Motivation
An explicit Green function is valuable when the geometry permits one, but existence should not depend on discovering such a formula. Nor should a discontinuous source automatically exclude a solution. A second derivative may fail to be continuous even when the function and its first derivative have a perfectly meaningful energy.
The weak formulation moves one derivative onto a test function. Its natural unknown is a function with square-integrable first derivatives, and its natural boundary condition is a closure property of that function space. This changes what we mean by a solution, so we must specify the space before using integration by parts or minimizing an energy.
The resulting existence proof has a concrete structure. A lower energy bound gives a finite infimum; a midpoint identity makes a minimizing sequence Cauchy; completeness supplies a minimizer; and arbitrary variations give the equation. The same geometry explains finite-dimensional approximation. None of these steps alone proves that the solution has continuous second derivatives or pointwise boundary values.
Definitions
Let be nonempty, bounded and open, and let all functions be real-valued. Choose a box of side length containing . Integrals and equality of Sobolev functions are understood up to sets of volume measure zero. We assume the usual completeness of .
A function has weak derivative if
Write and . The space consists of these functions, with squared norm . Weak derivatives are unique: a locally integrable difference with zero integral against every smooth compactly supported test is zero almost everywhere, as follows by mollification.
Definition 5.2.1
(The zero-boundary energy space and weak Poisson equation)
Define as the closure of in . Set
For , a zero-Dirichlet weak solution of is satisfying
Membership in the closure encodes the boundary condition here. We do not define it by arbitrarily assigning zero values on the boundary, nor assume a trace theorem for every bounded open set. For smooth functions, weak and ordinary derivatives agree by integration by parts against compactly supported tests.
Theorem/Proposition
Theorem 5.2.2
(Poincaré turns gradient energy into a complete norm)
For every ,
Consequently is an inner product on this space, its energy norm is equivalent to the inherited norm, and the space is complete in energy norm. The box constant is sufficient, not asserted optimal. No boundary smoothness or connectedness is needed.
Theorem 5.2.3
(Existence, minimization and stability for square-integrable sources)
For every there is a unique weak solution. It is exactly the unique minimizer of over . For every competitor ,
The estimates are
For two sources, the same bounds apply to the solution difference with the source difference on the right. Existence here is in the specified energy space; no classical regularity is concluded.
Theorem 5.2.4
(Galerkin approximation is an energy projection)
Let be a finite-dimensional subspace of . There is a unique satisfying the weak equation for every test in . It obeys
Also . If the spaces are nested and their union is dense in energy norm, their approximations converge to in energy norm and hence in . No mesh-dependent rate follows without a further approximation estimate.
Proof sketch or proof idea
Proof
(Completeness and the Poincaré inequality)
First establish the space facts instead of presuming completeness of the gradient norm. If a sequence is Cauchy in , its functions and weak derivatives have limits. Passing to the limit in the test identity shows that the derivative limits are weak derivatives of the function limit. Thus is complete, and its closed subspace is complete in the inherited norm.
For a smooth compactly supported function, extend by zero to the containing box. On each line parallel to the first coordinate axis, integrate its first derivative from the left box face, where the extension is zero. Cauchy–Schwarz bounds the squared value by the box length times the line integral of the squared derivative. Integrating once more along that axis and then over the other coordinates gives
For dimension one there are no other coordinates to integrate. The zero extension is smooth because the original support is compactly inside the open set. Passing to the defining closure proves the estimate for all zero-boundary functions. It makes the gradient norm nondegenerate, and
Norm equivalence transfers completeness and proves the first theorem.
Proof
(Existence through the midpoint identity)
For existence, write . Cauchy–Schwarz and Poincaré give
The infimum is finite. Choose a minimizing sequence whose energies tend to it. The quadratic lower bound also bounds their energy norms. More decisively, expansion gives the exact midpoint identity
The left side is at least the infimum. It follows that
Completeness gives a limit in the zero-boundary space. Both the quadratic energy and the source functional are continuous in energy norm, so the limit attains the infimum. Boundedness alone would not have given this strong convergence; the midpoint identity supplies it.
Proof X-ray
The missing existence step is convergence, not formal variation
Differentiating an energy at an assumed minimizer can only find its necessary equation. It does not produce a minimizer. Here the lower bound creates a finite target value, the midpoint identity controls distances between approximate minimizers, and completeness produces the limit. Only after that construction do variations identify the PDE.
Once a minimizer exists, a one-parameter expansion identifies its equation.
Proof
(Variations give the weak equation and stability)
For any direction in the space and any real scalar,
A minimum for both signs of sufficiently small scalar forces the linear coefficient to vanish. This is the weak equation. Conversely, expanding the energy of any competitor about a weak solution gives the exact energy-gap identity in the theorem. It proves global minimization and uniqueness. Testing with the solution itself and applying the source bound proves its energy estimate, with the zero solution handled directly if division by its norm would be needed. Poincaré gives the second estimate. Subtraction proves stability.
A classical solution that also belongs to the zero-boundary energy space satisfies the weak equation first for compactly supported smooth tests and then for all energy tests by density. Conversely, if a weak solution is in the interior and the source is continuous there, the test identity forces the classical equation pointwise. These implications do not supply that extra smoothness or boundary continuity.
Concept lensStructural
The boundary condition is a constraint on admissible variations
The energy contains the source and first derivatives. The fixed boundary enters through the space in which competitors and variations are allowed. Changing that space changes the problem, even if the written interior energy is unchanged. On a rough domain, its closure definition is the statement we have proved, without an unstated pointwise trace interpretation.
