Motivation
A transport equation on a finite interval receives information through one boundary and loses it through the other. Periodic cancellation no longer applies. A conservative numerical method must describe the incoming flux, the outgoing flux and the resulting change of total mass. Giving a second independent boundary value at the outflow generally changes the physical problem rather than improving the approximation.
There is also a distinction between a nodal value and the average over a cell. Both can satisfy an upwind-looking recurrence, but their exact boundary residuals are different. A cell average represents material spread over an interval; the incoming flux is a time average at an interface. Treating it as an imaginary exterior cell average without analysis can hide a boundary error.
We prove separate statements for point samples and finite-volume averages of the scalar equation , . The hypotheses are smooth and compatible, the spatial grid is uniform, and the boundary is inflow/outflow. The results concern this linear scalar model only, not general nonlinear conservation laws.
Definitions
Definition 12.3.1
(Inflow problem and nodal upwind scheme)
On , prescribe and the inflow value . Assume a classical solution exists on the closed space-time rectangle; this includes the needed corner compatibility. There is no independently prescribed value at . Set , , and .
For nodal samples at , , use
Definition 12.3.2
(Finite-volume upwind scheme with incoming time averages)
For finite volumes take cells , exact averages , and numerical incoming average
unless an approximation to that integral is explicitly used. The upwind finite-volume fluxes are at the left boundary and at the right face of cell . Thus
For cell vectors write . Their discrete mass satisfies the exact balance
The outgoing value is computed from the last cell; it is not overwritten by a separate boundary prescription.
Concept lensGeometric
Reconstruct, transport exactly, then average
The finite-volume update is an exact transport step applied to a piecewise-constant profile. Take the value on each cell, move the profile a distance to the right, and fill the strip with the exact inflow, . Averaging over cell gives . Over the strip the inflow averages to , so the first cell receives . This is why the time average, not a point value, is the natural incoming datum.
Theorem/proposition
Theorem 12.3.3
(Nonperiodic nodal stability and convergence)
The nodal update is monotone. With exact inflow data, initial maximum error and a smooth exact solution,
If the prescribed boundary samples have error at most on the interval, replace by . No outflow datum is required for this estimate.
Theorem 12.3.4
(Cell averages converge in discrete spatial )
For the finite-volume scheme with exact incoming time averages, initial cell-average error , and the same smooth solution,
where
If the incoming average has error , add . In particular left-endpoint sampling of a inflow adds at most . This proves first-order smooth convergence for cell averages in the stated norm; it is a separate argument from the nodal theorem.
Proof sketch or proof idea
Proof
(Nodal error: convex weights and the boundary row)
The nodal interior error equation has the same convex weights as the scheme, plus the exact-start residual. Its residual bound is the time Taylor remainder plus the backward-space difference remainder . The first row receives the boundary error instead of an unknown previous component. Hence one step's maximum error is at most the larger of the previous interior error and the current boundary error, plus times the residual bound. Induction proves the first theorem. The outgoing row uses the same one-sided stencil and needs no value beyond .
Proof
(Cell averages: a nonexpansive matrix and a localized residual)
For cell averages, let be the homogeneous-inflow matrix. All entries are nonnegative; every column sums to one except the last, which sums to . The triangle inequality therefore gives . The missing last-column weight is material leaving through the outflow, not a loss of conservation accounting.
For cells , view their exact averages as samples at cell centers of the function obtained by averaging over an interval of length . On the region used by these backward differences, that interval stays inside the physical domain. Averaging commutes with time and space derivatives, preserves the transport equation, and does not increase the bounds on or . The nodal residual argument thus applies to those cells. Their total weighted residual is at most the first term of .
The first cell requires its own calculation. Let be the exact outgoing face average. Integrating the PDE over the first cell and the time interval gives
Subtract the numerical flux expression evaluated at the exact old cell average. The first-cell scaled residual is exactly . Compare both quantities through : averaging distances across the cell gives , and averaging time distances gives . Therefore
This contribution is small in the weighted spatial norm even though the unweighted first-cell residual need not tend to zero. Combining all cells proves . The error recurrence and nonexpansiveness yield .
If the supplied incoming average differs by , the update perturbation is supported only in the first cell and has size . Its weighted norm is . Summing gives the additional term in the theorem. For , the average-value difference is at most , proving the final assertion. Every boundary term is accounted for rather than silently assigned the periodic residual estimate.
