4
$\begingroup$

Let a pair of random variables $(X, Y)$ over a finite product space $\mathcal{X}\times \mathcal{Y}$ be given. The conditional expectation operator is defined as $$(T_Yf) (y):=\mathbb{E}[f(X)|Y=y],$$ where $f$ is a real-valued function acting on $\mathcal{X}$.

It is well known the the operator $T$ is contractive in $L^p$-norm for $p\geq 1$, i.e., $$||Tf||_p\leq ||f||_p,$$ which can be easily proved using convexity of the map $t\mapsto t^p$ and Jensen's inequality.

I am looking at the map $p\mapsto \frac{||Tf||_p}{||f||_p}$ for $p\geq 1$. I am trying to see if this map is monotonic or not. Any idea?

$\endgroup$
4
  • $\begingroup$ Does this monotonicity break for an arbitrary contractive operator? $\endgroup$
    – Suvrit
    Mar 17, 2015 at 1:16
  • $\begingroup$ @suvrit, I focus only on conditional expectation operator. For arbitrary contractive operator, I guess, the monotonicity breaks. $\endgroup$ Mar 17, 2015 at 2:53
  • $\begingroup$ I asked, because it is not clear from the outset, where would the fact that $T$ is a conditional expectation operator play a role for monotonicity --- it seems to be used mostly for the contractivity. On the other hand, even non-contractive operators can lead to a monotonic ratio, so maybe your hypothesis does play a role... $\endgroup$
    – Suvrit
    Mar 17, 2015 at 3:11
  • $\begingroup$ By "over a finite product" you probably mean "with values in"? $\endgroup$ Mar 17, 2015 at 7:50

1 Answer 1

3
$\begingroup$

What if $X=\{1,2\}$ and $Y=\{1,2,3\}$. Equip both $X$ and $Y$ with the uniform probability measure. Set $f(x,1)=1$, $f(x,2)=x$ and $f(x,3)=2$. Then $Tf(x,1)=1$; $Tf(x,2)=1.5$ and $Tf(x,3)=2$.

Notice that $\|f\|_1=\|Tf\|_1=1.5$ and $\|f\|_\infty=\|Tf\|_\infty=2$, but $\|f\|_2>\|Tf\|_2$.

$\endgroup$
1
  • $\begingroup$ Thanks for your answer. Now we know the map $p\mapsto \frac{||Tf||_p}{||f||_p}$ is not monotonic. How about the map $p\mapsto \sup_{f\neq 0}\frac{||Tf||_p}{||f||_p}$? $\endgroup$ Mar 17, 2015 at 19:02

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.