betaFeedback

WorksPhysics and Astronomy

Resonant Amplification of Demographic Noise in Stochastic Common-Pool Resource Dynamics

Abstract

Common-pool resources are vulnerable to overexploitation, yet many social-ecological systems exhibit persistent oscillations. We introduce a stochastic eco-evolutionary model in which harvesting strategies and resource availability shape one another. In the infinite population limit the system approaches a stable state through damped oscillations. Demographic noise continually excites these modes, generating instead robust macroscopic quasi-cycles, whose frequencies depend on the noise ratio, the relative timescale between strategy and resource, and the harvesting pressure. The model reproduces the distinct frequency peaks, observed in empirical global fish-landings data for high- and low-income countries, showing that persistent environmental fluctuations can emerge naturally from demographic noise.

Common-pool resources (CPRs) are finite and nonexcludable shared resources. These characteristics make them vulnerable to overexploitation, potentially leading to the so-called “tragedy of the commons” [1]. As highlighted by Kobeissi et al. [2] and others [3, 4], in CPRs individual incentives promote overexploitation over coordinated collective benefit, ultimately driving resource depletion. However, this tragedy is not inevitable. The seminal works by Nobel Prize laureate Elinor Ostrom [5] alongside recent theoretical and behavioral experiments [6], demonstrates that introducing shared objectives can override short-term self-interest and sustain cooperation. Real-world CPR dynamics are governed by complex human-ecological feedbacks. Classical examples includes the alternating dominance of sardine and anchovy stocks observed in the Pacific [7, 8], as well as historical fishery collapses, where intensive harvesting interacts with ecological cycles [9, 10]. When resource-use strategies depend on both social interactions and environmental states, their coevolution is naturally captured by coupling evolutionary game theory (EGT) with resource dynamics [6, 11, 12, 13, 14]. The resulting, so-called eco-evolutionary (ECO-EVO) models comprise three key components: game structure, resource dynamics, and environment-dependent strategy adoption. Crucially, in ECO-EVO models the payoff matrix is no longer static; incentives to cooperate (preserve) or defect (overexploit) shift dynamically with resource availability. A standard formulation represents these state-dependent payoffs as a linear combination of games corresponding to depleted and abundant environments [11, 12, 13]. Despite their simplicity, ECO-EVO models exhibit a rich dynamical repertoire, including bistability, strategy coexistence, and sustained oscillations characterizing recurrent tragedy-of-the-commons cycles. Most ECO-EVO models couple deterministic replicator equations with resource dynamics. However, empirical observables like fish productivity [15, 16], forests growth and supply [17] and freshwater systems [18] indicates that deterministic dynamics alone cannot account for observed resource dynamics. While previous works have highlighted resonantly amplified quasi-cycles driven by demographic noise [19] or explored extrinsic environmental fluctuations [20], they decoupled stochasticity from the intricate feedback between strategy evolution and resource dynamics. Recently, this feedback has been integrated into numerical models exploring knowledge feedback and network topology [21], and into system-level dynamical frameworks to derive analytical conditions for stochastic asymptotic stability [22]. However, these approaches formulate stochasticity directly at the macroscopic scale by introducing independent noise processes. Consequently, the role of intrinsic, finite-size stochasticity within this bidirectional feedback system remains largely unexplored. Here, we bridge this gap, by introducing a stochastic ECO-EVO model of common-pool resource dynamics in which, rather than postulating macroscopic noise, strategy updates, resource harvesting, and resource regeneration are all treated as coupled stochastic processes at the individual level. Starting from a microscopic description, we systematically derive the macroscopic (Langevin) dynamics of the coupled agent–resource system, allowing us to move beyond purely numerical treatments and quantify analytically how demographic noise reshapes the stability and sustainability of the coupled social-ecological system.

We also show that our ECO-EVO model reproduces the variation in oscillation frequencies observed in fish-landings data across countries, linking this variation to differences in harvesting pressure and adaptation timescales between high- and low-income countries.

Stochastic eco-evolutionary model.

Consider a population of NN agents, where each individual adopts one of two strategies: cooperation, cc, or defection dd. These behaviors determine how individuals interact with one another and how they exploit a shared, renewable resource, which coevolves with the population of strategies. In line with standard treatments in chemical kinetics [23], we model the resource as a discrete quantity to facilitate the stochastic description that follows. Hence, at any given time, the system is defined by a population of CC cooperators and DD defectors, where C+D=NC+D=N, together with a resource level R∈[0,1,2,…,K]R\in[0,1,2,\ldots,K], where KK denotes the carrying capacity. To quantify how, in our individual-based model, different interactions impact the resource, we introduce the harvest-interaction matrix E={(ecc,ecd),(edc,edd)}E=\{(e_{cc},e_{cd}),(e_{dc},e_{dd})\}, whose entry eije_{ij} represents the per-capita extraction rate exerted by an individual of strategy i∈{c,d}i\in\{c,d\} when interacting with one of strategy j∈{c,d}j\in\{c,d\}. Following Tu et al. [6], cooperators are defined by extraction rates 0<ecc<b0<e_{cc}<b, with bb being the intrinsic resource regeneration rate, which allow the resource to renew faster than it is consumed. Defectors, by contrast, satisfy edd>be_{dd}>b, so the resource becomes extinct at stationarity if all agents choose to defect. The total extraction rates contributed by a cooperator and a defector in a well-mixed population of size NN are, respectively:

hc(R,C,D)=RN(N−1)[ecc(C−1)+ecdD],h_c(R,C,D)=\frac{R}{N(N-1)}\left[e_{cc}(C-1)+e_{cd}D\right],
hd(R,C,D)=RN(N−1)[edcC+edd(D−1)].(1)h_d(R,C,D)=\frac{R}{N(N-1)}\left[e_{dc}C+e_{dd}(D-1)\right]. \tag*{(1)}

where we have assumed that the harvested units scale with the available resource RR, and the normalization 1/[N(N−1)]1/[N(N-1)] reflects the even distribution of RR across all pairwise interactions, weighted by the extraction coefficients eije_{ij}. The resource depletion is then modeled as a stochastic process occurring at a total rate hcC+hdDh_cC+h_dD,

r→hcC+hdDvr\xrightarrow{h_cC+h_dD}v

where the lower-case letters denote the type of species occupying a given site: rr represents a unit of resource, and vv represents an empty (available) site, as would naturally arise in a spatial formulation. Resource regeneration is described as a microscopic birth process occurring at rate bb,

r+v→br+r.r+v\xrightarrow{b}r+r.

Formally, this leads naturally to logistic growth. The finite carrying capacity KK prevents unbounded growth, as the resource can only occupy the K−rK-r currently empty sites.

Strategy updates occur also stochastically and depend both on agents’ interactions and on the current resource level, thereby incorporating the environmental feedback. Following the knowledge-feedback mechanism introduced in [21], combined with imitation dynamics, agents are more likely to switch to defection when others are overexploiting the resource and the resource stock is abundant. Conversely, resource scarcity together with widespread cooperation increases the propensity to adopt a cooperative strategy. Specifically, we assume that the rate of switching to defection is proportional to the resource stock RR, whereas the rate of switching to cooperation increases when resources are scarce and most agents contribute to their preservation. The two strategy-update processes are therefore c→νR′Ndc\xrightarrow{\frac{\nu R'}{N}}d and d→ν(1−R′)+C/Ncd\xrightarrow{\frac{\nu(1-R')+C/N}{}}c where ν\nu sets the strategy-update timescale, R′=R/[K(1−ecc/b)]R'=R/[K(1-e_{cc}/b)] is the level of resource RR rescaled by the effective carrying capacity [21], and (u)+≡max⁡(0,u)(u)_+\equiv\max(0,u) prevents negative rates for R′>1R'>1 (see SM). The dynamics of the stochastic ECO-EVO model can be investigated numerically using the Gillespie algorithm [24]. Figure 1(a) reports the fraction of cooperators xx and the resource level ρ\rho (solid line) as functions of time for a single realization of the process. Results clearly reveal sustained oscillations with a frequency whose ν\nu dependence is shown Figure 1(b)(red symbols).

Stochastic ECO-EVO model trajectories and peak-frequency dependence

Figure 1. (a) A single trajectory of the stochastic ECO-EVO model with K=1000K=1000, N=200N=200, and ecc=ecd=0.5e_{cc}=e_{cd}=0.5, edd=edc=1.5e_{dd}=e_{dc}=1.5, b=ν=1b=\nu=1, numerically simulated (solid lines), reveals sustained quasi-cycles in the cooperator fraction xx and the resource level ρ\rho. In the infinite-population limit of Eq. (2) the system undergoes damped oscillations (dash-dotted lines) around the interior fixed point given by (4) (horizontal dotted line). (b) Peak frequency fpeakf_{\mathrm{peak}} of the power spectrum of ρ\rho as a function of the timescale ν\nu. The solid line is the LNA prediction; red symbols show the mean over n=100n=100 Gillespie runs of length tmax⁡=500t_{\max}=500 (after discarding an initial transient), with error bars indicating one standard deviation. Other parameters as in (a).

Analytical description.

To uncover the origin of these oscillations, we derive an analytical description of our model within the master-equation framework [23]. From the microscopic processes of the model we first construct nonnegative transition rates between states (C,D,R)(C,D,R), with transitions outside the physical domain 0≤C,D≤N0\le C,D\le N and 0≤R≤K0\le R\le K forbidden. These rates determine the master equation for the state probability P(C,D,R,t)P(C,D,R,t) of the system to have CC cooperators, DD defectors and RR resource level at time tt. Introducing the cooperator fraction x=C/Nx=C/N and the scaled resource level ρ=R/K\rho=R/K, the diffusion approximation yields the Itô-Langevin equations

dρdt=ρ[b(1−ρ)−g(x)]+Bρ(ρ,x)K ηρ(t),(2)\frac{d\rho}{dt}=\rho\left[b(1-\rho)-g(x)\right]+\sqrt{\frac{B_\rho(\rho,x)}{K}}\,\eta_\rho(t), \tag*{(2)}
dxdt=νx(1−x)(1−2ρ′)+Bx(ρ,x)N ηx(t),\frac{dx}{dt}=\nu x(1-x)(1-2\rho')+\sqrt{\frac{B_x(\rho,x)}{N}}\,\eta_x(t),

where ν\nu control the relative timescale of strategy updates and resource dynamics, ρ′=ρ/(1−ecc/b)\rho'=\rho/(1-e_{cc}/b), and

g(x)=eccx2+(ecd+edc)x(1−x)+edd(1−x)2g(x)=e_{cc}x^2+(e_{cd}+e_{dc})x(1-x)+e_{dd}(1-x)^2

is the harvest pressure. In the diffusion terms, the coefficients obtained from the second jump moments of the microscopic dynamics, are

Bρ(ρ,x)=ρ[b(1−ρ)+g(x)],B_\rho(\rho,x)=\rho\left[b(1-\rho)+g(x)\right],
Bx(ρ,x)=νx(1−x).B_x(\rho,x)=\nu x(1-x).

while ηρ\eta_\rho and ηx\eta_x are independent Gaussian white noises with zero mean correlations ⟨ηi(t)ηj(t′)⟩=δijδ(t−t′)\langle\eta_i(t)\eta_j(t')\rangle=\delta_{ij}\delta(t-t'). Resource and population fluctuations therefore scale as K−1/2K^{-1/2} and N−1/2N^{-1/2}, respectively. Eq. (2) holds for ρ′≤1\rho'\leq1; the full expressions for ρ′>1\rho'>1 and derivation details are given in the SM

Infinite-population limit.

