On Thu, 11 Nov 2021 10:26:30 GMT, Laurent Bourgès <lbour...@openjdk.org> wrote:
>> You made an amazing job ! >> Your bigdecimal impl looks good. >> I will play with your test experiment... asap. >> I think numerical accuracies are related to the dynamic /magnitude of >> values: points. >> The test should evaluate(max error) depending on the quick length(curve) ~ >> manhattan norm(curve arms), so small curves have less numerical error >> whereas huge curve (10^50 length) means points are wide spread and coeffs... >> solver...points are less accurate. >> >> I will try making sampled control points vary in log scale: 10^-6 to 10^38 >> (float max) so curve length will vary in huge range and determine the >> histogram of max(error) / length ratio. > > see condition number = magnitude range on such plot: > https://github.com/JuliaMath/AccurateArithmetic.jl/blob/master/test/figs/sum_accuracy.svg I just pushed a commit ( 40bda06 ) that addresses the machine error problem to my satisfaction. It includes a unit test that failed before that commit and now passes. At your convenience let me know your thoughts. (That commit continues to focus on very small doubles; it may (?) need help to address very large double errors.) ------------- PR: https://git.openjdk.java.net/jdk/pull/6227
