In my class I am tasked to prove that given any $X,Y$ r.v. $Var(E[X|Y]) \le Var(X)$ with only the properties of the conditional expected value.

I looked up the proof for the total variance law but I don't want to use it since we have not yet defined conditional variance.

my attempt to prove this goes as follows:

$$0 \le Var(X) = E(X^2) - E(X)^2 = E(E(X^2|Y)) - E(E(X|Y))^2 = Var(E(X|Y)) + E(E(X^2|Y)) - E(E(X|Y)^2) $$

so this reduces to prove that $E(E(X^2|Y)) \ge E(E(X|Y)^2)$.

To prove this it seemed simple at first, call $G$ the sigma algebra on which $Y$ is defined

then $$E[E[X^2 | G]1_{G_1}] = E[X^21_{G_1}] \ge E[X1_{G_1}]^2 = E[E[X | G]1_{G_1}]^2 \quad \forall{G_{1}} \in G$$

But this is not what I want even if it's close. How could I adjust my reasoning?

Any other paths to the solution are welcome.


1 Answer 1


You need the conditional form of Jenson's inequality:

$$f(E[X|Y])\leq E[f(X)|Y],$$

whenever $f$ is convex. You can find a proof here.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.