📄 evolutionary biology

Unlocking a flexible set of phylogenetic models for discrete and continuous trait evolution using discretized stochastic diffusion

This paper extends the "discretized diffusion approximation" method, originally developed for Brownian motion with reflective limits, to unlock a flexible suite of new phylogenetic models for analyzing both discrete and continuous trait evolution, including threshold, semi-threshold, and joint models where evolutionary rates depend on co-evolving traits.

Original authors: Revell, L. J., Alencar, L. R. V., Alfaro, M. E., Dain, J., Hill, N. J., Jones, M., Martinet, K. M., Romero-Alarcon, V., Harmon, L. J.

Published 2026-08-17

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Revell, L. J., Alencar, L. R. V., Alfaro, M. E., Dain, J., Hill, N. J., Jones, M., Martinet, K. M., Romero-Alarcon, V., Harmon, L. J.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve the mystery of how life changes over millions of years. You have a family tree of animals, like a giant, branching diagram showing who is related to whom. Your job is to figure out how specific traits—like the size of a beak, the length of a tail, or the color of a feather—evolved along the branches of that tree. For a long time, scientists have used a mathematical tool called "Brownian motion" to model this. Think of Brownian motion like a drunk person stumbling randomly through a foggy park. They don't have a destination; they just take steps in random directions. Over time, the distance they wander from the starting point grows, but you can't predict exactly where they will end up, only the probability of where they might be. This "random walk" has been the standard way to guess how traits evolve.

However, real life isn't always a random walk in an open park. Sometimes, there are walls. A lizard can't grow a tail longer than its body, or a bird's beak can't get so big it breaks its skull. These are "boundaries" or "constraints." For a long time, trying to do the math for a random walk that hits a wall and bounces back was incredibly difficult, almost impossible, for large family trees. It was like trying to calculate the exact path of that drunk person if the park had invisible, bouncy walls everywhere. Because the math was so hard, scientists often just ignored the walls, which led to wrong guesses about how fast traits were changing. This paper is about a clever new way to solve that math puzzle, opening the door to understanding evolution in much more complex and realistic ways.

The Paper: Unlocking Evolutionary Models with a Digital Ladder

This paper, written by a team of evolutionary biologists, is essentially a masterclass in how to use a clever mathematical trick to solve problems that were previously too hard to crack. The authors are building on a method introduced by Boucher and Démery in 2016, which they describe as a "discretized diffusion approximation." To understand what that means, imagine you are trying to measure the smooth, flowing movement of a river. It's continuous; the water flows without stopping. But if you wanted to simulate that river on a computer that only understands whole numbers, you might chop the river into tiny, distinct steps or "bins." You'd say, "The water is here, then it moves to the next bin, then the next."

The authors show that if you make these steps small enough—so small they are almost invisible—the "stepped" simulation becomes indistinguishable from the real, smooth river. They call this the "discretized diffusion approximation." It's like using a digital ladder to climb a smooth slide; if the rungs are close enough together, you can climb the slide just as smoothly as if you were sliding down it.

The big news here is that this trick works for all kinds of evolutionary scenarios that previously had no working math formulas. The authors don't just explain the trick; they use it to unlock five new ways to model how traits evolve.

First, they revisit Bounded Brownian Motion. This is the "drunk person with walls" scenario. In the past, if a trait hit a limit (like a maximum size), the math broke down. Using their ladder trick, the authors can now simulate a trait bouncing off walls. They ran simulations showing that if you ignore these walls, you will systematically underestimate how fast a trait is evolving. It's like thinking a runner is slow because they keep hitting a wall and bouncing back, when in reality, they are sprinting but just can't get past the barrier. Their new method fixes this, giving accurate estimates of evolutionary speed even when traits are constrained.

Next, they tackle the Threshold Model. Imagine a trait that is actually continuous underneath but looks discrete on the surface. Think of a lizard's "liability" to have a certain number of toes. Maybe the underlying biology is a smooth scale of development, but once it crosses a certain "threshold," the lizard is born with either 4 toes or 5 toes. The paper shows how to use their ladder method to figure out where those invisible thresholds are, even when we only see the final toe count. They simulated data and found that their method could correctly identify these hidden thresholds and distinguish them from other random evolutionary patterns.

Then, they introduce a brand new idea called the Semi-Threshold Model. This is a mix of the two previous ideas. Imagine a trait like "horn length." It can grow and shrink smoothly, but if it shrinks to zero, the horn is gone. You can't measure a negative horn length; it just becomes "absent." However, the potential to grow a horn (the liability) might still be evolving underneath, even if the horn is gone. The authors created a model for this "stuck" state. They simulated data where traits were either measurable or "censored" (stuck at zero or a maximum), and their method successfully estimated the evolutionary rates, whereas standard methods failed.

The paper also explores Joint Evolution, where two different types of traits influence each other.

