arXiv:math/0006015v1 [math.

AP] 2 Jun 2000

Uniqueness of solutions to Hamilton-Jacobi equations

arising in the Calculus of Variations
Gianni Dal Maso∗ and Hélène Frankowska†

Abstract.
We prove the uniqueness of the viscosity solution to the Hamilton-Jacobi equation
associated with a Bolza problem of the Calculus of Variations, assuming that the
Lagrangian is autonomous, continuous, superlinear, and satisfies the usual convexity
hypothesis. Under the same assumptions we prove also the uniqueness, in a class
of lower semicontinuous functions, of a slightly different notion of solution, where
classical derivatives are replaced only by subdifferentials. These results follow from
a new comparison theorem for lower semicontinuous viscosity supersolutions of the
Hamilton-Jacobi equation, that is proved in the general case of lower semicontinuous
Lagrangians.

Key words. Discontinuous Lagrangians, Hamilton-Jacobi equations, viability the-

ory, viscosity solutions.

AMS-MOS subject classification: 49L20 (primary), 49L25 (secondary).

1 Introduction
Let us consider a Bolza problem of the Calculus of Variations
Z t 
(1.1) min L(y(s), y ′ (s))ds + ϕ(y(t)) : y ∈ W 1,1 (0, t; IRn ), y(0) = x ,
0

where the final cost ϕ: IRn 7→ IR+ ∪ {+∞} is lower semicontinuous and the La-
grangian L: IRn × IRn 7→ IR+ is locally bounded and lower semicontinuous. We
assume also that L(x, ·) is convex for every x ∈ IRn and that the following Tonelli
∗
SISSA, via Beirut 2-4, 34014 Trieste, Italy
†
CNRS, ERS2064, Centre de Recherche Viabilité, Jeux, Contrôle, Université de Paris-Dauphine,
75775 Paris Cx 16, France

1
type coercivity assumption is satisfied: there exists a function Θ: IRn 7→ IR+ such
that
Θ(u)
(1.2) lim = +∞ , ∀ (x, u) ∈ IRn × IRn , L(x, u) ≥ Θ(u) .
|u|→∞ |u|

These assumptions guarantee the existence of absolutely continuous minimizers (see,

e.g., [7]). The classical Lagrange problem (with the fixed final condition y(t) =
z) may be reduced to the form (1.1) by simply setting ϕ(z) := 0 and ϕ := +∞
elsewhere.
The value function V : IR+ × IRn 7→ IR+ ∪ {+∞} for the Bolza problem (1.1) is
defined by
Z t 
V (t, x) := min L(y(s), y ′ (s))ds + ϕ(y(t)) : y ∈ W 1,1 (0, t; IRn ), y(0) = x .
0

Under our assumptions on L and ϕ the value function is lower semicontinuous on

IR+ × IRn and locally Lipschitz on IR+
⋆ × IRn , where IR⋆ := IR \ {0}. Moreover it
+ +
satisfies the initial condition

∀ x ∈ IRn , lim inf V (h, y) = ϕ(x) .

h→0+, y→x

If V is smooth and L is continuous, then V is a classical solution to the Hamilton-

Jacobi equation
⋆ n
 Vt + H(x, −Vx ) = 0 in IR+ × IR ,


(1.3)
V (0, ·) = ϕ in IRn ,


with Hamiltonian H: IRn × IRn 7→ IR defined by

(1.4) H(x, p) := sup (hp, ui − L(x, u)) ,

u∈IRn

where h·, ·i denotes the scalar product in IRn . In other words, H(x, ·) is the Legendre-
Fenchel conjugate of L(x, ·).
It is well known that (1.3) may not have smooth solutions, even if H and ϕ
are smooth. To overcome this difficulty, different notions of generalized solutions
have been proposed. The notion of viscosity solution can be introduced by means
of subdifferentials and superdifferentials.
We recall that the subdifferential ∂− W (x) of a function W : IRn 7→ IR ∪ {+∞} at
a point x ∈ dom(W ) is defined by

