Motivation
A curve can carry several unknown coordinates. For a path in the plane, changing one coordinate may change the cost of motion in the other; for an energy describing two interacting fields, changing their difference may cost energy even when neither field has a preferred absolute value. Treating each component as a separate scalar optimization problem would discard these interactions. The correct generalization keeps the functional scalar and lets its argument be vector-valued.
Constraints introduce a second kind of dimension. A vector curve with two components can satisfy one integral constraint, two constraints, or many. Conversely, one scalar curve can satisfy several independent moment conditions. The number of multipliers is determined by the number of independent scalar constraints, rather than by the number of coordinates of the curve. Understanding this distinction is more useful than memorizing a longer Euler–Lagrange formula.
We develop a classical fixed-endpoint theory and prove its normal multiplier condition. Knowledge of the scalar first variation, integration by parts and the finite-dimensional implicit function theorem is sufficient. Pointwise state constraints and inequality constraints require additional arguments and are explicitly distinguished below.
Definitions
Definition 4.1.1
(A scalar functional of a vector curve)
Fix , integers , and vectors . Let
Assume is real-valued and on an open neighborhood of the relevant triples, with . Gradients are column vectors. An admissible direction belongs to , meaning both endpoint vectors are zero. Weak local extrema use the norm .
Only curves whose triples lie in the integrand's domain are compared. The open-domain assumption and compact interval ensure that all sufficiently small perturbations of a given curve stay in that domain.
The output is one real number. Consequently its first variation is also scalar:
The two dot products contain every component. They do not imply that splits into a sum of independent component costs. Differentiation under the integral is justified for each fixed direction by continuity of the derivatives on a compact neighborhood of the perturbed triples, just as in the scalar case.
Definition 4.1.2
(Integral constraints and their normal rank condition)
For , take real-valued integrands and require
At a feasible curve , the linearized constraint map is
The normal rank condition is that this map is onto. Equivalently, there exist endpoint-zero directions for which the matrix is invertible.
The equivalence follows directly from finite-dimensional linear algebra. If the map is onto, choose directions whose images are the coordinate basis of . Conversely, an invertible image matrix lets linear combinations of those directions produce any desired constraint derivative. This condition is about first-order independence at the particular candidate. Merely writing different formulas for the constraints does not establish it.
Theorem/Proposition
Theorem 4.1.3
(Vector Euler–Lagrange necessity)
Suppose is an unconstrained weak local extremum. Then
as a vector identity for . Equivalently, there is one equation for every . These are necessary stationary equations.
The total derivative contains the velocity Hessian multiplying . This Hessian is an matrix, and solving explicitly for the accelerations requires its invertibility. The theorem assumes regularity of the candidate; it does not infer that regularity merely from membership in the comparison class.
Theorem 4.1.4
(Normal multipliers for finitely many integral constraints)
Let be a weak local extremum of among feasible curves in , and assume the normal rank condition. There is a unique vector of real constants such that
If also , define . Then
The integral equalities must still be imposed when determining the curve and multipliers. They are not replaced by the differential equations.
The multipliers are constants because there are finitely many scalar integral equalities. If instead a constraint is imposed separately at every , its first-order restrictions are pointwise, and the present theorem does not apply. Similarly, one-sided directions from inequality constraints cannot be treated by simply declaring every first variation zero.
Proof sketch or proof idea
Proof
(From vector variations to a coupled system)
At an unconstrained extremum, . Integrate each component by parts; the endpoint vector is zero, so
Choose , with the th coordinate vector and any scalar endpoint-zero function. The scalar fundamental lemma makes the th coefficient zero. Repeating for every coordinate gives the vector identity. Independent test directions extract equations; they do not remove coupling inside their coefficients.
Proof
(The normal multiplier theorem by an actual feasible correction)
Choose with invertible matrix . For any direction , consider
The integrands' regularity makes continuously differentiable for small parameters. At its derivative in is . The finite-dimensional implicit function theorem supplies a differentiable with and . Thus this family preserves the full nonlinear constraints, not merely their linearization.
Differentiation gives . Let . Constrained stationarity along the corrected family gives
Define by . This vector is fixed by the chosen correcting directions, so the same multipliers work for every . If two multiplier vectors worked, their difference would annihilate the image of , which is all of ; hence they are equal. Finally apply the vector first-variation argument to the augmented integrand .
Proof X-ray
Why a first-order tangent is not the whole proof
A direction satisfying only preserves the constraints to first order. A straight perturbation can still violate a nonlinear constraint at second order. The implicit function theorem supplies an exactly feasible nearby curve family. The rank condition permits that correction; stationarity along the corrected family supplies the multiplier identity. No infinite-dimensional implicit function theorem is needed here, because the correction uses only finitely many parameters.
