Evolutionary Dynamics

Evolutionary dynamics is the mathematical study of how the frequencies of competing types (genes, strategies, phenotypes, replicators) change over time under selection, mutation, recombination, and stochastic drift. Key result: in deterministic infinite populations, selection drives frequencies according to the replicator equation, whose rest points and stability connect directly to Nash equilibria; in finite populations, fixation probabilities and the Moran process govern fate. Applies to population genetics, game theory, ecology, epidemiology, cancer biology, and cultural evolution. Founded by Fisher, Wright, Haldane (1920s–30s) and extended by Maynard Smith, Price, Hofbauer & Sigmund, and Nowak.

Evolutionary dynamics describes the temporal change in the composition of a population of self-replicating entities subject to differential reproductive success. Precisely, given a set of types indexed $i = 1, \dots, n$ with frequencies $x_{i}$ summing to one, the theory specifies a vector field $\overset{x}{˙} = F (x)$ (deterministic limit) or a Markov chain (finite populations) encoding how selection on fitness and stochastic sampling reshape the simplex of frequencies. It matters because it unifies Darwinian natural selection, the algebra of games, and the probability of stochastic extinction within a single dynamical-systems and stochastic-process framework. A surprising fact: natural selection — often imagined as relentlessly optimizing — can produce non-convergent limit cycles, chaos, and outcomes that are demonstrably worse for every individual (the evolution of cooperation, the tragedy of the commons, and evolutionary suicide all emerge from the same equations).

The Core Insight

Replication plus selection equals a flow on the probability simplex in which each type's growth rate is its fitness relative to the population mean. Whenever fitness depends on the frequencies themselves — i.e., the payoff to a strategy depends on what everyone else is doing — the dynamics become a game, and the fixed points of the flow coincide with the equilibria of that game. Finiteness adds noise: drift can fix a deleterious type or eliminate an advantageous one, so "survival of the fittest" becomes a statement about probabilities, not certainties.

Formal Statement

The Replicator Equation

Let the population consist of $n$ types with frequencies collected in $x = (x_{1}, \dots, x_{n})$ living on the simplex

S_{n} = {x \in R^{n} : x_{i} \geq 0, \sum_{i = 1}^{n} x_{i} = 1} .

⟳the constraints are nonnegativity of frequencies and normalization;

S_{n}

is an

(n - 1)

-dimensional standard simplex.

Let $f_{i} (x)$ be the fitness (per-capita reproductive rate) of type $i$ , and let

\overset{ˉ}{f} (x) = j = 1 \sum n x_{j} f_{j} (x)

be the mean population fitness. The replicator equation is

\overset{x}{˙}_{i} = x_{i} (f_{i} (x) - \overset{ˉ}{f} (x)), i = 1, \dots, n .

⟳start from unnormalized abundances

\overset{n}{˙}_{i} = n_{i} f_{i}

with total

N = \sum n_{j}

; differentiate

x_{i} = n_{i} / N

using the quotient rule and substitute

\dot{N} / N = \overset{ˉ}{f}

Here $f_{i}$ may be frequency-independent (constant — pure selection) or frequency-dependent (a game). In the linear game case $f_{i} (x) = (A x)_{i} = \sum_{j} a_{ij} x_{j}$ , where $A = (a_{ij})$ is the payoff matrix giving the payoff to type $i$ against type $j$ .

Evolutionarily Stable Strategy (ESS)

A strategy $x^{*} \in S_{n}$ is an ESS if for every $x \neq = x^{*}$ in a neighborhood,

x^{* ⊤} A x > x^{⊤} A x .

Equivalently (Maynard Smith's two conditions): for all

x \neq = x^{*}

x^{⊤} A x^{*} \leq x^{* ⊤} A x^{*}, and if equality holds then x^{⊤} A x < x^{* ⊤} A x .

⟳the equilibrium (first) condition is Nash; the stability (second) condition handles neutral alternative best replies by requiring the resident to beat the mutant in mutant-dominated populations. Expand

\overset{ˉ}{f}

of a perturbed population

(1 - ϵ) x^{*} + ϵ x

and require it to invade.

Folk Theorem of Evolutionary Game Theory

For the replicator dynamics [3] with linear fitness $f = A x$ on $S_{n}$ :

If $x^{*}$ is a Nash equilibrium, it is a rest point of [3].
If $x^{*}$ is the limit of an interior orbit as $t \to \infty$ , it is a Nash equilibrium.
If $x^{*}$ is Lyapunov stable, it is a Nash equilibrium.
If $x^{*}$ is an ESS, it is asymptotically stable.

The converse of (4) fails: asymptotic stability does not imply ESS for $n \geq 3$ in mixed strategies (though it does for the single-population symmetric case under certain conditions).

Equivalence: Replicator and Lotka–Volterra

The replicator equation on $S_{n}$ in $n$ variables is topologically equivalent (via a diffeomorphism removing one dimension) to the generalized Lotka–Volterra equation in $n - 1$ variables:

\overset{y}{˙}_{k} = y_{k} (r_{k} + l = 1 \sum n - 1 b_{k l} y_{l}), k = 1, \dots, n - 1.

⟳set

y_{k} = x_{k} / x_{n}

, differentiate the ratio, substitute [3], and absorb constants; the map

S_{n}^{\circ} \to R_{> 0}^{n - 1}

is the Hofbauer transformation.

Why It Holds / Derivation

Deriving the Replicator Equation

Step 1 — Unnormalized growth. Let

n_{i} (t) \geq 0

be the abundance (number) of type

i

, growing at per-capita rate

f_{i}

\overset{n}{˙}_{i} = n_{i} f_{i} (x) .

⟳this is the defining assumption of exponential per-capita reproduction with rate equal to fitness.

Step 2 — Total population. Let

N = \sum_{j} n_{j}

. Summing [7],

\dot{N} = j \sum \overset{n}{˙}_{j} = j \sum n_{j} f_{j} = N \overset{ˉ}{f} (x),

using

x_{j} = n_{j} / N

\sum_{j} n_{j} f_{j} = N \sum_{j} x_{j} f_{j} = N \overset{ˉ}{f}

⟳factor out

N

and recognize the frequency-weighted mean.

Step 3 — Quotient rule. With

x_{i} = n_{i} / N

\overset{x}{˙}_{i} = \frac{n ˙ _{i} N - n _{i} N ˙}{N ^{2}} = \frac{n _{i}}{N} (\frac{n ˙ _{i}}{n _{i}} - \frac{N ˙}{N}) = x_{i} (f_{i} - \overset{ˉ}{f}) .

⟳divide numerator and denominator by

N

and substitute [7] and [8].

Step 4 — Invariance of the simplex. Check

\sum_{i} \overset{x}{˙}_{i} = \sum_{i} x_{i} f_{i} - \overset{ˉ}{f} \sum_{i} x_{i} = \overset{ˉ}{f} - \overset{ˉ}{f} = 0

, so

\sum_{i} x_{i}

is conserved; and

x_{i} = 0 \Rightarrow \overset{x}{˙}_{i} = 0

, so each face is invariant. Hence

S_{n}

is forward-invariant.

■

ESS implies Asymptotic Stability under Replicator Dynamics

Step 1 — Candidate Lyapunov function. For an interior ESS

x^{*}

, define the relative entropy (Kullback–Leibler divergence)

V (x) = i : x_{i}^{*} > 0 \sum x_{i}^{*} ln \frac{x _{i}^{*}}{x _{i}} \geq 0,

with

V (x) = 0

iff

x = x^{*}

(Gibbs inequality).

⟳KL divergence is nonnegative and vanishes only at equality of distributions; it is the natural information-geometric distance on the simplex.

Step 2 — Time derivative. Differentiate along [3]:

\dot{V} = - i \sum x_{i}^{*} \frac{x ˙ _{i}}{x _{i}} = - i \sum x_{i}^{*} (f_{i} - \overset{ˉ}{f}) = - (x^{* ⊤} A x - x^{⊤} A x) .

