🔢 mathematics

Two-sided estimates of Lyapunov exponents for Milstein schemes of non-autonomous stochastic differential equations

This paper establishes sharp two-sided estimates for the Lyapunov exponents of Milstein schemes applied to non-autonomous stochastic differential equations, demonstrating that the stabilizing effect of multiplicative noise observed in continuous systems is preserved in the discretized setting.

Original authors: Vu Thi Hue, Bao Quoc Tang, Do Duc Thuan

Published 2026-08-11

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Vu Thi Hue, Bao Quoc Tang, Do Duc Thuan

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a giant, chaotic dance floor where everything is constantly jiggling. In the world of physics and engineering, we often try to predict how things move using neat, tidy equations. But in reality, there's always a little bit of "static" or random noise—like a sudden gust of wind or a random bump—that throws a wrench in the works. This is the realm of Stochastic Differential Equations (SDEs). Think of them as the rulebook for how a system behaves when it's being nudged by random forces, like a leaf floating down a river that's also being tossed by a storm.

Usually, we think of this random noise as a nuisance. If a bridge is wobbling, we want the wind to stop, not blow harder. But here's the twist: sometimes, adding more noise actually stops the wobbling! It sounds like magic, but it's a real phenomenon called stochastic stabilization. Imagine a spinning top that's about to fall over. If you shake the table just right, the vibration might actually keep the top spinning upright longer than if the table were perfectly still. Scientists call the measure of how fast something grows or shrinks over time its Lyapunov exponent. If this number is negative, the system is stable (calm); if it's positive, it's unstable (chaotic).

Now, here's the catch: we can't solve these complex equations with a pencil and paper for real-world problems. We have to use computers to approximate the answers, breaking time into tiny, discrete steps. This is where numerical schemes come in. They are like taking a smooth video of the dancing top and turning it into a flipbook of individual frames. The big question is: when we turn the smooth, continuous world into a choppy, digital one, do we accidentally break the magic? Does the computer simulation still show that the noise is stabilizing the system, or does the digital "pixelation" mess it up?

This paper, written by Vu Thi Hue, Quoc Bao Tang, and Do Duc Thuan, dives deep into this exact problem. They focus on a specific, high-precision method for creating those digital flipbooks called the Milstein scheme. While other methods exist, the Milstein scheme is special because it pays extra attention to the "second-order" effects of the noise—the subtle, curved ways the randomness interacts with the system. The authors wanted to know: if we use this fancy Milstein method, can we prove mathematically that it preserves the stabilizing effect of the noise, just like the real, continuous world does?

The team didn't just guess; they built a rigorous mathematical fortress around the problem. They tackled two main scenarios: linear equations (where the rules are straightforward and proportional) and nonlinear equations (where the rules get messy and the system reacts in complex, unpredictable ways). Their goal was to find "two-sided estimates" for the Lyapunov exponents. Think of this as trying to trap a wild animal between two fences. They wanted to prove that the true value of the stability lies strictly between a calculated upper fence and a lower fence.

What they found is quite reassuring for anyone building simulations. They proved that for the Milstein scheme, the "digital" Lyapunov exponents stay very close to the "real" ones, provided the time steps (the frames in the flipbook) are small enough. Specifically, they showed that the noise-induced stabilization isn't just a fluke of the continuous world; it survives the transition to the digital world. They derived precise formulas that act as upper and lower bounds, effectively saying, "No matter how you run this simulation, the stability will be at least this good and no worse than this."

One of their clever tricks involved looking at the specific "second-order" term in the Milstein equation. This term is like a hidden safety net that catches errors other methods might miss. By exploiting this term, they could use advanced math (Taylor expansions) to bound the behavior of logarithmic functions, which are the tools used to measure growth rates. They showed that even when the system is pushed to its limits, the Milstein scheme respects the physics of the situation.

The paper also ventured into the wilder territory of nonlinear equations, where the noise doesn't just push the system but changes the rules of the game itself. Even here, they managed to construct similar safety fences, proving that the stabilization effect holds up, though the math gets significantly more intricate. They didn't just stop at theory, though. They ran computer simulations to visualize their findings. In one example, they took a system that was naturally unstable (like a bridge about to collapse) and showed that with a small amount of noise, it stayed shaky. But when they cranked up the noise intensity in their simulation, the system suddenly calmed down and decayed to zero, exactly as their formulas predicted.

In short, this paper provides a strong mathematical guarantee that the Milstein scheme is a "structurally preserving" tool. It means that when engineers and scientists use this method to model complex, noisy systems—from financial markets to biological cells—they can trust that the digital model won't lie to them about whether the system is stable or chaotic. The magic of noise-induced stabilization isn't lost in the translation from the smooth, continuous world to the pixelated, digital one; it's right there, captured in the math.

Technical Summary: Two-sided estimates of Lyapunov exponents for Milstein schemes of non-autonomous stochastic differential equations

Problem Formulation
The paper addresses the preservation of stochastic stabilisation effects when moving from the continuum setting of stochastic differential equations (SDEs) to their discretised numerical approximations. While the stabilising effect of multiplicative noise on SDEs is a well-established phenomenon (citing foundational works such as [Kap51], [Kha69], and [ACW83]), it is not guaranteed that numerical schemes preserve the almost sure exponential stability or instability of the exact solution. Previous research has largely focused on establishing upper bounds for the Lyapunov exponents of numerical schemes (e.g., Euler-Maruyama and Milstein) to demonstrate stability preservation. However, obtaining sharp lower bounds, which are necessary to precisely characterize the large-time behavior and confirm that the discretised system does not artificially dampen or amplify dynamics, has remained an open challenge.

