Estimation of denormalization errors (Q1105983)

The authors consider the denormalization errors of floating-point addition by computer. Denormalization errors arise when the computer equalizes the orders of two addends shifting, and perhaps rounding, the mantissa of the addend of lower order. The expected value and the variance of such error is at first calculated for the sum of two numbers. The errors and the addends are considered as random variables with suitable distributions. Then the authors examine the sum of many addends in two different cases: when the addends are arbitrarily ordered and when they are sorted in a particular way, i.e. at every step the two smallest numbers are replaced by their sum. In the first case the authors calculate an upper and a lower bound of the expected value of the final error, in the second case they calculate an upper bound of the final error and arrive to confirm the well known conclusion that the method with sorting is more advantageous.