The missing step in an abstract existence theorem
The direct method asks for weak compactness and weak lower semicontinuity. For a squared derivative norm, the second requirement is immediate from Hilbert-space geometry. For an integral depending nonlinearly on both height and slope, it needs a proof.
In one dimension bounded derivative energy makes the heights converge uniformly along a subsequence. The slopes generally converge only weakly. We will use strong height convergence to control the height argument and convexity to handle the weak slope argument. Convexity in the height is not required. This distinction permits existence for nonconvex potentials, but it does not supply uniqueness.
The admissible space and growth conditions
Fix , , and endpoint heights . Use the absolutely continuous representatives in
The affine interpolant belongs to this class. Assume is on , where is an open interval containing . Its second and third arguments are the height and the slope .
Definition 7.2.1
(A quadratic-growth, slope-convex class)
Assume that is convex for each fixed , and that constants , give the following bound for all and all real :
For every assume there is a finite constant such that, whenever and ,
These local-in-height bounds hold for all real slopes. They ensure that the functional and its first variation are integrable on the stated class. They are explicit hypotheses, not consequences of convexity alone.
Every admissible height satisfies
Thus a derivative-bounded sequence has uniformly bounded heights and a common modulus of continuity. Arzelà–Ascoli gives a uniformly convergent subsequence. A weakly convergent subsequence of the derivatives identifies that limit as an function: pass to the limit in
using the test function . Uniform convergence retains both endpoint values.
Prove the lower-limit inequality
Lemma 7.2.2
(Lower-limit inequality under strong heights and weak slopes)
Under the hypotheses of the definition, if uniformly on and in , then .
Proof
(Freeze the height, then use a supporting line in the slope)
The weakly convergent derivatives are bounded in , and the heights lie in a common range . The mean value theorem in the height argument gives
For the remaining slope argument, the supporting-line inequality of a differentiable convex function gives, almost everywhere,
The coefficient belongs to by the slope-growth bound, so it is a legitimate fixed test for weak derivative convergence and the integral of the last term tends to zero. Integrating and taking the lower limit proves .
Proof X-ray
Use different convergence for the two arguments
The proof splits into a height change at fixed slope and a slope change at fixed height, and each piece is paired with the convergence it can bear. Uniform height convergence is paired with the growth estimate: the slopes need not converge pointwise or remain uniformly bounded, so the bound must be quadratic in and is integrated against the bounded . Weak slope convergence is paired with one fixed supporting slope evaluated at the limit; convexity is exactly what turns a nonlinear difference into a linear one, and only a linear expression passes to a weak limit. The inequality, not an equality, is all that survives, which is why the conclusion is a lower limit. No supporting plane in the pair was used, so joint convexity was never assumed. Replacing both steps by “the integrand is continuous” would incorrectly pass weak derivatives through a nonlinear function.
Theorem 7.2.3
(Existence for the stated height-dependent integral class)
Under the hypotheses above attains its infimum on . A minimizing sequence has a subsequence whose heights converge uniformly and whose derivatives converge weakly in to a minimizing height.
Proof
(Coercive bound, compactness, and the lemma)
The affine comparison has finite energy and the lower bound makes the infimum finite. Choose a sequence approaching it. The same bound,
makes the sequence derivative-bounded. The compactness argument of the previous section supplies an admissible limit, and the lower-limit inequality gives minimality. This constructs a minimizer before solving any Euler equation.
The lemma also proves sequential weak lower semicontinuity on the endpoint class . For a general weakly convergent sequence, take a subsequence realizing its lower limit and then its uniformly convergent subsequence. Its limit is the same weak limit, and the lemma gives the required conclusion for the original lower limit.
Common mistake
(Applying before the heights are bounded)
The growth constants are only local in height. Before using one must know that all heights of the sequence lie in one range ; here that comes from the derivative bound and . Without it, the estimate for the height change has no uniform constant.