In the large-system limit, (K,N→∞(K,N\to\infty at fixed K/N)K/N), fluctuations vanish, and Eq. (2) reduces to two coupled deterministic equations. These admit three boundary fixed points and one interior equilibrium in the admissible state space Ω=[0,1]×[0,1]\Omega=[0,1]\times[0,1]. The boundary fixed points are:

(ρ∗,x∗)∈{(0,0),(0,1),(1−eccb,1)}.(3)(\rho^*,x^*)\in\left\{(0,0),(0,1),\left(1-\frac{e_{cc}}{b},1\right)\right\}. \tag*{(3)}

The first two correspond to resource extinction under pure defection and pure cooperation, whereas the third describes a fully cooperative population sustaining the resource (ecc<be_{cc}<b). Because edd>be_{dd}>b, the corresponding pure-defection state (1−edd/b,0)(1-e_{dd}/b,0) is unphysical (ρ∗<0\rho^*<0), leaving (ρ∗=0,x∗=0)(\rho^*=0,x^*=0) as the sole equilibrium on the x=0x=0 boundary (the “tragedy of the commons”). The interior equilibrium is

(ρ∗,x∗)=(12(1−eccb),x~),(4)(\rho^*,x^*)=\left(\frac{1}{2}\left(1-\frac{e_{cc}}{b}\right),\tilde{x}\right), \tag*{(4)}

where x~∈(0,1)\tilde{x}\in(0,1) satisfies

12(1−eccb)=1−g(x~)b.(5)\frac{1}{2}\left(1-\frac{e_{cc}}{b}\right)=1-\frac{g(\tilde{x})}{b}. \tag*{(5)}

Because defectors increase harvesting (g(x)≥eccg(x)\geq e_{cc}), the resource decreases whenever ρ′≥1\rho'\geq1 (ρ˙≤0\dot{\rho}\leq0). Trajectories are thus confined to ρ′≤1\rho'\leq1, and since the interior equilibrium sits at ρ′∗=1/2\rho'^*=1/2 and excursions into ρ′>1\rho'>1

are exponentially rare (see SM). At (ρ∗,x∗)(\rho^*,x^*), the Jacobian has strictly negative trace, τ=−bρ∗<0\tau=-b\rho^*<0, excluding both a Hopf bifurcation and a center. More generally, the Bendixson–Dulac theorem rules out periodic orbits in the interior of Ω\Omega for the deterministic dynamics [25, 26] (see SM). The equilibrium is therefore always stable, and approached either monotonically or through damped oscillations, depending on whether τ2>4Δ\tau^2>4\Delta (stable node) or τ2<4Δ\tau^2<4\Delta (stable spiral), where Δ\Delta is the determinant of the Jacobian. This spiral relaxation is illustrated by the solutions of the coupled deterministic equations, reported in Figure 1(a) (dash-dotted lines). Figure 2 further illustrates how the nature of the interior fixed point changes as a function of the extraction rate of two defectors, edde_{dd}, and of that of two cooperators, ecce_{cc}. Without loss of generality, we assume homogeneous interaction strengths: edc=edde_{dc}=e_{dd} and ecd=ecce_{cd}=e_{cc} [21]. Two different values of the relative timescale ν\nu are considered. When strategy update is slow compared to the resource dynamics (ν≪1\nu\ll1, e.g., ν=0.05\nu=0.05), the parameter space supporting a stable spiral is exceedingly narrow (Figure 2(a)). In this regime, the resource rapidly approaches its stationary value, while strategies evolve too slowly to induce any significant feedback.

Characterization of the interior fixed point in the infinite-population limit

Figure 2. Characterization of the interior fixed point in the infinite-population limit of Eq. (2). Shown are the cases of two relative timescales: (a) ν=0.05\nu=0.05 and (b) ν=1\nu=1. Axes denote the rescaled extraction rates edd/be_{dd}/b and ecc/be_{cc}/b, with cross-interactions fixed to ecc=ecde_{cc}=e_{cd} and edd=edce_{dd}=e_{dc}. The dashed line edd=(b+ecc)/2e_{dd}=(b+e_{cc})/2 separates the physically admissible region (x∗≥0x^*\geq0) from the nonphysical regime. The solid line marks the transition to complex eigenvalues. Within the stable spiral regime, the heatmap encodes f0f_0 the frequency of damped oscillations.

In contrast, as ν\nu increases (e.g., ν=1\nu=1), strategies react more rapidly to variations in resource abundance. This has two effects: it significantly expands the parameter region where the equilibrium is a stable spiral, and also increases the associated oscillation frequency (Figure 2(b)). Such frequency, f0=12πΔ−τ2/4f_0=\frac{1}{2\pi}\sqrt{\Delta-\tau^2/4}, is determined by the imaginary part of the Jacobian eigenvalues, and it is reported by the color gradient in Figure 2.

Finite population.

Having ruled out periodic orbits in the infinite-population limit, we now relax this assumption. As illustrated in Figure 1 (a), finite-size stochasticity transforms the damped oscillations of the deterministic dynamics into sustained macroscopic fluctuations. This phenomenon, known as resonant amplification [19, 27], arises because intrinsic noise continuously excites the coupled resource–strategy modes and prevents the system from settling into its stable equilibrium.

Unlike previous eco-evolutionary models [11, 12], where strategy update rules yield periodic orbits even in the deterministic limit, here persistent quasi-cycles emerge only as a finite-size effect, sustained entirely by noise. To quantify the frequency of these noise-induced quasi-cycles, we employ the Linear Noise Approximation (LNA) [27, 28] to analytically derive the system’s power spectrum (see SM). We find that the spectrum of the resource exhibits a resonant peak at fpeak=ωρ∗/2πf_{\mathrm{peak}} = \omega_{\rho}^{*}/2\pi, with

ωρ∗2=−n+(n+Δ)2−nτ2.\omega_{\rho}^{*2} = -n + \sqrt{(n+\Delta)^2 - n\tau^2}.

Because this frequency is governed by the noise-to-coupling ratio n∝νK/Nn \propto\nu K/N, it depends on the relative timescale of strategic adaptation, ν\nu. Figure 1 (b) plots fpeakf_{\mathrm{peak}} as a function of ν\nu, demonstrating agreement between the theoretical LNA prediction and the frequencies extracted from numerical Gillespie simulations.

Applications to real data of global fish landings.

Surrogate-data tests and Kramers-Moyal diffusion estimates [29, 30] applied to fish-landing time series from different countries [31] reveal that their fluctuations are not compatible with purely deterministic nonlinear dynamics, which points to intrinsic stochasticity (see SM). To compare our model with the data, we interpret NN as the number of effective harvesting units (e.g. industrial fleets or vessels) and KK as the number of discrete biomass patches or quota units. Reported landings are treated as a Poisson sample of the resource, i.e. proportional to resource availability. Figure 3(a-Figure 3d) shows the landing time series and their corresponding spectra. For 162 time series the spectral maximum lies at f=0.016 yr−1f = 0.016\,\mathrm{yr}^{-1}, panel c of 3, which is the lowest resolved frequency (1/63 yr−11/63\,\mathrm{yr}^{-1}) of the observation window. For 11 time series maximum lies at a higher frequency of f=0.032 yr−1f = 0.032\,\mathrm{yr}^{-1}, panel d of 3. Figure 3(e-Figure 3f) compares the aggregated empirical spectra for High- and Low-income countries against our theoretical framework (analytical LNA) alongside numerical Langevin and Gillespie simulations. Both groups are reproduced by varying the noise ratio nn which depends on the extraction rates, the adaptation rate ν\nu and the ratio K/NK/N. High-income, industrialized fisheries (Figure 3(e)) combine intensive harvesting (ecc=ecd=0.9e_{cc} = e_{cd} = 0.9, edd=edc=1.8e_{dd} = e_{dc} = 1.8) with slower behavioral adaptation (ν=0.24 yr−1\nu= 0.24\,\mathrm{yr}^{-1}), and are run by

Temporal and spectral characterization of fish-landing dynamics

Figure 3. Temporal and spectral characterization of fish-landing dynamics. (a,b) Representative fish-landing time series (tons/year). (c,d) Normalized power spectra for all time series (light-gray lines) show dominant spectral peaks at f=0.016 yr−1f = 0.016\,\mathrm{yr}^{-1} (c) and f=0.032 yr−1f = 0.032\,\mathrm{yr}^{-1} (d). Solid curves denote the spectra for (c) high-income (teal) and (d) low-income (orange) countries. Income classifications follow World Bank definitions [31]. (e,f) Spectra of empirical landings for (e) high-income (teal) and (f) low-income (orange) countries are compared to : analytical LNA (dashed green), Langevin integration (dotted blue, mean ±1\pm1 s.e.m.) and Gillespie simulations (solid red, mean ±1\pm1 s.e.m.). The parameters adopted are (e) ecc=0.9e_{cc} = 0.9, edd=1.8e_{dd} = 1.8, ν=0.24 yr−1\nu= 0.24\,\mathrm{yr}^{-1}, K/N=100K/N = 100 and (f) ecc=0.5e_{cc} = 0.5, edd=1.5e_{dd} = 1.5, ν=0.38 yr−1\nu= 0.38\,\mathrm{yr}^{-1}, K/N=100K/N = 100. Simulated trajectories are downsampled to match the annual temporal resolution of the empirical data, prior to spectral estimation.

few large operators (small NN, K/N=10000K/N = 10000). The system is then overdamped and the analytical spectrum decays monotonically from the lowest resolved frequency f=0.016 yr−1f = 0.016\,\mathrm{yr}^{-1}, as observed. Low-income, artisanal fisheries (K/N=100K/N = 100), (Figure 3(f)) combine lighter extraction (ecc=ecd=0.5e_{cc} = e_{cd} = 0.5, edd=edc=1.5e_{dd} = e_{dc} = 1.5) with faster, local behavioral feedback (ν=0.38 yr−1\nu= 0.38\,\mathrm{yr}^{-1}). Here the LNA predicts an interior resonance at fpeak=0.032 yr−1f_{\mathrm{peak}} = 0.032\,\mathrm{yr}^{-1}, in agreement with the simulations and the empirical spectrum.

