Motivation
A derivative condition changes more than the first row of a Dirichlet matrix. Boundary nodes now become unknowns, pure Neumann problems have a constant nullspace, and the natural discrete integral gives half weight to the endpoint control volumes. Ignoring one of these changes can produce a singular inconsistent system or an unjustified energy proof.
We construct a symmetric flux-balance discretization of on an interval. Robin conditions add endpoint energy; pure Neumann conditions leave a constant mode and require a flux balance. The same construction also illustrates an important accuracy lesson: the endpoint residual divided by its small control-volume weight is only first order, yet the integrated residual acts on test functions strongly enough to prove second-order global convergence.
These are complete statements for this one-dimensional constant-coefficient model with exact smooth data. They do not establish arbitrary multidimensional boundary methods.
Definitions
Definition 11.1.1
(Robin conditions and the weighted grid)
On , prescribe outward-normal Robin conditions
Neumann means both coefficients vanish. Put , , , and include all nodal unknowns. Endpoint control volumes have weights ; interior weights are . Write , , and .
Definition 11.1.2
(Half-cell flux-balance system)
The actual equations are
Denote this system by . It is a flux balance using endpoint-half trapezoidal loads. Its symmetric bilinear form is
For pure Neumann data the continuous compatibility condition is . Exact nodal quadrature need not preserve this identity. Set and use the projected load , which has zero sum. Fix uniqueness by requiring , the prescribed continuous mean of the exact solution. For Robin data with , use the original load without this projection.
Concept lensStructural
Row scaling changes the inner product, not the method
One may divide every row by its control-volume weight and write an operator . This normalized matrix is generally not symmetric in the ordinary Euclidean inner product. It is self-adjoint in the weighted inner product, because . The corresponding compatibility test is the weighted test against constants, which again gives . A code using normalized rows and an analysis using integrated loads must translate between these conventions rather than compare differently scaled residual numbers.
Theorem/proposition
Theorem 11.1.3
(Solvability and uniform discrete coercivity)
If at least one Robin coefficient is positive, is positive definite and the system is uniquely solvable. If, for example, , then
For pure Neumann data, the nullspace is exactly the constants. A load is solvable exactly when its sum is zero; the mean constraint then makes the solution unique. On the weighted mean-zero subspace ,
Here the squared edge norm is . Constants are independent of , for fixed physical Robin coefficients.
Theorem 11.1.4
(Second-order nodal convergence despite endpoint residual scaling)
Let solve the stated model, with exact and exact boundary data. For fixed Robin coefficients with positive sum, the unprojected numerical solution has . For pure Neumann conditions assume continuous compatibility and prescribe the exact continuous mean ; the projected-load solution also has . The constants depend on , derivative bounds and the fixed coercivity parameters, not on the mesh.
Proof sketch or proof idea
Proof
(Energy positivity and a discrete Poincaré argument)
Every term of is nonnegative. Zero gradient makes all nodal values equal; a positive endpoint coefficient then makes that constant zero. This proves positive definiteness. To obtain its uniform bound, write . Cauchy–Schwarz gives . Sum with weights, whose total is , and use . This gives the displayed . If only the right coefficient is positive, reflect the interval.
With no Robin terms, gradient energy has exactly the constant nullspace. Symmetry makes the range the orthogonal complement of that nullspace in the ordinary coefficient pairing, so zero load sum is necessary and sufficient. For a weighted mean-zero vector, each is a weighted average of differences . Every difference is at most in magnitude. Therefore , proving the endpoint trace estimate and, after summing, the stated Poincaré estimate. This also proves uniqueness and stability on the constrained subspace.
Proof
(Dual pairing of the integrated residual)
Now insert exact nodal values and let be the integrated residual. Interior Taylor expansion gives , where . At the left endpoint, substitute and ; the derivative and second-derivative terms cancel. Thus
Dividing these by would give an endpoint estimate. We instead bound their action on a test vector. For Robin error , . Interior terms satisfy
For , the endpoint trace is bounded by : control by its Robin energy and use . Combine the residual estimates and coercivity to obtain . Division, including the trivial zero-error case separately, gives energy error and then the weighted nodal norm bound. No pointwise residual norm was substituted for this dual estimate.
For pure Neumann error, . Let and . Testing by removes the projection term exactly, because . The preceding residual argument now uses the mean-zero trace and Poincaré bounds, giving . Finally the numerical mean is , while the trapezoidal mean of exact samples differs from the continuous mean by at most . Hence and .
The projection is consistent: since continuous compatibility holds, by trapezoidal quadrature. It repairs a small discrete imbalance rather than pretending incompatible continuous data define a solution.
