betaFeedback

WorksComputer Science

Equilibrium Statistics and Stability in Random Two-Population and Three-Population Evolutionary Games

Abstract

Multi-population matrix evolutionary games provide a framework for studying adaptation and strategic interactions among heterogeneous agents. When payoffs are uncertain, however, analysing a specified game does not reveal how frequently different evolutionary outcomes occur or how their probabilities depend on modelling assumptions. We study two- and three-population games with two strategies per population and both within- and between-population pairwise interactions, with payoffs fixed during each evolutionary trajectory. For two populations, we derive equilibrium-existence probabilities and analytical stability results for generic games and three structured classes: diversity, exclusive, and avoidance games. For generic two-population games, independent Gaussian sampling of payoff entries and of replicator coefficients produces different equilibrium statistics. Interaction structure further distinguishes coexistence from attraction: diversity and exclusive games share the same interior-existence probability, but only exclusive games admit asymptotically stable interior equilibria. Avoidance games admit transversely attracting equilibrium families without locally asymptotically stable individual points. For generic three-population games under independent payoff sampling, we obtain existence probabilities for all equilibrium types and joint existence-and-stability probabilities for boundary equilibria. Monte Carlo simulations support the analytical predictions and quantify multistability, while phase portraits illustrate initial-condition dependence. Under the same Gaussian payoff-sampling protocol, simulations show a lower frequency of fully mixed equilibria but a higher frequency of multistability for three populations than for two.

Introduction

Collective adaptation in heterogeneous multi-agent systems depends on both the incentives facing individual groups and the interactions that connect them. The importance of coupled populations is well established in ecological dynamics [1, 2, 3]. In strategic settings, comparable interdependence arises among governments, enterprises, and communities in resource provision and environmental governance [4, 5, 6, 7], and among developers, users, regulators, and media in AI governance [8, 9, 10]. Multi-population evolutionary games describe adaptation among groups with different roles [11], while networked extensions highlight the effects of within- and between-group interactions [12, 13, 14, 15]. Evolutionary dynamics also connects collective-behaviour modelling with computational applications such as sensor identification [16, 17]. When interaction costs and benefits are uncertain, however, analysing one specified game cannot establish how representative its outcomes are [18].

Existing research shows that evolutionary outcomes depend on how interactions and adaptation jointly determine the relative success of competing strategies. Even with an unchanged strategy set, differences in interaction structure can change whether cooperation and fairness persist [19, 20, 21, 22, 23]. Persistence, however, is not synonymous with convergence: coexistence can occur at an attracting mixed equilibrium or through sustained oscillations, while multiple attractors can make the realised outcome depend on initial conditions [24, 25, 26, 27, 28]. Analyses of asymmetric games further show that equilibria need not be isolated: continuous equilibrium sets require their own stability analysis [29, 30]. Alongside these dynamical insights, random-game theory quantifies equilibrium occurrence across uncertain payoff realisations [31, 32, 33, 34]. Connections between equilibrium equations and random polynomial systems make it possible to characterise the expected number of interior equilibria and the distribution of their counts [35, 36, 37]. Together, these advances distinguish three questions: which equilibria are possible, how frequently they occur, and whether they attract the dynamics.

However, connecting these questions for populations with both within- and between-population interactions requires a consistent treatment of uncertainty, interaction structure, and stability. Interior-equilibrium counts alone do not describe boundary outcomes, and coexistence does not necessarily imply attraction. Moreover, specifying uncertainty in payoff entries is not equivalent to specifying it directly in replicator coefficients: shared payoff entries induce dependence between the coefficients, even when the entries themselves are independent. Restrictions on which interactions contribute to payoffs can further change both equilibrium frequencies and whether equilibria are isolated or form continuous families. These issues concern variation across fixed games, distinct from stochastic forcing in population dynamics or changes of game during evolution [38, 39, 40]. Addressing them within a common framework would provide benchmarks when precise payoffs are unavailable and clarify which conclusions depend on the assumed representation of uncertainty, rather than on interaction structure alone.

To address these questions, we study generalised evolutionary games with two strategies per population under Gaussian uncertainty, focusing on two and three populations. For generic two-population games, we compare independent payoff sampling, which starts from the game matrix, with independent coefficient sampling, which starts from the replicator equations but omits the dependence induced by shared payoffs. We then impose zero-block constraints to obtain diversity, exclusive, and avoidance games, sampling only the active payoff entries. Gaussian probability calculations and the Kac–Rice formula [41] yield equilibrium-existence probabilities and analytical stability results. For three populations, we retain pairwise interactions and derive existence probabilities for all candidate types and joint existence-and-stability probabilities for boundary candidates; interior stability is assessed numerically. Monte Carlo simulations test the predictions and quantify multistability, while phase portraits illustrate initial-condition dependence and non-equilibrium attraction. Payoffs remain fixed during each trajectory, so the randomness describes variation across games rather than temporal noise.

In summary, the main contributions of this paper are as follows:

  1. We derive equilibrium-existence and joint existence-and-stability probabilities for generic two-population games under payoff and coefficient sampling, showing how uncertainty representation changes equilibrium statistics.

  1. We determine how three interaction structures change equilibrium occurrence and stability: diversity games have no asymptotically stable interior equilibrium, exclusive games can support stable interior coexistence, and avoidance games admit transversely attracting equilibrium families without asymptotically stable individual points.

  1. We extend the generic analysis to three populations under payoff sampling, deriving existence probabilities for all candidate types and joint existence-and-stability probabilities for boundary candidates. Simulations quantify multistability and three-population interior stability, while two-population phase portraits illustrate initial-condition dependence.

Section II defines the models and numerical protocol. Sections III and IV present two- and three-population results. Section V reports numerical results and Section VI concludes. Proofs, additional simulations, and model extensions are provided in the Supplementary Material.

Models and Methods

General setting

Consider mm infinitely large, well-mixed populations with strategy sets Sj={1,…,nj}S_j=\{1,\ldots,n_j\} [11], [26]. Let xijx_{ij} be the frequency of strategy ii in population jj, with ∑i∈Sjxij=1\sum_{i\in S_j}x_{ij}=1, and write xj=(xij)i=1nj\mathbf{x}_j=(x_{ij})_{i=1}^{n_j} and x=(x1,…,xm)\mathbf{x}=(\mathbf{x}_1,\ldots,\mathbf{x}_m). If πij(x)\pi_{ij}(\mathbf{x}) is the strategy payoff, the population average payoff and replicator dynamics are

πˉj(x)=∑i∈Sjxijπij(x),(1)\bar{\pi}_j(\mathbf{x})=\sum_{i\in S_j}x_{ij}\pi_{ij}(\mathbf{x}), \tag*{(1)}
x˙ij=xij(πij(x)−πˉj(x)),1≤j≤m,i∈Sj.(2)\dot{x}_{ij}=x_{ij}\bigl(\pi_{ij}(\mathbf{x})-\bar{\pi}_j(\mathbf{x})\bigr),\qquad1\leq j\leq m,\quad i\in S_j. \tag*{(2)}

Thus, a strategy increases in frequency when its payoff exceeds its population average. The frequency constraints reduce the system to ∑j=1m(nj−1)\sum_{j=1}^{m}(n_j-1) independent coordinates.

When each population has two strategies, label them s2i−1s_{2i-1} and s2is_{2i}, and let xix_i denote the frequency of the former. With x=(x1,…,xm)\mathbf{x}=(x_1,\ldots,x_m) and i=1,…,mi=1,\ldots,m,

fi(x):=πs2i−1(x)−πs2i(x),(3)f_i(\mathbf{x}):=\pi_{s_{2i-1}}(\mathbf{x})-\pi_{s_{2i}}(\mathbf{x}), \tag*{(3)}
x˙i=xi(1−xi)fi(x).(4)\dot{x}_i=x_i(1-x_i)f_i(\mathbf{x}). \tag*{(4)}

Each encounter involves two players from the same or different populations. Let βrs\beta_{rs} be the payoff to strategy srs_r against sss_s. With equal interaction weights, the expected payoff for r=1,…,2mr=1,\ldots,2m is

πsr(x)=∑j=1m[βr,2j−1xj+βr,2j(1−xj)].(5)\pi_{s_r}(\mathbf{x})=\sum_{j=1}^{m}\left[\beta_{r,2j-1}x_j+\beta_{r,2j}(1-x_j)\right]. \tag*{(5)}

Averaging over populations instead would multiply all payoffs by 1/m1/m and only rescale time.

A state x∗∈[0,1]m\mathbf{x}^{*}\in[0,1]^m is an equilibrium if and only if

xi∗(1−xi∗)fi(x∗)=0,i=1,…,m.(6)x_i^{*}(1-x_i^{*})f_i(\mathbf{x}^{*})=0,\qquad i=1,\ldots,m. \tag*{(6)}

Interior equilibria satisfy fi(x∗)=0f_i(\mathbf{x}^{*})=0 for every ii; boundary equilibria require the full condition (6). Pure coordinates represent one strategy, while mixed coordinates represent coexistence of both strategies in that population.

Unless otherwise specified, stability below means local asymptotic stability relative to the feasible state space: nearby feasible states remain nearby and converge to the equilibrium.

Two-population two-player two-strategy games

For m=2m=2, the strategy sets are {s1,s2}\{s_1,s_2\} and {s3,s4}\{s_3,s_4\}. The diagonal and off-diagonal blocks of B=(βrs)B=(\beta_{rs}) describe within- and between-population payoffs, respectively:

B=s1s2s3s4s1β11β12β13β14s2β21β22β23β24s3β31β32β33β34s4β41β42β43β44(7)B= \begin{array}{c|cccc} &s_1&s_2&s_3&s_4\\ s_1&\beta_{11}&\beta_{12}&\beta_{13}&\beta_{14}\\ s_2&\beta_{21}&\beta_{22}&\beta_{23}&\beta_{24}\\ s_3&\beta_{31}&\beta_{32}&\beta_{33}&\beta_{34}\\ s_4&\beta_{41}&\beta_{42}&\beta_{43}&\beta_{44} \end{array} \tag*{(7)}

Write x=x1x=x_1 and y=x2y=x_2 for the frequencies of s1s_1 and s3s_3, respectively. Substituting (5) into (3) gives

f1(x,y)=a1x+a2y+a3,f2(x,y)=b1x+b2y+b3,\begin{aligned} f_1(x,y)&=a_1x+a_2y+a_3,\\ f_2(x,y)&=b_1x+b_2y+b_3, \end{aligned}

where

a1=β11−β21+β22−β12,a2=β13−β23+β24−β14,a3=β12−β22+β14−β24,b1=β31−β41+β42−β32,b2=β33−β43+β44−β34,b3=β32−β42+β34−β44.\begin{aligned} a_1&=\beta_{11}-\beta_{21}+\beta_{22}-\beta_{12},\\ a_2&=\beta_{13}-\beta_{23}+\beta_{24}-\beta_{14},\\ a_3&=\beta_{12}-\beta_{22}+\beta_{14}-\beta_{24},\\ b_1&=\beta_{31}-\beta_{41}+\beta_{42}-\beta_{32},\\ b_2&=\beta_{33}-\beta_{43}+\beta_{44}-\beta_{34},\\ b_3&=\beta_{32}-\beta_{42}+\beta_{34}-\beta_{44}. \end{aligned}

Thus (4) becomes

x˙=x(1−x)f1(x,y),y˙=y(1−y)f2(x,y).(8)\begin{aligned} \dot{x}&=x(1-x)f_1(x,y),\\ \dot{y}&=y(1-y)f_2(x,y). \tag*{(8)} \end{aligned}

Three-population two-player two-strategy games

For m=3m = 3, let xx, yy, zz be the frequencies of s1s_1, s3s_3, s5s_5 in the strategy pairs {s1,s2}\{s_1, s_2\}, {s3,s4}\{s_3, s_4\}, and {s5,s6}\{s_5, s_6\}. All blocks of B∈R6×6B \in\mathbb{R}^{6 \times6} are retained, with payoffs given by (5).

Let u=(x,y,z)T\boldsymbol{u} = (x,y,z)^{T} and write

f(u)=Cu+d,C=(a1a2a3b1b2b3c1c2c3),d=(a4b4c4).(9)f(\boldsymbol{u}) = C\boldsymbol{u} + \boldsymbol{d}, \qquad C = \begin{pmatrix} a_1 & a_2 & a_3 \\ b_1 & b_2 & b_3 \\ c_1 & c_2 & c_3 \end{pmatrix}, \qquad\boldsymbol{d} = \begin{pmatrix} a_4 \\ b_4 \\ c_4 \end{pmatrix}. \tag*{(9)}

The payoff-to-coefficient mapping can be expressed compactly as

Cij=β2i−1,2j−1−β2i,2j−1−β2i−1,2j+β2i,2j,C_{ij} = \beta_{2i-1,2j-1} - \beta_{2i,2j-1} - \beta_{2i-1,2j} + \beta_{2i,2j},
di=∑j=13(β2i−1,2j−β2i,2j),i,j∈{1,2,3}.(10)d_i = \sum_{j=1}^{3}(\beta_{2i-1,2j} - \beta_{2i,2j}), \qquad i,j \in\{1,2,3\}. \tag*{(10)}

Thus the replicator dynamics on [0,1]3[0,1]^3 is

x˙=x(1−x)(a1x+a2y+a3z+a4),y˙=y(1−y)(b1x+b2y+b3z+b4),z˙=z(1−z)(c1x+c2y+c3z+c4).(11)\begin{aligned} \dot{x} &= x(1-x)(a_1x+a_2y+a_3z+a_4),\\ \dot{y} &= y(1-y)(b_1x+b_2y+b_3z+b_4),\\ \dot{z} &= z(1-z)(c_1x+c_2y+c_3z+c_4). \tag*{(11)} \end{aligned}

Deterministic three-party games with two strategies per party have also been analysed in [29], [30]. Unlike those models, the present extension retains pairwise interactions, so the payoff differences remain affine in the population frequencies.

Random payoffs

The replicator dynamics depend on the payoff entries only through the coefficients of the payoff differences. Uncertainty can therefore be specified either in the underlying payoffs or directly in these coefficients. We distinguish two Gaussian sampling models:

  • Payoff sampling (microscopic uncertainty): active payoff entries are sampled independently from N(0,1)\mathcal{N}(0,1), and the replicator coefficients are then calculated from them. For structured games, excluded payoff blocks are fixed at zero.

  • Coefficient sampling (macroscopic uncertainty): the coefficients of the generic replicator equations are sampled independently from N(0,1)\mathcal{N}(0,1).

Derived coefficients are generally correlated and need not have equal variance; independent coefficient sampling does not preserve this dependence. Sampled parameters remain fixed throughout each game’s dynamics.

Numerical protocol

Monte Carlo simulations use M=106M = 10^6 independent game realisations per setting under the sampling assumptions above. Comparisons across interaction structures share the same base payoff draws. Equilibrium candidates are computed algebraically and retained when all free coordinates lie strictly in (0,1)(0,1).

For isolated candidates, numerical local asymptotic stability requires every eigenvalue of the full replicator Jacobian to satisfy

Re⁡λ<−τ,τ=10−12+10−10max⁡(1,∥J∥∞).\operatorname{Re}\lambda< -\tau, \qquad\tau= 10^{-12} + 10^{-10}\max(1,\lVert J\rVert_{\infty}).

Real parts within [−τ,τ][-\tau,\tau] are flagged for separate assessment. Structurally nonhyperbolic cases are classified using the analytical results.

All reported frequencies use the full MM realisations as their denominator. For candidate ii, EiE_i denotes existence and SiS_i local asymptotic stability; thus P(Ei∩Si)\mathbb{P}(E_i \cap S_i) is a joint probability. The counts NN and NSN_S denote existing and locally asymptotically stable equilibria. For avoidance games, they instead count equilibrium families and transversely attracting families, with transverse attraction denoted by TT. Grouped entries report mean per-candidate frequencies; where shown, SD denotes the between-candidate standard deviation. Implementation details are provided in the Supplementary Material.

Two-population analytical results

We analyse generic games under both sampling models and diversity, exclusive, and avoidance games under independent Gaussian sampling of their active payoffs. Detailed proofs and expanded special-case formulas are provided in the Supplementary Material.

Generic games

In the nondegenerate case, there are four vertex candidates, four open-edge candidates, and one interior candidate. The latter five require their free coordinates to lie in (0,1)(0,1).

