Abstract: This paper evaluates the performance and non-stiff order of two classes of exponential integrators—Lawson and ETD—for solving the nonlinear Schrödinger equation, highlighting the critical role of data regularity.
Introduction
The nonlinear Schrödinger (NLS) equation is a cornerstone of modern physics, describing phenomena ranging from fiber-optic pulses to deep-sea gravity waves. Numerically, the NLS equation is a "stiff" problem: the linear dispersive part evolves much faster than the nonlinear part. Standard numerical solvers often struggle with this stiffness, requiring tiny time steps to remain stable.
Exponential integrators offer a solution by solving the fast linear part exactly using matrix exponentials, allowing for much larger steps. However, their performance on dispersive, non-parabolic problems like the NLS equation depends heavily on the specific "flavor" of the integrator used.
Methods: Lawson vs. ETD
The study compares two specific fourth-order methods:
- Lawson4: An "Integrating Factor" method that uses a change of variables to "cancel out" the linear part before applying a standard Runge–Kutta scheme.
- ETD4RK: An "Exponential Time Differencing" method that approximates the nonlinear term with a polynomial and integrates the resulting expression exactly.
To test these, the NLS equation is semi-discretized in space using a Discrete Fourier Transform (DFT). This turns the partial differential equation into a system of ordinary differential equations in the frequency domain, where the linear operator becomes a simple diagonal matrix.
Key Insights: The Regularity Trap
The most significant finding is that the "best" method depends entirely on the input data's regularity (smoothness).
- Lawson Sensitivity: Lawson4's accuracy is directly tied to the regularity of the potential function $V(x)$. If the potential is "rough," Lawson4 loses its fourth-order accuracy, with its "stiff order" dropping significantly.
- ETD Sensitivity: Conversely, ETD4RK is remarkably robust regarding the potential but is highly sensitive to the smoothness of the initial condition. If the starting wave function isn't smooth enough, ETD4RK's performance degrades.
Key Results and Accuracy
In ideal conditions—smooth potential and smooth initial state—Lawson4 was roughly 100 times more accurate than ETD4RK at comparable step sizes.
As shown in the figure above, both methods exhibit fourth-order convergence initially, but Lawson4 maintains a lower error profile until hitting a "rounding error floor" at approximately $10^{-10}$. This floor is likely due to the challenges of computing the complex $\phi$-functions required for ETD schemes with high precision.
Numerical Stability and $\phi$-functions
Evaluating the functions required for ETD schemes is computationally tricky. These $\phi$-functions act as averaging filters, but their numerical implementation is prone to instability.
The figure above illustrates the behavior of these functions in the frequency domain. While they help damp high-frequency modes, rounding errors in their calculation limit the long-term accuracy of ETD methods.
Conclusion
The choice of integrator for the NLS equation is not absolute. If you have a very smooth potential, Lawson4 is the superior choice due to its efficiency and accuracy. However, if the potential is rough but the starting state is smooth, ETD4RK is the safer bet. Future work remains to bridge the theoretical gap in "stiff order" analysis for dispersive equations.
Summary updated