When indifference between identical bets rules out correlation effects

Consider three equally likely cards. Bet C pays 0, 50 and 100 on the respective cards; bet L pays 50, 100 and 0. Each bet offers the same distribution of prizes. But L wins the head-to-head comparison on two cards, while C wins by a larger amount on the remaining card.

Within the paper’s framework, strictly preferring either bet reveals sensitivity to how the two alternatives’ outcomes are paired, rather than to their separate prize distributions. That distinction is the starting point of Bang-Bang Correlation Sensitivity: A Local-to-Global Dichotomy for Correlation Preference, by Yi-Chun Chen, Soo Hong Chew and Xinhan Zhang. The November 3, 2026 arXiv preprint, version 2611.05188v1, presents theoretical results and a proposed experimental test, not new experimental findings.

The demanding meaning of indifference

The central result concerns a finite set of possible outcomes. A lottery has “full support” when every outcome in that set has positive probability. A coupling specifies how the outcomes of two lotteries occur together while preserving each lottery’s own probabilities.

The authors show that, under three axioms, indifference between two copies of just one full-support lottery under every possible coupling rules out correlation sensitivity at every full-support lottery over the same outcome set. Conversely, if correlation sensitivity exists at one such lottery, every full-support lottery permits a nonindifferent comparison between copies with identical distributions.

The word “every” matters. Indifference in the three-card example alone is not generally enough to establish the theorem’s premise. Nor does the result say that a correlation-sensitive person must prefer one alternative in every comparison. It says that sensitivity cannot disappear entirely at one full-support lottery while remaining elsewhere in the interior domain.

Three assumptions do the work

Continuity requires weak preference to survive limits of joint distributions. Completeness requires each binary comparison to be resolved in at least one direction, allowing indifference. Betweenness imposes restrictions on mixtures of comparisons that keep the first alternative’s lottery fixed: mixtures must preserve specified patterns of acceptance, rejection and indifference.

These are substantive restrictions, not findings about everyone’s behavior. The theorem does not require transitivity—the familiar condition that preference for A over B and B over C implies preference for A over C. It also does not require the stronger independence axiom used in the paper’s benchmark correlation expected utility model.

The authors’ proof first constructs local linear evaluations, then uses optimal-transport duality and consistency across linked comparisons to connect those evaluations. The resulting restriction is why they describe correlation sensitivity as “bang-bang”: under the assumptions, the interior domain cannot contain both sensitive and insensitive regions.

Separating pairing effects from lottery effects

In the sensitive branch, the three axioms characterize what the authors call correlation weighted utility, including comparisons involving lotteries without full support. Its evaluation combines two components: a state-by-state comparison term that can respond to changes in pairing, and a bilinear term depending only on the marginal lotteries.

This separates correlation effects from departures from classical independence. Changing the coupling while holding both prize distributions fixed changes only the first component.

The insensitive branch requires an additional axiom, projective independence, to obtain Fishburn’s skew-symmetric bilinear representation. Strong independence instead reduces correlation weighted utility to Lanzani’s correlation expected utility. Combining that stronger independence with correlation insensitivity yields expected utility. Correlation insensitivity alone therefore does not establish the classical expected-utility model.

An experimental connection, not a completed test

The paper proposes giving the same participants both identical-distribution choices and HILO comparisons, which mix alternatives with common consequences, on a shared three-outcome support.

The direction of the identical-distribution preference predicts how two diagnostic HILO choices respond as dependence moves from maximal positive dependence toward independence. Favoring C predicts movement toward the choices associated with standard fanning-out behavior; favoring L predicts the reverse. These probability predictions additionally assume a nondecreasing response rule that remains stable across coupling treatments.

The proposed design could distinguish pairing effects from marginal effects that remain under maximal positive dependence. Its implementation must avoid changing visual event-splitting alongside the coupling, which would confound a pure-correlation comparison.

The scope remains important. The broader outcome-space extension retains finitely supported lotteries, and insensitivity on one support need not survive an enlarged support. The authors also construct a copula-based model without imposed betweenness in which local insensitivity coexists with sensitivity elsewhere. Their message is consequently conditional but sharp: under the stated axioms, complete local indifference carries unusually strong implications for choices beyond that lottery.