Figure 6.

Mean squared error in the approximation to the true posterior, as a function of the number of alignment samples. Shown for the pairwise globin example. Although the pair-HMM involves neighbour-dependent terms (leading to an affine gap penalty), the mean-field approximation leads to a better estimate of the true posterior until around 1000-2000 samples are taken. This is due to the presence of intersections between paths in the alignment DAG, which allows for a higher effective sample size to be obtained from the same number of alignments.