🌀 nonlinear sciences

Redefining Fitness: Inference, Information and Phase Transitions in Evolutionary Dynamics

This paper resolves fundamental issues in evolutionary theory by redefining fitness as a Bayesian likelihood, demonstrating that natural selection acts to maximize the mutual information between population structure and environmental statistics, thereby establishing information maximization as the governing principle of evolution.

Original authors: Luís MA Bettencourt, Brandon J Grandison, Jordan T Kemp

Published 2026-07-16✓ Author reviewed ⓘ

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Luís MA Bettencourt, Brandon J Grandison, Jordan T Kemp

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Game of Guessing

Imagine you are playing a high-stakes video game where the rules change every time you press a button. Sometimes the ground is lava, sometimes it's ice, and sometimes it's made of jelly. To win, you don't just need to be strong; you need to be a master guesser. You have to figure out, "If I see a red cloud, is the ground about to turn to lava?" If you guess right, you survive and your character grows stronger. If you guess wrong, you lose a life. This is the essence of evolution, but instead of video game characters, we are talking about living things like bacteria, beetles, or birds.

For over a hundred years, scientists have tried to measure how good a living thing is at this game using a concept called fitness. In the old school of thought, fitness was a bit like a scoreboard that only looked at the final score: "How many babies did you have?" If you had lots of babies, you were "fit." If you had none, you weren't. But this scoreboard had a weird problem. It was circular, like saying "You are a good runner because you win races, and you win races because you are a good runner." It also had a glitch: imagine a butterfly that is perfectly camouflaged but gets eaten by a bird just by bad luck before it can lay eggs. The old scoreboard would say it had zero fitness, even though it was actually a great survivor. It couldn't tell the difference between being good at the game and just getting lucky.

This is where a new idea comes in, one that treats evolution less like a scoreboard and more like a detective solving a mystery. Scientists are now asking: What if being "fit" isn't just about having babies, but about having a really good mental model of the world? What if evolution is actually a giant process of learning and guessing? This is the question a team of researchers set out to answer, using math that sounds like it belongs in a spy novel but actually explains how life adapts to a changing universe.

The Detective's Guide to Evolution

In this new story, the authors propose a radical shift: Fitness is not a score; it's a prediction.

Imagine every type of animal or plant carries a tiny, invisible "survival manual" in its DNA. This manual isn't a list of instructions like "grow wings" or "make poison." Instead, it's a set of probabilities. It says things like, "If the environment is dry, there is an 80% chance I will survive," or "If the environment is wet, there is only a 10% chance." The authors call this a Bayesian likelihood. In plain English, it's just a fancy way of saying: "How well does my specific body type predict what the world is going to do next?"

The paper argues that nature is essentially a massive, slow-motion game of Bayesian inference. This is a type of math used by detectives and AI to update their beliefs based on new evidence.

The Old Way: "I survived, so I must be fit." (This is circular and ignores luck).
The New Way: "I have a model that predicts the environment well. Because my model is accurate, I am more likely to survive and reproduce."

The authors show that when you look at evolution through this lens, three big problems disappear:

The Circularity Problem: You don't need to know who had the most babies to know who is fit. You just look at how well their "manual" predicts the environment.
The Mismatch Problem: The lucky butterfly that got eaten didn't have a bad model; it just had bad luck. The model (the fitness) is about the probability of survival, not the single outcome.
The Prediction Problem: Because fitness is now a model, we can actually predict what will happen in the future. If the environment changes, we can see which "manuals" will become outdated and which will become the winners.

The Information Superhighway

Here is the most mind-bending part of the discovery. The authors found that when a population evolves over a long time, it isn't just trying to have more babies. It is trying to maximize information.

Think of the environment as a noisy radio station broadcasting a secret code. The different types of animals in a population are like different radios trying to tune into that station.

If a radio is static-filled (a bad model), it can't hear the code.
If a radio is clear (a good model), it hears the code perfectly.

The paper proves that natural selection acts like a filter that keeps the clearest radios and throws away the static-filled ones. Over time, the population evolves until the "noise" between the animals and their environment is gone. In math terms, the population structure becomes perfectly aligned with the environmental statistics. The authors show that the growth rate of a population is directly linked to Mutual Information—a measure of how much the population "knows" about its world.