Proof X-ray
Why the boundary's apparent first order does not settle global order
The proof turns on how the endpoint residual is measured. Read pointwise, divided by its weight is . In the proof it is never divided: it enters only through the pairing , and is bounded by , through the Robin term or, in the Neumann case, through the mean-zero trace estimate. The endpoint contribution is therefore , the same size as the interior contribution, whose residuals are summed over about nodes. Three hypotheses are used. Symmetry gives the identity . Exact consistency of the endpoint balance cancels the and terms; dropping the half-cell load , for example, leaves and the same argument yields only first order. Coercivity from the first theorem converts the energy bound into a bound in . In the Neumann case the test vector must be , not : only the mean-zero part annihilates the projection term , and the discarded constant is estimated separately by quadrature.
Worked examples
Worked example 11.1.5
(A complete Robin matrix)
Let , , and exact solution . Then , , , and
Multiplication verifies . This quadratic is reproduced exactly because its third and fourth derivatives vanish, including in the endpoint balance. The matrix is genuinely constrained by Robin terms; no artificial fixed endpoint value was imposed.
Worked example 11.1.6
(The prescribed mean differs from the sample mean)
Use the same quadratic with pure Neumann data . The load sums to zero already. The exact continuous mean is , but the weighted mean of is . Since constants do not change any flux, imposing the continuous mean gives
The error is the constant . It is not a failure of solvability or flux accuracy; it is the explicitly predicted quadrature effect of the chosen normalization. Fixing a sample mean instead would define a different normalization.
Worked example 11.1.7
(Correcting an actual discrete incompatibility)
Take , so , and mean . On the two-cell grid the raw load is , whose sum is . The singular Neumann matrix cannot solve that raw system.
Here , and subtracting gives . The mean-constrained solution is . Direct multiplication by the tridiagonal Neumann matrix with diagonal , adjacent entries and corner entries zero gives that projected load; its weighted mean is . The compatibility correction and mean constraint perform different jobs.
Why the weights and the compatibility check are structural
The endpoint equation follows by integrating the differential equation over a half control volume. At the left, the outward derivative is prescribed at zero, while the interior interface flux is approximated by the first slope . The endpoint forcing is integrated with a half-cell weight. At the right the orientation reverses. This derivation explains the signs and the endpoint weights together; changing only one of them would break the symmetric energy identity.
The nullspace also explains why a solver tolerance is not a compatibility repair. An incompatible right side cannot be made exactly solvable by iterating longer. The projection changes the load in a specified direction and then the mean constraint selects one solution. For already compatible discrete data its projection coefficient is zero, so no repair is needed. For genuinely incompatible continuous data, however, a small-quadrature-error argument is unavailable. For instance with zero outward flux at both ends has no continuous solution, since the integrated balance would require . Projecting that load would remove its entire mean and solve a different physical problem. The convergence theorem explicitly starts from a compatible exact solution.
Perturbations in supplied data can be analyzed in the same energy pairing. If nodal forcing changes by and the endpoint flux data by , the new load is . Its action on a test vector is at most . The established trace and coercivity estimates therefore bound the solution perturbation. Second-order data errors preserve the stated rate; coarse boundary measurements can instead dominate it. For a pure Neumann problem the same argument applies after projection on the mean-zero space, and the prescribed mean has its own constant-mode error. These are distinct sources of error, not an unexplained defect of the boundary stencil.
There is an equivalent direct way to compute the projected Neumann solution. Solve the augmented system with equations and . Multiplying the first equation by gives , because the weights sum to . Substitution recovers exactly the projected problem. To prove uniqueness of this augmented solve, take a homogeneous difference: its multiplier is zero by the same constant test, its displacement is then constant by the nullspace result, and its zero prescribed mean makes that constant zero. Thus the augmentation is nonsingular even though alone is singular. The extra scalar represents the load correction, not an additional physical unknown in the differential equation. This construction is particular to the stated compatibility and mean convention; adding it to a coercive Robin problem would impose an unnecessary new condition.
Midpoint flux weights keep the matrix symmetric, but coercivity and endpoint consistency must be rechecked.
A path to generalization
What changes for a variable conductivity?
Midpoint flux weights keep the matrix symmetric, but coercivity and endpoint consistency must be rechecked.
For with , weight each edge by a midpoint value and write the boundary conditions for the flux . The form plus the Robin terms is still symmetric, and the Neumann nullspace is still the constants. The coercivity constants now depend on , and the endpoint Taylor cancellation must be redone, because the half cell sees a varying .
