I am using PARDISO to compute a Schur complement matrix. I am testing with real matrix meaning that I am using complex numbers with 0 imaginary parts. When the matrix size is small, the imaginary part of the result is zero but when the size increases, I get a small imaginary part. For example, for a 110000X11000 matrix and a Schur matrix size of about 100x100, the imaginary part is about 1e-6 to 1e-5 smaller than the real part. I would assume the imaginary part should be always 0. I know that numeric error can exist but I would think it is impossible to have anything than 0 by combining zero numbers.Also, a 1e-6 accuracy in double precision does not look that good. Am I wrong? If it is needed I can provide an example. I will need to make a simplified code for that.
Thanks,
Marc