It's as if evolution is a student studying for a test. The student who memorizes the textbook (the environment) perfectly will get an A (high fitness). The student who guesses randomly will fail. The paper suggests that the ultimate goal of natural selection is to turn the population into the most accurate "student" possible, maximizing the information it holds about its surroundings.

The Game of Life: Cooperation and Tipping Points

To prove this works, the authors applied their new "prediction" math to three classic problems in biology, and the results were surprisingly clear.

1. The Task Switching Game
Imagine a group of workers who can do two jobs: Job A or Job B. The boss (the environment) randomly switches between needing Job A or Job B.

The Result: The paper shows that the workers will naturally evolve to split their time exactly in proportion to how often the boss asks for each job. If the boss asks for Job A 70% of the time, the population will evolve so that 70% of the workers are ready for Job A. They become a perfect mirror of the environment's probability.

2. The Prisoner's Dilemma (The Cooperation Puzzle)
This is a famous game where two people can either cooperate or betray each other. Usually, betrayal wins in the short term, but cooperation is better for the group in the long run.

The Result: The authors found that cooperation only emerges if the players can "read" each other. If two players are completely random and independent (they don't know what the other will do), betrayal wins. But if there is a link between them—like they are related, or they play the game many times, or they can signal each other—cooperation becomes the winning strategy.
The paper uses a "phase diagram" (like a weather map for biology) to show exactly when cooperation takes over. It turns out that cooperation is like a tipping point. If the connection between players is strong enough, the whole group suddenly flips from "everyone betrays" to "everyone cooperates." This happens because the players are building a shared model of each other, increasing the "information" between them.

3. Group Dynamics
What happens when you have groups of people, and they interact with their own group and other groups?

The Result: The math shows that cooperation thrives inside a group if the members are closely linked (like family or friends). However, if the group interacts too much with outsiders who are "random" or untrustworthy, the whole system can collapse back into betrayal. The paper suggests that for large-scale cooperation (like in human societies) to work, groups need to develop strong internal signals and perhaps even "punish" those who break the rules, creating a tight loop of information and trust.

Why This Matters

The authors aren't just playing with numbers; they are offering a new toolkit. By treating fitness as a predictive probability, they can solve problems that were previously too messy to figure out. They can draw maps (phase diagrams) that show exactly when a population will switch from one behavior to another, just like water turning into ice.

They also show that the old idea of "fitness" was too narrow. It focused on the result (babies), but this new view focuses on the cause (the ability to predict). It suggests that evolution is a form of learning. Every time a species adapts, it isn't just getting stronger; it's getting smarter about its world.

The paper concludes that this approach unifies many different fields. The same math that explains how a bacteria finds food also explains how a stock market reacts to news, how a brain learns a language, and how a society decides to cooperate. It turns out that nature, data science, and human behavior might all be running on the same operating system: the drive to reduce uncertainty and maximize information.

So, the next time you see a bird building a nest or a person sharing a secret, remember: they aren't just acting on instinct. They are running a complex, invisible calculation, trying to guess the future, and in doing so, they are playing the ultimate game of life.

Technical Summary: Redefining Fitness: Inference, Information and Phase Transitions in Evolutionary Dynamics

Problem Statement
The paper addresses fundamental conceptual and practical limitations in the standard definition of fitness as "relative reproductive success." The authors identify three persistent difficulties with this definition:

Circularity: Defining a type as "fitter" because it reproduces more, and reproducing more because it is "fitter," offers no causal explanation for adaptation or predictive power.
Mismatch: Defining fitness strictly by reproductive output fails to distinguish between a phenotype's inherent adaptive capacity (propensity) and the stochastic outcome of a specific event (e.g., a well-adapted individual dying by chance).
Prediction: The standard definition struggles to predict when a type will be favored in future, changing environments.

While the propensity interpretation of fitness attempts to resolve these issues by viewing fitness as a probabilistic disposition, it has lacked a rigorous mathematical framework that connects it to the dynamics of selection. Furthermore, the connection between natural selection and statistical inference, previously noted as a formal mapping, has remained largely a mathematical curiosity without being operationalized as a tool for solving concrete evolutionary problems.

Methodology
The authors operationalize the mathematical equivalence between the replicator equation (under pure selection in asexual haploid populations) and Bayes' theorem. They propose a redefinition of fitness where the fitness of a type $s$ in an environmental state $e$ , denoted $w_s(e)$ , is proportional to a Bayesian likelihood function:
$w_s(e) = \alpha(e) p(e|s)$
Here, $p(e|s)$ represents the conditional probability of the environment given the phenotype (the type's model of the environment), and $\alpha(e)$ is an environment-dependent normalization factor.

