2025-11-23T09:16:16.957000

Reversibility, covariance and coarse-graining for Langevin dynamics: On the choice of multiplicative noise

Ayala, Dirr, Pavliotis et al.
We study the interplay between reversibility, geometry, and the choice of multiplicative noise (in particular Itô, Stratonovich, Klimontovich) in stochastic differential equations (SDEs). Building on a unified geometric framework, we derive algebraic conditions under which a diffusion process is reversible with respect to a Gibbs measure on a Riemannian manifold. The condition depends continuously on a parameter $λ\in [0,1]$ which interpolates between the conventions of Itô ($λ= 0$), Stratonovich ($λ= \frac 1 2$) and Klimontovich ($λ= 1$). For reversible slow-fast systems of SDEs with a block-diagonal diffusion structure, we show, using the theory of Dirichlet forms, that both reversibility and the Klimontovich noise interpretation are preserved under coarse-graining. In particular, we prove that the effective dynamics for the slow variables, obtained via projection onto a lower-dimensional manifold, retain the Klimontovich interpretation and remain reversible with respect to the marginal Gibbs measure/free energy. Our results provide a flexible variational framework for modeling coarse-grained reversible dynamics with nontrivial geometric and noise structures.
academic

Reversibility, covariance and coarse-graining for Langevin dynamics: On the choice of multiplicative noise

Basic Information

  • Paper ID: 2511.03347
  • Title: Reversibility, covariance and coarse-graining for Langevin dynamics: On the choice of multiplicative noise
  • Authors: Mario Ayala (TU München), Nicolas Dirr (Cardiff University), Grigorios A. Pavliotis (Imperial College London), Johannes Zimmer (TU München)
  • Classification: math.PR (Probability Theory)
  • Publication Date: November 6, 2025 (arXiv preprint)
  • Paper Link: https://arxiv.org/abs/2511.03347

Abstract

This paper investigates the interplay between reversibility, geometric structure, and the choice of multiplicative noise interpretation (specifically Itô, Stratonovich, and Klimontovich conventions) in stochastic differential equations (SDEs). Based on a unified geometric framework, the authors derive algebraic conditions for reversibility of diffusion processes with respect to Gibbs measures on Riemannian manifolds. These conditions depend continuously on a parameter λ∈0,1, which interpolates between Itô (λ=0), Stratonovich (λ=1/2), and Klimontovich (λ=1) conventions. For reversible slow-fast systems with block-diagonal diffusion structure, the authors employ Dirichlet form theory to prove that both reversibility and Klimontovich noise interpretation are preserved under coarse-graining. In particular, the effective dynamics of slow variables obtained through projection onto a lower-dimensional manifold preserve the Klimontovich interpretation and remain reversible with respect to the marginal Gibbs measure/free energy.

Research Background and Motivation

Core Problem

The core problem addressed in this paper is: How should one choose the interpretation of stochastic integrals (Itô, Stratonovich, or Klimontovich) in the multiplicative noise (state-dependent noise) setting to ensure that the diffusion process is reversible with respect to a given Gibbs measure?

Importance of the Problem

  1. Foundations of Statistical Mechanics: Reversibility is a central concept in equilibrium statistical mechanics, requiring not only the existence of an invariant measure but also satisfaction of the detailed balance condition, corresponding to zero entropy production rate.
  2. Sampling Algorithms: In MCMC methods and stochastic gradient descent, reversible Langevin dynamics serve as fundamental tools.
  3. Physical Modeling: Multiplicative noise naturally appears in fluctuating hydrodynamics, active matter models, and other physical systems.
  4. Convergence Rate Optimization: Optimizing the diffusion matrix σ can accelerate convergence to equilibrium, which is crucial in sampling problems.

Limitations of Existing Approaches

  1. Restriction to Additive Noise: The classical overdamped Langevin equation dXt = -∇V(Xt)dt + √2dWt applies only to constant diffusion matrices.
  2. Dilemma with Itô Interpretation: Under the Itô interpretation, multiplicative noise almost always destroys reversibility unless extremely stringent harmonic coordinate conditions are satisfied.
  3. Lack of Unified Framework: The relationships between different noise interpretations and their effects on reversibility have lacked systematic study.
  4. Unknown Preservation Under Coarse-graining: Whether noise interpretation and reversibility are preserved under coarse-graining in multiscale systems remains unclear.

