| paper | Optimizing Viscous Democracy |
| authors | Ben Armstrong, Shiri Alouf-Heffetz, Nimrod Talmon |
| venue | IJCAI 2024 |
| filed under | voting · theory |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | medium |
| authors would recognise it | yes |
Theorem 1
Given rational type masses \(\mu_t\), competence values \(q_t\), an allowed typed delegation graph \(H\), and \(\rho>0\), choose real masses \(r_t\) and \(y_{t,u}\) satisfying \(r_t+\sum_{u\in N_H(t)}y_{t,u}=\mu_t\), \(\sum_t r_t=\rho\), nonnegativity, and acyclic positive support, maximizing atomless weighted-majority accuracy: \(1\) if \(\sum_t q_t r_t>\rho/2\), \(0\) if it is below \(\rho/2\), and \(1/2\) at equality.
An exchangeable high-multiplicity typed delegation network with role masses \(\mu_t\), competences \(q_t\), direct-voting masses \(r_t\), delegation-flow masses \(y_{t,u}\), an acyclic type-level support, and asymptotic weighted-majority accuracy as objective.
The fixed-\(\rho\) law-of-large-numbers objective is not literally Theorem 1's variable-size finite-sample accuracy and strips away most nontrivial delegation structure.
fatal: False
The mirror covers Theorem 1 at \(\alpha=0\) and proposes a secondary extension of Theorem 2 at \(\alpha=1/m\); it leaves the simulations, qualitative \(\alpha^\ast\) examples, and cited Corollary 1 outside the main mirror.
There is a defensible mirror, although I would not claim that every result in the paper continuizes equally well. My strongest anchor is Theorem 1; Theorem 2 gives a more ambitious secondary question.
The regime is a large deliberation platform or DAO organised into repeated communities. A type \(t\) records everything relevant to delegation: competence \(q_t\), permitted delegation channels, and network role. There are \(\tau\) such roles and \(N\gg\tau\) participants, with mass \(\mu_t\) of type \(t\). Repeating each role many times is the high-multiplicity operation; it is not a claim that every arbitrary social network has few types.
Let \(H=(T,E_H)\) be the typed delegation network. A continuum delegation plan chooses direct-voting mass \(r_t\) and delegated mass \(y_{t,u}\), satisfying
\[ r_t+\sum_{u\in N_H(t)}y_{t,u}=\mu_t. \]
The positive-support directed graph of \(y\) must be acyclic. Thus a type cohort may split its mass between voting directly and delegating through several permitted channels, exactly as many indistinguishable agents can do in a high-multiplicity population.
Write \(P_{t,u}=y_{t,u}/\mu_t\) for the induced delegation proportions and \(g_t=r_t/\mu_t\). If \(v^{(0)}=\mu\) and \(v^{(\ell+1)}=v^{(\ell)}P\), then the effective weight carried by type \(t\) is
\[ W_t=g_t\sum_{\ell\ge 0}\alpha^\ell v_t^{(\ell)}. \]
The sum is finite because the delegation support is acyclic. This is the same distance-dependent attenuation as in the paper, but applied to mass rather than named voters.
For a diffuse continuum of independent competence signals, the finite-election winner probability has a law-of-large-numbers limit. I use
\[ \operatorname{Acc}_\infty(r,y)= \begin{cases} 1,&\sum_t q_tW_t>\frac12\sum_tW_t,\\ \frac12,&\sum_t q_tW_t=\frac12\sum_tW_t,\\ 0,&\sum_t q_tW_t<\frac12\sum_tW_t. \end{cases} \]
A fixed direct-voting mass \(\sum_t r_t=\rho>0\) avoids the degenerate solution of retaining an arbitrarily small number of highly competent voters. This is also faithful to the paper’s fixed-delegator-rate experiments.
My lead problem is therefore:
\[ \textbf{Mass-ODG}_0(\rho). \]
Given rational \(H,q,\mu,\rho\), find a feasible acyclic mass delegation plan \((r,y)\) with \(\sum_t r_t=\rho\) maximizing \(\operatorname{Acc}_\infty\).
This is a genuine continuous population problem: the \(r_t\) and \(y_{t,u}\) are real masses, and the input size depends on the number of types rather than the number of repeated participants. It mirrors the paper’s Theorem 1, “For \(\alpha=0\), ODG is in P,” which is proved in this paper. At \(\alpha=0\), \(W_t=r_t\), so delegation affects only which mass remains active. On a connected typed network, any discarded mass can be routed toward a direct-voting type using a forest, so the optimization reduces to retaining the highest-competence mass, with partial retention of the marginal type if necessary. Sorting the types and checking the rational threshold gives a polynomial algorithm. Thus this is a natural Class A mirror of the paper’s tractable boundary.
The paper’s full ODG allows the number of gurus to vary; the fixed-\(\rho\) version is the high-multiplicity analogue of fixing \(k\) in the proof of Theorem 1. One can also study the outer problem of selecting \(\rho\), but fixing the active fraction is the cleaner and more meaningful population-scale formulation.
My secondary anchor is Theorem 2, proved here: “For \(\alpha=1/m\), \(m\in\mathbb N\), ODG is NP-Hard.” The corresponding question is
\[ \textbf{Mass-ODG}_{1/m}(\rho), \]
with exactly the same instance and solution definition, but with \(\alpha=1/m\), and with the objective \(\operatorname{Acc}_\infty\).
I would not claim that Theorem 2’s NP-hardness transfers directly. Its X3C reduction relies on an individual element voter choosing one of three incident set voters. In the continuum, the mass of that element type can split fractionally among all three. That destroys the integral exact-cover argument. This is precisely why the question is interesting rather than a mechanical restatement.
