Error estimation of floating-point summation and dot product (Q766234)

The author introduces the improvement of the well-known Wilkinson-type estimates bounds for the error of standard floating-point recursive summation and dot product by a factor 2 smaller. According to the error analysis, the author derives new error bounds for these two operations, which are nice by omitting higher order terms and close to what would be expected and are easy to be computed in floating-point rounding. Meanwhile, the new bounds fill some theoretical gap, demonstrate the power of the unite in first place-concept and do not require directed rounding. Except the theoretical merit, the bounds introduced in the paper may be useful in applications whose rigorous bounds are mandatory and directed rounding is not available such as computer-associated proofs and interval libraries for scalar, vector and matrix operations.