Lemma 1. Assume that the denominators below are nonzero, as holds almost surely under either generic sampling model. The equilibrium candidates of (8) and the corresponding Jacobian eigenvalues are as follows. An open-edge or interior candidate is an equilibrium only when all of its free coordinates lie strictly in (0,1)(0,1).

  • at A1=(0,0)A_1 = (0,0): λ1=a3\lambda_1 = a_3 and λ2=b3\lambda_2 = b_3.

  • at A2=(0,1)A_2 = (0,1): λ1=a2+a3\lambda_1 = a_2 + a_3 and λ2=−(b2+b3)\lambda_2 = -(b_2 + b_3).

  • at A3=(1,0)A_3 = (1,0): λ1=−(a1+a3)\lambda_1 = -(a_1 + a_3) and λ2=b1+b3\lambda_2 = b_1 + b_3.

  • at A4=(1,1)A_4 = (1,1): λ1=−(a1+a2+a3)\lambda_1 = -(a_1 + a_2 + a_3) and λ2=−(b1+b2+b3)\lambda_2 = -(b_1 + b_2 + b_3).

  • at A5=(0,−b3b2)A_5 = (0,-\frac{b_3}{b_2}): λ1=−b3(b2+b3)b2\lambda_1 = -\frac{b_3(b_2+b_3)}{b_2} and λ2=a3b2−a2b3b2\lambda_2 = \frac{a_3b_2-a_2b_3}{b_2}.

  • at A6=(1,−b1+b3b2)A_6 = (1,-\frac{b_1+b_3}{b_2}): λ1=−(b1+b3)(b1+b2+b3)b2\lambda_1 = -\frac{(b_1+b_3)(b_1+b_2+b_3)}{b_2}, λ2=a2b1−a1b2−a3b2+a2b3b2\lambda_2 = \frac{a_2b_1-a_1b_2-a_3b_2+a_2b_3}{b_2}.

  • at A7=(−a3a1,0)A_7 = (-\frac{a_3}{a_1},0): λ1=−a3(a1+a3)a1\lambda_1 = -\frac{a_3(a_1+a_3)}{a_1} and λ2=−a3b1+a1b3a1\lambda_2 = \frac{-a_3b_1+a_1b_3}{a_1}.

  • at A8=(−a2+a3a1,1)A_8 = (-\frac{a_2+a_3}{a_1},1): λ1=−(a2+a3)(a1+a2+a3)a1\lambda_1 = -\frac{(a_2+a_3)(a_1+a_2+a_3)}{a_1} and λ2=a2b1+a3b1−a1b2−a1b3a1\lambda_2 = \frac{a_2b_1+a_3b_1-a_1b_2-a_1b_3}{a_1}.

  • at A9=(x∗,y∗)=(a2b3−a3b2a1b2−a2b1,a3b1−a1b3a1b2−a2b1)A_9 = (x^*,y^*) = (\frac{a_2b_3-a_3b_2}{a_1b_2-a_2b_1},\frac{a_3b_1-a_1b_3}{a_1b_2-a_2b_1}), λ1=−b+Δ2\lambda_1 = \frac{-b+\sqrt{\Delta}}{2}, λ2=−b−Δ2\lambda_2 = \frac{-b-\sqrt{\Delta}}{2}, where b=−[a1x∗(1−x∗)+b2y∗(1−y∗)]b = -[a_1x^*(1-x^*)+b_2y^*(1-y^*)], Δ=[a1x∗(1−x∗)−b2y∗(1−y∗)]2+4a2b1x∗(1−x∗)y∗(1−y∗)\Delta= [a_1x^*(1-x^*)-b_2y^*(1-y^*)]^2+4a_2b_1x^*(1-x^*)y^*(1-y^*).

The following theorem compares per-candidate existence and joint existence-and-stability probabilities under the two sampling models.

Theorem 1 (Equilibrium analysis of generic two-population two-player two-strategy random games).

(1) Coefficient sampling (macroscopic uncertainty). Suppose that a1,a2,a3,b1,b2,b3a_1,a_2,a_3,b_1,b_2,b_3 are independent N(0,1)\mathcal{N}(0,1) random

variables. Each vertex A1,…,A4A_1,\ldots,A_4 always exists and is locally asymptotically stable with probability 1/41/4. Each of A5A_5 and A7A_7 exists with probability 1/41/4 and exists and is locally asymptotically stable with probability 1/161/16. For each of A6A_6 and A8A_8, the corresponding probabilities are

pedge=1πarctan⁡12≈0.1959,pedge4≈0.0490.p_{\mathrm{edge}}=\frac{1}{\pi}\arctan\frac{1}{\sqrt{2}}\approx0.1959,\qquad\frac{p_{\mathrm{edge}}}{4}\approx0.0490.

The interior candidate A9A_9 exists with probability 1/121/12 and exists and is locally asymptotically stable with probability 1/481/48.

(2) Payoff sampling (microscopic uncertainty). Suppose that all payoff entries βij\beta_{ij} are independent N(0,1)\mathcal{N}(0,1) random variables. Each vertex A1,…,A4A_1,\ldots,A_4 always exists and is locally asymptotically stable with probability 1/41/4. Each open-edge candidate A5,…,A8A_5,\ldots,A_8 exists with probability 1/31/3 and exists and is locally asymptotically stable with probability 1/121/12. The interior candidate A9A_9 exists with probability

pint=2arctan⁡(2)−arctan⁡(22)π≈0.21635,p_{\mathrm{int}}=\frac{2\arctan(\sqrt{2})-\arctan(2\sqrt{2})}{\pi}\approx0.21635,

and exists and is locally asymptotically stable with probability pint/4≈0.05409p_{\mathrm{int}}/4\approx0.05409.

Under either sampling model, the conditional probability of local asymptotic stability of the interior equilibrium, given its existence, is 1/41/4. Numerical results below test these predictions and quantify the distributions of equilibrium counts within individual games. Idea of proof: The calculations separate the candidates according to their location. Vertices always exist, so their stability probabilities follow from the signs of the two Jacobian eigenvalues. For open-edge candidates A5,…,A8A_5,\ldots,A_8, existence requires the free coordinate to lie in (0,1)(0,1), and stability additionally requires both eigenvalues to be negative. These events are evaluated using the relevant joint Gaussian distributions, retaining the dependence induced by the chosen sampling model. For the interior candidate A9A_9, the Kac–Rice formula [41] gives its existence probability through the expected number of zeros of the affine payoff-difference system. Independent sign changes of the two payoff-difference equations yield the interior-stability probabilities. Detailed proofs are provided in the Supplementary Material.

The Supplementary Material further examines how changing the relative weights of within- and between-population interactions affects equilibrium statistics in a scaled payoff extension of the generic model.

We now keep the payoff-sampling assumption fixed and vary which interactions contribute to payoffs. Diversity games retain only between-population contributions, exclusive games retain only within-population contributions, and avoidance games retain contributions only within population 1. These restrictions isolate the roles of the two interaction types and reveal how they affect both equilibrium occurrence and attraction.

Diversity games

Diversity games retain only between-population payoffs: both diagonal blocks of BB are zero, giving a1=b2=0a_1=b_2=0.

This structure includes two-strategy representations of role-separated games such as ultimatum, trust, and dictator games:

(00β13β1400β23β24β31β3200β41β4200)\begin{pmatrix} 0 & 0 & \beta_{13} & \beta_{14}\\ 0 & 0 & \beta_{23} & \beta_{24}\\ \beta_{31} & \beta_{32} & 0 & 0\\ \beta_{41} & \beta_{42} & 0 & 0 \end{pmatrix}

and the replicator dynamics becomes

x˙=x(1−x)(a2y+a3),\dot{x}=x(1-x)(a_2y+a_3),
y˙=y(1−y)(b1x+b3).\dot{y}=y(1-y)(b_1x+b_3).

where a3=β14−β24a_3=\beta_{14}-\beta_{24}, and b3=β32−β42b_3=\beta_{32}-\beta_{42}. Under the stated payoff-sampling model, a2b1≠0a_2b_1\ne0 almost surely, and open-edge equilibria occur with probability zero. The candidates therefore consist almost surely of the four vertices and one interior candidate, denoted here by A5A_5. The latter is feasible only when both coordinates lie strictly in (0,1)(0,1). Their coordinates and Jacobian eigenvalues are:

  • at A1=(0,0)A_1=(0,0): λ1=a3\lambda_1=a_3, λ2=b3\lambda_2=b_3.

  • at A2=(0,1)A_2=(0,1): λ1=a2+a3\lambda_1=a_2+a_3, λ2=−b3\lambda_2=-b_3.

  • at A3=(1,0)A_3=(1,0): λ1=−a3\lambda_1=-a_3, λ2=b1+b3\lambda_2=b_1+b_3.

  • at A4=(1,1)A_4=(1,1): λ1=−(a2+a3)\lambda_1=-(a_2+a_3), λ2=−(b1+b3)\lambda_2=-(b_1+b_3).

  • at A5=(−b3b1,−a3a2)A_5=\left(-\frac{b_3}{b_1},-\frac{a_3}{a_2}\right), λ1=−a3(a2+a3)b3(b1+b3)a2b1\lambda_1=-\sqrt{\frac{a_3(a_2+a_3)b_3(b_1+b_3)}{a_2b_1}}, λ2=a3(a2+a3)b3(b1+b3)a2b1\lambda_2=\sqrt{\frac{a_3(a_2+a_3)b_3(b_1+b_3)}{a_2b_1}}.

Lemma 2. Suppose that the active payoff entries in the two off-diagonal blocks are independent N(0,1)\mathcal{N}(0,1) random variables, while the diagonal blocks are zero. Then:

  1. Each vertex A1,…,A4A_1,\ldots,A_4 always exists and is locally asymptotically stable with probability $1/4.

  1. The interior candidate A5A_5 exists with probability 1/41/4. When it exists, it is either a centre or a saddle, and is never locally asymptotically stable.

A centre has nearby closed trajectories and does not attract them. The Supplementary Material applies this structure to an ultimatum game, including stability of its continuous equilibrium set.

Exclusive games

Exclusive games retain only within-population payoffs: the off-diagonal blocks of BB are zero, giving a2=b1=0a_2=b_1=0, a3=β12−β22a_3=\beta_{12}-\beta_{22}, and b3=β34−β44b_3=\beta_{34}-\beta_{44}:

(β11β1200β21β220000β33β3400β43β44)\begin{pmatrix} \beta_{11} & \beta_{12} & 0 & 0\\ \beta_{21} & \beta_{22} & 0 & 0\\ 0 & 0 & \beta_{33} & \beta_{34}\\ 0 & 0 & \beta_{43} & \beta_{44} \end{pmatrix}

and the dynamics is

x˙=x(1−x)(a1x+a3),\dot{x}=x(1-x)(a_1x+a_3),
y˙=y(1−y)(b2y+b3).\dot{y}=y(1-y)(b_2y+b_3).

Assume a1b2≠0a_1b_2\ne0, as holds almost surely under the stated payoff-sampling model. The equilibrium candidates and their Jacobian eigenvalues are listed below. An open-edge or interior candidate is feasible only when all its free coordinates lie strictly in (0,1)(0,1).

  • at A1=(0,0)A_1=(0,0): λ1=a3\lambda_1=a_3, λ2=b3\lambda_2=b_3.

  • at A2=(0,1)A_2=(0,1): λ1=a3\lambda_1=a_3, λ2=−b2−b3\lambda_2=-b_2-b_3.

  • at A3=(1,0)A_3=(1,0): λ1=−a1−a3\lambda_1=-a_1-a_3, λ2=b3\lambda_2=b_3.

  • at A4=(1,1)A_4=(1,1): λ1=−a1−a3\lambda_1=-a_1-a_3, λ2=−b2−b3\lambda_2=-b_2-b_3.

  • at A5=(0,−b3b2)A_5=\left(0,-\frac{b_3}{b_2}\right): λ1=a3\lambda_1=a_3, λ2=−b3(b2+b3)b2\lambda_2=-\frac{b_3(b_2+b_3)}{b_2}.

  • at A6=(1,−b3b2)A_6=\left(1,-\frac{b_3}{b_2}\right): λ1=−a1−a3\lambda_1=-a_1-a_3, λ2=−b3(b2+b3)b2\lambda_2=-\frac{b_3(b_2+b_3)}{b_2}.

  • at A7=(−a3a1,0)A_7=\left(-\frac{a_3}{a_1},0\right): λ1=−a3(a1+a3)a1\lambda_1=-\frac{a_3(a_1+a_3)}{a_1}, λ2=b3\lambda_2=b_3.

  • at A8=(−a3a1,1)A_8=\left(-\frac{a_3}{a_1},1\right): λ1=−a3(a1+a3)a1\lambda_1=-\frac{a_3(a_1+a_3)}{a_1}, λ2=−b2−b3\lambda_2=-b_2-b_3.

  • at A9=(−a3a1,−b3b2)A_9=\left(-\frac{a_3}{a_1},-\frac{b_3}{b_2}\right), λ1=−a3(a1+a3)a1\lambda_1=-\frac{a_3(a_1+a_3)}{a_1}, λ2=−b3(b2+b3)b2\lambda_2=-\frac{b_3(b_2+b_3)}{b_2}.

Lemma 3. Suppose that the active payoff entries in the two diagonal blocks are independent N(0,1)N(0,1) random variables, while the off-diagonal blocks are zero. Then:

  1. Each vertex A1,…,A4A_1,\ldots,A_4 always exists and is locally asymptotically stable with probability 1/41/4.

  2. Each open-edge candidate A5,…,A8A_5,\ldots,A_8 exists with probability 1/21/2 and exists and is locally asymptotically stable with probability 1/81/8.

  3. The interior candidate A9A_9 exists with probability 1/41/4. It exists and is locally asymptotically stable with probability 1/161/16, and exists and is a saddle with probability 1/81/8.

Avoidance games

Avoidance games retain only payoffs within population 1: all blocks of BB except the upper-left block are zero, leading to a2=b1=b2=b3=0a_2=b_1=b_2=b_3=0 and a3=β12−β22a_3=\beta_{12}-\beta_{22}:

(β11β1200β21β220000000000)(12)\begin{pmatrix} \beta_{11} & \beta_{12} & 0 & 0\\ \beta_{21} & \beta_{22} & 0 & 0\\ 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 \tag*{(12)} \end{pmatrix}

The resulting dynamics is

x˙=x(1−x)(a1x+a3),\dot{x}=x(1-x)(a_1x+a_3),
y˙=0.(13)\dot{y}=0. \tag*{(13)}

Thus, population 2 retains its initial strategy frequency, y(t)=y(0)y(t)=y(0), while population 1 follows a scalar replicator equation. The scalar equilibrium candidates are

A1=0,A2=1,A3=−a3a1,A_1=0,\qquad A_2=1,\qquad A_3=-\frac{a_3}{a_1},

where A3A_3 is admissible only when it lies in (0,1)(0,1).

Each admissible scalar equilibrium AiA_i corresponds to a continuous family of equilibria in the full system:

{(Ai,y):0≤y≤1}.\{(A_i,y):0\le y\le1\}.

If AiA_i attracts nearby scalar states, trajectories converge to (Ai,y(0))(A_i,y(0)). We call this transverse attraction: different initial yy coordinates lead to different limiting points, so no single point attracts a full two-dimensional neighbourhood.

Lemma 4. In the avoidance game, suppose that the four active payoff entries β11,β12,β21,β22\beta_{11},\beta_{12},\beta_{21},\beta_{22} are independent N(0,1)N(0,1) random variables. The following statements hold almost surely, with scalar local asymptotic stability understood relative to [0,1][0,1] for the xx equation in (17).

  1. A1=0A_1=0 and A2=1A_2=1 always exist. They are stable if and only if a3<0a_3<0 and a1+a3>0a_1+a_3>0, respectively, each with probability 1/21/2.

  1. A3=−a3/a1A_3=-a_3/a_1 exists in (0,1)(0,1) if and only if a3(a1+a3)<0a_3(a_1+a_3)<0, with probability 1/21/2. Given its existence, A3A_3 is stable if and only if a3>0a_3>0. Thus it exists and is stable if and only if a3>0>a1+a3a_3>0>a_1+a_3, with joint probability $1/4; its conditional stability probability given existence is 1/21/2.

