One function can compare many paths
A stationary equation describes an individual curve. A function of time and terminal position can organize an entire family of endpoint problems. If its derivatives satisfy a Hamilton–Jacobi equation, the chain rule along a comparison path turns a pointwise square into an action bound. Equality identifies the paths calibrated by that function.
This description requires care when the best path changes. Different starting points can reach the same terminal point with equal minimum cost. The value remains perfectly meaningful, but a derivative that would identify one preferred slope may cease to exist. We will prove both the smooth construction and an explicit example where smooth initial data produce this obstruction.
Initial cost and terminal value
Let be -Lipschitz, with . For and , define
We use absolutely continuous representatives of the paths. The initial point is free and pays the cost ; the final point is fixed. The variable is time along a path, while is its final time. This differs from fixing both spatial endpoints of one problem.
Definition 17.1.1
(A classical verification function)
For , a classical verification function is satisfying
Derivatives are continuous up to the time boundary, with one-sided time derivatives there. A calibrated path satisfies along its interval.
The Hamiltonian here is , the conjugate of the kinetic energy . The equation is nonlinear even though the minimizing paths for a fixed starting point will be straight.
Verification is an exact square identity
Theorem 17.1.2
(A smooth Hamilton–Jacobi function gives a lower bound)
If such an exists, then for every with ,
Consequently . If a calibrated path ending at exists, it attains the value and .
Proof
(Apply the chain rule along a comparison path)
The path has compact image, so the derivatives of are bounded along it. The absolutely continuous chain rule and the equation give, almost everywhere,
Integrate and use . This proves the identity. Equality holds precisely when the velocity agrees with the gradient almost everywhere; a calibrated path therefore attains the lower bound.
Proof X-ray
Keep the bound and its attainment separate
The identity has three inputs, each used once. The Hamilton–Jacobi equation replaces by , and only this substitution lets the chain-rule derivative be rewritten as a difference of squares: completing the square in is the key move. The initial condition is what allows the starting point to stay free, because the value of at is exactly the cost the path pays there, so no endpoint term is left over. Compactness of the path's image bounds along it and justifies the absolutely continuous chain rule. The lower bound then holds for every path, but attainment is a different statement: it needs a path whose velocity equals almost everywhere on all of . Continuity of alone guarantees neither such a solution on the whole interval nor its uniqueness, so constructing the path is part of the proof.
The inequality does not construct a calibrated path. Continuity of alone is not a global existence or uniqueness theorem for its terminal-value differential equation. In the smooth construction below, the required paths will be written explicitly.
Reduce the path problem to its starting point
Theorem 17.1.3
(The attained quadratic value formula)
For every and ,
Every minimizing path is a constant-velocity segment from a minimizing origin , and every such segment is minimizing. In particular,
Proof
(Minimize the kinetic energy first, then the origin)
For a path starting at , Cauchy–Schwarz gives
Equality requires constant velocity almost everywhere, hence the unique segment . Minimizing over yields the formula as an infimum. Since , its scalar objective tends to infinity as . Continuity then gives an attaining origin. The equality case also proves the stated classification of optimal paths.
Choosing proves the upper bound. For the lower bound, use and complete the scalar square in :
This formula is often called the Hopf–Lax formula for the free quadratic action. The proof here includes its attainment and admissible paths. Its bounds imply uniform convergence as , despite the possibility that itself is unbounded.
The value is also -Lipschitz in position. Given a minimizing origin for , shift that origin by when comparing with . The displacement in the quadratic cost is unchanged, while the initial cost changes by at most . Reversing the comparison gives .
Characteristics and the smooth interval
Suppose first that is and a stationary path is . Vary its initial point but keep its terminal point fixed. Integration by parts in its first variation gives the boundary coefficient and interior equation . Hence every such stationary path has the form
The last equation can have several solutions. Stationarity of the scalar origin cost is equivalent to it; global minimality is a separate question.
Common mistake
(Forgetting that the starting point is free)
Every segment ending at has zero acceleration, so the interior equation alone selects nothing. The origin is free and pays , and it is the natural condition that singles out the characteristics. Fixing instead would pose a different problem, one fixed-endpoint problem for each origin.
