Computing projections for the Karmarkar algorithm (Q806966)

For computing the projection of a vector onto the nullspace of a matrix some alternatives are considered. Especially the sparsity of the computed matrices is investigated. The author works with an extended but indefinite matrix which is usually sparser than the originally used normal matrix. For the solution process a subroutine from Harwell is modified. Computational results are given for twelve test problems in the Netlib test set. These results indicate a computational advantage for factoring the extended matrix.