Motivation
A local triangle stiffness matrix is only one ingredient of a finite-element method. A convergence proof must explain why the assembled space approximates the exact solution, why the weak problem is stable and which geometric constants remain controlled as the mesh changes. Thin triangles can make interpolation gradients large even when nodal values look harmless.
We develop continuous piecewise-affine approximation for the Dirichlet Poisson equation on a polygon. The numerical proof is complete relative to explicitly stated analytical prerequisites: Sobolev traces and point evaluation on a reference triangle, a reference polynomial approximation lemma, and, only for the stronger result, a dual elliptic regularity assumption. Naming these inputs prevents a regularity theorem from being hidden inside an error assertion.
The results concern a specified conforming linear method, not every finite-element variant or unstructured-mesh algorithm.
Definitions
Definition 5.5.1
(Weak Dirichlet Poisson problem)
Let be a bounded connected polygonal domain with the Poincaré inequality on . Solve , , with zero boundary trace. Set and ; the energy norm is .
Definition 5.5.2
(Shape-regular conforming triangulation and linear element space)
A conforming triangulation covers the polygon by nondegenerate triangles meeting only at complete common edges or vertices. There are no hanging nodes. Its largest diameter is , and shape regularity means for every triangle, where is its inradius and is fixed throughout refinement. Let be continuous functions affine on each triangle and zero on the boundary. These functions belong to even though gradients can jump across edges.
For a continuous sufficiently regular function, matches its vertex values. The local nodal basis consists of barycentric coordinates. We use exact coefficient/load integrals; quadrature perturbations are not silently included in the theorem.
Below denotes one fixed global gradient-interpolation constant obtained in the scaling proof. Best approximation gives the energy estimate with that same constant, which is then used again for dual interpolation.
The analytical prerequisites used below are:
- The Poincaré inequality and completeness of in the gradient norm, together with the Hilbert-space representation theorem.
- On the fixed reference triangle , the embedding , making vertex evaluation continuous.
- The degree-one reference polynomial approximation lemma, often called the Bramble–Hilbert lemma: .
These are named analytical inputs, not results proved by the local matrix calculation. The following argument derives the numerical error estimate from them.
Theorem/proposition
Theorem 5.5.3
(Conforming Galerkin existence and energy convergence)
The weak Poisson problem and its conforming discrete problem have unique solutions. Suppose additionally the exact solution belongs to and has the continuous representative needed for nodal interpolation. On a shape-regular family,
The constant also includes the fixed reference-element analytical constants; it does not depend on mesh size. Poincaré then gives an error of order without any dual regularity assumption.
Theorem 5.5.4
(Second-order error under an explicit dual assumption)
Assume further that every has a zero-boundary weak dual solution of satisfying . Then
The regularity estimate is an additional analytical premise, not asserted for arbitrary polygons with reentrant corners.
Proof sketch or proof idea
Proof
(Coercivity, orthogonality and affine scaling)
Poincaré makes a complete inner product on , and is continuous there. The representation theorem gives the unique weak solution. Restricting to the finite-dimensional conforming space gives a positive definite interior stiffness system and a unique . Subtraction gives . Consequently, for every ,
Thus energy error is the best available energy approximation. It remains to prove that approximation is small; this is where mesh geometry enters.
On the reference triangle, the interpolation map is bounded because its three coefficients are continuous vertex evaluations and its basis functions are fixed. It reproduces every affine polynomial. Therefore, for any ,
after taking the infimum. This is an actual consequence of the stated embedding and polynomial approximation input, rather than an unexplained interpolation claim.
Map to a physical triangle by , and write . Shape regularity gives and . The chain rule and change of variables yield
Combine the inequalities: the determinant cancels, while . Sum over all triangles and use . Continuity makes vertex values well defined; the two shared vertex values then determine the same affine trace from both sides of every shared edge. Zero boundary trace makes boundary nodal values zero, so . Best approximation with proves the first theorem.
Proof
(Duality argument for the estimate)
For the result, put and solve the stated dual problem with . Its weak equation and Galerkin orthogonality give
Divide if the error is nonzero and apply the energy estimate. This proves the second theorem and displays precisely where dual regularity is used.
Proof X-ray
Which facts belong to analysis, geometry and algebra?