W (y) − W (x) − hp, y − xi

 
∂− W (x) := p ∈ IRn : lim inf ≥0 .
y→x |y − x|

2
while the superdifferential ∂+ W (x) is defined by ∂+ W (x) := −∂− (−W )(x).
A continuous function W : IR+⋆ × IRn 7→ IR is said to be a viscosity solution to the

Hamilton-Jacobi equation (1.3) if the following conditions are satisfied:

⋆
(1.5) ∀ (t, x) ∈ IR+ × IRn , ∀ (pt , px ) ∈ ∂− W (t, x), pt + H(x, −px ) ≥ 0 ,
⋆
(1.6) ∀ (t, x) ∈ IR+ × IRn , ∀ (pt , px ) ∈ ∂+ W (t, x), pt + H(x, −px ) ≤ 0 .
In [8] and [9] the uniqueness of a bounded uniformly continuous viscosity solution of
(1.3) is proved under some assumptions on H, which imply in particular that H is
continuous and H(·, p) is uniformly continuous for every p ∈ IRn . In [10, Theorem
4.5] we proved the following result, that can be applied also to unbounded solutions.

Theorem 1.1 Assume that L is continuous. Let W : IR+ ⋆ × IRn 7→ IR be a locally

+
Lipschitz viscosity solution of (1.3) which satisfies the initial condition

(1.7) ∀ x ∈ IRn , lim inf W (h, y) = ϕ(x) .

h→0+,y→x

⋆ × IRn .
Then W = V on IR+

To describe the new uniqueness results, we introduce the notion of contingent

directional derivatives of a function W : IRn 7→ IR ∪ {+∞}. These are defined, for
every x ∈ dom(W ) and for every u ∈ IRn , by

W (x + hv) − W (x)
(1.8) D↑ W (x)(u) := lim inf .
h→0+ h
v→u

The main result of this paper is the following theorem, which shows that the
value function V is the unique viscosity solution in the larger class of continuous
functions whose contingent derivatives satisfy the following very weak assumptions:

(1.9) ∀ (t, x) ∈ dom(W ), t > 0, D↑ W (t, x)(0, 0) = 0 ,

(1.10) ∀ (t, x) ∈ dom(W ), t > 0, ∃ u ∈ IRn , D↑ W (t, x)(−1, u) < +∞ .

Theorem 1.2 Assume that L is continuous. Let W : IR+ ⋆ × IRn 7→ IR be a continu-

+
ous viscosity solution of (1.3) which satisfies (1.9), (1.10), and the initial condition
(1.7). Then W = V on IR+ ⋆ × IRn .

3
 In [10, Theorem 4.4] we considered also a different notion of generalized solu-
tion and we proved the following uniqueness result in the class of locally Lipschitz
functions.
Theorem 1.3 Assume that L is continuous. Let W : IR+ ⋆ × IRn 7→ IR be a locally
+
Lipschitz function which satisfies the initial condition (1.7) and solves the Hamilton-
Jacobi equation (1.3) in the following sense :
⋆
(1.11) ∀ (t, x) ∈ IR+ × IRn , ∀ (pt , px ) ∈ ∂− W (t, x), pt + H(x, −px ) = 0 .
⋆ × IRn .
Then W = V on IR+
In this paper we shall prove the following result, which provides uniqueness in
the larger class of lower semicontinuous functions W satisfying (1.9) and (1.10).
Theorem 1.4 Assume that L is continuous. Let W : IR+ ⋆ × IRn 7→ IR ∪ {+∞}
+
be a lower semicontinuous function which satisfies the initial condition (1.7), the
technical conditions (1.9) and (1.10), and solves the Hamilton-Jacobi equation (1.3)
in the sense of (1.11). Then W = V on IR+ ⋆ × IRn .

In the proof of Theorem 1.4 we use the following comparison result, which follows
immediately from [10, Theorem 4.1, Remark 4.2 and Proposition 4.3].
Theorem 1.5 Assume that L is continuous. Let W : IR+ ⋆ × IRn 7→ IR ∪ {+∞} be
+
a lower semicontinuous function which satisfies the initial condition (1.7). Suppose
that W is a subsolution of the Hamilton-Jacobi equation (1.3) in the following sense:
⋆
(1.12) ∀ (t, x) ∈ IR+ × IRn , ∀ (pt , px ) ∈ ∂− W (t, x), pt + H(x, −px ) ≤ 0
⋆ × IRn .
Then W ≤ V on IR+
In the proof of Theorems 1.2 and 1.5 we use also a very general comparison result
(Theorem 1.7) for lower semicontinuous viscosity supersolutions of the Hamilton-
Jacobi equation (1.3). To our knowledge the strongest result in this direction,
dealing with possibly discontinuous Lagrangians, is the following theorem proved
in [10, Theorem 5.1], where the notion of supersolution is given by using contingent
inequalities.
Theorem 1.6 Let W : IR+ × IRn 7→ IR+ ∪ {+∞} be a lower semicontinuous function
which satisfies the initial condition W (0, ·) = ϕ. Suppose that W is a supersolution
of (1.3) in the following sense :
∀ (t, x) ∈ dom(W ), t > 0, ∃ u ∈ IRn , D↑ W (t, x)(−1, u) ≤ −L(x, u) .
Then W ≥ V on IR+ × IRn .

4
 The comparison result for viscosity supersolutions we are going to prove is the
following theorem, where we need the additional assumptions (1.9) and (1.10).

Theorem 1.7 Let W : IR+ × IRn 7→ IR+ ∪ {+∞} be a lower semicontinuous function
which satisfies (1.9) and (1.10). Suppose that W (0, ·) = ϕ and that W is a viscosity
supersolution of the Hamilton-Jacobi equation (1.3), i.e., W satisfies (1.5). Then
W ≥ V on IR+ × IRn .

2 Preliminaries
Let K ⊂ IRn be a nonempty subset and x ∈ K. The contingent cone TK (x) to K at
x is defined by
dist(x + hv, K)
v ∈ TK (x) ⇐⇒ lim inf = 0.
h→0+ h
The negative polar cone T − to a subset T ⊂ IRn is given by

T − := {v ∈ IRn : ∀ w ∈ T , hv, wi ≤ 0} .

When K is convex , then [TK (x)]− coincides with the the usual normal cone NK (x)
of convex analysis.
The epigraph Epi(W ) of a function W : IRn 7→ IR ∪ {+∞} is defined by

Epi(W ) := {(x, r) ∈ IRn × IR : r ≥ W (x)} .

We shall need the following version of Rockafellar’s result (see [12]).

h i−
Lemma 2.1 Let x ∈ dom(W ) and let (p, 0) ∈ TEpi(W ) (x, W (x)) be such that
p 6= 0. Then there exist xε converging to x (as ε → 0+) and
h i−
(pε , qε ) ∈ TEpi(W ) (xε , W (xε ))

converging to (p, 0) as ε → 0+ such that qε < 0 for every ε > 0.

A closed subset K of IRn is called a viability domain of a set-valued map G: K ❀
IRn if for every x ∈ K
G(x) ∩ TK (x) 6= ∅ .
The following statement summarizes several versions of the viability theorem (see [2]).

Theorem 2.2 (Viability) Let K ⊂ IRn be a closed set and let G: K ❀ IRn be an
upper semicontinuous set-valued map with compact convex values. The following
conditions are equivalent:

5
(a) K is a viability domain of G;

(b) G(x) ∩ co TK (x) 6= ∅ for every x ∈ K;

(c) for every x ∈ K there exist ε > 0 and a solution y: [0, ε[ 7→ K to the Cauchy
problem
y ′ (t) ∈ G(y(t)) ,

(2.1)
y(0) = x .

The equivalence (a) ⇐⇒ (b) was proved in [11]. This proof was simplified in [2,
page 85]. The fact that (a) ⇐⇒ (c) was first proved by Bebernes and Schuur in [6].
A proof can be found in [3] or [2].

The next theorem allows to deal with some unbounded set-valued maps with
closed convex values. As usual, BR denotes the closed ball with centre 0 and radius
R.

Theorem 2.3 Let K ⊂ IRn be a closed set and let G: K ❀ IRn be an upper semicon-
tinuous set-valued map with closed convex values. We assume that for every x ∈ K
there exists R > 0 such that for all small h > 0

dist(x + hG(x), K) = dist(x + h(G(x) ∩ BR ), K) .

Then the following statements are equivalent:

(a) K is a viability domain of G;

(b) for every x ∈ K and for every p ∈ [TK (x)]− we have inf hp, ui ≤ 0.
u∈G(x)

Theorem 2.3 is a direct consequence of the following more technical result, which
will be crucial in the proof of Theorem 1.7.

Theorem 2.4 Let K ⊂ IRn be a closed set, let x ∈ K, and let G: K ❀ IRn be a
set-valued map with non-empty closed convex values. Assume that there exists R > 0
such that for all small h > 0

(2.2) dist(x + hG(x), K) = dist(x + h(G(x) ∩ BR ), K) .

Assume also that the support function, defined by

σ(y, p) := sup hp, ui ,

u∈G(y)

6
satisfies the following upper semicontinuity condition at x: for every ε > 0 there
exists a neighbourhood U of x such that

(2.3) σ(x, p) + ε|p| > σ(y, p)

for every y ∈ U ∩ K and for every p in the set

NG(x) (BR ) := {p ∈ IRn : ∃ u ∈ G(x) ∩ BR , p ∈ NG(x) (u)} .

Finally, assume that for every y ∈ K in a suitable neighbourhood of x we have

(2.4) σ(y, −p) ≥ 0

for every p ∈ [TK (y)]− such that −p ∈ NG(x) (BR ). Then G(x) ∩ TK (x) ∩ BR 6= ∅.

From the proof given below it follows that the same result holds if [TK (y)]− is
replaced by the set of proximal normals to K at y.

Proof — Let us define the function g: IR+ 7→ IR+ by

1
g(h) := dist(x + hG(x), K)2 .
2
Observe that g is locally Lipschitz around zero and g(0) = 0. For all small h > 0 let
us consider uh ∈ G(x)∩BR and xh ∈ K such that dist(x+hG(x), K) = |x+huh −xh |.
Then xh → x when h → 0+.
Consider h > 0 such that g′ (h) does exist and fix any u ∈ G(x). Since G(x) is
convex, for all nonnegative h, k we have (h + k)G(x) = hG(x) + kG(x). Therefore

1 |x + huh + ku − xh |2 − |x + huh − xh |2
g ′ (h) ≤ lim = hph , ui ,
2 k→0+ k
where ph := x + huh − xh . Consequently

(2.5) g′ (h) ≤ inf hph , ui = − sup h−ph , ui = −σ(x, −ph ) .

u∈G(x) u∈G(x)

As xh is a point of K with minimum distance from x + huh , the vector ph is a

proximal normal to K, therefore it belongs to [TK (xh )]− . On the other hand, uh is
the point of G(x) with minimum distance from (xh −x)/h. Thus −ph ∈ NG(x) (uh ) ⊂
NG(x) (BR ). By (2.4) we have

(2.6) − σ(xh , −ph ) ≤ 0 .

7
 By the uniform upper semicontinuity (2.3) of σ for every ε > 0 there exists
hε > 0 such that for 0 < h < hε

(2.7) − σ (x, −ph ) ≤ −σ (xh , −ph ) + |ph |ε ≤ −σ (xh , −ph ) + hRε ,

where the last inequality follows from the fact that |ph | ≤ hR. From (2.5), (2.6),
and (2.7) we obtain that g′ (h) ≤ hRε for every h ∈ (0, hε ) at which the derivative
g′ (h) exists.
Integrating g′ we deduce that for 0 < h < hε
Rh2
g(h) ≤ ε.
2
This implies that for 0 < h < hε
dist(x + huh , K) dist(x + hG(x), K) √
= ≤ Rε .
h h
Let hi → 0+ be a sequence such that uhi converges to some u ∈ G(x). From the
very definition of contingent cone we deduce that u ∈ TK (x). ✷

3 Proof of the comparison theorem

This section is devoted to the proof of the new comparison theorem for viscosity
supersolutions.

Proof of Theorem 1.7 — We first claim that H is locally bounded. Indeed for
all x ∈ IRn we have H(x, p) ≥ −L(x, 0), thus H is locally bounded from below. On
the other hand, H(x, p) ≤ Θ∗ (p), where Θ∗ denotes the Legendre-Fenchel conjugate
of Θ. Since the function Θ has a superlinear growth, the convex function Θ∗ takes
only finite values, so it is locally bounded. This shows that H is also locally bounded
from above.
Consequently, the function H(x, ·) is locally Lipschitz with respect to p, locally
uniformly with respect to x. By (1.2) H(·, p) is upper semicontinuous with respect
to x. These two properties together imply that for every x ∈ IRn , for every M > 0,
and for every ε > 0 there exists a neighbourhood U of x such that

(3.1) ∀ y ∈ U , ∀ p ∈ BM , H(x, p) + ε > H(y, p) .

Let us define the set-valued map G: IRn ❀ IR × IRn × IR with closed convex values
by
G(x) := {(−1, u, −L(x, u) − ρ) : ρ ≥ 0, u ∈ IRn } .

8
Let K := Epi(W ), let (t, x) ∈ dom(W ) with t > 0, and let z := (t, x, W (t, x)). We
want to show that all assumptions of Theorem 2.4 are satisfied (here z plays the role
of x). To prove (2.2) we show that there exists R > 0 such that for all small h > 0

∃ u ∈ IRn , |(−1, u, −L(x, u))| ≤ R ,

dist ((t, x, W (t, x)) + h(−1, u, −L(x, u)), K) = dist ((t, x, W (t, x)) + hG(x), K) .

As W ≥ 0, it is easy to see that for every h > 0 there exist uh ∈ IRn and zh =
(th , xh , rh ) ∈ IR × IRn × IR, with rh ≥ W (th , xh ), such that

|(t, x, W (t, x)) + h(−1, uh , −L(x, uh )) − (th , xh , rh )|

= dist((t, x, W (t, x)) + hG(x), K) ≤ h(L(x, 0) + 1) .

If W (t, x)−hL(x, uh )−rh > 0, by increasing rh we could make W (t, x)−hL(x, uh )−

rh = 0, which would contradict the definition of distance. Therefore W (t, x) −
hL(x, uh ) − rh ≤ 0, hence

hL(x, uh ) + W (th , xh ) − W (t, x) ≤ hL(x, uh ) + rh − W (t, x) ≤ h(L(x, 0) + 1) .

We claim that uh is bounded for all h > 0 small enough. Assume by contradiction
that there exist hi → 0+ such that |uhi | → ∞.

Case 1. Assume first that for a subsequence, still denoted by hi , we have hi |uhi | ≥
c for some c > 0. Since W (th , xh ) ≥ 0, we have

hi L(x, uhi ) − W (t, x) ≤ hi (L(x, 0) + 1) ;

dividing by hi |uhi | and taking the limit we get

L(x, uhi )
lim sup < +∞ ,
i→∞ |uhi |

which contradicts (1.2).

Case 2. It remains to consider the case hi |uhi | → 0. Since

max {|th + h − t|, |xh − x − huh |} ≤ h(L(x, 0) + 1),

we deduce that for some v ∈ IRn and for a subsequence, still denoted by hi , we have
th i − t xh i − x
lim =0 and lim = v.
i→∞ hi |uhi | i→∞ hi |uhi |

9
Furthermore,

hi L(x, uhi ) + W (thi , xhi ) − W (t, x) ≤ hi (L(x, 0) + 1) ;

dividing by hi |uhi | and taking the limit yields

W (thi , xhi ) − W (t, x)

lim = −∞ .
i→∞ hi |uhi |

Hence D↑ W (t, x)(0, v) = −∞, which implies D↑ W (t, x)(0, 0) = −∞ (see [5]). This
contradicts (1.9) and completes the proof of our claim.

Thus |uh | is uniformly bounded when h > 0 is small. As L is locally bounded,

there exists R > 0 such that |(−1, uh , −L(x, uh ))| ≤ R for all small h > 0.
Observe next that, if (pt , px , q) ∈ NG(x) (BR ), then q ≥ 0. Moreover, if q = 0,
then px = 0; if q > 0, then there exists u ∈ BR such that px /q ∈ ∂u L(x, u), the
subdifferential of L(x, ·) at u. As L is locally bounded, this implies that there exists
a constant M such that px /q ∈ BM for every (pt , px , q) ∈ NG(x) (BR ) with q > 0.
To prove (2.3) it is enough to show that for every ε > 0 there exists a neighbour-
hood U of x such that

(3.2) sup (−pt + hpx , ui − qL(x, u)) + ε|(pt , px , q)|

u∈IRn

> sup (−pt + hpx , ui − qL(y, u))

u∈IRn

for every y ∈ U and for every (pt , px , q) ∈ NG(x) (BR ). If q = 0, then px = 0 and
(3.2) is trivial. If q > 0, then (3.2) can be written as
px px
−pt + qH(x, ) + ε|(pt , px , q)| > −pt + qH(y, ) ,
q q

which follows easily from (3.1).

Let us check that (2.4) holds true. Fix (s, y, r) ∈ Epi(W ), with s > 0, and
h i−
(pt , px , q) ∈ TEpi(W ) (s, y, r) . Since (0, 0, 1) ∈ TEpi(W ) (s, y, r), we have q ≤ 0.
Therefore (2.4) is equivalent to

(3.3) sup (pt + h−px , ui + qL(y, u)) ≥ 0 .

u∈IRn
h i−
If q < 0, then (pt /|q|, px /|q|, −1) ∈ TEpi(W ) (s, y, r) , hence (pt /|q|, px /|q|) ∈
∂− W (s, y) (see [5, page 249]) and we deduce (3.3) from (1.5). If q = 0 and

10
p 6= 0, then the supremum in (3.3) is +∞. If q = 0 and px = 0, then (pt , 0, 0) ∈
hx i−
TEpi(W ) (s, y, r) . By (1.10) there exists u ∈ IRn such that z := D↑ W (s, y)(−1, u) <
+∞. As Epi(D↑ W (s, y)(·, ·)) = TEpi(W ) (s, y, W (s, y)), the vector (−1, u, z) belongs
to TEpi(W ) (s, y, W (s, y)), which is contained in TEpi(W ) (s, y, r). By the definition of
h i−
TEpi(W ) (s, y, r) we obtain −pt ≤ 0, which yields (3.3) when q = 0 and px = 0.
From Theorem 2.4 we deduce that
 
G(x) ∩ TEpi(W ) (t, x, W (t, x)) 6= ∅ .

As Epi(D↑ W (t, x)(·, ·)) = TEpi(W ) (t, x, W (t, x)), we obtain that

∃ u ∈ IRn , D↑ W (t, x)(−1, u) ≤ −L(x, u).

This and Theorem 1.6 imply that W ≥ V on IR+ × IRn . ✷

Remark 3.1 From the proof of Theorem 1.7 we see that condition (1.10) yields,
for t > 0, h i −
(pt , 0, 0) ∈ TEpi(W ) (t, x, W (t, x)) =⇒ pt ≥ 0 .
If, in addition, the subsolution inequality (1.12) is satisfied, and
h i−
(pt , 0, 0) ∈ TEpi(W ) (t, x, W (t, x)) ,

then by Rockafellar’s Lemma 2.1 there exist (ti , xi ) → (t, x) and (pit , pix , qi ) ∈
h i−
TEpi(W ) (ti , xi , W (ti , xi )) , with qi < 0, such that (pit , pix , qi ) → (pt , 0, 0). Then
(−pit /qi , −pix /qi ) ∈ ∂− W (ti , xi ) and from (1.12) we obtain −pit /qi − L(xi , 0) ≤ 0,
hence pit + qi L(xi , 0) ≤ 0. Taking the limit we get pt ≤ 0. This shows that (1.10)
and (1.12) together imply the following geometric condition for t > 0:
h i−
(pt , 0, 0) ∈ TEpi(W ) (t, x, W (t, x)) =⇒ pt = 0 .

4 Proofs of the uniqueness results

We begin with the proof of the uniqueness theorem for viscsity solutions.

Proof of Theorem 1.2 — By Theorem 1.7 we have W ≥ V . Recall that the

hypograph of W is defined by
⋆
Hyp(W ) := {(t, x, r) ∈ IR+ × IRn × IR : r ≤ W (t, x)} .

11
Define the closed set

K := Hyp(W ) ∪ (IR− × IRn × IR) .

We claim that for all (t, x, r) ∈ K

(4.1) ∀ u ∈ IRn , (−1, u, −L(x, u)) ∈ co TK (t, x, r) .

It is enough to prove it in the case t > 0 and r = W (t, x), since the other cases are
evident. Fix u ∈ IRn . Then by (1.6)

(4.2) ∀ (pt , px ) ∈ ∂+ W (t, x), pt + h−px , ui − L(x, u) ≤ 0 .

We want to prove that

h i−
(4.3) ∀ (−pt , −px , q) ∈ THyp(W ) (t, x, W (t, x)) pt + h−px , ui − qL(x, u) ≤ 0 .

When q > 0 we have (pt /q, px /q) ∈ ∂+ W (t, x), thus (4.3) follows from (4.2). By
h i−
Lemma 2.1, applied to −W , if (0, 0, 0) 6= (−pt , −px , 0) ∈ THyp(W ) (t, x, W (t, x)) ,
h i−
then for some (ti , xi ) → (t, x) and (−pit , −pix , qi ) ∈ THyp(W ) (ti , xi , W (ti , xi )) , with
qi > 0, we have (pit , pix , qi ) → (pt , px , 0). So

pit + h−pix , ui − qi L(xi , u) ≤ 0.

Taking the limit we get pt + h−px , ui ≤ 0, which concludes the proof of (4.3).
By the separation theorem, (4.1) follows from (4.3). Since the lower set-valued
limit of contingent cones is equal to Clarke’s tangent cone (see for instance [5]), from
⋆ × IRn ,
the continuity of L we deduce that, for all (t, x) ∈ IR+

(−1, u, −L(x, u)) ∈ CHyp(W ) (t, x, W (t, x)).

Fix ε > 0. Then it is not difficult to check that

 
∀ u ∈ IRn , (−1, u, −L(x, u) − ε) ∈ Int CHyp(W ) (t, x, W (t, x)) .

By [4, Proposition 13, p. 425] this yields

W (t − h, x + hv) − W (t, x)
(4.4) lim inf ≥ −L(x, u) − ε.
h→0+, v→u h
We have to show that for all t > 0, x ∈ IRn , V (t, x) ≥ W (t, x). Let y be a
minimizer of the Bolza problem (1.1). Since L is continuous, y ′ ∈ L∞ (0, t; IRn ) by

12
[1]. Consider a sequence of continuous functions ui : [0, t] 7→ IRn which is bounded
in L∞ (0, t; IRn ) and converges to y ′ almost everywhere in [0, t], and let ti → 0+ and
xi → y(t) be such that W (ti , xi ) → ϕ(y(t)). Define
Z t−ti
yi (s) := xi − ui (τ )dτ, ∀ s ∈ [0, t − ti ].
s

and yi (s) := xi for s > t − ti . Then yi converges to y uniformly in [0, t]. Fix i and
set ψ(s) := W (t − s, yi (s)) for s ∈ [0, t − ti ]. By (4.4) for every s ∈ [0, t − ti [ we have
ψ(s + h) − ψ(s)
(4.5) lim sup ≥ −L(yi (s), ui (s)) − ε.
h→0+ h
Consider the system
 ′ ′
 (α (s), z (s)) = (1, −L(yi (s), ui (s)) − ε), s ≥ 0

(α(0), z(0)) = (0, W (t, yi (0))) .



where we have set ui (s) := ui (t) for all s ≥ t − ti . It has the unique solution
 Z s 
(α(s), z(s)) := s, W (t, yi (0)) − L(yi (τ ), ui (τ ))dτ − εs .
0

According to Theorem 2.2 and (4.5), this solution is viable in Hyp(ψ)∪([t − ti , +∞[×
IR), i.e., for all s ∈ [0, t − ti ] we have (α(s), z(s)) ∈ Hyp(ψ). Thus for all s ∈ [0, t − ti ]
Z s
W (t − s, yi (s)) ≥ W (t, yi (0)) − L(yi (τ ), ui (τ ))dτ − εs .
0

In particular
Z t−ti
W (t, yi (0)) ≤ W (ti , xi ) + L(yi (τ ), ui (τ ))dτ + ε(t − ti ) .
0

Since the functions L and W are continuous, taking the limit we get
Z t
W (t, x) − εt ≤ ϕ(y(t)) + L(y(τ ), y ′ (τ ))dτ V (t, x) .
0

Finally, as ε → 0+ we obtain W (t, x) ≤ V (t, x). ✷

Proof of Theorem 1.4 — The inequality W ≥ V is proved in Theorem 1.7,

while the inequality W ≤ V follows from Theorem 1.5. ✷

13
References
[1] AMBROSIO L., ASCENZI O. & BUTTAZZO G. (1989) Lipschitz regularity
for minimizers of integral functionals with highly discontinuous integrands, J.
Math. Anal. Appl. 142, 301-316.

[2] AUBIN J.-P. (1991) Viability Theory, Birkhäuser, Boston.

[3] AUBIN J.-P. & CELLINA A. (1984) Differential Inclusions, Grundlehren

Math. Wiss. 264, Springer-Verlag, Berlin.

[4] AUBIN J.-P. & EKELAND I. (1984) Applied Nonlinear Analysis, Wiley
& Sons, New York.

[5] AUBIN J.-P. & FRANKOWSKA H. (1990) Set-Valued Analysis, Birkhäu-

ser, Boston.

[6] BEBERNES J.W. & SCHUUR J.D. (1970) The Ważewski topological method
for contingent equations, Ann. Mat. Pura Appl. 87, 271-280.

[7] CESARI L. (1983) Optimization Theory and Applications. Problems

with Ordinary Differential Equations, Appl. Math. 17, Springer-Verlag,
Berlin.

[8] CRANDALL M.G. & LIONS P.-L. (1983) Viscosity solutions of Hamilton-
Jacobi equations, Trans. Amer. Math. Soc. 277, 1-42.

[9] CRANDALL M.G. & LIONS P.-L. (1986) On existence and uniqueness of so-
lutions of Hamilton-Jacobi equations, Nonlinear Anal. 10, 342-370.

[10] DAL MASO G. & FRANKOWSKA (to appear) Value functions for Bolza prob-
lems with discontinuous Lagrangians and Hamilton-Jacobi inequalities, ESAIM
Control Optim. Calc. Var..

[11] GUSEINOV H.G., SUBBOTIN A.I. & USHAKOV V.N. (1985) Derivatives for
multivalued mappings with application to game-theoretical problems of control,
Problems Control and Inform. 14, 155-168.

[12] ROCKAFELLAR T. (1981) Proximal subgradients, marginal values and aug-

mented Lagrangians in nonconvex optimization, Math. Oper. Res. 6, 424-436.

14