Adaptive Time & Order Algorithm

• Adaptive Time and Order Algorithm is a dynamic numerical strategy that adjusts timestep and convergence order based on rigorous a posteriori error estimators.
• It employs cost models to decide between refining the time-step or elevating the order, optimizing overall computational efficiency in simulations.
• Widely applied in areas like PDEs, CFD/FSI, and model reduction, it significantly reduces computational cost while maintaining stability and prescribed accuracy.

An adaptive time and order algorithm is a dynamic strategy for numerical computation in which both the time discretization (stepsize) and the algebraic or convergence order of the underlying method are locally and automatically adjusted, based on rigorous a posteriori error estimates and cost models, to achieve prescribed accuracy with controlled computational complexity. This class of algorithms underpins recent advances in simulation of rough differential equations, PDEs, variable-order diffusions, CFD/FSI systems, and model reduction, with distinct algorithmic realizations reflecting the structure of each problem.

1. Core Principles and Error Estimation

Adaptive time and order algorithms are fundamentally built around rigorous, local a posteriori error estimators that decompose the global error into contributions from individual timesteps and, in some cases, order choices. For rough differential equations, as in the adaptive log-ODE methodology, the key is an exact error representation formula for any quantity of interest $g(y_T)$:

$$g(y_T) - g(\bar y_T) = \sum_{k=0}^{n-1} \left(\int_0^1 \Psi(t_k, \bar y_{t_k} + s\, e_k) s\, ds \right) e_k$$

where $e_k$ is the local one-step error and $\Psi$ is a dual "backward" solution measuring sensitivity to perturbations. This allows direct computation of the global error from local defect contributions (Bayer et al., 2023).

In variable-order and linear multistep schemes, estimators are usually based on divided differences of the numerical solution over several timesteps, yielding accurate proxies for the local truncation error at each step (Qin et al., 2022, Li et al., 2024, Quintana-Murillo et al., 2024). In pressure-dominated flow, the difference between a BDF2 step and a BDF3-approximate step via a Newton correction provides a heuristic but remarkably effective error estimate for timestep selection (Prusak et al., 2024).

2. Adaptive Workflow: Time Refinement vs. Order Elevation

The prototypical adaptive workflow involves, for each timestep or interval, the strategic choice between time-step refinement (halving the interval) or increasing the algebraic order of the integrator. For instance, in the adaptive log-ODE algorithm for RDEs, the core criterion is to compare the computational cost of halving step size versus raising the order. Refinement by a factor $m$ reduces local error by $m^{-(N+1)/p}$, while increasing order from $N$ to $N+1$ reduces the error proportionally to $\omega(I)^{1/p}$, where $\omega$ is a control of roughness (Bayer et al., 2023). The algorithm computes

$$m = \left(\frac{a_N}{a_{N+1}}\right)^{p/(N+1-p)}\, \omega(I)^{-1/(N+1-p)}$$

and prefers refinement if $m \leq \rho_N$ (cost ratio for raising the order), otherwise increases the order.

This adaptive decision mechanism is critical for efficiency, particularly in contexts with highly nonuniform or local regularity in the solution or its driving data, ensuring that computational resources are concentrated where the dynamics are most challenging.

3. Algorithmic Realizations in Key Application Domains

Adaptive time and order algorithms have specialized variants across application domains:

• Rough Differential Equations: The adaptive log-ODE method leverages error representation and local cost modeling to decide between grid refinement and order elevation on each subinterval, ensuring near-optimal complexity for any tolerance (Bayer et al., 2023).
• Linear Multistep/Filter-Based Schemes: In parabolic systems and coupled Stokes/Darcy models, a variable-step LMM (e.g., theta-scheme or BDF2) is augmented by adaptive time filters that lift the local convergence order (e.g., to second or third order), with error control via next-order divided differences and robust stepsize adaptation (Qin et al., 2022, Li et al., 2024).
• Fractional and Variable-Order Diffusions: For fractional Caputo subdiffusion problems, step-doubling adaptive strategies iteratively adjust the time mesh to enforce local error below a user-prescribed tolerance, shown to decrease total CPU time by orders of magnitude relative to fixed-timestep approaches (Quintana-Murillo et al., 2024).
• CFD and FSI: For pressure-dominated CFD and fluid-structure interaction, adaptive time algorithms based on BDF2/BDF3 with a single Newton-type correction exploit the difference as a heuristic estimator, feeding a classical PI controller for time-step selection (Prusak et al., 2024).
• Space-Time Model Reduction: Modern reduced order modeling algorithms adaptively select both temporal and spatial discretization parameters, exploiting randomized SVD and leave-one-out error estimation to control model order in large-scale data-driven system identification (Pelling et al., 10 Jun 2025).

