Core Concepts
Even with moderate overlaps in the inputs of sloppy or accurate double-word addition algorithms, these algorithms can still guarantee error bounds of O(u^2(|a| + |b|)) in faithful rounding. Under certain conditions, the accurate algorithm can achieve a relative error bound of O(u^2) in the presence of moderate input overlaps.
Abstract
The content discusses the robustness of double-word addition algorithms, specifically the sloppy add (Algorithm 4) and accurate add (Algorithm 5) algorithms. It demonstrates that these algorithms can maintain reliable performance even when there is moderate overlap in the inputs.
Key highlights:
The authors prove that the Fast2Sum operation in both Algorithm 4 and Algorithm 5 remains valid even when the inputs have moderate overlap, as long as the overlap is within a certain range.
For the sloppy add algorithm (Algorithm 4), the authors provide a tight relative error bound of 3u^2 + O(u^3) when the input operands have opposite signs and the overlap is not severe (r(xh, yh) ≤ 1/2).
For the accurate add algorithm (Algorithm 5), the authors show that the relative error bound is (3o + 15)u^2 + O(u^3), where o is the maximum of the overlap factors for the two input operands.
The authors also discuss the application of these findings in double-word multiplication and interval arithmetic, demonstrating that the sloppy add algorithm can be safely used in place of the accurate add algorithm in many cases, with negligible precision costs but significant performance gains.