| paper | Tracking Truth by Weighting Proxies in Liquid Democracy |
| authors | — |
| venue | AAMAS 2022 |
| filed under | voting · theory |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c22r1) |
| judge confidence | medium |
| authors would recognise it | yes |
Theorem 3
statement extracted from the paper’s text layer
Given a finite type set \(T\), rational masses \(\mu\) with \(λ\mu_t∈ℤ\), accuracies \(q_t∈(1/2,1)\), and a connected type graph \(H\), form the cohort blow-up with \(λ\mu_t\) agents of type \(t\), identical accuracy, and uniform delegation access to adjacent cohorts. Each type chooses a feasible stochastic delegation kernel \(D\), inducing expected per-agent weights \(r_D(t)\). Find \(D\) maximizing \(Q_{λ,μ}(D)=\Pr[Σ_t r_D(t)(2X_t−λ\mu_t)>0]+\frac12\Pr[Σ_t r_D(t)(2X_t−λ\mu_t)=0]\), where \(X_t\sim\operatorname{Binomial}(λ\mu_t,q_t)\) independently.
A type-level mass-routing model on a regular cohort blow-up: μ is population mass, \(D_{tu}\) is delegation share, surplus and deficit mass are routed along \(H\), and the objective is the paper's expected-weight majority accuracy after lifting to λ\mu_t agents.
For fixed \(q_t>1/2\), \(Q_{λ,μ}(D)→1\) for every nondegenerate profile as \(λ→∞\), so the proposed finite-resolution objective has no informative atomless limit and the error-exponent rescue is not part of the paper.
fatal: False
The mirror covers the centralized optimal truth-tracking result, Algorithm 2, and Theorem 3 for cohort-regular high-multiplicity networks; it leaves decentralized equilibria, price of anarchy, limit-weight results, and experiments aside.
The strongest honest case is a Class A mirror of Theorem 3, and I would use no second anchor. Theorem 3 is proved in this paper: “Algorithm 2 outputs an element of \(\arg\max_{D\in\mathcal D}q_D\).” The surrounding discussion explicitly identifies Algorithm 2 as polynomial-time in the number of agents. The paper contains no named NP-hardness or parameterized-hardness result, so there is no hardness anchor to prefer.
My lead problem is Cohort-Weighted Delegation Truth Tracking\(_\infty\).
A type \(t\) is a cohort of agents sharing the same prediction accuracy \(q_t\in(1/2,1)\) and the same network role: they may delegate only to themselves or to agents in the neighbouring cohorts specified by a connected type graph \(H\). The type therefore includes both competence and delegation access; it is not merely an accuracy label.
An instance consists of a finite type set \(T\), rational masses \(\mu_t\) with \(\sum_t\mu_t=1\), accuracies \(q_t\), and the connected graph \(H\). The regime is \(\lambda\mu_t\) agents of type \(t\), where \(\lambda\) is large and \(\tau=|T|\) is fixed or moderate. Think of a large organisation, federation, or consultation network containing many repeated cohorts with the same expertise and communication position.
A decision is a type-level weighted delegation kernel \(D\). Each type \(t\) chooses shares \(D_{tu}\) over \(u=t\) and neighbouring types \(u\in H(t)\); delegation to a type \(u\) is distributed uniformly over the agents in that cohort. This is exactly the paper’s weighted-delegation action after quotienting symmetric clones by type. Transitive delegation remains present.
Let \(r_D(t)\) be the expected voting weight per member of type \(t\), obtained from the paper’s expected-weight construction, Equations (2)–(3), on the induced \(\lambda\)-agent clone instance. By symmetry, every member of a type has the same expected weight. If \(X_t\sim\operatorname{Binomial}(\lambda\mu_t,q_t)\) independently, define
\[ Q_{\lambda,\mu}(D) = \Pr\!\left[ \sum_{t\in T} r_D(t)\bigl(2X_t-\lambda\mu_t\bigr)>0 \right] +\frac12 \Pr\!\left[ \sum_{t\in T} r_D(t)\bigl(2X_t-\lambda\mu_t\bigr)=0 \right]. \]
This is the paper’s group-accuracy objective \(q_D\), written after aggregating identical agents by type. The problem is:
\[ \text{given }(T,\mu,q,H,\lambda),\quad \text{find a feasible }D\text{ maximizing }Q_{\lambda,\mu}(D). \]
A solution is the type-level delegation kernel \(D\), or an equivalent mass-routing certificate, together with its induced type weights \(r_D(t)\).
The expected answer is tractability. Put
\[ a_t=\log\frac{q_t}{1-q_t}, \qquad \bar a=\sum_{u\in T}\mu_u a_u. \]
The clone expansion of Theorem 3’s target weights gives the per-member target
\[ r_t^\star=\frac{a_t}{\bar a}. \]
Thus type \(t\) should retain total effective mass \(\mu_t r_t^\star\). Types with \(r_t^\star<1\) export surplus mass \(\mu_t(1-r_t^\star)\); types with \(r_t^\star>1\) receive mass \(\mu_t(r_t^\star-1)\). Since
\[ \sum_t\mu_t r_t^\star=\sum_t\mu_t=1, \]
the exports and imports balance. Connectedness of \(H\) lets us route the surplus along paths, exactly as Algorithm 2 routes weight between individual agents. Cycles can be cancelled, leaving an acyclic type-level flow. If \(F_{tu}\) is the routed mass from \(t\) to \(u\), with incoming mass \(I_t\), then the type-level delegation shares can be recovered by
\[ D_{tu}=\frac{F_{tu}}{\mu_t+I_t} \]
for \(u\neq t\), with the remaining share retained at \(t\). This uses only \(O(\tau^2)\) flow variables and avoids enumerating \(\lambda\) agents. The finite clone lift shows why the resulting profile is still the authors’ problem, not an arbitrary weighted-voting relaxation. The expected classification is therefore Class A: a polynomial-time mass-routing algorithm, subject to the usual numerical-precision treatment of the logarithms.
