LBFGSMIN

For large problems, the BFGS algorithm may not be feasible, since it requires storing a $ k\times k$ matrix. The LBFGS (''L'' is for ''limited memory'') method is able to store all information for updates in vectors, which substantially reduced memory requirements. It may also be faster than the BFGS algorithm in some circumstances, since it may use fewer floating point operations per iteration, depending upon the size of the problem. While it will usually require more iterations that BFGS, it may be faster if each iteration is faster. Liu and Nocecal $ 1989$ is a reference.



Søren Hauberg 2008-04-29