Evidence page · genetics and linguistics

A community of five million people descends from about a hundred.

The Vysya of Andhra Pradesh and Telangana carry one of the most extreme founder events recorded in any human population — more extreme than the Ashkenazi Jewish bottleneck. They are also a middle-ranking merchant caste, not an oppressed one, and that is the part worth pausing on. This page sets out when South Asian communities stopped exchanging marriage partners, what the genetics can and cannot say about why, and what the same period did to the Dravidian languages.

The Vysya founder event.

the numbers

For an estimated three thousand years the Vysya have experienced negligible gene flow from neighbouring groups. The consequence is medical: a roughly hundredfold elevated rate of butyrylcholinesterase deficiency, a metabolic disorder causing severe reactions to common anaesthetics — a recessive variant that drifted to high frequency because there was too little variation for it to be diluted. For comparison, the Ashkenazi Jewish founding population is estimated at around 350 individuals. The Vysya bottleneck is tighter.

But the Vysya are not an oppressed caste.

and that is the finding, not an aside

When the mixing stopped.

a window that opened and closed

The Vysya are early, but they are not exceptional in kind. The wider pattern is a demographic transformation with a beginning and an end:

Why it stopped: what the genetics does and does not say.

the limit of the method

The Indus question.

what the dates do and do not settle

Testing the migration hypothesis properly.

the timing objection is weak; the ancestry test is not

A reasonable reconstruction runs: the Indus cities decline around 1800 BCE; trading families move gradually east and south over many generations, mixing as they go; they reach the Andhra coast, establish themselves, and only then seal into endogamy around 1000 BCE. The timing objection to this is weak. Eight hundred years is more than thirty generations, and a founder event dates the end of mixing, not the origin of a population. Nothing in the chronology forbids a slow migration.

But the hypothesis makes a prediction that can be checked. If a community descends from northwestern migrants who arrived and then sealed, it should carry more northwestern ancestry than its neighbours — a higher West Eurasian or Iranian-related fraction, preserved by the very isolation that produced the bottleneck.

What the regional comparison would have to show

None of these has been reported. What is reported at the community level is descriptive rather than quantified — Reddy and Kamma cluster together with a broadly similar ancestry balance; Kapu shows more internal regional variation than tightly endogamous communities like Kamma or Velama; coastal, Rayalaseema and Telangana profiles differ, with inland and highland populations sometimes carrying higher southern-component proportions. That is a landscape with real structure in it, and nobody has published the numbers.

So the migration hypothesis is currently unfalsified rather than supported. It makes a clear prediction, the data to test it is largely collected, and the test has not been carried out. That is a different situation from a hypothesis that has been checked and failed, and this page does not conflate the two.

Other Telugu groups that sealed.

the Vysya are not alone in the region

The founder-event survey names several communities from the Telugu-speaking region, which bears directly on whether this was one community's history or a regional process:

Five communities from one linguistic region, spanning merchant, landholding, warrior and labouring positions, all showing strong founder events. This is not the isolation of one unusual group. It is a regional pattern in which many communities sealed, and the seals did not correlate with rank.

Who imposed the languages.

a claim worth stating precisely

Languages spread with power, not by consensus, and the population of any modern linguistic state was never uniformly composed of that language's speakers. The administrative record supports the general shape of this. The Telugu term Reddi — earlier Raddi, Rattoḍi, Rattakuḍi, connected to Rāṣṭrakūṭa — denoted village headmen responsible for organising cultivation and collecting taxes, and from the seventh century members of these families held important administrative posts in kingdoms. The Kapu were landholders who also served as military commanders under Vijayanagara.

The Dravidian family and its branches.

Kolipakam et al. 2018

A Bayesian phylogenetic analysis of first-hand lexical data from native speakers dates the Dravidian family to approximately 4,500 years old — about 80 varieties, 220 million speakers. And its internal structure is not what the geography suggests.

The authors' own caution matters here. They report "considerable uncertainty with regard to the relationships between the main branches," and a differently constrained analysis returns a different higher-order topology. The four branches are secure; the order in which they separated is not.

How firm is 4,500 years?

the interval the headline dropped

What that interval contains is the whole of the question:

All three sit inside the published interval. The evidence as it stands does not choose between "Dravidian split during the Indus period" and "Dravidian split because of its collapse." Anyone asserting either with confidence is reading the midpoint as though it were the finding.

