Abstract
The 'star paradox' in phylogenetics is the tendency for a particular resolved tree to be sometimes strongly supported even when the data is generated by an unresolved ('star') tree. There have been contrary claims as to whether this phenomenon persists when very long sequences are considered. This note settles one aspect of this debate by proving mathematically that there is always a chance that a resolved tree could be strongly supported, even as the length of the sequences becomes very large.
Introduction: The Star Paradox
In phylogenetics, researchers aim to reconstruct the "tree of life" using DNA sequences. A common challenge arises when evolution occurs rapidly, causing multiple lineages to appear to diverge simultaneously from a single point. This results in an unresolved "star tree," or polytomy, denoted as $T_0$.
Ideally, statistical models should reflect uncertainty when analyzing data generated by a star tree. However, Bayesian methods often exhibit a "star paradox," where they assign a posterior probability near 100% to one specific, incorrect resolved tree (such as $T_1, T_2,$ or $T_3$).
A long-standing debate in the field questioned whether this paradox was merely a small-sample artifact. Some argued that with infinite data, posterior probabilities would eventually equalize to 1/3 for each resolution. Steel and Matsen's work provides a formal mathematical proof that this optimistic view is incorrect: the paradox is a fundamental feature of the Bayesian framework in this context, not a sample-size error.
Methods
The authors utilize a rigorous mathematical proof focusing on the simplest case: three taxa evolving under a 2-state symmetric Markov process. Their approach includes:
- Reasonable Priors: Assuming equal prior probability (1/3) for the three resolved trees and continuous, non-zero probability density for branch lengths.
- Stochastic Analysis: Analyzing counts of site patterns (nucleotide configurations across species).
- The "Function of the Limit" Fallacy: Highlighting that the limit of a function of random variables (like posterior probability) does not necessarily equal the function of the limit of those variables.
- Lemma 2.2: Proving that the ratio of posterior probabilities between two trees can remain arbitrarily large with non-vanishing probability, even as sequence length ($n$) approaches infinity.
Key Insights
- Persistent Stochastic Error: As DNA sequences grow longer, random fluctuations (stochastic error) do not simply "average out" to favor the star tree. Instead, these fluctuations are sufficient to bias Bayesian calculations toward a specific resolved tree.
- Robustness of the Paradox: The proof applies to a broad class of "reasonable" priors, including standard exponential and gamma priors used in common phylogenetics software.
- Infinite vs. Long Finite Sequences: Previous simulations suggesting the paradox disappears often used "ideal" data (exact expected frequencies). In reality, data is stochastic. This paper proves that for any finite sequence—no matter how long—the probability of a "false" high-confidence result remains.
Results: Theorem 2.1
The central finding is Theorem 2.1: For any confidence level (e.g., 99%) and any resolved tree, the probability that the Bayesian model supports that tree with that level of confidence does not converge to zero as sequence length goes to infinity.
In practical terms, no amount of data can guarantee that a high posterior probability isn't an artifact of the star paradox. The rate of false inferences does not vanish even as $n \to \infty$.
Conclusion and Implications
While the proof is restricted to a simple 2-state model with three taxa, the underlying logic suggests the phenomenon extends to more complex scenarios. The paper does not suggest Bayesian phylogenetics is "broken" but highlights a mathematical reality that practitioners must consider. Future research may focus on developing methods to distinguish between true evolutionary signals and artifacts of the star paradox.
Summary updated