  1. Scalar local asymptotic stability of AiA_i is equivalent to transverse attraction of its corresponding equilibrium family. Every point (Ai,y∗)(A_i,y^*) in an attracting family is Lyapunov stable relative to [0,1]2[0,1]^2. No individual point (Ai,y∗)(A_i,y^*) is locally asymptotically stable in the full two-dimensional system, because it has arbitrarily close distinct equilibria in the same family.

Three-population analytical results

We next examine how the equilibrium statistics extend to three-population games. Here we return to the generic model, retain all within- and between-population payoff blocks, and consider independent payoff sampling. Related deterministic three-population models have been studied in [4, 5, 6]. The state space is now a cube. Its vertices, open edges, open faces, and interior correspond to zero, one, two, and three populations maintaining both strategies, respectively. This classification extends the vertex-edge-interior description of the two-population state square. We first identify the candidates and their eigenvalues, and then give their existence and stability probabilities.

The full Jacobian is

Jij(u)=δij(1−2ui)fi(u)+ui(1−ui)Cij,(14)J_{ij}(\mathbf{u})=\delta_{ij}(1-2u_i)f_i(\mathbf{u})+u_i(1-u_i)C_{ij}, \tag*{(14)}

where δij=1\delta_{ij}=1 when i=ji=j and 0 otherwise. Equilibrium points and their eigenvalues are listed below, with i,j,k∈{0,1}i,j,k\in\{0,1\}. Every starred coordinate is determined separately for its candidate and must lie strictly in (0,1)(0,1). The formulas assume that the relevant principal submatrices of CC are nonsingular, as holds almost surely under independent Gaussian payoff sampling.

(i) Vertices. The eight points u∗=(i,j,k)\mathbf{u}^*=(i,j,k) always exist. Their eigenvalues are

λ1=(1−2i)f1(i,j,k),\lambda_1=(1-2i)f_1(i,j,k),
λ2=(1−2j)f2(i,j,k),\lambda_2=(1-2j)f_2(i,j,k),
λ3=(1−2k)f3(i,j,k).\lambda_3=(1-2k)f_3(i,j,k).

(ii) Open-edge points. The twelve candidates have the forms (x∗,i,j)(x^*,i,j), (i,y∗,j)(i,y^*,j), and (i,j,z∗)(i,j,z^*). For u∗=(x∗,i,j)\mathbf{u}^*=(x^*,i,j),

x∗=−a2i+a3j+a4a1,x^*=-\frac{a_2i+a_3j+a_4}{a_1},

and the point exists on its open edge if and only if 0<x∗<10<x^*<1. Its eigenvalues are λ1=a1x∗(1−x∗)\lambda_1=a_1x^*(1-x^*), λ2=(1−2i)f2(x∗,i,j)\lambda_2=(1-2i)f_2(x^*,i,j), and λ3=(1−2j)f3(x∗,i,j)\lambda_3=(1-2j)f_3(x^*,i,j).

EquilibriumExistence: P(Ei)\mathbb{P}(E_i)Existence and stability: P(Ei∩Si)\mathbb{P}(E_i \cap S_i)
TypePointTheorySimulationTheorySimulation
VertexA1=(0,0)A_1=(0,0)11.00001/4=0.25001/4=0.25000.2500
A2=(0,1)A_2=(0,1)11.00001/4=0.25001/4=0.25000.2497
A3=(1,0)A_3=(1,0)11.00001/4=0.25001/4=0.25000.2494
A4=(1,1)A_4=(1,1)11.00001/4=0.25001/4=0.25000.2510
Open edgeA5 (x=0)A_5\ (x=0)1/3≈0.33331/3 \approx0.33330.33331/12≈0.08331/12 \approx0.08330.0833
A6 (x=1)A_6\ (x=1)1/3≈0.33331/3 \approx0.33330.33321/12≈0.08331/12 \approx0.08330.0834
A7 (y=0)A_7\ (y=0)1/3≈0.33331/3 \approx0.33330.33321/12≈0.08331/12 \approx0.08330.0831
A8 (y=1)A_8\ (y=1)1/3≈0.33331/3 \approx0.33330.33281/12≈0.08331/12 \approx0.08330.0835
InteriorA9=(x∗,y∗)A_9=(x^*,y^*)pint≈0.2163p_{\mathrm{int}}\approx0.21630.2163pint/4≈0.0541p_{\mathrm{int}}/4\approx0.05410.0541

Table 1. Theoretical and empirical equilibrium probabilities in generic two-population games under microscopic uncertainty. Notes: Analytical values follow from Theorem 1.

The other two families follow by the corresponding permutations of coordinates and payoff-difference equations.

(iii) Open-face points. The six candidates have the forms (x∗,y∗,k)(x^*,y^*,k), (x∗,k,z∗)(x^*,k,z^*), and (k,y∗,z∗)(k,y^*,z^*). For u∗=(x∗,y∗,k)u^*=(x^*,y^*,k),

\begin{pmatrix}x^*\\y^*\end{pmatrix} =- \begin{pmatrix}a_1&a_2\\b_1&b_2\end{pmatrix}^{-1} \begin{pmatrix}a_3k+a_4\\b_3k+b_4\end{pmatrix}. \tag*{(15)}\end{pmatrix}

It exists on the open face if and only if 0<x∗,y∗<10<x^*,y^*<1. With Dxy=diag⁡(x∗(1−x∗),y∗(1−y∗))D_{xy}=\operatorname{diag}(x^*(1-x^*),y^*(1-y^*)) and

K=Dxy(a1a2b1b2),K=D_{xy}\begin{pmatrix}a_1&a_2\\b_1&b_2\end{pmatrix},

the eigenvalues are

λ1,2=tr⁡K±(tr⁡K)2−4det⁡K2,\lambda_{1,2}=\frac{\operatorname{tr}K\pm\sqrt{(\operatorname{tr}K)^2-4\det K}}{2},
λ3=(1−2k)f3(x∗,y∗,k).\lambda_3=(1-2k)f_3(x^*,y^*,k).

The other two families again follow by corresponding permutations.

(iv) Interior point. The candidate u∗=(x∗,y∗,z∗)=−C−1du^*=(x^*,y^*,z^*)=-C^{-1}d exists if and only if u∗∈(0,1)3u^*\in(0,1)^3. Set D=diag⁡(ur∗(1−ur∗):r=1,2,3)D=\operatorname{diag}(u_r^*(1-u_r^*):r=1,2,3). Its three eigenvalues are the roots, counted with multiplicity, of det⁡(λI3−DC)=0\det(\lambda I_3-DC)=0.

These four classes give 27 candidate locations and exhaust all equilibria under the stated nonsingularity assumption; only the eight vertices are guaranteed to exist. When no eigenvalue has zero real part, a candidate is locally asymptotically stable relative to [0,1]3[0,1]^3 if and only if all three eigenvalues have negative real parts. At a boundary equilibrium, this condition includes perturbations away from the boundary, not just perturbations along it. The corresponding derivations are provided in the Supplementary Material.

For a specified candidate with kk free coordinates, let EE denote existence and SS denote local asymptotic stability relative to [0,1]3[0,1]^3. Write pk=P(E)p_k=\mathbb{P}(E) and qk=P(E∩S)q_k=\mathbb{P}(E\cap S) for the per-candidate probabilities; k=0,1,2,3k=0,1,2,3 corresponds to a vertex, open-edge point, open-face point, or interior point, respectively.

Lemma 5 (Probabilities under microscopic uncertainty). Suppose that all 36 payoff entries βij\beta_{ij} are independent N(0,1)\mathcal{N}(0,1) variables. For each specified candidate, the following statements hold.

(i) Each vertex always exists and is locally asymptotically stable with probability 1/81/8:

p0=1,q0=18.p_0=1,\qquad q_0=\frac{1}{8}.

(ii) Each open-edge point exists with probability p1p_1 and exists and is locally asymptotically stable with probability q1q_1, where

p1=1πarccos⁡(23)≈0.2677,p_1=\frac{1}{\pi}\arccos\left(\frac{2}{3}\right)\approx0.2677,
q1=18πarccos⁡(23)≈0.0335.q_1=\frac{1}{8\pi}\arccos\left(\frac{2}{3}\right)\approx0.0335.

