Motivation
A shortest-path question asks for a curve, not a single real number. Even after a curve is represented as a graph , its values at different points cannot be optimized independently: the integrand may depend on the slope, and changing one part of the graph affects how it joins the rest. The calculus of variations turns this coupled problem into an equation that a candidate curve must satisfy.
The central device is to put a one-parameter family through a candidate. Ordinary one-variable calculus then supplies a necessary condition. The substantial work is identifying which families are admissible, justifying differentiation under the integral, and extracting an equation from the resulting integral identity. This process produces stationary curves. Whether they minimize the functional is a separate question addressed by second variation and sufficient conditions.
We work on a finite interval , with , and first keep both endpoint values fixed. All integrals in this section are ordinary integrals of continuous functions; later problems posed in weaker function spaces are not covered by the conclusions of this section.
Definitions
Definition 1.1.1
(Functional, admissible class, and local minimum)
Let be on an open neighborhood of the triples under consideration. The functional is
A weak local minimum satisfies for all admissible with , for some , where . A strong local minimum uses instead, with the same admissible class. A global minimum compares with every admissible curve.
A strong minimum is therefore a weak minimum: closeness in the norm implies closeness in the uniform norm. The converse need not hold because a small-amplitude perturbation can have a large slope. These terms describe the neighborhood used for comparison, not the size of the objective value. Existence is also not part of the definition: specifying an infimum does not produce a curve attaining it.
Definition 1.1.2
(Admissible direction and first variation)
For fixed endpoints an admissible direction is with . Set . The first variation in that direction is
The derivatives of are evaluated at . A curve is stationary when this expression vanishes for every admissible direction.
For a fixed direction, the relevant triples for small lie in a compact subset of the open domain of . Continuity of its derivatives bounds the differentiated integrand uniformly. This justifies differentiating the integral. At a local minimum, the scalar function has a minimum at zero, so its derivative vanishes. This uses both positive and negative small . With inequality constraints, some directions are one-sided, and equality of the derivative to zero requires reconsideration.
Theorem/Proposition
Theorem 1.1.3
(Fundamental lemma and Euler–Lagrange necessity)
If a continuous function on satisfies for every function vanishing at the endpoints, then on .
Consequently, if , , and is a weak local extremum in , then
This is the Euler–Lagrange equation. It is a necessary stationarity condition, not a general minimum test.
The derivative in this equation is a total derivative along the curve. In particular,
The equation can be solved explicitly for where . If that coefficient vanishes, the resulting equation may be degenerate. We have assumed regularity of the curve to state this classical form; it does not follow just from the definition of the admissible class. More general weak formulations require their own regularity arguments.
Theorem 1.1.4
(Beltrami identity for an autonomous integrand)
If is independent of and a curve satisfies Euler–Lagrange, then
for a constant along the curve. This identity often reduces a second-order equation to a first-order relation. It does not in general replace Euler–Lagrange on portions where .
Theorem 1.1.5
(Normal multiplier for one integral constraint)
Let , with . Suppose is a constrained local extremum of with fixed endpoints and . If there exists an endpoint-zero direction with , then some real number satisfies
for every endpoint-zero direction . If , it therefore satisfies Euler–Lagrange for .
The nonzero derivative is a constraint qualification. Without it, the normal multiplier conclusion has not been established. The multiplier is constant because the constraint is one scalar integral, rather than a separate pointwise condition at each . The sign convention chosen for the augmented functional also determines the sign of the multiplier.
Counterexample mode
A degenerate constraint has no normal multiplier
The qualification cannot simply be omitted. On with zero endpoints, minimize subject to . Continuity forces the only feasible curve to be , so it is a constrained minimum. But for every direction, while equals for . No constant makes vanish. The multiplier theorem is not contradicted: its nondegeneracy hypothesis fails at the only feasible point, which is exactly why that hypothesis is stated.
Proof sketch or proof idea
Proof
(The fundamental lemma and the first variation)
If at an interior point, continuity gives a closed subinterval on which . Choose on and zero outside. It is , nonnegative, and positive inside that subinterval, so , a contradiction. Apply the same argument to to exclude negative values. Continuity then gives zero at the endpoints too.
At a local extremum, the first variation is zero. Integration by parts gives
The boundary term vanishes because the endpoints are fixed, and the fundamental lemma gives the differential equation. Notice that endpoint restrictions were used on , not imposed on .
Proof X-ray
What each proof step actually uses
Two different ideas carry the proof. The fundamental lemma is a localization argument: the bump is and vanishes at both ends of , so it is itself an admissible direction, and continuity of is what supplies an interval of constant sign. The second idea is integration by parts, which moves the derivative from onto ; this is the one place where the hypothesis on is used, because must exist and be continuous for the lemma to apply. Fixed endpoints enter only through , which removes the boundary term; with a free endpoint the same calculation leaves a natural condition instead. Differentiability of along the family needs a smooth integrand on a compact neighborhood of the perturbed triples. The argument uses only that the derivative at vanishes, from both signs of , so it applies equally to minima and maxima and proves nothing about which occurs.
