Causality is Geometry: The Korach Manifold Separation Principle for Stochastic Dynamical Systems & World Models

Read full text Summarise with VIRAT
Authors: Jordan Korach
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

The modeling and control of stochastic, non-holonomic dynamical systems remains fundamentallybottlenecked by the assumption of known, ambient Euclidean geometry. While recent extensions ofReproducing Kernel Hilbert Space (RKHS) Koopman operator theory to Stochastic Differential Equationshave established rigorous Feynman-Kac semigroups, their truncation error bounds scale with ambientdimension D, triggering the curse of dimensionality. Furthermore, discovering causal structuralconstraints in such systems without interventional access remains an open challenge. In this work, weintroduce the Korach Manifold Separation Principle (KMSP), a unified framework that factorizes chaoticstochastic dynamics into three invariant structures: topology, structural causality, and directional flow. Bycombining spectral graph-theoretic manifold discovery with Nyström RKHS Koopman approximations,we prove that truncation error is strictly bounded by the intrinsic dimension d ≪ D of the discoveredproduct manifold, regularized by a novel Bochner-Ricci asymmetric stability criterion. Crucially, weformulate the Korach Causal-Geometric Equivalence Theorem, demonstrating that InverseReinforcement Learning (IRL) on Riemannian manifolds recovers an undirected structural metric g_ijwhich, when paired with the asymmetric Koopman Jacobian, recovers the exact directed causal DAG ofthe underlying data-generating process. We validate our approach on stochastic Kuramoto oscillators,Lorenz-96 turbulence, and synthetic non-linear systems, establishing a strictly intrinsic, geometricallyaware foundation for global optimal control

Fields of Study & Subjects

Model Reduction and Neural Networks Control Systems and Identification Mathematical Dynamics and Fractals

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint