Eurogenes K36 is the fine-grained one: thirty-six components with names like Fennoscandian, North Sea, Iberian, Italian, North Caucasian, Armenian, Arabian. It is the calculator behind the colourful "K36 maps" people post, and it is simultaneously the most detailed and the most over-read tool in the GEDmatch family. Both things follow from the same design choice.
The general rules for any admixture calculator apply here with interest; this post is about what raising K to 36 buys and costs.
What K = 36 actually does#
A component calculator holds K allele-frequency profiles and explains your genome as a mixture of them. Raise K and the clustering carves the same reference data into thinner slices: with thirteen components northwest Europe is one profile, with thirty-six it becomes several neighbouring ones — North Sea, Fennoscandian, North Atlantic and their kin. The slices are genuinely there in the data, in the sense that the clustering found them; whether they are separable in one person's genome is a different question, and mostly the answer is no.
Neighbouring K36 components are heavily correlated. A genome with real ancestry from one North Sea-adjacent population will scatter weight across three or four adjacent components, in proportions that shift run to run and chip to chip. The author of the most careful independent guide to the Eurogenes project put it bluntly: the K36 results become more refined and less confident at once. That is not a flaw someone could patch — it is the trade K buys, the same reason the same person's K13 and K36 profiles do not contradict each other even though they look nothing alike.
The right way to read it#
Read K36 output as a texture, never as an itemised list:
- Cluster the components yourself. Sum the Scandinavian-adjacent ones, the Iberian-adjacent ones, the Caucasus-adjacent ones. The sums are far more stable than any single line, and they are the level at which the calculator is actually informative.
- Ignore the tail. Fifteen components at 0.3–2% are the optimiser distributing noise across thin profiles. In a 36-way split the noise floor rises to a meaningful fraction of the small entries, and no error bars exist to flag which are real — none ever do in this family.
- Treat the labels as peaks, not borders. "Italian" is where that profile is modal in the training references, not a statement that some ancestor was Italian. Third-party K36 maps interpolate your component scores over modern political geography — attractive, and two full steps removed from your genome: once by the clustering, once by the cartography.
When K36 is the wrong tool#
If the question is "which populations am I closest to, at fine grain" — the question K36 maps appear to answer — a plain distance ranking answers it directly, without forcing your genome through thirty-six correlated profiles. The free G25 distance tool ranks reference populations by straight Euclidean distance across six eras, and the PCA viewer shows the neighbourhood itself. For mixture percentages against dated, ancient references rather than modern regional constructs, the era-scoped Global25 admixture calculators are the modern equivalent of what K36 was reaching for in 2013.
And if a K36 line is about to become a claim — "the Armenian component proves a Caucasus ancestor" — that is a hypothesis for a method with a test attached. A qpAdm model can reject a proposed source; a 36-component optimiser never can.
K36 remains the most entertaining calculator in the family, and the entertainment is legitimate. The trouble only ever starts when the thirty-six numbers get promoted from texture to fact.
Terms used here are defined in the glossary.