This mirror covers the paper’s centralized optimal-truth-tracking result, specifically Theorem 3 and its Algorithm 2. I would not count Theorem 4 as a second computational anchor: it gives a structural characterization of equilibria, but not a named complexity classification. It does, however, motivate a follow-up population problem concerning decentralized type-level Nash equilibria.
The weakest point is that the literal atomless limit is not completely benign. As \(\lambda\to\infty\), independent voting noise averages out and group accuracy tends to \(0\) or \(1\), so the finite-resolution objective \(Q_{\lambda,\mu}\), or an associated large-deviation error exponent, is more informative than a bare atomless probability. This makes the proposal an extension of the theorem rather than a purely literal replacement of \(n\) by a continuum.
A second limitation is structural: arbitrary individual delegation networks cannot be recovered from type masses alone. The mirror is persuasive for cohort-regular networks—many agents sharing the same accuracy and access role—but not for networks whose essential difficulty lies in individual identities or path histories. Further questions include disconnected type graphs, continuum versions of the paper’s limit-weight model, and decentralized weighted-delegation equilibria.
The sole anchor is Theorem 3, and it is a genuine named algorithmic result. The negative case therefore cannot deny that. It must show that the proposed “continuization” is either not population-continuous or collapses into a restatement.
The first problem is that the paper already has continuous decision variables. A weighted delegation profile \(D\) is an \(n\times n\) stochastic matrix with real entries, and Algorithm 2 is already a fractional mass-routing construction over the individual network. Replacing \(n\) agents by \(\lambda\mu_t\) clones and renaming their total weight \(\mu_t\) does not by itself continuize the population. It merely normalizes an algorithm the paper already gives.
The proposed type quotient is also exact only under a strong extra symmetry assumption. A type-level kernel can be lifted to clones only if, for example, every member of one cohort has the same access to every member of an adjacent cohort, and “retaining weight” means delegating to oneself rather than uniformly to one’s own cohort. Otherwise the same \((\mu,q,H)\) admits different individual networks and different transitive-delegation behaviour. Preserving arbitrary networks requires an identity-level graph, or a graphon-like relational kernel, in addition to the type distribution. That is no longer the paper’s finite-type problem. Restricting to regular cohort blow-ups is coherent, but it turns the proposed result into Algorithm 2 with supplies and demands scaled by \(\mu\).
More seriously, the literal population limit destroys the paper’s objective. For fixed type weights \(r_D(t)\),
\[ \mathbb{E}\!\left[\sum_t r_D(t)(2X_t-\lambda\mu_t)\right] = \lambda\sum_t \mu_t r_D(t)(2q_t-1), \]
while the variance is \(O(\lambda)\). Since the proposed setting has \(q_t>1/2\), every profile retaining a positive amount of voting weight has correctness probability converging to \(1\) as \(\lambda\to\infty\). Thus the atomless limit cannot distinguish the optimal routing in Theorem 3: almost every nondegenerate delegation policy ties at accuracy \(1\). The only genuinely different limiting cases are policies that lose essentially all weight through cycles, which is a pathology rather than a meaningful optimization landscape.
The proponent’s rescue—retain finite \(\lambda\), or optimize a large-deviation error exponent—is possible, but it changes the question. Finite \(\lambda\) is not an atomless society, and an error exponent is a reliability objective absent from the paper. Moreover, the optimal weight vector is already supplied by Theorem 1:
\[ r_t^\star=\frac{\log(q_t/(1-q_t))} {\sum_u\mu_u\log(q_u/(1-q_u))}. \]
Once connectedness permits routing, the rest is an ordinary transshipment calculation. The proposed continuous problem contributes no new pricing problem, complexity boundary, or population-dependent combinatorics; it compresses a theorem whose solution is already closed-form.
A graphon or measure-valued repair would be more genuinely continuous, but then the paper’s notion of a guru becomes problematic: in an atomless population, an individual proxy has measure zero. One must replace individual gurus and their weighted majority by a new operator-level delegation and voting model. That could be interesting, but it would be a new mean-field liquid-democracy theory, not a continuous mirror of Theorem 3.
So the strongest negative conclusion is that Theorem 3 yields, at most, a narrow high-multiplicity restatement for specially constructed regular cohort networks. Its direct limit is trivial; its finite version is already solved by the paper’s fractional algorithm; and its genuinely continuous repair changes the model. I would not claim that no researcher could ever make the resulting mean-field problem worthwhile, but the submitted anchor does not establish a worthwhile population continuization for this paper.
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.