Introduction

The emergence and evolution of cooperation is central to understanding the evolution and human activity-associated dynamics. One of the most popular theoretical frameworks that is used to shed light on such issues is evolutionary game theory. Game theory has also been successfully applied in diverse fields such as evolutionary biology and psychology1, computer science and operations research2,3, political science and military strategy4,5, cultural anthropology6, ethics and moral philosophy7, economics8,9, traffic flow research10,11 and public health12. When preferences and goals of participating agents are in conflict, game theory can explain and predict interactive decisions13. The central aim of game theory research is to determine conditions needed for cooperation to emerge between egoistic individuals14,15,16. Two of the most famous models for game theory include the prisoner's dilemma (PD) and public goods game (PGG)17. While the PD for a pairwise interaction attracted the attention of biologists and social scientists, PGG for group interactions was the focus of studies in experimental economics18. The PGG was often studied to identify effects of collective action arising from joint group decisions. Although sometimes the group interactions can be modeled as repeated simple pair interactions as with the PD, the most fundamental unit of the game is irreducible multi-agent nature13,19,20. The PGG offers valuable insight into prevailing socioeconomic problems such as pollution, deforestation, mining, fishing, climate control and environmental protection13. In identifying potential solutions to these issues, PGGs with various strategies13,17,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47 have been suggested and studied. Economists have mainly studied PGG with two strategies, C and D, in which all agents participate and share a single common pool21,22,23,24.

In this report, we focus on a PGG with voluntary participation25 in which three strategic players (cooperators (C), defectors (D) and loners (L)) are considered. Each C contributes c to the common pool, whereas D attempts to exploit the resource at no cost. Then, each C gets the payoff PC as PC = rcnC/(nC + nD) − c, whereas each D obtains PD as PD = rcnC/(nC + nD). Here, nC (nD) denotes the number of C's (D's) participating in the game and r(>1) is the multiplication factor, which describes synergistic effects of cooperation. In contrast, L refuses to participate in the game and relies only on private payoff σ. In this report the condition, 0 < σ < c(r − 1), is imposed25.

Recently, the spatial evolutionary PGG (SEPGG) has been intensively studied to understand how steady-state strategies emerge on various structures and to identify characteristic features of such steady-state strategies17,25,26,28,29,30,31,32,33. In the SEPGG, each agent is assigned to a node on a lattice or network. In a unit game of the SEPGG, only a randomly selected agent and its linked neighbors participate26. Then, in each update of the SEPGG, a randomly selected agent i adopts the strategy of a randomly selected neighbor j of i with a transition probability fij that depends on payoffs Pi and Pj17. The SEPGG studies have revealed interesting results such as cyclic dominance25,27, transition nature26 and payoff distribution28. The effects of underlying topology on the SEPGG properties17,28,29,30,31,32,33 have also been found, such as the spatial reciprocity on diluted networks34 and multiplex networks35,36,37,38,39,40.

Since the SEPGG on regular lattices and sparse networks has considered only local interactions, the number of participants in a unit game centered at a node i cannot exceed ki + 1, where ki is the degree (or coordination number) of i. Thus, the SEPGG on sparse networks is hardly a theoretical model of real-life examples with very large participants such as taxes, provision levels, tolls, user fees, etc.48. Such cases involving public resources which anyone can overuse can be mapped into “tragedy of the commons” problem49,50. However, SEPGG in which all agents participate in a unit game has been rarely studied. Thus, we focus on SEPGG with very large participants.

In this report, the SEPGG with three strategies on a complete graph (CG) and dense complex networks is considered to understand the SEPGG with large participants. The CG is a simple undirected graph in which any node on the graph is linked to all other nodes. Thus, the number of links on the CG is N(N − 1)/2, where N is the number of nodes. In the SEPGG on the CG, all agents participate in a unit game. From analytically exact rate equations of the SEPGG on the CG, two stationary states depending on r and N are found. The state with only L agents (or L-state) is stable for low . The state with only D agents (or D-state) is stable for high . r* at which the crossover from the L-state to the D-state occurs is analytically obtained and also confirmed by numerical simulation. In the SEPGG on the CG, a C-dominant state cannot be stable even for very high r. These stationary states on the CG are very peculiar compared to the C-dominant state (or C-state) on regular lattices and sparse networks for very high r28,30,31,32,33. The L-state on the CG is also very peculiar in the sense that the L-state occurs only for σ > c(r − 1) in the PGG game with the well-mixed population26, whereas the L-state on the CG occurs even when 0 < σ < c(r − 1) or r is quite high.

