Despite its general design intended for arbitrary optmization problem, the Limited Memory BFGS ``off the shelf'' code was successful in minimizing the Huber misfit function. The results achieved by the two algorithms are significantly different so that a direct comparison of cost is difficult. The Huber/BFGS combination tends to require about twice as much CPU time per iteration as the Least squares/CG.

Stanford Exploration Project

4/20/1999