I've just started to study random processes and I'm trying to solve the following problem:

Let $W(t)$ be a Brownian motion with filtration $F(t)$ generated by $ W(t)$ (i.e., $F(t)=\sigma \left( W(s)\right) $, $s \in [0,t]$).

  • Is the process $|W(t)|$ a sub martingale? A super martingale? Or neither?
  • Is $|W(t)|$ a Markov process?

Unfortunately, I can't see how to do that using the definition of sub/super martingale.. Any help will be appreciated!

  • 2
    $\begingroup$ Use Jensens inequality for the absolute value function. $\endgroup$ – Lutz Lehmann Feb 17 '14 at 0:19
  • 2
    $\begingroup$ Conditional Jensen en.wikipedia.org/wiki/Conditional_expectation#Basic_properties $\endgroup$ – ir7 Feb 17 '14 at 0:23
  • 1
    $\begingroup$ To rule out the other, being both a sub and super martingale implies being a martingale. What would $E[|W(t)|]$ be if $|W(t)|$ were a martingale? What would that imply? I would give a hint about the Markov property, but it depends on what definition you're using. $\endgroup$ – Chris Janjigian Feb 17 '14 at 0:37
  • 1
    $\begingroup$ @Chris Janjigian Jensen inequality for conditional expectations yields $|E[W(t)|F(s)]| \le E[|W(t)||F(s)]$. Thus, since $W(t)$ is a martingale, the left side is equal to $|W(s)|$. So finally, I have the following inequality: $E[|W(t)||F(s)] \ge |W(s)|$ - which tells that |W(t)| is submartingale. Answering your question - if |W(t)| was martingale then $E[|W(t)||F(s)] = |W(s)|$. $\endgroup$ – mouse Feb 17 '14 at 1:41
  • 1
    $\begingroup$ The hint was about the unconditional expectation $E[|W(t)|]$, not the conditional expectation. For the Markov property, notice that the filtration generated by $|W(t)|$ is a strict subset of the filtration generated by $W(t)$. Think about the tower rule. $\endgroup$ – Chris Janjigian Feb 17 '14 at 3:45

Concerning you first question, $|W|$ is a submartingale:

In general if $X$ is a martingale and $\phi: \mathbb R \rightarrow \mathbb R$ is a convex funtion such that $\mathbb E[\phi(X_t)] < \infty$, then a simple application of the conditional Jensen's inequality shows that $$ \mathbb{E}\bigl[\phi(X_t)\mid \mathcal{F}_s \bigr] \geq \phi \Bigl(\mathbb{E}\bigl[X_t\mid \mathcal{F}_s \bigr] \Bigr) = \phi(X_s), \quad \text{a.s.} $$ Hence, the process $(\phi(X_t))_{t \geq 0}$ is a submartingale.

In your case $X=W$ and $\phi(x)=|x|$. It is easy to check that $\phi$ is convex and $E[|W_t|] < \infty$. Therefore $|W|$ is a submartingale.

Edit: the last part is incorrect as pointed out in the comments.

Concerning the Markov Property just keep in mind that $x \mapsto |x|$ is a measurable function (because it is continuous). Therefore the Markov property of $W$ implies the Markov property of $|W|$.

  • 3
    $\begingroup$ It's not generally true that a measurable (or even continuous) function of a Markov process is a Markov process. For example, take simple random walk $X_n$ on $\mathbb{Z}$ and $Y_n = f(X_n)$ where $f(-2) = 2$ and $f(x) = x$ otherwise. Then $$1/2 = P(Y_3 = 3 \mid Y_2 = 2, Y_1 = 1) \ne P(Y_3 = 3 \mid Y_2 = 2, Y_1 = -1) = 0.$$ $\endgroup$ – Nate Eldredge Dec 14 '16 at 14:57
  • 1
    $\begingroup$ The "submartingale" bit is good, but this justification of the Markov property is very bad. Try it with $f(x)= \max \{0,x\}$, for fun. $\endgroup$ – D. Thomine Dec 14 '16 at 15:02
  • $\begingroup$ uuu, yeah thank you for pointing that out, I made a silly mistake $\endgroup$ – Cettt Dec 14 '16 at 15:08
  • $\begingroup$ @NateEldredge It is my understanding that the function needs only to be measurable and bounded to imply the Markov property. Would it be allowable to say $|B_t| _=^d \sup_{0 \le s \le t} B_s$ sup of continuous function over finite interval is bounded, and sup B_s is measurable. So by the Markov property $|B_t|$ is Markov. I don't like it because it seems to be pulling at the ear with the wrong hand but I am weak at maths so cant see how to do it otherwise. $\endgroup$ – Novice May 30 '19 at 7:11
  • $\begingroup$ @Novice: That really doesn't make any sense to me. I would suggest that you try to give a precise statement of the theorem you are trying to use, and then see whether its hypotheses are satisfied. $\endgroup$ – Nate Eldredge May 30 '19 at 13:59

To show the Markov property there is a direct approach.

Let $u \in \mathcal{B}_b(\mathbb{R})$, and $s,t \ge 0$. Then by the independent and stationary increment properties of BM:

\begin{align*} \mathbb{E}[u(|B_{t+s}|) | \mathcal{F_s}] &= \mathbb{E}[u(|B_{t+s}-B_s +B_s|)| \mathcal{F_s}]\\ &= \mathbb{E}u(|B_{t+s}-B_s +y|)\bigg|_{y=B_s}\\ &= \mathbb{E}u(|B_t +y|)\bigg|_{y=B_s} \end{align*}

By symmetry of BM we get that, for $g_t(z)$ normal denisty,

\begin{align*} \mathbb{E}[u(|B_{t+s}|) | \mathcal{F_s}] &= \frac{1}{2} \left[\mathbb{E}u(|B_t +y|)\bigg|_{y=B_s} + \mathbb{E}u(|B_t -y|)\bigg|_{y=B_s} \right]\\ &= \frac{1}{2} \left[ \int_{-\infty}^{+\infty}\left(u(|z+y|) + u(|z-y|)\right)g_t(z)dz \right]\bigg|_{y=B_s}\\ &= \frac{1}{2} \left[ \int_{-\infty}^{+\infty}u(|z|)(g_t(z+y) + g_t(z-y))dz \right]\bigg|_{y=B_s}\\ &= \int_{0}^{+\infty}u(|z|)(g_t(z+y) + g_t(z-y))dz \bigg|_{y=B_s} \end{align*}

The integral is independent of $s$. Hence $\mathbb{E}[u(|B_{t+s}|) | \mathcal{F_s}]$ is a function of $|B_s|$.


\begin{equation*} \mathbb{E}[u(|B_{t+s}|) | \mathcal{F_s}] = g(B_s) \end{equation*}

for some Borel function $g$.

It follows, by the tower property for conditional expectations, that

\begin{align*} \mathbb{E}[u(|B_{t+s}|) | X_s] &= \mathbb{E}\left[ \mathbb{E}[u(|B_{t+s}|) | \mathcal{F_s}|X_s\right]\\ &= \mathbb{E}[g(B_s)|X_s]\\ &=g(B_s). \end{align*}

Hence we conclude that

\begin{equation*} \mathbb{E}[u(|B_{t+s}|) | \mathcal{F_s}] = \mathbb{E}[u(|B_{t+s}|) | X_s]. \end{equation*}

Thus $|B_t|$ is Markov.

  • 1
    $\begingroup$ There are some mismatched parentheses that make this hard to read. $\endgroup$ – Nate Eldredge May 30 '19 at 14:00
  • $\begingroup$ Thanks Nate for the remark. Can you let me know if this is now correct. Best wishes. $\endgroup$ – Novice Jun 1 '19 at 16:45
  • $\begingroup$ The first two lines still look wrong, with the parentheses mismatched around the absolute values. $\endgroup$ – Nate Eldredge Jun 1 '19 at 18:22
  • $\begingroup$ why can we have $$\mathbb{E}\Big[h(|W_{s+t}-W_{s}+W_{s}|)\ \Big|\ \mathcal{F}_{s}\Big]=\mathbb{E}h(|W_{s+t}-W_{s}+y|)\Big|_{y=W_{s}}?$$ $\endgroup$ – JacobsonRadical Jun 6 '20 at 19:28

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.