Flip the question: from solving to minimizing
Everything in this rung has had the same shape: you are handed an equation like the elliptic equation -Laplacian u = f and asked to find the function u that satisfies it. The calculus of variations flips the question around. Instead of asking which function makes this equation true, it asks which function makes some number as small as possible. That number is an energy: you feed in a candidate function and out comes a single real value measuring how 'expensive' that candidate is. The winner — the cheapest function — turns out to solve the PDE, not by coincidence but by an exact mechanism we are about to see.
The model energy is the Dirichlet energy E[u] = (1/2) integral |grad u|^2 - integral f u, taken over functions u that match the prescribed boundary data. Read the first piece as a roughness penalty: |grad u|^2 is large wherever u is steep, so this term punishes wiggling and rewards smoothness. The second piece is a reward for lining up with the source f. A minimizer of E is the function that balances 'be as flat as you can' against 'lean toward where f wants you' — exactly the tug-of-war that the equation -Laplacian u = f describes. This is the precise sense behind Dirichlet's principle: solving Poisson's equation and minimizing the Dirichlet energy are the same problem wearing two costumes.
The Euler-Lagrange equation: where the slope is zero
How do you find the minimum of ordinary calculus? You set the derivative to zero. The calculus of variations does exactly this, but on a space of functions instead of a space of numbers. At a minimizer u, nudge it by a tiny test bump: replace u with u + e*v, where v is a smooth function vanishing on the boundary (so the boundary data stays fixed) and e is a small number. Now E[u + e*v] is just an ordinary function of the single number e, and if u is the minimum it must bottom out at e = 0. So the ordinary derivative d/de of E[u + e*v], evaluated at e = 0, must vanish — for every allowed bump v.
Carry out that derivative for the Dirichlet energy and you get integral (grad u . grad v) - integral f v = 0 for all v. Integrate the first term by parts (the boundary term dies because v vanishes there) and it becomes integral (-Laplacian u - f) v = 0 for all v. Here is the magic step. If a fixed function multiplied against every test function v always integrates to zero, that function must itself be zero everywhere — there is nowhere for it to hide. So -Laplacian u - f = 0, i.e. -Laplacian u = f. The minimizer satisfies the PDE. This recovered equation is the Euler-Lagrange equation of the energy.
A second derivative tells minimum from saddle
Be honest about one thing: setting the first variation to zero only says u is a critical point, the function-space analogue of a flat spot. In ordinary calculus a flat spot can be a minimum, a maximum, or a saddle. The same caution applies here, and it matters: many famous PDEs are Euler-Lagrange equations of energies that have no minimum at all — minimal surfaces and harmonic maps live at saddle points, found by min-max games rather than by descending to a floor. So a critical point of the energy is a solution candidate, and you still owe an argument for why the one you want exists.
What makes the Dirichlet energy so well-behaved is that |grad u|^2 is a bowl: as a function of u it curves upward in every direction, so it has a genuine bottom, not a saddle. Technically the second variation is integral |grad v|^2, which is never negative — the energy is convex. Convexity is the function-space version of 'this is a valley, not a pass', and it is exactly the structure that guarantees a critical point really is the minimum. This convexity is the same fact you saw as coercivity in guide 1: the energy grows without bound as ||u|| grows, so its graph really does cup upward and trap a lowest point.
The direct method: catching a minimizer that exists
Now the genuinely hard question. We know that if a minimizer exists it solves the PDE. But does a minimizer exist? Naively it should: the energy is bounded below, so it has an infimum, so just take functions whose energy marches down toward that infimum and grab the limit. This sounds airtight and it nearly cost the subject its life — the historical scandal is that Dirichlet's principle was used for decades before anyone noticed the limit might not exist, because a sequence of functions can squirm toward the bottom while refusing to converge to any actual function. Weierstrass found explicit counterexamples. The naive argument has a real hole.
The repair is the direct method of the calculus of variations, and it is the variational payoff of the whole earlier rung. The trick is to stop demanding ordinary convergence and accept weak convergence in a Sobolev space instead. A minimizing sequence with bounded energy is bounded in that space, and bounded sets in a Sobolev space are weakly compact — so a subsequence does converge weakly to some limit u. Weak convergence is loose enough that the limit is guaranteed to exist, yet the energy is lower semicontinuous under it (the convex |grad u|^2 term cannot jump up in the limit, only possibly down), which is exactly enough to force E[u] to be the infimum. The limit is the minimizer. Existence, rescued.
- Write the energy E[u] and confirm it is bounded below and coercive (grows as ||u|| grows), so a finite infimum exists and minimizing sequences cannot run off to infinity.
- Take a minimizing sequence u_n with E[u_n] decreasing toward the infimum; coercivity makes it bounded in the Sobolev space.
- Extract a weakly convergent subsequence u_n -> u (bounded sets are weakly compact) — this is the step the naive argument was missing.
- Use weak lower semicontinuity (from convexity) to conclude E[u] <= liminf E[u_n] = infimum, so u attains the minimum and, by Euler-Lagrange, solves the PDE as a weak solution.
From weak minimizer to classical solution — and the eigenvalue echo
The direct method hands you a winner u living in a Sobolev space — a weak solution, not yet a smooth, classical one. But you already own the upgrade. The elliptic regularity theory of guide 3 says that a weak solution of a nice elliptic equation with smooth data is automatically smooth, and you bootstrap derivative by derivative until it is as classical as you like. So the full elliptic story now reads as a clean assembly line: minimize the energy (this guide) to get existence, then regularize (guide 3) to get smoothness. Variations supply the solution; regularity polishes it.
Variations also explains the eigenvalues of guide 4. The principal eigenvalue is not pulled out of thin air — it is the minimum of a quotient. Minimize the Rayleigh quotient R[u] = integral |grad u|^2 / integral u^2 and the minimum value is the first eigenvalue lambda_1, while the minimizer is the first eigenfunction; the Euler-Lagrange equation of that quotient is precisely -Laplacian u = lambda u. The higher eigenvalues come from the same idea repeated with min-max, which is exactly the min-max characterization you met in guide 4. So eigenvalues are variational too — every one of them is the bottom of some constrained energy.
Two honest edges to take up the ladder. First, the direct method's smooth ending relied on convexity: the energy was a bowl. Strip that away — minimal surfaces, harmonic maps, equations with the wrong sign — and minimizers may fail to exist, may be non-unique, or may form singularities, and you genuinely need the harder min-max and mountain-pass machinery to find the saddle-point critical points instead. Second, the whole correspondence works because the operator was an Euler-Lagrange equation of some energy; not every PDE is. The honest scope of this beautiful method is: equations that are the gradient of an energy, ideally a convex one — a huge and central class, but not all of them.