Derived allele frequency spectra in each population. The derived allele frequency spectrum (the proportion of SNPs of each possible derived allele frequency) is shown for each of the HapMap populations, after discovery of SNPs in two reads of the same ancestry. The YRI spectrum is based on SNPs ascertained in both the Cor17109 and the Cor17119 libraries; the CEU spectrum is based on the Cor7340 and the HuAA libraries; and the CHB+JPT spectrum is based on the Cor11321 and the HuFF libraries. SNPs ascertained in individuals of the same ancestry are pooled together (Supplementary Fig. 1 online), as are allele frequency data from the two East Asian populations, CHB and JPT (Supplementary Fig. 2 online). Also shown is the expected derived allele frequency spectrum for a population of constant size throughout history and the same ascertainment scheme. Although all spectra are biased by discovery in two chromosomes, they are comparable because the bias is identical for all spectra (we account for this bias in our analyses; see Methods).

Discovery of SNPs by comparing two sequencing reads from an individual of known ancestry. SNPs useable for analysis are identified as sites where one read matches the reference sequence and the other does not (an arrowhead of either type indicates a mismatch compared with the reference sequence; see Methods). The leftmost SNP is not ascertained because there is only one read, and the rightmost SNP is not ascertained because both reads share the same allele.

Modeling provides an excellent fit to the observed allele frequency spectra. (a,b) CEU (a) and CHB+JPT (b) data compared with that predicted by the models. Model 1 allows for one bottleneck in the history of each population, and model 2 allows for two bottlenecks (Supplementary Methods). For presentation (not actual analysis), the spectra are divided into 15 bins. Model 2 provides a better fit to the data, with mean squared error (averaged over bins) reduced to 62% of the model 1 value in CEU and 47% of the model 1 value in CHB+JPT.