Uniform-in-time diffusion approximations for multiscale stochastic systems

1. Problem Statement

The paper addresses the rigorous mathematical analysis of fully coupled multiscale stochastic differential equations (SDEs) of the form:
$$\begin{cases} dX^\varepsilon_t = \varepsilon^{-2}b(X^\varepsilon_t, Y^\varepsilon_t)dt + \varepsilon^{-1}c(X^\varepsilon_t, Y^\varepsilon_t)dt + \sqrt{2}\varepsilon^{-1}\sigma(X^\varepsilon_t, Y^\varepsilon_t)dW_t, \\ dY^\varepsilon_t = F(X^\varepsilon_t, Y^\varepsilon_t)dt + \varepsilon^{-1}H(X^\varepsilon_t, Y^\varepsilon_t)dt + \sqrt{2}G(X^\varepsilon_t, Y^\varepsilon_t)dW_t, \end{cases}$$
where $X^\varepsilon$ represents the fast component (evolving on the $\varepsilon^{-2}$ scale) and $Y^\varepsilon$ the slow component. The parameter $\varepsilon \to 0$ quantifies the time-scale separation.

Key Limitations of Existing Theory:

1. Finite-Time Horizon: Most existing results only prove convergence on finite intervals $[0, T]$, failing to characterize long-time behavior or stationary distributions.
2. Marginal Focus: Prior work typically analyzes only the marginal law of the slow variable $Y^\varepsilon$, ignoring the joint law $(X^\varepsilon, Y^\varepsilon)$, which is crucial for thermodynamic observables (e.g., energy, entropy).
3. Restrictive Assumptions: Classical theories often require bounded coefficients, uniform ellipticity, and high regularity, excluding important physical models like degenerate Langevin systems or those with distributional (irregular) drifts.

The central question, highlighted by Freidlin, is: How does the stationary distribution of the original multiscale system relate to that of the homogenized limit, and do the limits $\varepsilon \to 0$ and $t \to \infty$ commute?

2. Methodology

The authors develop a novel framework based on quantitative error decomposition and weighted functional analysis.

A. Functional Framework

• Weighted Hölder Spaces: To handle unbounded coefficients and moment growth, the authors introduce anisotropic weighted Hölder spaces $C^\alpha_\rho C^\beta_\omega$. These spaces allow coefficients to grow polynomially while maintaining control over regularity.
• Polynomial Weights: They utilize a class of weights $W_N$ that are closed under algebraic operations and satisfy local translation estimates, essential for handling non-compact state spaces.

B. Analytic Tools

• Parameter-Dependent Poisson Equations: The core of the proof relies on solving Poisson equations for the fast dynamics with the slow variable as a parameter. The authors establish regularity estimates for the solution $\Phi(x, y)$ with respect to the parameter $y$, showing that if coefficients are Hölder continuous, the corrector inherits specific regularity.
• Infinite-Horizon Kolmogorov Equations: They derive long-time decay estimates for the semigroup associated with the frozen fast dynamics. Crucially, they prove that algebraic (polynomial) mixing rates are sufficient for uniform-in-time convergence, relaxing the need for exponential mixing often required in previous literature.
• Error Decomposition: The joint law error is decomposed into three parts:
  1. Slow Motion Error: Convergence of the slow variable to the homogenized limit.
  2. Conditional Equilibrium Error: The deviation of the fast variable from its local invariant measure $\mu_y$ conditioned on the slow variable.
  3. Initial Layer: A transient term capturing the rapid relaxation of the fast variable at $t \sim O(\varepsilon^2)$.

3. Key Contributions

1. Quantitative Uniform-in-Time Joint Law Approximation

The paper establishes that the joint law of $(X^\varepsilon_t, Y^\varepsilon_t)$ converges uniformly in time to a skew-product measure:
$$\text{Law}(X^\varepsilon_t, Y^\varepsilon_t) \approx \mu_{\bar{Y}_t}(dx) \otimes \text{Law}(\bar{Y}_t)(dy)$$
where $\mu_y$ is the conditional invariant measure of the fast process given the slow state $y$, and $\bar{Y}_t$ is the homogenized slow process.

