Figure 4:

Distance between the true and estimated invariant subspaces computed with MASE (scaled and unscaled ASE) and ASE of the mean of the graphs for a sample of graphs distributed according to a multilayer SBM with 3 communities. On the left panel, all graphs have the same expected matrix. The three methods perform almost the same, reducing the error as the sample size increases. On the right panel, the connection probabilities of the SBM are chosen uniformly at random for each graph. ASE on the average adjacency matrix performs poorly, while MASE still improves with a larger sample size.