Re: Square root of a negative rral value

On 2008-11-06 00:37:50 -0400, Terence <tbwright@xxxxxxxxx> said:

Well, yes, there was a 100 by100 matrix inversion, and a triple matrix
multiplication, and values wandered from about 0.4 up to 10**11, so it
might not be considered to be a stable calculation.
But the input is real-world data.

I was limited to REAL*8.
I accept the good point that 0.0 is a better default than the SQRT of
the absolute value, since the sample standard deviations SHOULD be
positive.

Some folks would take the view that the negative square root is
a hint that there are problems to be solved. Not shoved under the rug
or papered over or forgotten or otherwise ignored.

Anyway it works now, but I wonder just how many more pre-IBM360
algorithms have real-time problmes?? My best guide to "practical
methods" is the UK 1961 National Physica Laboratory White Paper on
Modern Computing Methods - often meaning hand-cranked calculators!

Why not take a look at Ake Bjorck's SIAM book on Least Squares. It
is much much much more modern than NPL's book on Modern Computing
Methods published by HMSO in the 1960s. A lot of its stuff is based
on fixed point with double length accumulation, the fi_2 mode.



.