4. Stability, Convergence, and Complexity Guarantees

These adaptive algorithms are analyzed through a blend of discrete stability theory—energy estimates, CFL-type constraints, a priori and a posteriori bounds—and rigorous complexity analysis. For log-ODE solvers, total computational cost to reach error $\mathrm{TOL}$ satisfies

$$\mathrm{Cost}(\mathrm{TOL}) = O\left(\mathrm{TOL}^{-\frac{p}{N+1-p}}\right)$$

recovering the optimal scaling as $N$ increases (Bayer et al., 2023). For time-filter approaches applied to variable-step LMMs (including BDF2-TF), unconditional nonlinear stability is established under mild step-ratio conditions (e.g., $\tau_n \leq 1.0315$), and convergence rates up to third order in both time and space are demonstrated for smooth solutions (Qin et al., 2022, Li et al., 2024). In the fractional setting, the adaptive L1 scheme is unconditionally stable and achieves first order in the maximal step size (Quintana-Murillo et al., 2024).

5. Pseudocode Architectures and Implementation Aspects

Although details differ, a generic pseudocode pattern emerges: (i) initialize grid, order, and method parameters; (ii) at each step, solve for the current approximation with current step size and order; (iii) compute the local a posteriori error estimator; (iv) based on estimator, decide to accept, refine, or elevate order (and possibly reject and repeat); (v) update grid, order, and continue. Concrete variants include the log-ODE "mark-then-refine/raise" procedure (Bayer et al., 2023), LMM+filter pipelines (Qin et al., 2022, Li et al., 2024), and step-doubling routines for variable-order diffusions (Quintana-Murillo et al., 2024).

Auxiliary decisions, such as adaptive stepsize clamping, safety factors, and step acceptance/rejection, are ubiquitously used to ensure robustness, especially in stiff or highly transient regimes. Importantly, the added overhead (e.g., backward dual RDE, filter vector updates) is typically $O(1)$ per step and negligible relative to the main PDE/DDE solve.

6. Numerical Performance and Scenario-Dependent Behavior

Empirical studies consistently report dramatic efficiency gains and robust error control. In log-ODE integration, adaptive time-order refinement reduces computational work by one to two orders of magnitude versus uniform schemes, with the error curve quantitatively tracking the sharp theoretical estimator (Bayer et al., 2023). In variable-order fractional diffusion, adaptive time selection leads to solution of long-time problems ($t \sim 10^3$–$10^4$) with $\ll 1\%$ of the steps required by a fixed-$\Delta t$ method, maintaining error near the target throughout (Quintana-Murillo et al., 2024). For BDF2-TF in 3D Stokes-Darcy systems, the third-order adaptive scheme achieves high accuracy at lower computational cost compared to either plain BDF2 or BDF3, and remains stable under realistic variable-step regimes (Li et al., 2024). In all contexts, the adaptivity mechanism prevents catastrophic stepsize selections and guarantees local error accuracy.

7. Limitations, Practical Recommendations, and Extensions

Adaptive time and order algorithms rely crucially on accurate a posteriori error estimation and calibrated cost models. In some strongly stiff or highly irregular problems, underestimated local error or inappropriate order selection can lead to instability or over-refinement; thus, adaptive controllers (e.g., safety factors, bounds on step size ratios) are carefully tuned for each application (Qin et al., 2022, Li et al., 2024, Prusak et al., 2024). The overall computational gain depends on the granularity and efficiency of estimator computation vis-à-vis the main solve, but in all documented cases, the overhead is substantially outweighed by the reduction in the total number of steps or function evaluations.

There is ongoing research to further automate adaptivity across all discretization parameters (time, order, space, even method family), including node-wise selection in PDE contexts (Malheiro et al., 2021), adaptive randomized linear algebra for large-scale ROM (Pelling et al., 10 Jun 2025), and advanced error/regularity quantification in nonsmooth or non-Markovian driving signals (Bayer et al., 2023, Quintana-Murillo et al., 2024). The general principle of minimizing total computational work given a global error constraint via local, adaptive, data-driven decision rules is now a central paradigm in scientific computation.