Proof
(Galerkin orthogonality and best approximation)
For the finite-dimensional problem, choose a basis . Its coefficient matrix has entries and its load entries are the corresponding source integrals. The quadratic form of any nonzero coefficient vector is the strictly positive energy of its basis combination, so the symmetric matrix is positive definite and invertible. Subtracting its test equations from the exact equation gives orthogonality. For every ,
This proves best approximation. Density and nesting then make the best error tend to zero: choose one close competitor and retain it in all larger spaces. The energy-gap formula follows from the already proved minimization identity.
Worked examples
Worked example 5.2.5
(A constant source on the unit interval)
For and , take . It has zero endpoints and . To prove it belongs to , choose smooth cutoffs that vanish near the endpoints, equal one outside endpoint strips of width , and have derivative bounded by a constant times . In those strips the function is bounded by a constant times and its derivative is bounded. Hence
The function error also tends to zero in , so the compactly supported smooth products converge in . Integration by parts and density now identify the unique weak solution. Direct calculation gives and . The energy-gap identity compares it with every energy competitor, not merely smooth ones.
Worked example 5.2.6
(A discontinuous source with a nonclassical second derivative)
On the unit interval let equal one to the left of the midpoint and zero to its right. The solution is
Values and first derivatives agree at the midpoint; the second derivative jumps. Cutoffs as above, followed by smooth mollification of the compactly supported functions, prove energy-space membership. Integrating by parts on the two halves produces no interface term because the first derivatives match. The identity holds first for smooth compact tests and then for every energy test by density and continuity of both integrals. Thus uniqueness verifies the weak solution. Its squared energy norm is . This solution is not at the midpoint, although its weak first derivative and energy are well defined. The source's value at the single midpoint does not affect the weak equation.
Worked example 5.2.7
(One basis function gives a best approximation, not an exact solution)
Let on the unit interval. The exact solution is . Choose spanned by . Then
The exact squared energy norm is and the projected squared energy norm is , so orthogonality gives squared error and energy-functional gap . The residual source is not zero pointwise. Its integral against the chosen test direction is zero, which is exactly the finite-dimensional equation.
Counterexample mode
A single successful test is not the weak equation
In the preceding example, the approximate solution has , not the source . Define the load residual as . Testing with happens to give equality. Testing the residual with instead gives the positive integral of , equal to . This is an admissible zero-boundary test. Galerkin orthogonality in a proper subspace does not imply the equation against every energy test.
The existence theorem also depends on the admissible class, as the next computation shows.
Counterexample mode
Removing the boundary constraint can destroy the minimum
Minimize the same functional over all of with source one. Constant competitors have zero gradient and energy , which tends to negative infinity. There is no minimizer. Equivalently, the proposed weak equation tested with the constant one would require zero to equal the positive volume. The zero-boundary constraint was essential to coercivity; replacing it changes the compatibility and uniqueness questions.
Common mistakes
Common mistake
(Assigning pointwise values to energy-space functions)
Do not identify a Sobolev equivalence class with arbitrary values assigned on a boundary or at isolated points. The present boundary condition is the closure defining the zero-boundary space. A general trace theorem and boundary regularity require additional work.
Common mistake
(Ignoring the interface term of a derivative jump)
An almost-everywhere ordinary second derivative does not capture every distributional second derivative: jumps in a first derivative create interface terms.
Common mistake
(Treating a bounded minimizing sequence as convergent)
A coercive bound only bounds a minimizing sequence. The midpoint argument, rather than a bare boundedness claim, establishes convergence here. No compact embedding or unproved classical regularity is hidden in the proof.
Summary
Poincaré makes first-derivative energy a complete norm on the zero-boundary space. The quadratic midpoint identity constructs an energy minimizer, and arbitrary variations identify the unique weak Poisson solution. The same inner-product geometry gives stability, an exact energy gap and Galerkin best approximation. These are existence and approximation statements in an explicit function space, with classical regularity kept as a separate question.
Exercises
Quick check
(On , solve the zero-boundary problem with constant source . Compute the squared gradient norm and minimum functional value.)
Rescale the unit-interval quadratic solution.
Solution
The solution is . Its derivative is , so
Endpoint cutoffs justify energy-space membership, and the theorem supplies global minimality and uniqueness.
Quick check
(On , let . Compute the integral of for a smooth compactly supported test . Why is not weakly harmonic despite having ordinary second derivative zero away from the midpoint?)
Integrate on the two halves and retain the interface term.
Solution
The first derivative is one on the left and minus one on the right. Thus
This is not zero for all tests. The distributional source is twice a point evaluation at the midpoint, often written as a Dirac point source. It is not an source, so it lies outside the source class of the existence theorem, but its test identity makes the missed interface contribution explicit.
Quick check
(For nested Galerkin spaces , prove an exact relation among the squared errors of their approximations and their difference. Deduce monotonicity of the error and functional values.)
Use energy orthogonality against the larger space.
Solution
The difference belongs to the larger space, so
Therefore the error cannot increase. The functional gap is half that squared error, so the approximate functional values decrease toward the exact minimum whenever the nested union is dense. Nesting alone gives monotonicity, not convergence to zero error without density.
Quick check
(On , prescribe endpoint values and with source one. Use an affine lift to find the solution and describe its admissible competitor class.)
Keep variations zero at the endpoints while the unknown lies in an affine translate.
Solution
Take . The solution is , and its competitors lie in . On this affine class explicitly define
The affine lift has constant derivative whose integral against a zero-boundary test derivative is zero, first for compact tests and then by density. Thus the correction solves the same homogeneous-boundary weak equation. Variations stay in the zero-boundary space, and expansion about this solution gives the same squared-gradient energy gap for this functional on the affine class. This example uses an explicit lift, not a claim that arbitrary rough boundary data always admit one.