Theorem 17.1.4
(A noncrossing interval gives a classical solution)
Assume in addition that and for some . Define if , and if . For , the map
is a strictly increasing bijection of . Write its inverse as . The origin is the unique global minimizer, and
Thus is on this open time strip and solves . Its unique optimal path is , for .
The time is guaranteed by the chosen lower curvature bound; it is not asserted to be the largest possible classical existence time for every initial function.
Proof
(Invert the origin map and differentiate the attained minimum)
We have . Since a differentiable -Lipschitz function has , the difference is bounded by . Therefore is onto as well as strictly increasing. The implicit function theorem makes jointly .
The second derivative in of the scalar objective is . Its already attained minimum is consequently unique, at the solution of . The relation gives the displayed expression for the value.
When differentiating the original objective at , every term containing a derivative of is multiplied by . This gives and . Since these first derivatives are , is . Differentiating gives and the stated curvature formula. Also , so as the first derivatives extend continuously to and , locally uniformly in . Thus is up to the initial time and qualifies for verification on every closed strip ending before .
Finally, at earlier times the same origin satisfies . Hence along the explicit path, , with the value at supplied by . This constructs a calibrated path throughout the interval, not just an endpoint solution.
Proof X-ray
Where the curvature bound enters, and why the origin is never differentiated twice
The bound is used twice, with the same threshold : it makes positive, so the origin map is injective, and it makes the scalar objective strictly convex, so the stationary origin is the unique minimizer rather than one branch among several. Surjectivity needs a different hypothesis, the Lipschitz bound, through . The key move in computing and is that the derivatives of multiply the stationarity expression , which vanishes; is differentiated only for . The final paragraph is not decoration: it constructs the calibrated path that the verification theorem needs for attainment.
Before crossing, is and differentiating the PDE gives . Along the constructed characteristics, stays constant. The denominator in the curvature formula records a possible obstruction to continuing this smooth description.
Dynamic programming remains meaningful without derivatives
Proposition 17.1.5
(Dynamic programming for the quadratic value)
For and ,
and the minimum over is attained.
Proof
(Complete the square in the intermediate point)
Insert the origin formula for and minimize over the pair . For each fixed , completing a square gives
The intermediate point minimizing this expression is . The remaining minimum is exactly the origin formula at time . Both inner and outer minima are attained: the initial formula gives an attaining , and the displayed intermediate point gives . This is an infimum over a product, not an exchange of an infimum with a supremum.
The value also has a useful curvature inequality after differentiability fails, together with a two-sided control in time.
Proposition 17.1.6
(Semiconcavity and monotonicity in time)
For and ,
Proof
(An infimum of affine functions, then the value bounds restarted at time )
At fixed ,
is concave, since it is an infimum of affine functions of . Indeed, each affine function at a convex combination is at least the same combination of the two infima; taking its infimum preserves that inequality. Its second difference is therefore nonpositive, and the second difference of is exactly , which gives the first inequality. For the second, apply the earlier value bounds to the dynamic-programming formula, whose initial function is still -Lipschitz.
The first inequality is called semiconcavity with constant . It allows corners; it does not assert a classical second derivative. Thus the variational evolution remains controlled in both time and space without choosing a classical gradient at every point.
Concept lensGeometric
A minimum chooses among branches
The origin formula compares complete action values. The characteristic equation identifies stationary origins. While the origin map is one-to-one and the scalar objective strictly convex, these descriptions select the same unique path. After crossing, several stationary branches can arrive together and their costs must be compared. A corner in the value records that branch selection, rather than an ambiguity in the definition of the minimum.
Worked examples
Worked example 17.1.7
(Linear initial cost: every characteristic has the same slope)
Let for a real constant . The minimizing origin is , and
The formula is classical for every positive time because . Its gradient is , and the calibrated path has that constant velocity. The cost of the initial point and the kinetic action combine to the displayed value. A negative value is possible because the initial cost is not required to be nonnegative; the Lipschitz growth bound still keeps the minimum finite.
