Accuracy of sinh and complex cos/sin #247
Conversation
| return x + (x * x * x) / 6 + (x * x * x * x * x) / 120; | ||
| } else { | ||
| return (Math.exp(x) - Math.exp(-x)) / 2; | ||
| } |
There was a problem hiding this comment.
Style note: You have a return statement in the if, so you don't need the else.
if (Math.abs(x) < 1) {
return x + (x * x * x) / 6 + (x * x * x * x * x) / 120;
}
return (Math.exp(x) - Math.exp(-x)) / 2;
There was a problem hiding this comment.
I think both versions are ok. @BigFav version is more compact, but on the other hand, an explicit else shows very clearly that the second return is only occurring in an else case (there are rules of thumb that say that every if should have an else prevents cases similar to these tricky fall-though cases of switches). Both are fine with me.
There was a problem hiding this comment.
I'll leave it as is for now, but would be happy with whatever you decide.
|
Good job catching this case. Most of my comments are style related (with the only exception of the failed test case), so hopefully this will be merged pretty smoothly. |
|
Thanks @pavpanchekha, nice improvement! Can you please fix this round-off error making the unit tests fail? After that I will merge your PR. |
The fix is to test that the imaginary part is nonzero, not that it is exactly 1e-100. In some software / hardware combos, it seems to return 1.0000000000000001e-100.
|
Thanks a lot @pavpanchekha |
Accuracy of sinh and complex cos/sin
|
This 5th degree taylor expansion of sinh is quite far off when you get close to -1 or 1. For example, for x = 0.99, it’s off by about 1.87e-4. I assume that’s dramatically more error than you want in a math library. For a graph of the error, here’s Wolfram Alpha: http://www.wolframalpha.com/input/?i=sinh%28x%29+-+%28x+%2B+%281%2F6%29+*+x%5E3+%2B+%281%2F120%29+*+x+%5E+5%29+from+-1+to+1 The GNU Scientific Library uses the degree 17 Taylor expansion for computing sinh in the range [-1, 1]: |
The
sinandcosfunction for complex arguments, and thesinhfunction for real arguments, are inaccurate when the inputs are very small. This is becauseMath.exp(x) - Math.exp(-x)returns zero for small x, instead of the more accurate2x.This patch replaces
sinhby a Taylor expansion when the input is small, which increases accuracy.