⟳

\overset{x}{˙}_{i} / x_{i} = f_{i} - \overset{ˉ}{f}

; then

\sum_{i} x_{i}^{*} f_{i} = x^{* ⊤} A x

and

\sum_{i} x_{i}^{*} \overset{ˉ}{f} = \overset{ˉ}{f} = x^{⊤} A x

Step 3 — Sign from the ESS condition. By the ESS inequality [4],

x^{* ⊤} A x > x^{⊤} A x

for

x \neq = x^{*}

in a neighborhood, hence

\dot{V} < 0

there. By Lyapunov's theorem

x^{*}

is asymptotically stable.

■

John Maynard Smith | 1920–2004

Trained first as an aeronautical engineer, Maynard Smith turned to biology under J.B.S. Haldane. With George Price he introduced the evolutionarily stable strategy in the 1973 Nature paper "The Logic of Animal Conflict," importing von Neumann–Morgenstern game theory into Darwinian biology. His 1982 book Evolution and the Theory of Games established the field. Crucially, ESS reframed evolution: an equilibrium is not what is best for the species but what cannot be invaded by a rare mutant — a strictly individual-selectionist, dynamical notion that dissolved much loose "good of the species" thinking.

Building Intuition

Picture the simplex $S_{n}$ as the arena. A population is a single point; evolution moves it. The replicator equation says each coordinate grows in proportion to how much it beats the average. Above-average types swell; below-average types shrink; the average itself rises whenever fitness is constant (Fisher's theorem). Geometrically, the flow is a "fitness gradient" warped by the metric of the simplex.

Now make fitness frequency-dependent and the landscape deforms as you walk on it — like trying to climb a hill that sinks where you step. This is why frequency-dependent selection can produce stable coexistence (negative frequency dependence, the rare type wins) or oscillation (rock–paper–scissors, no type is ever best). The ESS is the place where, having arrived, no rare invader can do better — a fixed point that is also a "no regret" condition.

In finite populations, replace the smooth flow with a noisy walk on a lattice of integer counts. Now the destiny of a type is a gambler's ruin problem: it either fixes (reaches frequency 1) or vanishes. Selection biases the coin; drift keeps flipping it.

Replicator flow on the 1-simplex for a coordination game: two stable rest points at the corners separated by an unstable interior equilibrium $x^\ast$.

Worked Examples

Hawk–Dove Game: Finding the ESS and the Equilibrium Frequency

Given. Two strategies contest a resource of value

V = 2

. Cost of injury in an escalated fight

C = 6

. Payoff matrix (row = focal, column = opponent):

| | Hawk | Dove | |---|---|---| | Hawk | $(V - C) /2 = - 2$ | $V = 2$ | | Dove | $0$ | $V /2 = 1$ |

So $A = (- 2 0 21)$ .

Step 1 — Pure strategies. Hawk is not an ESS since Hawk-vs-Hawk payoff

- 2 <

Dove-vs-Hawk payoff

0

. Dove is not an ESS since Hawk invades Dove (Hawk-vs-Dove

2 >

Dove-vs-Dove

1

). So the ESS is mixed. Step 2 — Indifference (interior Nash). Let

p

= frequency of Hawk. A mixed ESS requires equal payoffs to Hawk and Dove:

payoff(Hawk) = - 2 p + 2 (1 - p), payoff(Dove) = 0 \cdot p + 1 \cdot (1 - p) .

Set equal:

- 2 p + 2 - 2 p = 1 - p \Rightarrow - 4 p + 2 = 1 - p \Rightarrow - 3 p = - 1.

Step 3 — Solve.

p^{*} = \frac{1}{3} .

General formula:

p^{*} = V / C = 2/6 = 1/3.

✓ Step 4 — Stability. With

V < C

this interior point is an ESS (since the off-diagonal advantage gives negative frequency dependence: rare Hawks do well, common Hawks do badly).

p_{Hawk}^{*} = \frac{1}{3} (dimensionless frequency), an ESS

Moran Process: Fixation Probability of an Advantageous Mutant

Given. A finite population of

N = 100

individuals, two types: resident (fitness 1) and mutant with relative fitness

r = 1.05

(5% advantage). One mutant appears. Use the constant-fitness Moran process. Step 1 — Fixation probability formula. For a single mutant of fitness

r

in a Moran population of size

N

, the probability that it fixes is

ρ = \frac{1 - 1/ r}{1 - 1/ r ^{N}} .

