Abstract
This paper is concerned with the differences between deterministic and randomized finite-memory decision rules. It is shown that for any hypothesis-testing problem there exists a b < CO such that, for any B, the optimal deterministic rule with B + b bits in memory has a lower error probability than the optimal randomized rule with B bits in memory. Suboptimal deterministic rules with this property are de- monstrated. These deterministic rules lose at most b bits. Thus for large memories the fraction of memory, measured in bits, lost by deterministic rules is negligible.

This publication has 5 references indexed in Scilit: