betaFeedback

WorksPhysics and Astronomy

Stochastic Diversity Stabilization in Nonlinear Competitive Dynamics: A Statistical Mechanics Approach to Fluctuation Suppression and Finite-Size Scaling

Abstract

Stochastic forcing can prevent diversity collapse in competitive dynamical systems, yet the quantitative conditions and the dynamical mechanism responsible remain poorly characterized. We address both questions in a class of nonlinear stochastic replicator dynamics on the N-dimensional probability simplex, parameterized by a nonlinearity exponent gamma and additive Gaussian noise of intensity sigma. Under the baseline interaction, the minimum noise required to sustain diversity decays as a power law, sigma_c ~ N^(-alpha); replacing the baseline interaction with a superlinear variant shifts alpha from about 1.1 to about 2.0, in agreement with a linear-stability prediction alpha = (gamma+2)/2, identifying alpha as a model-specific quantity rather than a universal index. Scanning the (gamma, sigma) plane reveals a smooth dynamical crossover: the fluctuation susceptibility chi_H = N Var(H) peaks at gamma ~ 0.7 (chi_H ~ 0.6) and is suppressed at high gamma, with no divergence near gamma ~ 1.4; matched finite-size analysis finds no systematic divergence of the peak with N, ruling out a genuine second-order phase transition. The entropy relaxation time tau_relax(gamma) remains approximately constant (about 2.9 time units) across the full range of gamma, even as chi_H decreases sixfold; this decoupling between relaxation timescale and fluctuation amplitude implies that suppression arises from convergence of diverse pathways to similar endpoints rather than from accelerated relaxation along a single channel. These findings identify nonlinear fluctuation suppression via pathway multiplicity – distinct from both noise-induced transitions and equilibrium criticality – as the operative stabilization pattern in this class of nonequilibrium stochastic systems.

Introduction

Background

In competitive systems comprising multiple interacting species or agents, deterministic dynamics generically drive the system toward a state of reduced diversity: one or a few dominant components come to monopolize the available “resource” space while the rest are suppressed. Such diversity-collapse dynamics arise across a broad class of nonlinear stochastic systems constrained to a probability simplex, including species extinction in ecological communities [1, 2], fixation in finite-population evolutionary games [3], winner-take-all convergence in multi-agent learning and market-share competition, and weight concentration in neural network ensembles. Understanding the conditions under which stochastic forcing can counteract this ordering tendency is therefore a question of broad relevance to nonlinear science [4, 5]. This phenomenon, commonly referred to as diversity collapse or competitive exclusion, is a robust prediction of mean-field replicator dynamics [1] and related evolutionary game-theoretic frameworks. On the NN-dimensional simplex ΔN−1\Delta^{N-1}, the replicator equation assigns a positive feedback to above-average components and a negative feedback to below-average ones, creating a self-reinforcing ordering tendency that terminates in a corner state corresponding to the extinction of all but a single type. Such collapse dynamics are well-documented in theoretical ecology, population genetics, and multi-agent learning, and represent a fundamental tension with the empirical observation that diversity is remarkably persistent across biological and social systems.

A natural candidate mechanism for resisting this ordering tendency is stochasticity. Random fluctuations, whether arising from finite population size, environmental variability, or intrinsic demographic noise, perturb the deterministic flow and can, under appropriate conditions, prevent the system from becoming trapped at a corner of the simplex. This stabilizing role of noise has been anticipated on theoretical grounds in the context of noise-induced transitions [4, 5], supported by population-level models in evolutionary game theory [1, 3], and demonstrated numerically in ecological models where demographic stochasticity sustains species coexistence that deterministic dynamics would otherwise eliminate [2]. In the statistical physics language, the question can be reframed as whether additive stochastic forcing is capable of shifting the system away from the ordered (low-entropy) absorbing-like state into a disordered (high-entropy) stationary regime, in analogy with noise-induced nonequilibrium phase transitions [5].

Despite this body of work, two important aspects of stochastic diversity stabilization remain insufficiently understood. First, the quantitative parameter dependence of the stabilizing effect has not been systematically characterized: specifically, how the noise–entropy response curve evolves with system size NN, what role the deterministic dynamics play at large NN, and whether any observed size-dependent pattern is robust to changes in the microscopic interaction rule. Second, the dynamical mechanism by which stronger nonlinearity in the competitive interaction suppresses collective fluctuations has not been identified. Addressing these two gaps is the primary aim of the present work, which approaches diversity stabilization as a problem in nonlinear stochastic dynamics on constrained state spaces rather than in terms of equilibrium phase transitions.

Present work

To address these questions, we study a class of stochastic replicator-like systems defined on the NN-dimensional probability simplex, in which the competitive interaction is parameterized by a nonlinearity exponent γ\gamma and the dynamics are driven by additive Gaussian noise of strength σ\sigma. Three main contributions are reported.

(i) Size-dependent noise–entropy response and deterministic self-stabilization. We systematically measure the noise–entropy response curve ⟨H⟩(σ)\langle H\rangle(\sigma) across system sizes N∈{5,8,10,15,20,30,50}N \in\{5, 8, 10, 15, 20, 30, 50\} and reveal a qualitative transition in stabilization mechanism: for small NN the entropy at σ=0\sigma= 0 is well below the high-diversity regime, so external noise is the primary stabilizing agent; for large NN the deterministic dynamics alone sustain a high-entropy stationary state, and noise plays only a secondary modulating role. In the small-NN regime (N≤15N \le15) the noise intensity at the steepest response point decreases approximately as σc∼N−α\sigma_c \sim N^{-\alpha}; however, this approximate scaling breaks down at large NN, and the exponent α\alpha shifts from ≈1\approx1 to ≈2\approx2 as the nonlinearity γ\gamma increases, in agreement with the analytical prediction α=(γ+2)/2\alpha= (\gamma+ 2)/2 (§ 3.2–§ 3.3, Appendix § 9), confirming that it reflects the specific balance of the baseline replicator interaction rather than a universal law.

(ii) Nonlinear phase structure and fluctuation analysis. We construct a dynamical phase diagram in the (γ,σ)(\gamma,\sigma) parameter plane and show that the system undergoes a smooth crossover rather than a sharp phase transition as γ\gamma increases. To test for criticality, we compute the fluctuation susceptibility χH=NVar⁡(H)\chi_H = N \operatorname{Var}(H), where HH is the normalized Shannon entropy, across the full parameter space. The susceptibility peaks at γ≈0.7\gamma\approx0.7 (χH≈0.6\chi_H \approx0.6) and is suppressed at high γ\gamma; no divergence is observed near γ≈1.4\gamma\approx1.4, where a naive visual inspection of the phase diagram might suggest a structural change. Matched finite-size analysis at N∈{10,20,40}N \in\{10, 20, 40\} finds no systematic divergence of the peak susceptibility with NN, in direct contradiction to expectations near a genuine second-order phase transition [6, 7]. These findings establish that the high-γ\gamma regime is not a critical state but an entropy-saturated nonlinear phase in which collective fluctuations are suppressed by the nonlinear interaction rather than amplified.

(iii) Relaxation dynamics and mechanistic interpretation. We measure the macroscopic entropy relaxation time τrelax\tau_{\mathrm{relax}} as a function of γ\gamma using a stationary-distribution criterion appropriate for noisy systems [8, 9]. The resulting τrelax(γ)\tau_{\mathrm{relax}}(\gamma) curve is approximately flat, with mean values between 2.5 and 3.2 time units across the full γ\gamma range and no systematic dependence on nonlinearity strength. This insensitivity of the relaxation timescale stands in sharp contrast to the sixfold decrease of χH\chi_H over the same γ\gamma range, revealing a decoupling between the rate at which individual trajectories reach stationarity and the degree of ensemble-level fluctuation suppression. The decoupling implies that strong nonlinear coupling suppresses collective fluctuations not by accelerating convergence toward a single deep attractor, but by routing different trajectories through diverse microscopic pathways that nonetheless converge to similar macroscopic endpoints—a mechanism we term pathway-multiplicity-driven ensemble decorrelation.

The remainder of the paper is organized as follows. Section 2 defines the model and all observables. Sections 3–5 present the three groups of numerical results. Section 6 discusses the findings in relation to existing theoretical frameworks, and Section 7 summarizes the conclusions.

Model and Methods

Stochastic dynamics on the simplex

We consider a system of NN interacting components whose states are represented by the vector x=(x1,x2,…,xN)\mathbf{x} = (x_1,x_2,\ldots,x_N) constrained to the (N−1)(N-1)-dimensional probability simplex,

ΔN−1={x∈RN | xi≥0 ∀i,∑i=1Nxi=1}.(1)\Delta^{N-1} = \left\{ \mathbf{x} \in\mathbb{R}^{N} \,\middle|\, x_i \ge0\ \forall i,\quad\sum_{i=1}^{N} x_i = 1 \right\}. \tag*{(1)}

Each component xix_i can be interpreted as the relative abundance of type ii in a population, the market share of agent ii, or more generally the probability weight assigned to option ii in a finite mixture. The simplex geometry ensures that competition is inherently zero-sum: an increase in xix_i must be offset by a decrease elsewhere.

Baseline interaction

The deterministic driving force is taken from the class of replicator-like dynamics [1]. In the baseline model, the drift term for component ii is

fi(x)=xi(xi−⟨x⟩),⟨x⟩=1N∑j=1Nxj=1N,(2)f_i(\mathbf{x}) = x_i\left(x_i - \langle x\rangle\right),\qquad\langle x\rangle= \frac{1}{N}\sum_{j=1}^{N}x_j = \frac{1}{N}, \tag*{(2)}

where the equality ⟨x⟩=1/N\langle x\rangle= 1/N follows from the simplex constraint. The sign structure of fif_i is that of a positive feedback: components above the population mean gain a positive drift and those below the mean are driven further downward. Under purely deterministic dynamics (σ=0\sigma= 0), this creates a self-reinforcing ordering that terminates at a vertex of the simplex, corresponding to the extinction of all but one type—the diversity-collapse state described in § 1.1.

Generalized nonlinear interaction

To study how the strength of the competitive nonlinearity affects the collective behavior, we introduce the one-parameter family

fi(x;γ)=xi(xiγ−⟨xγ⟩),⟨xγ⟩=1N∑j=1Nxjγ,(3)f_i(\mathbf{x};\gamma) = x_i\left(x_i^\gamma- \langle x^\gamma\rangle\right),\qquad\langle x^\gamma\rangle= \frac{1}{N}\sum_{j=1}^{N}x_j^\gamma, \tag*{(3)}

where γ>0\gamma> 0 is the nonlinearity exponent. The baseline model (2) is recovered at γ=1\gamma= 1. For γ>1\gamma> 1 the positive feedback is superlinear: high-abundance components receive a disproportionately amplified advantage, making competitive exclusion mechanistically stronger. The exponent γ\gamma therefore serves as the primary control parameter governing the type and intensity of collective behavior investigated in § 4 and § 5.

