Motivation
Passing from one spatial dimension to two adds more than another difference quotient. Boundary values enter along four sides, the unknowns form a coupled linear system, and time-step restrictions depend on both spatial spacings. A convincing computation must explain how these pieces fit together. A small residual is useful only if the discrete boundary-value problem responds to residuals in a controlled way.
Two complementary arguments provide that control. A discrete maximum principle compares values point by point and gives a maximum-norm error estimate through a barrier. A discrete energy identity measures differences between neighboring nodes and explains why diffusion dissipates a squared norm. The first argument is particularly transparent for the five-point stencil; the second extends naturally to variational methods.
We use a rectangle and smooth exact solutions so the proof exposes numerical reasoning without hiding a regularity theorem. Curved domains, corner singularities, unstructured meshes and general variable coefficients require additional work.
Definitions
Definition 3.2.1
(Five-point Dirichlet scheme on a rectangle)
Let , , , with integers . Nodes are . Interior indices satisfy , . Define the negative discrete Laplacian
For with Dirichlet data , the scheme is at interior nodes and at boundary nodes. Boundary values are known numbers. Moving a term from the left to the right adds to the load. A consistent ordering of interior nodes turns these equations into a matrix system; it does not change the underlying stencil.
Definition 3.2.2
(Discrete inner product, energy and backward Euler step)
For zero boundary vectors, use the discrete inner product and norm . The discrete energy measures squared neighboring differences. For heat evolution , , backward Euler is
with the appropriate new-time boundary values included. Here denotes the time step and denote spatial spacings. Their roles must remain distinct when studying refinement.
Theorem/proposition
Theorem 3.2.3
(Two-dimensional comparison and Poisson convergence)
If in the interior and on the boundary, then everywhere. Consequently the Dirichlet system is uniquely solvable. If solves , has continuous pure fourth derivatives on the closed rectangle, and boundary/forcing samples are exact, then
Here means the exact nodal samples. The bound concerns this smooth rectangular Dirichlet problem, not arbitrary domains or weak solutions.
Theorem 3.2.4
(Discrete energy and backward Euler heat convergence)
For zero boundary vectors, unless . The backward Euler heat scheme is uniquely solvable for every . With zero forcing it satisfies
Its homogeneous-boundary step map is also nonexpansive in maximum norm. If the exact solution has bounded on , and its time-dependent Dirichlet data and forcing are sampled exactly, then
No coupling condition between and the spatial spacings is needed for this estimate.
Proof sketch or proof idea
Proof
(Minimum propagation and a quadratic barrier)
Suppose a negative minimum of occurs at an interior node. Each term is nonpositive, so there. Together with , all these terms vanish; every neighbor has the same negative value. Repeating along the connected rectangular grid reaches a boundary node, contradicting its nonnegative value. This propagation step matters: a non-strict inequality at a minimum alone would not yet be a contradiction.
Apply comparison to and for a homogeneous problem to obtain uniqueness. The interior system is square, so injectivity gives existence for every load. For an error bound, define
The centered differences differentiate these quadratics exactly, giving . Also on the boundary and . Taylor's theorem in each coordinate gives residual with
The error has zero boundary values and . Both and have nonnegative boundary values and nonnegative discrete Laplacian. Comparison gives , proving the stated estimate. The barrier is allowed to be positive on the boundary: it only needs to dominate the zero boundary error there.
Proof X-ray
How the barrier turns a residual into an error
The proof has two stages with different hypotheses. Comparison uses only the signs of the stencil, a positive center weight and nonpositive neighbor weights, together with connectedness: every interior node is joined to the boundary by a chain of neighbors. The propagation step is the non-obvious part. At a negative interior minimum, forces every neighbor difference to vanish, so the minimum spreads to the neighbors and eventually reaches the boundary, where . A block of interior nodes cut off from the boundary could hold a negative constant undetected. The second stage needs a barrier, a grid function with . The quadratic works because centered second differences are exact on quadratics, and its maximum is the stability constant. The key move is to compare with rather than with zero: turns the residual bound into a pointwise error bound. Smoothness enters only through , and only the pure fourth derivatives and appear; this is why the polynomial of the first worked example, of total degree four but with vanishing pure fourth derivatives, is reproduced exactly.
The heat-equation theorem combines the energy identity with a pointwise bound for the implicit step.
