Consider a probability space $(\Omega, \mathcal B, \mathbb P)$, and 2 random variables $X,Y:\Omega\to\mathbb R$

Conditional expectation $E(X|Y)=E(X|\sigma(Y))$ can be defined as a function that is

  1. $\sigma(Y)$ -measurable. I.e. there cannot be a subset $A$ of $\Omega$ on which $Y$ maps to a unique value, but where the conditional expectation does not.

  2. $E(E(X|Y)\cdot 1_A)=E(X\cdot 1_A)$ For all $A\in \sigma(Y)$.

How do we prove that this defines a function that is unique almost-everywhere?

  • $\begingroup$ I think it is defined differently in Las Vegas. $\endgroup$ – Michael Sep 8 '18 at 15:28

For uniqueness: Suppose $Z, Z'$ are two functions (random variables) that satisfy 1 and 2. Set $A = \{\omega: Z(\omega) \ge Z'(\omega)\}$. Since $Z,Z'$ are both $\sigma(Y)$ measurable, the same is true of $A$. Hence $E[Z 1_A] = E[X 1_A] = E[Z' 1_A]$ by 2. Subtracting, we have $E[(Z-Z') 1_A] = 0$. But $(Z-Z') 1_A$ is nonnegative everywhere, so we conclude $(Z-Z') 1_A = 0$ almost everywhere. That is to say that $Z \le Z'$ almost everywhere. By symmetry, we also get $Z' \le Z$ almost everywhere, so that $Z=Z'$ almost everywhere.

Existence is more difficult. The proofs I've seen either use the Radon-Nikodym theorem, or the Riesz representation theorem in Hilbert space. Any measure-theoretic probability book will have a proof.

| cite | improve this answer | |
  • $\begingroup$ Could you formally explain the step "so we conclude $(Z-Z') 1_A = 0$ almost everywhere"? I get the intuition, but I don't know how to formally state it. $\endgroup$ – user56834 Sep 10 '18 at 3:13
  • 1
    $\begingroup$ @Programmer2134: Theorem: Suppose $W$ is a nonnegative measurable function and $E[W] = 0$. Then $W=0$ almost everywhere. Proof: let $n$ be a positive integer and consider the set $A_n = \{ \omega : W(\omega) \ge 1/n\}$. Then $E[W] \ge P(A_n)/n$ (Markov's inequality). Hence $P(A_n) = 0$. Now $\{\omega : W(\omega) > 0\} = \bigcup_n A_n$ so by countable additivity this set also has measure zero. $\endgroup$ – Nate Eldredge Sep 10 '18 at 3:54

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.