When a smooth transport solution breaks
The linear transport equation moves every feature at one speed. In the nonlinear conservation law
the characteristic speed is . Where a larger value sits to the left of a smaller value, the rear can catch the front. A classical solution can then develop an infinite slope, so a numerical method should not rely on the solution remaining differentiable. This equation is called Burgers' equation. We first account for mass over a cell and then ask what selects a physically relevant discontinuity.
Definition 12.2.1
(Conservative update and Lax–Friedrichs flux)
Let cells have width , time step , and . On a periodic grid, approximates the average of in cell at time . Integrating the equation over a cell says that its mass changes by incoming minus outgoing flux. A conservative update has the form
The same interface flux is used by both neighboring cells. Its consistency condition is : a constant state then remains constant. We take the global Lax–Friedrichs flux, whose parameter is fixed (letting vary from interface to interface would give the local Lax–Friedrichs, or Rusanov, flux):
The first term averages physical fluxes; the second transfers information from high to low values. We choose , where bounds the absolute values encountered. This parameter bounds the characteristic speed on . It is fixed across interfaces and time in the proof below; an adaptive choice needs its own argument.
Conservation and a maximum principle
Theorem 12.2.2
(A bounded conservative update)
Suppose the periodic initial data satisfy , take , and assume . Then (1)–(2) preserve the sum and the bounds for every step, in exact arithmetic. More generally, each new value lies between the minimum and maximum of the three old neighboring values.
Proof
(Monotone in every argument)
Write the update as . On the cube ,
Thus is nondecreasing in each argument there; since the cube is a product of intervals, one coordinate can be moved at a time without leaving it. If and , consistency gives and , so . Induction keeps every value in , justifying the derivative signs at every later step. Summing (1) over a periodic grid cancels each interface flux with itself, proving conservation. Multiplying the sum by gives discrete mass.
Proof X-ray
Where each hypothesis enters
The step restriction is used once, for : it keeps the weight of the cell's own old value nonnegative. The bound is used for the two neighbor derivatives, since needs and needs .
Those signs are known only on the cube , so the proof is an induction. The bound at step justifies monotonicity at step , and monotonicity yields the bound at step . Consistency enters only to evaluate at constant triples. Conservation uses neither hypothesis; it needs only the shared interface flux and periodicity.
Neither property proves convergence to an entropy solution. A stable-looking discrete profile is not yet an error estimate for a shock.
The two hypotheses on are easy to conflate in practice.
Common mistake
(Choosing the dissipation below the data bound)
The parameter must bound on all values that occur, not merely satisfy . With data in , and , the old triple gives and , so the new center value is . It exceeds the largest old value : the derivative is negative at .
A complete update
Worked example 12.2.3
(One pulse on four periodic cells)
Take , and . The fluxes at the four right interfaces, in order, are
Applying (1), including the last-to-first periodic interface, gives
The sum remains , and all values stay between and . The pulse has spread: the dissipative flux improves monotonicity but also smears a sharp front. This single step illustrates the mechanism; it is not a convergence proof.
Worked example 12.2.4
(Removing the dissipation creates an undershoot)
Replace (2) by the centered flux , which is still consistent and conservative, and repeat the pulse step with and . The right-interface fluxes are
Cell one receives and loses , so it stays at . Cell two gains , cell three is unchanged, and cell four loses :
The sum is still , but the fourth value is negative although every old value was in . Conservation alone therefore does not prevent a new undershoot; the dissipative term in (2) is what makes the update monotone.
Which discontinuity is admissible?
For a jump moving at speed , write . Integrating the conservation law across the moving jump yields the Rankine–Hugoniot relation
For Burgers' flux and unequal states, . Relation (3) is necessary for this piecewise-constant weak solution, but it admits both an upward and a downward jump. It does not alone choose the physically relevant solution.
Definition 12.2.5
(Entropy pair and the entropy condition at a jump)
For Burgers' equation take the entropy pair , , for which . Across a jump moving at speed , the distribution has coefficient . The jump satisfies the entropy condition for this pair when .
Proposition 12.2.6
(Jump speed and entropy sign for Burgers' equation)
Let be the states of a jump satisfying (3) with . Then
Hence the jump satisfies the entropy condition for this pair if and only if .
Proof
(Factor out the jump)
Write and . Relation (3) reads , and division by gives . Then
The bracket equals , so the coefficient is . It is nonpositive exactly when .
Worked example 12.2.7
(Testing a downward shock)
Take , . Relation (3) gives , and directly
in agreement with . The downward shock passes the test. Its speed lies strictly between the characteristic speeds on the left and on the right, so characteristics from both sides run into the shock rather than out of it.
Counterexample mode
The jump condition alone does not select the solution
It is tempting to accept any piecewise-constant function whose jump satisfies (3). Take the increasing Riemann data , with the jump at , and the jump moving at . Then , so (3) holds and this is a weak solution. Its entropy coefficient is , so it violates the entropy condition. The entropy solution for is instead the rarefaction fan: for , for , and for . Inside the fan, . The repaired statement requires both (3) and the entropy inequality.
This calculation checks one entropy pair for these two specific jumps; a general existence, uniqueness or numerical convergence theorem requires additional hypotheses and arguments. A small residual away from the jump cannot establish entropy admissibility at the jump.
Practice with solutions
Quick check
(Verify that the flux (2) is consistent, that is, for every .)
Set in (2).
Solution · Answer
The difference term is zero and , so a constant grid is unchanged.
Quick check
(For periodic data with and , compute one step of (1)–(2).)
Compute the three right-interface fluxes first.
Solution · Answer
The right-interface fluxes are , and . Incoming-minus-outgoing flux therefore gives . Its sum is .
Quick check
(Take left state and right state . Find the shock speed and entropy coefficient.)
Use (3), then .
Solution · Answer
. Since and , the coefficient is , in agreement with . The shock is admissible.
Quick check
(For , , and , does the bounded-update theorem apply?)
Compute .
Solution · Answer
, so its hypothesis fails. This does not prove that every particular grid fails; it removes the stated guarantee.