GNU Scientific Library - Bugs: bug #38548, Rounding issues in gsl-histogram...
You are not allowed to post comments on this tracker with your current authentication level.
bug #38548: Rounding issues in gsl-histogram with integer numbers
Submitter: | Christian Leitold <cleitold> | ||
Submitted: | Tue 19 Mar 2013 03:53:57 PM UTC | ||
Category: | Accuracy problem | Severity: | 3 - Normal |
Operating System: | Debian 7.0 amd64 | Status: | None |
Assigned to: | None | Open/Closed: | Open |
Release: | 1.15 |
Attached Files
file #27634: varlist.dat added by cleitold (1KiB - application/x-ns-proxy-autoconfig - Sample data, use e. g. xmin = 0, xmax = n = 60)
Depends on the following items: None found
Items that depend on this one: None found
Carbon-Copy List
Follows 1 latest change.
Date | Changed by | Updated Field | Previous Value | => | Replaced by |
---|---|---|---|---|---|
2013-03-19 | cleitold | Attached File | - | Added varlist.dat, #27634 |
Powered by Savane 3.13-f8d8.
Corresponding source code
A rounding problem occurs in gsl-histogram when you try to make a histogram from pure integer data, where it is appropriate to choose a bin size of 1. I have tracked this down to the file histogram/init.c, where the array range is filled with the bin limits. There, in make_uniform one has
double f1 = ((double) (n-i) / (double) n);
double f2 = ((double) i / (double) n);
range[i] = f1 xmin + f2 xmax;
In the case of e. g. xmin = 0.0, xmax = 60.0, n = 60, this (on my
system, gcc 4.6, standard make command from the file archive) leads to
range[31] = 31.000000000000003552713679
where it should be, without rounding error,
range[31] = 31.000000000000000000000000
This then leads to the number 31 being put into bin number 30, instead of 31. A bunch of uniformely distributed integer data would then lead to a peak of double the average height at 30, and no entry at 31. I have attached a sample list of integer data where the problem occurs (varlist.dat). As discussed in the corresponding thread [1] on the mailing list, a proposed solution is
double dx = (xmax-xmin) / (double) n;
double f1 = xmin + i*dx;
double f2 = xmax - (n-i) * dx;
range[i] = (f1 + f2) * 0.5;
As I understand it, in particular after reading [2], there are certain special cases where floating-point arithmetic is just as accurate and error-free as integer arithmetic. This should be fulfilled here, when xmin, xmax, n are real integer values (even though represented as doubles), and thus dx (=1), f1 and f2 are integer values as well.
[1] http://lists.gnu.org/archive/html/bug-gsl/2013-02/msg00006.html
[2] http://en.wikipedia.org/wiki/Double-precision_floating-point_format