Research Motivation

The authors' starting point is to establish a unified geometric framework that systematically characterizes the algebraic conditions for reversibility under different noise interpretations, and to prove that the Klimontovich interpretation possesses special preservation properties under coarse-graining. This provides a theoretical foundation for modeling reversible dynamics with nontrivial geometry and noise structure.

Core Contributions

  1. Unified Reversibility Condition: The authors derive a parametrized algebraic condition (2λ-1)∇^c·(σσ^T) = 2λσ(∇^c·σ^T) that completely characterizes reversibility of the SDE with respect to Gibbs measures, where λ∈0,1 interpolates between different noise interpretations.
  2. Precise Characterization of Three Noise Types:
    • Itô Noise (λ=0): Requires vanishing covariant divergence ∇^c·(σσ^T)=0, equivalent to harmonic coordinate condition, extremely restrictive.
    • Stratonovich Noise (λ=1/2): Requires σ∇^c·σ^T=0, a noise incompressibility condition.
    • Klimontovich Noise (λ=1): Always satisfied in one dimension; automatically satisfied in higher dimensions for diagonal or block-diagonal σ.
  3. Coarse-graining Preservation Theorem: For slow-fast systems with independent noise sources, both Klimontovich noise interpretation and reversibility are preserved in the averaging limit, with effective dynamics remaining reversible with respect to the marginal Gibbs measure.
  4. Dirichlet Form Method: Using Kuwae-Shioya's Mosco convergence theory, the authors provide a rigorous variational framework for coarse-graining of multiscale systems, offering an elegant alternative to classical generator methods.
  5. Geometric Perspective: The diffusion matrix M=σσ^T is interpreted as a Riemannian metric g=M^{-1}, providing a unified treatment of noise and geometric corrections within this geometric framework.

Method Details

Task Definition

Input:

  • Smooth confining potential function V: ℝ^d → ℝ (satisfying lim_{|x|→∞}V(x)=+∞)
  • Smooth invertible matrix field σ: ℝ^d → GL(d) (volatility matrix)
  • Noise parameter λ∈0,1

Output: Determine whether the SDE dXt=−σ(Xt)σT(Xt)∇V(Xt)dt+2σ(Xt)◦λdWtdX_t = -σ(X_t)σ^T(X_t)∇V(X_t)dt + \sqrt{2}σ(X_t) ◦_λ dW_t is reversible with respect to the Gibbs measure G(dx)=1ZVe−V(x)volM(dx)G(dx) = \frac{1}{Z_V}e^{-V(x)}\text{vol}_M(dx) where vol_M(dx)=√(ω_M(x))dx is the Riemannian volume element, ω_M(x)=det(M^{-1}(x)).

Constraints: σ must satisfy uniform ellipticity condition, i.e., there exists C>0 such that ξ^T M(x)ξ ≥ C‖ξ‖² for all x and ξ.

Theoretical Framework

1. Geometric Setup

The inverse M^{-1}(x) of the diffusion matrix M(x)=σ(x)σ^T(x) is interpreted as a Riemannian metric g(x)=M^{-1}(x), making (ℝ^d,g) a Riemannian manifold. On this manifold:

  • Levi-Civita Connection: The unique torsion-free connection compatible with the metric, satisfying ∇^c_ℓ g_=0
  • Christoffel Symbols: Γijk=12gkℓ(∂igjℓ+∂jgiℓ−∂ℓgij)Γ^k_{ij} = \frac{1}{2}g^{kℓ}(∂_i g_{jℓ} + ∂_j g_{iℓ} - ∂_ℓ g_{ij})
  • Covariant Divergence: For a matrix field A, the row covariant divergence is defined as (∇c⋅A)j=∂iAji+ΓikiAjk(∇^c·A)^j = ∂_i A^{ji} + Γ^i_{ik}A^{jk}

