Evolutionary Entropy: Multiple Perspectives

• Evolutionary entropy is defined as an entropy functional on evolving systems, with distinct interpretations in demography, stochastic dynamics, thermodynamics, and computational optimization.
• In demographic models, it quantifies the temporal dispersion of reproduction through normalized reproductive distributions and generation time, thereby predicting life-history traits.
• In stochastic and algorithmic frameworks, it measures the uncertainty of state transitions and guides optimization, linking evolutionary dynamics to principles of entropy.

Evolutionary entropy denotes several non-equivalent constructions across population biology, stochastic evolutionary dynamics, thermodynamic theory, and computational methods. In demographic theory it is a demographic invariant measuring the temporal organization of reproduction; in finite-population stochastic models it is the entropy rate of the Markov chain generated by mutation, selection, and drift; in thermodynamic and self-organization frameworks it measures the number and diversity of metabolic cycles or the cooperativity of energy-transducing networks; and in computational work it is used as an objective over solution populations, sensori-motor states, or correspondence matrices (Oliveira, 24 Jun 2026, Harper, 2013, Demetrius et al., 2020, Nikfarjam et al., 2021). The common element is an entropy functional attached to an evolving system, but the underlying state space, normalization, and interpretation vary substantially.

1. Terminological scope and principal meanings

Across the literature represented here, the expression refers to distinct objects rather than a single invariant. In Demetrius-style structured-population theory, evolutionary entropy is a life-history quantity tied to the distribution of reproductive contribution across ages or stages (Buescu et al., 20 Jun 2026). In Markovian population models such as Moran and Wright-Fisher processes, it is the stationary average of transition entropies and therefore a measure of long-run pathwise randomness (Harper, 2014). In thermodynamic generalizations, it is defined from the number and diversity of metabolic cycles or from entropy-like functionals arising through coarse-graining and invariant measures (Demetrius et al., 2020). In evolutionary computation and robotics, entropy is attached to distributions over solution features or sensori-motor states and functions directly as a search objective (Delarboulas et al., 2010).

Usage family	State space	Representative formulation
Structured populations	Reproductive ages or stages	$H=-\sum_j p_j\log p_j \,/\, T$
Finite-population stochastic processes	Population compositions in a Markov chain	$H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$
Thermodynamic/self-organization theory	Metabolic cycles, macrostates, or coarse-grained variables	$H=\bar S/T$ or entropy from invariant measures
Computational evolutionary methods	Tour segments, sensori-motor states, or transport plans	$H(P)$ over features, or $H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$

A crucial terminological boundary is that some papers on entropy and evolution are explicitly not about specialized notions of evolutionary entropy. Klauber’s "Evolution and Earth's Entropy" is about ordinary physical thermodynamic entropy and states that it is "not about specialized notions such as 'evolutionary entropy' in mathematical biology" (Klauber, 2010). This distinction matters because arguments about the second law, solar flux, and planetary entropy balance belong to a different literature from demographic or information-theoretic evolutionary entropy.

2. Demographic entropy in age- and stage-structured populations

In the Demetrius framework, evolutionary entropy is a demographic invariant that measures the temporal organization of a population’s life cycle (Oliveira, 24 Jun 2026). For age-structured populations, the fundamental construction uses the growth-adjusted reproductive distribution. If $p_j$ is the probability that the mother of a randomly chosen newborn belongs to age class $j$, then

$$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$

Here $T$ is generation time and $H$ measures the temporal dispersion of reproductive contribution rather than lifetime replacement or asymptotic growth alone (Buescu et al., 20 Jun 2026).

The 2026 age-structured analysis sharpens this interpretation by showing that the relevant object is the normalized post-maturity reproductive distribution. Under Euler–Lotka normalization, if $H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$0 denotes the growth-discounted survivorship weight and $H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$1 fertility, then $H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$2 and $H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$3. The paper’s reduction principle states that evolutionary entropy and generation time are invariant under multiplicative rescaling of survivorship and fertility on the reproductive interval. Accordingly, "the relevant entropy is determined not by absolute survivorship, fertility, or juvenile mortality, but by the normalized post-maturity reproductive distribution" (Buescu et al., 20 Jun 2026). For constant fertility on a reproductive interval $H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$4, the entropy takes the explicit form

$H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$5

with

$H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$6

The same paper gives explicit formulas for finite and open-group Leslie models, including geometric reproductive tails. In the geometric regime, governed by $H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$7, it proves a sharp threshold determined by

$H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$8

separating populations with a unique finite entropy-maximizing endpoint from those whose entropy increases toward an asymptotic value determined solely by the age at first reproduction (Buescu et al., 20 Jun 2026). This converts reproductive lifespan into an entropy-optimization problem: for fixed $H(P)= -\sum_i \pi_i \sum_j T_{ij}\log T_{ij}$9, the reproductive endpoint $H=\bar S/T$0 can be studied through the maximization of $H=\bar S/T$1.

