Motivation
Euler–Lagrange describes what happens inside a smooth extremal. It does not determine how the curve meets a free boundary or whether two smooth pieces may join at a corner. Those questions are encoded in the boundary terms discarded when both endpoints were fixed.
A free endpoint height and a moving endpoint location are different freedoms: with the terminal abscissa fixed only a vertical variation is allowed, while an endpoint sliding along a target curve has related horizontal and vertical displacements. A corner has the same two freedoms inside the interval.
We continue with , , for scalar curves that are on smooth pieces and continuous across finitely many admitted corners. The conditions below are necessary for stationarity, not certificates of a minimum.
Definitions
Definition 2.1.1
(Momentum and the boundary covector)
Along a smooth curve write
If an endpoint has displacement in the plane, the boundary contribution at the right endpoint is
The left endpoint carries the negative of this expression. Here is the total change of the endpoint ordinate, including the effect of moving its abscissa.
Some accounts use the Hamiltonian . The expression is a covector because it acts linearly on the allowed displacement; it is not in general the Euclidean normal to the extremal.
Definition 2.1.2
(A sliding endpoint and an interior corner)
A right endpoint constrained to the graph is described by and , with . Its first-order displacement satisfies .
An interior corner at joins two smooth arcs with the same value but possibly different one-sided slopes and . A movable corner permits independent small changes in and . At a prescribed fixed interface, is not an available variable.
Theorem/Proposition
Theorem 2.1.3
(Natural boundary conditions and transversality)
Suppose a smooth stationary curve has fixed left endpoint and satisfies Euler–Lagrange in its interior.
If is fixed and is free, then . If the right endpoint instead slides freely along , with its abscissa interior to the allowed parameter range, then
With an additional terminal cost , , the sliding condition becomes
at the endpoint. At fixed with free terminal height it becomes .
“Free” means that both signs of small variations in the relevant variable are allowed. At the edge of a permitted parameter interval only one-sided variations may exist, and an inequality can replace the equality.
Theorem 2.1.4
(Weierstrass–Erdmann conditions for a movable corner)
Suppose a continuous piecewise stationary curve has one interior corner, and the admissible class permits independent movements of its location . For the same smooth integrand on the two arcs, Euler–Lagrange holds on each open arc and
The first equality is momentum continuity; the second is continuity of . If the corner abscissa is prescribed and cannot move, the boundary calculation supplies only the first equality.
If the problem requires globally curves, a slope jump is excluded regardless of these conditions. When corners are allowed, the equalities can still exclude them: if is injective, momentum continuity forces .
Proof sketch or proof idea
Proof
(The moving-endpoint calculation)
Assume is near with , and is jointly on a common neighborhood of , with fixed. Write , , and . The partial derivatives , are evaluated along , and .
Step 1: a moving upper limit. depends on through the integrand and through the interval. Leibniz's rule records both:
At the chain rule and the commuting of and (joint smoothness) give
Step 2: integration by parts. Since is and (differentiate in ),
No cancellation happens at : need not vanish.
Step 3: from to the endpoint displacement. holds the abscissa fixed at , whereas follows the actual endpoint as moves. The endpoint rises or falls merely by sliding along the graph, and the chain rule makes the difference exact:
Substituting and collecting coefficients gives the moving-endpoint formula
On an extremal the integral vanishes, leaving , in terms of the quantities that endpoint conditions relate.
For fixed and free height, while is arbitrary, so . On the target graph, , and arbitrary gives . Finally, the differential of the terminal cost is , which yields the stated modified formulas.
The same bookkeeping, applied on each side of an interior point, gives the corner conditions.
Proof
(Why a corner gives two conditions)
Split the integral at and apply the preceding boundary calculation separately to the two arcs. Once Euler–Lagrange holds on each arc, the right boundary contribution of the left arc is . The left boundary contribution of the right arc has the opposite sign. Thus
When both displacement variables are independently allowed, each coefficient must vanish; the subsection on corner conditions without smooth arcs below constructs such families and removes the assumption on the arcs. If is fixed, for every admissible family, so the second coefficient is not tested.
Concept lensGeometric
The boundary condition acts on a tangent
The covector formulation also works when the target is not conveniently a graph. Suppose a regular target is parametrized by , with , and the terminal parameter lies in an open interval. Its allowed displacement is , so stationarity requires ; a terminal cost adds . A vertical target line, and , recovers ; a horizontal target tests instead.
For the length integrand, is the unit tangent to the graph. Its action on the target tangent is then the Euclidean dot product, and transversality becomes perpendicular intersection. This geometric interpretation is derived from the functional, rather than assumed in advance.
Perpendicularity does not decide which stationary endpoint is closest. From to an endpoint on the unit circle , the best curve is a segment of length . Both and satisfy transversality, but gives the global minimum , while gives and is not even a local minimum. Boundary equations only eliminate candidates.
