The distribution of admixture tract lengths has received considerable attention, in part because it can be used to infer the timing of past gene flow events between populations. It is commonly assumed that these lengths can be modeled as independently and identically distributed (iid) exponential random variables. This assumption is fundamental for many popular methods that analyze admixture using hidden Markov models. We compare the expected distribution of admixture tract lengths under a number of population-genetic models to the distribution predicted by the Wright-Fisher model with recombination. We show that under the latter model, the assumption of iid exponential tract lengths does not hold for recent or for ancient admixture events and that relying on this assumption can lead to false positives when inferring the number of admixture events. To further investigate the tract-length distribution, we develop a dyadic interval-based stochastic process for generating admixture tracts. This representation is useful for analyzing admixture tract-length distributions for populations with recent admixture, a scenario in which existing models perform poorly.T HERE has been interest in analyzing population genomic data by using methods that partition an admixed individual's genome into blocks originating from different ancestral populations. An early version of the popular program Structure (Falush et al. 2003) accomplished this with a hidden Markov model (HMM), indexed along the genome, with hidden states corresponding to the ancestral population each position was inherited from. The contiguous blocks of the genome inherited from a population are called "admixture/ migrant tracts/segments," depending on the context. For consistency, we use the term "admixture tract" in this article. Admixture tracts are unobservable, and their existence can be inferred only from genomic data. The process of doing so is called "admixture deconvolution" or "ancestry painting" and has been used in a number of different contexts, such as in admixture mapping for identifying human disease-associated genes (Hoggart et al. 2003;Reich et al. 2005), populationgenetic inferences aimed at understanding human ancestry (Bryc et al. 2010;Henn et al. 2012), or identifying regions affected by natural selection (Tang et al. 2007).The technique of using HMMs to partition an individual's genome into admixture tracts has been used in subsequent methods. Hoggart et al. (2003) and Smith et al. (2004) used HMMs for inferring admixture tracts with the purpose of admixture mapping and controlling for population stratification, similar to the method of Falush et al. (2003). More recent publications have focused on admixture deconvolution for more general population-genetic purposes, such as Tang et al. (2006) and Sundquist et al. (2008).In HapMix (Price et al. 2009), the HMM model of Li and Stephens (2003) for modeling linkage disequilibrium is extended to include admixture between two populations. HapMix uses a genotype-based state space and so do...