On Thu, 27 Jul 2017, Sven Schreiber wrote: We had a (simple-minded) check for PSD-ness but it was in the wrong place in the code. That's now fixed in git and snapshots, so psdroot should now fail instead of producing garbage ("factors" containing NaNs). And it will produce a valid factor in at least some cases where cholesky fails. Example: <hansl> set verbose off set seed 321457 # n = dimension of matrix, r = rank n = 8 r = 6 # with the above psdroot works, cholesky does not matrix A = zeros(n, n) loop i=1..r -q x = muniform(n, 1) A += x*x' endloop # take a peek at eigenvalues eval eigensym(A) print "psdroot" catch L = psdroot(A) if $error == 0 print L eval A - L*L' else print "psdroot failed" endif print "cholesky" catch L = cholesky(A) if $error == 0 print L eval A - L*L' else print "cholesky failed" endif </hansl> However, it's now clear to me that psdroot() doesn't really do what it advertises. There's new lapack-based code in place in libgretl that should do a proper job on rank-deficient matrices but it needs more refinement and testing before I replace the original. Allin