Assume now that E is a complete separable metric space, and that the sample paths of X are right-continuous with left limits; in particular (21) is valid. In this situation all of the previously displayed formulae have versions involving a conditioning on , as indicated in various settings by Kac [33], Ray [55] and Dynkin [18, 19]. Such conditioning can be achieved in great generality using h-processes. To illustrate, the h-process version of (27), as formulated in [19], takes a simple form due to cancellation of the h-factors in the product of Green's kernels. See also Proposition (5.14) of [51] for a discrete time example, and [1] for the formula obtained this way for the mean exit time of a diffusion on an interval conditioned to exit at a specified boundary point. The effect of conditioning on is simplest for the special class of Markov killing times introduced in the following definition:
More formally, assuming T has been set up as a stopping time relative to a suitable enlargement of the filtration , the assumption is that, under for every , the process is an -martingale. Equivalently,
for every positive -predictable process Z and every positive -measurable function f.
The above definition makes sense, and the obvious analog of (29) is valid, if a general PCAF is substituted for . For example, the last exit time from a subset B of E will be ``killing with rate dA(t)'' (for a suitable PCAF A) provided X is a strong Markov process with quasi-left-continuous sample paths (a ``Hunt process''). In this context the PCAF A is naturally associated with the so-called ``equilibrium distribution'' on B; see [10, 26, 25, 28]. Note that a first passage time into a set B will be of this form only if the first passage occurs at the time of a jump of X. In particular, a predictable Markov killing time T, such as the hitting time of a set for a process with continuous paths, will not be of this form.
Proof. For n = 0 the result is the special case in (29):
For general n we proceed by induction, as in the proof of (4). Thus, define
and notice that . Using as an abbreviation of , we have
the final equality following from (13) and (29). Thus, , and (30) follows by induction on n. Remark. Formula (30) holds also with either or or both replaced by general PCAFs. A formula could also be obtained with instead of , assuming the existence of a Lévy system for the jumps of X and that T is a jump time. See e.g. [4, 52].
Example. The special case of the above proposition for an independent exponential time, when for all x, is already evident in Kac [33]. Then is the resolvent operator of the semi-group of X. After cancelling the common factor of on both sides, the result is as follows: for arbitrary ,
While the existence of left limits was assumed in the previous proposition, it is easily shown that no such hypothesis is required for (32).