Conclusions

Real-world common-pool resources frequently exhibit complex oscillations amidst the tragedy of the commons [2]. In this Letter we introduced a stochastic ECO-EVO modeling framework to study this dual behavior. Notice that the strategy evolution of our model, in the quasi-stationary limit for slow resource dynamics, reduces to a replicator equation where resource feedback yields a Chicken (Hawk-Dove) game (see SM). By stabilizing a mixed strategy, this mechanism averts resource depletion despite a persistent local temptation to defect, providing a realistic bio-economic representation [13]. Furthermore, demographic noise resonantly amplifies the deterministic relaxation towards this state, sustaining macroscopic quasi-cycles. We found that the frequency of these fluctuations is controlled by the strategy-update timescale ν\nu and the noise ratio nn. By matching these parameters to the extraction and behavioral characteristics of high- and low-income countries, our model was able to successfully reproduce the frequency peaks found in global fish-landing data. Ultimately, our findings highlight the limits of purely deterministic models or of other approaches that impose independent noise at the macroscopic scale, showing that intrinsic finite-size stochasticity and the timescales of human strategy adaptation must be treated as essential and intertwined drivers of socio-ecological sustainability.

Acknowledgments

We thank Sandro Azaele and Giacomo Frigerio for useful discussions. S. S. acknowledges financial support from the MUR - PNC (DD n. 1511 30-09-2022) Project No. PNC0000002, DigitAl lifelong pRevEntion (DARE). V.L. ackowledges NERC funding (NE/Y001184/1).

Data Availability

The code to reproduce all simulations and figures is available on GitHub at https://github.com/SusannaCaruso/Resonant-Amplification-of-Demographic-Noise-in-Stochastic-CPR-Dynamics. The fish landing data used in this study are open-source and publicly available from the FAO [31].

From the Microscopic Processes to the Langevin equation

Master Equation

From the microscopic processes, we define the transition rates T(y∣z)T(\mathbf{y}\mid\mathbf{z}) representing the probability per unit time of transitioning from state z\mathbf{z} to state y\mathbf{y}. Given the total population N=C+DN=C+D and resource carrying capacity KK, the non-zero transition rates are:

T(C,D,R+1∣C,D,R)=bR(1−RK),T(C,D,R−1∣C,D,R)=hc(C,D,R)C+hd(C,D,R)D,T(C+1,D−1,R∣C,D,R)=ν(1−R′)++DCN,T(C−1,D+1,R∣C,D,R)=νR′CDN,R′≡RK(1−ecc/b)≥0,(6)\begin{aligned} T(C,D,R+1\mid C,D,R)&=bR\left(1-\frac{R}{K}\right),\\ T(C,D,R-1\mid C,D,R)&=h_c(C,D,R)C+h_d(C,D,R)D,\\ T(C+1,D-1,R\mid C,D,R)&=\nu(1-R')^{+}+\frac{DC}{N},\\ T(C-1,D+1,R\mid C,D,R)&=\nu R'\frac{CD}{N},\qquad R'\equiv\frac{R}{K(1-e_{cc}/b)}\geq0, \tag*{(6)} \end{aligned}

with (u)+=max⁡(0,u)(u)^{+}=\max(0,u). Here, Rc=K(1−ecc/b)R_c=K(1-e_{cc}/b) represents the effective carrying capacity of the resource under a fully cooperative population (C=N,D=0C=N,D=0). Following the knowledge-feedback mechanism [21], agents assess resource abundance relative to this cooperative baseline: the propensity to defect scales with R′=R/RcR'=R/R_c, whereas the scarcity-driven incentive to cooperate scales with 1−R′1-R'. Because the carrying capacity is K>RcK>R_c, transient states with R∈(Rc,K]R\in(R_c,K] correspond to R′∈(1,(1−ecc/b)−1]R'\in(1,(1-e_{cc}/b)^{-1}]. In such states of abundance above the cooperative baseline, there is no perceived scarcity, and the operator (1−R′)+(1-R')^{+} sets the d→cd\to c transition rate to zero. Since hc,d≥0h_{c,d}\geq0 by construction, all transition rates are strictly non-negative across the entire state space and vanish at the physical boundaries C,D∈{0,N}C,D\in\{0,N\} and R∈{0,K}R\in\{0,K\}.

The corresponding master equation governing the probability distribution P(C,D,R,t)P(C,D,R,t) reads:

ddtP(C,D,R,t)=[(E−R−1)T(C,D,R+1∣C,D,R)+(E+R−1)T(C,D,R−1∣C,D,R)+(E+CE−D−1)T(C+1,D−1,R∣C,D,R)+(E−CE+D−1)T(C−1,D+1,R∣C,D,R)]P(C,D,R,t),(7)\begin{aligned} \frac{d}{dt}P(C,D,R,t)=\big[& (\mathcal{E}_{-}^{R}-1)T(C,D,R+1\mid C,D,R)\\ &+(\mathcal{E}_{+}^{R}-1)T(C,D,R-1\mid C,D,R)\\ &+(\mathcal{E}_{+}^{C}\mathcal{E}_{-}^{D}-1)T(C+1,D-1,R\mid C,D,R)\\ &+(\mathcal{E}_{-}^{C}\mathcal{E}_{+}^{D}-1)T(C-1,D+1,R\mid C,D,R)\big]P(C,D,R,t), \tag*{(7)} \end{aligned}

where the step operators act as EX±f(X)=f(X±1)\mathcal{E}_{X}^{\pm}f(X)=f(X\pm1).

To obtain the continuum limit, we introduce the intensive variables:

x=CN,1−x=DN,ρ=RK,(8)x=\frac{C}{N},\qquad1-x=\frac{D}{N},\qquad\rho=\frac{R}{K}, \tag*{(8)}

with step increments Δx=1N\Delta x=\frac{1}{N} and Δρ=1K\Delta\rho=\frac{1}{K}. Under this scaling, the transition rates per unit volume/carrying capacity scale as:

T(C,D,R+1∣C,D,R)=Kwρ+(x,ρ)=Kbρ(1−ρ),T(C,D,R−1∣C,D,R)=Kwρ−(x,ρ)=Kρg(x),T(C+1,D−1,R∣C,D,R)=Nwx+(x,ρ)=Nν(1−ρ′)+x(1−x),T(C−1,D+1,R∣C,D,R)=Nwx−(x,ρ)=Nνρ′x(1−x),(9)\begin{aligned} T(C,D,R+1\mid C,D,R)&=K w_{\rho}^{+}(x,\rho)=Kb\rho(1-\rho),\\ T(C,D,R-1\mid C,D,R)&=K w_{\rho}^{-}(x,\rho)=K\rho g(x),\\ T(C+1,D-1,R\mid C,D,R)&=N w_{x}^{+}(x,\rho)=N\nu(1-\rho')^{+}x(1-x),\\ T(C-1,D+1,R\mid C,D,R)&=N w_{x}^{-}(x,\rho)=N\nu\rho'x(1-x), \tag*{(9)} \end{aligned}

where g(x)=eccx2+(ecd+edc)x(1−x)+edd(1−x)2g(x)=e_{cc}x^{2}+(e_{cd}+e_{dc})x(1-x)+e_{dd}(1-x)^{2} and ρ′=ρ1−ecc/b\rho'=\frac{\rho}{1-e_{cc}/b}.

The discrete step operators translate into continuous shift operators:

E±ρ→ϵ±Δρ,ϵ±Δρf(ρ)=f(ρ±Δρ)\mathcal{E}^{\rho}_{\pm} \to\epsilon^{\pm\Delta\rho}, \qquad\epsilon^{\pm\Delta\rho}f(\rho)=f(\rho\pm\Delta\rho)
E±xE±d→ϵ±Δx,ϵ±Δxf(x)=f(x±Δx).\mathcal{E}^{x}_{\pm}\mathcal{E}^{d}_{\pm} \to\epsilon^{\pm\Delta x}, \qquad\epsilon^{\pm\Delta x}f(x)=f(x\pm\Delta x).

Expanding the operators to second order in Δx=1/N≪1\Delta x=1/N\ll1 and Δρ=1/K≪1\Delta\rho=1/K\ll1:

ϵ±Δρf(ρ)−1=±Δρ∂∂ρf(ρ)+Δρ22∂2∂ρ2f(ρ)+O(Δρ3f(ρ)),\epsilon^{\pm\Delta\rho}f(\rho)-1=\pm\Delta\rho\frac{\partial}{\partial\rho}f(\rho)+\frac{\Delta\rho^{2}}{2}\frac{\partial^{2}}{\partial\rho^{2}}f(\rho)+O(\Delta\rho^{3}f(\rho)),
ϵ±Δxf(x)−1=±Δx∂∂xf(x)+Δx22∂2∂x2f(x)+O(Δx3f(x)).(10)\epsilon^{\pm\Delta x}f(x)-1=\pm\Delta x\frac{\partial}{\partial x}f(x)+\frac{\Delta x^{2}}{2}\frac{\partial^{2}}{\partial x^{2}}f(x)+O(\Delta x^{3}f(x)). \tag*{(10)}

Substituting these expansions into the master equation yields:

∂∂tP(x,ρ,t)=(−1K∂∂ρ+12K2∂2∂ρ2)[Kwρ+(x,ρ)P]+(+1K∂∂ρ+12K2∂2∂ρ2)[Kwρ−(x,ρ)P]+(−1N∂∂x+12N2∂2∂x2)[Nwx+(x,ρ)P]+(+1N∂∂x+12N2∂2∂x2)[Nwx−(x,ρ)P].(11)\begin{aligned} \frac{\partial}{\partial t}P(x,\rho,t)={}&\left(-\frac{1}{K}\frac{\partial}{\partial\rho}+\frac{1}{2K^{2}}\frac{\partial^{2}}{\partial\rho^{2}}\right)\left[K w_{\rho}^{+}(x,\rho)P\right]\\ &+\left(+\frac{1}{K}\frac{\partial}{\partial\rho}+\frac{1}{2K^{2}}\frac{\partial^{2}}{\partial\rho^{2}}\right)\left[K w_{\rho}^{-}(x,\rho)P\right]\\ &+\left(-\frac{1}{N}\frac{\partial}{\partial x}+\frac{1}{2N^{2}}\frac{\partial^{2}}{\partial x^{2}}\right)\left[N w_{x}^{+}(x,\rho)P\right]\\ &+\left(+\frac{1}{N}\frac{\partial}{\partial x}+\frac{1}{2N^{2}}\frac{\partial^{2}}{\partial x^{2}}\right)\left[N w_{x}^{-}(x,\rho)P\right]. \tag*{(11)} \end{aligned}

Simplifying the scale factors NN and KK and grouping by derivative orders leads directly to the two-dimensional Fokker–Planck equation:

∂P∂t=−∂∂x(AxP)−∂∂ρ(AρP)+12∂2∂x2(BxxP)+12∂2∂ρ2(BρρP),(12)\frac{\partial P}{\partial t}=-\frac{\partial}{\partial x}(A^{x}P)-\frac{\partial}{\partial\rho}(A^{\rho}P)+\frac{1}{2}\frac{\partial^{2}}{\partial x^{2}}(B^{xx}P)+\frac{1}{2}\frac{\partial^{2}}{\partial\rho^{2}}(B^{\rho\rho}P), \tag*{(12)}

where the drift coefficients AA and diffusion coefficients BB are:

Ax(ρ,x)=wx+−wx−=νx(1−x)[(1−ρ′)++ρ′],A^{x}(\rho,x)=w_{x}^{+}-w_{x}^{-}=\nu x(1-x)\left[(1-\rho')^{+}+\rho'\right],
Aρ(ρ,x)=wρ+−wρ−=bρ(1−ρ)−ρg(x),A^{\rho}(\rho,x)=w_{\rho}^{+}-w_{\rho}^{-}=b\rho(1-\rho)-\rho g(x),
Bxx(ρ,x)=1N(wx++wx−)=νNx(1−x)[(1−ρ′)++ρ′],B^{xx}(\rho,x)=\frac{1}{N}(w_{x}^{+}+w_{x}^{-})=\frac{\nu}{N}x(1-x)\left[(1-\rho')^{+}+\rho'\right],
Bρρ(ρ,x)=1K(wρ++wρ−)=1K[bρ(1−ρ)+ρg(x)].(13)B^{\rho\rho}(\rho,x)=\frac{1}{K}(w_{\rho}^{+}+w_{\rho}^{-})=\frac{1}{K}\left[b\rho(1-\rho)+\rho g(x)\right]. \tag*{(13)}

Evaluating the positive-part operator (1−ρ′)+(1-\rho')^{+} explicitly in the two regions ρ′≤1\rho'\le1 (i.e., ρ≤ρc≡1−ecc/b\rho\le\rho_{c}\equiv1-e_{cc}/b) and ρ′>1\rho'>1 (ρ>ρc\rho>\rho_{c}), the strategy drift AxA^{x} and unscaled diffusion factor Bx(ρ,x)≡NBxx(x,ρ)B_{x}(\rho,x)\equiv NB^{xx}(x,\rho) take the piecewise form

Ax(x,ρ)={νx(1−x)(1−2ρ′)for ρ′≤1,−νx(1−x)ρ′for ρ′>1,(14)A^{x}(x,\rho)= \begin{cases} \nu x(1-x)(1-2\rho') & \text{for }\rho'\le1,\\ -\nu x(1-x)\rho' & \text{for }\rho'>1, \end{cases} \tag*{(14)}
Bx(ρ,x)={νx(1−x)for ρ′≤1,νx(1−x)ρ′for ρ′>1.(15)B_{x}(\rho,x)= \begin{cases} \nu x(1-x) & \text{for }\rho'\le1,\\ \nu x(1-x)\rho' & \text{for }\rho'>1. \end{cases} \tag*{(15)}

Derivation of the Coupled Langevin System (Itô Interpretation)

To derive the continuous stochastic dynamics corresponding to the multivariate Fokker–Planck equation, we consider the state vector z=(x,ρ)T\mathbf{z}=(x,\rho)^{\mathsf{T}}. The general multivariate Fokker–Planck equation takes the form:

∂P(z,t)∂t=−∑i∈{x,ρ}∂∂zi[Ai(z)P(z,t)]+12∑i,j∈{x,ρ}∂2∂zi∂zj[Bij(z)P(z,t)].(16)\frac{\partial P(\mathbf{z},t)}{\partial t} =-\sum_{i\in\{x,\rho\}}\frac{\partial}{\partial z_i}\left[A^{i}(\mathbf{z})P(\mathbf{z},t)\right] +\frac{1}{2}\sum_{i,j\in\{x,\rho\}}\frac{\partial^{2}}{\partial z_i\partial z_j}\left[B^{ij}(\mathbf{z})P(\mathbf{z},t)\right]. \tag*{(16)}

where Ai(z)A^i(z) denotes the deterministic drift vector and Bij(z)B^{ij}(z) is the symmetric diffusion tensor. Because transitions in the microscopic coordinate xx and the resource density ρ\rho proceed via statistically independent jump events, the off-diagonal cross-diffusion coefficients vanish identically (Bxρ=Bρx=0B^{x\rho}=B^{\rho x}=0). The diffusion matrix is therefore diagonal:

B(x,ρ)=(Bxx(x,ρ)00Bρρ(x,ρ))=(νNx(1−x)[(1−ρ′)++ρ′]00ρK[b(1−ρ)+g(x)]).(17)\mathbf{B}(x,\rho)=\begin{pmatrix} B^{xx}(x,\rho) & 0 \\ 0 & B^{\rho\rho}(x,\rho) \end{pmatrix}=\begin{pmatrix} \frac{\nu}{N}x(1-x)\left[(1-\rho')^{+}+\rho'\right] & 0 \\ 0 & \frac{\rho}{K}\left[b(1-\rho)+g(x)\right] \end{pmatrix}. \tag*{(17)}

Under the Itô interpretation, the Fokker–Planck equation maps directly to a system of stochastic differential equations [23, 28]. The noise amplitude matrix σ(z)\boldsymbol{\sigma}(z) satisfies B(z)=σ(z)σT(z)\mathbf{B}(z)=\boldsymbol{\sigma}(z)\boldsymbol{\sigma}^{\mathsf{T}}(z). Choosing the diagonal representation σii(z)=Bii(z)\sigma_{ii}(z)=\sqrt{B^{ii}(z)}, the system reads in Itô Langevin form:

dxdt=Ax(x,ρ)+Bxx(x,ρ) ηx(t),dρdt=Aρ(x,ρ)+Bρρ(x,ρ) ηρ(t).(18)\begin{aligned} \frac{dx}{dt} &= A^x(x,\rho)+\sqrt{B^{xx}(x,\rho)}\,\eta_x(t),\\ \frac{d\rho}{dt} &= A^\rho(x,\rho)+\sqrt{B^{\rho\rho}(x,\rho)}\,\eta_\rho(t). \tag*{(18)} \end{aligned}

where ηx(t)\eta_x(t) and ηρ(t)\eta_\rho(t) are mutually independent standard Gaussian white noise processes with zero mean and covariance ⟨ηi(t)ηj(t′)⟩=δijδ(t−t′)\langle\eta_i(t)\eta_j(t')\rangle=\delta_{ij}\delta(t-t').

Substituting the explicit drift and diffusion expressions yields the general coupled Langevin equations valid across the entire state space Ω=[0,1]2\Omega=[0,1]^2:

dxdt=νx(1−x)[(1−ρ′)++ρ′]+νx(1−x)[(1−ρ′)++ρ′]N ηx(t),dρdt=ρ[b(1−ρ)−g(x)]+ρ[b(1−ρ)+g(x)]K ηρ(t).(19)\begin{aligned} \frac{dx}{dt} &= \nu x(1-x)\left[(1-\rho')^{+}+\rho'\right]+\sqrt{\frac{\nu x(1-x)\left[(1-\rho')^{+}+\rho'\right]}{N}}\,\eta_x(t),\\ \frac{d\rho}{dt} &= \rho\left[b(1-\rho)-g(x)\right]+\sqrt{\frac{\rho\left[b(1-\rho)+g(x)\right]}{K}}\,\eta_\rho(t). \tag*{(19)} \end{aligned}

In the region ρ≤ρc≡1−ecc/b\rho\le\rho_c\equiv1-e_{cc}/b (i.e., ρ′≤1\rho'\le1), Eq. (19) reduces to Eq. (2) of the main text, or equivalently in Itô differential notation:

dx=νx(1−x)(1−2ρ′) dt+νx(1−x)N dWtx,dρ=ρ[b(1−ρ)−g(x)] dt+ρ[b(1−ρ)+g(x)]K dWtρ.(20)\begin{aligned} dx &= \nu x(1-x)(1-2\rho')\,dt+\sqrt{\frac{\nu x(1-x)}{N}}\,dW_t^x,\\ d\rho&= \rho\left[b(1-\rho)-g(x)\right]\,dt+\sqrt{\frac{\rho\left[b(1-\rho)+g(x)\right]}{K}}\,dW_t^\rho. \tag*{(20)} \end{aligned}

where WtxW_t^x and WtρW_t^\rho denote independent standard Wiener processes.

Domain of Validity of (2) of the Main Text

In the main text, Eq. (2) is presented in its simplified form corresponding to the region ρ≤ρc=1−ecc/b\rho\le\rho_c=1-e_{cc}/b (or equivalently ρ′≤1\rho'\le1), where the positive-part operator simply reduces to (1−ρ′)+=1−ρ′(1-\rho')^{+}=1-\rho'. Here we show why the system resides in this region and why the cutoff at ρ=ρc\rho=\rho_c does not alter any of our analytical results.

First, consider how the resource ρ\rho evolves across the state space. Because cooperators extract the resource at the lowest rate (ecc≤ecd,edce_{cc}\le e_{cd},e_{dc} and ecc<b<edde_{cc}<b<e_{dd}), the total harvest pressure g(x)g(x) is minimized when the population consists solely of cooperators (x=1x=1), where g(1)=eccg(1)=e_{cc}. Whenever defectors are present (x<1x<1), the harvest pressure is higher:

g(x)=eccx2+(ecd+edc)x(1−x)+edd(1−x)2≥ecc[x2+2x(1−x)+(1−x)2]=ecc.(21)g(x)=e_{cc}x^2+(e_{cd}+e_{dc})x(1-x)+e_{dd}(1-x)^2\ge e_{cc}\left[x^2+2x(1-x)+(1-x)^2\right]=e_{cc}. \tag*{(21)}

Using the definition ρc=1−ecc/b\rho_c=1-e_{cc}/b (so that ecc=b(1−ρc)e_{cc}=b(1-\rho_c)), we can bound the deterministic rate of change of the resource as

ρ˙=ρ[b(1−ρ)−g(x)]≤ρ[b(1−ρ)−ecc]=bρ(ρc−ρ),(22)\dot{\rho}=\rho\left[b(1-\rho)-g(x)\right]\le\rho\left[b(1-\rho)-e_{cc}\right]=b\rho(\rho_c-\rho), \tag*{(22)}

with strict inequality whenever x<1x<1. This simple bound has two direct physical consequences: (i) Trajectories are confined to ρ≤ρc\rho\le\rho_c: Whenever the resource exceeds the cooperative carrying capacity (ρ>ρc\rho> \rho_c), the right-hand side of the inequality above is negative, meaning ρ˙<0\dot{\rho} < 0 regardless of the strategy composition xx. Even along the threshold line ρ=ρc\rho= \rho_c, the resource decreases (ρ˙<0\dot{\rho} < 0) for any x<1x < 1. Therefore, any deterministic trajectory that starts inside the domain Ωc=[0,ρc]×[0,1]\Omega_c = [0,\rho_c] \times[0,1] can never leave it. Conversely, if the system is initialized in a state of abundant resource with ρ(0)>ρc\rho(0) > \rho_c and x(0)<1x(0) < 1, the resource steadily declines and because abundance promotes defection (x˙<0\dot{x} < 0), harvesting intensifies during the transient, pulling the system into the region ρ≤ρc\rho\le\rho_c in finite time.

(ii) All equilibria and local stability properties are unaffected: Within Ωc\Omega_c, the general Langevin system in Eq. (19) is identical to Eq. (2) of the main text. All physically admissible fixed points lie inside Ωc\Omega_c. In particular, the coexistence equilibrium (ρ∗,x∗)(\rho^*,x^*) sits right in the middle of this domain at ρ∗=ρc/2\rho^* = \rho_c/2 (so that ρ′∗=1/2\rho'^* = 1/2). Consequently, local expansions around the interior equilibrium, including the Jacobian stability analysis and the Linear Noise Approximation (LNA), depend solely on the smooth dynamics within ρ<ρc\rho< \rho_c.

(iii) Stochastic fluctuations into ρ>ρc\rho> \rho_c are exponentially rare at large KK: In a finite system, demographic noise causes ρ\rho to fluctuate around the interior equilibrium ρ∗=ρc/2\rho^* = \rho_c/2 with a typical amplitude that shrinks as 1/K1/\sqrt{K}. Because the threshold ρ=ρc\rho= \rho_c lies at a fixed distance Δρ=ρc/2\Delta\rho= \rho_c/2 from the equilibrium, crossing it requires a fluctuation many standard deviations above the mean. The probability of such a large excursion decays exponentially with KK [23, 28], meaning that after the initial transient the system remains inside ρ≤ρc\rho\le\rho_c virtually all the time.

Linear Stability Analysis and Jacobian

Within the domain ρ≤ρc\rho\le\rho_c (ρ′≤1\rho' \le1), the Jacobian of the deterministic system is given by

J(ρ,x)=(∂ρ˙∂ρ∂ρ˙∂x∂x˙∂ρ∂x˙∂x),(23)J(\rho,x)=\begin{pmatrix} \frac{\partial\dot{\rho}}{\partial\rho} & \frac{\partial\dot{\rho}}{\partial x}\\ \frac{\partial\dot{x}}{\partial\rho} & \frac{\partial\dot{x}}{\partial x} \end{pmatrix}, \tag*{(23)}

with entries

∂ρ˙∂ρ=b(1−2ρ)−g(x),\frac{\partial\dot{\rho}}{\partial\rho}=b(1-2\rho)-g(x),
∂ρ˙∂x=−ρg′(x)=−ρ[2eccx+(ecd+edc)(1−2x)−2edd(1−x)],\frac{\partial\dot{\rho}}{\partial x}=-\rho g'(x)=-\rho\left[2e_{cc}x+(e_{cd}+e_{dc})(1-2x)-2e_{dd}(1-x)\right],
∂x˙∂ρ=−2νx(1−x)1−ecc/b,\frac{\partial\dot{x}}{\partial\rho}=-\frac{2\nu x(1-x)}{1-e_{cc}/b},
∂x˙∂x=ν(1−2x)(1−2ρ′),\frac{\partial\dot{x}}{\partial x}=\nu(1-2x)(1-2\rho'),

where g(x)=eccx2+(ecd+edc)x(1−x)+edd(1−x)2g(x)=e_{cc}x^2+(e_{cd}+e_{dc})x(1-x)+e_{dd}(1-x)^2.

Evaluated at the interior equilibrium (ρ∗,x∗)(\rho^*,x^*), where b(1−ρ∗)=g(x∗)b(1-\rho^*)=g(x^*) and ρ′∗=1/2\rho'^*=1/2 (so that Jxx=0J_{xx}=0), the Jacobian reduces to

J=(−bρ∗−ρ∗g′(x∗)−a0),a=2νx∗(1−x∗)1−ecc/b>0.(24)J=\begin{pmatrix} -b\rho^* & -\rho^*g'(x^*)\\ -a & 0 \end{pmatrix}, \qquad a=\frac{2\nu x^*(1-x^*)}{1-e_{cc}/b}>0. \tag*{(24)}

The trace and determinant of JJ are therefore

τ≡Tr⁡J=−bρ∗<0,Δ≡det⁡J=−aρ∗g′(x∗)>0,(25)\tau\equiv\operatorname{Tr}J=-b\rho^*<0,\qquad \Delta\equiv\det J=-a\rho^*g'(x^*)>0, \tag*{(25)}

where Δ>0\Delta>0 follows from g′(x∗)<0g'(x^*)<0 (since increasing the fraction of cooperators reduces the overall harvest pressure). Because τ<0\tau<0 and Δ>0\Delta>0, the interior fixed point is always locally asymptotically stable. Its eigenvalues are given by

λ±=τ2±iΩ,Ω=Δ−τ24.(26)\lambda_{\pm}=\frac{\tau}{2}\pm i\Omega,\qquad \Omega=\sqrt{\Delta-\frac{\tau^2}{4}}. \tag*{(26)}

Accordingly, the interior equilibrium is a stable node if τ2>4Δ\tau^2>4\Delta and a stable focus/spiral if τ2<4Δ\tau^2<4\Delta (complex conjugate eigenvalues with negative real part). Since τ≠0\tau\ne0 at the interior fixed point, the linearized system never has a center and never undergoes a Hopf bifurcation, oscillations can only be sustained by noise (quasi-cycles), see also § 1.5.

Exclusion of periodic orbits

Having classified the local stability of the interior fixed point, we now prove that the deterministic (drift) dynamics of system (2) in the main text cannot admit limit cycles or closed periodic orbits. As shown in Sec S1 C, since ρ˙<0\dot{\rho}<0 everywhere in ρ>ρc\rho>\rho_c, any hypothetical periodic orbit must lie within the region ρ≤ρc\rho\leq\rho_c. In the open, simply connected interior D=(0,1)×(0,ρc)⊂R2\mathcal{D}=(0,1)\times(0,\rho_c)\subset\mathbb{R}^2 (and more generally on (0,1)2(0,1)^2), the deterministic drift vector field F=(F(ρ,x),G(ρ,x))\mathbf{F}=(F(\rho,x),G(\rho,x)) is defined by the planar system:

x˙=F(ρ,x)=νx(1−x)(1−2ρ′),ρ˙=G(ρ,x)=ρ[b(1−ρ)−g(x)].(27)\begin{aligned} \dot{x} &= F(\rho,x)=\nu x(1-x)(1-2\rho'),\\ \dot{\rho} &= G(\rho,x)=\rho\left[b(1-\rho)-g(x)\right]. \tag*{(27)} \end{aligned}

To rule out the existence of periodic solutions, we employ the Bendixson–Dulac theorem [25, 26] using the Dulac multiplier:

ϕ(ρ,x)=1ρx(1−x).(28)\phi(\rho,x)=\frac{1}{\rho x(1-x)}. \tag*{(28)}

The function ϕ(x,ρ)\phi(x,\rho) is of class C∞C^\infty, strictly positive, and non-vanishing everywhere in D\mathcal{D}. According to the Bendixson-Dulac criterion, if the weighted divergence:

∇⋅(ϕF)=∂(ϕF)∂x+∂(ϕG)∂ρ(29)\nabla\cdot(\phi\mathbf{F})=\frac{\partial(\phi F)}{\partial x}+\frac{\partial(\phi G)}{\partial\rho} \tag*{(29)}

does not change sign and is not identically zero on any open subset of D\mathcal{D}, then the system contains no closed periodic orbits entirely contained in D\mathcal{D}. Evaluating each term explicitly:

∂∂x(ϕF)=∂∂x[ν(1−2ρ′)ρ]=0,(30)\frac{\partial}{\partial x}(\phi F)=\frac{\partial}{\partial x}\left[\frac{\nu(1-2\rho')}{\rho}\right]=0, \tag*{(30)}
∂∂ρ(ϕG)=∂∂ρ[b(1−ρ)−g(x)x(1−x)]=−bx(1−x).(31)\frac{\partial}{\partial\rho}(\phi G)=\frac{\partial}{\partial\rho}\left[\frac{b(1-\rho)-g(x)}{x(1-x)}\right]=-\frac{b}{x(1-x)}. \tag*{(31)}

Therefore, the divergence simplifies to:

∇⋅(ϕF)=−bx(1−x).(32)\nabla\cdot(\phi\mathbf{F})=-\frac{b}{x(1-x)}. \tag*{(32)}

Given that b>0b>0 and x(1−x)>0x(1-x)>0 for all x∈(0,1)x\in(0,1), it follows that ∇⋅(ϕF)<0\nabla\cdot(\phi\mathbf{F})<0 strictly across the entire domain D\mathcal{D}. Consequently, by the Bendixson–Dulac theorem, the deterministic system does not admit any periodic orbits or limit cycles. Remarkably, this holds even if in (S25) (1−2ρ′)(1-2\rho') is replaced by the full expression [(1−ρ′)+−ρ′][(1-\rho')_+-\rho'] on (0,1)2(0,1)^2.

Linear-Noise Approximation: Analytical Power Spectrum

Linearization around the Interior Fixed Point

To quantify the stochastic quasi-cycles analytically, we apply the Linear Noise Approximation (LNA) [23, 28] around the interior equilibrium (ρ∗,x∗)(\rho^*,x^*). This local expansion requires that the model parameters admit a feasible interior equilibrium x∗∈(0,1)x^* \in(0,1), which sits at ρ∗=(1−ϵc/b)/2=ρc/2\rho^*=(1-\epsilon_c/b)/2=\rho_c/2 (i.e., ρ′∗=1/2\rho^{\prime*}=1/2, hence ρ′<1\rho^\prime<1), and the system sizes K,N≫1K,N\gg1 are sufficiently large that fluctuations around (ρ∗,x∗)(\rho^*,x^*) remain small, of order O(K−1/2,N−1/2)O(K^{-1/2},N^{-1/2}).

Let z=(δρ,δx)⊤\mathbf{z}=(\delta\rho,\delta x)^{\top} denote the fluctuations around the interior fixed point (ρ∗,x∗)(\rho^*,x^*), where g(x∗)=b(1−ρ∗)g(x^*)=b(1-\rho^*). To leading order in K−1/2K^{-1/2} and N−1/2N^{-1/2}, the linearized Langevin dynamics reads

z˙=Jz+ξ(t),⟨ξ(t)ξ(t′)⊤⟩=B δ(t−t′),(33)\dot{\mathbf{z}}=J\mathbf{z}+\boldsymbol{\xi}(t),\qquad\langle\boldsymbol{\xi}(t)\boldsymbol{\xi}(t^{\prime})^{\top}\rangle=B\,\delta(t-t^{\prime}), \tag*{(33)}

where JJ is the Jacobian evaluated at (ρ∗,x∗)(\rho^*,x^*), and the noise covariance (diffusion) matrix B=B(ρ∗,x∗)B=\mathbf{B}(\rho^*,x^*) accounting for demographic noise in both variables is

B=(Bρ00Bx)=(2bρ∗(1−ρ∗)K00νx∗(1−x∗)N),(34)B=\begin{pmatrix}B_\rho&0\\0&B_x\end{pmatrix}=\begin{pmatrix}\dfrac{2b\rho^*(1-\rho^*)}{K}&0\\0&\dfrac{\nu x^*(1-x^*)}{N}\end{pmatrix}, \tag*{(34)}

where we used bρ∗(1−ρ∗)+ρ∗g(x∗)=2bρ∗(1−ρ∗)b\rho^*(1-\rho^*)+\rho^*g(x^*)=2b\rho^*(1-\rho^*).

Spectral Matrix and Resonant Peak Frequency

To compute the power spectral density matrix, we take the Fourier transform z~(ω)=∫−∞∞z(t)e−iωt dt\widetilde{\mathbf{z}}(\omega)=\int_{-\infty}^{\infty}\mathbf{z}(t)e^{-i\omega t}\,\mathrm{d}t. The spectral matrix is given by

S(ω)=M−1B(M−1)†,M=iωI−J,(35)S(\omega)=M^{-1}B(M^{-1})^\dagger,\qquad M=i\omega I-J, \tag*{(35)}

with

M−1=1det⁡M(iω−JxxJρxJxρiω−Jρρ),det⁡M=(Δ−ω2)−iωτ,(36)M^{-1}=\frac{1}{\det M}\begin{pmatrix}i\omega-J_{xx}&J_{\rho x}\\J_{x\rho}&i\omega-J_{\rho\rho}\end{pmatrix},\qquad\det M=(\Delta-\omega^2)-i\omega\tau, \tag*{(36)}

such that

∣det⁡M∣2=(Δ−ω2)2+ω2τ2≡D(ω).(37)|\det M|^2=(\Delta-\omega^2)^2+\omega^2\tau^2\equiv\mathcal{D}(\omega). \tag*{(37)}

Using Jxx=0J_{xx}=0 and Jρρ=τJ_{\rho\rho}=\tau, the elements of the spectral matrix evaluate to:

Sρρ(ω)=ω2Bρ+Jρx2BxD(ω),(38)S_{\rho\rho}(\omega)=\frac{\omega^2B_\rho+J_{\rho x}^2B_x}{\mathcal{D}(\omega)}, \tag*{(38)}
Sxx(ω)=(ω2+τ2)Bx+Jxρ2BρD(ω),(39)S_{xx}(\omega)=\frac{(\omega^2+\tau^2)B_x+J_{x\rho}^2B_\rho}{\mathcal{D}(\omega)}, \tag*{(39)}
Sρx(ω)=iωJxρBρ−(iω+τ)JρxBxD(ω)=Sxρ∗(ω).(40)S_{\rho x}(\omega)=\frac{i\omega J_{x\rho}B_\rho-(i\omega+\tau)J_{\rho x}B_x}{\mathcal{D}(\omega)}=S_{x\rho}^*(\omega). \tag*{(40)}

Because Sρρ(ω)S_{\rho\rho}(\omega) is an even function of ω\omega, we find its local extrema for ω>0\omega>0 by introducing the variable u=ω2≥0u=\omega^2\geq0. In terms of uu, Eq. (38) becomes

Sρρ(u)=uBρ+Jρx2Bx(Δ−u)2+uτ2.(41)S_{\rho\rho}(u)=\frac{uB_\rho+J_{\rho x}^2B_x}{(\Delta-u)^2+u\tau^2}. \tag*{(41)}

Setting c≡τ2−2Δc\equiv\tau^2-2\Delta and defining the noise-to-coupling ratio n≡Jρx2BxBρn\equiv\frac{J_{\rho x}^2B_x}{B_\rho}, this expression simplifies to

Sρρ(u)=Bρu+nu2+cu+Δ2.(42)S_{\rho\rho}(u)=B_\rho\frac{u+n}{u^2+cu+\Delta^2}. \tag*{(42)}

The stationary points satisfy dSρρdu=0\frac{dS_{\rho\rho}}{du}=0, which yields the quadratic equation

u2+2nu−(Δ2−nc)=0.(43)u^2 + 2nu - (\Delta^2 - nc) = 0. \tag*{(43)}

Using Δ2−nc=Δ2−n(τ2−2Δ)=(n+Δ)2−nτ2\Delta^2 - nc = \Delta^2 - n(\tau^2 - 2\Delta) = (n+\Delta)^2 - n\tau^2, the unique physical root is

ωρ∗2=u∗=−n+(n+Δ)2−nτ2.(44)\omega_{\rho}^{*2} = u^* = -n + \sqrt{(n+\Delta)^2 - n\tau^2}. \tag*{(44)}

Analogously, the power spectrum of the strategies, Sxx(ω)S_{xx}(\omega), peaks at

ωx∗2=−n~+(n~+Δ)2−n~τ2,(45)\omega_{x}^{*2} = -\tilde{n} + \sqrt{(\tilde{n}+\Delta)^2 - \tilde{n}\tau^2}, \tag*{(45)}

where n~=τ2+Jρx2BρBx\tilde{n}=\tau^2+\frac{J_{\rho x}^2B_{\rho}}{B_x}. A peak at ωρ>0\omega_{\rho}>0 exists iff

Δ2>n(τ2−2Δ);(46)\Delta^2 > n(\tau^2 - 2\Delta); \tag*{(46)}

otherwise SS decreases monotonically from ωρ=0\omega_{\rho}=0.

Comparison with the Deterministic Frequency $\Omega = |\operatorname{Im}\lambda|`

Substituting u=Ω2≡Δ−τ2/4u=\Omega^2\equiv\Delta-\tau^2/4 into Eq. (43) with c=τ2−2Δc=\tau^2-2\Delta shows that the LNA peak frequency fpeak=ωρ∗/(2π)f_{\mathrm{peak}}=\omega_{\rho}^*/(2\pi) coincides with the deterministic damped oscillation frequency f0=Ω/(2π)f_0=\Omega/(2\pi) at the critical noise ratio

nc=Δ−τ28=Ω2+τ28.(47)n_c=\Delta-\frac{\tau^2}{8}=\Omega^2+\frac{\tau^2}{8}. \tag*{(47)}

In the presence of demographic noise, the observed resonant frequency ωρ∗\omega_{\rho}^* is shifted relative to the deterministic frequency Ω\Omega depending on τ\tau and the relative intensity nn of the two noise sources:

  • Resource-dominated noise (n→0n\to0, n<ncn<n_c, i.e., Jρx2Bx≪BρJ_{\rho x}^2B_x\ll B_{\rho}): Taking n→0n\to0 in Eq. (44) gives

ωρ∗→Δ=Ω2+τ24>Ω.(48)\omega_{\rho}^*\to\sqrt{\Delta}=\sqrt{\Omega^2+\frac{\tau^2}{4}}>\Omega. \tag*{(48)}

In this limit, demographic noise shifts the spectral peak to a higher frequency than the deterministic frequency Ω\Omega.

