The Ladder of Emergence E: Information Cost Landscape and Gradient Flow — The Economic Geometry of Model Selection

Fuente: Zenodo
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: zhou, changzheng, zhou, ziqing
Format: Recurso digital
Veröffentlicht: Zenodo 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866901403140620288
author zhou, changzheng
zhou, ziqing
author_facet zhou, changzheng
zhou, ziqing
contents <p>The core challenge of model selection lies in achieving an optimal balance be<br>tween descriptive complexity and predictive power. From the intersection of infor<br>mation theory and differential geometry, this paper constructs a rigorous mathe<br>matical framework of “cognitive economics” for this trade-off. The entire system<br>begins by recognizing the fundamental tension between classical Kolmogorov com<br>plexity and continuous geometric tools, bridging them via the minimum description<br>length principle, constructing the model space as the union of parametric statisti<br>cal manifolds. On this manifold, the descriptive complexity and predictive mutual<br>information of a model are unified into a smooth economic cost functional. This<br>paper rigorously proves that on this cost landscape, the natural gradient flow with<br>Fisher information as metric possesses local existence and uniqueness, and under<br>the assumptions of real analyticity of the cost and compactness of the physical sub<br>manifold, it converges to local minima via the Łojasiewicz inequality. Paradigm<br>emergence is interpreted as the gradient flow being obstructed at a low-dimensional<br>critical submanifold, then restarting through embedding into a higher-dimensional<br>Grassmannian tube. Subsequently, the three failure modes of hypothesis testing<br>are encoded into a sequential information filtration, and a rigorous channel separa<br>tion argument is given to prove that under a finite information budget, a three-layer<br>filtration is the minimal complete structure that can distinguish logical falsehood,<br>statistical unreliability, and insufficient precision. Finally, within the frameworks<br>of metric measure spaces and optimal transport, the pointwise gradient flow is lifted<br>to a Wasserstein gradient flow of probability measures, and the discrete paradigm<br>transition is quantified as the optimal stochastic control of a Schrödinger bridge<br>problem, whose minimum relative entropy gives the information cost and proba<br>bility of the transition. This framework provides a unified geometric and analytic<br>language for the cognitive economics of theory choice.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_20019264
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle The Ladder of Emergence E: Information Cost Landscape and Gradient Flow — The Economic Geometry of Model Selection
zhou, changzheng
zhou, ziqing
model selection; cost landscape; information geometry; natural gradient f low; minimum description length; information filtration; optimal transport
<p>The core challenge of model selection lies in achieving an optimal balance be<br>tween descriptive complexity and predictive power. From the intersection of infor<br>mation theory and differential geometry, this paper constructs a rigorous mathe<br>matical framework of “cognitive economics” for this trade-off. The entire system<br>begins by recognizing the fundamental tension between classical Kolmogorov com<br>plexity and continuous geometric tools, bridging them via the minimum description<br>length principle, constructing the model space as the union of parametric statisti<br>cal manifolds. On this manifold, the descriptive complexity and predictive mutual<br>information of a model are unified into a smooth economic cost functional. This<br>paper rigorously proves that on this cost landscape, the natural gradient flow with<br>Fisher information as metric possesses local existence and uniqueness, and under<br>the assumptions of real analyticity of the cost and compactness of the physical sub<br>manifold, it converges to local minima via the Łojasiewicz inequality. Paradigm<br>emergence is interpreted as the gradient flow being obstructed at a low-dimensional<br>critical submanifold, then restarting through embedding into a higher-dimensional<br>Grassmannian tube. Subsequently, the three failure modes of hypothesis testing<br>are encoded into a sequential information filtration, and a rigorous channel separa<br>tion argument is given to prove that under a finite information budget, a three-layer<br>filtration is the minimal complete structure that can distinguish logical falsehood,<br>statistical unreliability, and insufficient precision. Finally, within the frameworks<br>of metric measure spaces and optimal transport, the pointwise gradient flow is lifted<br>to a Wasserstein gradient flow of probability measures, and the discrete paradigm<br>transition is quantified as the optimal stochastic control of a Schrödinger bridge<br>problem, whose minimum relative entropy gives the information cost and proba<br>bility of the transition. This framework provides a unified geometric and analytic<br>language for the cognitive economics of theory choice.</p>
title The Ladder of Emergence E: Information Cost Landscape and Gradient Flow — The Economic Geometry of Model Selection
topic model selection; cost landscape; information geometry; natural gradient f low; minimum description length; information filtration; optimal transport
url https://doi.org/10.5281/zenodo.20019264