Drift as Deformation
Abstract
This manuscript examines behavioral drift in learned systems through a structural lens rather than a purely geometric one. Instead of treating drift as displacement in latent space, the work proposes interpreting drift as deformation relative to directions that preserve system identity. The core idea is to distinguish structurally stiff dimensions, those enforced by agreement between independent observers, from elastic dimensions where variation is absorbed without semantic consequence. Drift is then characterized by how perturbations project onto these stiffness-weighted directions, revealing when change constitutes genuine structural deformation rather than benign motion. The paper develops this perspective conceptually and formally, introducing an agreement-based energy functional that induces a resistance landscape prior to metric geometry. This ordering allows geometry to be interpreted relative to structural constraints, rather than assumed a priori. The work is intended as a foundation for future empirical and geometric analysis of drift, particularly in systems where identity, role, or semantics must be preserved over time.
Fields of Study & Subjects
Topological and Geometric Data Analysis
Neural Networks and Applications
Domain Adaptation and Few-Shot Learning
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 2 February 2026
- Language
- EN
- Document type
- Preprint