Sharp ULP rounding error bound for the hypotenuse function
Let \(m= m(z)\) and \(e= e(z)\) be the mantissa and the exponent of \(z\in\mathbb{R}\), respectively. That is, \(| z|= m\times 2^e\), \(e\in \{0,\pm 1, \pm 2,\dots\}\), \({1\over 2}\leq m<1\). For a given precision \(p\in \{1,2,3,\dots\}\), let \(\text{ulp}(z)\), \(z\in\mathbb{R}\), be the function defined by \(\text{ulp}(z)= 2^{e-p}\), \(\text{ulp}(0)= 0\). The paper under review presents a ULP (units in the last place) bound for the hypotenuse function \(z= \sqrt{x^2+ y^2}\) in case of the calculations be performed according to the algorithm given below \[ x_1= x,\quad x_2=y,\quad x_3= x_1\times x_1= x^2,\quad x_4= x_2\times x_2= y^2, \] \[ x_5= x_3+ x_4= x^2+ y^2,\quad x_6= \sqrt{x_5}= z. \] The main result of the paper is the following Theorem: \(|\overline z- z|< \alpha\text{ ulp}(z)\), where \(\alpha= {5\sqrt 3+6\over 12(1- 7\times 2^{-p})}\), \(p\in\{3,4,\dots\}\). Numerical tests are reported in the paper to conclude that the bound \(\alpha= 1.222\), \(p\geq 15\), practically cannot be improved.
This page was built for publication: Sharp ULP rounding error bound for the hypotenuse function
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4240586)
