Motivation
Solving Euler–Lagrange produces a stationary curve. The next task is to determine whether nearby curves increase the functional. Ordinary calculus suggests examining a second derivative, but a functional has infinitely many directions. A positive coefficient of in the integrand does not by itself control negative contributions depending on . The interval length and endpoint restrictions also matter.
There are two complementary ways to prove a minimum. A local argument expands the functional to second order and controls the remainder uniformly over small perturbations. A global argument compares the integrand at two curves using convexity. The first can work without global convexity; the second can prove more than a local result when its stronger hypotheses hold. We keep the two kinds of conclusion separate.
This section treats fixed-endpoint scalar problems on , with . The integrand is on an open neighborhood of the relevant triples, and the stationary reference curve is . Perturbations satisfy . Where a derivative of a coefficient is used, its additional regularity is stated explicitly.
Definitions
Definition 3.1.1
(Second variation and its quadratic form)
Along put
The second variation is
If , integration by parts gives , where . We call coercive in the endpoint-zero class if some satisfies for every such .
Coercivity supplies a uniform quantitative bound. Merely knowing separately for each nonzero does not give that bound automatically in an infinite-dimensional space. The word “positive” must therefore be interpreted carefully when it is used in a sufficient condition. The derivative norm is adequate here because zero endpoint values control the size of itself.
Why positive curvature in every direction can still fail
Consider endpoint-zero curves in and
The zero curve is stationary, and its second variation is . This is strictly positive for every nonzero admissible : continuity makes nonzero on an interval. Yet these separate positive values do not supply a uniform derivative bound. For ,
Thus no can satisfy the coercivity inequality for all directions. The normalization of is irrelevant to that ratio, since both terms are quadratic.
In fact zero is not even a weak local minimum. Put . Both and tend to zero, but direct integration gives
Proof X-ray
Where the finite-dimensional analogy breaks
For each fixed direction the positive quadratic term eventually dominates the fourth-order term as its amplitude tends to zero. The required amplitude, however, can depend on the direction. The oscillating sequence changes direction while approaching zero, so no single neighborhood is justified by those separate tests. Coercivity supplies the uniform bound that the Taylor-remainder argument needs.
Removing the negative fourth-power term gives , which has the same second variation at zero and a unique global minimum there. Thus this quadratic form alone decides neither conclusion. The example does not say that every noncoercive second variation fails to give a minimum; it shows why a proof needs more than positivity direction by direction. Here is zero along the reference curve, so this example does not contradict results that assume a strictly positive strengthened Legendre coefficient.
Concept lensAnalytic
Coercivity as a positive Rayleigh quotient
When and is continuous, coercivity is equivalent to the weaker-looking condition
taken over endpoint-zero . One direction is . Conversely, for ,
and a small makes the total coefficient of nonnegative. The example above has : its quotient is , yet coercivity fails.
Definition 3.1.2
(Weierstrass excess function)
At a fixed and reference slope , define
It measures how far the integrand at a comparison slope lies above its tangent line at . If is convex, then for all slopes in its convex domain.
Nonnegativity of this excess along one extremal is a slope comparison at the same height . A competitor usually has a different height too. For a general height-dependent integrand, an excess inequality alone does not account for that change. A field of extremals or another comparison argument is needed for the classical general strong sufficiency theorem; we do not assume such a theorem without constructing its hypotheses.
Theorem/Proposition
Theorem 3.1.3
(Necessary second variation and Legendre condition)
If is a weak local minimum, then for every endpoint-zero direction . In particular,
The latter is the scalar Legendre necessary condition. Neither it nor nonnegative second variation alone is a sufficient minimum test.
Theorem 3.1.4
(Coercive second variation gives a strict weak minimum)
Suppose is stationary and for some and every endpoint-zero . Then is a strict weak local minimum: for all sufficiently small nonzero in ,
The integrand hypothesis is understood on a neighborhood of the compact reference curve of triples, so its Hessian is uniformly continuous on a smaller compact neighborhood.
Theorem 3.1.5
(A positive Jacobi solution supplies coercivity)
Assume , , is continuous, and on . If there exists a function on the entire closed interval satisfying
then is coercive in the endpoint-zero class. In particular the preceding sufficient condition applies. Positivity on the closed interval is part of this statement; a solution vanishing at an endpoint cannot simply be substituted into this proof.
