The conditional probability at w = 0.1 might just be 0. Thems the breaks.
Exactly, replace w = pm.Beta('w', alpha=20., beta =20.) with w = 0.25. The sampler started complaining about convergence when I did it, which is normal because we know we’re in a very flat region of the likelihood space (i.e. there’s not much information out there for the sampler to go on). I imagine if you put in w=0.1 it will crash and burn. But it’s worth a try!