Have you tried an alternative parameterization (halfnormal, then rescaling)? If the initial energy is still bad, then the error is occurring somewhere else in the model. If not, then the spike at 1 in the beta dist is wreaking havoc (it might if the matrix is somewhat large).