What next?

The unimpressive improvement in the image quality seems to be due to two factors. First, the introduced slowness change was too low in spatial frequency. The initial slowness model was already well determined for features at this scale. In addition, 2-D tests showed that image quality was most improved when allowing smaller scale changes to the velocity. Relaxing the smoothness constraint (by both decreasing $\epsilon$ in fitting goals (4) and iterating more) is called for. The second weakness seems to be poor constraint of the lower portion of the model. The obvious answer is to pick more reflectors. Both of these changes have been made, but unfortunately time constraints did not allow the results to be included in this paper.