Remark 8 Similar to Algorithm 1, in the inner iterations of
Algorithm 2 any initial values '(0),w(0) can be used. By
our experience, the corresponding values for these variables
obtained at the previous outer iteration are usually a good
initial state for warm-starting the inner layer computations
because they generally do not vary significantly after the first
few outer iterations, and so Algorithm 2 can converge much
faster.