I think that Dmitri overstates his case a bit. This multiplication in observation space works for some algorithms, not for others. Ordinary least squares regression is somewhat of an exception here. Logistic regression is a simple counter-example.
It is still useful to have a vector weight and it helps users. It may be useful in some situations to also all a full correlation matrix, but I haven't had a need for that yet. On Sun, Feb 22, 2009 at 11:24 AM, Dimitri Pourbaix <pourb...@astro.ulb.ac.be > wrote: > Either one considers the full weighting matrix (including potential > correlation between observations) or one does not account for any weight > at all. By premultiplying both the function matrix and the observation > vector by the square root of the weight matrix, one can forget about it > completely in the rest of the computation. > -- Ted Dunning, CTO DeepDyve