Why estimate the error after solving?
An a priori estimate describes how an approximation should improve as the mesh becomes fine. Its right side often contains a derivative of the unknown exact solution. An a posteriori estimate instead uses the computed function, the specified differential equation and the current mesh. Such an estimate can provide an error certificate for a particular calculation and suggest where further resolution may help. The word “posteriori” does not make every residual a certificate: a stability argument must connect that residual to the norm of error we want.
We establish that connection for the interval Poisson equation. The result is deliberately restricted to a model for which every step and constant can be checked. It does not establish a complete adaptive finite-element theory.
The problem, mesh and computable quantities
Definition 5.4.1
(Weak interval Poisson problem)
Let and prescribe exact endpoint values . The weak problem seeks with , satisfying
Here consists of square-integrable functions with a square-integrable weak first derivative. In one dimension each such function has an absolutely continuous representative, so its nodal values are well defined. The zero subscript specifies zero endpoint traces. The derivative norm is a norm on ; integrating from the left endpoint proves a Poincaré inequality. Subtracting the affine boundary lift and applying the Hilbert-space representation theorem gives unique solvability. Moreover the equation implies weakly, so . Our reliability proof needs only and the weak equation, rather than a bound on its unknown second derivative.
Choose any finite partition , with and . No uniformity or bound on neighboring length ratios is assumed. Choose an integer . The trial functions are continuous, polynomial of degree at most on each cell, and have the prescribed endpoints; is their homogeneous test space. Let satisfy the Galerkin equations with exact stiffness and load integration. The constrained stiffness matrix is positive definite, so exists uniquely.
Definition 5.4.2
(Energy error, cell residual and indicators)
For the Galerkin solution , the energy error and cell residual are
The second derivative is taken separately inside each cell. The local indicators and their total are
For the cellwise second derivative is zero; it does not include a distribution concentrated at a node. These quantities are computable when the cellwise integrals of the squared residual are available. The Galerkin load vector alone does not supply them: in particular, it does not determine . If their evaluation is approximate, a rigorous numerical certificate additionally needs bounds for that evaluation error. A sampled approximation of an integral is not automatically an upper bound.
Where the nodal jumps went
Lemma 5.4.3
(Residual representation and Galerkin cancellation)
For let , let be the derivative jump at the interior node , and let be the continuous piecewise-linear function taking the same values as at every mesh node. Then
If satisfies the Galerkin equations exactly, then for every
Proof
(Proof by cellwise integration by parts)
On each cell , integration by parts gives . Summing over the cells, the right boundary term of the left cell and the left boundary term of the right cell combine at into . Endpoint contributions vanish because . Adding gives the first formula.
The interpolant is continuous, linear on each cell and zero at both endpoints, so it belongs to , even when the method uses degree . Galerkin orthogonality gives , hence . Apply the first formula to : it is zero at every node, so every jump term disappears.
Common mistake
(Declaring the derivative jumps to be zero)
A continuous finite-element function generally has a jump in its derivative at each interior node, so the jump terms in the residual representation are not zero in themselves. In the nonuniform example below . The jumps drop out of the final formula only because the test function vanishes at the nodes, and that step is available only after the Galerkin equations have removed .
This is a special advantage of interval elements: point evaluation is continuous on , and an exact nodal interpolant is available. The same step cannot simply be transferred to two-dimensional functions, where vertex values need not be defined. Higher-dimensional residual estimators usually retain edge or face jump terms and require a different interpolation argument.
Concept lensAnalytic
The energy error as the size of the residual
The weak equation turns the residual into . By Cauchy–Schwarz, with equality at ,
So is the dual norm of a computable functional, and any bound for all bounds . The reliability proof, run with an arbitrary in place of , gives exactly such a bound with . Restricted to , the same supremum is the algebraic term below.
