The proof is essentially just a formalization of Kac's [33] original argument for space-time BM. Variations appear in the proofs of similar results in [38, 50, 51, 19]. Let Y denote the killed process with state space defined by if t<T and if , where is a cemetery state. In terms of Y, the killing time T is just the hitting time of . It is therefore enough to prove the result for T the hitting time of a point in the state space E of X.
We assume, without any real loss of generality, that the sample space is equipped with a family of shift operators , such that for all . Furthermore, we assume that there is a filtration on to which X is adapted and with respect to which X has the simple Markov property:
for all , all non-negative -measurable functions F, and all non-negative functions on that are measurable with respect to .
The key property of a hitting time T of X is that it is a terminal time [8, 59]; that is, an -stopping time T with the property on the event . The basic inductive step which allows Kac's moment formula (4) to be pushed from n to n+1 involves the following identity, which holds for an arbitrary -stopping time T. Let be the pre-T occupation kernel defined by
for an arbitrary non-negative measurable v. If is an initial distribution, then is the measure
which describes the expected amount of time X spends in various subsets F of E up to time T. Call the pre-T occupation measure for . In case T is a Markov killing time of X, G is the potential kernel derived from the killed process, as discussed in Section 1. But the above definition (11) of G makes sense, and the following identity is valid, for an arbitrary stopping time T:
Occupation measure identity [38, 50, 51] For each initial distribution on E, each non-negative -measurable , and each non-negative -measurable f,
The assumed measurability of implies that is -measurable, because is -measurable. Thus the left side of (13) is well defined. Now Fubini's theorem shows that
which by the Markov property (10) and the -measurability of is equal to
Taken together, (12), (14), and (15) yield (13).
Notice that the proof of (13) required no strong Markov property of X. So the occupation measure identity holds without any assumptions about the state space of X or path properties of X beyond the joint measurability of as a function of t and .
Turning to the proof of (4), observe that
where
Because T is a terminal time, the obvious change of variables in (17) leads to
Thus (4) follows by induction from the occupation measure identity (13).