Changing variables requires an inverse, not just a new letter
The Euler–Lagrange equation differentiates momentum, . It is tempting to treat as an independent variable and replace the second-order equation by a first-order system. That step is legitimate only when the momentum relation determines velocity. A nonzero derivative can give a local inverse without giving every momentum a velocity or selecting one global branch.
Velocity convexity provides a particularly useful global answer. The quantity is maximized over velocities, and its value becomes the Hamiltonian. Differentiating at the maximizing velocity produces the canonical equations. The maximization matters: the associated action in state and momentum variables is a stationary principle and can be a saddle even when the original configuration-space action has a strict minimum.
We prove the transformation under a uniform convexity hypothesis, then test its limits. One nonlinear example will have an independently proved global minimum; another will illustrate a coupled canonical system. These comparisons keep equivalence of differential equations separate from existence and minimality of a boundary-value solution.
The velocity Legendre transform
Definition 15.1.1
(Uniform convexity and the conjugate Hamiltonian)
Let be an open interval and be open. Assume and that one constant satisfies
On a fixed compact interval with , the configuration action is
for admissible curves with prescribed endpoint values. All compared curves must stay in . The state is , velocity is , and momentum is . At fixed position and state, define the Legendre conjugate
All vectors and pairings are real and Euclidean. The supremum definition is meaningful even outside the convex setting, but finiteness, attainment and smoothness must then be checked separately.
Theorem 15.1.2
(A global velocity inverse and a smooth Hamiltonian)
Under these assumptions, for every there is exactly one velocity with . It is the unique maximizing velocity in the definition of . The inverse is jointly in its variables, is finite and , and
The matrix inequalities are in the quadratic-form order. At fixed , is Lipschitz in momentum with constant .
Proof
(Use coercivity for existence and monotonicity for uniqueness)
Fix . Integrating the velocity Hessian along a line gives
Thus this continuous function tends to infinity as . It has a minimum: outside a sufficiently large ball its value exceeds its value at zero, and a closed ball is compact in this finite-dimensional velocity space. At a minimizer its gradient vanishes, giving .
For any velocities , the same Hessian bound yields
This proves uniqueness. Applying Cauchy–Schwarz to the left side also gives the claimed inverse Lipschitz bound. Strict convexity makes the minimizer of the unique maximizer of the conjugate expression.
The implicit function theorem applies to , since its velocity derivative is the invertible Hessian. It gives a local inverse; uniqueness makes all such local inverses agree, producing the joint inverse on the stated domain.
Now write . On differentiating, every term containing a derivative of cancels because . This proves the three first-derivative identities. Their right sides are , so . Differentiating in gives the Hessian formula and its bounds.
Proof X-ray
Three steps precede the first-order system
The proof uses three separate ingredients, and the implicit function theorem is only the last. Coercivity, from the uniform bound integrated along a line, gives every momentum at least one velocity; the exponential counterexample below shows that a merely positive Hessian can leave momenta uncovered. Monotonicity, the same bound integrated between two velocities, gives at most one, so the local inverses produced by the implicit function theorem are restrictions of a single global rather than competing branches. Only then does the cancellation behind work everywhere: it uses , the defining property of the one maximizer. The canonical equations later re-express the original stationary equation; none of these steps compares actions of neighboring curves.
Common mistake
(Reading the curvature bound in the wrong direction)
The upper bound on is worth reading in the correct direction. A lower bound on the curvature of gives an upper bound on the curvature of its conjugate. It does not give a uniform positive lower bound on without additional hypotheses. For example has inverse curvature , which tends to zero at large momenta.
With a single smooth Hamiltonian in hand, the conjugacy can also be read as an inequality.
Theorem 15.1.3
(The dual gap is nonnegative and has a precise equality case)
For every comparison velocity ,
The gap on the left is zero exactly when , equivalently . In particular the conjugacy recovers the original integrand by maximizing over momentum.
Proof
(Expand around the maximizing velocity)
Apply the strong convexity bound for with base velocity and tangent , then substitute . This proves the gap estimate. To recover at a fixed , the inequality gives an upper bound on , and choosing attains it. For the specific quadratic , the gap is exactly . The quantitative estimate is then an equality for every , while the gap itself is zero only at .
Counterexample mode
A positive Hessian alone need not cover every momentum
For , the Hessian is positive at every velocity, but takes only positive values. For the maximizing velocity is and . At the supremum is zero, approached as , and is not attained. For , letting makes .
Thus local invertibility does not guarantee a global velocity for every momentum. Uniform convexity is a convenient sufficient global hypothesis, not a claim about the weakest possible hypotheses for each individual integrand.
Canonical equations
Theorem 15.1.4
(Euler–Lagrange and the canonical system are equivalent)
Under the Legendre assumptions, a curve staying in on its interval in satisfies Euler–Lagrange if and only if its momentum forms a solution of
Conversely, any pair satisfying this system with has and the displayed momentum relation, so it gives a classical Euler–Lagrange curve. Along a canonical solution,
Thus an autonomous integrand has constant Hamiltonian along its extremals. The equivalence assumes an existing trajectory staying in ; it makes no general existence or boundary-value solvability claim and does not certify a minimum.
Proof
(Translate each equation through the inverse )
For a classical Euler–Lagrange curve, the inverse relation gives , and . Conversely, makes the canonical velocity satisfy . Because is and is , the equation for makes . The other canonical equation is then Euler–Lagrange. Finally the chain rule gives .
