Motivation
A second-variation argument compares a stationary curve with perturbations whose heights and slopes are both small. A strong minimum asks for more: a curve can stay close in height while developing very large slopes. To control those competitors, a useful strategy is to compare the full integrand with a total derivative, without discarding a Taylor remainder.
The total derivative integrates to endpoint data. If the comparison is exact on a candidate curve, those endpoints supply a lower bound that the candidate attains. Such a function is called a calibration here. It can certify a minimum even when the integrand is not jointly convex in height and slope. The construction also explains why a nonnegative slope excess along one curve is insufficient: the comparison must account for changes in height across a whole region.
We work with a scalar first-order functional and fixed endpoint abscissas. The rectangular-strip assumptions below let us construct the comparison potential explicitly. The construction is not a general existence theorem for such fields; the direct method is a separate approach to existence.
Definitions
Definition 8.1.1
(Comparison curves and a calibration)
Let , let be an open interval, and let be on an open set containing . Compare curves with values in and fixed heights , . Their cost is .
A potential is a calibration inequality on this strip if
for every , and . A candidate is calibrated when equality holds along its position and slope at every .
All slopes are included in the inequality. A bound valid only for slopes close to the candidate would not automatically control the high-slope curves in a strong neighborhood. Each individual competitor has finite cost because its position and slope form a compact curve inside the stated domain.
Concept lensAnalytic
The calibration inequality as a Hamilton–Jacobi inequality
Put , allowing the value . Requiring the inequality for every slope is then exactly
and a calibrated candidate is one whose slope attains the supremum with equality. For the field potential constructed below, nonnegative excess says that attains the supremum when , so : the potential solves the Hamilton–Jacobi equation .
Definition 8.1.2
(Slope field, boundary coefficients and excess)
Let be on a neighborhood of the strip. Define
The field compatibility condition is , where these are partial derivatives at fixed height or position. At a comparison slope , the Weierstrass excess is
It compares slopes at the same height. The field supplies such a comparison throughout the strip, rather than only on one reference graph.
The expression defines the field's Hilbert integral along a path. We will prove that compatibility makes it an exact differential on the rectangle, so the integral depends only on endpoints. No path-independence claim for arbitrary domains with holes is needed.
Theorem/Proposition
Theorem 8.1.3
(A calibrated candidate minimizes the full functional)
If satisfies the calibration inequality and a comparison curve is calibrated, then every comparison curve in the strip satisfies
Thus is a global minimum within this strip. It is also a strong local minimum among all fixed-endpoint curves: its compact image has a positive uniform neighborhood contained in , so sufficiently small height perturbations remain in the comparison region without any slope restriction.
Theorem 8.1.4
(A compatible field produces the comparison potential)
Assume is as above, on the strip, and for every allowed position, height and real slope. Then there exists a potential with , , and every comparison curve has the exact identity
Any curve satisfying the endpoint data is therefore a minimum. If the excess is strictly positive whenever , that candidate is the unique minimum in the strip, hence a strict strong local minimum.
Theorem 8.1.5
(Velocity convexity supplies the excess inequality)
For fixed , if for every real , then . If everywhere, the inequality is strict for . More quantitatively, gives .
These assumptions concern the entire comparison segment of slopes, not only the reference slope. No convexity in the height variable is required by the compatible-field theorem.
Proof sketch or proof idea
Proof
(Integrate the calibration inequality)
For the first theorem, the ordinary chain rule gives
Integrate the pointwise inequality and use the common endpoints. A calibrated candidate attains this boundary value exactly. The last strong-local conclusion uses compactness only to keep nearby heights in the strip; it does not suppose their derivatives are small.
The second theorem needs a potential, and the proof builds one explicitly.
Proof
(Construct the Hilbert potential and compare all competitors)
Choose and put
The paths used in this formula stay inside the rectangle. We have , and differentiation in , compatibility and the fundamental theorem give
By the definition of excess, . Substitute a comparison curve and integrate its total derivative to obtain the exact identity. If , its excess vanishes, so it attains the lower bound.
For strictness, equality of the integrals makes the continuous nonnegative excess vanish everywhere. The strict-excess assumption gives . Two such curves with the same left value must coincide. To see this directly, bound by on the compact height interval containing both graphs. Their difference satisfies . With , we have and ; multiplying by shows . Hence , proving uniqueness.
Why is this an extremal field? Differentiating the definitions at fixed gives . Thus compatibility is equivalent to . Along any integral curve that exists in the strip,
That is Euler–Lagrange. The field is therefore a compatible family of stationary directions across nearby heights. The reverse practical task—constructing a nonsingular field around an arbitrary extremal—is additional work, not a consequence we assume without proof.
Proof
(Taylor's formula with integral remainder)
Velocity convexity follows from the one-variable Taylor formula with integral remainder:
This proves all three excess claims. A positive second derivative at one slope alone would not justify the same inequality for arbitrarily distant slopes.
Proof X-ray
The two jobs in a strong comparison
The excess inequality compares arbitrary slopes at one height. Compatibility makes those height-by-height tangent expressions assemble into one endpoint potential. Both jobs are needed: without compatibility, the integral of can depend on the whole competitor. Once it is an exact derivative, the full nonnegative excess remains, rather than a small-derivative Taylor approximation.
Worked examples
Worked example 8.1.6
(An exponential field for a quadratic height cost)
Let , choose , and work on any finite interval. Then , and . Take . Directly,
For endpoint data , , the curve is calibrated. For every fixed-endpoint competitor,
It is the unique global minimum. The boundary compatibility is essential to attainment: when the final height is different, the same inequality remains true but this first-order field supplies no calibrated candidate with those endpoints. A lower bound need not be sharp for every boundary-value problem.
Worked example 8.1.7
(A nonconvex double-well potential with a global certificate)
Take
Here , , so compatibility holds, and
Along a curve, the first term is . On , with and endpoint heights in their increasing order, satisfies . It is therefore the unique global minimizer among all real-valued competitors with those endpoints.
Writing , its value is , approaching as . This last limit is a calculation of finite-interval values, not an unproved whole-line existence theorem. The height potential has second derivative , negative near zero, so joint convexity fails along part of the candidate. The calibration proves a minimum without that stronger convexity assumption.
Worked example 8.1.8
(A position-dependent field and potential)
On , let
Then , , and . A potential is , and the exact identity is
The field curves are . For heights , , this curve uniquely minimizes the action, with value . The family fills each vertical line without intersections, and each curve satisfies . The term is essential here; retaining only would give the wrong comparison.
Common mistakes
Counterexample mode
Nonnegative excess alone can coexist with a saddle
For , choose . Its excess is at every height and slope. But , , so and : compatibility fails away from zero. On with zero endpoint heights, the zero curve is stationary, yet gives . Thus a slope comparison alone does not prove even a weak minimum. There is no endpoint-only potential whose derivatives are these incompatible .
Common mistake
(Treating a certificate as an existence proof)
A certificate is not an existence method unless a curve attaining it has also been constructed with the required endpoints. Nor does failure to find one particular field prove that the problem has no minimum. The exact comparison is conditional on a valid field and a matching integral curve; it should not be turned into an unsupported necessary criterion.
Common mistake
(Convexity at the field slope is not enough)
The velocity-convexity theorem needs along the whole segment of slopes between and , not only at the field slope . For and the constant field , one has . Yet at the excess is , because . Such a field gives no calibration inequality.
Summary
A calibration bounds the full action by endpoint data. A compatible slope field constructs its potential explicitly on a rectangular strip, while velocity convexity supplies nonnegative Weierstrass excess for all slopes. The result controls competitors close only in height and can certify nonconvex height problems. Endpoint compatibility, field compatibility and excess positivity perform different tasks and must each be checked.
Exercises
Quick check
(Minimize , where and , with and . Choose a calibrated field that fits arbitrary endpoint heights, and find the exact minimum.)
Try for a constant determined by the endpoints.
Solution
Put and . For , we have and , so compatibility holds. Take
The excess is , and the calibrated curve is . It has the required endpoints and is unique. Its minimum is . Using only would attain its zero lower bound only when the endpoint mismatch vanished.
Quick check
(For the double-well functional on , reverse the endpoint heights to and . Find a sharp calibration and prove the unique minimum.)
The sign of the slope field can change even though the energy has the same formula.
Solution
Use and . Then . The curve solves this first-order equation and fits the reversed endpoints. With , the boundary potential gives the same minimum . Strict excess and the proved initial-value uniqueness argument give the unique global minimizer. The increasing field gives a valid but unsharp lower bound for these data.
Quick check
(For with and , the potential gives the lower bound . Is it attained? Find the actual minimum and explain the difference.)
Solve the classical equation and evaluate the action using along its solution.
Solution
Equality for would require , and the left height zero forces , incompatible with the final height one. The Euler–Lagrange solution is . Its action is . For any endpoint-zero perturbation , integration by parts gives the exact gap , proving a unique global minimum. No valid inequality failed; this particular calibration did not have an attaining curve for the chosen endpoints.