--- In Nostratica@yahoogroups.com, "H.M. Hubey" <hubeyh@...> wrote:

> Then at worst you'd have two roots for every word, and there would
> be no patterns in the sound changes from the protolanguage to the
> daughter languages. If the change is a Markov process you'd expect
> to find patterns. And therein lies the answer. Ultimately
> the answer is in probability theory.

I don't think so. For every occurrance of X in language A that
corresponds to a Y in language B you'd have a phoneme Z (which is
chosen to be similar to both X and Y) in Proto-AB. In this way you'd
have completely consistent diachronic changes. I suppose probability
theory combined with typology is the ultimate answer.