Theorem 3.1.6
(Joint convexity gives a global comparison)
Suppose that for every fixed , the function is differentiable and convex on a convex domain containing the triples of and every competitor being compared. If satisfies Euler–Lagrange, then it is a global minimizer among those fixed-endpoint competitors. If the integrand is strictly convex in for each , the minimizer is unique.
Proof sketch or proof idea
Proof
(Localize the second variation with a bump)
At a minimum, has nonnegative second derivative, giving . To obtain Legendre, suppose at an interior point. Choose a nonzero smooth bump supported in and put . For small , the leading term is
The mixed term is and the height term . Hence , a contradiction. Continuity handles the endpoints. Each bump is an admissible direction; multiplying it by a sufficiently small scalar makes it an arbitrarily small perturbation.
The sufficient conditions need a uniform estimate rather than a sign in each direction.
Proof
(Uniform remainder control and the Jacobi identity)
For endpoint-zero , Cauchy–Schwarz gives , and hence . Taylor expansion of at the reference triples, uniformly on their compact neighborhood, yields
where as . The linear part is zero by stationarity. Choose the neighborhood so that . Coercivity gives the desired lower bound.
For the Jacobi criterion set . Expanding the square and using gives
The endpoint term vanishes because and stays positive. Thus . Let , , , and . Since has zero endpoints,
These inequalities give an explicit positive coercivity constant. This proof uses a nonsingular positive solution rather than silently dividing by a Jacobi solution at one of its zeros.
Proof X-ray
The substitution that turns the second variation into a square
The decisive move is writing . The cross terms in are absorbed by the exact derivative, and the equation is exactly what makes the remaining terms cancel, leaving . Positivity of on the closed interval is used three times: to divide by , to make the boundary term vanish, and to give in the final bound; then turns the square into coercivity. If vanished at an endpoint, the boundary term would become a limit to be checked, as in the Poincaré derivation below.
Proof
(Integrate the supporting-plane inequality)
For convexity, the supporting-plane inequality gives
After integration the right side is zero by Euler–Lagrange and the endpoint conditions. If strict convexity holds and is nonzero, continuity gives an interval where the pair is nonzero, so the integral inequality is strict. This proves a global result directly and does not require a small norm.
Worked examples
Worked example 3.1.7
(Interval length can change the answer)
Consider with zero endpoints. The zero curve is stationary and for every . The sharp endpoint-zero Poincaré inequality is
One derivation applies the square identity above on with and , then lets . For endpoint-zero , near zero and near , so the boundary terms tend to zero. The nonnegative square integral yields the inequality.
For , Poincaré gives ; for , directly . Thus throughout , zero is the unique global minimum because the functional is quadratic. If , every multiple of has energy zero: the minimum is not strict. If , substituting that same sine gives negative energy, so zero is not a local minimum. Legendre alone did not distinguish these cases. The Jacobi solution with , has its first subsequent zero at when , identifying the same threshold.
Worked example 3.1.8
(A nonlinear globally convex integrand)
Let and prescribe . Its Hessian in is diagonal with entries and , so it is strictly convex everywhere. The zero curve satisfies Euler–Lagrange. The global convexity theorem therefore proves uniqueness; directly, every term in is nonnegative and the term forces equality only at zero.
For nonzero endpoint data, the same convexity argument proves that any classical Euler–Lagrange solution would be the unique global minimizer. It does not by itself establish existence of such a classical solution. Existence requires an additional argument and a specified function space.
Worked example 3.1.9
(Slope excess proves the shortest graph comparison)
Take and fixed endpoint data. The straight graph has constant slope . Convexity gives
For a competitor , integrate with . The tangent term has constant coefficient, so its integral is . Hence . Strict slope convexity forces equality only when , and the endpoints then give .
This is a full global comparison because the integrand is independent of height and , so no omitted height term remains. It also proves strong local minimality, since a global minimum in the chosen class is a minimum in every smaller comparison neighborhood.
