One chain of four with huge nleapfrog, tiny step size

In the model I’m currently working on (a multivariate state-space model), three out of four chains ran with no problems (cmdstan 2.16, 1000 warmup, 1000 sampling iterations completed in about two hours). The fourth chain has only managed 30 iterations in about 17 hours, and it’s now on a step size of 1E-9, and 1048580 leapfrog steps per iteration. The log probability is much lower than for the other chains. I’d set the max tree depth to 20, which was more than enough for the other chains, but this chain is hitting 20. I left everything else on the default settings. I’m guessing that it’s got stuck in a region of parameter space far from the mode. Is there anything obvious I can do other than use stronger priors?

I would plot the state of the system implied by the stuck chain and see if it’s a reasonable state. That’ll give you good hints on how to adjust the priors. It may be that a specific prior just didn’t mean what you think it does.