Consider $X$ a nonnegative random variable on some probability space $(\Omega, \mathcal{F}, \mathbb{P})$. We are given that

$$ A = \left\{ \omega : X(\omega) > 0 \right\} $$

and that $\mathcal{G}$ is a sub-sigma-field of $\mathcal{F}$. Can we conclude that $\mathbb{E}(X \ |\ \mathcal{G}) > 0$ for almost all $\omega \in A$?

I think the answer is no, but I'm having trouble coming up with a counterexample. One thing I tried was

$$ \Omega = \left\{ a, b , c \right\}, \; \mathcal{G} = \left\{ \emptyset, \Omega, \{a, b\}, \{c\}\right\} $$ with $X(\omega) = \mathbf{1}_{\{c\}}$ and $\mathbb{P}(\{c\}) = 0$, $\mathbb{P}(\{ a \}) = \mathbb{P}(\{ b \}) = 1/2$. Here, obviously $A = \{ c \}$ and

$$ \int_A \mathbb{E}(X | \mathcal{G}) \text{d}\mathbb{P} = \int_A X \text{d} \mathbb{P} = 0, \text{ since } \mathbb{P}(A) = 0$$

However, we can set $\mathbb{E}(X \ |\ \mathcal{G})$ equal to any positive number above for $\omega \in A$, so we cannot disprove the claim. Any help?

More generally, any pointer to resources with "pathological" cases and counterexamples in conditional probability and conditional expectation would be more than helpful.

  • $\begingroup$ There is no need to assume $X \geq 0$. Defining $A = \{\omega : X(\omega)>0\}$ is enough to conclude $E[X|\mathcal{G}]>0$ for all $\omega \in A$ except for a set of probability measure 0. $\endgroup$ – Michael May 21 '18 at 19:00

Let $X$ be any random variable (possibly one that can take negative values) and let $A = \{\omega : X(\omega) >0\}$. You can indeed conclude $E[X|\mathcal{G}]>0$ for all $\omega \in A$ except possibly a set of probability measure 0 (note that I have replaced your "almost all" with "except for a set of probability measure 0").

Suppose $B \subseteq A$, $P[B]>0$, and $E[X|\mathcal{G}]\leq 0$ for all $\omega \in B$. We want to reach a contradiction. Note that $B \cap\{\omega : X(\omega)>1/n\} \nearrow B$. So there is a positive integer $m$ such that $$P[B \cap \{\omega : X(\omega)>1/m\}] >0 \quad (Eq. *)$$
So \begin{align} 0&\overset{(a)}{\geq}\int_B E[X|\mathcal{G}]dP \\ &\overset{(b)}{=} \int_B X dP \\ &\overset{(c)}{\geq} \int_B (1/m)1_{\{X>1/m\}}dP \\ &= (1/m)P[B \cap \{\omega : X(\omega)>1/m\}] \\ &\overset{(d)}{>}0 \end{align} where (a) holds because $E[X|\mathcal{G}]\leq 0$ for all $\omega \in B$; (b) holds by definition of conditional expectation; (c) holds because $X(\omega) \geq (1/m)1_{\{X>1/m\}}$ for all $\omega \in B \subseteq A$; (d) holds by (Eq. *). The conclusion $0>0$ is the desired contradiction. $\Box$

However, a "counter-example" that relates to expectation is that it is not always the case: $$ E[X+Y]=E[X]+E[Y]$$ This equation is true whenever the right-hand-side does not lead to an undefined case of $\infty - \infty$. It fails, for example, if $E[X]=\infty$ and $Y=1-X$.

| cite | improve this answer | |

First of all, recall that $X \geq 0$ implies $\mathbb{E}(X \mid \mathcal{G}) \geq 0$ almost surely. Moreover, we know that

$$\forall G \in \mathcal{G}: \quad \int_G X \, d\mathbb{P} = \int_G \mathbb{E}(X \mid \mathcal{G}) \, d\mathbb{P};$$

in particular, we find for $G := \{\mathbb{E}(X \mid \mathcal{G}) = 0\}$

$$0 = \int_G \mathbb{E}(X \mid \mathcal{G}) \, d\mathbb{P} = \int_G X \, d\mathbb{P}.$$

As $X \geq 0$ this implies

$$X=0 \quad \text{almost surely on $\{\mathbb{E}(X \mid \mathcal{G})=0\}$}$$


$$\{\mathbb{E}(X \mid \mathcal{G})=0\} \subseteq \{X=0\} \quad \text{up to a null set}.$$

Taking the complement, we conclude that

$$\{X>0\} \subseteq \{\mathbb{E}(X \mid \mathcal{G})>0\} \quad \text{up to a null set},$$

and this proves the assertion.

| cite | improve this answer | |
  • 1
    $\begingroup$ (+1, but minor comment on first line): $X\geq 0 \implies E[X|\mathcal{G}]\geq 0$ [for all $\omega$ except possibly a set of measure 0]. $\endgroup$ – Michael May 21 '18 at 19:05
  • $\begingroup$ @Michael Thanks for your upvote and the comment; I've just fixed it. $\endgroup$ – saz May 21 '18 at 19:08

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.