Anonymous user
Talk:Statistics/Basic: Difference between revisions
→Wrong emphasis in 'Extra'?: robustness is always a reasonable extra-credit thing on RC
No edit summary |
(→Wrong emphasis in 'Extra'?: robustness is always a reasonable extra-credit thing on RC) |
||
Line 10:
:: Making it numerically stable, that's challenging. It's easy enough if you have a small number of values of all about the same scale, but that's not always the case. –[[User:Dkf|Donal Fellows]] 17:11, 2 July 2011 (UTC)
::: That's actually rarely a problem in the real world. If a distribution is narrow, scale difference is small; if distribution is wide, losing some precision on really small numbers wouldn't affect either average or stddev. It probably will be a concern only when you have a few very large numbers and a lot of smaller ones (say < 10^-16 relatively in abs, but about 10^16 in quantity), but what kind of physical measurement would give a distribution like that? In any event, I didn't say anything about that in the task; the distribution used is uniform, it really can't get much simpler than that. --[[User:Ledrug|Ledrug]] 17:35, 2 July 2011 (UTC)
:::: The distribution comes up quite often when working with quantities that follow a power law (i.e., where they are distributed more evenly in log space) which is actually quite often. In any case, the warning about such things is relevant because someone ''will'' copy the code on this page and use it unwisely; there are whole legions of fools who want to program by cut-n-paste only and without any thought for side conditions, but even so it is still something that we should note for our own consciences. Write robust code for extra credit! –[[User:Dkf|Donal Fellows]] 18:03, 2 July 2011 (UTC)
| |||