CS 581, Project Proposal

Sybil-Resistant Liquid Democracy under Stake-Weighted Voting

Relevant Prior Work

Liquid democracy allows voters to either vote directly or transitively delegate their voting power. [1] study an epistemic model in which voters have heterogeneous probabilities of choosing the correct alternative. They ask whether delegation can sometimes improve accuracy while asymptotically doing no harm relative to direct voting. They show that local delegation rules cannot guarantee both properties, while a nonlocal mechanism that explicitly limits accumulated weight can do so under additional assumptions.

Subsequent work studies concentration more directly. Berinsky et al. quantify how much delegated weight may accumulate on a voter without destroying the epistemic benefit of liquid democracy, and show that several natural random delegation processes satisfy suitable concentration bounds [2]. Armstrong, Alouf-Heffetz, and Talmon instead introduce viscous democracy, in which voting weight decays with delegation distance. They show computational hardness of finding optimal delegations and give simulation evidence that stronger attenuation reduces the emergence of super-voters and can improve collective accuracy [3].

A separate blockchain-governance literature studies the tension between pseudonymity and unequal stake. Mohan, Khezr, and Berg formalize Sybil resistance for token-weighted voting and show that using token holdings alone cannot simultaneously achieve Sybil resistance and avoid plutocratic concentration. Their bond-voting mechanism introduces time commitment as a second scarce resource [4]. More recently, Bennett et al. show that concave wallet-level token weighting is vulnerable to false-name splitting: under broad conditions an optimal splitter can recover voting power that grows asymptotically linearly with total stake [5].

These questions have largely been studied in different models. A recent survey of computational liquid democracy emphasizes substantial variation in how existing work models delegation, voting, participation, power, and dynamics, and organizes these choices into a common taxonomy [6]. The goal of this project is to study how pseudonymous stake, nonlinear stake weighting, nonlinear delegation attenuation, concentration, and epistemic accuracy interact within one model.

Model

A voter with stake 𝑠 initially receives voting weight 𝐺(𝑠), where 𝐺 is increasing and 𝐺(0)=0. Every delegation step applies an increasing attenuation function 𝐹 satisfying

𝐹(0)=0,0≀𝐹(π‘₯)≀π‘₯.

A contribution traveling 𝑑 delegation edges therefore has effective weight

𝐻𝑑(𝑠)=𝐹○𝑑(𝐺(𝑠)).

The mechanism should ideally balance three objectives:

Preliminary Results

Source-side Sybils

Suppose an economic owner with stake

𝑆=𝑠1+…+π‘ π‘˜

splits it among several identities. At a common delegation depth 𝑑, source-side Sybil resistance is exactly

𝐻𝑑(𝑆)β‰₯βˆ‘π‘–=1π‘˜π»π‘‘(𝑠𝑖).

Thus superadditivity of 𝐻𝑑 is the natural false-name-resistance condition; convexity together with 𝐻𝑑(0)=0 is a sufficient condition.

Convex source weighting and concave delegation attenuation are not inherently incompatible. Fix a maximum delegation depth 𝐷 and 0<π‘ž<1, and define

𝐹(π‘₯)=(1+π‘₯)π‘žβˆ’1,𝐺(𝑠)=(1+𝑠)π‘žβˆ’π·βˆ’1.

Then

𝐻𝑑(𝑠)=(1+𝑠)π‘žπ‘‘βˆ’π·βˆ’1.

For every 𝑑≀𝐷, 𝐻𝑑 is convex and hence source-Sybil-resistant. Meanwhile its exponent decreases toward 1 as delegation depth grows, with

𝐻𝐷(𝑠)=𝑠.

The mechanism therefore moves from strongly convex weighting toward proportional weighting as a contribution travels farther:

𝐺=𝐻0→𝐻1→…→𝐻𝐷=identity.

Delegation progressively removes the superlinear premium received by large stakes without introducing profitable source-side splitting.

Bounded whale amplification

The preceding construction has an unbounded ratio 𝐺(𝑠)𝑠, but this is not necessary. For 0<𝛼<1, define

𝐹𝛼(π‘₯)=𝛼π‘₯+(1βˆ’π›Ό)(1βˆ’π‘’βˆ’π‘₯),

let 𝐽=πΉπ›Όβˆ’1, and set

𝐺=𝐽○𝐷.

Then

𝐻𝑑=𝐽○(π·βˆ’π‘‘)

is convex for every 𝑑≀𝐷 and satisfies

1≀𝐻𝑑(𝑠)π‘ β‰€π›Όβˆ’(π·βˆ’π‘‘).

Thus there exist source-Sybil-resistant mechanisms in which voting power is never more than a constant factor larger than proportional stake. The bound also decreases with delegation depth and reaches 1 at depth 𝐷.

This leaves a nontrivial optimization problem. For

𝐻𝑑=𝐹○𝑑○𝐺,

what is the smallest achievable sequence 𝐢0,…,𝐢𝐷 satisfying

𝐻𝑑(π‘₯+𝑦)β‰₯𝐻𝑑(π‘₯)+𝐻𝑑(𝑦),𝐻𝑑+1(𝑠)≀𝐻𝑑(𝑠),𝐻𝑑(𝑠)𝑠≀𝐢𝑑?

The constructions above prove feasibility, but not optimality.

Receiver-side Sybils

Concavity creates a different false-name attack when it is applied to aggregate weight at the receiving identity. If one delegate receives π‘₯+𝑦, its transformed weight is

𝐹(π‘₯+𝑦).

If the same controller can create two receiving identities and arrange for them to receive π‘₯ and 𝑦, its total becomes

𝐹(π‘₯)+𝐹(𝑦).

For strictly concave 𝐹 with 𝐹(0)=0,

𝐹(π‘₯)+𝐹(𝑦)>𝐹(π‘₯+𝑦).

Thus the same concavity that reduces marginal returns to accumulated delegated weight creates a receiver-side Sybil incentive if each pseudonymous receiver gets a fresh application of 𝐹. This manipulation occurs after the original stake transformation, so changing 𝐺 alone cannot remove it.

This separates two structurally different attacks:

source Sybil→splitting the original stake,receiver Sybil→splitting the identity at which concavity is applied.

A central open question is whether provenance-preserving delegation, aggregation before nonlinear transformation, or a weaker identity primitive can prevent receiver-side splitting without losing the useful effects of concave attenuation.

Epistemic Question

The final objective is to extend the epistemic analysis of liquid democracy [1], [2] to this stake-weighted setting.

Suppose underlying decision-maker 𝑖 has competence

𝑝𝑖=Pr[𝑖 votes correctly]

and final effective weight π‘Šπ‘–. If

π‘Œπ‘–βˆΌBernoulli(𝑝𝑖),

then the total correct voting weight is

𝑋=βˆ‘π‘–π‘Šπ‘–π‘Œπ‘–.

Delegation can improve the expected outcome by moving weight toward more competent voters, but it also concentrates many votes onto fewer independent decisions. Existing epistemic work shows that excessive concentration can destroy the accuracy benefit of delegation [1], [2].

Pseudonymity creates an additional issue: several wallets may correspond to one underlying decision-maker and therefore need not represent independent signals. Consequently, wallet-level decentralization is not necessarily information decentralization.

The main question is whether analogues of positive gain and do-no-harm can be obtained when voters have heterogeneous stake, voting power is transformed by 𝐺 and 𝐹, and economic actors may control multiple pseudonymous identities.

Research Goal

The preliminary results establish that:

The project will characterize how general these phenomena are. The main goals are to determine the optimal tradeoff between source-Sybil resistance, delegation attenuation, and bounded whale power; characterize the additional information needed to prevent receiver-side Sybils; and determine when these mechanisms retain the epistemic advantages of delegation.

Appendix: Technical Details

Finite-Depth Construction

For

𝐹(π‘₯)=(1+π‘₯)π‘žβˆ’1

and

𝐺(𝑠)=(1+𝑠)π‘žβˆ’π·βˆ’1,

iteration gives

𝐹○𝑑(π‘₯)=(1+π‘₯)π‘žπ‘‘βˆ’1,

and therefore

𝐻𝑑(𝑠)=(1+𝑠)π‘žπ‘‘βˆ’π·βˆ’1.

Since 𝑑≀𝐷 and 0<π‘ž<1,

π‘žπ‘‘βˆ’π·β‰₯1,

so 𝐻𝑑 is convex. At 𝑑=𝐷,

𝐻𝐷(𝑠)=𝑠.

Bounded-Amplification Construction

For

𝐹𝛼(π‘₯)=𝛼π‘₯+(1βˆ’π›Ό)(1βˆ’π‘’βˆ’π‘₯),

we have

𝐹𝛼′(π‘₯)=𝛼+(1βˆ’π›Ό)π‘’βˆ’π‘₯,

and hence

𝛼≀𝐹𝛼′(π‘₯)≀1.

Moreover,

𝐹𝛼″(π‘₯)=βˆ’(1βˆ’π›Ό)π‘’βˆ’π‘₯<0,

so 𝐹𝛼 is increasing and strictly concave. Its inverse 𝐽=πΉπ›Όβˆ’1 is therefore increasing and convex.

By the inverse-function rule,

𝐽′(𝑦)=1𝐹𝛼′(𝐽(𝑦)),

so

1≀𝐽′(𝑦)≀1𝛼.

Since 𝐽(0)=0,

𝑠≀𝐽(𝑠)≀𝑠𝛼.

Choosing 𝐺=𝐽○𝐷 gives

𝐻𝑑=𝐽○(π·βˆ’π‘‘),

and iterating the preceding inequality yields

𝑠≀𝐻𝑑(𝑠)β‰€π›Όβˆ’(π·βˆ’π‘‘)𝑠.

References

Bibliography

  • [1] A. Kahng, S. Mackenzie, and A. D. Procaccia, β€œLiquid Democracy: An Algorithmic Perspective,” Journal of Artificial Intelligence Research, vol. 70, pp. 1223–1252, 2021, doi: 10.1613/jair.1.12261.
  • [2] A. J. Berinsky et al., β€œTracking Truth with Liquid Democracy,” Management Science, vol. 71, no. 9, pp. 7821–7839, 2025, doi: 10.1287/mnsc.2023.02470.
  • [3] B. Armstrong, S. Alouf-Heffetz, and N. Talmon, β€œOptimizing Viscous Democracy,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI 2024), 2024, pp. 2643–2650. doi: 10.24963/ijcai.2024/292.
  • [4] V. Mohan, P. Khezr, and C. Berg, β€œVoting with Time Commitment for Decentralized Governance: Bond Voting as a Sybil-Resistant Mechanism,” Management Science, vol. 70, no. 12, pp. 8709–8734, 2024, doi: 10.1287/mnsc.2023.01536.
  • [5] A. Bennett, P. Vander Vos, D. V. Le, and M. Belenkiy, β€œConcave is the New Linear: The Impossibility of Anti-Plutocratic DAO Governance.” 2026. doi: 10.48550/arXiv.2605.18990.
  • [6] D. Grossi, A. Nitsche, and G. Papasotiropoulos, β€œModeling Liquid Democracy: A Survey of the (Computational) Social Choice Literature,” in Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence (IJCAI-ECAI 2026), 2026.