More specifically, the time evolution of the SEPGG on the CG for high r is shown to have the following stages. In early time, the numbers of both C and L agents decrease, whereas the number of D agents hardly varies. Eventually, the D-state becomes stable. Hence, the time evolution of the SEPGG for high r describes key processes to the “tragedy of the commons” very well49,50, because the key processes are the following processes: First, the most of agents overuse the public resource in the commons as defector. Then, the overuse of the public resource will ruin it.

Ref. 26 revealed that the dominant state on sparse networks for high r is the C-state. Hence, we investigate crossover behaviors of the L-state or the D-state on dense networks such as the CG to a C-state on sparse networks by numerical simulation. For low r, first the crossover from the L-state to a D-state occurs and the D-state successively crosses over to a C-state as mean-degree 〈k〉 decreases. Furthermore, the D-state for moderate 〈k〉 remains even in the limit N → ∞. We also quantitatively find that cooperation gradually increases as the number of participants or 〈k〉 decreases, which is the origin of two crossovers. Hence, the crossovers for low r describe how the enhanced cooperation on sparse networks with low 〈k〉 overcomes “tragedy of the commons”, resulting in the C-state. For high r the direct crossover from the D-state to the C-state occurs. This direct crossover is nearly the same as that from the D-state to the C-state for low r.

Results

SEPGG on the complete graph

From fij in Eq. (11) using {Pi} on the CG, exact rate equations of densities on the CG are written as

and

where , etc.

To obtain stationary states from general initial configurations with , and , early time behaviors of ρC, ρD and ρL must be considered. Early time behaviors of ρC, ρD and ρL are determined based on competition between two terms of Eqs. (1)– (3), respectively. As ρCρD tanh(−βc/2) ≤ 0 in Eq. (1) and ρCρD tanh(βc/2) ≥ 0 in Eq. (2) for any non-negative ρC, ρD, β and c, two distinctive steady states are achievable depending on the value of ρC. When , in Eq. (1), in Eq. (2) and in Eq. (3). Thus, and , which make in Eq. (1) and after some time. From these relations we find that the state of {, , } appears when . Similarly, when , and , which also make in Eq. (3) and after some time. As a result, when , the state of {, , } appears. We call this state the D-state. In contrast, when , and , which make in Eq. (2) and after some time. Thus, the state of {, , } appears. We call this state the L-state. As the D-state or the L-state appears depending on the condition , we now examine the stability of the D-state based on rate equations (1)– (3). If the D-state is unstable, the L-state should be stable.

In the D-state with {, , }, the rate equation (1) becomes

because . By solving Eq. (4) for time t, we obtain

Similarly, the rate equation (3) also becomes

When , and

As ρC decreases with t, the condition for the D-state breaks down for t > t*. From the Eq. (5) and the condition with , . Therefore, on the CG with N → ∞, the L-state is the only stationary state. However, on the CG with finite N, the nonzero-minimum of ρL is 1/N and thus ρL = 0 if ρL(t) < 1/N. Therefore, if , then ρL(t > t*) = 0 and the D-state is still the stationary state. These results mean that the SEPGG on the CG with finite N has the following stationary state. For , the D-state becomes stable, where

or

More specifically, this D-state for high r or has never been found on regular lattices and sparse networks. As emphasized in our introductory remarks, this state also describes “tragedy of the commons” very well. In contrast, for , the L-state becomes stable. This L-state for has never been found on regular lattices and sparse networks either. The L-state is also anomalous and surprising, because no body remains as an active participant in the PGG for . No C-dominant stationary state is found on the CG even for high r. Compared to the C-dominant stationary states on a square lattice17,26 and on sparse networks28,30,31,32,33 for high r, the stationary states on the CG are unique and anomalous.

In Fig. 1, ρC(t), ρD(t) and ρL(t) from a single run of simulation on the CG with N = 105 are plotted. ρC(t) and ρL(t) decay exponentially in the early time regardless of the r value. For , the time dependences of ρC(t) and ρL(t) are sustained throughout and the stationary D-state eventually appears as shown in Fig. 1(a) (r = 2000). In contrast, when , ρL(t) increases after some time or for t > t* and the L-state eventually appears as shown in Fig. 1(b) (r = 60). Hence, simulation data presented in Fig. 1 exactly reproduce the analytical results of rate equations (1)– (3). More specifically, early time behaviors of ρL ~ exp(−t) and are confirmed by fittings to simulation data as shown in Figs. 1(a) and 1(b). Furthermore, the crossover time t* for r = 60 is t* = 8.86 in Fig. 1(b), which is nearly identical to t* obtained from .

Figure 1
figure 1

Simulation results of the SEPGG on the CG.

Plots of ρC(t), ρD(t) and ρL(t) of the SEPGG with c = 1, σ = 1 and β = 1 from a single simulation run with N = 105. The dotted horizontal line denotes the value of 1/N. (a) When r = 2000, the stationary D-state appears. By fitting the data to Eqs. (5) and (7), ρC ~ exp(~αCt) with (solid line) and ρL ~ exp(−αLt) with αL = 1.00(2)(~1.0) (dash-dotted line) are obtained. (b) When r = 60, the stationary L-state eventually appears. The vertical dashed line denotes the value of . By the fitting, ρC ~ exp(−αCt) with (solid line) and ρL ~ exp(−αLt) with αL = 0.97(4)(~1.0) (dash-dotted line) are obtained for t < t*.

Full size image

When and in the limit of N → ∞, the time dependences of ρC, ρD and ρL on the CG shown in Fig. 1(b) effectively present the process to the anomalous L-state with no active participants. The process means the following three steps. First, most agents defect one another. C then changes his strategy to D and ρC(t) decreases. Thus, D cannot receive enough payoff50, causing ρD(t) to decrease and ρL(t) to increase. Finally, most agents become L, as no one remains in the commons. Consequently, the stationary L-state eventually appears for .

To analyze the dependence of stationary states on the multiplication factor r, , and are obtained from simulations for various N and r by averaging over 1,000 realizations. Simulation results of and for various N and r are shown in the insets of Fig. 2. As shown in insets of Fig. 2, the crossover value of r, i.e., r*, from the stationary L-state to the stationary D-state increases with N as expected from Eq. (9). More specifically, and in Fig. 2 exactly depend on the single scaling parameter r0 defined as . The scaling behaviors confirm that the L-state crosses over to the D-state at as Eq. (9).

Figure 2
figure 2

Simulation results of the SEPGG on the CG for various r and N.

Plots of (a) and (b) against for N = 103, 104, 105 and 106. c = 1, σ = 1 and β = 1 are used. Inset of (a): Plots of against r. Inset of (b): Plots of against r.

Full size image

Crossover from the behavior on dense networks to that on sparse networks

A dense network is a network in which the mean-degree 〈k〉 satisfies 〈k〉 ∝ N51. For example, the CG is a typical dense network, as 〈k〉 = N − 1 in the CG. In a sparse network, 〈k〉 = finite51. In the SEPGG on the CG, either the L-state or the D-state is stable depending on r and N and the C-dominant state cannot be stable. In contrast, the C-dominant state is stable for relatively high r in the SEPGG on sparse networks such as random networks30,33 and two dimensional square lattices17,26. Therefore, it is interesting to study how crossover from the L-state and the D-state on dense networks to the C-dominant state on sparse networks occurs for given values of r and N.

We first investigate how the L-state on dense networks crosses over to the C-dominant state on sparse networks. Since the L-state is stable for low r0 on the CG as shown in Fig. 2, the crossover behaviors for low r0 are studied by simulations on random networks with 〈k〉. For a given N and 〈k〉, , and are obtained by averaging over 2,000 realizations. Typical crossover behaviors for r0 = 0.3 are shown in Fig. 3. As shown in Fig. 3(a), two crossovers occur successively as 〈k〉 decreases. The L-state is stable when 〈k〉 is quite high. The C-state of {, , } is stable when 〈k〉 is low enough. For moderate 〈k〉 the D-state is stable. Therefore, for low r0, the stationary state is first changed from the L-state to a D-state and crossover from the D-state to a C-state occurs as 〈k〉 decreases.