⟳solve the gambler's-ruin recursion for the birth–death chain whose ratio of down-to-up transition rates is

1/ r

, with absorbing boundaries at

0

and

N

Step 2 — Plug in.

1/ r = 1/1.05 = 0.95238

. Then

1 - 1/ r = 0.047619

1/ r^{N} = (0.95238)^{100}

. Compute

ln (0.95238) = - 0.048790

, times

100 = - 4.8790

, so

1/ r^{100} = e^{- 4.8790} = 0.007600

. Then

1 - 1/ r^{100} = 0.99240

. Step 3 — Divide.

ρ = \frac{0.047619}{0.99240} = 0.04798.

Step 4 — Interpret. A mutant with a 5% advantage fixes only ~4.8% of the time — about the naive selection coefficient

s = 0.0476 \approx (r - 1) / r

. Compare neutral case

ρ_{neutral} = 1/ N = 0.01

. So selection multiplies the fixation chance roughly fivefold but most beneficial mutants are still lost to drift.

ρ \approx 0.048 (dimensionless probability)

Common Misconceptions

"Evolution maximizes fitness / the population always climbs uphill." This is true only for frequency-independent selection of a single locus (Fisher's Fundamental Theorem, and only for the additive component of variance). With frequency dependence the mean fitness can decrease over time, orbits can cycle (rock–paper–scissors), and the dynamics can converge to a state where everyone is worse off (the Prisoner's Dilemma drives the population to mutual defection). There is no global Lyapunov function in general; "fitness landscapes" as static hills are a frequency-independent fiction. "ESS, Nash equilibrium, and asymptotically stable rest point are the same thing." They form a strict hierarchy: ESS

\Rightarrow

asymptotically stable

\Rightarrow

Nash

\Rightarrow

rest point, and none of the reverse implications hold in general. A rest point need not be Nash (e.g., a saddle); a Nash equilibrium need not be asymptotically stable (a center, as in matching pennies); and an asymptotically stable point need not be an ESS once

n \geq 3

Rock–Paper–Scissors: No ESS, Perpetual Cycling

Take the zero-sum cyclic payoff matrix

A = 01 - 1 - 1 01 1 - 1 0 .

The unique Nash equilibrium is the centroid

x^{*} = (1/3, 1/3, 1/3)

, but it is not an ESS — there is no neighborhood where

x^{* ⊤} A x > x^{⊤} A x

since the game is zero-sum and the equilibrium is a center. Under replicator dynamics the orbits are closed periodic curves around

x^{*}

(the function

\prod_{i} x_{i}

is conserved), so the population cycles forever Rock→Paper→Scissors→Rock and never settles. If the matrix is perturbed so the diagonal sum is slightly positive,

x^{*}

becomes asymptotically stable but still not an ESS — concretely refuting "asymptotically stable

\Rightarrow

ESS."

Connections

Evolutionary dynamics sits at a crossroads of several fields. The replicator equation is formally a Lotka–Volterra Equations system after a coordinate change [6], tying selection to predator–prey ecology. Its fixed-point theory is exactly Nash Equilibrium of the underlying game, making it the dynamic counterpart of static Game Theory. The KL-divergence Lyapunov function [10] reveals an Information Geometry structure — the replicator flow is gradient ascent in the Shahshahani (Fisher) metric. In finite populations the Moran Process and Wright–Fisher Model connect to Markov Chains and gambler's-ruin theory, while diffusion limits link to the Fokker–Planck Equation via Kimura's diffusion approximation. The adaptive-dynamics extension feeds into Bifurcation Theory through evolutionary branching.

