It appears that the effective sample size is allowed to be greater than the number of samples taken here: https://github.com/pymc-devs/pymc3/blob/master/pymc3/diagnostics.py#L262. For noisy autocorr. estimates, maybe this should be clipped at the maximum number of samples like this: ``` python return min(n * m, int(m * n / (1. + 2 * rho[1:t].sum()))) ``` ?