Figure 3
figure 3

Simulation results of the SEPGG on random networks for .

(a) Plots of , and against 〈k〉 for N = 16000. c = 1, σ = 1 and β = 1 are used. (b) Plots of against 〈k〉 for N = 4000, 8000, 16000 and 32000. Here, and are not shown, because for high 〈k〉 and for low 〈k〉. (c) Plots of 〈k〉1 and 〈k〉2 against N. The straight lines denotes fittings of with and with to corresponding data. (d) Plot of Δ 〈k〉 (≡〈k〉1 − 〈k〉2) against N. (e) Plot of against with ν1 in (c). (f) Plot of against with ν2 in (c).

Full size image

The stability of the D-state for moderate 〈k〉 in the limit N → ∞ is studied using the following methods. From simulation data of , and as in Figs. 3(a) and 3(b), we first obtain 〈k〉1 at which relations and hold simultaneously. We also obtain 〈k〉2 at which and hold. For example, dependences of 〈k〉1 and 〈k〉2 on N for r0 = 0.3 are shown in Fig. 3(c). The dependence of Δ 〈k〉 (≡〈k〉1 − 〈k〉2) is also shown in Fig. 3(d). As shown in Fig. 3(d), Δ 〈k〉 increases monotonically with N, guaranteeing the stability of the D-state for moderate 〈k〉 in the limit N → ∞. Furthermore, as shown in Fig. 3(c), 〈k〉1 and 〈k〉2 satisfy power laws and . By fitting these power laws to data presented in Fig. 3(c), crossover exponents are obtained as ν1 = 0.898(2), ν2 = 0.520(2). The result ν1 > ν2 also guarantees the stability of the D-state for moderate 〈k〉. The crossover property from the L-state to the D-state presented in Fig. 3(b) is adequately described by the single exponent ν1 obtained in Fig. 3(c). for higher 〈k〉 and various N are plotted against the scaling variable with the obtained ν1 as in Fig. 3(e), which shows that for higher 〈k〉 is a function of the single scaling variable . As shown in Fig. 3(f), crossover from the D-state to the C-state also satisfies the scaling property that for lower 〈k〉 is a function of the single scaling variable with the obtained exponent ν2. Using the same method ν1's and ν2's for various low r0(<1) are obtained as shown in Fig. 4. Because ν1 > ν2 in Fig. 4, the D-state for moderate 〈k〉 and low r0(<1) is stable in the limit N → ∞.

Figure 4
figure 4

Plots of exponents ν1 and ν2 against r0.

c = 1, σ = 1 and β = 1 are used. In the limit N → ∞, the D-state for moderate 〈k〉 is stable, because ν1 > ν2.

Full size image

Furthermore, the dependences of , and on 〈k〉 for low r0 in Fig. 3(a) are quite similar to the time dependences of ρC(t), ρD(t) and ρL(t) on the CG for low r0 shown in Fig. 1(b). In Fig. 1(b), initially there are enough Cs. As t increases, D governs the system. Finally L dominates, because D cannot receive enough payoff. Likewise, in Fig. 3(a), for low 〈k〉 there are also enough Cs. For moderate 〈k〉 D governs the system. When 〈k〉 becomes high enough, L dominates. Hence, it is very interesting to compare dynamical behaviors on the CG to static crossover behaviors depending on 〈k〉.