A minimizer has an absolutely continuous momentum
For any , the curve remains in the endpoint class. Its heights stay in a bounded range for small . Differentiation under the integral gives
To justify it, the height term is dominated by a constant times . The slope term is dominated by a constant times . Both are . Dominated convergence therefore applies, without assuming has a classical second derivative.
Put
The identity says in distributions. Consequently has the absolutely continuous representative
For a direct justification, subtract the displayed primitive from . Its integral against is zero for all smooth compactly supported . Every smooth compactly supported zero-integral test is such a derivative; subtracting a fixed unit-integral test then shows that the difference is constant almost everywhere. This is an integrated Euler equation, not a formal differentiation of a merely measurable slope.
Common mistake
(Differentiating the momentum pointwise too early)
For an minimizer, is only an class, so writing pointwise is not yet meaningful. What the first variation gives is the integrated identity above; pointwise differentiation becomes legitimate only after a regularity argument such as the one that follows.
When the weak minimum is a classical extremal
Theorem 7.2.4
(A nondegenerate slope Hessian upgrades the minimizer to )
Assume additionally is and, for each bounded height range, there is such that
Every minimizer in is on and satisfies
classically, with the prescribed endpoints.
Proof
(Invert the momentum, then bootstrap twice)
For fixed , the function is strictly increasing onto the whole real line. Its derivative is at least , so it tends to positive and negative infinity at the respective ends. It has an inverse , continuous and locally by the implicit function theorem. In particular,
The minimizing height is continuous and its momentum representative is continuous, so is continuous. It equals the weak derivative almost everywhere. Integrating this equality first proves . Now is continuous, so . Composing the inverse with makes of class , hence , and holds pointwise.
Proof X-ray
One Hessian bound, used twice, uniformly in the slope
The proof never differentiates ; it starts from the integrated identity, which already makes the momentum continuous although is only an class. The bound is then used twice. It makes a bijection of , and uniformity over all slopes matters because is not yet known to be bounded. It also supplies the nonzero derivative that the implicit function theorem needs for a inverse . Each pass of the bootstrap gains one derivative: continuous gives continuous , hence continuous , hence and . In the degenerate-Hessian exercise below the inverse still exists and is continuous, but it is not differentiable at zero, so the chain stops at .
This is a regularity conclusion from the weak minimum, not an assumption used in its existence proof, and the two-step argument avoids presupposing the regularity it is intended to prove.
Concept lensAnalytic
Existence, regularity and uniqueness need separate arguments
Coercivity and slope convexity produce an minimum. The momentum identity gives additional regularity, and a nondegenerate inverse upgrades it to a classical curve. Uniqueness needs another property, such as suitable joint convexity. None of these conclusions should be borrowed from a different step without its hypotheses.
The three examples below test each of these separations in turn, starting from the simplest case.
Worked example 7.2.5
(Edge case: no height dependence)
Take on with endpoints . All hypotheses hold with , , and . The momentum is and , so the integrated Euler equation gives almost everywhere. The endpoints force , the affine interpolant, with . Uniqueness here comes for free: any admissible with has , because .
Worked example 7.2.6
(A nonlinear slope law can have a classical minimum without an explicit solution)
On fix , and take
The integrand is bounded below by . On each bounded height range its value and have the required quadratic bounds, while
Thus a globally minimizing curve exists. It solves
with the stated endpoints. The proof does not need a closed-form solution of this nonlinear boundary-value equation. It also does not assert uniqueness or that every solution of this equation minimizes.
Worked example 7.2.7
(Strict convexity in slope does not guarantee a unique minimum)
Take zero endpoints on and
Every hypothesis holds, including , so at least one classical global minimum exists. Yet for it cannot be the zero curve. The Taylor formula through degree three has fourth-derivative remainder, giving for between zero and . Since , this proves . Therefore the trial satisfies
At , the right side is . Since , a minimizer is nonzero. The functional is even in , so its negative is a distinct minimizer. The supporting-line proof required slope convexity only; it never promised uniqueness for a nonconvex height dependence.
