Conditional Expectation and the Tower Rule

If linearity of expectation is the most-used tool in this course, conditioning is the second. The tower rule (law of total expectation) says you can compute an expectation in stages:

$$E[X] = E\big[\,E[X \mid Y]\,\big]$$

Average the conditional averages, weighted by how likely each condition is. Almost every "two-stage randomness" question is this identity applied once.

The basic move

"Roll a die, then flip that many fair coins. Expected heads?"

Condition on the roll: $E[\text{heads} \mid N = n] = n/2$. Then average over the roll:

$$E[\text{heads}] = E[N/2] = \frac{3.5}{2} = 1.75.$$

No enumeration of 6 × outcomes — the stage structure does the work. The general pattern for a random sum $S = X_1 + \cdots + X_N$ (random $N$ independent of the $X_i$) is Wald's identity:

$$E[S] = E[N] \cdot E[X]$$

Instant applications: expected total order flow (random number of trades × average trade size), total claims (random claim count × average claim), total P&L over a random number of bets. Variance has its own two-stage version (the law of total variance, $\operatorname{Var}(X) = E[\operatorname{Var}(X|Y)] + \operatorname{Var}(E[X|Y])$) — "within" plus "between" variance, worth knowing by name.

Conditional expectation as best prediction

$E[X \mid Y]$ is a random variable — a function of $Y$ — and it is the best predictor of $X$ given $Y$ in the least-squares sense. Regression (from the statistics course) is the linear approximation of exactly this object; regression to the mean is a statement about it. The martingale definition from earlier in this course ("$E[M_{t+1} \mid \mathcal{F}_t] = M_t$") now reads naturally: the best prediction of tomorrow is today.

Two properties used constantly, worth stating precisely:

• Taking out what is known: $E[f(Y) X \mid Y] = f(Y)\, E[X \mid Y]$.
• Independence drops the condition: if $X \perp Y$, then $E[X \mid Y] = E[X]$.

Conditioning as a problem-solving reflex

The craft is choosing what to condition on — the first event, the first success, the identity of a maximum, a symmetry-breaking piece of information:

• First-step conditioning built every waiting-time recurrence in this course (flips until HH, ant on the cube).
• Condition on the source: "Two dealers quote; one is sharp (spread 2), one is wide (spread 10), you can't tell which is which. Expected spread of a random quote?" — $\tfrac12(2) + \tfrac12(10) = 6$. Trivial with conditioning, murky without.
• The two-envelope paradox is a cautionary tale: conditioning done sloppily ("the other envelope has $2x$ or $x/2$, each with probability ½") smuggles in an improper prior and "proves" you should always switch. Diagnosing where that argument breaks — the conditional probabilities aren't ½ for any proper prior — is a genuinely strong interview answer.

The interview version

"A miner faces three doors: one exits in 2 hours; one loops back after 3 hours; one loops back after 5. He picks uniformly at random each time, with no memory. Expected time to exit?" — Condition on the first door: $E = \tfrac13(2) + \tfrac13(3 + E) + \tfrac13(5 + E)$, so $E = \tfrac{10}{3} + \tfrac23 E$, giving $E = \mathbf{10}$ hours. One equation, because conditioning on the first choice resets the problem. When a question has stages, layers, or a "then" in it — condition first, compute second.