Real-World Applications

Antibiotic and antiviral resistance. Replicator and Moran models predict the spread of resistant strains under drug pressure; fixation probabilities [12] guide dosing strategies to keep resistant mutants below the selection threshold (used in HIV combination therapy and tuberculosis treatment design).
Cancer evolution. Tumors are clonal populations under selection; the Moran process on tissue structures models the accumulation of driver mutations and the emergence of therapy resistance (Nowak's work on colorectal cancer and chronic myeloid leukemia).
Evolution of cooperation. Replicator dynamics with the Prisoner's Dilemma, plus mechanisms (kin selection, direct/indirect reciprocity, network reciprocity, group selection), explain altruism — the "five rules for cooperation" framework, applied from microbial public-goods to human institutions.
Vaccination and epidemiology. Game-theoretic models of voluntary vaccination treat individuals as strategists balancing infection risk against vaccine cost; the ESS vaccination level explains free-riding and herd-immunity shortfalls.
Animal conflict and behavior. The Hawk–Dove model and its descendants quantify ritualized fighting, territory contests, and sex-ratio evolution (Fisher's $1 : 1$ ratio is itself an ESS argument).
Cultural and language evolution. Replicator–mutator equations model the spread of conventions, grammars, and memes; Nowak's language dynamics derive coherence thresholds for shared grammar.
Algorithmic learning and AI. Multiplicative-weights/regret-matching algorithms are discrete-time replicator dynamics; understanding their convergence (and failure to converge in zero-sum games) informs the training of adversarial systems and learning in games.

Historical Development

1922–1930 — R.A. Fisher. Diffusion treatment of gene frequencies; the Fundamental Theorem of Natural Selection (1930) linking the rate of fitness increase to additive genetic variance.
1931 — Sewall Wright. Introduces the adaptive landscape and the role of random drift in small populations; the Wright–Fisher model of finite-population sampling.
1932 — J.B.S. Haldane. The Causes of Evolution; quantitative selection coefficients and the mathematical synthesis of Mendelism and Darwinism.
1958 — P.A.P. Moran. Introduces the Moran birth–death process, giving a clean finite-population Markov chain with tractable fixation probabilities [12].
1962 — Motoo Kimura. Diffusion approximation for fixation probabilities; later (1968) the neutral theory of molecular evolution, foregrounding drift.
1970 — George R. Price. The Price equation, a covariance identity decomposing evolutionary change into selection and transmission terms — the most general statement of evolutionary dynamics.
1973 — Maynard Smith & Price. "The Logic of Animal Conflict" defines the ESS; game theory enters biology.
1978 — Taylor & Jonker; Hofbauer, Schuster, Sigmund. Formulate the replicator equation as a dynamical system and prove the link to ESS and Lotka–Volterra.
1998–2006 — Hofbauer & Sigmund; Nowak. Evolutionary Games and Population Dynamics (1998) and Nowak's Evolutionary Dynamics (2006) consolidate the field, adding evolutionary graph theory (2005, fixation on networks) and the unification of finite/infinite, deterministic/stochastic frameworks.

Subtleties & Limits

Finite-population, structured, and non-equilibrium frontiers. Several deep questions remain open or only partially resolved:

Evolutionary graph theory. Fixation probabilities depend on population structure; certain graphs ("amplifiers of selection," e.g., stars) boost $ρ$ while others suppress it. A full classification of which graphs amplify versus suppress for arbitrary games (including frequency-dependent fitness) is incomplete, and the existence of "superamplifiers" robust to mutation remains debated.
The strategy-of-evolution problem. For games with $n \geq 4$ strategies, replicator dynamics can exhibit chaos (Hofbauer–Sigmund); characterizing parameter regions of chaotic versus convergent behavior, and what this means biologically, is unsettled.
Stochastic stability and timescale separation. When mutation, selection, and drift act on disparate timescales, the long-run distribution (stochastically stable states, à la Foster–Young / Fudenberg–Imhof) can differ from any deterministic prediction; rigorous results exist only for special classes.
Multilevel selection and the Price equation closure. The Price equation is exact but not closed (the transmission term requires its own dynamics); a fully predictive multilevel theory reconciling individual and group selection is still contested.
Eco-evolutionary feedback. Coupling population size dynamics to frequency dynamics (eco-evolutionary models) breaks the simplex-invariance assumption; generic conditions for evolutionary branching, coexistence, and evolutionary suicide under such feedback are an active research area in adaptive dynamics.