For a standard Brownian motion $W_t$ and a given crossing probability $\alpha < 1$, I want to have a boundary function $f(t) > 0$, such that the probability that $W_t$ ever crosses the boundary is bounded by the given probability, i.e. $$ P(|W_t| > f(t), \exists t > 0) \le \alpha $$ By Law of the iterated logarithm, any $f(t) = O(\sqrt{t\log\log t})$ does not meet this condition because $$ \lim\sup_{n\rightarrow \infty} \frac{W_t}{\sqrt{t\log\log t}} = \sqrt{2} $$ if I get it right. For my case, I would like an $f(t) = o(t)$. In other words, I want to have a function such that $$ \lim_{t\rightarrow \infty}\frac{f(t)}{t} = 0 $$.


If $f(t)=b+t$, where $b>0$, then $ P(W_t > f(t)$ for some $t > 0)=e^{-2b}$, so if you choose $b=-\log\sqrt{\alpha/2}$, then $$ P(|W_t| > f(t)\hbox{ for some }t > 0)\le 2P(W_t > f(t)\hbox{ for some }t > 0)=2e^{-2b}=\alpha. $$

  • $\begingroup$ Can you prove that exponential bound? (I believe it, I'm just not sure of the proof.) $\endgroup$ – Ian Jul 6 '16 at 16:42
  • $\begingroup$ Just use the "gambler's ruin" formula for $X_t:=W_t-t$ (a Brownian motion with unit negative drift): $\Bbb P[T_a<T_b]={S(0)-S(a)\over S(b)-S(a)}$, where $T_a:=\inf\{t:X_t=a\}$ is the hitting time of $a$, etc., and $S(x):=e^{2x}$ is a scale function for $X$. In the limit $a\to-\infty$ you get $\Bbb P[T_b<\infty]=\lim_{a\to-\infty}\Bbb P[T_b<T_a]=S(0)/S(b)=e^{-2b}$. $\endgroup$ – John Dawkins Jul 6 '16 at 19:40

Fix $\alpha \in (0,1)$. Let $M_t := \sup_{s \leq t} W_s$ and $m_t := \inf_{s \leq t} W_s$ be the running maximum and mimimum, respectively. Then

$$\mathbb{P}(\exists t \in [0,n]: |W_t|>c_n) = \mathbb{P}(\{M_n>c_n\} \cup \{m_n < -c_n\})$$

for any constant $c_n>0$ and $n \in \mathbb{N}$. Since both $M_n$ and $m_n$ are real-valued, we can choose $c_n>0$ sufficiently large such that

$$\mathbb{P}(\exists t \in [0,n]: |W_t|>c_n) = \mathbb{P}(\{M_n>c_n\} \cup \{m_n < -c_n\}) \leq (1-\alpha) \alpha^n.$$

If we define

$$f(t) := \sum_{n=1}^{\infty} c_n 1_{(n-1,n]}(t)$$


$$\mathbb{P}(\exists t>0: |W_t|>f(t)) \leq \sum_{n \in \mathbb{N}} \mathbb{P}(\exists t \in [0,n]: |W_t|>c_n) \leq (1-\alpha) \sum_{n \in \mathbb{N}} \alpha^n = \alpha$$

which shows that $f$ does the job.


  • The distribution of $M_n$ and $m_n$ is known; in fact, $M_n \sim -m_n \sim |W_n|$ (se e.g. Brownian Motion - An Introduction to Stochastic Processes by Schilling & Partzsch, Chapter 6, for a proof). This allows us to compute bounds for the constants $c_n$.

  • This construction of $f$ works, more generally, for any stochastic process $(X_t)_{t \geq 0}$ with continuous sample paths.

  • $\begingroup$ While perfectly reasonable, I think without an explicit characterization of or at least estimate for the distribution of $M_n,m_n$, it is not really what the OP was going for. $\endgroup$ – Ian Jul 6 '16 at 16:41
  • $\begingroup$ @Ian The distribution of $M_n$ and $m_n$ is known, that's hardly a problem. (The OP hasn't specified what he's looking for so....) $\endgroup$ – saz Jul 6 '16 at 16:48
  • $\begingroup$ It's known, sure, but is it known to the OP? Perhaps a reference to that effect would be useful, so that the OP can get a feel for the scale of $c_n$. $\endgroup$ – Ian Jul 6 '16 at 16:50
  • $\begingroup$ @Ian I have added a reference. Thanks for your feedback. $\endgroup$ – saz Jul 7 '16 at 5:51

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.