Hadamard Well-Posedness

Three demands, not one

The first two guides of this rung handed you the cast: the three boundary conditions, and the Cauchy data that says which type of equation needs which kind of side information. You now know how to write down a problem — an equation in the interior, plus data on the boundary or the initial slice. But writing a problem down is not the same as asking a sensible question. Jacques Hadamard, around 1902, drew the line precisely. A problem is a well-posed problem when it passes three tests at once, and ill-posed the moment it fails any one of them.

HADAMARD WELL-POSEDNESS  =  all three at once

  1. EXISTENCE       there is at least one solution
  2. UNIQUENESS      there is at most one solution
  3. STABILITY       small change in data  ->  small change in solution
                     (continuous dependence on the data)

  fail any one  ->  ILL-POSED

Take them in turn. Existence says there really is a solution: you have not over-constrained the problem into emptiness. Recall the Neumann compatibility condition from guide 1 — demand a net outflow that does not balance and no steady solution exists at all; that is an existence failure. Uniqueness says there is at most one: you have not under-constrained it, leaving a whole family of answers. Laplace's equation with no boundary data has an infinite ocean of solutions; pin the boundary and the ocean drains to one. So far, so reasonable — these two are the demands anyone would think to make.

The third demand is the one that bites

The third demand, stability, is the subtle one — and the reason Hadamard's name is remembered. It asks for continuous dependence on the data: if you nudge the boundary or initial data by a tiny amount, the solution may move only by a correspondingly tiny amount. Phrase it as a promise: feed in nearly-equal data, get out nearly-equal answers. Without this, a unique solution that exists is still worthless, because the data you actually measure always carries error — a thermometer rounds, a sensor jitters. If a hair's-breadth wobble in the input can swing the output wildly, the exact answer to your idealized problem tells you nothing about the real one.

Make the promise quantitative and you arrive at the working tool of the trade. Stability is usually a statement of the form: the size of the difference of two solutions is bounded by a constant times the size of the difference of their data, measured in some sensible norm. Such a bound — solution controlled by data, with a constant that does not blow up — is called an a priori estimate. "A priori" because you can prove the bound before you ever construct the solution; it is a property of the equation, not of any particular answer. Continuous dependence is exactly an a priori estimate in disguise, and the next two guides are built around finding them.

Stability gives you uniqueness for free

Here is a small but lovely structural fact that ties the three demands together. A genuine stability estimate already contains uniqueness — you do not have to prove uniqueness separately. Watch the logic. Suppose u and v both solve the same problem with identical data. Apply the stability bound to the pair: the size of u - v is at most a constant times the size of (data of u) - (data of v). But the data are identical, so that difference is exactly zero. A nonnegative quantity bounded above by zero must itself be zero, so u - v = 0, hence u = v.

So among Hadamard's three, stability is the heavyweight: nail it and uniqueness comes along as a free corollary. Existence remains a genuinely separate promise — building a solution and bounding it are different labors, as guide 1 already warned (the maximum principle hands you uniqueness almost for nothing, yet says nothing about whether any solution exists). This is why working mathematicians chase a priori estimates so relentlessly. One good estimate can quietly settle two of the three conditions at a stroke, leaving only the constructive question of existence to handle on its own.

1. To test well-posedness of a problem, check the three demands in turn — existence, uniqueness, stability.
2. Try for a stability (a priori) estimate first: solution size bounded by data size, with a non-exploding constant.
3. If you get it, uniqueness falls out automatically (equal data forces equal solutions), so only existence is left to settle.
4. If no such estimate can possibly hold — small data swinging the solution wildly — the problem is ill-posed, and you stop trusting any answer it gives.

When the third demand fails

Now the cautionary half, the part that made Hadamard worry in the first place. Two famous problems exist and are unique, yet shatter stability — and the next guide of this rung dissects them in full. The first is the backward heat equation: take the heat equation u_t = k u_xx and run it backwards in time, asking for the temperature in the past given the temperature now. Forward in time, heat diffusion is the great smoother — every jagged initial profile melts into a bland curve. Run the film backwards and you are demanding that a bland curve un-melt into spikes, and the finer the spike, the more violently the data must have differed an instant earlier.

Quantify it and the failure is total. A high-frequency wiggle in the current temperature of size epsilon, of wavenumber n, traced backward a fixed time, grows like epsilon times e^(k n^2 t) — and that factor races to infinity as n grows. So data differing by less than any thermometer could ever detect can correspond to past states that differ by an unbounded amount. No constant on earth can bound the solution by the data; continuous dependence is dead, and with it well-posedness. This is the price of type-dependent well-posedness: the heat equation is well-posed forward, ill-posed backward. Same equation, opposite arrow of time, opposite verdict.

Hadamard's own example was the second one: prescribing Cauchy data for Laplace's equation. As guide 2 explained, Cauchy data — both the value and the normal derivative on a curve — is the natural data for hyperbolic equations like the wave equation, where it is perfectly well-posed. Hand the same kind of data to the elliptic Laplace equation, though, and disaster: Hadamard exhibited data as small as (1/n) sin(n x) on a line, whose unique harmonic continuation grows like (1/n) sin(n x) sinh(n y), exploding as n increases. Vanishingly small data, unboundedly large solution. The lesson the next guide will hammer home: ellipticity is destiny, and Cauchy data on an elliptic equation is the wrong question.

Ill-posed is not the same as wrong

One honest correction before you leave, because beginners over-read the word "ill-posed." It does not mean the problem is forbidden, meaningless, or false. The backward heat equation is exactly what a doctor faces reconstructing where a tumor started from how it looks today, and Cauchy-Laplace problems arise genuinely in geophysics and medical imaging. These are real, important questions. "Ill-posed" is a verdict only about raw stability: as posed, with arbitrary noisy data, the answer cannot be trusted. The cure is regularization — adding extra information (the answer is smooth, or small, or matches a prior model) that tames the explosion. A whole branch of applied mathematics, inverse problems, lives precisely here.

Carry forward one clean picture. Hadamard well-posedness is a three-legged stool — existence, uniqueness, continuous dependence — and a stool stands only on all three legs. The first two legs feel obvious; the third, stability, is the one that secretly does the heavy lifting, hands you uniqueness for free, and separates a question you can actually compute from one that merely looks like mathematics. The next guide drops the backward heat and Cauchy-Laplace problems under a microscope to watch stability die; the one after builds the energy estimates that prove it lives when the problem is posed right.