• Convergence Rate: The error bound is of order $O(\varepsilon^\beta)$ (depending on the regularity of coefficients in the slow variable) plus a transient term $\ell_0(t/\varepsilon^2)$ describing the initial relaxation.
• Generality: The result holds for unbounded, irregular coefficients, degenerate structures, and weakly mixing dynamics.

2. Commutativity of Limits and Stationary Measures

The authors prove that for a broad class of observables, the limits commute:
$$\lim_{\varepsilon \to 0} \lim_{t \to \infty} \mathbb{E}[\phi(Z^\varepsilon_t)] = \lim_{t \to \infty} \lim_{\varepsilon \to 0} \mathbb{E}[\phi(Z^\varepsilon_t)]$$
Consequently, the invariant measure $\nu_\varepsilon$ of the original system converges weakly to the skew-product $\mu_y(dx)\bar{\mu}(dy)$, where $\bar{\mu}$ is the invariant measure of the homogenized equation. This provides a rigorous answer to Freidlin's question.

3. Relaxation of Regularity Assumptions

The framework accommodates:

• Unbounded Coefficients: Via polynomial weights.
• Degenerate Diffusion: The full system need not be non-degenerate; ergodicity can be transferred from the fast to the slow component via corrector terms.
• Distributional Drifts: In the periodic homogenization application, the fast drift can be a distribution with negative Hölder regularity ($\alpha \in (-1, -1/2)$), provided it is divergence-free.

4. Main Results and Applications

Application I: Uniform Averaging Principle

For classical fast-slow systems ($c \equiv H \equiv 0$), the authors derive a uniform-in-time averaging principle that improves upon finite-time results by allowing unbounded coefficients and weak mixing rates.

Application II: Uniform Smoluchowski–Kramers (S-K) Approximation

The paper applies the theory to the degenerate Langevin system with state-dependent friction $A(y)$:
$$\varepsilon^2 \ddot{Y}^\varepsilon_t = -\nabla U(Y^\varepsilon_t) - A(Y^\varepsilon_t)\dot{Y}^\varepsilon_t + \sqrt{2}\sigma(Y^\varepsilon_t)\dot{W}_t$$

• Result: Proves uniform convergence of the joint position-scaled velocity law.
• Thermodynamic Implications:
  • Energy: Derives a generalized equipartition theorem where kinetic energy is regulated by the friction matrix via a Lyapunov equation.
  • Entropy Production: Shows that the macroscopic rescaled dissipation becomes independent of the non-symmetric friction matrix, depending only on noise intensity.
  • Free Energy: Identifies a Kullback–Leibler divergence term arising from the renormalization of fast degrees of freedom into the macroscopic free energy landscape.

Application III: Periodic Homogenization with Distributional Drifts

The authors tackle the singular regime where the fast drift $b$ is a distribution (negative Hölder regularity).

• Method: They employ a periodic Zvonkin transformation to algebraically eliminate the singularity, converting the problem into a regular Hölder framework.
• Result: Establishes the first uniform-in-time homogenization result for SDEs with distributional drifts, achieving an $O(\varepsilon)$ convergence rate.

5. Significance

• Theoretical Breakthrough: This work resolves the long-standing open problem of characterizing the joint stationary distribution of multiscale systems and proving the commutativity of limits.
• Physical Relevance: By providing uniform-in-time error bounds, the results validate the use of reduced (homogenized) models for long-time prediction and steady-state thermodynamics, which is critical for applications in molecular dynamics, climate science, and financial mathematics.
• Robustness: The ability to handle degenerate and irregular coefficients significantly broadens the scope of applicable physical models, moving beyond idealized assumptions of uniform ellipticity and boundedness.
• Numerical Implications: The rigorous justification of the skew-product structure supports the development of efficient multiscale numerical methods (e.g., Heterogeneous Multiscale Methods) for stationary regimes.

In summary, this paper provides a comprehensive, quantitative, and uniform-in-time theory for multiscale stochastic systems, bridging the gap between transient dynamics and stationary thermodynamics under highly general and physically realistic conditions.