Motivation
A wave equation stores both displacement and velocity. Giving only the initial displacement does not determine its solution, and a two-level numerical history must encode both pieces of data correctly. This is the first difference from a heat step: a stable recurrence with an inaccurate first step can retain only first-order accuracy, despite using centered second differences thereafter.
The second difference is the role of energy. Heat dissipates a norm; an unforced wave exchanges kinetic and potential energy. A useful discrete energy must reflect that exchange. Leapfrog conserves a cross-time expression involving adjacent displacement gradients. That expression is not manifestly positive, so conservation alone does not establish a useful stability bound. A strict CFL margin supplies mesh-independent coercivity.
We prove these facts for a one-dimensional constant-speed wave equation with homogeneous Dirichlet boundaries, then derive a finite-time error estimate for a smooth exact solution. Periodic mode examples clarify the CFL endpoint without claiming that every endpoint computation is unstable. The theory is linear; nonlinear hyperbolic problems are not covered.
Definitions
Definition 12.1.1
(Leapfrog scheme with a Taylor start)
Consider
with , initial displacement and velocity . Let , , , , , and . Fix at every level. For interior vectors use and its induced norm. Write on all edges , with edge inner product . The endpoints are included in these differences.
The centered wave recurrence, often called leapfrog here, is
The initial levels are not arbitrary. With exact samples , and , use
at interior nodes, keeping the endpoint values zero. This is a second-order Taylor start using the PDE to replace initial acceleration. The velocity is associated with a half time; the average displacement is .
Concept lensAlgorithmic
Leapfrog as alternating velocity and displacement updates
The same method can be run with the half-step velocity as a stored unknown:
Eliminating recovers the three-level recurrence exactly, because . In this form the Taylor start is a half-length velocity update, , followed by . The initial velocity data enter only through that half step.
Definition 12.1.2
(Cross-time energy and a uniform strict CFL margin)
Define the cross-time energy
A uniform strict CFL hypothesis means with the same throughout refinement. Set . This is stronger than requiring a separate positive margin on each finite grid that might vanish as the grid is refined.
Theorem/proposition
Theorem 12.1.3
(Conservation becomes stability through coercivity)
Under the uniform strict CFL hypothesis, the energy satisfies
For zero forcing it is exactly conserved. With forcing it obeys
Given a bounded initial displacement norm and the initial-energy data bound derived below, displacement and half-step velocity have finite-time, mesh-independent bounds in terms of those data and the accumulated forcing. Conservation without the coercivity condition is not the claimed stability theorem.
Theorem 12.1.4
(Second-order finite-time convergence with a proper start)
Suppose a classical solution exists on the closed space-time rectangle, satisfies the homogeneous boundary conditions, and has bounded continuous , with having bounded fourth derivative. Use exact initial/forcing samples and the Taylor start. Under , the maximum over of is . Constants may depend on and these derivative bounds, but not on the mesh. The statement assumes the boundary/initial compatibility implicit in the existence of this smooth solution.
Proof sketch or proof idea
Proof
(Energy identity, coercivity and the forced bound)
Summation by parts with zero endpoints gives . Apply it to the recurrence and take the inner product with . The kinetic difference uses ; the gradient terms telescope across adjacent times. The resulting exact identity is
To interpret it, substitute and into the energy:
Here , summed over all edges with zero endpoints, proves . Thus the velocity coefficient is at least , proving coercivity. It controls half-step velocity and the gradient of the time-average, rather than claiming that the cross-gradient term alone is positive.
Cauchy–Schwarz in the forced identity gives
Divide by the sum of square roots; if both vanish the resulting inequality is already trivial. Summing proves the square-root energy bound. Also , so . This proves a displacement bound without needing a separate estimate for each time level's gradient.
Proof X-ray
Why the energy must be rewritten before it is used
The proof has three moves, and the CFL margin enters in only one of them. First, the test vector is chosen because it makes both parts telescope. The kinetic part telescopes through , and the potential part telescopes through summation by parts, which uses the zero endpoints. This gives exact conservation for every , so conservation by itself carries no information about the time step.
Second, and decisively, the cross term is rewritten with and . That exposes a negative contribution . The inverse inequality bounds it by , and only the strict margin leaves the positive coefficient . If were allowed to shrink with , every later constant would grow.
Third, the forced identity closes the argument. Cauchy–Schwarz together with gives an increment of bounded by , which sums over the steps. The convergence proof reuses exactly this estimate on the error equation, with the consistency residual as forcing.
The convergence theorem applies this stability estimate to the error between the computed and exact samples.
Proof
(Residual, start error and accumulation)
For consistency, insert exact nodal values into the scaled recurrence. Its residual is
Centered Taylor remainders give
Put . Its boundary is zero, its forcing is , and . The Taylor start differs from the exact first time sample only by the spatial acceleration error and cubic time remainder, so
Since , its initial energy is exactly ; no uncontrolled derivative of the starting error is needed. The forced stability result then gives for the half steps needed to reach . Summing error velocities yields
For bounded steps the right side is , proving the theorem. This proof shows separately how the initial velocity representation and every later residual affect the answer.
Worked examples
Worked example 12.1.5
(An entire fixed-end grid and its conserved energy)
Take , , , hence . With interior displacement , zero velocity and forcing, the proper start is . The next complete interior vector is ; both endpoints remain zero at every level.
At the first half time the velocity is , whose squared norm is . Edge gradients of and are and , with edge inner product five. Thus .
At the next half time the velocity is , squared norm , and the cross-gradient inner product is . Hence . The potential cross term is now negative, yet the full energy remains positive and unchanged. Replacing the cross product by a single-time gradient square would destroy this exact identity.
Worked example 12.1.6
(A resolved mode and a bad start)
For , , the interior vector has eigenvalue nine. With , a modal coefficient satisfies . A zero-velocity Taylor start gives , hence . Its numerical angle is , and .
For a general resolved mode, . If one instead sets , omitting initial acceleration, the solution becomes . The added amplitude is for a fixed smooth mode as . At generic finite times this is a genuine first-order error. A second-order recurrence does not repair a first-order velocity initialization.
For a spatial wave number , and the dispersion relation is on the resolved principal branch. It measures phase error, even though the discrete energy is conserved.
Worked example 12.1.7
(Exceeding CFL and interpreting its endpoint)
On , the mode has eigenvalue . Choose , so and . Its recurrence is , with roots . One root has magnitude greater than one. Even a proper zero-velocity start gives , exciting the growing mode.
For fixed-end grids the largest eigenvalue is , strictly below . Thus is still positive-energy on each fixed finite Dirichlet grid, but its coercivity margin tends to zero with refinement. That distinction is why the theorem states a uniform strict margin; the periodic endpoint case is examined under Common mistakes.
Boundary treatment and the meaning of data
A fixed end means setting its displacement at every time, not advancing it with an interior stencil requiring an undefined external node. Interior rows next to an endpoint use the prescribed boundary value. Nonzero Dirichlet data can be handled by a lift interpolating the two endpoint values. Then has zero boundaries and forcing , with initial displacement and velocity . For an affine spatial lift, . Smooth-data estimates require enough derivatives of the lift and compatible initial data.
Reflecting or absorbing conditions are different problems. Neumann endpoints alter the spatial operator and its natural weighted energy; an absorbing condition permits energy to leave through the boundary. The zero-end summation-by-parts identity does not automatically prove stability for either. Likewise, time-dependent forcing should be sampled at the recurrence's central time. Quietly substituting a different time quadrature changes the consistency calculation.
Initial energy can be bounded by the supplied data
The conserved quantity is useful only if its initial value is itself controlled. Insert the Taylor start into the original cross-time expression and collect terms. The mixed products between initial velocity and discrete acceleration cancel, leaving the exact identity
Consequently . The upper bound uses only the natural initial velocity norm, displacement gradient and scaled initial forcing. It has no inverse power of . The negative term should not be discarded when proving nonnegativity; that job was already done by strict-CFL coercivity. Here discarding it is legitimate because an upper bound is wanted.
This also clarifies what “stable initial data” means for a wave problem. A high-frequency displacement with fixed nodal amplitude can have a large gradient energy. A large resulting velocity is then not automatically a violation of a stability estimate expressed in terms of that gradient. To compare a family of increasingly fine computations, bound the actual norms on the right, rather than only the largest initial displacement value.
The initial formula is more than a computational convenience. It represents the initial derivative through a second-order displacement history while preserving the algebra needed for this energy balance. Alternative starts can be analyzed, but they must supply a controlled initial error energy. The error theorem uses the particularly simple fact that its exact displacement start makes the first cross-gradient error term vanish.