(iii) Each open-face point exists with probability p2p_2 and exists and is locally asymptotically stable with probability q2q_2, where

p2=2πarctan⁡(126)≈0.1282,p_2=\frac{2}{\pi}\arctan\left(\frac{1}{2\sqrt{6}}\right)\approx0.1282,
q2=14πarctan⁡(126)≈0.0160.q_2=\frac{1}{4\pi}\arctan\left(\frac{1}{2\sqrt{6}}\right)\approx0.0160.

(iv) The interior point exists with probability

p3=2∫0∞ϕ(t)[2Φ(t3)−1]3 dt≈0.0922.(16)p_3=2\int_0^\infty\phi(t)\left[2\Phi\left(\frac{t}{\sqrt{3}}\right)-1\right]^3\,\mathrm{d}t\approx0.0922. \tag*{(16)}

where ϕ\phi and Φ\Phi are the standard normal density and distribution function.

The proof is given in the Supplementary Material. The probabilities q0,q1,q2q_0,q_1,q_2 are joint, not conditional, probabilities; the corresponding conditional stability probabilities equal 1/81/8. The lemma does not give an analytical value for the interior-stability probability q3q_3, which is assessed numerically.

Numerical results

The numerical results support the analytical distinction between equilibrium existence and attraction, and show how it relates to multistability and initial-condition dependence. The comparison covers generic two-population games, the three interaction structures, and generic three-population games. Additional coefficient-sampling results and model-specific dynamical examples are provided in the Supplementary Material.

Generic two-population games

Equilibrium statistics depend on whether uncertainty is specified at the payoff or coefficient level. Under independent payoff sampling, the interior equilibrium exists with frequency 0.2163 in Table I, consistent with the analytical value pint≈0.21635p_{\mathrm{int}}\approx0.21635. Independent coefficient sampling instead gives 1/12≈0.08331/12\approx0.0833, supported by the supplementary simulations. Likewise, the four open-edge candidates share existence probability 1/31/3 under payoff sampling, whereas coefficient sampling produces two pairs with distinct probabilities. Thus, the two uncertainty descriptions define different evolutionary benchmarks.

Under payoff sampling, a specified equilibrium is locally asymptotically stable with probability 1/41/4 conditional on its existence, regardless of whether it lies at a vertex, on an open edge, or in the interior. The joint existence-and-stability probabilities are therefore 1/41/4, 1/121/12, and pint/4≈0.0541p_{\mathrm{int}}/4\approx0.0541, respectively, consistent with Table I. Equilibrium existence is consequently substantially more common than attraction.

Notes: In (b), AiA_i denotes a scalar equilibrium from Lemma 4; each corresponding family is {(Ai,y):0≤y≤1}\{(A_i,y):0\le y\le1\}. For A3∈(0,1)A_3\in(0,1), the family includes boundary endpoints at y=0,1y=0,1. The generic-game estimates use a separate realization set from Table I; pintp_{\mathrm{int}} is defined in Theorem 1(2).

Notes: Theory is given by Lemma 5; the interior-existence value evaluates its exact integral. Simulation entries are mean ±\pm between-candidate SD, omitted for the single interior candidate. The dash denotes an analytical interior-stability probability not established here.

Multistability occurs in more than one third of payoff-sampled games, although a single stable equilibrium is the most common outcome. This concerns the coexistence of attractors within the same game, which is not determined by the probabilities for individual candidates alone. In Table II, 34.62% of games have at least two locally asymptotically stable equilibria, 64.65% have exactly one, and 0.74% have none. The mean counts are 5.5489 existing equilibria and 1.3874 stable equilibria. Five equilibria is the most frequent total count (31.03%), illustrating that most existing equilibria need not be attractors.

Empirical probabilityNumber of equilibria, nn
0123456789
P^(N=n)\widehat{\mathbb{P}}(N=n)00000.22220.31030.27030.11840.05110.0277
P^(Ns=n)\widehat{\mathbb{P}}(N_{\mathrm{s}}=n)0.00740.64650.30440.03490.006900000

Table II. Empirical Equilibrium-Count Distributions in Generic Two-Population Games Under Microscopic Uncertainty. Notes: Both marginal distributions use the same payoff draws as Table I. Zero denotes no observed occurrences; N<4N<4 is impossible because all four vertices always exist. Rounding may cause row sums to differ slightly from one.

Coexisting stable equilibria make the long-term outcome depend on the initial strategy frequencies even when payoffs are fixed. Panels (d)–(f) of Figure 1 display two, three, and four stable equilibria, respectively. Each attracts a different set of initial states, forming its basin of attraction. In particular, panel (d) admits both a stable vertex and a stable interior equilibrium, so initial conditions can select between pure-strategy and mixed-strategy outcomes. Panels (b) and (c) instead have a unique stable equilibrium on the boundary and in the interior, respectively.

Representative phase portraits and initial-condition dependence

Figure 1. Representative phase portraits and initial-condition dependence. Panels (a)–(f) contain Ns=0,1,1,2,3,4N_{\mathrm{s}}=0,1,1,2,3,4 locally asymptotically stable equilibria, respectively. Panel (a) has an attracting boundary heteroclinic cycle, not a stable equilibrium or an interior limit cycle. The unique stable equilibrium is on the boundary in (b) and in the interior in (c); (d) exhibits coexistence of a stable vertex and a stable interior equilibrium. Black filled circles denote stable equilibria, open circles denote saddles, and crosses denote unstable sources. No neutrally stable equilibria occur in these samples. Black arrows indicate the vector field. Pale colours in (b)–(f) identify numerically estimated basins of attraction within each panel; darker curves illustrate trajectories from selected initial states, marked by small squares. The examples were selected from payoff-sampled games to illustrate distinct dynamical regimes; their selection does not represent the relative frequencies of those regimes. Writing θ=(a1,a2,a3,b1,b2,b3)\theta=(a_1,a_2,a_3,b_1,b_2,b_3), the parameter sets are: (a) θ=(1.5632,−2.2136,0.3277,2.2314,0.6897,−1.4731)\theta=(1.5632,-2.2136,0.3277,2.2314,0.6897,-1.4731); (b) θ=(1.5693,2.2535,−6.6246,−3.6760,−4.3716,2.1825)\theta=(1.5693,2.2535,-6.6246,-3.6760,-4.3716,2.1825); (c) θ=(−0.9148,4.0690,−1.5678,−2.8839,−2.0114,2.4812)\theta=(-0.9148,4.0690,-1.5678,-2.8839,-2.0114,2.4812); (d) θ=(−2.1943,−6.5550,2.5883,0.5775,0.7098,−0.4233)\theta=(-2.1943,-6.5550,2.5883,0.5775,0.7098,-0.4233); (e) θ=(2.3992,2.2562,−3.3708,2.6826,2.5720,−1.3808)\theta=(2.3992,2.2562,-3.3708,2.6826,2.5720,-1.3808); (f) θ=(4.6340,0.2002,−2.5835,−0.0118,1.5367,−0.7956)\theta=(4.6340,0.2002,-2.5835,-0.0118,1.5367,-0.7956). The listed parameter values are rounded to four decimal places.

The absence of a stable equilibrium does not preclude an attracting dynamical outcome. In Figure 1(a), trajectories approach a boundary heteroclinic cycle: they repeatedly pass near successive boundary saddle points rather than converge to a single equilibrium. The interior equilibrium is unstable, so this recurrent switching is distinct from attraction to a stable coexistence point.

Effects of interaction structure

Interaction structure also changes the availability of partially mixed outcomes, in which one population maintains both strategies while the other uses a pure strategy. For each specified open edge, the equilibrium-existence probability is zero in diversity games, 1/31/3 in generic games, and 1/21/2 in exclusive games. The corresponding joint existence-and-stability probabilities are zero, 1/121/12, and 1/81/8, consistent with Table III.

