On the interchange of derivative and expectation for likelihood ratio derivative estimators

Sufficient conditions for the validity of interchange between derivative and expectation, in the context of likelihood ratio gradient estimation, were given in L'Ecuyer 1990. The aim of this paper is to shed additional light on these conditions and introduce specific variants of them, which are often easier to check. Sufficient conditions for the derivative estimator to have finite moments up to a given order are also given and illustrated by examples. In particular, we give an example of an unbiased derivative estimator which satisfies the interchange conditions but which has infinite variance.