Hi, I wonder if there is some agreed-upon best way to reconstruct the matrix of a positive definite operator A using "sampling" (like in tomography). More in detail I want to do this: I have many small sets of basis functions. The sets are in general not orthogonal. I compute matrix elements <i|A|j>, where |i> and |j> belong to the same "set". In other words, in the non-orthogonal basis I know certain diagonal blocks of A, while the other elements are unknown. I want to determine an estimate of the off diagonal elements. One way of reconstructing A is to simply take any orthogonal basis for the union of all basis functions, and then work with that. However, the orthogonal basis is not unique. My question is if there is a best way of doing this?