Universal Closest Refinement on Measurable Bipartite Relations

Abstract: We study the universal closest refinement problem on measurable bipartite relations over standard Borel spaces. Given prescribed side measures, the feasible class consists of finite refinement plans concentrated on the relation and carrying one fixed marginal. The main question is whether this highly nonunique class nevertheless contains a mathematically distinguished class of refinements. We show that the correct one-sided extremal criterion is level-optimal maximin, a levelwise maximin principle formulated through truncation and overflow profiles. We then prove that this structure is exactly the one selected by convex refinement: every minimizer of a strictly convex refinement criterion is level-optimal maximin, while every level-optimal maximin refinement minimizes the full class of relevant proper lower semicontinuous convex divergence functionals. Proportional response then identifies the opposite-side partner and yields a universally closest refinement pair. Our main theorem shows that every such pair is universally closest among all feasible pairs for every divergence satisfying the data-processing inequality under measurable post-processing, and conversely that every universally closest pair has this structure. We also prove a converse paired characterization for strictly convex closest pairs. Finally, we give an equilibrium-theoretic characterization of level-optimal maximin pairs through a naturally associated continuum economy with measure-valued commodities. The measure-theoretic theory requires new tools beyond the finite case, including disintegration, measurable selection, measurable max-flow/min-cut duality, augmentation arguments, and a symmetric density decomposition separating absolutely continuous and singular components.

• The paper establishes a level-optimal maximin framework that uniquely characterizes divergence-minimizing bipartite refinements.
• It utilizes measure-theoretic tools such as disintegration and symmetric density decompositions to extend finite-space techniques to continuous settings.
• The study reveals a duality between convex optimization and Walrasian equilibrium, providing a unified theory for optimal allocation in bipartite systems.

Universal Closest Refinement on Measurable Bipartite Relations: Theory, Characterizations, and Equilibrium

Overview and Problem Setting

This paper develops a comprehensive measure-theoretic theory for the universal closest refinement problem on measurable bipartite relations between standard Borel spaces. Given base measures on two standard Borel spaces and a measurable relation between them, one studies finite refinement plans concentrated on the relation and satisfying a fixed marginal constraint. The central question is the structural selection problem: does a mathematically distinguished element or class of elements exist within the large, highly nonunique feasible set of refinements?

The work builds on prior finite-space theory [CX25] but crucially addresses qualitatively new challenges arising in general measure spaces, including the necessity of tools such as disintegration, measurable selection, and a careful separation of absolutely continuous and singular components in payload marginal decompositions. The extension is motivated both by intrinsic mathematical interest but also by applications in areas such as privacy composition, where the relevant probability measures and divergence functionals are typically not discrete [ZSCM23, ZZCS24, DRS22].

One-Sided Extremal Criterion: Level-Optimal Maximin

The paper demonstrates that the correct structural invariant is a one-sided criterion termed level-optimal maximin. Instead of classical pointwise or local density-based criteria, the analysis proceeds by studying, for each truncation level $t$, the maximal admissible mass that can be packed below this level in the opposite-side marginal; overflow profiles (amount exceeding $t$) encode the relevant extremality properties. The level-optimal maximin condition asserts that for each $t > 0$, the truncated payload matches the supremum attainable among all feasible refinements constrained by density cap $t$. This replaces combinatorial decomposition arguments from the finite case with a functional-analytic, levelwise extremal principle, reflecting the richer behavior of continuous spaces.

Universal Closest Pairs and Convex Refinement

The main theorems establish a duality between this level-optimal maximin structure and convex variational problems:

• Minimizers of strictly convex refinement criteria (specifically, divergence minimization over feasible marginals) are precisely the level-optimal maximin refinements.
• Conversely, any level-optimal maximin refinement is simultaneously optimal for every proper convex lower semicontinuous divergence functional over the feasible class.

