The Thermodynamics of Learning: Recognition Cascades, Information-Geometric Action Bounds, and Cross-Domain Predictions from Unified Spectral Structure

Summarise with VIRAT
Authors: Duggan Matthews
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

We formulate learning as natural gradient flow on a statistical manifold and show that learning trajectories obey an action-budget inequality determined by the spectral structure of the environment. Using normal-form reduction we demonstrate that generic codimension-1 recognition events obey a universal quartic cost law: the Fisher action required for structural reorganisation scales as the fourth power of the information-geometric displacement, with explicit coefficient 1/12 from the cusp normal form. Combining the quartic law with the action budget yields a cascade constraint governing the total structural reorganisation achievable under finite spectral budget — learning systems are driven toward cascades of smaller recognition events rather than single large reorganisations. We then identify a cross-domain regularity: the same quartic recognition law, cascade constraint, and conservation structure appear independently in machine learning (grokking transitions in neural networks), cosmological phase transitions (symmetry breakings from Planck to recombination), cellular biology (differentiation and de-differentiation under energetic constraint), multi-agent prediction (collective convergence toward compatibility), quantum biology (anomalous coherence timescales in living systems), and gravitational dynamics (entanglement maintenance and geometric coherence). We derive seven cross-domain predictions that follow from the assumption that this regularity is structural rather than coincidental, and from no single-domain theory alone: shared quartic exponents across neural networks and cellular differentiation; universal cascade bounds across training and cosmology; incompatibility cost predicting cancer treatment resistance; meta-ignorance diagnostics for failed clinical trials; preserved spectral activation ordering across systems; biological quantum coherence scaling with regulatory bandwidth via the Mutualism bifurcation; and life enriching global geometric coherence through nonlocal entanglement maintenance. Each prediction is testable on existing data or near-term experiments. If the predictions fail, the cross-domain regularity is coincidence and the domains require separate explanations. If the predictions succeed, the domains are governed by a common spectral operator and the separation between physics, biology, machine learning, and cosmology is not in the thing being studied but in the studying. The observation channels are many. The operator is one. The predictions will tell us which.

Fields of Study & Subjects

Complex Systems and Dynamics Neural Networks and Applications Advanced Thermodynamics and Statistical Mechanics

Publication details

Journal / source
OSF Preprints (OSF Preprints)
Publication date
20 March 2026
Language
EN
Document type
Preprint