Strong and weak convergence rates for slow-fast system driven by multiplicative Lévy noises

This paper establishes optimal strong and weak convergence rates for slow-fast systems driven by multiplicative $\alpha$-stable Lévy noises by overcoming challenges in exponential ergodicity and gradient estimates through coupling and spatial periodic methods, while also deriving explicit formulas for tangent maps induced by nonlinear immersions.

The Gist

Imagine you are trying to predict the path of a hiker (the "slow" system) walking through a dense, chaotic forest.

In this forest, the hiker is being pushed around by two forces:

1. The Wind (Slow Noise): A gentle, steady breeze that changes direction slowly.
2. The Swarm of Bees (Fast Noise): A frantic, hyper-active swarm of bees buzzing around the hiker at incredible speeds, constantly bumping into them and changing their direction instantly.

In the real world, the hiker's path is a mess of slow drifts and sudden, jarring jumps caused by the bees. It's impossible to write a simple formula to predict exactly where the hiker will be in an hour because the bees are too chaotic.

The Big Idea: The "Average" Forest
Mathematicians have a trick called the Averaging Principle. Instead of tracking every single bee, they ask: "If the bees are so fast, what is their average effect on the hiker?"

They create a new, imaginary "Average Forest" where the chaotic bees are replaced by a smooth, gentle wind that represents the average push of the swarm. The hiker in this imaginary forest moves much more smoothly.

The Problem with "Standard" Math
Most previous studies assumed the bees were just random bumps (like standard Brownian motion). But in this paper, the authors look at a more dangerous scenario: Multiplicative Jump Noise.

Think of it this way:

• Standard Noise: The bees bump the hiker with the same force, no matter where the hiker is.
• Multiplicative Noise (This Paper): The bees are smarter. If the hiker is near a cliff, the bees push harder. If the hiker is in a valley, they push softer. The "bumpiness" depends on the hiker's current location.

This makes the math incredibly hard. It's like trying to predict the path of a surfer where the size of the waves changes depending on exactly where the surfer is standing on the board.

What This Paper Achieves
The authors, Qiu-Chen Yang and Kun Yin, managed to solve this difficult puzzle. They proved exactly how fast the "Real Hiker" (with the crazy bees) converges to the "Average Hiker" (with the smooth wind) as the bees get faster and faster.

They found two types of convergence rates:

1. Strong Convergence (The "Path" Match):

   • Analogy: If you watch a video of the Real Hiker and the Average Hiker side-by-side, how close are their paths at every single second?
   • The Result: They found a specific speed limit for this match. It depends on how "jumpy" the bees are (a parameter called $\alpha$). The faster the bees jump, the slower the paths match up perfectly. They derived a precise formula for this speed.

2. Weak Convergence (The "Outcome" Match):

   • Analogy: You don't care about the exact path. You just want to know: "What is the probability the hiker ends up in the North field vs. the South field?"
   • The Result: This is easier to predict. Even if the paths wiggle differently, the final destination statistics match up very quickly. They proved this happens at the fastest possible speed (Order 1).

The Secret Weapons Used
To solve this, the authors had to invent some new mathematical tools:

• The "Coupling" Method: Imagine taking two hikers starting at different spots in the forest and magically forcing them to walk together as fast as possible. If they can be forced to meet quickly, it proves the system is stable. They used this to show the "fast bees" settle down quickly.
• The "Heat Kernel" Expansion: This is like analyzing the "fingerprint" of the bees' movement. By looking at the microscopic details of how the bees jump, they could predict the smooth, large-scale behavior.
• The "Tangent Map" (The Geometry Trick): The bees move on a sphere (like directions on a globe). The authors had to figure out how to stretch and squeeze this sphere mathematically without tearing it. They derived a new formula for how these shapes bend, which was crucial for proving the math works.

Why Does This Matter?
This isn't just about hikers and bees. These "Slow-Fast" systems appear everywhere:

• Finance: A stock price (slow) reacting to high-frequency trading algorithms (fast jumps).
• Chemistry: A slow chemical reaction driven by fast, random molecular collisions.
• Physics: Particles moving through a fluid with turbulent, jumping eddies.

The Bottom Line
This paper takes a very messy, unpredictable real-world problem (where the chaos depends on your location) and gives us a precise, reliable way to simplify it. It tells engineers and scientists exactly how much error they will make if they ignore the chaos and just use the "average" model, and it proves that for many practical purposes, the average model is an excellent approximation.

Technical Summary

Here is a detailed technical summary of the paper "Strong and Weak Convergence Rates for Slow-Fast System Driven by Multiplicative Lévy Noises" by Qiu-Chen Yang and Kun Yin.

1. Problem Statement

The paper investigates the averaging principle for multiscale slow-fast stochastic differential equations (SDEs) driven by multiplicative $\alpha$-stable Lévy noises.

The system is modeled by the following equations where $\varepsilon \to 0$ represents the time-scale separation:
$$\begin{cases} dX^\varepsilon_t = b(t, X^\varepsilon_t, Y^\varepsilon_t)dt + \delta_1(t, X^\varepsilon_t, Y^\varepsilon_t)dL^1_t, & X^\varepsilon_0 = x \\ dY^\varepsilon_t = \frac{1}{\varepsilon}f(X^\varepsilon_t, Y^\varepsilon_t)dt + \frac{1}{\varepsilon^{1/\alpha_2}}\delta_2(X^\varepsilon_t, Y^\varepsilon_t)dL^2_t, & Y^\varepsilon_0 = y \end{cases}$$
Key Challenges:

• Multiplicative Noise: Unlike previous studies that assumed additive noise (constant diffusion coefficients), this model features jump coefficients $\delta_1$ and $\delta_2$. This introduces significant difficulties in proving the existence of invariant measures, establishing exponential ergodicity, and deriving crucial gradient estimates for the "frozen" fast process.
• Heavy-Tailed Noise: The driving noises $L^1_t$ and $L^2_t$ are isotropic $\alpha$-stable processes ($1 < \alpha_1, \alpha_2 < 2$), which lack finite second moments and require non-local operator techniques.
• Regularity: The coefficients depend on time and space, requiring specific Hölder regularity assumptions to derive convergence rates.

2. Methodology

The authors employ a combination of probabilistic and analytic techniques to overcome the challenges posed by multiplicative jumps:

• Two Approaches for Exponential Ergodicity:

  1. Coupling Method: Used for the general (non-periodic) case. The authors construct a Markov coupling operator for the non-local generator of the frozen fast process. They utilize a Lyapunov-type function and the partially dissipative condition of the drift $f$ to prove exponential contractivity in the $L^p$-Wasserstein distance.
  2. Spatial Periodic Method: Used for the periodic case. By imposing compactness and spatial periodicity on coefficients, they project the process onto a compact quotient space ($\mathbb{R}^{d_2}/\Lambda$). They apply Doeblin's theorem by establishing a lower bound on the transition probability density.

• Heat Kernel Asymptotic Expansion:
  To handle the multiplicative noise $\delta_2$, the authors derive the existence and regularity of the transition probability density $p(t, y, Y)$ for the frozen process. They use asymptotic expansions of the heat kernel relative to a process with constant coefficients to obtain necessary gradient estimates ($\nabla_y p$). This is critical for bounding the martingale terms in the strong convergence analysis.

• Non-local Poisson Equations (Corrector Equations):
  To quantify the convergence rates, the authors construct corrector equations (non-local Poisson equations) of the form:
  $$\mathcal{L}_2 u(t, x, y) + (g(t, x, y) - \bar{g}(t, x)) = 0$$
  where $\mathcal{L}_2$ is the generator of the frozen fast process. They establish regularity estimates (boundedness and gradient bounds) for the solution $u$.

• Mollification:
  Since the coefficients may not be smooth enough for direct application of Itô's formula, the authors use mollification techniques to approximate the coefficients and derive error bounds dependent on the Hölder regularity exponent $v$.

3. Key Contributions

1. Extension to Multiplicative Jump Coefficients:
   This is the first work to establish convergence rates for slow-fast systems with multiplicative $\alpha$-stable noises. Previous works were limited to additive noise or constant diffusion coefficients. The paper resolves the technical hurdle of deriving invariant measures and ergodicity when the jump intensity depends on the state.

2. Optimal Strong Convergence Rate:
   Under sufficient Hölder regularity ($v$) of the time-dependent coefficients, the authors prove the optimal strong convergence rate of order:
   $$\varepsilon^{1 - \frac{1}{\alpha_2}}$$
   (Specifically, the rate is $\varepsilon^{p(1 - \frac{1}{\alpha_2})}$ in $L^p$ norm when $v \ge 1$). This matches the optimal rate found in additive noise models but is achieved under much more complex conditions.

3. Optimal Weak Convergence Rate:
   The paper establishes a weak convergence rate of order 1 (i.e., $O(\varepsilon)$), which is optimal. This is derived using Kolmogorov equations and the regularity of the corrector solution.

4. Geometric Analysis of Tangent Maps:
   A novel technical contribution is the derivation of an explicit formula for the tangent map and its Jacobian determinant induced by a nonlinear immersion on the sphere $S^{d-1}$. This geometric result is essential for transforming the spherical measure in the presence of multiplicative noise and proving the uniform ellipticity of the transition density.

5. Dual Ergodicity Proofs:
   The paper provides two distinct proofs for the exponential ergodicity of the frozen process: one via coupling (for general dissipative systems) and one via spatial periodicity (for periodic systems), broadening the applicability of the averaging principle.

4. Main Results

• Theorem 2.1 (Strong Convergence):
  For $p \in [1, \alpha_1 \wedge \alpha_2)$, assuming $\delta_1$ is time-dependent only (additive in the slow component for the strong rate proof) and coefficients have regularity $v$:
  $$\mathbb{E}\left[ \sup_{t \in [0,T]} |X^\varepsilon_t - \bar{X}_t|^p \right] \le C \varepsilon^{p \left[ (\frac{v}{\alpha_2}) \wedge (1 - \frac{1 \vee (\alpha_1 - v)}{\alpha_2}) \right]}$$
  When $v \ge 1$, this simplifies to the optimal order $\varepsilon^{1 - 1/\alpha_2}$.

• Theorem 2.2 (Weak Convergence):
  For test functions $\phi \in C^{2+\gamma}_b$, the weak error satisfies:
  $$\sup_{t \in [0,T]} |\mathbb{E}\phi(X^\varepsilon_t) - \mathbb{E}\phi(\bar{X}_t)| \le C \left( \varepsilon^{\frac{v}{\alpha_2}} + \varepsilon^{1 - \frac{\alpha_1 - v}{\alpha_2}} \right)$$
  When $v = \alpha_1 = \alpha_2$, the rate is $O(\varepsilon)$.

• Theorem 2.3 (Exponential Contractivity):
  The frozen equation satisfies exponential contractivity in the $L^p$-Wasserstein distance:
  $$W_p(\delta_{y_1}P_t, \delta_{y_2}P_t) \le C e^{-\beta t} |y_1 - y_2|$$
  This holds under the partially dissipative condition using the coupling method.

• Theorem 2.4 (Tangent Map):
  Provides the explicit Jacobian determinant for the map $F(\hat{\omega}) = \frac{A^{-1}\hat{\omega}}{|A^{-1}\hat{\omega}|}$, showing $|J_F(\hat{\omega})| = \frac{|\det(A^{-1})|}{|A^{-1}\hat{\omega}|^d}$. This is crucial for the change of variables in the Lévy measure analysis.

5. Significance

• Theoretical Advancement: The paper bridges a gap in the literature by moving from additive to multiplicative Lévy noise in multiscale systems. This is a significant step toward modeling real-world phenomena where noise intensity often depends on the system state (e.g., in finance or physics).
• Methodological Rigor: The derivation of gradient estimates via heat kernel asymptotics for multiplicative stable processes is a sophisticated contribution that enables the proof of strong convergence rates, which are notoriously difficult in the presence of jumps.
• Practical Implications: The established convergence rates ($1 - 1/\alpha_2$for strong,$1$ for weak) provide quantitative error bounds for numerical simulations and model reduction techniques (averaging) in systems driven by heavy-tailed noise, guiding the choice of time steps and approximation schemes.
• Geometric Insight: The explicit calculation of the Jacobian for the nonlinear immersion on spheres offers a tool for future research involving state-dependent jump measures and geometric stochastic analysis.

In summary, this paper successfully extends the averaging principle to a more complex and realistic class of stochastic systems, providing rigorous proofs for optimal convergence rates and introducing new geometric tools to handle multiplicative jump coefficients.