Iterative Local Voting (ILV)

• Iterative Local Voting (ILV) is a decision-making process that enables voters to incrementally adjust outcomes by making local modifications in complex, high-dimensional spaces.
• The method restricts voters to improve within a bounded neighborhood—using norm-ball constraints or local dominance in discrete domains—ensuring convergence and fairness.
• Empirical evaluations reveal that ILV reduces cycles and biases, often converging to socially optimal or median outcomes, thereby enhancing collective welfare.

Iterative Local Voting (ILV) is a collective decision-making procedure for high-dimensional or multi-issue domains in which voters incrementally steer the outcome by locally modifying candidate solutions. Unlike classical voting, which usually aggregates full preference reports in one round, ILV restricts voters to local improvement steps—either in a discrete combinatorial space (e.g., issues with candidate sets) or a continuous space (e.g., resource allocation). Theoretical analysis and empirical studies of ILV uncover convergence properties, equilibrium characterizations, fairness guarantees, and behavioral regularities under diverse domain models and interaction protocols (Garg et al., 2017, Kavner et al., 2023).

1. Domain Formalization and Algorithmic Structure

ILV operates over either:

• Discrete multi-issue domains: $\mathcal{D} = \prod_{i=1}^p D_i$ where $D_i$ is the candidate set for issue $i$ (Kavner et al., 2023).
• Continuous high-dimensional spaces: $\mathcal{X} \subset \mathbb{R}^M$ convex, closed, and bounded (Garg et al., 2017).

Each voter is associated with a utility function $f_v(x)$ (continuous) or strict ranking $R_j$ over $\mathcal{D}$ (discrete). The core workflow:

For continuous domains (Garg et al., 2017):

1. Initialize $x_0 \in \mathcal{X}$ and set radius $r_0$.
2. At each round $t$, select a random voter $D_i$0.
3. Ask $D_i$1 for their maximally preferred point within a norm ball of radius $D_i$2 around $D_i$3.
4. Update to $D_i$4 (optionally project to $D_i$5).
5. Reduce $D_i$6 according to a step-size schedule.

For multi-issue discrete domains (Kavner et al., 2023):

1. Fix an initial profile $D_i$7.
2. At each step, schedule an agent $D_i$8 and issue $D_i$9.
3. Agent $i$0 considers all single-issue deviations $i$1 and evaluates local dominance (see Section 2).
4. If local dominance is found, $i$2 switches $i$3 accordingly.
5. Repeat until no locally dominating deviation exists for any agent/issue.

2. Local Dominance and Bounded Rationality

ILV's refinement over classical improvement dynamics arises from a bounded rationality principle: voters evaluate options only within local uncertainty sets and without precise probabilistic beliefs.

For discrete domains (Kavner et al., 2023):

• Each agent $i$4 maintains an uncertainty set $i$5 over possible score vectors $i$6 for all issues.
• Local dominance: $i$7 locally dominates $i$8 if (i) for all $i$9, $\mathcal{X} \subset \mathbb{R}^M$0 and (ii) for some $\mathcal{X} \subset \mathbb{R}^M$1, $\mathcal{X} \subset \mathbb{R}^M$2.

For continuous domains (Garg et al., 2017):

• Local moves are restricted by norm-balls; a voter chooses the local maximizer inside the allowed neighborhood.
• This can be interpreted either as explicit support for local maxima or as a noisy subgradient, reflecting bounded information and effort.

This framework models realistic behavior in settings where full knowledge is unavailable or cognitive effort is limited.

3. Convergence Theorems and Equilibrium Properties

The convergence properties of ILV depend critically on domain structure, uncertainty, and preference dependencies.

Continuous spaces (Garg et al., 2017):

• Spatial (Lp) utility, dual norm restriction: For $\mathcal{X} \subset \mathbb{R}^M$3 and $\mathcal{X} \subset \mathbb{R}^M$4-norm neighborhoods with $\mathcal{X} \subset \mathbb{R}^M$5, ILV converges almost surely to the social welfare maximizer for $\mathcal{X} \subset \mathbb{R}^M$6.
• Weighted Euclidean/utilitarian utilities: The limit point is the unique welfare maximizer.
• Additive decomposability with $\mathcal{X} \subset \mathbb{R}^M$7 neighborhoods: Converges to the per-coordinate medians (coordinate-wise, Pareto-optimal, midrange fairness).

Multi-issue discrete domains (Kavner et al., 2023):

• Cycles possible: General ILV dynamics may not converge, even with binary issues and fixed uncertainty radii.
• Sufficient conditions for convergence:
  • O-legal preferences: If rankings' dependency on issues is structured (each marginal ranking depends only on prior issues in some order $\mathcal{X} \subset \mathbb{R}^M$8), dynamics are acyclic.
  • Alternating uncertainty: Agents set tighter uncertainty on the issue in play and looser on others, preventing cycling.
• Empirical finding: Even modest uncertainty (larger $\mathcal{X} \subset \mathbb{R}^M$9) in agents' models sharply reduces cycling frequency and typically ensures convergence in practice.

Summary of theoretical conditions:

Setting	Convergence Guarantee	Reference
Continuous, spatial with dual norms	Social optimum, almost surely	(Garg et al., 2017)
Discrete, no uncertainty or structure	Cycles possible	(Kavner et al., 2023)
Discrete, O-legal or alternating $f_v(x)$0	Finite-time convergence	(Kavner et al., 2023)

4. Fairness, Welfare, and Societal Outcomes

ILV exhibits natural welfare and fairness properties aligned with the chosen utility and neighborhood model.

• Social welfare maximization: For $f_v(x)$1 utility, converges to the $f_v(x)$2 minimizing total distance to voter ideals—utilitarian optimality.
• Pareto optimality and robustness: With decomposable utility and $f_v(x)$3 neighborhoods, the outcome is the vector of coordinate-wise medians (robust to outliers, "midrange fairness").
• Majority core, no-envy: In all cases, outcomes are in the "majority core"—no individual can locally deviate to improve, aligning with group stability and group no-envy.

Empirical studies confirm that strategic local adjustment in ILV generally improves aggregate welfare over truthful voting, and convergence points are often consistent and robust across initializations (Kavner et al., 2023).

5. Empirical Evaluation and Implementation Insights

ILV's practical behavior has been investigated through large-scale experiments in budget allocation and extensive computational simulations.

Mechanical Turk experiment (continuous ILV) (Garg et al., 2017):

• Voters (≈2,000) allocated a hypothetical US federal budget across $f_v(x)$4 categories with either $f_v(x)$5, $f_v(x)$6, or $f_v(x)$7 norm constraints.
• $f_v(x)$8 neighborhoods yielded rapid, consistent convergence to nearly identical limit points, approximating the median budget in each category.
• $f_v(x)$9 and $R_j$0 produced multiple distinct equilibria.
• User interface (slider anchoring) exerted substantial bias on full-report settings, but constrained ILV responses mitigated this bias.
• Many voters used less than their available credits in $R_j$1, indicating large local indifference regions or flat utility zones.

Simulation studies (discrete multi-issue ILV) (Kavner et al., 2023):

• As uncertainty radius $R_j$2 increases, cycle frequency falls and equilibrium is quickly reached.
• Welfare gains (Borda sum) from iterative dynamics typically outpace those from truthful profiles, but these gains decrease slightly as uncertainty increases.

6. Limitations, Open Problems, and Extensions

ILV's theoretical and empirical analyses highlight several subtleties and unresolved challenges:

• Non-convergence in complex settings: For multi-candidate issues or when preference dependencies are arbitrary, cycles can persist even with uncertainty refinements (Kavner et al., 2023).
• Implicit utilities: Discrete ILV typically relies on ordinal (not cardinal) preferences; thus, welfare results are indirect.
• Scheduler design: Adaptive selection of agents and issues could accelerate convergence in future implementations.
• Strategic and privacy considerations: Partial reporting, incentives, and privacy-preserving mechanisms remain open directions.
• Extension to simultaneous or nonatomic updating: Asynchronous/asymmetric protocols and population models may further affect convergence and welfare.

7. Relation to Broader Voting and Collective Optimization Literature

ILV connects strategic voting under bounded rationality (as in local-dominance theory (Meir et al., 2014)) with iterative improvement heuristics common in optimization and market design. Its stochastic subgradient interpretation (continuous domains) aligns with distributed optimization and learning under local oracle access (Garg et al., 2017). The equilibrium definitions and dominance relations generalize classical Nash, ordinal dominance, and core concepts suited to high-dimensional or combinatorial settings. Empirical findings corroborate long-standing descriptive principles such as Duverger's law, by capturing genuinely emergent patterns of behavior without full-information assumptions.

---

Key References:

• "Iterative Local Voting for Collective Decision-making in Continuous Spaces" (Garg et al., 2017)
• "Convergence of Multi-Issue Iterative Voting under Uncertainty" (Kavner et al., 2023)
• "A Local-Dominance Theory of Voting Equilibria" (Meir et al., 2014)