To recover more standard expressions of the F-K formula, as presented in Section III.19 of Rogers-Williams [57], let be the semigroup derived from by killing X with state-dependent rate v(x). So if T is the associated Markov killing time then
Summing formula (32) weighted by yields an expression for the resolvent of this semigroup:
where is the resolvent of the semi-group of X. Some mild regularity on f and v are required to make sense of the second equality in (34), but when rearranged as
the formula holds as an identity of bounded positive kernels for arbitrary and non-negative measurable v. This is formula (III.19.5) of [57], which is one of the three forms of the F-K formula presented by Rogers-Williams. As they remark, and as shown by the above argument, (35) is a robust form of the F-K formula which is valid with no hypotheses on the underlying Markov process X beyond jointly-measurable paths.
To interpret the F-K formula as a statement relating the infinitesimal generators of and , let us recall that the weak infinitesimal generator of (say) is the operator defined by
on the domain comprising those functions f for which the pointwise limit indicated in (36) exists and . See [15], where it is shown that for each , the resolvent operator is an injective mapping of (the class of bounded -measurable functions) onto , and . Briefly, . Viewed in these terms, (35) amounts to the most common presentation of the F-K formula:
the domain consisting of those functions for which uv is a bounded function. In the context of symmetric Markov processes, (37) can be reformulated in terms of ``Dirichlet forms''; see Section 6.1 of [24].
A third form of the F-K formula noted in [57] (III.19.7) is the following variant of (35):
which can be understood with almost no calculation due to the following
Probabilistic interpretation of (38). Consider , the minimum of T and an independent exponential time with rate . From (34), the Green's operator for X killed at time is . Since killing at time is the same as killing with rate function , the distribution of on is found by an easy variation of the ``last exit'' formula (31):
Thus (38) comes from integrating with respect to the decomposition
for an arbitrary additive functional .
Another probabilistic interpretation of (38). Multiplication of both sides of (38) by yields an identity of Markov kernels which may be understood in another way. By the companion of (39) with instead of ,
and by (39) and the memoryless property of ,
Adding these two distributions we arrive at , which is the distribution of .
Remarks. For a constant function v, say for all x, (38) reduces to the resolvent identity:
So the above arguments give two simple probabilistic interpretations of this identity involving the minimum of two independent exponential variables and and analysis of according to whether or . Since it is obvious that , the resolvent identity also yields (35) in this special case. For a general v, comparison of (35) and (38) establishes the identity of kernels
where the first identity was interpreted probabilistically above.
There is also a probabilistic interpretation of the second identity in (44), involving the idea of resurrection of X after the killing time T, as considered in [46, 22]. Note that (44) is the Laplace transformed version of the following identity:
and that if is an initial distribution, then
for any measurable . The first equality in (45) is evident from evaluation of this expectation by conditioning on . The second equality in (45) is obtained by the following construction. Assume for simplicity that v is bounded. Given X, let where are the points of a Poisson process on with intensity . The right hand expression in (45) arises from evaluating the expectation (46) by conditioning on , where Replacing the fixed time t by gives a similar interpretation of the second equality in (44). From this perspective, the middle and right hand expressions in (44) and (45) are seen to be typical ``first entrance'' and ``last exit'' decompositions.