Linear stability of the uniform fixed point

The uniform state xi=1/Nx_i = 1/N for all ii is a fixed point of the deterministic dynamics (3) for any γ>0\gamma> 0. Linearizing around this point, the Jacobian of the drift restricted to the simplex tangent space takes the form

Jij∣1/N=γN−γ(δij−1N)=γN−γPij,(4)\left.J_{ij}\right|_{1/N} = \gamma N^{-\gamma}\left(\delta_{ij} - \frac{1}{N}\right) = \gamma N^{-\gamma}P_{ij}, \tag*{(4)}

where P=I−N−111TP = I - N^{-1}\mathbf{1}\mathbf{1}^{\mathsf{T}} is the orthogonal projector onto the zero-sum subspace (the tangent space of ΔN−1\Delta^{N-1}). On this subspace PP acts as the identity, so JJ has a single eigenvalue

λ(γ,N)=γN−γ(5)\lambda(\gamma, N) = \gamma N^{-\gamma} \tag*{(5)}

with multiplicity N−1N - 1. Since λ>0\lambda> 0, the uniform state is linearly unstable: small perturbations grow at rate λ\lambda, driving the system toward the simplex vertices (diversity collapse).

Two features of (5) connect directly to the numerical results reported below: (i) at fixed NN, the instability rate decreases rapidly with γ\gamma (e.g., from 0.05 at γ=1\gamma= 1 to 0.005 at γ=2\gamma= 2 for N=20N = 20), providing a first-principles explanation for the self-stabilization of the high-entropy state at high γ\gamma (§ 4); (ii) at fixed γ\gamma, λ→0\lambda\to0 as N→∞N \to\infty, explaining why large systems self-stabilize without noise (§ 3.2). The generalized analytical bound σc(N,γ)\sigma_c(N,\gamma) derived from (5) is presented in Appendix § 9.

Stochastic differential equation and numerical integration

Additive Gaussian noise of uniform intensity σ≥0\sigma\geq0 is superimposed on the deterministic drift, yielding the Itô stochastic differential equation

dxi=fi(x;γ) dt+σ dWi(t),i=1,…,N,(6)dx_i = f_i(x;\gamma)\,dt + \sigma\,dW_i(t), \qquad i = 1,\ldots,N, \tag*{(6)}

where Wi(t)W_i(t) are independent standard Wiener processes [8, 9]. Equation (6) is written in the Itô sense; the simplex projection described below introduces an implicit reflecting boundary that maintains the probabilistic interpretation of xx throughout the simulation. Following the standard formulation of stochastic processes on bounded state spaces [8, 9], equation (6) is integrated numerically using the Euler–Maruyama scheme [10, 11],

xi(t+Δt)=xi(t)+fi(x(t);γ) Δt+σΔt ξi,ξi∼i.i.d.N(0,1),(7)x_i(t+\Delta t)=x_i(t)+f_i(x(t);\gamma)\,\Delta t+\sigma\sqrt{\Delta t}\,\xi_i,\qquad\xi_i \overset{\mathrm{i.i.d.}}{\sim}\mathcal{N}(0,1), \tag*{(7)}

with time step Δt=10−3\Delta t=10^{-3}. After each update, simplex membership is enforced by a two-step projection: (i) all negative entries are clipped to a small positive floor ε=10−10\varepsilon=10^{-10}, and (ii) the vector is renormalized to unit sum,

xi←max⁡(xi,ε)∑j=1Nmax⁡(xj,ε).(8)x_i \leftarrow\frac{\max(x_i,\varepsilon)}{\sum_{j=1}^{N}\max(x_j,\varepsilon)}. \tag*{(8)}

This projection preserves the simplex constraint at every step without modifying the qualitative dynamics, as verified by the insensitivity of all reported results to the exact value of ε\varepsilon over several orders of magnitude. Constrained SDEs on the simplex admit several well-posed numerical treatments: elimination of one component via the constraint xN=1−∑i=1N−1xix_N=1-\sum_{i=1}^{N-1}x_i and integration in reduced coordinates (e.g. the log-ratio coordinates used by Imhof [12] for stochastic replicator dynamics on ΔN−1\Delta^{N-1}), and step-wise projection back onto the feasible set after each Euler–Maruyama update (e.g. Bubeck, Eldan & Lehec [13], who show that projected Langevin schemes discretize the corresponding reflected diffusion in the sense of Lions & Sznitman [14]). In the simplex interior these approaches integrate the same reduced dynamics; we adopt the projection scheme (8) because it treats all NN components symmetrically and handles the boundary uniformly via the floor ε\varepsilon. The floor also prevents the xiln⁡xi→0x_i\ln x_i\to0 term from producing numerical underflow in the entropy calculation; since ε=10−10\varepsilon=10^{-10} contributes at most εln⁡ε≈2.3×10−9\varepsilon\ln\varepsilon\approx2.3\times10^{-9} per component to the raw entropy sum, the systematic bias on HH is negligible at all σ\sigma values explored.

Observables

Normalized Shannon entropy

The diversity of the system state x\mathbf{x} is quantified by the normalized Shannon entropy

H(x)=−1ln⁡N∑i=1Nxiln⁡xi,(9)H(\mathbf{x})=-\frac{1}{\ln N}\sum_{i=1}^{N}x_i\ln x_i, \tag*{(9)}

where the prefactor (ln⁡N)−1(\ln N)^{-1} maps the range of HH onto the unit interval: H=0H=0 corresponds to a fully ordered (single-dominant) corner state, and H=1H=1 to the uniform distribution xi=1/Nx_i=1/N for all ii. The normalization is essential when comparing results across system sizes, because the raw Shannon entropy −∑xiln⁡xi-\sum x_i\ln x_i grows with NN at fixed diversity level, introducing a spurious size dependence that would systematically bias estimates of the critical noise intensity σc\sigma_c toward smaller values for larger NN. All entropy values reported in this paper refer to the normalized form (9).

Critical noise intensity

For a given system size NN and nonlinearity exponent γ\gamma, the critical noise intensity σc\sigma_c is defined as the smallest value of σ\sigma for which the long-time average of HH exceeds a prescribed diversity threshold HthrH_{\mathrm{thr}},

σc(N,γ)=min⁡{σ≥0:⟨H⟩ss>Hthr},(10)\sigma_c(N,\gamma)=\min\{\sigma\ge0:\langle H\rangle_{\mathrm{ss}}>H_{\mathrm{thr}}\}, \tag*{(10)}

where ⟨⋅⟩ss\langle\cdot\rangle_{\mathrm{ss}} denotes a steady-state time average taken after a burn-in period of at least 3×1033\times10^3 integration steps. We set Hthr=0.7H_{\mathrm{thr}}=0.7 throughout; results are qualitatively insensitive to moderate variations of this threshold. The quantity ⟨H⟩ss\langle H\rangle_{\mathrm{ss}} is a deterministic quantity obtained by averaging over both time (post-burn-in trajectory) and independent trials; the threshold condition therefore involves no ambiguity from instantaneous stochastic fluctuations. The quantity σc\sigma_c measures the minimum stochastic forcing needed to sustain a diverse stationary state against the ordering tendency of the deterministic drift.

Fluctuation susceptibility

To probe collective fluctuations and test for critical behavior, we define the fluctuation susceptibility

χH=NVar⁡(H),(11)\chi_H=N\operatorname{Var}(H), \tag*{(11)}

in direct analogy with the magnetic susceptibility χ∝NVar⁡(m)\chi\propto N\operatorname{Var}(m) used in finite-size scaling analysis of spin systems [6, 7]. Here Var⁡(H)\operatorname{Var}(H) is the variance of H(x(t))H(\mathbf{x}(t)) computed over the steady-state trajectory. Physically, a large χH\chi_H indicates that independent realizations of the dynamics produce widely different entropy levels—the system is sensitive to initial conditions and has not settled into a unique macroscopic state. A small χH\chi_H indicates that all realizations converge to a similar entropy, signaling robust self-organization. The factor of NN rescales the susceptibility so that a genuine second-order phase transition would manifest as a divergence χH∼Nη/ν\chi_H \sim N^{\eta/\nu} in the thermodynamic limit, where η\eta and ν\nu are standard critical exponents, while a smooth crossover or finite-size artefact would produce a χH\chi_H that either saturates or decreases with increasing NN [6, 7]. This distinction is exploited in § 4 to determine whether the structural changes observed in the phase diagram correspond to true critical behavior.

Entropy relaxation time

The relaxation time τrelax\tau_{\mathrm{relax}} measures how quickly the macroscopic entropy H(t)H(t) reaches its stationary distribution from a random initial condition drawn uniformly on ΔN−1\Delta^{N-1}.

A natural but physically incorrect approach would be to declare relaxation complete once the microscopic state vector ceases to change appreciably, e.g., when 1N∑i∣xi(t)−xi(t−Δt)∣<εmicro\frac{1}{N}\sum_i \lvert x_i(t)-x_i(t-\Delta t)\rvert< \varepsilon_{\mathrm{micro}}. In a noise-driven system, however, individual components xix_i are permanently agitated by the Wiener increments in (6) and never reach a microscopic fixed point [8, 9]. Applying such a criterion causes the simulation to run until the maximum allowed time in all parameter regimes, producing a spurious flat τrelax(γ)\tau_{\mathrm{relax}}(\gamma) curve that carries no physical information (see Appendix § 8).

We therefore adopt a macroscopic convergence criterion based on the stability of the running entropy average. The simulation trajectory is partitioned into non-overlapping windows of nw=100n_w=100 integration steps. At each window boundary we compute the window-averaged entropy Hˉk\bar{H}_k for the current window kk and the preceding window k−1k-1. Relaxation is declared when the absolute difference ∣Hˉk−Hˉk−1∣\lvert\bar{H}_k-\bar{H}_{k-1}\rvert remains below a tolerance εH=10−4\varepsilon_H=10^{-4} for ns=5n_s=5 consecutive window pairs,

τrelax=Δt⋅t∗,t∗=min⁡{t:∣Hˉk(t)−Hˉk−1(t)∣<εH for ns consecutive k}.(12)\tau_{\mathrm{relax}}=\Delta t\cdot t^*, \qquad t^*=\min\left\{t:\left\lvert\bar{H}_k(t)-\bar{H}_{k-1}(t)\right\rvert<\varepsilon_H\ \text{for }n_s\text{ consecutive }k\right\}. \tag*{(12)}

This criterion is immune to microscopic noise fluctuations because H(x)H(\mathbf{x}) is smooth, NN-fold averaged function of the state: its running mean stabilizes as soon as the macroscopic distribution reaches stationarity, regardless of the ongoing microscopic agitation [8, 9]. The choices nw=100n_w = 100, εH=10−4\varepsilon_H = 10^{-4}, and ns=5n_s = 5 were verified to produce qualitatively identical τrelax(γ)\tau_{\mathrm{relax}}(\gamma) curves when varied by factors of two in either direction; detailed sensitivity checks are provided in Supplementary Material S5.

