🌀 nonlinear sciences

Social welfare optimisation under institutional reward and punishment

This paper develops a welfare-centric framework for institutional incentives in social dilemmas, demonstrating that optimal reward or punishment schemes often differ significantly from those merely minimizing cost or maximizing cooperation frequency, and providing analytical conditions and algorithms to identify these welfare-maximizing strategies.

Original authors: Van An Nguyen, Vuong Khang Huynh, Huu Loi Bui, Hai Anh Ha, Quang Dung Le, Tan Dat Nguyen, Ngoc Ngu Nguyen, Zhao Song, Manh Hong Duong, Le Hong Trang, The Anh Han

Published 2026-06-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Van An Nguyen, Vuong Khang Huynh, Huu Loi Bui, Hai Anh Ha, Quang Dung Le, Tan Dat Nguyen, Ngoc Ngu Nguyen, Zhao Song, Manh Hong Duong, Le Hong Trang, The Anh Han

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a large group of people (or computer agents) trying to decide whether to help each other or look out only for themselves. In game theory, this is called a "social dilemma." Usually, the selfish choice wins because it's the easiest path. To fix this, an outside authority (like a government or a system administrator) steps in to offer rewards for helping or punishments for hurting.

For a long time, scientists have asked: "How much money should the authority spend to get the most cooperation?" They focused on two goals: getting people to cooperate as much as possible, and spending as little money as possible.

This paper asks a different, more holistic question: "What actually makes the whole group happiest and most successful, once we subtract the cost of the rewards or punishments?" The authors call this Social Welfare. Think of it as the "net profit" for the entire society.

Here is a breakdown of their findings using simple analogies:

1. The "Net Profit" Problem

Imagine a town where the mayor wants everyone to recycle.

The Old Way: The mayor asks, "How much should I pay people to recycle so that 90% of them do it, while spending the least amount of tax money?"
The New Way (This Paper): The mayor asks, "How much should I pay so that the total happiness of the town is maximized, after we subtract the cost of the payments?"

The paper finds that the "best" amount to spend for net happiness is often very different from the "best" amount to spend just to get people to cooperate. Sometimes, spending a little more to get slightly more cooperation actually lowers the town's total happiness because the cost of the reward outweighs the benefit.

2. The Two Tools: The Carrot vs. The Stick

The study compares two methods:

The Carrot (Reward): Giving a bonus to those who cooperate.
The Stick (Punishment): Taking money away from those who don't cooperate.

The authors discovered that rewards are generally better at maximizing the group's total happiness, provided the reward system is efficient.

The Analogy: Imagine you are trying to get a dog to sit.
- Reward: You give a treat. The dog gets happy, you spend a treat, but the dog learns well.
- Punishment: You scold the dog. The dog stops misbehaving, but it's stressed, and you wasted energy scolding.
- The Finding: Unless the "stick" is incredibly effective at a very low cost, the "carrot" almost always leaves the whole system (dog + owner) in a better state. The paper provides a mathematical formula to tell you exactly when the stick might be better than the carrot (spoiler: it's very rare).

3. The "Sweet Spot" and the "Phase Transition"

One of the most interesting findings is that there isn't just one simple rule for how much to pay. The relationship between how much you pay and how happy the group gets is like a rollercoaster, not a straight line.

The "Phase Transition": Imagine you are tuning a radio. At first, turning the dial (increasing the reward) makes the signal (happiness) get clearer. But if you turn it too far, the signal gets static and worse.
The authors found that depending on how "strong" the agents are (how much they care about the reward) and how "efficient" the reward is, the system can suddenly jump from a state where "more money = more happiness" to a state where "more money = less happiness."
They identified specific "tipping points." If you cross a certain threshold of selection intensity (how strongly agents react to incentives), the optimal strategy changes completely.

4. The "Zero or Target" Rule

The paper proves a surprising simplification: The best amount of money to spend is usually one of two things:

Zero: It's actually better to do nothing than to try to fix things with a clumsy reward system.
A Specific Target: If you do spend money, there is a very specific, calculable "sweet spot" amount. You don't need to guess; the math tells you exactly where the peak of the happiness hill is.

They even created a simple algorithm (a step-by-step recipe) that a computer can use to find this exact number instantly.

5. Why This Matters

The paper concludes that optimizing for "cooperation" and optimizing for "happiness" are not the same thing.

The Trap: If a policy maker only looks at "How many people are cooperating?" they might spend a fortune on rewards that technically work but actually drain the society's resources, leaving everyone worse off in the end.
The Solution: By focusing on Social Welfare (Total Benefit minus Total Cost), we can find policies that are truly beneficial. The authors show that the "best" incentive for cooperation is often much lower than the "best" incentive for maximizing the cooperation rate.

In summary: The paper argues that when designing rules for groups (whether human societies or AI systems), we shouldn't just ask "How do we get the most people to play nice?" We should ask, "How do we make the whole group's life better, accounting for the cost of the rules?" Often, the answer is to spend less than you think, or to use rewards instead of punishments.

Technical Summary: Social Welfare Optimisation under Institutional Reward and Punishment

Problem Statement
Existing literature on institutional incentives for promoting cooperation in multi-agent systems and evolutionary games typically frames the design problem as a bi-objective optimization: minimizing the institutional cost while achieving a target long-run frequency of cooperation. However, these frameworks largely overlook a holistic metric: social welfare, defined as the total population payoff net of institutional expenditure. The authors posit that schemes optimized solely for cost-efficiency or cooperation frequency may be suboptimal or even detrimental to overall societal well-being. This paper addresses the gap by analytically studying the optimization of social welfare under costly institutional incentives (both rewards and punishments) in finite, well-mixed populations playing social dilemmas.