We thus now focus on the time dependence of ρC(t), ρD(t) and ρL(t) for various 〈k〉 to understand crossover behaviors for low r0 in Fig. 3(a). The time dependences of ρC, ρD and ρL for moderate 〈k〉 are shown in Fig. 5(a) and those for low 〈k〉 are shown in Fig. 5(b). For high 〈k〉, the time dependence is nearly identical to that on the CG shown in Fig. 1(b). For moderate 〈k〉 and high 〈k〉, ρC and ρL decrease, but ρD increases in early time. However, the stationary state is strongly affected by the subsequent time dependence of ρC. If 〈k〉 is quite high or if , ρC decays quickly and ρD cannot receive enough payoff. As a result, ρL increases for t > t* and the stationary L-state appears as explained in Fig. 1(b). In contrast, for moderate 〈k〉 or , ρC(t) decreases relatively slowly and ρL(t) never have a chance to increase reversely before the time at which ρL(t) ≤ 1/N [see Fig. 5(a)]. This means that the cooperation is effectively enhanced for moderate 〈k〉 and D receives enough payoff until L disappears due to the enhanced cooperation. This first crossover is quite similar to the crossover from the L-state in Fig. 1(b) to D-state in Fig. 1(a) on the CG. For low 〈k〉 or , ρC(t) never decreases as on sparse networks28,30,31,32,33 [see Figs. 5(b)] and . Hence, the crossover from the D-state to the C-state (or C-dominant state) occurs for 〈k〉 ~ 〈k〉2 as 〈k〉 decreases.

Figure 5
figure 5

Time dependence of ρC(t), ρD(t) and ρL(t) on random networks with N = 16000 for r0 = 0.3.

Plots of ρC(t), ρD(t) and ρL(t) (a) for moderate 〈k〉 ( = 30) and (b) for low 〈k〉 ( = 10). (a) For moderate 〈k〉 ( = 30), ρD increases with t, whereas ρC and ρL decreases. Finally, the stationary D-state emerges. (b) For low 〈k〉 ( = 10), ρC increases with t, whereas ρD and ρL decreases. Finally, the stationary C-state appears. The time dependences for high 〈k〉 are not shown, because they are nearly the same as those shown in Fig. 1(b).

Full size image

The two crossovers for low r0 thus derive from a gradual increase of cooperation as the number of participants (or 〈k〉) decreases. Therefore, the crossovers that describe the disappearance of both the anomalous state with no active participants and “tragedy of the commons” quantitatively show that agents in the larger group hardly cooperate relative to those in the smaller group45,46. However, this dependence on the group size is not necessarily accurate, because a recent study on PGG44 reported that increasing the group size does not necessarily lead to mean-field behaviors.

Finally, we study the crossover from the D-state to a C-state for high r0(>1). Typical crossover behaviors for high r0 are shown in Fig. 6(a). As shown in Fig. 6(a), for high r0( = 10), the D-state is stable when 〈k〉 is quite high. The C-state is stable when 〈k〉 is low enough. Therefore, for high r0, the direct crossover from the D-state to the C-state occurs as 〈k〉 decreases. To analyze the dependence of this direct crossover on N, for various N are obtained by simulation as shown in Fig. 6(b). The dependence of the direct crossover on N can be obtained by the ansatz , where at 〈k〉3 both and hold. From the dependence of 〈k〉3 on N, is obtained for r0 = 10. This direct crossover satisfies the scaling property that is a function of the single scaling variable with ν3 = 0.51. As shown in Fig. 6(d), ν3's for various high r0(>1) are obtained using the same method. The data in Fig. 6(d) show that the value of ν3 increases as r0 increases. As the D-state is always stable on the CG or dense networks with 〈k〉 ∝ N, the upper bound of ν3 should be equal to 1. We also confirm that the time dependences of ρC(t), ρD(t) and ρL(t) for high r0 are nearly the same as those in Fig. 1(a) for high 〈k〉 and as those in Fig. 5(b) for low 〈k〉, respectively. Hence, this direct crossover is nearly identical to the second crossover from the D-state to the C-state for low r0.

Figure 6
figure 6

Simulation results of the SEPGG on random networks for .

(a) Plots of , and against 〈k〉 for N = 16000. c = 1, σ = 1 and β = 1 are used. The stationary state is changed from the D-state to a C-state as 〈k〉 decreases. for any 〈k〉. (b) Plots of against 〈k〉 for N = 4000, 8000, 16000 and 32000. (c) Plot of against with . (d) Plot of ν3 against 1/r0.

Full size image

Discussion

In summary, we have studied the SEPGG on the CG and complex dense networks to understand behaviors of the SEPGG with very large participants. By analyses of the rate equations, we have shown that the L-state of {, , } is stable on the CG for r < r* with . In contrast, the D-state of {, , }, representing “tragedy of the commons”, is stable for r > r*. These analytic results on the CG have been confirmed by simulation.