Proof X-ray
One cell decides the norm
The two theorems use different norms, and the reason is a single cell. Stability in comes from the column sums of : each column sums to at most one because the scheme is conservative and the missing weight leaves through the outflow face. The hypothesis is used only to make every entry nonnegative. Interior consistency is borrowed from the nodal proof by reading cell averages as samples of the averaged function; this needs the averaging intervals of the cells and of their upwind neighbors to lie inside .
The key move concerns the first cell. Replacing the interface value by the cell average leaves a scaled residual of order one; for the affine solution in the worked examples it is . An accumulation argument in the maximum norm would add at each of up to steps, giving only a bound of size about , which does not tend to zero. In the same residual carries the weight , so its total contribution is . The argument closes through and nonexpansiveness. It does not claim a vanishing maximum residual, and it does not deduce a maximum-norm theorem from the L1 bound.
Worked examples
Worked example 12.3.5
(Incoming and outgoing mass on a complete grid)
Take and cell vector . With first incoming average two, , and mass increases from one to two. With the next incoming average zero, and mass remains two because the previous outgoing cell was zero.
One further zero-inflow step gives , with mass . The loss equals . The sum is not constant, but the conservative boundary-flux balance is exact at every step.
Worked example 12.3.6
(An affine solution exposes the boundary-cell residual)
For , let and . On the same grid the exact initial cell averages are , and the first incoming time average is . One numerical step gives , while the exact averages at are . The error is confined initially to the first cell.
For general with , its first-cell residual is , which does not vanish at fixed . Its weighted contribution is only , exactly the mechanism used in the proof. Nodal upwinding with exact old-time boundary values would reproduce this affine solution at every node; changing the meaning of the unknown changes the boundary calculation.
Worked example 12.3.7
(An outflow value must not be imposed twice)
Take three unit cells, , , initial averages and zero incoming data. The next vector is , mass two. Inflow contributes mass , and outflow removes , so the mass change is minus one.
If the last cell were instead overwritten by an artificial zero boundary condition, the vector would become , losing another unit with no declared flux. At Courant one, the correct scheme is the exact cell shift, and the first cell equals the incoming time average. An additional outflow assignment breaks that exact transport structure.
Interpreting the result and the initialization
The cell-average theorem starts from actual integrals. Initializing instead at cell midpoints introduces a separate quadrature error; for a initial profile it is at most in each cell, so it can be added to . A general rough profile has no such derivative estimate. The same numerical array should therefore be labeled by what it represents, not called both a nodal sample and a cell average interchangeably.
The first-cell residual bound is deliberately robust rather than sharp for every solution. Some profiles or Courant choices create cancellation. At Courant one the method shifts exact cell averages exactly, but that special fact is not needed to prove convergence for the whole interval . The absolute residual bound provides a uniform statement without relying on accidental cancellation.
The norm has physical spatial weight. Dividing by the number of cells or omitting can obscure the difference between a localized boundary defect and a defect present everywhere. The proof estimates the integral-sized error . It says nothing about a nonlinear flux whose coefficients depend on the evolving solution; that setting needs a different stability and solution theory.
For negative transport speed, the geometry reverses: data enter at the right boundary, and the right state is used in the upwind flux. Reversing only a sign while leaving the incoming side fixed would solve the wrong initial-boundary problem. The present positive-speed proof can be reflected in space, but it is not a theorem for time-varying speeds that change sign.
Two computations with the same incoming data have a homogeneous boundary difference even when each physical solution is nonzero at the boundary. This is the setting of the matrix estimate. If the incoming data differ, the separate injection term must remain. The proof does not count it as another interior component with an invented value outside the domain.
Nonexpansiveness is also not strict contraction at every step. A single-signed error, for example a nonnegative error supported well inside the interval, can retain its whole L1 mass before reaching the outflow. At Courant one a homogeneous-inflow error is shifted exactly; only the part crossing the last face disappears. A bound with constant one is therefore natural, and claiming uniform strict decay at every step would be false. Once the boundary injection and residual are included, the error estimate describes both transport of existing error and addition of new error. These are different mechanisms even though both appear in the same numerical array.
