|
1 | | -The GSL Bugs Database is at http://savannah.gnu.org/bugs/?group=gsl |
2 | | - |
3 | | -This file was generated from it at Thu Jul 18 10:23:38 2013 |
4 | | - |
5 | | ------------------------------------------------------------------------- |
6 | | -BUG-ID: 28267 |
7 | | -STATUS: Open/Postponed |
8 | | -CATEGORY: Accuracy problem |
9 | | -SUMMARY: poor convergence region for gsl_sf_hyperg_1F1 |
10 | | - |
11 | | -The magnitude of the error is greater than the value itself for |
12 | | -a<<0, b>0, x>>0 |
13 | | - |
14 | | -gsl_sf_hyperg1F1(-3.78e+01, 2, 1.035e+02) => -7.00055e+18 +/- 3.77654e+19 |
15 | | - |
16 | | ------------------------------------------------ |
17 | | -From: Weibin Li < [email protected]> |
18 | | - |
19 | | -Subject: [Bug-gsl] gsl_sf_hyperg_1F1 |
20 | | -Date: Mon, 30 Nov 2009 15:26:11 +0100 |
21 | | - |
22 | | -Hi, guys |
23 | | - |
24 | | -I experienced bugs with gsl_sf_hyperg_1F1. The version is gsl_1.12, but |
25 | | -the testing is also done with gsl_1.13, using a mac with version 10.5.8 |
26 | | -and hp-workstation with gnome_2.24.1. |
27 | | - |
28 | | -#include<stdlib.h> |
29 | | -#include<stdio.h> |
30 | | -#include<math.h> |
31 | | -#include "gsl/gsl_sf_hyperg.h" |
32 | | - |
33 | | -int main(int argc, char **argv) |
34 | | -{ |
35 | | - int ii; |
36 | | - double Ri; |
37 | | - double v0; |
38 | | - |
39 | | - gsl_sf_result r; |
40 | | - for(ii = 10; ii< 140;ii++) |
41 | | -{ |
42 | | - Ri = 2013; |
43 | | - v0 =38.86871 +ii*2.5e-10; |
44 | | - |
45 | | - gsl_sf_hyperg_1F1_e(1.0-v0,2,2.0*Ri/v0,&r); |
46 | | - printf("%lg\t%14.13g\t%14.13g\n",v0,r.val,r.err); |
47 | | - |
48 | | - } |
49 | | -return 0; |
50 | | -} |
51 | | - |
52 | | -The output of above code shows an extremely large error. by increasing |
53 | | -Ri, the relative error decreases. Is there any idea to fix this? |
54 | | - |
55 | | -Thank you very much. |
56 | | - |
57 | | -Best |
58 | | - |
59 | | -Weibin |
60 | | - |
61 | | --It's inherently difficult to compute the value in this region either way as there is massive cancellation in both the series and the Kummer transformed series.I cannot find any algorithm which handles this case.Confirmed |
62 | | - |
63 | | ------------------------------------------------------------------------- |
64 | | -BUG-ID: 21828 |
65 | | -STATUS: Open/Confirmed |
66 | | -CATEGORY: Performance |
67 | | -SUMMARY: suboptimal performance of gsl_fdfsolver_lmsder |
68 | | - |
69 | | -From: "Alexander Usov" < [email protected]> |
70 | | - |
71 | | -Subject: [Help-gsl] Strange performance of gsl_fdfsolver_lmsder |
72 | | -Date: Wed, 24 Oct 2007 20:45:01 +0200 |
73 | | - |
74 | | -Hi all, |
75 | | - |
76 | | -I am currently working on the problem involving source extraction from |
77 | | -astronomical images, which essentially boils down to fitting a number of |
78 | | -2d gaussians to the image. |
79 | | - |
80 | | -One of the traditionally used fitters in this field is a Levenberg-Marquardt, |
81 | | -which gsl_fdfsolver_lmsder is and implementation of. |
82 | | - |
83 | | -At some moment I have notices that for the bigger images (about 550 |
84 | | -pixels, 20-30 parameters) gsl's lmsder algorithm spends a large fraction |
85 | | -of the run-time (about 50%) doing household transform. |
86 | | - |
87 | | -While looking around for are different minimization algorithms I have made |
88 | | -a surprising finding that original netlib/minpack/lmder is almost twice faster |
89 | | -that that of gsl. |
90 | | - |
91 | | -Could anyone explain such a big difference in performace? |
92 | | - |
93 | | --- |
94 | | -Best regards, |
95 | | - Alexander. |
96 | | - |
97 | | -_______________________________________________ |
98 | | -Help-gsl mailing list |
99 | | - |
100 | | -http://lists.gnu.org/mailman/listinfo/help-gsl |
101 | | - |
102 | | - |
103 | | -From: Brian Gough < [email protected]> |
104 | | -To: "Alexander Usov" < [email protected]> |
105 | | - |
106 | | -Subject: Re: [Help-gsl] Strange performance of gsl_fdfsolver_lmsder |
107 | | -Date: Thu, 25 Oct 2007 21:57:08 +0100 |
108 | | - |
109 | | -At Wed, 24 Oct 2007 20:45:01 +0200, |
110 | | -Alexander Usov wrote: |
111 | | -> At some moment I have notices that for the bigger images (about 550 |
112 | | -> pixels, 20-30 parameters) gsl's lmsder algorithm spends a large fraction |
113 | | -> of the run-time (about 50%) doing household transform. |
114 | | -> |
115 | | -> While looking around for are different minimization algorithms I have made |
116 | | -> a surprising finding that original netlib/minpack/lmder is almost twice faster |
117 | | -> that that of gsl. |
118 | | -> |
119 | | -> Could anyone explain such a big difference in performace? |
120 | | - |
121 | | -I have a vague memory that there was some quantity (Jacobian?) that |
122 | | -MINPACK only computes fully at the end, but in GSL it is accessible to |
123 | | -the user at each step so I felt I had to update it on each iteration |
124 | | -in the absence of some alternate scheme. Sorry this is not a great |
125 | | -answer but I am not able to look at it in detail now. |
126 | | - |
127 | | --- |
128 | | -Brian Gough |
129 | | - |
130 | | -_______________________________________________ |
131 | | -Help-gsl mailing list |
132 | | - |
133 | | -http://lists.gnu.org/mailman/listinfo/help-gsl |
134 | | - |
135 | | -1824 |
136 | | - |
0 commit comments