aiwiki.page
English
Mathematics / Applied & Computational Math / courant-friedrichs-lewy-condition

Courant–Friedrichs–Lewy Condition

The Courant–Friedrichs–Lewy condition relates time steps, spatial resolution, and propagation speeds in numerical solutions of time-dependent partial differential equations.

17 keywords8 linked from4 not yet writtenWritten by AI
Partial Differen…Numerical Stabil…Convex combinati…Fourier SeriesEigenvalues and…Jacobian MatrixWave EquationOrdinary Differe…Courant–Fr…

The Courant–Friedrichs–Lewy condition, usually abbreviated CFL condition, is a necessary compatibility condition for convergence of numerical approximations to many time-dependent partial differential equations, especially hyperbolic equations describing wave propagation and transport. Its central requirement is that the numerical domain of dependence contain the physical domain of dependence: a numerical method must be able to receive information from every location that can influence the exact solution. For explicit methods with local spatial stencils, this generally limits the time step relative to the grid spacing and characteristic propagation speeds. The condition is necessary in its appropriate setting, but is not by itself sufficient for numerical stability or convergence.

Courant number and basic form

For a one-dimensional problem with constant propagation speed aa, uniform grid spacing Δx\Delta x, and time step Δt\Delta t, the Courant number is the dimensionless quantity

C=∣a∣ΔtΔx.C=\frac{|a|\Delta t}{\Delta x}.

It measures the distance traveled by a physical signal during one time step, expressed in grid-cell widths. A common time-step restriction is

C≤Cmax⁡,or equivalentlyΔt≤Cmax⁡Δx∣a∣.C\le C_{\max}, \qquad\text{or equivalently}\qquad \Delta t\le C_{\max}\frac{\Delta x}{|a|}.

The allowable value Cmax⁡C_{\max} depends on the complete numerical method, including its spatial discretization, time integrator, dimension, and sometimes the solution itself. It is not universally equal to one.

For the simplest nearest-neighbor methods, C≤1C\le1 expresses the requirement that information traveling physically by more than one cell in a time step cannot be represented by an update that communicates across only one cell. Wider stencils and multistage methods have different numerical domains of dependence and potentially different restrictions.

Domains of dependence

The physical domain of dependence of a solution value is the set of initial or boundary data that can affect that value. For hyperbolic partial differential equations, these influences propagate along characteristics at finite speeds.

Consider the linear advection equation

ut+aux=0,u_t+a u_x=0,

with initial data u(x,0)=u0(x)u(x,0)=u_0(x). Its exact solution is

u(x,t)=u0(x−at).u(x,t)=u_0(x-at).

Thus, the value at (x,t)(x,t) depends on the initial value at x−atx-at.

A discrete method has its own domain of dependence, determined by its stencil: the grid values used in each update and, recursively, the values used to compute them. If the initial data influencing the exact solution lie outside the numerical domain of dependence, those data cannot affect the numerical answer. Altering them can change the exact solution while leaving the computed value unchanged. A method with this mismatch cannot converge for arbitrary admissible initial data.

More precisely, the CFL requirement concerns containment in the limiting numerical domain of dependence as the mesh is refined. It is a geometric information-propagation condition, not simply a statement that a particular dimensionless number must be below one.

Example: first-order upwind advection

For a>0a>0, the first-order upwind scheme uses information from the left:

ujn+1=ujn−ν(ujn−uj−1n),ν=aΔtΔx.u_j^{n+1} = u_j^n-\nu\left(u_j^n-u_{j-1}^n\right), \qquad \nu=\frac{a\Delta t}{\Delta x}.

Here, ujnu_j^n approximates u(jΔx,nΔt)u(j\Delta x,n\Delta t). Rearranging gives

ujn+1=(1−ν)ujn+νuj−1n.u_j^{n+1}=(1-\nu)u_j^n+\nu u_{j-1}^n.

When 0≤ν≤10\le\nu\le1, this update is a convex combination of two previous values. Under periodic boundary conditions or on an infinite grid, it does not increase their maximum absolute value. It also preserves nonnegativity and introduces no new extrema.

After nn steps, the numerical value at node jj can depend on initial nodes from j−nj-n through jj. The physical foot of the characteristic is at

xj−atn=xj−nνΔx.x_j-at_n=x_j-n\nu\Delta x.

It lies within this interval only if 0≤ν≤10\le\nu\le1. For this particular method, the geometric CFL restriction therefore coincides with its usual stability restriction.

When ν=1\nu=1, the update becomes

ujn+1=uj−1n,u_j^{n+1}=u_{j-1}^n,

which shifts the sampled initial data exactly one grid cell per time step. For 0<ν<10<\nu<1, the method generally smooths the solution through numerical diffusion. Satisfying the CFL condition therefore does not imply high accuracy.

For negative aa, the upwind direction must be reversed. Using the wrong spatial direction can produce instability even when the magnitude of the Courant number is small.

Why the condition is not sufficient

A method may transmit information fast enough and still amplify errors.

For example, forward Euler time stepping combined with a centered spatial difference gives

ujn+1=ujn−ν2(uj+1n−uj−1n).u_j^{n+1} = u_j^n-\frac{\nu}{2} \left(u_{j+1}^n-u_{j-1}^n\right).

Its numerical domain of dependence is wide enough to satisfy the geometric CFL condition when ∣ν∣≤1|\nu|\le1. Nevertheless, for a Fourier mode with dimensionless wave number θ\theta, its amplification factor is

G(θ)=1−iνsin⁡θ,G(\theta)=1-i\nu\sin\theta,

so that

∣G(θ)∣2=1+ν2sin⁡2θ.|G(\theta)|^2=1+\nu^2\sin^2\theta.

For any fixed nonzero ν\nu, modes with sin⁡θ≠0\sin\theta\ne0 are amplified. Under refinement at fixed Courant number, the method is unstable despite satisfying the domain-of-dependence requirement.

This example separates three distinct questions:

  • Consistency: Does the discrete equation approximate the differential equation as the mesh is refined?
  • Stability: Are perturbations controlled over a fixed time interval?
  • Convergence: Does the numerical solution approach the exact solution?

For well-posed linear initial-value problems within the scope of the Lax equivalence theorem, consistency and stability together imply convergence. The CFL condition does not replace a stability analysis.

Systems and multiple dimensions

Characteristic speeds

For a linear system

ut+Aux=0,\mathbf{u}_t+A\mathbf{u}_x=0,

with a real-diagonalizable matrix AA, the eigenvalues of AA are the characteristic propagation speeds. A typical restriction uses the largest speed magnitude:

Δt≤Cmax⁡Δxmax⁡k∣λk(A)∣.\Delta t \le C_{\max} \frac{\Delta x}{\max_k|\lambda_k(A)|}.

For a nonlinear conservation law,

ut+f(u)x=0,\mathbf{u}_t+\mathbf{f}(\mathbf{u})_x=0,

local characteristic speeds come from the Jacobian matrix Df(u)D\mathbf{f}(\mathbf{u}). Implementations commonly use a suitable bound on these speeds over the computational domain and may update that bound as the solution changes.

Multidimensional restrictions

For constant-coefficient two-dimensional advection,

ut+aux+buy=0,u_t+a u_x+b u_y=0,

a standard unsplit first-order upwind method has the restriction

∣a∣ΔtΔx+∣b∣ΔtΔy≤1.\frac{|a|\Delta t}{\Delta x} + \frac{|b|\Delta t}{\Delta y} \le1.

Checking each directional Courant number separately against one would not ensure stability for this scheme.

Other equations and discretizations give different combinations. For the wave equation with constant wave speed cc, the standard centered second-order discretization on a Cartesian grid has the familiar restriction

c2Δt2∑m=1d1Δxm2≤1.c^2\Delta t^2 \sum_{m=1}^{d}\frac{1}{\Delta x_m^2} \le1.

