On the stability of utilitarian aggregation

1. Problem Statement

The paper addresses the robustness of Harsanyi's (1955) Aggregation Theorem. Harsanyi's theorem states that if a collective utility function ($u_0$) satisfies the Pareto unanimity principle (if all individuals prefer $x$ to $y$, the group must prefer $x$ to $y$) and the Pareto indifference principle (if all are indifferent, the group is indifferent), then $u_0$ must be a linear weighted sum of individual von Neumann-Morgenstern (vN-M) utilities:
$$u_0(x) = \sum_{i=1}^N a_i u_i(x) + b$$
where $a_i \geq 0$.

The Core Issue: In real-world applications (e.g., expert panels, social choice with incomplete information), preferences are often approximations, and the strict Pareto conditions may be violated slightly due to measurement error, uncertainty, or misspecification. The paper asks: If the Pareto conditions are only approximately satisfied (bounded violations), does the aggregation rule remain approximately utilitarian?

2. Methodology and Framework

Setting

• Domain: The set of alternatives $X$ is a convex subset of a real vector space.
• Utilities: The paper assumes vN-M utilities, meaning $u_0$ and $u_i$ are affine mappings (preserving mixtures: $u(\lambda x + (1-\lambda)y) = \lambda u(x) + (1-\lambda)u(y)$).
• Agents: A Decision Maker (DM) with utility $u_0$ and $N$ individuals with utilities $u_1, \dots, u_N$.

Approximate Pareto Axioms

The author introduces $\epsilon$-Pareto conditions, where $\epsilon \geq 0$ quantifies the degree of violation.

• $\epsilon$-Semistrong Pareto: If $u_i(x) \geq u_i(y)$ for all $i$, then $u_0(x) \geq u_0(y) - \epsilon$.
  • Interpretation: If everyone prefers $x$, the DM considers $x$ "near-optimal" relative to $y$, allowing for a small utility gap of $\epsilon$.
• $\epsilon$-Pareto Indifference: If $u_i(x) = u_i(y)$ for all $i$, then $|u_0(x) - u_0(y)| \leq \epsilon$.
• Sequential $\epsilon$-Strong Pareto: A stronger condition involving limits of sequences to ensure strictly positive weights in the aggregation.

Approximate Utilitarian Aggregation

The paper defines $u_0$ as approximately utilitarian if it is close to a linear aggregator $w(x) = \sum a_i u_i(x) + b$. Two equivalent characterizations are provided:

1. Oscillation Bound: $u_0(x) = \sum a_i u_i(x) + e(x)$, where the oscillation of the error term $e$ is bounded by $\epsilon$ ($\omega_e(X) \leq \epsilon$).
2. Supremum Norm Bound: $u_0$ is within distance $\epsilon/2$ of a utilitarian aggregator in the supremum norm: $\|u_0 - w\|_\infty \leq \epsilon/2$.

3. Key Contributions and Results

Main Theorems

The paper establishes equivalence theorems linking the violation of Pareto axioms to the distance from a utilitarian aggregator.

• Theorem 1 (Duality Formula): Under continuous vN-M utilities and a compact domain, there exists a duality between the maximum violation of the Semistrong Pareto condition and the minimum distance to the set of utilitarian aggregators.
  $$\min_{w \in W^+} \|u_0 - w\|_\infty = \frac{1}{2} \max \{ u_0(y) - u_0(x) : (x,y) \in V \}$$
  where $V$ is the set of pairs where individuals unanimously prefer $x$ but the DM prefers $y$ (or is indifferent). This proves that the "error" in aggregation is exactly half the maximum intensity of the Pareto violation.

• Theorem 2 & 3 (General Equivalence):

  • Theorem 2: In compact settings, $\epsilon$-Semistrong Pareto is equivalent to $u_0$ being $\epsilon$-close to a utilitarian aggregator with non-negative weights ($W^+$).
  • Theorem 3: This result is extended to general vN-M settings (without requiring compactness) using an extension argument and limiting processes. It confirms that small violations of Pareto principles lead to aggregation rules that are close to linear.

• Theorem 4 (Indifference): $\epsilon$-Pareto Indifference is equivalent to $u_0$ being approximately utilitarian with weights in $W$ (allowing negative weights, though the context usually implies non-negative).

• Theorem 5 & 6 (Positive Weights): To ensure strictly positive weights ($W^{++}$), the paper introduces Sequential $\epsilon$-Strong Pareto.

  • Theorem 5: Sequential $\epsilon$-Strong Pareto is equivalent to $u_0$ being approximately utilitarian with strictly positive weights.
  • Theorem 6: In settings where $X$ is the convex hull of finitely many points (e.g., finite lotteries), the standard $\epsilon$-Strong Pareto is sufficient (the sequential condition is not needed).

Applications to Belief Aggregation

The paper applies these results to the Anscombe-Aumann framework (Subjective Expected Utility):

• Proposition 2: If individuals have similar tastes (bounded by $\epsilon_1$) and the DM satisfies $\epsilon_2$-Likelihood Dominance, then:
  1. The DM's taste function is approximately a weighted sum of individual tastes.
  2. The DM's probability measure is approximately a linear combination of individual priors ($P_0 \approx \sum \lambda_i P_i$).
• Significance: This relaxes Mongin's (1998) result, showing that a "probability dictator" is not strictly necessary if tastes are sufficiently similar and Pareto conditions are approximate.

4. Technical Significance

1. Stability of Harsanyi's Theorem: The paper provides a quantitative stability result (answering Ulam's question for this context). It proves that the class of utilitarian aggregation rules is stable under small perturbations of the Pareto axioms. The distance from the ideal linear form is linearly bounded by the magnitude of the axiom violations.
2. Duality and Minimax: The proofs rely heavily on the Minimax Theorem (Sion/Kneser) and duality in convex analysis. The author frames the Pareto conditions as efficiency problems in the space of utility difference profiles.
3. Handling Misspecification: The framework offers a decision-theoretic justification for "satisficing" behavior. If individual preferences are uncertain or misspecified, the DM can adopt an approximately utilitarian rule without inheriting the specific biases of the misspecified inputs (unlike strict Pareto aggregation, which would force the DM to adopt the bias).
4. Extension to Infinite Dimensions: While the main theorems handle finite agents, the discussion suggests how best approximation methods in normed spaces could extend these results to infinite societies (Zhou, 1997), provided compactness assumptions are met.

5. Conclusion

Leandro Nascimento's paper rigorously demonstrates that Harsanyi's aggregation theorem is robust. Even when the strict Pareto unanimity and indifference principles are violated by a small parameter $\epsilon$, the resulting collective utility function remains close to a weighted sum of individual utilities. The deviation is bounded by $\epsilon/2$ in the supremum norm. This provides a theoretical foundation for using utilitarian aggregation in practical settings where preferences are estimated, noisy, or only partially known.