No label defined (Q4367765)

The authors consider solution methods for the equality constrained quadratic programming problem. In the case when the Hessian \(G\) of the objective function is positive definite and extremely sparse (sparsity more than 95\%) and the number of constraints are at least 10\% of the number of the unknown there is an interesting approach, known as splitting algorithm, which consists in transforming the original problem into a sequence of equality constrained quadratic subproblems with \(D\) as objective matrix (\(D\) is a symmetric positive definite matrix satisfying the relation \(G= D+ H\), a \(P\)-regular splitting). When the number of constraints is reasonably small, elimination (direct) methods are effective solvers for least distance problems. In order to show the behaviour of the splitting method, the authors report results, carried out on a Cray C90 on a test problems arising in constrained bivariate interpolation. The authors also analyze the numerical stability of the direct elimination method following a backward error analysis procedure. They observe that the bounds of data perturbations do not depend on condition numbers of the matrices \(G\) and \(C\) (matrix of linear equality constraints), thus potential instabilities are not present in the bound of the algorithmic error.