Hmmmmmm… I’m using cmdstan for my sins (ie I don’t have a version of RStan using 2.17).
But I’m confused as to why the init needs to be so small - the “true” value of sigma is 2 (the posterior mean from the full model is slightly higher), which I would’ve expected to not be badly scaled. On the log scale, the correct posterior from sigma (the only parameter) is in (-0.52,1.1).
My experience so far is that they either fail completely immediately or never move. But again - this is a very easy zero finding problem (the minimum of a log-concave density), so I’m not sure what’s wrong. It also decouples so you can plot the curves and they’re not that weird… (almost linear in a broad neighbourhood of the zero)