Each ingredient of the energy proof does one job. Coercivity, from the Poincaré inequality, is spent on solvability and on the Pythagorean identity; it is why the energy error equals a best-approximation error, with constant one. Conformity is spent twice: makes Galerkin orthogonality legitimate, and the shared vertex values on each edge put back into , so that it can serve as a competitor. The analytical inputs act only on the fixed reference triangle. The embedding makes bounded, and because reproduces affine functions, for every ; this is the move that lets the Bramble–Hilbert lemma replace a full norm by the seminorm. Shape regularity is spent in a single line, , after the determinant factors of the two changes of variables cancel. The skinny-triangle example shows this line failing: there grows much faster than . The duality proof reuses the same interpolation constant, now for , and spends the dual regularity hypothesis only on . Assembling a stiffness matrix establishes none of these steps.
Worked examples
Worked example 5.5.5
(Assemble every node of a four-triangle mesh)
Split the unit square into four triangles by connecting its corners to . On the bottom triangle, ordered as its two corners and , the basis gradients are , and , and its area is . Its local matrix is
Order the global corners cyclically, then the center. Adding all four element contributions gives
for constant forcing one. Before boundary constraints this matrix has the constant null vector. With all four corner values zero, only the center equation remains: , so . The discrete energy and load work both equal . Boundary reactions are retained as flux balance, not incorrectly imposed as additional free-node equations.
Concept lensGeometric
Stiffness entries read from the angles
For linear elements the local matrix has a coordinate-free description. If vertices of face the angle at the third vertex, then , and each diagonal entry is minus the sum of the other entries in its row. In the bottom triangle the corner–corner entry is because the angle at is , and each corner–center entry is because the opposite angle is . In the skinny triangle below, the angle at tends to , and the entry coupling with equals .
Worked example 5.5.6
(An exact smooth manufactured forcing)
Take , whose forcing is . It is smooth and has zero square-boundary trace. On the bottom triangle use reference coordinates , , with Jacobian determinant and center basis .
The local center load is the reference integral of . Using gives . Symmetry gives total center load , hence . The exact center value is , so the nodal difference is . The full method has a convergence theorem; it does not promise exact nodal values for every polynomial forcing.
Worked example 5.5.7
(Skinny geometry destroys a uniform interpolation constant)
Take triangle vertices and . Its nodal interpolant is . Thus the second component of the gradient error has magnitude . The area is , giving
The diameter stays of order one, so a uniform estimate with would have a right side tending to zero and a left side bounded away from zero. Shape regularity cannot simply be deleted. This is a counterexample to that uniform local constant, not a claim that every mesh containing any narrow triangle necessarily fails to converge.
What the global theorem requires of implementation
A conforming boundary edge has two zero nodal values, so its affine trace is zero along the whole edge. A shared interior edge has the same two nodal values from both adjacent elements, so the assembled function is continuous there. These simple facts establish membership in ; continuity of gradients is not required. Hanging nodes need constraints or a different space analysis and are excluded here.
The geometric constant is a family property. Every single nondegenerate finite mesh has some finite worst aspect ratio, but that alone does not keep the interpolation constant bounded as a sequence is refined. The theorem fixes and lets decrease. Element counts or a visually dense mesh cannot substitute for that condition.
Exact integration is also explicit. On the constant-load example the local integrals are elementary, while the manufactured forcing needs higher polynomial moments. A quadrature rule that does not integrate the required degree changes the load and may change the bilinear form. Its consistency error can be analyzed, but exact Galerkin orthogonality no longer applies without that perturbation term. A local matrix copied from a reference triangle must also be transformed by the physical Jacobian; copying it unchanged onto a different geometry is not assembly of the stated method.
Finally, the dual regularity premise is logically separate from primal solvability. The weak solution exists on the assumed polygon by coercivity, but corner singularities may prevent the regularity used for the displayed rates. The first theorem explicitly assumes primal , and the stronger estimate additionally assumes the dual bound. No unproved corner-regularity statement is being smuggled into either conclusion.
Counterexample mode
A reentrant corner can remove second-derivative regularity
Claim: on every polygon a weak solution with forcing lies in . Take the L-shaped domain , polar angle , and . It is harmonic and vanishes on both edges at the corner, with integrable but , so is a multiple of . With a smooth cutoff equal to one for and zero for , solves for a smooth supported in , yet . The repair is the theorems' own premise: primal and the dual bound are assumed, or proved for the given domain.
A direct derivation of the reference affine approximation input
For this degree-one setting, the named polynomial approximation lemma can itself be reduced to ordinary mean-zero Poincaré estimates on the fixed reference triangle. Let be its centroid, and define as the mean of plus the mean gradient of dotted with . Then has zero mean, and each component of its gradient also has zero mean. Apply Poincaré to each first derivative: the gradient error is bounded by a fixed constant times the Hessian seminorm. Apply it again to : its norm is bounded by a fixed constant times that gradient error. The second derivatives of are zero, so the norm of is bounded by a fixed multiple of . This proves the displayed reference approximation lemma in the present setting.
