Budget aggregation is a process in which citizens vote by declaring their individual ideal budget allocation, and a pre-determined rule aggregates all votes into a single outcome. Recent theoretical work has proposed various aggregation rules, along with impossibility results for satisfying desirable axioms simultaneously. These analyses rely on assumptions about how voters evaluate non-ideal allocations, yet such assumptions have not been empirically validated on human subjects. We present a framework for empirically testing hypotheses about human utility functions using simple pairwise comparisons. We introduce a modular, open-source polling system that, after eliciting a subject's ideal allocation, presents carefully generated pairs of non-ideal alternatives. Different pair-generation algorithms allow testing various properties of utility functions. Using this framework, we conduct polls with hundreds of participants. The results show that standard utility models, including $\ell_1$, $\ell_2$, and Leontief, fail to capture human preferences, as very few participants behave consistently with any single model. In contrast, we find strong empirical support for more general properties, such as star-shaped, multi-dimensional single-peaked, and peak-linear preferences. We also find that most participants exhibit asymmetries both with respect to sign (gains vs. losses) and issue, contradicting any utility model based on an $\ell_p$ metric. These findings suggest that developing practical budget-aggregation mechanisms requires more flexible models of human utility functions.

翻译：预算聚合是一种公民通过声明个人理想预算分配进行投票，并由预定规则将所有投票汇总为单一结果的流程。近期理论研究提出了多种聚合规则，以及同时满足理想公理的不可能性结论。这些分析依赖于选民如何评估非理想分配的假设，但此类假设尚未在人类受试者中得到实证验证。我们提出一个框架，通过简单的成对比较对关于人类效用函数的假设进行实证检验。我们引入一个模块化、开源投票系统，在获取受试者理想分配后，系统会呈现精心生成的非理想替代方案对。不同的方案生成算法可检验效用函数的多种属性。利用该框架，我们对数百名参与者进行了投票实验。结果表明，标准效用模型（包括ℓ₁、ℓ₂和列昂惕夫型）无法捕捉人类偏好，因为极少有参与者的行为与任何单一模型一致。相反，我们发现了对更一般属性（如星形、多维单峰和峰值线性偏好）的强实证支持。同时发现多数参与者在符号（得失）和议题维度上均表现出不对称性，这否定了任何基于ℓₚ度量的效用模型。这些发现表明，开发实用的预算聚合机制需要更灵活的人类效用函数模型。