1: \begin{abstract}

2: We propose a gradient-based method for quadratic programming problems with a

3: single linear constraint and bounds on the variables.

4: Inspired by the GPCG algorithm for bound-constrained convex

5: quadratic programming [J.J. Mor\'e and G. Toraldo, SIAM J.~Optim.~1, 1991],

6: our approach alternates between two phases until convergence: an

7: identification phase, which performs gradient projection iterations until

8: either a candidate active set is identified or no reasonable progress is

9: made, and an unconstrained minimization phase, which reduces the objective

10: function in a suitable space defined by the identification phase, by

11: applying either the conjugate gradient method or a recently proposed

12: spectral gradient method. However, the algorithm differs from GPCG not only

13: because it deals with a more general class of problems,

14: but mainly for the way it stops the minimization phase.

15: This is based on a comparison between a measure

16: of optimality in the reduced space and a measure of bindingness

17: of the variables that are on the bounds, defined by extending the concept

18: of proportional iterate, which was proposed by some authors for box-constrained problems.

19: If the objective function is bounded, the algorithm converges to a stationary point

20: thanks to a suitable application of the gradient projection method in the identification

21: phase. For strictly convex problems, the algorithm converges to the optimal solution in a finite

22: number of steps even in case of degeneracy. Extensive numerical

23: experiments show the effectiveness of the proposed approach.

24: \end{abstract}

25: