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

The paper investigates the preservation of stochastic stabilisation in the Milstein numerical scheme for non-autonomous stochastic differential equations (SDEs). The authors establish sharp two-sided estimates for Lyapunov exponents, which describe the large-time behaviour of dynamical systems, in both the almost sure and p-moment senses.

The paper begins by considering the linear non-autonomous SDE:

dX(t) = a(t)X(t)dt + αX(t)dW(t),

with the Milstein approximation defined by the stochastic difference equation:

Xh(tn+1) = Xh(tn)[1 + ΔAn - (α2/2)h + αΔWn + (α2/2)(ΔWn)2].

The main results for this linear case are presented in Theorems 1.1 and 1.2. Theorem 1.1 provides almost sure estimates for the Lyapunov exponent, giving both an upper bound:

lim sup (1/tn) logXh(tn) ≤ lim sup (1/n) Σ log1+ΔAk/h + (α2/2) lim sup (1/n) Σ 1/(1+ΔAk)2 - α2 lim inf (1/n) Σ 1/(1+ΔAk)2 + O(h),

and a lower bound:

lim inf (1/tn) logXh(tn) ≥ lim inf (1/n) Σ log1+ΔAk/h + (α2/2) lim inf (1/n) Σ 1/(1+ΔAk)2 - α2 lim sup (1/n) Σ 1/(1+ΔAk)2 + O(h(1/2)).

Theorem 1.2 provides analogous p-moment estimates for 0 < p < 1, with upper and lower bounds for lim sup and lim inf of (1/tn) log EXh(tn) p, respectively.

The paper then extends these results to nonlinear non-autonomous SDEs of the form:

dX(t) = a(t)X(t)dt + f(t, X(t))dW(t),

under Assumption (F) on the function f. The Milstein scheme for this nonlinear problem is given by:

Xh(tn+1) = Xh(tn)[1 + ΔAn + γ1n ΔWn + (1/2)γ2n((ΔWn)2 - h)],

where γ1n = f(tn, Xh(tn))/Xh(tn) and γ2n = f(tn, Xh(tn))∂x f(tn, Xh(tn))/Xh(tn).

Theorems 1.3 and 1.4 provide almost sure and p-moment estimates for the Lyapunov exponents of this nonlinear scheme, respectively. For instance, Theorem 1.3 gives the upper bound:

lim sup (1/tn) logXh(tn) ≤ lim sup (1/n) Σ log1+ΔAk/h + (f2/2) lim sup (1/n) Σ 1/(1+ΔAk)2 - f2 lim inf (1/n) Σ 1/(1+ΔAk)2 + O(h),

and the lower bound:

lim inf (1/tn) logXh(tn) ≥ lim inf (1/n) Σ log1+ΔAk/h + (f2/2) lim inf (1/n) Σ 1/(1+ΔAk)2 - f2 lim sup (1/n) Σ 1/(1+ΔAk)2 + O(h(1/2)).

The key ideas of the analysis include:

1. Exploiting the second-order stochastic noise term from the Milstein scheme to obtain precise estimates for Taylor expansions of logarithmic and power functions.

2. Establishing lower bounds for the random process Nk (defined as Nk = 2(1+ΔAk)(Yk - (α2/2)h) + (Yk - (α2/2)h)2 for the linear case) using Lemma 3.2, which shows Nk > -((m-1)/m)(1+ΔAk)2 almost surely for a suitable m > 1.

3. Computing precise estimates for the expectations E(Nk j) for j = 1, 2, 3 in Lemma 3.1.

4. Using a truncation technique for Taylor expansions (Lemma 2.3) to bound logarithmic and power functions from both above and below.

5. Estimating the expectations of remainder terms ξρ(Nk) and Ψρ(Nk) in Lemmas 3.3 and 3.4.

The paper also includes numerical simulations in Section 5 that illustrate the theoretical results, showing that sufficiently large noise intensities can stabilise unstable deterministic problems, and that the stabilisation effect is preserved in the discretised setting.

The authors conclude that their results demonstrate the preservation of stabilisation from the continuum setting to the discretised setting, and that the estimates recover the precise Lyapunov exponent for the continuum problem as the step size h tends to zero.

Improvements to AI Systems Based on This Paper:

1. Adaptive Step-Size Control for SDE Solvers

• Improvement: Integrate the sharp two-sided Lyapunov exponent estimates (Theorems 1.1–1.4) into an AI-driven numerical solver that automatically adjusts the Milstein step size h to preserve stochastic stabilisation.

• What the improved AI can do: Given a non-autonomous SDE with unknown drift and noise coefficients, the AI can predict whether the discretised solution will remain stable (e.g., 1 over t n X h(t n) < 0) and choose h to ensure the upper bound stays negative, preventing spurious instability in long-time simulations.

2. Early-Warning System for Loss of Stability in Stochastic Dynamical Systems

• Improvement: Use the lower-bound estimates (with O(h 1/2) terms) to train a classifier that detects when a numerical scheme is about to lose stabilisation (i.e., when the Lyapunov exponent crosses zero).

• What the improved AI can do: Monitor real-time simulation data from financial, biological, or engineering systems modelled by SDEs, and issue alerts before divergence occurs, even when the underlying drift is time-varying and only partially observed.

3. Hyperparameter Optimisation for Stochastic Neural Differential Equations

• Improvement: Apply the p-moment estimates (Theorem 1.2 and 1.4) to guide the selection of noise intensity and drift coefficients in neural SDEs used for generative modelling or reinforcement learning.

• What the improved AI can do: Automatically tune the noise schedule in a neural SDE to ensure both sample diversity (large p-moment growth) and long-term stability (bounded Lyapunov exponent), avoiding mode collapse or explosion during training.

4. Robustness Certification for Discretised Stochastic Controllers

• Improvement: Leverage the almost sure upper bound to certify that a feedback controller (encoded as a(t)) preserves stabilisation under discretisation, even with unknown but bounded noise terms f(t,x).

• What the improved AI can do: Given a control policy for a stochastic system, the AI can verify—before deployment—that the Milstein-discretised closed-loop system will not destabilise, and can recommend minimal noise intensities f* to guarantee stability margins.

5. Data-Efficient Estimation of Lyapunov Exponents from Noisy Trajectories

• Improvement: Use the explicit formulas for E(N k j) (Lemma 3.1) and the truncation technique (Lemma 2.3) to design a Bayesian or neural estimator that infers the true Lyapunov exponent from short, discretised trajectories.

• What the improved AI can do: From limited observations of a stochastic process, the AI can output a confidence interval for the true Lyapunov exponent, correcting for discretisation bias using the O(h) and O(h 1/2) corrections, enabling faster and more accurate stability classification in scientific discovery.

6. Automatic Theorem-Guided Code Generation for SDE Simulations

• Improvement: Embed the paper’s lemmas (e.g., lower bound on N k, Taylor truncation bounds) into a symbolic AI that generates provably stable Milstein solver code for arbitrary user-defined SDEs.

• What the improved AI can do: A user inputs a non-autonomous SDE; the AI outputs a custom solver with pre-computed step-size bounds and error estimates that guarantee the preservation of stochastic stabilisation, eliminating the need for manual numerical analysis.

The stabilising effect of multiplicative noise for stochastic differential equations, though counterintuitive, has been observed and investigated extensively in last decades. In practice, it is desirable to know if such stabilisation holds also for the discretised setting. In this paper, we address this problem by means of sharp upper and lower estimates for Lyapunov exponents of Milstein schemes for non-autonomous stochastic differential equations. These estimates provide precise large time behaviour in both almost sure and p-moment sense. In particular, our results show the preservation of stabilisation from the continuum setting to the discretised setting. One main idea of our analysis is to exploit the second order term concerning the stochastic noise from the Milstein scheme to obtain precise estimates for Taylor expansions of logarithmic and power functions.