4
$\begingroup$

Consider some differentiable function $f(X,Y)$, where $X$ and $Y$ are scalar RVs. Assume that $\frac{\partial^2 f}{\partial X^2}>0$ but that nothing is known about the rest of the Hessian of $f$. Can we nonetheless make the following claim, which is effectively Jensen's inequality in "one direction"? $$ E[f(X,Y)]>E[f(E[X],Y)] $$

$\endgroup$

1 Answer 1

3
$\begingroup$

Yes, that is true. For any fixed $y$, define $g_y(X) = f(X, y)$. Note that $g_y$ is a function only of $X$, and is strictly convex in it. Therefore, for any $y$,

$$ E_X[f(X, y)] = E_X[g_y(X)] > g_y(E[x]) = f(E_X[X], y). $$

So, for any $y$,

$$ E_X[f(X, y)] > f(E_X[X], y). $$

However, the LHS and RHS are RVs of $Y$, and you can write

$$ E_Y\left[ E_X[f(X, y)] \right] > E_Y \left[ f(E_X[X], y) \right]. $$

$\endgroup$
2
  • $\begingroup$ Just to be clear, to complete the argument, we'd apply the law of iterated expectations as follows: $E[f(X,y)] > f(E[X],y)\Rightarrow E[E[f(X,Y)\vert Y=y]]>E[E[f(E[X],Y)\vert Y=y]],$ which yields the expression given in the original post? $\endgroup$
    – sunga
    Jul 1, 2017 at 2:52
  • $\begingroup$ @sunga Exactly. The inequality holds for any $y$, but that means the LHS and RHS are random variables in $Y$. The law of iterated expectations completes it. $\endgroup$
    – Ami Tavory
    Jul 1, 2017 at 3:02

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.