  • Strategy-dominated noise (n→∞n\to\infty, n>ncn>n_c, i.e., Bρ≪Jρx2BxB_{\rho}\ll J_{\rho x}^2B_x): Expanding the square root in Eq. (44) for large nn, (n+Δ)2−nτ2=n+Δ−τ22+O(n−1)\sqrt{(n+\Delta)^2-n\tau^2}=n+\Delta-\frac{\tau^2}{2}+O(n^{-1}), yields

ωρ∗→Δ−τ22=Ω2−τ24<Ω.(49)\omega_{\rho}^*\to\sqrt{\Delta-\frac{\tau^2}{2}}=\sqrt{\Omega^2-\frac{\tau^2}{4}}<\Omega. \tag*{(49)}

Here, demographic noise shifts the resonant frequency below the deterministic value (and suppresses the peak entirely if Δ≤τ2/2\Delta\leq\tau^2/2).

Closed Form in Terms of the Microscopic Parameters

For the symmetric cross-interaction choice used in the main text, ecd+edc=ecc+edde_{cd}+e_{dc}=e_{cc}+e_{dd}, the harvest function simplifies to the linear form g(x)=eccx+edd(1−x)g(x)=e_{cc}x+e_{dd}(1-x), with constant derivative g′(x)=ecc−edd<0g'(x)=e_{cc}-e_{dd}<0. The interior fixed point is given explicitly by

ρ∗=b−ecc2b,x∗=χ2,χ≡2edd−ecc−bedd−ecc∈(0,2),(50)\rho^*=\frac{b-e_{cc}}{2b},\qquad x^*=\frac{\chi}{2},\qquad\chi\equiv\frac{2e_{dd}-e_{cc}-b}{e_{dd}-e_{cc}}\in(0,2), \tag*{(50)}

so that x∗(1−x∗)=χ(b−ecc)/[4(edd−ecc)]x^*(1-x^*)=\chi(b-e_{cc})/[4(e_{dd}-e_{cc})]. The Jacobian trace and determinant then read

τ=−b−ecc2,Δ=νx∗(1−x∗)(edd−ecc)=νχ(b−ecc)4,(51)\tau=-\frac{b-e_{cc}}{2},\qquad\Delta=\nu x^*(1-x^*)(e_{dd}-e_{cc})=\frac{\nu\chi(b-e_{cc})}{4}, \tag*{(51)}

while the deterministic frequency and resource noise intensity become

Ω=b−ecc44νχ−(b−ecc),Bρ=b2−ecc22bK.(52)\Omega= \frac{\sqrt{b-e_{cc}}}{4}\sqrt{4\nu\chi-(b-e_{cc})}, \qquad B_{\rho} = \frac{b^{2}-e_{cc}^{2}}{2bK}. \tag*{(52)}

Consequently, the noise-to-coupling ratios take the compact forms

n=KNγΔ,n~=τ2+NKΔγ,withγ≡(b−ecc)(edd−ecc)2b(b+ecc).(53)n = \frac{K}{N}\gamma\Delta, \qquad\tilde{n} = \tau^{2} + \frac{N}{K}\frac{\Delta}{\gamma}, \qquad\text{with}\qquad\gamma\equiv\frac{(b-e_{cc})(e_{dd}-e_{cc})}{2b(b+e_{cc})}. \tag*{(53)}

Thus, K/N≫1K/N \gg1 corresponds to strategy-dominated demographic noise, whereas K/N≪1K/N \ll1 corresponds to resource-dominated demographic noise.

Time series analysis

To assess the presence of stochastic dynamics in the data obtained from [31], we performed a preliminary analysis of the time series using two complementary techniques: surrogate-data testing and the empirical estimation of the Kramers-Moyal diffusion coefficient.

The dataset consists of 229 time series spanning 63 years. These include both country-level estimates and aggregated classifications provided by the FAO and World Bank, namely four income groups (high-income, upper-middle-income, lower-middle-income, and low-income countries) and seven World Bank geographical regions.

Prior to the analysis, incomplete and non-informative time series were removed. Specifically, we discarded all series containing more than 30% zero entries (37 time series) as well as one constant series.

To apply the surrogate-data testing procedure, we first removed the long-term trend associated with the global increase in fishing activity observed over recent decades. To this end, the original series were transformed into log-returns. Following this transformation, additional time series containing more than 20% missing values (NaN) were discarded. The remaining series were subsequently standardized, yielding a final dataset composed of 188 time series.

The surrogate-data analysis tests whether the observed fluctuations are compatible with a stochastic process or instead reflect non linear deterministic systems. In particular, we tested the null hypothesis that the data were generated by a monotonic nonlinear transformation of linearly filtered noise.

As test statistic, we employed the nonlinear prediction error (NLPE), which quantifies how accurately neighboring points in the reconstructed phase space predict the future evolution of a trajectory. The surrogate-testing procedure consists in comparing the value of the test statistic computed from the empirical data with that obtained from surrogate series preserving both the autocorrelation structure (equivalently, the power spectrum) and the probability distribution of the original data. For each empirical time series, we generated 100 surrogate realizations. The nonlinear prediction error was computed using an embedding dimension m=3m = 3 [29], delay t=1t = 1, and k=3k = 3 nearest neighbors.

Under these parameters, only 2 out of the 188 time series exhibited a test statistic significantly different from that of the corresponding surrogate ensemble. In contrast, the remaining 186 series showed no statistically significant deviation from the surrogate data (median pp-value: p=0.465p = 0.465). To corroborate this result, we estimated the Kramers–Moyal coefficient (diffusion coefficient D(2)D^{(2)}) directly from the standardized time series following [30]. Consistent with the surrogate test, we found a strictly positive diffusion coefficient across the dataset, with a median value D(2)=0.2956D^{(2)} = 0.2956 and a mean value ⟨D(2)⟩=0.3989\langle D^{(2)} \rangle= 0.3989. Although both methods are subject to finite-sample effects due to the 63-year observation window, together they provide empirical support for modeling fishery fluctuations as an intrinsically stochastic process. Beyond confirming their stochastic nature, spectral analysis of the 188 valid fish-landing time series reveals that the majority cluster into two main characteristic frequency groups. Specifically, 162 out of 188 series exhibit a dominant spectral peak at the lowest resolvable frequency of the 63-year observation window, f=1/63 yr−1≈0.016 yr−1f = 1/63\,\mathrm{yr}^{-1} \approx0.016\,\mathrm{yr}^{-1} (0.01587 yr−10.01587\,\mathrm{yr}^{-1}), corresponding to a monotonically decaying (overdamped) spectrum. A second coherent group, comprising 1 out of 188 series (including the low-income country aggregate), instead displays an interior resonant peak at f=2/63 yr−1≈0.032 yr−1f = 2/63\,\mathrm{yr}^{-1} \approx0.032\,\mathrm{yr}^{-1} (0.03175 yr−10.03175\,\mathrm{yr}^{-1}). The emergence of the two primary spectral groups points to distinct underlying eco-evolutionary feedbacks, which we link in the main text to differences in harvesting pressure, strategy-adaptation timescales and noise ratio.

Mapping to the standard replicator equation

Usually, eco-evolutionary models with environmental feedback explicitly incorporate strategy dynamics into a standard replicator equation by assuming a payoff matrix, or a linear combination, that varies according to the resource dynamics. Crucially, in our framework, the payoff matrix is not fixed beforehand. Instead, we demonstrate that by considering the mean-field system, i.e., assuming zero noise in the Langevin system (2) of the main text and analyzing the local behavior around the interior fixed point, we can directly map our model to a standard replicator equation.

Because the strategy dynamics are explicitly coupled to the resource, such a mapping holds in the quasi-stationary regime where the resource relaxes to its non-trivial equilibrium ρ∗(x)=1−g(x)/b\rho^*(x) = 1 - g(x)/b for a given strategy composition xx.

We begin by recalling the classical formulation of replicator dynamics for two interacting strategies. Consider a well-mixed population in which individuals engage in pairwise interactions described by the payoff matrix

Π=(RSTP).(54)\Pi= \begin{pmatrix} R & S \\ T & P \end{pmatrix}. \tag*{(54)}

Let x∈[0,1]x \in[0,1] denote the fraction of individuals adopting the first strategy. The expected payoffs of the two strategies are

π1(x)=Rx+S(1−x),π2(x)=Tx+P(1−x),(55)\pi_1(x) = Rx + S(1-x), \qquad\pi_2(x) = Tx + P(1-x), \tag*{(55)}

and the mean population payoff is

πˉ(x)=xπ1(x)+(1−x)π2(x).(56)\bar{\pi}(x) = x\pi_1(x) + (1-x)\pi_2(x). \tag*{(56)}

The replicator equation governing the evolution of xx is then given by

dxdt=x(π1(x)−πˉ(x))=x(1−x)(π1(x)−π2(x)).(57)\frac{dx}{dt} = x\left(\pi_1(x)-\bar{\pi}(x)\right) = x(1-x)\left(\pi_1(x)-\pi_2(x)\right). \tag*{(57)}

Substituting the explicit expressions for the payoffs yields

dxdt=x(1−x)(Ax+B),(58)\frac{dx}{dt} = x(1-x)(Ax+B), \tag*{(58)}

where

A=R−S−T+P,B=S−P.A = R-S-T+P,\qquad B = S-P.

We now compare Eq. (58) with the strategy dynamics of the deterministic system,

dxdt=νx(1−x)(1−2ρ′),(59)\frac{dx}{dt} = \nu x(1-x)(1-2\rho'), \tag*{(59)}

where ρ′\rho' denotes the rescaled resource level. In general, since ρ′\rho' evolves dynamically and depends on xx, the strategy dynamics cannot be written as a standard replicator equation with constant payoff entries. However, when the resource relaxes to a non-null stationary state ρ=ρ∗\rho=\rho^*, defined by

ρ∗=1−g(x)b,(60)\rho^* = 1-\frac{g(x)}{b}, \tag*{(60)}

where g(x)g(x) is a function solely of xx, the strategy equation can be evaluated at this stationary resource level. Substituting ρ∗\rho^* into the strategy dynamics yields

dxdt=νx(1−x)(1−21−g(x)b1−eccb).(61)\frac{dx}{dt} = \nu x(1-x)\left(1-2\frac{1-\frac{g(x)}{b}}{1-\frac{e_{cc}}{b}}\right). \tag*{(61)}

After straightforward algebraic manipulations, this equation can be rewritten as

dxdt=ν~x(1−x)(Qx2+Ax+B),(62)\frac{dx}{dt} = \tilde{\nu}x(1-x)(Qx^2+Ax+B), \tag*{(62)}

where

Q=2b(ecc−ecd−edc+edd),A=2b(ecd+edc−2edd),Q = \frac{2}{b}(e_{cc}-e_{cd}-e_{dc}+e_{dd}),\qquad A = \frac{2}{b}(e_{cd}+e_{dc}-2e_{dd}),
B=2eddb−eccb−1,ν~=ν1−eccb.B = \frac{2e_{dd}}{b}-\frac{e_{cc}}{b}-1,\qquad\tilde{\nu} = \frac{\nu}{1-\frac{e_{cc}}{b}}.

