Power in Liquid Democracy

Yuzhe Zhang, Davide GrossiAAAI 2021 · DOI

Abstract

Liquid democracy is a hybrid voting model that allows voters to transitively delegate their ballots to more competent peers. While promising for expertise-driven decision-making, it raises concerns about the concentration of power. This paper introduces the Delegative Banzhaf (DB) index, a novel axiomatic measure of power in delegative systems. Through game-theoretic modeling and simulations, we identify a "Connectivity Penalty" where denser social networks lead to higher power inequality. Conversely, we uncover a "Power-Greedy Paradox": when agents prioritize their own influence, they shorten delegation chains, leading to a more egalitarian distribution of power than in a purely altruistic society.

Introduction

Liquid democracy bridges the gap between direct and representative democracy. In this system, a voter can either cast their own ballot or delegate it to a trusted "guru." Crucially, this delegation is transitive: if voter 1 delegates to voter 2, and voter 2 delegates to voter 3, voter 3 wields the combined weight of all three.

Comparison of delegation structures: (a) a transitive chain ($d_1$) and (b) direct delegation to a single guru ($d_2$).

As illustrated in Figure 2, these structures can range from simple direct delegations to complex multi-step chains. While these chains allow for the aggregation of expertise, they also create "middlemen" whose influence has historically been difficult to quantify. This paper provides the first formal mathematical tool to measure the power of these intermediaries.

Methodology

The researchers developed the Delegative Banzhaf (DB) index by generalizing the classical Banzhaf index, which calculates the probability of a voter being "decisive" in a winning coalition. The DB index is the unique measure satisfying key fairness axioms, most notably the "Bloc Principle," which states that if two agents merge, their combined power should equal the sum of their individual powers.

To understand how power distributions evolve, the authors modeled agents as strategic actors maximizing a utility function: $$Utility = Power \times Accuracy$$ Here, agents seek to delegate to competent gurus (Accuracy) while maintaining their own influence (Power). The parameter $\alpha$ controls the weight given to individual power. Simulations were conducted on synthetic random networks, varying the probability of connection ($p$) and using Iterated Better Response Dynamics (IBRD) to find stable states.

Key Insights and Results

The Power of the Middleman

The DB index reveals that delegators are not merely passive conduits. An agent positioned closer to a guru in a long chain typically holds more power than those further away, as they control a larger "pipe" of transitive votes. If a middleman withdraws their support, the guru loses a significant block of influence, giving the middleman substantial leverage.

The Connectivity Penalty

Counter-intuitively, a more connected social network can lead to less democratic outcomes. As the probability of links between agents ($p$) increases, it becomes easier for the population to identify and funnel votes toward a few "super-gurus."

Gini coefficient of power distribution as a function of network connectivity ($p$). Higher connectivity leads to greater power inequality.

As shown in Figure 6, the Gini coefficient—a measure of inequality where 0 is perfect equality and 1 is total concentration—rises as connectivity increases. This suggests that "soft" constraints on social network density may be necessary to prevent extreme power imbalances.

The "Power-Greedy" Paradox

The most striking finding is the role of the power-seeking parameter $\alpha$. When agents are purely motivated by accuracy, they delegate freely, inadvertently creating massive power centers. However, when agents are "power-greedy" (high $\alpha$), they become reluctant to delegate to avoid losing their own DB index. This individualistic "hoarding" of votes prevents the formation of long delegation chains—which drop from an average length of 5.5 to 0.5—and results in a more decentralized, equal distribution of power across the system.

Discussion and Limitations

The study highlights that the structure of social networks and the strategic motivations of voters are critical to the health of liquid democracy. However, calculating the DB index is computationally "hard" (NP-complete), suggesting that real-world participants likely use heuristics rather than exact calculations. Future work should explore how "bounded rationality" affects these dynamics and investigate the sensitivity of these results to different voting quotas and dynamic agent attitudes.

Summary updated

BibTeX
@inproceedings{zhang2021power,
  author = {Yuzhe Zhang and Davide Grossi},
  title = {{Power in Liquid Democracy}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i6.16729},
}
Figure