I have problem in proving a martingale property. Assume a random variable $\tau$ with a exponential distribution with parameter $\lambda$.

Then define $Y_t = 1_{\{\tau \leq t \}}$. We then can prove $E Y_t=1 - e^{-\lambda t}$, the distributional property of $\tau$. We then define $M_t := Y_t - (1-e^{-\lambda t})$, so that $E M_t=0$ for all $t \geq 0 $.

Now the problem is in proving that the process $M$ is a martingale with respect to the natural filtration $\mathbb{F}$ of $Y$, denoted $(\mathcal{F}^Y_t)_{t \geq 0 }$.

I've started with the expression $$E[M_t-M_s|\mathcal{F}^Y_s]$$ with the objective to prove that it is equal to zero. By definition of $M$ we can rewrite this to $$E[ Y_t-Y_s+e^{-\lambda t}-e^{-\lambda s}|\mathcal{F}^Y_s].$$

Since the exponential terms are deterministic the objective is now equivalent to proving that $$E[ Y_t-Y_s|\mathcal{F}^Y_s]=-e^{-\lambda t}+e^{-\lambda s}.$$

By properties of the indicator we reduce this to \begin{equation} P[s<\tau\leq t|\mathcal{F}^Y_s]=-e^{-\lambda t}+e^{-\lambda s}. \end{equation}

Here is where I'm (maybe too late) uncertain of my approach. Since $\tau$ is fixed for a given $\omega$ (we view all random variables as measurable functions off a probability space), in the conditioning we are only interested if $Y_s$ is either one or zero, since this gives us the information if $\tau$ is larger than or less than $s$.

Hence we can write \begin{equation} P[s<\tau\leq t|\mathcal{F}^Y_s]=P[s<\tau\leq t|\tau \leq s] + P[s<\tau\leq t|\tau > s]. \end{equation}

This should make sence, but the first term is zero and the second is equal to $-e^{-\lambda t}+e^{-\lambda s}/e^{-\lambda s}$ via standard methods of integration and using the PDF of a exponential distributed RV.

What am I doing wrong here? I know the martingale property should hold. Should I introduce in the conditioning $\tau \leq u$ and integrate over all $u$ from zero to $s$?


The compensator of $Y$ is the $(\mathcal F_s^Y)$-adapted process $t\mapsto\lambda\min(t,\tau)$, which has expected value $1-e^{-\lambda t}$. That is, $M_t:=1_{\{\tau\le t\}}-\lambda\min(t,\tau)$ is a martingale.

Added detail: The idea is that the process to subtract from $Y$ to make it a martingale (the compensator of $Y$) should be the integral from $0$ to $t$ of the "infinitesimal" conditional expectation $$ \eqalign{\Bbb E[Y_{t+dt}-Y_t|\mathcal F_t^Y] &=\Bbb P[t<\tau\le t+dt|\mathcal F_t^Y]\cr &=1_{t<\tau}\lambda\,dt.\cr } $$ The second equality here uses the memoryless property of the exponential distribution: $\{t<\tau\}$ is $\mathcal F_t^Y$-measurable, and on this event, $\tau-t$ is independent of $\mathcal F_t^Y$ with the exponential law of parameter $\lambda$; thus, on $\{t<\tau\}$, the conditional probability that $\tau-t$ is at most $dt$ is equal to $1-e^{-\lambda \cdot dt}=\lambda\cdot dt$ (neglecting smaller order terms). Thus, the compensator of $Y$ ought to be $$ \int_0^t 1_{\{s<\tau\}}\lambda\,ds=\int_0^{t\wedge\tau}\lambda\,ds=\lambda\min(t,\tau). $$ and so $M_t:=Y_t-\lambda\min(t,\tau)$ ought to be a martingale. It's pretty straightforward at this point to use the memorylessness of the exponential random variable $\tau$ to check that $M$ is indeed an $(\mathcal F_t^Y)$ martingale.

  • $\begingroup$ Can you give me a reference for this or else elaborate your statement with an proof? Also, how is this intuïtive? $\endgroup$ – Tim Huijgens Apr 19 '16 at 7:18
  • $\begingroup$ Thanks for your explanation. Is this method with infinitesimal conditional expectation usual for determining the compensator? $\endgroup$ – Tim Huijgens Apr 19 '16 at 16:35
  • $\begingroup$ I'm not sure how "usual" it is; it's my personal heuristic, but where I got it I no longer remember... $\endgroup$ – John Dawkins Apr 20 '16 at 19:40

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.