JOVANA
Explore Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

The Weak Formulation of a Boundary-Value Problem

We have the spaces; now we put them to work. Multiply a boundary-value problem by a test function, integrate by parts once, and the second-order PDE collapses into a single, far gentler statement: a bilinear form equals a linear form, for every test function. That one move is the doorway to existence proofs and to the entire finite element method.

Why the classical problem is too demanding

Take the cleanest model elliptic problem there is: find u with -Laplacian u = f inside a region, and u = 0 on the boundary. This is the Dirichlet problem for Poisson's equation — the steady temperature of a plate held cold at its rim while heat is pumped in at rate f. Asking for a classical solution means asking for a u you can differentiate twice, everywhere, and have u_xx + u_yy land exactly on f at every single point. That is a heavy demand.

The trouble is that perfectly reasonable physics violates it. Let f have a jump — a heat source switched on over half the plate and off over the other half — or let the region have a sharp re-entrant corner, like an L-shaped room. Then the true temperature is continuous and physically unambiguous, yet its second derivatives blow up or fail to exist right at the seam. There is no classical solution, even though everyone agrees what the answer should be. We met exactly this rigidity at the start of the distributions rung; the spaces H^k and H^1_0 from the last two guides are the cure built specifically for it.

Multiply by a probe, integrate by parts once

Here is the single move that drives everything. Suppose for a moment u is a classical solution. Take any smooth probe v that vanishes on the boundary, multiply the equation -Laplacian u = f by v, and integrate over the whole region. On the left you have the integral of (-Laplacian u) v; on the right, the integral of f v. So far nothing has changed — we have only smeared the pointwise equation against a single weight. The magic is in what we do to the left side next.

Integrate by parts — in higher dimensions this is the divergence theorem, Green's first identity. It moves one derivative off u and onto v: the integral of (-Laplacian u) v becomes the integral of (grad u) dot (grad v), plus a boundary term. Because our probe v is zero all along the boundary, that boundary term vanishes completely. What is left is breathtakingly more symmetric: the integral of grad u dotted with grad v equals the integral of f v. Notice the count: u now carries only one derivative, not two, and v carries one as well. The second-order problem has become a first-order one in disguise.

classical :  -Laplacian u = f   in Omega,    u = 0   on boundary

multiply by probe v (v = 0 on boundary), integrate over Omega:

   integral of (-Laplacian u) v  =  integral of f v

Green / integrate by parts  (boundary term drops since v = 0 there):

   integral of grad(u) . grad(v)  =  integral of f v     for every such v

  ^ one derivative on u, one on v -- both live happily in H^1
The whole derivation on one card: the second-order equation becomes a balanced first-order identity once the boundary term is killed by v = 0.

Choosing the right home for u and v

Look hard at the surviving identity, integral of grad u dot grad v = integral of f v. To even write the left side we only need grad u and grad v to be square-integrable — that is, we need u and v to have one weak derivative in L^2. That is precisely the membership card of the Sobolev space H^1. And the demand that v be zero on the boundary, read through the trace from the last guide, is exactly the defining property of H^1_0, the closure of the compactly supported smooth functions. So the natural home for both the unknown u and the probes v is H^1_0. The boundary condition u = 0 is no longer an extra equation tacked on — it is built into the very space we search in.

This relocation is the quiet genius of the method. A classical solution had to be twice differentiable; a member of H^1_0 only needs one weak derivative, and it is allowed corners, kinks, and the mild blow-ups that real data forces. The earlier guides paid the dues for this: H^1_0 is complete, so limits of approximate solutions stay inside it, and the Poincaré inequality guarantees that on a bounded region the gradient alone controls the function. We will lean on exactly that fact next.

The bilinear form and the weak solution

Now we name the two sides. Write a(u, v) for the integral of grad u dot grad v, and write L(v) for the integral of f v. The object a(u, v) is a bilinear form: linear in u when v is held fixed, and linear in v when u is held fixed — it eats two H^1_0 functions and returns a number, the way a dot product does. The object L(v) is a linear form: feed it one probe, get one number. The entire boundary-value problem has now been distilled into a single clean sentence.

A function u in H^1_0 is called a weak solution if a(u, v) = L(v) holds for every v in H^1_0. This is the weak formulation of the Dirichlet problem. Read it as a balance condition: u is the configuration that, when you probe its gradient against any allowed test direction, responds exactly as the source f does against that same direction. No second derivatives appear anywhere. The reformulation is faithful, too — if a weak solution happens to be smooth, you can integrate by parts in reverse and recover -Laplacian u = f pointwise, so we have lost nothing real, only the excess rigidity.

  1. Start from the classical PDE plus its boundary condition, here -Laplacian u = f with u = 0 on the boundary.
  2. Multiply by an arbitrary test function v that vanishes on the boundary, and integrate over the region.
  3. Integrate by parts once (Green's identity); the boundary term dies because v = 0 there, leaving the integral of grad u dot grad v = the integral of f v.
  4. Read off the natural space H^1_0 — one weak derivative for u and v, and zero trace builds in the boundary condition.
  5. Define a(u,v) and L(v), and declare u a weak solution when a(u,v) = L(v) for all v in H^1_0.

Why this shape is exactly what we wanted

The point of bending the problem into 'a(u, v) = L(v) for all v' is not tidiness — it is that this shape is solvable by a theorem. The bilinear form for our model has two priceless properties. It is bounded: a(u, v) is no bigger than a constant times the H^1 sizes of u and v, because it is just an integral of a product. And it is coercive: a(u, u) is the integral of |grad u|^2, which by the Poincaré inequality is bounded below by a constant times the full H^1_0 norm of u. Coercivity is the abstract echo of an energy that cannot vanish unless u itself does.

These two words — bounded and coercive — are the precise hypotheses of the Lax-Milgram theorem, which guarantees that there exists exactly one u in H^1_0 with a(u, v) = L(v) for all v. That is the headline result of the very next guide, so we stop just short of proving it; but you can already see why the weak formulation was worth the trouble. It converted a PDE that might have no classical solution into an abstract equation in a complete space whose solvability is a single, checkable structural fact about a bilinear form.

Honest edges and the road ahead

Be honest about three things. First, the clean drop of the boundary term depended on v being zero on the boundary; a Neumann condition instead leaves the boundary term alive and folds the flux data into L(v), which is why the weak formulation of a Neumann problem looks different and carries a compatibility condition. Second, coercivity is not automatic — for a general elliptic operator you need genuine uniform ellipticity and control of the lower-order terms, and without coercivity Lax-Milgram simply does not apply, so existence can fail or uniqueness can break. The method is powerful, not universal.

Third, and most important to keep straight: a weak solution is not yet a classical one. Lax-Milgram will hand you a u in H^1_0 — a function with one square-integrable derivative and nothing more guaranteed. It need not be twice differentiable, and on a domain with a sharp corner it genuinely will not be. Whether that weak u is secretly smooth, so that -Laplacian u = f holds pointwise after all, is a separate question with its own machinery: elliptic regularity, the subject of guide 5. The weak formulation buys existence cheaply; regularity is the bill you pay afterward to climb back to the classical statement.

Step back and savour the reshaping. We began with a stern pointwise demand that real data can render unsatisfiable, and ended with a single balanced identity in a complete Hilbert space — solvable by one structural theorem, minimizable as an energy, and discretizable into a matrix. The next guide spends that setup, proving existence and uniqueness outright via Lax-Milgram. After that, regularity earns the smoothness back. The weak formulation is the hinge the whole rung turns on.