Proof
(Differentiate the proposed first integral)
For Beltrami, differentiate along a curve. The result is . Both terms vanish under the stated assumptions. Conversely, constancy with yields only . Dividing by at a zero is not legitimate. For example, with , every constant curve satisfies the Beltrami identity but none satisfies Euler–Lagrange, which would require .
Proof
(Correct a line perturbation back onto the constraint)
For the multiplier statement, apply the ordinary implicit function theorem to near . Its derivative in is nonzero, so there is a differentiable maintaining the constraint. Differentiating gives . Constrained stationarity now gives
Thus works for every . This is why an arbitrary straight-line perturbation need not itself preserve the integral constraint: the small correction in the direction restores it.
There is a useful way to recover some regularity without assuming a second derivative of at the outset. Suppose only that a curve is stationary. Put , , and . These three functions are continuous, and is continuously differentiable. The first variation and integration by parts applied to , rather than to the possibly nondifferentiable , give
for every endpoint-zero direction. Let and let be its average over the interval. The continuous function has integral zero, so its integral from to defines a legitimate endpoint-zero function . Using this particular direction gives . Thus is constant. It follows that is actually and .
This integrated argument proves the Euler–Lagrange equation in momentum form even before a second derivative of the curve is available. To go further, suppose is nonzero along the curve. The implicit function theorem locally solves for the slope as a function . Since , , and the newly regular momentum are , the identity makes continuously differentiable. Hence the curve is locally. This explains why a nondegeneracy hypothesis can connect a weak stationarity identity to the classical equation. Without that hypothesis, the final regularity step has not been justified.
Another useful check is invariance under a total derivative. If is , replacing by adds along a curve, whose integral is the same for every competitor in the fixed-endpoint class. Stationary curves and minimum comparisons are therefore unchanged, and inserting the extra terms into Euler–Lagrange shows the mixed derivatives cancel, a useful algebra check. When an endpoint can move, however, the added term becomes a terminal cost: interior equations still agree, but natural boundary conditions can change, as the next example shows.
The same interior equation, different endpoint choices
Worked example 1.1.6
(A total derivative becomes a terminal cost)
Take curves on with , and consider
because . First fix . The extra term is the same number for every competitor. The affine curve minimizes the remaining energy: for , the endpoints of are zero, and
Equality requires and hence ; the unique minimum is .
Now keep but leave free. A variation must vanish at but can have any value at . For a stationary candidate, integration by parts gives
Interior variations give ; then arbitrary endpoint values give the natural condition . Together with , these give . This calculation finds a candidate, not by itself a minimum. For every admissible competitor, complete the square instead:
Equality holds only for , and the left endpoint then forces . Thus the unique minimum is . By contrast, without the added derivative the free-endpoint minimum is . The interior equation is unchanged, but the endpoint condition and answer change. Here “free endpoint” means a free endpoint height at fixed , not a moving value of .
Quick check
(Let and leave free. Minimize over .)
Write as a total derivative.
Solution · Answer
Since , this is . Stationarity gives and , hence . For , and the cross term equals
Consequently , with equality only at . The unique minimum is . The boundary term is part of the energy, so it must remain in the comparison.
When the integrand is affine in the slope
Earlier we noted that the standard division argument solves the Euler–Lagrange equation for where . The most basic case in which this fails everywhere is an integrand that is affine in the slope,
with and of class on an open subset of the -plane containing the graphs of the curves under consideration. Here . The total-derivative integrands above have exactly this form, but they are a special case. The general case behaves very differently, and it is worth seeing why.
Theorem 1.1.7
(The Euler–Lagrange condition loses its derivatives)
Let be stationary for under all endpoint-zero directions. Then
This is a relation between and the height alone. No derivative of appears, and no assumption on the curve is needed.
Proof
(Integrate by parts; the slope terms cancel)
To prove it, note that the momentum is already a function of along any curve, with derivative . So we may integrate by parts directly, with no need for the integrated argument used above:
The boundary term vanished because . The terms cancel exactly, and that cancellation is the whole point. The fundamental lemma then gives the displayed relation.
The cancellation has two very different consequences.
- If on an open rectangle (or any convex open set) containing the graphs, the form has a potential with and . Then is the total derivative studied above, every admissible curve is stationary, and for every competitor.
- If is not identically zero, stationarity still forces the graph into its zero set. That pointwise constraint can leave no freedom to meet the endpoints, as in the example below; it need not do so. When the zero set fails to contain either or , no admissible curve is stationary. Every local extremum in the norm has vanishing first variation, so then has no -weak local minimum or maximum in the class.
There is also a geometric reading. Let be curves with the same endpoints, and suppose are on an open neighborhood of the entire closed strip . Then
To see this, the fundamental theorem of calculus expresses as the integral of . Put . Differentiating gives . The endpoint values of are zero, so integration yields the formula. It remains valid if the graphs touch infinitely often; no simple-boundary assumption is needed.