2. Generator Analysis

The infinitesimal generator of SDE (8) is: LMf=ΔMf+B⋅∇f\mathcal{L}_M f = Δ_M f + B·∇f where:

  • Laplace-Beltrami Operator: ΔMf=1ωM∂i(ωMMij∂jf)Δ_M f = \frac{1}{\sqrt{ω_M}}∂_i(\sqrt{ω_M}M^{ij}∂_j f)
  • Corrected Drift: Bi=−Mij(∇V)j+2λ(∂jMji−σiℓ∂kσkℓ)+MkjΓkjiB^i = -M^{ij}(∇V)_j + 2λ(∂_j M^{ji} - σ^{iℓ}∂_k σ^{kℓ}) + M^{kj}Γ^i_{kj}

The second term is the noise correction (depending on λ), and the third term is the geometric correction.

3. Reversible Generator

The reversible generator for the Gibbs measure G(dx) (following BGL13) is: LGf=ΔMf+Γ(−V,f)\mathcal{L}_G f = Δ_M f + Γ(-V,f) where Γ(-V,f)=-M^{ij}(∂_j V)∂_i f is the carré-du-champ operator. Using the contraction property of Christoffel symbols: ΓikjMik=(∇c⋅(σσT))jΓ^j_{ik}M^{ik} = (∇^c·(σσ^T))^j the reversible generator can be written as: LGf=Mij∂ij2f+(∇c⋅(σσT))j∂jf−Mij(∂iV)∂jf\mathcal{L}_G f = M^{ij}∂²_{ij}f + (∇^c·(σσ^T))^j∂_j f - M^{ij}(∂_i V)∂_j f

Main Theorem (Theorem 3.1)

Statement: G(dx) is reversible with respect to X_t (i.e., the generator is self-adjoint in L²(G)) if and only if: (2λ−1)∇c⋅(σσT)=2λσ(∇c⋅σT)(2λ-1)∇^c·(σσ^T) = 2λσ(∇^c·σ^T)

Proof Strategy:

  1. First Step: Write the coordinate representation of the reversible generator L_G (as above)
  2. Second Step: Using the noise transformation formula (56), express L in a form containing λ corrections: Lf=Mij∂ij2f+2λ[(∇c⋅(σσT))−σ(∇c⋅σT)]j∂jf−Mij(∂iV)∂jf\mathcal{L}f = M^{ij}∂²_{ij}f + 2λ[(∇^c·(σσ^T)) - σ(∇^c·σ^T)]^j∂_j f - M^{ij}(∂_i V)∂_j f
  3. Third Step: Using metric compatibility (17), prove that the geometric correction term equals the covariant divergence: −ΓikjMik=(∇c⋅(σσT))j-Γ^j_{ik}M^{ik} = (∇^c·(σσ^T))^j
  4. Fourth Step: Compare coefficients of L=L_G to obtain that reversibility is equivalent to: 2λ[(∇c⋅(σσT))−σ(∇c⋅σT)]=(∇c⋅(σσT))2λ[(∇^c·(σσ^T)) - σ(∇^c·σ^T)] = (∇^c·(σσ^T)) Rearranging yields the theorem condition.

Key Lemma (Proposition 2.3): Proves that [∇c⋅(σσT)]j−[σ(∇c⋅σT)]j=[∇⋅(σσT)]j−[σ(∇⋅σT)]j[∇^c·(σσ^T)]^j - [σ(∇^c·σ^T)]^j = [∇·(σσ^T)]^j - [σ(∇·σ^T)]^j i.e., the Christoffel symbol terms completely cancel in this combination, allowing the condition to be expressed using Euclidean divergence.

Technical Innovations

  1. Introduction of Row Covariant Divergence: The standard covariant divergence of the (2,0)-tensor M=σσ^T vanishes identically due to metric compatibility. The authors' row covariant divergence captures the correct geometric analogue.
  2. Unified Interpolation Framework: A single parameter λ unifies three noise interpretations, making the reversibility condition continuously dependent on λ, revealing intrinsic connections between them.
  3. Bridge Between Geometry and Stochastic Analysis: Connects the choice of stochastic integral (an analytical problem) with covariant operations in Riemannian geometry (a geometric problem).
  4. Clarification of Graham's Covariant Form: Shows that under the weaker condition σ∇^c·σ^T=0, the Stratonovich SDE is equivalent to the Itô SDE, improving upon Diósi 2024's result.

