Stochastic rounding
From Wikipedia, the free encyclopedia
In statistics, stochastic rounding refers to a method of rounding numbers such that when rounding a fraction that falls exactly between two integers (e.g. 3.5), the decision on whether to round up or down is random (or pseudorandom). Stochastic rounding prevents totals of large numbers of calculations from drifting up or down, but it can also make the results of a statistical analysis difficult to verify or replicate.