Worked example 1.1.8
(One candidate line, and an infimum that is not attained)
On , consider curves with and , and
Here , , and . By the theorem, a stationary curve must satisfy at every point. That line meets the right endpoint only when .
An exact identity settles every case without assuming that a minimizer exists. For any admissible , integrate the slope term by parts using and :
Case . Then . Equality forces the continuous function to vanish, so is the unique global minimizer and . The line was found from the necessary condition. Its minimality comes from the identity, which also agrees with the strip formula: for with , the swept integral is .
Case . Every admissible curve differs from at . By continuity it differs on a nontrivial interval near , so strictly. The bound is still sharp. The smooth curves satisfy both endpoint conditions, and
So the infimum is and no admissible curve attains it. The sequence follows the line on most of the interval and then makes up the endpoint mismatch in a thin layer near . The endpoint contribution is fixed, and the squared-error integral of that layer tends to zero. Taking with shows that is also unbounded above. With no stationary curve available, has no -weak local extremum in this class.
When vanishes identically, the Euler–Lagrange condition need not produce a two-parameter family for the boundary values to select from, so one must check directly whether the prescribed endpoints lie on the stationary set; if they do not, even a finite infimum may fail to be attained.
Quick check
(For with and , find all stationary curves and the value of .)
Compare with .
Solution · Answer
Here , so every admissible curve is stationary. Indeed , so for every competitor, and every admissible curve is a global minimizer and maximizer at once.
Quick check
(For with , find the stationary curve and decide whether it is a minimum.)
Integrate by parts.
Solution · Answer
Now , so the only candidate is , and it is admissible. Integrating by parts, , so . Equality holds only for , which is therefore the unique global minimizer.
Worked examples
Worked example 1.1.9
(Fixed endpoints and minimum Dirichlet energy)
Minimize with , , and . Euler–Lagrange gives , hence
To establish a minimum, write any competitor as , where . Expanding yields
The cross term is zero because is constant and . Equality forces and then . Thus the curve is the unique global minimizer, with minimum . The equation found the candidate; the exact difference proved the claim.
Worked example 1.1.10
(An energy with a restoring term)
For , , and , the equation is . The two-dimensional solution family is , and the endpoint conditions give
For any endpoint-zero , integration by parts and cancel the linear terms. Therefore
This is strictly positive for nonzero , proving a unique global minimum. A common sign error would give ; checking the derivative of immediately detects it. Along the actual curve, Beltrami says is constant, which also follows from .
Worked example 1.1.11
(Prescribed area produces a parabola)
Minimize subject to and . The constraint derivative is , nonzero for , so the normal multiplier condition applies. For , Euler–Lagrange gives .
The endpoint conditions give . Its integral is , so
If satisfies the same constraints, then and has zero endpoints. The cross term becomes . The objective gap is again , so this is the unique global constrained minimizer, for either sign of .
Common mistakes
Common mistake
(Stationary does not mean least)
For with zero endpoints, satisfies Euler–Lagrange. But gives , whereas gives . Both families approach zero in . The stationary curve is neither a local minimum nor a local maximum.
Common mistake
(Imposing natural conditions at fixed endpoints)
Endpoint values must be substituted after solving the equation, and their restrictions must already have been used when deriving it. Do not add a condition such as in a fixed-endpoint problem. There the boundary variations vanish, so the boundary term gives no such condition. Free endpoints require a different calculation. Also, do not treat writing down the differential equation as a proof that a smooth solution exists: the equation, the boundary conditions, the admissible class, and minimality are separate steps of the argument.
Summary
An admissible perturbation turns a functional into an ordinary scalar function of its parameter. First variation, integration by parts, and the fundamental lemma give Euler–Lagrange under explicit classical regularity assumptions. Autonomy yields a first integral, and a regular scalar integral constraint yields a constant multiplier. Each device narrows the candidates; an objective comparison or a sufficient condition is still needed to prove minimization.
Exercises
Quick check
(For with , find the stationary curve and decide whether it minimizes .)
Complete the square after writing a competitor as the candidate plus an endpoint-zero function.
Solution
The equation is , so . For , the linear contribution is . Thus , with equality only for . The minimum value is : integrate over . A negative linear term in an integrand does not by itself make the functional unbounded below.
Quick check
(For , test the curve against both Beltrami constancy and Euler–Lagrange, and explain the difference.)
Keep the factor when differentiating the proposed first integral.
Solution
For the constant curve, is constant. However, . Beltrami constancy gives , which here reads and supplies no Euler–Lagrange information. It is a necessary consequence of the equation, with a weaker converse at zero velocity.
Quick check
(Minimize with zero endpoints and . Find the multiplier for the convention and the minimum value.)
The sign of the multiplier depends on the stated convention.
Solution
The constraint qualification holds because . The calculation in the worked example gives , , and . For every admissible perturbation with zero integral, the exact objective gap is , proving uniqueness. A multiplier of would instead correspond to writing the augmented functional as .