What the previous guide left on the table
By the end of the last guide you had performed a small act of alchemy. The boundary-value problem -Laplacian u = f with u = 0 on the edge — a statement about second derivatives that demands u be twice differentiable — had become a weak formulation: find u in the space H^1_0 such that the integral of (grad u) dotted with (grad v) equals the integral of f times v, for every test function v in H^1_0. You wrote it compactly as B(u,v) = L(v) for all v. The whole problem now lives inside one Hilbert space, demanding only one derivative of u, in the weak sense.
But rewriting a question is not answering it. We have a beautiful equation B(u,v) = L(v); does a u that satisfies it actually exist? And if one does, is it the only one, or could two different functions both pass every test? These are the existence and uniqueness questions, and a clever rearrangement does not settle them — a vivid-looking formula can still describe nothing real. The triumph of this guide is a single theorem that answers both at once, and the price of admission is astonishingly low: you only need to check two inequalities about the bilinear form B.
Two inequalities are all the theorem asks
Here is the whole machine in one breath. Work in a Hilbert space H — for us H = H^1_0 with the Sobolev inner product. Suppose B(u,v) is a bilinear form and L(v) is a bounded linear functional on H. If B meets two conditions, then there exists exactly one u in H with B(u,v) = L(v) for all v. The first condition is boundedness (also called continuity): there is a constant C with |B(u,v)| no bigger than C times the norm of u times the norm of v — B cannot blow up. The second is coercivity: there is a constant alpha greater than 0 with B(u,u) at least alpha times the norm of u squared — B has a genuine floor under it. That is the entire hypothesis.
Lax-Milgram, in three lines: bounded : |B(u,v)| <= C * ||u|| * ||v|| (B never explodes) coercive : B(u,u) >= alpha * ||u||^2, alpha > 0 (B has a floor) ===> exactly one u in H solves B(u,v) = L(v) for every v.
Feel why these two pull in opposite directions and together pin everything down. Boundedness says B is not too big — it is a continuous, well-behaved object, so small changes in the inputs make only small changes in the output. Coercivity says B is not too small — feeding a function to itself can never give back something puny compared to that function's size, so B can never collapse a nonzero u down to nothing. A form squeezed between a ceiling and a floor like this behaves almost exactly like an honest inner product, and an inner product is precisely the structure a Hilbert space is built to handle.
Coercivity is the crux — and Poincare is what supplies it
Of the two conditions, boundedness is usually the easy one — for our Dirichlet Laplacian it follows from the Cauchy-Schwarz inequality in a line. Coercivity is where the real work, and the real interest, lives. For B(u,u) equal to the integral of |grad u|^2, coercivity demands that the integral of |grad u|^2 control the full H^1 norm — gradient part and the size of u itself. But at first glance B(u,u) only sees the gradient; it is blind to a function that is large yet flat. What rescues us is a theorem you met two guides ago: the Poincare inequality.
Poincare says that on H^1_0 — functions vanishing on the boundary — the norm of u itself is bounded by a constant times the norm of grad u. Because a function pinned to zero around the rim cannot wander far without building up a steep gradient somewhere, its size is controlled by its slope. That single inequality lets the gradient-only quantity B(u,u) dominate the whole H^1_0 norm, delivering coercivity with a concrete constant. Notice how much rides on the boundary condition: it is exactly the u = 0 constraint, baked into H^1_0, that makes Poincare true and the whole argument run. This is the deep reason the homogeneous Dirichlet problem is the cleanest case in the theory.
Why it works: the energy picture and Riesz
You do not have to take existence on faith; there is a picture. When B is symmetric — B(u,v) = B(v,u), true for our Laplacian — coercivity makes B(u,u) an honest energy, and solving B(u,v) = L(v) is exactly the same as minimizing the energy functional J(u) = (1/2) B(u,u) - L(u) over all of H^1_0. Coercivity means this energy bowl curves upward in every direction and shoots to infinity far out, so it has a genuine bottom; that unique lowest point is the solution. This is the variational heart of the matter — the modern, rigorous version of Dirichlet's old principle that nature settles into the state of least energy.
For the symmetric case, the engine under the hood is the Riesz representation theorem: in a Hilbert space, every bounded linear functional is just an inner product against one fixed element. Coercivity lets you treat B(u,v) itself as a new inner product on H (boundedness and coercivity are exactly what guarantee it is one), and then L(v) — being bounded — must be that new inner product against some unique u. That u is your solution, and Riesz hands it over already proved unique. The general, possibly non-symmetric case needs a short extra argument (a contraction-mapping fixed point), but the symmetric story above is the soul of it.
From a proof of existence to an actual approximation
Lax-Milgram proves a solution exists, but it lives in an infinite-dimensional space — not yet something a computer can hold. The bridge is the Galerkin method, and it is disarmingly natural: instead of testing against all of H^1_0, pick a finite-dimensional subspace V_n (say, n carefully chosen basis functions) and demand B(u_n, v) = L(v) only for v in V_n. That collapses the infinite family of equations into an n-by-n linear system you can actually solve. The same two inequalities that gave existence on the whole space give it on every subspace too — coercivity is inherited — so each finite problem has a unique solution u_n, no extra hypotheses required.
Then comes the payoff that makes the whole edifice worth building. As you enrich the subspaces — let n grow so V_n fills out more and more of H^1_0 — the approximations u_n converge to the true weak solution u, and coercivity even hands you an error estimate (this is the celebrated Cea's lemma): u_n is within a fixed constant of the best possible approximation from V_n. Choose those basis functions to be little tent-shaped functions on a mesh and Galerkin becomes the finite element method, the workhorse that simulates bridges, airframes, and heat flow in the real world. The abstract two-inequality theorem and the engineer's mesh are the same idea wearing different clothes.
Honest limits, and what comes next
Be clear about what Lax-Milgram does and does not promise. It requires coercivity, and many real operators are not coercive — add a large enough zero-order term to make B(u,u) able to go negative and the floor is gone. When coercivity fails the theorem is silent, but not the theory: the Fredholm alternative takes over, trading the unconditional 'always exactly one solution' for a sharp either-or — either the problem has a unique solution, or the homogeneous problem has nontrivial solutions and existence then needs the data to be orthogonal to them. Lax-Milgram is the clean, coercive corner of a larger landscape, not the whole of it.
And one honesty about what kind of solution you have won. Lax-Milgram delivers a weak solution living in H^1_0 — a function with merely one weak derivative, which need not be twice differentiable, so it is not yet a classical solution in the original sense of '-Laplacian u = f holds at every point'. We have existence and uniqueness, but in a deliberately enlarged space. The natural and pressing question — does this weak solution, when the data f and the domain are smooth, secretly turn out to be smooth after all, so it solves the PDE in the old classical way? — is exactly the subject of the final guide in this rung, on regularity. Lax-Milgram opened the door; regularity tells us how nice the room inside really is.