Varying a function on a domain
In one variable the Euler–Lagrange equation contains a single derivative . For a function on a domain, a perturbation can change its slope in each coordinate direction. Those derivatives combine into a divergence.
Definition 18.1.1
(Integral functional and compactly supported variations)
Let be a bounded open set, and let , , be . For a sufficiently smooth , consider
A compactly supported variation of is a family with and real. The function is a local minimizer with respect to these variations if, for each such , for all of both signs with sufficiently small.
These variations do not alter near the boundary, so boundary values play no role in the first result. We write for the partial derivative in and for the gradient in .
Theorem 18.1.2
(Euler–Lagrange equation in divergence form)
Assume is finite, , the composed vector field is on , and is a local minimizer with respect to every compactly supported variation. Then
Proof
(Differentiate under the integral, then apply the fundamental lemma)
Fix with support in a compact set . Outside the integrands of and agree, so the difference is an integral over . On , the mean value theorem and uniform continuity of on a compact set containing the values give uniform convergence of the difference quotient. This justifies differentiation under the integral. Because minimizes a differentiable function of on an open interval, its derivative vanishes there. Integration by parts, with no boundary term because vanishes near , gives
All derivatives of here are evaluated at . The function is continuous. The fundamental lemma states that a continuous with for every vanishes identically: if , a nonnegative bump supported where and positive at gives a positive integral. Hence (1) holds.
Proof X-ray
Why the boundary disappears
Each hypothesis has one job. The two signs of turn minimality into an equation; with one-sided variations, as under a constraint, only an inequality would follow. The regularity of , together with compact support, licenses differentiation under the integral. The regularity of the composed field is what allows its divergence to be formed and makes continuous, which the fundamental lemma needs.
The boundary disappears because the test function vanishes in a neighborhood of the boundary, not merely on it. Integration by parts therefore has no boundary contribution. Only stationarity was used, so (1) is necessary for the stated minimum and holds for maxima and saddles alike; it is not by itself sufficient. Allowing nonzero boundary variations is a different problem: it introduces boundary terms and leads to natural boundary conditions under suitable regularity assumptions.
The divergence form is easiest to read in a case with a variable coefficient.
Worked example 18.1.3
(A conductivity equation with a source)
Let be positive and , and take
Then and , so (1) reads , that is,
The coefficient stays inside the divergence. Expanding gives ; dropping the term , as if could be pulled out, would be wrong unless is constant. For and the equation reduces to Laplace's equation.
The Dirichlet principle
Take . Then and , so (1) becomes : stationary points are harmonic. The minimum statement needs boundary values and a class of competitors.
Theorem 18.1.4
(Dirichlet principle for a given harmonic function)
Let have a sufficiently smooth boundary, let be harmonic, and let have the same boundary values as . For ,
with equality only for . Thus is the unique minimizer in this classical competitor class.
Proof
(Green's identity and the exact energy gap)
Put , so on . Green's identity gives
Expanding therefore yields the exact identity (2). Equality forces everywhere by continuity. On each connected component is constant, and its zero boundary values force that constant to vanish. The argument starts with an existing harmonic function; it is not an existence proof for arbitrary boundary data.
Worked example 18.1.5
(Linear boundary data on the unit disk)
On the unit disk , prescribe boundary values . The affine function is harmonic with these values, and , so .
Test (2) on the competitor , which is smooth and equals on the circle. Here and . The cross term is by symmetry, and in polar coordinates
Hence . A direct computation agrees: integrates to over .
Common mistake
(Forgetting that the boundary values are fixed)
The zero function has Dirichlet energy on the unit disk, but it does not refute the previous example: it has boundary values , not . Every comparison in the Dirichlet principle is made inside the class of functions with the prescribed boundary values, and the vanishing boundary term in Green's identity uses exactly this.
Adding a potential
For
with , the Euler–Lagrange equation is .
Theorem 18.1.6
(Convex potentials make stationary points global minima)
Assume the smooth-domain hypotheses of the Dirichlet principle, let be convex, let solve , and let have the same boundary values. Then
Proof
(Convexity absorbs the linear term)
Put . Convexity gives pointwise. Green's identity, the zero boundary values of and the PDE give . Hence
The boundary condition removes the boundary term and convexity controls the potential; both are used.
Convexity cannot be dropped, as the following one-dimensional case shows.
Counterexample mode
An Euler–Lagrange equation alone does not give a minimum
Take , , zero boundary values and the concave potential . The zero function solves the Euler–Lagrange equation . For the competitor , which has the same boundary values,
for every . So zero is stationary but not even a local minimum. With a convex the theorem above rules this out; stationarity and minimality are different claims.
When the potential is quadratic, the convexity inequality is an identity and the energy gap can be computed exactly.
Worked example 18.1.7
(A quadratic potential in one dimension)
Take , , and boundary values , . The equation with these values has the solution , and
For the straight competitor ,
Since exactly, the gap equals with , here about , and it is at least as the theorem states. In one dimension Green's identity is just integration by parts with two boundary points.
Practice and solutions
Quick check
(Derive the Euler–Lagrange equation for .)
Compute and first.
Solution · Answer
Here and , so (1) gives , that is, . This is the case of the potential section; since this is convex, a classical solution with prescribed boundary values is the global minimizer among same-boundary competitors.
Quick check
(For , what is the energy difference between a harmonic and a competitor with the same boundary values?)
Expand .
Solution · Answer
It is exactly . The cross term vanishes by Green's identity, because and on .
Quick check
(Why does a compactly supported test function produce no boundary term in the derivation of (1)?)
Look at where is nonzero.
Solution · Answer
Its support is a compact subset of , so vanishes on a neighborhood of . Integration by parts can be carried out on a region containing the support whose boundary lies where , and every boundary integral is zero.
Quick check
(Does the Euler–Lagrange equation alone prove a minimum?)
Recall the concave potential on .
Solution · Answer
No. With on the zero function solves the Euler–Lagrange equation, yet . An additional structural argument, such as the Dirichlet energy identity or convexity of , is needed to conclude minimality.