Céa’s lemma and quasi-optimality
Separate the space’s approximation capacity from the variational stability constant and the norm in which optimality holds.
On this page
The question for today
What you’ll understand
- Prove the M/alpha quasi-optimality bound for conforming coercive Galerkin approximation.
- Distinguish general norm bounds from exact energy best approximation.
- State the approximation and regularity ingredients needed for a rate.
A finite element error contains an approximation question and a stability question. Could the discrete space represent the exact solution accurately, and does solving its weak equations find a comparably good representative? Céa’s lemma separates these questions for a bounded coercive form. Its constant depends on the chosen norm and the variational operator, while its best-approximation term depends on the space and solution. A rate appears only after that approximation term has actually been bounded.
Orthogonality converts coercivity into an error bound
Let V be a real Hilbert space, let a be a bounded bilinear form with continuity constant M and coercivity constant alpha greater than zero, and let F be a bounded linear functional. The exact and conforming finite-dimensional problems use the same a and F. For the error e, coercivity bounds alpha times its squared norm by a(e,e). Subtract any discrete approximation v_h and use orthogonality to replace the second argument e by u−v_h inside that expression.
Continuity then bounds the result by M times the error norm times the chosen approximation error. Cancel a nonzero error norm and take the infimum over the discrete space; if the error is zero the conclusion is immediate. This proof does not require symmetry, but it does require conformity, coercivity on the full continuous space, and exact agreement of the forms. A nonsymmetric problem can still be quasi-optimal without being an orthogonal projection in an inner product induced by a.
Symmetry gives a sharper conclusion in a different norm
If a is also symmetric, it defines an energy inner product. Galerkin orthogonality gives a Pythagorean decomposition for the error of any other discrete candidate: its energy error squared equals the Galerkin energy error squared plus the energy distance between the two discrete functions. Consequently the Galerkin solution is the exact best approximation in the energy norm, with constant one. That is stronger and more specific than applying the general M-over-alpha estimate in a background norm.
The closest point in an energy metric need not be the closest point in Euclidean or unweighted H¹ distance. Coefficient anisotropy weights directions differently, moving the projection. Changing a norm also changes the continuity and coercivity constants, so quoting a numerical quasi-optimality constant without its norm is incomplete. A large M-over-alpha bound is an upper guarantee and need not be the actual error ratio for a particular right-hand side or subspace.
Convergence and rates require approximation information
For a family of spaces with uniform continuity and coercivity constants, approximation of every relevant solution by discrete functions implies convergence. To obtain an order-h gradient estimate with P1 elements, one can combine Céa with a valid H² interpolation estimate on a shape-regular mesh, assuming the exact solution really has that regularity. The theorem does not manufacture derivatives missing at a reentrant corner or an interface. A weaker regularity class changes the approximation term and possibly the attainable rate.
A second-order L² error estimate usually invokes a dual problem and its extra regularity through the Aubin–Nitsche argument; it does not follow merely by relabeling the energy norm. Likewise, improving discretization does not necessarily make the unpreconditioned algebraic matrix uniformly conditioned. The stiffness matrix on quasi-uniform meshes typically becomes harder to solve as h decreases, even when the variational quasi-optimality constant stays fixed. Approximation stability and solver conditioning are related but different analyses.
Work it through
Energy projection differs from Euclidean projection
In R² use a(v,w)=vᵀ diag(1,4) w, exact vector u=(1,0), and V_h=span{(1,1)}. Compute the Galerkin vector and compare its two error metrics.
Write u_h=c(1,1). Testing with (1,1) gives 5c=1, so u_h=(1/5,1/5). The Euclidean closest vector would instead be (1/2,1/2).
The Galerkin error is (4/5,−1/5), whose Euclidean norm squared is 17/25. The Euclidean best error squared is one half, confirming that the Galerkin vector is not Euclidean-best.
Its energy error squared is 16/25+4/25=4/5. Directly minimizing (1−c)²+4c² gives the same c=1/5. In Euclidean norm, M=4 and alpha=1 give the valid loose bound four times the best error.
Optimality belongs to the energy metric in this symmetric example; the general background-norm guarantee remains valid without being sharp.
The living notebook
Explore the model
Interactive lab · matrix
Read the norm constants of a small metric
Reset, set diagonal scale a=2 and shear b=0. Predict singular values two and one for the selected matrix B. For the quadratic metric A=BᵀB, infer eigenvalues four and one, hence M=4 and alpha=1 in Euclidean norm. Reset and set a=1,b=0 for an identity metric. Readouts describe the selected final matrix even during interpolation. This toy family does not report a finite element stiffness matrix. Changes retain time; Reset restores both controls and zero time.
- Area scale · det A
- 1.4
- Singular values · σmax, σmin
- 1.597, 0.876
- Condition number · κ₂(A)
- 1.823
The selected matrix is A = [[a,b],[0,1]]. Playback interpolates geometry between the identity and A; readouts always describe A. Its determinant is a and its singular values come from AᵀA. Positive a makes every supported matrix invertible. This two-dimensional family is not a general eigensolver or a learned network. Parameter changes pause playback and recompute the model at the current elapsed time. Reset restores both parameters and returns time to zero. Playback starts paused. Adjust either parameter, then choose Play or Step.
Look forVariational constants depend on the metric and declared norm; they are not the observed error ratio of every projection.
Pause & reason
Check your understanding
Your turn
Try it yourself
A conforming coercive problem has M=6 and alpha=2 in its stated V norm. A verified best-approximation bound is 0.03. What error upper bound follows? If refinement changes that approximation bound to 0.015, what new bound follows, and must the actual error halve?
Take a moment to make a prediction or work it out on paper. Open a hint when you need one.
Give me a hint
Apply the same M/alpha multiplier to each approximation bound; distinguish upper bounds from equalities.
Compare with the solution
The quasi-optimality multiplier is six divided by two, or three. The first error bound is therefore 0.09, and the refined bound is 0.045. The guaranteed upper bound halves, but Céa does not force the actual errors to equal either bound or to have ratio one half. They can be substantially smaller, and a particular exact solution may have unusually good approximation in one of the spaces.
Keep the idea
Take it with you
- Céa separates variational stability from the discrete space’s best-approximation error.
- Symmetric coercive Galerkin approximation is exactly optimal in its energy norm.
- Rates, dual-norm improvements, and algebraic conditioning need their own additional arguments.
Read further
Original sources and deeper explanations to follow at your own pace.
Optional progress, saved in this browser when available. Revisit any lesson at any time.
Lesson not marked complete.