sure the approximate square root L isn’t triangular, but the approximate Matrix A = L * L^T
is. Thus the eigenvalues M + 1 … N should be positive and close to 0. If 0 the matrix is singular.
It would be better to construct the matrix like this, that diag(A) = I. Eigenvalues of M+1 … N = eps >0.
I adapted following paper to be able to use distributions like logit, poisson, etc.
I discovered that these GLM stuff usually overfits. A noise parameter is need.
I think its called nugget parameter.
However, the eigenvalue/vector decomposition is slow and not very stable. This is
not the case, if we could construct a reasonable eigenvectors/-values decomposed
matrix beforehand.