Decentralized Fair Division

Fair division is typically framed from a centralized perspective. However, in practice resource allocation often occurs via decentralized networks. We study a decentralized variant of fair division inspired by altruistic dynamics observed in behavioral economics and other practical settings. We develop an approach for decentralized fair division and compare it with a centralized approach with respect to fairness and social welfare guarantees. Our decentralized model can be seen as a relaxation of previous models of sequential exchange, in light of impossibility results concerning the inability of those models to achieve desirable outcomes. We find that the two models of resource allocation offer contrasting fairness and social welfare guarantees, and map out how these guarantees depend on valuations and other model parameters. We further show conditions under which a mix of the two approaches outperforms either approach in isolation. Despite the simplicity of our decentralized model, we show that under appropriate conditions it can ensure high-quality allocative decisions in an efficient fashion.

💡 Research Summary

The paper tackles the largely unexplored problem of fair division in decentralized settings, where agents act altruistically, possess only local knowledge, and face cognitive constraints. The authors propose a stochastic, polynomial‑time decentralized exchange process that gradually reallocates goods among agents while allowing agents to sacrifice part of their own utility for the sake of overall welfare. Each agent i has an additive valuation v_i over goods and an inalienable endowment e_i representing utility that cannot be traded. Endowments are visible to agents but hidden from a centralized planner, reflecting real‑world situations where local communities know socioeconomic status or disaster impact better than a central authority.

The central benchmark is Nash social welfare, defined as the sum of logarithms of agents’ total utilities (valuation plus endowment). The paper compares two mechanisms: (1) a stylized centralized planner that knows all valuations but not endowments and can thus maximize Nash welfare over the goods only; (2) the proposed decentralized process that uses local information, altruistic behavior, and a simple update rule. The authors prove several theorems. Theorems 1‑3 show that the ratio of total valuations to total endowments determines which mechanism yields higher Nash welfare: when this ratio exceeds a critical threshold, the centralized approach dominates; below it, the decentralized process is superior. Theorem 4 demonstrates that a hybrid scheme—running the decentralized dynamics for a while and then switching to centralized optimization—can strictly improve welfare over either pure approach, especially when valuations and endowments are highly imbalanced.

A novel fairness notion, Endowment‑Relative Envy‑Freeness (EREF), is introduced. EREF relaxes traditional weighted envy‑freeness by requiring that agents with lower endowments never envy the bundles of agents with higher endowments. Theorems 5‑7 establish conditions under which the decentralized mechanism satisfies EREF.

Extensive simulations across a range of parameter settings (different endowment distributions, valuation asymmetries, numbers of goods, and network topologies) confirm the theoretical predictions. The empirical results show that the decentralized process often reaches near‑optimal Nash welfare when endowments are relatively uniform, while the centralized planner excels when endowments are large and heterogeneous. The hybrid approach consistently yields the best performance in mixed scenarios. Moreover, the simulations illustrate that EREF is achieved in practice under the identified conditions.

Overall, the work contributes a new paradigm for fair division that incorporates realistic behavioral assumptions and information asymmetries. By quantifying when decentralization is beneficial, providing a tractable algorithm, and defining a practical fairness criterion, the paper offers valuable guidance for designing resource‑allocation systems in disaster relief, community aid, and emerging digital platforms where centralized control is limited or undesirable. Limitations include the assumption of additive valuations and static endowments; future research directions involve extending to non‑additive utilities, dynamic endowments, richer models of altruism, and implementation on real decentralized infrastructures with privacy and security considerations.