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 -dimensional simplex , 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 , what role the deterministic dynamics play at large , 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 -dimensional probability simplex, in which the competitive interaction is parameterized by a nonlinearity exponent and the dynamics are driven by additive Gaussian noise of strength . Three main contributions are reported.
(i) Size-dependent noise–entropy response and deterministic self-stabilization. We systematically measure the noise–entropy response curve across system sizes and reveal a qualitative transition in stabilization mechanism: for small the entropy at is well below the high-diversity regime, so external noise is the primary stabilizing agent; for large the deterministic dynamics alone sustain a high-entropy stationary state, and noise plays only a secondary modulating role. In the small- regime () the noise intensity at the steepest response point decreases approximately as ; however, this approximate scaling breaks down at large , and the exponent shifts from to as the nonlinearity increases, in agreement with the analytical prediction (§ 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 parameter plane and show that the system undergoes a smooth crossover rather than a sharp phase transition as increases. To test for criticality, we compute the fluctuation susceptibility , where is the normalized Shannon entropy, across the full parameter space. The susceptibility peaks at () and is suppressed at high ; no divergence is observed near , where a naive visual inspection of the phase diagram might suggest a structural change. Matched finite-size analysis at finds no systematic divergence of the peak susceptibility with , in direct contradiction to expectations near a genuine second-order phase transition [6, 7]. These findings establish that the high- 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 as a function of using a stationary-distribution criterion appropriate for noisy systems [8, 9]. The resulting curve is approximately flat, with mean values between 2.5 and 3.2 time units across the full range and no systematic dependence on nonlinearity strength. This insensitivity of the relaxation timescale stands in sharp contrast to the sixfold decrease of over the same 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 interacting components whose states are represented by the vector constrained to the -dimensional probability simplex,
Each component can be interpreted as the relative abundance of type in a population, the market share of agent , or more generally the probability weight assigned to option in a finite mixture. The simplex geometry ensures that competition is inherently zero-sum: an increase in 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 is
where the equality follows from the simplex constraint. The sign structure of 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 (), 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
where is the nonlinearity exponent. The baseline model (2) is recovered at . For the positive feedback is superlinear: high-abundance components receive a disproportionately amplified advantage, making competitive exclusion mechanistically stronger. The exponent 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 for all is a fixed point of the deterministic dynamics (3) for any . Linearizing around this point, the Jacobian of the drift restricted to the simplex tangent space takes the form
where is the orthogonal projector onto the zero-sum subspace (the tangent space of ). On this subspace acts as the identity, so has a single eigenvalue
with multiplicity . Since , the uniform state is linearly unstable: small perturbations grow at rate , driving the system toward the simplex vertices (diversity collapse).
Two features of (5) connect directly to the numerical results reported below: (i) at fixed , the instability rate decreases rapidly with (e.g., from 0.05 at to 0.005 at for ), providing a first-principles explanation for the self-stabilization of the high-entropy state at high (§ 4); (ii) at fixed , as , explaining why large systems self-stabilize without noise (§ 3.2). The generalized analytical bound derived from (5) is presented in Appendix § 9.
Stochastic differential equation and numerical integration
Additive Gaussian noise of uniform intensity is superimposed on the deterministic drift, yielding the Itô stochastic differential equation
where 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 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],
with time step . After each update, simplex membership is enforced by a two-step projection: (i) all negative entries are clipped to a small positive floor , and (ii) the vector is renormalized to unit sum,
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 over several orders of magnitude. Constrained SDEs on the simplex admit several well-posed numerical treatments: elimination of one component via the constraint and integration in reduced coordinates (e.g. the log-ratio coordinates used by Imhof [12] for stochastic replicator dynamics on ), 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 components symmetrically and handles the boundary uniformly via the floor . The floor also prevents the term from producing numerical underflow in the entropy calculation; since contributes at most per component to the raw entropy sum, the systematic bias on is negligible at all values explored.