Counterexample mode
A smooth quadratic-growth integrand can still fail weak lower semicontinuity
Could the growth bounds alone give the lemma? On with zero endpoints take
It is smooth and nonnegative, has quadratic growth and a derivative of at most linear growth. But , so it is not convex. Let have slope and then on the two halves of each of equal periods, starting at zero in each period. Its height is at most , and its endpoint is zero.
The heights converge uniformly to zero and the derivatives converge weakly to zero in . To verify the latter, their zero mean on each period makes the integral against a continuous test tend to zero by uniform continuity; density of continuous functions and the uniform derivative bound give the result for every test. Thus converges weakly in to zero.
Nevertheless and . All the stated growth bounds remain available, but slope convexity is missing, and the weak lower-limit inequality fails. This example shows failure of that proof hypothesis, not nonexistence for every nonconvex integrand: already the single tent attains the minimum zero.
Practice with solutions
Exercise 7.2.8
(For fixed endpoint heights and , let . Verify existence, prove the minimizer is , and derive a quantitative energy gap proving uniqueness.)
SynthesisProof
Solution(4)
- 1
Check the existence hypotheses
Extend linearly beyond each endpoint using its endpoint value and derivative, giving a extension to an open interval. Completing the height square gives , while . All local growth hypotheses hold and , so the existence theorem applies.
- 2
Upgrade regularity through the momentum
Here and its weak derivative is . The primitive makes absolutely continuous, so ; continuity of and then makes continuous, hence and .
Because: The primitive representation of the momentum.
- 3
Expand the energy and use the weak equation
Key step
For any endpoint-zero , the weak equation cancels the linear terms, leaving
Because: is convex, so it lies above its tangent line.
- 4
Conclude uniqueness
∎Equality of objective values forces . The additional height convexity, not slope convexity alone, has given uniqueness.
Exercise 7.2.9
(For the zero-endpoint cosine-potential problem above, prove that zero is the unique global minimum when , and explain why the endpoint case needs more than a strictly positive quadratic estimate.)
CoreProof
Solution(4)
- 1
Compare with a quadratic
Because: and the fixed-endpoint Poincaré inequality .
- 2
Treat
The right side is strictly positive for nonzero , so zero is the unique minimum.
- 3
Treat the threshold
Key step
The quadratic bound is only semidefinite. But has and with the sign of , so for . A nonzero continuous is nonzero on an interval, so .
- 4
Conclude
∎for . Thus zero is still the unique minimum at the threshold.
Exercise 7.2.10
(On with zero endpoints put , and . With the inverse of , show that is the unique global minimum and is but not . Identify exactly which regularity hypothesis fails.)
SynthesisProof
Solution(4)
- 1
Verify the existence hypotheses
The function is convex and bounded below by , and . The local growth bounds follow from
together with .
Because: for .
- 2
Construct the candidate
is continuous, odd and strictly increasing with limits , so its inverse is continuous and odd. Oddness gives . Also is positive left of zero and negative right of it, and because . Consequently and .
- 3
Verify the Euler identity and certify the minimum
Key step
The momentum is , so . Integrating the tangent inequalities of and against this identity shows for every endpoint-zero . Strict convexity of forces equality only when the slopes agree, and the endpoints then force the same curve.
Because: Tangent inequalities of the convex functions and .
- 4
Locate the failure of regularity
∎Near zero , hence , so is continuous but not differentiable at zero. The integrand is and satisfies the existence assumptions, but : the positive uniform lower Hessian bound, not existence or uniqueness, is the missing hypothesis of the theorem.
Exact reach of the argument
This proves a scalar, one-dimensional quadratic-growth existence theorem with explicit differentiable growth bounds, plus a classical regularity upgrade under a separate velocity-Hessian condition. It does not prove every version of Tonelli's theorem, general -growth or nonsmooth-integrand results, multidimensional quasiconvex lower semicontinuity, constrained existence or regularity under degenerate Hessians. Convexity in height was unnecessary for existence, and uniqueness was proved only where a separate comparison supplied it.