… are hell. Heiko and I have been spending the day trying to get code to work for a numerically stable streaming update of a Cholesky factor in an adaptive Importance Sampling setting. This is less then trivial, also because it has to be done in a weighted fashion (because of Importance Sampling) and weights should be kept in logspace for stability. Thinking I will have to go through this again for an Eigendecompositon instad of Cholesky for a collaboration planned with Mathieu Gerber does not make things easier to stomach.
This day again strengthens my wish to move more in the direction of Judith-Russeau-statistics, i.e. away from computers, towards pen and paper.
PS: Does anybody have code solving exactly this problem (we have tons for almost this problem)?