_2$, $\int_{A_1} Y d\mu = \int_{A_1} \mathbb{1}_{A_2} d\mu = \int_{A_1 \cap A_2} 1 d\mu = \mu(A_1 \cap A_2)$.

Now, $Z = \mathbb{E}[X \mid Y]$ is the random variable that is measurable with respect to $\sigma(Y)$ (i.e, $\sigma(Z) \subset \sigma(Y)$), that agrees with the integral of X whenever we integrate it over any set in $\sigma(Y)$. (For the purposes of this answer, just believe me when I say that this indeed exists and is unique)

Now, tower rule somewhat follows:

$\mathbb{E}[\mathbb{E}[Z \mid X_1]\mid X_1, X_2] = \mathbb{E}[Z\mid X_1]$, since $\mathbb{E}[Z \mid X_1]$ is measurable with respect to $\sigma(X_1, X_2)$.

$\mathbb{E}[\mathbb{E}[Z \mid X_1, X_2]\mid X_1] = \mathbb{E}[Z\mid X_1]$, since for any set $A$ in $\sigma(X_1)$ (which is also in $\sigma(X_1, X_2)$, we have $\int_A \mathbb{E}[Z \mid X_1, X_2] d\mu = \int_A Z d\mu = \int_A \mathbb{E}[Z \mid X_1] d\mu$.

As such, we get the same tower rule:

$\mathbb{E}[\mathbb{E}[Z \mid X_1]\mid X_1, X_2]$ = $ \mathbb{E}[\mathbb{E}[Z \mid X_1, X_2]\mid X_1]$ = $ \mathbb{E}[Z\mid X_1]$.

$\endgroup$
  • $\begingroup$ Your definition of $\sigma(Y)$ can be misleading the way you use the term "measurable". I would suggest rephrasing it a bit. And the phrase "measurable by" sounds unusual, but isn't necessarily wrong. Good answer though! $\endgroup$ – jdods Dec 30 '17 at 12:58
  • 1
    $\begingroup$ Thanks :D. Changed that around a little bit; hopefully I made that clearer... Also yeah, I guess technical terminology is "measurable with respect to", which I guess is there for a reason. (also yeah apparently at some point I had said measurable with respect to a random variable which kinda didn't make sense). $\endgroup$ – E-A Dec 30 '17 at 14:05

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.