📊 statistics

Choosing optimal Strang splitting estimators of nonlinear stochastic differential equation models

This paper derives third-order bias error measures for Strang splitting schemes in nonlinear stochastic differential equations and demonstrates through simulation that the optimal splitting strategy for parametric inference depends on the specific model dynamics, with linearization around fixed points performing best for potential models and alternative splittings excelling in slow-fast excitable systems.

Original authors: Magnus Frederik Jensen, Johan Ravn Cornelius, Susanne Ditlevsen

Published 2026-07-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Magnus Frederik Jensen, Johan Ravn Cornelius, Susanne Ditlevsen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ 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 you are trying to predict the path of a leaf floating down a turbulent river. The water doesn't flow in a straight line; it swirls, eddies, and gets pushed by random gusts of wind. In the world of science, this is the challenge of modeling Stochastic Differential Equations (SDEs). These are mathematical recipes used to describe systems that change over time but are also jostled by random noise. You'll find them everywhere: from how neurons fire in a brain and how climate patterns shift, to how stock prices jump and how chemicals react. The problem is that for complex, twisting systems, we can't write down a perfect, neat formula to tell us exactly where the leaf will be next. The "transition density"—the probability map of where the leaf might go—is often a messy, unsolvable puzzle.

To get around this, scientists use "splitting" tricks. Think of it like navigating a complex maze. Instead of trying to solve the whole maze at once, you break it into two simpler parts: a straight, predictable hallway (the linear part) and a wild, twisting garden (the nonlinear part). You solve the hallway perfectly, then solve the garden perfectly, and stitch them together. This is called Strang splitting. It's a popular, fast, and clever way to estimate what's happening in these chaotic systems. But here's the catch: there isn't just one way to cut the maze. You could choose to make the "hallway" very long and the "garden" short, or vice versa. You could even choose to cut the maze in a totally different spot. Until now, nobody knew which cut was the best one to get the most accurate answer, especially when you only have a limited amount of data (a "finite sample").

This paper dives into that mystery. The authors, Magnus Frederik Jensen, Johan Ravn Cornelius, and Susanne Ditlevsen, ask a simple but crucial question: How do we choose the perfect way to split the math? They don't just guess; they derive new mathematical formulas to measure the "bias" (how far off the average guess is) and the "variance" (how scattered the guesses are) of these splitting methods. They found that while all these different ways of splitting eventually lead to the same answer if you wait forever, in the real world with limited data, the choice of split matters a huge amount.

The team tested their ideas on two famous mathematical playgrounds. The first is the double-well potential model, which is like a ball rolling in a landscape with two deep valleys separated by a hill. The second is the FitzHugh-Nagumo model, which mimics a neuron firing: it sits quietly, gets a little push, and then shoots off on a wild, fast excursion before settling back down.

Their simulations revealed a fascinating "one size does not fit all" rule. For the double-well model (the ball in the valleys), the best strategy was to split the math around the stable fixed points—the very bottom of the valleys where the ball naturally wants to rest. This "fixed point linearization" gave incredibly accurate results. However, for the FitzHugh-Nagumo model (the firing neuron), sticking to the stable resting point was a disaster. Because the neuron spends so much time on those wild, fast excursions far away from the resting spot, the "fixed point" method missed the action. Instead, the authors found that an adaptive splitting worked best. This is like a smart navigator that changes its strategy depending on where the system is right now: if the neuron is resting, use one map; if it's firing wildly, switch to a different map that understands the chaos.

The paper suggests that there is no single "magic bullet" splitting method. Instead, the optimal choice depends entirely on the shape of the system you are studying. If your system is a calm, potential-driven landscape, anchor your math to the stable valleys. But if your system is an excitable, fast-moving beast that loves to wander far from home, you need a flexible, adaptive approach that follows the action. By carefully choosing how to split the problem, scientists can get much sharper, more reliable estimates of how these complex, noisy worlds actually work.

Technical Summary: Choosing Optimal Strang Splitting Estimators of Nonlinear Stochastic Differential Equation Models

Problem Statement
Statistical inference for nonlinear multivariate stochastic differential equation (SDE) models is challenging because transition densities are rarely available in closed form, rendering exact likelihood-based inference intractable. While Strang splitting schemes have recently been introduced as a computationally efficient pseudo-likelihood approach for SDEs with additive noise, a critical unresolved issue remains: the splitting of the drift field into linear and nonlinear components is not unique. Although all admissible splittings yield the same asymptotic distribution for the estimator as the time step $h \to 0$ , the paper demonstrates that different splitting choices lead to dramatically different finite-sample behaviors regarding bias, variance, and numerical stability. The central problem addressed is how to define and identify an "optimal" splitting that minimizes approximation errors in finite samples.

Methodology
The authors develop a theoretical framework to analyze the local approximation errors of the Strang splitting scheme and propose criteria for selecting optimal splittings.

Theoretical Derivation of Error Measures:
The core theoretical contribution is the derivation of higher-order moment expansions for the one-step predictions generated by the Strang scheme. By expanding the moments of the Strang flow $\Phi^{[S]}_h$ and comparing them to the true diffusion process, the authors derive explicit expressions for the leading-order bias and variance terms.
- They prove that the bias of the Strang one-step prediction is of order $O(h^3)$ and depends explicitly on the chosen splitting parameters (the linearization matrix $A$ and vector $b$ ).
- Similarly, the difference in variance between the Strang approximation and the true process is also of order $O(h^3)$ and splitting-dependent.
- These error terms are expressed as functions $B_\theta(x)$ (bias) and $\Delta V_\theta(x)$ (variance difference), which depend on the derivatives of the nonlinear component $N(x)$ and the linearization center.
Optimality Criteria:
The paper proposes defining an optimal splitting as one that minimizes these error measures. Two primary strategies are investigated:
- Fixed Point Linearization: Linearizing the drift around stable fixed points of the system. This is particularly motivated for potential models where the process spends most of its time near stable equilibria.
- Adaptive/Local Linearization: Selecting the linearization center $c$ based on the current state of the process or minimizing the expected error over the invariant distribution. This includes piecewise schemes where different linearization centers are used in different regions of the state space.
Simulation Studies:
The methodology is tested on two distinct model classes:
- The Double-Well Potential Model: A one-dimensional gradient SDE representing a potential landscape with two stable fixed points.
- The Stochastic FitzHugh-Nagumo Model: A two-dimensional slow-fast excitable system where trajectories can undergo large excursions away from the stable fixed point.

Key Contributions and Results

Characterization of Finite-Sample Bias: The paper establishes that while Strang splitting is asymptotically unbiased, it possesses a systematic $O(h^3)$ bias in finite samples that is directly controlled by the choice of splitting. This explains why different splittings perform differently even when asymptotic theory suggests equivalence.
Model-Dependent Optimality:
- For Potential Models (Double-Well): The study finds that linearizing around the stable fixed points (specifically, using the fixed point within the basin of attraction where the process currently resides) yields transition densities with minimal local bias and variance. Simulation results confirm that this "fixed point linearization scheme" produces highly accurate parameter estimates, outperforming local linearization around the current state $x_t$ or random splittings.
- For Excitable Slow-Fast Systems (FitzHugh-Nagumo): The fixed point linearization scheme performs poorly when the system exhibits large excursions (e.g., in oscillatory or excitable regimes) because the process spends significant time far from the equilibrium. In these cases, an adaptive piecewise splitting scheme—linearizing around points on the stable slow manifolds rather than the fixed point—significantly improves the accuracy of the transition density and parameter estimation.
Practical Implementation: The authors demonstrate that for potential models, the fixed point linearization allows for an analytical maximization of the pseudo-likelihood with respect to the diffusion coefficient $\Sigma$ , reducing the dimensionality of the numerical optimization problem and improving robustness.

Significance and Claims
The paper claims that the choice of splitting is not merely a numerical detail but a central component of the inference procedure for nonlinear SDEs. The authors argue that:

Optimality is Context-Dependent: There is no universal "best" splitting. For gradient/potential systems, linearization around stable fixed points is a robust default choice. However, for systems with complex phase-space geometry (like slow-fast excitable systems), splittings must reflect the dominant dynamical structures (e.g., slow manifolds) rather than just equilibria.
Theoretical Justification for Empirical Success: The derived $O(h^3)$ error expansions provide a theoretical basis for the observed finite-sample performance differences, linking the quality of the parameter estimates directly to the accuracy of the transition density approximation.
Adaptive Strategies: The paper suggests that carefully designed adaptive splittings can significantly improve inference in nonlinear stochastic systems, particularly when the system dynamics are not well-represented by a single linearization point.

The authors conclude that while minimizing derived error measures directly can yield marginal improvements, the computational cost often outweighs the gains. Instead, leveraging the specific geometric structure of the model (fixed points for potentials, slow manifolds for excitable systems) offers a principled and practical path to optimal estimation.

Technical Summary: Choosing Optimal Strang Splitting Estimators of Nonlinear Stochastic Differential Equation Models

More like this