Fig. 5.

Linear shrinkage can improve performance under limited sampling. Spearman’s rank correlation coefficients between inferred and true selection coefficients, averaged over 200 simulations with the same setup as shown in Fig. 1a, are shown using different strengths of linear shrinkage, with (a) different sampling depths and (b) different sampling time intervals. Linear shrinkage of a proper strength improves inference results of our method. When the sampling time interval gets larger, the optimal strength increases as well, which is why we choose the $y$-axis to be $\gamma /\mathrm{\Delta }g$, the linear shrinkage strength divided by the sampling time interval. We find that a strength of 10 times time interval, $\gamma =10\mathrm{\Delta }g$, yields the highest Spearman’s $\rho$ consistently.