Observables
Normalized Shannon entropy
The diversity of the system state is quantified by the normalized Shannon entropy
where the prefactor maps the range of onto the unit interval: corresponds to a fully ordered (single-dominant) corner state, and to the uniform distribution for all . The normalization is essential when comparing results across system sizes, because the raw Shannon entropy grows with at fixed diversity level, introducing a spurious size dependence that would systematically bias estimates of the critical noise intensity toward smaller values for larger . All entropy values reported in this paper refer to the normalized form (9).
Critical noise intensity
For a given system size and nonlinearity exponent , the critical noise intensity is defined as the smallest value of for which the long-time average of exceeds a prescribed diversity threshold ,
where denotes a steady-state time average taken after a burn-in period of at least integration steps. We set throughout; results are qualitatively insensitive to moderate variations of this threshold. The quantity 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 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
in direct analogy with the magnetic susceptibility used in finite-size scaling analysis of spin systems [6, 7]. Here is the variance of computed over the steady-state trajectory. Physically, a large 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 indicates that all realizations converge to a similar entropy, signaling robust self-organization. The factor of rescales the susceptibility so that a genuine second-order phase transition would manifest as a divergence in the thermodynamic limit, where and are standard critical exponents, while a smooth crossover or finite-size artefact would produce a that either saturates or decreases with increasing [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 measures how quickly the macroscopic entropy reaches its stationary distribution from a random initial condition drawn uniformly on .
A natural but physically incorrect approach would be to declare relaxation complete once the microscopic state vector ceases to change appreciably, e.g., when . In a noise-driven system, however, individual components 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 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 integration steps. At each window boundary we compute the window-averaged entropy for the current window and the preceding window . Relaxation is declared when the absolute difference remains below a tolerance for consecutive window pairs,
This criterion is immune to microscopic noise fluctuations because is smooth, -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 , , and were verified to produce qualitatively identical 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 and the nonlinearity exponent are varied systematically as described below for each group of experiments; all other parameters are held fixed.
| Parameter | Symbol | Value |
| Integration time step | ||
| Total simulation steps | ||
| Burn-in steps | ||
| Window size (relaxation) | 100 steps | |
| Stable windows required | 5 | |
| Entropy convergence tol. | ||
| Simplex floor | ||
| Diversity threshold | 0.7 | |
| Independent trials | — | (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 and steps (total simulation time 60 units) to ensure macroscopic convergence; all other experiments use the baseline . The window parameters are defined in integration steps, so their physical durations scale with : e.g., 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 and the noise intensity is scanned over a uniform grid of with step . For each pair, five independent trials are run from distinct random initial conditions drawn uniformly on ; the reported is the mean over trials and over the post-burn-in trajectory. Note that lies in the deterministically self-stabilized regime (§ 3.2) and is therefore excluded from the power-law fit; only are used for the scaling analysis. The universality test uses the same grid with the interaction replaced by the superlinear form (3) at .
Phase diagram and fluctuation analysis (§ 4). The nonlinearity exponent and noise intensity are scanned jointly over (14 uniformly spaced values) and (35 uniformly spaced values), with fixed. Twenty independent trials are run per parameter pair; the fluctuation susceptibility is computed from the pooled post-burn-in entropy time series across all trials. For the finite-size robustness check, the same grid is repeated at , also with twenty trials per point. Numerical stability with respect to the integration step is verified by repeating the grid at .
Relaxation dynamics (§ 5). The nonlinearity exponent is varied over (15 uniformly spaced values) at fixed and , a value within the range where the susceptibility peak is observed (§ 4). For these runs a coarser time step 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 value to obtain reliable estimates of the mean and standard error of ; 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 ( vs , Supplementary S5); (ii) finite-size scaling across (Supplementary S6); (iii) sensitivity of to the window parameters and , verified to produce qualitatively identical curves when varied by factors of two; and (iv) insensitivity of the simplex-floor parameter over several orders of magnitude, with the resulting entropy bias bounded at . 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 , . Figure 1a shows representative trajectories in the two-dimensional simplex for a deterministic run () and a stochastic run (), both started from the near-uniform initial condition . Under deterministic dynamics the system follows a direct path toward a vertex, reaching the diversity-collapse state within approximately integration steps. Under stochastic dynamics the trajectory explores the interior of the simplex throughout the full simulation window of steps, with no discernible drift toward any vertex.

