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

This paper proposes the Path-Integral Unavoidable Canonical Nonlinearity (PUCN) framework to quantify the cumulative information-geometric cost between arbitrary discrete and continuous distributions, including those with different supports, by utilizing $e$-mixture and harmonic mixture techniques to decompose total canonical nonlinearity into intrinsic discretization costs and residual contributions.

The Gist

The Big Picture: Mapping a Rough Terrain

Imagine you are trying to describe a complex, bumpy landscape (like a mountain range with valleys and peaks) using a smooth, flat map. In physics, this "landscape" is the world of atoms in a material (like an alloy), and the "map" is the mathematical model scientists use to predict how those atoms behave.

Usually, scientists try to approximate this bumpy, discrete world (where atoms sit in specific spots) using a smooth, continuous curve (like a bell curve or a Gaussian distribution). This works well in many cases, but it's never perfect. The "error" or the "distortion" caused by trying to force a smooth curve onto a bumpy reality is what this paper calls Canonical Nonlinearity (CN).

The Problem: Two Types of "Distortion"

The author, Koretaka Yuge, points out that previous methods for measuring this distortion were like trying to measure the error of a photo without knowing if the error came from the camera lens or the subject itself.

1. The "Lens" Error (Discretization): Even if the real world were perfectly smooth, turning it into a digital grid (pixels) creates a small amount of error. This is unavoidable. The paper calls this Unavoidable Canonical Nonlinearity (UCN). It's like the pixelation you see when you zoom in too far on a photo; it happens just because you are using a grid.
2. The "Subject" Error (Non-Gaussianity): Sometimes, the real world isn't smooth at all; it's jagged and weird. If the atoms are arranged in a way that doesn't look like a nice bell curve, that's a second type of error.

The Limitation: The old method (UCN) could only measure the "Lens Error" for a single, specific smooth curve. It couldn't easily compare two different landscapes, especially if they looked totally different (had different "supports"). It was like having a ruler that could only measure one specific type of wood, but couldn't tell you the difference between oak and pine.

The Solution: The "Path-Integral" Journey

To fix this, the author introduces a new tool called Path-Integral UCN (PUCN).

Think of this not as a static ruler, but as a hiking journey.

• The Old Way: You tried to measure the distance between two cities by looking at a single snapshot. If the terrain changed drastically between them, your measurement was confusing.
• The New Way (PUCN): You imagine a hiker walking a path from City A (Distribution 1) to City B (Distribution 2). As the hiker walks, they constantly measure the "cost" of the terrain under their feet.

How the Hiker Walks (The Two Rules)

To make this journey mathematically sound, the hiker follows two specific rules:

1. The "Exponential" Path (The E-Mixture):
   The hiker must stay on a specific type of road that respects the fundamental laws of thermodynamics (the "exponential family"). Imagine this as walking on a highway that always curves in a way that makes sense for the physics of the system. This ensures the hiker doesn't take a shortcut that breaks the laws of nature.

2. The "Harmonic" Step (The M-Mixture):
   As the hiker walks, the size of their "footprint" (the discretization cell) changes. If the terrain gets rougher, the footprint needs to adjust. The paper suggests a "harmonic" way to adjust this.

   • Analogy: Imagine you are walking on a grid of stepping stones. If the stones get bigger or smaller as you walk, you don't just average their sizes. You adjust your step so that the uncertainty of where you might step is balanced perfectly. This is the "harmonic mixture."

The Result: A Clear Breakdown

By using this "hiking path" method, the PUCN allows scientists to break down the total error into two clear parts:

1. The Intrinsic Cost (UCN): The cost just for turning the smooth world into a grid (the "Lens Error").
2. The Residual Cost: The extra cost because the actual world is weird and jagged, not smooth (the "Subject Error").

Why This Matters

Before this paper, scientists were struggling to separate these two errors. They were like a chef trying to taste a soup but couldn't tell if the saltiness came from the salt they added or the natural saltiness of the vegetables.

With PUCN, the chef can now say: "Okay, 30% of the saltiness is just because I used a specific type of pot (discretization), and 70% is because the vegetables were naturally salty (non-Gaussian nature)."

Summary in One Sentence

This paper invents a new mathematical "hiking path" that allows scientists to measure exactly how much error comes from simply turning a smooth world into a digital grid, versus how much error comes because the real world is actually weird and jagged, even when comparing two completely different systems.

Technical Summary

1. Problem Statement

In the statistical thermodynamics of classical discrete systems (e.g., substitutional alloys), the mapping from microscopic interatomic interactions to thermodynamic equilibrium configurations is inherently nonlinear. This phenomenon is known as Canonical Nonlinearity (CN).

