JacobiSVD.solve() boils down to:

Code: Select all

    dst = dec().matrixV().leftCols(diagSize)
        * invertedSingVals.asDiagonal()
        * dec().matrixU().leftCols(diagSize).adjoint()
        * rhs();

Would it make a performance improvement if everything except rhs was pre-multiplied, so it does not have to be multiplied each time solve() is called?

If yes, would that significantly decrease numerical stability?

yes this would significantly reduce numerical stability, and performance wise this would only make sense if the rhs have a lot of columns, and/or solve is called fot a large number of rhs. Otherwise premultiplying would be considerably slower.