Fig. 1

a Cross-method comparison of single-threaded wall time for generating m/4 samples from the m-dimensional MVB distribution on a log-log scale. The proposed rb_dplr algorithm scales linearly in sample size and problem dimension. Both rb_dplr and the unstructured variant rb_unstr outperform existing methods including haplosim [3], three methods implemented by Kruppa et al. (kruppa-*) [1], and GenOrd [15] Solid lines reflect linear splines with fixed knots fitted to numerical experiment results and dashed lines reflect extrapolations. b Drawing genotypes directly from their equilibrium distribution under AM. Comparison of heritabilities in synthetic genotype/phenotype data generated using rb_dplr to sample from the appropriate MVB distribution, versus the forward-time approach of Border et al. [8]. Best-fit lines and standard-errors summarize variation across 100 replicates with 2000 haploid causal variants for 8000 individuals, for phenotypes with panmictic heritability cross-mate phenotypic correlation both fixed to 0.5