Reliability with an explicit constant
Theorem 5.4.4
(A one-dimensional residual upper bound)
Under the stated interval, forcing, exact endpoint and exact Galerkin integration assumptions, on any finite partition and for any fixed degree ,
This is a reliability bound: the right side is a guaranteed upper bound in the derivative norm in exact arithmetic. The constant is valid and need not be sharp. No lower bound, optimal mesh rate or adaptive convergence conclusion is part of this statement.
Proof
(Proof with an explicit local constant)
First prove a local inequality without importing an unknown interpolation constant. If and , then for , Cauchy–Schwarz applied on both sides of gives
Add and rearrange to obtain . Integrating in , and using , proves
Next, on a cell, . This is the average of , so the constant and its mean-zero remainder are orthogonal in . In particular,
Put . The weak equation and the residual lemma now give
If , the result already holds. Otherwise division proves the theorem.
Proof X-ray
Where each hypothesis is spent
Each hypothesis is used at one identifiable line. Exact endpoint values put in , so the weak equation may be tested with , giving . Exact Galerkin integration is spent on , and the jumps then disappear because vanishes at every node, which uses that point values are defined on . The constant comes from the local inequality applied to on each cell, and the cell projection identity replaces by without any interpolation constant.
The argument closes because the unknown appears once on each side, so division leaves a bound containing only the data, the mesh and the computed function, although the unknown error served as the test function. Remove the Galerkin hypothesis and the term survives. The hat counterexample below has and error two; the algebraic term introduced after it is exactly the missing line restored.
A calculation that separates nodes from functions
Worked example 5.4.5
(Constant forcing on a nonuniform mesh)
Take , zero endpoints and linear elements on nodes . The interior hat has stiffness and load , hence its coefficient is . The exact solution is , and . The computed slopes are and .
On a cell of length , the exact derivative differs from its cell average by a centered linear function. Its squared integral is . Therefore
The guaranteed squared bound is , twice the actual squared error. Its effectivity ratio, meaning the upper bound divided by the actual error, is in this example. A nonsharp certificate can still be useful. Exact nodal values have not made the function exact: between nodes a chord differs from the curved solution.
The jump is . The hat value at the interior node is one, so . This checks the jump sign and exhibits the discrete cancellation directly.
For this constant-coefficient interval problem the linear Galerkin solution is, in fact, the nodal interpolant of for any forcing. To see why, each test derivative is constant on a cell, while . Thus for every linear test function, and discrete uniqueness gives . This identity is special to this model. It does not mean that nodal agreement controls the error without a within-cell estimate.
What an indicator can suggest
The two contributions in the example are and . The longer cell contributes of the total. A bulk marking rule chooses a set with for a specified . For , the longer cell alone qualifies. The choice of such a set is an algorithmic decision, not an extra conclusion of reliability.
Worked example 5.4.6
(Bisecting the marked cell)
Bisect the longer cell, giving nodes and cell lengths , and solve again. With and linear elements the residual is , so , and the Galerkin solution is again the nodal interpolant, so each cell contributes to . Hence
Compared with , the squared indicator falls by the factor . Bisecting only the shorter cell would instead give , a factor : one new node placed by the indicator achieves almost all of the available reduction, and one placed against it achieves almost none.
This calculation explains the value of local information in this example. It does not prove efficiency for general forcing or convergence of a repeated solve–estimate–mark–refine procedure. Such claims require additional arguments. In particular, the reliability theorem alone supplies no lower bound comparing with .
For linear elements , so these indicators depend only on the forcing and mesh, despite being derived from the residual of the solved problem. This simplification belongs to the one-dimensional model. For quadratic elements the computed generally matters. With constant forcing the exact solution is quadratic and belongs to that trial space, giving both and under exact integration.
Counterexamples and two error budgets
Counterexample mode
Zero volume residual without a Galerkin solve
Set , zero endpoints and nodes . The exact solution is zero. Let a proposed approximation be the hat function with midpoint value one. Its cell second derivatives vanish, so its volume indicator is zero. Its derivative norm is , so its error is two. The jump is , and : it does not satisfy the Galerkin equations. The theorem's discrete-solve hypothesis is essential. Conversely, jump terms alone cannot certify error: with and a single linear cell there are no interior jumps and , yet .
An algebraic residual can restore the missing information. For a basis of , let ; this matrix is positive definite.
Proposition 5.4.7
(Certificate for an inexact discrete solve)
Let be any conforming trial function with the exact endpoint values, let be the indicator formed with , and let
with when . Then
Proof
(Proof by keeping the discrete residual)
Let , the residual functional of ; its cell residual is and its jumps use . For we have and , so Cauchy–Schwarz in the inner product gives . Put and . Then . The first term is at most . The second is handled exactly as in the reliability proof and is at most . The cell projection identity gives both and , so ; divide when .
This is a mathematical certificate, but evaluating may itself require a solve or a separately certified bound. In the hat counterexample , and , recovering the exact error.
Common mistake
(Replacing by the Euclidean residual)
A small does not by itself make small. One has , with equality when is an eigenvector for the smallest eigenvalue. For equal linear cells and , which is about . So can exceed by a factor of about , and the factor grows under refinement. Use itself or a justified eigenvalue bound.
Worked example 5.4.8
(A bound that splits into two budgets)
Take , zero endpoints and nodes , and suppose an iterative solver returns the midpoint value instead of the Galerkin value . Then has cell residual on both cells, so . The residual vector is , and , so . The certificate is
For the exact error, Galerkin orthogonality splits into energy-orthogonal parts, where . The first part has squared norm , and the second has squared norm . The error is therefore , below the certificate. The term accounts for the unfinished solve and for the discretization.
There is a different issue when a calculation replaces by cell averages . Define using and define the data oscillation
Evaluating this oscillation likewise requires the integrals of , beyond just the load vector. The triangle inequality in the weighted product of cell spaces gives . If solves the original Galerkin problem exactly, then . If instead it solves the modified problem, it need not be Galerkin for ; use the algebraic term formed with the original as well. This distinction prevents changing the problem silently while evaluating its error.
Practice
Quick check
(For , zero endpoints and linear elements on equal cells, find , the exact energy error, and the reliability upper bound.)
Integrate the squared centered linear derivative error on one cell, then sum.
Solution · Answer
With , . The exact solution is , and its linear interpolant is the Galerkin solution. Each cell contributes to the squared derivative error. Thus and the upper bound is . Their ratio is for every . This establishes that ratio only for this example, rather than proving general estimator efficiency.
Quick check
(On nodes with , can the short cell alone satisfy the bulk rule with ? If both cells are bisected instead, by what factor does change?)
Compare squared contributions rather than their square roots.
Solution · Answer
The short-cell fraction is , so it cannot qualify alone; the long-cell fraction is . Bisecting both cells gives lengths and , one quarter of the original , because every is divided by eight while the number of cells doubles. Bisecting the long cell alone already gave the factor with one new node instead of two. The exact squared errors have the same ratios here, since for this constant-forcing example.
Quick check
(Use one linear cell, zero endpoints and . Find , the cell-average forcing, the oscillation and the exact energy error. What fails if the oscillation is discarded?)
The homogeneous one-cell linear test space is zero dimensional. Solve the continuous equation explicitly.
Solution · Answer
The only admissible one-cell linear trial function is zero, so . The cell average of is zero, giving , while . The exact solution is , whose derivative has norm . The full bound is and is valid. Dropping oscillation would incorrectly certify zero error. Here even the modified and original discrete problems coincide because their test space is zero; the missed information lies entirely within the cell.
Quick check
(On nodes take and a proposed midpoint value . Compute , , and the exact error. Explain why the algebraic term has the right scale.)
The constrained stiffness matrix has the single entry four.
Solution · Answer
Write . Since the cell residual is zero, . The original load is zero, , and . Thus . Since the exact solution is zero and , the actual error is also . The algebraic term recovers precisely the contribution that nodal interpolation could not cancel for this unsolved discrete equation.