Discrete-Dependent Continuous: Imagine a fish that lives in either open water or a reef. The paper suggests that the rate at which the fish's body shape changes might depend on which habitat it's in. In open water, maybe body shape changes slowly; in a reef, it changes fast. The authors built a model where the "ladder" for the continuous trait (body shape) has different rungs depending on the discrete trait (habitat). Their simulations showed this method could accurately figure out how fast the fish were evolving in each habitat, doing a better job than older methods that tried to guess the habitat history first and then measure the shape.
Continuous-Dependent Discrete: This is the reverse. Imagine a reptile that lays eggs or gives live birth. The paper suggests that the rate at which a species switches from eggs to live birth might depend on the temperature (a continuous trait). If it's hot, maybe the switch happens faster. They modeled this using a "sigmoidal" curve (an S-shape) to show how the temperature changes the odds of switching. While they didn't run huge simulations for this one due to computer limits, they showed that the math works and can fit the data.

Finally, they present a Multi-Trend Model. This is for situations where traits don't just wander randomly; they have a direction. Imagine a virus spreading across a map. It might drift randomly, but if the host animal changes (a discrete trait), the virus might suddenly start drifting north or south (a trend). The authors showed how to model this "drift with a direction" that changes based on the host. They simulated this on non-uniform trees and found their method could accurately estimate both the speed of the drift and the direction of the trend.

What the Paper Does and Doesn't Do

The authors are very clear about what they have achieved. They have provided proof of concept through simulations that this "ladder" method works for these specific models. They showed that when they simulated data with known rules, their method could find those rules back with reasonable accuracy. For example, they demonstrated that ignoring boundaries leads to wrong answers, and their method fixes that. They also showed that their joint models are more accurate than the old "two-step" method of guessing one trait first and then the other. However, they explicitly note that these analyses are preliminary and intended to demonstrate the feasibility of the approach rather than serve as a comprehensive investigation of statistical properties.

They also explicitly rule out the idea that this is just a way to turn continuous data into simple categories. They emphasize that they are not saying "let's just call big lizards 'large' and small ones 'small' and analyze them that way." That would be a mistake. Their method uses the "ladder" only as a temporary mathematical tool to calculate the probability of the smooth, continuous reality. The final answer is always about the continuous trait, not the bins.

They also suggest that this approach is just the tip of the iceberg. They mention that this method could likely be used for "hidden-rate" models (where we don't see the trait causing the change) or even for traits that jump suddenly (discontinuous jumps), but they haven't fully tested those yet. They admit that for some of the more complex models, like the continuous-dependent discrete trait, they haven't run massive simulation studies yet because the math is heavy on the computer. They recommend that future researchers test those specific limits.

In short, this paper doesn't claim to have solved every mystery of evolution. Instead, it hands the scientific community a powerful, flexible new tool—a way to approximate complex, messy evolutionary processes that were previously too hard to calculate. It suggests that by using this "discretized diffusion" trick, we can finally build models that respect the walls, the thresholds, and the hidden connections of the natural world, leading to a clearer picture of how life has changed over time.

Technical Summary: Unlocking a Flexible Set of Phylogenetic Models for Discrete and Continuous Trait Evolution Using Discretized Stochastic Diffusion

Problem Statement
The practical utility of phylogenetic comparative methods often depends on the ability to fit mathematical models that accurately capture evolutionary processes. However, many biologically realistic models for trait evolution lack tractable analytic solutions for their likelihood functions, particularly when dealing with bounded continuous traits, threshold models, or complex joint evolution of discrete and continuous characters. For instance, the "bounded Brownian motion" model (Boucher and Démery 2016), which describes evolution on a flat adaptive landscape with hard constraints, possesses an analytic probability function that is intractable for reasonably-sized phylogenies. Similarly, standard threshold models require the integration of high-dimensional multivariate normal distributions, a task that becomes computationally impractical as the number of taxa increases. Consequently, researchers are often forced to rely on simplified models or two-step estimation procedures that may introduce bias or fail to capture the full complexity of the evolutionary process.

Methodology: The Discretized Diffusion Approximation
The core methodological contribution of this paper is the application and expansion of the "discretized diffusion approximation," a technique originally identified by Boucher and Démery (2016) and extended by Boucher et al. (2018). This approach leverages the mathematical convergence of a continuous-time, symmetric, ordered Markov chain (a random walk on a lattice) to continuous stochastic diffusion (Brownian motion) as the number of lattice states approaches infinity (or the bin width $\delta$ approaches zero).

The procedure involves:

Discretization: The continuous trait space is divided into $n$ bins of width $\delta$ .
Markov Chain Construction: A continuous-time Markov chain is defined on this grid. Transition rates between adjacent bins are set to $q_c = \sigma^2 / (2\delta^2)$ , where $\sigma^2$ is the instantaneous diffusion rate. For models with trends ( $\mu$ ), the rates become asymmetric ( $q_+$ and $q_-$ ).
Likelihood Calculation: The probability of the discretized data ( $x'$ ) is computed efficiently using Felsenstein's pruning algorithm (1981), which is standard for discrete character evolution.
Density Approximation: The probability density of the original continuous data, $f(x)$ , is approximated by $\hat{f}_\delta(x) = p(x') / (\delta^N)$ , where $N$ is the number of taxa. As $\delta \to 0$ , this approximation converges to the true probability density.

