A small height change can test a distant slope
A weak minimum compares curves whose heights and slopes are both close to the reference curve. A strong minimum must also withstand curves that stay close in height while making a different slope on a short interval. A positive second variation only examines slopes near the reference slope. The Weierstrass necessary condition tests every finite comparison slope, including those outside that neighborhood.
The relevant perturbation has two parts. A short leg uses the desired slope change and creates a small height displacement. A longer, much gentler leg restores the original height. The restoring leg must be accounted for: an isolated slope change generally breaks the endpoint condition. Stationarity cancels its first-order cost, leaving the excess function as the leading action change on the short leg.
We will also smooth the corners. A broken-line drawing is useful for finding the calculation, but it is not a valid competitor when the admissible class is . The smoothing argument below preserves the leading action change and keeps all slopes bounded. Neither unbounded slopes nor a growth assumption at infinite velocity is needed to detect a negative excess at one fixed finite slope.
Comparison classes and slope excess
Definition 14.1.1
(Strong and weak minima in a fixed-endpoint class)
Let , let be open, and assume is on an open set containing . For curves with specified endpoint values, put
A reference curve is a strong local minimum if some gives for every admissible with . A weak local minimum requires this only when . We assume in the theorem below.
“Strong” names the stronger minimizing property. The height-only topology is weaker than the topology: at a fixed radius its ball contains the corresponding ball, so more curves must satisfy the minimum comparison. The compact image of has positive clearance inside , so sufficiently small height perturbations stay in the integrand's domain. All velocities are allowed by the stated domain assumption; a separate problem with bounded admissible velocities would need a correspondingly restricted test.
Definition 14.1.2
(The vector Weierstrass excess)
At a fixed position and height, with reference velocity and comparison velocity , define
The dot denotes the Euclidean pairing in . The excess compares the graph of the velocity integrand with its tangent hyperplane at the reference velocity. It vanishes at .
The position and height are held fixed in this definition. It is therefore a velocity comparison, not a comparison of the full functional at two different curves. This distinction will matter when we discuss sufficiency.
Common mistake
(Comparing integrand values instead of the tangent)
It is tempting to test a distant slope by the sign of . That difference omits the term , which the restoring leg pays for through stationarity. For on with , the comparison slope gives , yet the excess is and the line is the unique global minimum by the exact criterion below. The correct test is the excess, which compares with its tangent at the reference velocity.
The necessary condition
Theorem 14.1.3
(Every strong minimum has nonnegative excess at every slope)
Under the preceding assumptions, if is a strong local minimum, then
This is the Weierstrass necessary condition for the stated unconstrained-velocity, fixed-endpoint problem. The conclusion includes finite velocities far from .
First, ordinary endpoint-zero variations give stationarity: for any fixed such , sufficiently small positive and negative multiples stay in the strong neighborhood. Differentiating gives . The classical Euler–Lagrange equation follows componentwise because and have the stated regularity. Along the reference curve write . Then and .
The identity also holds for continuous piecewise endpoint-zero fields: integration by parts on the pieces cancels internal boundary terms because and are continuous. For the intermediate calculation we use the same action integral for these piecewise curves. Strong minimality will be invoked only after replacing them by smooth competitors.
Proof
(Build a short excursion and a gentle return)
Fix and . Put and . If the excess is zero. Otherwise choose a small positive , and define a piecewise affine endpoint-zero field by
Take small enough that its support is inside . The short leg has derivative ; the return leg has derivative . The latter exactly restores the height, since its length is . Moreover and its slopes remain bounded as .
For this intermediate curve subtract the zero linear variation. If denotes the resulting integrand, then
On the short leg, , , and uniformly, while . Thus its integrand converges uniformly to . Its integral is that excess times , plus .
On the return leg both increments are small. A uniformly bounded Hessian on a compact neighborhood gives
The constants can depend on the fixed comparison slope. Only a compact set of velocities is used. Consequently the broken-line calculation gives
It remains to justify using this calculation in the comparison class.