Simulation parameters

All numerical experiments share a common baseline configuration summarized in Table 1. The system size NN and the nonlinearity exponent γ\gamma are varied systematically as described below for each group of experiments; all other parameters are held fixed.

ParameterSymbolValue
Integration time stepΔt\Delta t10−310^{-3}
Total simulation stepsTT6×1036 \times10^{3}
Burn-in stepsTburnT_{\mathrm{burn}}3×1033 \times10^{3}
Window size (relaxation)nwn_w100 steps
Stable windows requirednsn_s5
Entropy convergence tol.εH\varepsilon_H10−410^{-4}
Simplex floorε\varepsilon10−1010^{-10}
Diversity thresholdHthrH_{\mathrm{thr}}0.7
Independent trials—≥5\ge5 (up to 20)

Table 1. Baseline simulation parameters used throughout this work. Parameters varied in specific experiments are stated in the corresponding results sections.

Note. The relaxation analysis (§ 5) uses Δt=0.01\Delta t = 0.01 and T=6000T = 6000 steps (total simulation time 60 units) to ensure macroscopic convergence; all other experiments use the baseline Δt=10−3\Delta t = 10^{-3}. The window parameters (nw,ns,εH)(n_w, n_s, \varepsilon_H) are defined in integration steps, so their physical durations scale with Δt\Delta t: e.g., nw=100n_w = 100 steps corresponds to 0.1 time units in the baseline experiments and 1.0 time unit in the relaxation analysis.

Critical noise scaling (§ 3). The system size is varied over N∈{5,8,10,15,20}N \in\{5, 8, 10, 15, 20\} and the noise intensity is scanned over a uniform grid of σ∈[0,0.20]\sigma\in[0, 0.20] with step Δσ=0.005\Delta\sigma= 0.005. For each (N,σ)(N, \sigma) pair, five independent trials are run from distinct random initial conditions drawn uniformly on ΔN−1\Delta^{N-1}; the reported ⟨H⟩ss\langle H \rangle_{\mathrm{ss}} is the mean over trials and over the post-burn-in trajectory. Note that N=20N = 20 lies in the deterministically self-stabilized regime (§ 3.2) and is therefore excluded from the power-law fit; only N∈{5,8,10,15}N \in\{5, 8, 10, 15\} are used for the scaling analysis. The universality test uses the same grid with the interaction replaced by the superlinear form (3) at γ∈{1.0,1.2,1.4,1.6,1.8,2.0}\gamma\in\{1.0, 1.2, 1.4, 1.6, 1.8, 2.0\}.

Phase diagram and fluctuation analysis (§ 4). The nonlinearity exponent and noise intensity are scanned jointly over γ∈[0.5,2.0]\gamma\in[0.5, 2.0] (14 uniformly spaced values) and σ∈[0,0.08]\sigma\in[0, 0.08] (35 uniformly spaced values), with N=20N = 20 fixed. Twenty independent trials are run per parameter pair; the fluctuation susceptibility χH\chi_H is computed from the pooled post-burn-in entropy time series across all trials. For the finite-size robustness check, the same (γ,σ)(\gamma, \sigma) grid is repeated at N∈{10,20,40}N \in\{10, 20, 40\}, also with twenty trials per point. Numerical stability with respect to the integration step is verified by repeating the N=20N = 20 grid at Δt=5×10−4\Delta t = 5 \times10^{-4}.

Relaxation dynamics (§ 5). The nonlinearity exponent is varied over γ∈[0.25,2.0]\gamma\in[0.25, 2.0] (15 uniformly spaced values) at fixed N=20N = 20 and σ=0.03\sigma= 0.03, a value within the range where the susceptibility peak is observed (§ 4). For these runs a coarser time step Δt=0.01\Delta t = 0.01 is used to achieve a total simulation time of 60 time units (6000 steps), ensuring that the macroscopic entropy has ample time to reach stationarity; the coarser step is justified by the convergence check in Supplementary S5. Twenty independent trials are run per γ\gamma value to obtain reliable estimates of the mean and standard error of τrelax\tau_{\mathrm{relax}}; raw trial-level distributions are also reported to characterize the trial-to-trial variability.

Numerical validation summary. The robustness of all reported results has been verified through four independent checks, which are documented in the Supplementary Materials: (i) integration time-step convergence (Δt=10−3\Delta t = 10^{-3} vs 5×10−45 \times10^{-4}, Supplementary S5); (ii) finite-size scaling across N∈{10,20,40}N \in\{10, 20, 40\} (Supplementary S6); (iii) sensitivity of τrelax\tau_{\mathrm{relax}} to the window parameters nwn_w and εH\varepsilon_H, verified to produce qualitatively identical curves when varied by factors of two; and (iv) insensitivity of the simplex-floor parameter ε=10−10\varepsilon= 10^{-10} over several orders of magnitude, with the resulting entropy bias bounded at <10−8< 10^{-8}. Together, these checks confirm that the conclusions of Sections 3–5 are not artefacts of specific numerical choices.

Noise-induced stabilization and size-dependent diversity response

Basic phenomenology: noise arrests diversity collapse

We begin by establishing the qualitative contrast between deterministic and stochastic dynamics on the simplex at N=3N = 3, γ=1\gamma= 1. Figure 1a shows representative trajectories in the two-dimensional simplex for a deterministic run (σ=0\sigma= 0) and a stochastic run (σ=0.08\sigma= 0.08), both started from the near-uniform initial condition x0≈(1/3,1/3,1/3)\mathbf{x}_0 \approx(1/3, 1/3, 1/3). Under deterministic dynamics the system follows a direct path toward a vertex, reaching the diversity-collapse state within approximately 2×1032 \times10^3 integration steps. Under stochastic dynamics the trajectory explores the interior of the simplex throughout the full simulation window of 5×1035 \times10^3 steps, with no discernible drift toward any vertex.

Basic phenomenology of stochastic diversity stabilization

Figure 1. Basic phenomenology of stochastic diversity stabilization (N=3N = 3, γ=1\gamma= 1). (a) Representative trajectories on the two-dimensional simplex for deterministic dynamics (σ=0\sigma= 0, left) and stochastic dynamics (σ=0.08\sigma= 0.08, right). The deterministic trajectory collapses to a vertex within ∼2×103\sim2 \times10^{3} steps; the stochastic trajectory explores the simplex interior throughout 5×1035 \times10^{3} steps. (b) Steady-state mean normalized entropy ⟨H⟩ss\langle H\rangle_{\mathrm{ss}} as a function of noise strength σ\sigma. The monotonic increase confirms that noise systematically promotes diversity; saturation near ⟨H⟩ss≈0.73\langle H\rangle_{\mathrm{ss}} \approx0.73 reflects the geometric ceiling of the N=3N = 3 simplex.

The dependence of long-time diversity on noise intensity is quantified in Figure 1b, which shows the steady-state mean entropy ⟨H⟩ss\langle H\rangle_{\mathrm{ss}} as a function of σ\sigma for N=3N = 3, γ=1\gamma= 1. The curve rises monotonically from ⟨H⟩ss≈0\langle H\rangle_{\mathrm{ss}} \approx0 at σ=0\sigma= 0 and saturates near ⟨H⟩ss≈0.73\langle H\rangle_{\mathrm{ss}} \approx0.73 for σ≳0.20\sigma\gtrsim0.20. The monotonic character confirms that noise acts as a diversity-promoting force throughout the range explored, consistent with the general theory of noise-induced stabilization [4, 5]. The saturation reflects the finite-entropy ceiling imposed by the N=3N = 3 simplex geometry rather than any dynamical transition.

Size-dependent noise–entropy response

To characterize how the role of noise changes with system size, we compute the full noise–entropy response curve ⟨H⟩ss(σ)\langle H\rangle_{\mathrm{ss}}(\sigma) for each N∈{5,8,10,15,20,30,50}N \in\{5, 8, 10, 15, 20, 30, 50\} under the baseline interaction (γ=1\gamma= 1).

Figure 2 shows all seven response curves together. Two qualitatively distinct behaviors are visible. For small NN (N≤15N \le15), the entropy at zero noise is substantially below the high-diversity region (⟨H⟩∣σ=0≈0.44\langle H\rangle|_{\sigma=0} \approx0.44–0.720.72), and increasing σ\sigma produces a pronounced, steep rise in ⟨H⟩\langle H\rangle: noise is the primary agent driving the system away from the ordering tendency of the deterministic drift. For large NN (N≥20N \ge20), the entropy at σ=0\sigma= 0 is already elevated (⟨H⟩∣σ=0≈0.80\langle H\rangle|_{\sigma=0} \approx0.80–0.850.85), and the response to increasing σ\sigma is shallow: the deterministic dynamics alone sustain a high-diversity stationary state, and noise plays only a secondary modulating role. This crossover in stabilization mechanism—from noise-dominated at small NN to deterministically self-stabilized at large NN—is the central finding of this section and connects to the large-NN self-organization discussed in evolutionary game dynamics [3].

Noise–entropy response curves for system sizes N = 5, 8, 10, 15, 20, 30, and 50

Figure 2. Noise–entropy response curves ⟨H⟩ss(σ)\langle H\rangle_{\mathrm{ss}}(\sigma) for system sizes N∈{5,8,10,15,20,30,50}N \in\{5,8,10,15,20,30,50\} under the baseline interaction (γ=1\gamma= 1; 10 trials per point). Dashed vertical lines mark the inflection point σc\sigma_c (location of steepest response) for each NN. Small systems (N≤15N \le15, dark tones) show a pronounced noise-driven entropy rise from a low baseline, indicating that external noise is the primary stabilizing agent. Large systems (N≥20N \ge20, light tones) already maintain high entropy at σ=0\sigma= 0; their response to noise is shallow, reflecting deterministic self-stabilization. The crossover between these two regimes is the central result of this section.

Within the noise-dominated small-NN regime, the noise intensity at the steepest point of the response curve (the inflection point σc\sigma_c, defined as argmax⁡σd⟨H⟩/dσ\operatorname{argmax}_{\sigma} \mathrm{d}\langle H\rangle/\mathrm{d}\sigma) decreases with NN. For N∈{5,8,10,15}N \in\{5, 8, 10, 15\} a preliminary power-law fit to the inflection points gives

σc(N)≈CN−α,α≈1.4(N≤15 only),(13)\sigma_c(N) \approx C N^{-\alpha}, \qquad\alpha\approx1.4 \quad(N \le15\ \text{only}), \tag*{(13)}

as detailed above (four data points, N∈{5,8,10,15}N \in\{5, 8, 10, 15\}; R2=0.96R^2 = 0.96 on the log-log fit). (13) is offered as an empirical description of the N≤15N \le15 data only; we do not extrapolate it to larger NN, where the system self-stabilizes and σc\sigma_c loses physical meaning (see below). This exponent is obtained from the inflection-point definition of σc\sigma_c (the noise intensity at steepest response). An alternative threshold-crossing definition (⟨H⟩ss>Hthr\langle H\rangle_{\mathrm{ss}} > H_{\mathrm{thr}}), used in the universality test of § 3.3 for computational efficiency, yields α≈1.1\alpha\approx1.1 at γ=1\gamma= 1; the two values (1.4 vs 1.1) are of the same order, and the key qualitative result—that α\alpha depends on the interaction form—is robust to the definition choice (§ 3.3). This approximate scaling breaks down at larger NN. At large NN, the instability rate of the uniform fixed point λ=γN−γ\lambda= \gamma N^{-\gamma} (equation (5)) vanishes: for γ=1\gamma= 1, λ=1/N→0\lambda= 1/N \to0, so the deterministic ordering tendency weakens and the system self-stabilizes without noise. The analytical bound in Appendix B confirms this trend.

