Figure 2: Loglikelihood structures of shift-scale and location-scale selfmodeling coefficients. The left pane (a) shows the structure for the (c,d) selfmodeling coefficients, compared to the right pane(b) which shows the (c,m) location-scale transformation. For a function with a single dominant peak, the location-scale transformation has a much lower correlation between c and m and thus the correspondingMCMC algorithm has superior mixing properties.