The stage-structured extension replaces Leslie matrices by irreducible Lefkovitch matrices and identifies entropy with the entropy rate of the Markov chain induced by the Perron–Frobenius normalization

$H=\bar S/T$2

If $H=\bar S/T$3 is the stationary distribution of $H=\bar S/T$4, then

$H=\bar S/T$5

For pure Lefkovitch matrices, the paper derives the closed form

$H=\bar S/T$6

with a decomposition

$H=\bar S/T$7

The first term measures uncertainty from transition or reproductive pathways, and the second measures uncertainty from stage retention versus advancement (Oliveira, 24 Jun 2026). In this framework, discrete-time growth decomposes as

$H=\bar S/T$8

so entropy appears as one component of long-run demographic performance, complementary to reproductive potential (Oliveira, 24 Jun 2026).

Empirically, the 2026 age-structured study tests the theory on 130 animal species and reports that predicted and observed reproductive medians coincide exactly for about $H=\bar S/T$9 of species, nearly $H(P)$0 are within one reproductive class, more than $H(P)$1 within two classes, and more than $H(P)$2 within three classes (Buescu et al., 20 Jun 2026). This makes evolutionary entropy a predictive life-history descriptor rather than a purely formal quantity.

3. Entropy rate in stochastic evolutionary processes

A different lineage defines evolutionary entropy as the entropy rate of the stationary Markov chain generated by mutation, selection, and stochastic drift. For a finite Markov chain with stationary distribution $H(P)$3 and transition probabilities $H(P)$4, the entropy rate is

$H(P)$5

This is interpreted as a measure of long-run variation because it combines where the process spends its time with how random its local transitions are (Harper, 2014).

In the multidimensional Moran and more general incentive processes with mutation, the population state is a count vector $H(P)$6 on the $H(P)$7-simplex, and the process becomes irreducible once mutation is present (Harper, 2014). The paper proves a universal upper bound depending only on the number of types: $H(P)$8 For the neutral landscape, entropy rate is monotonically increasing in the mutation probability $H(P)$9, and for the Moran/Fermi process with fixed $H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$0,

$H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$1

By contrast, for $H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$2-replicator and $H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$3-Fermi incentive processes with $H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$4,

$H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$5

The biological interpretation is direct: mutation sustains long-run variability, whereas strong directional selection typically suppresses it by pushing the process toward fixation states (Harper, 2014).

The two-type analysis of Moran and Wright–Fisher processes develops the same idea under the heading "inherent randomness" (Harper, 2013). Here the entropy rate is again

$H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$6

and is used to compare update rules, mutation regimes, and fitness landscapes. For the Moran process, a major analytic result is the population-size-independent upper bound

$H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$7

The extremal local transition law is

$H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$8

which has entropy $H(\mathbf W)=-\sum_{ij}W_{ij}\log W_{ij}$9 (Harper, 2013). For Wright–Fisher, the one-step transition law is binomial, and the maximal entropy rate at fixed $p_j$0 scales asymptotically as

$p_j$1

These papers emphasize that neutral evolution tends to maximize entropy, while decreasing mutation drives entropy to zero. They also distinguish sharply between boundary mutation and uniform mutation. Under boundary mutation, stationary mass can remain concentrated near fixation states, lowering entropy even in large populations. Under uniform mutation, the stationary distribution can concentrate near interior states of maximal local transition entropy (Harper, 2013). Entropy rate therefore quantifies neither fixation probability alone nor stationary diversity alone, but the stationary unpredictability of the evolutionary trajectory.

