On monotone function closure of perfect and statistical zero-knowledge

Assume we are given a language $L$ with an honest verifier perfect zero-knowledge proof system. Assume also that the proof system is a $leq 3$ move Arthur-Merlin game. The class of such languages includes all random self-reducible language, and also any language with a perfect zero-knowledge non-interactive proof. We show that such a language satisfies a certain closure property, namely that languages constructed from $L$ by applying certain monotone functions to statements on membership in $L$ have perfect zero-knowledge proof systems. The new set of languages we can build includes $L$ itself, but also for example languages consisting of $n$ words of which at least $tleq n$ are in $L$. A similar closure property is shown to hold for the complement of $L$ and for statistical zero-knowledge. The property we need for the monotone functions used to build the new languages is that there are efficient secret sharing schemes for their associated access structures. This includes (but is not necessarily limited to) all monotone functions with polynomial size monotone formulas.