ODE solver behavior change from Math v3.2.0 to v3.3.0

[Niko]

@wds15

But yes, AM did give me nothing useful unless tolerances where set very tight in the old scheme.

This means what happened exactly? Stan just kept lowering the stepsize trying to achieve some acceptance rate, but failed?

@wds15

I also did some simulations where showed to me that with the new setting you can basically see it as more conservative approach and setup your tolerances somewhat looser to get the same speed as before (and similar actual precision).

Yeah, I believe that.

@wds15

How come you get back to this?

I was worried that weird things could happen for models involving adaptive ODE-solvers. @jtimonen found this thread. Before that we had been discussing whether it’s sensible to consider the sensitivities in the stepsize adaptation. I believe it is not, partially because (I believe) this makes it impossible to recover the exact ODE trajectory considered during HMC sampling using CmdStan.