Resilient Liquid Democracy: Mitigating Voting Power Imbalances via Secure Delegation Networks

Protecting the Flow of Votes in Liquid Democracy

Traditional online voting where you can delegate your vote to experts often leads to a few people holding all the power. These concentrated nodes are easy targets for manipulation. This new system uses special encryption to hide who you are delegating to until the vote is over. This prevents popularity contests and makes the system much harder to break if a major delegate fails.

The researchers propose a mechanism that addresses a fundamental tension in modern governance. They seek to route expertise to the right places without creating a fragile, centralized oligarchy.

The Fragility of Expert Routing

Liquid democracy is a hybrid voting model. It balances direct participation with specialized knowledge. In this system, a voter can choose to vote directly on an issue. Alternatively, they can delegate their voting power to a "delegate"—someone they trust to make an informed decision. This creates a network where influence flows from the many to the few.

However, most current implementations suffer from two structural flaws. First, they are "transparent." This means delegation choices are visible as they happen. Much like a social media trend, this visibility triggers herding. Voters flock to already-popular delegates rather than searching for true expertise. Second, this transparency makes coercion easy. If a person's delegation is public, they can be pressured to prove they followed orders. Finally, these systems are brittle. If a highly-backed delegate becomes inactive, a massive chunk of the electorate's power simply vanishes. This leads to significant "vote loss," where the intended democratic will is never recorded.

Building a Secure Delegation Network

To solve these issues, the authors introduce a protocol. It separates the act of choosing a delegate from the act of revealing that choice. The system relies on two primary technical pillars: Timed-Release Encryption (TRE) and ranked multi-delegation.

As shown in, the process is divided into a "Voting Phase" and a "Reveal Phase." During the voting phase, participants do not publish their choices in the clear.

Instead, they use TRE. This is a cryptographic method that locks information behind a timer. Think of it like a digital time capsule. The contents are mathematically hidden from everyone until a specific moment in time ($T_r$) is reached. To implement this, the authors use a decentralized committee and Shamir’s Secret Sharing (a method of splitting a secret into multiple parts). This "sealed formation" ensures that no one can observe popularity trends while the election is still ongoing.

To handle the risk of delegate failure, the authors replace the traditional single-delegate model with a ranked system. Instead of picking just one person, a voter provides an ordered list of preferred delegates. They also provide a "personal fallback ballot." This is a direct vote that is only counted if all chosen delegates fail to participate. This transforms delegation from a single point of failure into a resilient routing network.

The Recoverable-Gap Law

The researchers tested this mechanism using four diverse datasets. These ranged from Swiss municipal budgeting to a massive survey of 60,000 US voters. Their findings challenge the assumption that delegation is a universal good for accuracy.

The study finds that delegation does not always improve the quality of a decision. Instead, the authors report that whether delegation helps follows a "recoverable-gap law." They define this gap ($\gamma$) as the amount of accuracy lost when certain groups of people abstain from voting. Specifically, $\gamma = 1 - (\text{abstention accuracy})$.

The authors demonstrate that delegation only improves representational accuracy when abstention is both large and systematically unrepresentative. This means the people staying home must be fundamentally different from those participating. If the people who abstain are already a good statistical match for the rest of the population, the gap is small. In these cases, delegation can actually be harmful. Importing the views of a "competence elite" can pull the outcome away from the actual will of the electorate. This relationship is visualized in .

It shows that expertise-based delegation only begins to pay off once the gap exceeds a certain threshold.

Beyond accuracy, the paper highlights a massive structural benefit. In a "transparent" regime, the desire to follow popular leaders causes voting power to concentrate heavily. The authors use an abstract model to show this can cause the system to collapse into an oligarchy. In such a state, a single "super-delegate" controls nearly all voting share [, Figure 5].

By sealing the formation phase, the authors report that power remains distributed across many local delegates. This makes the system much harder to manipulate.

Perhaps the most striking result concerns system robustness. The authors modeled "targeted attacks" where an adversary tries to disable the most influential delegates. In a standard single-delegate system, this resulted in a 30.2% loss of votes [Table 4]. This means nearly one-third of the electorate would be disenfranchised. However, the authors report that their ranked delegation with fallback ballots slashed this loss to just 2.7%. This effectively preserves the democratic intent even under duress. This result was replicated across twenty different real-world elections .

Limits of the Seal

While the proposed mechanism offers significant protections, the authors define its boundaries. The "sealed" nature of the system only protects voters during the formation phase. Once the reveal time $T_r$ is reached and the encryption is lifted, all delegations become public and auditable. This means the system does not protect against coercion that occurs after the votes are revealed.

Furthermore, the authors note that their empirical models rely on "proxies" for expertise. These include survey responses or engagement levels rather than a perfect measurement of true knowledge. They also admit that their data-grounded model is "two-level." It looks at voters delegating to experts but does not fully capture the complexities of "transitive" delegation. This is where experts delegate to other experts. Such a dynamic could further concentrate power in ways their current models might understate.

Figures from the paper