This method allows for the computation of likelihoods for models where analytic solutions are unavailable, provided a corresponding Markov chain representation can be constructed.

Key Contributions and New Models
The paper demonstrates the versatility of this estimation technique by applying it to five specific models, four of which are presented as novel or significantly extended applications to phylogenetic comparative biology:

Bounded Brownian Motion (Reflective Limits): The paper re-establishes the utility of the discretized diffusion approximation for Brownian motion with reflective boundaries. It confirms that this method avoids the systematic underestimation of the diffusion rate ( $\sigma^2$ ) that occurs when fitting unbounded models to bounded data.
The Threshold Model: The authors apply the approximation to the classic threshold model from evolutionary quantitative genetics, where an unobserved continuous "liability" evolves via Brownian motion and crosses fixed thresholds to produce discrete states. The method replaces the need for high-dimensional numerical integration or Monte Carlo sampling, allowing for efficient model fitting and parameter estimation.
The "Semi-Threshold" Model: A new model is introduced where a trait is observable as a continuous value within a specific range but is censored (recorded as a fixed value) outside that range, while the underlying liability continues to evolve. The paper shows how to adjust the likelihood calculation (dividing by $\delta^m$ rather than $\delta^N$ , where $m$ is the number of uncensored tips) to handle this mixed data type.
Discrete-Character-Dependent Continuous Trait Evolution: The paper presents a joint model where the rate of Brownian motion for a continuous trait ( $\sigma^2$ ) depends on the state of a co-evolving discrete character. Unlike traditional two-step approaches (stochastic mapping followed by rate fitting), this method calculates the joint likelihood directly, resulting in less biased estimates of rate heterogeneity, particularly when discrete transition rates are high.
Continuous-Character-Dependent Discrete Trait Evolution: A novel inverse model is proposed where the transition rates of a discrete trait depend on the value of a continuous trait (modeled via a sigmoidal function). While extensive simulation was not feasible due to computational constraints (large matrix exponentiation), the paper demonstrates the construction of the infinitesimal generator matrix and provides a maximum likelihood fit to simulated data.
Multi-Trend Directional Model: A discrete-character-dependent model is developed where both the diffusion rate ( $\sigma^2$ ) and the linear trend ( $\mu$ ) of a continuous trait vary as a function of a discrete character state. This model is shown to be identifiable on non-ultrametric trees and allows for the estimation of directional trends that switch based on discrete regimes.

Results
Through simulation studies, the authors demonstrate that the discretized diffusion approximation yields reliable parameter estimates:

Bounded Models: Fitting the bounded model correctly recovers $\sigma^2$ and root states ( $x_0$ ), whereas unbounded models systematically underestimate $\sigma^2$ when data are generated under bounds.
Threshold Models: The method successfully discriminates between threshold models and ordered Mk models, with estimated thresholds centering on generating values.
Joint Models: In the discrete-character-dependent continuous trait model, joint estimation via the discretized diffusion approximation provides unbiased estimates of state-specific rates ( $\sigma^2_a, \sigma^2_b$ ) and transition rates ( $q_{ab}$ ). Crucially, it outperforms the traditional two-step stochastic mapping approach, which tends to underestimate the ratio of rates between regimes, especially when discrete transitions are frequent.
Semi-Threshold and Directional Models: Preliminary analyses show reasonable estimation properties for the new semi-threshold and multi-trend models, though the authors note these require further comprehensive study.

Significance and Claims
The authors position this work as a "concept piece" intended to unlock a flexible set of phylogenetic models that were previously inaccessible due to intractable likelihood functions. They argue that the discretized diffusion approximation is a powerful, general-purpose tool that transforms problems previously deemed unsolvable into ones that can be addressed via standard numerical optimization.

The paper claims that the Boucher and Démery (2016) approach, while mathematically rooted in stochastic process theory (e.g., Ethier and Kurtz 1986; Risken 1996), was novel in phylogenetics and has been underappreciated. By expanding its application to threshold, semi-threshold, and complex joint evolution models, the authors assert that this method "barely scratches the surface" of potential applications. They suggest that the approach could be extended to hidden-rate models, Ornstein-Uhlenbeck processes with varying parameters, and models with absorbing boundaries or discontinuous jumps.

The authors conclude with a metaphorical observation: while the "law of the instrument" warns that having a single tool (a hammer) makes every problem look like a nail, their exploration suggests an "inverse law" for phylogenetics—where many complex evolutionary hypotheses (nails) can indeed be effectively addressed by this specific approximation (the hammer). The methods are implemented in the R package phytools, making these advanced models accessible to the broader community.

Technical Summary: Unlocking a Flexible Set of Phylogenetic Models for Discrete and Continuous Trait Evolution Using Discretized Stochastic Diffusion

More like this