Bayesian Polarization in Multidimensional States

• Bayesian polarization is the divergence of beliefs among rational agents caused by processing public signals over multidimensional state spaces.
• The phenomenon highlights that while joint posteriors converge onto an identified set, agents’ marginal beliefs can diverge under coordinatewise stochastic order.
• Persistent polarization hinges on signal structure conditions—spanning, balanced, and non-compensatory—that also translate into divergences in policy or action outcomes.

Bayesian polarization is the phenomenon in which belief updating by Bayesian agents—given public or shared signals about a multidimensional state—results in a divergence of beliefs (“polarization”) across the population. While classic impossibility theorems restrict polarization under first-order stochastic dominance in one-dimensional settings, recent advances reveal that the extension to multidimensional state spaces enables new forms and loci of polarization. Theoretical developments have characterized both possibility and impossibility results depending on the stochastic order employed and the structure of the signals agents observe, with significant implications for both epistemic modeling and observed political polarization.

1. Multidimensional Bayesian Belief Polarization

In one-dimensional settings, Bayesian updating with common signals cannot generate persistent belief divergence in the sense of first-order stochastic dominance. For multidimensional state spaces Θ = Θ₁ × Θ₂ × … × Θₖ with k ≥ 2, Bayesian agents’ beliefs are represented by joint probability distributions over Θ, with updates driven by incoming signals about the environment. The central insight is that, while agents’ joint posteriors must “collapse” onto the identified set (the collection of states consistent with observed signals), their marginal posteriors on each coordinate can diverge. Formally, define the coordinatewise stochastic order (Editor’s term) on the product space: Pˡ ⪯₍cw₎ Pᴴ if, for each coordinate i, the marginal posterior of agent L is stochastically dominated by that of agent H.

The main result establishes that coordinatewise polarization—Qˡ ⪯₍cw₎ Pˡ ⪯₍cw₎ Pᴴ ⪯₍cw₎ Qᴴ—can occur for all marginals, even after infinitely many signals, provided the identified set left after learning (e.g., Γ = {(min₁, ..., minₖ), (max₁, ..., maxₖ)}) is “diagonal” or otherwise appropriately structured. Each agent’s marginal beliefs can then diverge further (in the first-order stochastic dominance sense) from the other’s, despite observing the same data and applying Bayes’ rule. This stands in sharp contrast to the one-dimensional impossibility result previously established by Baliga et al. (2013).

2. Stochastic Dominance, Orders, and Limits of Polarization

First-order stochastic dominance (FOSD) between probability measures P and Q on a lattice (e.g., ℝ or ℝᵈ) is defined via E_P[u] ≤ E_Q[u] for all bounded, increasing (with respect to the usual partial order) measurable functions u. In one dimension, the impossibility of persistent polarization under FOSD arises from Bayesian posteriors’ averaging properties and information aggregation: both agents’ posteriors must move toward the same value as signals accumulate. In higher dimensions, the structure of dominance must be clarified.

Three stochastic orders are considered:

• Full stochastic order: Pˡ ≼₍st₎ Pᴴ requires Pˡ(U) ≤ Pᴴ(U) for all upper sets U ⊆ Θ (where U is closed upward in each coordinate).
• Upper orthant order: Pˡ ≼₍uo₎ Pᴴ requires the above only for U that are upper orthants, i.e., U = {θ : θ₁ > a₁, θ₂ > a₂, …, θₖ > aₖ} for some thresholds vector a.
• Coordinatewise order: Pˡ ≼₍cw₎ Pᴴ if and only if Pˡᵢ ≼₍st₎ Pᴴᵢ for marginals on each coordinate i.

The key impossibility result is that, even in the multidimensional case, persistent polarization is ruled out under the full stochastic order: the joint posteriors of Bayesian agents cannot diverge in this strong sense after observing common signals. However, coordinatewise polarization remains possible and generically persists in the infinite-signal limit under appropriate signal structures. There is also a “short-run” possibility for polarization in the intermediate (upper orthant) order; one-shot updates can result in divergence, but this cannot persist indefinitely due to posterior convergence on the identified set.

3. Signal Structure and Necessary/Sufficient Conditions for Persistent Polarization

Persistent polarization requires that, after any number of signals, the identified set Γ (the set of states with nonzero posterior mass) satisfies specific geometric criteria. In two-dimensional state spaces (Θ = {x₁, x₂} × {y₁, y₂}), polarization for all marginals is possible if and only if Γ (and its complement) is:

• Spanning: Attains min and max in each coordinate.
• Balanced: Does not lie below or above any threshold rectangle; i.e., does not include (exclude) all states strictly below (above) some threshold.
• Non-compensatory: Does not permit compensation of low values in one coordinate for high values in another; i.e., not only the cross-diagonal set, but the “diagonal” set.

For example, in Θ = {x₁, x₂} × {y₁, y₂}, the set {(x₁, y₁), (x₂, y₂)} admits strong coordinatewise polarization, as the marginal posteriors can diverge simultaneously on each axis. These conditions (precisely characterized in the paper’s theorem for 2D) ensure that the marginal posteriors for high and low types move apart along both axes, even after an infinite sequence of signals.

If Γ fails any of these conditions (e.g., is not spanning or is compensatory), coordinatewise polarization is precluded in the limit, and agents' beliefs coalesce.

4. Short-Run Polarization and Intermediate Stochastic Orders

Transience in polarization can occur in the short run when employing the intermediate upper orthant order. A single public signal that is not fully revealing can induce one-shot polarization (i.e., Qˡ ≼₍uo₎ Pˡ ≼₍uo₎ Pᴴ ≼₍uo₎ Qᴴ), but as the number of signals increases and the identified set shrinks, joint convergence supersedes this effect, and limit polarization according to the upper orthant order is impossible. This resolves the tension between observed short-run divergence and theoretical long-run convergence under strong stochastic dominance.

5. Implications for Polarization in Actions

Translating belief polarization into action polarization depends on the agents’ utility structure. If payoffs are additively separable across policy dimensions, then coordinatewise polarization in beliefs generates action polarization: agents' optimal decisions (e.g., voting for a candidate or supporting a policy) can diverge more as marginal beliefs pull apart. If payoffs depend on joint events or payoff functions with strong complementarities, the impossibility of full multidimensional stochastic order polarization limits the divergence in joint actions. Thus, the signal structure, the geometry of the identified set, and the agents' payoff functions all interact to determine whether observed actions will exhibit polarization.

A practical implication is that “balanced” or “partial” information signals—those that are non-compensatory and spanning—can themselves increase issue-by-issue polarization when priors differ, even as overall joint event beliefs converge with more information. This insight is relevant for interpreting polarization in empirical political economy where public signals (e.g., media reports) are rarely completely revealing.

6. Theoretical and Observational Significance

The main contributions of recent research on Bayesian polarization in multidimensional spaces are:

• Extending the classic impossibility theorems on stochastic dominance polarization, originally formulated for unidimensional beliefs, to the settings where each agent holds beliefs over multiple interrelated issues.
• Providing a precise geometric characterization of informational conditions (on the identified set and signal structure) that enable or preclude persistent coordinatewise polarization.
• Introducing the role of stochastic order hierarchy (full, upper orthant, coordinatewise) for stratifying types and duration of possible polarization.
• Connecting the theoretical distinctions to practical consequences in action polarization, especially when utility is separable.

Empirically, this framework rationalizes how sophisticated agents, observing the same public evidence, can persistently diverge on specific dimensions of complex issues—mirroring observed fractures in multidimensional political space—while still converging in joint beliefs in the limit.

7. Mathematical Formulation and Key Results

Let Θ = Θ₁ × Θ₂ × ... × Θ_d.

• Agents L and H have initial priors Pˡ, Pᴴ with Pˡ ⪯₍cw₎ Pᴴ.
• After observing a public signal s, their posteriors are Qˡ, Qᴴ.
• Coordinatewise polarization: Qˡ ⪯₍cw₎ Pˡ ⪯₍cw₎ Pᴴ ⪯₍cw₎ Qᴴ, i.e., for all coordinates i, Qˡᵢ ≼₍st₎ Pˡᵢ ≼₍st₎ Pᴴᵢ ≼₍st₎ Qᴴᵢ.

Necessary and sufficient conditions for polarization in two dimensions are:

• Γ non-compensatory, balanced, and spanning (as defined above).

For full stochastic order, polarization is impossible in both one-shot and limit, as the lattice structure prevents divergence on all upper sets. For coordinatewise (and one-shot upper orthant), polarization is possible with the proper structure of the identified set and signals.

---

Bayesian polarization, in the modern theoretical sense, refers to the phenomenon whereby multidimensionality in the state space and signal structure create scope for persistent divergence in marginal beliefs—even among rational agents observing the same public information and applying Bayes’ rule—subject to geometric and informational conditions on the signal structure, and with sharp limits imposed by global stochastic dominance. This framework provides both formal clarification of the mechanisms underlying empirical polarization and a vocabulary for distinguishing among subtler types of belief and action divergence.