Conversations on Contextuality

In the form of a dialogue (imitating in style Lakatos’s Proof and Refutation), this chapter presents and explains the main points of the approach to contextuality dubbed Contextuality-by-Default.

💡 Research Summary

**
The paper presents the Contextuality‑by‑Default (CbD) theory through a dialogic format reminiscent of Lakatos’s “Proof and Refutation.” Two characters—an Expositor who explains the theory and an Interlocutor who questions and probes the concepts—guide the reader through the foundational ideas, technical definitions, and implications of CbD.

The first part establishes the basic ontology of CbD: every measurement is identified by a pair of labels, an “object” (what is measured) and a “context” (the conditions under which it is measured). Consequently, the same object measured in different contexts is represented by distinct random variables. For example, two yes/no questions (A and B) can be presented either verbally (V) or in writing (W). This yields four variables: R_VA, R_WA, R_VB, and R_WB. Variables recorded within the same context (e.g., R_VA and R_VB) have a joint distribution, whereas variables from mutually exclusive contexts (e.g., R_VA and R_WA) are stochastically unrelated—they have no joint distribution and therefore no meaningful probability of being simultaneously true.

The second segment introduces the notion of a coupling. For a set of stochastically unrelated variables, a coupling is a new set of jointly distributed random variables (denoted with a tilde, e.g., \tilde R_VA) that preserve the original marginal distributions. All possible couplings are considered; the set is typically infinite except in degenerate cases where marginal probabilities are 0 or 1. Among these couplings, a maximal coupling maximizes the probability that two variables representing the same object in different contexts are equal. For binary variables, the maximal equality probability is 1 − |p_V − p_W|, where p_V and p_W are the marginal probabilities of “yes” in the two contexts. For continuous variables, any joint distribution with the prescribed marginals qualifies as a coupling, and maximal equality may be less than one if the marginals differ.

The third part explains how CbD defines (non)contextuality. For each object, one computes the maximal equality probability between its contextual versions. Then one asks whether there exists at least one global coupling of all variables such that, simultaneously for every object, the equality events achieve their maximal probabilities. If such a global coupling exists, the system is non‑contextual; if not, the system is contextual. This criterion reduces contextuality to a feasibility problem: can the collection of maximal pairwise equality constraints be satisfied jointly?

The dialogue also explores variations in experimental design. If the survey is altered so that each participant receives only one question, then each variable belongs to a distinct context, and all variables become pairwise stochastically unrelated. Nevertheless, the labeling of objects and contexts remains essential, and the CbD rules still apply. The conversation further clarifies that stochastic unrelatedness does not imply statistical independence; independence is a property of a specific coupling, not of the original variables.

Finally, the authors contrast CbD with traditional contextuality analyses in quantum physics, such as Bell‑type inequalities or the Kochen‑Specker theorem. Traditional approaches often assume the existence of a joint distribution for all measurements and derive contradictions. CbD, by contrast, starts from the empirically observable marginal distributions, acknowledges that many variables are inherently non‑jointly distributed, and systematically examines all admissible couplings. This makes CbD applicable not only to quantum experiments but also to behavioral and social‑science data where contextual influences are pervasive.

In summary, CbD provides a rigorous, probabilistic framework consisting of four steps: (1) label each measurement by object and context; (2) treat variables from different contexts as stochastically unrelated; (3) construct all possible couplings and identify maximal pairwise equality probabilities; (4) test the compatibility of these maximal constraints across the whole system. The existence or absence of a compatible global coupling determines non‑contextuality or contextuality, respectively. This methodology offers a unified way to assess contextual effects across diverse scientific domains.