The initial errors in the nodal and cell-average theorems are also measured in different units. The first asks for the worst nodal deviation; the second asks for the spatial total of the absolute deviation. If the cell averages are initialized by midpoint quadrature, each cell's error is second order, but the sum must still be multiplied by the cell length to give a total bound that matches the length of the interval. Omitting this weight can make the same smooth error appear to grow as the number of nodes increases.
A first-cell residual of fixed size does not pile up indefinitely at the same location without moving. Error is carried downstream from the inflow region, and each new injection is counted with the spatial size of one cell. The proof does not need to follow every path cell by cell, because the bound for a nonnegative matrix already controls the total of everything transported; but it gives no additional conclusion that every individual cell error is small to the same order. A pointwise analysis would need more structure; the L1 theorem cannot simply be renamed.
On an open interval, conservation is an accounting balance: the new stock equals the old stock plus what comes in minus what goes out. That the outgoing flux uses the old last cell is part of the definition of this explicit algorithm; using the new value or a time average instead gives a different time scheme, which needs its own analysis. All time labels are kept here so that an actual worksheet can be compared line by line with the proof, rather than guessed from the name of the scheme.
The exact incoming time average requires only the integral of the known boundary function, not any knowledge of the future interior solution. It is therefore a flux computable from the given data. If only discrete observations are available, the boundary quadrature term of the theorem records the resulting accuracy explicitly, instead of the data being assumed exact.
Common mistakes
Common mistake
(An invented exterior neighbor for the first cell)
Boundary consistency should not be checked by pretending the first cell has an available interior neighbor outside the interval. Its actual incoming flux must be used, and the first-cell residual must be computed from the exact flux balance over that cell, as in the proof.
Common mistake
(Point value versus interval average at the inflow)
The nodal formula uses the boundary value ; the exact finite-volume balance uses the interval average . Either may be approximated, but their residuals differ. For they differ by at every step, which is exactly the boundary quadrature term in the cell-average theorem.
Common mistake
(Expecting constant mass on an open interval)
Conservation on an open interval means that the change of equals times incoming minus outgoing value, not that the total is constant. In the complete-grid example the mass falls from to in one step, and the loss is precisely the outgoing flux .
The smooth assumptions also exclude incompatible corners and shocks. The argument supplies first-order error for this scalar linear problem, not general nonlinear finite-volume convergence or entropy selection. Nor does conservation of a single total show that every cell is accurate after finitely many time steps.
Counterexample mode
Upwinding without the Courant restriction
Because the nodal update draws only on the upwind side, it can look stable for every time step. Take , zero inflow, and . Then
The maximum norm doubles in one step and both new values leave the old range ; the weight is negative. The exact value is fixed at the inflow corner, outside the stencil , so the scheme cannot see the characteristic it must follow. The repair is the hypothesis : both weights are then nonnegative, and the foot of the characteristic lies in .
Summary
One inflow condition supplies the incoming characteristic information. Nodal upwinding has a maximum-norm error proof; finite-volume upwinding needs a separate cell-average boundary calculation. A localized residual and the discrete spatial L1 norm yield first-order convergence, with incoming quadrature errors and outgoing mass explicitly counted.
Exercises
Quick check
(With , update cell vector using incoming average one. Check the mass balance.)
The outflow uses the old last cell.
Solution · Answer
The vector is , with mass three. The old mass is also three, and . Conservation accounts for the boundary exchange even though individual cells change.
Quick check
(For , and , derive the first-cell scaled residual directly.)
Compare the old cell average and the time-averaged value at .
Solution · Answer
The old first-cell average is , and the outgoing time average is . Their difference is . Multiplying by gives . Multiplication by the cell weight makes its absolute contribution .
Quick check
(If , compute its exact incoming average and the cumulative L1 boundary-error bound when is used instead.)
The error in each incoming average is .
Solution · Answer
The exact average is . Left-endpoint sampling differs by . Each step injects weighted error at most ; over at most steps the bound is . This is a boundary quadrature contribution, separate from the interior residual.
Quick check
(Prove the homogeneous-inflow matrix is nonexpansive in by summing its columns.)
Identify the single column with a missing outflow weight.
Solution · Answer
The old value in column contributes to its own next cell and to the next one, total one. The last column contributes only , since the other part leaves the domain. Nonnegative entries and the triangle inequality therefore give . An inhomogeneous incoming value is an added first-cell load, not part of this matrix.