Fast and accurate out-of-core PCA framework for large scale biobank data

Performance on the East Asian data from 1000 Genomes Project. PCA performance of different software with various numbers of inferred PCs (K) based on 2,000,000 SNPs and 400 individuals from four East Asian populations. (A) Number of epochs used by each software. * Not out-of-core, only allows for in-core computation with this number of iterations. (B) Two estimated PCs K = 2 for each methods including the true full SVD. (C) Memory usage as a function of K (log scale at x-axis). (D) Accuracy (MEV) compared to the true full-rank SVD as a function of K (log scale at x-axis). (E) Convergence of PCAone and PCAone_{H + Y} shown as accuracy per epoch.