Proof
(Summation by parts and the implicit step)
For energy, one-dimensional summation by parts in every row and column gives
The edge sums include boundary edges. Vanishing energy makes all differences zero; the zero boundary then makes the whole vector zero. Thus is symmetric positive definite on interior vectors, and is invertible. Take its time equation's inner product with and use to obtain the energy identity.
For the maximum-norm statement, let , with zero boundary. At a positive maximum of , , so there. Apply the same argument to . Therefore . Inserting exact heat samples produces residual
bounded by the parenthesized expression in the theorem. Exact boundary sampling makes the error boundary zero even for nonzero time-dependent physical data. The error equation and inverse bound give . Sum in time to complete the proof.
Concept lensAnalytic
Energy and maximum norms see different features
The energy proof uses symmetry and an inner product; it controls the collective size of the grid vector. The comparison proof uses the signs of individual stencil weights; it controls every node. A different discretization can preserve an energy estimate while losing a maximum principle. These are related but distinct properties, so a squared-norm stability claim should not be silently promoted to pointwise monotonicity.
Worked examples
Worked example 3.2.5
(A complete four-interior-node Poisson solve)
On the unit square take , so both spacings are . Order the interior values as . With zero boundary data, the matrix is
Choose , so . Each of the four load entries is . Symmetry suggests equal unknowns, and substituting gives , hence . These are exactly the sampled solution values. Although has total degree four, its pure fourth derivatives in each coordinate vanish; the five-point stencil is therefore exact here. A mixed fourth-degree term does not by itself generate a truncation error for this operator.
Worked example 3.2.6
(Boundary elimination on a one-node interior)
Use the unit square with spacings , leaving only the center unknown . Give its left, right, bottom and top neighbors values , and set at the center. The equation is , so .
With instead, , the average of its four neighbors. This illustrates the discrete harmonic mean-value property. A positive source changes that average by ; a maximum principle for supplies a lower bound from the boundary, rather than asserting that every positive-source solution stays below its boundary maximum. The sign of the operator and of the source must be checked.
Worked example 3.2.7
(The same heat grid has low and high modes)
On the four-node grid of the first example, samples of are all and form an eigenvector with eigenvalue . Backward Euler with , multiplies this vector by ; forward Euler multiplies it by .
For forward Euler the weights are , , where and . Nonnegativity requires and proves maximum-norm nonexpansiveness. At the low mode still has factor , but the highest mode has eigenvalue and factor . Looking only at the smooth initial mode would miss an unstable perturbation. Backward Euler gives factors and , both positive and below one.
Boundary conditions change the algebra
For prescribed Dirichlet values, the proof above is complete. Neumann data prescribe an outward normal derivative instead. At the left boundary, a centered ghost relation for is , hence . The outward derivative is there; forgetting this sign imposes a different condition. Substitution changes both the stencil and load.
A pure Neumann Laplacian has constants in its nullspace. Its load must satisfy a discrete flux compatibility condition, and a mean value must be fixed for uniqueness. Boundary control volumes naturally produce half/quarter cell weights in a discrete energy inner product. Reusing the unweighted Dirichlet energy proof after changing endpoint rows is unjustified. Robin data , with , add a boundary contribution proportional to ; nonnegative such terms support coercivity, and a positive contribution can remove the constant nullspace. These observations explain the structure, but no general Neumann/Robin accuracy theorem is claimed here.
Tensor modes and what unconditional stability preserves
With zero Dirichlet data, the rectangular-grid vectors are eigenvectors, for , . Substituting into each difference quotient gives eigenvalue . Products appear because the two coordinate operators act independently, and their eigenvalues add. These vectors form a basis of the interior space by the one-dimensional sine orthogonality in each coordinate.
Consequently the largest eigenvalue grows like the sum of the inverse squared spacings. The explicit maximum-norm weight condition is a convenient mesh-uniform sufficient condition; for a particular finite grid the exact squared-norm stability condition is . These statements use different norms and should be labeled accordingly. A successful computation of one mode says nothing about a high mode introduced by roundoff or a boundary mismatch.
Counterexample mode
Squared-norm stability without nonnegativity
Claim: if the forward Euler heat step satisfies , it also keeps nonnegative data nonnegative. On the four-node grid of the first worked example the eigenvalues are ; take , so every factor lies in . Start from in the ordering . Then and
The norm decreased, yet a value became negative, since the center weight is . Nonnegativity needs the weight condition , here .