This covers a broad class of divergences (including those satisfying the data-processing inequality, channel-DPI), and, strongly, a single refinement simultaneously minimizes all such divergences, not just a specific one. The paper leverages a reduction to "hockey-stick" divergences (mixtures of elementary truncation functionals) to show that levelwise optimality suffices for universal optimality over all channel-DPI divergences.

Proportional response construction enables the canonical pairing of one-sided refinements: given a level-optimal maximin refinement on one side, a reciprocal partner on the other side can be uniquely constructed via reverse conditional structure. The resulting pair is called a universally closest refinement pair—they are simultaneously optimal for all convex divergences in the class.

Main Numerical/Structural Claim: The set of universally closest refinement pairs is exactly parametrized by pairs of mutual proportional responses between level-optimal maximin refinements. No additional structures outside this class can achieve universal optimality for the full class of divergences satisfying channel-DPI.

Measure-Theoretic Tools and Existence

Passage to the measure-theoretic setting necessitates an array of technical tools:

• Disintegration of measures on standard Borel spaces for the construction of conditional plans.
• Measurable versions of max-flow/min-cut duality for feasible mass transport and cut-based capacity analysis.
• Augmentation arguments (generalizing finite augmenting-path techniques) to support variational principles and strict convexity gaps.
• Symmetric density decompositions separating absolutely continuous and singular parts, foundational for both the analysis and the economic equilibrium interpretation.

Existence results for optimal refinement plans are obtained only under significant regularity, including closure of the relation, reflecting genuine analytic difficulties absent in the finite case. Examples in the Appendix illustrate failure of existence in the absence of closedness.

Equilibrium Characterization and Continuum Economies

A key contribution is an equilibrium-theoretic characterization: the class of level-optimal maximin pairs corresponds exactly to Walrasian equilibria in a continuum economy naturally induced by the bipartite relation and measures.

In this framework:

• Agents correspond to the two sides of the bipartite relation.
• Commodities are finite measures (allowing for both trade and retention).
• The refinement plans are interpreted as allocations; the relation encodes feasibility constraints on bilateral trade.
• The equilibrium prices encode, via a specific transform of marginal densities, the optimal "exchange rates" on measure-valued commodities for each agent.

The analysis provides both directions:

1. Every level-optimal maximin pair gives rise to a Walrasian equilibrium in the associated economy.
2. Conversely, every Walrasian equilibrium can be transformed, via a canonical extraction procedure, into a level-optimal maximin refinement pair.

Importantly, mutual proportional response is not necessary for equilibrium, further distinguishing the pairwise optimality structure for divergence minimization from the (weaker) structural requirements of economic equilibrium.

Implications and Prospects

The results establish a unified measure-theoretic theory of feasible refinements in bipartite structures with prescribed marginals, capturing both the analytic core (level-optimal maximin) and dual/convex-analytic structure. The universal closestness result is particularly strong: optimal divergence-minimizing couplings are structurally unique up to their absolutely continuous part and valid for all convex divergences from a broad class. This provides a robust and explicit selection principle in a setting where naive local or pointwise optimality notions fail, and aligns with structural results in comparison of experiments theory and information processing.

Possible future developments include:

• Extension to infinite or nonstandard spaces (failure, for example, of separability or closeness conditions would require new techniques).
• Application to areas such as privacy and information theory, particularly where universal optimality across divergence families is needed.
• Investigation of computational analogues and tractable algorithms for large, structured measurable relations.
• Deeper study of uniqueness and rigidity phenomena, especially for singular components and the role of equilibrium selection.

Conclusion

This work provides a rigorous and complete account of the universal closest refinement problem for measurable bipartite relations. By isolating level-optimal maximin as the governing structural invariant, relating it to convex optimization, and establishing its equivalence with economic equilibrium formulations, the paper creates a robust theoretical foundation for refinement planning in general measure spaces. These results both generalize and transcend their finite counterparts, forming a core reference point for subsequent analysis of measure-theoretic transport, information structure comparison, and economic allocation under bipartite constraints.