Abstract :
This paper introduces a technique for transforming partitioned quasi-Newton algorithms into H-form algorithms. The resulting algorithms have essentially the same space requirements as the standard partitioned QN method, but involve only a global matrix–vector multiplication (rather than a global solution by conjugate gradients) at each iteration. Results demonstrate that the method, which is highly suitable for parallelization, is competitive with other quasi-Newton methods in minimizing partially separable polynomial functions of large dimension.