ModelEntity typeExistence: P(E)\mathbb{P}(E)Joint: P(E∩S)\mathbb{P}(E \cap S) or P(E∩T)\mathbb{P}(E \cap T)
TheorySimulationTheorySimulation
* (a) Isolated equilibria: SS denotes local asymptotic stability
GenericVertices11.00001/40.2499
Open-edge points1/30.33311/120.0833
Interior pointpintp_{\mathrm{int}}0.2160pint/4p_{\mathrm{int}}/40.0539
DiversityVertices11.00001/40.2501
Open-edge points00.000000.0000
Interior point1/40.249700.0000
ExclusiveVertices11.00001/40.2500
Open-edge points1/20.50071/80.1251
Interior point1/40.25081/160.0629
* (b) Avoidance: TT denotes transverse attraction of the corresponding equilibrium family
AvoidanceFamilies at x=A1,A2x=A_1,A_211.00001/20.5001
Family at x=A3x=A_31/20.50101/40.2504

Table III. Effect of interaction structure on equilibrium existence and stability

With the payoff-sampling assumption held fixed, interaction structure changes stability even when the probability of interior coexistence remains the same. Diversity and exclusive games both have interior-existence probability 1/41/4, supported by the frequencies 0.2497 and 0.2508 in Table III. Yet diversity games admit no asymptotically stable interior equilibrium, whereas exclusive games have joint interior-stability probability 1/161/16, with simulated frequency 0.0629. Interior coexistence alone therefore does not identify an attracting mixed-strategy outcome. Detailed candidate-wise probabilities and equilibrium-count distributions are reported in Tables S3 and S4 of the Supplementary Material.

Avoidance changes the attracting object from an isolated equilibrium to a continuous family. The family associated with the interior scalar equilibrium A3∈(0,1)A_3 \in(0,1) exists with frequency 0.5010 and exists and is transversely attracting with frequency 0.2504, consistent with the probabilities 1/21/2 and 1/41/4 in Lemma 4. Boundary families have mean transverse-attraction frequency 0.5001. None of these families contains an individually locally asymptotically stable point of the full two-dimensional system. Detailed candidate-wise results and equilibrium-count distributions are provided in the Supplementary Material.

Three-population numerical results

Within the three-population model, a specified equilibrium candidate is less likely to exist when it requires more populations to maintain both strategies. In Table IV, the per-candidate existence frequencies decrease from 1 at vertices to 0.2678 on open edges, 0.1284 on open faces, and 0.0923 in the interior, consistent with Lemma 5. Interior coexistence is substantially more frequent than stable interior coexistence: the joint interior-stability frequency is only 0.0096. For each boundary candidate, the analytical joint stability probability is one eighth of its existence probability. Equivalently, its conditional probability of local asymptotic stability, given existence, is 1/8. No corresponding analytical relation is established here for the interior candidate.

TypeCandidatesExistenceExistence and stability
TheorySimulationTheorySimulation
Vertex81.00001.0000±0.00001.0000\pm0.00000.12500.1251±0.00040.1251\pm0.0004
Open edge120.26770.2678±0.00030.2678\pm0.00030.03350.0335±0.00020.0335\pm0.0002
Open face60.12820.1284±0.00040.1284\pm0.00040.01600.0161±0.00010.0161\pm0.0001
Interior10.09220.0923—0.0096

Table IV. Theoretical and empirical equilibrium probabilities in generic three-population games under microscopic uncertainty

Under the independent Gaussian payoff model with equal interaction weights, multistability is more frequent in the three-population sample than in the two-population sample, despite the lower probability of a fully mixed equilibrium. Games with at least two locally asymptotically stable equilibria account for 41.42% of Table V, compared with 34.62% in Table II. A single stable equilibrium remains the most common outcome (57.28%), whereas 1.30% of games have none. These count distributions characterize multistability, not global convergence: games without a stable equilibrium may still possess non-equilibrium attractors. Additional three-dimensional trajectories are provided in the Supplementary Material.

(a) Total number of equilibria
nn891011121314151617
P^(N=n)\widehat{\mathbb{P}}(N=n)0.05820.10330.15430.15690.14990.11730.08730.06340.04120.0263
nn18192021222324252627
P^(N=n)\widehat{\mathbb{P}}(N=n)0.01780.00980.00590.00370.00190.00130.00070.00030.00020.0001
(b) Number of locally asymptotically stable equilibria
nn0123456789–27
P^(Ns=n)\widehat{\mathbb{P}}(N_s=n)0.01300.57280.32850.06740.01560.00210.00050.0001<0.0001< 0.00010

Table V. Empirical Equilibrium-Count Distributions in Generic Three-Population Games Under Microscopic Uncertainty. Notes: The same experiment underlies Table IV. Panel (a) splits one distribution across two ranges. The last entry of (b) aggregates 9≤Ns≤279 \leq N_s \leq27; the positive entry <0.0001< 0.0001 corresponds to 17 games.

Conclusion

We establish probabilistic benchmarks for equilibrium existence and stability in two-population games, with a generic three-population extension. Our results show that equilibrium statistics depend both on interaction structure and on how uncertainty is represented. In particular, independent sampling of payoff entries differs from independent sampling of replicator coefficients, because shared payoff entries induce dependence among the coefficients. The structured two-population games illustrate why equilibrium existence and attraction must be distinguished. Diversity and exclusive games have the same probability of an interior equilibrium, but differ in whether stable interior coexistence is possible. In avoidance games, equilibria may form continuous families that attract trajectories transversely, even though individual points on those families are not asymptotically stable. This distinction is essential to stability analyses of asymmetric games, where equilibrium sets need not be isolated [29, 30].

Under the Gaussian payoff ensembles considered here, fully mixed equilibria occur less frequently in three-population games than in two-population games, whereas multistability is more frequent. These findings are benchmarks for the specified uncertainty models, rather than universal properties of evolutionary games. Payoffs remain fixed during each trajectory; thus, the probabilities describe variation across games, not stochastic forcing or changes in the game during evolution, which are studied separately [38, 39, 40]. Future work could investigate how the results change under correlated or empirically calibrated payoffs, and develop analytical stability results for interior equilibria in three-population games.

Acknowledgment

Z. Song and T.A.H. were supported by EPSRC (grant no. EP/Y00857X/1). M.H.D is supported by EPSRC (grant no. EP/Y008561/1).

