Belief-Level Aggregation Methods

• Belief-Level Aggregation is a framework to combine individual probabilistic forecasts into a unified group belief, using cumulative distribution functions and threshold events.
• It introduces level-strategyproof conditions ensuring that strategic misreports by agents only push the aggregate belief further from their true values.
• Applications include incentive-compatible extensions to voting systems like majority judgment, with specific implementations such as the middlemost and proportional-cumulative aggregators.

Belief-level aggregation refers to the set of quantitative and qualitative methods for combining individual agents’ probabilistic beliefs or credences into a single group-level belief distribution or decision-making surrogate. This article surveys core definitions, representative aggregation rules, foundational axioms (and the associated impossibility results), operational characterizations, recent methodological advances (particularly level-strategyproof mechanisms), and highlights new incentive-compatible extensions to classic voting rules within uncertain environments.

1. Definitions and Formalization

Let $\Lambda$ denote a totally ordered set of outcomes (e.g., grades, states, policy alternatives), and let $\mathcal{P} = \Delta(\Lambda)$ be the space of all Borel probability measures over $\Lambda$ (the belief simplex). An $n$-agent belief profile is $(p_1,\dots, p_n) \in \mathcal{P}^n$, where $p_i$ encodes expert $i$'s probabilistic forecast. A probability aggregation function (PAF) is a mapping

$\psi: \mathcal{P}^n \to \mathcal{P}$

which determines the society’s or system’s aggregate belief.

For practical mechanism design, attention focuses on aggregation at individual “level-events,” $E(a) = \{x \in \Lambda : x \leq a\}$ (threshold events), and on the image of the probability measure at each threshold, regarded as the cumulative distribution function (CDF) $P(a) = p(E(a))$.

2. Level-Strategyproofness and Incentive Compatibility

The axiom of level-strategyproofness (Level-SP) is a focal point for aggregation under strategic reporting. Let $p = (p_1, p_{-i})$ be a belief profile, $p_i'$ any alternative report by agent $i$, and $a \in \Lambda$ any level. A PAF $\psi$ is level-SP if, for any $i$, $p$, $p_i'$ and $a$:

• If $p_i(E(a)) < \psi(p)(E(a))$, then $\psi(p)(E(a)) \leq \psi(p_{-i}, p_i')(E(a))$.
• If $p_i(E(a)) > \psi(p)(E(a))$, then $\psi(p)(E(a)) \geq \psi(p_{-i}, p_i')(E(a))$.

This is equivalent to: $$|\psi(p_{-i}, p_i')(E(a)) - p_i(E(a))| \geq |\psi(p)(E(a)) - p_i(E(a))|$$ which states that any strategic misreport at level $a$ can only bring the aggregate further from the agent’s true belief, not closer.

Moving to the CDF domain, every PAF induces a CDF-aggregation function $\Psi$ on $\mathcal{C}^n$ (the space of all CDFs over $\Lambda$). Level-SP translates precisely to uncompromisingness at every $a$: $$|\Psi(P_{-i}, P_i')(a) - P_i(a)| \geq |\Psi(P)(a) - P_i(a)|$$ for all $i$, $a$.

Level-SP implies incentive compatibility for any utility that is single-peaked in CDF-space (e.g., $u_i(p) = -\|\pi(p) - \pi(p_i)\|_{L_r(\Lambda)}$), a distinct domain from standard single-peakedness over distributions.

3. Structure, Characterization, and Impossibility Theorems

A general axiomatic characterization (Max–Min formula) of level-SP PAFs is available (Laraki et al., 2021). For every coalition $S \subseteq N = \{1, \dots, n\}$, there exists a phantom function $f_S:\Lambda \to [0,1]$ such that regularity, monotonicity, and boundary conditions are met:

• $S \subseteq T \implies f_S(a) \leq f_T(a)$,
• $f_S$ is weakly increasing and right-continuous,
• $\lim_{a \to \inf \Lambda} f_\emptyset(a) = 0$, $\lim_{a \to \sup \Lambda} f_N(a) = 1$.

The aggregation at level $a$ is: $$\Psi(P)(a) = \max_{S \subseteq N} \min\{ f_S(a), \min_{i \in S} P_i(a) \}$$ Unanimity holds if $f_\emptyset \equiv 0$, $f_N \equiv 1$.

In the anonymous case, $f_S$ depends only on $|S|$; thus, with phantoms $f_0 \leq f_1 \leq \cdots \leq f_n$: $$\Psi(P)(a) = \text{median}\left(P_1(a), \dots, P_n(a), f_0(a), \dots, f_n(a)\right)$$

Impossibility results show that strengthening level-SP (e.g., by requiring independence across outcomes, strong plausibility, or full diversity) forces the aggregation rule to be dictatorial unless very weak criteria are used. For $\Lambda$ an interval, level-SP + independence implies that only dictatorial rules are admissible.

4. Practically Useful Level-SP Methods

Two non-dictatorial, anonymous, incentive-compatible families arise:

4.1 Middlemost-Cumulative Aggregator

For $n$ agents, set $k_0 = \lceil (n+1)/2 \rceil$ and define

$$\Psi^\text{mid}(P)(a) := \text{median}(P_1(a), \dots, P_n(a))$$

That is, for each threshold, take the median belief across agents. In the anonymous median formula, set $f_0 = \cdots = f_{k_0-1} = 0$, $f_{k_0} = \cdots = f_n = 1$, so $$\Psi^\text{mid}(P)(a) = \text{median}(P_i(a), 0, \ldots, 0, 1, \ldots, 1)$$. This rule is certainty- and plausibility-preserving, resistant to manipulation, but does not mix beliefs in “dominated” profiles (it simply selects an agent’s view for the output).

4.2 Proportional-Cumulative Aggregator

With equal weights ($w_i=1/n$), the proportional-cumulative aggregator inserts uniform phantoms at $1/n, 2/n,\ldots,(n-1)/n$: $$\Psi^\text{prop}(P)(a) = \text{median}\left(P_1(a),\dots,P_n(a),\frac{1}{n},\ldots,\frac{n-1}{n}\right)$$ Equivalently: $$\Psi^\text{prop}(P)(a) = \sup \left\{y \in [0,1] \mid \frac{1}{n}\#\{i : P_i(a) \geq y\} \geq y \right\}$$ This rule mixes beliefs proportionally across agent ranks at each threshold. It maintains anonymity and certainty but not full plausibility; within intervals where no phantom lies between CDF values, “flattening” may occur.

5. Applications and Generalizations: Voting under Uncertainty

Proportional-cumulative aggregation provides principled, incentive-compatible extensions to various majoritarian voting procedures, especially where agents can submit probabilistic judgments on discrete grades or approvals (Laraki et al., 2021). For $\Lambda = \{0,1\}$ (approval voting), each expert reports $p_i(1)$, so

$$\Psi^\text{prop}(P)(1) = \frac{1}{n} \sum_{i=1}^n p_i$$

This reproduces classical majority voting under deterministic inputs and extends naturally to probabilistic approval.

On finite grade scales $\Lambda$, majority judgment methods (MJ) can be extended by aggregating the grade-CDFs via $\Psi^\text{prop}$ at each threshold, then applying lexicographic tie-breaking as usual. If agents supply only Dirac delta beliefs, the classical MJ/approval outcomes are recovered.

6. Example Computations

Consider $\Lambda = \{1,2,3\}$, $n=3$, agents with beliefs:

• $p_1 = \delta_1$
• $p_2 = \frac{1}{2} \delta_1 + \frac{1}{2} \delta_3$
• $p_3 = \delta_3$

Their CDFs $P_i(a)$ are:

• $P_1(a) = [1,1,1]$
• $P_2(a) = [0.5,0.5,1]$
• $P_3(a) = [0,0,1]$
• Middlemost rule: median at $a=1,2$ is $0.5$, at $a=3$ is $1$. Aggregated CDF is $[0.5, 0.5, 1]$.
• Proportional rule: insert $1/3$, $2/3$ as phantoms. At $a=1$, sorted values are $\{0, 1/3, 0.5, 2/3, 1\}$; median is $0.5$. Aggregated CDF matches the middlemost rule in this symmetric case, but in “dominated” profiles they differ.

7. Limitations and Normative Insights

Any substantial strengthening of diversity or plausibility along with Level-SP collapses to dictatorship. No fully strategyproof aggregator can simultaneously mix beliefs, maintain independence across outcomes, and respect every agent’s support. These limits are robust and are established formally via axiom interactions (Laraki et al., 2021).

The proportional-cumulative and middlemost-cumulative families are therefore the two principal non-dictatorial, anonymous, level-strategyproof choices. Proportional-cumulative aggregation, in particular, yields seamless and incentive-compatible extensions of approval voting, majority rule, and majority judgment to settings where experts/voters provide true probabilistic distributions over the alternatives.

• See "Level-strategyproof Belief Aggregation Mechanisms" (Laraki et al., 2021) for all formal statements, proofs, and toy computations.