It is well-known that, in globally hyperbolic spacetimes, the time separation function $\tau$ (aka Lorentzian distance function) enjoys the following property: fix a point $p$ and a point $q \in I^-(p)$. Then if $\gamma : [0,1] \to M$ is a past-directed causal curve starting at $q$, the function
\begin{equation} \lambda \mapsto \tau(p,\gamma(\lambda)) \qquad (*) \end{equation}
is *strictly* increasing. A proof of this fact follows simply by noting that, in this context: (a) a maximal causal curve between $p$ and any point in $I^-(p)$ exists and is necessarily a smooth timelike geodesic; (b) due to the reverse triangle inequality, one only needs to worry about the case where $\gamma$ is an achronal null geodesic segment. The conclusion then easily follows (see e.g. the proof of Proposition 3.1 in http://arxiv.org/abs/1010.5032).

Suppose one were interested in studying (time-oriented) spacetimes *with smooth boundary* and suppose also that the boundary is everywhere timelike. Then one could fairly straightforwardly come up with a notion of global hyperbolicity for such spacetimes: namely strong causality plus, as usual, compactness of the sets $J^-(x) \cap J^+(y)$ (where causal and chronological curves are defined in the same way as in the case with no boundaries, say using piecewise smooth curves). This has in fact been already done and the causal theory of such spacetimes was studied by D. A. Solis in his PhD thesis (University of Miami, 2006). See also Section 2.2 in http://arxiv.org/abs/0808.3233. It turns out that globally hyperbolic spacetimes with timelike boundary are causally simple, the time-separation function is continuous, and there exist maximal *continuous* causal curves between any two causally related points.

To get a flavour of the kind of result that does *not* still hold in this category, notice that it is no longer true that if $q \in J^-(p) \setminus I^-(p)$ then every causal curve from $p$ to $q$ is an achronal null (pre)geodesic. Consider for instance (1+2)-dimensional Minkowski space with a timelike cylinder removed, a point $q$ on the boundary and a point $p$ (either on the boundary or in the interior) "behind" the cylinder. It is clear that for some such $p$ there are causal but not timelike curves to $q$, but also that none of these curves are null geodesics. On the other hand, one can prove the weaker result that if $\gamma : [0,1] \to M$ is a causal past-directed curve from $p$ to $q$ with $\gamma(0,1) \subset \mathrm{Int} \, M$ then either $q \in I^-(p)$ or $\gamma$ is a smooth null (pre)geodesic. Similarly, the "smooth geodesic" bit in remark (a) in my first paragraph clearly fails to hold here.

Having said all that (!): can anyone think of an example showing that the function defined by $(*)$ should not be expected to be strictly increasing if one allows for timelike boundaries (but under the assumption of "global hyperbolicity")? I should say that I don't necessarily believe that a counterexample exists; but it is clear IMHO that the strategy outlined in my first paragraph does not straightforwardly adapt, and I am not aware of any alternative methods.

with timelike boundary, given two chronologically related points $q\ll p$, is it possible to have a maximal, past-directed causal curve $\gamma:[0,1]\rightarrow M$ such that $\gamma(0)=q$, $\gamma(1)=p$ and $\gamma|_{[\epsilon,1]}$ is a(n achronal) null geodesic for some $\epsilon\in(0,1)$? It is clear that strict monotonicity of $(*)$ holds iff the answer is negative, since any segment of $\gamma$ must be maximal (non-strict monotonicity always holds by the reverse triangle inequality). $\endgroup$ – Pedro Lauridsen Ribeiro Sep 29 '15 at 18:00without boundary$N$, such that $U$ is mapped to a causally convex subset of $N$. But this hardly seems to me like the most helpful of criteria! $\endgroup$ – Umberto Lupo Sep 29 '15 at 21:56