Path-Integral Formulation of Unavoidable Canonical Nonlinearity: Dynamic Discretization Cost over Variable Supports

TL;DR

This paper introduces the Path-Integral UCN (PUCN), a novel method to quantify the information-geometric cost of discretization in classical systems, capturing differences between diverse distributions and supports.

Contribution

It develops a path-integral framework for unavoidable canonical nonlinearity, enabling explicit decomposition and comparison of discretization costs across systems.

Findings

PUCN quantifies the geometric cost between different distributions.

It allows decomposition of total CN into UCN and residual contributions.

The method captures costs between states with different supports.

Abstract

In the statistical thermodynamics of classical discrete systems, the map from microscopic interactions to thermodynamic equilibrium configurations generally exhibits complex nonlinearity, known as "canonical nonlinearity" (CN). While conventionally characterized by the Kullback-Leibler (KL) divergence, this approach inevitably misses intrinsic nonlinearities arising from the discretization of continuous Gaussian families themselves. This intrinsic effect of unavoidable CN (UCN), has recently been quantified within a transport-information-geometric framework. However, the UCN is fundamentally limited to evaluating the discretization-induced cost for a single continuous distribution. It therefore does not capture the information-geometric cost between a continuous Gaussian reference and an actual discrete distribution, nor between states with fundamentally different supports, making it conceptually unclear how to decompose the overall CN. To address this limitation, we propose the Path-Integral UCN (PUCN) quantifying the cumulative information-geometric cost between distinct distributions. The PUCN adopts a path by (i) retaining the canonical distribution as an exponential family via the $e$-mixture (geometric mean) of the base measure, leading to an arithmetic mixture of the Fisher metric as the CN standard, and (ii) enforcing covariant changes in the discretization cell through the harmonic mixture of its second-moment matrix $M$, reflecting the uncertainty in parameter variations on the statistical manifold. The resulting PUCN provides a flexible measure of the geometric cost between arbitrary states, including those with essentially different supports. This formulation enables an explicit quantification of CN between different CDOS systems and a natural decomposition of the total CN into the UCN and a residual contribution, which has not been clearly separated in existing approaches.