We have also studied crossover behaviors from the L-state or the D-state on dense networks to the C-dominate state on sparse networks by numerical simulation on random networks with a mean degree 〈k〉. For r < r*, the L-state first crosses over to a D-state and successively this D-state crosses over to a C-state as 〈k〉 decreases. We have investigated the dependence of the crossovers on N for low r0 using the ansatz and , where the L-state is stable for , the D-state is stable for and the C-state is stable for . From the numerical simulations, ν1 and ν2 have been obtained. Since ν1 > ν2 for r < r*, we have found that the D-state for moderate 〈k〉 is stable even in the limit N → ∞. We have also studied the time dependences of ρC, ρD and ρL on random networks with 〈k〉 to understand the crossover behaviors for r < r*. For moderate 〈k〉, the D-state is stable, because ρC decreases relatively slowly. For low 〈k〉, cooperation is enhanced and the C-state is stable. The two crossovers for r < r* derive from a gradual increase of cooperation as the number of participants (or 〈k〉) decreases. The crossovers thus show how the enhanced cooperation on sparse networks with low 〈k〉 produces the C-state, overcoming both the anomalous state with no active participants and “tragedy of the commons” for low r0.

For high r0, the D-state is stable when 〈k〉 is high. The C-state is stable when 〈k〉 is low. Therefore, for high r0, the direct crossover from the D-state to the C-state occurs as 〈k〉 decreases. The dependence of the direct crossover on N has been also analyzed by the ansatz , where the D-state appears for and the C-state appears for . From the numerical simulations, ν3 has been obtained. The value of ν3 increases to 1 as r0 increases, because the D-state always appears on the CG or dense networks with 〈k〉 ∝ N. The crossovers thus describe how the enhanced cooperation on sparse networks with low 〈k〉 overcomes “tragedy of the commons” and makes the C-state for high r0.

Finally, the cyclic dominance in Ref. 25 can also be found for very low r and 〈k〉. For example, for r0 = 0.1, the crossover from the C-state to the cyclic dominance occurs at on the network with the size N = 104. This crossover behavior is not explained quantitatively here, because the crossover occurs only on sparse networks.

Methods

Let us define the SEPGG model on a given graph or network in detail. Each agent is assigned to a node on the network. Variable si of the agent on node i represents the strategy of i. The si is a cooperator (C), defector (D) or loner (L). The number of agents with a given strategy is denoted as , and , where N is the size of the network.

In each update of SEPGG on the network, an agent i is randomly selected. Then, the payoff Pi of i depends on the strategies of ki + 1 participants, where ki is the degree of i. If ni,C is the number of agents with C, ni,D is the number of agents with D and ni,L is the number of agents with L among the ki + 1 participants, ni,C + ni,D + ni,L = ki + 1. Pi is thus given by

Here, c is the cost contributed by a C to the common pool, r(>1) is the multiplication factor and σ is the fixed payoff of a L26. We impose the condition 0 < σ < c(r − 1) as in Ref. 25. Even if only one active participant remains, the payoff of the agent still follows Eq. (10). Then, the strategy of i is updated through the comparison of Pi with Pj of a randomly selected neighbor j among ki neighbors in order to select a better strategy. If si ≠ sj, the agent i stochastically adopts the strategy sj of the neighbor j with transition probability fij. We use

as in Ref. 17. Here β(≥0) controls the amount of noise. In each update of SEPGG, the payoffs in fij of Eq. (11) on regular lattices and sparse networks depend on the configuration of all the agents at the time of the update. In contrast, Pi in fij on the CG depends only on si and NC, ND and NL, of the strategies on the entire graph, because all agents participate in each unit game. The payoff {Pi} on the CG is thus written as

where the densities ρC(≡NC/N), ρD(≡ND/N) and ρL(≡NL/N) are used. To confirm the analytic results, simulations are performed for various N and r. Here, we mainly report the results of simulations with , c = 1, σ = 1 and β = 1. Simulations with various combinations of , , , c, σ and β are tested and nearly identical results are obtained.