This is an exercise from Karatzas and Shreve.

Find a $(Y_s)_{s \in [0,1]}$ progressively measurable such that

  1. $ 0 < \int_0^1 Y_s ^2 ds < \infty$ almost surely, and

  2. $\int _0^1 Y_s dW_s = 0$ almost surely

where $W$ is a Brownian Motion.

Not having much luck so far. I know $Y$ cannot be deterministic and that the integral cannot be a true martingale but that is about it.


Here is a solution sketch. By making a time change, it is enough to find an $X$ where $\int_0^\infty X_r\,dB_r=0$ a.s. The geometric Brownian motion $dX = X\,dB$ goes to zero, but the problem is that when we write it as an integral, $X_t=X_0+\int_0^tX_r\,dB_r$, and set $X_0=0$, then $X_t=0$ for all $t$. So we might try a modification of this SDE: \[ dX = \left({\frac1{1+t} + X}\right)\,dB,\quad X_0=0. \] If $Z=1/(1+t)+X$, then we should have $\int_0^\infty Z_r\,dB_r=0$ a.s.

We then want to make a time change: \[ U_t = X_{t/(1-t)} = \int_0^{t/(1-t)} Z_r\,dB_r = \int_0^t V_s\,dM_s, \] where \[ V_t = Z_{t/(1-t)} = 1 - t + X_{t/(1-t)}, \] and $M_t=B_{t/(1-t)}$. Lastly, we observe that we can write \[ M_t = \int_0^t \frac1{1-s}\,dW_s, \] where $W$ is another Brownian motion. Hence, if we define \[ Y_t = \frac1{1-t}V_t = 1 + \frac1{1-t}X_{t/(1-t)}, \] then we have $U_t=\int_0^t Y_s\,dW_s$, and $U_1=0$ a.s.


I cannot think of an easy proof that $X_t\to0$. For purposes of the exercise, there are other nonzero integrands satisfying $\int_0^\infty Z_r\,dB_r=0$ a.s. A simpler, although less interesting example, is $\int_0^\infty 1_{[0,T]}(r)\,dB_r=0$ a.s., where $T=\inf\{t\ge1:B_t=0\}$.

As for proving $X_t\to0$, the intuition was based on the fact that the solution to $dX = (\varepsilon + X)\,dB$ is $X=(X_0+\varepsilon)G-\varepsilon$, where $G_t = \exp(B_t - t/2)$ is the geometric Brownian motion. In this case, the solution tends to $-\varepsilon$. However, the solution to $dX = (1/(1+t) + X)\,dB$ with $X_0=0$, is \[ X_t = G_t - \frac1{1+t} - G_t\int_0^t \frac1{(1+s)^2}G_s\,ds. \] From here the proof would be easy if the process $H_t=G_t\int_0^t G_s^{-1}\,ds$ were almost surely bounded above. Unfortunately, this does not seem to be the case. Writing $dH=dt+H\,dB$ and doing some crude mental calculations with Feller's test for explosions, I believe it can be shown that $\liminf H=0$ a.s. and $\limsup H=\infty$ a.s. So it seems a more delicate analysis is required to prove that $X_t\to0$ (assuming it is even true).

| cite | improve this answer | |
  • $\begingroup$ Thanks for your answer. Do you have a proof that $X_t \to 0$ as $t \to \infty$ ? $\endgroup$ – mathman Mar 2 '12 at 11:46
  • $\begingroup$ I am sorry, but no. In the edit above, I offer an alternative solution to the problem which avoids this, and also some commentary on the question about proving $X_t\to0$. $\endgroup$ – Jason Swanson Mar 6 '12 at 7:32
  • $\begingroup$ Thanks again for your effort. One point, though, if we go down your route of trying to change time by mapping $[0,1]$ to $[0, \infty)$, will we not end up with a process which is not progressively measurable? For example, if I understand you correctly, you think that $Y_s := 1_{[0,T]} (\frac{s}{1-s}) \frac{1}{1-s}$ is a solution, but I don't think it is progressively measurable. $\endgroup$ – mathman Mar 8 '12 at 10:32
  • $\begingroup$ If $\tau=\inf\{s\ge1/2:M_s=0\}$, then $Y_s=1_{[0,\tau]}(s)\frac1{1-s}$. This process is adapted to $\{\mathcal{F}^M_t\}$, and therefore also adapted to $\{\mathcal{F}^W_t\}$. It also has left-continuous sample paths. It is therefore progressively measurable with respect to $\{\mathcal{F}^W_t\}$. $\endgroup$ – Jason Swanson Mar 8 '12 at 19:43
  • $\begingroup$ I agree that $Y_t$ is $\{\mathcal{F}^M_t\}$ adapted but if $M_t = W_{t/(1-t)}$ and $t/(1-t) >t$ then why is $Y_t$ adapted to $\{\mathcal{F}^W_t\}$ ? $\endgroup$ – mathman Mar 9 '12 at 14:40

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.