Well, the suggestion was not to drop support for Diagonal(Fill(...)) but simply to convert them to ScalMat when a FiniteGP is constructed. The main advantage is that Diagonal(Fill(...)) would hit the same internal optimizations as ScalMat but we only have to implement them for ScalMat. Similar to eg observations as Matrix but internally we only have to deal with vectors.
Originally posted by @devmotion in #236 (comment)
Well, the suggestion was not to drop support for Diagonal(Fill(...)) but simply to convert them to ScalMat when a FiniteGP is constructed. The main advantage is that Diagonal(Fill(...)) would hit the same internal optimizations as ScalMat but we only have to implement them for ScalMat. Similar to eg observations as Matrix but internally we only have to deal with vectors.
Originally posted by @devmotion in #236 (comment)