| paper | Federated Assemblies |
| authors | Daniel Halpern, Ariel D. Procaccia, Ehud Shapiro, Nimrod Talmon |
| venue | AAAI 2025 |
| filed under | multiwinner · multiwinner |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | yes |
Theorem 2
statement extracted from the paper’s text layer
Given a rooted laminar tree, rational masses of exact single-leaf types, and rational assembly mass \(a\) with \(M_v\) at least \(a\), construct a randomized assignment of measurable sets \(A_v\) of mass \(a\) satisfying inheritance, pointwise inclusion probability \(a/M_v\), and child intersection quota \(a M_c/M_f\) both in expectation and in every realization.
Exact leaf-signature types with rational mass, a positive assembly mass scaling with population, and randomized inherited mass subsets as decision variables; the objective is exact feasibility and constructive sampling.
A nondegenerate limit requires scaling assembly size with population, producing a macroscopic mass committee rather than the paper's finite deliberative panel; the proponent acknowledges this semantic shift but does not show that practitioners would accept it.
fatal: False
The mirror covers Theorems 1 and 2 directly and gives a high-multiplicity limit of Theorem 3, but leaves finite-seat rounding, arbitrary-DAG feasibility, dynamic over-time fairness, and the paper's empirical results untouched.
The positive case is real, but it is a mirror of federated sortition—not of voting. The paper has no named NP-hardness, W[1]-hardness, or explicit “in P” theorem. Its relevant anchors are three named constructive theorems, all proved by the authors: Theorems 1 and 2 in the paper itself, and Theorem 3 with its proof relegated to the full version.
The cleanest high-multiplicity model is this. Let \(\mathcal L\) be the leaf assemblies and let a type be an exact signup signature \(L\subseteq\mathcal L\): two people have the same type precisely when they belong to the same leaves. Give each realized type \(L\) rational mass \(\mu_L\). Thus the population is the atomless measure space
\[ \Omega=\bigsqcup_L \bigl(\{L\}\times[0,\mu_L)\bigr). \]
For a node \(v\), let \(\Omega_v\) contain the types signed up for at least one descendant leaf, with mass \(M_v\). Let \(a\) be the target assembly mass, requiring \(M_v\ge a\) for every node.
A randomized continuous assembly assignment is a distribution over measurable sets \(A_v\subseteq\Omega_v\), each of mass \(a\), satisfying inheritance
\[ A_f\subseteq\bigcup_{c\in\mathrm{CHILDREN}(f)}A_c. \]
Individual representation means that almost every member of \(\Omega_v\) is selected with probability \(a/M_v\). For a federation \(f\) and child \(c\), define the paper’s weighted child mass
\[ W_{c,f}=\sum_{L:L\cap D(c)\neq\varnothing} \frac{\mu_L}{|\{d\in\mathrm{CHILDREN}(f):L\cap D(d)\neq\varnothing\}|}, \]
and \(q_{c,f}=W_{c,f}/M_f\). Ex ante child representation requires
\[ \mathbb E[|A_f\cap A_c|]\ge a q_{c,f}, \]
while the continuous ex-post version requires this inequality in every realization.
This is computationally finite in representation. For each type \(L\), consider the finite collection of node-selection patterns \(S\) satisfying inheritance: whenever \(f\in S\), at least one child of \(f\) is also in \(S\). A realization is described by masses \(z_{L,S}\). Assembly sizes, intersections, and representation probabilities are all linear expressions in these variables. A finite convex combination of such pattern allocations is therefore a certificate for the randomized solution. Clearing denominators and scaling all population masses and the assembly mass by \(k\) recovers a finite high-multiplicity instance. The assembly itself remains an all-or-nothing measurable subset in each realization; this is not fractionalizing a policy outcome.
My lead anchor is Theorem 2, proved in the paper: Algorithm 2 satisfies individual representation, ex ante child representation, and ex post child representation on laminar instances.
The corresponding problem is Continuous Laminar Federated Sortition. The input is a rooted tree \(G\), rational masses \(\mu_\ell\) at its leaves, an assembly mass \(a\), and the assumption that each person belongs to exactly one leaf. The task is to output a randomized inherited assignment satisfying the three conditions above. Since child populations are disjoint, \(W_{c,f}=M_c\). The continuous algorithm is exact:
Every child receives exactly its continuous quota, so ex-post representation is exact rather than rounded down. Along a root-to-leaf path, the inclusion probabilities telescope to \(a/M_f\), giving individual representation. This is a direct, author-recognizable mirror of Theorem 2 and should be Class A: the solution is an explicit rational mass-flow construction.
The regime is plausible for a large federated civic platform or a hierarchy of municipal assemblies: millions of participants, perhaps dozens or hundreds of leaf communities, and many repeated members of each leaf-signature type. Population and assembly mass must scale together when taking the clone limit; holding a finite \(n\)-person panel fixed while population tends to infinity would make the normalized problem degenerate.
A second, broader anchor is Theorem 1, proved in the paper: Algorithm 1 satisfies individual representation and ex ante child representation.
Its mirror is Continuous DAG Ex-Ante Sortition. The input is an arbitrary finite DAG, rational type masses \(\mu_L\), and assembly mass \(a\). The task is to output an inherited randomized assignment satisfying individual and ex-ante child representation; ex-post quotas are not required. The continuous analogue of Algorithm 1 assigns a common random priority to the population and takes the top mass \(a\) in every node population. Equivalently, one can use the finite type-pattern formulation above.
If \(c\) is a child of \(f\), then \(\Omega_c\subseteq\Omega_f\), so anyone selected among the top \(a\) mass of \(\Omega_f\) is also selected among the top \(a\) mass of \(\Omega_c\). The expected intersection is \(aM_c/M_f\), which dominates \(aW_{c,f}/M_f\). This is again Class A, with a direct global-priority construction. It covers only the first two guarantees, but it extends beyond laminar graphs.
