When Delegators Can Walk Away: Capture Cost in Liquid Democracy with Costly Participation

Online organizations that vote with tokens, such as the decentralized autonomous organizations (DAOs) that govern many blockchain protocols, have a turnout problem. Average token turnout sits near 5%. As a result, an attacker needs only about 2.6% of the total supply to control the votes that are actually cast (SoK: Attacks on DAOs). A common proposed fix is liquid democracy, in which each holder can vote directly or hand their weight to someone they trust, and that person can pass it on again. Empirical Measurements on Pricing Oracles and Decentralized Governance for Stablecoins recommends it for MakerDAO. The theory of liquid democracy, however, was built for a world where nobody abstains. The mechanisms designed to stop a few "super-voters" from collecting most of the weight were all analyzed assuming every delegator stays in the game. These include damping weight as it travels down a chain (Optimizing Viscous Democracy) and letting each voter name two possible delegates (The Fluid Mechanics of Liquid Democracy).

This proposal asks what happens to those mechanisms when each token holder can rationally walk away. We model every holder as choosing among three actions: vote directly and pay a transaction fee plus the effort of becoming informed, delegate and pay a smaller fee, or abstain for free. The choice depends on how likely the holder's weight is to change the outcome. We then measure centralization as the cost of capturing the organization: what it costs an outsider to buy a winning block of active tokens, and how few existing participants must collude to form a majority. Static inequality scores computed over all tokens do not capture this. Our central claim is that endogenous abstention can reverse the ranking of anti-concentration mechanisms. A rule that shrinks super-voters when everyone participates may, once delegators can leave, push small holders out entirely. That lowers active weight and makes capture cheaper.

Core Claim

Abstention and delegation interact through what the abstention literature calls the pivotal filter: a rational agent participates only when the expected effect of their weight on the outcome exceeds the cost. Delegation lowers the cost of staying in, so it should raise turnout. But a delegator only gains from delegating if the recipient is likely to be decisive, and the most decisive recipients are already the large holders. Delegation should therefore raise the cost of buying control while lowering the number of entities that must be bribed or must collude. Mechanisms that make delegated weight less valuable, such as viscosity decay or caps, also lower the delegator's expected benefit. Under costs they can convert delegation into abstention. We expect this to reduce capture cost on both axes at once. The prediction is falsifiable, and if it holds, it overturns recommendations derived under full participation.

For holder $i$ with stake $s_i$ choosing action $a$, which is vote, delegate to neighbor $j$, or abstain, the payoff is

$$U_i(a) = \pi_i(a)\,\Delta_i - c_i(a),$$

where $\pi_i(a)$ is the probability that the weight $i$'s action places on the ballot flips the outcome, $\Delta_i$ is how much $i$ cares about the outcome, and $c_i(a)$ is the fee plus information effort. Abstention costs nothing and has $\pi_i = 0$. Given the resulting active weights $w_k$ of terminal voters, we report two capture costs. The first is $K_{\text{buy}}$, the market price of the tokens needed to exceed half of the total active weight. The second is $K_{\text{coll}}$, the smallest number of active voters whose combined weight exceeds half. $K_{\text{coll}}$ is an abstention-aware Nakamoto coefficient (the Nakamoto coefficient being the minimum number of entities that together control a majority).

Relation to Prior Work

The closest work is The Cost Perspective of Liquid Democracy: Feasibility and Control. It assigns voting and delegation costs within a delegation graph and proves that a central organizer choosing who abstains, under a budget and with power caps, faces NP-complete problems. We keep its cost structure but remove the organizer: agents choose for themselves. We use the organizer's optimum only as an upper-bound benchmark on small instances.

On the abstention side, Equilibria of Plurality Voting with Abstentions shows that small voting costs can eliminate pure equilibria in simultaneous play, and that deciding whether one exists is NP-complete, while sequential play always has a subgame-perfect equilibrium. This result constrains our method. We do not search for pure equilibria. Instead we study noisy best-response dynamics, in which agents usually pick the better action but sometimes err, and a sequential variant. Group Wisdom at a Price motivates the information-cost term: costly information caps the active pool and breaks the guarantee that larger electorates get more accurate. It considers only voting versus abstaining, with no delegation.

On the delegation side, Liquid Democracy: An Algorithmic Perspective and A Contribution to the Critique of Liquid Democracy show that local delegation can do worse than direct voting because super-voters emerge. The viscosity and two-choice papers propose fixes. All four assume delegators participate, and we relax exactly that.

On the empirical side, Liquid Democracy in DPoS Blockchains shows that in EOS the top 4% of delegatees hold nearly half the delegated weight. The Hidden Shortcomings of (D)AOs documents delegate Nakamoto coefficients of 1 or 2 and a "governance tax" in gas fees. Voting Behaviour and Power in Online Democracy shows that super-voters exercised their theoretical pivotal power in only 1 of 3 cases. None of these papers models why participants abstain, so their numbers serve as calibration and validation targets rather than as competing models.

Research Program

We first build the model on directed trust graphs. These include Erdős–Rényi and Barabási–Albert synthetic graphs, following How Should We Vote?, and empirical delegation graphs reconstructed from EOS and Steem using the methodology of Li et al. Stakes are drawn from empirical DAO distributions. Agents estimate pivot probability by Monte Carlo sampling of other agents' current actions. Their prior over how proposals pass includes the empirical yes-bias of 0.71 that Kling et al. use in their Beta2 power index, which models voting power with a Beta distribution fitted to observed approval behavior. We run logit best-response dynamics until the distribution of actions stabilizes. In parallel we run sequential play on small instances, where a subgame-perfect equilibrium is guaranteed and can be computed by backward induction.

