| paper | Voter Priming Campaigns: Strategies, Equilibria, and Algorithms |
| authors | Jonathan Shaki, Yonatan Aumann, Sarit Kraus |
| venue | AAAI 2025 |
| filed under | voting · manipulation |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | yes |
Theorem 5
statement extracted from the paper’s text layer
Given a finite type set \(T\), rational masses \(\mu_t\) summing to \(1\), rational \(q_{t,i}(c)\), \(s_{t,i}(0)\), common sensitivities \(\rho_i\), issues \(I\), and budgets \(W^A,W^B\), compute \(i,j\in I\) such that \(w^A=W^Ae_i\) and \(w^B=W^Be_j\) form a pure Nash equilibrium for \(r_\mu(c,w)=P_\mu(c,w)/\sum_{c'}P_\mu(c',w)\), where \(P_\mu(c,w)=\sum_{t\in T}\mu_t\sum_{i\in I}q_{t,i}(c)(s_{t,i}(0)+\rho_iw_i)/\sum_{k\in I}(s_{t,k}(0)+\rho_kw_k)\).
A two-candidate parliamentary priming game over voter types \(t\), masses \(\mu_t\), and aggregate coefficients \(A_c(\mu)=\sum_t\mu_t\sum_iq_{t,i}(c)s_{t,i}(0)\) and \(B_{c,i}(\mu)=\sum_t\mu_tq_{t,i}(c)\). With \(W^*=W^A+W^B\), use \(Q_i^c(\mu)=A_c(\mu)/W^*+\rho_iB_{c,i}(\mu)\), and each candidate chooses a budget allocation \(w^c\) maximizing its vote share.
The proponent’s displayed \(Q_i^c(\mu)\) is algebraically incorrect: the baseline term must sum over all issues and be independent of \(i\), namely \(A_c(\mu)/W^*\); the mirror requires this repair.
fatal: False
The mirror covers Theorem 5’s two-candidate parliamentary focused-equilibrium existence result and its \(O(|I|^2+|V|)\) algorithm, with voter sums replaced by type-weighted sums. It leaves the general parliamentary and presidential results, including Theorems 2–4 and 6–12, unmirrored.
My strongest positive case is a Class A mirror of Theorem 5, which I regard as the paper’s best anchor.
Theorem 5, proved in this paper rather than cited, states that in the parliamentary setting with two candidates there is always a pure Nash equilibrium with focused investments, and that Algorithm 2 computes one in time \(O(|I|^2+|V|)\). This is a genuinely computational result: it identifies a restricted equilibrium structure and gives an algorithm.
The appropriate continuous population is a large electorate divided into a relatively small number of exact campaign-response types. A type \(t\) contains the complete information used by the model: the issue-specific quality scores \(q_{t,i}(c)\), the initial saliences \(s_{t,i}(0)\), and any induced preference ordering. The ranking alone is insufficient here, since the paper’s voting probabilities depend on the magnitudes of \(q\) and \(s\). Two voters have the same type exactly when they are indistinguishable in all these quantities.
The society is a rational distribution \(\mu\) over a finite type set \(T\), where \(\mu_t\) is the fraction of voters of type \(t\), \(\sum_t\mu_t=1\). The natural regime is a national or regional election with \(N\) very large and \(\tau=|T|\) small or moderate: for example, many voters in each demographic–geographic cohort have the same calibrated issue-quality and salience profile. This is not merely saying that voters have similar opinions. It says that the campaign model itself treats each cohort as behaviorally identical, which is exactly the high-multiplicity assumption.
There are two candidates \(A\) and \(B\), issues \(I\), issue sensitivities \(\rho_i\), and budgets \(W^A,W^B\). Candidate \(c\) chooses an investment vector \(w^c\in\mathbb{R}_{\ge 0}^{I}\) with \(\sum_i w_i^c=W^c\); let \(w=w^A+w^B\). For type \(t\), the salience of issue \(i\) is \(s_{t,i}(0)+\rho_iw_i\), and its voting probability for candidate \(c\) is
\(p_t(c,w)=\sum_{i\in I}q_{t,i}(c)\dfrac{s_{t,i}(0)+\rho_iw_i}{\sum_{j\in I}(s_{t,j}(0)+\rho_jw_j)}\).
The aggregate expected vote share is obtained by integrating over the population:
\(P_\mu(c,w)=\sum_{t\in T}\mu_t p_t(c,w)\),
and
\(r_\mu(c,w)=\dfrac{P_\mu(c,w)}{\sum_{c'\in\{A,B\}}P_\mu(c',w)}\).
Each candidate’s utility is \(r_\mu(c,w)\), exactly as in the paper’s parliamentary setting.
The continuous problem is therefore:
Two-Candidate Continuous Parliamentary Priming Equilibrium. Given the finite type table, rational masses \(\mu_t\), rational \(q_{t,i}(c)\), \(s_{t,i}(0)\), \(\rho_i\), and budgets \(W^A,W^B\), output issues \(i,j\in I\) such that \(w^A=W^Ae_i\) and \(w^B=W^Be_j\) form a pure Nash equilibrium. That is, for every feasible alternative allocation \(x^A\), \(r_\mu(A,w^A,w^B)\ge r_\mu(A,x^A,w^B)\), and symmetrically for \(B\).
This is not a weakened priming problem. It retains the paper’s two candidates, arbitrary issue qualities, voter abstention, salience amplification, budgets, vote-share utilities, and focused-equilibrium requirement. Only the electorate representation changes: the finite sum over named voters becomes a weighted sum over repeated types.
Theorem 1 explains why the mirror should be tractable. Replacing the voter sum by the population average gives
\(Q_i^c(\mu)=\dfrac{1}{W^A+W^B}\sum_{t\in T}\mu_tq_{t,i}(c)s_{t,i}(0)+\rho_i\sum_{t\in T}\mu_tq_{t,i}(c)\),
and the same algebra yields \(r_\mu(c,w)=\dfrac{Q^c(\mu)\cdot w}{Q^*(\mu)\cdot w}\). Thus the continuous society is represented by the same sufficient statistics as the discrete electorate, but the input size is the number of types rather than the number of voters. Algorithm 2 can then be run on these aggregated coefficients. Under rational input and exact arithmetic, the expected running time is polynomial in the type-table size and \( |I| \), roughly \(O(|I|^2+\tau |I|)\) for two candidates.
The expected classification is therefore Class A. The tractability comes from exactly the structural ingredients highlighted by the programme: the winning objective reduces to linear aggregate coefficients, and Theorem 5 shows that an equilibrium can be searched for among single-issue investments. Continuization does not create the structural theorem; rather, it gives the theorem its natural high-multiplicity formulation and removes irrelevant dependence on the number of repeated voters.
The authors should recognise this as their own problem. The paper already assumes that the electorate is large enough for expected vote counts to be treated as exact, and Theorem 1 explicitly eliminates individual-voter detail. A distribution over voter-response types is consequently closer to the paper’s native model than an artificial reinterpretation would be. For an atomic election with \(n_t\) voters of type \(t\), setting \(\mu_t=n_t/N\) recovers the same vote shares exactly. Conversely, a rational \(\mu\) can be implemented by a sufficiently large finite electorate.
This mirror deliberately covers only Theorem 5. It does not claim that all of the paper has been continuized. Natural follow-up questions include whether the all-candidate parliamentary equilibrium of Theorem 4 admits an FPT or approximation theory in the continuous type model; whether the exponential search in Algorithm 1 can be reduced using the aggregate \(Q\)-vectors; and how equilibrium behaves when \(\mu\) is estimated rather than known exactly. The presidential nonexistence results, especially Theorems 9 and 12, would also raise the useful boundary question of whether their nonexistence examples persist for diffuse rather than finitely supported populations.
The weakest point is that Theorem 1 has already performed most of the relevant aggregation. For this anchor, the continuous mirror may look like a mathematically clean high-multiplicity restatement rather than a new algorithmic breakthrough. The original paper’s running time is already linear in the number of voters, so the asymptotic gain is representational rather than dramatic. If a referee demands that a continuous mirror must change the complexity classification, this case is vulnerable. But that is a limitation of novelty, not of faithfulness: the paper’s own model, assumptions, and strongest two-candidate algorithm extend exactly to a distribution over repeated voter types.
The strongest case against the proposed mirror is that Theorem 5 has already removed the population from the computational object.
The paper’s own Theorem 1 shows that, for fixed campaign parameters, the entire electorate enters only through the aggregate quantities
\[ A_c(\mu)=\sum_{t\in T}\mu_t\sum_{i\in I}q_{t,i}(c)s_{t,i}(0) \]
and
\[ B_{c,i}(\mu)=\sum_{t\in T}\mu_t q_{t,i}(c). \]
Indeed,
\[ r_\mu(c,w)= \frac{A_c(\mu)+\sum_i \rho_i B_{c,i}(\mu)w_i} {\sum_{c'}A_{c'}(\mu)+\sum_i \rho_i B_{c',i}(\mu)w_i}. \]
Since every candidate invests exactly \(W^*\) in total, this can be written as the paper does, with
\[ Q_i^c(\mu)=\frac{A_c(\mu)}{W^*}+\rho_iB_{c,i}(\mu), \]
and then \(r_\mu(c,w)=\frac{Q^c(\mu)\cdot w}{Q^*(\mu)\cdot w}\).
Thus the distribution \(\mu\) is not part of the equilibrium problem except through a low-dimensional sufficient statistic. The candidates’ feasible actions, deviations, utilities, and equilibrium conditions contain no voter-level or type-level structure. Algorithm 2 is simply run on the resulting \(Q\)-vectors. Whether \(Q\) came from ten named voters, ten million repeated voters, or an abstract population measure is computationally invisible.
This also exposes a minor error in the positive case. Its displayed definition of \(Q_i^c(\mu)\) uses the issue-specific term \(\sum_t\mu_tq_{t,i}(c)s_{t,i}(0)\). The first term must instead contain the sum over all issues, \(A_c(\mu)/W^*\), and therefore be independent of \(i\). With that correction, the mirror works—but precisely as a weighted restatement of Theorem 1, not as a new population-continuous formulation.
The proposed national-cohort story is perfectly legitimate as a high-multiplicity encoding. It is not vulnerable to the usual objection that prices or preferences belong to individuals: exact duplicates are precisely what high multiplicity means. But it does not produce a population-level computational question. The paper already assumes a sufficiently large electorate that only expected aggregate votes matter, and already proves that individual voter data can be discarded. The continuous society therefore contributes compression of repeated input rows, not a new mass-allocation, separation, robustness, or approximation problem.
Trying to improve the mirror does not help while preserving the paper’s model. A type space containing richer \(q\)- and \(s\)-profiles still collapses to the same \(A\) and \(B\) statistics. In fact, the set of attainable aggregate statistics is a convex hull in dimension \(O(|C||I|)\), so arbitrary distributions add no essential structure beyond a small finite representation. Conversely, if one makes salience sensitivity type-dependent, introduces nonlinear response curves, or models heterogeneous exposure to campaigns, then the population may matter nonlinearly. But that changes the behavioural model on which Theorem 5 depends; it is a new priming game, not a continuous mirror of the theorem. The same objection applies to stronger attempts based on Theorem 4 or the presidential nonexistence results: with the paper’s primitives, they are all properties of the same finite-dimensional \(Q\)-game; with altered primitives, they are no longer results being mirrored.
The proponent’s only genuine gain is representational: a rational type table can encode a large cloned electorate more succinctly. That is a respectable high-multiplicity observation, but it is unusually weak here because the source paper has already performed the relevant aggregation and already gives an algorithm linear in the number of voters. The proposed “continuous problem” asks Algorithm 2 to solve exactly the object that Theorem 1 has already exposed.
I would therefore make the negative case as follows: Theorem 5 has a faithful continuous encoding, but not a worthwhile ChoCo mirror. The population is mathematically present only before the paper’s own sufficient-statistic reduction and disappears from the computational problem afterward. This is not an airtight universal impossibility claim—one could build a worthwhile heterogeneous-response extension—but that extension would be a new model rather than a mirror of this paper’s named result.
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.