Coarse-graining Theory (Section 4)

Slow-Fast System Setup

Consider a two-dimensional system (generalizing to higher dimensions is straightforward):

dX^n_t = -σ²_1(X^n_t,Y^n_t)∂_x V(X^n_t,Y^n_t)dt + \sqrt{2}σ_1(X^n_t,Y^n_t)◦_K dW¹_t \\ dY^n_t = -nσ²_2(X^n_t,Y^n_t)∂_y V(X^n_t,Y^n_t)dt + \sqrt{2n}σ_2(X^n_t,Y^n_t)◦_K dW²_t \end{cases}$$ where: - X is the slow variable, Y is the fast variable - Noise sources W¹ and W² are independent (corresponding to block-diagonal diffusion matrix) - Both use Klimontovich interpretation (λ=1) - Joint Gibbs measure is μ(dxdy) = Z^{-1}_V e^{-V(x,y)}dxdy ### Averaging Limit As n→∞, classical results [Kha68] show that X^n_t converges to the effective process: $$d\bar{X}_t = \bar{b}(\bar{X}_t)dt + \sqrt{2}\bar{σ}_1(\bar{X}_t)dW_t$$ where the effective coefficients are: $$\bar{b}(x) = \frac{1}{Z_V(x)}\int_ℝ [∂_x σ²_1(x,y) - σ²_1(x,y)∂_x V(x,y)]e^{-V(x,y)}dy$$ $$\bar{σ}²_1(x) = \frac{1}{Z_V(x)}\int_ℝ σ²_1(x,y)e^{-V(x,y)}dy$$ and the marginal measure is μ_∞(dx) ∝ ∫_ℝ e^{-V(x,y)}dy. ### Preservation Theorem **Key Calculation**: Through the differential relation $$∂_x(\bar{σ}²_1(x)) = \bar{b}(x) + \bar{σ}²_1(x)\frac{∂_x Z_V(x)}{Z_V(x)}$$ the effective SDE can be rewritten in Klimontovich form: $$d\bar{X}_t = -\bar{σ}²_1(\bar{X}_t)∂_x\ln\left(\int_ℝ e^{-V(\bar{X}_t,y)}dy\right)dt + \sqrt{2}\bar{σ}_1(\bar{X}_t)◦_K dW_t$$ This directly proves: **Klimontovich noise interpretation is preserved under averaging**, and the effective dynamics remain reversible with respect to the marginal Gibbs measure μ_∞. ### Dirichlet Form Method #### Mosco Convergence Framework The authors employ Kuwae-Shioya [2003] theory, studying convergence of Dirichlet forms on varying Hilbert spaces: 1. **Pre-limit Spaces**: $$H¹_n(μ) = \{f∈L²(ℝ²,μ) : \int_{ℝ²}∑_{ij}M^{ij}_n(∂_i f)(∂_j f)μ(dxdy)<∞\}$$ where M_n = diag(σ²_1, nσ²_2) 2. **Limit Space**: $$H¹(μ_∞) = \{f∈L²(ℝ,μ_∞) : \int_ℝ \bar{σ}²_1(x)(∂_x f)²μ_∞(dx)<∞\}$$ 3. **Embedding Maps**: Φ_n: H¹(μ_∞)→H¹_n(μ) defined by (Φ_n f)(x,y)=f(x) **Key Property (Remark 4.1)**: Φ_n is an isometric embedding preserving Dirichlet forms: $$\mathcal{E}_n(Φ_n f) = \mathcal{E}(f), \quad ∀f∈H¹(μ_∞)$$ **Main Result (Theorem 4.1)**: The Dirichlet form sequence {(E_n,D(E_n))} converges in the Mosco sense to (E,D(E)), ensuring convergence of semigroups and resolvents, thus rigorously proving the averaging principle. ### Higher-Dimensional Generalization For (d+m)-dimensional systems, all results directly generalize as long as noise sources are independent (block-diagonal structure). The effective diffusion parameter becomes a matrix-valued Bochner integral: $$\bar{σ}_1(x)\bar{σ}^T_1(x) = \frac{1}{Z_V(x)}\int_{ℝ^m}σ_1(x,y)σ^T_1(x,y)e^{-V(x,y)}dy$$ For σ_1(x,y)=UΛ(x,y)U^T (U constant orthogonal, Λ diagonal), the authors verify through explicit calculation (Section 4.7) that the Klimontovich condition (55) still holds after averaging. ## Experimental Setup **Note**: This is a pure theoretical mathematics paper containing no numerical experiments. All results are theorems and propositions obtained through rigorous mathematical proofs. ### Theoretical Verification Methods 1. **One-Dimensional Case Analysis** (Section 1.1): Through explicit generator calculations, verify that only Klimontovich interpretation (λ=1) ensures reversibility for arbitrary σ(x). 2. **Geometric Examples**: - Diagonal matrices σ=diag(σ_1,...,σ_d) automatically satisfy Klimontovich condition - Block-diagonal structures preserve reversibility - Constant orthogonal transformations σ=UDU^T (U constant) preserve the condition 3. **Counterexample Construction**: If U=U(x) depends on position, the Klimontovich condition typically fails, showing that rotation of noise directions introduces geometric distortion. ## Related Work ### Stochastic Differential Equation Theory 1. **Itô Integral** [Itô44]: Ensures the integral is a martingale, foundational for weak solution and convergence theory. 