I noticed that the relax_and_split() method in the permanent magnet optimization tools was .pop()'-ing keys such as nuand the regularization terms out ofkwargs`, even though these should actually be passed to the convex MwPGP algorithm in the code block:

algorithm_history, _, _, m = convex_step(
    A_obj=pm_opt.A_obj,
    b_obj=pm_opt.b_obj,
    ATb=ATb,
    m_proxy=np.ascontiguousarray(m_proxy.reshape(pm_opt.ndipoles, 3)),
    m0=np.ascontiguousarray(m.reshape(pm_opt.ndipoles, 3)),  # note updated m is new guess
    m_maxima=mmax,
    **kwargs
)

This causes a bug where if you optimize with relax_and_split(), the default value of nu = 1e100 is used since this key has been removed from the kwargs dict when it is passed on to the convex MwPGP method. This pull request changes it to a get() method for the relevant keys instead.