Methodology
The study employs evolutionary game theory within finite populations of size $N$ .

Game Models: The analysis covers the Donation Game (DG) and the Public Goods Game (PGG).
Dynamics: Population evolution follows the Fermi strategy update rule with selection intensity $\beta$ . The dynamics are modeled as an absorbing Markov chain over states representing the number of cooperators.
Incentive Mechanisms:
- Reward: An institution pays a per-capita cost $\theta$ to cooperators, increasing their payoff by $a\theta$ , where $a$ is the reward efficiency.
- Punishment: An institution pays a per-capita cost $\theta$ to punish defectors, decreasing their payoff by $\hat{a}\theta$ , where $\hat{a}$ is the punishment efficiency.
Welfare Formulation: The authors derive explicit analytical expressions for the expected total social welfare ($SW$ for reward, $\widehat{SW}$ for punishment). These expressions aggregate the payoffs of all players across transient states, weighted by expected visitation rates derived from the fundamental matrix of the Markov chain, and subtract the total institutional costs.
Optimization: The core mathematical task is to maximize these welfare functions with respect to the incentive level $\theta$ . The authors analyze the qualitative behavior of these non-convex functions, deriving conditions for monotonicity, phase transitions, and the location of global optima.

Key Contributions and Results

Analytical Characterization of Social Welfare:
The paper derives closed-form expressions for expected social welfare under both reward and punishment. It identifies that the behavior of the welfare function depends critically on the incentive efficiency ( $a$ or $\hat{a}$ ) and the selection intensity ( $\beta$ ).
- Phase Transitions: The authors identify parameter regimes where social welfare exhibits a single optimal incentive level versus regimes where it becomes non-monotonic with multiple local optima. Specifically, for efficient rewards ( $a > 1$ ) and high selection intensity ( $\beta > \beta^*$ ), the welfare function can display complex phase transitions with multiple extrema.
- Zero-Sum vs. Non-Zero-Sum: In the zero-sum transfer case ( $a=1$ ), welfare is unimodal (increasing then decreasing). In non-zero-sum cases ( $a \neq 1$ ), the monotonicity undergoes qualitative shifts based on efficiency and selection strength.
Localization of Optimal Incentives:
A central theoretical result is that any welfare-maximizing incentive $\theta^*$ is either zero or concentrated around a specific, analytically determined target value ( $\theta_\infty = -\delta/a$ for rewards, $\hat{\theta}_\infty = -\delta/\hat{a}$ for punishments).
- Convergence Rates: The paper proves that as selection intensity $\beta \to \infty$ , the optimal incentive converges to these targets. For rewards, the convergence rate is $O(1/(a\beta))$ . For punishments, the convergence is slightly slower, $O(\log \beta / (\hat{a}\beta))$ , due to an additional logarithmic factor.
- Algorithmic Solution: Based on these localization properties, the authors provide an efficient algorithm (Algorithm 1) using a restricted interval grid search with upper and lower bounds to compute the optimal incentive numerically.
Reward vs. Punishment Comparison:
The study derives close-formed conditions under which institutional reward outperforms punishment in terms of social welfare for any given budget.
- Efficiency Threshold: The authors prove that if the reward efficiency $a$ is sufficiently high relative to the punishment efficiency $\hat{a}$ (specifically satisfying a condition involving harmonic numbers and population size), the institution can always adjust the investment cost such that reward yields higher social welfare than punishment.
- Dominance: In many standard parameter regimes, reward mechanisms are shown to dominate punishment in maximizing social welfare, unless punishment efficiency is exceptionally high relative to the benefit-to-cost ratio of the game.
Divergence from Cost-Optimization:
The paper demonstrates a systematic gap between incentives optimized for cost or cooperation frequency and those that maximize social welfare.
- Trade-offs: Maximizing social welfare often requires incentive levels that differ significantly from the minimum cost required to meet a cooperation threshold. In efficient transfer regimes ( $a > 1$ ), welfare maximization favors larger incentives, creating a conflict with cost minimization. Conversely, in inefficient regimes, the objectives may align, favoring the smallest feasible incentive.

Significance and Claims
The authors claim that their work provides a more realistic and policy-relevant framework for engineering cooperative behavior in biological, social, and artificial systems. By shifting the focus from mere cooperation frequency or cost minimization to social welfare, the study challenges the prevailing assumption that cost-efficient or high-cooperation schemes are inherently optimal for society.

The results suggest that:

Institutional designs should be judged not just by the level of cooperation induced, but by the net benefit to the population after accounting for the costs of intervention.
There are specific parameter regimes where maximizing cooperation or minimizing cost leads away from welfare-maximizing policies.
Reward mechanisms, when sufficiently efficient, generally offer a superior path to social welfare compared to punishment, provided the efficiency ratio meets specific theoretical thresholds.

The paper concludes that while it offers a rigorous welfare-centric analytical framework, future work is needed to explore adaptive/hybrid mechanisms, global characterizations of welfare landscapes, and heterogeneous incentive designs dependent on individual roles or local contributions.