Optimal Batching Under Computation Budget.

Batching methods operate by dividing data into batches and conducting inference by aggregating estimates from batched data. These methods have been used extensively in simulation output analysis and, among other strengths, an advantage is the light computation cost when using a small number of batches. However, under computation budget constraints, it is open to our knowledge which batching approach among the range of alternatives is statistically optimal, which is important in guiding procedural configuration. We show that standard batching, but also certain carefully designed schemes using uneven-size batches or overlapping batches, are large-sample optimal in the sense of so-called uniformly most accurate unbiasedness from a dual view of hypothesis testing.