General Error Accumulation Relation

Abstract

This document contains a comprehensive evaluation of the preprint "General Error Accumulation Relation" by Stamelos Loutsos (September 2026). The original work addresses a fundamental problem in computational neuroscience and reinforcement learning: how the microscopic Bellman/Temporal‑Difference update rules relate to the macroscopic empirical Loutsos relation, which explicitly accounts for sensorimotor delay.The evaluated paper introduces a novel theoretical framework that bridges these two descriptive levels. Its central contribution is a general scaling law showing that the scaling exponent of accumulated prediction error is determined entirely by the full autocorrelation function of the error process. This result generalises earlier exponential‑decay models and explains anomalous, sub‑diffusive scaling observed in oscillatory neural systems such as ring attractors. The theory is rigorously validated on synthetic TD(0) learners and on real mouse behavioural data from 39 sessions, achieving strong correlations and very low prediction errors, which confirms the universality of the accumulation mechanism.This evaluation critically assesses the strengths and limitations of the work. It highlights the theoretical rigour, the multi‑level empirical validation, and the conceptual clarity offered by the proposed three‑level hierarchy—ranging from early learning dynamics, through Bellman equilibrium, to residual temporal correlations that persist even after convergence. At the same time, it identifies open questions, including the treatment of non‑stationary processes, the sensitivity to the learning‑rate parameter, and the asymptotic nature of the derivation that links the general relation to the specific linear Loutsos form.This assessment is intended as a supplementary resource for researchers working on delayed reinforcement learning, neural error processing, and the statistical physics of learning systems. It offers a balanced perspective that acknowledges both the significant contributions and the remaining challenges, while also suggesting directions for future extensions and practical applications.