Next, we calibrate the cost parameters, gas price relative to stake and information effort, to observed turnout and delegation shares. We use Uniswap, ENS and Compound-style governors (single-hop delegation; Feichtinger et al.), MakerDAO (Gu et al.), and EOS and Steem (transitive delegation). LiquidFeedback, where fees are zero, isolates the information-cost term.

Finally, we use the calibrated model to compare mechanisms under endogenous abstention: plain transitive delegation, viscosity with $\alpha \in \{0.25, 0.5, 1\}$, two-choice delegation, weight caps, square-root stake weighting, and gas-subsidized delegation. For each mechanism we report $K_{\text{buy}}$, $K_{\text{coll}}$, turnout, and truth-tracking accuracy on binary issues with a hidden correct answer. On small graphs we also report the gap to the central organizer's optimum.

Baselines

The naive reference point is the static capture cost used in current audits. It treats observed turnout as fixed and computes the tokens needed to beat it, which gives the roughly 2.6% of supply reported in SoK: Attacks on DAOs. This baseline predicts that any mechanism leaving token holdings unchanged also leaves capture cost unchanged. Our model beats it if it predicts measurable shifts in capture cost under mechanism changes that this baseline misses, and if those shifts match observed cross-platform differences.

The strongest existing mechanism-level competitor is viscous delegation evaluated under full participation. Optimizing Viscous Democracy reports 10% to 20% accuracy gains over standard liquid democracy for $\alpha \in [0.25, 0.5]$ on ER and BA graphs. We reproduce that setting exactly, using the same graphs, competence distributions and accuracy metric, and confirm the reported range before switching on abstention. The comparison is then the same mechanism with and without endogenous abstention, at matched graph size and number of simulation runs. Two-choice delegation is treated the same way, using the confluent-flow setup of The Fluid Mechanics of Liquid Democracy. That paper reports asymptotic bounds rather than a dataset number, so its benchmark is the predicted collapse of maximum weight.

For concentration, the strongest non-delegation competitor is square-root stake weighting. How Does Stake Distribution Influence Consensus? reports an average Gini reduction of 37.16% and a 101.04% increase in the liveness Nakamoto coefficient across 10 proof-of-stake chains. We apply it to the same DAO stake distributions, with and without Sybil splitting priced at the calibrated gas cost.

For predicting who actually exercises power, the strongest baseline is the Beta2 index from Kling et al. It reports a squared error of 0.627 against the Banzhaf index's 1.320, and a perplexity of 76.6 against 297.9, on LiquidFeedback. Our model beats it only if it lowers squared error on the same proposals and data split.

The key ablation is the identical model with abstention removed: the abstain action is deleted, and agents can only vote or delegate. If mechanism rankings do not change under this ablation, our central claim is false.

Why not just…?

Why not just make voting free, as off-chain gasless voting does? Fees are only part of the cost. Future of Algorithmic Organization finds participation-weighted Gini coefficients above 0.9 in financialized DAOs even when Snapshot voting is included. Information effort remains, and Group Wisdom at a Price shows that this effort alone caps active participation. Our model tests this directly: setting gas to zero while keeping effort costs should leave $K_{\text{coll}}$ low. LiquidFeedback, which had no fees and still developed a power law with exponent 1.38, is the real-world check.

Why not just fit a regression of turnout and concentration on gas price and stake? This is the real threat. Within a single platform, a reduced-form regression may predict turnout as well as our model does. It cannot, however, say what happens under a delegation rule the platform has never used. We therefore make out-of-mechanism prediction the decisive test. Both the regression and the structural model are fitted on EOS, which is almost entirely single-hop. Both then predict Steem's chain depth (up to 5 hops) and its lower representativeness. We also fit both on single-hop DAOs and predict EOS's top-4% share. If the regression matches the structural model on these transfers, we will drop the equilibrium machinery and report the simpler model.

Why not just use square-root weighting and skip delegation? It flattens the weight distribution among people who vote, but it does nothing about who abstains. It is also vulnerable to splitting stake across fake identities: $k$ accounts each holding $s/k$ receive combined weight $\sqrt{k}\cdot\sqrt{s}$. The comparison above prices that splitting at calibrated gas costs. If square-root weighting still achieves the highest $K_{\text{coll}}$ once Sybil splitting is priced in, we will report that delegation mechanisms are dominated.

Proposed Experiments and Decisive Checks

The first check is the reversal test. On ER and BA graphs, with and without abstention, we compare viscosity at $\alpha \in [0.25, 0.5]$ against $\alpha = 1$. We look at accuracy, $K_{\text{buy}}$ and $K_{\text{coll}}$, sweeping cost from zero up to the level at which Group Wisdom at a Price predicts total abstention. Our hypothesis is supported if viscosity's advantage shrinks or reverses as cost grows, and if the mechanism is the one we propose: delegators at long chain distances switching to abstention.

The second check is calibration fidelity. The model should reproduce the reported targets: turnout below 5% with fewer than 20 deciding voters in MakerDAO, delegate Nakamoto coefficients of 1–2 in the governor-style DAOs, and the top 4% holding about half the weight in EOS.

The third check is out-of-mechanism transfer, as described above, with the regression as the competitor.

The fourth check is whether the model's predicted exercised power on LiquidFeedback reproduces the observed 1-in-3 rate at which super-voters flipped outcomes they could have flipped, and whether it beats the Beta2 squared error of 0.627.

The fifth check uses small graphs of a few dozen agents. There we compute the central organizer's optimum from The Cost Perspective of Liquid Democracy by exhaustive search, which measures how much capture cost decentralized choice gives up relative to it.

A null result on the reversal test would still be useful. It would mean that the full-participation design recommendations can be safely carried over to high-abstention organizations.