4
$\begingroup$

Let $X$ be a random variable on the probability space $(\Omega, \mathcal{F}, P)$ and $\mathcal{A} \subset \mathcal{B} \subset \mathcal{F}$ be a $\sigma$-subalgebras.

I want to prove that if $$ \mathrm{E}[e^{u X}|\mathcal{A}] = \mathrm{E}[e^{u X}|\mathcal{B}] $$ holds for any $u \in \mathbb{C}$ then $P(X\in \Gamma|\mathcal{A}) = P(X\in \Gamma|\mathcal{B})$ for any borel $\Gamma$. Any hints?

$\endgroup$
1
$\begingroup$

I would do something like that: First, your assumptions imply that $\mathbb{E}[e^{uX}] < \infty ~\forall u \in \mathbb{R}$. let $\forall u \in \mathbb{R}:~f_a(u) := \mathbb{E}[e^{u X} | \mathcal{A}]$ and $f_b(u) := \mathbb{E}[e^{u X} | \mathcal{B}]$. Using your assumption, one can see that up to a countable number of modifications the function $f_a$ is $\mathcal{C}^{\infty}(\mathbb{R})$ and that: $$ \mathbb{P}(dw) ~a.s.~~f_a^{(n)}(u)= \mathbb{E}(X^n e^{u X} | \mathcal{A})$$

The same apply to $f_b$. From $f_a = f_b$ is follows that $f_a^{(n)}(0) = f_b^{(n)}(0)$. Thus, $\mathbb{P}(dw) \text{ a.s. },~\mathbb{E}[X^n | \mathcal{A}] = \mathbb{E}[X^n | \mathcal{B}]$ where the event of probabilty zero can be choosen independently of $n$.

It thus follows that, if $f$ is a polynomial on $\mathbb{R}$, $\mathbb{P}(dw) \text{ a.s. },~\mathbb{E}[f(X) | \mathcal{A}] = \mathbb{E}[f(X) | \mathcal{B}]$. The result can be extended to any continuous function $f$ with a compact support. It is then not hard to conclude, as the indicator function of any closed set can be arbitrary approximated by a continuous function.

$\endgroup$
  • $\begingroup$ Sorry can you explain why this process have smooth modification? $\endgroup$ – qwenty May 7 '18 at 20:49
  • 1
    $\begingroup$ You can apply the Kolmogorov continuity criteria (en.wikipedia.org/wiki/Kolmogorov_continuity_theorem). Using Cauchy-Schwartz inequality, you can show, that if $x,y \in [-A, A]$ for some $A > 0$, then $\mathbb{E} |f_a(x) - f_b(y)|^2 \leq C_A |x-y|^2$ so there exists a continuous modification of $f_a$. The same apply to all the derivatives of $f_a$ $\endgroup$ – Adri May 9 '18 at 11:57
  • $\begingroup$ @Andri i feel stupid but still can't see how to show that $\mathbb{E} |f_a(x) - f_a(y)|^2 \leq C_A |x-y|^2$ $\endgroup$ – qwenty May 9 '18 at 19:58
  • $\begingroup$ $\mathbb{E} |f_a(x) - f_a(y)|^2 \leq \mathbb{E}|e^{x X} - e^{y X}|^2$ by Cauchy-Schwartz, and you can use for instance: $\forall a, b \geq 0,~ |e^{a} - e^b| \leq e^{max(a,b)} |a-b| $ such that: $\mathbb{E} |f_a(x) - f_a(y)|^2 \leq \mathbb{E} e^{2 A X} X^2 |x - y|^2$, assuming $X$ is a non negative random variable. $\endgroup$ – Adri May 10 '18 at 18:01
2
+25
$\begingroup$

1. A subtlety here is that we are given uncountably many random variables, which is problematic in principle. For instance, it is impossible to discuss the continuity of $f_{\mathcal{A}}(u) = \mathbf{E}[e^{uX}\mid\mathcal{A}]$ because it requires you to look at uncountably many random variables. (This kind of situation is typical when dealing with stochastic processes such as Brownian motion. A typical solution there is to realize the law of a stochastic process on a suitable path space.)

2. Assume that $X$ is real-valued and $e^{u \lvert X \rvert}$ is integrable for all $u>0$. Then there exists a continuous modification of $f_{\mathcal{A}}$, which we again denote by $f_{\mathcal{A}}$. Of course, each value $f_{\mathcal{A}}(u)$ of this modification is a version of $\mathbf{E}[e^{uX}\mid\mathcal{A}]$.

Let $f_{\mathcal{B}}$ be the continuous modification of $u\mapsto\mathbf{E}[e^{uX}\mid\mathcal{B}]$ as well. Then by the continuity, the event $\{ f_{\mathcal{A}}(u) = f_{\mathcal{B}}(u) \text{ for all } u \in \mathbb{C} \}$ is measurable with probability $1$. Using these versions, we can initiate the conditional version of Fourier transform to show the claim.

Let $\mathcal{S}(\mathbb{R})$ the Schwartz space. Since the Fourier transform $\varphi \mapsto \int e^{i\xi u}\varphi(u) \, du$ is an isomorphism on $\mathcal{S}(\mathbb{R})$, we may consider its inverse transform $\varphi \mapsto \check{\varphi}$. Then for each $\varphi \in \mathcal{S}(\mathbb{R})$ and $A \in \mathcal{A}$,

\begin{align*} \mathbf{E}\left[ \left( \int \check{\varphi}(u)f_{\mathcal{A}}(iu) \, du \right) \mathbf{1}_{A}\right] = \int \check{\varphi}(u) \mathbf{E}[e^{iuX}\mathbf{1}_A] \, du = \mathbf{E}\left[ \varphi(X)\mathbf{1}_A \right] \end{align*}

and hence $\int \check{\varphi}(u)f_{\mathcal{A}}(iu) \, du = \mathbf{E}[\varphi(X)\mid\mathcal{A}]$ almost surely. By the indistinguishability of $f_{\mathcal{A}}$ and $f_{\mathcal{B}}$, it follows that $\mathbf{E}[\varphi(X)\mid\mathcal{A}] = \mathbf{E}[\varphi(X)\mid\mathcal{B}]$ almost surely.

Finally, for each Borel $\Gamma \subseteq \mathbb{R}$, pick $\varphi_n \in \mathcal{S}(\mathbb{R})$ such that $\varphi_n(X) \to \mathbf{1}_{\Gamma}(X)$ converges in $L^1(\mathbf{P})$. Then it follows that

  • $\mathbf{E}[\varphi_n(X) \mid \mathcal{A}] \to \mathbf{P}[X \in \Gamma \mid \mathcal{A}]$ in $L^1(\mathbf{P})$, and
  • $\mathbf{E}[\varphi_n(X) \mid \mathcal{B}] \to \mathbf{P}[X \in \Gamma \mid \mathcal{B}]$ in $L^1(\mathbf{P})$.

Therefore the limit must coincide in $L^1(\mathbf{P})$ and hence $\mathbf{P}$-a.s.

$\endgroup$
  • $\begingroup$ Sorry can you explain why this process have continuous modification? $\endgroup$ – qwenty May 7 '18 at 20:49
  • $\begingroup$ @user2715119, Sorry for the late reply. The idea is that you can collect $\mathbf{E}[e^{uX}\mid\mathcal{A}]$ for $u$ in some fixed countable dense subset of $\mathbb{C}$ and then use to interpolate them. The proof will not be so different from that of Kolmogorov's continuity theorem. $\endgroup$ – Sangchul Lee May 9 '18 at 12:38

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.