4. Thermodynamic and self-organizing formulations

Another major use of evolutionary entropy is explicitly thermodynamic. In "Evolutionary entropy and the Second Law of Thermodynamics," evolutionary entropy is defined as "a statistical measure which describes the number and diversity of metabolic cycles in a population of replicating organisms" (Demetrius et al., 2020). In the age-structured form,

$p_j$2

The same paper formulates a "Fundamental Theorem of Evolution": in systems open to the input of energy and matter, evolutionary entropy increases when the energy source is scarce and diverse, and decreases when the energy source is abundant and singular (Demetrius et al., 2020). Its central thesis is that, in an appropriate limit, this theorem becomes the second law of thermodynamics.

"Self-Organization, Evolutionary Entropy and Directionality Theory" generalizes this program to collective behavior and self-assembly (Demetrius, 2023). There evolutionary entropy is a statistical measure of cooperativity and also "describes the rate at which a network of interacting metabolic units convert an external energy source into mechanical energy and work." The paper again writes

$p_j$3

but interprets $p_j$4 as the number and diversity of interaction cycles and $p_j$5 as cycle time. Within Directionality Theory, self-assembly is analyzed as a variation–selection process, and the main principle is the "Entropic Principle of Self-Organization": equilibrium states maximize evolutionary entropy contingent on the production rate of the external energy source (Demetrius, 2023). In the matrix setting, growth, entropy, and reproductive potential satisfy

$p_j$6

A complementary thermodynamic program appears in "Thermodynamics of Evolution and the Origin of Life" (Vanchurin et al., 2021). Starting from a maximum-entropy principle constrained by a loss function, it derives a canonical ensemble

$p_j$7

with free energy

$p_j$8

The paper then defines "biological temperature" as the measure of stochasticity of the evolutionary process and "evolutionary potential" as the amount of evolutionary work required to add a new trainable variable: $p_j$9 This construction is extended to a grand-potential formalism and used to model the origin of life as a phase transition between an ensemble of molecules and an ensemble of organisms (Vanchurin et al., 2021).

At a more abstract level, "Why does entropy drive evolution equations?" argues that entropy drives evolution equations because it characterizes the invariant measure of an underlying stochastic dynamics after coarse-graining (Peletier, 8 Mar 2026). In the GENERIC framework, deterministic evolution takes the form

$j$0

with entropy $j$1 driving the irreversible part of the motion. The paper’s "Basic Answer" is that entropy is either equal to or a remnant of an invariant measure, and that both the entropy functional and its dynamical role are generated together by coarse-graining (Peletier, 8 Mar 2026). This is closely aligned with the description-dependent lesson of the Yule-type growth study, which shows that fine-grained entropy $j$2, coarse-grained entropy $j$3, and their time dependence depend on whether one tracks the full configuration or only a size distribution, and whether size is treated as continuous or discrete (Goh et al., 2016).

5. Ordinary thermodynamic entropy, evolution, and the second law

A distinct but frequently conflated literature addresses whether biological evolution conflicts with the second law of thermodynamics. These papers are explicitly about ordinary thermodynamic entropy rather than demographic or Markovian evolutionary entropy.

Bunn’s "Evolution and the second law of thermodynamics" treats Earth as part of a larger closed system containing Sun, Earth, and radiation, and states the relevant condition as

$j$4

Using $j$5 and $j$6, the paper estimates

$j$7

It then places a deliberately generous upper bound on the entropy decrease needed to build all present biomass,

$j$8

and concludes that sunlight can compensate this in

$j$9

which is less than a year (Bunn, 2009).

Klauber’s "Evolution and Earth’s Entropy" makes a different but complementary point. It emphasizes that Earth itself, as an open system, can undergo a local entropy decrease because incoming solar heat arrives effectively at a higher temperature than outgoing terrestrial radiation. Using

$$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$0

and an average day/night temperature variation of about $$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$1, the paper estimates a local entropy decrease of about $$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$2 of Earth’s entropy throughput. With Styer’s throughput rate

$$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$3

this gives approximately

$$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$4

which Klauber rounds as greater than

$$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$5

This dwarfs Styer’s $$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$6 and Bunn’s $$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$7 estimates for entropy decrease associated with evolution (Klauber, 2010). The significance for the present topic is primarily terminological: these arguments concern standard thermodynamic bookkeeping, not the specialized evolutionary-entropy theories of mathematical biology.

6. Computational and algorithmic uses

In computational research, entropy is often used as an explicit objective for evolutionary search rather than as a demographic invariant. "Entropy-Based Evolutionary Diversity Optimisation for the Traveling Salesperson Problem" defines population diversity by a high-order entropy measure over contiguous $$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$8-node tour segments (Nikfarjam et al., 2021). If $$H=\frac{Y}{T},\qquad Y=-\sum_{j=1}^{\omega} p_j\log p_j,\qquad T=\sum_{j=1}^{\omega} j\,p_j.$$9 is the frequency of a segment $T$0 in a population $T$1 of size $T$2, the contribution is

$T$3

and total entropy is

$T$4

The theory proves that maximum entropy is achieved if and only if the segment-frequency gap $T$5 is zero or one. Experimentally, the method improves over edge-based diversity measures especially for large populations and longer segments (Nikfarjam et al., 2021).

In evolutionary robotics, entropy becomes an intrinsic fitness. "Open-Ended Evolutionary Robotics: an Information Theoretic Approach" defines the sensori-motor stream

$T$6

and computes the entropy of clustered sensori-motor states: $T$7 This yields a "curiosity instinct" when applied to the current individual and a "discovery instinct" when the state archive is inherited across generations (Delarboulas et al., 2010). The paper’s point is that entropy can function as a self-computable fitness in embedded evolution without external task rewards.

A further computational use appears in cross-species cell-type matching. "Unsupervised Evolutionary Cell Type Matching via Entropy-Minimized Optimal Transport" introduces OT-MESH, which first solves entropy-regularized optimal transport

$T$8

and then minimizes the entropy of the resulting transport plan,

$T$9

through iterative cost-matrix refinement (Qiao, 30 May 2025). Here entropy is not a direct measure of evolution but of ambiguity in the inferred correspondence matrix: lower transport-plan entropy means sharper and more interpretable cross-species homology hypotheses. On a macaque peripheral-versus-foveal bipolar-cell benchmark, OT-MESH achieved Sparseness $H$0, Entropy $H$1, and ARI $H$2, whereas standard OT without MESH had Sparseness $H$3, Entropy $H$4, and ARI $H$5 (Qiao, 30 May 2025).

7. Broader conceptual and speculative extensions

Several papers extend entropy-centered evolutionary language beyond standard demographic, population-genetic, or optimization settings. "Evolutionary interpretations of entropy model for correspondence matrix calculation" derives an entropy-linear programming problem for origin–destination matrices from stochastic chemical kinetics with detailed balance and from population games / logit dynamics (Gasnikov et al., 2015). In the zero-noise limit, it proposes an equilibrium-selection rule that chooses, among path decompositions of a fixed edge-flow equilibrium, the one minimizing

$H$6

Here entropy is the macroscopic footprint of decentralized stochastic adaptation rather than a biological state variable.

"Triadic Conceptual Structure of the Maximum Entropy Approach to Evolution" is explicitly conceptual and philosophical (Herrmann-Pillath et al., 2010). It combines Peircean semiotics with maximum entropy, Maximum Power, and Maximum Entropy Production, treating evolution as the emergence of information-carrying structures that maximize information capacity and energy-gradient exploitation. The paper’s contribution is not a new computable entropy statistic but a triadic architecture linking sign, object, interpretant, and thermodynamic directionality.

Some thermodynamic extensions are openly speculative. "The Entropy Principle and the Influence of Sociological Pressures on SETI" proposes that intelligent technological civilization may be favored because it can increase entropy rapidly on planetary scales, and tests this idea with Monte Carlo realization techniques (Bozhilov et al., 2010). The same speculative thermodynamic style appears in "External Entropy Production and Human Evolution toward Multi-body Life," which distinguishes internal entropy production within the body from external entropy production generated by tools, fire, and cooperation, models brain size $H$7, interacting group size $H$8, and external entropy production

$H$9

and interprets the result as the emergence of "multi-body life" (Sawada et al., 17 Jun 2026). These works widen the semantic field of evolutionary entropy, but they do so by moving away from standard demographic or population-genetic usage toward broad nonequilibrium and civilizational interpretations.

Taken together, these literatures show that evolutionary entropy is best understood as a family of entropy constructions defined on evolving biological or algorithmic systems. In one family it is a demographic invariant of reproductive timing; in another it is a stationary entropy rate of mutation–selection–drift dynamics; in another it is a thermodynamic or coarse-grained functional associated with metabolic cycles, invariant measures, or self-organization; and in yet another it is an explicit optimization objective for exploration, diversity, or correspondence inference. The shared vocabulary is real, but the mathematical object and biological meaning depend entirely on the underlying dynamical framework.