Proof
(Smooth the three corners without changing the leading action)
Choose a nonnegative, even, smooth function supported in with . Such a function is obtained by normalizing for and setting it to zero outside. Extend the broken-line field by zero and set
This field is smooth. For sufficiently small, its enlarged support is still inside , so it preserves the endpoints. Nonnegativity and unit mass give and when . The derivative is the average of the bounded piecewise constant derivative.
Evenness gives , so convolution preserves every affine segment exactly wherever the averaging interval crosses no corner. Thus and , together with their derivatives, agree outside the three corner neighborhoods of radius . The total length of these neighborhoods is at most .
Both curves and their bounded velocities lie in one fixed compact subset of the domain, for all small . Boundedness of there therefore gives
The smooth curve is a legitimate competitor and its height tends uniformly to the reference height. Strong minimality makes its action difference nonnegative. Dividing its asymptotic action difference by and passing to the limit proves the excess inequality at . Continuity extends it to both endpoints, with any fixed . This proves the theorem.
Proof X-ray
Why the restoring leg disappears only after stationarity
The return leg has a first-order cost of the same order as the short leg; its slope is small, but it acts over a longer interval. Euler–Lagrange cancels the linear variation over the whole perturbation. Only then does the return contribute a quadratic remainder of order smaller than the short-leg length. The all-slope excess is what survives, and the final smoothing keeps this leading term intact.
Corollary 14.1.4
(Legendre is the infinitesimal part of the excess condition)
At every point of a strong minimum, the velocity Hessian is positive semidefinite:
To see this, use in the excess and expand: . Dividing the nonnegative excess by and letting gives the inequality. The Weierstrass condition controls all finite , whereas this corollary controls only the tangent quadratic form.
The smoothed needles need not approach the reference curve in : on the interior of the short leg their derivative change is exactly . Thus a weak minimum does not permit this all-slope argument. The obstruction can occur with a uniformly bounded slope excursion; it does not require derivatives tending to infinity.
A setting where the condition is also sufficient
Theorem 14.1.5
(An exact criterion for an autonomous affine curve)
Suppose the integrand is alone, defined and on , and the reference curve has constant velocity . Among all fixed-endpoint curves, it is a strong local minimum if and only if for every . In that case it is a global minimum.
If the excess is strictly positive for every , the global minimum is unique.
Necessity is the theorem just proved. For sufficiency, write a competitor as with zero endpoint values. Since is constant,
Under the strict condition, equality forces everywhere by continuity, and the endpoints give . No global convexity of is assumed; only a global supporting tangent at this particular reference velocity is needed.
Not by itself; the classical sufficient condition asks for nonnegative excess throughout a field of extremals.
A natural next question
Does nonnegative excess suffice beyond affine curves?
Not by itself; the classical sufficient condition asks for nonnegative excess throughout a field of extremals.
Not in general: the counterexample below has nonnegative excess along a stationary curve that is not even a weak minimum. The proof above needed constant, so that the linear term integrated to zero. The classical replacement embeds the extremal in a suitable field of extremals around its graph. If the excess, taken at the field's slope, is nonnegative at every point of that neighborhood, the action difference is an integral of excess, and a strong minimum follows.
Worked examples
Worked example 14.1.6
(Classifying affine curves for a double-well velocity energy)
On take and the affine curve , with endpoint values zero and . Every such curve is stationary. Direct factorization gives
If , the excess is nonnegative for all , so the affine curve is a global minimum. If and , the choice gives ; if , take . Thus the affine curve is a strong local minimum exactly when .
For the strict excess criterion gives uniqueness. At , equality requires the competitor's continuous derivative to take only values in . It must therefore be constant, and the endpoint difference selects the value . These boundary cases are also unique global minima.
There is nevertheless a range of strict weak minima that fail the strong test. If , then . On a small slope neighborhood, the excess is bounded below by a positive constant times , proving a strict weak minimum after integration. But its excess at the distant slope is negative. For example gives while .
For , the second variation is negative for every nonzero endpoint-zero direction, excluding even a weak minimum. A practice question below handles by a cubic decrease; reflection of curves gives the same conclusion at because is even. Thus the affine curve is a weak local minimum exactly when , and these weak minima are strict.
