Local CFL Constraint in Adaptive Methods

• Local CFL constraint is a condition enforcing stability in explicit time integration by setting local time step limits based on local mesh size and characteristic speeds.
• It enables adaptive, efficient time-stepping in multi-dimensional PDE solvers while ensuring global causality and discrete conservation across neighboring regions.
• Its practical implementation in local time-stepping algorithms leads to significant computational speedups without compromising accuracy or stability.

A local CFL (Courant–Friedrichs–Lewy) constraint refers to a stability condition for explicit time integration schemes in numerical PDEs, imposed independently at discrete subdomains (cells, patches, elements, or nodes) rather than globally. The local CFL constraint establishes bounds on the admissible time step in each spatial region based on its local mesh size and local characteristic speeds. Its correct definition, enforcement, and implications are foundational in high-performance, adaptive, and multi-physics numerical methods for hyperbolic and wave-like equations, notably when combined with local time-stepping (LTS) strategies and mesh refinement. The theory and practice of local CFL constraints address both discrete stability and the preservation of physical propagation properties, as well as their limitations in complex coupled algorithms.

1. Mathematical Definition and Canonical Role

In explicit schemes for hyperbolic PDEs, the CFL constraint is a necessary condition for stability, typically written globally as

$$\Delta t \leq C_{\rm CFL} \frac{h}{\lambda_{\max}},$$

where $h$ is the mesh spacing and $\lambda_{\max}$ the maximum local wavespeed. For nonuniform meshes or spatially varying coefficients, the local CFL constraint instead assigns a bound at each cell or element: $$\Delta t_i \leq C_{\rm CFL} \frac{h_i}{\lambda_{\max,i}}$$ where $h_i$ and $\lambda_{\max,i}$ are the local mesh size and maximum characteristic speed in region $i$ (Lilly et al., 2024, Gnedin et al., 2018).

For general conservation laws in multiple dimensions (e.g., shallow water, compressible flow), $\Delta t_i$ depends on the minimum spacing $\min\{\Delta x_i, \Delta y_i, \Delta z_i\}$ and the maximum eigenvalue of the flux Jacobian. This local approach permits significant flexibility in adaptive discretizations, allowing large steps in regions with coarse mesh or slow waves and small steps elsewhere.

2. Local CFL Constraints in Local Time-Stepping Algorithms

In LTS, distinct $\Delta t_i$ are assigned to cells or groups of cells so that each evolves at a time scale compatible with its local stability. The local CFL constraint becomes operational: $$\Delta t_{p} \leq C_{\rm CFL} \frac{\Delta x_p}{\lambda_{\max,p}}$$ for patch $p$ (Gnedin et al., 2018, Lilly et al., 2024, Ricardo et al., 9 Jul 2025). Algorithmic enforcement requires coordination among neighboring cells to prevent causality violations. In the FB-LTS scheme for shallow water equations, $\Delta t_i$ is quantized into step-levels, ensuring

$$\text{CFL}_i = \frac{(|\mathbf{u}| + \sqrt{gH})\,\Delta t_i}{\Delta x_i} \leq \text{CFL}_{\max}$$

with per-cell steps grouped for efficient subcycling and interface synchronization (Lilly et al., 2024).

Table: Local CFL constraint definitions in various contexts

Method/Class	Local CFL Formulation	Reference
Finite volume / FV-LTS	$$\Delta t_p \leq C_{\rm CFL} \frac{\Delta x_p}{\lambda_{\max,p}}$$	(Gnedin et al., 2018)
Adaptive DG/ADER-DG	$$\nu := \|\mathbf{u}\|_\infty\Delta t / h \leq \nu_{\max}$$	(Ricardo et al., 9 Jul 2025)
Shallow water FB-LTS	$$\Delta t_i \leq \text{CFL}_{\max} \frac{\Delta x_i}{|\mathbf{u}|+\sqrt{gH}}$$	(Lilly et al., 2024)

See main text for notation details.

3. Synchronization and Causal Consistency

A subtlety of local CFL constraints is the necessity to maintain global causality: waves originating from one region must accurately propagate into adjacent regions without unphysical delays. Merely setting each $\Delta t_i$ according to its local bound can violate this if independent time evolution allows neighboring patches to be out of synchronization.

To prevent this, it is enforced that for any two neighboring patches $p$ and $q$ (with $p$ about to advance),

$t_q^{n+1} \leq t_p^{n+1}$

where $t_p^{n+1} = t_p^n + \Delta t_p$ is the end time of $p$'s step. If $q$'s planned step would overrun $t_p^{n+1}$, its step is shortened ("step-drop") or it is forced to advance in interlock with $p$ ("pull-back") so that no patch gets ahead of its neighbors (Gnedin et al., 2018). This protocol ensures strict adherence to the local CFL limit and correct discrete propagation.