Universality test: model-specificity of the scaling exponent

The power-law ((13)) raises a natural question: does the exponent α≈1.4\alpha\approx1.4 reflect a universal property of diversity stabilization on the simplex, or is it specific to the baseline interaction ((2))? To answer this, we replace the baseline drift with the superlinear form ((3)) at γ>1\gamma> 1 and repeat the σc(N)\sigma_c(N) measurement for the same set of system sizes, averaging over five independent trials per parameter point to suppress single-run noise.

Figure 3a shows σc(N)\sigma_c(N) for the superlinear interaction at γ=2\gamma= 2. In contrast to the baseline model, the decay is steeper: σc\sigma_c drops from ≈0.10\approx0.10 at N=5N = 5 to ≈0.015\approx0.015 at N=20N = 20, yielding a fitted exponent α≈2.0\alpha\approx2.0, steeper than the baseline value. The curve is monotonically decreasing and well described. by a power law throughout. The key observation is not a breakdown of the power-law structure itself, but a systematic shift in the exponent: stronger nonlinear coupling steepens the σc(N)\sigma_c(N) decay, meaning that larger systems become self-stabilizing more rapidly under superlinear interactions.

Two-panel plot showing critical noise intensity versus system size and fitted scaling exponent versus interaction nonlinearity

Figure 3. Model-specificity of the critical noise scaling. (a) Critical noise intensity σc\sigma_c versus system size NN for the superlinear interaction (3) at γ=2\gamma= 2 (five trials per point). The curve is monotonically decreasing with fitted exponent α≈2.0\alpha\approx2.0, steeper than the baseline value α≈1.1\alpha\approx1.1. (b) Fitted scaling exponent α\alpha as a function of the nonlinearity parameter γ\gamma. The exponent remains near ≈1\approx1 for γ≤1.6\gamma\le1.6 (with a non-monotonic dip at γ=1.2\gamma= 1.2), then rises to ≈2.0\approx2.0 at γ=2.0\gamma= 2.0; α\alpha is model-dependent and sensitive to the interaction nonlinearity.

To characterize how the exponent evolves as the nonlinearity is increased continuously from the baseline, we fit σc(N)\sigma_c(N) at each value of γ∈{1.0,1.2,1.4,1.6,1.8,2.0}\gamma\in\{1.0, 1.2, 1.4, 1.6, 1.8, 2.0\} and plot the resulting α(γ)\alpha(\gamma) in Figure 3b. The exponent remains near α≈1\alpha\approx1 for γ≤1.6\gamma\le1.6 (ranging from 1.08 at γ=1.0\gamma= 1.0 to 1.28 at γ=1.6\gamma= 1.6, with a non-monotonic dip to 0.99 at γ=1.2\gamma= 1.2), then rises sharply to α≈2.0\alpha\approx2.0 at γ=2.0\gamma= 2.0. The overall trend is an increase from ≈1.1\approx1.1 to ≈2.0\approx2.0, but the low-γ\gamma plateau and the non-monotonicity at γ=1.2\gamma= 1.2 indicate that the γ\gamma-dependence of α\alpha is not a simple smooth function.

These results indicate that α\alpha is a model-dependent quantity that varies with the functional form of the competitive interaction, rather than a universal critical index in the sense of renormalization-group universality classes [15, 16]. The sensitivity of α\alpha to γ\gamma indicates that the finite-size scaling of σc\sigma_c reflects the specific way the nonlinear drift balances positive feedback against the simplex constraint at each interaction strength. The structured γ\gamma-dependence of α\alpha is partially captured by the analytical bound derived in Appendix B: linearization of the drift at the uniform fixed point yields the prediction αtheory(γ)=(γ+2)/2\alpha_{\mathrm{theory}}(\gamma) = (\gamma+ 2)/2 (equation (B.4)), which increases linearly from 1.5 at γ=1\gamma= 1 to 2.0 at γ=2\gamma= 2. The numerical values lie systematically below this prediction at low γ\gamma (where the instability is strong and linearization is least accurate) but converge to it at high γ\gamma (α≈2.0\alpha\approx2.0 vs αtheory=2.0\alpha_{\mathrm{theory}} = 2.0 at γ=2\gamma= 2), consistent with the weakening of the instability rate λ=γN−γ\lambda= \gamma N^{-\gamma} that makes the linearized approximation progressively more accurate.

Nonlinear phase structure and absence of collective criticality

Dynamical phase diagram

We now turn to the broader parameter space spanned by the nonlinearity exponent γ\gamma and the noise strength σ\sigma. For each point on a 14×3514 \times35 grid covering γ∈[0.5,2.0]\gamma\in[0.5, 2.0] and σ∈[0,0.08]\sigma\in[0, 0.08], we compute the steady-state mean normalized entropy ⟨H⟩ss\langle H \rangle_{\mathrm{ss}} at N=20N = 20, averaging over five independent trials and the post-burn-in trajectory. The result is displayed as a heat map in Figure 4a.

Four-panel phase structure and fluctuation analysis plots

Figure 4. Phase structure and fluctuation analysis at N=20N = 20. (a) Dynamical phase diagram in the (γ,σ)(\gamma, \sigma) plane; color encodes ⟨H⟩ss\langle H\rangle_{\mathrm{ss}}. The transition from low to high entropy is a smooth crossover. (b) Heat map of the entropy variance Var⁡(H)\operatorname{Var}(H) in the (γ,σ)(\gamma, \sigma) plane. The global maximum occurs at low γ≈0.7\gamma\approx0.7; variance decreases as γ\gamma increases. (c) Fluctuation susceptibility χH=NVar⁡(H)\chi_H = N\operatorname{Var}(H) map; peak value χH≈0.6\chi_H \approx0.6 at (γ,σ)≈(0.7,0.002)(\gamma, \sigma) \approx(0.7, 0.002). (d) Peak susceptibility χHpeak\chi_H^{\mathrm{peak}} versus nonlinearity γ\gamma (maximum over all σ\sigma at each γ\gamma; trials = 20). The curve falls on average from ≈0.6\approx0.6 at γ≈0.7\gamma\approx0.7 to ≈0.10\approx0.10 at γ=2.0\gamma= 2.0; no divergence or pronounced peak is observed near γ≈1.4\gamma\approx1.4, ruling out a critical phase transition at that location.

The heat map reveals a systematic dependence of the stationary diversity on both control parameters. At low nonlinearity (γ≲0.8\gamma\lesssim0.8) the entropy is generally depressed across the entire range of σ\sigma explored, with the most pronounced ordering occurring at small σ\sigma where the deterministic collapse tendency is least counteracted. As γ\gamma increases beyond approximately 1.0–1.4, the entropy rises progressively and eventually saturates near its maximum value for γ≳1.5\gamma\gtrsim1.5, largely irrespective of σ\sigma. In this high-γ\gamma regime the system maintains a high-diversity stationary state even in the absence of noise (σ≈0\sigma\approx0), indicating that the superlinear interaction itself acts as a self-stabilizing mechanism that substitutes for external stochastic forcing.

The transition between the low-entropy and high-entropy regimes is continuous and gradual throughout the parameter plane. No sharp boundary, discontinuous jump, or narrow coexistence region is visible in fig:4a; the equal-entropy contours run as smooth, broadly spaced curves with no sign of bunching or abrupt curvature change. Automated detection of the maximum entropy gradient over the full grid yields a value of 0.041 per unit γ\gamma, confirming the absence of any sharp structural feature. The grid spacing in γ\gamma is approximately 0.115, so reported peak positions carry an intrinsic resolution uncertainty of ±0.06\pm0.06; this does not affect any qualitative conclusion but should be borne in mind when comparing peak locations across experiments.

More generally, the precise location and amplitude of the χH\chi_H peak are sensitive to the random-seed ensemble and the parameter grid, and vary by ∼0.2\sim0.2 in γ\gamma and ∼20%\sim20\% in amplitude across independent runs. The conclusions of this section depend only on two robust features that are stable across all runs: (i) χH\chi_H does not diverge with NN, and (ii) χH\chi_H is strongly suppressed for γ≳1.5\gamma\gtrsim1.5. The system therefore exhibits a smooth crossover rather than a phase transition as γ\gamma is varied, a distinction with important consequences for the fluctuation analysis presented in § 4.2.

Fluctuation susceptibility analysis

The smooth crossover seen in Figure 4a raises the question of whether the structural change near γ≈1.0\gamma\approx1.0–1.41.4 nevertheless harbors genuine critical behavior at the level of fluctuations. In equilibrium statistical mechanics, a second-order phase transition that appears continuous in the order parameter can still produce a divergent susceptibility at the critical point [16, 6]. We therefore compute the full fluctuation landscape χH(γ,σ)=NVar⁡(H)\chi_H(\gamma, \sigma) = N \operatorname{Var}(H) over the same 14 ×\times 35 parameter grid used in § 4.1.

Entropy variance and susceptibility maps. Figure 4b shows the variance Var⁡(H)\operatorname{Var}(H) as a heat map in the (γ,σ)(\gamma, \sigma) plane. The global maximum of Var⁡(H)\operatorname{Var}(H) is located in the low-γ\gamma region near γ≈0.7\gamma\approx0.7, σ≈0.03\sigma\approx0.03, and the variance decreases progressively as γ\gamma increases toward 2.0. The susceptibility map χH=NVar⁡(H)\chi_H = N \operatorname{Var}(H) shown in Figure 4c reproduces the same spatial structure—as expected, since the two quantities differ only by a factor of N=20N = 20—with a peak value of χH≈0.6\chi_H \approx0.6 at (γ,σ)≈(0.7,0.002)(\gamma, \sigma) \approx(0.7, 0.002). Notably, neither map shows any concentrated bright region near γ≈1.4\gamma\approx1.4, the location where the phase diagram changes appearance most visibly.

Peak susceptibility as a function of γ\gamma. To make this point precise, Figure 4d shows χHpeak(γ)\chi_H^{\mathrm{peak}}(\gamma), defined as the maximum of χH\chi_H over all σ\sigma values at each fixed γ\gamma. The curve attains its global maximum of ≈0.6\approx0.6 at γ≈0.7\gamma\approx0.7 and decreases on average as γ\gamma increases, falling to ≈0.10\approx0.10 by γ=2.0\gamma= 2.0 (a reduction by a factor of approximately six). Local revivals appear at several intermediate γ\gamma values, but all remain well below the primary peak and are associated with the stochastic variability of the low-γ\gamma regime identified in § 4.3. There is no peak, kink, or inflection point of any significance at γ≈1.4\gamma\approx1.4.

