在使用SciPy的scipy.special.ellipeinc
和ellipkinc
时,似乎存在一些数值不稳定的岛屿。例如,
>>> from scipy.special import ellipkinc
>>> ellipkinc(0.9272952180016123, 0.68359375000000011)
nan
>>> ellipkinc(0.9272952180016123, 0.6835937500000002)
2.0518660200390668
>>> ellipkinc(0.9272952180016123, 0.68359375)
1.0259330100195332
>>> ellipkinc(0.9272952180016123, 0.68359374)
1.0259330081166262
这发生在k ^ 2.sin ^ 2(phi)接近0.3的情况下,但这里的椭圆积分本身并没有什么异常,所以可能它是一个数值的东西。我不太了解这个算法的内部工作原理来说明错误,那么我最好的选择是什么?我想到了:
round(0.68359375000000011,8)
说,但这肯定会减慢我的代码?
答案 0 :(得分:1)
(这是对问题的扩展评论,而不是对如何解决问题的答案。)
这看起来像ellipkinc
中的错误。我得到一个浮点值,其中函数返回nan
,以及四个相邻浮点值,其中函数返回两次"正确"值:
In [91]: phi = 0.9272952180016123
In [92]: mbad = 0.68359375000000011
In [93]: m = np.nextafter(mbad, 0)
In [94]: mvals = []
In [95]: for j in range(10):
....: mvals.append(m)
....: m = np.nextafter(m, 1)
....:
In [96]: mvals
Out[96]:
[0.68359375,
0.68359375000000011,
0.68359375000000022,
0.68359375000000033,
0.68359375000000044,
0.68359375000000056,
0.68359375000000067,
0.68359375000000078,
0.68359375000000089,
0.683593750000001]
In [97]: ellipkinc(phi, mvals)
Out[97]:
array([ 1.02593301, nan, 2.05186602, 2.05186602, 2.05186602,
2.05186602, 1.02593301, 1.02593301, 1.02593301, 1.02593301])