Figure 1. Basic phenomenology of stochastic diversity stabilization (, ). (a) Representative trajectories on the two-dimensional simplex for deterministic dynamics (, left) and stochastic dynamics (, right). The deterministic trajectory collapses to a vertex within steps; the stochastic trajectory explores the simplex interior throughout steps. (b) Steady-state mean normalized entropy as a function of noise strength . The monotonic increase confirms that noise systematically promotes diversity; saturation near reflects the geometric ceiling of the simplex.
The dependence of long-time diversity on noise intensity is quantified in Figure 1b, which shows the steady-state mean entropy as a function of for , . The curve rises monotonically from at and saturates near for . 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 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 for each under the baseline interaction ().
Figure 2 shows all seven response curves together. Two qualitatively distinct behaviors are visible. For small (), the entropy at zero noise is substantially below the high-diversity region (–), and increasing produces a pronounced, steep rise in : noise is the primary agent driving the system away from the ordering tendency of the deterministic drift. For large (), the entropy at is already elevated (–), and the response to increasing 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 to deterministically self-stabilized at large —is the central finding of this section and connects to the large- self-organization discussed in evolutionary game dynamics [3].

Figure 2. Noise–entropy response curves for system sizes under the baseline interaction (; 10 trials per point). Dashed vertical lines mark the inflection point (location of steepest response) for each . Small systems (, dark tones) show a pronounced noise-driven entropy rise from a low baseline, indicating that external noise is the primary stabilizing agent. Large systems (, light tones) already maintain high entropy at ; 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- regime, the noise intensity at the steepest point of the response curve (the inflection point , defined as ) decreases with . For a preliminary power-law fit to the inflection points gives
as detailed above (four data points, ; on the log-log fit). (13) is offered as an empirical description of the data only; we do not extrapolate it to larger , where the system self-stabilizes and loses physical meaning (see below). This exponent is obtained from the inflection-point definition of (the noise intensity at steepest response). An alternative threshold-crossing definition (), used in the universality test of § 3.3 for computational efficiency, yields at ; the two values (1.4 vs 1.1) are of the same order, and the key qualitative result—that depends on the interaction form—is robust to the definition choice (§ 3.3). This approximate scaling breaks down at larger . At large , the instability rate of the uniform fixed point (equation (5)) vanishes: for , , 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 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 and repeat the measurement for the same set of system sizes, averaging over five independent trials per parameter point to suppress single-run noise.
Figure 3a shows for the superlinear interaction at . In contrast to the baseline model, the decay is steeper: drops from at to at , yielding a fitted exponent , 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 decay, meaning that larger systems become self-stabilizing more rapidly under superlinear interactions.

Figure 3. Model-specificity of the critical noise scaling. (a) Critical noise intensity versus system size for the superlinear interaction (3) at (five trials per point). The curve is monotonically decreasing with fitted exponent , steeper than the baseline value . (b) Fitted scaling exponent as a function of the nonlinearity parameter . The exponent remains near for (with a non-monotonic dip at ), then rises to at ; 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 at each value of and plot the resulting in Figure 3b. The exponent remains near for (ranging from 1.08 at to 1.28 at , with a non-monotonic dip to 0.99 at ), then rises sharply to at . The overall trend is an increase from to , but the low- plateau and the non-monotonicity at indicate that the -dependence of is not a simple smooth function.
These results indicate that 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 to indicates that the finite-size scaling of reflects the specific way the nonlinear drift balances positive feedback against the simplex constraint at each interaction strength. The structured -dependence of is partially captured by the analytical bound derived in Appendix B: linearization of the drift at the uniform fixed point yields the prediction (equation (B.4)), which increases linearly from 1.5 at to 2.0 at . The numerical values lie systematically below this prediction at low (where the instability is strong and linearization is least accurate) but converge to it at high ( vs at ), consistent with the weakening of the instability rate 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 and the noise strength . For each point on a grid covering and , we compute the steady-state mean normalized entropy at , averaging over five independent trials and the post-burn-in trajectory. The result is displayed as a heat map in Figure 4a.

Figure 4. Phase structure and fluctuation analysis at . (a) Dynamical phase diagram in the plane; color encodes . The transition from low to high entropy is a smooth crossover. (b) Heat map of the entropy variance in the plane. The global maximum occurs at low ; variance decreases as increases. (c) Fluctuation susceptibility map; peak value at . (d) Peak susceptibility versus nonlinearity (maximum over all at each ; trials = 20). The curve falls on average from at to at ; no divergence or pronounced peak is observed near , 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 () the entropy is generally depressed across the entire range of explored, with the most pronounced ordering occurring at small where the deterministic collapse tendency is least counteracted. As increases beyond approximately 1.0–1.4, the entropy rises progressively and eventually saturates near its maximum value for , largely irrespective of . In this high- regime the system maintains a high-diversity stationary state even in the absence of noise (), 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 , confirming the absence of any sharp structural feature. The grid spacing in is approximately 0.115, so reported peak positions carry an intrinsic resolution uncertainty of ; 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 peak are sensitive to the random-seed ensemble and the parameter grid, and vary by in and in amplitude across independent runs. The conclusions of this section depend only on two robust features that are stable across all runs: (i) does not diverge with , and (ii) is strongly suppressed for . The system therefore exhibits a smooth crossover rather than a phase transition as 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 – 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 over the same 14 35 parameter grid used in § 4.1.
Entropy variance and susceptibility maps. Figure 4b shows the variance as a heat map in the plane. The global maximum of is located in the low- region near , , and the variance decreases progressively as increases toward 2.0. The susceptibility map shown in Figure 4c reproduces the same spatial structure—as expected, since the two quantities differ only by a factor of —with a peak value of at . Notably, neither map shows any concentrated bright region near , the location where the phase diagram changes appearance most visibly.
Peak susceptibility as a function of . To make this point precise, Figure 4d shows , defined as the maximum of over all values at each fixed . The curve attains its global maximum of at and decreases on average as increases, falling to by (a reduction by a factor of approximately six). Local revivals appear at several intermediate values, but all remain well below the primary peak and are associated with the stochastic variability of the low- regime identified in § 4.3. There is no peak, kink, or inflection point of any significance at .
Absence of critical divergence: a key negative result. The foregoing observations constitute a direct test of the criticality hypothesis. If the crossover near were a genuine second-order phase transition, the susceptibility should exhibit a pronounced maximum or divergence at the transition point, growing with system size as in the spirit of finite-size scaling [16, 6]. The data show the opposite: the susceptibility peaks at with and falls by a factor of approximately six by ; the value at is less than one-third of the low- maximum. While the susceptibility peak at may superficially resemble a critical point, its lack of divergence with (§ 4.3) and the insensitivity of the relaxation timescale to (§ 5) are inconsistent with equilibrium or nonequilibrium universality classes [15, 16]. We conclude that no critical phase transition occurs along the axis within the parameter range explored. Instead, the system enters an entropy-saturated nonlinear phase for , 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 ), the peak susceptibilities at 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 increases. This is inconsistent with the divergent growth expected near a genuine second-order phase transition [6, 7]. Furthermore, in all three cases the high- region () shows strongly suppressed , 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 (see also Table 2). Regime I (): the deterministic drift biases the system toward low-entropy states; is large () but does not diverge with , identifying a finite-size effect rather than true criticality. Regime II (): a smooth crossover in which the stationary probability mass redistributes from the ordering corners toward the simplex interior; declines without any local extremum. Regime III (): the mean entropy saturates near unity independently of , and 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 ( range) | |||
| I: ordering-prone () | large () | ||
| II: crossover (–) | declining | ||
| III: entropy-saturated () | small () |
Table 2. Summary of the three dynamical regimes. : peak fluctuation susceptibility over at each . : mean relaxation time (20 trials). : 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 defined in equation (12) as a function of at fixed and , a noise level within the range where the susceptibility peak is observed.
Figure 5a shows computed from five independent trials. The curve fluctuates between 2.2 and 3.4 time units with no systematic trend as increases from 0.25 to 2.0.

Figure 5. Entropy relaxation time as a function of nonlinearity at , . (a) Mean of five independent trials; the curve fluctuates between 2.2 and 3.4 time units with no systematic dependence on . (b) Mean one standard error over twenty independent trials. The grand mean is across all ; the curve shows no monotonic trend, in contrast to the sixfold variation of over the same range (Figure 4d). (c) Boxplot of per-trial relaxation times. The interquartile range is comparable across all values, confirming that neither the central tendency nor the variability of depends on nonlinearity strength.
To obtain statistically reliable estimates, Figure 5b shows averaged over twenty independent trials, with error bars representing one standard error. The mean curve confirms the absence of a systematic trend: fluctuates between 2.5 and 3.2 with a grand mean of approximately 2.9 across all values, and no monotonic increase or decrease is discernible. The trial-to-trial spread is substantial throughout (standard deviation –), reflecting persistent variability at all nonlinearity strengths.
Figure 5c displays the raw per-trial relaxation times as a boxplot. At every 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 range, confirming that the variability structure is insensitive to nonlinearity strength.
The central finding of this section is therefore negative: does not mirror the dramatic -dependence of . While the susceptibility decreases approximately sixfold from low to high (§ 4.2), the relaxation timescale remains essentially constant.
Decoupling between relaxation rate and fluctuation suppression
The insensitivity of to , combined with the strong -dependence of , 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 at high 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: drops sixfold while remains unchanged at 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 and decreasing has a natural interpretation in terms of the law of total variance. Conditioning on the microscopic relaxation pathway ,
Because each trajectory takes approximately the same time to reach stationarity regardless of , the rate at which individual trajectories explore the configuration space is not enhanced by nonlinearity. What changes is the endpoint distribution: at high , different trajectories—following qualitatively different microscopic paths (as evidenced by the persistent trial-to-trial variability in )—converge to macroscopic states with similar entropy values. In the language of equation (14), the second term is suppressed because different pathways produce similar conditional means , 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 to be independent of while decreases. If instead were found to decrease proportionally with (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 and across the three dynamical regimes.
Mechanistic notes. All regimes: The relaxation timescale is approximately constant across all three regimes, while 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 column, which varies sixfold across regimes, and the column, which remains constant. This decoupling is the primary evidence for the pathway-multiplicity mechanism: fluctuation suppression in the high- 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- scaling exponent 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 to , with a non-trivial -dependence. The nonlinearity exponent governs the type of collective behavior—ordering-prone at low , smoothly crossing over to entropy-saturated at high —without producing a divergent susceptibility or a -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 plays the role of environmental fluctuation strength and the normalized entropy corresponds directly to the Shannon diversity index routinely measured in that field; the low-, 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-, 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 (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 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 . The present system shares the qualitative feature that noise promotes a more uniform (high-entropy) stationary distribution, and the critical noise intensity 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 , and its apparent location in the 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 , and this divergence can be detected even in small systems by tracking how the peak susceptibility grows with . The results of § 4.3 demonstrate that this diagnostic fails in the present system: matched-trial (trials = 20) measurements at yield peak values of 0.7, 0.6, and 0.5—values within the same order of magnitude, with a slight decrease with . This failure of FSS is itself informative—it identifies the fluctuation peak at low 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 . 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 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 , 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- trend— as —but the mechanism differs. In evolutionary game theory the noise is intrinsic (demographic), scales as , and its effect diminishes as grows because the law of large numbers suppresses relative fluctuations. Here the noise is external and additive, and the diminishing requirement for noise as increases reflects a growing self-organizing capacity of the deterministic simplex dynamics, not a dilution of stochastic effects. The power-law exponent quantifies this self-organizing efficiency for the baseline replicator interaction, and its sensitivity to the interaction form (§ 3.3) suggests that the large- self-stabilization is a model-specific property rather than a generic feature of competitive dynamics on the simplex.
Limitations and outlook
The scaling exponent applies only for ; at larger the system self-stabilizes. The analytical prediction (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 and —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 to obtain a sharper 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 scaling is well-defined.
Conclusion
We have studied stochastic diversity stabilization in a class of replicator-like competitive dynamics on the -dimensional probability simplex, parameterized by an additive noise strength and a nonlinearity exponent . Under the baseline linear replicator interaction, the critical noise intensity required to sustain a diverse stationary state decays as a power law, , implying that sufficiently large systems self-stabilize without external noise. Replacing the baseline with a superlinear variant shifts the exponent from to as increases, in quantitative agreement with the linear-stability prediction and establishing that is a model-dependent quantity rather than a universal critical index. Scanning the full parameter plane reveals a smooth dynamical crossover from an ordering-prone, noise-dependent regime at low to an entropy-saturated, self-organizing phase at high , with no divergent fluctuation susceptibility at any point along the transition: the peak of is located at , well separated from the visual crossover region near , and matched finite-size analysis finds no systematic divergence with , ruling out a genuine second-order phase transition in the thermodynamic limit. Relaxation dynamics complete the picture: the entropy relaxation time is approximately constant at time units across the full 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 . The joint behavior of and 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 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,
with . 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 contains a stochastic contribution of order – for the parameter values used here, which permanently exceeds 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 steps, corresponding to simulation time .
Applying this criterion to the same parameter settings used in Section 5 (, , ) produces a perfectly flat for all —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 is two orders of magnitude below the ceiling, confirming genuine convergence to stationarity (Figure 5).
Analytical bound on
We derive a scaling bound on the critical noise intensity as a function of both system size and nonlinearity exponent , using the linear stability result (5) from § 2.1.
Step 1: Heuristic stationary variance
The uniform fixed point is linearly unstable with instability rate (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 ), the diffusive spreading (intensity ), and the confining effect of the simplex boundaries to obtain the characteristic displacement scale, yielding a variance estimate
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 and . All conclusions drawn from this appendix rest solely on the resulting scaling exponent and not on the prefactor: the simplex boundary, not the linearized rate , sets the true confining scale, so the factor should be read as fixing the - and -dependence rather than the absolute coefficient.
Step 2: Entropy lower bound
A second-order Taylor expansion of the normalized Shannon entropy (9) around gives, using ,
Requiring and inverting:
The effective scaling exponent is therefore
which predicts a linear increase of with .
Comparison with numerical results
Equation (B.4) makes two testable predictions. First, for the baseline interaction (), ; the numerical estimates (inflection-point definition) and (threshold-crossing definition) are of the same order of magnitude, with the discrepancy attributable to finite- effects and the heuristic nature of (B.1). Second, for , ; the threshold-crossing fit gives , in close quantitative agreement. More generally, the predicted linear trend captures the observed increase of with (Figure 3b), with the analytical values systematically above the numerical ones at low and converging at high . This convergence at high is consistent with the weakening of the instability rate , 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.

Figure S4. Fluctuation-based dynamical boundary , defined as the noise intensity at which is maximized over at each fixed (). The boundary is approximately stable for but becomes highly erratic for , 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- 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 .

Figure S5. Numerical stability check: comparison of results at (left column) and (right column), both at . Top row: steady-state mean normalized entropy in the plane; color ranges differ by less than 2%. Bottom row: fluctuation susceptibility ; the qualitative structure—peak at low , suppression at high —is reproduced at both time steps. The shift in peak position reflects seed- and grid-dependent variability documented in the main text, not numerical error.

Figure S6. Finite-size verification of fluctuation suppression. Each column corresponds to a system size (left to right). Top row: steady-state mean normalized entropy in the plane. Bottom row: fluctuation susceptibility . All three scans use trials = 20. Peak values are (, ), (, ), and (, ); no systematic increase with is observed, ruling out critical divergence [6, 7]. In all three panels, is strongly suppressed at , confirming size-independent fluctuation suppression in the high-nonlinearity regime.
Fluctuation-based dynamical boundary
For each value of the nonlinearity exponent , the noise intensity at which the fluctuation susceptibility attains its maximum over defines a fluctuation-based dynamical boundary in the plane. Figure S4 shows this boundary at .
In the low- regime (), the boundary is relatively stable, with confined to a narrow low-noise band and showing no clear trend. For , the boundary becomes highly erratic, with 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- regime is uniformly low and nearly flat across all (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 . To verify that the conclusions are not sensitive to this choice, we repeat the phase diagram and fluctuation susceptibility analysis at and compare the two outputs directly. Figure S5 shows the mean entropy (top row) and the susceptibility (bottom row) for both time steps at .
The mean entropy maps (top row) are qualitatively identical: both show the same progression from low entropy at small to high, -independent
entropy at large , with the same approximate location of low-entropy pockets near . The color ranges differ by less than 2% in the dynamic range.
The susceptibility maps (bottom row) confirm that the spatial structure of —peak at low , suppression at high —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 to achieve a total simulation time of 60 units within 6000 steps. A separate convergence check is not required for the relaxation observable, for two reasons. First, is defined via a running-window average of the macroscopic entropy over 100-step windows (= 1.0 time unit at ), which smooths out step-level integration errors by averaging over individual entropy contributions. Second, the measured 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 ), not step-level precision. To confirm this directly, we repeated the relaxation analysis at (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 is at and 2.81 at , a relative difference of only 3%. Individual points differ by up to , 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–). The near-cancellation of these fluctuations in the grand mean indicates that the systematic effect is negligible once trial-level scatter is averaged out. The qualitative conclusion—that is approximately constant across —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 persists as the system size grows, we repeat the susceptibility scan at . Figure S6 shows the mean entropy (top row) and the fluctuation susceptibility (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:
| peak | ||
| 10 | 0.7 | 1.08 |
| 20 | 0.6 | 0.7 |
| 40 | 0.5 | 0.62 |
Table 3.
The three values lie within the same order of magnitude and do not show a systematic increase with ; indeed, the peak decreases slightly from to . A genuine second-order phase transition would require to grow as [6, 7]; the absence of such growth rules out critical behavior.
High- suppression is size-independent.. In all three panels (Figure S6, bottom row), the susceptibility in the high- regime () is strongly suppressed relative to the low- 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 values (from at to at ), consistent with the dynamical variability discussed in the main text and not indicative of a fixed critical point.
References
- [1]J. Hofbauer, K. Sigmund, Evolutionary Games and Population Dynamics, Cambridge University Press, Cambridge, 1998. doi:10.1017/CB09781139173179.
- [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]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]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]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]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]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]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]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]P. E. Kloeden, E. Platen, Numerical Solution of Stochastic Differential Equations, Springer, Berlin, 1992. doi:10.1007/978-3-662-12616-5.DOI
- [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]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]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]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]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]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]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]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]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]F. Ritort, P. Sollich, Glassy dynamics of kinetically constrained models, Advances in Physics 52 (4) (2003) 219–342. doi:10.1080/0001873031000093582.DOI
- [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]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]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]V. Privman (Ed.), Finite Size Scaling and Numerical Simulation of Statistical Systems, World Scientific, Singapore, 1990. doi:10.1142/1011.DOI
- [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]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