Weird default tolerance in testing function

[mfherbst]

This default rtol of 1e8 strikes me as odd.

AtomsCalculators.jl/src/submodules/Testing.jl

Line 31 in 2ab8742

function test_forces(sys, calculator; force_eltype=nothing, rtol=1e8, kwargs...)

I have not read through the full mechanics, but I'd expect something <1, so I suspect it's a typo and 1e-8 is meant.

[cortner]

good catch. And it should probably be sqrt(precision)

[tjjarvinen]

Thanks for noticing this.

The reason for the correct value 1e-8 is that only very special methods have accuracy more than 8 significant numbers. So there should be no need smaller tolerance in practive.

On the other hand there can be issues with multithreading and even different implementations of calculators for different calls (forces vs energy_forces etc.) that cause small variations. So, it is better to have relatively loose tolerance here.