¥ See http://afni.nimh.nih.gov/doc/misc/3dDeconvolveSummer2004/
¥ Equation solver: Gaussian elimination to compute R matrix pseudo-inverse was
replaced by SVD (like principal components)
H Advantage: smaller sensitivity to
computational errors
H ÒCondition numberÓ and Òinverse errorÓ
values are printed at program startup, as
measures of accuracy of pseudo-inverse
H Condition number < 1000 is good
H Inverse error < 1.0e-10 is good
¥ 3dDeconvolve_f program can be used
to compute in single precision (7 decimal
places) rather than double precision (16)
H For better speed, but with lower
numerical accuracy
H Best to do at least one run both ways to
check if results differ significantly (SVD solver should be
safe)