The phase-space stationary principle
On a fixed compact interval , consider pairs with prescribed endpoint values of and no endpoint restrictions on . The phase-space action is
For directions with , its first variation is
The boundary term is zero because , without fixing the endpoint values of . Momentum variations have no differentiated term and supply no additional endpoint condition. Independent compactly supported tests give the canonical equations; conversely those equations make every such first variation zero. Thus stationarity of this phase-space action is another precise variational formulation of the same equations. If a height at one end were wholly free, with no endpoint cost or other endpoint constraint, its unrestricted boundary displacement would instead require at that end.
The dual gap also gives for each pair. Equality holds exactly when the momentum relation is satisfied everywhere. For a curve , its momentum is and attains this equality. With the curve held fixed, the operation recovers the original density by maximization in momentum. It does not turn the phase-space action into a joint minimization problem or justify exchanging minimization and maximization.
Counterexample mode
A configuration minimum can become a phase-space saddle
Take and on with zero endpoint heights. Then has the unique global minimum . Its Hamiltonian is . The canonical equations and , together with the zero endpoint heights, force , so this is the unique stationary phase pair and .
Keeping and taking for a nonzero function gives . On the other hand, choose a nonzero function with zero endpoints, and take , . Both phase variables are , and . These pairs approach in the phase topology. The phase action has a saddle, despite strict global minimality of the original action.
The same inequality explains which half of the phase action carries a maximum.
Concept lensAnalytic
Eliminating momentum is a maximum at fixed velocity
The dual gap says that lies below and touches it at the correct momentum. The momentum variation therefore selects a maximum of that expression. The configuration variable still carries its own variational comparison. Treating both selections as minima reverses one half of the duality and obscures why the phase action can be a saddle.
Worked examples
Worked example 15.1.5
(A coupled quadratic system with two frequencies)
Take , and let
The eigenvalues of are and , so the theorem applies with . Here and
The canonical equations are and , equivalent to . The eigenvectors and have frequencies and . Decomposing the initial height into those two modes, the data , give
Direct substitution gives . Although the initial second height is zero, it generally becomes nonzero because both modes evolve with different frequencies. The momentum components are coupled to both velocity components through . Replacing momentum by velocity without that matrix would lose this two-frequency motion. The positive Hamiltonian is a conserved phase-space quantity; it is not by itself a certificate that every stationary arc minimizes its fixed-endpoint action.
Common mistake
(Writing the Hamiltonian in velocity variables)
In the coupled example the energy equals only after substituting , which gives . Writing instead would give , , hence , which is not the Euler–Lagrange equation .
Worked example 15.1.6
(A nonlinear state-dependent metric with an exact global minimum)
Take and consider . The velocity Hessian is , and
The canonical equations are and . Eliminating momentum gives .
To prove a minimum with prescribed , , , introduce
Since and , one has as and as . Thus is a smooth bijection of the real line with smooth inverse. For , the action becomes . Cauchy–Schwarz gives
Equality holds exactly when is affine. The unique global minimum is therefore
This curve is smooth and satisfies the Euler equation. Its Hamiltonian is the constant . The global comparison, rather than canonical stationarity alone, proved existence and uniqueness of this boundary-value minimum.
Worked example 15.1.7
(A conjugate without a closed form)
Take and , so and is strictly increasing onto . The theorem gives a Hamiltonian although is the real root of a cubic. At the root is , so
within the bound . The dual gap at is . At large momenta and : this edge case shows that the upper bound on is the only uniform one.
Practice
Quick check
(Compute the velocity conjugate of for every real momentum . Which momenta have a maximizing finite velocity, which have an unattained finite supremum, and which give an infinite supremum?)
The Hessian is positive everywhere, but the momentum map has a bounded range.
Solution
The length density has a restricted smooth momentum domain. The momentum is , with range . For , solving gives . The full conjugate is
For , every finite velocity gives , while taking velocity of matching sign and magnitude tending to infinity makes the expression tend to zero. Thus the supremum is zero and is not attained. For , the same sequence makes it tend to positive infinity. The smooth canonical momentum domain is the open interval, not its endpoints or the whole real line.
Quick check
(For , examine the momentum relation at and the global conjugate there. Prove that is not differentiable at zero without solving the cubic relation for general .)
Evaluate the supremum at the two velocities and . Distinguish local algebraic branches from a global maximizing velocity.
Solution
Several momentum branches meet a nonsmooth conjugate. Here , so at zero momentum the algebraic velocities are . The supremum is , attained at , whereas the branch would give . Thus solving the momentum equation alone does not select the conjugate maximizer.
The conjugate is finite for every fixed because at large velocity. Evaluating at and gives . If a derivative existed at zero, its right difference quotients would force it to be at least one, while its left difference quotients would force it to be at most minus one. This is impossible. The positive local Hessians at give separate local inverse branches, not a single smooth global Hamiltonian through zero momentum.
Quick check
(For scalar states in and , with real and , derive and the canonical equations. Recover the oscillator equation and its energy using the correct momentum.)
The term linear in velocity is a total derivative, but it changes the momentum coordinate.
Solution
A boundary derivative shifts momentum without changing the equation. The momentum is , so
The canonical system is and . Hence . Its classical curves are , with , and the conserved Hamiltonian is . The boundary derivative changes the canonical coordinate, while the correctly reconstructed oscillator equation and energy stay the same.