The endpoint half weight does not mean that the endpoint matters less; it reflects the different length of the control volume it represents. The vector norm approximates the continuous integral, while the matrix pairing keeps the actual balance of each control volume. Together these two structures determine how the residual enters the energy estimate. Dividing the endpoint rows by their weights without changing the inner product accordingly can lead one to conclude, from an unsymmetric ordinary matrix, that the original method has no energy structure.
In the quadratic example the fluxes are exact, yet a constant nodal error appears; this shows directly that pure Neumann data determine the solution only up to a constant. What is prescribed is the continuous mean, while what the discrete solution can compute directly is the trapezoidal mean; the difference between them is covered by the quadrature bound. Prescribing instead the mean of the known exact samples would make the nodes agree exactly in this example, but that changes the normalization data and cannot be used to refute the constant offset predicted by the theorem.
The test vector in the energy error estimate also cannot be chosen freely while the constraint is ignored. In the Robin case the error can be tested against itself, because the whole space is coercive; in the Neumann case the mean must be removed first, and only testing with the mean-zero part makes the projection term vanish. This order shows that the nullspace is not a cosmetic adjustment made after solving but a core ingredient of the stability and convergence proof. Adding the constant component back at the end gives the full nodal-norm conclusion.
The two ends have opposite outward normals. The signs of the flux data should be written from the physical boundary first and only then enter the discrete equations; they must not be flipped afterwards to make a particular numerical answer look right.
Common mistakes
Common mistake
(Sign of the left outward derivative)
The left outward derivative is , not . Reversing that sign changes both the continuous compatibility condition and the endpoint balance, so the discrete system no longer approximates the stated problem.
Common mistake
(A small nonzero load sum is still incompatible)
A pure Neumann load must satisfy its discrete condition exactly. A small nonzero sum is still outside the range of the singular matrix, and no solver tolerance or longer iteration makes such a system exactly solvable.
Pinning a node before the incompatibility is resolved hides this failure instead of removing it.
Counterexample mode
Pinning a node does not repair an incompatible Neumann load
Claim: replacing one equation of the singular Neumann system by a pin always yields a solution of the Neumann discretization. Take the raw load from the worked example on correcting an incompatibility, on two cells, , whose sum is , and pin in place of the first row. The remaining rows and give and . The discarded first row now reads instead of : the left flux balance fails by exactly the load imbalance. The pinned system is always solvable precisely because the deleted equation absorbs the mismatch. A pin is a legitimate normalization only after the load has zero sum; the deleted row then holds automatically, since .
Common mistake
(Robin coefficients that shrink with the mesh)
A positive but mesh-dependent Robin coefficient may give constants that deteriorate as it tends to zero. The theorem fixes physical coefficients. The Neumann limit is handled on a constrained space, not by claiming the Robin coercivity constant stays uniform in that limit. Approximate boundary data and quadrature introduce additional dual residuals and require their own accuracy budget.
A second-order nodal error is also not the same as a second-order gradient error of the piecewise linear function relative to the continuous solution. The energy used here is the discrete energy of the vector of nodal differences, and interpolation error must not be overlooked.
Summary
Half-cell flux balance produces a symmetric boundary method with a natural weighted norm. Robin terms remove the constant mode; pure Neumann data require compatibility and a mean. Integrated residuals, energy coercivity and endpoint trace control prove second-order convergence, even though a naively scaled endpoint residual is only first order.
Exercises
Quick check
(Prove that the pure Neumann matrix annihilates constants and has no other null vectors.)
Use the edge energy rather than computing a determinant.
Solution · Answer
Every difference of a constant is zero, hence its energy and matrix image vanish. Conversely implies , so every adjacent pair agrees. Connectedness of the interval then makes constant.
Quick check
(Explain why projecting the load does not by itself determine a unique Neumann solution.)
Separate range compatibility from nullspace normalization.
Solution · Answer
Projection makes the load orthogonal to the constant nullspace, guaranteeing existence. If solves it, so does for every . A prescribed weighted mean fixes that remaining constant; these are independent operations.
Quick check
(For on , choose . What is on any uniform grid?)
The trapezoidal rule is exact on affine functions.
Solution · Answer
The discrete integral of is exactly , so its sum with is zero. Thus at every uniform resolution. This does not say every exact smooth compatible load has zero quadrature imbalance.
Quick check
(Why cannot the Robin estimate with remain uniform as when the other coefficient is zero?)
Test a constant vector.
Solution · Answer
For , the weighted norm squared is while energy is only . Any inequality needs , which diverges. At zero coefficient the constant must instead be removed by a mean constraint.