Figure 1.

A schematic diagram of TeraStructure, stochastic variational inference for the Pritchard-Stephens-Donnelly (PSD) model. The algorithm maintains an estimate of the latent population proportions for each individual. At each iteration it samples SNP measurements from the large database, infers the per-population frequencies for that SNP, and updates its idea of the population proportions. This is much more efficient than algorithms that must iterate across all SNPs at each iteration.