Absence of critical divergence: a key negative result. The foregoing observations constitute a direct test of the criticality hypothesis. If the crossover near γ≈1.4\gamma\approx1.4 were a genuine second-order phase transition, the susceptibility χH\chi_H should exhibit a pronounced maximum or divergence at the transition point, growing with system size as χH∼Nη/ν\chi_H \sim N^{\eta/\nu} in the spirit of finite-size scaling [16, 6]. The data show the opposite: the susceptibility peaks at γ≈0.7\gamma\approx0.7 with χH≈0.6\chi_H \approx0.6 and falls by a factor of approximately six by γ=2.0\gamma= 2.0; the value at γ≈1.4\gamma\approx1.4 is less than one-third of the low-γ\gamma maximum. While the susceptibility peak at γ≈0.7\gamma\approx0.7 may superficially resemble a critical point, its lack of divergence with NN (§ 4.3) and the insensitivity of the relaxation timescale to γ\gamma (§ 5) are inconsistent with equilibrium or nonequilibrium universality classes [15, 16]. We conclude that no critical phase transition occurs along the γ\gamma axis within the parameter range explored. Instead, the system enters an entropy-saturated nonlinear phase for γ≳1.5\gamma\gtrsim1.5, characterized by high mean entropy and simultaneously low collective fluctuations—a combination inconsistent with criticality but consistent with a nonlinear smoothing mechanism, as elaborated in § 4.4.

Finite-size verification

A finite-size check of the fluctuation suppression result is provided in Supplementary Material S6. Using matched trial counts (trials = 20 for all NN), the peak susceptibilities at N∈{10,20,40}N \in\{10, 20, 40\} are 0.7, 0.6, and 0.5 respectively (see Supplementary Material S6). The values remain within the same order of magnitude across all three sizes, with no systematic divergence as NN increases. This is inconsistent with the divergent growth χHpeak∼Nη/ν\chi_H^{\mathrm{peak}} \sim N^{\eta/\nu} expected near a genuine second-order phase transition [6, 7]. Furthermore, in all three cases the high-γ\gamma region (γ≳1.5\gamma\gtrsim1.5) shows strongly suppressed χH\chi_H, confirming the fluctuation suppression identified in § 4.2 regardless of system size.

Three-regime dynamical picture

The results of § 4.1–§ 4.3 define three regimes as a function of γ\gamma (see also Table 2). Regime I (γ≲0.8\gamma\lesssim0.8): the deterministic drift biases the system toward low-entropy states; χH\chi_H is large (≈0.6\approx0.6) but does not diverge with NN, identifying a finite-size effect rather than true criticality. Regime II (0.8≲γ≲1.40.8 \lesssim\gamma\lesssim1.4): a smooth crossover in which the stationary probability mass redistributes from the ordering corners toward the simplex interior; χH\chi_H declines without any local extremum. Regime III (γ≳1.5\gamma\gtrsim1.5): the mean entropy saturates near unity independently of σ\sigma, and χH\chi_H is strongly suppressed—a hallmark of fluctuation suppression by the nonlinear deterministic coupling rather than by noise [4]. The dynamical consequences of this suppression are analyzed in § 5.

Regime (γ\gamma range)χHpeak\chi_{\mathrm{H}}^{\mathrm{peak}}⟨τ⟩\langle\tau\ranglestd⁡(τ)\operatorname{std}(\tau)
I: ordering-prone (≲0.8\lesssim0.8)large (≈0.6\approx0.6)∼2.9\sim2.9∼0.8\sim0.8
II: crossover (0.80.8–1.41.4)declining∼2.9\sim2.9∼0.6\sim0.6
III: entropy-saturated (≳1.5\gtrsim1.5)small (≲0.15\lesssim0.15)∼2.9\sim2.9∼0.9\sim0.9

Table 2. Summary of the three dynamical regimes. χHpeak\chi_{\mathrm{H}}^{\mathrm{peak}}: peak fluctuation susceptibility over σ\sigma at each γ\gamma. ⟨τ⟩\langle\tau\rangle: mean relaxation time (20 trials). std⁡(τ)\operatorname{std}(\tau): trial-to-trial standard deviation.

Relaxation dynamics and pathway multiplicity

Insensitivity of the relaxation timescale to nonlinearity

The fluctuation analysis of § 4 characterizes the stationary state but does not reveal how the system reaches it. We now measure the entropy relaxation time τrelax\tau_{\mathrm{relax}} defined in equation (12) as a function of γ\gamma at fixed N=20N = 20 and σ=0.03\sigma= 0.03, a noise level within the range where the susceptibility peak is observed.

Figure 5a shows τrelax(γ)\tau_{\mathrm{relax}}(\gamma) computed from five independent trials. The curve fluctuates between 2.2 and 3.4 time units with no systematic trend as γ\gamma increases from 0.25 to 2.0.

Three panels showing entropy relaxation time as a function of nonlinearity

Figure 5. Entropy relaxation time τrelax\tau_{\mathrm{relax}} as a function of nonlinearity γ\gamma at N=20N = 20, σ=0.03\sigma= 0.03. (a) Mean of five independent trials; the curve fluctuates between 2.2 and 3.4 time units with no systematic dependence on γ\gamma. (b) Mean ±\pm one standard error over twenty independent trials. The grand mean is ≈2.9\approx2.9 across all γ\gamma; the curve shows no monotonic trend, in contrast to the sixfold variation of χH\chi_H over the same γ\gamma range (Figure 4d). (c) Boxplot of per-trial relaxation times. The interquartile range is comparable across all γ\gamma values, confirming that neither the central tendency nor the variability of τrelax\tau_{\mathrm{relax}} depends on nonlinearity strength.

To obtain statistically reliable estimates, Figure 5b shows τrelax\tau_{\mathrm{relax}} averaged over twenty independent trials, with error bars representing one standard error. The mean curve confirms the absence of a systematic trend: ⟨τrelax⟩\langle\tau_{\mathrm{relax}}\rangle fluctuates between 2.5 and 3.2 with a grand mean of approximately 2.9 across all γ\gamma values, and no monotonic increase or decrease is discernible. The trial-to-trial spread is substantial throughout (standard deviation ≈0.5\approx0.5–1.31.3), reflecting persistent variability at all nonlinearity strengths.

Figure 5c displays the raw per-trial relaxation times as a boxplot. At every γ\gamma the interquartile range spans approximately [2.0, 3.5], with occasional outliers reaching 5–7 time units. The boxplot widths are comparable across the full γ\gamma range, confirming that the variability structure is insensitive to nonlinearity strength.

The central finding of this section is therefore negative: τrelax\tau_{\mathrm{relax}} does not mirror the dramatic γ\gamma-dependence of χH\chi_H. While the susceptibility decreases approximately sixfold from low to high γ\gamma (§ 4.2), the relaxation timescale remains essentially constant.

Decoupling between relaxation rate and fluctuation suppression

The insensitivity of τrelax\tau_{\mathrm{relax}} to γ\gamma, combined with the strong γ\gamma-dependence of χH\chi_H, constitutes the most informative result of the relaxation analysis. The decoupling between these two observables rules out a simple potential-deepening mechanism for fluctuation suppression and instead points to a pathway-level explanation.

What the decoupling rules out. If the suppression of χH\chi_H at high γ\gamma arose from a progressively deeper effective potential well around the high-entropy stationary state—analogous to increasing the spring constant in a harmonic trap—then the relaxation time should decrease proportionally: a deeper well implies faster convergence. The data show the opposite: χH\chi_H drops sixfold while τrelax\tau_{\mathrm{relax}} remains unchanged at ≈2.9\approx2.9 time units. This rules out any mechanism in which the stationary distribution narrows because the individual-trajectory dynamics converge more rapidly to a single preferred macrostate.

What the decoupling implies.. The combination of constant τrelax\tau_{\mathrm{relax}} and decreasing χH\chi_H has a natural interpretation in terms of the law of total variance. Conditioning on the microscopic relaxation pathway ω\omega,

Var⁡(H)=E[Var⁡(H∣ω)]+Var⁡(E[H∣ω]).(14)\operatorname{Var}(H) = \mathbb{E}\left[\operatorname{Var}(H \mid\omega)\right] + \operatorname{Var}\left(\mathbb{E}[H \mid\omega]\right). \tag*{(14)}

Because each trajectory takes approximately the same time to reach stationarity regardless of γ\gamma, the rate at which individual trajectories explore the configuration space is not enhanced by nonlinearity. What changes is the endpoint distribution: at high γ\gamma, different trajectories—following qualitatively different microscopic paths (as evidenced by the persistent trial-to-trial variability in τrelax\tau_{\mathrm{relax}})—converge to macroscopic states with similar entropy values. In the language of equation (14), the second term Var⁡(E[H∣ω])\operatorname{Var}(\mathbb{E}[H \mid\omega]) is suppressed because different pathways ω\omega produce similar conditional means E[H∣ω]\mathbb{E}[H \mid\omega], even though the pathways themselves are diverse.

We term this pathway-multiplicity-driven decorrelation: strong nonlinear coupling does not accelerate the dynamics or deepen a single basin of attraction, but rather creates many approximately equivalent routes to the high-entropy interior of the simplex. The resulting diversity of pathways, combined with the similarity of their endpoints, suppresses the ensemble variance without affecting the trajectory-level relaxation rate.

Falsifiability. The decoupling-based argument makes a specific, testable prediction: fluctuation suppression via pathway multiplicity requires τrelax\tau_{\mathrm{relax}} to be independent of γ\gamma while χH\chi_{\mathrm{H}} decreases. If instead τrelax\tau_{\mathrm{relax}} were found to decrease proportionally with χH\chi_{\mathrm{H}} (i.e., both observables tracking the same control parameter), this would indicate a potential-deepening mechanism rather than pathway multiplicity, and the present interpretation would be falsified. The data support the decoupling scenario.

Mechanistic summary

Table 2 summarizes the joint behavior of χH\chi_{\mathrm{H}} and τrelax\tau_{\mathrm{relax}} across the three dynamical regimes.

Mechanistic notes. All regimes: The relaxation timescale ⟨τrelax⟩≈2.9\langle\tau_{\mathrm{relax}}\rangle\approx2.9 is approximately constant across all three regimes, while χH\chi_{\mathrm{H}} varies sixfold. This decoupling implies that fluctuation suppression in Regime III does not arise from faster convergence (deeper potential well) but from pathway-multiplicity-driven ensemble decorrelation: diverse microscopic trajectories converge to similar macroscopic endpoints, reducing the between-trajectory variance without affecting the individual relaxation rate.