• The Limitation of Existing Methods: Traditionally, CN is quantified using the Kullback-Leibler (KL) divergence between a real discrete system's configurational density of states (CDOS) and a reference discretized Gaussian distribution. However, this approach conflates two distinct sources of nonlinearity:
  1. The intrinsic non-Gaussianity of the physical system.
  2. The geometric distortion caused purely by the discretization of a continuous Gaussian distribution.
• The Gap: While recent work identified the "Unavoidable Canonical Nonlinearity" (UCN)—the cost intrinsic to discretization itself—it is limited to evaluating a single continuous distribution. It cannot:
  • Measure the geometric cost between a continuous Gaussian reference and an actual non-Gaussian discrete distribution.
  • Handle cases where the discrete states possess fundamentally different supports (configurational spaces).
  • Provide a unified framework to decompose total CN into intrinsic discretization costs and residual non-Gaussian contributions.

2. Methodology

The author proposes the Path-Integral Unavoidable Canonical Nonlinearity (PUCN) to overcome these limitations. The methodology involves constructing a cumulative path on the statistical manifold connecting two distinct distributions ($P_0$ and $P_1$).

Core Concepts

• UCN Basis: The UCN ($\zeta$) is defined as the expected KL divergence under a discretization map, represented geometrically as the trace of the product of the discretization cell's second-moment matrix ($M$) and the Fisher information metric ($\Omega$):
  $$\zeta = \text{Tr}(M\Omega)$$
• Path-Integral Construction: The PUCN ($\zeta_P$) is defined as the integral of the local discretization cost along a path parameterized by $\lambda \in [0,1]$:
  $$\zeta_P = \int_0^1 d\lambda \, \text{Tr}[M(\lambda)\Omega(\lambda)]$$

Path Construction Strategy

To ensure the path is physically and geometrically meaningful, the author imposes specific constraints on how the distributions and discretization cells evolve:

1. Statistical Manifold Path (The Distribution):

   • The path must remain within the exponential family to maintain consistency with the canonical structure.
   • This is achieved via an e-geodesic (geometric mean of the base measure).
   • For the Fisher metric $\Omega(\lambda)$ (inverse covariance matrix), an arithmetic interpolation is used:
     $$\Omega(\lambda) = (1-\lambda)\Omega(0) + \lambda\Omega(1)$$

2. Discretization Cell Path (The Support):

   • The evolution of the discretization cell matrix $M(\lambda)$ is determined by the uncertainty in parameter variations.
   • By requiring that the covariance of parameter variations corresponds to the inverse Fisher metric, the author derives a harmonic interpolation for $M(\lambda)$:
     $$M(\lambda) = \left[ (1-\lambda)M^{-1}(0) + \lambda M^{-1}(1) \right]^{-1}$$

Robustness

Unlike KL-divergence-based methods which are sensitive to the choice of continuous embedding, the PUCN relies on the trace form ($\text{Tr}(M\Omega)$). This form originates from optimal transport theory in the infinitesimal limit, providing a purely geometric characterization that is stable across different embedding schemes.

3. Key Contributions

• Generalization of UCN: Extends the concept of unavoidable nonlinearity from a single distribution to a path connecting arbitrary distributions, even those with different supports.
• Decomposition of Canonical Nonlinearity: Provides a rigorous mathematical framework to separate total CN into two distinct components:
  1. Intrinsic Discretization Cost (UCN): The cost arising solely from the discretization of a Gaussian reference.
  2. Residual Contribution: The cost arising from the deviation of the actual system from the Gaussian reference (non-Gaussianity).
• New Geometric Measure: Introduces a flexible metric for information-geometric costs that does not require the distributions to share the same support, a significant advancement over previous KL-divergence approaches.

4. Results

• Mathematical Formulation: The paper successfully derives the PUCN formula (Eq. 6) and demonstrates its decomposition (Eq. 7):
  $$\zeta_P = \zeta_{\text{UCN}} + \int_0^1 d\lambda \, \text{Tr}[M, \Delta\Omega(\lambda)]$$
  Here, the second term quantifies the non-Gaussian contribution on the discrete support.
• Consistency: The formulation ensures that when the path connects a continuous Gaussian to its discretized version, the result recovers the standard UCN. When connecting to a non-Gaussian system, the residual term captures the specific structural nonlinearity of that system.
• Applicability: The framework allows for the explicit quantification of CN between different Cluster Expansion (CDOS) systems, facilitating comparisons between systems with fundamentally different lattice constraints.

5. Significance

• Theoretical Clarity: Resolves the conceptual ambiguity in existing literature regarding the separation of "discretization artifacts" from "physical non-Gaussianity."
• Unified Framework: Offers a unified approach to treat discretization effects and structural deviations within information geometry.
• Practical Utility: Enables more accurate thermodynamic modeling of complex materials (like alloys) by isolating the geometric cost of the lattice discretization from the actual physics of the interactions.
• Future Directions: The framework opens new avenues for analyzing optimal paths that minimize geometric costs and extending these concepts to non-Gaussian continuous references.

In summary, this paper advances the field of statistical thermodynamics by providing a robust, path-integral-based tool to quantify and decompose the geometric costs inherent in modeling discrete physical systems, distinguishing between the unavoidable costs of discretization and the intrinsic nonlinearity of the material itself.