Why do computed nodes sometimes agree exactly?
In a one-dimensional Poisson problem, a piecewise-linear finite-element curve can pass through the exact solution at every mesh node while differing from it between nodes. This is more than a coincidence for a constant diffusion coefficient and exact integration. It is also a warning: agreement at nodes is a surprisingly weak test of the full approximation. We will prove the identity and use it to derive error bounds without assuming a uniform mesh.
Problem and interpolation
Let , , , and . For , let satisfy
Let be the continuous piecewise-linear functions on this mesh with zero endpoint values. The Galerkin solution satisfies the same equation for every ; all integrals are exact. The one-dimensional representative of an function is continuous, so its nodal values are meaningful. Write for the piecewise-linear interpolant of these values. For this model, weak regularity gives and almost everywhere: indeed, the weak equation identifies its distributional second derivative as .
Theorem 3.4.1
(A special nodal identity)
For every such and every finite mesh, . In particular, at every node.
Proof
(Element-by-element Galerkin orthogonality)
On each , is constant and . The fundamental theorem for the one-dimensional Sobolev representative therefore gives
Summing over and using the weak equation for with gives , so satisfies the Galerkin equation. The Galerkin solution is unique: the difference of two solutions satisfies , so is constant, and the zero boundary condition makes it zero. Thus . Neither constant forcing nor equal element lengths was used.
Proof X-ray
Which hypotheses make the cancellation exact
Three features of the model are used, each at one place. First, point values of exist because functions are continuous in one dimension, so is defined. Second, the coefficient multiplying is the constant and is constant on each element; together they reduce the element integral to a difference of endpoint values, which vanishes because and agree at the nodes. This is the key move. Third, the load enters only through the weak equation of , so the exact load functional is what the Galerkin system must contain. Each failure has a concrete form below: a variable coefficient breaks the second step, and midpoint quadrature breaks the third. In two dimensions the second step has no analogue, since an element integral of a gradient against a constant vector is a boundary integral over the element edges, not a difference of nodal values.
Explicit error estimates
Theorem 3.4.2
(Energy and estimates on arbitrary meshes)
For the same problem,
The constants here are deliberately simple; they are not asserted to be sharp.
Proof
(Local bounds from a zero-mean derivative)
On an interval of length , put . Since , the mean of is zero. For almost every ,
Here the slope of is constant and absolute continuity of justifies the difference formula. Integrating the squared bound over yields . Also , so Cauchy–Schwarz gives . Sum the squared local estimates, use and , and substitute . No mesh-ratio assumption is needed in this one-dimensional estimate.
Worked example 3.4.3
(Exact errors for constant forcing)
For on a uniform mesh of elements, and . On an element with local coordinate , the interpolation error is . Thus
Adding the identical contributions gives and . The nodes are exact even though both errors are nonzero. Halving halves the energy error and quarters the error in this example.
Worked example 3.4.4
(A nonconstant load on two elements)
Take and nodes . The exact solution is . Integrating the middle hat against gives load , and its stiffness entry is . Therefore . The identity applies even though the load is not constant.
Worked example 3.4.5
(A nonuniform mesh and the one-element limit)
Take with nodes , so , and . The single interior hat has stiffness and load , half the base of the hat. Hence , and indeed . The element-by-element error formula of the first example gives
so the energy error is about , below the bound . In the extreme case of one element there is no interior node, and . The identity still holds, and the energy error is , again below .
When the identity fails
Counterexample mode
Inexact load integration
In the last example, applying midpoint quadrature separately on the two elements gives the load , not . The computed middle value is then , rather than . The algebraic system can be solved exactly and still miss the exact nodal value because it represents a different load functional.
The explorer below solves the piecewise-linear system exactly on a chosen mesh and compares every nodal value with , for exact and midpoint loads and for a variable coefficient.
For variable diffusion, the same cancellation need not hold: the test derivative is constant on each element, but the coefficient multiplying it is not. For instance, solve with , . The exact solution is . On two equal elements, exact assembly gives left and right stiffness weights and , so the interior value is , not . Nonzero boundary data are incorporated into the trial space here. Thus the earlier theorem cannot be transferred to variable coefficients merely because the functions are piecewise linear.
The theorem assumes the exact Galerkin system, not floating-point arithmetic, approximate quadrature, or an iterative solve stopped at a tolerance. It also does not establish analogous nodal exactness in two dimensions. The measured algebraic residual and the continuous discretization error answer different questions.
Common mistake
(Reading nodal agreement as a small error)
A plot of computed against exact values at the nodes shows zero error for every mesh in this model, whatever the mesh size. The approximation error lives between the nodes: in the constant-forcing example the energy error is . Convergence must be measured in a norm, such as the energy or norm, not by nodal agreement.
Exercises with solutions
Quick check
(For and nodes , find the sole interior Galerkin value and explain why no matrix computation is necessary.)
Solution · Answer
The exact solution is . By the nodal identity, . As a check, the stiffness is and the load is , giving .
Quick check
(With uniform , how do the exact errors in the constant-forcing example change under two successive halvings of ?)
Solution · Answer
The energy error is multiplied by twice and the error by twice. The final errors are and of their initial values.
Quick check
(For , find and its value at . State what the theorem predicts for the two-element Galerkin value.)
Solution · Answer
Integrating with gives , so . With exact integration the Galerkin value at is also . Directly: the hat load is and the stiffness is .
Quick check
(Why does a plot that compares only computed and exact nodal values fail to measure the energy error in the first worked example?)
Solution · Answer
Every nodal error is zero, whereas the derivative error between nodes has squared integral over the whole domain.