The key feature of Table 2 is the contrast between the χH\chi_{\mathrm{H}} column, which varies sixfold across regimes, and the ⟨τ⟩\langle\tau\rangle column, which remains constant. This decoupling is the primary evidence for the pathway-multiplicity mechanism: fluctuation suppression in the high-γ\gamma regime is not a consequence of faster or more deterministic relaxation, but of the convergence of diverse relaxation pathways to similar macroscopic endpoints. Together with the smooth crossover established in § 4.1 and the absence of critical fluctuations confirmed in § 4.2, this defines nonlinear fluctuation suppression via pathway multiplicity as the operative stabilization pattern in this class of stochastic competitive systems.

Discussion

Synthesis of main findings

The small-NN scaling exponent α\alpha is a model-specific quantity tied to the baseline replicator interaction (2), not a universal critical index [15, 16, 17]: replacing the interaction with a superlinear variant shifts the exponent from ≈1.1\approx1.1 to ≈2.0\approx2.0, with a non-trivial γ\gamma-dependence. The nonlinearity exponent γ\gamma governs the type of collective behavior—ordering-prone at low γ\gamma, smoothly crossing over to entropy-saturated at high γ\gamma—without producing a divergent susceptibility or a γ\gamma-dependent relaxation timescale [18, 15, 5].

The core finding is that strong nonlinearity suppresses collective fluctuations while leaving the relaxation timescale unchanged—a decoupling that rules out potential-deepening mechanisms and instead points to pathway-multiplicity-driven ensemble decorrelation as the operative stabilization pattern. This pattern—nonlinear fluctuation suppression via pathway multiplicity—is distinct from noise-induced stabilization [4, 19] (which operates through stochastic forcing) and from glassy dynamics [20, 21, 22] (where suppression accompanies slowing of relaxation). This decoupling between relaxation rate and fluctuation suppression has not been previously reported in simplex-constrained competitive dynamics; existing accounts attribute diversity stabilization either to noise [4, 5] or to deterministic attractors [1, 3], but not to an ensemble decorrelation effect operating independently of both. The regime structure identified here (Table 2) is expected to arise in other competitive-exclusion dynamics on the simplex, such as Lotka–Volterra systems and cyclic competition models, provided the interaction preserves the simplex constraint. Beyond these model-level extensions, the regime structure admits a qualitative correspondence with empirically measurable quantities in real competitive systems, offered here as an interpretive mapping rather than a quantitative fit. In microbial community ecology, for instance, the noise intensity σ\sigma plays the role of environmental fluctuation strength and the normalized entropy HH corresponds directly to the Shannon diversity index routinely measured in that field; the low-γ\gamma, noise-dependent regime then maps onto the empirically documented situation in which environmental variability sustains species coexistence that deterministic competition would eliminate, while the high-γ\gamma, self-stabilizing regime maps onto communities whose diversity is maintained by strong niche structure largely independently of external noise. An analogous reading applies to market-share dynamics, with 1−HHI1 - \mathrm{HHI} (one minus the Herfindahl–Hirschman concentration index) as the diversity observable. These are qualitative correspondences intended to locate the model’s regimes within known phenomenology; quantitative calibration against specific datasets is left to future work.

Relation to existing theoretical frameworks

Noise-induced transitions.. The theory of noise-induced transitions developed by Horsthemke and Lefever [4] identifies conditions under which additive or multiplicative noise shifts the stationary distribution of a nonlinear system, sometimes creating new phases that have no deterministic counterpart. In that framework, the noise strength σ\sigma plays the role of a bifurcation parameter, and a transition occurs when the stationary probability density changes its qualitative shape—for instance, from unimodal to bimodal—at a critical σc\sigma_c. The present system shares the qualitative feature that noise promotes a more uniform (high-entropy) stationary distribution, and the critical noise intensity σc(N)\sigma_c(N) identified in § 3 is formally analogous to a noise-induced transition threshold. However, the fluctuation analysis of § 4 reveals a key difference: the “transition” is not accompanied by any divergence of the susceptibility χH\chi_H, and its apparent location in the (γ,σ)(\gamma, \sigma) plane shifts continuously with system size rather than converging to a well-defined thermodynamic limit. Similarly, Van den Broeck and colleagues showed that noise can induce genuine nonequilibrium phase transitions with associated critical behavior in spatially extended systems [5]; the present model, being a mean-field system on the simplex with no spatial structure, does not develop the long-range correlations necessary to sustain such a transition. The diversity stabilization observed here is therefore better described as a noise-biased smooth crossover [23] than as a noise-induced phase transition.

Finite-size scaling theory.. Standard finite-size scaling [6, 7, 24] provides a systematic framework for extrapolating finite-system measurements to the thermodynamic limit near a genuine second-order transition. At a critical point, the susceptibility scales as χ∼Nη/ν\chi\sim N^{\eta/\nu}, and this divergence can be detected even in small systems by tracking how the peak susceptibility grows with NN. The results of § 4.3 demonstrate that this diagnostic fails in the present system: matched-trial (trials = 20) measurements at N∈{10,20,40}N \in\{10, 20, 40\} yield peak values of 0.7, 0.6, and 0.5—values within the same order of magnitude, with a slight decrease with NN. This failure of FSS is itself informative—it identifies the fluctuation peak at low γ\gamma as a finite-size pseudo-critical artefact, present because small systems have relatively large boundary-to-bulk ratios and are easily destabilized by fluctuations, but diluted as the interior bulk behavior comes to dominate at larger NN. The implication for numerical studies of similar systems is methodological: a susceptibility peak observed at a single system size is not sufficient evidence for a phase transition; monotonic growth of the peak with NN is the necessary condition, and this test should be applied routinely.

Evolutionary game dynamics and large-population limits. In finite-population evolutionary game theory, stochastic effects arising from demographic noise are known to alter the long-run frequency distribution of strategies relative to the deterministic replicator prediction [1, 3]. Taylor and Nowak [3] showed that in finite populations the probability of fixation—the analog of diversity collapse in the present model—depends sensitively on population size NN, and that for certain game structures larger populations are effectively more deterministic, converging toward the replicator equilibrium. The present model exhibits a superficially similar large-NN trend—σc→0\sigma_c \to0 as N→∞N \to\infty—but the mechanism differs. In evolutionary game theory the noise is intrinsic (demographic), scales as N−1/2N^{-1/2}, and its effect diminishes as NN grows because the law of large numbers suppresses relative fluctuations. Here the noise is external and additive, and the diminishing requirement for noise as NN increases reflects a growing self-organizing capacity of the deterministic simplex dynamics, not a dilution of stochastic effects. The power-law exponent α\alpha quantifies this self-organizing efficiency for the baseline replicator interaction, and its sensitivity to the interaction form (§ 3.3) suggests that the large-NN self-stabilization is a model-specific property rather than a generic feature of competitive dynamics on the simplex.

Limitations and outlook

The scaling exponent α\alpha applies only for N≤15N \le15; at larger NN the system self-stabilizes. The analytical prediction αtheory=(γ+2)/2\alpha_{\mathrm{theory}} = (\gamma+ 2)/2 (Appendix § 9) captures the trend; a rigorous bound awaits nonlinear analysis of the boundary effects.

The two main conclusions—absence of a phase transition and decoupling between τrelax\tau_{\mathrm{relax}} and χH\chi_H—are established numerically. Analytical derivations (e.g. via Fokker–Planck analysis of the stationary distribution) would strengthen them and are a priority for future work. The decoupling is an empirical observation consistent with pathway multiplicity; a constructive proof that pathway multiplicity implies decoupling remains open, and the falsifiability criterion in § 5.2 provides a testable prediction for future studies.

Natural extensions include: (i) Fokker–Planck analysis [25] of the stationary density pss(x)p_{\mathrm{ss}}(\mathbf{x}) to obtain a sharper α\alpha prediction; (ii) large-deviation theory [26] to characterize diversity collapse as a rare event; (iii) graph-structured interactions to test whether pathway multiplicity survives under spatial heterogeneity; and (iv) systematic variation of interaction classes (asymmetric matrices, frequency-dependent selection) to delineate the basin for which σc(N)\sigma_c(N) scaling is well-defined.

Conclusion

We have studied stochastic diversity stabilization in a class of replicator-like competitive dynamics on the NN-dimensional probability simplex, parameterized by an additive noise strength σ\sigma and a nonlinearity exponent γ\gamma. Under the baseline linear replicator interaction, the critical noise intensity required to sustain a diverse stationary state decays as a power law, σc∼N−α\sigma_c \sim N^{-\alpha}, implying that sufficiently large systems self-stabilize without external noise. Replacing the baseline with a superlinear variant shifts the exponent from α≈1.1\alpha\approx1.1 to ≈2.0\approx2.0 as γ\gamma increases, in quantitative agreement with the linear-stability prediction α=(γ+2)/2\alpha= (\gamma+ 2)/2 and establishing that α\alpha is a model-dependent quantity rather than a universal critical index. Scanning the full (γ,σ)(\gamma,\sigma) parameter plane reveals a smooth dynamical crossover from an ordering-prone, noise-dependent regime at low γ\gamma to an entropy-saturated, self-organizing phase at high γ\gamma, with no divergent fluctuation susceptibility at any point along the transition: the peak of χH=NVar⁡(H)\chi_H = N\operatorname{Var}(H) is located at γ≈0.7\gamma\approx0.7, well separated from the visual crossover region near γ≈1.4\gamma\approx1.4, and matched finite-size analysis finds no systematic divergence with NN, ruling out a genuine second-order phase transition in the thermodynamic limit. Relaxation dynamics complete the picture: the entropy relaxation time τrelax(γ)\tau_{\mathrm{relax}}(\gamma) is approximately constant at ≈2.9\approx2.9 time units across the full γ\gamma range, even as the susceptibility decreases sixfold. This decoupling between relaxation rate and fluctuation amplitude rules out a simple potential-deepening mechanism and implies that the system reaches the high-entropy stationary state via a broad ensemble of diverse microscopic pathways whose endpoints converge at high γ\gamma. The joint behavior of χH\chi_H and τrelax\tau_{\mathrm{relax}} supports a unified mechanistic interpretation: strong nonlinear coupling suppresses collective entropy fluctuations not by accelerating convergence toward a single attractor, but by generating pathway multiplicity that decorrelates trajectory endpoints at the ensemble level—a mechanism we term nonlinear fluctuation suppression via pathway multiplicity, which is distinct from both noise-induced transitions and glassy dynamics.

Failure of the microscopic relaxation criterion

As noted in Section 2.2, the entropy relaxation time τrelax\tau_{\mathrm{relax}} must be defined via a macroscopic convergence criterion rather than a microscopic one. This appendix documents the failure of the microscopic approach in order to justify the choice made in equation (12).