Worked example 14.1.7
(Coordinate slope tests miss a coupled velocity excursion)
For a two-component curve on with zero endpoints, let
At the zero curve, and . Also , so every excess test confined to one coordinate axis passes. Yet . The vector Weierstrass condition fails, so zero is not a strong local minimum.
It is a strict weak minimum: since , for one has . A weak neighborhood of radius at most bounds the perturbed velocity accordingly, so the action increases unless the perturbation is zero. The decreasing needle changes the two velocity components together. A positive velocity Hessian and all coordinate-axis comparisons did not test that finite coupled excursion.
The exact affine criterion also settles cases where the excess is easy to sign, and its strict form is where uniqueness can be lost.
Worked example 14.1.8
(Arc length, and a null integrand as the edge case)
For the length integrand (scalar, ) and an affine reference of slope , multiply the excess by :
The Cauchy–Schwarz inequality for the vectors and gives , with equality only when the vectors are parallel, that is, when . The excess is therefore strictly positive for , and the exact criterion shows that the segment is the unique global minimizer of length among graphs with the same endpoints.
Edge case: for the excess vanishes identically, since is its own tangent line. The criterion still applies, so every affine curve is a global minimum. The strict hypothesis fails, and uniqueness fails with it: for every competitor, so all admissible curves have the same action. Nonnegative excess gives minimality; only strict positivity rules out other minimizers.
Necessity is not a general sufficiency theorem
Counterexample mode
Nonnegative excess can coexist with a decreasing height direction
For on with zero endpoints, the zero curve is stationary and its excess is for every comparison slope. The velocity Hessian is even strictly positive. Nevertheless,
Thus zero is not even a weak minimum. The height term has disappeared from the same-height excess calculation, but it has not disappeared from the action. The autonomous affine sufficiency theorem cannot be used here. A valid general sufficient comparison needs additional structure, such as a compatible field or a suitable coercive second variation.
The counterexample shows what the excess cannot see: the height dependence of the integrand. What it does see is a statement about the velocity graph alone.
Concept lensGeometric
A supporting tangent is stronger than positive curvature at one point
Legendre asks about curvature at the reference velocity. Weierstrass asks whether the tangent hyperplane there lies below every finite comparison velocity. A nonconvex function can have positive local curvature while falling below that tangent elsewhere; it can also have a globally supporting tangent at selected velocities without being convex everywhere. The double-well example exhibits both possibilities.
Practice
Quick check
(Let be and replace by . Prove that the Weierstrass excess is unchanged and that fixed-endpoint strong and weak minima are unchanged.)
This added integrand is a total derivative along a curve. Check the excess directly as well as the integral.
Solution · Guided solution
Write . Its velocity gradient is . At each fixed position and height, the added terms are affine in velocity: the velocity-independent term and the velocity-linear term cancel in the excess, giving . Neither term is assumed constant in position or height. The chain rule also gives
The endpoint expression is identical for every competitor, so every action difference, and hence either local-minimum notion, is unchanged. The assumed regularity keeps the modified integrand in the stated class.
Quick check
(For on with zero endpoints, show that zero is a strict weak minimum but not a strong minimum. Compute the broken needle's action for the comparison slope , then justify using smooth competitors.)
Use a short leg of length and a return leg of length . The slopes remain bounded.
Solution · Guided solution
Factor . For , it is at least , proving strict weak minimality at zero. But , so the all-slope excess condition fails.
For the broken needle, the short slope is and the return slope is . Thus
Smoothing with width changes the action by , so while . These are admissible smooth decreasing competitors. Their slopes remain bounded rather than tending to infinity.
Quick check
(In the double-well example take , where the Legendre coefficient vanishes. Put and let be its endpoint-zero primitive. Is the affine curve a weak minimum?)
Compute the third and fourth moments of and use the exact polynomial action difference.
Solution · Guided solution
Take . Then and both endpoints vanish. Trigonometric orthogonality gives
The linear term integrates to zero because . At , the quadratic term vanishes, and exact expansion gives
This is negative for sufficiently small negative . These perturbations tend to zero in , so the affine curve is not a weak minimum. Checking a nonnegative, here zero, Legendre coefficient did not control the first nonzero higher-order term.