I have put up a few posts warning readers to be careful of confusing PCA plots with real genetic variation. PCA plots are just ways to capture variation in large data sets and extract out the independent dimensions. Its great at detecting population substructure because the largest components of variation often track between population differences, which consist of sets of correlated allele frequencies. Remeber that PCA plots usually are constructed from the two largest dimensions of variation, so they will be drawn from just these correlated allele frequency differences between populations which emerge from historical separation and evolutionary events. Observe that African Americans are distributed along an axis between Europeans and West Africans. Since we know that these are the two parental populations this makes total sense; the between population differences (e.g., SLC24A5 and Duffy) are the raw material from which independent dimensions can pop out. But on a finer scale one has to be cautious because the distribution of elements on the plot as a function of principal components is sensitive to the variation you input to generate the dimensions in the first place.
I can give you a concrete example: me. I showed you my 23andMe ancestry painting yesterday. I didn’t show you my position on the HGDP data set because I’ve shared genes with others and I don’t want to take the step of displaying other peoples’ genetic data, even if at a remove. But, I have reedited some “demo” screenshots and placed where I am on the plot to illustrate what I’m talking about above. The first shot is my position on the two-dimensional plot of first and second principal components of genetic variation from the HGDP data set.