The methodology proceeds by:

Mapping Dynamics to Inference: Treating type frequencies as prior probabilities and environmental observations as data, the update of frequencies over time is shown to be exactly equivalent to Bayesian updating.
Analyzing Log-Fitness: The authors shift focus from fitness ( $w$ ) to log-fitness ( $\ln w$ ), which corresponds to the growth rate ( $\gamma$ ). They demonstrate that statistically independent contributions to fitness multiply, while their logarithms add.
Deriving Information Quantities: By averaging log-fitness over environmental histories, the authors derive that the quantity maximized by selection is the mutual information $I(E, S)$ between the population structure and environmental statistics.
Polynomial Expansion: They express log-fitness as a polynomial in binary phenotypic variables. They analyze how the coefficients of these polynomials change sign as environmental parameters vary, identifying these sign changes as evolutionary phase transitions.

Key Contributions and Results
The paper applies this framework to three canonical problems, deriving phase diagrams and cooperation conditions in each:

Task Switching in Stochastic Environments:
- Result: The authors show that the long-run selected population distribution mirrors the environmental probability ( $f(s) \to p(e)$ ).
- Mechanism: Selection drives the population to maximize the mutual information $I(S, E)$ , effectively making the population a perfect predictor of the environment. The state maximizing selection is the one that fully explains the environmental variation.
Evolutionary Games (Prisoner's Dilemma and Mixed Games):
- Result: The log-fitness framework recovers and extends Hamilton-type rules for the emergence of cooperation.
- Mechanism:
  - In statistically independent populations (random matching), cooperation never emerges (defection is the Nash equilibrium).
  - In well-mixed populations (where agents share the same marginal probability), a phase transition occurs. If benefits exceed costs ( $B > C$ ), cooperation is globally optimal; otherwise, a "tipping point" (fitness valley) exists, requiring a critical initial frequency of cooperators to reach the global maximum.
  - In conditionally dependent populations (e.g., kin selection, reciprocity), the condition for cooperation is $rB + d f_{1|1} > C$ , where $r$ is the regression of one agent's strategy on another's. This unifies kin selection, direct reciprocity, and punishment as mechanisms that generate mutual information ( $I(S_1, S_2) > 0$ ) between agents.
- Mixed Games: When environments switch between competitive (Prisoner's Dilemma) and cooperative games, the probability of the cooperative environment acts as a control parameter (analogous to temperature), shifting the phase boundary and enabling cooperation to emerge even when it would fail in a static competitive environment.
Selection in Group-Structured Populations:
- Result: The framework generalizes to structured populations by distinguishing between in-group and out-group interactions.
- Mechanism: In-group interactions drive cooperation (via high relatedness/reciprocity), while out-group interactions often select for defection. The authors derive conditions under which cooperation can be stabilized within groups despite external pressure, and how larger-scale cooperative behaviors can evolve through the expansion of group markers and collective identities.

Significance and Claims
The paper claims that redefining fitness as a Bayesian likelihood resolves the circularity, mismatch, and prediction problems by grounding adaptation in a causal, probabilistic model of the environment rather than a retrospective record of reproductive success.

The primary significance of this work is the operationalization of the selection-inference mapping into a constructive analytical tool. The authors demonstrate that:

Information Maximization is the governing principle of natural selection in variable environments, as selection amplifies types with the most accurate environmental models.
Phase Transitions in evolutionary dynamics can be identified analytically by examining the sign changes in the polynomial coefficients of log-fitness, driven by environmental statistics.
Unification: Diverse mechanisms for the emergence of cooperation (kin selection, reciprocity, spatial structure) are unified as alternative means of generating mutual information within populations.

The authors note that this framework is exact for pure selection in discrete-time asexual haploid populations. They acknowledge that other forces (mutation, drift, recombination) modify these conclusions but suggest that the likelihood definition of fitness remains robust, with other forces adding complexity to the phase space (e.g., mutation acting as a Markov transition operator). The work positions evolutionary dynamics as a form of statistical learning, suggesting that evolution, Bayesian inference, and statistical learning share a common computational substrate.

The Great Game of Guessing

The Detective's Guide to Evolution

The Information Superhighway

The Game of Life: Cooperation and Tipping Points

Why This Matters

More like this