On nonnegative unbiased estimators (Q2343962)

The authors study the existence of algorithms that generate almost surely estimators that are non-negative. They show that given a non-constant function \(f:\mathbb R\to \mathbb R_+\) and a sequence of unbiased estimators of \(\lambda\in \mathbb R\), then there does not exist generally an algorithm giving almost surely a nonnegative unbiased estimator of \(f(\lambda)\in \mathbb R_+\). This means that the ability to sample an unbiased estimator \(X\) of a quantity \(\lambda\) is not enough to obtain a nonnegative estimator of \(f(\lambda)\). However, when there is added an additional information such that almost sure lower or upper bounds on \(X\) are available, then the required algorithms (schemes, factories) might exist. The case when \(f\) is increasing and the support of \(X\) is \([a,\infty)\) remains partly unsettled. Even when an \(f\)-factory for a given problem exist, then the paper indicates that in some contexts nonnegative unbiased estimators cannot be obtained, and thus the pseudo-marginal approach cannot be applied. The assumed assumptions are consistent with some recent Monte Carlo methods suggested for large data sets that take an advantage of almost sure bounds to bypass the evaluation of the full likelihood. The authors also discuss implementable schemes based on a Bernoulli factory or on a random truncation of infinite series. The discussed algorithms terminate with probability one, however, the expected computational time is not necessarily finite.