References

  1. [1]A. J. Lotka, “Elements of physical biology,” Science Progress in the Twentieth Century (1919-1933), vol. 21, no. 82, pp. 341–343, 1926.
  2. [2]V. Volterra, “Fluctuations in the abundance of a species considered mathematically,” Nature, vol. 118, no. 2972, pp. 558–560, 1926.DOI
  3. [3]M. K. A. Gavina, T. Tahara, K.-i. Tainaka, H. Ito, S. Morita, G. Ichinose, T. Okabe, T. Togashi, T. Nagatani, and J. Yoshimura, “Multi-species coexistence in lotka-volterra competitive systems with crowding effects,” Scientific reports, vol. 8, no. 1, p. 1 198, 2018.DOI
  4. [4]M. Zhang and Z. Kong, “A tripartite evolutionary game model of emergency supplies joint reserve among the government, enterprise and society,” Computers & Industrial Engineering, vol. 169, p. 108132, 2022.DOI
  5. [5]X. Zhao, Y. Bai, L. Ding, and L. Wang, “Tripartite evolutionary game theory approach for low-carbon power grid technology cooperation with government intervention,” IEEE Access, vol. 8, pp. 47 357–47 369, 2020.DOI
  6. [6]K. Zhou, Q. Wang, and J. Tang, “Evolutionary game analysis of environmental pollution control under the government regulation,” Scientific Reports, vol. 12, no. 1, p. 474, 2022.DOI
  7. [7]S. Encarnação, F. P. Santos, F. C. Santos, V. Blass, J. M. Pacheco, and J. Portugali, “Paradigm shifts and the interplay between state, business and civil sectors,” Royal Society open science, vol. 3, no. 12, p. 160753, 2016.
  8. [8]A. Alalawi, P. Bova, T. Cimpeanu, A. Di Stefano, M. H. Duong, E. F. Domingos, T. A. Han, M. Krellner, N. B. Ogbo, S. T. Powers et al., “Trust as regulation? discerning users are vital to build trust and effective ai regulation,” Applied Mathematics and Computation, vol. 508, p. 129627, 2026.
  9. [9]N. Balabanova, A. Bashir, P. Bova, A. Buscemi, T. Cimpeanu, H. C. da Fonseca, A. Di Stefano, M. H. Duong, E. F. Domingos, A. Fernandes et al., “Media and responsible ai governance: a game-theoretic and llm analysis,” Philosophical Transactions of the Royal Society A, 2026, in press.arxiv.org/abs/2503.09858
  10. [10]H. C. Da Fonseca, A. Fernandes, Z. Song, T. Cimpeanu, N. Balabanova, A. Bashir, P. Bova, A. Buscemi, A. Di Stefano, M. H. Duong et al., “Can media act as a soft regulator of safe ai development? a game theoretical analysis,” in Artificial Life Conference Proceedings 37, vol. 2025, no. 1. MIT Press, 2025, p. 90.arxiv.org/abs/2509.02650
  11. [11]P. D. Taylor, “Evolutionarily stable strategies with two types of player,” Journal of applied probability, vol. 16, no. 1, pp. 76–83, 1979.DOI
  12. [12]A. Liu, L. Wang, G. Chen, and X. Guan, “Heterogeneously networked evolutionary games with intergroup conflicts,” IEEE Transactions on Cybernetics, vol. 54, no. 10, pp. 5684–5695, 2024.DOI
  13. [13]V. Vargas-Panesso, N. Quijano, L. F. Giraldo, and J. Barreiro-Gomez, “Modeling strategic intercommunity connections in evolutionary games,” IEEE Transactions on Cybernetics, 2026.
  14. [14]T. Cimpeanu, A. Di Stefano, C. Perret, and T. A. Han, “Social diversity reduces the complexity and cost of fostering fairness,” Chaos, Solitons & Fractals, vol. 167, p. 113051, 2023.DOI
  15. [15]V. A. Vargas-Pérez, M. Chica, Z. Song, and T. A. Han, “Out-earning or outnumbering: Two routes to cooperation in the fixed-role trust game on structured populations,” Knowledge-Based Systems, p. 117282, 2026.
  16. [16]S. Tan, Y. Wang, Y. Chen, and Z. Wang, “Evolutionary dynamics of collective behavior selection and drift: Flocking, collapse, and oscillation,” IEEE transactions on cybernetics, vol. 47, no. 7, pp. 1694–1705, 2016.DOI
  17. [17]A. Yazidi, M. A. Pinto-Orellana, H. Hammer, P. Mirtaheri, and E. Herrera-Viedma, “Solving sensor identification problem without knowledge of the ground truth using replicator dynamics,” IEEE transactions on cybernetics, vol. 52, no. 1, pp. 16–24, 2020.DOI
  18. [18]M. H. Duong and T. A. Han, “Evolutionary dynamics with random payoff matrices,” Europhysics Letters, vol. 149, no. 3, p. 32001, 2025.arxiv.org/abs/2410.22010
  19. [19]D. Jia, I. Romić, L. Shi, Q. Su, C. Liu, J. Liu, P. Holme, X. Li, and Z. Wang, “Stochastic evolutionary agency and prosociality are inextricably linked in economic games,” Nature Human Behaviour, vol. 9, no. 12, pp. 2620–2631, 2025.
  20. [20]D. Jia, X. Dai, J. Xing, P. Tao, Y. Shi, and Z. Wang, “Asymmetric interaction preference induces cooperation in human-agent hybrid game,” Science China Information Sciences, vol. 68, 04 2025.DOI
  21. [21]H. Wei, J. Zhang, C. Zhang, and M. Cao, “Indirect reciprocity enhances collective cooperation on weighted networks,” IEEE Transactions on Cybernetics, 2025.DOI
  22. [22]H. Guo, Z. Song, M. Perc, X. Li, and Z. Wang, “Third-party intervention of cooperation in multilayer networks,” IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 53, no. 11, pp. 6646–6657, 2023.DOI
  23. [23]Z. Song, T. Cimpeanu, C. Shen, and T. A. Han, “Evolution of fairness in hybrid populations with specialised ai agents,” Philosophical Transactions of the Royal Society A, 2026, in press.arxiv.org/abs/2602.18498
  24. [24]C. Shen, Z. Song, X. Wang, L. Shi, M. Perc, Z. Wang, and J. Tanimoto, “Evolutionary dynamics of reputation-based voluntary prisoner’s dilemma games,” IEEE Transactions on Cybernetics, pp. 1–12, 2026.arxiv.org/abs/2601.16643
  25. [25]Z. Wang, Z. Song, C. Shen, and S. Hu, “Emergence of punishment in social dilemma with environmental feedback,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 10, 2023, pp. 11 708–11 716.
  26. [26]J. Bauer, M. Broom, and E. Alonso, “The stabilization of equilibria in evolutionary game dynamics through mutation: mutation limits in evolutionary games,” Proceedings Mathematical, Physical, and Engineering Sciences, vol. 475, no. 2231, p. 20190355, 2019.DOI
  27. [27]Z. Wang, C. Mu, S. Hu, C. Chu, and X. Li, “Modelling the dynamics of regret minimization in large agent populations: a master equation approach,” in Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, L. D. Raedt, Ed. International Joint Conferences on Artificial Intelligence Organization, 7 2022, pp. 534–540, main Track. [Online]. Available: https://doi.org/10.24963/ijcai.2022/76
  28. [28]T. Galla and J. D. Farmer, “Complex dynamics in learning complicated games,” Proceedings of the National Academy of Sciences, vol. 110, no. 4, pp. 1232–1236, 2013.DOI
  29. [29]S. Song, Q. Pan, and M. He, “Study on the interior equilibrium point of a special class of 2× 2× 2 asymmetric evolutionary games,” Royal Society Open Science, vol. 11, no. 7, p. 231960, 2024.DOI
  30. [30]S. Song, Q. Pan, X. Gao, and M. He, “Stability analysis of evolutionary dynamics of 2× 2× 2 asymmetric games,” Proceedings of the Royal Society A, vol. 480, no. 2282, p. 20230478, 2024.DOI
  31. [31]T. Galla, “Two-population replicator dynamics and number of nash equilibria in matrix games,” EPL (Europhysics Letters), vol. 78, no. 2, p. 20005, 2007.DOI
  32. [32]T. A. Han, A. Traulsen, and C. S. Gokhale, “On equilibrium properties of evolutionary multi-player games with random payoff matrices,” Theoretical Population Biology, vol. 81, no. 4, pp. 264–272, 2012.DOI
  33. [33]M. H. Duong and T. A. Han, “Statistics of the number of equilibria in random social dilemma evolutionary games with mutation,” The European Physical Journal B, vol. 94, no. 8, p. 171, 2021.DOI
  34. [34]M. H. Duong et al., “Random evolutionary games and random polynomials,” Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 481, no. 2319, 2025.arxiv.org/abs/2304.13831
  35. [35]C. S. Gokhale and A. Traulsen, “Evolutionary games in the multiverse,” Proceedings of the National Academy of Sciences, vol. 107, no. 12, pp. 5500–5504, 2010.DOI
  36. [36]M. H. Duong and T. A. Han, “Analysis of the expected density of internal equilibria in random evolutionary multi-player multi-strategy games,” Journal of Mathematical Biology, vol. 73, no. 6, pp. 1727–1760, 2016.pubmed.ncbi.nlm.nih.gov/27107868
  37. [37]M. H. Duong, H. M. Tran, and T. A. Han, “On the distribution of the number of internal equilibria in random evolutionary games,” Journal of Mathematical Biology, vol. 78, no. 1, p. 331, 2018.DOI
  38. [38]M. Hong Duong, G. A. Pavliotis, and J. Tugaut, “Multi-species mckean–vlasov dynamics in non-convex landscapes,” Nonlinearity, vol. 39, no. 10, p. 105003, 2026.arxiv.org/abs/2507.07617
  39. [39]T.-J. Feng, J. Mei, C. Li, X.-D. Zheng, S. Lessard, and Y. Tao, “Stochastic evolutionary stability in matrix games with random payoffs,” Physical Review E, vol. 105, no. 3, p. 034303, 2022.DOI
  40. [40]M. Feng, B. Pi, L.-J. Deng, and J. Kurths, “An evolutionary game with the game transitions based on the markov process,” IEEE transactions on systems, man, and cybernetics: systems, vol. 54, no. 1, pp. 609–621, 2023.DOI
  41. [41]J.-M. Azaïs and M. Wschebor, Level sets and extrema of random processes and fields. John Wiley & Sons, 2009.

Paper details

vlnc:W_sz8hc24e
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents