Plans

The initial commit is designed mostly to reproduce the results in Julia from https://github.com/ClaireYLin/regularized-field-map-estimation

Planned improvements:

• refine data (or weight) scaling in b0scale using new derivations
• store initial cost and then cost after each update in b0map when in track mode
• validate gradient / cost consistency using autodiff
• more in-place operations?
• factor the b0map code into smaller units
• ultimately the nonlinear CG with monotone line search could be generalized and useful elsewhere
• request that lldl support units (unlikely). so perhaps use reinterpret?
• timing test for whole-brain 2mm isotropic, e.g., 112 x 112 x 72
• https://github.com/korbinian90/ROMEO.jl for unwrapping

Start with the initial commit and inspect b0scale, b0map, and lldl; the issue lists these entry points but does not select a single task. Before coding, choose one improvement and define its validation, such as gradient/cost consistency or whole-brain timing, because the current issue has no concrete done criterion.