Motivation
A soap film suggests that a surface with small area should have some kind of balance. The precise local condition is zero mean curvature. This condition is meaningful even where no physical film exists, and it does not by itself say that the surface has the least area among every possible spanning surface. Graphs give a setting in which we can calculate both the local balance and a precise comparison theorem.
The useful distinction is between an equation satisfied by one surface and an inequality comparing it with competitors. We first derive the equation from curvature, then derive it again from area. Convexity will supply the comparison, while a second derivative of area will explain the local cost of changing the height. The helicoid and Scherk graph show that this balance allows genuinely curved surfaces.
Definitions
Definition 2.5.1
(Minimal surface and fixed-boundary vertical variation)
A regular surface is called minimal when its mean curvature vanishes at every point. For a graph this is a local differential condition on its height function. A fixed-boundary vertical variation changes the height over a fixed planar domain without moving its boundary heights. These variations form the precise competitor class used below.
Basic differentiation, integration by parts and the fundamental-form formulas are the prerequisites; no existence theorem for a boundary-value problem is assumed.
Let be open and let . Parametrize its graph by . Write , , and . Since , the upward normal is . The graph is regular everywhere: the vertical component of this cross product is one.
The coefficients of the first and second fundamental forms are
Here is a coefficient of the second form, not the height function .
Theorem 2.5.2
(The minimal surface equation for graphs)
The mean curvature of the graph of with respect to the upward normal satisfies
Thus the graph is minimal exactly when
Proof
(Mean curvature and the divergence form)
Here and , so , with , gives the first expression for . For the second, differentiate both components:
and add. Since , is equivalent to (1).
The denominators are positive, so no solutions were lost by clearing them. Reversing changes the sign of but leaves equation (1) unchanged. If the gradient is small, its nonlinear terms are small relative to the displayed second derivatives, suggesting the Laplace equation as a linear approximation, not as an equivalent equation.
Counterexample mode
A harmonic height need not be minimal
Claim to test: since (1) linearizes to the Laplace equation, every harmonic has a minimal graph. Take . Then , but , , and , so the left side of (1) equals , which is not identically zero on any open set. The repair is to keep the nonlinear terms: writing (1) as , a harmonic is minimal exactly when the bracket vanishes as well.
First variation and fixed-boundary comparison
Let be a bounded domain with piecewise smooth boundary, and assume . The graph area is
For vanishing on , set . Differentiation under the integral and integration by parts yield
Concept lensAnalytic
Minimality as stationarity of area
In particular, every minimal graph is stationary for these vertical fixed-boundary variations. Conversely, stationarity for all smooth compactly supported forces the continuous divergence to vanish: if it had a nonzero value at a point, a nonnegative bump in a small neighborhood of constant sign would make the integral nonzero. Hence stationarity in this class is equivalent to (1).
Stationarity is a first-order condition. For graphs it nevertheless implies a comparison with every competitor having the same boundary values.
Theorem 2.5.3
(Area comparison among graphs)
Under the assumptions above, if satisfies (1) and has on , then . Equality holds only when on every connected component of that meets the boundary.
Proof
(Supporting planes of the area integrand)
Put for . Its Hessian satisfies
by Cauchy–Schwarz. Integrating the derivative of along the line segment shows the strict supporting-plane inequality
with equality exactly when . Apply it pointwise with and . The integral of the linear term vanishes by integration by parts, because on the boundary and satisfies (1). This proves the comparison. If equality holds, continuity and strictness imply everywhere in ; the difference is constant on each connected component, and its boundary value makes the constant zero.
Proof X-ray
Convexity in the gradient does the work
The key move is to view area as and to use convexity of in the gradient variable, not in the height. Cauchy–Schwarz gives a positive Hessian, hence a supporting plane at that lies below everywhere. Equation (1) is used exactly once: after integration by parts the linear term becomes , and the boundary condition removes the boundary term. Strictness of the supporting-plane inequality and continuity give in the equality case, and the boundary values fix the constant on each component meeting . Because the argument compares gradients pointwise over one domain, it says nothing about an immersed competitor that is not a graph over , or about a graph with different boundary values.
Second variation: the cost of a height change
The first variation has a zero linear term at a minimal graph. That leaves an important question: is the next term positive, negative, or zero? A positive quadratic term means that small changes in this particular direction cost area. We can calculate it directly, keeping the planar domain and boundary values fixed throughout. In the formulas below, integrals without an indicated measure use planar area measure on .
Theorem 2.5.4
(Positive second variation for graph variations)
Let be a bounded connected domain with piecewise smooth boundary, let , and let vanish on . For ,
If is not identically zero, then . When is minimal, as well, so every such nonzero vertical variation increases area for all sufficiently small nonzero .
Proof
(Differentiate the area integrand twice)
At each point put and . Differentiating once gives . A second differentiation at zero gives
The numerator is at least by Cauchy–Schwarz. Gradients are bounded on the compact closure, and the square-root denominator is at least one, so differentiating under the integral is justified by a uniform bound for in a small closed interval.
If the integral were zero, its continuous nonnegative integrand would vanish everywhere. Thus in . Connectedness would make constant, and its zero boundary value would force that constant to be zero. This proves strict positivity for a nonzero variation. For minimal , Taylor expansion in the single real parameter gives , which proves the final statement. The sufficiently small range can depend on the chosen .
Notice that the second-derivative formula itself does not require minimality. Minimality is used to cancel the first derivative. A nonstationary graph can have positive second derivative in a direction and still decrease initially because its nonzero linear term has the opposite sign. Also, this is a formula for vertical graph variations. It is not a statement of the general second-variation formula for arbitrary normal variations of an immersed surface.
A quantitative comparison when slopes are bounded
The same Hessian gives a useful strengthening of the earlier comparison. Suppose is minimal, on the boundary, and throughout . Every gradient on the segment joining them has norm at most by the triangle inequality. For , apply the one-variable Taylor formula with integral remainder to :
Indeed the remainder is the integral of the Hessian quadratic form with weight for ; the weight integrates to one half. After integration over , the linear term vanishes by the same boundary integration by parts as before. Therefore
This inequality measures area excess by a squared difference of gradients. The common slope bound is part of its hypothesis, not a universal constant that works for all graphs. In particular, two minimal graphs with the same boundary values must agree: apply the original comparison in both directions and then its equality case. This proves uniqueness if a solution exists. It constructs no solution and does not permit infinite boundary heights.
Worked example 2.5.5
(A bump above a tilted plane)
On the unit disk let and . The boundary heights remain unchanged for . Here is constant and . The second-variation integrand becomes
Polar coordinates give : for example, integrate over and . Consequently
The initial area is and the first derivative is zero because the plane is minimal and the variation vanishes on its boundary. For a horizontal plane the second derivative is , so the leading area excess is . This computes the local cost of the bump rather than merely asserting that the area becomes larger.
Worked examples
A helicoid as a local graph
Worked example 2.5.6
(A minimal graph with negative Gaussian curvature)
Fix and take on the half-plane . With ,
Substituting these derivatives into (1) gives zero: the two linear second-derivative terms cancel, as do the remaining terms. In polar coordinates with , the graph is with , a patch of a helicoid. This is a graph on the stated half-plane, not a globally single-valued graph over the punctured plane: increasing the angle by changes the height by .
Its Gaussian curvature is
The determinant formula follows from . Thus zero mean curvature does not imply a flat surface. The helicoid is also ruled: for fixed , varying traces a straight line, although the present graph patch contains only a portion of each ruling.
Scherk's minimal graph
Margin Note
Why try a sum of one-variable functions
For the mixed term disappears, and (1) separates as , so both sides are one constant. With constant one, and .
Worked example 2.5.7
(Separate variables and canceling nonlinear terms)
On , put
The cosines are positive on , so the function is smooth there. We have , , , and . The left side of (1) becomes
Hence this nonplanar graph is minimal. Its height diverges near portions of the boundary of the square. The fixed-boundary theorem applies on a smaller bounded smooth domain whose closure lies inside ; it does not license inserting these infinite boundary values into the preceding theorem.
Common mistakes
Common mistake
(Zero mean curvature is not zero Gaussian curvature)
Mean curvature is half the sum of the principal curvatures; Gaussian curvature is their product. Opposite nonzero principal curvatures cancel in the sum but give a negative product. The helicoid calculation is an explicit instance. One should not replace the word minimal by flat, even locally, without checking which curvature vanishes.
Common mistake
(A local graph does not remove a global branch problem)
The helicoid height uses one angular branch on a half-plane. Moving once around the missing axis changes the height, so the same expression does not define a single-valued height on the punctured plane. The surface itself can extend beyond the chosen graph patch. Regularity of a surface and the existence of one global graph representation are different questions.
Exercises
Quick check
(Prove that is minimal and determine its Gaussian curvature.)
Solution · Answer
All second derivatives vanish, so (1) holds and the determinant formula gives .
Quick check
(For with , determine when the graph is minimal.)
Solution · Answer
Equation (1) reduces to . Thus is affine on each connected interval and the graph is a plane patch. A curved cylindrical graph is not minimal.
Quick check
(On the unit disk compare the areas of and for .)
Solution · Answer
Both have zero boundary values and the plane is minimal. Since is not identically zero, the equality case above gives . No integration of the square root is needed.
Quick check
(What relation holds between the principal curvatures at a point of a minimal graph?)
Solution · Answer
Their sum is , so they are and for some real . Consequently . Equality at one point does not force the whole connected surface to be a plane; the statement is pointwise.
Quick check
(For the bump in the worked example, calculate when . Explain why the positivity of this derivative alone would not suffice without stationarity.)
Solution · Answer
Since , the formula gives . The coefficient of in the area expansion is therefore . A nonzero first derivative would dominate this quadratic term for sufficiently small of one sign. Here the fixed-boundary first variation is zero, so the positive second derivative does give a strict local increase along this variation.
Quick check
(On the unit disk, with and , find an explicit valid for both gradients and use (2) to give a positive numerical expression bounding the area excess from below.)
Solution · Answer
We have and , so take . Moreover . Equation (2) yields . This lower bound is valid for every nonzero , not just for the small-parameter Taylor regime.
Quick check
(What are the eigenvalues of parallel and perpendicular to a nonzero vector ? Explain why an unbounded slope prevents a slope-independent positive lower bound for this quadratic form.)
Solution · Answer
In the direction of , the Hessian has eigenvalue ; in a perpendicular direction it has eigenvalue . Substitute these directions into the displayed Hessian formula. Both eigenvalues are positive at every finite slope, but approach zero as tends to infinity. On slopes bounded by , the smaller one is at least , exactly the bound used in (2). At , both eigenvalues equal one.