For a pointwise contrast, at every forces at every . The coordinate tests are then generally inadmissible. One must account for the tangent space at each position; position-dependent multipliers can arise in a separate theory. An integral condition instead requires only one scalar equality and, when its derivative is nonzero, has one constant multiplier. Similar notation should not obscure these different feasible sets.
Worked examples
Worked example 4.1.5
(Coupled velocities and an exact minimum)
On , minimize
with and . The momenta are and , so the equations are and . Their matrix is invertible and positive definite: . Thus both accelerations vanish and the endpoints give .
For any endpoint-zero vector perturbation , the linear term in the objective difference is , since is constant. Hence
Equality implies , then . The straight vector curve is therefore the unique global minimizer, and . The cross term affects both momenta even though the final path is straight.
Worked example 4.1.6
(Height coupling separates into sum and difference)
Take , with endpoint vectors and . The coupled equations are
Set and . Addition gives and subtraction gives . The endpoints become and . Therefore
This change of variables diagonalizes the interaction; it does not assume that either original component is independently minimizing. Integration by parts using the coupled equations cancels the linear terms, leaving the exact gap
It is positive for nonzero endpoint-zero , proving a unique global minimum.
Worked example 4.1.7
(Two integral constraints with coupled kinetic energy)
Minimize with both endpoint vectors zero and
Let . Correcting directions and give
For , the augmented equation is . The solution and constraints give
To check the signs, put . Then and . Both multipliers are constant; their values are fixed only after imposing the two integral equalities.
A feasible perturbation has , so . The exact gap is . Since is positive definite, the candidate is the unique global constrained minimizer, with
Common mistakes
Counterexample mode
A constrained minimum with no normal multiplier
Use zero endpoint vectors and require . Continuity and nonnegativity force the sole feasible curve to be . It therefore minimizes on the feasible set. But at zero, for every direction, whereas is not identically zero. No constant makes vanish. The rank condition fails exactly where the normal theorem would need it.
Common mistake
(Inverting a singular correcting matrix)
Rank failure does not always prevent multipliers from existing. Redundant constraints and have dependent derivatives; multipliers, if present, can be nonunique because only their combination enters the augmented integrand. Remove redundancy or analyze degeneracy explicitly rather than inverting a singular correcting matrix.
A multiplier rule that also puts a multiplier on the objective, which may then be zero.
A natural next question
What survives when the rank condition fails?
A multiplier rule that also puts a multiplier on the objective, which may then be zero.
At a constrained extremum there are always constants with on . Otherwise would map onto , and the implicit-function correction above, with added to the corrected quantities, would give feasible curves with on both sides of . In the counterexample, and work. Normality is exactly what forces .
Common mistake
(A positive velocity Hessian is not a minimum test)
Nor does a positive velocity Hessian alone prove a minimum. For with zero endpoints, zero is stationary and . The direction lowers the objective, while raises it. The height term matters. The exact nonnegative gaps in the worked examples supplied the missing sufficient comparisons.
Summary
Vector variations produce one Euler–Lagrange equation per state component, with coupling retained inside the gradients. Independent scalar integral constraints contribute constant multipliers when their linearized map is onto. The finite-dimensional implicit function theorem constructs feasible variations and proves the normal multiplier identity. Solving the augmented equations still requires endpoints and constraint values, and proving a minimum still requires a separate comparison.
Exercises
Quick check
(Minimize with zero endpoint vectors and . Verify normality, find the multiplier for , and prove the minimum.)
The gradient of the constraint integrand is .
Solution
The direction has constraint derivative , so normality holds. The augmented equations are . Zero endpoints give . The constraint yields , hence and . Its energy is . For any feasible perturbation, the cross term is , leaving . This proves a unique global minimum, not merely stationarity.
Quick check
(For a scalar zero-endpoint curve on , take and . Show that their derivatives are independent using the directions and . Explain why two constraints are possible with only one state component.)
Calculate all four entries of the correcting matrix.
Solution
The correcting matrix is
Thus the derivative onto is surjective. The space of scalar curve variations is infinite-dimensional; it is not a one-dimensional space simply because each function value is scalar. Two independent moments therefore pose no dimension contradiction. In fact the invertible matrix also constructs feasible curves for any prescribed pair of moment values by combining these two directions.
Quick check
(For , prescribe endpoint vectors and . Derive the system, use the conserved momentum of , and prove your candidate is the unique global minimum.)
The integrand is independent of but not of . Keep those partial derivatives distinct.
Solution
The momenta are and . Since , write . But , so the other equation is . The endpoint conditions give
Their energy is . The stationary linear terms vanish for endpoint-zero perturbations, and the exact gap is . Equality implies , then ; next and . Thus the candidate is the unique global minimizer. Conserved momentum is not the same as constant velocity when that momentum contains another coordinate.