One sign, three different worlds
By now you have done the hard parts of this rung. You wrote the general second-order linear PDE A u_xx + 2B u_xy + C u_yy + (lower order) = 0, you computed the discriminant B^2 - AC, and you watched the canonical forms emerge once you rode along the characteristics. So far the type has felt like a sorting label — a tidy way to file an equation under elliptic, parabolic, or hyperbolic. This guide is about the payoff: that label is not a filing convenience, it is the law that governs the equation. The sign of one number, B^2 - AC, secretly decides what data you are allowed to give, how fast effects travel, and whether the solution will smooth out or stay jagged.
Keep the three model equations in front of you as anchors, because everything below is really a story about them. The elliptic model is Laplace's equation Laplacian u = 0 (here u_xx + u_yy = 0): no time at all, a pure equilibrium. The parabolic model is the heat equation u_t = k u_xx: time moves only forward, and things diffuse. The hyperbolic model is the wave equation u_tt = c^2 u_xx: time is reversible and signals march. These three are not arbitrary examples — by the canonical-form result, every second-order linear PDE looks locally like one of them. Understand these three and you understand the whole forest.
type sign of B^2 - AC model equation characteristics time? ---------- ---------------- ------------------- --------------- ------------ elliptic B^2 - AC < 0 u_xx + u_yy = 0 none (complex) no time parabolic B^2 - AC = 0 u_t = k u_xx one family forward only hyperbolic B^2 - AC > 0 u_tt = c^2 u_xx two families reversible
The type tells you which data is allowed
Here is the first and deepest consequence. A PDE on its own has far too many solutions — recall that a general solution carries arbitrary functions, not just arbitrary constants, so you need data to single one out. But you cannot give just any data; each type wants a particular kind, and giving the wrong kind makes the problem ill-posed. A problem is well-posed (Hadamard's three demands) when a solution exists, is unique, and depends continuously on the data — wiggle the data a little, the solution wiggles only a little. Match the data to the type and you get all three; mismatch it and you lose at least one, often catastrophically.
The pairing follows directly from the type. Elliptic equations describe equilibrium with no time direction, so they want boundary data on the entire boundary of a closed region — a Dirichlet problem (fix u on the edge) or a Neumann problem (fix its normal slope). Parabolic equations have a one-way time, so they want an initial-boundary-value problem: the starting profile in time, plus boundary conditions in space, marched forward only. Hyperbolic equations have a reversible time and finite-speed signals, so they want a Cauchy problem in the proper sense: prescribe both u and its time derivative u_t at the initial instant, then march forward — or backward, it does not care.
Two cautionary tales: what mismatching costs
It is one thing to say mismatched data is ill-posed; it is far more memorable to watch it fail. Take the most famous example, due to Hadamard. Laplace's equation u_xx + u_yy = 0 is elliptic, so it wants boundary data — but suppose we stubbornly force Cauchy data on it, prescribing u and u_y along the line y = 0 as if it were a wave. Choose the data u(x,0) = 0 and u_y(x,0) = sin(n x) / n, which is tiny for large n — it shrinks toward zero everywhere as n grows. The solution that fits is u(x,y) = sin(n x) sinh(n y) / n^2. As n grows, sinh(n y) explodes like e^(n y), so the solution becomes enormous from data that is essentially nothing. Continuous dependence is destroyed: an imperceptible change in the data produces an unbounded change in the answer.
The second tale is the backward heat equation — running diffusion in reverse. The heat equation forward in time is beautifully well-posed, but try to ask 'what did the temperature look like one second ago?' and you are running a smoothing process backward, which means un-smoothing. A coffee cup that has cooled to room temperature could have started from countless different earlier states; reconstructing the sharp past from the blurred present is unstable, because the tiniest measurement noise in the present, expanded backward, blows up. This is why you can pour cream into coffee but cannot un-pour it: diffusion is irreversible, and the math says so before any physics does.
Finite speed, infinite speed, and smoothing
The type also dictates how fast effects travel and whether sharp features survive — and the wave and heat equations sit at opposite extremes. The hyperbolic wave equation has finite propagation speed: a disturbance at one point cannot be felt anywhere until a signal travelling at speed c has had time to arrive. This gives a genuine domain of dependence — the value at (x, t) depends only on initial data within distance c*t — and the carrier curves are exactly the real characteristic families you found from B^2 - AC > 0. Disturbances ride along characteristics at a strict speed limit.
The parabolic heat equation is the opposite. Its solution from a point spike is the heat kernel, a Gaussian that is strictly positive for every x the instant t > 0 — so warming one spot raises the temperature everywhere immediately. That is infinite propagation speed, and it comes bundled with an instant smoothing effect: even jagged, discontinuous initial data becomes infinitely smooth the moment time starts. The wave equation does the reverse: a sharp corner in the initial shape stays a sharp corner forever, carried along the characteristics without ever being blurred. Hyperbolic preserves features; parabolic erases them — and that difference is, once again, the sign of B^2 - AC speaking.
When the type itself changes — and when it breaks
Two honest complications keep the picture from being too tidy. First, when the coefficients A, B, C depend on x and y, the sign of B^2 - AC can flip from region to region — so the same equation can be elliptic in one place and hyperbolic in another, an equation of mixed type. The classic is the Tricomi equation y u_xx + u_yy = 0, here B = 0 and the discriminant is -y: elliptic where y > 0, hyperbolic where y < 0, parabolic exactly on y = 0. It genuinely models transonic flow, where air switches from subsonic to supersonic. The type is local, and on the dividing line the rules of well-posedness change with it.
Second, this whole guide lives in the world of linear equations, where the type is fixed by coefficients you can read off in advance. Step into nonlinear equations and the equation's character can depend on the solution itself, and beautiful behaviour can break in finite time. A perfectly smooth solution of a scalar conservation law like Burgers' equation u_t + u u_x = 0 can steepen until its slope becomes vertical and a shock forms — after which no classical solution exists at all. Past the shock you must allow weak solutions, and these are not unique without an extra entropy condition to pick the physical one. The clean type-decides-everything story is the first floor of a much taller building.
It is worth naming the open horizon, too, so you keep an honest perspective. The classification above is fully understood for second-order linear equations, but some of the most important nonlinear PDEs of physics remain partly mysterious. Whether smooth solutions of the three-dimensional Navier-Stokes equations stay smooth for all time, or can blow up, is one of the famous unsolved Millennium Problems — a reminder that beyond this clean rung, vast and genuinely open country still lies ahead. Classification is the map you carry into it.