Meaningless recursion in pytram Estimator base class

[franknoe]

in the Estimator base class we have:

    @property
    def pi_i(self):
        return np.exp(-self.f_i)

    @property
    def f_i( self ):
        return -np.log(self.pi_i)

and

    @property
    def pi_K_i(self):
        return np.exp(-self.f_K_i)

    @property
    def f_K_i(self):
        return -np.log(self.pi_K_i)

which are endless recursions and don't do anything. I guess they are effectively never called because they are overridden in the base class, but then at least one of the two should have no implementation and raise a NotImplementedError when called.

[cwehmeyer]

Depending on the estimator derived from this (abstract) class, it is one or the other version which is overridden. I think this approach is reasonable since we do not need to implement things twice.

[franknoe]

I understand the logic, but I probably wouldn't do it that way because it's dangerous when new Estimators are developed without understanding this logic.
I think a good default behavior is to expect that the estimator computes the energies and always use the energies as a default internal representation as this is numerically stable. The stationary distributions can then be computed robustly from the energies. So one would implement the stationary distributions in the base class and only give an empty implementation of the energy-properties there that has to be overriden.

Anyway this is low-prio.