When checking the energy by hand, it helps to label every quantity with its time location: displacements live on integer levels, velocities and average displacements on half-integer levels, and the forcing work averages the half-step velocities before and after a level. Treating all of these as quantities at one time makes it easy to miswrite the cross gradient as a single-level square. The first worked example lists every edge precisely so that this distinction can be checked number by number rather than only through abstract matrix notation.
The spatial sums also use two index sets: there are interior nodes but edges, including the two edges at the ends. Zero endpoint values do not mean that the boundary edges can be dropped; the slopes next to the endpoints still store potential energy. This is also why the fixed-end energy proof cannot be transferred directly to periodic, reflecting or absorbing problems. Finally, the identities of this unit hold in exact arithmetic. Floating-point or evaluation errors can be estimated separately as additional forcing or update perturbations; a small energy drift should neither be read at once as an error in the derivation nor be ignored altogether while claiming that the actual computation conserves energy exactly.
Common mistakes
Common mistake
(A first-order start under a second-order recurrence)
Do not claim second-order accuracy after using without checking the start. Omitting the initial acceleration changes a resolved mode from to , and is of order . The centered recurrence then propagates this first-order error faithfully.
Common mistake
(Treating the conserved expression as a norm)
The cross-time energy is conserved for every , but its potential part can be negative: in the complete fixed-end example it equals at the second half step. Positivity and control of the velocity come from the coercivity estimate, not from conservation. Boundary edges must also be kept in the summation by parts; the slopes next to the fixed ends store potential energy.
Common mistake
(Stability is not resolution)
The CFL bound controls a numerical propagation mechanism; it does not certify accuracy for unresolved wavelengths. Energy conservation controls a quadratic expression and does not imply exact phase. The highest grid frequencies can have significant dispersion even when every algebraic computation is correct.
Counterexample mode
Conservation at the CFL endpoint is not stability
One might expect to suffice, since the energy is still conserved at . On an even periodic grid the checkerboard mode has , so its coefficients satisfy , whose characteristic root is double. From , the solution is . The displacement stays below on a fixed interval, but the half-step velocity grows like at near . The energy cannot see this growth: per unit mode norm,
for every . The energy is conserved but no longer coercive. The repair is the uniform strict margin . This is loss of coercivity, not a claim that every borderline displacement diverges.
Finally, the smooth convergence theorem does not establish a sharp error rate for discontinuous displacement, point impulses or nonlinear waves. Those settings need separate notions of solution and approximation; the constant-coefficient model of this unit cannot stand in for the whole theory of hyperbolic equations.
Summary
Leapfrog needs both a sound start and a sound propagation estimate. The PDE supplies second-order initial acceleration. A cross-time energy is conserved, and a uniform strict CFL margin makes it coercive. Applying its forced estimate to the exact-sample residual proves finite-time second-order displacement convergence. Borderline modes and incorrect starts show precisely what the hypotheses prevent.
Exercises
Quick check
(Take , initial interior vector , zero velocity and forcing. Compute both the proper first step and the second step.)
This mode has spatial eigenvalue 27.
Solution · Answer
Here , so the proper first coefficient is . The second is . Thus and , with zero endpoints. Using the lowest-mode eigenvalue nine would produce the wrong answer for this sign-changing vector.
Quick check
(On the periodic checkerboard at , verify from , and compute the half-step velocity magnitude.)
The characteristic root is a repeated minus one.
Solution · Answer
Substitution into cancels the linear factors, and the two starting values agree. The difference . For near this grows like , despite bounded initial half-step velocity and bounded finite-time displacement. This verifies the stated coercivity failure.
Quick check
(Under the strict-margin energy estimate, suppose an error has , and forcing norm at most . Derive the finite-time displacement bound.)
First bound the square root of error energy, then sum velocities.
Solution · Answer
Initial error energy is at most , so . Coercivity bounds every half-step error velocity by . Since , summing at most increments gives . These are absolute forcing bounds, with no cancellation assumption.
Quick check
(Transform boundary data , to homogeneous data with an affine spatial lift. Find the new forcing and both initial data.)
Differentiate the lift in time before writing the velocity data.
Solution · Answer
Take . Then and , so solves the homogeneous-boundary equation with forcing . Since , its initial displacement and velocity are still . Their endpoint compatibility must agree with the original smooth solution; subtracting a lift does not create missing regularity.