Interpretting `Warning: Close-to-zero parameters found` warning

[TorkelE]

When I try to computer intervals, I sometimes get these warnings:

 ^t^l Warning: Close-to-zero parameters found when using :LN_NELDERMEAD.
 ^t^t @ LikelihoodProfiler ~/.julia/packages/LikelihoodProfiler/Qi97K/src/method_lin_extrapol.jl:34
 ^t^l Warning: Close-to-zero parameters found when using :LN_NELDERMEAD.
 ^t^t @ LikelihoodProfiler ~/.julia/packages/LikelihoodProfiler/Qi97K/src/method_lin_extrapol.jl:34
 ^t^l Warning: Close-to-zero parameters found when using :LN_NELDERMEAD.
 ^t^t @ LikelihoodProfiler ~/.julia/packages/LikelihoodProfiler/Qi97K/src/method_quadr_extrapol.jl:35
 ^t^l Warning: Close-to-zero parameters found when using :LN_NELDERMEAD.
 ^t^t @ LikelihoodProfiler ~/.julia/packages/LikelihoodProfiler/Qi97K/src/method_quadr_extrapol.jl:35
 ^t^l Warning: Close-to-zero parameters found when using :LN_NELDERMEAD.
 ^t^t @ LikelihoodProfiler ~/.julia/packages/LikelihoodProfiler/Qi97K/src/cico_one_pass.jl:32
 ^t^l Warning: Close-to-zero parameters found when using :LN_NELDERMEAD.
 ^t^t @ LikelihoodProfiler ~/.julia/packages/LikelihoodProfiler/Qi97K/src/cico_one_pass.jl:32
 ^t^l Warning: Close-to-zero parameters found when using :LN_NELDERMEAD.

Here I compute the same interval using all methods. Is it that the warning happens for each method, and then in both directions?

Exactly what does this warning mean, and how worried should I be about it?

[ivborissov]

We included this warning because we have seen cases, when :LN_NELDERMEAD local_alg fails to find a solution when starting from ~zero initial values. I suggest you confirm the obtained results with another local_alg

[TorkelE]

Makes sense. Is this recorded somewhere in the output solution (i.e. is it possible to check that this have happened)?

[ivborissov]

No, unfortunately we don't record it, so it is just a warning. It doesn't mean "wrong result", it's more like "the algorithm may be sensitive to values close to zero". But you are right, if you are examining many parameters (e.g. in a loop) it is not that convenient to deal with warnings. So maybe it is worth to add special retcodes, will look at it

[TorkelE]

Sounds good. Yeah, some kind of record would be useful (although admittedly my application is also a bit niche). Feel free to closet his if you want :+1: