I have both a more general question (concerning stopping times), and then a more specific application (as described in the title).

Let $(\Omega,\mathcal{F},(\mathcal{F}_t)_{t \geq 0},\mathbb{P})$ be a filtered probability space, which we can assume to satisfy the "usual conditions" if necessary.

Let $(X_t)_{t \geq 0}$ be a real-valued progressively measurable homogeneous Markov process over the above filtered probability space, and assume that $(X_t)$ has continuous sample paths.

Let $A \in \mathcal{B}(\mathbb{R})$ be a Borel set such that $$ \mathbb{P}(\exists \, t \geq 0 \textrm{ s.t. } X_t \in A) \ = \ 1. $$

Q1. Does there necessarily exist a stopping time $\tau:\Omega \to [0,\infty]$ such that $$ \mathbb{P}(X_\tau \in A) \ > \ 0 \, ? $$

Now I fear that the answer to Q1 is no (although I'd like it to be yes), since even in general, the existence of measurable selections is a highly non-trivial issue. Nonetheless, I am hoping that the particular claim which I wanted to prove using a positive answer to Q1 might still be true:

Suppose our Markov process above satisfies the following additional properties:

  1. $X_0$ is (almost surely) equal to some constant value $\xi\,$;
  2. for almost all $\omega$ the set $\{X_t(\omega):t \geq 0\}$ is equal to the whole of $\mathbb{R}\,$;
  3. $(X_t)_{t \geq 0}$ is described by transition probabilities $P_t(x,\cdot)$ satisfying the following type of continuity: for every bounded continuous $g:X \to \mathbb{R}$ the map $(t,x) \mapsto \int_\mathbb{R} g(y) \, P_t(x,dy)$ is continuous.

Q2. Is it necessarily the case that the only Borel set $B \in \mathcal{B}(\mathbb{R})$ satisfying:

  • $\xi \in B$
  • $P_t(x,B)=1$ for all $x \in B, \, t \geq 0$
  • $P_t(x,B)<1$ for all $x \in \mathbb{R} \setminus B, \, t \geq 0$

is $\mathbb{R}$ itself?

If the answer to Q1 is yes, then I'm sure the answer to Q2 must also be yes: supposing otherwise for a contradiction, just take $A$ in Q1 to be $X \setminus B$ and use the strong Markov property.

  • $\begingroup$ I've just edited the question to slightly weaken Q1. (It was stronger than necessary before.) I would assume that the answer to Q1 is a well-known fact - does anyone know? $\endgroup$ Mar 28, 2015 at 0:20
  • $\begingroup$ I've also edited Q2. (In its previous form, the answer was trivially no, as pointed out in Nate Eldredge's answer below.) $\endgroup$ Mar 28, 2015 at 11:42

1 Answer 1


The answer to Q1 is Yes, and the answer to Q2 is No.

If $\mathbb{P}(X_0 \in A) > 0$ then we are done by taking $\tau = 0$. So suppose $\mathbb{P}(X_0 \in A^c) = 1$.

Call $y$ a right endpoint of $A$ if $y \in A$ and there exists $x < y$ with $(x,y) \subset A^c$. Let $A^+$ be the set of all right endpoints of $A$; then $A^+$ is countable. Likewise let $A^-$ be the set of all left endpoints, which is also countable. Finally, let $D \subset A$ be countable and dense in $A$. Set $C = A^+ \cup A^- \cup D$.

Claim. If $a < b$ and $[a,b] \cap A \ne \emptyset$ then $[a,b] \cap C \ne \emptyset$. That is, any nontrivial closed interval that meets $A$ must meet $C$.

Proof of claim. If $(a,b) \cap A \ne \emptyset$ then by density $(a,b)$ contains a point of $D$. Otherwise, we have $(a,b) \subset A^c$. Then we must have either $a \in A$, in which case $a \in A^-$, or else $b \in A$, in which case $b \in A^+$. In all cases $[a,b]$ meets $C$.

We now know that for almost every $\omega$, we have that $X_0(\omega) \notin A$ and there exists $t$ such that $X_t(\omega) \in A$. Suppose $X_0(\omega) < X_t(\omega)$; then by the claim, $[X_0(\omega), X_t(\omega)]$ intersects $C$, so by continuity there exists $s$ such that $X_s(\omega) \in C$. The same holds if $X_0(\omega) > X_t(\omega)$.

Enumerate $C$ as $C = \{c_1, c_2, \dots\}$ and let $\sigma_n = \inf\{t : X_t = c_n\}$. We have just shown that $\mathbb{P}\left(\bigcup_{n=1}^\infty \{\sigma_n < \infty\}\right) = 1$. So by countable additivity we may choose $N$ so large that $\mathbb{P}\left(\bigcup_{n=1}^N \{\sigma_n < \infty\}\right) > 0$. Set $\sigma = \sigma_1 \wedge \dots \wedge \sigma_N$; then $\mathbb{P}(\sigma < \infty) > 0$. Since $\{c_1, \dots, c_N\}$ is closed, on $\{\sigma < \infty\}$ we have $X_\sigma \in \{c_1, \dots, c_N\}$. Choose $M < \infty$ large enough that $\mathbb{P}(\sigma < M) > 0$. Finally set $\tau = \sigma \wedge M$. Then $X_\tau$ is well defined, and on the event $\{\sigma < M\}$ we have $X_\tau = X_\sigma \in \{c_1, \dots, c_N\} \subset A$. So $\mathbb{P}(X_\tau \in A) \ge \mathbb{P}(\sigma < M) > 0$. (In fact by choosing $N,M$ sufficiently large we can make $\mathbb{P}(X_\tau \in A)$ arbitrarily close to 1.)

For Q2, let $X_t$ be standard Brownian motion, which satisfies properties 1,2,3 with $\xi = 0$. Let $B = \mathbb{R} \setminus \{1\}$, so that $0 \in B$. Since Gaussian measure is absolutely continuous to Lebesgue measure, $P_t(x,B) = 1$ for any $x$ and any $t > 0$, and $P_0(x,B) = 1$ when $x \in B$. So Q2 is not satisfied. (We could have chosen $B$ to be any set of full Lebesgue measure that contains 0.)

Something seems to be wrong with your "strong Markov property" logic.

  • $\begingroup$ Thank you very much for this answer. I realise I was getting things muddled in my head as I wrote Q2. I've now corrected it (so I think it now asks the right thing!). Presumably the answer to Q2 is now yes - and I probably don't actually need Q1 to prove it: Suppose it's false, with $B$ being a counterexample. By Fubini, the set of times $t$ at which $X_t\in B$ is almost surely a Lebesgue-full set. But fix $c\in\mathbb{R}\setminus B$ and let $\tau$ be the hitting time associated to $\{c\}$; the strong Markov property at $\tau$ (+ Fubini) should contradict this. Is this right? $\endgroup$ Mar 28, 2015 at 10:32
  • $\begingroup$ PS: Just out of curiosity, does anyone know if the answer to Q1 generalises to higher dimensions, or even to general Polish spaces? $\endgroup$ Mar 28, 2015 at 11:40

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.