2. **Stratonovich Integral** [Str66]: Preserves classical chain rule, natural in physical modeling. 3. **Klimontovich Integral** [Kli90, Hän78]: Consistent with Fick's diffusion law, corresponding Fokker-Planck equation [ER25]. ### Reversibility and Non-equilibrium Statistical Mechanics 1. **Detailed Balance Condition** [BGL13, Pav14]: Self-adjoint generator equivalent to zero entropy production. 2. **Entropy Production Theory** [MN03, CP23]: Projection of irreversible component determines entropy production rate. 3. **Fluctuation Relations** [JQQ04]: Gallavotti-Cohen type fluctuation theorems. ### Geometric Stochastic Analysis 1. **Graham's Covariant Form** [Gra77, Gra85]: First proposed covariant formulation of Fokker-Planck equation. 2. **Diffusions on Manifolds** [Hsu02, Wei72]: Laplace-Beltrami operator on Riemannian manifolds. 3. **Diósi's Covariance Condition** [Dió24]: This paper generalizes and improves his Stratonovich-Itô equivalence result. ### Sampling and Optimization 1. **Preconditioned Langevin Dynamics** [Che+24]: Connections with Fisher-Rao gradient flow. 2. **Spectral Gap Optimization** [Lel+25]: Optimizing σ to maximize convergence rate. 3. **Stochastic Gradient Descent** [LTE19, DZ20]: Fokker-Planck perspective. ### Multiscale and Coarse-graining 1. **Averaging Principle** [Kha68, PS08]: Classical slow-fast system theory. 2. **Dirichlet Form Method** [BR14]: Studying averaging through Dirichlet forms. 3. **Mosco Convergence** [KS03, Kol06]: Spectral convergence on varying Hilbert spaces. 4. **Two-Scale Convergence** [DDP23]: Brownian motion in periodic potentials. ### Data-Driven Methods 1. **Learning Diffusion Matrices** [Zhu+25]: Learning reversible/irreversible components from trajectory data. 2. **Discovering Evolution Equations** [Li+19, Hua+25]: Learning macroscopic operators from particle data. ## Conclusions and Discussion ### Main Conclusions 1. **Decisive Role of Noise Choice**: In the multiplicative noise setting, the choice of stochastic integral interpretation (Itô, Stratonovich, Klimontovich) fundamentally affects reversibility: - Itô interpretation is too restrictive (requires harmonic coordinates) - Klimontovich interpretation is most permissive (always holds in 1D, holds for common structures in higher dimensions) - Stratonovich is intermediate 2. **Superiority of Klimontovich Noise**: - Naturally preserved under coarse-graining - Effective dynamics automatically preserve reversible structure - Marginal Gibbs measure remains in Gibbs form 3. **Unifying Power of Geometric Perspective**: Interpreting the diffusion matrix as a Riemannian metric, the covariant divergence condition provides intrinsic geometric characterization independent of coordinate choice. 4. **Elegance of Dirichlet Form Approach**: The Mosco convergence framework provides a variational perspective for multiscale analysis, avoiding technical complexities of classical generator methods. ### Limitations 1. **Requirement of Block-Diagonal Structure**: The coarse-graining preservation theorem requires independent noise sources for slow and fast variables (block-diagonal M_n), excluding coupled noise cases. 2. **Smoothness Assumptions**: Requires C² smoothness of σ and V; theory does not apply to non-smooth or degenerate cases (e.g., boundary conditions). 3. **Uniform Ellipticity**: Requires the diffusion matrix to be uniformly positive definite, excluding cases where some degrees of freedom are "frozen." 4. **Specificity to Gibbs Measures**: Theory specifically targets Gibbs-form measures; extension to more general invariant measures (e.g., non-equilibrium steady states) requires modification. 