Posted by Don Visco on June 19, 1998 at 15:30:36:
I need to accurately invert a 108 x 108 real,
non-symmetric matrix. Currently I am using the
LU decomposition routine in Numerical Recipes but
when I check the inverse I find that some of the
off-diagonal elements are not zero. In fact,
the last two rows have elements which are in the thousands.
When I ran the code using the -d16 switch on F90, the
problem went away but it took about 15 times longer.
I tried iterative improvement but that doesn't help.
Any suggestions on a robust algorithm which does not require
the -d16 switch. Please e-mail me directly if you don't mind.