10.10 Divergences

We will now explore divergences, a diagnostic that is exclusive to NUTS, as it is based on the inner workings of the method and not a property of the generated samples. Divergences are a powerful and sensitive method that indicate the sampler has most likely found a region of high curvature in the posterior that cannot be explored properly. A nice feature of divergences is that they usually appear close to the problematic parameter space region, and thus we can use them to identify where the problem may be.

Let’s discuss divergences with a visual aid:

Figure 10.14: Pair plot for selected parameters from models model_c and model_nc

As you can see, Figure 10.14 shows the following three subplots:

• The left subplot: We have a scatter plot for two parameters of model model_c; namely, one dimension of the parameter b (we just picked one at random – feel free to pick a different one), and the logarithm of the parameter a. We take the logarithm because...