Correct posteriors but many divergences

[ricardoV94]

Those look close enough. The finite differences become less reliable with smaller epsilon. Usually I look at 1e-5 or 1e-6