5. **Non-trivial Klimontovich Condition in High Dimensions**: In higher dimensions, the Klimontovich condition (35) remains a non-trivial algebraic constraint, not satisfied for all σ. ### Future Directions 1. **Coarse-graining with Coupled Noise**: Study averaging for non-block-diagonal diffusion matrices, possibly requiring modified Klimontovich interpretation. 2. **Non-equilibrium Extensions**: Extend theory to non-equilibrium steady states with external forces or circulation flows. 3. **Numerical Algorithms**: Develop new numerical schemes based on Klimontovich discretization, preserving reversibility at discrete level. 4. **Learning Framework**: Simultaneously learn σ, V, and optimal λ from data, enabling data-driven reversible modeling. 5. **Infinite-Dimensional Extensions**: Extend theory to SPDEs and infinite-dimensional manifolds (e.g., probability measure spaces). 6. **Quantum Analogues**: Explore connections with path integral representations in quantum mechanics. ## In-Depth Evaluation ### Strengths 1. **Theoretical Depth and Originality**: - First unified parametrized reversibility condition, revealing continuous spectrum of three noise interpretations - Preservation of Klimontovich noise under coarse-graining is novel discovery with profound physical significance - Introduction of row covariant divergence elegantly resolves the problem of standard covariant divergence vanishing identically 2. **Mathematical Rigor**: - All theorems have complete proofs with accurate citations - Application of Mosco convergence demonstrates sophisticated functional analysis techniques - Use of metric compatibility and geometric properties is elegant 3. **Framework Unification**: - Organically combines stochastic analysis, Riemannian geometry, non-equilibrium statistical mechanics, and multiscale analysis - Theory is consistent from one dimension to higher dimensions, from single scale to multiscale - Dirichlet form method provides complementary perspective to classical approaches 4. **Physical Intuition**: - Interpretation of σ as "noise landscape," with small σ suppressing fluctuations and large σ dominated by noise - Geometric meaning of covariant divergence condition is clear (harmonic coordinates, incompressibility, etc.) - Coarse-graining preservation aligns with spirit of second law of thermodynamics 5. **Writing Quality**: - Clear structure progressing from motivation to theory to applications - One-dimensional example (Section 1.1) provides intuitive entry point - Complete appendices including noise transformation formulas and Mosco convergence background ### Weaknesses 1. **Limited Application Scope**: - Lacks numerical experiments verifying theoretical predictions (while a mathematics paper, simple examples would strengthen persuasiveness) - Block-diagonal assumption limits applicability to real systems (many physical systems have coupled noise) - Boundary conditions and constrained systems not discussed 2. **Technical Complexity**: - Covariant derivatives and Christoffel symbol calculations may be barriers for non-specialists - High abstraction of Mosco convergence, insufficient connection to physical intuition - Motivation for some proof steps (e.g., Proposition 2.3) could be clearer 3. **Insufficient Dialogue with Physics Literature**: - Graham's work [Gra77, Gra85], though cited, lacks deep comparative analysis - Specific connections to fluctuating hydrodynamics [MRZ25] only briefly mentioned - Missing links to experimentally observable quantities (e.g., entropy production measurements) 4. **Discussion of Open Problems**: - For σ not satisfying Klimontovich condition, do other noise interpretations preserve coarse-graining? - Does non-block-diagonal case admit partial preservation? - How does reversibility behave under discrete-time approximation? ### Impact Assessment 1. **Contribution to Mathematics**: - Provides new tools for intersection of stochastic analysis and Riemannian geometry (row covariant divergence) - Extends application