Computational cost

As it stands, the cost of computing a weighting function of this kind is twice the cost of a single migration. Add the cost of the migration itself, and this approach is 25% cheaper than running two iterations of conjugate gradients, which costs two migrations per iteration.

However, the bandwidth of the weighting functions is much lower than that of the migrated images. This allows considerable computational savings, as modeling and remigrating a narrow frequency band around the central frequency produces similar weighting functions to the full bandwidth. Repeating the first experiment (${\bf m}_{\rm ref}={\bf m}_1$) with half the frequencies gives a $\mbox{NSD}=0.147$ - the same as before within the noise-level of the experiment.