14
$\begingroup$

Let $(\Omega,\mathcal F,\mu)$ be a probability space. It is well-known that if $\mathcal A$ is a sub-$\sigma$-algebra of $\mathcal F$, $p\geqslant 1$ and $X$ is an element of $\mathbb L^p$ which takes real values, then $$\tag{1} \lVert\mathbb E[X\mid\mathcal A]\rVert_p\leqslant \lVert X\rVert_p.$$

We now consider for $p\gt1$ the weak $\mathbb L^p$-space, denoted by $\mathbb L^{p,\infty}$, which consists of all the real valued random variables such that $\lVert X\rVert_{p,\infty}^*:= \left(\sup_{t\gt 0 }t^p\mu\{|X|\gt t\}\right)^{1/p} $ is finite. It is known that $\lVert \cdot \rVert_{p,\infty}^*$ is not a norm because the triangle inequality fails. However, defining $$\lVert X\rVert_{p,\infty}:=\sup\left\{\mu(A)^{1/p-1}\mathbb E[| X|\cdot\mathbb 1_{A} ], A\in\mathcal F,\mu(A)\gt 0 \right\},$$ then $\lVert X\rVert_{p,\infty}$ is a norm and $$\tag{2} \lVert X\rVert_{p,\infty}^*\leqslant \lVert X\rVert_{p,\infty}\leqslant \frac p{p-1}\lVert X\rVert_{p,\infty}^*.$$


Questions:

  1. Does the inequality $$\lVert\mathbb E[X\mid\mathcal A]\rVert_{p,\infty} \leqslant \lVert X\rVert_{p,\infty}$$ hold in general?
  2. If not, is there a norm $N_p$ on $\mathbb L^{p,\infty}$ which is equivalent to $\lVert\cdot\rVert_{p,\infty} $ and for which the inequality $$N_p\left(\mathbb E[X\mid\mathcal A]\right)\leqslant N_p(X) $$ takes place?

Motivations. In general, subadditive sequences have good properties. If $T\colon\Omega\to\Omega$ is a measure preserving map and $\mathcal M$ is a sub-$\sigma$-algebra of $\mathcal F$ such that $T \mathcal M\subset \mathcal M$, then defining $S_n(f):=\sum_{j=0}^{n-1} f\circ T^i$ ($f$ is $\mathcal M$-measurable), the sequence $\left(\lVert\mathbb E\left[S_n(f)\mid \mathcal M\right]\rVert_p\right)_{n\geqslant 1} $ is subadditive for any $p\geqslant 1$. A similar result for weak $\mathbb L^p$ norms would be nice.


For the first question, we can show that $$t\mu\{\mathbb E[X\mid\mathcal A]>t\}^{1/p} \leqslant \frac 1{\mu\{\mathbb E[X\mid\mathcal A]>t\}^{1-1/p}}\mathbb E[f\mathbb{1}\{\mathbb E[X\mid\mathcal A]>t\} ],$$ hence $$\lVert \mathbb E[X\mid\mathcal A]\rVert_{p ,\infty}^*\leqslant \lVert f\rVert_{p,\infty},$$ and it follows from (2) that $$\lVert \mathbb E[X\mid\mathcal A]\rVert_{p ,\infty}\leqslant \frac p{p-1} \lVert f\rVert_{p,\infty},$$ so the question is actually whether we can get rid of the factor $\frac p{p-1}$.

We can try to show this when $f$ is a simple function, but I fail to see how to apply efficiently convexity inequalities.

$\endgroup$
  • $\begingroup$ en.wikipedia.org/wiki/Interpolation_space#Real_interpolation $\endgroup$ – Alexander Shamov Jan 28 '15 at 20:32
  • $\begingroup$ @AlexanderShamov Indeed, it seems that my question is related to interpolation of Banach spaces. However, I fail to see how to apply one of the methods of interpolation in the link. For example, the connexion of the $\lVert \cdot\rVert_{p ,\infty}$ with the obtained norm in the interpolated space is not clear (to me). Do you have a precise idea in how to apply the theorem? Anyway I will study the book by Bennett and Sharpley. $\endgroup$ – Davide Giraudo Jan 28 '15 at 22:10
  • 3
    $\begingroup$ If I'm not mistaken, in the case $X_0 = L^1$, $X_1 = L^\infty$, $\theta = 1 - \frac{1}{p}$, the functional $K(x, t, X_0, X_1)$ defined in that Wikipedia article coincides with $\sup_{\mu[A] = t} \mathsf{E} |x| \mathsf{1}[A]$. Which makes the $\Vert \cdot \Vert_{p,\infty}$ norm exactly equal to the $K(\theta,\infty)$ interpolation norm. $\endgroup$ – Alexander Shamov Jan 29 '15 at 15:42
  • $\begingroup$ You are right and i can conclude from Theorem 1.11 301 of Bennett and Shapley. Maybe you could post this as an answer. $\endgroup$ – Davide Giraudo Jan 29 '15 at 16:51
  • 3
    $\begingroup$ The Calderon-Mitiagin interpolation theorem says that all rearrangement invariant function spaces are exact interpolation spaces between $L_1$ and $L_\infty$. This is Theorem 2.2 in Bennett-Sharpley and can be found in many other books as well. $\endgroup$ – Bill Johnson Mar 30 '15 at 0:05
1
$\begingroup$

Yes. You can modify your definition of the norm as: $$\|X \|_{p,\infty}=\sup [(\int_\Omega \phi )^{1/p-1}\mathbb{E}(|X|\phi), \text{ } 0\leq \phi \leq 1] $$ Then $$\|\mathbb{E}(X|\mathcal{A}) \|_{p,\infty}=\sup [(\int_\Omega \mathbb{E}(\phi|\mathcal{A}))^{1/p-1}\mathbb{E}(|X|\mathbb{E}(\phi|\mathcal{A})), 0\leq \phi \leq 1] $$

And we can conclude because $0\leq \phi \leq 1 \Rightarrow 0\leq \mathbb{E}(\phi|\mathcal{A})\leq 1$

$\endgroup$
-1
$\begingroup$

my typesetting was so bad it wasn't worth posting my answer but the suggestion was to apply neyman pearson lemma to show that you can replace sup over sets A with only sets of form A=(X>C), then compute norm of E(X|A). I did assume a r.v.s had continuous distributions if convenient.

$\endgroup$
  • $\begingroup$ Thanks for your reply. Could you include the statement of Neyman Pearson's lemma? $\endgroup$ – Davide Giraudo Jan 27 '15 at 18:10
  • $\begingroup$ en.wikipedia.org/wiki/Neyman%E2%80%93Pearson_lemma in this case take likelihood ratio proportional to X, then it says maximize E(X; A) over P(A) = c by choosing A = (X > A), which is the set where the likelihood ratio, X, is large. $\endgroup$ – mike Jan 27 '15 at 18:50
  • 3
    $\begingroup$ Could you please give more details? For the moment I fail to see the connection between likelihood ratios and the context of my question. $\endgroup$ – Davide Giraudo Jan 28 '15 at 16:04

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.