Introduction
A finite-difference formula turns a differential equation into algebra, but the resulting algebraic system still has to be solved. This section follows both steps for the Dirichlet Poisson problem on a square. It distinguishes three objects: the exact function, the exact solution of the grid equations, and an approximate iterate. The first two differ through discretization; the last two differ through the linear solver.
Motivation
Suppose the boundary temperature is prescribed and a steady heat source acts inside a thin plate. After suitable normalization, a model is . At an interior node the five-point stencil compares its value with four neighbors. A boundary value is known data, even when it appears next to an interior unknown. We will prove that the resulting system has a unique solution, bound the grid error, and explain why a simple iterative solver can become slow on a fine grid.
Definitions and setup
Definition 3.5.1
(Five-point Dirichlet problem)
Let , let , and put . At grid points , prescribe whenever or is or . For define
The discrete problem is , with . It is convenient to regard boundary entries as part of the displayed stencil; when a matrix acts only on interior unknowns, move known boundary terms to the right. The maximum norm below includes boundary nodes. For zero boundary vectors, the interior Euclidean norm is .
Definition 3.5.2
(Jacobi iteration)
Starting from any interior vector , keep the boundary values fixed and update all interior points simultaneously, using only values from step on the right:
This is the interior equation solved for the center value, with the neighbors taken from the previous iterate.
Theorem/proposition
Theorem 3.5.3
(Maximum principle and second-order error)
If at every interior node and on the boundary, then throughout the grid. Consequently the discrete Dirichlet problem has exactly one solution. Suppose solves with boundary data , and , . If boundary values and samples of are exact, then
Theorem 3.5.4
(Jacobi convergence on the square)
For fixed , Jacobi iteration for the interior system converges from any initial interior vector. If is the iterate minus the exact discrete solution, then
The factor depends on the grid and tends to one as ; this bound is not a mesh-independent iteration count.
Proof sketch or proof idea
Proof
(Neighbor propagation and a one-directional barrier)
For the maximum principle, suppose a positive maximum occurs at an interior node. The index grid is connected through neighbor steps, so every interior node is joined to the boundary by such a path. Since every neighbor is no greater, minus their sum is nonnegative. The hypothesis makes it nonpositive, so all four neighbors attain the same maximum. Repeating along a path to the boundary contradicts nonpositive boundary values. Applying this result to the difference of two solutions and its negative proves uniqueness. The interior system is square, so its trivial kernel also proves existence.
Centered Taylor expansions in the coordinate directions give, for sampled ,
Let and . Although is not zero on the horizontal sides, it is nonnegative on the entire boundary. It depends only on , so its -differences vanish, and the -difference is exact on the quadratic ; hence exactly. Thus and on the boundary; the same holds for . The maximum principle gives . This proves the stated constant. The argument also says that a zero-boundary grid vector with satisfies .
Proof X-ray
A barrier that need not vanish on the boundary
The comparison argument needs only one-sided boundary information. Since on the boundary and there, both and are nonpositive on the boundary, whether or not vanishes. That freedom is what allows a barrier depending on alone: the five-point operator applied to it reduces to the one-dimensional second difference, which is exact on quadratics, so no new error term appears.
The hypotheses enter at three places. Smoothness of supplies the consistency bound . Exact boundary data make vanish on the boundary. Connectedness of the grid through neighbor steps lets a plateau of maximal values propagate until it meets the boundary. The maximum principle is applied twice, to and to , because the residual bound is two-sided. The constant is from Taylor's theorem times .
The Jacobi estimate rests on a different tool: an orthogonal basis of eigenvectors in place of a comparison argument.
Proof
(Sine modes diagonalize the Jacobi error)
Subtract the discrete solution, which is a fixed point of the Jacobi update. The boundary error is zero, and the interior error satisfies . The sine arrays , , form an orthogonal basis of the interior vectors. The Jacobi error matrix sends each array to itself times
The largest absolute value is , attained at both and . Because and the sines vanish at and , the boundary zeros are respected. Expanding in the orthogonal basis, each coefficient is multiplied by at every step, and Parseval's identity proves the norm estimate. The high-frequency mode can alternate in sign while decaying; convergence is not the same as monotone pointwise change.