For comparison, Crank–Nicolson on homogeneous heat evolution is . Taking its inner product with proves that the squared norm cannot increase. Its factor on an eigenmode is . When is large, this factor approaches minus one. The energy inequality remains true while an unresolved component alternates and is only weakly damped. Backward Euler instead sends that factor toward zero. This calculation explains a method-selection issue without claiming a full Crank–Nicolson error theorem.
Finally, time-dependent boundary data are not a once-only initialization. At every implicit step their new-time contributions enter the right side. If those values are lagged at the old time, the algorithm solves a different boundary discretization; the exact-sample error equation proved above no longer applies unchanged. Boundary treatment is part of the numerical method and its order.
Boundary perturbations can be separated from interior residuals. If the Poisson error has boundary magnitude at most and interior residual bounded by , comparison with gives . A constant has zero discrete Laplacian, so the same barrier proof applies. This estimate shows why boundary sampling accuracy is part of the overall convergence rate, rather than an implementation detail outside the analysis.
Unequal spacings in the two directions must both be refined. If while stays fixed, the term in the theorem does not vanish, so this is not complete convergence for the continuous two-dimensional problem. The matrix has interior unknowns; a corner node lies on two sides but is still a single boundary value, not something to count twice. When planning memory and checking boundary elimination, these simple index relations prevent the same physical value from being counted twice.
The connectedness used by the comparison principle also matters in practice. If some coupling coefficients are deleted, or different boundary conditions are imposed on disconnected subgrids, minimum propagation can proceed only along the connections that remain. Each connected component needs suitable boundary control before the constant degree of freedom can be excluded in the same way. This is a structural hypothesis of the uniqueness proof, not merely a pictorial convenience.
Common mistakes
Common mistake
(Checking only one direction's step restriction)
Do not use the one-dimensional condition while ignoring : the center coefficient contains both. A condition sufficient for all meshes can be stricter than the exact spectral condition for one fixed small grid.
Common mistake
(Residuals of differently scaled equations)
Also, a residual in the unscaled matrix equation differs by spatial powers from the residual of an equation multiplied by ; the inverse estimate must use the same normalization.
Common mistake
(Unconditional stability is not unconditional accuracy)
Unconditional time stability does not mean arbitrary time steps are accurate. The theorem still contains an contribution. Likewise, backward Euler's energy dissipation does not establish second-order time accuracy. A Crank–Nicolson discretization can have an energy bound yet retain sign-alternating high-frequency components; energy and pointwise positivity must be assessed separately.
Summary
The five-point Dirichlet operator has a comparison principle, a positive energy and a uniquely solvable linear system. A quadratic barrier converts its smooth-solution residual into an explicit second-order spatial error bound. Backward Euler adds a nonexpansive time step and yields first-order time convergence without a spatial coupling restriction. Boundary type, data sampling and norm remain part of every statement.
Exercises
Quick check
(On with , , find the largest time step giving nonnegative forward Euler heat weights for .)
Use both coordinate contributions.
Solution · Answer
The condition is . Thus , giving . Checking only the restriction would allow , which makes the combined center coefficient negative.
Quick check
(On the unit square, use the barrier theorem when , and . Simplify the error bound.)
The two derivative bounds contribute separately.
Solution · Answer
Here , and the derivative expression is . Hence . This bound assumes exact boundary and forcing samples and the smoothness stated in the theorem; errors in those inputs would add further contributions.
Quick check
(Verify the discrete summation-by-parts identity for three interior values with in one dimension.)
Expand the edge squares and compare coefficients.
Solution · Answer
The left side before division by the spacing is . The edge sum is , which expands to exactly the same expression. Including both boundary edges is essential.
Quick check
(A pure Neumann discrete Laplacian is symmetric, annihilates constants and has no other null vectors. Prove the necessary load compatibility condition and explain how fixing the mean restores uniqueness.)
Test the equation with the constant vector.
Solution · Answer
If , symmetry and give . Thus the load must have zero sum in this symmetric normalization; a weighted control-volume formulation uses the corresponding weighted sum. If two solutions have the same prescribed mean, their difference is a constant null vector with zero mean, hence zero. On the mean-zero subspace, symmetry and the stated nullspace condition imply invertibility onto the mean-zero loads.