Abstract
<jats:p>Local ancestry inference assigns each position along an admixed chromosome to a source population, underpinning admixture mapping, ancestry-specific association testing and admixture dating. Validation is almost exclusively on continentally divergent sources (Hudson's FST ≈ 0.1) and coalescent simulations; we examine both restrictions. Across FST from 0.0022 to 0.243 we compare five methods—two likelihood baselines, RFMix, FLARE and a dilated convolutional network—on identical sites with exact ground truth, and on 11 real 1000 Genomes pairs. Three findings follow. First, a feasibility floor: at FST = 0.0022 no method exceeds 0.575, and CHB/CHS at FST = 0.00042 yields at best 0.551. Pairs motivating fine-scale analysis, such as northern versus southern Han, fall below it. Second, per-site accuracy conceals a failure of tract structure: the most accurate method per site produces 78.8× too many tracts, implying an admixture time 61.2× too old, which Viterbi decoding removes at no cost to accuracy (+0.0002). Third, the simulated lead does not survive real data, and the deficit is one of input more than of the architectures we varied: attention, state-space layers, capacity, objective and self-supervised pretraining each move accuracy by at most 0.006, while supplying the haplotype information the released tools receive recovers +0.031 on 8 of 8 pairs below FST = 0.04 and nothing above it—necessary but not sufficient, since the network still trails on 10 of 11 pairs. Two quantities usually held fixed matter more than architecture: the statistic summarising reference matching, and reference panel size, which no method is near saturating.</jats:p>