Concept lensAlgorithmic
Jacobi as explicit pseudo-time stepping
Apply the explicit forward-time scheme to the auxiliary heat problem with the given boundary data, using the five-point Laplacian and ratio :
With , the largest ratio for which all five weights are nonnegative, the center weight vanishes and . The step is then exactly the Jacobi update. Solving the Poisson system by Jacobi is marching a heat equation to its steady state with the largest step that keeps the weights nonnegative. The pseudo-time step is , while the slowest smooth mode decays on a time scale independent of , so the number of iterations grows like . Indeed .
Worked examples
Worked example 3.5.5
(A complete one-node solve)
Let , impose zero boundary values, and take . Then . At the only interior node, and the stencil gives , so , exactly the sampled value. Here both fourth pure derivatives vanish. Exactness for this polynomial does not imply exactness for general smooth solutions.
Worked example 3.5.6
(An algebraic residual and the error it controls)
Let have the correct boundary values, and suppose on interior nodes. If is the exact grid solution, then has zero boundary and . This is a solver error bound, not a bound on alone. Add the discretization bound to it. If a solver reports the residual of the unscaled matrix , divide that residual by before using this estimate.
Worked example 3.5.7
(Why convergence can be slow)
With , the slowest Jacobi factor is . With it is approximately . Thus one hundred iterations reduce a worst-case error component by only about on the finer grid. The common phrase “Jacobi converges” is correct for each fixed grid, but says little about the cost of refining it. Faster methods require their own analysis.
Common mistakes
Common mistake
(Dropping boundary values from the right-hand side)
When the matrix acts only on interior unknowns, each boundary neighbor of an interior node contributes a known term to the right-hand side. Omitting these terms solves the problem with zero boundary data, which is a different boundary-value problem, not a slightly less accurate version of the same one.
Common mistake
(Mixing new and old values in a Jacobi sweep)
Jacobi uses only values from iteration on the right. Overwriting the array in place, so that some neighbors already hold iteration , produces a Gauss–Seidel sweep. That is a different method, and the factors above no longer describe it.
Counterexample mode
A small update does not certify a small error
A common stopping rule claims that once falls below a tolerance, the iterate is that close to the discrete solution. Take and an error in the slowest mode, so . Since the update equals ,
The update passes a tolerance of , while the error is about two thousand times larger than the update. The repair uses the spectrum: the error matrix is symmetric with eigenvalues of magnitude at most , so every eigenvalue of is at least , and
In the example this bound gives , so it is sharp.
The error theorem also does not cover nonsmooth solutions, inexact data, irregular domains or a non-square mesh without a new argument.
Summary
The five-point operator has a discrete maximum principle. Together with a quadratic barrier, it provides uniqueness and a second-order nodal error bound under explicit smoothness assumptions. Jacobi converges on each fixed square grid, but its sharp worst-case factor approaches one under refinement. Discretization and solver errors need separate accounting.
Exercises
Quick check
(On the unit square take , and boundary values . Find the interior value and explain why it is exact.)
Use all four boundary neighbors.
Solution · Answer
The neighbors have values , , , , so their average is . The harmonic function has precisely these boundary values and its samples satisfy the stencil; uniqueness makes the numerical value exact.
Quick check
(Suppose , , and . Give the theorem's discretization bound. If a correct-boundary approximate solution satisfies on interior nodes, bound the total nodal error.)
Combine two different bounds using the triangle inequality.
Solution · Answer
The discretization error is at most . The algebraic error is at most . Hence the total nodal error is at most , provided the theorem's smoothness and exact-data assumptions hold.
Quick check
(For , find the Jacobi factor for the mode . What happens to this error mode after two iterations?)
The sign of a factor matters as well as its magnitude.
Solution · Answer
. Each iteration reverses the mode's sign and halves its amplitude; after two iterations it is of its initial value. A negative factor does not prevent convergence when its magnitude is below one.