Suppose now that X is a Markov chain with finite state space and T is a finite Markov killing time for X. By obvious reductions there is no loss of generality involved in the following
Assumption. The state space of X is where E is finite, is an absorbing state, and for all .
Let denote the substochastic semi-group of X restricted to E, and view and all other operators as matrices indexed by E, for example . Then is just
is the usual Q-matrix of the chain killed at time T. Recall that where is the operator of multiplication by v. It is clear that there exists b>0 such that for all v with the matrix is invertible. So
and the F-K formula (7) can be restated as follows: There exists b>0 such that for all v with
For let . Since , formula (49) determines the joint moment generating function of the . The second expression in (49) for this m.g.f. appears as formula (4) in Kingman [39], and again in Puri [54]. Kingman noted as a consequence that the joint m.g.f. is a ratio of two multilinear forms in v(x), , and that the marginal distribution of each is a mixture of a point mass at zero and an exponential distribution on . Kingman raised the problem, which is apparently still open, of characterizing which joint distributions can appear as the joint distributions of such occupation times of a transient finite state chain. For some study of particular examples see [36, 40].
Every killing time of a finite state chain is easily seen to be of the form assumed in Proposition 1 for the killing rate function where Q is the Q-matrix of the killed chain. From (31)
so the assumption that for all x implies . Formula (30) now yields expressions for the conditional moments of given . This leads to the following sharper form of the F-K formula for chains. Formula (52) is a variant of Theorem 2.1 of Dynkin [18]. See also Sections I.27 of [57] and IV.22 of [56] for related presentations.
Proof. Formula (51) results from summing (30) weighted by -1/n!, and using QG = -I. To derive (52) from (51), take to be a point mass at x and f to be the indicator of a single point y, and cancel the common factor of k(y). Formula (52) for can also be understood probabilistically by consideration of ; that is, the probability that the chain starting at x ends with left limit y at time T having survived the additional killing at rate v. By conditioning on ,
On the other hand, by conditioning on the same probability equals
Comparing the two results yields (52). Note the parallel between (52) and the more obvious formula
where is the transition function of the chain killed at time T, and is the same for instead of T, namely the semigroup with Q-matrix instead of Q, where . Compare with formula 2.6.6 of Itô-McKean [30] in Kac's original Brownian setting. A common generalization of these formulae in an abstract setting could be given using h-processes, but this is left to the reader.