The estimate that disagrees, and by a factor of three

The 4,500-year figure is not the only published estimate, and the paper says so itself: “Contrary to other work by Pagel et al., our findings suggest this younger age rather than their Proto-Dravidian estimate of around 13 000 years ago. Future work should investigate the disparity in clock rates here.” Two further estimates sit between them — Fuller places the differentiation of the Dravidian languages around 6,000 years ago on reconstructed crop vocabulary, and Southworth puts the diversification of the North, Central and South branches at 4,500–4,000 years ago, tied to the Southern Neolithic. A field whose published estimates run from about 4,000 to about 13,000 years has not settled this, and a page that reports only the youngest of them is not reporting the state of the question.

And the calibration has a known structural bias

The dates are anchored on first written attestations — Tamil's earliest lithic inscription of about 254 BCE is used to constrain the South I group to be at least 2,250 years old. Methodologists describe this as sound in principle: a protolanguage is necessarily older than its attested descendant. But it has a consequence.

The branching order is also unstable — a correction

So the honest position on Telugu is narrower than the striking version. Telugu is certainly not within South I — it does not belong with Tamil, Malayalam and Kannada, and the intuition that it is a different order of separation is correct. Whether it is the family's deepest division or its shallowest major one is not resolved by this data.

What the eastward hypothesis would need

A scenario in which an early-branching group moved east and later south — through the Gangetic plain and down into the peninsula — is not excluded, and the modern distribution of Kuṛux and Malto in Jharkhand and Bengal is exactly the kind of residue such a movement would leave. But the linguistic dating cannot test it, because a root age does not locate anything in space.

What would test it is the same thing missing everywhere else on this question: ancient DNA and dated archaeology from the eastern and central corridors, at the resolution that exists for the northwest. That survey has not been done, and until it is, the eastward route and the peninsular route remain equally unfalsified.

A test of that tree, from the dictionary.

and it disagrees, for an instructive reason

The claim can be checked independently. Counting shared etymological root numbers across 26 Dravidian languages and computing overlap gives this:

At branch level it works: every branch is most similar to itself, and similarity falls off with distance. But at language level it breaks.

Iranian-related ancestry.

what is established, and the part that is not

The component usually called "Iranian farmer-related" is the largest single element in the ancestry of the Indus world and of everyone descended from it:

• The Indus Periphery Cline individuals carry ~45–82% Iranian farmer-related and ~11–50% AASI ancestry, with negligible Anatolian farmer admixture.
• ASI — the southern ancestral component — is a minimum of ~55% Indus Periphery Cline ancestry. ANI is ~70%.
• Every mainland Indian population carries it, north and south. It predates the steppe arrival by millennia and is the shared inheritance underneath the north–south gradient.
• The lineage is deep: it shares a common ancestor with ancient Iranians more than 12,000 years ago. It is not evidence of recent migration from Iran.

The honest limits.

• Sources: the Vysya founder size, date and BChE rate, and the 81-of-263 founder-event survey, from the published founder-event literature — snippet level; the survey is not named on this page, and is recorded as an unidentified source. The platform’s own endogamy-clock page reports the Vysya bottleneck with error bars (144 ± 27 generations, roughly ± 780 years); the point figures used here should be read with that spread; the 4,200–1,900 BP mixture window from Moorjani et al. 2013 (AJHG); Dravidian family age and branch structure from Kolipakam et al. 2018 (Royal Society Open Science 5:171504, doi:10.1098/rsos.171504) — abstract, indexed full-text excerpts and citation statements read here; the topology result and the authors’ statement of uncertainty are verified, the 95% interval is not and has been withdrawn; Indus Periphery proportions from Narasimhan et al. 2019 (Science) — snippet level. The overlap matrix is computed here from shared etymological root numbers; dedr_roots.json ships with the site.
• The 4,500-year Dravidian date is a point estimate — mean 4,650, median 4,433 — inside a 95% interval of roughly 3,000–6,500 years, spanning the pre-Indus, Mature Indus and post-Indus periods alike, and the authors state they cannot exclude a significantly older root. Published estimates elsewhere run from about 4,000 years (Southworth) to about 13,000 (Pagel et al.). Calibration is anchored on first written attestation, which sets a floor rather than measuring language age. The interval is corroborated at secondary level; the paper has not been opened in full.
• The branch order is unresolved. The best-supported model puts Telugu opposite Tamil at the primary split; a differently constrained model recovers the traditional topology. The four branches are secure; their sequence is not.
• Nothing here connects the Vysya to Indus-era traders. The chronology does not exclude it; the test that would settle it needs a within-Andhra ancestry comparison that has not been published.
• Ancestry components are artefacts of the model and reference panels chosen, not natural categories, and are used only for relative comparison between populations analysed together.
• Community-level Iranian-ancestry figures for South India are not published at that resolution and none are given.