Maybe I misunderstand, but I think that's what they're doing, even if they don't use that terminology; I don't have the time to read the paper and code and deta

Maybe I misunderstand, but I think that's what they're doing, even if they don't use that terminology; I don't have the time to read the paper and code and detail right now, but from a 5-second glance that's an RP method. They multiply a given matrix by a random matrix, do computations on the result (apparently some power method that efficiently floats bigger eigenvalues in the small base), and project back to the original space.