1 | r""" |
---|
2 | Show numerical precision of $sin(x)/x$. |
---|
3 | """ |
---|
4 | |
---|
5 | import numpy as np |
---|
6 | import scipy.special |
---|
7 | from sympy.mpmath import mp |
---|
8 | #import matplotlib; matplotlib.use('TkAgg') |
---|
9 | import pylab |
---|
10 | |
---|
11 | mp.dps = 150 # number of digits to use in estimating true value |
---|
12 | |
---|
13 | SHOW_DIFF = True # True if show diff rather than function value |
---|
14 | LINEAR_X = False # True if q is linearly spaced instead of log spaced |
---|
15 | |
---|
16 | def mp_gamma(vec): |
---|
17 | """ |
---|
18 | Direct calculation using sympy multiprecision library. |
---|
19 | """ |
---|
20 | return [_mp_fn(mp.mpf(x)) for x in vec] |
---|
21 | |
---|
22 | def _mp_fn(x): |
---|
23 | """ |
---|
24 | Helper funciton for mp_j1c |
---|
25 | """ |
---|
26 | #return mp.gamma(x) |
---|
27 | return mp.loggamma(x) |
---|
28 | |
---|
29 | def np_gamma(x, dtype): |
---|
30 | """ |
---|
31 | Direct calculation using scipy. |
---|
32 | """ |
---|
33 | x = np.asarray(x, dtype) |
---|
34 | return scipy.special.gammaln(x) |
---|
35 | #return scipy.special.gamma(x) |
---|
36 | |
---|
37 | def lanczos_gamma(x, dtype): |
---|
38 | coeff = np.asarray(( |
---|
39 | 76.18009172947146, -86.50532032941677, |
---|
40 | 24.01409824083091, -1.231739572450155, |
---|
41 | 0.1208650973866179e-2,-0.5395239384953e-5), dtype) |
---|
42 | |
---|
43 | x = np.asarray(x, dtype) |
---|
44 | tmp = x + np.asarray(5.5, dtype) |
---|
45 | tmp -= (x + np.asarray(0.5, dtype))*np.log(tmp) |
---|
46 | ser = np.ones_like(x)*np.asarray(1.000000000190015, dtype) |
---|
47 | for k,c in enumerate(coeff): |
---|
48 | ser += c/(x + np.asarray(k+1, dtype)) |
---|
49 | return -tmp+np.log(np.asarray(2.5066282746310005, dtype)*ser/x); |
---|
50 | |
---|
51 | def plotdiff(x, target, actual, label): |
---|
52 | """ |
---|
53 | Plot the computed value. |
---|
54 | |
---|
55 | Use relative error if SHOW_DIFF, otherwise just plot the value directly. |
---|
56 | """ |
---|
57 | if SHOW_DIFF: |
---|
58 | err = np.clip(abs((target-actual)/target), 0, 1) |
---|
59 | pylab.loglog(x, err, '-', label=label) |
---|
60 | else: |
---|
61 | limits = np.min(target), np.max(target) |
---|
62 | pylab.loglog(x, np.clip(actual,*limits), '-', label=label) |
---|
63 | |
---|
64 | def compare(x, precision): |
---|
65 | r""" |
---|
66 | Compare the different computation methods using the given precision. |
---|
67 | """ |
---|
68 | target = np.asarray(mp_gamma(x), 'double') |
---|
69 | direct = np_gamma(x, precision) |
---|
70 | approx = lanczos_gamma(x, precision) |
---|
71 | plotdiff(x, target, direct, 'scipy '+precision) |
---|
72 | plotdiff(x, target, approx, 'sasmodels '+precision) |
---|
73 | pylab.xlabel("x (arbitrary)") |
---|
74 | if SHOW_DIFF: |
---|
75 | pylab.ylabel("relative error") |
---|
76 | else: |
---|
77 | pylab.ylabel("ln(gamma(x))") |
---|
78 | pylab.loglog(x, target, '-', label="true value") |
---|
79 | if LINEAR_X: |
---|
80 | pylab.xscale('linear') |
---|
81 | |
---|
82 | def main(): |
---|
83 | r""" |
---|
84 | Compare accuracy of different methods for computing $3 j_1(x)/x$. |
---|
85 | :return: |
---|
86 | """ |
---|
87 | if LINEAR_X: |
---|
88 | qr = np.linspace(1,1000,2000) |
---|
89 | else: |
---|
90 | qr = np.logspace(-3,5,400) |
---|
91 | pylab.subplot(121) |
---|
92 | compare(qr, 'single') |
---|
93 | pylab.legend(loc='best') |
---|
94 | pylab.subplot(122) |
---|
95 | compare(qr, 'double') |
---|
96 | pylab.legend(loc='best') |
---|
97 | pylab.suptitle('ln gamma') |
---|
98 | |
---|
99 | if __name__ == "__main__": |
---|
100 | main() |
---|
101 | pylab.show() |
---|