The collision we saw coming
In the previous guide we ran the characteristics of Burgers' equation u_t + u u_x = 0 and watched something unsettling: because the carrying speed is the value u, tall parts of the profile outrun short parts, so the fast characteristics catch the slow ones and the characteristics cross. This guide is about what happens at and after that collision — the part the smooth theory cannot describe. We are no longer asking how to build a solution; we are asking how a solution dies, and what takes its place.
Let us make the collision concrete with one initial profile. Take data that starts at u = 1 on the far left, ramps smoothly down to u = 0 on the far right — a falling step, gentle at first. Each point launches a straight characteristic x = x0 + u(x0) t. The points sitting on the high left part throw lines of slope-speed 1; the points on the low right part throw nearly vertical, slow lines. The fast lines lean rightward into the slow ones, and somewhere ahead they must intersect.
Two crossing characteristics carry two different remembered values of u to the same point (x, t). A single-valued function cannot honour both. Geometrically, the solution surface would have to fold over and stand vertically above one spot — the graph wants to become multivalued, like a breaking ocean wave curling over on itself. Nature does not actually deliver three values of the water height at once, so the smooth picture has to give way to something new: a jump.
Reading the breaking time off the data
Here is the honest and slightly surprising fact: the solution stays perfectly smooth right up to a definite instant, and then its derivative becomes infinite all at once. Up to that instant nothing looks wrong — the value u stays neatly between 0 and 1, bounded the whole way. What blows up is not the height but the slope u_x. This is gradient blow-up: bounded value, infinite gradient. Keeping those two failure modes apart is the single most useful habit on this rung.
We can pin the moment down with a one-line calculation. Differentiate the constant value along a characteristic and you find that the spatial slope obeys a Riccati-type law: where the initial profile is falling — that is, where the initial slope u'(x0) is negative — neighbouring characteristics converge and u_x races toward minus infinity. The first time any of them collides is the breaking time, and it is read straight off the data.
characteristic from x0: x = x0 + u0(x0) t, u stays = u0(x0)
slope along it: u_x = u0'(x0) / (1 + u0'(x0) t)
blows up when 1 + u0'(x0) t = 0
first breaking time t* = -1 / min over x0 of u0'(x0)
(only where u0' < 0)Past the breaking time: weak solutions
After the breaking time the equation u_t + u u_x = 0 simply stops making sense, because u_x no longer exists at the jump. To keep going we have to change what we mean by a solution. The trick is to step back to the conservation-law form, u_t + (u^2/2)_x = 0, and recall that this whole subject came from conservation laws: the differential equation is only a convenient stand-in for an integral bookkeeping statement that the total amount of stuff in any interval changes only through its flux across the two ends.
The integral statement is perfectly happy with a discontinuous u, because it never asks for u_x at all. Formalising this gives a weak solution: a function that may have jumps but still satisfies the conservation law when tested against any smooth probe function, the derivatives all moved onto the smooth probe by integration by parts. A weak solution agrees with the classical one wherever the latter exists, and quietly extends the theory through the jump. This is the same move you will see again with scalar conservation laws and, much later, in Sobolev spaces.
So the way past the catastrophe is to allow jumps — but a jump is not free. The integral bookkeeping that defines a weak solution does not let the front move at any speed it likes; conservation across the jump fixes how fast it travels. That constraint is the next thing to nail down.
How fast does the jump move? Rankine-Hugoniot
Picture the shock as a moving front with state u_L just behind it and u_R just ahead, travelling at speed s. In a small time step the front sweeps over a thin slab; the amount of conserved quantity it gains must equal the net flux carried in across its two faces. Setting gain equal to net flux gives the Rankine-Hugoniot condition s (u_R - u_L) = f(u_R) - f(u_L): speed times jump-in-value equals jump-in-flux. It is just conservation, applied across a discontinuity instead of through a smooth region.
For Burgers the flux is f(u) = u^2/2, so f(u_R) - f(u_L) = (u_R^2 - u_L^2)/2 = (u_R + u_L)(u_R - u_L)/2. Dividing by the jump u_R - u_L gives a clean answer: s = (u_L + u_R)/2. The shock travels at the average of the two states it separates. So if a value-1 region runs into a value-0 region, the wall between them moves at speed one-half — slower than the back, faster than the front, exactly as it must to keep eating both.
Which jump is the right one? The entropy condition
Here is the sting in the tail, and the reason this rung matters beyond bookkeeping. Once we admit jumps, the weak solution is no longer unique. Consider data that rises, say u = 0 on the left and u = 1 on the right. One option spreads it into a smooth rarefaction wave that fans the characteristics apart. But a second option also satisfies the equation in the weak sense and even satisfies Rankine-Hugoniot: leave the data as a standing jump from 0 to 1, an 'expansion shock'. Both are weak solutions. Only one is right.
We rule out the impostor with an extra physical demand called the entropy condition. In words: a genuine shock must compress — characteristics on both sides must run into it, never out of it. The expansion shock fails this, because its characteristics flow away from the front, leaving a region no characteristic reaches; that telltale gap is exactly what the rarefaction fan fills. Adding the entropy condition restores uniqueness: among all the weak solutions, exactly one is admissible.
Where does such a rule come from, rather than being pulled out of a hat? From a tiny dose of honesty about the model. Real fluids and traffic have a little diffusion; the truthful equation is u_t + u u_x = epsilon u_xx with a small epsilon > 0, which is smooth and has unique solutions forever. Take the limit epsilon to 0 and the smooth profile sharpens into a jump — but only the compressive, entropy-satisfying jumps survive. This selection principle is vanishing viscosity, and it is the cleanest justification for the entropy condition: the right weak solution is the ghost of a barely-viscous one.
What this rung leaves you with
Look back over the whole arc of these five guides. We began with the transport equation, where information just slides along straight roads; we learned the method of characteristics; we let the roads bend with the cargo and reached genuine nonlinearity; we asked which data curves are admissible. This last guide closes the loop with the deepest lesson: a perfectly smooth nonlinear law can destroy its own smoothness in finite time, and the only way forward is to widen what we accept as a solution — then re-narrow it with a physical principle.
That two-step rhythm — widen to weak solutions, then pick the right one with an admissibility condition — is one of the great recurring patterns of modern PDE theory, and you will meet it dressed in heavier clothing later on. It is worth naming the honest caveat now: weak solutions buy existence at the cost of uniqueness, and recovering uniqueness always takes an extra idea from outside the bare equation. For nonlinear conservation laws that idea is the entropy condition; in other corners of the subject it wears other names.
If you carry one image away, make it this: a falling ramp steepening into a vertical wall, then the wall settling into a clean front that marches at the average speed of the two states it divides. That single picture holds gradient blow-up, shock formation, Rankine-Hugoniot, and the entropy selection all at once — the entire drama of crossing characteristics in one breaking wave.