| paper | Fine-Grained Liquid Democracy for Cumulative Ballots |
| authors | — |
| venue | AAMAS 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 8
statement extracted from the paper’s text layer
Given candidates \(C\), finite voter types \(T\), rational population masses \(\mu\in\mathbb{Q}_{\ge 0}^{T}\) with \(\sum_{t\in T}\mu_t=1\), type partitions \(\mathcal{S}_t\), bundle budgets \(b_{t,S}\), defaults \(\mathbf d_{t,S}\), weights \(w_{t,S}\), and role delegates \(\delta(t,S)\in T\), find an \(\epsilon\)-strong approximation to a resolved profile \(x=(x_t)_{t\in T}\) satisfying \(\sum_{c\in C}x_{t,c}=1\), \(\sum_{c\in S}x_{t,c}=b_{t,S}\), and \(x_{t,S}=b_{t,S}\frac{\mathbf d_{t,S}+w_{t,S}x_{\delta(t,S),S}}{\left\|\mathbf d_{t,S}+w_{t,S}x_{\delta(t,S),S}\right\|_1}\) for every \(t,S\), together with the aggregate ballot \(z_c=\sum_{t\in T}\mu_t x_{t,c}\).
A role-symmetric high-multiplicity FGLD model with \(\tau=|T|\) complete types, type-level delegation, per-type cumulative ballots \(x_t\), WCC fixed-point constraints, and population-weighted aggregate support \(z\).
The paper delegates to named voters, whereas the mirror delegates to roles or types, and \(\mu\) does not affect the WCC fixed-point equations; thus the transfer is a symmetric quotient rather than an automatic relaxation of every finite instance.
fatal: False
The mirror covers the WCC branch of Theorem 8, including \(\epsilon\)-strong approximation parameterized by the largest type-level SCDC, but leaves the other proportionality notions, existence results, meta-theorems, and open hardness questions untreated.
There is a credible, though deliberately narrow, continuous mirror here. I would lead with the paper’s WCC resolution problem and anchor it on Theorem 8, proved in this paper. The theorem gives an \( \epsilon \)-strong approximation in time \( \operatorname{polylog}(\mathbf w,\mathbf d,1/\epsilon)\cdot s^{O(s)}\cdot\operatorname{poly}(n,m) \), where \(s\) is the size of the largest strongly connected delegation component. This is a genuine parameterized-computational result, not merely an existence theorem. I would not use Theorem 7 as a second anchor: it is the Renegar-based subroutine from which Theorem 8 is built.
The mirror is a role-based, high-multiplicity version of fine-grained liquid democracy for cumulative ballots. Let \(C\) be the \(m\) candidates and let \(T\) be a finite set of complete voter types. A type \(t\) specifies its candidate partition \(\mathcal S_t\), bundle budgets \(b_{t,S}\), default vote \(\mathbf d_{t,S}\), confidence weight \(w_{t,S}\), and delegate role \(\delta(t,S)\). Two voters have the same type only if they agree on all of these data. The society is a rational distribution \( \mu\in\mathbb Q_{\ge 0}^{T} \) with \( \sum_t\mu_t=1 \).
The intended regime is a large recurring electorate: for example, a participatory-budgeting platform with \(10^5\) or \(10^6\) participants but only dozens or hundreds of recurring profiles of issue bundles, default allocations, confidence levels, and delegate roles. The relevant agents are not arbitrary named voters. They are members of repeated organizational or demographic cohorts whose entire delegation behaviour is interchangeable. If delegation histories are genuinely individual-specific, those histories must be included in the type, and the mirror correctly loses its compression benefit.
I would call the problem Typed-WCC-FGLD\(_\infty\). Its decision variable is a resolved cumulative ballot \(x_t\in\mathbb R_{\ge0}^{m}\) for each type \(t\), interpreted as the common per-capita ballot of that type. It must satisfy \( \sum_{c\in C}x_{t,c}=1 \) and \( \sum_{c\in S}x_{t,c}=b_{t,S} \) for every bundle \(S\in\mathcal S_t\). Under WCC, it must also satisfy, for every \(t\) and \(S\),
\(x_{t,S}=\dfrac{\mathbf d_{t,S}+w_{t,S}x_{\delta(t,S),S}}{\|\mathbf d_{t,S}+w_{t,S}x_{\delta(t,S),S}\|_1}\,b_{t,S}.\)
The objective is the paper’s own objective: resolve all delegations with zero WCC regret. In the computational version, given rational \( \epsilon>0 \), the required output is an \( \epsilon \)-strong approximation: a profile \(x\) within \( \epsilon \) in \( \ell_\infty \)-distance of some exact WCC fixed point \(x^\star\). The population-level resolved support is then reported as \(z_c=\sum_{t\in T}\mu_t x_{t,c}\). Thus \(x_t\) resolves the delegation network, while \(\mu\) determines how much each resolved role contributes to the collective ballot.
This is recognisably the authors’ problem. The mirror retains cumulative unit ballots, bundle budgets, proportional delegation, defaults, confidence weights, transitive delegation, and cyclic dependencies. It does not replace the delegation problem with an unrelated welfare objective. The only structural change is that repeated voters are represented by one type-level variable.
The high-multiplicity bridge is also concrete. If \( \mu_t=p_t/Q \), create \(p_t\) clones of type \(t\). In a symmetric role-based delegation network, every clone of type \(t\) receives the same resolved ballot \(x_t\), and the aggregate ballot is exactly \(Q^{-1}\sum_t p_t x_t\). Conversely, a finite clone population with this role symmetry collapses to the stated \( \mu \)-instance. Individual ballots remain unit ballots; \( \mu_t \) is population mass, not fractionalising one voter.
The expected classification is Class A, parameterized. Construct the type-level delegation graph on \(T\times C\), with an edge from \((t,c_1)\) to \((u,c_2)\) whenever \(c_1,c_2\in S\) and \(\delta(t,S)=u\). Let \(s_\infty\) be the largest strongly connected component. The WCC constraints are quadratic semialgebraic constraints in only \( \tau m \) variables. The paper’s Renegar formulation and reverse-topological SCDC decomposition should therefore give an algorithm with running time \( \operatorname{polylog}(\mathbf w,\mathbf d,1/\epsilon)\cdot s_\infty^{O(s_\infty)}\cdot\operatorname{poly}(\tau,m,L) \), where \(L\) is the input encoding length. In particular, bounded cyclic delegation structure gives polynomial time independently of the total population size. This is precisely the kind of high-multiplicity gain the continuization programme seeks.
The scope is intentionally limited. This mirrors the WCC branch of Theorem 8, not every result in the paper. It does not claim a continuum analogue of the paper’s unresolved PPAD-hardness question, nor of its remarks on strategic voting or social welfare. Proposition 1’s nonexistence result for EP-T is relevant background but not a computational anchor here. Theorem 9 supports existence for continuous best responses, but I would not present it as a separate complexity result.
The weakest point is that \( \mu \) does not enter the local WCC fixed-point equations; it affects multiplicity and the aggregate readout \(z\), but not which type-level fixed points exist. An opponent can therefore argue that this is a compressed delegation-role model with a partly passive population distribution. That criticism is real. The defence is that the paper itself studies delegation resolution before aggregation, so its anchored computational object is naturally insensitive to population weights. The mirror still replaces a potentially enormous named-voter system by a distribution over complete delegation types and gives a population-level collective ballot as output. A more strongly mass-sensitive variant—where a voter delegates to a \( \mu \)-weighted pool of delegate types—would be interesting, but it would be a new extension rather than something Theorem 8 already proves.
The resulting follow-up questions are whether type-level WCC solutions round cleanly to finite clone populations, whether population-weighted delegate pools remain fixed-parameter tractable, and whether a genuinely mass-sensitive aggregate-selection problem introduces continuum-specific hardness.
The strongest negative case is not that Theorem 8 lacks computational content. It does not: it is a genuine parameterized algorithmic result. The problem is that its parameter \(n\) counts vertices in an identity-sensitive delegation network, not members of an exchangeable population. That distinction prevents the proposed continuization from being a population mirror in the programme’s sense.
The paper’s primitive delegation is a point relation \(\delta(v,S)\in V\): voter \(v\) delegates to a particular voter. To make two voters the same complete type, one must include enough information about that relation—and recursively about the delegate’s behaviour—to ensure that they are interchangeable in the fixed-point problem. In a generic delegation network this information is essentially a rooted position in the graph, so the number of types is \(\tau\approx n\). The alleged high-multiplicity compression disappears.
The proponent’s construction avoids this only by imposing a strong role symmetry: every voter of type \(t\) delegates to a role \(u\), and all members of role \(u\) are assumed to have the same resolved ballot. Under that extra assumption, the equations
\[ x_{t,S} = b_{t,S} \frac{\mathbf d_{t,S}+w_{t,S}x_{u,S}} {\left\|\mathbf d_{t,S}+w_{t,S}x_{u,S}\right\|_1} \]
are a valid quotient of the original equations. But \(\mu\) does not occur in them. It only appears in the post hoc readout
\[ z_c=\sum_t\mu_t x_{t,c}. \]
Since the paper explicitly does not study aggregation, welfare, or the resulting collective decision, \(z\) is not part of the computational problem being mirrored. The proposed “continuous society” is therefore passive metadata attached to a smaller delegation game. Theorem 8 has been transferred by symmetry reduction, not by continuizing the population.
The natural large-electorate examples expose the problem. Suppose a mass of followers delegates to one guru. Preserving the paper’s semantics leaves a measure-zero individual whose ballot can determine the ballot of a positive-mass cohort. That is an atomic delegation network embedded in a continuum, not an ordinary distribution over exchangeable voter types. If there are many individually meaningful gurus, their identities must be retained and \(\tau\) grows with \(n\). If followers instead delegate to a population-level pool, then the model has changed.
One might try that better model explicitly, replacing point delegation by a kernel \(\kappa_{t,S,u}\):
\[ x_{t,S} = b_{t,S} \frac{\mathbf d_{t,S} +w_{t,S}\sum_u\kappa_{t,S,u}x_{u,S}} {\left\|\mathbf d_{t,S} +w_{t,S}\sum_u\kappa_{t,S,u}x_{u,S}\right\|_1}. \]
This could be an interesting mean-field theory, but it is not the high-multiplicity relaxation of the paper’s FGLD. In the original model, different followers delegating to different delegate types are different complete types. Moreover, because WCC is nonlinear, averaging delegates and then applying WCC is generally different from applying WCC to each delegate and averaging the results. The finite-clone bridge therefore does not determine this equation; it is a new institutional rule.
The same objection applies to Theorem 9’s spatial-voting generalization. Continuous ballots or spatial actions are continuity of the action space, not of the population. A distribution over such agents still faces the point-delegation problem, while making delegation population-weighted again creates a new model.
The negative case is not airtight. A platform with millions of followers and a small, genuinely role-symmetric set of delegates could reasonably accept the proponent’s quotient, and then Theorem 8 would yield a useful compressed solver. But that is best described as an algorithm for symmetric delegation-role networks whose aggregate mass is subsequently reported—not as a computational mirror in which the society itself becomes the continuous object. On the paper’s actual computational result, no stronger population-sensitive mirror is supplied without changing what delegation means.
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.