Automatic linear correction of rounding errors (Q5952346): Difference between revisions

The standard analysis of the effect of rounding errors in floating point arithmetic is based on the estimation of computed value \(fl(.)\) with respect to the elementary errors bounded by the so-called epsilon machine. A first order approximation of the global rounding error accumulating elementary rounding errors is obtained using some numerical differentiations. The principle of the proposed method is to apply also \(fl(.)\) estimation to above global rounding error and then to correct two times the computed value. The proposed automatic algorithm enhances the stability of some numerical algorithms and its efficiency is close to double the precision of the computation.