Where the last guide left us
In the previous guide we took a uniformly elliptic operator in divergence form — think of the model -Laplacian u = f on a domain with u = 0 on the boundary — and refused to differentiate twice. Instead we multiplied by a test function v, integrated by parts once, and landed on a single clean statement: find u in the Sobolev space so that B(u, v) = L(v) for every v. The left side B is a bilinear form that eats two functions and returns a number; the right side L is a linear functional carrying the data f. A weak solution is exactly any u that makes this hold for all test v.
That reformulation is honest progress, but it is not yet an answer. The weak formulation only rephrases the question; it does not promise that a u answering it exists, nor that there is just one. The whole job of this guide is to supply that promise. The remarkable thing is that we can do it abstractly — purely from the shape of B and L on the Sobolev space — without ever writing down a formula for u and without yet knowing whether u is twice differentiable. Existence first; smoothness is the next guide's problem.
Lax-Milgram: coercivity buys existence outright
Here is the cleanest case. Suppose B is bounded — |B(u, v)| stays no bigger than a constant times the size of u times the size of v — and B is coercive — B(u, u) is at least a positive constant times the squared size of u. Coercivity is the analytic muscle: it says the form can never collapse, that pushing u to have a large Sobolev norm forces B(u, u) to be large too. Under exactly these two conditions, the Lax-Milgram theorem declares that for every right-hand side L there is one and only one u with B(u, v) = L(v) for all v. Existence and uniqueness, in one stroke.
Why is this believable? When B is also symmetric, B(u, u) is literally the energy, and the weak solution is the unique function that minimizes the energy functional — that is Dirichlet's principle from the previous guide. Lax-Milgram is the upgrade that drops the symmetry assumption: even for operators with a first-order drift term, where there is no energy to minimize, the same coercivity still pins down a unique u. The proof is the Riesz representation theorem wearing a slightly looser coat. You do not need the proof to use it; you need to check two inequalities.
For the model -Laplacian u = f with zero boundary data, coercivity is not free — it is a gift from geometry. The form is B(u, v) = integral of (gradient u dot gradient v), so B(u, u) is the integral of |gradient u|^2, which controls only the gradient, not u itself. The Poincare inequality is exactly what rescues us: on a bounded domain, a function that vanishes on the boundary cannot be large unless its gradient is large, so the integral of |gradient u|^2 already controls the full Sobolev norm. That single inequality is what makes the Laplacian coercive, and hence what makes Lax-Milgram fire.
When coercivity fails: the Fredholm alternative
Coercivity is generous but not universal. Add a zero-order term and consider -Laplacian u - c u = f with c a positive constant. Now B(u, u) = integral of |gradient u|^2 minus c times integral of u^2, and for large c the negative piece can overwhelm the positive one — coercivity is gone, and Lax-Milgram has nothing to grip. This is not a technicality. When c hits an eigenvalue of the Laplacian, the homogeneous problem -Laplacian u - c u = 0 has a genuine nonzero solution, so uniqueness must fail. The theory has to be honest about this.
The rescue is the Fredholm alternative, and it is one of the most satisfying dichotomies in analysis. It says: either the homogeneous problem has only the zero solution, in which case the inhomogeneous problem -Laplacian u - c u = f has exactly one solution for every f — just like a matrix that is invertible; or the homogeneous problem has nonzero solutions, in which case f must satisfy a finite list of orthogonality conditions for any solution to exist, and when it does, the solution is non-unique by exactly the homogeneous solutions. Existence and uniqueness rise and fall together, governed by a finite-dimensional kernel.
This is the exact infinite-dimensional echo of the linear algebra fact that for a square matrix A, either A x = b is uniquely solvable for all b, or A has a nontrivial null space and A x = b is solvable only when b is orthogonal to the null space of the transpose. The miracle is that an elliptic operator on an infinite-dimensional Sobolev space still obeys this finite-dimensional law. The hidden engine is compactness: the Rellich-Kondrachov theorem makes the relevant inverse operator a compact operator, and compact operators are precisely the ones that keep the matrix-like dichotomy intact.
Two existence theorems, side by side
------------------------------------------------------
B coercive (Lax-Milgram)
for EVERY f -> exactly one weak solution u
model: -Lap u = f, u = 0 on boundary
B not coercive (Fredholm alternative)
EITHER homogeneous has only u = 0
-> exactly one u for every f
OR homogeneous has nonzero solutions phi_1..phi_k
-> u exists only if L(phi_i) = 0 for all i
-> and then u is unique up to + span{phi_i}
model: -Lap u - c u = f, c near an eigenvalue
------------------------------------------------------How the proof actually catches a solution
It is worth seeing the constructive spirit, because it explains why we never needed a formula. The idea is to solve the problem in finitely many dimensions first and then pass to the limit. Pick a growing list of basis functions, demand B(u, v) = L(v) only against those finitely many v, and you are left with an ordinary linear system — a matrix equation you can solve. As you allow more and more basis functions, these approximate solutions get trapped: coercivity forces them to stay bounded, and a bounded sequence in a Sobolev space has a weakly convergent subsequence. The limit is the weak solution.
That recipe is the Galerkin method, and it is far more than a proof device — it is the backbone of the finite element method that engineers run every day. The same weak formulation that lets us prove existence is the formulation a computer actually discretizes. Building the theory on B(u, v) = L(v) rather than on a pointwise PDE is what makes the bridge from abstract existence to real computation so short.
- Multiply the PDE by a test function and integrate by parts once to reach B(u, v) = L(v) — the weak formulation.
- Check whether B is bounded and coercive on the Sobolev space (Poincare often supplies the coercivity).
- If coercive, invoke Lax-Milgram: one unique weak solution for every right-hand side, done.
- If not coercive, run the Fredholm alternative: check the homogeneous kernel, then test L against it to decide solvability.
- Only now, with existence secured, hand the weak solution to elliptic regularity to ask whether it is smooth.
Honest fine print
Three honesties before you climb on. First, everything here rests on uniform ellipticity: the matrix of coefficients must stay positive-definite by a fixed margin everywhere, with no degeneracy. Lose that and the bilinear form can fail to be bounded or coercive, and these theorems simply do not apply — the same way running the heat equation backwards or posing Cauchy data for Laplace's equation breaks well-posedness. The structure of the operator, not just its order, is what licenses the method.
Second, a weak solution is genuinely weaker than a classical one. It need only have a single derivative in an averaged, integrated sense — it lives in a Sobolev space, not in the world of twice-differentiable functions. So proving existence does NOT yet mean we have solved the original PDE pointwise. That gap is real, and closing it is exactly the work of elliptic regularity: for nice data and a nice domain, the weak solution turns out to be as smooth as the data allows. But that is a theorem to be earned, not an assumption to be made.