4. Numerical Stability and Mesh Independence

Traditionally, the maximal global time step in explicit schemes is governed by the stiffest (smallest) cell—the minimal $h_i/\lambda_{\max,i}$. Several modern LTS schemes show, both theoretically and numerically, that with appropriate local CFL enforcement, the effective time step in non-stiff regions may be much larger, with global stability and conservation preserved.

For example, with stabilized leapfrog local time-stepping (LF-LTS) and Chebyshev-type damping, the necessary stability condition becomes

$\Delta t \leq C h_c$

where $h_c$ is the coarse mesh diameter—and crucially independent of the smallest $h_f$ in fine mesh regions. The modified update operator preserves symmetry, energy, and optimal convergence rates under this mesh-independent CFL (Grote et al., 2020). In high-order DG contexts, local-implicit globally-explicit ADER-DG advances each element with a CFL number $\nu_{\max}=1/\sqrt{d}$—sharp, and independent of polynomial order $N$ (Ricardo et al., 9 Jul 2025).

5. Conservation, Accuracy, and Algorithmic Implementation

Proper local CFL constraint implementation in LTS or adaptive schemes must ensure not only discrete stability, but also discrete conservation (e.g., mass, energy), and preservation of order of accuracy:

• Conservation: Predictor–corrector coupling at time-level boundaries enforces a unified state at interfaces, ensuring discrete flux balance. For instance, in shallow water TRiSK with FB-LTS, fluxes at inter-stage boundaries are computed from a single state, yielding exact conservation of global mass and absolute vorticity (Lilly et al., 2024).
• High-Order Accuracy: In stabilized LF-LTS, explicit three-term recurrence structures and damped Chebyshev polynomials preserve second-order accuracy and energy conservation under local stepping (Grote et al., 2020).
• Bookkeeping: Implementation involves quantization into step-levels (e.g., two-level for coarse/fine), management of active/inactive step lists, and communication (e.g., MPI messages for pull/drop in distributed-memory) to guarantee neighbor synchronization (Lilly et al., 2024, Gnedin et al., 2018).

6. Broader Context, Limitations, and Nonlinear Programming Analogue

Local constraints ("CFL constraints") also arise in other domains under a different guise. In capacitated facility location LP relaxations, imposing "local" constraints—even with many layers of lift-and-project or flow-cover inequalities—cannot achieve a constant integrality gap unless genuinely global (i.e., unbounded-size-support) constraints are introduced (Kolliopoulos et al., 2013). This illustrates that while local constraints can be optimal in some explicit PDE solvers, in integer programming/local polyhedral relaxations for combinatorial optimization, they can be fundamentally insufficient.

A plausible implication is that in both PDEs and optimization, strict locality may suffice for numerical stability but not always for global exactness or competitiveness—demanding careful analysis in each setting.

7. Numerical Examples, Efficiency, and Practical Impact

Benchmarks demonstrate significant efficiency gains with rigorous local CFL enforcement:

• MPAS-Ocean with FB-LTS on TRiSK meshes achieves up to $10\times$ speedups over global RK4 under strong solution equivalence, while maintaining conservation and high order (Lilly et al., 2024).
• In patch-based fluid dynamics with strict local CFL, speedups of $3\times$–$10\times$ are reported relative to globally uniform stepping, with negligible overhead and identical convergence/shock properties to standard solvers (Gnedin et al., 2018).

These findings validate that formulating, verifying, and enforcing appropriate local CFL constraints—beyond naively local step selection—are vital for both theoretical and practical advances in high-fidelity time-dependent simulation.

---

References:

(Grote et al., 2020) Grote, Michel, Sauter. "Stabilized leapfrog based local time-stepping method for the wave equation" (Gnedin et al., 2018) Gnedin, Semenov, Kravtsov. "Enforcing the Courant-Friedrichs-Lewy Condition in Explicitly Conservative Local Time Stepping Schemes" (Lilly et al., 2024) Mills et al. "Local Time-Stepping for the Shallow Water Equations using CFL Optimized Forward-Backward Runge-Kutta Schemes" (Ricardo et al., 9 Jul 2025) Merklinger et al. "Scalable ADER-DG Transport Method with Polynomial Order Independent CFL Limit" (Kolliopoulos et al., 2013) Kolliopoulos, Moysoglou. "Sherali-Adams gaps, flow-cover inequalities and generalized configurations for capacity-constrained Facility Location"