Worked examples
Worked example 2.1.5
(A free terminal height for a coercive quadratic energy)
Minimize
with free. Euler–Lagrange gives , while the natural condition is because . Solving gives
For any competitor , only is required. The linear term is
The gap is , so the candidate is the unique global minimum. Without the natural condition, the boundary term at would remain and this argument would fail.
Worked example 2.1.6
(Shortest graph from a point to a sloping line)
For the length functional , let the initial point be and the terminal point slide along with . Euler–Lagrange gives constant slope . Since
transversality gives . The endpoint equation is , hence , and on .
The extremal tangent is perpendicular to the target tangent . For global minimality, every admissible graph is at least as long as the distance between its endpoints, which is minimized at and attained by the segment. The minimum length is .
Worked example 2.1.7
(A genuine minimizing corner in a nonconvex problem)
On , allow continuous piecewise curves with , and take
The curve has slopes and . Its energy is zero, and the nonnegative integrand proves it is a global minimizer in this class. Along either arc, and . Both corner conditions hold despite the slope jump.
There is no contradiction with the injectivity observation: takes the same value at and . In a globally problem this curve is not admissible.
Corner conditions without assuming smooth arcs
The corner argument above assumed each arc is and that a family moving the corner horizontally exists. In a problem posed over broken curves, we know only that the minimizer is continuous and piecewise . The first condition below comes from integrating the first variation once; the second from an explicit inner variation, which slides the graph sideways. That variation is small in the uniform norm but not in slope, which decides which kind of minimum the second condition needs.
Definition 2.1.8
(Broken curves and their two neighborhoods)
A function on is piecewise , written , if it is continuous and there is a partition such that is on each closed subinterval , using one-sided derivatives at its ends. The one-sided slopes and therefore exist. A point where they differ is a corner. The curve is admissible if , , and every one-sided triple lies in the open set on which is .
An admissible is a strong local minimum if for every admissible with , for some . It is a weak local minimum if this holds whenever, in addition, at every point where both curves are differentiable.
Every strong local minimum is a weak one, because the weak neighborhood is smaller.
Theorem 2.1.9
(Lemma of du Bois-Reymond for piecewise continuous functions)
Let be piecewise continuous on : continuous on each closed piece of some partition once its one-sided limits are used at the partition points. If
for every with , then equals a constant at every point where it is continuous.
Proof. Let be the average of over and put . This is continuous and piecewise , though generally not , and . Since , the hypothesis gives
A nonnegative function that is continuous on each open piece and has zero integral vanishes on each open piece. At a partition point where is continuous, follows by taking limits.
Theorem 2.1.10
(Integrated Euler–Lagrange equations and the Weierstrass–Erdmann conditions)
Let be admissible, and write and along , using one-sided slopes at partition points.
Part 1. If is a weak local minimum, there is a constant such that
at every point where is differentiable, and the one-sided values satisfy the same identity at each partition point. Consequently at every corner. Also, is on each closed arc with .
Part 2. If is a strong local minimum, there is a constant such that
in the same sense. Consequently at every corner. If does not depend on , then is one constant along the whole broken curve, across its corners.
No second derivative of is assumed; regularity of an arc is a separate question, answered locally in the previous note under . The right-hand sides are integrals of bounded piecewise continuous functions, hence continuous in ; this one observation yields both corner conditions.
Proof
(Part 1: vertical variations)
Let vanish at both endpoints. The one-sided triples of form a compact subset of : on each closed arc they are the image of a continuous map. Since is open, some neighborhood of also lies in . Hence is admissible for all sufficiently small . It also lies in the weak neighborhood, because both and are uniformly small. So has a local minimum at , with both signs of available. Differentiating on each piece of a common partition of and gives
Put . It is continuous and piecewise with on each open piece. Integration by parts is valid for products of such functions, and , so . Therefore for all such . The lemma gives wherever is continuous. On each closed arc, is continuous, so the identity also holds for the one-sided values. Because is continuous, the two one-sided values of at a corner both equal . Finally, is on each closed arc with derivative .
Part 2 needs a different family, one that moves the corner horizontally.
Proof
(Part 2: the inner variation that moves corners)
Let satisfy , and put . For , the function is continuous, has piecewise derivative , and fixes and . Hence it is an increasing bijection of onto itself whose inverse is piecewise . Define the competitor by
The point of the graph above moves horizontally to at unchanged height, so a corner at moves to . The curve is continuous, piecewise , has the original endpoint values, and
Admissibility and closeness. The triple of at differs from the triple of at by , uniformly in , so it lies in for small by the compactness argument of Part 1. With and ,
So is in every strong neighborhood of once is small.
