I am trying to understand the pricing of barrier options, and am considering the Brownian motion $\mathrm{d}X_t=a\mathrm{d}t+b\mathrm{d}W_t$, $a$ and $b$ constant. I am trying to:

  1. derive the distribution of $X_{T/2}$ given $X_T$ and $X_0$;

2i) prove that the probability $\mathbb{P}\left(\inf_{[t_1,t_2]}X_t<L\Big|X_{t_1},X_{t_2}\right)=\mathrm{exp}\left[-\frac{2(X_{t_1}-L)^+(X_{t_2}-L)^+}{b^2(t_2-t_1)}\right]$; and

ii) find $\mathbb{P}\left(\sup_{[t_1,t_2]}X_t>U\Big|X_{t_1},X_{t_2}\right)$.

For 1, is there a better way about it than to calculate the conditional density? I tried using the approach that one would use for a driftless Brownian motion, but ended up with multiple cross-terms that I can't get rid of. For 2i and ii, how would one derive them? The texts I have read mention the reflection principle, and I understand the conditional density can be imagined as the fraction of which paths hit the barrier, but I really don't understand much about it, particularly where the $(\cdot)^+$ operators enter. Any help and rigour is much appreciated. Thanks!


For part 1 of your question, the short answer is no, calculating conditional density is a looong way of doing it. Possible but not the easiest. Here is the sketch for a shorter version. We note that $(X_{T/2},X_{T})$ is a jointly Gaussian vector with mean $\mu = (X_0 + aT/2,X_0 + aT)$, and the variance-covariance matrix $$ \begin{pmatrix} b^2 T/2 & b^2 T/2 \\ b^2 T/2 & b^2 T \end{pmatrix} $$

A conditional distribution of one element of a Gaussian vector on another on is Gaussian. So the conditional distribution of $X_{T/2} \vert X_T$ is Gaussian. The mean and the variance of this distribution can be expressed in terms of $X_T, X_0, \mu,\Sigma$. Details can be found in many places, for example here. The conditional mean in particular is just a linear regression formula so easy to remember

$$ \mathrm{E}(X_{T/2} \vert X_T) = X_0 + aT/2 + \beta (X_T - aT - X_0) $$ where $$ \beta = \frac{b^2 T/2}{b^2T} = \frac{1}{2} $$ so it simplifies to $$ \mathrm{E}(X_{T/2} \vert X_T) = (X_0 + X_T)/2 $$ The conditional variance is calculated along the similar lines (see the link above)

Interestingly the conditional mean and, in fact, the whole conditional distribution does not depend on the drift $a$ and is the same as for the standard Brownian motion with $a=0$, the so-called Brownian bridge.

Q2 is quite a bit more involved, you should look it up in any decent stochastic calculus textbook (Karatzas & Shreve is my favourite). As you can see the right-hand side is independent of the drift $a$. Our discussion for Q1 demonstrates (if falls somewhat short of a formal proof) why it is the case -- once you "pin" the start and the end of a Brownian motion, the fact that it has (constant) drift is not relevant anymore.

As to your specific point as to what $(...)^+$ terms are doing in the formula. This is just a convenient shortcut to have one formula for different configurations of $X_{t_1}$, $X_{t_2}$ and $L$. For example, if $X_{t_1} < L$ then the lhs is trivially 1, and the rhs is $1$ as well because $(X_{t_1) - L)^+ =0 $ in this case.

  • $\begingroup$ Thanks so much for your answer! Before I accept, may I ask which pages in Karatzas and Shreve cover question 2? $\endgroup$ – user107224 Mar 18 at 0:14
  • $\begingroup$ In my copy it is Karatzas and Shreve, Brownian Motion and Stochastic Calculus, Second Edition, 1991, section 4.3.C, p265, eq (3.40). If you have a different edition (or a different book!) look for "Brownian bridge, Maximum of" in the index $\endgroup$ – piterbarg Mar 18 at 0:19
  • $\begingroup$ @piterbarg: That was a beautiful answer and made sense. ( once I went to the other for the conditional formulae of the bivariate normal. I always forget them ). But, if you don't mind, what's the intuition for why the answer doesn't depend on $X_0$. Is that because the knowledge of the value of $X_T$ kinds of cancels the usefulness of knowing $X_0$ ? Thanks. $\endgroup$ – mark leeds Mar 18 at 0:50
  • $\begingroup$ @markleeds ah sorry it does depend on X_0, I assumed X_0 = 0 implicitly. The general case follows if we consider X_{T/2} - X_0, X_T-X_0 instead. I'll fix that when I have a moment $\endgroup$ – piterbarg Mar 18 at 7:24
  • $\begingroup$ @piterbarg another silly question but does the reflection principle apply even with drift? also for the infimum case do I just plonk in minus times a supremum? $\endgroup$ – user107224 Mar 18 at 14:21

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.