{ "name": "ErfApprox", "summary": "The error function by Abramowitz and Stegun's formula 7.1.26, whose stated absolute error is below 1.5e-7 for every X.", "tags": [ "special functions", "approximation", "statistics" ], "sources": [ { "title": "Abramowitz and Stegun, Handbook of Mathematical Functions (National Bureau of Standards, 1964), section 7.1, formula 7.1.26", "url": "https://archive.org/details/HandbookOfMathematicalFunctions" } ], "cases": [ { "call": "ErfApprox(0.0)" }, { "call": "ErfApprox(0.5)" }, { "call": "ErfApprox(-1.0)" }, { "call": "ErfApprox(2.0)" }, { "call": "ErfApprox(3.5)" }, { "call": "ErfApprox(0.0451)" }, { "call": "ErfApprox(1.0e-5)" } ], "references": [ { "lang": "python", "level": "plain", "file": "ref/python/plain.py", "note": "The same formula." }, { "lang": "python", "level": "libm", "file": "ref/python/libm.py", "note": "Python's math.erf, which uses the platform C library's erf: the approximation must stay within its stated 1.5e-7 of it. ErfApprox(0.0451) sits next to the largest error, about 1.39e-7.", "abs": 1.5e-07 }, { "lang": "mojo", "level": "plain", "file": "ref/mojo/plain.mojo", "note": "The same formula in Mojo. The compiled program calls Mojo's own exp(), which in Mojo 1.2.0 is off by up to about 6.5e-12 of the value on [-16, 0], so the last digits are not the pseudo's. The result is 1 - Poly * exp(-A * A), so the difference is small in absolute terms (at most 1.2e-13 over 640,000 values of X between -8 and 8, many of them near 0, computed at run time by Mojo 1.2.0), but not in relative terms near 0, where erf is small: at X = 1.0e-5 the difference is about 8e-12 of the value.", "abs": 2e-13 } ] }