Recognizing Conceptual Units from Stable Compression in Modular Proof Worlds

Read full text Summarise with VIRAT
Authors: Hoang Duy Nguyen
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper develops a foundational theory for recognizing conceptual units in modular proof environments. Its core claim is that true conceptual units are not merely recurrent fragments, short descriptions, or locally useful lemmas. Instead, they are reusable proof-theoretic blocks whose identity remains stable under compression, structural closure, and recomposition across new proof contexts. The central problem is one of identifiability: when does a near-optimal compressed library recover the same conceptual structure as the hidden true module basis of the domain? The paper argues that pure compression is insufficient because three major pathologies can imitate good structure: over-fragmentation, where a true module is split into smaller recurring pieces; over-coarsening, where several true modules are merged into a large block that fits training data but fails under new recombinations; and off-axis substitution, where a different but superficially competitive basis replaces the true conceptual axes. To address this, the paper builds a layered framework. It begins with modular proof worlds, where hidden true modules are modeled not just as proof fragments, but as structured objects with external interfaces, inferential cores, occurrence profiles, and compositional roles. Training and test settings use the same hidden module alphabet, but differ in their composition graphs, making recompositional generalization a central requirement. A major contribution is the interface-and-closure layer. The paper distinguishes a minimal frame, which exposes only boundary information necessary for structural composition, from richer interfaces that may leak internal organization. It then defines minimal structural closure as the intersection of structural constraints forced in all admissible realizations of a refinement. This prevents contingent interface conveniences from being treated as free structure. Because admissible interface gauges may still vary, the theory works with equivalence classes of frames and isolates a regular frontier of minimal frames whose closure behavior is stable under small perturbations. On top of this layer, the paper defines regularized connection length and regularized cohesion. Regularized connection length measures how much non-free structural coordination is still required to make a refinement act like the original module in a given outer context, after factoring out what the regularized closure provides for free. Regularized cohesion then asks whether every genuine non-equivalent refinement of a module still incurs positive expected residual coordination burden across the module’s context distribution. In this way, a module counts as conceptually coherent when all real decompositions continue to pay a nontrivial reconnection cost even after admissible gauge choices are optimized. The paper then introduces multi-context inseparability, combining positive regularized cohesion with repeated use across diverse contexts. This supports the key anti-fragmentation bridge, called D3-final: if a module is genuinely cohesive and broadly reused, then replacing it by non-equivalent fragments creates a positive compression penalty that grows linearly with the number of occurrences, up to bounded lower-order terms. These ideas are assembled into a three-gap recognition architecture. Over-fragmentation is blocked by the anti-fragmentation gap derived from regularized cohesion. Over-coarsening is blocked by a recompositional generalization gap, since overly large blocks lose flexibility on novel test compositions. Off-axis substitution is blocked by a replacement gap induced by interface separation and bounded overlap, which prevent the true conceptual axes from being rotated into an alternative basis without detectable cost. The final target result is a recognition theorem from stable compression. Informally, it states that in a sufficiently structured modular proof world, any candidate library that is near-optimal under a global objective, stable under admissible representational perturbations, and strong under recomposition on novel composition graphs must belong to the same conceptual equivalence class as the hidden true library. The theorem does not seek recovery of one exact syntax or one unique interface. Rather, it seeks recovery up to stable conceptual equivalence: the same external role geometry, closure behavior, reuse profile, and recompositional function. The paper is explicit that the theory is not yet theorem-complete. Its main unresolved core is no longer definitional vagueness, but a concentrated mathematical problem: proving a genuine anti-fragmentation gap in a natural and sufficiently rich example class. Its present achievement is to isolate that hard core sharply and to replace the vague slogan that “compression discovers concepts” with a much stronger claim: conceptual units are the structures that survive simultaneous tests of near-optimality, stability, and recompositional generalization once fragmentation, coarsening, and basis substitution are all ruled out by explicit gap arguments.

Fields of Study & Subjects

Logic programming and type systems Computability AI Algorithms Mathematics Computing and Information Processing

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
28 March 2026
Language
EN