Cyclic Distinction–Integration Learning via Linear Coherence Bottlenecks

This paper introduces a cyclic learning framework in which representational enrichment arises through alternating phases of distinction and integration constrained by a linear coherence bottleneck. The model formalizes learning as a stability–plasticity process operating under bandwidth limitations, yielding predictable enrichment per cycle and an analytically tractable optimal period.The framework is presented conservatively as a representation-level learning mechanism, independent of specific biological or cognitive interpretations. Connections to dimensionality control, compression–decompression dynamics, and training stability in artificial learning systems are discussed.This work is exploratory and theoretical in nature and is intended to support further empirical investigation and ablation in machine learning and related domains.