The argument uses a single fixed triangle. Its constants are not asserted for arbitrary degenerating physical shapes; the affine scaling step is precisely what transfers the estimate with a controlled shape parameter. The Sobolev embedding needed to evaluate vertices remains an explicitly named analytical prerequisite. It cannot be replaced by saying that an arbitrary weak function has nodal values. Separating these two inputs makes clear why the proof is valid for the assumed solution but is not a complete interpolation theory for all data.
Nonzero boundary functions require a compatible lift before this homogeneous trial space and its error proof can be applied.
Here the seminorm counts only second derivatives, whereas the full norm also counts the function and its first derivatives. The reference approximation lemma first removes the affine part, and only then is the full norm of the remainder controlled by the second-order seminorm. Without subtracting a polynomial no such inequality is possible: a constant function has zero second-order seminorm but nonzero full norm. This is exactly where reproduction of affine functions does its work in the interpolation proof.
The shape scaling must track area and derivatives together. The area contributes the determinant, the gradient contributes the inverse matrix, and the second derivatives contribute two factors of the forward matrix. Only after these factors are written out one by one do the cancellation of the determinant and the single remaining power of the mesh length become visible. Saying merely that the elements are small, so the error is small, misses the counterexample in which a thin shape makes the inverse matrix large.
The five-node example also deliberately lists the full matrix before eliminating the boundary. Zero row sums of the global matrix express that a constant has zero gradient energy; after the Dirichlet values are imposed, the number of unknowns drops and the interior system becomes positive definite. The local, the unconstrained global and the constrained interior matrices are three different objects with different nullspace statements, and the singularity of one cannot be used to deny uniqueness for another; this distinction matters especially when checking an assembly in practice.
Common mistakes
Common mistake
(Semidefinite global matrix versus definite interior matrix)
Do not confuse the semidefinite unconstrained global matrix with the positive definite Dirichlet interior matrix. Before boundary conditions are imposed, the assembled matrix annihilates constant vectors; every row of the five-node matrix sums to zero. Only after the Dirichlet values are fixed does the remaining interior matrix become positive definite, because zero energy forces a constant function and the boundary values make that constant zero.
Common mistake
(An exact patch test is not an approximation proof)
Do not interpret an exact affine patch test as proof of approximation for arbitrary functions. Conversely, an inexact nodal value in the manufactured-forcing example does not contradict convergence.
Common mistake
(Nodal values of an function)
Nodal interpolation needs meaningful point values; arbitrary functions in two dimensions need not have them. Alternative averaging interpolants can reduce that regularity requirement, but they are different tools. The present proof names the embedding it uses.
Higher-order, nonconforming, adaptive or mixed methods require their own spaces and estimates. What is completed here is a specific conforming linear method; a single local matrix is not the whole of finite-element theory.
Summary
A complete triangular Galerkin error argument combines named analytical prerequisites, shape-regular affine scaling, coercivity and orthogonality. The five-node assembly provides an actual numerical method, while the manufactured solution and skinny-triangle counterexample explain what its theorem does and does not promise. Second-order accuracy requires an explicit additional dual regularity bound.
Exercises
Quick check
(Use constant forcing two in the five-node mesh. Find the center value and the discrete energy.)
The stiffness is unchanged and the load doubles.
Solution · Answer
The center load is , so . The energy is , and the load work is . Boundary corner values remain prescribed zero.
Quick check
(Identify the exact step in the proof that uses the extra dual regularity assumption.)
Energy approximation alone gives only one power of the mesh size.
Solution · Answer
After orthogonality, interpolation gives . Replacing by is the extra step. Without that premise, the displayed second-order conclusion has not been proved.
Quick check
(Prescribe the affine boundary function with forcing zero on the same mesh. Verify the center equation reproduces the exact center value.)
The four corner values sum to four.
Solution · Answer
The corner values are and the center value is one. The center row is , giving exactly . After subtracting the affine boundary lift, the difference lies in the homogeneous space; the affine function belongs to the trial space with those traces, so Galerkin uniqueness gives the same conclusion.
Quick check
(In the skinny-triangle example set . Find the interpolant's derivative and a lower bound on the energy interpolation error.)
Use the area as well as the large gradient.
Solution · Answer
The derivative of the interpolant is , while that of is zero. The area is , so the gradient-error norm is at least . The small area does not make this component vanish in the degenerating family.