Convergence properties of a class of reduced gradient algorithms in linearly constrained minimization