\documentclass[leqno,11pt]{article}\usepackage{latexsym}\renewcommand{\theequation}{\thesection.\arabic{equation}}\makeatletter\@addtoreset{equation}{section}\makeatother\mathchardef\emptyset="001F\newcommand{\R}{I\hspace{-1.5mm}R}\newcommand{\scriptR}{I\hspace{-0.9mm}R}\newcommand{\N}{I\hspace{-1.4mm}N}\newcommand{\scriptN}{I\hspace{-0.8mm}N}\newtheorem{Theorem}{Theorem}[section]\newtheorem{Definition}[Theorem]{Definition}\newtheorem{Proposition}[Theorem]{Proposition}\newtheorem{Lemma}[Theorem]{Lemma}\newtheorem{Corollary}[Theorem]{Corollary}\newtheorem{Remark}[Theorem]{Remark}\textwidth=15cm\date{}\title{AutonomousIntegral Functionals with Discontinuous\\Nonconvex Integrands:Lipschitz Regularity of Minimizers,\\DuBois-Reymond Necessary Conditions,\\and Hamilton-Jacobi Equations}\author{Gianni Dal Maso\thanks{ {\it SISSA, via Beirut 2-4,34014 Trieste, Italy \hfill\break\null\hspace{2\parindent} e-mail:\/} {\tt dalmaso@sissa.it}}\hspace{.1cm}and H\'el\`ene Frankowska\thanks{ {\it CNRS, CREA, Ecole Polytechnique, 1,Rue Descartes, 75005 Paris, France \hfill\break\null\hspace{2\parindent} e-mail:\/}{\tt franko@poly.polytechnique.fr}}}\begin{document}\maketitle\begin{abstract}This paper is devoted to theautonomous Lagrange problem of the calculus of variations witha discontinuous Lagrangian. We prove that every minimizer isLipschitz continuous if the Lagrangian is coercive andlocally bounded. The main difference with respect tothe previous works in the literature isthat we do not assume thatthe Lagrangian is convex in the velocity. We also show that, under someadditional assumptions, the DuBois-Reymond necessary conditionstill holds in the discontinuous case. Finally, we applythese results to deduce that the value function of the Bolzaproblem is locally Lipschitz and satisfies (in a generalized sense)a Hamilton-Jacobi equation.\end{abstract}\vspace{ 5 mm}{\small\noindent {\bf Key words.} Discontinuous Lagrangians, nonconvexintegrands, Lipschitz minimizers, DuBois-Reymond necessary conditions,Hamilton-Jacobi equations.\vspace{ 5 mm}\noindent{\bf AMS Mathematics Subject Classification 2000:}  49N60 (primary), 49K05,49L25 (secondary).}\newpage\section{Introduction}\label{intro}In this paper we study the Lipschitz continuity of the solutionsto the Lagrange problem of the calculus ofvariations\begin{equation}\label{dm300}\min \left\{ \int_{a}^{b}L(y(t),y'(t))dt \;|\;y \inW^{1,1}(a,b; \R^n),\;\; y(a)=x_a,\;\; y(b)= x_b\right\},\end{equation}where the Lagrangian $L:\R^n \times \R^n \rightarrow \R_+$  is   a Borel function having a superlinear growth with respectto the second variable, i.e.,there exists a function $ \Theta \colon  \R^n \rightarrow \R_+ $,with\begin{equation}   \label{dm302}\lim_{|u| \rightarrow  \infty } \frac{\Theta (u)}{|u|} = + \infty,\end{equation}such that\begin{equation}    \label{dm304}\forall \; (x,u) \in \R^n \times \R^n, \;\; L (x,u) \geq \Theta(u) .\end{equation}We assume also that$L$ is bounded in a neighborhood of each point of $\R^n \times\{0\}$, i.e.,\begin{equation} \label{dm301}   \forall \; x_0\in\R^n,\; \exists\; M>0,\;\exists \; r>0,\; \forall \;(x,u) \in B (x_0,r) \times B (0,r),\; L ( x,u) \leq M,\end{equation}where $B (x_0,r)$ is the closed ball with center $x_0$ and radius~$r$.The existence of a solution to (\ref{dm300}) is an easy consequence ofthe direct method of the calculus of variations when the functional\begin{equation}\label{eq150}{\cal L}_a^b(y ):=\int_{a}^{b} L(y(t),y'(t))dt\end{equation}is  sequentially weakly lowersemicontinuous  on $W^{1,1}(a,b;\R^n)$.By the classical results of Olech \cite{Olech} and Ioffe\cite{Ioffe},a standard assumption for the semicontinuity of ${\cal L}_a^b$ is that$L$ is lower semicontinuous on $\R^n\times\R^n$ and $L(x,\cdot)$ isconvex on $\R^n$ for every $x\in \R^n$, but these conditions are notnecessary for the lower semicontinuity of ${\cal L}_a^b$ (see, e.g.,\cite{DG-But-DM}).Recently Amar, Bellettini and Venturini haveproved in \cite{amar} that any integral functional of the form(\ref{eq150}), satisfying suitable growth conditions,has a lower semicontinuous envelope $\overline{\cal L}{}_a^b$that can be writtenas\begin{equation}\label{calL}\overline{\cal L}{}_a^b(y)=\int_{a}^{b} L^+(y(t),y'(t))dt\,,\end{equation}where $L^+$ is an integrand depending on $L$(see (\ref{feb30})). If $L$ is continuous, then$L^+$ coincides with the convexification $L_0$ of $L$ with respect to $u$,but, if $L$ is discontinuous, one can prove only that thefunction $L^+(x,\cdot)$ is convex for a.e.\ $x\in\R^n$, and there areexamples where $L^+(\cdot,u)$ is not lower semicontinuous on $\R^n$.This shows that there are problems of the form(\ref{dm300}) which admit a solution even if $L$ is not convex in $u$(nor lower semicontinuous in $x$), and provides a motivation for thestudy of the Lipschitz continuity of the solutions of (\ref{dm300})without convexity hypotheses.If $L(x,\cdot)$ is convex for every $x\in\R^n$,it was proved by Ambrosio, Ascenzi, and Buttazzo in \cite{ambr}that every minimizer of (\ref{dm300}) is Lipschitz continuous. This kind ofresults goes back to Tonelli \cite{t1,t2} for smooth Lagrangians, andis the first step to prove, under some additional conditions on $L$,that all minimizers are smooth(see, e.g., \cite[Section 2.6]{ces83}). Note that, in general,when the Lagrangian is timedependent, the problem may have no Lipschitz minimizer (see\cite{bm85} and \cite{cla-vin}).The aim of Section~\ref{main} of the present paperis to show that the convexity hypothesis can be removedfrom \cite{ambr}. Assuming only (\ref{dm302}),  (\ref{dm304}), and(\ref{dm301}), we prove that all minimizers of (\ref{dm300})are still Lipschitz continuous (Theorem~\ref{dm306}), and provide anestimate of the Lipschitz constant if, in addition, $L$ is locally bounded(Theorem~\ref{dm306a}).  If $L$ is continuous,then every minimizer $y$ of (\ref{dm300}) is also a minimizer ofthe same problem with $L$ replaced by its convexification $L_0$ withrespect to $u$, so thatthe Lipschitz continuity of $y$ follows from \cite{ambr}. But, if$L$ is discontinuous,we can only say (under suitable growth conditions) that $y$ is a minimizerof (\ref{dm300}) with $L$ replaced by $L^+$, and we know that$L^+(x,\cdot)$ is convex only for a.e.\ $x\in\R^n$. For this reason we cannot apply the results of \cite{ambr}. On the other hand,the proof of \cite{ambr} isbased on an extensionof the DuBois-Reymondnecessary condition, which is not always validwhen $L(x,\cdot)$ is not convex. Therefore we need different arguments.As in \cite{ambr}, we begin by proving (Lemma~\ref{dm309})that if $y$ is a minimizer of(\ref{dm300}), then the function $\psi(t):=t$ is a minimizerof the problem\begin{displaymath}\min \left\{ \int_{a}^{b}f(t,\psi'(t))dt \;|\;\psi \inW^{1,1}(a,b),\;\; \psi(a)=a, \;\; \psi(b)=b\right\},\end{displaymath}where\begin{equation}\label{dm307} f (t,v):= \left\{ \begin{array}{lll} L  ( y(t), {y' (t)}/{v} )v & \mbox{ if}  &v >\frac{1}{2},\\\noalign{\vskip 3pt}  + \infty &  \mbox{ if}  &v \leq \frac{1}{2}.\end{array} \right.\end{equation}Then we show (Lemma~\ref{dm374}) that $\psi(t):=t$ is a minimizerof the problem\begin{displaymath}\min \left\{ \int_{a}^{b}f_0(t,\psi'(t))dt \;|\;\psi \inW^{1,1}(a,b),\;\; \psi(a)=a, \;\; \psi(b)=b\right\},\end{displaymath}where$f_0= \overline {\rm co}_v f$ is the lower semicontinuous convexenvelope of $f$ with respect to $v$.This implies (Lemma~\ref{dm376}) that there existsa constant  $c \in \R$ such that\begin{displaymath}d^{\,l}_vf_0 (t,1) \leq c \leq d^{\,r}_v f_0 (t,1) \;\; \mbox{for a.e.}\;\; t \in [ a,b],\end{displaymath}where $d^{\,l}_v$ and $d^{\,r}_v$ denote the left and rightderivatives with respect to $v$.These inequalities, together with (\ref{dm302}),  (\ref{dm304}), and(\ref{dm301}), are used to obtaina bound on the Lipschitz constant of $y$(Theorem~\ref{dm306}), which is locallyuniform (with respect to the data of the problem) if $L$ is locally bounded(Theorem~\ref{dm306a}).In Section~\ref{DuB} we obtainsome extensions of the DuBois-Reymondnecessary condition.When $L(x,\cdot)$ is not convexthis condition is not always satisfied, and wepropose some additional assumptions on $L$,which hold true, for instance, when $L(x,\cdot)$ is semiconvexor differentiable.Under these assumptionswe show (Theorems~\ref{dm3pr70} and~\ref{dm363})that, if $y$ is aminimizer,then thereexists a constant  $c \in \R$ such that\begin{displaymath}   c \in  L (y (t),y' (t))-   \left\langle   \partial _{u} ^-L (y (t),y' (t)), y' (t)\right\rangle \;\;\mbox{ for a.e.}\;\; t \in [ a,b],\end{displaymath}where $ \partial _{u} ^-L (y (t),y' (t))$ is the subdifferentialof $L(y(t), \cdot)$ at $y'(t)$. More general results of this kind(Lemma~\ref{dm363n} and Theorem~\ref{dm363x}) are obtained withdifferent generalized gradients of $L$.Finally, in Section~\ref{Hamilton} we apply the Lipschitz regularity ofminimizers to study the value function of the Bolza problem:\begin{equation} \label{value}   V(t,x):= \inf \left\{\int_{0}^{t} L (y (s), y' (s))ds + \varphi  (y(t)) \;|\; y\in W^{1,1}(0,t;\R^n),\  y (0)=x \right\} ,\end{equation}where $ \varphi \colon  \R^n \to \R_+ \cup \{+ \infty\}$, $\varphi \not\equiv+ \infty $, and $L$ is locally bounded, not necessarily convexwith respect to the second variable.Let $\R_+^{\star}:=\{t\in\R \;|\; t>0\}$ and let$H$ be the Hamiltonian associatedwith~$L$, defined by\begin{equation}\label{H}   H (x,p):= \sup_{u \in \scriptR^n} \left( \left\langle   p,u\right\rangle   -L(x,u) \right),\end{equation}i.e., $H (x, \cdot )$ is the Legendre-Fenchel transform of $L (x,\cdot )$.Assuming that for all $(t,x)\in \R_+^{\star} \times \R^n$ the infimum in (\ref{value}) is attained,we prove that $V$ is locally Lipschitz on $\R_+^{\star} \times\R^n$ (Theorem~\ref{liploc})and solves the Hamilton-Jacobi equation\begin{equation}\label{HJ}V_t + H(x,-V_x)=0\end{equation}in a generalized sense (Theorem~\ref{thm2}).When$\varphi$  is lower semicontinuous, we also provide a comparisonresult for lower semicontinuoussubsolutions of~(\ref{HJ}), whichcharacterizes thevalue function as the maximal lower semicontinuous subsolutionof~(\ref{HJ}) (Theorem~\ref{thm1}).We conclude the paper with two results(Theorems~\ref{dm391} and~\ref{dm391b}) which show the relationshipsbetween minimizers of (\ref{value}) and contingent derivatives of thevalue function.\section{Lipschitz Regularity of Minimizers}\label{main}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel function,let  $[a,b]$ be a bounded closed interval in $\R$, and let  $y \in W^{1,1}(a,b; \R^n)$ be a function such that\begin{equation}   \label{dm305}\int_{a}^{b}L (y (t), y' (t))dt \leq \inf_{z \in S(y)}\int_{a}^{b} L (z (t), z' (t))dt  < + \infty,\end{equation}where$S(y):=\{z \in W^{1,1} (a,b; \R^n) \; | \; z (a)= y (a),\;\;z (b)= y (b)\}$.The main results of this section are the following two theorems.\begin{Theorem} \label{dm306}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel functionwhich satisfies   (\ref{dm302}),  (\ref{dm304}), (\ref{dm301}), andlet  $y \in W^{1,1} (a,b; \R^n)$ be a function which satisfies(\ref{dm305}). Then $y$ is Lipschitz continuous.\end{Theorem}When $L$ is locally bounded on $\R^n \times \R^n $, we obtain auniform estimate of the Lipschitz constant of every minimizer.\begin{Theorem} \label{dm306a}Let $\Theta\colon  \R^n \to \R_+$ be a function satisfying(\ref{dm302}), let$\Psi \colon  \R_+ \to \R_+$ be a nondecreasing function, and let$A$, $B$, $\alpha$, $\beta >0$. Then there exists a constant $K=K(\Theta, \Psi, A,B, \alpha,\beta)$ with thefollowing property: if $L\colon \R^n \times \R^n \rightarrow \R_+$ isany Borel functionsatisfying (\ref{dm304}) and\begin{equation}\label{dm379}\forall \; R>0, \quad \sup_{|x| \leq R, \,|u|\le R} L(x,u) \leq \Psi(R),\end{equation}and $y \in W^{1,1} (a,b; \R^n)$ satisfies (\ref{dm305}) and\begin{eqnarray}   \label{dm369a}   &\displaystyle\vphantom{\int_{a}^{b}}\inf_{a \leq t \leq b} |y(t)| \leq A ,\\   \label{dm370a}&\displaystyle\int_{a}^{b} L(y(t),y'(t))dt \leq B ,\\   \label{dm371}   &\displaystyle\vphantom{\int_{a}^{b}}\alpha \leq  b-a \leq \beta ,\end{eqnarray}then $y$ is Lipschitz continuous with Lipschitz constant bounded by~$K$.\end{Theorem}To prove Theorems \ref{dm306} and \ref{dm306a}we need some technical lemmas.Let us fix a function $y\in W^{1,1}(a,b;\R^n)$which satisfies (\ref{dm305}).As in  \cite{ambr} we use the auxiliary function $f \colon [a,b]\times \R \to  [0, + \infty ]$ defined by (\ref{dm307}),which turns out to be  $ {\cal L}_1 \times {\cal B}_1$-measurable,where $ {\cal L}_1  $  and $ {\cal B}_1$ denote the$\sigma$-algebras of Lebesgue measurable subsets of  $[a,b]$ and ofBorel subsets of  $\R$, respectively. {}From(\ref{dm307})  and (\ref{dm305}) it follows that\begin{equation}   \label{dm308}\int_{a}^{b} f (t,1) dt < + \infty .\end{equation}The following lemma is well known(see, e.g., \cite[p.~46]{ces83}). We write the proof only to give aself-contained presentation of the arguments used to obtainTheorems \ref{dm306} and \ref{dm306a}.\begin{Lemma}\label{dm309}We have\begin{equation}   \label{dm310}\int_{a}^{b}f (t,1)dt \leq \int_{a}^{b}f (t, \psi ' (t))dt\end{equation}for every  $\psi \in W^{1,1} (a,b)$ such that  $ \psi (a)=a$ and  $\psi(b)=b$.\end{Lemma}\noindent{\bf Proof} --- \hspace{ 2 mm}Let us fix $\psi \in W^{1,1}(a,b)$, with  $ \psi (a)=a$ and  $\psi (b)=b$, such that the righthand side of   (\ref{dm310}) is finite. Then $\psi ' (t) >\frac{1}{2}$ for almost all  $ t \in [ a,b]$. Thus  $ \psi $ isincreasing and  $| \psi (t)- \psi  (s)| \geq \frac{1}{2} |t-s|$for all $s,t \in [ a,b]$. Therefore the inverse function $\psi^{-1} \colon  [ a,b] \to  [ a,b]$ is Lipschitz continuous with Lipschitzconstant 2. These properties imply that\begin{equation}   \label{dm311}N \in {\cal L}_1, \;\; |N|=0  \Longleftrightarrow \psi (N) \in{\cal L}_1,\;\; | \psi  (N)| =0 ,\end{equation}where  $|\cdot |$ denotes the Lebesgue measure, and that  $z\colon = y\circ \psi ^{-1}$ belongs to $W^{1,1} (a,b;\R^n)$ and satisfies $z(a)=y (a)$ and  $z (b)= y (b)$. Using  (\ref{dm311}) and the chainrule one proves easily that $z' (t)=y' ( \psi ^{-1} (t))/ \psi ' (\psi ^{-1} (t))$ for a.e.\ $t \in [ a,b]$. Thus, after the changeof variables  $s= \psi ^{-1}  (t)$, one gets\begin{displaymath}   \int_{a}^{b} L (z (t),z' (t))dt= \int_{a}^{b}L ( y (s), {y'(s)}/{ \psi ' (s)} ) \psi ' (s) ds ,\end{displaymath}which, together with  (\ref{dm305}) and  (\ref{dm307}), yields(\ref{dm310}).$\; \; \Box$\vspace{ 5 mm}If  $g\colon  \R \rightarrow  [ 0, + \infty ]$ is an arbitrary function, itslower semicontinuous convex envelope $ \overline {\rm co} \, g \colon  \R\rightarrow  [ 0, + \infty ]$ is, by definition,  the greatest lowersemicontinuous convex function which is less than or equal to  $g$. It iswell known that the epigraph of  $\overline {\rm co} \, g$ is the closed convexhull of the epigraph of  $g$, and that, if  $ \overline {\rm co} \,g$ is finitein a neighborhood of some point  $v \in \R$, then\begin{equation}   \label{dm312}\overline {\rm co}\, g (v)= \inf \left\{ \lambda g(v_1) + (1-\lambda ) g (v_2)\;|\;\;  ( \lambda ,v_1, v_2) \in A (v) \right\},\end{equation}where  $A (v)$ is the set of all triples $ ( \lambda ,v_1, v_2) \in \R^3$with $ 0 < \lambda  <1,\; g (v_1) < + \infty , \; g (v_2) < + \infty ,$ and$v= \lambda v_1+ (1- \lambda )v_2$.Let us return to the function  $f$ defined in  (\ref{dm307}), andlet $f_0= \overline {\rm co}_v f$ be its lower semicontinuousconvex envelope with respect to $v$. We observe that for every  $t\in [a,b]$, the function $v \mapsto  f_0 (t,v)$ is continuous on$( \frac{1}{2}, + \infty )$, since it is convex and finite on thisset.Let us prove that for every $v \in ( \frac{1}{2}, + \infty)$ thefunction $t \mapsto f_0 (t,v)$ is Lebesgue measurable. Given  $\gamma \in \R$, by  (\ref{dm312}) the set $\{t \in [ a,b] \;| \;f_0 (t,v) < \gamma \}$ is the projection onto  $ [ a,b]$ of theset of all points $ (t, \lambda , v_1,v_2) \in [ a,b] \times (0,1)\times ( \frac{1}{2}, + \infty ) \times ( \frac{1}{2}, + \infty )$such that $v = \lambda v_1 + (1 - \lambda )v_2$ and $ \lambda f(t,v_1) + (1- \lambda ) f (t,v_2) < \gamma $. As this set belongsto the $ \sigma$-algebra  $ {\cal L} _1 \times {\cal B}_1\times{\cal B}_1\times {\cal B}_1$, from the projection theorem (see,e.g., \cite[Theorem 8.3.2]{af90sva}) we conclude that $ \{t \in [a,b]\;|\; f_0 (t,v) < \gamma \}$ is Lebesgue measurable. Thisproves that $t \to f_0 (t,v)$ is Lebesgue measurable, and hence$f_0$ is a Carath\'{e}odory function on  $ [ a,b] \times (\frac{1}{2}, + \infty )$.\\vspace{ 5 mm}The following lemma is usually proved when $f$ is continuous in $v$, orsatisfies some growth condition.  We give here a detailed proofto show that we do not need any additional hypothesis.\begin{Lemma}\label{dm374}We have\begin{equation}\label{dm313}\int_{a}^{b}f_0(t,1)dt \leq   \int_{a}^{b}f (t,1)dt \leq \int_{a}^{b} f_0 (t, \varphi  (t))dt\end{equation}for every  $ \varphi \in L^1 (a,b)$  with\begin{equation}\label{dm314}   \int_{a}^{b} \varphi  (t) dt =b-a .\end{equation}In particular,\begin{equation}    \label{dm326}f (t,1)=f_0 (t,1) \;\; \mbox{for a.e.\ }t \in  [ a,b].\end{equation}%for almost all $t \in [a,b]$.\end{Lemma}\noindent{\bf Proof} --- \hspace{ 2 mm} The first inequality in (\ref{dm313})follows from the fact that $f_0(t,1)\le f(t,1)$ for every$t\in[a,b]$. To prove the second inequalitywe argue by contradiction. Assume that there exists $ \varphi  \in L^1(a,b)$, satisfying  (\ref{dm314}), such that\begin{equation} \label{dm315}   \int_{a}^{b} f_0 (t, \varphi  (t))dt < \int_{a}^{b} f (t,1)dt .\end{equation}As  $f_0 (t,v) = + \infty $ for  $v < \frac{1}{2}$, from  (\ref{dm308})and  (\ref{dm315}) we obtain that $ \varphi  (t) \geq \frac{1}{2}$ fora.e.\ $t \in  [ a,b]$. Changing, if needed, $ \varphi $ on a set of measurezero, we may assume that this inequality is satisfied for every  $t \in  [a,b]$. If we replace  $ \varphi  (t)$ by  $ \frac{1}{2}  \varphi  (t) +\frac{1}{2}$, we obtain a new function, still denoted by  $ \varphi (t)$,which continues to fulfill  (\ref{dm314}) and  (\ref{dm315}) (byconvexity), and, in addition, satisfies the improved inequality $ \varphi(t) \geq \frac{3}{4}$ for every  $t \in  [ a,b]$. As $f_0$is a Carath\'{e}odoryfunction on  $ [ a,b] \times ( \frac{1}{2}, + \infty )$, the function$t\mapsto f_0(t,\varphi(t))$ is measurable.Let us fix  $ \varepsilon >0$ such that\begin{equation}   \label{dm316}\int_{a}^{b} [ f_0 (t, \varphi  (t)) + \varepsilon ]dt \; <\;\int_{a}^{b} f (t,1) dt .\end{equation}For every  $t \in  [ a,b]$ we define $ A_{ \varepsilon } (t)$ as the set ofall triples  $ ( \lambda ,v_1,v_2) \in \R^3$ such that  $0 < \lambda <1$,$v_1 > \frac{1}{2}$, $v_2 > \frac{1}{2}$,$\lambda v_1+ (1- \lambda )v_2 =\varphi  (t)$, and\begin{displaymath}   \lambda f (t,v_1) + (1- \lambda )f (t,v_2) < f_0 (t, \varphi  (t)) +\varepsilon .\end{displaymath}By  (\ref{dm312}) this set is nonempty for every  $t \in  [ a,b]$.{}From the measurability properties of  $f$ and$t\mapsto f_0(t,\varphi(t))$ we deducethat the graph of the set-valued map $t \leadsto A_{ \varepsilon }(t)$ belongs to $ {\cal L} _1 \times {\cal B}_1\times {\calB}_1\times {\cal B}_1$. By the Aumann--von Neumann selectiontheorem (see, e.g., \cite[Theorem III.22]{cava}) there existLebesgue measurable functions $ \mu , \varphi _1, \varphi _2$ on$ [ a,b]$ such that $ ( \mu (t), \varphi _1 (t), \varphi _2 (t))\in A_{ \varepsilon } (t)$ for every  $t \in [ a,b]$. {}From thedefinition of  $A_{ \varepsilon } (t)$ and from  (\ref{dm316}) wededuce that\begin{eqnarray}   \label{dm317}&\displaystyle\vphantom{\int_{a}^{b}}\forall \; t \in [ a,b],\;\; \; 0 < \mu  (t) <1,\; \; \; \varphi _1 (t) >\frac12,\; \; \;\varphi _2 (t) >\frac12,\\   \label{dm318}&\displaystyle\vphantom{\int_{a}^{b}}\forall \; t \in [ a,b],\;\; \; \mu  (t) \varphi _1 (t)+ (1 - \mu(t)) \varphi _2 (t) = \varphi (t) ,\\   \label{dm319}&\displaystyle\int_{a}^{b} [ \mu  (t) f (t, \varphi _1 (t)) + (1- \mu(t)) f (t,\varphi _2 (t))] dt < \int_{a}^{b}f (t,1)dt .