The Right Electorate Size for Attacking Delegated Votes

How much voting weight must a coordinated group add before it can probably swing a vote? Theory answers this for elections of independent, equally weighted voters: the group needs about $\sqrt{n}$ votes, because random swings in the tally among $n$ independent voters are about that size. Real governance does not look like that. In token-governed online organizations (DAOs, decentralized autonomous organizations), most holders abstain. Those who do vote often hand their vote to a proxy, who may pass it on again. Voting power ends up concentrated in a few heavy accounts. This proposal replaces the raw voter count with an effective electorate size computed from the weights that actually vote, after abstention and delegation. The claim has two parts. First, this quantity sets the attack threshold whenever no single account dominates the tally's randomness. Second, when one account does dominate, which heavy-tailed delegation data suggests is common, the success curve stops being smooth. Bribing one proxy then becomes cheaper than buying tokens.

The practical stakes are concrete. The standard security estimate for DAOs is deterministic. At 5% turnout, an attacker needs a majority of active votes, roughly 2.6% of token supply (SoK: Attacks on DAOs). That estimate fits openly malicious proposals that everyone else votes against. It says nothing about contested decisions, such as parameter changes or grant allocations, where the attacker only has to outweigh uncertainty about how everyone else will vote. It also says nothing about whether capital or bribery of a proxy is the cheaper route. A threshold law that covers both routes and can be computed from on-chain data would tell protocol designers which lever matters.

Core Claim

The proposal starts from the smooth phase transition of A Smooth Transition from Powerlessness to Absolute Power. That paper proves that for independent, identically distributed voters, a coalition of size $k = c\sqrt{n}$ succeeds with a probability $q(c)$ that rises smoothly in $c$. It relies on the central limit theorem, and the file explicitly lists the independence assumption as the open limitation. We generalize to weighted voting units. Let $w_i$ be the total weight cast by terminal voter $i$: their own tokens plus everything delegated to them, counted only among those who did not abstain. Let $W=\sum_i w_i$. If each unit's vote is an independent random choice, the tally's random swing has standard deviation proportional to $\sqrt{\sum_i w_i^2}$. That suggests two quantities:

$$ n_{\text{eff}} = \frac{W^2}{\sum_i w_i^2}, \qquad L = \frac{\max_i w_i^2}{\sum_i w_i^2}. $$

Here $n_{\text{eff}}$ is the effective number of independent voters, and $L$ is the share of tally variance held by the single heaviest account.

Our first prediction concerns the case where $L \to 0$, which is the Lindeberg condition under which the weighted central limit theorem holds. In that case, success probability against generalized scoring rules collapses onto the same smooth curve $q(c)$. The scaling variable is $c = k_w / \sqrt{\sum_i w_i^2}$, where $k_w$ is the coalition's added weight. The coalition's required share of active weight therefore scales as $1/\sqrt{n_{\text{eff}}}$, not $1/\sqrt{n}$.

Our second prediction concerns the case where $L$ stays bounded away from zero. Then the curve is no longer Gaussian, and it jumps at the weights of the largest accounts. Capturing a single proxy with share $w_{\max}/W$ moves the outcome by more than the tally's typical fluctuation once $w_{\max}/W \gtrsim 1/\sqrt{n_{\text{eff}}}$, which is the same as $L \gtrsim 1/n_{\text{eff}}$.

Delegation separates control from ownership: a proxy does not own what it casts. So in this regime, bribing a proxy can cost far less capital than buying an equivalent token stake. The empirical motivation is direct. Delegated power in the Pirate Party's LiquidFeedback follows a power law with reported exponent 1.38 (Voting Behaviour and Power in Online Democracy). Under either reading of that exponent, it sits below 2, the tail index at which the weighted central limit theorem fails. In EOS the top 4% of delegatees hold nearly half the delegated weight (Liquid Democracy in DPoS Blockchains).

Abstention does real work in this model rather than serving as background. Costly participation decides which weights are active. The jury-theorem analysis of Group Wisdom at a Price shows that effort costs cap the number of active voters. In On Rational Delegations in Liquid Democracy, rational agents who can delegate for free pile weight onto gurus. Together these imply a design trade-off governed by the cost of delegating. Cheap delegation pulls abstainers back in, which raises $W$, but it also concentrates weight, which lowers $n_{\text{eff}}$ and raises $L$. We expect the attack cost to be non-monotonic in delegation cost, and we aim to locate its peak.

Research Program

The theoretical part has three pieces. The first extends the hyperplane-rule analysis of A Smooth Transition from Powerlessness to Absolute Power to independent weighted units under the Lindeberg condition. This is a fairly direct generalization; its value lies in fixing the correct scaling variable. The second characterizes the limit when weights are heavy-tailed with tail index below 2. In that case the tally behaves like a sum dominated by its largest terms, and $q$ becomes a mixture over the votes of the top few accounts. We will bound the jump sizes in terms of $L$. The third adds a cost model for attacks. Token purchase costs are proportional to weight. Proxy capture costs are a separate bribe price that does not scale with delegated weight. We derive the regime boundary at which capture dominates.

The modeling part endogenizes the active weights. We extend the delegation game of On Rational Delegations in Liquid Democracy, which assumes delegation costs nothing, by adding a positive delegation cost alongside the voting effort cost. This is the decentralized setting the file notes neither that paper nor The Cost Perspective of Liquid Democracy covers. The latter studies a central planner choosing assignments under a budget, and it treats control as a planner's NP-hard combinatorial problem rather than as a probabilistic threshold. We run iterated best response to reach an equilibrium and measure $n_{\text{eff}}$, $L$, and the attack cost at that equilibrium.

The work also differs from the variance-preservation results of When is liquid democracy possible?. That paper shows certain graph topologies keep enough tally variance for local delegation to beat direct voting on accuracy. We study a different question, adversarial capture, and our quantity $L$ is closely related to the variance concentration that paper says must be avoided. Our results should therefore be read as the security-side counterpart of its accuracy-side condition.

Baselines

The naive reference point is the unweighted law from A Smooth Transition from Powerlessness to Absolute Power, applied with $n$ set to the number of active addresses. This is what one would do by taking the existing theory off the shelf. It ignores weights entirely, and it is the baseline that most directly isolates our key ingredient. We will run it on exactly the same active sets and proposals as our model. It wins if it predicts the 50%-success coalition weight as well as $n_{\text{eff}}$ does.

The strongest existing practical method is the deterministic active-majority threshold of SoK: Attacks on DAOs. That paper reports that at the average 5% turnout across 26 DAOs, an attacker needs roughly 2.6% of total supply. Alongside it we use the delegate Nakamoto coefficient from The Hidden Shortcomings of (D)AOs: the number of entities needed to hold a majority, reported as low as 1 or 2 for some DAOs. Both baselines are computed from the same snapshots we use. We count our method as beating them on contested proposals if it predicts the empirically required coalition weight with lower median absolute log-ratio error. For proxy capture specifically, the test is whether we correctly identify, proposal by proposal, whether single-proxy capture or token purchase is cheaper.

For the strongest synthetic comparison, no reported scalar result was found on Lacuna for attack-success curves on delegation graphs. The synthetic benchmark of On Rational Delegations in Liquid Democracy reports accuracy only graphically. Our reference point there is therefore direct Monte Carlo simulation, which serves as ground truth rather than as a competitor.

There is also a key-ingredient ablation that replaces endogenous abstention with fixed, exogenous turnout matched to the same mean. This tests whether the trade-off created by delegation cost depends on agents responding to costs, or whether a turnout number alone would suffice.

Why not just…?

Why not just use the Nakamoto coefficient or the active-majority calculation? These answer the worst-case question: how many entities hold a majority if everyone else opposes the attacker. For proposals the whole community would reject, that is the right answer, and we do not claim to improve on it there. For contested proposals the attacker only needs to beat uncertainty, and the deterministic numbers overstate the cost by a factor we can measure. Replays of proposals with near-even margins will show whether that factor is large. If it turns out small, the proposal's practical value is limited to the proxy-capture regime.

Why not just simulate attacks by Monte Carlo on the data and skip the theory? This is a real threat. Monte Carlo will produce correct curves for any given snapshot, and we use it as ground truth. The theory earns its place only if two quantities computed in closed form, $n_{\text{eff}}$ and $L$, reproduce those curves. That would let designers reason about mechanism changes without re-simulating. If the collapse onto $q(c)$ fails in experiment A below, or if $L$ fails to predict the capture regime in experiment C, the closed-form claim is wrong and only the empirical replay survives.

Why not just switch to square-root stake weighting? How Does Stake Distribution Influence Consensus? reports that weighting by the square root of stake cuts Gini by 37.16% on average across 10 chains. The same authors warn that splitting stake across fake accounts undoes it. Within our analysis, square-root weighting is simply one transformation of the $w_i$. We will compute $n_{\text{eff}}$ and $L$ with and without it, and with simulated splitting of stake across fake accounts, to see whether it actually moves an organization out of the capture regime.

Proposed Experiments

Experiment A: does the scaling collapse hold? This is synthetic. We draw $n$ from $10^2$ to $10^5$ voting units with weights whose tail index ranges from 1.1 to 3. Votes are independent and uniform over three candidates under plurality and Borda. We measure success probability against coalition weight by Monte Carlo. The prediction is that curves plotted against $c = k_w/\sqrt{\sum w_i^2}$ collapse for tail index above 2. Below 2, they should show jumps whose sizes are bounded by $L$. Plotting against $k/\sqrt{n}$ should fail to collapse in both regimes.

Experiment B: does delegation cost create a trade-off? This is the endogenous version. We use the 250-node Erdős–Rényi, Watts–Strogatz, and Barabási–Albert graphs from On Rational Delegations in Liquid Democracy, extended with a voting cost and a delegation cost. We sweep delegation cost and record $W$, $n_{\text{eff}}$, $L$, and the Monte Carlo attack cost under both attack routes. The decisive check is whether attack cost peaks at an intermediate delegation cost, and whether the peak shifts as the theory predicts across topologies. Barabási–Albert graphs should peak earliest.

Experiment C: does the theory predict real attack costs? This is empirical replay. We use the 21-DAO governance corpus of The Hidden Shortcomings of (D)AOs, the EOS and Steem delegation networks of Liquid Democracy in DPoS Blockchains, and the LiquidFeedback archive. For each proposal we compute $n_{\text{eff}}$ and $L$ over the weights that actually voted. We then resample sincere votes from each account's historical voting behaviour, using a block bootstrap across proposals to keep correlation between voters, and obtain empirical attack curves.

The decisive check has two parts. First, the theory's predicted 50%-success weight should beat both the unweighted $\sqrt{n}$ baseline and the deterministic thresholds in log-ratio error on contested proposals. Second, proposals with large $L$ should be exactly those where capturing the top proxy alone flips the outcome with probability above one half.

Experiment D: a calibration check against observed margins. Across LiquidFeedback proposals, the spread of realized margins around the reported 0.71 approval bias should scale with $\sqrt{\sum w_i^2}$ rather than with $\sqrt{n}$. The reported finding that super-voters reversed outcomes in only about a third of the cases where they could have (Voting Behaviour and Power in Online Democracy) serves as a reality check. Our model counts capture capability, not willingness to use it. If the empirical attack curves in pseudonymous DAOs track our predictions while LiquidFeedback's realized reversals fall well below them, that gap is a direct measurement of how much social accountability suppresses manipulation.