What we are hunting for
In the previous guide you pinned down Hadamard's three demands for a well-posed problem: a solution must exist, be unique, and depend continuously on the data. That third demand is the fragile one, and it is the whole story of this guide. A problem can have a perfectly good unique solution on paper and still be useless in practice, because real data always carries noise, and if a hair of noise can be amplified without limit, the answer you compute has nothing to do with the answer you wanted.
When that happens the problem is ill-posed. The word is precise, not a vague complaint: ill-posed means at least one of Hadamard's three conditions breaks, and in the two famous cases below the one that breaks is always continuous dependence on the data. Keep one slogan ready as you read: existence and uniqueness are about whether an answer is there; continuous dependence is about whether that answer is trustworthy.
And there is a deeper pattern underneath. From the classification rung you learned that an equation's type — elliptic, parabolic, or hyperbolic — dictates which data is well-posed. Ill-posedness is almost always the punishment for pairing data with the wrong type of equation: giving an elliptic equation the initial-value data that belongs to a hyperbolic one, or asking a parabolic equation to run against its built-in arrow of time. This is type-dependent well-posedness seen from its dark side.
Example one: the backward heat equation
Start with the request that sounds most reasonable. You have a temperature profile now; what did it look like an hour ago? Forward in time the heat equation u_t = k u_xx is the gold standard of a well-posed problem. Reversing the question means solving the backward heat equation — equivalently, running u_t = k u_xx with time pointing the other way — to reconstruct the past from the present. It looks innocent. It is a trap.
The cleanest way to see the explosion is through Fourier modes, the engine you met when the heat equation first showed its irreversibility. On an interval each mode sin(n pi x / L) evolves forward by the factor e^(-k (n pi / L)^2 t): the decay of Fourier modes carries n^2 in the exponent, so high frequencies are crushed fast. Going backward you must divide by that factor — multiply by e^(+k (n pi / L)^2 t) — and that exponent grows without bound as n grows. The forward equation's gentlest act becomes the backward equation's catastrophe.
Make it a tiny worked example. Suppose the true present is perfectly flat, u = 0. Now corrupt it by an unmeasurably small ripple of amplitude epsilon at high frequency n: the data error is epsilon, microscopic. Reconstruct the past one unit of time earlier and that ripple gets multiplied by e^(+k (n pi / L)^2), a number that is astronomical once n is large. Choose n big enough and an epsilon = 10^(-10) ripple in the present becomes a mountain in the inferred past. Tiny cause, unbounded effect — continuous dependence is dead, and the problem is ill-posed.
Example two: Cauchy data for Laplace
The second classic is Hadamard's own, and it teaches a different lesson: the type of equation cares which data you bring. Laplace's equation Laplacian u = u_xx + u_yy = 0 is elliptic, and an elliptic equation is the happy home of boundary data — fix u all the way around a closed region (the Dirichlet problem) and you get one good, stable solution. But suppose you instead try to march it forward from one edge, prescribing both the value and the normal slope along a single line. That is Cauchy data — the value-and-derivative pair that hyperbolic equations like the wave equation live on — and for Laplace it is poison.
Hadamard built an explicit booby trap to show it. Take the strip with y >= 0, demand u = 0 and u_y = (1/n) sin(n x) along the line y = 0. As n grows the data shrinks toward zero everywhere — its amplitude 1/n vanishes, and so do all its slopes, so this is an arbitrarily small perturbation of the all-zero data, whose solution is plainly u = 0. Yet the solution forced by Laplace's equation is u(x,y) = (1/n^2) sin(n x) sinh(n y), and sinh(n y) grows like e^(n y)/2. For any fixed height y > 0, the prefactor (1/n^2) e^(n y) blows up as n grows. Data of size 1/n, solution of size e^(n y)/n^2: the ratio is unbounded.
data on the line y = 0: u = 0, u_y = (1/n) sin(n x) <- amplitude 1/n -> 0
Laplace forces the solution: u(x,y) = (1/n^2) sin(n x) sinh(n y)
sinh(n y) ~ e^(n y) / 2
at any fixed height y > 0: |u| ~ e^(n y) / (2 n^2) -> infinity as n grows
vanishing CAUSE (1/n) -> exploding EFFECT (e^(n y)/n^2) = continuous dependence FAILSCompare this with the wave equation to feel why type is the culprit. Hand the hyperbolic equation u_tt = c^2 u_xx the same kind of Cauchy data — an initial displacement and an initial velocity — and the modes evolve by sines and cosines, things of size one that never blow up; that Cauchy problem is well-posed. Hand the elliptic Laplace equation that very same data and the modes evolve by sinh and cosh, which grow exponentially. Same data, opposite fate, and the only thing that changed was a sign in the equation — the difference between u_xx + u_yy and u_xx - u_yy — which is exactly the difference between elliptic and hyperbolic.
Why existence and uniqueness are not enough
A subtle trap lurks here, and it is worth naming so you never fall into it. The Cauchy-Kovalevskaya theorem guarantees that, for an analytic equation with analytic Cauchy data, a unique analytic solution exists near the data surface. Laplace's equation with the analytic data above satisfies every hypothesis — so existence and uniqueness genuinely hold. The solution is right there, exact and unique. Cauchy-Kovalevskaya simply says nothing about continuous dependence, and that is the leg the problem is missing.
This is the punchline of Hadamard's whole framework. A problem can clear two of the three bars — existence and uniqueness — with room to spare and still be worthless, because the missing third bar, continuous dependence, is the one that connects mathematics to measurement. Existence says the answer is somewhere; uniqueness says there is only one; continuous dependence says you can actually find it from imperfect data. Drop the third and the first two are a cruel joke: a unique answer you can never reliably compute.
The honest map: which data fits which type
Step back and the two cautionary tales fold into one rule of thumb. A PDE's general solution is fixed not by a few constants but by arbitrary functions, and the side data is what pins those functions down — so the data must match the equation's type or the matching goes haywire. Elliptic equations (Laplace, Poisson) want boundary data wrapped all the way around a region; that is well-posed, and prescribing Cauchy data on an open arc is not. Hyperbolic equations (the wave equation) want Cauchy data — value and velocity — at an initial time; that is well-posed, with information travelling at finite speed.
Parabolic equations (the heat equation) want initial data plus boundary data, and they want it pointing the right way in time — forward, with the arrow of diffusion, never against it. The backward heat equation breaks this last clause, the Cauchy-Laplace problem breaks the elliptic clause, and both are punished by the same loss of continuous dependence. The single sentence to carry away: the elliptic/parabolic/hyperbolic label is not bookkeeping — it is a binding contract about which data you are allowed to supply.
One last honesty check, so you do not over-learn the lesson. Ill-posed problems are not forbidden territory — they are everywhere in real applications (medical imaging, geophysics, every inverse problem worth solving). The point is not to avoid them but to recognise them, and then to add the extra information — smoothness bounds, prior assumptions, penalty terms — that turns the wild problem into a nearby tame one you can actually compute. The next guide builds the positive tool behind the well-posed cases: an energy estimate, the clean inequality that proves uniqueness and continuous dependence for the equations that deserve them.