An equation alone is not a problem
In the previous guide you met the three classic side conditions you can hang on a boundary — Dirichlet (fix the value), Neumann (fix the normal flux), and Robin (a weighted mix). But we never asked the prior question: how much data do we actually need, and where? The honest starting point is that a PDE on its own is wildly under-determined. It is not a problem yet; it is only half of one.
Here is the heart of the matter, and it is genuinely different from ordinary differential equations. Recall from the very first rung that the general solution of a PDE involves arbitrary functions, not arbitrary constants. The plain transport equation u_t + c u_x = 0 has general solution u(x,t) = f(x - c t), where f is any differentiable function whatsoever. An ODE leaves you a few numbers to fix; a PDE leaves you an entire function's worth of freedom to fix. That is why side data has to be a function spread over a curve or surface, not just a handful of values.
Initial data, boundary data, Cauchy data
It helps to name the kinds of data cleanly. Initial data prescribes the unknown (and sometimes its time derivative) on a single time slice, say t = 0, and we then evolve forward. Boundary data prescribes a condition along the spatial edge of the region — the Dirichlet, Neumann, or Robin conditions from last time — for all relevant times. And Cauchy data is the most demanding: it gives both the value u and its first normal derivative on a chosen curve or surface at once, like specifying both where a string is and how fast every point of it is moving.
The amount of data you need is tied to the order of the time-derivative. The heat equation u_t = k u_xx is first order in t, so one initial slice u(x, 0) is exactly the right amount. The wave equation u_tt = c^2 u_xx is second order in t, so you must give two slices: initial displacement and initial velocity, u(x, 0) and u_t(x, 0). That pair is Cauchy data on the line t = 0 — value and normal (here, time) derivative together. So "Cauchy data" and "initial data for a second-order-in-time equation" describe the same object when the carrier curve is a time slice.
EQUATION order in t data you must give on t = 0 ---------------------------------------------------------------- u_t = k u_xx (heat) 1 u(x,0) [one slice] u_tt = c^2 u_xx (wave) 2 u(x,0) and u_t(x,0) [Cauchy data] Laplace u_xx + u_yy = 0 -- data on the WHOLE boundary, not a slice
Why each type wants its own kind of data
This is the punchline the classification rung was building toward: the type dictates which data is well-posed. The discriminant B^2 - A C does not just sort equations for tidiness — it tells you, in advance, whether to feed the equation an initial slice or a closed boundary, and giving the wrong kind is not merely awkward, it is mathematically broken. Let us walk the three model equations and see each marriage of type and data.
- Hyperbolic (wave, u_tt = c^2 u_xx): finite speed, real characteristics, so signals ride along the characteristics x ± c t. Give Cauchy data — position u(x,0) and velocity u_t(x,0) — on the line t = 0, which is non-characteristic, and d'Alembert's formula hands you a unique solution that depends only on the domain of dependence, the little interval [x - c t, x + c t]. Initial / Cauchy data is exactly right here.
- Parabolic (heat, u_t = k u_xx): first order in time, infinite speed, smoothing. Give one initial slice u(x,0) plus boundary data for the sides of a rod, and the solution exists, is unique, and gets smoother as it evolves FORWARD. Asking for u_t(x,0) too would be one condition too many — the equation already fixes u_t from u_xx.
- Elliptic (Laplace, u_xx + u_yy = 0): no real characteristics, no propagation, pure equilibrium. There is no time direction to march along, so initial data on a slice makes no sense; instead you wrap data all the way around a closed boundary — Dirichlet or Neumann — and the interior settles into one smooth harmonic balance. Boundary data is the only well-posed choice.
Notice the pattern: hyperbolic and parabolic equations march forward from data on an open slice (an evolution problem), while elliptic equations have no marching direction and need data sealing the whole rim (a boundary-value problem). The number of real characteristic families you counted last rung — two, one, or none — is exactly what decides this. Real tracks for information mean you can pose data across them and evolve; no real tracks mean you must instead surround the region.
When data sits on a curve: characteristic or not?
Initial data on a flat time slice is the friendly special case. The general setting prescribes Cauchy data on a curve — value and normal derivative along some arc — and now a subtle condition appears that the slice hid. The curve must be non-characteristic: it must not run tangent to the characteristic directions of the equation. If it does, the equation cannot use the data to pry out the higher derivatives off the curve, and the data becomes either contradictory or powerless.
A tiny picture. For transport u_t + c u_x = 0 the characteristics are the lines x - c t = const. If you try to prescribe u along one of those very lines, you have just specified u along a curve on which u is forced to be constant — so either your data is constant (and tells you nothing about neighbouring characteristics) or it is non-constant (and contradicts the equation outright). Prescribe u along t = 0 instead — a line that crosses every characteristic transversally — and each characteristic carries its value off cleanly. Non-characteristic placement is what lets the data propagate.
The wrong data: a preview of ill-posedness
To feel why matching data to type matters, look at the two famous mismatches the rest of this rung dissects. First, prescribing Cauchy data for Laplace's equation. Laplace is elliptic; it wants boundary data, not value-and-normal-derivative on an open arc. Hadamard's classic example takes tiny boundary data of size 1/n that wiggles n times, and the corresponding harmonic solution grows like sinh(n y)/n^2, which blows up as n grows. An imperceptible change in the data produces an unbounded change in the solution. Existence and uniqueness can hold, yet the problem is still ill-posed.
Second, the backward heat equation — running the heat equation backward in time. Forward in time, heat smooths: a Fourier mode of frequency n decays like e^(-k n^2 t), so high-frequency wrinkles vanish fast. Reverse the clock and those same modes grow like e^(+k n^2 t), so the tiniest high-frequency speck of noise in your final data explodes. This is why you cannot reliably reconstruct yesterday's temperature from today's: diffusion erases the fine detail, and no stable inverse can put it back. Forward is fine, backward is ill-posed — the irreversibility is built into the parabolic type.
Both disasters share one diagnosis: the data was the wrong kind for the type. The fix is not cleverness but matching — boundary data for elliptic, forward initial data for parabolic, non-characteristic Cauchy data for hyperbolic. The next guide makes "well-posed" precise with Hadamard's three demands (existence, uniqueness, continuous dependence), and the two after that confront the ill-posed examples head-on and show how an energy estimate can prove uniqueness when the data is right. For now, carry this: choosing the data is half of stating a PDE problem, and the type tells you which half to choose.