Worked example 17.1.8
(A smooth initial cost produces two minimizing origins)
Take . It is one-Lipschitz and satisfies . Before time one, the classical theorem applies. At , its minimizing origin is zero and
Now fix and consider the even scalar cost . On , the ratio decreases strictly from one to zero. For a proof, the numerator has derivative and starts at zero, making the derivative of the ratio negative. There is exactly one with .
Because , the cost decreases on and increases on . For , . Evenness now proves that its only global minima are and . Hence there are two optimal paths, the segments from those origins to zero. In the exact state , we have and ; the two optimal arrival velocities are and .
For a nearby terminal point , either minimizing origin at zero remains an admissible trial origin, so
If a spatial derivative at zero existed, the first inequality with would force it to be at most , whereas the second with would force it to be at least . This is impossible. Smooth initial data have led to a nondifferentiable value, while the attained minimum, spatial Lipschitz bound, dynamic programming and semiconcavity all remain valid.
Worked example 17.1.9
(Edge case: a concave corner that never smooths)
Let , which is -Lipschitz but not differentiable at zero. Take . On the branch the objective is minimized at , with value . On the branch the objective is , which is at least its minimum over all real , namely at ; this exceeds . By symmetry,
At the two origins both attain , with arrival velocities . The corner of persists for every positive time, is concave, consistent with semiconcavity, and away from zero . Contrast the convex corner in the practice problems below, which is smoothed at once into near the origin.
Common mistake
(Using the smooth formulas past the crossing time)
The formula is valid only while is a bijection. For at it gives , which blows up as ; for the value has a corner at , so neither nor exists there. Beyond the guaranteed interval the origin must be chosen by comparing actions, not by solving .
Counterexample mode
A surviving characteristic can have the wrong action
In the cosine example at and , the stationary origin still solves . Its constant path has action one and satisfies both the interior Euler equation and initial natural condition. Yet , and the two minimizing origins have strictly smaller action. Continuing this stationary characteristic does not continue the minimizing branch. The value itself is not differentiable at this terminal point, so its global classical verification gradient is unavailable.
Practice with solutions
Quick check
(Let . Find the minimizing origin and the value for every , ; decide whether the value is and for positive time, and verify the Hamilton–Jacobi equation wherever its first derivatives are defined.)
Minimize over the branches and separately.
Solution
The scalar objective is strictly convex. Minimizing its two quadratic branches gives, with ,
Inside the interval, and ; outside, and . The first derivatives agree at , so is jointly for and solves the equation there, including those interfaces. The spatial second derivative jumps from to zero, so it is not at the interfaces. The initial cost does not satisfy the smooth-initial-data theorem, but the general value formula and this direct calculation apply.
Quick check
(Replace the kinetic energy by , with real , keeping an -Lipschitz initial cost. Derive the value formula and the classical Hamilton–Jacobi equation, find the velocity that gives equality in verification, and compute the value for .)
Remove the drift with .
Solution
Set . Its initial point is unchanged, its terminal point is and its kinetic cost is the free cost of . Therefore
The Hamiltonian is , so a classical verification function satisfies . Completing the square along a path gives the gap integrand ; equality requires velocity . For linear initial cost,
The calibrated path has velocity . Shifting the spatial frame changes the transport term in the PDE and the selected velocity together.
Quick check
(Drop the Lipschitz hypothesis and take . Minimize the origin expression directly for , and , distinguishing an infinite negative infimum from a finite attained value.)
Watch the sign of the coefficient of .
Solution
The origin objective is
For it has unique minimum at , with value . At it becomes : the value is zero at , attained by every origin, while it is for . For the negative leading coefficient makes the infimum for every . Each fixed finite-origin path still has a finite action, so those negative infinite infima are not attained. This initial function is outside the global Lipschitz theorem; smoothness alone does not prevent the initial reward at distant origins from overwhelming the kinetic cost.
The continuation established here
For the quadratic action, the value formula defines a finite attained minimum globally in time under the stated growth assumption. A noncrossing origin map supplies a classical PDE solution, while competing minimizing origins can destroy differentiability. Dynamic programming and semiconcavity remain variational statements after that event. General Hamiltonians, state-dependent actions, viscosity-solution existence/comparison/uniqueness, and a full shock-structure theorem have not been proved here.