Truth-Revealing Participatory Budgeting

Qishen Han, Artem Ivaniuk, Edith Elkind, Lirong XiaarXiv 2026 · arXiv

TL;DR

This paper shifts the study of Participatory Budgeting (PB) from a traditional subjective preference model to an epistemic one, where voters possess noisy, private signals about the objective quality and feasibility of projects. The authors reveal that while standard voting rules perfectly identify the best projects when costs are equal, varying project costs introduce fundamental mathematical barriers to information aggregation. Furthermore, they prove that honest voting is almost never a stable game-theoretic equilibrium.

What Problem Does This Paper Address?

Participatory Budgeting (PB) empowers citizens to directly decide how to allocate public funds. Historically, social choice theorists have studied PB through an axiomatic lens: assuming voters have fixed, known subjective preferences, how do we design fair and proportional aggregation rules?

However, real-world public projects frequently suffer from implementation failures—such as unexpected cost overruns, safety hazards, or simple failure to deliver intended outcomes—which voters cannot foresee with certainty. Instead of having fixed preferences, voters actually hold noisy information shaped by campaigns, expert reviews, and media.

This paper addresses the epistemic problem of PB: can common voting rules successfully aggregate these noisy individual signals to discover and elect the objectively "best" (highest quality) projects under a strict budget constraint?

Methods

The authors construct a novel epistemic framework for PB and analyze it using both asymptotic probability theory and game theory:

  • The Epistemic Framework: Each project $j$ has a hidden ground-truth quality $L_j$. Voters do not observe this quality directly but receive independent binary signals ($0$ or $1$, representing negative or positive impressions). The information structure is quality dominant, meaning higher-quality projects are inherently more likely to generate positive signals.
  • Utility Models: The authors evaluate two distinct objective welfare functions: Normal utility (the sum of elected project qualities) and Cost-proportional utility (where a project's societal impact scales linearly with its cost).
  • Voting Rules Analyzed: The study tests several standard PB rules, including Approval Voting (AV), AV per cost, Proportional Approval Voting (PAV), Greedy Cover, Phragmén's rule, and the Method of Equal Shares (MES).
  • Mathematical Analysis: The authors use the Law of Large Numbers (via Hoeffding's inequality) to analyze the asymptotic performance ratio (expected utility vs. optimal utility) as the number of voters $n \to \infty$. They also model voting as a Bayesian game to determine if informative (honest) voting constitutes a Bayes-Nash Equilibrium (BNE).

Datasets

The paper does not utilize real-world empirical datasets. Instead, the authors evaluate their framework using synthetic data generated for numerical simulations.

  • They generate 10,000 random instances per experimental setting.
  • Parameters varied include the number of agents ($n$ up to 100), a fixed set of alternatives ($m = 8$), varying cost ratios ($\alpha$), and different budget constraints ($B$).

Key Insights and Ideas

  • The Unit Cost Utopia: When all projects have an identical cost ($\alpha = 1$), the budget constraint simplifies to a simple cardinality constraint. In this scenario, the Condorcet Jury Theorem extends beautifully: as the voter pool grows, almost all common PB rules perfectly aggregate the noisy signals to select the optimal projects.
  • The Cost Dispersion Barrier: The moment project costs vary, a fundamental barrier emerges. The ratio between the highest and lowest project costs ($\alpha$) dictates the limits of truth-tracking. If costs are completely unconstrained, no voting rule can guarantee an asymptotic performance ratio better than $1/2$. Price dispersion fundamentally disrupts the voting rules' ability to balance cost against quality.
  • Honesty is Exceptionally Rare: In a strategic setting, voters shouldn't simply vote their signals. The authors prove that the mathematical conditions required for honest voting to be a Bayes-Nash Equilibrium form a Lebesgue zero-measure set. Voters are almost always incentivized to strategically manipulate their ballots based on how they expect others to vote.
  • The MES+AV Synergy: Rules designed purely for fairness, like the Method of Equal Shares (MES), often perform poorly in truth-tracking because they can leave budget unspent. However, completing the leftover budget of MES using standard Approval Voting (MES+AV) surprisingly yields the strongest empirical truth-revealing capability under general costs.

Key Results

  • Unit Cost Perfection: With unit costs ($\alpha = 1$), rules like AV, PAV, Phragmén, and MES+AV rapidly converge to a perfect performance ratio of 1.0 as the number of agents increases.
Figure 1: Performance with unit cost (α = 1)
  • General Cost Performance: When costs vary ($\alpha = 5$), performance remains strictly below 1.0 but comfortably exceeds the theoretical worst-case upper bound ($\approx 0.57$). Empirically, MES+AV achieves the highest performance.
Figure 2: Performance with general cost (α = 5)
  • Degradation via Cost Ratio: As the cost ratio $\alpha$ increases, the performance of the voting rules systematically degrades, confirming that price dispersion is the primary driver of information aggregation failure.
Figure 3: Performance on different α (AV rule, B = 7)
  • BNE Rarity: In numerical emulations of the strategic equilibrium condition under unit costs, honest voting was a BNE in only 0.7% of sampled information structures for 5 alternatives, dropping to 0.2% for 6 alternatives.

Limitations and Future Work

  • Homogeneous Utilities: The model assumes all voters share the exact same ground-truth utility function (i.e., they agree on what defines "quality"). In reality, voters have deeply conflicting, subjective values alongside their noisy information.
  • Binary Signal Simplification: The framework relies on simple binary signals ($0$ or $1$), which simplifies the nuanced, multi-dimensional information voters encounter in real life.
  • Strategic Behavior Constraints: The game-theoretic analysis of strategic voting is largely limited to the unit-cost setting; finding stable equilibria under general cost constraints remains an open challenge.
  • Future Outlook: The authors propose unifying the epistemic and axiomatic approaches to explore "informed participatory budgeting," where agents have heterogeneous utility functions conditioned on the underlying ground truth.

Summary updated

Abstract

Participatory Budgeting (PB) is commonly studied from an axiomatic perspective, where the aim is to design procedurally fair and economically efficient rules for voters with full information regarding their preferences. In contrast, we take an epistemic perspective and consider a framework where PB projects have different levels of underlying quality, indicating how well the project will take effect, which cannot be directly observed before implementation. Agents with noisy information cast votes to aggregate their information, and aim to elect a high-quality set of projects. We evaluate the performance of common PB rules by measuring the expected utility of their outcomes, compared to the optimal set of projects. We find that the quality of approximation improves as the range of project costs shrinks. When projects have unit cost, these common rules can identify the ``best'' set with probability converging to 1. We also study whether strategic agents have incentives to honestly convey their information in the vote. We find that it happens only under very restrictive conditions. We also run numerical experiments to examine the performance of different rules empirically and support our theoretical findings.
BibTeX
@misc{han2026truth,
  author = {Qishen Han and Artem Ivaniuk and Edith Elkind and Lirong Xia},
  title = {{Truth-Revealing Participatory Budgeting}},
  year = {2026},
  eprint = {2601.17538},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.17538},
}
Figure