The question uniqueness leaves on the table
Three guides back you met Hadamard's three demands for a well-posed problem: a solution must exist, it must be unique, and it must depend continuously on the data. Existence you have been chasing all rung — separation of variables, Fourier series, fundamental solutions all build a solution by hand. Continuous dependence is the stability you watched collapse in the ill-posed examples. But uniqueness has been the quiet one. We keep asserting the Dirichlet problem has one answer, the heat equation has one answer — yet how would you actually rule out a sneaky second solution hiding somewhere?
Here is the standard trick, and it is worth tattooing on the inside of your eyelids: to prove uniqueness for a linear problem, suppose there are two solutions and subtract them. Call them u1 and u2, both solving the same equation with the same data, and set w = u1 - u2. By the superposition principle — linearity, which you have leaned on since the first rung — w solves the homogeneous version of the equation, with zero data: zero source, zero boundary values, zero initial values. If you can show the only solution of that all-zeros problem is w = 0 identically, then u1 = u2 and the answer was unique all along. Uniqueness becomes a single sharp question: can a difference with no data of its own be anything but nothing?
Energy: one number that cannot lie
How do you prove a whole function w(x) — possibly wiggling all over a region — is identically zero? Checking it point by point is hopeless. The energy method answers with a beautiful piece of bookkeeping: collapse the entire function into one non-negative number, the energy, defined as the integral of w^2 (and sometimes its derivatives squared) over the region. The single most important property of this number is that it is non-negative, and it equals zero if and only if w is zero everywhere. So if you can corner the energy and force it to be zero, the function itself has nowhere left to hide — it must vanish.
You have actually already felt this idea twice on the ladder, dressed in physical clothes. The conservation of wave energy told you that a vibrating string's total energy — kinetic plus potential, the integral of u_t^2 + c^2 u_x^2 — stays exactly constant in time. The conservation of total heat told you that, with insulated ends, the integral of u over a rod never changes. Both are statements about an integral of a solution evolving in a controlled way. The energy method simply turns that physical instinct into a uniqueness weapon: choose the right integral, differentiate it in time, and use the equation to discover whether it can grow.
The heat equation, retired in four lines
Let us run the method on the heat equation u_t = k u_xx and watch it work. Suppose two solutions share the same initial temperature and the same Dirichlet boundary values, so their difference w satisfies w_t = k w_xx with w = 0 at both ends and w = 0 at the start. Define the energy E(t) as the integral of (1/2) w^2 across the rod. It starts at E(0) = 0 because w starts at zero. Now differentiate E in time, slip the time derivative under the integral, replace w_t by k w_xx from the equation, and integrate by parts. The boundary term dies because w = 0 at the ends, and what survives is a clean negative quantity.
E(t) = integral of (1/2) w^2 dx >= 0, E(0) = 0
E'(t) = integral of w w_t dx
= integral of w (k w_xx) dx (use the equation)
= -k integral of (w_x)^2 dx <= 0 (integrate by parts; ends vanish)
so E(t) <= E(0) = 0 and E(t) >= 0 ==> E(t) = 0 ==> w = 0Read the punchline slowly, because it is the whole game. The energy E(t) starts at zero, its rate of change E'(t) is never positive, so E can only stay put or sink — yet it can never sink below zero. Squeezed between a ceiling of zero and a floor of zero, E(t) must equal zero for all time. Therefore w is zero everywhere, u1 equals u2, and the heat problem has exactly one solution. Notice the a priori estimate hiding in plain sight: we bounded the size of w before ever writing down a formula for it, using only the equation and the data. That is the signature of every energy argument — control without a solution in hand.
Same idea, the wave equation and Laplace too
The lovely thing is how little changes for the wave equation u_tt = c^2 u_xx. Here you do not use the integral of w^2 itself; you use the physical energy E(t), the integral of (1/2)(w_t^2 + c^2 w_x^2), kinetic plus potential. Differentiate it, substitute the equation, integrate by parts, and the clamped (or insulated) ends kill the boundary term — but this time E'(t) comes out exactly zero, not negative. The wave equation conserves energy rather than dissipating it. Yet the conclusion is just as decisive: if w starts at rest with zero shape and zero velocity, then E(0) = 0, E stays at 0 forever, and since E is a sum of squares, both w_t and w_x must vanish — so w is constant, and the zero boundary data pins that constant to zero.
Even Laplace's equation Laplacian u = 0 bows to the same move, with no time variable at all. For two solutions of the Dirichlet problem agreeing on the boundary, the difference w is harmonic with w = 0 on the whole boundary. Form the Dirichlet energy, the integral of the gradient-squared of w over the region. Integrate by parts once: the boundary term vanishes because w = 0 there, and the interior term vanishes because Laplacian w = 0. So the integral of the gradient-squared is zero, which forces the gradient itself to be zero, so w is constant — and the zero boundary value makes that constant zero. This is the energy face of Dirichlet's principle: the harmonic function is the one of least energy, and least energy leaves no room for a rival.
From uniqueness to stability — and why backward heat fails
The same energy that proves uniqueness quietly upgrades to Hadamard's third demand, continuous dependence on the data. Run the heat argument again, but now let u1 and u2 start from slightly different initial data instead of identical data. The difference w no longer starts at zero, so E(0) is some small positive number — yet E'(t) is still <= 0, so E(t) <= E(0) for all later time. In words: the gap between two solutions can only shrink as the heat equation runs forward. A small error in the initial data stays small forever. That is stability, falling straight out of the same one-line inequality. On bounded domains a sharper tool, the Poincare inequality, even pushes this further, forcing the energy to decay exponentially.
Now turn the clock around and you see precisely why the backward heat equation from the previous guide was a disaster. Running diffusion in reverse flips the sign in front of k, and the energy inequality flips with it: instead of E'(t) <= 0 you get E'(t) >= 0, and worse, high-frequency wiggles make E explode. The very dissipation that gave forward heat its rock-solid stability becomes uncontrolled amplification when you reverse time. The energy method does not just prove the good problems are well-posed; by the same calculation, with the inequality pointing the wrong way, it diagnoses exactly why the bad ones are not. Uniqueness and stability are two readings of a single, honest accounting of energy.
Where uniqueness genuinely breaks: shocks and entropy
It would be dishonest to leave you thinking uniqueness is always this clean. The whole subtract-and-take-energy argument rested on linearity — the difference of two solutions solving a homogeneous problem — and that crutch snaps the moment the equation is nonlinear. Recall the rude surprise from the conservation-laws rung: a smooth solution of a nonlinear scalar conservation law can develop a shock in finite time, a genuine discontinuity, after which no classical solution exists at all. Past the shock you must allow non-smooth weak solutions — and here uniqueness collapses outright. The same initial data can spawn several different weak solutions, all obeying the equation in the weak sense. Energy alone no longer picks the winner.
Nature, of course, picks exactly one. The repair is to add an extra admissibility rule that the physical solution must obey, the entropy condition — a one-way inequality saying, in effect, that characteristics may crash into a shock but never spray out of one, mirroring the second law of thermodynamics where entropy only increases. Demand it, and uniqueness is miraculously restored: among all the weak solutions, exactly one is the entropy solution, and it is the one you get as the friction-free limit of a tiny bit of viscosity. So uniqueness is not free and not automatic. For nice linear problems an energy estimate hands it to you in four lines; for nonlinear conservation laws you must earn it with a deeper physical principle.
And at the very frontier, even existence-and-uniqueness is wide open. Whether the 3D Navier-Stokes equations always have a unique smooth solution for all time is the famous Navier-Stokes millennium problem — unsolved, with a million-dollar prize attached. The energy method does give a bound on one integral of the velocity, but that single estimate is not strong enough to rule out a blow-up in three dimensions, where the energy can pile into ever-finer swirls faster than the bound can control. Keep the honest picture in view: energy estimates are a workhorse you will reach for constantly, but they are not a master key that opens every uniqueness question.