Phase transition on the convergence rate of parameter estimation under an Ornstein-Uhlenbeck diffusion on a tree

Cécile Ané, Lam Si Tung Ho, Sebastien RocharXiv 2014 · arXiv

Abstract

In evolutionary biology, the Ornstein-Uhlenbeck (OU) process is a standard model for trait evolution, capturing both random drift and stabilizing selection. However, the shared ancestry of species introduces complex correlations that challenge standard statistical assumptions. This paper investigates the fundamental limits of parameter estimation in this context, identifying a critical "phase transition" where the tree's branching rate relative to selection strength determines whether the estimation of the ancestral optimum is statistically efficient. While the estimation of the mean trait value is highly sensitive to tree topology, we demonstrate that covariance parameters can be estimated consistently using a novel contrast-based approach.

Introduction

A central goal in phylogenetic comparative methods is to reconstruct the evolutionary history of continuous traits, such as body size or metabolic rate. The Ornstein-Uhlenbeck (OU) process is frequently employed for this purpose, modeled as a stochastic differential equation where a trait is pulled toward a selection optimum ($\mu$) with a strength ($\alpha$). Unlike independent data points, species on a phylogenetic tree share evolutionary history, meaning that adding more species does not necessarily provide proportional increases in information. This study explores how the growth dynamics of the tree itself dictate the convergence rates of our statistical estimators.

The Problem: Phylogenetic Signal and Information Limits

Standard statistical theory suggests that as the number of samples ($n$) increases, the error in estimating a mean should decrease at a rate of $1/\sqrt{n}$. In phylogenetics, this is complicated by the "phylogenetic signal"—the correlation between species due to shared ancestry. Previous research established that if a tree's total height is bounded, the mean trait value cannot be estimated perfectly even with infinite species. This paper addresses the more realistic scenario where the tree continues to grow in height as new species are added, asking whether the increase in time-depth is sufficient to overcome ancestral correlations.

Methodology

The authors utilize three primary mathematical frameworks to analyze estimation efficiency:

  1. Maximum Likelihood Estimation (MLE): The study rigorously analyzes the standard MLE for the selection optimum $\mu$, determining the conditions under which it remains consistent.
  2. Electrical Network Analogy: By mapping the phylogenetic tree to an electrical circuit, the authors equate the variance of the trait estimator to the "effective resistance" of the tree. This allows the application of probability theory tools from the "reconstruction problem" to establish bounds on estimation error.
  3. Contrast-Based Estimation: To estimate the selection strength ($\alpha$) and stationary variance ($\gamma$), the authors propose a new estimator based on "contrasts." By selecting pairs of species with non-overlapping evolutionary paths, they generate independent observations that bypass the correlation issues inherent in the full tree structure.

The Phase Transition in Trait Estimation

The core finding of the paper is the identification of a phase transition at the threshold $\Lambda = 2\alpha$, where $\Lambda$ represents the tree's growth rate and $\alpha$ represents the selection strength.

Subcritical Regime ($\Lambda < 2\alpha$)

When selection is strong relative to the branching rate, the correlations between species decay quickly enough that the tips provide relatively independent information. In this regime, the standard $\sqrt{n}$ convergence rate for the mean trait value is maintained.

Supercritical Regime ($\Lambda > 2\alpha$)

If the tree branches too rapidly, the species are "trapped" by their shared history. The tree grows faster than selection can pull the traits toward the optimum, causing the "effective sample size" to scale with the tree's height rather than the number of species. In this regime, the variance of the MLE for $\mu$ scales as $\Omega(n^{-2\alpha/\Lambda})$, which is significantly slower than the standard $1/n$ rate.

Key Results and Practical Implications

The study demonstrates a fundamental split in the difficulty of parameter estimation. While estimating the center of the process (the optimum $\mu$) is highly sensitive to the tree's branching structure, estimating the dynamics of the process (the parameters $\alpha$ and $\gamma$) is much more robust. The authors prove that their novel contrast-based estimator achieves $\sqrt{n}$-consistency regardless of the growth rate, provided the tree contains sufficient internal nodes at various age bands.

For practitioners, these results suggest a "design rule": if a lineage is in the supercritical regime (where speciation rate $\lambda > 2\alpha$), adding more contemporary species to the study provides diminishing returns. Instead, researchers should prioritize increasing the time-depth of the tree by including deeper, older lineages or fossil data to improve the accuracy of ancestral trait reconstruction.

Limitations and Future Directions

While providing a rigorous theoretical foundation, the study assumes the phylogenetic tree (topology and branch lengths) is known without error. In empirical studies, tree estimation itself is a significant source of uncertainty. Furthermore, the model assumes constant parameters across the tree. Future research could extend these findings to "mosaic" models, where different clades experience different selection regimes, or to non-ultrametric trees that explicitly include extinct lineages.

Summary updated

Abstract

Diffusion processes on trees are commonly used in evolutionary biology to model the joint distribution of continuous traits, such as body mass, across species. Estimating the parameters of such processes from tip values presents challenges because of the intrinsic correlation between the observations produced by the shared evolutionary history, thus violating the standard independence assumption of large-sample theory. For instance Ho and Ané \cite{HoAne13} recently proved that the mean (also known in this context as selection optimum) of an Ornstein-Uhlenbeck process on a tree cannot be estimated consistently from an increasing number of tip observations if the tree height is bounded. Here, using a fruitful connection to the so-called reconstruction problem in probability theory, we study the convergence rate of parameter estimation in the unbounded height case. For the mean of the process, we provide a necessary and sufficient condition for the consistency of the maximum likelihood estimator (MLE) and establish a phase transition on its convergence rate in terms of the growth of the tree. In particular we show that a loss of $\sqrt{n}$-consistency (i.e., the variance of the MLE becomes $Ω(n^{-1})$, where $n$ is the number of tips) occurs when the tree growth is larger than a threshold related to the phase transition of the reconstruction problem. For the covariance parameters, we give a novel, efficient estimation method which achieves $\sqrt{n}$-consistency under natural assumptions on the tree.
BibTeX
@misc{ane2014phase,
  author = {Cécile Ané and Lam Si Tung Ho and Sebastien Roch},
  title = {{Phase transition on the convergence rate of parameter estimation under an Ornstein-Uhlenbeck diffusion on a tree}},
  year = {2014},
  eprint = {1406.1568},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1406.1568},
}

Loading related content...

Figure