Course contentsShow
← Mathematics
Lesson 2936 of 3,12158. Joint and Limit BehaviorFree lesson

Iterated Expectations and the Tower Property

Apply the law of iterated expectations to compute expectations by conditioning on intermediate random variables.

Iterated Expectations and the Tower Property

You'll learn how to compute expectations by "peeling back" layers of conditioning using the tower property of iterated expectations.

The Core Idea

The tower property (also called the law of iterated expectations) states that if you condition on an intermediate random variable and then take expectations again, you get back the original unconditional expectation:

E[X]=E[E[X∣Y]]E[X] = E[E[X \mid Y]]

In plain English: the expectation of a conditional expectation equals the unconditional expectation. Think of it like a telescope that can collapse: you can zoom into finer detail (conditioning on YY), then zoom back out (taking expectation over YY), and arrive right where you started.

Why It Works

When you compute E[X∣Y]E[X \mid Y], you get a function of YY — call it g(Y)g(Y). Now g(Y)g(Y) is itself a random variable. Taking its expectation means averaging over all possible values of YY, weighted by their probabilities. This averaging process reconstructs the Law of Total Probability for expectations.

When to Use It

The tower property shines when direct computation of E[X]E[X] is hard, but:

  1. You can find E[X∣Y]E[X \mid Y] for some cleverly chosen YY
  2. Then compute E[E[X∣Y]]E[E[X \mid Y]] more easily

Example context: Suppose XX counts total claims and YY counts the number of policyholders. Finding E[X∣Y=n]E[X \mid Y = n] (expected claims given nn policyholders) might be straightforward, then you average over the distribution of YY.

Extension

This extends to chains: E[X]=E[E[E[X∣Y]∣Z]]E[X] = E[E[E[X \mid Y] \mid Z]] when you condition on multiple variables sequentially — hence the name "tower."

Key Takeaway: The tower property lets you break complex expectations into manageable pieces by conditioning on intermediate random variables, then "averaging out" that conditioning to recover the original expectation.