How Should We Vote? A Comparison of Voting Systems within Social Networks

The strongest positive case is a qualified one: the paper’s liquid-democracy problem has a credible continuous mirror, but the paper fails the programme’s strict anchor gate.

The supplied paper contains no numbered Theorem, Lemma, Corollary, or Proposition asserting a computational result. Definitions 1 and 2 only define active voters and \(\rho\)-noisy sortition. Its only relevant statement is the unnumbered sentence in Section 3.1 that optimal delegations are NP-hard, explicitly citing Caragiannis and Micha (2019). The cited paper’s actual anchor is Theorem 2, which proves that approximating the optimal delegation value within an additive \(1/16\) is NP-hard; it is proved there, not here. [Caragiannis and Micha, Theorem 2](https://www.cs.toronto.edu/~emicha/papers/liquid_democracy.pdf) Thus, formally, this paper has no eligible in-paper computational anchor. What follows is the strongest provisional case built around that cited hardness result.

My lead mirror is Continuous Optimal Typed Delegation. An instance consists of a finite set of complete types \(T\), rational masses \(\mu_t\) with \(\sum_t\mu_t=1\), rational competences \(q_t\in[0,1]\), and a typed social-network relation \(H\subseteq T\times T\). A type includes not only competence but also the complete pattern of permitted delegation edges; two agents of the same type therefore have identical competence and identical delegation possibilities. The relation may be symmetric to match the paper’s undirected network model, or directed if one wants to retain the “follows” semantics used in the cited hardness construction. An integer \(K\) may bound the number of active representatives, and \(\eta\) is an accuracy threshold.

A solution chooses at most \(K\) active representative slots \(r_1,\ldots,r_\ell\), each occupied by an agent of some type \(u_j\). It chooses direct-voting masses \(d_t\ge0\), and delegation masses \(x_{t,j,p}\ge0\), where \(p\) is an allowed path in \(H\) from source type \(t\) to representative type \(u_j\). The conservation constraint is \(d_t+\sum_{j,p}x_{t,j,p}=\mu_t\). Thus every unit of population either votes directly or delegates transitively to one representative. The weight of representative \(j\) is \(w_j=\sum_{t,p}x_{t,j,p}\).

Let \(Y_j\) be independent with \(\Pr[Y_j=1]=q_{u_j}\). In the atomless limit, the direct-voting population contributes its deterministic law-of-large-numbers margin \(\sum_t d_tq_t\), while the finitely many representatives retain stochastic votes. The decision problem asks whether there is a feasible delegation plan satisfying \(\Pr[\sum_t d_tq_t+\sum_jw_jY_j\ge\tfrac12]\ge\eta\), with ties won by the correct alternative. The optimization version maximizes this probability.

This is recognizably the authors’ problem: it retains the binary ground truth, competence values, social-network restrictions, transitive delegation, weighted majority, active representatives, and probability of finding the truth. The fractional-looking \(x\)-variables do not mean that an individual delegates fractionally. For rational masses, clearing denominators produces many indistinguishable clones, and the flows can be realized by assigning different clones to different representatives. The continuous formulation is the high-multiplicity limit: individual root votes have weight \(O(1/N)\), while direct-vote fluctuations disappear as \(N\) grows. Exact treatment of knife-edge ties and the finite-\(N\) self-vote terms would be part of the discrete–continuous dictionary.

A plausible regime is a very large online or institutional community organized into repeated competence-network roles: for example, users sharing a credential band, topic channel, and access policy, with the same calibrated competence and the same role-level follow graph. Millions of agents can then occupy a few dozen or few hundred types. This is much stronger than merely saying that voters have similar competence. The network-access pattern is part of the type, as the programme requires. It is not a mirror for arbitrary Facebook graphs whose neighborhoods are all unique; in that setting the number of types grows with the population.

For the unrestricted typed-network problem, I would initially expect Class B, with hardness transferring. The cited reduction’s important combinatorics live in the variable-and-clause routing topology and in the relative weights of repeated supernodes, not merely in the names of individual leaves. Repeated leaves can become rational masses, while the backbone roles and their allowed delegation paths remain in \(H\). The constant gap between \(3/4\) and \(11/16\) suggests that a carefully normalized reduction could survive. This is a conjectured classification, not a result of the supplied paper. For fixed or highly regular \(H\), the problem could instead become Class A through flow, dynamic programming, or configuration methods; determining that boundary would be a worthwhile continuation.

The mirror covers only the optimal-delegation/ground-truth part of the paper. It does not claim to continuize the empirical comparisons of direct democracy, Liquid Better, Liquid Max, \(\rho\)-noisy sortition, simulated annealing, or the Monte Carlo methodology. Those are experiments, not named computational results, and should not be promoted into anchors.

The weakest point is that delegation is unusually identity- and topology-sensitive. An atomless population erases the identity of an individual root and makes direct-vote randomness self-average. Preserving the problem therefore requires a typed blow-up of the network and a finite set of macroscopic representative slots; it is not a literal replacement of every arbitrary \(n\)-vertex graph by type marginals. If a referee demands an exact mirror of arbitrary individual-level networks and exact finite-\(N\) probabilities, this proposal fails. But for repeated-role social networks—the kind of high-multiplicity scenario the programme explicitly permits—it remains an author-recognizable continuous version of their central computational question.