See Peter et al., 2020 for the description of the dataset. (A) displays the recovered graph under the edge parameterization. (B) displays the recovered graph under the node parameterization. Both parameterization have their own regularization parameters λ and α, but these parameters are not on the same scale. We set and for the node parameterization which is seen to yield similar results to those in Peter et al., 2020. For the edge parameterization, we set and so that the resulting graph reveals similar geographic structure to the node parameterization. We also set the lower bound . From the plots, it is worth noting two important distinctions: (1) We see the migration surfaces shown in (B) recover sharper edge features while the migration surfaces in (A) are overall smoother. This is attributed to the fact that node parameterization has its own additional regularization effect on the edge weights, and in order to achieve similar degree of regularization strength for the edge parameterization, it needs a higher regularization parameters, which results in more blurring edges than the node parameterization. (2) When measuring correlation of the estimated allele frequencies among nodes, we find that Deme B is the node with the second highest correlation to Deme A, whereas Deme C (and nearby demes) is not as much correlated to Deme A compared to Deme B. Panel (A) reflects this feature by exhibiting a corridor between Deme A and Deme B and reduced gene-flow beneath that corridor. This reduced gene-flow disappears in (B), even if the regularization parameters are varied over a range of values. Additionally, Deme D is most highly correlated to Deme E, F, and G, and this is implicated by a long-range corridor connecting those demes appearing in Panel (A) while not shown in (B). These results suggest that the form of the node parameterization is perhaps too strong and in this case limits the model’s ability to capture desirable geographic features that are subtle to detect.