Application of Prior is Inconsistent

[cortner]

The way we apply the prior is different across different scenarios. This creates confusion. Fundamentally, the question seems to be whether the prior is part of the model or part of the solver. In other words should it rescale the basis in ACEpotentials, or should it rescale the lsq system inside of ACEfit? There is no rush to address this, but in a future version it would be nice to become more systematic about this.

[wcwitt]

In some ways, I like rescaling the basis, such that the prior is completely decoupled from the solver.

However, we have also discussed solvers that might adjust the prior (e.g. a BLR that optimizes the prior strength). So maybe the prior has to go to ACEfit?

[cortner]

I'm really not sure. I do think that there is a "true prior" which lives in the model. (And in a near-future rewrite I want this to simply become a rescaling of the radial basis!)

Maybe we just need to accept the fact it can be in both. And for some solvers there can be an "adjusted prior".