Explore the consistency of results

As we cannot reach the globally optimal partition, we may want to check the consistency of results from run to run. For example, we might want to calculate the description length of the data at a single point, \((K_a, K_b)\), repeatedly, without running through the whole heuristic. One can use,

oks.compute_and_update(ka, kb, recompute=True)

We then check an internal variable, which faithfully stores the description length (or entropy) that we have computed,

oks.bookkeeping_dl[(ka, kb)]

We retain a bookkeeping of other useful observables as well. They are bookkeeping_dl, bookkeeping_e_rs, and bookkeeping_mb. When we run this repeatedly, we are accessing the precision of the inference engine. Note that unless the underlying graph is super structured (e.g., bipartite cliques), the resulting description lengths vary.