My tentative expectation is that \(\textbf{Mass-ODG}_{1/m}(\rho)\) is a Class C candidate: the type-level delegation proportions create products along paths, while the requirement that the positive-support graph remain acyclic is a global combinatorial constraint. Those features survive continuization even though the particular X3C integrality gadget does not. It may turn out to be tractable through a nonlinear-flow or configuration formulation, but if it remains hard, the hardness would be genuinely continuum-specific rather than inherited from population indivisibility.
The authors should recognise both mirrors as their problem. The candidates, competence values, network restrictions, transitive delegation, acyclicity, guru weights, viscosity parameter, and weighted-majority epistemic objective are all retained. Only the population representation changes from individual voters to masses of indistinguishable voter roles. I would not use Corollary 1 as a third anchor: it is cited from Caragiannis and Micha, and its \(\alpha=1\) hardness has the same integrality-transfer problem without adding an independent continuous question.
The weakest point is the accuracy semantics. Literal finite-population group accuracy is not preserved unchanged in an atomless society: independent votes concentrate, so the probability becomes the threshold functional \(\operatorname{Acc}_\infty\). If the authors insist that a guru must be a named individual with one shared random signal, then a mixed atomic–continuum model is needed, and the result is less purely population-continuous. The second weakness is that arbitrary networks may have \(\tau\) nearly equal to \(N\); the mirror is strongest for repeated network roles, not for a committee in which every voter has a unique neighbourhood.
The resulting programme is still substantial: determine whether fractional mass splitting makes \(\textbf{Mass-ODG}_{1/m}(\rho)\) polynomial, identify restrictions such as trees or bounded-treewidth type networks, and establish finite-\(N\) approximation bounds between the discrete ODG problem and its continuum limit. That is a meaningful continuous computational-social-choice agenda anchored directly in Theorems 1 and 2.
Neither anchor survives as a worthwhile mirror. They share a more fundamental problem: ODG is not an electorate-profile problem. Its computational object is a labelled, acyclic subgraph. A voter’s identity matters because it determines both her competence and her exact neighbourhood.
A finite type space cannot preserve that structure in general. If a type records the precise neighbourhood needed by ODG, then distinct graph vertices usually have distinct types, so \(\tau\) grows with \(N\) and there is no high-multiplicity compression. If several vertices are identified as one “network role,” the resulting type graph is a blow-up of the original graph and permits different delegation plans. Repeated communities do not fix this: preserving which members belong to the same community requires a coupling or community label, which again scales with the number of communities. Allowing every copy of a role to delegate to every copy of another role is a new exchangeable network model, not a population representation of the paper’s \(G\).
The proposed acyclicity constraint exposes the problem especially clearly. Take copies ordered as
\[ A_1\to B_1\to A_2\to B_2\to\cdots\to A_k\to B_k. \]
Every individual delegation graph is acyclic, but the aggregate type support contains both \(A\to B\) and \(B\to A\). Thus type-level acyclicity excludes valid high-multiplicity realizations. Allowing the type cycle does not solve it: the same aggregate matrix can represent a literal cycle, which is forbidden, or an arbitrarily long alternating chain, whose viscous weight depends on its length. One must add individual identities or a rank/depth coordinate. If depth grows with \(N\), the finite-type population mirror has disappeared; if depth is bounded, one has imposed a new bounded-depth problem.
Theorem 1 is not rescued by fixing \(\rho\). At \(\alpha=0\), delegation paths have zero weight, so the network, transitivity, and viscosity all vanish. The finite theorem concerns selecting named voters and computing a finite-sample majority probability. In the proposed atomless model with independent signals and positive direct mass \(\rho\), the law of large numbers turns that probability into a threshold:
\[ \operatorname{Acc}_\infty= \mathbf{1}\!\left[\sum_t q_t r_t>\frac12\sum_t r_t\right] \]
apart from the tie case. The “optimization” is simply retaining the highest-competence mass until \(\rho\) is reached. That is a continuous knapsack calculation, not a mirror of the paper’s delegation problem. Removing the fixed \(\rho\) permits vanishingly small active populations; preserving the finite-sample probability requires atomic gurus or finite-\(N\) corrections. In either repair, the genuinely nontrivial individual component returns.
Theorem 2 is even less compatible with continuization. Its X3C reduction uses the identity of every element and set voter: each \(x_i\) has its own incidence triple, and each \(S_j\) has its own membership pattern. Those incidences are exactly what determine the feasible delegations. A type that preserves them is essentially a named vertex, so the reduction has no multiplicity regime. Repeating the gadget only repeats copies around a still-unique X3C core.
Collapsing the element voters into a common type and allowing mass to split among the three set types is a legitimate fractional network model, but it is no longer the paper’s one-delegate-per-voter problem represented at high multiplicity. It has replaced the incidence-level delegation choice by a flow variable. That may be an interesting new optimization problem, but it is not anchored by Theorem 2 in the way claimed. Giving each type a shared signal instead of independent signals would preserve nontrivial accuracy, but then the real decision-makers are finitely many type-level experts, not an atomless population.
The honest weakness in this negative case is that one can deliberately define a fixed, exchangeable role template with bounded delegation depth and call its mass-flow problem a continuous social-choice model. I would not claim that such a model is impossible or useless. I would claim that it is a new typed network-flow programme requiring its own motivation. For the paper’s actual computational results, \(\alpha=0\) degenerates into deterministic mass selection, while \(\alpha=1/m\) depends on precisely the identities, incidence relations, and global acyclicity that a finite population distribution forgets. On the paper as written, neither theorem supplies a credible ChoCo mirror.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.