Optional Sampling Theorem for non-positive sub-martingale

56 Views Asked by At

I'm trying to reconstruct the proof of this theorem from my lecture note. Could you please verify if my proof is correct or contains logical mistake? Thank you so much!

Let $\left(\Omega, \mathcal{F},(\mathcal{F}_{n}), \mathbb P\right)$ be a filtered probability space. If $(X_{n})$ is a non-positive sub-martingale and $S \le T$ are finite stopping times. Then $\mathbb E [ X_T | \mathcal F_S ] \ge X_S$.


My attempt:

First, we need the following lemmas:

$\textbf{Lemma 1}$ If $(X_{n})$ is a sub-martingale w.r.t $(\mathcal{F}_{n})$ and $\phi: \mathbb R \to \mathbb R$ is non-decreasing convex. Then $(\phi(X_{n}))$ is also a sub-martingale w.r.t $(\mathcal{F}_{n})$.

$\textbf{Lemma 2}$ If $(X_{n})$ is a bounded sub-martingale w.r.t $(\mathcal{F}_{n})$ and $S \le T$ are finite stopping times. Then $(X_S, X_T)$ is also a sub-martingale w.r.t $(\mathcal{F}_{S}, \mathcal{F}_{T})$.

$\textbf{Lemma 3}$ If $\mathcal G$ is a sub $\sigma$-algebra of $\mathcal F$ and $X:\Omega \to \mathbb R$ is integrable. Then $$\mathbb E[X | \mathcal G] \ge 0 \quad \text{a.s} \iff \forall \Lambda \in \mathcal G: \mathbb E [\mathbf{1}_\Lambda X] \ge 0$$

For each $m \in \mathbb N$, we define a map $\phi_m: \mathbb R \to \mathbb R, \quad x \mapsto \max \{x, -m\}$. It's easy to verify that $\phi_m$ is non-decreasing convex and $\phi_m (x) \downarrow\ x$ as $m \to \infty$. By $\textbf{Lemma 1}$, $(\phi_m(X_{n}))_{n \in \mathbb N}$ is a sub-martingale.

Because $\phi_m$ is non-decreasing and bounded from below by $-m$, and $X_n$ is bounded from above by $0$, $(\phi_m(X_{n}))_{n \in \mathbb N}$ is bounded. Then, by $\textbf{Lemma 2}$, $(\phi_m(X_{0}),\phi_m(X_{S}))$ is a sub-martingale w.r.t $(\mathcal F_0, \mathcal F_S)$ and thus $-\infty < \mathbb E [X_{0}] \le \mathbb E [\phi_m(X_{0})] \le \mathbb E [\phi_m(X_{S})]$.

By $\textbf{Monotone Convergence Theorem}$, we get $\mathbb E [\phi_m(X_{S})] \to \mathbb E [X_S]$ as $m \to \infty$. Hence $\mathbb E [X_{0}] \le\mathbb E [X_S] \le 0$. Thus $X_S$ is integrable. Similarly, $X_T$ is integrable.

By $\textbf{Lemma 2}$ again, we get $(\phi_m(X_{S}),\phi_m(X_{T}))$ is a sub-martingale w.r.t $(\mathcal F_S, \mathcal F_T)$ and thus $\mathbb E[\phi_m(X_{T})-\phi_m(X_{S}) | \mathcal F_S] \ge 0$. By $\textbf{Lemma 3}$, this is equivalent to $$\forall \Lambda \in \mathcal F_S: \mathbb E [\mathbf{1}_\Lambda \phi_m(X_{T})] \ge \mathbb E [ \mathbf{1}_\Lambda \phi_m(X_{S}) ]$$

By $\textbf{Monotone Convergence Theorem}$ again, we get $\mathbb E [\mathbf{1}_\Lambda \phi_m(X_{S})] \to \mathbb E [\mathbf{1}_\Lambda X_S]$ and $\mathbb E [\mathbf{1}_\Lambda \phi_m(X_{T})] \to \mathbb E [\mathbf{1}_\Lambda X_T]$ as $m \to \infty$. As a result, $\forall \Lambda \in \mathcal F_S: \mathbb E [\mathbf{1}_\Lambda X_{T}] \ge \mathbb E [ \mathbf{1}_\Lambda X_{S} ]$. By $\textbf{Lemma 3}$ again, we get $\mathbb E [X_{T} - X_{S} | \mathcal F_S] \ge 0$.