BOBYQA violates bounds constraints by roundoff error

[JasmineTC]

What happened?

Hi, I'm using LN_BOBYQA to minimise a function inside a loop and feeding result of the previous minimisation into the next iteration. Overwhelming this works fine. However in one case the results of the minimisation is a floating point* amount out of bounds and so I get a "nlopt.nlopt.invalid_argument" error.

I can probably just bypass this problem by rounding the result (e.g. numpy.rint) but this leads to precision loss and overhead for a rare edge case.

*bound is 1e-4, result is 9.999999999999999e-05

How to reproduce the issue?

Unfortunately code too complicated to distil into useful MWE.

Version

NLOPT: 2.9.1
Python: 3.13.5
Numpy: 2.2.6

Operating System

Linux

Installation media

pip

Additional Context

No response

[stevengj]

Apparently there is a rounding error in the BOBYQA algorithm; I agree that this should be fixed.

As a workaround, I would use something like numpy.clip rather than rounding. clip should be harmless.

[JasmineTC]

I've put together a medium working example here: https://github.com/JasmineTC/MWEforNLOPT. There's a docker file included for reproducibility.