Adaptive temporal control in numerical solutions aims to balance efficiency and accuracy when multiple time scales coexist. Adaptive time-stepping adjusts step size based on estimated local truncation error and stability monitors, but in multiscale partial differential equations this local focus interacts nontrivially with long-range error accumulation. Erik Hairer, Université de Genève, and Gerhard Wanner, Université de Genève, established foundational principles for variable-step control that remain central to understanding these tradeoffs. Local correctness does not guarantee faithful long-term dynamics when unresolved fast scales feed back into slow modes.
Mechanisms of error amplification
When a time integrator shortens or lengthens steps in response to instantaneous error indicators, it changes the discrete sampling of fast oscillations and nonlinear couplings. Two principal mechanisms drive error growth. The first is stiffness driven by disparate eigenvalues: explicit methods face a CFL condition that forces tiny steps for stability, while implicit methods damp stiff modes but introduce solver and consistency errors. Randall LeVeque, University of Washington, has discussed how discretization choices interact with wave propagation and stability limits. The second mechanism is resonance between step-periodic sampling and intrinsic multiscale oscillations, producing systematic phase errors that accumulate into large global error over long simulation times. Error estimators tuned to pointwise residuals can miss structure transported across scales, so adaptive controllers may underresolve critical dynamics.
Practical consequences and mitigation
Consequences range from quantitatively wrong long-time statistics in turbulence and climate models to qualitatively incorrect pattern formation in materials simulations. Weinan E, Princeton University, developed multiscale frameworks such as the Heterogeneous Multiscale Method that explicitly separate scales and couple integrators, reducing blind spots of purely local adaptivity. Alfio Quarteroni, EPFL, has emphasized coupling spatial multiscale discretizations with time-stepping strategies that honor conservation laws and invariants to limit drift. Effective mitigation combines multirate methods that use different step sizes for different components, error estimators aligned with slow manifold projections, and asymptotic-preserving schemes that retain correct limits as scale separation increases. Monitoring conserved quantities and coarse-grained observables often reveals growing global error before pointwise indicators do. For applications with human or environmental stakes, such as coastal flood forecasting or seismic hazard assessment, these safeguards are essential to avoid decisions based on spurious numerical artifacts.