A natural first attempt is to declare relaxation complete when the point-wise change in the state vector falls below a threshold,

1N∑i=1N∣xi(t)−xi(t−Δt)∣<εmicro,(15)\frac{1}{N}\sum_{i=1}^{N}\left|x_i(t)-x_i(t-\Delta t)\right| < \varepsilon_{\mathrm{micro}}, \tag*{(15)}

with εmicro=10−4\varepsilon_{\mathrm{micro}} = 10^{-4}. In a noise-free system this criterion correctly identifies the approach to a fixed point. In the presence of additive Wiener noise, however, the increment xi(t)−xi(t−Δt)x_i(t)-x_i(t-\Delta t) contains a stochastic contribution of order σΔt∼10−3\sigma\sqrt{\Delta t}\sim10^{-3}–10−210^{-2} for the parameter values used here, which permanently exceeds εmicro=10−4\varepsilon_{\mathrm{micro}} = 10^{-4} regardless of how close the macroscopic distribution is to stationarity [8, 9]. As a consequence, criterion (A.1) never triggers, and every simulation run terminates at the maximum allowed time T=6000T=6000 steps, corresponding to simulation time t=60t=60.

Applying this criterion to the same parameter settings used in Section 5 (N=20N=20, σ=0.03\sigma=0.03, γ∈[0.25,2.0]\gamma\in[0.25,2.0]) produces a perfectly flat τrelax=60\tau_{\mathrm{relax}}=60 for all γ\gamma—the maximum allowed simulation time—confirming that the microscopic criterion carries no physical information in any noise-driven system. The macroscopic entropy criterion (12) resolves this issue: the resulting τrelax≈2.9\tau_{\mathrm{relax}}\approx2.9 is two orders of magnitude below the ceiling, confirming genuine convergence to stationarity (Figure 5).

Analytical bound on σc(N,γ)\sigma_c(N,\gamma)

We derive a scaling bound on the critical noise intensity as a function of both system size NN and nonlinearity exponent γ\gamma, using the linear stability result (5) from § 2.1.

Step 1: Heuristic stationary variance

The uniform fixed point xi=1/Nx_i = 1/N is linearly unstable with instability rate λ=γN−γ\lambda= \gamma N^{-\gamma} (equation (5)). In a purely linear system, this instability would cause unbounded growth; in the present model, the simplex boundaries provide a reflecting constraint that confines the dynamics and permits a stationary distribution. As a heuristic estimate, we balance the outward drift (rate λ\lambda), the diffusive spreading (intensity σ\sigma), and the confining effect of the simplex boundaries to obtain the characteristic displacement scale, yielding a variance estimate

Var⁡(xi)∼σ22λ=σ2Nγ2γ.(16)\operatorname{Var}(x_i) \sim\frac{\sigma^2}{2\lambda} = \frac{\sigma^2 N^\gamma}{2\gamma}. \tag*{(16)}

Equation (B.1) is an order-of-magnitude estimate, not a rigorous result: the uniform point is unstable, so the standard Ornstein–Uhlenbeck stationary-variance formula does not strictly apply, and the true stationary distribution is shaped by the nonlinear drift and the simplex boundaries. The estimate is nevertheless useful because it correctly captures the parametric dependence on NN and γ\gamma. All conclusions drawn from this appendix rest solely on the resulting scaling exponent αtheory=(γ+2)/2\alpha_{\mathrm{theory}} = (\gamma+ 2)/2 and not on the prefactor: the simplex boundary, not the linearized rate λ\lambda, sets the true confining scale, so the 1/(2λ)1/(2\lambda) factor should be read as fixing the NN- and γ\gamma-dependence rather than the absolute coefficient.

Step 2: Entropy lower bound

A second-order Taylor expansion of the normalized Shannon entropy HH (9) around E[xi]=1/N\mathbb{E}[x_i] = 1/N gives, using ∣∂2H/∂xi2∣1/N=N/ln⁡N\left|\partial^2 H/\partial x_i^2\right|_{1/N} = N/\ln N,

E[H(x)]≳1−N22ln⁡NVar⁡(xi)=1−σ2Nγ+24γln⁡N.(17)\mathbb{E}[H(\mathbf{x})] \gtrsim1 - \frac{N^2}{2\ln N}\operatorname{Var}(x_i) = 1 - \frac{\sigma^2 N^{\gamma+2}}{4\gamma\ln N}. \tag*{(17)}

Requiring E[H]≥Hthr\mathbb{E}[H] \ge H_{\mathrm{thr}} and inverting:

σc≲2γ(1−Hthr)ln⁡NNγ+2∼γ N−(γ+2)/2(ln⁡N)1/2.(18)\sigma_c \lesssim2\sqrt{\frac{\gamma(1-H_{\mathrm{thr}})\ln N}{N^{\gamma+2}}} \sim\sqrt{\gamma}\,N^{-(\gamma+2)/2}(\ln N)^{1/2}. \tag*{(18)}

The effective scaling exponent is therefore

αtheory(γ)=γ+22,(19)\alpha_{\mathrm{theory}}(\gamma) = \frac{\gamma+ 2}{2}, \tag*{(19)}

which predicts a linear increase of α\alpha with γ\gamma.

Comparison with numerical results

Equation (B.4) makes two testable predictions. First, for the baseline interaction (γ=1\gamma= 1), αtheory=3/2=1.5\alpha_{\mathrm{theory}} = 3/2 = 1.5; the numerical estimates α≈1.4\alpha\approx1.4 (inflection-point definition) and α≈1.1\alpha\approx1.1 (threshold-crossing definition) are of the same order of magnitude, with the discrepancy attributable to finite-NN effects and the heuristic nature of (B.1). Second, for γ=2\gamma= 2, αtheory=2.0\alpha_{\mathrm{theory}} = 2.0; the threshold-crossing fit gives α≈2.0\alpha\approx2.0, in close quantitative agreement. More generally, the predicted linear trend α=(γ+2)/2\alpha= (\gamma+ 2)/2 captures the observed increase of α\alpha with γ\gamma (Figure 3b), with the analytical values systematically above the numerical ones at low γ\gamma and converging at high γ\gamma. This convergence at high γ\gamma is consistent with the weakening of the instability rate λ=γN−γ\lambda= \gamma N^{-\gamma}, which makes the linearized approximation more accurate when the uniform state is only marginally unstable.

Declaration of Competing Interest

The author declares that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data Availability Statement

All simulation code (Python scripts for Figure 1–5 and Supplementary Figure S4–S6), raw output data, and analysis notebooks are publicly available at https://github.com/ouruocun-dotcom/stochastic-diversity-stabilization. The repository includes fixed random seeds, parameter files, and instructions to reproduce every figure in this paper.

Fluctuation-based dynamical boundary

Figure S4. Fluctuation-based dynamical boundary σ∗(γ)\sigma^{*}(\gamma), defined as the noise intensity at which χH\chi_{H} is maximized over σ\sigma at each fixed γ\gamma (N=20N = 20). The boundary is approximately stable for γ≲1.2\gamma\lesssim1.2 but becomes highly erratic for γ≳1.3\gamma\gtrsim1.3, where the susceptibility landscape is nearly flat and the apparent maximum is dominated by statistical noise. This instability is a consequence of fluctuation suppression in the high-γ\gamma regime (the fluctuation analysis of the main text) rather than a physical feature of the system.

Acknowledgements

The author would like to express sincere gratitude to the host institution for the academic environment and support provided during this research.

This work was conducted independently by the author. The author used an AI language model (Claude, Anthropic) for assistance with manuscript writing and language editing. All scientific content, experimental design, data analysis, and conclusions are the sole responsibility of the author.

Supplementary Material

This document contains three supplementary figures referenced in the main text. Figure S4 provides the fluctuation-based dynamical boundary extracted from the susceptibility landscape of the main text. Figure S5 documents the numerical stability of the main results with respect to the integration time step. Figure S6 presents the matched-trial finite-size verification of fluctuation suppression across N∈{10,20,40}N \in\{10, 20, 40\}.

Numerical stability check comparing results at two time steps

Figure S5. Numerical stability check: comparison of results at Δt=10−3\Delta t = 10^{-3} (left column) and Δt=5×10−4\Delta t = 5 \times10^{-4} (right column), both at N=20N = 20. Top row: steady-state mean normalized entropy ⟨H⟩ss\langle H\rangle_{\mathrm{ss}} in the (γ,σ)(\gamma,\sigma) plane; color ranges differ by less than 2%. Bottom row: fluctuation susceptibility χH=NVar⁡(H)\chi_H = N \operatorname{Var}(H); the qualitative structure—peak at low γ\gamma, suppression at high γ\gamma—is reproduced at both time steps. The shift in peak γ\gamma position reflects seed- and grid-dependent variability documented in the main text, not numerical error.

Finite-size verification of fluctuation suppression

Figure S6. Finite-size verification of fluctuation suppression. Each column corresponds to a system size N∈{10,20,40}N \in\{10,20,40\} (left to right). Top row: steady-state mean normalized entropy ⟨H⟩ss\langle H\rangle_{\mathrm{ss}} in the (γ,σ)(\gamma,\sigma) plane. Bottom row: fluctuation susceptibility χH=NVar⁡(H)\chi_H = N\operatorname{Var}(H). All three scans use trials = 20. Peak values are ≈0.7\approx0.7 (N=10N = 10, γ≈1.1\gamma\approx1.1), ≈0.6\approx0.6 (N=20N = 20, γ≈0.7\gamma\approx0.7), and ≈0.5\approx0.5 (N=40N = 40, γ≈0.6\gamma\approx0.6); no systematic increase with NN is observed, ruling out critical divergence [6, 7]. In all three panels, χH\chi_H is strongly suppressed at γ≳1.5\gamma\gtrsim1.5, confirming size-independent fluctuation suppression in the high-nonlinearity regime.

Fluctuation-based dynamical boundary

For each value of the nonlinearity exponent γ\gamma, the noise intensity σ∗(γ)\sigma^{*}(\gamma) at which the fluctuation susceptibility χH=NVar⁡(H)\chi_{H} = N \operatorname{Var}(H) attains its maximum over σ\sigma defines a fluctuation-based dynamical boundary in the (γ,σ)(\gamma, \sigma) plane. Figure S4 shows this boundary at N=20N = 20.

In the low-γ\gamma regime (γ≲1.2\gamma\lesssim1.2), the boundary is relatively stable, with σ∗\sigma^{*} confined to a narrow low-noise band and showing no clear trend. For γ≳1.3\gamma\gtrsim1.3, the boundary becomes highly erratic, with σ∗\sigma^{*} swinging abruptly between near-zero values and the upper edge of the scanned range without settling at any stable value. This instability does not reflect a physical phase boundary; it arises because the susceptibility in the high-γ\gamma regime is uniformly low and nearly flat across all σ\sigma (see Figure 4c of the main text), so the location of its maximum is dominated by statistical noise rather than any structural feature. The behavior is consistent with the fluctuation suppression of the entropy-saturated regime identified in the main text, and underscores that the fluctuation-based boundary is not a meaningful diagnostic in the entropy-saturated phase.

