Consider a filtration $\mathcal{F}_t$ and let $Y$ be integrable on the same probability space. Then $M_t=\mathbb{E}[Y|\mathcal{F_t}]$ is a martingale.
In Chapter 3, Exercise 8, Øksendal asks us to show a converse statement (3.8b) leveraging a Corollary (C.7) to show that if an $\mathcal{F}_t$-martingale is uniformly $L^p(\mathbb{P})$ for some $p>1$ then there exists an integrable $Y$ for which $M_t=\mathbb{E}[Y|\mathcal{F_t}]$.
I'm finding that I require the additional condition that $M_t$ is continuous; am I missing something here, or is the book simply missing the condition?
By (C.7), with continuity, we have that an integrable $Y$ exists such that $M_t\mathbin{\overset{ L^1 }{\longrightarrow}}Y$. Then, notice that $\mathbb{E}[Y-M_t|\mathcal{F}_t]=0$ if $\mathbb{E}[1_H(Y-M_t)]=0$ is for all events $H\in \mathcal{F}_t$.
But $1_H(M_s-M_t)\mathbin{\overset{ L^1 }{\longrightarrow}}1_H({Y-M_t})$ as $s\rightarrow\infty$. So $\mathbb{E}[{1_H(Y-M_t)}]=\lim_{s\rightarrow\infty}\mathbb{E}[{1_H(M_s-M_t)}]$ where for $s\ge t$
$$
\mathbb{E}[1_H({M_s-M_t})]=\mathbb{E}[1_H\mathbb{E}[{M_s-M_t}|{\mathcal{F}_t}]]=0\,\,.
$$
The continuity assumption seems load-bearing. Without continuity, let $\epsilon$ be a Rademacher random variable, $M_t=\epsilon 1\{t\ge 1\}$, and an unnaturally large constant filtration $\mathcal{H}_t=\sigma(\epsilon)$, then $M_t=0\neq \epsilon=\mathbb{E}[{\epsilon}|{\mathcal{H}_{t}}]$ for any $t<1$. Is it possible to construct a counterexample with a natural filtration?
Best Answer
There is an important property of stochastic processes that seems to be missing from Okesndal's book, or at least if it is there I couldn't find it. Specifically, a sub-martingale $M$ with respect to a filtration $(\mathcal F_t)$ satisfying the usual conditions has a right-continuous modification if and only if $t \mapsto \mathbb{E}[M_t]$ is right-continuous. In particular, if $M$ is a martingale, then $M$ has a right-continuous modification.
By considering this modification and using Theorem C.6 instead of Corollary C.7, we can prove the statement in the problem without using any additional assumptions.