Source-linked AI summary
Analysis of split exponential integrators for semilinear problems with noncommuting operators
Luca Battestini, Marco Caliari, Fabio Cassini
TL;DR
Computing exponential-like functions for stiff semilinear problems can be costly, motivating rigorous analysis of splitting for noncommuting operators. The paper derives splitting-error representations and proves convergence for first- and second-order split exponential integrators. Experiments validate the analysis and report substantial computational savings against unsplit methods.
Problem
For stiff semilinear problems, exponential integrators require costly exponential-like matrix functions evaluated at the combined operator L = A + B, motivating split approximations.
Source: §1 · p. 2Method
The paper derives error representations for splitting with noncommuting operators and proves convergence of split exponential Euler and second-order exponential Runge–Kutta methods in a strongly continuous semigroup framework.
Source: §2 · p. 3, §4 · p. 15Results
Experiments on nonlinear dispersive and diffusion–reaction equations validated the theory and showed significant savings; matrix-oriented split schemes were about two orders of magnitude faster than unsplit methods.
Source: §3.2 · p. 15, §4 · p. 16Takeaways & Limitations
The split methods provide a rigorously analyzed alternative whose matrix-oriented implementation can substantially reduce computational time compared with unsplit methods in the reported experiments.
Source: §4 · p. 15, §3.2 · p. 15The analysis relies on suitable contraction bounds; shift and renorming can address their failure only in some cases, such as when A generates a strongly continuous semigroup and B is bounded.
Source: §2.1 · p. 5
Abstract
from arXiv · showhide
Splitting the exponential and exponential-like functions can significantly reduce the overall computational cost of exponential integrators, providing a strong motivation for their rigorous analysis. In this paper, we analyze splitting errors for two noncommuting unbounded operators and derive new representations that include the popular Lie--Trotter and Strang cases. Moreover, within the abstract framework of strongly continuous semigroups, we prove the convergence of a split version of the exponential Euler method and a second-order exponential Runge--Kutta method for semilinear problems. Numerical experiments on a nonlinear dispersive equation and a diffusion--reaction equation demonstrate the theoretical findings, stability properties, and efficiency of the proposed split exponential integrators compared to the state of the art.
1 Introduction
Stiff semilinear problems make exponential-like matrix functions costly to compute, motivating split approximations. The paper analyzes splitting errors and convergence for first- and second-order split exponential integrators, then compares their performance with unsplit methods.
1 Introduction
- For stiff semilinear problems, evaluating exponential-like functions at L = A + B can be costly, motivating approximations that use A and B separately.Such problems arise from spatial discretizations of semilinear PDEs; the operators can be large-norm matrices, and standard explicit time-marching schemes require small time steps.Source: §1 · p. 2
- The paper derives new error representations for splitting exponential-like functions with noncommuting operators and analyzes convergence of first- and second-order split exponential integrators.The studied schemes are ERK1L and ERK2L, based on split versions of exponential Euler and exponential Runge–Kutta methods.Source: §1 · p. 2
- The numerical section compares the split integrators' performance with unsplit methods and concludes the paper with possible future developments.The introduction also previews numerical examples and discussion of stability properties.Source: §1 · pp. 2–3
2 Splitting approximations and exponential integrators
The section derives splitting-error representations for noncommuting operators and uses them to analyze first- and second-order split exponential integrators. Under stated semigroup and regularity assumptions, it establishes global-error bounds for ERK1L and ERK2L.
2 Splitting approximations and exponential integrators
- The abstract analysis assumes A, B, and L = A + B generate strongly continuous semigroups on a Banach space.The section presents the schemes and convergence analysis within this semigroup framework.Source: §2 · p. 3
2.1 Split exponential integrator ERK1L
- For v in D1 the local splitting error is bounded by Cτ, while for v in D2 it is bounded by Cτ^2.The error representation has total operator degree two; for ℓ=0 it gives a Lie–Trotter splitting error with products of A and B in rightmost positions.Source: §2.1 · pp. 4–5
- Under Assumptions 1–3, Theorem 2.3 states a global-error bound for scheme ERK1L (8); Assumption 2 requires ∥etA∥≤1 and ∥etB∥≤1 for 0≤t≤T.Source: §2.1 · pp. 5–6
2.2 Split exponential integrator ERK2L
- ERK2L replaces the exponential using Strang splitting and uses second-order approximations for φ1 and φ2, while the exponential requires a higher-order approximation.The paper obtains ERK2L by inserting the splitting approximations into ETD2RK.Source: §2.2 · pp. 7–8
- The ERK2L Strang-splitting error representation has total degree three in A and B and differs from earlier representations.The representation supports a local error bound when v belongs to D3.Source: §2.2 · pp. 8–9
- Under Assumptions 1, 2, and 4, the authors establish a global-error bound for ERK2L.Assumption 4 requires u(t) in D3, g(u(t)) in D2, and bounded first and second Fréchet derivatives of g.Source: §2.2 · p. 9
3 Numerical examples
The examples validate the split integrators on a nonlinear dispersive equation and a diffusion–reaction equation. For the diffusion–reaction problem, matrix-oriented split methods attain different accuracy levels faster than unsplit methods, despite requiring more time steps.
3.1 Nonlinear dispersive equation
- The dispersive-equation experiments confirm the expected first- and second-order convergence of ERK1L and ERK2L, respectively.The problem uses 300 Fourier modes per direction and runs to T = 1.Source: Figure 1 · p. 11, §3.1 · p. 12
3.2 Enzyme kinetics of Michaelis–Menten type
- The diffusion–reaction test discretizes space with second-order finite differences and uses separate actions of matrix functions of A and B to reduce computation.The split approach computes small matrix functions and uses scalar quantities and level 3 BLAS.Source: §3.2 · pp. 12–13
- At T = 0.25, ERK1L damps the corner oscillations produced by method (13) on the diffusion–reaction problem.Both methods use 40 time steps on a (200, 200) spatial discretization.Source: §3.2 · p. 14, Figure 2 · p. 14
- With ERK2L, the diffusion–reaction solution approaches the zero stationary state over time.The simulation uses (100, 100) spatial points and τ = 1/3200.Source: Figure 3 · p. 14, §3.2 · p. 15
- About two orders of magnitude speedup: matrix-oriented split schemes outperform unsplit methods at all tested accuracy levels, despite requiring more time steps.The comparison uses ERK1L and ERK2L against ETD1RK and ETD2RK, respectively.Source: Figure 4 · p. 15, §3.2 · p. 15
4 Conclusions
The paper establishes splitting-error representations for two noncommuting unbounded operators and proves convergence for split first- and second-order exponential integrators. Experiments on nonlinear dispersive and diffusion–reaction equations validate the theory and show computational savings against unsplit methods.
4 Conclusions
- New elementary error representations for two noncommuting unbounded operators include Lie–Trotter and Strang splittings, with rigorous convergence proofs for split first-order exponential Euler and second-order exponential Runge–Kutta schemes.Source: §4 · p. 15
- Experiments on nonlinear dispersive and diffusion–reaction equations validated the theoretical results and showed significant computational savings versus unsplit state-of-the-art methods.Source: §4 · p. 16
- Future research directions include extending semigroup error analysis to higher-order exponential Runge–Kutta methods and multi-term operator splittings.Source: §4 · p. 16