Irrelevant and independent natural extension for sets of desirable gambles

The results in this paper add useful tools to the theory of sets of desirable gambles, a growing toolbox for reasoning with partial probability assessments. We investigate how to combine a number of marginal coherent sets of desirable gambles into a joint set using the properties of epistemic irrelevance and independence. We provide formulas for the smallest such joint, called their independent natural extension, and study its main properties. The independent natural extension of maximal coherent sets of desirable gambles allows us to define the strong product of sets of desirable gambles. Finally, we explore an easy way to generalise these results to also apply for the conditional versions of epistemic irrelevance and independence. Having such a set of tools that are easily implemented in computer programs is clearly beneficial to fields, like AI, with a clear interest in coherent reasoning under uncertainty using general and robust uncertainty models that require no full specification.

💡 Research Summary

The paper makes a substantial contribution to the theory of sets of desirable gambles, a flexible framework for reasoning under partial probability assessments. It begins by highlighting a gap in the existing literature: while the theory provides tools for handling a single marginal coherent set of desirable gambles, it lacks a systematic method for combining several marginal sets into a joint model that respects epistemic relationships. To fill this gap, the authors introduce two key concepts—epistemic irrelevance and epistemic independence—and use them as guiding principles for constructing joint sets.

Epistemic irrelevance is defined as the condition that learning the outcome of one variable does not alter the desirability judgments about gambles on another variable. In the language of desirable gambles, this translates into a preservation condition: the conditional set of desirable gambles for a variable, given any event concerning the other variable, must be contained in the original marginal set. Epistemic independence is the symmetric strengthening of this condition, requiring mutual irrelevance.

Building on these notions, the authors define the “independent natural extension” (INE). Given n marginal coherent sets D₁,…,Dₙ, the INE is the smallest coherent set that contains every gamble whose restriction to each marginal coincides with a gamble in the corresponding conditional set. In other words, it is the minimal joint extension that simultaneously satisfies the coherence axioms and the independence constraints. The paper provides rigorous proofs that the INE enjoys three fundamental properties: (i) coherence (no sure loss), (ii) preservation of the original marginals (marginal consistency), and (iii) minimality (no proper subset can satisfy both coherence and independence).

A special case is examined when each marginal is a maximal coherent set. In this situation the INE can be expressed as the “strong product” of the marginals. The strong product is larger than the ordinary product of sets because it enforces independence explicitly, yet it remains the greatest set that preserves maximality of each component. This construction yields a new algebraic operation on maximal sets, enriching the toolbox for robust uncertainty modeling.

The authors do not stop at unconditional relationships. They extend the framework to conditional epistemic irrelevance and independence, showing that the same construction works when the irrelevance statements are conditioned on an event A. The conditional INE retains the same coherence, marginal consistency, and minimality properties, which makes the approach suitable for dynamic or sequential decision problems where information arrives over time.

From a computational perspective, the paper demonstrates that the INE can be formulated as a linear programming problem. The constraints that define the extension are linear inequalities derived from the marginal desirability conditions. Consequently, existing LP solvers can compute the INE efficiently when the marginals are themselves tractable (e.g., described by a polynomial number of linear constraints). This algorithmic tractability is emphasized as a major advantage for applications in artificial intelligence, where large‑scale uncertain models must be processed quickly.

In the discussion and conclusion, the authors outline several avenues for future work. These include extending the theory to continuous outcome spaces, integrating the INE with other imprecise probability frameworks such as possibility theory or belief functions, and applying the strong product to multi‑stage decision problems like partially observable Markov decision processes. Overall, the paper delivers a mathematically rigorous yet practically implementable method for constructing joint coherent models from marginal assessments, thereby broadening the applicability of sets of desirable gambles to a wide range of AI and decision‑theoretic contexts.