\end{eqnarray}For every  $k \geq 2$ let\begin{displaymath}   E^{ (k)}:= \left\{ t \in  [ a,b]\;|\; \mu (t) \in ( \frac{1}{k+1},\frac{1}{k}] \cup [ \frac{k-1}{k}, \frac{k}{k+1}) \right\} .\end{displaymath}As $0< \mu (t) <1$, the interval  $[ a,b]$ is the union of the sets$E^{ (k)}$, which are pairwise disjoint. As  $\varphi\in L^1(a,b)$and $f(\cdot,1)\in L^1(a,b)$ by (\ref{dm308}),from(\ref{dm318}) and (\ref{dm319}) we obtain that for every  $k \geq2$ the functions $ \varphi _i (t)$ and $f (t, \varphi _i (t))$belong to $L^{1} (E^{ (k)})$.By the Lyapunov theorem (see, e.g.,  \cite[Theorem 8.6.3 andProposition 8.6.2]{af90sva}) there exist two disjoint measurablesets $ E^{ (k)}_{1}$ and  $ E^{ (k)}_{2}$, with  $E^{ (k)}_{1}\cup E^{ (k)}_{2}=E^{ (k)}$, such that\begin{eqnarray}    \label{dm321}    &\displaystyle\sum_{i=1}^{2} \int_{E^{ (k)}_{i}}^{} \varphi _i (t) dt = \int_{E^{(k)}}^{} [  \mu  (t) \varphi _1 (t) + (1- \mu  (t)) \varphi _2 (t)] dt,\\    \label{dm322}    &\displaystyle\sum_{i=1}^{2} \int_{E^{ (k)}_{i}}^{} f(t, \varphi _i (t)) dt =\int_{E^{ (k)}}^{} [  \mu  (t) f(t,\varphi _1 (t)) + (1- \mu(t))f(t, \varphi _2 (t))] dt .\end{eqnarray}Let $\displaystyle   E_1:= \bigcup_{k=2}^{ \infty }E^{ (k)}_{1}$ and $   \displaystyle E_2:=   \bigcup_{k=2}^{\infty }E^{ (k)}_{2}$.By  (\ref{dm314}), (\ref{dm318}),   (\ref{dm319}), (\ref{dm321}),(\ref{dm322}) we obtain\begin{eqnarray}   \label{dm323}&  \displaystyle\int_{E_1}^{} \varphi _1 (t)dt + \int_{E_2}^{} \varphi _2 (t) dt=\int_{a}^{b} \varphi  (t)dt =b-a,\\    \label{dm324}  &    \displaystyle\int_{E_1}^{} f (t, \varphi _1 (t))dt + \int_{E_2}^{}f (t, \varphi_2 (t)) dt < \int_{a}^{b} f (t,1)dt .\end{eqnarray}Let $ \varphi_3 \in L^{1} (a,b)$ be the function defined by $ \varphi_3:= \varphi_1$ on  $E_1$ and  $ \varphi_3 := \varphi _2$ on  $E_2$, andlet  $ \psi $ be the primitive of  $ \varphi_3$ with  $ \psi  (a)=a$. By(\ref{dm323}) we have also  $ \psi  (b)=b$, while (\ref{dm324}) gives\begin{displaymath}   \int_{a}^{b}f (t, \psi ' (t))dt < \int_{a}^{b} f (t,1)dt,\end{displaymath}which contradicts   (\ref{dm310}) and concludes the proof of(\ref{dm313}).As  $f_0 \leq f$ and $f(\cdot,1)\in L^1(a,b)$ by (\ref{dm308}),if we take $ \varphi  \equiv 1$ in  (\ref{dm313}) weget~(\ref{dm326}).$\; \; \Box$\begin{Remark}{\rm As  $v\mapsto f_0 (t,v)$ is convex and finite,     for every $t \in  [a,b]$ there exist the limits\begin{eqnarray}\label{dm327}&\displaystyled^{\,l}_v f_0 (t,1):= \lim_{v \rightarrow 1-}  \frac{f_0 (t,v)-f_0(t,1)}{v-1} = \sup_{v<1} \frac{f_0 (t,v)-f_0 (t,1)}{v-1},\\\label{dm328}&\displaystyled^{\,r}_v f_0 (t,1):= \lim_{v \rightarrow 1+}  \frac{f_0 (t,v)-f_0(t,1)}{v-1} = \inf_{v>1} \frac{f_0 (t,v)-f_0 (t,1)}{v-1},\end{eqnarray}and we have $ - \infty < d^{\,l}_v f_0 (t,1) \leq d^{\,r}_v f_0 (t,1) < +\infty $.}\end{Remark}For the sake of completeness, we give now a new elementary proof of aparticular case of Theorem 3.1 of  \cite{ambr}.\begin{Lemma}\label{dm376}There exists a constant  $c \in \R$ such that\begin{equation}   \label{dm329}d^{\,l}_vf_0 (t,1) \leq c \leq d^{\,r}_v f_0 (t,1) \;\; \mbox{for a.e.}\;\; t \in [ a,b].\end{equation}\end{Lemma}\noindent{\bf Proof} --- \hspace{ 2 mm}We argue by contradiction. If (\ref{dm329}) does not hold, then thereexists $ \alpha \in \R$ such that\begin{displaymath}   \mathop{\rm ess \, sup }_{t \in  [ a,b]}d^{\,l}_v f_0 (t,1) > \alpha >\mathop{\rm ess \, inf } _{t \in[ a,b]}d^{\,r}_v f_0 (t,1) .\end{displaymath}Then the sets\begin{displaymath}   A_l := \left\{ t \in [ a,b] \;| \; d^{\,l}_v f_0 (t,1) > \alpha  \right\}\;\; \& \;\;  A_r :=\left\{ t \in [ a,b] \;| \; d^{\,r}_v f_0 (t,1) <\alpha  \right\}\end{displaymath}are disjoint and have positive measure. By  (\ref{dm327}) for every  $t \inA_l$ the set\begin{displaymath}   E_l (t):= \left\{ v \in (  \frac{1}{2},1 ) \;|\; \frac{f_0(t,v) -f_0 (t,1)}{v-1} > \alpha  \right\}\end{displaymath}is nonempty. Since the graph of the set-valued map $t \leadsto E_l(t)$ belongs to $ {\cal L}_1 \times {\cal B}_1$, by the Aumann--vonNeumann selection theorem (see, e.g.,  \cite[Theorem III.22]{cava})there exists a measurable function  $ \delta _l \colon  A_l \to (0,\frac{1}{2})$ such that\begin{equation}    \label{dm330}\forall \; t \in A_l,\;\; f_0 (t,1- \delta _l (t)) -f_0 (t,1) < -\alpha \delta _l (t).\end{equation}Similarly, using  (\ref{dm328}) we can prove that there exists ameasurable function  $ \delta _r \colon  A_r \to (0, \frac{1}{2})$such that\begin{equation}    \label{dm331}\forall \; t \in A_r,\;\; f_0 (t,1+ \delta _r (t))-f_0 (t,1) <\alpha \delta _r (t).\end{equation}Let us define\begin{equation}    \label{dm332}c_l:= \left[  \int_{A_l}^{} \delta _l (t) dt\right]^{-1}\;\;\&\;\;c_r:= \left[  \int_{A_r}^{} \delta _r (t) dt\right]^{-1},\end{equation}and let  $ \varphi  (t):=-c_l \delta _l (t)$ for $t \in A_l,\; \varphi(t)=c_r \delta _r (t)$ for $t \in A_r$, and  $ \varphi  (t)=0$ otherwise.Then by  (\ref{dm332}) we have  $ \int_{a}^{b} \varphi  (t) dt=0$, and,by   (\ref{dm313}), for every  $ \varepsilon >0$ this implies\begin{displaymath}   \int_{a}^{b} \left[  f_0 (t,1+ \varepsilon  \varphi  (t)) - f_0(t,1)\right] dt \geq 0 ,\end{displaymath}which is equivalent to\begin{equation}    \label{dm333}\int_{A_l}^{} \left[  f_0 (t,1- \varepsilon c_l \delta _l (t)) -f_0 (t,1)\right] dt + \int_{A_r}^{} \left[  f_0 (t,1+ \varepsilonc_r \delta _r (t)) - f_0 (t,1)\right] dt \geq 0 .\end{equation}By the monotonicity property of the difference quotient of a convexfunction, using   (\ref{dm330}) we obtain for  $ \varepsilon c_l <1$\begin{equation}   \label{dm334}f_0 (t,1- \varepsilon c_l \delta _l (t))-f_0 (t,1) \leq  \varepsilon c_l \left[  f_0 (t, 1- \delta _l (t))- f_0 (t,1)\right]  < - \alpha  \varepsilon c_l \delta _l (t)\end{equation}for every $t \in A_l$.Similarly,  for $ \varepsilon c_r < 1$ we obtain, using (\ref{dm331}),\begin{equation} \label{dm3e33}f_0 (t,1+ \varepsilon c_r \delta _r (t))-f_0(t,1) \leq \varepsilon c_r \left[ f_0 (t,1+  \delta _r (t))-f_0 (t,1)   \right]   < \alpha  \varepsilon c_r \delta _r (t)\end{equation}for every $t\in A_r$.{}From (\ref{dm332})--(\ref{dm3e33}) it follows that\begin{displaymath}   0 < - \alpha  \varepsilon c_l \int_{A_l}^{} \delta _l (t) dt + \alpha\varepsilon c_r \int_{A_r}^{} \delta _r (t) dt=0.\end{displaymath}This contradiction proves  (\ref{dm329}).$\; \; \Box$\vspace{5 mm}\noindent{\bf Proof of Theorem \ref{dm306}} --- \hspace{ 2 mm} By(\ref{dm328}) and (\ref{dm329}) there exists a constant $c\in \R$ such thatfor a.e.\ $t \in [ a,b]$ and every$ \varepsilon  \in (0,1)$ we have\begin{displaymath}   c \leq d^{\, r}_v f_0 (t,1) \leq \frac{f_0 (t, 2- \varepsilon ) -f_0(t,1)}{1- \varepsilon },\end{displaymath}hence$   (1- \varepsilon )c \leq f_0 (t,2- \varepsilon ) -f_0 (t,1)$,which implies\begin{equation}   \label{dm3e34}(1- \varepsilon )c + \varepsilon f_0 (t,1) \leq f_0 (t,2-\varepsilon ) - (1- \varepsilon )f_0 (t,1) .\end{equation}By convexity we have\begin{equation}   \label{dm3e35}f_0 (t,2- \varepsilon ) \leq  \varepsilon f_0 ( t, {1}/{\varepsilon } ) + (1- \varepsilon )f_0 (t,1),\end{equation}so that we obtain from  (\ref{dm3e34}) and  (\ref{dm3e35})\begin{equation}   \label{dm3e36}(1- \varepsilon ) c + \varepsilon f_0 (t,1) \leq \varepsilon f_0( t, {1}/{ \varepsilon } ) \leq  \varepsilon f (t,{1}/{ \varepsilon }) .\end{equation}By  (\ref{dm304}) and  (\ref{dm307}) for every  $v>0$ we have\begin{equation}   \label{dm3e37}f (t,v) \geq L  ( y (t), {y' (t)}/{v} )v \geq\Theta ( {y' (t)}/{v} )v \geq \overline{\rm co}\,\Theta ( {y' (t)}/{v} )v,\end{equation}where $\overline{\rm co}\, \Theta $ is the lower semicontinuous convexenvelope of  $\Theta$, which still satisfies (\ref{dm302}). Sincethe function $v \mapsto \overline{\rm co}\, \Theta ( {y' (t)}/{v})v$ is convex for $v>0$, from (\ref{dm3e37}) we deduce that\begin{equation}   \label{dm3e38}\forall \; v>0,\;\; f_0 (t,v) \geq \overline{\rm co}\, \Theta({y' (t)}/{v} )v .\end{equation}{}From  (\ref{dm307}),  (\ref{dm3e36}), and  (\ref{dm3e38}) weobtain\begin{equation}   \label{dm3e39}(1- \varepsilon )c + \varepsilon \, \overline{\rm co}\, \Theta (y' (t) ) \leq L (y (t), \varepsilon y' (t))\end{equation}for a.e.\ $t \in [ a,b]$ and every  $ \varepsilon  \in (0,1)$.Let us now fix  $t \in [ a,b]$ such that  (\ref{dm3e39}) holds and$|y' (t)| >2$. For $ \nu \in (0,1]$ let  $ \varepsilon  (t) =\frac{ \nu }{|y' (t)|} < \frac{1}{2}$. By  (\ref{dm3e39}) we have\begin{equation}   \label{dm3e40}\min \{c,0\} + \nu \frac{\overline{\rm co}\, \Theta ( y' (t) )}{|y'(t)|} \leq L \left( y (t), \nu \frac{y' (t)}{|y' (t)|} \right).\end{equation}Since $y \in W^{1,1} (a,b; \R^n)$, there exists $R>0$ such that\begin{equation}   \label{dm3e41}\forall \; t \in [ a,b],\;\;|y (t)| \leq R.\end{equation}Since $B(0,R)$ is compact, from  (\ref{dm301}) we know that\begin{equation}   \label{dm3e42}\exists \; M>0,\; \exists \; r \in (0,1],\;\; \forall \; u \in B(0,r),\;\; L (y(t),u) \leq M .\end{equation}Choosing $ \nu =r$, from  (\ref{dm3e40}),  (\ref{dm3e41}), and(\ref{dm3e42}) we get\begin{equation}   \label{dm3e43}\min \{c,0\} + r \frac{\overline{\rm co}\, \Theta \left( y' (t) \right)}{|y'(t)|} \leq M.\end{equation}Since $\overline{\rm co}\, \Theta$ satisfies   (\ref{dm302}), by(\ref{dm3e43}) there exists a constant $C =C(\Theta, c, r,M) \geq 2$,depending only on $\overline{\rm co}\, \Theta$,$c$, $r$,  and $M$, such that for a.e.\ $t \in [ a,b]$ with $|y' (t)| >2$we have\begin{equation}   \label{dm3e44}|y' (t)| \leq C.\end{equation}As $C \geq 2$, inequality (\ref{dm3e44}) holds also when  $|y'(t)| \leq 2$. $\; \; \Box$\vspace{5 mm}\noindent{\bf Proof of Theorem \ref{dm306a}} --- \hspace{ 2mm} By Lemma \ref{dm376} there exists $c$ such that\begin{equation}   \label{dm377}d^{\,l}_vf_0 (t,1) \leq c \leq d^{\,r}_v f_0 (t,1) \;\; \mbox{for a.e.}\;\; t \in [a,b] .\end{equation}By (\ref{dm304}) and(\ref{dm370a})  we have\begin{equation}\label{123}\int_{a}^{b} \Theta (y'(t)) dt \leq B ,\end{equation}which, by (\ref{dm302}), gives\begin{displaymath}\int_{a}^{b} |y'(t)| dt\le M_1,\end{displaymath}for a constant $M_1=M_1(\Theta, B, \beta)>0$. This inequality,together with (\ref{dm369a}), yields $|y (t)| \leq R $ for every$t\in[a,b]$,with $R=R(\Theta, A,B,\beta)=A+M_1$.We next provide an estimate of $c$ from below.{}From (\ref{123}) and (\ref{dm371}) it follows that\begin{equation}   \label{dm378}\alpha \; \mathop{\rm ess\,inf}_{t \in [a,b]} \Theta (y'(t)) \leq\int_{a}^{b} \Theta (y'(t)) dt \leq B .\end{equation}By(\ref{dm302}) there exists $M_2=M_2(\Theta, \alpha, B)$ such that for a set$\Omega_y \subset [a,b]$ of positive measure\begin{equation}   \label{dm380}\forall \; t \in \Omega_y,\;\; |y'(t)| \leq M_2 .\end{equation}This implies that\begin{equation}   \label{dm381}\forall \; t \in \Omega_y,\;\; L(y(t), \textstyle \frac{4}{3}y'(t))\leq \Psi (R+ 2 M_2) .\end{equation}Since by (\ref{dm327})\begin{displaymath}c \geq \sup_{v <1}  \frac{f_0(t,v)-f_0(t,1)}{v-1}\end{displaymath}and since $f_0(t,1) \geq 0$ for almost all $t\in [a,b]$, we get, setting $v= \frac{3}{4}$,\begin{equation}   \label{dm375}\textstylec \geq -4f(t,  \frac{3}{4})= -3L(y(t),\frac{4}{3} y'(t)) \geq -3\Psi (R+ 2 M_2) .\end{equation}We now return to the last part of the proof of Theorem \ref{dm306}with $\nu=r=1$ and $M=\Psi(R+1)$. As the constant $C$ whichappears in (\ref{dm3e44}) depends on $c$ in a decreasing way, and$c\ge -3\Psi (R+ 2 M_2)$, it is enough to set $K=C(\Theta, -3\Psi(R+ 2 M_2), 1, \Psi(R+1))$.$\;\;\Box$\section{DuBois-Reymond  Necessary Conditions}\label{DuB}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel function,let $y \in W^{1,1} (a,b; \R^n)$ be a function which satisfies(\ref{dm305}), and let$g \colon  [ a,b] \times \R \rightarrow [0, + \infty ]$ be the functiondefined by\begin{equation}   \label{dm3e45}g ( t,v):=  \left\{ \begin{array}{lll} L (y (t), vy' (t)) & \mbox{if\ \ }  0<v<2,\\ + \infty  & \mbox{  otherwise.} &\end{array} \right.\end{equation}By  (\ref{dm307}) we have $f (t,v)=g (t, \frac{1}{v})v$ for every$v \ne 0$. Let $g_0 := \overline {\rm co}_v g $ be the lowersemicontinuous convex envelope of  $g$ with respect to  $v$. Asthe functions $v \mapsto f_0 (t, v)$ and  $v \mapsto g_0 (t,\frac{1}{v})v$ are lower semicontinuous and convex for $v>0$, wededuce that\begin{displaymath}   \forall \;  t \in [ a,b], \; v>0,\;\; f_0 (t,v)=g_0 \left( t,{1}/{v} \right)v .\end{displaymath}Therefore we obtain from  (\ref{dm326})\begin{equation}   \label{dm3e46}g_0 (t,1)=g (t,1)=L (y (t), y' (t)) \;\; \mbox{ for a.e.}\;\; t \in [ a,b].\end{equation}Furthermore $t \mapsto g_0 (t ,v)$ is measurable. To prove this factit is enough to adapt the arguments used for $f_0$in the proof of Lemma \ref{dm309}.Let us define $d^{\, l}_vg_0 (t,1)$ and  $d^{\,r}_vg_0 (t,1)$ as in(\ref{dm327}) and (\ref{dm328}). It is easy to prove that\begin{displaymath}   d^{\,l}_v f_0 (t,1)=g_0 (t,1)- d^{\,r}_vg_0 (t,1)\;\;\&\;\;   d^{\,r}_v f_0 (t,1)=g_0 (t,1)- d^{\,l}_vg_0 (t,1).\end{displaymath}Therefore, by  (\ref{dm329}) there exists a constant $c \in \R$ such that\begin{equation}   \label{dm3e47}d^{\, l}_v g_0 (t,1)\le L (y (t), y' (t))-c\le d^{\,r}_v g_0 (t,1)\;\; \mbox{  for a.e.\ }t \in [ a,b].\end{equation}Notice that if  $u \mapsto L (y (t),u)$ is differentiable at $y' (t)$,with gradient $ \nabla _u L (y (t),y' (t))$, then by (\ref{dm3e45})the function $v \mapsto g (t,v)$ has a derivative at  $v=1$ which isequal to$ \left\langle    \nabla _u L (y (t),y' (t)),y'(t)\right\rangle $.By (\ref{dm3e46}) this implies\begin{equation}   \label{dm3e48}d^{\,l}_vg_0 (t,1)=d^{\, r}_v g_0 (t,1)= \left\langle   \nabla _uL (y(t),y' (t)),y' (t)\right\rangle,\end{equation}and from  (\ref{dm3e47}) and (\ref{dm3e48}) we obtain theDuBois-Reymond  necessary condition\begin{displaymath}L (y (t),y' (t))- \left\langle   \nabla _u L (y (t),y' (t)),y'(t)\right\rangle =c\;\; \mbox{for a.e.}\;\; t \in [ a,b] .\end{displaymath}Our aim is to derive similar results when  $L (y (t), \cdot )$ is notdifferentiable.All our extensions of theDuBois-Reymond  necessary condition(Theorems~\ref{dm3pr70}, \ref{dm363},and~\ref{dm363x})are based on the following lemma.\begin{Lemma}\label{dm363n}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel function,let  $y \in W^{1,1} (a,b; \R^n)$ be a function which satisfies(\ref{dm305}), and let$\psi\colon [a,b] \times \R^n \to\R$ be a Carath\'eodory function, with $\xi\mapsto \psi(t,\xi)$convex and positivelyhomogeneous of degree one, such that for a.e.\ $t \in [ a,b]$\begin{equation}\label{dm360'n}  -d^{\,l}_vg_0 (t,1)\le \psi(t, -y' (t)) \quad\&\quadd^{\,r}_vg_0 (t,1)\le \psi(t, y' (t)).  \end{equation}Then there exist a constant  $c \in \R$ and a measurable function$p \colon  [ a,b] \rightarrow \R^n$ such that for a.e.\ $t \in [ a,b]$\begin{eqnarray}   \label{dm362*}& p (t) \in \partial_\xi \psi(t,0) ,\\\label{dm3a60*}& L (y (t),y' (t))- \left\langle   p (t), y' (t)\right\rangle = c ,\end{eqnarray}where $\partial_\xi \psi(t,0)$ denotes the subdifferential of the convexfunction $\psi(t,\cdot)$ at~$0$.\end{Lemma}\noindent{\bf Proof} --- \hspace{ 2 mm} Since $\psi(t, \cdot)$ is convex andpositively homogeneous of degree one,\begin{equation}\label{dm3650}\forall \; \xi \in \R^n,\;\; \max_{q \in \partial _{\xi} \psi(t,0)}\langle q,\xi\rangle = \psi(t,\xi).\end{equation}  By  (\ref{dm3e47}) and (\ref{dm360'n})  there exists a constant  $c$ such that for a.e.\ $t \in [ a,b]$\begin{displaymath}  -\psi(t,-y'(y))\leq L (y (t),y' (t))-c \leq \psi(t,y'(t)),\end{displaymath}so that  (\ref{dm3650}) implies\begin{displaymath}  \min _{q \in \partial _{\xi} \psi(t,0) } \langle q,y'(t)\rangle  \leq L (y (t),y' (t)) - c  \leq  \max _{q\in \partial _{\xi} \psi(t,0) } \langle q,y'(t)\rangle .\end{displaymath}  Let us fix $t \in [ a,b]$ such that these inequalities aresatisfied. The set $\partial _{\xi} \psi(t,0)$ being convex, we deduce that forsome $q \in\partial _{\xi} \psi(t,0)$ we have $ \langle q,y'(t)\rangle= L (y(t),y' (t)) - c$.  For every $t \in [ a,b]$ let\begin{equation}   \label{dm368} B (t):= \left\{ q \in \partial_\xi \psi(t,0) \;|\;\left\langle   q,y' (t)\right\rangle  =L (y (t),y' (t)) -c \right\}.\end{equation}By the previous argument  $B (t) \ne \emptyset $ for  a.e.\ $t \in[ a,b]$. The graph of the set-valued map $t \leadsto B (t)$ is theintersection of the sets  $B_1$ and $B_2$ defined by\begin{eqnarray*}& B_1:= \left\{  (t,q) \in [ a,b] \times \R^n\;|\; \left\langle   q,y'(t)\right\rangle =L (y (t),y' (t))-c  \right\} ,\\  & B_2 :=  \left\{  (t,q) \in [ a,b] \times \R^n\;|\;  \forall \; \xi \in \R^n,\;\; \left\langle   q, \xi \right\rangle\leq \psi (t, \xi )  \right\}.\end{eqnarray*}Clearly  $B_1$ and   $B_2 $ belong to $ {\cal L}_1 \times {\calB}_n$, where ${\cal B}_n$ denotes the $\sigma$-algebraof all Borel subsets of $\R^n$. This implies that the graph of theset-valued map $t\leadsto B (t)$ defined by (\ref{dm368}) belongs to  ${\cal L}_1\times {\cal B}_n$, and by the Aumann-von Neumann selectiontheorem (see \cite[Theorem III.22]{cava}), there exists ameasurable function $p \colon  [ a,b] \to \R^n$ such that  $p (t)\in B (t)$ for a.e.\ $t \in [ a,b]$. Then (\ref{dm362*}) and(\ref{dm3a60*}) follow from (\ref{dm368}).  $\; \; \Box$\vspace{5 mm}Let $L_0:=\overline {\rm co}_u\, L$ be the lower semicontinuousconvex envelopeof  $L$ with respect to  $u$. Then $ (t,u) \mapsto L_0 (y (t),u)$ is aCarath\'eodory function. This can be verified as in the case of $f_0$(proof of Lemma  \ref{dm309}),taking convex combinations of $n+1$ vectors.     \begin{Theorem}\label{dm3pr70}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel functionandlet  $y \in W^{1,1} (a,b; \R^n)$ be a function which satisfies(\ref{dm305}).Suppose that for a.e.\ $t \in [ a,b]$\begin{eqnarray}   \label{ex1}& L (y (t),y' (t))=L_0 (y (t),y' (t)),\\   \label{ex2}   &   -d^{ \,l}_v g_0 (t,1) \le d_u L_0 (y (t),y' (t)) (-y' (t)),\\   \label{ex3}   &    d^{ \,r}_v g_0 (t,1) \le d_u L_0 (y (t),y' (t)) (y' (t)) ,\end{eqnarray}where $d_u L_0(x,u)(\xi)$ denotes thedirectional derivative of the convex function $L_0(x,\cdot)$ at $u$ inthe direction $\xi$.Then there exist a constant $c \in \R$ and a measurable function $p\colon [a,b] \rightarrow \R^n$ such that a.e.\ $t \in [ a,b]$\begin{eqnarray}   \label{dm3e550}& p (t) \in \partial_u L_0 (y (t),y' (t)),\\   \label{dm3e560}& L_0 (y (t),y' (t))- \left\langle   p (t),y' (t)\right\rangle =c.\end{eqnarray}Consequently,\begin{equation}\label{constH0}\langle p(t),y'(t)\rangle - L(y(t),y'(t)) =\sup_{u\in\scriptR^n} (\langle p(t),u\rangle -L(y(t),u)) = -c\end{equation}for a.e.\ $t \in [ a,b]$.\end{Theorem}\begin{Remark}{\rm Since $g_0(t,v)\ge L_0(y(t),vy'(t))$ for a.e.\$t\in [a,b]$ and for every $v\in\R$,if (\ref{ex1}) holds, then by (\ref{dm3e46})\begin{displaymath}  -d^{ \,l}_v g_0 (t,1) \ge d_u L_0 (y (t),y' (t)) (-y' (t))\quad\&\quad    d^{ \,r}_v g_0 (t,1) \ge d_u L_0 (y (t),y' (t)) (y' (t)) ,\end{displaymath}so that  (\ref{ex2}) and  (\ref{ex3}) are actually equalities.Assumptions (\ref{ex1})--(\ref{ex3}) are satisfied, for instance,if $g_0(t,v)= L_0 (y (t),v y' (t))$for a.e.\ $t \in [ a,b]$ and every  $v\in\R$.}\end{Remark}\begin{Remark}{\rm If $H$ is the Hamiltonian associated to $L$,defined in (\ref{H}), then,  by (\ref{constH0}),$H(y(t),p(t))=-c$ for almost all $t\in [a,b]$.The function $p$ corresponds to the co-state of optimalcontrol theory. In other words, we proved that the Hamiltonianis constant along the optimal trajectory/co-state pair $(y,p)$. In thecase of smooth Hamiltonians this isindeed a well known property ofoptimal trajectories of autonomousBolza control problems.}\end{Remark}\noindent{\bf Proof of Theorem \ref{dm3pr70}} --- \hspace{ 2 mm}The result follows from Lemma~\ref{dm363n}, taking$\psi(t,\xi):=d_u L_0 (y (t),y' (t))(\xi)$. Indeed, the convexityof $ L_0 (y (t),\cdot)$ implies that  $\psi(t,\cdot)$ is convex and$\partial_\xi\psi(t,0)=\partial_u L_0 (y (t),y' (t))$for a.e.\ $t \in [ a,b]$.Equality (\ref{constH0}) follows from (\ref{ex1}), (\ref{dm3e550}),and (\ref{dm3e560}).   $\; \; \Box$   \vspace{ 5mm}To state further extensions of the DuBois-Reymond necessary condition, weneed to recall several notions of generalized derivatives.Let $\varphi \colon \R^m\to \R \cup \{+ \infty \}$. The subdifferentialof $\varphi$ at $x\in \mbox{\rm dom}(\varphi)$ is defined by\begin{equation}\label{subd3}\partial^- \varphi(x):=\left\{p\in \R^n\;|\; \liminf_{y\to x}\frac{ \varphi(y)- \varphi(x)-\langlep,y-x\rangle}{|y-x|} \geq 0\right\} .\end{equation}An equivalent definition of subdifferential uses the lowercontingent derivatives of $\varphi$ defined by\begin{equation}\label{dirder}\forall \; u \in \R^m,\;\; D_{\uparrow } \varphi (x) (u) := \liminf_{\textstyle {h \rightarrow 0+ \atop v \rightarrow u}} \frac{\varphi(x+hv) - \varphi (x)}{h} .\end{equation}Then,\begin{equation}   \label{subd1}   p \in \partial^- \varphi (x)  \;\; \Longleftrightarrow \;\; \forall \; v\in\R^n, \;\; \left\langle   p,v\right\rangle \leq D_{\uparrow } \varphi(x) (v) .\end{equation}(see, e.g., \cite{af90sva}). The upper contingentderivative of $\varphi$ at $x$ isdefined by\begin{displaymath}\forall \; u \in \R^m,\;\; D _{\downarrow }  \varphi (x) (u) := \limsup_{\textstyle {h \rightarrow 0+ \atop v \rightarrow u}} \frac{\varphi(x+hv) - \varphi (x)}{h} .\end{displaymath}The superdifferential $\partial^+ \varphi (x)$ of $\varphi$ at $x$ is definedby $ \partial^+ \varphi (x):=- \partial^- (-\varphi) (x)$ or, equivalently, by\begin{equation}   \label{supd1}   p \in \partial^+ \varphi (x)  \;\; \Longleftrightarrow \;\; \forall \; v\in\R^n, \;\; \left\langle   p,v\right\rangle \geq D_{\downarrow } \varphi(x) (v) .\end{equation}We use also the lower Dini directional derivative, defined by\begin{equation}    \label{dm3e49}\forall \; u \in \R^m,\;\; d^- \varphi(x) ( u ):=\liminf_{h \rightarrow 0+} \frac{\varphi (x+hu  ) -\varphi (x)}{h}.\end{equation}\vspace{ 5mm}Let us return to the Lagrangian $L$ considered at the beginning ofthis section. Partial derivatives and partial differentials of $L$with respect to $u$are defined inthe usual way: given $ x \in \R^n $,we consider the function $\varphi (\cdot):=L(x,\cdot)$, and set $ D _{\uparrow u} L (x,u) := D _{\uparrow } \varphi (u) $,$\partial _u^- L (x,u):=\partial^- \varphi(u)$,$ D _{\downarrow u} L (x,u) := D _{\downarrow } \varphi (u)$, $\partial _u^+ L (x,u):= \partial^+ \varphi (u)$, and$d^-_u L(x,u):= d^-\varphi(u)$.\begin{Remark}\label{rem90}{\rmBy  (\ref{dm3e45}) and(\ref{dm3e46}) for a.e.\ $t \in [ a,b]$ and every $v \in (0,2)$ wehave\begin{displaymath}   g_0 (t,v) \leq g (t,v) =L (y (t),vy' (t))\quad\&\quad g_0 (t,1)=g (t,1)=L (y(t),y' (t)),\end{displaymath}which implies\begin{equation}        \label{dm364} \ \ \ \-d^{\,l}_v g_0 (t,1) \leq  d_{u}^- L ( y (t),y' (t)) (-y' (t)) \quad\&\quad    d^{\,r}_vg_0 (t,1) \leq d_{u}^-L (y (t),y' (t)) (y' (t)).\end{equation}Therefore the conclusions of Lemma~\ref{dm363n} continue to hold if(\ref{dm360'n}) is replaced by\begin{equation}        \label{dm360z}        d_{u}^-L (y (t),y' (t)) (\pm y' (t))\le \psi(t, \pm y' (t))\end{equation}for a.e.\ $t \in [ a,b]$.}\end{Remark}\vspace{ 5mm}For every $(x,u)\in\R^n\times\R^n$ let$\xi\!\mapsto \overline {\rm co} \, D _{\uparrow u} L(x,u)(\xi)$ bethe lower semicontinuous convex envelope of the function\$\xi\mapsto D _{\uparrow u} L(x,u)(\xi)$.\begin{Theorem}\label{dm363}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel function andlet  $y \in W^{1,1} (a,b; \R^n)$ be a function which satisfies(\ref{dm305}).Suppose that for a.e.\ $t \in [ a,b]$\begin{eqnarray}\label{dm360'}&   -d^{ \,l}_v g_0 (t,1) \le \overline {\rm co} \,D_{\uparrow u} L (y (t),y'(t))(-y' (t)),\\   \label{dm360'b}   &    d^{ \,r}_v g_0 (t,1) \le \overline {\rm co} \,D_{\uparrow u} L (y (t),y'(t))(y' (t)),\\\label{dm360''}& \forall \;\xi \in \R^n,\;\;\; \overline {\rm co}\, D _{\uparrow u}L (y (t),y'(t))(\xi)\in\R.\end{eqnarray}Then there exist a constant  $c \in \R$ and a measurable function$p \colon  [ a,b] \rightarrow \R^n$ such that\begin{eqnarray}   \label{dm362}& p (t) \in \partial _{u} ^-L (y (t),y' (t)) \;\; \mbox{ fora.e.\ } t \in [ a,b],\\\label{dm3a60}&   L (y (t),y' (t))- \left\langle   p (t), y' (t)\right\rangle = c \;\;\mbox{ for a.e.\ } t \in [ a,b].\end{eqnarray}\end{Theorem}\begin{Remark}\label{rem105}{\rm By (\ref{dm364}) inequalities(\ref{dm360'}) and (\ref{dm360'b}) are satisfied if\begin{displaymath}     d_{u}^-L (y (t),y' (t)) (\pm y' (t))\le \overline {\rm co} \,D_{\uparrow u} L (y (t),y'(t))(\pm y' (t)),\end{displaymath}for a.e.\ $t \in [ a,b]$.This shows that (\ref{dm360'})--(\ref{dm360''})are always satisfied if $L(y(t),\cdot)$ isdifferentiable at $y'(t)$.}\end{Remark}\begin{Remark}\label{rem106}{\rmWe recall thata function $\varphi \colon  \R^n \to  \R$ is called {\emsemiconvex\/} if there exists $\omega \colon \R_{+} \times {\bfR}_{+} \to  \R_{+}$, satisfying          \begin{displaymath}\label{hp002}          \forall \;  r \leq R,  \; \forall \;  s \leq S, \;  \; \omega (r,s)\leq \omega (R,S) \;   \; \&   \; \;  \lim_{s\rightarrow 0+}\omega(R,s) \;  =  \;  0,          \end{displaymath}such that for every $R>0$,  $\lambda \in [0,1]$, and all          $ x, y \in B(0,R)$          \begin{displaymath}          \varphi (\lambda x + (1-\lambda )y) \leq \lambda \varphi (x)  +(1-\lambda )\varphi (y)  +  \lambda (1-\lambda )|x-y| \; \omega(R,|x-y|) .          \end{displaymath}Observe that every convex function  is semiconvex (with $\omega $equal to zero). Furthermore, if   $\varphi \colon \R^{n}\to {\bfR}$ is continuously differentiable, then it is semiconvex.Using standard arguments of convex analysis (see, e.g.,\cite[p.~25]{aubconvex})one can prove that every semiconvex function $\varphi$ islocally Lipschitz. Furthermore, for every $v  \in \R^n$          \begin{displaymath}          D _{\uparrow } \varphi (x) (v)  = \lim_{ h\rightarrow 0+}         \frac{\varphi (x+hv ) -\varphi (x)}{h},          \end{displaymath}          and the function $v \mapsto  D _{\uparrow }\varphi (x) (v)$ is convex (see, e.g., \cite[Theorem 3.9]{cf91cha}).These facts, together with Remark~\ref{rem105}, show thatassumptions (\ref{dm360'})--(\ref{dm360''}) of Theorem \ref{dm363}are always satisfiedwhen $L(x,\cdot)$ is semiconvex.}\end{Remark}  \noindent{\bf Proof of Theorem \ref{dm363}} --- \hspace{ 2 mm} Let us define$\psi(t,\xi):= \overline {\rm co} \,D_{\uparrow u} L (y (t),y'(t))(\xi)$. Using the projection theoremit is possible to check the measurability with respect to~$t$.Notice that $\partial_\xi\psi(t,0)\subset \partial_u^- L (y (t),y'(t))$for a.e.\ $t\in  [ a,b]$. Indeed, if$q\in \partial_\xi\psi(t,0)$, then $\langle q,\xi\rangle\le\psi(t,\xi)\le  D_{\uparrow u} L (y (t),y'(t))(\xi)$ for every$\xi\in\R^n$, hence$q\in \partial^-_u L (y (t),y'(t))$by  (\ref{subd1}).The conclusion follows then fromLemma~\ref{dm363n}.$\;\; \Box$\begin{Remark}{\rmTheorem~\ref{dm3pr70} has stronger assumptions andstronger conclusions than Theorem~\ref{dm363}.Indeed, as  $L \geq L_0$, it follows from(\ref{ex1}) that\begin{equation}\label{eq700}     d_u L _0(y (t),y'(t))(\xi)=D_{\uparrow u} L _0(y (t),y'(t))(\xi)\leD_{\uparrow u} L (y (t),y'(t))(\xi)\end{equation}for every $\xi\in\R^n$. Since$\xi\mapsto d_u L _0(y (t),y'(t))(\xi)$ is convex, we conclude that\begin{displaymath}d_u L _0(y (t),y'(t))(\pm y'(t))\le  \overline {\rm co} \,D_{\uparrow u} L (y (t),y'(t))(\pm y' (t)).\end{displaymath}This shows that(\ref{ex1})--(\ref{ex3}) imply (\ref{dm360'})--(\ref{dm360''}).On the other hand, (\ref{subd1}) and (\ref{eq700})yield $ \partial ^- _u L_0 (y(t),y' (t)) \subset \partial _u ^- L (y (t),y' (t))$ for a.e.\ $t\in  [ a,b]$. Therefore  (\ref{ex1}), (\ref{dm3e550}), and(\ref{dm3e560}) imply (\ref{dm362}) and (\ref{dm3a60}).}\end{Remark}\begin{Theorem}\label{dm363x}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel function andlet  $y \in W^{1,1} (a,b; \R^n)$ be a function which satisfies(\ref{dm305}).Suppose that $L(x,\cdot)$ is locally Lipschitz continuous for every$x\in\R^n$.Then there exist a constant  $c \in \R$ and a measurable function$p \colon  [ a,b] \rightarrow \R^n$ such thatfor a.e.\ $t \in [ a,b]$\begin{eqnarray}   \label{dm362**}& p (t) \in \partial_u L (y (t),y' (t)),\\\label{dm3a60**}& L (y (t),y' (t))- \left\langle   p (t), y' (t)\right\rangle = c,\end{eqnarray}where $\partial_u L(x,u)$ denotes the Clarke generalized gradient of$L(x,\cdot)$ at~$u$.\end{Theorem}\noindent{\bf Proof} --- \hspace{ 2 mm}Let us define\begin{displaymath}     \psi (t,\xi)= \limsup _{\textstyle {h \rightarrow 0+ \atop u \rightarrow y'(t)}} \frac{L(y (t), u+h\xi) - L (y (t),u)}{h} .\end{displaymath}It is known that $\psi (t,\cdot)$ is convex and that  $\partial_\xi\psi (t,0)$ is theClarke generalized gradient of $L(y(t),\cdot)$ at $y'(t)$  (see \cite{Cla}).Since $d^-_{u} L (y (t), y' (t)) (\xi)\le \psi(t,\xi)$, the resultfollows fromLemma~\ref{dm363n} and Remark~\ref{rem90}.$\;\;\Box$.   \vspace{5 mm}   Replacing subdifferential by superdifferential we get another   extension of the DuBois-Reymond  necessary condition, which is   meaningful only at those points $t \in [ a,b]$ for which   $\partial _u^+ L (y (t), y' (t))\neq\emptyset$.\begin{Proposition}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel functionandlet  $y \in W^{1,1} (a,b; \R^n)$ be a function which satisfies(\ref{dm305}).There exists a constant  $c \in \R$ such that\begin{displaymath}  \forall\;   p \in \partial _u^+ L (y (t), y' (t)),\;\;\;L (y (t),y' (t))- \left\langle   p,y' (t)\right\rangle =c\end{displaymath}for a.e.\ $t \in [ a,b]$.\end{Proposition}\noindent{\bf Proof} --- \hspace{ 2 mm} {}From (\ref{dm3e46}) we have\begin{eqnarray*}& d^{\,r}_v g_0 (t,1) \leq D _{\downarrow u} L (y (t), y' (t)) (y' (t)),\\&  d^{\,l}_v g_0 (t,1) \geq -D _{\downarrow u} L (y (t), y' (t)) (-y'   (t)).\end{eqnarray*}These inequalities and(\ref{dm3e47}) imply thatthere exists a constant  $c$ such that for a.e.\ $t \in [a,b]$\begin{displaymath}-D _{\downarrow u} L (y (t), y' (t)) (-y'   (t)) \leq L (y (t),y' (t)) -c \leqD _{\downarrow u} L (y (t), y' (t)) (y' (t)),\end{displaymath}andwe deduce from (\ref{supd1}) that for all $p \in \partial _u^+ L(y (t), y' (t))$\begin{displaymath}   \left\langle -p, -y' (t)  \right\rangle \leq L (y (t),y' (t)) -c  \leq\left\langle   p, y' (t)\right\rangle,\end{displaymath} ending the proof.$\; \; \Box$\section{Hamilton-Jacobi Inequalities}\label{Hamilton}Let  $ \varphi \colon \R^n \to \R_+ \cup \{ + \infty \}$ with$\varphi \not\equiv+ \infty $.Given $T>0$ and $y_0 \in \R^n$, let us consider theBolza problem:\begin{displaymath}\min_y \int_{0}^{T} L (y(s), y' (s))ds + \varphi  (y(T))\end{displaymath}over all absolutely continuous functions $y \in W^{1,1} (0,T;\R^n)$ satisfying the initial condition $y (0)= y_0$. The dynamicprogramming approach associates with this problem the family ofproblems ($t\geq 0$, $x\in\R^n$):\begin{displaymath}\min_y \int_{0}^{t} L (y (s), y' (s))ds + \varphi  (y(t))\end{displaymath}over all absolutely continuous functions $y \in W^{1,1} (0,t;\R^n)$ satisfying $y (0)= x$. The corresponding value function$V\colon  \R_+\times \R^n \to  \R_+ \cup \left\{ + \infty  \right\}$ is definedby (\ref{value}).\begin{Proposition}\label{prop15}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel functionand let $ \varphi \colon \R^n \to \R_+ \cup \{ + \infty \}$ with$\varphi \not\equiv+\infty$.Then $V(0,x)=\varphi(x)$ for every $x\in\R^n$.Furthermore, if $L$ satisfies   (\ref{dm302}) and (\ref{dm304}), and$\varphi$ is lower semicontinuous, then\begin{equation}\label{eq115}\liminf_{\textstyle {t \rightarrow 0+ \atop x \rightarrow x_0}} V(t,x)\ge\varphi(x_0)\end{equation}for every $x_0\in\R^n$.\end{Proposition}\noindent{\bf Proof} --- \hspace{ 2 mm} Fix $x _0\in \R^n$ andlet $t_i \to 0+$, $x_i \to x _0$ be such that\begin{displaymath}\liminf_{\textstyle {t \rightarrow 0+ \atop x \rightarrow x_0}} V(t,x)= \lim_{i \to \infty} V(t_i,x_i).\end{displaymath}If the above limit is infinite, then our claim follows. If this limitis finite,then we consider $y_i \in W^{1,1}(0,t_i;\R^n)$ suchthat $y_i(0)=x_i$ and\begin{displaymath}\int_{0}^{t_i} L (y_i (s), y'_i(s))ds +\varphi  (y_i(t_i)) \leq V(t_i,x_i) + \frac{1}{i}.\end{displaymath}By (\ref{dm304}), since $\varphi \geq 0$,for some $M>0$ we have $\int_{0}^{t_i} \Theta( y'_i(s))ds \leq M$ forevery~$i$. Setting $y_i'(s)=0$ for  $s \in (t_i,1]$, we deduce from (\ref{dm302}) that the functions $y_i'$ are  equiintegrable and therefore the functions $y_i$ are equicontinuous.  Since $t_i \to 0+$ and $y_i(0)=x_i\to x_0$, we get $y_i(t_i)\to x_0$.On the other hand, since $L \geq 0$,  we have$\varphi  (y_i(t_i)) \leq V(t_i,x_i) + \frac{1}{i}$.Taking the lower limit and usingthe lower semicontinuity of $\varphi$ we conclude the proof.$\;\;\Box$\vspace{5 mm}We recall that $\R_+ ^{\star}:=\{t\in\R \;|\; t>0\}$.In this section we often assume the following hypotheses:\vspace{ 5 mm}\noindent({\bf H1}) for every $(t,x) \in \R_+^\star \times \R^n$ the infimum in(\ref{value}) is attained,\medskip\noindent({\bf H2}) $L$ is  locally bounded and satisfies(\ref{dm302}) and~(\ref{dm304}).\vspace{ 5 mm}\noindentIt is easy to see that ({\bf H2}) implies that$0 \leq V(t,x) < \infty$ for all $(t,x) \in \R_+^\star \times \R^n$.\begin{Remark}\label{rem10}{\rm If for every $t>0$ the functional\begin{equation}\label{eq50}{\cal L}_0^t(y) :=\int_{0}^{t} L(y(s),y'(s))ds\end{equation}is  sequentially weakly lowersemicontinuous  on $W^{1,1}(0,t;\R^n)$ and $ \varphi$ is lowersemicontinuous, then from (\ref{dm302}) and (\ref{dm304})it follows that  ({\bf H1})is satisfied. Furthermore, arguing as in \cite[Proof ofProposition 3.1]{dm-fra}, we can show that in this case $V$ is lowersemicontinuous on $\R_+ \times \R^n$.}\end{Remark}\begin{Lemma}\label{minlip}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel functionand let $ \varphi \colon \R^n \to \R_+ \cup \{ + \infty \}$ with$\varphi \not\equiv+\infty$. Assume that $({\bf H1})$ and $({\bfH2})$ are satisfied.Then, given $ (t_0,x_0) \in \R_+^{ \star } \times \R^n$and $0<\delta<t_0$,there exists $r>0$ such that for all $ (t,x) \in B ( (t_0,x_0),\delta )$  everyminimizer $ y ( \cdot; t, x)$ of  (\ref{value})  is $r$-Lipschitz.\end{Lemma}\noindent{\bf Proof} --- \hspace{ 2 mm} Consider $y_0 \in \R^n$, with$\varphi(y_0)<+\infty$,and set $z(s)= x+ \frac{s}{t}(y_0-x)$. If $y(\cdot; t,x)$ is aminimizer of (\ref{value}), we obtain\begin{displaymath}\int _{0}^{t}L(y(s;t,x),y'(s;t,x))ds \leq \varphi (y_0) + t \sup_{s \in [0,t]}L (z(s),(y_0-x)/t ).\end{displaymath}Since $L$ is locally bounded, for every$0<\delta<t_0$there exists a constant $M_\delta>0$ such that\begin{displaymath}\sup_{s \in [0,t]} L (z(s),(y_0-x)/t )\le M_\delta\end{displaymath}for every   $ (t,x) \in B ( (t_0,x_0), \delta )$. The conclusion followsnow from Theorem~\ref{dm306a}.   $\; \; \Box$\begin{Theorem}\label{liploc}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel functionand let $ \varphi \colon \R^n \to \R_+ \cup \{ + \infty \}$ with$\varphi \not\equiv+\infty$. Assume that $({\bf H1})$ and $({\bfH2})$ are satisfied.  Then $V $ is locally Lipschitz on$\R_+^{ \star } \times \R^n$. If $\varphi$ is lower semicontinuouson $\R^n$, then $V $ is lower semicontinuous  on${\R_+\times \R^n}$.\end{Theorem}\noindent{\bf Proof} --- \hspace{ 2 mm}The Lipschitzcontinuity is proved in \cite[Corollary 3.4]{dm-fra}.For the reader's convenience we repeat here the proof.  Fix  $(t_0,x_0) \in \R_+^{ \star } \times \R^n$.By Lemma\ref{minlip}, there exist $r>0$ and $\delta >0$ such that for all $ (t,x)\inB ( (t_0,x_0), \delta )$  every minimizer $ y ( \cdot; t, x)$ of(\ref{value}) is $r$-Lipschitz. We may assume that $5\delta<t_0$.Let $(t_1, x_1)$ and $(t_2, x_2)$ be two distinct points of$B ( (t_0,x_0), \delta )$, let $h_1:=|t_1-t_2| + |x_1-x_2|$, and$s_1:= h_1-t_1+t_2$. Let $u_1\in\R^n$ besuch that $y(s_1;t_2,x_2)= x_1+h_1u_1$. Then $0< h_1<t_1$,$0\le s_1 \le 2h_1$, and\begin{equation}\label{dm301*}|u_1| \leq \frac{|y(s_1;t_2,x_2)-x_2|}{h_1} +\frac{|x_2-x_1|}{h_1} \leq 2r + 1.\end{equation}Let $y_1\colon [0,t_1]\mapsto\R^n$ be the function defined by\begin{displaymath}  y_1 (s)=\cases{x_1+su_1&if $0\le s\le h_1$,  \cr  \noalign{\vskip 3pt}  y(s-t_1+t_2; t_2,x_2)&if $h_1\le s \le t_1$.  \cr}\end{displaymath}Then\begin{eqnarray*}& \displaystyleV(t_1,x_1)\leq \int_0^{t_1} L(y_1(s), y_1^\prime(s))ds +\varphi(y_1(t_1))=\\& \displaystyle=\int_0^{h_1} L(x_1+s u_1, u_1)ds+\int_{s_{1}}^{t_{2}} L(y(s; t_2,x_2), y'(s; t_2,x_2))ds+\varphi(y(t_{2}; t_2,x_2)).\end{eqnarray*}As $s_{1}=h_{1}-t_{1}+t_{2}\geq 0$ and $L\geq 0$, we obtain\begin{displaymath}V(t_1,x_1) \leq \int_0^{h_1} L(x_1+s u_1, u_1)ds +V(t_2,x_2).\end{displaymath}Since $L$ is locally bounded, it follows from (\ref{dm301*}) thatthere exists a constant $M$, depending only on $L$, $t_0$, $x_0$,$\delta$, and $r$, such that\begin{displaymath}V(t_1,x_1)-V(t_2,x_2) \leq Mh_1=M(|t_1-t_2| + |x_1-x_2|).\end{displaymath}Exchanging the roles of $(t_1,x_1)$ and $(t_2,x_2)$ we get that$V$ in $M$-Lipschitz on $B ((t_0,x_0), \delta )$.If $\varphi$ is lower semicontinuous,  then $V$is lower semicontinuous at all points of $\{0\}\times\R^n$ byProposition~\ref{prop15}. The lower semicontinuity on$\R_+^\star\times\R^n$ is a consequence of the localLipschitz continuity.$\; \; \Box$\vspace{ 5 mm}We recall the definition of the function $L ^+(x,u)$ used in\cite{dm-fra} to study Hamilton-Jacobi equation (\ref{HJ}) arisingfrom adiscontinuous Lagrangian:\begin{equation}\label{feb30}L^+(x,u) := \limsup_{h \rightarrow 0+} \frac{1}{h} \inf_y \left\{\int_{-h}^{0} L (y (s),y' (s))ds  \;|  \; y (-h)=x-hu,\; y (0)=x\right\}.   \end{equation}Similarly we define the function $L^-(x,u)$ by\begin{equation}\label{dm384}L^-(x,u) := \liminf_{h \rightarrow 0+} \frac{1}{h} \inf_y \left\{\int_{0}^{h} L (y (s),y' (s))ds   \;|  \; y (h)=x+hu,\; y (0)=x\right\} .   \end{equation}\begin{Proposition} \label{eq---}Assume that $L$ is locally bounded. If $u_h\to u$ as $h\to0+$, then\begin{eqnarray}\label{eq20}&\displaystyleL^+ (x,u) =\limsup_{h \rightarrow 0+}\frac{1}{h} \inf _y\left\{ \int_{-h}^{0} L (y ( s ),y' ( s )) d s\;|\; y (-h)=x-hu_h  ,\;  y (0) = x\right\},\\\label{eq21}&\displaystyleL^- (x,u) =\liminf_{h \rightarrow 0+}\frac{1}{h} \inf _y\left\{ \int_{0}^{h} L (y ( s ),y' ( s )) d s\;|\; y (h)=x+hu_h  ,\;  y (0) = x\right\}.\end{eqnarray}\end{Proposition}\noindent{\bf Proof} --- \hspace{ 2 mm}The following proof is a slight modification of the proof of\cite[Proposition~3.6]{dm-fra}.Let us fix $ (x,u)$ and $u_h$ as in the statement of theproposition, and let $ \overline L (x,u)$ be the right hand sideof  (\ref{eq20}).We want to show that $\overline L(x,u)\leq L^+(x,u)$.For every $h>0$ let $\varepsilon_{h}= |u_h-u|$and let $y_{h}$ be such that $y_h(-(1-\varepsilon_{h})h)=x-(1-\varepsilon_{h})h u $,  $y_h (0)=x$, and\begin{eqnarray*}   &\displaystyle   \int_{-(1-\varepsilon_{h})h}^{0} L(y_h ( s ),y'_h ( s )) d s   -h^2\le   \\     &\displaystyle  \le \inf_y \left\{ \int_{-(1-\varepsilon_{h})h}^{0}  L(y ( s ),y' ( s )) d s  :y(-(1-\varepsilon_{h})h)=x-(1-\varepsilon_{h})h u  ,\;  y (0)=x\right\}.\end{eqnarray*}We extend  $y_h$ on the interval$ [-h,-(1-\varepsilon_{h})h]$ by theaffine function satisfying $y_h (-h)= x-h u_h$ and$y_h(-(1-\varepsilon_{h})h)=x-(1-\varepsilon_{h})h u$.Since on this interval the derivative of $y_h$ is equal to$ (u_h- (1-\varepsilon_{h}) u)/\varepsilon_{h}$, which is uniformlybounded,we deduce that for some $M>0$ and all $h>0$,\begin{displaymath}  \int_{-h}^{ 0} L (y_h (s),y_h' (s))ds \leq\int_{-(1-\varepsilon_{h}) h}^{0}L (y_h(s),y_h' (s))ds + \int_{-h}^{ -(1-\varepsilon_{h}) h} M ds .\end{displaymath}Dividing by $h$ and taking the upper limitwhen $h \rightarrow 0+$we get $\overline L (x,u) \leq L^+ (x,u)$. The opposite inequality canbe proved in the same way.The proof of (\ref{eq21}) is similar.$\; \; \Box$  \begin{Remark} \label{rem1}{\rm{}From the previous proposition it follows that, if $h_i\to 0+$,$u_i\to u$, and$y_i\in W^{1,1}(0,h_i)$ satisfies $y_i(0)=x$ and $y(h_i)=x+h_iu_i$, then\begin{equation}\label{eq41}L^-(x,u) \le \liminf_{i \rightarrow \infty}\frac{1}{h_i} \int_{0}^{h_i} L (y_i ( s ),y'_i ( s )) d s.\end{equation}}\end{Remark}\begin{Proposition} \label{prop2}Assume that $L$ is locally bounded.For every $y\in W^{1,1}(0,t;\R^n)$we have $L^+(y(s),y'(s)) \le L(y(s),y'(s))$and $L^-(y(s),y'(s)) \le L(y(s),y'(s))$for a.e.\ $s\in [0,t]$. If $y$ is a minimizer of (\ref{value}), then$L^+(y(s),y'(s))=L^-(y(s),y'(s))=L(y(s),y'(s))$ for a.e.\ $s\in [0,t]$.\end{Proposition}\noindent{\bf Proof} --- \hspace{ 2 mm}Assume first $y\in W^{1,\infty}(0,t;\R^n)$.Since $L$ is locally bounded, the function $s \mapsto\psi (s):= \int_{0}^{s} L (y (\tau),y' (\tau))d\tau$ is absolutelycontinuous. Let$s\in [0,t]$ be such that both $ \psi ' (s)$ and $y' (s)$ do exist and$\psi ' (s) = L ( y (s),y' (s)) $. Since$u_{h}=(y(s)-y(s-h))/h$ converges to $y'(s)$ as $h\to 0+$,from Proposition~\ref{eq---} we obtain\begin{displaymath}L^+(y(s),y' (s)) \leq \lim_{h\to 0+} \frac{1}{h} \int_{s-h}^{s} L ( y(\tau),y' (\tau))d\tau = \psi ' (s) = L ( y (s),y'(s)) ,\end{displaymath}which concludes the proof of the inequality$L^+(y(s),y'(s)) \le L(y(s),y'(s))$ when $y$ is Lipschitz.If $y\in W^{1,1}(0,t;\R^n)$, we can apply a Lusintype approximation theorem for Sobolev functions (see, e.g.,\cite[Theorem~3.10.5]{Zie}), which asserts that for every$\varepsilon>0$ there exist $y_\varepsilon\in W^{1,\infty}(0,t;\R^n)$and an open set $U_\varepsilon$ such that$|U_\varepsilon|<\varepsilon$ and $y_\varepsilon(s)=y(s)$ for all$s\in [0,t]\setminus U_\varepsilon$. As $y'_\varepsilon(s)=y'(s)$ fora.e.\  $s\in [0,t]\setminus U_\varepsilon$, we obtain that$L^+(y(s),y'(s)) \le L(y(s),y'(s))$and $L^-(y(s),y'(s)) \le L(y(s),y'(s))$for a.e.\ $s\in [0,t]\setminus U_\varepsilon$. Since $\varepsilon$ isarbitrary, these inequalities hold for a.e.\ $s\in [0,t]$.If $y$ is a minimizer of (\ref{value}), then for every $s\in (0,t)$and every $h\in (0,s)$ we have\begin{equation}\label{eq400}\ \ \ \ \ \ \ \ \ \ \int_{s-h}^{s} L (y ( \tau ),y' ( \tau )) d \tau =\inf _z\left\{ \int_{-h}^{0} L (z (\tau),z' (\tau)) d\tau\;|\; z (-h)\!=\!y(s-h)  ,\;  z (0)\!= \!y(s)\right\}.\end{equation}Let us fix a Lebesgue point $s\in (0,t)$for the function $L(y(s),y'(s))$such that $(y(s)-y(s-h))/h\to y'(s)$ as $h\to 0+$. If we divide  both sides of(\ref{eq400}) by~$h$, the left hand side tends to $L(y(s),y'(s))$while theright hand side tends to $L^+(y(s),y'(s))$ thanks toProposition~\ref{eq---}, applied with $x:=y(s)$, $u:=y'(s)$, and$u_h:=(y(s)-y(s-h))/h$. Therefore$L^+(y(s),y'(s))=L(y(s),y'(s))$ for a.e.\ $s\in [0,t]$.The proof for $L^-(y(s),y'(s))$ issimilar.$\; \; \Box$\vspace{5 mm}   Let us define\begin{eqnarray}\label{eq200}&\displaystyle H ^+(x,p):= \sup_{u \in \scriptR^n} \left(\left\langle   p,u\right\rangle   -L^+(x,u) \right),\\\label{eq201}&\displaystyle   H ^-(x,p):= \sup_{u \in \scriptR^n} \left( \left\langle   p,u\right\rangle   -L^-(x,u) \right) .\end{eqnarray}\begin{Theorem}\label{thm2}     Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel functionand let $ \varphi \colon \R^n \to \R_+ \cup \{ + \infty \}$ with$\varphi \not\equiv+\infty$. Assume that $({\bf H1})$ and $({\bfH2})$ are satisfied.Then the value function $V$ satisfies the following twoinequalities:\begin{eqnarray} \label{dm360}&\forall \; (t,x) \in \R_+^{ \star } \times \R^n ,\;\; \exists \; u   \in \R^n,\;\;  D _{\uparrow}V (t,x) (-1,u)  \leq -L^- (x,u),\\\label{dm360c}   &\forall \; (t,x) \in \R_+ \times \R^n , \;   \forall \; u \in \R^n,\;\; D _{\downarrow}V (t,x) (1,-u) \leq L^+   (x,u).\end{eqnarray}   Consequently, $V$ is a supersolution to (\ref{HJ}) on$\R_+^{ \star } \times \R^n $  when $H$ is replaced by $H^-$,i.e.,\begin{equation} \label{supersol}   \forall \; (t,x) \in \R_+^{ \star } \times \R^n ,\;\; \forall \; (p_t,p_x)\in\partial^-V (t,x),\;\;p_t+ H^-(x,-p_x) \geq 0,\end{equation}   and $V$ is a subsolution to (\ref{HJ}) on$\R_+ \times \R^n $  when $H$ is replaced by $H^+$, i.e.,\begin{equation} \label{dm360d}   \forall \; (t,x) \in \R_+ \times \R^n , \;   \forall \; (p_t, p_x) \in \partial^- V (t,x),\;\; p_t + H^+ (x,-p_x)  \leq   0 .\end{equation}\end{Theorem}\noindent{\bf Proof} --- \hspace{ 2 mm}   Let  $t>0, \; x \in \R^n$, and let $y$ be aminimizer of  (\ref{value}).  By Theorem \ref{dm306} $y ( \cdot)$ is Lipschitz. By minimality for all $0<h \leq t$ we have\begin{displaymath}   V (t,x)= V (t-h, y (h))+ \int_{0}^{h}L (y (s), y' (s))ds .\end{displaymath}Consider $h_i \rightarrow 0+$ such  that for some $u \in \R^n$,$u_i:=(y (h_i)-x)/h_i\rightarrow u$. Then $y(h_i)=x+h_iu_i$, and(\ref{eq41}) yields\begin{displaymath}   D _{\uparrow }V (t,x) (-1,u) \leq -\limsup_{i \rightarrow  \infty }\frac{1}{h_i} \int_{0}^{h_i} L (y (s), y' (s))ds \leq -L^-(x,u) ,\end{displaymath}which proves (\ref{dm360}).Let  $ (p_t,p_x) \in \partial^- V (t,x)$. Then, by (\ref   {subd1}),\begin{displaymath}- p_t + \left\langle   p_x, u\right\rangle \leq    D _{\uparrow }V (t,x) (-1,u),\end{displaymath}hence  $p_t + \left\langle   -p_x, u\right\rangle -L^- (x,u) \geq0$. By (\ref{eq201}) this inequality gives (\ref{supersol}).To prove inequality (\ref{dm360c}), we fix any $u \in \R^n$ andlet $h_i \rightarrow 0+$, $u_i \rightarrow u$.{}From the definition of $V$ it follows that\begin{eqnarray*}&V(t+h_i,x-h_iu_i)-V (t,x) \leq\\&\displaystyle\le  \inf_y \left\{ \int _{-h_i}^{0}L(y(s),y'(s))ds \;|\;  y(-h_i)=x-h_iu_i, y(0)=x \right\}.\end{eqnarray*}Then we divide by $h_i$ and passto the upper limit as $i\to\infty$. Taking (\ref{eq20}) into account we obtain(\ref{dm360c}). To prove (\ref{dm360d}) it is enough to apply(\ref{subd1}), (\ref{dm360c}), and~(\ref{eq200}). $\; \; \Box$\begin{Theorem}\label{thm1}Let $L\colon \R^n \times \R^n \rightarrow \R_+$ be a Borel function,let $ \varphi \colon \R^n \to \R_+ \cup \{ + \infty \}$be a lower semicontinuousfunction with$\varphi \not\equiv+\infty$,  andlet $W\colon  \R_+\times \R^n \to \R \cup \{ + \infty \}$ be a lower semicontinuousfunction which satisfies the initial condition $W (0, \cdot )=\varphi$. Assume that $({\bf H1})$ and $({\bfH2})$ are satisfied. If $W$ is a subsolution of the Hamilton-Jacobi equation(\ref{HJ}), in the sense that\begin{equation} \label{feb45}   \forall \; (t,x) \in \mbox{\rm dom}(W), \;   \forall \; u \in \R^n,\;\; D _{\downarrow }W (t,x) (1,-u) \leq L^+ (x,u) ,\end{equation}then $W\le V$ on $\R_+ \times \R^n$.Therefore the value function $V$ is the greatestlower semicontinuous function $W$ which satisfiesinequality~(\ref{feb45}) and the initial condition$W(0,\cdot)=\varphi$.\end{Theorem}We shall use the following well known lemma (see, e.g.,\cite[Chapter~5, Section~2, Exercise~3]{Roy}). For completeness wegive here an elementary proof. To simplify notation, in the case offunctions $f$ of one real variable the lower Dini derivative $d^-f(t)(1)$,defined by (\ref{dm3e49}), is denoted by $d^-f(t)$.\begin{Lemma}\label{lemma1}Let $[a,b]$ be a bounded closed interval in $\R$ and let$f\colon [a,b]\to\R\cup\{+\infty\}$ be a lower semicontinuousfunction such that $d^-f(t)\le 0$ for every $t\in [a,b)$ with$f(t)<+\infty$.Then $f(b)\le f(a)$.\end{Lemma}\noindent {\bf Proof} --- \hspace{ 2 mm} For every $\varepsilon>0$let us consider the lower semicontinuous function$f_\varepsilon(t):= f(t)-\varepsilon t$, which  satisfies thestronger inequality $d^-f_\varepsilon(t)\le -\varepsilon$ forevery $t\in [a,b)$ with $f_\varepsilon(t)<+\infty$. We claim that$f_\varepsilon(b)\le f_\varepsilon(a)$. If this inequality is notsatisfied for some $\varepsilon>0$, then the infimum of$f_\varepsilon$ is attained at some $t_\varepsilon\in [a,b)$ andwe have $f_\varepsilon(t_\varepsilon)<+\infty$. Since$d^-f_\varepsilon(t_\varepsilon)\le -\varepsilon$, there exists$s_\varepsilon\in (t_\varepsilon, b)$ such that$f_\varepsilon(s_\varepsilon)< f_\varepsilon(t_\varepsilon)$,contradicting the minimality of $t_\varepsilon$ and proving ourclaim. Taking the limit in the inequality $f_\varepsilon(b)\lef_\varepsilon(a)$ as $\varepsilon\to0+$ we conclude the proof. $\;\; \Box$\vspace{5 mm}An alternative proof of the following corollary can be found in\cite[Chapter~6, Exercise~10]{Let}.\begin{Corollary}\label{cor1}Let $[a,b]$ be a bounded closed interval in $\R$ and let$f\colon [a,b]\to\R\cup\{+\infty\}$ be a lower semicontinuousfunction with $f(a)<+\infty$.Suppose that there exists a constant $M\in \R$ such that$d^-f(t)\le M$for every $t\in [a,b)$ with $f(t)<+\infty$. Then$f(t)<+\infty$ for every $t\in [a,b]$.Suppose, in addition, that for some $g\in L^1(a,b)$we have$d^-f(t)\le  g(t)$for a.e.\ $t\in [a,b]$.Then\begin{equation}\label{eq5}f(b)-f(a) \le\int_a^b g(t)dt.\end{equation}\end{Corollary}\noindent{\bf Proof} --- \hspace{ 2 mm} By Lemma~\ref{lemma1} the function$f_M(t):=f(t)-Mt$ is nonincreasing. Therefore $f(t)\lef(a)+M(t-a)<+\infty$ for every$t\in[a,b]$. If $d^-f(t)\le  g(t)$for a.e.\ $t\in [a,b]$, then $f'_M(t)\le g(t)-M$ fora.e.\ $t\in [a,b]$, and the Lebesgue theorem on derivatives ofmonotone functions yields\begin{displaymath}f_M(b)-f_M(a)\le \int_a^b f'_M(t)dt \le \int_a^b (g(t)-M)dt,\end{displaymath}which implies~(\ref{eq5}).$\; \; \Box$\vspace{5 mm}\noindent{\bf Proof of Theorem~\ref{thm1}} --- \hspace{ 2 mm} Let us fix$t>0$, $x\in \R^n$, and let $y$ bea minimizer of (\ref{value}).It is Lipschitz continuous by Theorem~\ref{dm306}. Let us define$\gamma(s):=W(s,y(t-s))$. Then $\gamma$ is lower semicontinuous on$[0,t]$ and $\gamma(0)=\varphi(y(t))<+\infty$.Let us fix $s\in [0,t)$ with $\gamma(s)<+\infty$.Consider a sequence $h_i\to0+$such that\begin{equation}\label{eq6}d^-\gamma(s)=\lim_{i\to \infty} \frac{\gamma(s+h_i)-\gamma(s)}{h_i}.\end{equation}We can write\begin{equation}\label{eq7}\gamma(s+h_i)-\gamma(s)= W(s+h_i, y(t-s)-h_i u_i)-W(s, y(t-s)),\end{equation}with $u_i:=-(y(t-s-h_i)-y(t-s))/h_i$. Passing to a subsequence we mayassume that $u_i$ converges in $\R^n$ to some vector $u$, whose normis bounded by theLipschitz constant of $y$. {}From (\ref{feb45}), (\ref{eq6}), and (\ref{eq7})it follows that\begin{equation}\label{eq8}d^-\gamma(s)\leL^+(y(t-s),u).\end{equation}Since the function $L^+$ is locally bounded, we conclude that thereexists a constant $M$ such that$d^-\gamma(s)\le M$for every $s\in [0,t)$ with $\gamma(s)<+\infty$.By Corollary~\ref{cor1} this implies that $\gamma(s)<+\infty$for every $s\in [0,t]$.If the derivative $y'(t-s)$ exists, then $u=y'(t-s)$. Therefore (\ref{eq8})gives$d^-\gamma(s)\leL^+(y(t-s),y'(t-s))$ for a.e.\ $s\in[0,t]$,which, together with Proposition~\ref{prop2}, yields$d^-\gamma(s)\leL(y(t-s),y'(t-s))$ for a.e.\ $s\in[0,t]$.By Corollary~\ref{cor1}we obtain\begin{displaymath}\gamma(t)\le \gamma(0) + \int_0^tL(y(t-s),y'(t-s))ds,\end{displaymath}which is equivalent to\begin{displaymath}W(t,x)\le\varphi(y(t)) + \int_0^t L(y(s),y'(s))ds.\end{displaymath}Since the righthand side is equal to $V(t,x)$, we have proved that $W(t,x)\le V(t,x)$.The last assertion of the theorem follows nowfrom Theorems~\ref{liploc} and~\ref{thm2}.$\; \; \Box$\vspace{5 mm}We conclude this section with some results which connect theminimizers of (\ref{value}) with the contingent derivatives of thevalue function.\begin{Theorem}\label{dm391}     Let $L\colon \R^n \times \R^n\rightarrow \R_+$ be a Borel function andlet $ \varphi \colon \R^n \to \R_+ \cup \{ + \infty \}$ with$\varphi \not\equiv+\infty$. Assume that $({\bf H1})$ and $({\bfH2})$ are satisfied.If $y$ is a minimizer of (\ref{value}),then\begin{eqnarray} \label{dm360x}& D _{\uparrow}V (t-s,y(s)) (-1,y'(s)) =D _{\downarrow}V (t-s,y(s)) (-1,y'(s))  = -L(y(s),y'(s)),\\\label{dm360y}   & D _{\uparrow}V (t-s,y(s)) (1,-y'(s))=   D _{\downarrow}V (t-s,y(s)) (1,-y'(s)) = L(y(s),y'(s))\end{eqnarray}for almost all $s \in [0,t]$.\end{Theorem}\noindent  {\bf Proof} --- \hspace{ 2 mm}Since $V$ is locallyLipschitz on $\R_+^{ \star } \times \R^n$ by Theorem~\ref{liploc},the function $\gamma(s):=V(t-s, y (s))$ is locally absolutelycontinuous on $[0,t)$.Fix $s \in (0,t)$ suchthat the derivatives $\gamma'(s)$ and $y'(s)$ exist.Let us prove that\begin{eqnarray}\label{eq502}& D _{\uparrow}V (t-s,y(s)) (-1,y'(s))=D _{\downarrow}V (t-s,y(s)) (-1,y'(s))=\gamma'(s),\\\label{eq503}& D _{\uparrow}V (t-s,y(s)) (1,-y'(s))=  D _{\downarrow}V (t-s,y(s)) (1,-y'(s))=-\gamma'(s).\end{eqnarray}Let $h_i\to0+$ and $u_i\to y'(s)$ such that\begin{equation}\label{eq520}\ \ \ \ \  D _{\downarrow}V (t-s,y(s)) (-1,y'(s))=\lim_{i\to\infty} \frac{V(t-s-h_i, y(s)+h_iu_i)-V(t-s, y(s))}{h_i} .\end{equation}As $V$ is Lipschitz near $(t-s, y(s))$, there exists a constant$M>0$ such that\begin{eqnarray}\label{eq521}& |V(t-s-h_i, y(s+h_i))- V(t-s-h_i, y(s)+h_iu_i)|\le\\\nonumber&\le M|y(s+h_i)-y(s)-h_iy'(s)|  + Mh_i|u_i-y'(s)|\end{eqnarray}for $i$ large enough. {}From (\ref{eq520}) and (\ref{eq521})we obtain the second equality in (\ref{eq502}). The other equalitiesin (\ref{eq502}) and (\ref{eq503}) can be obtained in the same way.On the other hand, by minimality, for all small $h>0$we have\begin{equation}\label{eq500} \displaystyle V(t-s-h, y (s+h))=V(t-s,y(s)) -\int_{s}^{s+h}L(y(\tau),y'(\tau))d\tau .\end{equation}Since $y$ is Lipschitz by Theorem~\ref{dm306}, the function $s\mapstoL(y(s),y'(s))$ is bounded. By the Lebesgue theorem (\ref{eq500})% and (\ref{eq501})implies that $\gamma'(s)=-L(y(s),y'(s))$ for a.e.\ $s\in [0,t]$.The conclusion follows now from (\ref{eq502}) and (\ref{eq503}).  $\; \; \Box$\begin{Theorem}\label{dm391b} Let $L\colon \R^n \times \R^n\rightarrow \R_+$ be a Borel functionandlet $ \varphi \colon \R^n \to \R_+ \cup \{ + \infty \}$be a lower semicontinuous function with$\varphi \not\equiv +\infty$. Assume that $({\bf H1})$ and $({\bfH2})$ are satisfied, and let $V$ be thevalue function given by (\ref{value}).Define the set-valued maps\begin{eqnarray*}& F(t,x):=\{u\in\R^n \;| \;  D _{\uparrow}V(t,x) (-1,u)  \leq -L (x,u)\},\\& G(t,x):=\{u\in\R^n \;| \;  D _{\downarrow}V(t,x) (1,-u)  \geq L (x,u)\}.\end{eqnarray*}Given $t>0$, $x\in\R^n$, and$y\in W^{1,1}(0,t;\R^n)$ with $y(0)=x$, the following conditions areequivalent:\begin{itemize}\item[(a)] $y$ is aminimizer of (\ref{value}),\item[(b)] $y' (s) \inF(t-s,y(s))$ for a.e. $s \in [0,t]$,\item[(c)] $y' (s) \inG(t-s,y(s))$ for a.e. $s \in [0,t]$.\end{itemize}\end{Theorem}\noindent{\bf Proof}  --- \hspace{ 2 mm}If $y$is a minimizer if (\ref{value}), then $y'(s) \in F(t-s,y(s))$ and$y'(s) \in G(t-s,y(s))$ fora.e.\ $s \in (0,t)$ by (\ref{dm360x}) and (\ref{dm360y}).If $y$ solves the differential inclusion $y' (s) \inF(t-s,y(s))$ for a.e. $s \in [0,t]$, wedefine$\gamma (s) :=V(t-s,y(s))$. Since $V$ is locally Lipschitz on$\R^\star_+\times\R^n$, the function $\gamma$ is locallyabsolutely continuous on $[0,t)$. Using (\ref{eq502})and the definition of $F(t-s,y(s))$, we obtain that$L(y(s),y'(s)) \leq -\gamma '(s)$ for almost all $s \in [0,t]$,which implies that $L (y(s),y'(s))$ is integrable on $[0,t-\varepsilon]$ for every $\varepsilon>0$. By integrating we obtain\begin{displaymath}\gamma (t- \varepsilon ) +\int_{0}^{t- \varepsilon}L (y(s),y'(s))ds \leq \gamma(0).\end{displaymath}As $\gamma (t- \varepsilon )=V(\varepsilon,y(t- \varepsilon))$ and$\gamma(0)=V(t,x)$,taking the lower limit as $\varepsilon\to 0$ and using Proposition~\ref{prop15}we get\begin{displaymath}\varphi (y(t))+ \int _{0}^{t} L (y(s),y'(s))ds \leqV(t,x).\end{displaymath}Consequently, $y$ is a minimizer of (\ref{value}).If $y$ solves the differential inclusion $y' (s) \inG({t-s},y(s))$ for a.e. $s \in [0,t]$,  we repeat the same proof,replacing (\ref{eq502}) with (\ref{eq503}).$\;\;\Box$\vspace{ 5 mm}\noindent {\bf Acknowledgements.} The work of Gianni Dal Maso ispart of the European Research Training Network ``Homogenizationand Multiple Scales'' under contract HPRN-2000-00109, and of theProject ``Calculus of Variations'', supported by SISSA and by theItalian Ministry of Education, University, and Research. GianniDal Maso wishes to thank the hospitality and support of theUniversity of Paris-Dauphine for two visits in May 2000 andJanuary 2001, during which the main results of this paper havebeen obtained.The work of H\'el\`ene Frankowska is part of theEuropean Research Training Network ``Evolution Equations forDeterministic and Stochastic Control Systems'', RTN-2002-12.H\'el\`ene Frankowska wishes to thank SISSA for the hospitality in thesummer 2002.\small\begin{thebibliography}{99}\bibitem {amar} AMAR M., BELLETTINI G. \& VENTURINI S. (1998) {\itIntegral representation of functionals defined on curves of\/}$W^{1,p}$, Proc. Roy. Soc. Edinburgh Sect. A {\bf 128}, 193-217.\bibitem {ambr} AMBROSIO L.,  ASCENZI O. \&  BUTTAZZO G. (1989) {\itLipschitz regularity for minimizers of integral functionals with highlydiscontinuous integrands\/}, J. Math. Anal. Appl. {\bf 142}, 301-316.\bibitem{aubconvex} AUBIN J.-P. (1993){\sc Optima and Equilibria},Grad. Texts in Math. 140, Springer-Verlag, Berlin.\bibitem{af90sva} AUBIN J.-P. \& FRANKOWSKA H. (1990) {\sc Set-ValuedAnalysis}, Birk\-h\"{a}u\-ser, Boston.\bibitem{bm85} BALL J. \& MIZEL V.J. (1990) {\it One dimensionalvariational problems whose minimizers do not satisfy theEuler-Lagrange equation}, Arch. Rational. Mech. Anal. {\bf 90} (1985),325-388.\bibitem{cf91cha}  CANNARSA P. \& FRANKOWSKA H. (1991) {\it Somecharacterizations of optimal trajectories in control theory}, SIAMJ. on Control and Optimization, 29, 1322-1347.\bibitem{cava}   CASTAING C. \& VALADIER M.   (1977){\sc Convex Analysisand Measurable Mutifunctions}, Springer-Verlag, Berlin.\bibitem{ces83}   CESARI   L.   (1983)   {\sc  Optimization   Theory   andApplications. Problems with Ordinary Differential Equations},Appl. Math. 17, Springer-Verlag, Berlin.\bibitem{Cla} CLARKE F.H. (1983){\sc  Optimization   and Nonsmooth Analysis}, Wiley-Inter\-science, NewYork.\bibitem{cla-vin} CLARKE F.H. \& VINTER R.B. (1985){\it Regularity properties of solutions to the basic problem in thecalculus of variations\/}, Trans. Amer.Math. Soc.{\bf 289}, 73-98.\bibitem {dm-fra} DAL MASO G. \& FRANKOWSKA (2000) {\itValue functions for Bolza problems with discontinuous Lagrangiansand Hamilton-Jacobi inequalities\/}, ESAIM Control Optim. Calc.Var. {\bf 5},  369-394.\bibitem {DG-But-DM} DE GIORGI E., BUTTAZZO G. \& DAL MASO G. (1983){\it On the lower semicontinuity of certain integral functionals\/},Atti Accad. Naz. Lincei Rend. Cl. Sci. Fis. Mat. Natur. (8) {\bf 74},274-282.\bibitem{Ioffe} IOFFE A.D. (1977){\it On lower semicontinuity of integral functionals\/},SIAM J. Control Optim.  {\bf 15}, 521-521 and 991-1000.\bibitem {Let} LETTA G. (1976){\sc  Teoria Elementare dell'Integrazione},Boringhieri, Torino.\bibitem{Olech} OLECH C. (1976){\it Weak lower semicontinuity of integral functionals\/}, J.Optim. Theory Appl. {\bf 19}, 3-16.\bibitem {Roy} ROYDEN H.L. (1969){\sc  Real Analysis},Collier Macmillan, New York.\bibitem{t1} TONELLI L. (1915) {\it Sur une m\'eth\^ode directedu calcul des variations}, Rend. Circ. Mat. Palermo {\bf 39}, 223-264.\bibitem{t2} TONELLI L. (1921) {\sc Fondamenti di Calcolo delle Variazioni},Vol. 1, 2, Zanichelli, Bologna.\bibitem{Zie} ZIEMER W.P. (1989) {\sc Weakly Differentiable Functions},Springer-Verlag, Berlin.\end{thebibliography}\normalsize\end{document}