Numerical stability with respect to integration time step

All main-text results use the integration time step Δt=10−3\Delta t = 10^{-3}. To verify that the conclusions are not sensitive to this choice, we repeat the phase diagram and fluctuation susceptibility analysis at Δt=5×10−4\Delta t = 5 \times10^{-4} and compare the two outputs directly. Figure S5 shows the mean entropy ⟨H⟩ss\langle H \rangle_{\mathrm{ss}} (top row) and the susceptibility χH\chi_{H} (bottom row) for both time steps at N=20N = 20.

The mean entropy maps (top row) are qualitatively identical: both show the same progression from low entropy at small γ\gamma to high, σ\sigma-independent

entropy at large γ\gamma, with the same approximate location of low-entropy pockets near (γ,σ)≈(0.6,0.00)(\gamma, \sigma) \approx(0.6, 0.00). The color ranges differ by less than 2% in the dynamic range.

The susceptibility maps (bottom row) confirm that the spatial structure of χH\chi_H—peak at low γ\gamma, suppression at high γ\gamma—is reproduced at both time steps. The peak location shifts between the two time steps, consistent with the seed- and grid-dependent peak-position variability documented in the main text, rather than indicating integration error. All qualitative conclusions of the main text are unaffected by this variation.

Note on the relaxation time step.. The relaxation analysis of the main text uses a coarser step Δt=0.01\Delta t = 0.01 to achieve a total simulation time of 60 units within 6000 steps. A separate Δt\Delta t convergence check is not required for the relaxation observable, for two reasons. First, τrelax\tau_{\mathrm{relax}} is defined via a running-window average of the macroscopic entropy over 100-step windows (= 1.0 time unit at Δt=0.01\Delta t = 0.01), which smooths out step-level integration errors by averaging over ∼N×100\sim N \times100 individual entropy contributions. Second, the measured τrelax≈2.9\tau_{\mathrm{relax}} \approx2.9 time units is nearly three times the window size, so the convergence criterion responds to the slow macroscopic dynamics rather than to the step-level numerical scheme. The key requirement for the relaxation experiment is sufficient total simulation time (60 units ≫τrelax≈2.9\gg\tau_{\mathrm{relax}} \approx2.9), not step-level precision. To confirm this directly, we repeated the relaxation analysis at Δt=0.005\Delta t = 0.005 (12000 steps, same total time of 60 units, with the running window widened to 200 steps to preserve the 1.0 time-unit window width). The grand mean over all γ\gamma is ⟨τrelax⟩=2.89\langle\tau_{\mathrm{relax}} \rangle= 2.89 at Δt=0.01\Delta t = 0.01 and 2.81 at Δt=0.005\Delta t = 0.005, a relative difference of only 3%. Individual γ\gamma points differ by up to ∼13%\sim13\%, but these differences are unsystematic—some points increase and others decrease between the two time steps—and lie within the trial-to-trial standard deviation reported in the relaxation analysis of the main text (std(τ)≈0.5(\tau) \approx0.5–1.31.3). The near-cancellation of these fluctuations in the grand mean indicates that the systematic Δt\Delta t effect is negligible once trial-level scatter is averaged out. The qualitative conclusion—that τrelax\tau_{\mathrm{relax}} is approximately constant across γ\gamma—holds identically at both time steps. This check is numerical (a grand-mean comparison) and is not presented as a separate figure; the reproducing script is provided in the code repository.

Finite-size verification of fluctuation suppression

To assess whether the fluctuation suppression observed at N=20N = 20 persists as the system size grows, we repeat the (γ,σ)(\gamma, \sigma) susceptibility scan at N∈{10,20,40}N \in\{10, 20, 40\}. Figure S6 shows the mean entropy ⟨H⟩ss\langle H\rangle_{\mathrm{ss}} (top row) and the fluctuation susceptibility χH=NVar⁡(H)\chi_H = N\operatorname{Var}(H) (bottom row) for all three sizes, each computed with trials = 20.

Results with matched trial counts.. All three system sizes were scanned with trials = 20. The peak susceptibilities are:

NNχHpeak\chi_H^{\mathrm{peak}}peak γ\gamma
100.71.08
200.60.7
400.50.62

Table 3.

The three values lie within the same order of magnitude and do not show a systematic increase with NN; indeed, the peak decreases slightly from N=10N = 10 to N=40N = 40. A genuine second-order phase transition would require χHpeak\chi_H^{\mathrm{peak}} to grow as Nη/νN^{\eta/\nu} [6, 7]; the absence of such growth rules out critical behavior.

High-γ\gamma suppression is size-independent.. In all three panels (Figure S6, bottom row), the susceptibility in the high-γ\gamma regime (γ≳1.5\gamma\gtrsim1.5) is strongly suppressed relative to the low-γ\gamma peak, irrespective of system size. This confirms that the fluctuation suppression identified in the fluctuation analysis of the main text is a robust feature of the high-nonlinearity regime and not a finite-size artefact.

Peak location variability.. The peak location shifts between NN values (from γ≈0.62\gamma\approx0.62 at N=40N = 40 to γ≈1.08\gamma\approx1.08 at N=10N = 10), consistent with the dynamical variability discussed in the main text and not indicative of a fixed critical point.

References

  1. [1]J. Hofbauer, K. Sigmund, Evolutionary Games and Population Dynamics, Cambridge University Press, Cambridge, 1998. doi:10.1017/CB09781139173179.
  2. [2]A. J. McKane, T. J. Newman, Predator-prey cycles from resonant amplification of demographic stochasticity, Physical Review Letters 94 (21) (2005) 218102. doi:10.1103/PhysRevLett.94.218102.DOI
  3. [3]C. Taylor, M. A. Nowak, Evolutionary game dynamics in finite populations, Bulletin of Mathematical Biology 66 (6) (2004) 1621–1644. doi:10.1016/j.bulm.2004.03.004.
  4. [4]W. Horsthemke, R. Lefever, Noise-Induced Transitions: Theory and Applications in Physics, Chemistry, and Biology, Springer, Berlin, 1984. doi:10.1007/3-540-36852-3.
  5. [5]C. Van den Broeck, J. M. R. Parrondo, R. Toral, Noise-induced nonequilibrium phase transition, Physical Review Letters 73 (25) (1994) 3395–3398. doi:10.1103/PhysRevLett.73.3395.
  6. [6]M. E. Fisher, M. N. Barber, Scaling theory for finite-size effects in the critical region, Physical Review Letters 28 (23) (1972) 1516–1519. doi:10.1103/PhysRevLett.28.1516.DOI
  7. [7]K. Binder, Finite size scaling analysis of Ising model block distribution functions, Zeitschrift für Physik B 43 (2) (1981) 119–140. doi:10.1007/BF01293604.DOI
  8. [8]N. G. van Kampen, Stochastic Processes in Physics and Chemistry, 3rd Edition, Elsevier, Amsterdam, 2007. doi:10.1016/B978-0-444-52965-7.X5000-4.DOI
  9. [9]C. W. Gardiner, Stochastic Methods: A Handbook for the Natural and Social Sciences, 4th Edition, Springer, Berlin, 2009, ISBN: 978-3-540-70712-7.
  10. [10]P. E. Kloeden, E. Platen, Numerical Solution of Stochastic Differential Equations, Springer, Berlin, 1992. doi:10.1007/978-3-662-12616-5.DOI
  11. [11]D. P. Landau, K. Binder, A Guide to Monte Carlo Simulations in Statistical Physics, 5th Edition, Cambridge University Press, Cambridge, 2021. doi:10.1017/9781108780346.DOI
  12. [12]L. A. Imhof, The long-run behavior of the stochastic replicator dynamics, Ann. Appl. Probab. 15 (1B) (2005) 1019–1045. doi:10.1214/105051604000000837.DOI
  13. [13]S. Bubeck, R. Eldan, J. Lehec, Sampling from a log-concave distribution with projected Langevin Monte Carlo, Discrete Comput. Geom. 59 (4) (2018) 757–783. doi:10.1007/s00454-018-9992-1.DOI
  14. [14]P.-L. Lions, A.-S. Sznitman, Stochastic differential equations with reflecting boundary conditions, Comm. Pure Appl. Math. 37 (1984) 511–537. doi:10.1002/cpa.3160370408.DOI
  15. [15]H. Hinrichsen, Non-equilibrium critical phenomena and phase transitions into absorbing states, Advances in Physics 49 (7) (2000) 815–958. doi:10.1080/00018730050198152.DOI
  16. [16]P. C. Hohenberg, B. I. Halperin, Theory of dynamic critical phenomena, Reviews of Modern Physics 49 (3) (1977) 435–479. doi:10.1103/RevModPhys.49.435.DOI
  17. [17]U. C. Täuber, Critical Dynamics: A Field Theory Approach to Equilibrium and Non-Equilibrium Scaling Behavior, Cambridge University Press, Cambridge, 2014. doi:10.1017/CBO9781139046213.DOI
  18. [18]W. Kob, H. C. Andersen, Testing mode-coupling theory for a supercooled binary Lennard-Jones mixture, Physical Review E 51 (5) (1995) 4626–4641. doi:10.1103/PhysRevE.51.4626.DOI
  19. [19]T. Biancalani, L. Dyson, A. J. McKane, Noise-induced bistable states and their mean switching time in foraging colonies, Physical Review Letters 112 (3) (2014) 038101. doi:10.1103/PhysRevLett.112.038101.DOI
  20. [20]F. Ritort, P. Sollich, Glassy dynamics of kinetically constrained models, Advances in Physics 52 (4) (2003) 219–342. doi:10.1080/0001873031000093582.DOI
  21. [21]J.-P. Bouchaud, Weak ergodicity breaking and aging in disordered systems, Journal de Physique I 2 (9) (1992) 1705–1713. doi:10.1051/jp1:1992238.DOI
  22. [22]L. Berthier, G. Biroli, Theoretical perspective on the glass transition and amorphous materials, Reviews of Modern Physics 83 (2) (2011) 587–645. doi:10.1103/RevModPhys.83.587.DOI
  23. [23]P. Ashwin, S. Wieczorek, R. Vitolo, P. Cox, Tipping points in open systems: bifurcation, noise-induced and rate-dependent examples in the climate system, Philosophical Transactions of the Royal Society A 370 (1962) (2012) 1166–1184. doi:10.1098/rsta.2011.0306.DOI
  24. [24]V. Privman (Ed.), Finite Size Scaling and Numerical Simulation of Statistical Systems, World Scientific, Singapore, 1990. doi:10.1142/1011.DOI
  25. [25]H. Risken, The Fokker–Planck Equation: Methods of Solution and Applications, 2nd Edition, Springer, Berlin, 1989. doi:10.1007/978-3-642-61544-3.
  26. [26]M. Assaf, B. Meerson, WKB theory of large deviations in stochastic populations, Journal of Physics A: Mathematical and Theoretical 50 (26) (2017) 263001. doi:10.1088/1751-8121/aa669a.DOI

Paper details

vlnc:W_r3xjtbe4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

Claude

View statement
Looking for related papers…
Contents