A separate issue is whether a minimizing curve exists at all. Consider the same comparison class on with endpoint values zero and one, and the smooth nonnegative integrand . For every fixed , this integrand is convex in , and it satisfies the Legendre inequality . Nevertheless its infimum is not attained in this class.
To see this, put on . For , define on , and on . Since , , and , this is a admissible curve, not merely a piecewise linear approximation. A change of variable gives
The integrand is nonnegative, so the infimum is zero. If an admissible curve attained zero, continuity and nonnegativity would imply everywhere. Thus for , forcing a constant curve and contradicting the distinct endpoint values. No admissible minimizer exists.
There is also no classical stationary curve satisfying both endpoints. Its momentum equation would be , so would be constant for . Because a curve has bounded derivative near zero, this constant must be zero, again contradicting the endpoints. Thus the global convexity comparison theorem has not failed: its assumption that a stationary reference curve is available is not met.
The example identifies what additional existence theory would need to control. The sequence concentrates its entire height change near the endpoint where the derivative weight degenerates. Its slopes grow without bound, and no subsequence converges in the norm to an admissible curve. Convexity controls comparisons once a candidate exists; it does not automatically provide compactness of minimizing sequences. The direct method addresses existence through suitable growth, compactness, closure of constraints, and lower semicontinuity in an explicitly chosen function space. A full development of that method is outside this section. Keeping existence and sufficiency separate prevents a correct comparison argument from being used to claim a curve that has never been constructed.
Common mistakes
Counterexample mode
A zero second variation may hide a decreasing fourth-order term
For with zero endpoints, is stationary and for every direction. But for every nonzero . Thus nonnegative second variation is only necessary. The coercive theorem cannot be applied because no positive bounds below by .
Common mistake
(Weak sufficiency does not control steep competitors)
A weak sufficient condition controls small heights and small slopes. It does not automatically control high-slope competitors that are close only in height. A proof of a strong minimum must address those competitors through a suitable comparison theorem, such as the global convexity argument above, rather than changing the name of the conclusion.
For a concrete distinction, take on the unit interval with zero endpoints. The zero curve has coercive second variation . When the uniform norm of the slope is smaller than one, the integrand is nonnegative, with equality only when the slope vanishes. Thus zero is a strict weak local minimum. But take . Its height tends uniformly to zero while its slopes grow. Integrating the squared and fourth-power cosines gives , which is negative for all sufficiently large integers. Therefore zero is not a strong local minimum. The Taylor remainder theorem remains correct: this sequence never enters its required small neighborhood in the norm that controls both the function and its derivative.
Summary
Second variation tests all admissible directions, and localization yields Legendre necessity. Coercivity plus a uniform Taylor remainder proves a strict weak minimum. A positive Jacobi solution on the closed interval is one way to verify that coercivity. Joint convexity gives a separate global comparison. The Weierstrass excess function makes slope comparisons explicit, but a general height-dependent strong sufficiency theorem requires additional structure beyond its nonnegativity along a single curve.
Exercises
Quick check
(For with zero endpoints, does the zero curve minimize ? Give a specific direction or an inequality.)
Compare with .
Solution
Since , Poincaré gives for nonzero endpoint-zero . The functional is exactly , so zero is its unique global minimum. The sign would reverse along the first sine direction only if the coefficient exceeded .
Quick check
(For , use a positive Jacobi solution and then give a direct coercivity constant.)
The Jacobi equation is .
Solution
Choose , which is positive on the closed interval and satisfies . The Jacobi criterion proves coercivity. Directly, , so works and is a simpler certificate here. A general theorem is useful only insofar as it verifies a hypothesis that the problem needs.
Quick check
(Construct a stationary functional at the zero curve whose second variation vanishes but which has a strict global minimum there. Compare it with the negative fourth-power counterexample.)
The admissible class consists of continuously differentiable curves with zero endpoints.
Solution
Take . Its first and second variations at zero vanish. Nevertheless, every nonzero continuous curve is nonzero on some open subinterval, so its fourth power has strictly positive integral. Zero is the unique global minimum. The sign-reversed functional is instead strictly smaller away from zero. The common vanishing quadratic approximation cannot distinguish them; the first nonzero higher-order term does.