The richer and more ambitious anchor is Theorem 3, proved by the authors with the proof placed in the full version: under regularity conditions, Algorithm 3 on semi-laminar instances gives individual representation, ex ante child representation, and approximate ex-post representation within one seat.
The corresponding problem is Continuous Semi-Laminar Federated Sortition. The input has the paper’s region tree \(R\), topic set \(T\), and rational masses \(\mu_{r,L}\), where \(L\subseteq T\) is the exact set of topics joined by a person in leaf region \(r\). The task is to output an inherited randomized assignment satisfying individual representation, ex-ante representation, and exact continuous ex-post quotas
\[ |A_f\cap A_c|\ge a q_{c,f}. \]
This is the high-multiplicity limit of the paper’s guarantee. Scale all masses and the assembly mass by \(k\). The discrete bound becomes
\[ \lfloor k a q_{c,f}\rfloor-1. \]
After dividing by \(k\), the deficit is at most \(2/k\), so it vanishes. The paper’s regularity assumptions remain meaningful as ratio conditions; absolute lower bounds such as \(n\ge 2/(\epsilon\delta)\) and \(|C_L|\ge2\) become automatic along the scaling sequence.
I would expect this regular semi-laminar problem to be Class A, because the paper’s machinery is already based on dependent rounding, Birkhoff–von Neumann decompositions, and convex programming. In the continuum, the integer-rounding obstruction disappears. The important open questions are whether exact feasibility holds without the regularity assumptions, whether a polynomial-size extended formulation exists, and whether fully general overlapping DAGs contain a continuum-specific hard case.
The mirror is strongest in the laminar and semi-laminar scenarios. The types are not an artificial simplification: \(L_i\) is already the paper’s complete equivalence signature. A realistic regime is a large digital federation in which millions of users share a moderate number of geographic/topic-membership signatures. If identity history, arrival time, personal trust links, or other named relationships become part of the constraints, they must be added to the type; if nearly every person becomes a distinct type, the high-multiplicity case disappears.
The weakest point is that the continuous version scales the assembly resource jointly with the population and replaces integer quota rounding by exact mass quotas. That is appropriate for a high-multiplicity limit, but less compelling for a traditional citizens’ assembly that always has, say, exactly 20 seats. I would therefore not claim that the whole paper has been continuized, or that arbitrary overlapping instances are tractable. The defensible claim is narrower: Theorems 1–3 have faithful continuous population analogues in natural large federated communities, with Theorem 2 providing the safest direct Class-A mirror and Theorem 3 offering a substantive high-multiplicity extension.
The strongest case against the mirror is that the paper’s computational object is an indivisible, finite deliberative panel. Its central difficulty is selecting an \(n\)-element subset while preserving integer child quotas, inheritance, and avoidance of duplicate representatives.
If population size grows while \(n\) remains fixed, the normalized assembly mass \(a=n/N\) tends to zero. Individual representation and all child quotas then vanish, so the limiting problem is degenerate. If \(n\) is scaled with \(N\) to keep \(a>0\), the output is no longer a citizens’ assembly in the paper’s sense: it is a positive-mass population allocation, potentially containing an uncountable number of members. That is the best charitable mass formulation, but it changes the semantic object rather than merely replacing multiplicities.
This defeats the proposed mirror of Theorem 1. The common-priority construction is mathematically valid after that re-modelling, but it is only a direct mass version of an already explicit finite sampling algorithm. It introduces no missing computational problem: there is no optimization, decision threshold, or unresolved representation question. Recasting it as a type-pattern LP makes the formulation look richer, but then the computational object is the newly invented distribution-over-patterns problem, not the theorem’s random-selection algorithm.
Theorem 2 has the same problem even more starkly. In the laminar case, the continuous construction simply sends mass \(aM_c/M_f\) down every tree edge and uses measure-preserving randomization within each type. The integer rounding and seat-allocation issues—the only technically substantive part of Algorithm 2—have disappeared. This is a legitimate fractional-flow corollary, but it is not a worthwhile new computational mirror unless one counts an immediate restatement of the finite theorem as sufficient.
Theorem 3 does not rescue the case. Scaling the instance by \(k\) and dividing the \(\lfloor kaq\rfloor-1\) guarantee by \(k\) does suggest an asymptotic exact-quota statement, but that is not yet an algorithmic result over rational masses. The paper gives no uniform complexity bound for constructing the limiting distribution. The proponent’s finite pattern formulation has exponentially many node-selection patterns, and no pricing or separation theorem is supplied. Thus there are only two possibilities: either the continuous problem collapses to a simple fractional feasibility construction, in which case the paper’s dependent-rounding challenge has been erased, or it becomes a new exponential-support configuration problem whose difficulty is not established by Theorem 3.
The exact-signature type model is not itself objectionable. Membership signatures really are the paper’s complete equivalence classes, and repeated signatures could occur in a large digital platform. The objection is narrower: the only nondegenerate version requires treating a deliberative committee as a macroscopic mass, while the fixed-panel version loses all substance. That makes this a poor candidate for a population-continuization programme aimed at discovering new computational landscapes.
This negative case is not airtight. Under ChoCo’s permissive high-multiplicity convention, the laminar mass-flow construction is a genuine, author-recognizable Class-A mirror, and the semi-laminar limit could be a useful extension. I therefore could not honestly defend the universal claim that no worthwhile scenario exists. The most defensible negative verdict is weaker: the paper offers clean continuous restatements, but no convincing new computational mirror beyond fractionalizing away the very integrality constraints that motivate its results.
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.