Equation (S57) has the same functional form as the standard replicator equation (S53) only in the special case Q=0Q=0. The condition Q=0Q=0 corresponds to a symmetric scenario in terms of per-capita extraction rates. In this case, the sum of the harvesting rates in a fully cooperative population (ecce_{cc}) and a fully defective population (edde_{dd}) is exactly equal to the sum of the harvesting rates in a mixed population (ecd+edce_{cd}+e_{dc}). Hence ecc+edd=ecd+edce_{cc}+e_{dd}=e_{cd}+e_{dc}, in other words, the total exploitation of the resource is balanced: the combined effect of the extreme populations (all cooperators or all defectors) produces the same total resource extraction as a population composed of a mix of strategies.

The parameter choice adopted in the main text satisfies this condition, enabling a consistent mapping to an effective payoff matrix. Under this assumption, the coefficients of the replicator equation can be identified as

R−S−T+P=2b(ecd+edc−2edd),S−P=2eddb−eccb−1.R-S-T+P = \frac{2}{b}(e_{cd}+e_{dc}-2e_{dd}),\qquad S-P = \frac{2e_{dd}}{b}-\frac{e_{cc}}{b}-1.

Since these relations provide two constraints for four payoff entries, two parameters remain free. Choosing SS and TT as independent parameters, the effective payoff matrix can be written as

Πeff=(eccb−1+TSTS+1+eccb−2eddb).(63)\Pi_{\mathrm{eff}} = \begin{pmatrix} \frac{e_{cc}}{b}-1+T & S\\ T & S+1+\frac{e_{cc}}{b}-\frac{2e_{dd}}{b} \end{pmatrix}. \tag*{(63)}

Classification of game regimes

We characterize the game-theoretic structure induced by the effective payoff matrix Πeff\Pi_{\mathrm{eff}} through the invasion quantities T−RT-R and P−SP-S, whose signs fully determine the phase diagram of the replicator dynamics. From the effective payoff matrix Πeff\Pi_{\mathrm{eff}}, these quantities are given by:

T−R=1−eccb,(64)T-R = 1-\frac{e_{cc}}{b}, \tag*{(64)}
P−S=1−2eddb+eccb.(65)P-S = 1-2\frac{e_{dd}}{b}+\frac{e_{cc}}{b}. \tag*{(65)}

The first difference, T−RT-R, quantifies the temptation to defect when all other players are cooperating. Because the model is defined on the domain ecc<be_{cc}<b, this difference is strictly positive (T−R>0T-R>0). This allows us to rule out game classifications characterized by a negative T−RT-R value, such as coordination and harmony games (Stag-Hunt and Harmony games).

As a result, the system is restricted to two game classes: the Chicken game and the Prisoner’s Dilemma. The parameter space edde_{dd} and ecce_{cc} are separated by the nullcline P−S=0P-S=0, which yields the boundary edd=ecc+b2e_{dd}=\frac{e_{cc}+b}{2}. Specifically the game is classified as a Chicken Game when edd>ecc+b2e_{dd}>\frac{e_{cc}+b}{2}, which is always satisfied for edd>be_{dd}>b, and the Prisoner Dilemma otherwise. Hence the Chicken game regime is always accessible and as pointed out by [13], when modeling shared common resources, the Chicken game often emerges as the more natural framework. In this context, an abundance of resources incentivizes defection; conversely, when resources are scarce and all other players defect, mutual defection results in the worst possible outcome.

References

  1. [1]G. Hardin, The Tragedy of the Commons, Science 162, 1243 (1968).
  2. [2]Z. Kobeissi, I. Mazari-Fouquer, and F. D. Ruiz-Balet, The tragedy of the commons: a Mean-Field Game approach to the reversal of travelling waves, Nonlinearity 37, 115010 (2024).arxiv.org/abs/2303.01365
  3. [3]D. Feeny, B. J. Berkes, B. J. McCay, and J. M. Acheson, The Tragedy of the Commons: Twenty-two years later, Human Ecology 18, 1 (1990).pubmed.ncbi.nlm.nih.gov/12316894
  4. [4]E. Ostrom, Tragedy of the Commons, in The New Palgrave Dictionary of Economics (Palgrave Macmillan, London, 2008) pp. 1–5.
  5. [5]E. Ostrom, The Challenge of Common-Pool Resources, Environment: Science and Policy for Sustainable Development 10.3200/ENVT.50.4.8-21 (2008).
  6. [6]C. Tu, P. D’Odorico, Z. Li, and S. Suweis, The emergence of cooperation from shared goals in the governance of common-pool resources, Nature Sustainability 6, 139 (2023).DOI
  7. [7]F. P. Chavez, J. Ryan, S. E. Lluch-Cota, and M. Ñiquen C., From Anchovies to Sardines and Back: Multidecadal Change in the Pacific Ocean, Science 299, 217 (2003).
  8. [8]G. Sugihara, R. May, H. Ye, C.-h. Hsieh, E. Deyle, M. Fogarty, and S. Munch, Detecting Causality in Complex Ecosystems, Science 338, 496 (2012).DOI
  9. [9]J. A. Hutchings, Collapse and recovery of marine fishes, Nature 406, 882 (2000).pubmed.ncbi.nlm.nih.gov/10972288
  10. [10]R. A. Myers, N. J. Barrowman, J. A. Hutchings, and A. A. Rosenberg, Population dynamics of exploited fish stocks at low population levels, Science 269, 1106 (1995).pubmed.ncbi.nlm.nih.gov/17755535
  11. [11]J. S. Weitz, C. Eksin, K. Paarporn, S. P. Brown, and W. C. Ratcliff, An oscillating tragedy of the commons in replicator dynamics with game-environment feedback, Proceedings of the National Academy of Sciences 113, E7518 (2016).pubmed.ncbi.nlm.nih.gov/27830651
  12. [12]A. R. Tilman, J. B. Plotkin, and E. Akçay, Evolutionary games with environmental feedbacks, Nature Communications 11, 915 (2020).DOI
  13. [13]H. Ito and M. Yamamichi, A complete classification of evolutionary games with environmental feedback, PNAS nexus 3, pgae455 (2024).DOI
  14. [14]C. Hauert, C. Saade, and A. McAvoy, Asymmetric evolutionary games with environmental feedback, Journal of Theoretical Biology 462, 347 (2019).DOI
  15. [15]N. Jonzén, J. Ripa, and P. Lundberg, A Theory of Stochastic Harvesting in Stochastic Environments, The American Naturalist 159, 427 (2002).DOI
  16. [16]C. N. K. Anderson, C.-h. Hsieh, S. A. Sandin, R. Hewitt, A. Hollowed, J. Beddington, R. M. May, and G. Sugihara, Why fishing magnifies fluctuations in fish abundance, Nature 452, 835 (2008).DOI
  17. [17]A. Yoshimoto, A new stochastic model for harvesting behavior with application to nonstationary forest growth and supply, Canadian Journal of Forest Research 26, 1967 (1996).DOI
  18. [18]A. Tabi, T. Siqueira, and J. D. Tonkin, Species interactions drive continuous assembly of freshwater communities in stochastic environments, Scientific Reports 14, 21747 (2024).DOI
  19. [19]A. J. Bladon, T. Galla, and A. J. McKane, Evolutionary dynamics, intrinsic noise, and cycles of cooperation, Physical Review E - Statistical, Nonlinear, and Soft Matter Physics 81, 066122 (2010).pubmed.ncbi.nlm.nih.gov/20866493
  20. [20]F. Bodin, G. Wang, and J. B. Plotkin, Eco-evolutionary games in noisy environments (2026), iSSN: 2692-8205 Pages: 2026.05.20.726658 Section: New Results.DOI
  21. [21]M. Saeediان, C. Tu, F. Menegazzo, P. D’Odorico, S. Azaele, and S. Suweis, Modelling co-evolution of resource feedback and social network dynamics in human-environmental systems, New Journal of Physics 26, 083004 (2024).
  22. [22]X. Zhao, K. Hu, Y. Tao, and L. Shi, Stochastic ecoevolutionary dynamics under coupled behavioral and environmental feedback, Applied Mathematics and Computation 520, 129959 (2026).DOI
  23. [23]N. G. V. K. van Kampen, Stochastic Processes in Physics and Chemistry (Elsevier, 2011).
  24. [24]D. T. Gillespie, A general method for numerically simulating the stochastic time evolution of coupled chemical reactions, Journal of Computational Physics 22, 403 (1976).DOI
  25. [25]I. Bendixson, Sur les courbes définies par des équations différentielles, Acta Mathematica 24, 1 (1901).DOI
  26. [26]A. Gasull and H. Giacomini, Effectiveness of the Bendixson–Dulac theorem, Journal of Differential Equations 305, 347 (2021).arxiv.org/abs/2101.03874
  27. [27]A. J. McKane and T. J. Newman, Predator-Prey Cycles from Resonant Amplification of Demographic Stochasticity, Physical Review Letters 94, 218102 (2005).DOI
  28. [28]C. W. C. W. . Gardiner, Stochastic methods : a handbook for the natural and social sciences (Berlin : Springer, 2009).
  29. [29]M. W. Small, Applied Nonlinear Time Series Analysis: Applications in Physics, Physiology and Finance, World Scientific Series on Nonlinear Science Series A Ser No. v.52 (World Scientific Publishing Co Pte Ltd, Singapore, 2005).
  30. [30]S. Siegert, R. Friedrich, and J. Peinke, Analysis of data sets of stochastic systems, Physics Letters A 243, 275 (1998).DOI
  31. [31]Food and Agriculture Organization of the United Nations (FAO) and World Bank, Capture fishery production, https://archive.ourworldindata.org/20260910-011030/grapher/capture-fishery-production.html (2026), processed by Our World in Data. Original data from World Development Indicators 129. Archived on September 10, 2026.

Paper details

vlnc:W_g6h7vew9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents