TRANSMISSION // 2609.13696v2

How information pools across multidimensional states

Source published

A mathematical study characterizes extreme distributions of posterior means in Bayesian persuasion, showing how information can pool within geometric regions—and why boundary effects matter.

AI-generated conceptual illustration for How information pools across multidimensional states
AI-generated conceptual illustration; not documentary evidence.

When a sender chooses what information to share, the receiver’s beliefs can be summarized by their average state, or posterior mean. Xin Wang’s theoretical preprint studies which distributions of these means are feasible and what special structure appears at the extreme points of that set. The answer depends on how multidimensional regions fit together, especially at their boundaries.

In Bayesian persuasion, a sender chooses a signal to influence a receiver’s beliefs and, ultimately, the receiver’s decisions. This paper studies a particular class of problems: the sender’s payoff depends on each posterior belief only through its mean. The state may have several dimensions—for example, a product may have both a value to a customer and a commission for a platform.

The central question is geometric: among all distributions of posterior means that a signal can generate, what do the extreme ones look like? An extreme point is a distribution that cannot be written as a nontrivial average of two different feasible distributions. Finding their structure can narrow the candidates relevant to an optimization problem. It does not, by itself, identify the best signal for every payoff.

From signals to a geometric problem

The paper uses a standard result from martingale transport theory. A distribution of posterior means is feasible exactly when it is dominated by the prior in convex order. Informally, the posterior means preserve the prior’s average but are less spread out in the sense that every convex function has a no-larger expectation under the posterior-mean distribution.

This turns the choice of signal into an optimization over probability distributions subject to a mathematical constraint. The paper then studies how feasible distributions are organized into convex regions, called components. Within a component, the posterior means can be related to the states through a martingale coupling: conditional on a posterior mean, the expected state is that mean.

The analysis draws on the component decomposition developed in martingale transport theory. Wang establishes a geometric characterization of convex-order feasibility under a condition that the prior assigns zero probability to the relative boundaries of the non-singleton components. Under that condition, convex-order dominance can be understood locally, component by component, rather than only across the whole state space.

What the extreme-point result says

With boundary mass ruled out, the paper gives a necessary-and-sufficient characterization of extreme feasible distributions. In almost every component, the conditional support of posterior means must be affinely independent: none of its points can be formed as an affine combination of the others. In a component of dimension d, that limits the conditional support to at most d + 1 points. A singleton component corresponds to revealing the state at that point.

The word “conditional” matters. Each individual component can have a small support even when there are infinitely many components overall. In multiple dimensions, the paper shows that there can be uncountably many non-singleton components of lower dimension. So a local bound on posterior means does not imply that the overall signal has only finitely many possible posterior means.

The paper illustrates this with a product-recommendation example. A platform observes two quantities, product value and commission, each uniformly distributed on the unit interval. The customer buys when the expected value is high enough relative to their cost, and the platform’s expected payoff depends on the product of the two posterior means. The paper shows that revealing the sum of the two quantities is optimal in this example, and that the optimal distribution of posterior means is unique. It is nevertheless nonatomic: it places mass along a continuum of points on the diagonal. Each of an uncountable collection of pooling regions contains just one posterior mean.

Why boundaries change the answer

The clean characterization relies on a specific regularity condition: the union of the relative boundaries of non-singleton components must have zero prior probability. When that condition fails, boundary points can connect the closures of different components. A feasible transport may then redistribute mass across regions in ways that local analysis does not capture.

In that case, affine independence within each component remains necessary for extremality, but it is no longer sufficient. The paper supplies an example in which each component meets the local condition but the overall distribution is not extreme. The limitation is therefore substantive, not just technical: boundary interactions can create additional ways to express a distribution as a mixture.

Scope and uncertainty

These are mathematical results, not an empirical test of how real senders or receivers behave. The paper’s main theorem characterizes extreme points under its boundary condition; without that condition, it provides necessary conditions rather than a complete characterization. And even when an optimum can be shown to occur at an extreme point, the geometric result alone does not say which extreme point maximizes a particular payoff. The paper’s product example establishes its own optimum with a separate payoff argument.

The source is an arXiv preprint by Xin Wang. A preprint is a public research manuscript and may not have undergone peer review. The results should therefore be read as the author’s theoretical contribution, not as settled empirical evidence. The paper also identifies open questions, including when economically relevant model assumptions guarantee that the boundary condition holds.

Reviewed cross-references

These connections were reviewed by a model, not independently fact-checked by a human. Similarity does not imply agreement; consult the original papers and the evidence below.

  • How much persuasion survives when information comes in separate parts?

    Shared method: Both papers use posterior-mean distributions and mean-preserving contractions to represent feasible signaling: Wang studies convex-order contractions of a multidimensional prior, while Jia, Wang, and Xu describe persuasion by parts as a sum of independent contractions. The feasible sets differ—the latter imposes product experiments across independent coordinates—so this is a methodological connection, not a shared result about optimality.

    Model review: · openai-codex/gpt-6-luna

    Evidence and provenance

    Selected excerpts, not complete papers. PDF extraction can omit figures or damage equations, tables and reading order; sources may be preprints.

    • 2609.13696v2 — Original paper

      Excerpt reference: static/papers/2609.13696v2.md:69-190
      Basis: Paper text
      Source SHA-256: 7afc56dae162d7c162e855415e2a6a8c610aa28d7eb4374746c2b9046fd08874

    • 2611.02845v1 — Original paper

      Excerpt reference: static/papers/2611.02845v1.md:693-749
      Basis: Paper text
      Source SHA-256: 68d2a2d19b6c44849b0e373d0cb8203d216116ef40436014c5298ecd7f012a59

Automatically matched by shared topic keywords. These links do not imply that the sources agree or establish a consensus.