The area problem is nonlinear even for a scalar height
A graph over a planar domain has more area when it tilts. The local area factor is not a quadratic function of its slope: it is the square root of one plus the squared slope. This difference changes the stationary equation. A harmonic height need not be stationary for area, although the Dirichlet energy remains a useful small-slope approximation.
There is also a favorable feature. The area factor is strictly convex in the gradient, and its first derivative is bounded. Together these facts allow a global comparison with Sobolev competitors, even when those competitors are not classically smooth. We will prove that comparison directly, rather than infer a minimum solely from the nonlinear PDE.
Fixing the boundary without assuming pointwise boundary values
Let be a bounded connected Lipschitz domain. For a scalar height , write
For smooth , this is the surface area of its graph. We use the same integral for . It is finite because and the domain has finite measure.
Definition 18.2.1
(The fixed Sobolev boundary class)
Fix and set
Here is the closure of in the norm. This closure condition is our precise zero-boundary convention for differences of competitors. The proof below uses it explicitly; it does not assign pointwise boundary values to every Sobolev representative. A minimizer means a minimum of the graph-area integral over this entire class.
The statement concerns graphs. It does not yet compare arbitrary parametrized surfaces, immersed surfaces, changes of topology or generalized surfaces.
The bounded flux gives the first variation
Define the gradient of the density by
Theorem 18.2.2
(A first variation valid for directions)
For and , the function is differentiable at zero, with
Every minimizer therefore satisfies
This last identity is the weak minimal-graph equation.
Proof
(Domination comes from the density, not a bound on the competitor's slope)
The bound on makes globally one-Lipschitz. For nonzero ,
The quotient converges pointwise almost everywhere to . Its displayed bound is integrable, so dominated convergence proves the formula. Both signs of remain in the affine admissible class. A minimum consequently has zero first derivative, in particular for compactly supported smooth directions.
If is also inside the domain, integration by parts against compact tests and the fundamental lemma identify the pointwise equation
Writing the coordinates as , its equivalent numerator is
The denominator removed from the divergence is . Thus this is not generally the Laplace equation.
Common mistake
(A minimal graph is not a harmonic graph)
Dropping the factor inside the divergence produces Laplace's equation, which is only the linearization at zero slope. The Scherk graph below solves the minimal-graph equation while off the diagonals . Conversely, a harmonic height need not be area-stationary, as the last counterexample of this page shows.
Weak stationarity is a global graph certificate
For gradient vectors , define the excess above the tangent plane by
The density has Hessian
Cauchy–Schwarz proves the inequality. In particular the Hessian is positive on every nonzero direction, so is strictly convex, , and equality holds exactly when .
Theorem 18.2.3
(Weak solutions and unique global graph minima coincide)
For a given , the following are equivalent: minimizes over ; and satisfies the weak equation for every compactly supported smooth test. Whenever such a function exists, it is unique up to equality almost everywhere. More precisely, every satisfies
when is a weak solution. This is not an existence theorem for arbitrary boundary data.
Proof
(Extend the weak test, then subtract the supporting tangent)
Necessity was proved by the first variation. For sufficiency put and choose converging to it in . The bounded flux gives
Hence the linear term against also vanishes. Subtracting that term from the finite energy difference gives the stated exact excess identity. Strict convexity shows that equality forces almost everywhere.
The zero-boundary condition then forces , not an arbitrary constant. One can see the needed Poincaré estimate directly: enclose the domain in , extend each compactly supported by zero, integrate its first-coordinate derivative along horizontal lines, and then integrate over the box. This gives . Passing to the limit gives the same bound for . Thus zero gradient implies zero function. The excess identity proves uniqueness of any two weak solutions with the same boundary class.
Proof X-ray
Why a compact test identity controls every admissible graph
The weak PDE initially knows only compactly supported smooth tests. The zero-boundary closure supplies their approximations to the difference of two competitors, and bounded flux permits passage to that difference in the integral. Strict convexity then turns the canceled linear term into a nonnegative global gap. Neither a classical boundary integral nor a bound on the competitor's gradient is needed. Existence of a weak solution remains a separate question.
Quantitative control needs a slope bound
Theorem 18.2.4
(A quadratic gap on bounded-gradient classes)
Suppose is a weak solution. If and both gradient norms are at most almost everywhere, then
For , the second variation is
Proof
(Taylor's formula along the segment of gradients)
The straight segment between two vectors of norm at most stays in that ball. Taylor's formula with integral remainder and the Hessian lower bound therefore give the lower excess estimate, with factor . The Hessian upper bound gives the upper estimate. Integrate and use the exact excess identity. The second-variation formula follows by differentiation with the integrable bound ; its Hessian is evaluated at the reference gradient.
Every finite slope has positive Hessian, which proves a strict global minimum when the boundary class fixes constants. Along increasingly large slopes, the smallest Hessian eigenvalue tends to zero. A common slope bound restores a uniform coefficient in the quadratic estimate. Removing that hypothesis while keeping the same coefficient would make a stronger, false assertion.
The slope hypothesis cannot simply be deleted: a single bump, scaled up, already defeats every uniform quadratic constant.
Counterexample mode
A unique global minimum without a uniform quadratic gap
For zero boundary data, is the unique area-minimizing graph. Choose a nonzero and let , . Since for ,
Their ratio tends to zero. No positive constant can bound the area gap below by that squared derivative norm for every smooth fixed-boundary competitor. This does not contradict the preceding bounded-gradient theorem: the gradients in this family are unbounded as the parameter grows.
Explicit nonlinear minimizers and a misleading harmonic candidate
Worked example 18.2.5
(An affine graph gives an exact area minimum)
Let , where . Its flux is constant, so its integral against a compact test gradient vanishes. The comparison theorem proves it is the unique minimum in , with
This assertion includes rough fixed-boundary graph competitors, not only smooth perturbations. The affine example has a simple geometry, but the proof's boundary cancellation is the same one used for curved graphs.
Worked example 18.2.6
(A Scherk graph is a curved global minimizer on a rectangle)
Choose and . Define
The function is smooth on a neighborhood of the closed rectangle. Its derivatives are , , , and . Hence the minimal-graph numerator is
It is therefore the unique global area minimum in its own fixed boundary class. Its gradient is bounded by , so the quantitative theorem also applies to competitors with that common bound. In general is not zero: a nonlinear minimal graph need not be harmonic. The restriction on the rectangle keeps the poles outside the closed domain; no finite-boundary-data theorem at a pole is asserted.
Worked example 18.2.7
(Adding a constant shows what the boundary class excludes)
Let be a weak solution and let be a constant. Then , so , and satisfies the same weak equation. Uniqueness is not contradicted, because . Indeed would have to lie in , and the Poincaré bound from the comparison proof would give .
For the linear term vanishes trivially and the excess is zero. If constants were admissible, uniqueness would therefore fail exactly by additive constants. On a smooth domain one would say that has boundary values shifted by ; the closure definition of records the same fact without pointwise boundary values.
Counterexample mode
A Dirichlet minimizer need not minimize graph area
On the unit disk take and prescribe its boundary class. This function is harmonic and minimizes Dirichlet energy, but its area flux has divergence
which is not identically zero. Choose a nonnegative smooth compactly supported that is positive somewhere , and set . Then is a smooth compact test and
For sufficiently small positive the area decreases while the boundary data stay fixed. The two minimization principles concern different integrands. A small-slope approximation can relate their energies without making their exact stationary equations identical.
Practice with solutions
Quick check
(For a radial height , derive the radial minimal-graph equation and prove that a radial solution on a full disk is constant. On an annulus , construct the solution with flux constant and inner boundary value zero, compute its area, and state its proved minimum property.)
This explains why a radial minimal graph can bend on an annulus but not smoothly through a disk center. Use the radial form of the divergence.
Solution · Guided solution
For positive radius, the radial divergence formula gives
Thus the expression in parentheses is a constant . For a solution smooth through the center it tends to zero as , so and . On the annulus the center is absent. For the specified positive constant,
The strict bound on makes the function smooth up to both boundary circles. Using polar coordinates and gives
The weak comparison theorem proves the unique minimum among graphs in this solution's own fixed Sobolev boundary class. It does not claim that every pair of annular boundary heights arises from a parameter in this range, or compare all nongraph surfaces spanning the circles.
Quick check
(For , find an exact expression and a bound for . Deduce a controlled area expansion for with , and explain what it does and does not say about the stationary equation.)
This quantifies the small-slope Dirichlet approximation. Rationalize the square-root difference.
Solution · Guided solution
Rationalizing the square-root difference gives
Set and integrate. The fourth-power assumption makes the error integrable, and
The Dirichlet energy is the leading quadratic correction for this scaled family. This is an energy approximation with a quantified remainder, not an exact replacement of the minimal-graph PDE by Laplace's equation at nonzero amplitude. The harmonic counterexample shows why that distinction matters.
Quick check
(Let be a constant symmetric positive definite matrix and replace the density by . Derive its flux and Hessian, prove the corresponding weak global graph comparison, find the minimum for affine boundary data, and evaluate the density for and the affine height .)
Changing the slope metric changes the flux. Use Cauchy–Schwarz in the inner product defined by .
Solution · Guided solution
Differentiation gives
Cauchy–Schwarz in the inner product bounds the numerator below by for nonzero . The flux is bounded by , since . These facts give the same dominated first variation, extension from compact tests, and strict excess comparison as above, now with weak equation . The argument is conditional on an admissible weak solution and makes no general existence claim.
An affine height has constant flux, so it is the unique minimum in its own boundary class, of value . For the stated matrix and height, and the density is , so the minimum is . The correct equation comes from this density's flux, not from retaining the Euclidean formula after changing the objective.
The limits of the graph theorem
We have proved a weak comparison and uniqueness theorem for an existing graph solution, explicit curved and affine examples, slope-dependent quantitative estimates, and the stated constant-metric extension. Existence and boundary regularity for general prescribed data, vertical limiting gradients, BV relaxation, nonparametric-to-parametric comparison, unrestricted surface topology and free-boundary minimal surfaces remain outside these arguments.