Differentiation. Change the variable to , so that :
The interval no longer depends on , and the integrand is smooth in on each piece of a common partition. Differentiating at gives , so
This has exactly the form of Part 1, with in place of and in place of . The same integration by parts and the same lemma give and the continuity of at corners. When , the integral is zero and throughout.
To first order the inner variation moves a corner by , and a vertical variation by ; these are the independent displacements in .
Proof X-ray
Why Part 2 asks for a strong minimum
Near a moved corner there is an interval of length about on which still uses the slope from one side of the old corner, while already uses the slope from the other side. On that short interval the slopes differ by roughly the full jump , however small is. Thus the inner family is uniformly close to but generally not close in slope, and only a strong local minimum is guaranteed to be no larger than it. The proof therefore does not establish the second condition for a merely weak local minimum. Part 1 uses only vertical variations, which are close in slope, so it holds for weak minima.
What the two conditions say about the slopes at a corner
At a corner, and are common to both sides, and only the slope jumps. Put . The tangent line to the graph of at is
Its slope is and its value at is . Hence the two Weierstrass–Erdmann conditions together say: the tangent lines to at the two corner slopes are the same line. If is strictly convex, a tangent line lies strictly below the graph at every other point, so no corner exists. More precisely, if were convex between and , a line tangent at both ends would lie below and coincide with the chord, while convexity keeps below the chord; so would be affine there. A double tangent at distinct slopes therefore requires to fail strict convexity between them.
Worked example 2.1.11
(Classifying every strong local minimum of a double-well slope energy)
On , with and , minimize over admissible curves
Here depends only on . Direct differentiation gives
Let be a strong local minimum. Since , the theorem gives constants with and at every point of differentiability.
Case 1: has a corner, with slopes . There is then a line tangent to at both. The quartic has leading coefficient and double roots at and , so , where and . Since is linear, the difference
must have degree at most one. Its coefficient is , so . Then it equals , whose coefficient forces . Thus , and , so . Now every slope of is a common root of and . The roots of are and those of are . Hence every slope lies in . On each arc, the derivative is continuous and takes only these two values, so it is constant there.
Case 2: has no corner. Then is continuous, and allows only the at most three roots of a cubic. So is constant, and the endpoints force with . This curve is not a strong local minimum. For each integer , let have slope on and slope on , for . Then , , and . Case 2 is therefore impossible.
So every strong local minimum is a broken line with slopes and whose slope- pieces have total length . Conversely, every such curve has and, since , is a global minimum. The straight line satisfies the integrated Euler–Lagrange equation but fails a comparison the corner conditions never test.
Common mistake
(Two ways to misuse the corner conditions)
“The arcs must first be shown to be .” No. Part 1 yields momentum continuity and the momentum form of Euler–Lagrange directly from the integrated first variation. regularity of an arc is a further conclusion that needs a hypothesis such as .
“Vertical variations also give continuity of .” No. Such variations never move a corner horizontally, which is why a prescribed interface yields only momentum continuity. Continuity of needs a variation that moves the corner, and the natural one is the inner variation.
Common mistakes
Common mistake
(Matching too much at a fixed interface)
Take , with , , allowing continuous piecewise smooth curves. Constant slopes on each side and momentum continuity give . The endpoint difference gives , so and . Here and , which differ. The interface is fixed, so this is allowed. Expanding the quadratic energy and using momentum continuity proves the curve is the unique global minimizer in this class.
For a general , transversality is the covector equation on allowed endpoint tangents, not Euclidean perpendicularity.
Summary
Boundary conditions come from available variations. A free height at a fixed abscissa tests momentum; a sliding endpoint tests the full boundary covector against the target tangent; a movable corner tests both its height and location, producing the two Weierstrass–Erdmann equalities; a prescribed interface tests only the height.
Exercises
Quick check
(Minimize with and free. Find the natural condition and the minimizer.)
Include the derivative of the terminal cost before solving for the slope.
Solution
Euler–Lagrange gives , so . The boundary condition is , giving . For an arbitrary with , the linear terms cancel by this condition and the gap is . Hence the minimizer is unique and global. Setting alone would omit part of the objective.
Quick check
(For the length integrand and a terminal graph , derive the condition on the endpoint slope . What does it say when ?)
Remember that these notes restrict competitors to graphs with finite derivative.
Solution
The condition is , or . When the target tangent is horizontal, no finite satisfies it: a perpendicular segment would be vertical and needs a different parametrization. So there is no interior stationary endpoint among finite-slope graphs; this is not a nonexistence theorem for parametrized curves.
Quick check
(Show that a stationary movable corner for must have matching one-sided slopes. Does the same argument apply to ?)
Inspect whether the momentum map is one-to-one.
Solution
For the quadratic integrand, , so directly gives . The second equality then follows as well. For , the momentum map is not injective: both slopes and give and . Thus the first argument does not apply, and the absolute-value minimizer above supplies a concrete allowable corner.