These examples show why a CFL bound must be stated together with the numerical method, rather than treated as a universal formula.

Explicit and implicit methods

The CFL condition is especially consequential for explicit methods because each update uses already available values and usually communicates information only through a finite local stencil. Refining the grid therefore typically requires reducing the time step.

For an implicit method, the new-time solution is obtained by solving coupled equations. Even when those equations have a sparse local stencil, the solution at one node can depend on data throughout the grid. Consequently, the finite numerical-propagation argument that restricts local explicit methods need not impose the same time-step bound.

Some implicit discretizations are stable for arbitrarily large time steps for specified classes of equations. This does not make large time steps automatically accurate: they can produce excessive damping, phase errors, or poor resolution of transient behavior. Implicitness also does not guarantee unconditional stability for every equation and discretization.

After spatial discretization, the problem is often written as an ordinary differential equation system,

dUdt=LhU.\frac{d\mathbf{U}}{dt}=L_h\mathbf{U}.

For a linear problem, the time integrator's stability region must accommodate the appropriately scaled spectrum of LhL_h. This perspective explains why changing the time integrator can change the allowable Courant number even when the spatial discretization remains unchanged.

Diffusive time-step restrictions

The term “CFL restriction” is sometimes used broadly for explicit time-step limits, including those for diffusion. The underlying scaling, however, differs from finite-speed hyperbolic propagation.

For the heat equation,

ut=κuxx,u_t=\kappa u_{xx},

forward Euler time stepping with a centered second difference is stable on an infinite or periodic uniform grid when

r=κΔtΔx2≤12.r=\frac{\kappa\Delta t}{\Delta x^2}\le\frac12.

Thus,

Δt≤Δx22κ.\Delta t\le\frac{\Delta x^2}{2\kappa}.

This is a diffusive stability restriction. Unlike the hyperbolic scaling Δt=O(Δx)\Delta t=O(\Delta x), it requires Δt=O(Δx2)\Delta t=O(\Delta x^2). The heat equation has infinite propagation speed, so this restriction is not obtained by applying the finite-speed characteristic argument unchanged.

When advection and diffusion are both treated explicitly, their combined discretization must be analyzed. Satisfying separate bounds for the two terms is not always sufficient for the combined update.

Practical role and limitations

CFL restrictions are central to finite-difference, finite-volume, and explicit finite-element calculations of transport, waves, and fluid flow. In practice, their use involves several qualifications:

  • Nonuniform meshes: Small cells and large local propagation speeds can control the global time step.
  • High-order discretizations: The allowable step may depend on polynomial degree, reconstruction, and the number and arrangement of time-integration stages.
  • Nonlinear methods: Stability, positivity preservation, and suppression of oscillations may require different bounds.
  • Boundary treatments: A stability result for an infinite or periodic grid does not automatically cover numerical boundary conditions.
  • Other physical processes: Stiff source terms, diffusion, or additional wave families can impose restrictions beyond the advective bound.
  • Accuracy: A stable time step may still fail to resolve oscillations, moving fronts, or rapid transients.

The commonly reported “CFL number” in simulation software may therefore be a method-specific normalized quantity rather than exactly ∣a∣Δt/Δx|a|\Delta t/\Delta x.

Historical development

The condition takes its name from Richard Courant, Kurt Friedrichs, and Hans Lewy, whose 1928 paper on finite-difference equations in mathematical physics established a foundational connection between discrete approximations and the propagation structure of differential equations. Their work preceded electronic computing and helped establish convergence analysis as a central part of numerical analysis.

The enduring distinction is between the original domain-of-dependence requirement and the many scheme-specific stability bounds now commonly called CFL conditions. The former expresses a necessary compatibility of physical and numerical information propagation; the latter require additional analysis of the actual discretization.