Hi @certik . Maybe I have not emphasised the following point enough:
The solver is provided by another library that cannot be modified by me. The interface of the solver is fixed as solver(objective), and that of the objective is fixed as objective(x, f).
The question is essentially: what if I have an objective whose signature is bla(x, f, hyper_parameter), where hyper_parameter is unknown before the program starts? In other words, what if the implementation yof the objective depends on hyperparameters that are unknown beforehand? How to optimise such an objective using solver?
See my code above for details. See a similar example of root finding by @ivanpribec in a previous discussion if mine is not clear enough; see also the example in Note 12.18 on page 290 of WD 1539-1 J3/10-007r1, F2008 Working Document. As an example mentioned in the standard, it is not a wired situation but a common use case.
My approach is to use an internal procedure as you mentioned. But this leads to executable stack, which is a security issue. It is not allowed on some systems, and it causes a segfault of MATLAB R2025a.
Another approach is to pass the hyperparameter using a module variable, but it is not thread-safe or recursion-safe?
Is it possible to have a standard-confirming implementation that is thread-safe without internal procedures?