Common value for RCOND in xGESLY, xGELSD, etc?

[jgpallero]

Hello,
Can anyone tell me a general value for using in DGELSY and DGELSD routines? I suppose that the value depends on the type of problem and real data, but I suppose too that for most problems a common value can be used. For example, in GNU Octave, for solving a least squares problem via DGELSD, a RCOND value of -1 is set (http://hg.savannah.gnu.org/hgweb/octave ... x.cc#l2430). This imply that via DLALSD the actual RCOND value will be the machine epsilon via DLAMCH('Epsilon') (http://www.netlib.org/lapack/double/dlalsd.f). Could be epsilon a good value for common problems?
Thanks

[Julien Langou]

RCOND = -1 is good enough. If your data is only approximate and you know the quality of your approximation, say the first 3 digits of my data is correct then all other digits are just placeholders and are meaningless, then it would make sense to set RCOND to 1e-3. By default, RCOND=-1 uses the machine precision. So, in double precision, it assumes all 16 digit of your initial data are relevant for you. It's kind of hard to summarize this in three sentences ... I hope this gives you some clues. Julien. (There are also people that tune RCOND to find a "gap" in the singular values.)