Active Low-Dimensionalization: Participation Ratio as a Proxy for Historical Assembly Depth

A tractable measure called Participation Ratio (PR), derived from eigenvalue distributions of neural trajectory covariance matrices, can distinguish biological from artificial information processing. Simulations show that history-dependent gain formation in biological systems actively reduces the dimensionality of state-space usage, converging toward a low-PR attractor state regardless of input diversity, consistent with the Maximum Entropy principle. In contrast, fixed gain in artificial systems passively reflects input diversity as high-dimensional spread. The authors propose PR as a practically computable, neurobiologically grounded index for comparing structural organization between biological and artificial systems, connecting to prior work on gain-mediated segregation/integration transitions, attractor manifold topology, and cortical gain modulation.

Study at a glance

Abstract

A central unresolved question in computational consciousness research concerns whether meaningful differences between biological and artificial systems can be quantified without recourse to computationally intractable measures such as Integrated Information Theory (IIT). We propose Participation Ratio (PR) — derived from the eigenvalue distribution of neural trajectory covariance matrices — as a tractable proxy for the effective dimensionality of state-space usage. We show through simulation that history-dependent (biological) gain formation and fixed (AI-style) gain produce qualitatively distinct PR signatures: biological gain actively low-dimensionalizes the state space in a manner robust to input diversity, while AI fixed gain passively reflects input diversity as high-dimensional spread. This active low-dimensionalization is interpreted via the Maximum Entropy principle (Savin & Tkacik, 2017): history-dependent gain formation imposes structured constraints on eigenvalue distributions, converging toward a low-PR attractor state regardless of environmental variability. The result connects to Shine et al. (2018) on gain-mediated segregation/integration transitions, Claudi et al. (2025) MADE framework for attractor manifold topology, and Ferguson & Cardin (2020) on cortical gain modulation mechanisms. We propose PR as a practically computable, neurobiologically grounded index for comparing the structural organization of biological and artificial information processing