The authors investigate this problem for both linear and nonlinear non-autonomous SDEs using the Milstein scheme. Specifically, they consider:

Linear non-autonomous SDEs: $dX(t) = a(t)X(t)dt + \alpha X(t)dW(t)$ .
Nonlinear non-autonomous SDEs: $dX(t) = a(t)X(t)dt + f(t, X(t))dW(t)$.

The primary objective is to derive two-sided (upper and lower) estimates for the Lyapunov exponents of the Milstein approximation, thereby determining if the numerical scheme preserves the exact stability region of the continuous system.

Methodology
The core of the analysis relies on exploiting the specific structure of the Milstein scheme, particularly its second-order term involving the stochastic noise, to obtain precise estimates for Taylor expansions of logarithmic and power functions.

Recursive Representation: The authors express the Milstein approximation $X_h(t_n)$ as a product of independent random variables. By taking logarithms, the Lyapunov exponent is transformed into a sum involving terms of the form $\log((1+\Delta A_k)^2 + N_k)$ , where $N_k$ is a random process derived from the noise increments and the scheme's coefficients.
Truncation of Taylor Expansions: A critical technical step involves bounding the functions $f_{\eta,r}(x) = (\eta+x)^r$ $f_{η, r} (x) = (η + x)^{r}$ and $g_\rho(x) = \log(\rho+x)$ $g_{ρ} (x) = lo g (ρ + x)$ from both above and below.
- Standard Taylor expansions provide upper bounds.
- Lower bounds are more challenging due to singularities near zero. The authors employ a truncation technique (Lemma 2.3) that restricts the domain of the expansion variable to avoid the singularity, utilizing a parameter $m > 1$ to establish valid lower bounds for $x$ in the range $(-m^{-1}m\eta, 0)$ .
Lower Bounds on Random Processes: To apply the lower bound truncation, the authors prove that the random process $N_k$ (and its nonlinear counterpart) is bounded from below almost surely by a negative constant times $(1+\Delta A_k)^2$ . This is achieved by analyzing the quadratic structure of the Milstein noise term and ensuring the step size $h$ is sufficiently small.
Expectation Estimates: The proofs involve calculating precise expectations of powers of the noise terms (e.g., $E[N_k^j]$ ) and remainder terms in the Taylor expansions. The authors utilize the properties of Brownian motion increments and the Kolmogorov strong law of large numbers to handle the asymptotic behavior of the sums.

Key Contributions and Results

Linear Non-Autonomous Equations:
- Theorem 1.1: Provides sharp upper and lower bounds for the almost sure Lyapunov exponent of the Milstein scheme. The bounds depend on the time-averaged drift $\Delta A_k$ and the noise intensity $\alpha$ . The lower bound includes a term involving the lim inf of the inverse squared drift, distinguishing it from previous upper-bound-only results.
- Theorem 1.2: Extends these results to $p$ -moment stability for $0 < p < 1$ , providing two-sided estimates for the growth rate of $E[|X_h(t_n)|^p]$ .
- Recovery of Continuum Limits: The authors demonstrate that as the step size $h \to 0$ , these discrete estimates converge to the known Lyapunov exponents of the continuous SDE, confirming the consistency of the method.
Nonlinear Non-Autonomous Equations:
- Under Assumption (F) regarding the growth and regularity of the diffusion coefficient $f(t, x)$ , the authors derive analogous results.
- Theorem 1.3: Establishes two-sided bounds for the almost sure Lyapunov exponent of the nonlinear Milstein scheme, involving the bounds $f_*$ and $f^*$ of the function $f(t,x)/x$ .
- Theorem 1.4: Provides corresponding $p$ -moment estimates for the nonlinear case.
Numerical Simulations:
- The paper includes numerical experiments validating the theoretical findings. Simulations on linear equations with bounded and unbounded coefficients, as well as a nonlinear equation, illustrate how increasing noise intensity stabilises unstable deterministic systems. The plots confirm that the Milstein scheme correctly captures the transition from instability to stability as predicted by the derived Lyapunov exponent bounds.

Significance and Claims
The paper claims to advance the field of structural-preserving numerical schemes by moving beyond one-sided (upper) estimates to provide sharp two-sided estimates for Lyapunov exponents. This is significant because:

It rigorously confirms that the Milstein scheme preserves the stochastic stabilisation effect from the continuum to the discretised setting for both linear and nonlinear non-autonomous systems.
It addresses a gap in the literature where lower bounds for Lyapunov exponents in discretised settings were previously unavailable or not well-posed.
The methodology, specifically the exploitation of the second-order noise term in the Milstein scheme to derive precise Taylor expansion bounds, is presented as a robust approach applicable to other problems.

The authors modestly note that while they provide explicit bounds for the Milstein scheme, the optimal bound for the Lyapunov exponent of the continuous nonlinear SDE itself remains an open problem due to implicit nonlinearities. Consequently, their results offer the sharpest available characterization for the discretised system within the current theoretical framework.

More like this