speedup over other methods across the sample size range (Fig. 2a and Supplementary Table 1), attaining a 14x speedup over SHAPEIT2 and a 12x speedup over HAPI-UR at N≈150,000. (Beagle was unable to phase 1% of the genome in 2 weeks at N≈150,000.) We note that (like other methods) Eagle has parameters that produce a trade-off in speed and accuracy (Online Methods); Eagle's ––fast mode achieved a further ≈2x speedup over the default while incurring only a slight loss of accuracy (Table 1). All methods exhibited superlinear but subquadratic scaling of running time with sample size, consistent with the presence of both linear and quadratic algorithmic components. (For a detailed discussion of the run time scaling of each of Eagle's algorithmic steps, see Online Methods and Supplementary Table 2.) We also observed that Eagle achieved modest (2–8x) savings in memory cost compared to other methods (Fig. 2b and Supplementary Table 1). All methods exhibited memory cost scaling roughly linearly with sample size.