of Dirichlet forms to multiscale problems - May inspire research on general tensor field divergence theory 2. **Contribution to Statistical Mechanics**: - Clarifies thermodynamic meaning of different noise interpretations - Provides new geometric perspective for non-equilibrium theory - Mathematical support for naturalness of Klimontovich noise 3. **Potential Impact on Sampling Algorithms**: - Can guide design of preconditioned Langevin algorithms preserving reversibility - Multiscale sampling can exploit coarse-graining preservation - Provides theoretical foundation for adaptive MCMC 4. **Practical Value**: - For applications requiring reversible dynamics design (molecular dynamics, materials science), provides clear design criteria - Coarse-graining theory applicable to complex system dimensionality reduction - Data-driven methods [Zhu+25] can leverage this framework 5. **Reproducibility**: - Pure theoretical results with high reproducibility - All theorems have proofs, enabling independent verification - Sufficient citations for tracing background ### Applicable Scenarios 1. **Molecular Dynamics Simulation**: Design Langevin thermostats preserving Gibbs distribution, especially for non-uniform temperature fields. 2. **Bayesian Sampling**: Design efficient MCMC algorithms on complex posterior distributions, accelerating convergence through σ optimization. 3. **Multiscale Modeling**: Derive macroscopic deterministic or stochastic equations from microscopic random models while preserving thermodynamic consistency. 4. **Active Matter**: Model fluctuating hydrodynamics of self-propelled particle systems, preserving energy-momentum conservation. 5. **Machine Learning**: Theoretical analysis of stochastic optimization algorithms, understanding effects of different discretization schemes on convergence. 6. **Mathematical Finance**: Design stochastic volatility models preserving no-arbitrage conditions. ## Key References 1. **[BGL13]** Bakry, Gentil, Ledoux. *Analysis and geometry of Markov diffusion operators*. Springer, 2013. (Classical reference for reversible diffusions) 2. **[Pav14]** Pavliotis. *Stochastic processes and applications*. Springer, 2014. (Stochastic processes textbook, frequently cited in this work) 3. **[KS03]** Kuwae & Shioya. "Convergence of spectral structures". *Comm. Anal. Geom.* 2003. (Foundation of Mosco convergence theory) 4. **[Gra77]** Graham. "Covariant formulation of non-equilibrium statistical thermodynamics". *Z. Phys. B* 1977. (First covariant Fokker-Planck equation) 5. **[Kha68]** Khasminskij. "On the principle of averaging". *Kybernetika* 1968. (Classical averaging principle) 6. **[CP23]** Costa & Pavliotis. "The entropy production of stationary diffusions". *J. Phys. A* 2023. (Modern entropy production theory) 7. **[Lel+25]** Lelièvre et al. "Optimizing the diffusion coefficient". *Math. Comp.* 2025. (Diffusion matrix optimization) 8. **[Dió24]** Diósi. "The covariant Langevin equation". *Rep. Math. Phys.* 2024. (Prior work generalized in this paper) --- **Overall Assessment**: This is a high-quality theoretical mathematics paper making substantial contributions at the intersection of stochastic analysis and Riemannian geometry. The main innovations are unified characterization of reversibility conditions under different noise interpretations and first proof of Klimontovich noise preservation under coarse-graining. The theory is rigorous, the framework elegant, with potential profound impact on statistical mechanics and sampling algorithms. Main limitations are restricted application scope (block-diagonal assumption) and lack of numerical verification. Recommended future work includes numerical implementation, coupled noise generalization, and experimental connections. The paper is suitable for publication in top-tier journals in probability theory or mathematical physics.